Stochastic Calculus for Fractional Brownian Motion and Related Processes

Lecture Notes in Mathematics Editors: J.-M. Morel, Cachan F. Takens, Groningen B. Teissier, Paris 1929 Yuliya S. Mis...

Author: Yuliya Mishura

53 downloads 1277 Views 3MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form

DOWNLOAD PDF

Lecture Notes in Mathematics Editors: J.-M. Morel, Cachan F. Takens, Groningen B. Teissier, Paris

1929

Yuliya S. Mishura

Stochastic Calculus for Fractional Brownian Motion and Related Processes

ABC

Yuliya S. Mishura Department of Mechanics and Mathematics Kyiv National Taras Shevchenko University 64 Volodymyrska 01033 Kyiv Ukraine [email protected]

ISBN 978-3-540-75872-3

e-ISBN 978-3-540-75873-0

DOI 10.1007/978-3-540-75873-0 Lecture Notes in Mathematics ISSN print edition: 0075-8434 ISSN electronic edition: 1617-9692 Library of Congress Control Number: 2007939114 Mathematics Subject Classification (2000): 60G15, 60G44, 60G60, 60H05, 60H07, 60H10, 60H40, 91B24, 91B28 c 2008 Springer-Verlag Berlin Heidelberg This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Cover design: design & production GmbH, Heidelberg Printed on acid-free paper 987654321 springer.com

Preface

For several decades the semimartingale processes were the best model in order to implement many ideas. The stochastic calculus for semimartingales and the general theory of stochastic processes, which are closely connected to the theory of stochastic integration and stochastic differential equations, were originated by N. Wiener (Wie23), P. Lévy (Le48), K. Itˆ o (Itˆ o42), (Itˆ o44), (Itˆ o51), A.N. Kolmogorov (Kol31), W. Feller (Fel36), J.L. Doob, M. Loéve, I. Gikhman and A. Skorohod (the list of related papers and books is very long and we do not mention it here in full). Those ideas were developed further by several authors, among them there are K. Bichteler (Bi81), C.S. Chou, P.A. Meyer and C. Stricker (CMS80), K.L. Chung and R.J. Williams (ChW83), C. Dellacherie (Del72), C. Dellacherie and P.A. Meyer (DM82), C. DoléansDade and P.A. Meyer (DDM70), H. F¨ ollmer (Fol81a), P.A. Meyer (Me76) and M. Yor (Yor76). These theoretical data were fruitfully discussed and summarized in the monographs of J. Jacod (Jac79), R. Elliott (Ell82), P.E. Kopp (Kop84), M. Métivier and J. Pellaumail (MP80), B. Øksendal (Oks03), P. Protter (Pro90). Limit theorems in the most general semimartingale framework were proved by J. Jacod and A.N. Shiryaev (JS87). A very convenient way to consider financial markets is to insert them into semimartingale models, as perfectly demonstrated by I. Karatzas and S. Shreve (KS98), A.N. Shiryaev (Shi99), F. Delbaen and W. Schachermayer (DS06). The Malliavin calculus for the Wiener process was presented in the books of P. Malliavin (Mal97) and D. Nualart (Nua95). However, in recent years the well-studied theory of semimartingales turns out to be insufficient in order to describe many phenomena. On one hand, telecommunication connections, asset prices and other objects have “long memory”. This effect cannot be described with the help of such processes as the Wiener process, which has independent increments and has no memory. On the other hand, the concept of turbulence in hydrodynamics can be described by self-similar fields with stationary (dependent) increments (A.M. Yaglom (Yag57), A. Monin and A.M. Yaglom (MY67) and A.M. Yaglom (Yag87)).

VI

Preface

A.N. Kolmogorov (Kol40) was the first to consider continuous Gaussian processes with stationary increments and with the self-similarity property; it means that for any a > 0 there exists b > 0 such that Law(X(at); t ≥ 0) = Law(bX(t); t ≥ 0). It turns out that such processes with zero mean have a special correlation function: 1 2H |s| + |t|2H − |t − s|2H , EX(t)X(s) = 2 where 0 < H < 1. A.N. Kolmogorov called such Gaussian processes “Wiener Spirals” (“Wiener screw-lines”). Later, when the papers of H.E. Hurst (Hur51) and H.E. Hurst, R.P. Black and Y.M. Simaika (HBS65), devoted to long-term storage capacity in reservoirs, were published, the parameter H got the name “Hurst parameter”. The stochastic calculus of the processes mentioned above originated with the pioneering work of B.B. Mandelbrot and J.W. van Ness (MvN68) who considered the integral moving average representation of X via the Wiener process on an infinite interval and called this process fractional Brownian motion (fBm). Note that B.B. Mandelbrot worked with fractional processes during a long period and his later results concerned the fractals and scaling were summarized in the book (Man97). Note also that it was proved in the paper (GK05) that the moving average representation of fBm is unique in the class of the right-continuous, nondecreasing concave functions on R+ . The first result where fBm appeared as the limit in the Skorohod topology of stationary sums of random variables was obtained by M. Taqqu (Taq75); another scheme of convergence to fBm in the uniform topology was considered in (Gor77). Spectral properties of fBm were studied by G. Molchan (Mol69), G. Molchan and J. Golosov (MG69), G. Molchan (Mol03), and later by K. Dzhaparidze and H. van Zanten (DvZ05), (RLT95), (SL95). The next intensive wave of interest in fBm arose in the 1990s. It can be explained by various applications of fBm and other long-memory processes in teletraffic, finances, climate and weather derivatives. The paper (DU95) was one of the first paper devoted to stochastic analysis for fBm. Note that fBm is neither a semimartingale (except the case H = 1/2 when it is a Brownian motion) nor a Markov process. However, it is closely connected with fractional calculus and can be represented as a “fractional integral” (with the help of a comparatively complicated hypergeometric kernel) via the Wiener process not only on infinite, but also on finite intervals. This was stated by I. Norros, E. Valkeila and J. Virtamo (NVV99) and C. Bender (Ben03a). Such a representation, together with the Gaussian property of fBm and the H¨ older property of its trajectories (fBm with Hurst index H is Hölder up to order H) permits us to create an interesting and specific stochastic calculus for fBm. The development of the theory of long-memory processes moved in several directions: stochastic integration, stochastic differential equations, optimal filtering, financial applications, statistical inference, from one side (these topics create the main points of this book) and a lot of other theoretical problems and applications, from

Preface

VII

the other side. In our Preface we mention for the most part the papers that are not mentioned and used in the text of the book but play a very important role in the development of the theory of long-memory processes. For example, series, spectral and wavelet analysis for fBm was considered in (AS96), (ALP02a), (Mas93), (Mas96), (RZ91), (DvZ05), (DF02), (DvZ04), (SL95), (Mac81b); local times, the Tanaka formula, the law of the iterated logarithm, maximal properties and the Kallianpur–Robbins law for fBm and related processes were studied in (Ber69), (CNT01), (HO02), (HP04b), (Sin97), (HOS05), (GRV03), (KK97), (KM96), (KO99), (Ros87), (KM96), (Kono96), (Sh96), (ElN93), (Tal96) and (Taq77). Furthermore, stochastic evolution equations driven by fBm were investigated in the papers (AG03), (CD01), (MN03), (TTV03) and some methods of construction of fBm were proposed in (Yor88) and (Sai92). R.J. Adler and G. Samorodnitsky (AS95) considered super processes connected to fBm. The Clark–Ocone theorem for fBm was established in (BE03) and (AOPU00); forward and symmetric integrals for fBm were constructed in (BO04), (CN02), (Zah02b) (note that the general theory of forward, backward and symmetric integrals was created by F. Russo and P. Vallois in (RV93), (RV95a), (RV95b), (RV98) and (RV00)). Detection and prediction problems were discussed in the papers (BP88), (GN96), (Dun06); the stochastic maximum principle for a controlled process governed by an SDE involving fBm was proved in (BHOS02); stochastic Fubini theorem for fBm was studied in (KM00); time rescaling for fBm was investigated in (Mac81a); Hausdorff measure and packing dimension connected to fBm were considered in (Tal95), (TX96), (Xiao91), (Xiao96), (Xiao97a), (Xiao97b); estimation of the parameters of long-memory processes, in particular, the estimates of the Hurst parameter are presented in (Ber94), (BGK06), (BG96), (BG98), (GR03a). Markov properties of some functionals connected with an fBm were considered in (CC98). Rough path analysis for fBm was studied in (CQ02) and some of its applications were considered in the manuscript (HN06); the properties of the Gaussian spaces generated by an fBm were established in (PT01); distribution of functionals connected with fBm was obtained in (CM96), (LN03), (ElN99) (Sin97), (Zha96), (Zha97); the Skorohod–Stratonovich integral for fBm was studied in (Dec01), (ALN01), (AMN01), (AN02); the properties of spectral exponent of fBm were established in (LP95); multi-parameter fractional Brownian fields were studied in (ENO02), (Kam96), (ALP02b), (Lind93), (Gol84), (KK99), (OZ01), (PT02a), (Tal95), (TV03), (TT03), (Tud03), (MisIl03), (MisIl04), (MisIl06), (Mur92); set-parametrized fractional Brownian fields have been studied in the papers (HM06a), (HM06b); asymptotic properties of two-dimensional fractional Brownian fields were considered in (BaNu06). The Malliavin calculus for fBm was developed in (Hu05), (Pri98), (Nua03), (Nua06); fBm in Hilbert space was constructed and investigated in (DPM02). The papers (HN04), (KLeB02), (AHL01), (ALN01), (CKM03) are devoted to stochastic fractional Ornstein–Uhlenbeck, Riesz–Bessel and Lévy type

VIII

Preface

processes. An interesting formula of transformation of fBm with Hurst index H into fBm with index 1 − H was obtained in (Jost06). Mention also the papers (DU98), (Daye03) and forthcoming book (BHOZ07). Note that fBm has a long-memory property only for H ∈ (1/2, 1). In the case H ∈ (0, 1/2) it is a process with short memory. The theory of such processes is quite different. FBm with H ∈ (0, 1/2) was studied in (ALN01), (AMN00), (AI04) and (CN05); simulation of fBm and various applications of fBm were considered in (CM95), (CM96), (Nor95), (Yin96), (Dun00), (Dun01), (DF02), (Seb95) and (Sin94). Fractional Brownian motion as a model of financial markets was proposed in a large number of papers. (See, for example, (AM06), (BE04), (BSV06), (BO02), (BH05), (Che01b), (Dun04), (EvH01), (EvH03), (Gap04), (HO03), (HOS03), (HOS05), (Rog97), (Sch99), (Shi01), (Sot01), (SV03), (WRL03), (WTT99), (Wyss00) and (Zah02a).) Financial markets with memory were considered in (AI05a), (AI05b), (INA07) and (IN07). Moreover, filtering and prediction problems were considered in (CD99), (INA06), (KKA98b), (LeB98), (KLeBR99), (KLeB99), (KLeBR00), (Dun06) and (GN96). In addition, some related applied problems were studied, e.g., in (MS99), (Nar98), (Nor95), (Nor97), (Nor99). An estimate of ruin probabilities for the models with the long-range dependence was studied in (Mis05), (HP04b). Statistical inferences for the processes related to fBm are a very extended area. The major contributions to this theory were made, among other authors, by M. Taqqu and P.M. Robinson. We mention here also the papers of P. Doukhan, A. Khezour and G. Lang (DKL03), L. Giraitis and P.M. Robinson (GR03b), and the papers (DH03), (HH03), (KS03), (MS03), (BLOPST03), (WTT99). Of course, our list of the papers devoted to the theory of fBm is not exhaustive. The book of P. Doukhan, G. Oppenheim, M. Taqqu (editors): Theory and Applications of Long-range Dependence (Birkh¨ auser, Boston 2003) contains papers devoted to different aspects of stochastic calculus for fractional Brownian motion and related processes. We mention, in particular, the papers of D. Surgailis (Sur03a), (Sur03b) and M. Maejima (Mae03), devoted to central and non-central limit theorems, where the asymptotic distribution is not the classical standard normal and the limit process is not the Wiener process. The processes of moving average type are obtained as the limiting ones for increasing sums of some stationary sequences that do not have finite variance. See also the papers (Ho96), (Dec03), (Do03), (Mol03), (PT03), (Taq03), (SW03) from this edition describing stochastic analysis and other aspects of the processes with long memory; papers concerning statistical problems were mentioned above. It is clear from the aforesaid descriptions and citations that there exists the urgent need to systematize the existing results devoted to fractional Brownian motion, to select the best of them (in the author’s opinion) and to present them in appropriate form. Also, some well-known results admit generalizations, and it can be done without great technical difficulties. The present book is devoted to the solution of these two problems. Of course, we cannot claim the complete presentation of all the results concerning fractional

Preface

IX

Brownian motion; it is impossible as the reader can see from aforesaid list. So, we choose only the following topics: Wiener and stochastic integration, Itˆ o formula, Fubini and Girsanov theorems, stochastic differential equations, filtering in the mixed Brownian–fractional-Brownian models, financial applications, some statistical inferences for fractional Brownian motion and the stochastic calculus of multi-parameter fractional Brownian processes. These fields coincide with the main directions of our own interest in the long-memory effect. The book consists of six chapters divided into 41 sections. Chapter 1 is devoted to the Wiener integration (when the integrand is nonrandom) with respect to fractional Brownian motion. Section 1.1 is devoted to the principal definitions from fractional calculus. We recall the notions of fractional integrals and derivatives both for finite and infinite intervals, formulate the Hardy– Littlewood theorem, give the Fourier transformation for fractional integrals and derivatives and calculate the values of some important fractional derivatives. Section 1.2 contains some elementary properties of fractional Brownian motion including the simplest spectral representations. Section 1.3 contains the Mandelbrot–van Ness representation of fractional Brownian motion via the Wiener process and some fractional kernels on real axes. These kernels are the prototypes for the future definition of the Wiener integration w.r.t. fBm. Sections 1.4 and 1.5 describe the construction of fractional Brownian motion and fractional noise on white noise space. Such space is convenient for applications since it is possible to consider mixed Brownian–fractional-Brownian processes and linear combinations of fractional Brownian motions with different Hurst indices on such space and to apply Wick calculus to them. It is proved that any fractional noise with H ∈ [1/2, 1] belongs to the Hida distribution space S ∗ (we establish the corresponding estimates for the negative norms). The relations between motion and noise are established as in the usual Wick calculus for the Wiener noise. In Section 1.6 we return to fBm on arbitrary space. The section contains the definition of the Wiener integral with respect to fBm and various relations between different “integrable spaces” related to fBm. Section 1.7 is devoted to (non) completeness of the Gaussian spaces generated by fBm, in connection with their norms. Section 1.8 contains the representation of fBm via the Wiener process on any finite interval [0, T ] and some representations for auxiliary processes. Sections 1.9 and 1.10 present moment estimates for Wiener integrals w.r.t. fractional Brownian motion. Using the conditions of continuity of the trajectories of Wiener integrals w.r.t. fBm (Section 1.11) we extend in Section 1.12 the upper moment estimates to solutions of very simple stochastic differential equations containing Wiener integrals. Section 1.13 contains the proof of the stochastic Fubini theorem for the Wiener integrals w.r.t. fractional Brownian motion. Section 1.14 deals with such Gaussian processes that can be transformed into martingales with the help of some kernels (fBm can be transformed into the Wiener process with the help of hypergeometric kernels). Section 1.15 is devoted to different convergence schemes, in which fBm is approximated by the sequence of semimartingales, and even

X

Preface

by the continuous processes with bounded variation. In the last case Wiener integrals w.r.t. fractional Brownian motion also can be approximated. Section 1.16 demonstrates the H¨ older properties of the Wiener integrals w.r.t. fractional Brownian motion. Section 1.17 contains some auxiliary estimates for fractional derivatives of fBm and for the Wiener integrals w.r.t. Wiener process via the Garsia–Rodemich–Rumsey inequality. Section 1.18 contains one- and two-sided bounds for power variations for fBm and Wiener integrals w.r.t. fBm. Section 1.19 contains the result stating that some conditions of quadratic variation of a stochastic process supply that this process is an fBm; it is kind of generalization of the Lévy theorem for the Wiener process. Section 1.20 concludes; it describes Wiener fields on the plane and related fractional integrals and derivatives. Chapter 2 is devoted to stochastic integration w.r.t. fractional Brownian motion and other aspects of stochastic calculus of fBm. There exist several approaches to stochastic integration w.r.t. fractional Brownian motion: pathwise integration, Wick integration, Skorohod integration, isometric integration and some others that are not mentioned here. Pathwise stochastic integration in fractional Sobolev-type spaces and in fractional Besov-type spaces is described in Section 2.1 and is generalized to fBm fields in Section 2.2. Wick integration is considered in Section 2.3 and is reduced to the integration w.r.t. white noise. Two approaches to the Skorohod integration and their connections with forward, backward and symmetric integration are discussed in Section 2.4. Isometric integration is the subject of section 2.5. The stochastic Fubini theorem and various versions of the Itˆ o formula and the Girsanov theorem are contained in Sections 2.6–2.8 which conclude Chapter 2. Chapter 3 is devoted to different properties of stochastic differential equations involving fBm. Section 3.1 contains the conditions of existence and uniqueness of solution of a “pure” stochastic differential equation containing a pathwise integral w.r.t. fBm and the estimates of its solution. Most of the theorems are stated in the spirit of the paper (NR00) but the results of Z¨ ahle (Zah99) on existence of local solutions are also presented since they are used later for construction of global solutions in the cases when other results cannot help. Some properties of SDEs with stationary coefficients including differentiability and local differentiability of the solutions are presented in Subsection 3.1.4. Existence and uniqueness of solutions of SDEs with two-parameter fractional Brownian fields is contained in Subsection 3.1.6. Semilinear “pure” and “mixed” SDEs are considered in detail in Subsections 3.1.5 and 3.2.1. The rate of convergence of Euler approximations of solutions of SDEs involving fBm is the subject to Section 3.4. SDEs with fractional white noise are considered in Section 3.3, and a detailed discussion of SDEs with additive Wiener integrals w.r.t. fBm is presented in Section 3.5. Chapter 4 is devoted to filtering problems in the mixed fractional models. Section 4.1 considers the case when the signal process is modeled by mixed stochastic differential equations involving both fractional Brownian motion and the Wiener process and the observation process is the sum of the fractional

Preface

XI

Brownian integral and the term of bounded variation. Optimal filtering in conditionally Gaussian linear systems with mixed signals and fractional Brownian observation is studied in Section 4.2. In these sections we consider only nonrandom integrands in all the stochastic integrals. In Section 4.3 we make an attempt to generalize the model and consider polynomial integrands depending on fBm. Chapter 5 is devoted to financial models involving fBm. In general, financial markets fairly often have a long memory and it is a natural idea to model them with the help of fBm or with the help of some of its modifications. Nevertheless, it is not so easy to do this because the market model is “good” when it does not admit arbitrage and the models involving fractional Brownian motion are not arbitrage-free. So, this chapter is devoted to some methods of construction of the long-memory arbitrage-free models and to the discussion of different approaches to this problem. In Section 5.1 we introduce the mixed Brownian–fractional-Brownian model and establish conditions that ensure the absence of arbitrage in such a model. In Section 5.2 we consider a fractional version of the Black–Scholes equation for the mixed Brownian-fractional Brownian model which contains pathwise integrals w.r.t. fBm, discuss possible applications of Wick products in fractional financial models and produce Black–Scholes equation for the fractional model involving Wick product w.r.t. fBm. Chapter 6 is devoted to the solution of some statistical problems involving fBm. The choice of the first problem which is solved in Sections 6.1 and 6.2 was evoked by some financial reasonings considered in Chapter 5. More exactly, we try to determine which of the two geometric Brownian motions from (5.2.6) serves as the better model for the real financial market, i.e. we test the complex hypothesis concerning the shifts in the geometric fBm; one of the shifts corresponds to the pathwise integral, and another to the Wick integral. In Section 6.3 we consider the existence and the properties of estimates of the shift parameter in different “pure” and “mixed” models involving fBm and, possibly, the Wiener process, which can be independent of or, conversely, “linearly dependent” on fractional Brownian motion. I am grateful to Esko Valkeila who invited me several times to Helsinki University during the period of 1997-2005 and presented a possibility for fruitful work and discussion of the problems connected to fractional Brownian motion and related topics. Also, I am grateful to David Nualart for inviting me to Barcelona University during 2001–2003 when we discussed the problems connected to stochastic differential equations involving fBm. My thanks to all my other coauthors, with whom we have written the series of papers devoted to the stochastic calculus for fractional Brownian motion, especially to Jean Memin, Alexander Kukush, Georgij Shevchenko and Taras Androshchuk. My special thanks to Murad Taqqu and Christian Bender for their useful suggestions concerning contents of the minicourse of the lectures devoted to the stochastic calculus for fBm that I delivered in Helsinki Technology University

XII

Preface

in May 2005. I wish to thank also Celine Jost who has carefully read a part of the text of this book and made a lot of improvements.

Kiev, April 24 2007

Yuliya Mishura

Contents

1

Wiener Integration with Respect to Fractional Brownian Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1 The Elements of Fractional Calculus . . . . . . . . . . . . . . . . . . . . . . . 1.2 Fractional Brownian Motion: Definition and Elementary Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 Mandelbrot–van Ness Representation of fBm . . . . . . . . . . . . . . . . 1.4 Fractional Brownian Motion with H ∈ ( 12 , 1) on the White Noise Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5 Fractional Noise on White Noise Space . . . . . . . . . . . . . . . . . . . . . 1.6 Wiener Integration with Respect to fBm . . . . . . . . . . . . . . . . . . . 1.7 The Space of Gaussian Variables Generated by fBm. . . . . . . . . . 1.8 Representation of fBm via the Wiener Process on a Finite Interval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.9 The Inequalities for the Moments of the Wiener Integrals with Respect to fBm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.10 Maximal Inequalities for the Moments of Wiener Integrals with Respect to fBm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.11 The Conditions of Continuity of Wiener Integrals with Respect to fBm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.12 The Estimates of Moments of the Solution of Simple Stochastic Differential Equations Involving fBm . . . . . . . . . . . . . 1.13 Stochastic Fubini Theorem for the Wiener Integrals w.r.t fBm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.14 Martingale Transforms and Girsanov Theorem for Longmemory Gaussian Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.15 Nonsemimartingale Properties of fBm; How to Approximate Them by Semimartingales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.15.1 Approximation of fBm by Continuous Processes of Bounded Variation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.15.2 Convergence B H,β → B H in Besov Space W λ [a, b]. . . . . 1.15.3 Weak Convergence to fBm in the Schemes of Series . . . .

1 1 7 9 10 12 16 24 26 35 41 54 55 57 58 71 71 73 78

XIV

Contents

1.16 H¨ older Properties of the Trajectories of fBm and of Wiener Integrals w.r.t. fBm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 1.17 Estimates for Fractional Derivatives of fBm and of Wiener Integrals w.r.t. Wiener Process via the Garsia–Rodemich– Rumsey Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 1.18 Power Variations of fBm and of Wiener Integrals w.r.t. fBm . . 90 1.19 Lévy Theorem for fBm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 1.20 Multi-parameter Fractional Brownian Motion . . . . . . . . . . . . . . . 117 1.20.1 The Main Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 1.20.2 H¨ older Properties of Two-parameter fBm . . . . . . . . . . . . . 117 1.20.3 Fractional Integrals and Fractional Derivatives of Two-parameter Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 118 2

Stochastic Integration with Respect to fBm and Related Topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 2.1 Pathwise Stochastic Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 2.1.1 Pathwise Stochastic Integration in the Fractional Sobolev-type Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 2.1.2 Pathwise Stochastic Integration in Fractional Besov-type Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128 2.2 Pathwise Stochastic Integration w.r.t. Multi-parameter fBm . . . 131 2.2.1 Some Additional Properties of Two-parameter Fractional Integrals and Derivatives . . . . . . . . . . . . . . . . . . 131 2.2.2 Generalized Two-parameter Lebesgue–Stieltjes Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 2.2.3 Generalized Integrals of Two-parameter fBm in the Case of the Integrand Depending on fBm . . . . . . . . . . . . . 136 2.2.4 Pathwise Integration in Two-parameter Besov Spaces . . 136 2.2.5 The Existence of the Integrals of the Second Kind of a Two-parameter fBm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 2.3 Wick Integration with Respect to fBm with H ∈ [1/2, 1) as S ∗ -integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 2.3.1 Wick Products and S ∗ -integration . . . . . . . . . . . . . . . . . . . 141 2.3.2 Comparison of Wick and Pathwise Integrals for “Markov” Integrands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 2.3.3 Comparison of Wick and Stratonovich Integrals for “General” Integrands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 2.3.4 Reduction of Wick Integration w.r.t. Fractional Noise to the Integration w.r.t. White Noise . . . . . . . . . . . . . . . . . 157 2.4 Skorohod, Forward, Backward and Symmetric Integration w.r.t. fBm. Two Approaches to Skorohod Integration . . . . . . . . 158 2.5 Isometric Approach to Stochastic Integration with Respect to fBm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162 2.5.1 The Basic Idea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162 2.5.2 First- and Higher-order Integrals with Respect to X . . . 164 2.5.3 Generalized Integrals with Respect to fBm . . . . . . . . . . . . 169

Contents

XV

2.6 Stochastic Fubini Theorem for Stochastic Integrals w.r.t. Fractional Brownian Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 2.7 The Itˆ o Formula for Fractional Brownian Motion . . . . . . . . . . . . 182 2.7.1 The Simplest Version . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182 2.7.2 Itˆ o Formula for Linear Combination of Fractional Brownian Motions with Hi ∈ [1/2, 1) in Terms of Pathwise Integrals and Itˆ o Integral . . . . . . . . . . . . . . . . . . 183 2.7.3 The Itˆ o Formula in Terms of Wick Integrals . . . . . . . . . . 184 2.7.4 The Itˆ o Formula for H ∈ (0, 1/2) . . . . . . . . . . . . . . . . . . . . 185 2.7.5 Itˆ o Formula for Fractional Brownian Fields . . . . . . . . . . . 186 2.7.6 The Itˆ o Formula for H ∈ (0, 1) in Terms of Isometric Integrals, and Its Applications . . . . . . . . . . . . . . . . . . . . . . 189 2.8 The Girsanov Theorem for fBm and Its Applications . . . . . . . . . 191 2.8.1 The Girsanov Theorem for fBm . . . . . . . . . . . . . . . . . . . . . 191 2.8.2 When the Conditions of the Girsanov Theorem Are Fulfilled? Differentiability of the Fractional Integrals . . . 193 3

Stochastic Differential Equations Involving Fractional Brownian Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 3.1 Stochastic Differential Equations Driven by Fractional Brownian Motion with Pathwise Integrals . . . . . . . . . . . . . . . . . . 197 3.1.1 Existence and Uniqueness of Solutions: the Results of Nualart and Rˇ a¸scanu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 3.1.2 Norm and Moment Estimates of Solution . . . . . . . . . . . . . 202 3.1.3 Some Other Results on Existence and Uniqueness of Solution of SDE Involving Processes Related to fBm with (H ∈ (1/2, 1)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204 3.1.4 Some Properties of the Stochastic Differential Equations with Stationary Coefficients . . . . . . . . . . . . . . . 206 3.1.5 Semilinear Stochastic Differential Equations Involving Forward Integral w.r.t. fBm . . . . . . . . . . . . . . . . . . . . . . . . . 220 3.1.6 Existence and Uniqueness of Solutions of SDE with Two-Parameter Fractional Brownian Field . . . . . . . . . . . . 223 3.2 The Mixed SDE Involving Both the Wiener Process and fBm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225 3.2.1 The Existence and Uniqueness of the Solution of the Mixed Semilinear SDE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225 3.2.2 The Existence and Uniqueness of the Solution of the Mixed SDE for fBm with H ∈ (3/4, 1) . . . . . . . . . . . . . . . 227 3.2.3 The Girsanov Theorem and the Measure Transformation for the Mixed Semilinear SDE . . . . . . . . 238 3.3 Stochastic Differential Equations with Fractional White Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240 3.3.1 The Lipschitz and the Growth Conditions on the Negative Norms of Coefficients . . . . . . . . . . . . . . . . . . . . . . 240 3.3.2 Quasilinear SDE with Fractional Noise . . . . . . . . . . . . . . . 241

XVI

Contents

3.4 The Rate of Convergence of Euler Approximations of Solutions of SDE Involving fBm . . . . . . . . . . . . . . . . . . . . . . . . . . . 243 3.4.1 Approximation of Pathwise Equations . . . . . . . . . . . . . . . . 244 3.4.2 Approximation of Quasilinear Skorohod-type Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255 3.5 SDE with the Additive Wiener Integral w.r.t. Fractional Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262 3.5.1 Existence of a Weak Solution for Regular Coefficients . . 263 3.5.2 Existence of a Weak Solution for SDE with Discontinuous Drift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266 3.5.3 Uniqueness in Law and Pathwise Uniqueness for Regular Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271 3.5.4 Existence of a Strong Solution for the Regular Case . . . . 272 3.5.5 Existence of a Strong Solution for Discontinuous Drift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274 3.5.6 Estimates of Moments of Solutions for Regular Case and H ∈ (0, 1/2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278 3.5.7 The Estimates of the Norms of the Solution in the Orlicz Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280 3.5.8 The Distribution of the Supremum of the Process X on [0, T ] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284 3.5.9 Modulus of Continuity of Solution of Equation Involving Fractional Brownian Motion . . . . . . . . . . . . . . . 287 4

Filtering in Systems with Fractional Brownian Noise . . . . . . 291 4.1 Optimal Filtering of a Mixed Brownian–Fractional-Brownian Model with Fractional Brownian Observation Noise . . . . . . . . . . 291 4.2 Optimal Filtering in Conditionally Gaussian Linear Systems with Mixed Signal and Fractional Brownian Observation Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295 4.3 Optimal Filtering in Systems with Polynomial Fractional Brownian Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298

5

Financial Applications of Fractional Brownian Motion . . . . . 301 5.1 Discussion of the Arbitrage Problem . . . . . . . . . . . . . . . . . . . . . . . 301 5.1.1 Long-range Dependence in Economics and Finance . . . . 301 5.1.2 Arbitrage in “Pure” Fractional Brownian Model. The Original Rogers Approach . . . . . . . . . . . . . . . . . . . . . . 302 5.1.3 Arbitrage in the “Pure” Fractional Model. Results of Shiryaev and Dasgupta . . . . . . . . . . . . . . . . . . . 304 5.1.4 Mixed Brownian–Fractional-Brownian Model: Absence of Arbitrage and Related Topics . . . . . . . . . . . . . 305 5.1.5 Equilibrium of Financial Market. The Fractional Burgers Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321 5.2 The Different Forms of the Black–Scholes Equation . . . . . . . . . . 322 5.2.1 The Black–Scholes Equation for the Mixed Brownian–Fractional-Brownian Model . . . . . . . . . . . . . . . . 322

Contents

XVII

5.2.2 Discussion of the Place of Wick Products and Wick– Itˆ o–Skorohod Integral in the Problems of Arbitrage and Replication in the Fractional Black–Scholes Pricing Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323 6

Statistical Inference with Fractional Brownian Motion . . . . . 327 6.1 Testing Problems for the Density Process for fBm with Different Drifts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327 6.1.1 Observations Based on the Whole Trajectory with σ and H Known . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329 6.1.2 Discretely Observed Trajectory and σ Unknown . . . . . . . 331 6.2 Goodness-of-fit Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335 6.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335 6.2.2 The Whole Trajectory Is Observed and the Parameters µ and σ Are Known . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335 6.2.3 Goodness-of-fit Tests with Discrete Observations . . . . . . 337 6.2.4 On Volatility Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 340 6.2.5 Goodness-of-fit Test with Unknown µ and σ . . . . . . . . . . 342 6.3 Parameter Estimates in the Models Involving fBm . . . . . . . . . . 343 6.3.1 Consistency of the Drift Parameter Estimates in the Pure Fractional Brownian Diffusion Model . . . . . . . . . . . . 344 6.3.2 Consistency of the Drift Parameter Estimates in the Mixed Brownian–fractional-Brownian Diffusion Model with “Linearly” Dependent Wt and BtH . . . . . . . . . . . . . . 349 6.3.3 The Properties of Maximum Likelihood Estimates in Diffusion Brownian–Fractional-Brownian Models with Independent Components . . . . . . . . . . . . . . . . . . . . . . 354

A

Mandelbrot–van Ness Representation: Some Related Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363

B

Approximation of Beta Integrals and Estimation of Kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 391

1 Wiener Integration with Respect to Fractional Brownian Motion

1.1 The Elements of Fractional Calculus Let α > 0 (and in most cases below α < 1 though this is not obligatory). Define the Riemann–Liouville left- and right-sided fractional integrals on (a, b) of order α by x 1 α f (t)(x − t)α−1 dt, (Ia+ f )(x) := Γ (α) a and α (Ib− f )(x)

1 := Γ (α)

b

f (t)(t − x)α−1 dt, x

respectively. α We say that the function f ∈ D(Ia+(b−) ) (the symbol D(·) denotes the domain of the corresponding operator), if the respective integrals converge for almost all (a.a.) x ∈ (a, b) (with respect to (w.r.t.) Lebesgue measure). The Riemann–Liouville fractional integrals on R are defined as x 1 α (I+ f )(x) := f (t)(x − t)α−1 dt, Γ (α) −∞ and α f )(x) := (I−

1 Γ (α)

∞

f (t)(t − x)α−1 dt,

x

respectively. α The function f ∈ D(I± ) if the corresponding integrals converge for a.a. α ), 1 ≤ p < α1 . x ∈ R. According to (SKM93), we have inclusion Lp (R) ⊂ D(I± Moreover, the following Hardy–Littlewood theorem holds. Theorem 1.1.1 ((SKM93)). Let 1 ≤ p, q < ∞, 0 < α < 1. Then the α operators I± are bounded from Lp (R) to Lq (R) if and only if 1 < p < α1 and q = p(1 − αp)−1 . This means, in particular, that for any 1 < p < α1 and p , there exists a constant Cp,q,α such that q = 1−αp

2


R

R

|f (u)||x − u|α−1 du

q1

q dx

≤ Cp,q,α f Lp (R) .

(1.1.1)

Fractional integration admits the following composition formulas for fractional integrals: α+β α+β α β α β Ia+ f = Ia+ f, Ib− Ib− f = Ib− f Ia+

for f ∈ L1 [a, b]. If α + β ≥ 1 then these equalities hold at any point x ∈ (a, b), otherwise they hold for a.a. x. Also, α+β α β I± I± = I± f

for f ∈ Lp (R), α, β > 0 and α + β < p1 . Let f ∈ Lp [a, b], g ∈ Lq [a, b], p, q ≥ 1 and p1 + 1q ≤ 1 + α, or let p > 1, q > 1 and p1 + 1q = 1 + α. Then we have the following integration-by-parts formula for fractional integralsintegrationby-parts formula!for fractional integrals:

b

b

α g(x)(Ia+ f )(x)dx = a

f (x)(Ibα− g)(x)dx.

a

Let f ∈ Lp (R), g ∈ Lq (R), p > 1, q > 1 and R

1 p

+

1 q

= 1 + α. Then

α g(x)(I+ f )(x)dx =

R

α f (x)(I− g)(x)dx.

(1.1.2)

Let C λ (T) be the set of Hölder continuous functions f : T → R of order λ, i.e. C λ (T) = f : T → R f λ := sup |f (t)| t∈T

+ sup |f (s) − f (t)|(t − s)−λ < ∞ . s,t∈T

α If α > 0 and αp > 1, then I± (Lp (R)) ⊂ C λ [a, b] for any −∞ < a < b < ∞ 1 and 0 < λ ≤ α − p .

The next result is evident. Lemma 1.1.2. Let 0 < α < 1, f ∈ Lp (R), 1 ≤ p < f (x) = 0 for a.a. x ∈ R.

1 α

α and I± f = 0. Then

α For p ≥ 1, denote by I± (Lp (R)) the class of functions f , that can be α ϕ for some presented as Riemann–Liouville integrals, more exactly, f = I± ϕ ∈ Lp (R), p ≥ 1. Lemma 1.1.2 ensures the uniqueness of such function ϕ. For 0 < α < 1 it coincides for a.a. x ∈ R with the left- (right-) sided Riemann– Liouville fractional derivative of f of order α. These derivatives are denoted by

1.1 The Elements of Fractional Calculus −α α (I+ f )(x) = (D+ f )(x) :=

d 1 Γ (1 − α) dx

−α α f )(x) = (D− f )(x) := (I−

d −1 Γ (1 − α) dx

and

x

−∞

∞

3

f (t)(x − t)−α dt, f (t)(t − x)−α dt,

x

respectively. α (Lp (R)) coincides with the class of those functions For p > 1, the class I± p f ∈ Lr (R), r = 1−αp , for which the integrals

x−ε

−∞

and

∞

(f (x) − f (t))(x − t)−α−1 dt

(f (x) − f (t))(t − x)−α−1 dt,

x+ε α respectively, converge in Lp (R) as ε → 0. Thus, for f ∈ I± (Lp (R)) with p > 1 the Riemann–Liouville derivatives coincide with the Marchaud fractional derivatives 1 α (D+ f )(x) := (f (x) − f (x − y))y −α−1 dy, Γ (1 − α) R+

and α − f )(x) := (D

1 Γ (1 − α)

R+

(f (x) − f (x + y))y −α−1 dy,

α respectively. If α > 0 and αp < 1, then I± (Lp (R)) ⊂ Lq (R) for 1q = p1 − α. The Riemann–Liouville fractional derivatives can be considered on any α (Lp [a, b]) of interval [a, b] ⊂ R in the following way: we introduce the class I± α α functions f that can be presented as f = Ia+ ϕ (f = Ib− ϕ) for ϕ ∈ Lp [a, b], p ≥ 1, where we denote x d 1 −α α f )(x) = (Da+ f )(x) = f (t)(x − t)−α dt, (Ia+ Γ (1 − α) dx a

and −α α (Ib− f )(x) = (Db− f )(x) = −

d 1 Γ (1 − α) dx

b

f (t)(t − x)−α dt,

x

α f respectively. In this case the Riemann–Liouville fractional derivatives Da+ α and Db− f admit the Weyl representation of fractional derivatives (we suppose that f = 0 outside (a, b)): 1 α f (x)(x − a)−α (Da+ f )(x) = Γ (1 − α) x +α (f (x) − f (t))(x − t)−α−1 dt · 1(a,b) (x), a

4


and 1 f (x)(b − x)−α Γ (1 − α) b +α (f (x) − f (t))(t − x)−α−1 dt · 1(a,b) (x),

α (Db− f )(x) =

x

respectively, where the convergence of the integrals holds pointwise for a.a. x ∈ (a, b) for p = 1 and in Lp [a, b] for p > 1. α ϕ for some According to (SKM93, Theorem 13.4), we have that f = Ia+ −α ϕ ∈ Lp [a, b], where 1 < p < ∞, if and only if f (x)(x − a) ∈ Lp [a, b] and

b

|ψε (x)|p dx < ∞,

sup ε>0

x−ε where ψε (x) = a and p ≥ 1. Then

a+ε

f (x)−f (t) (x−t)1+α dt,

α a + ε ≤ x ≤ b. Let f ∈ I± (Lp (R)), 0 < α < 1 α −α I± I± f = f ;

(1.1.3)

moreover, for f ∈ L1 (R) we have that −α α I± I± f = f.

(1.1.4)

0 We set I± f := f . The composition formula for fractional derivatives has the form β α+β α Da+ f = Da+ f, Da+

(1.1.5)

α+β where α ≥ 0, β ≥ 0 and f ∈ Ia+ (L1 (R)). α (Lp [a, b]) and g ∈ Also, under the assumptions 0 < α < 1, f ∈ Ia+ α Ib− (Lq [a, b]), 1/p + 1/q ≤ 1 + α we have the integration-by-parts formula for fractional derivatives b b α α (Da+ f )(x)g(x)dx = f (x)(Db− g)(x)dx. (1.1.6) a

a

α α f and Db− f exist, belong For 0 < α < 1 and f ∈ C 1 [a, b], the derivatives Da+ to Lr [a, b] for 1 ≤ r < 1/α, and have the form x 1 α −α −α f (t)(x − t) dt , Da+ f = f (a)(x − a) + Γ (1 − α) a

and α f Db−

1 = Γ (1 − α)

−α

f (b)(b − x)

b

−

−α

f (t)(t − x) x

respectively. Let the general indicator function be given by

dt ,

1.1 The Elements of Fractional Calculus

⎧ ⎪ ⎨ 1, 1(a,b) (t) = −1, ⎪ ⎩ 0,

5

a ≤ t < b, b ≤ t < a, otherwise.

Lemma 1.1.3. Let H ∈ (0, 12 ) ∪ ( 12 , 1) and α = H − 12 . Then, for all t ∈ R, we have the equality α 1(0,t) )(x) = (I−

1 α ((t − x)α + − (−x)+ ). Γ (1 + α)

Proof. Let H ∈ ( 12 , 1) and, for example, x < 0 < t (the other cases can be considered similarly). Then, ∞ 1 1(0,t) (u)(u − x)α−1 du Γ (α) x t 1 1 ((t − x)α − (−x)α ) . (u − x)α−1 du = = Γ (α) 0 Γ (α + 1)

α 1(0,t) )(x) = (I−

(1.1.7)

Let H ∈ (0, 12 ). According to the definition of the fractional derivative and (1.1.3), we must prove that ∞ α −α−1 ((t − u)α du = Γ (−α)Γ (α + 1)1(0,t) (x). (1.1.8) + − (−u)+ )(u − x) x

Let, for example, 0 < x < t. Then the left-hand side of (1.1.8) equals t (t − u)α (u − x)−α−1 du1(0,t) (x) x

= B(α + 1, −α)1(0,t) (x) = Γ (−α)Γ (α + 1)1(0,t) (x). The other cases can be considered similarly.

1 α α Remark 1.1.4. Obviously, (I+ 1(a,b) (x)) = Γ (1+α) ((b − x)α + − (a − x)+ ), −∞ < a < b < ∞. Let f ∈ L1 (R). The Fourier transform of f is defined as (Ff )(x) = f(x) = eixt f (t)dt. R

Denote by S(R) the class of smooth, i.e. infinitely differentiable, and rapidly decreasing functions. Theorem 1.1.5 ((SKM93)). (i) For any 0 < α < 1 and f ∈ L1 (R) it holds that α f ) = f(x) · (∓ix)−α , F(I± απi sign x . where (∓ix)α = |x|α exp ∓ 2

6


(ii) For any 0 < α < 1 and f ∈ S(R) it holds that −α F(I± f ) = f(x) · (∓ix)α .

For H ∈ (0, 1) we introduce the set FH := f ∈ L2 (R), f : R → R

with the norm ||f ||2FH =

R

2 −2α dx < ∞ |f (x)| |x| R

|f(x)|2 · |x|−2α dx.

Here and throughout the whole text α := H − 1/2 . The set FH will be considered in detail in Sections 1.6 and 1.7. We say that f is step function, or elementary function, if there exist a finite number of points tk ∈ R, 0 ≤ k ≤ n − 1, and ak ∈ R, 1 ≤ k ≤ n, such that n ak 1[tk−1 ,tk ) (t). f (t) = k=1

Lemma 1.1.6. Let f ∈ FH . Then there exists a sequence of step functions fn , such that f − fn FH → 0, n → ∞. Theorem 1.1.7 ((PT00b)). For H ∈ (0, 1), the set FH is a linear space with inner product (f, g)FH = g (x)|x|−2α dx, α = H − 1/2. f(x) R

Moreover, the set of elementary functions belongs to FH , and it is dense in FH . Proof. The first statement is evident. Furthermore, < a < for any 2−∞−2α (a,b) (x)| |x| dx = b < ∞, it holds that 1(a,b) ∈ FH , because R |1 ixb ixa 2 −2−2α |e − e | |x| dx, and the latter integral is equivalent to the conR vergent integral |x|−2−2α dx, in the neighborhood of ±∞, and equivalent to the convergent integral |x|−2α dx in the neighborhood of 0. Therefore, any step function belongs to FH . The second statement then follows from Lemma 1.1.6. Lemma 1.1.8 ((PT00b)). Let f ∈ L2 (R). Then, for any H ∈ (0, 1), there exists a sequence of step functions fn such that |f(x) − fn (x)|x|−2α |2 dx → 0, n → ∞. (1.1.9) R

1.2 Fractional Brownian Motion: Definition and Elementary Properties

7

Proof. Indeed, for ε > 0, put fε (x) := f(x)1{|x|>ε} . Then R |f(x) − fε (x)|2 dx → 0, ε → 0. Let H ∈ (0, 12 ). Then fε (x) = (f(x)|x|α 1{|x|>ε} )|x|−α = gε (x)|x|−α , where gε ∈ L2 (R), α = H − 1/2. Now (1.1.9) follows from Lemma

1.1.6. In the case H ∈ [ 12 , 1) the proof is similar.

1.2 Fractional Brownian Motion: Definition and Elementary Properties Let (Ω, F, P ) be a complete probability space. Definition 1.2.1. The (two-sided, normalized) fractional Brownian motion (fBm) with Hurst index H ∈ (0, 1) is a Gaussian process B H = {BtH , t ∈ R} on (Ω, F, P ), having the properties (i) B0H = 0, (ii) EBtH = 0, t ∈ R, (iii) EBtH BsH = 12 (|t|2H + |s|2H − |t − s|2H ), s, t ∈ R. Remark 1.2.2. Since E(BtH −BsH )2 = |t−s|2H and B H is a Gaussian process, it has a continuous modification, according to the Kolmogorov theorem. Indeed, n 22 nH . for all n ≥ 1 it holds that E|BtH − BsH |n = 1 Γ ( n+1 2 )|t − s| π2 Remark 1.2.3. For H = 1, we set BtH = Bt1 = tξ, where ξ is a standard normal random variable. Remark 1.2.4. It is possible to consider the fBm only on R+ (one-sided fBm) with evident changes in Definition 1.2.1. The characteristic function has the form n 1 H ϕλ (t) := E exp i λk Btk = exp − (Ct λ, λ) , 2 k=1

where Ct = (EBtHk BtHi )1≤i,k≤n and (·, ·) is the inner product on Rn . Therefore, it follows from item (iii) of Definition 1.2.1, that for any β > 0 1 2H (1.2.1) ϕλ (βt) = exp − β (Ct λ, λ) . 2 Definition 1.2.5. A stochastic process X = {Xt , t ∈ R} is called b-selfsimilar if d {Xat , t ∈ R} = {ab Xt , t ∈ R} in the sense of finite-dimensional distributions.

8


From Definition 1.2.5 and (1.2.1) it follows that B H is H-self-similar. Note that E(BtH −BsH )(BuH −BvH ) =

1 (|s−u|2H +|t−v|2H −|t−u|2H −|s−v|2H ). (1.2.2) 2

It follows from (1.2.2) that the process B H has stationary increments (evidently, it is not stationary itself). Let H = 12 . Then the increments of B H are non-correlated, and consequently independent. So B H is a Wiener process which we denote further by B or W . For H ∈ (0, 12 ) ∪ ( 12 , 1) and t1 < t2 < t3 < t4 , it follows from (1.2.2) for α = H − 1/2 that E(BtH4 − BtH3 )(BtH2 − BtH1 ) = 2αH

t2

t1

t4

(u − v)2α−1 du dv.

t3

Therefore, the increments are positively correlated for H ∈ ( 12 , 1) and negatively correlated for H ∈ (0, 12 ). Furthermore, for any n ∈ Z \ {0}, the autocovariance function is given by

1

n+1

H r(n) := EB1H (Bn+1 − BnH ) = 2αH

(u − v)2α−1 du dv 0

n

∼ 2αH|n|2α−1 ,

|n| → ∞.

If H ∈ (0, 12 ), then n∈Z |r(n)| ∼ n∈Z\{0} |n|2α−1 < ∞. ∞ If H ∈ ( 12 , 1), then n=1 |r(n)| ∼ n∈Z\{0} |n|2α−1 = ∞. In this case we say that fBm B H has of long-range dependence. For the spectral the property H − BnH , n ∈ Z , which is denoted by fH (λ), density function of XnH := Bn+1 it holds that (BG96; DvZ05), (0) |λ + 2πk|−2−2α , λ ∈ [−π, π] , fH (λ) = CH |eiλ − 1|2 k∈Z (0)

where CH is some constant depending on H. It is easy to see that fH (λ) ∼ CH |λ|2 |λ|−2−2α = CH |λ|−2α (0)

as λ → 0. Therefore, for H ∈ ( 12 , 1) it holds that fH (λ) → ∞ as λ → 0, and, for H ∈ (0, 12 ), it holds that f (λ) → 0 as λ → 0. According to (PT00b) and (ST94), B H admits the spectral representation d (1) = {BtH , t ∈ R} = {CH R (eitx − 1) (ix)−1 |x|−α dB(x), t ∈ R}, where B B1 + iB2 is a complex Gaussian measure with B1 (A) = B1 (−A), B2 (A) = for any Borel set A of −B2 (−A) and E(B1 (A))2 = E(B2 (A))2 = mesh(A) 2 12 (1) finite Lebesgue measure mesh(A) and CH = Γ (2H+1) sin(1/2·π(H+1/2)) . 2π

1.3 Mandelbrot–van Ness Representation of fBm

9

1.3 Mandelbrot–van Ness Representation of fBm Let W = {Wt , t ∈ R} be the two-sided Wiener process, i.e. the Gaussian process with independent increments satisfying EWt = 0 and EWt Ws = s ∧ t, 1 α s, t ∈ R. Evidently, W = B 2 . Denote kH (t, u) := (t − u)α + − (−u)+ , where 1 α = H − 2 . The following representation is due to Mandelbrot and van Ness (MvN68). H

H

Theorem 1.3.1. The process B = {B t , t ∈ R} defined by 1 1 H (2) ,1 , B t := CH kH (t, u)dWu , H ∈ 0, ∪ 2 2 R 1/2 2H sin πHΓ (2H) 1 − 12 (2) α α 2 where CH = , (1 + s) − s ds + = 2H Γ (H + 1/2) R+ has a continuous modification which is a normalized two-sided fBm. (2)

Remark 1.3.2. The constant CH is calculated in Appendix A. H

H

H

Proof. Evidently, B is a Gaussian process with B 0 = 0 and EB t = 0. Furthermore, it holds that for t > 0, t (2) 2 0 2 H kH (t, u)du + (t − u)2α du = t2H . E(B t )2 = CH −∞

0

For t < 0 we have that (2) 2 H E(B t )2 = CH

t

−∞

2 kH (t, u)du +

0

(−u)2α du = (−t)2H .

t

Furthermore, for h > 0, it holds that H B s+h

−

H Bs

=

(2) CH

s

−∞

kH (s + h, u) − kH (s, u) dWu +

s+h

kH (s + h, u) dWu =: I1 + I2 .

(1.3.1)

s

Note that the terms I1 and I2 on the right-hand side of (1.3.1) are independent, and the Wiener process W has stationary increments. Therefore, d

I1 = H

0

−∞ H

d kH (s, u) − kH (0, u) dWu , I2 =

h

kH (h, u)dWu , 0

H

and E(B s+h −B s )2 = E(B h )2 = h2H . By combining these results, we obtain that

10

1 Wiener Integration with Respect to Fractional Brownian Motion H

H

EB s B t =

H 2 H 1 H 2 H 2 E Bs + E Bt − E Bt − Bs 2 1 = |t|2H + |s|2H − |t − s|2H . 2

(1.3.2)

The proof follows immediately from Definition 1.2.1 and Remark 1.2.2.

Define the operator H M± f := (3)

(3)

α f, H ∈ (0, 12 ) ∪ ( 12 , 1), CH I± f, H = 12 ,

(1.3.3)

(2)

where CH = CH Γ (H + 12 ). Corollary 1.3.3. It follows from Lemma 1.1.3 and Theorem 1.3.1, that for any H ∈ (0, 1) the process H BtH = (M− 1(0,t) )(s)dWs (1.3.4) R

is a normalized fractional Brownian motion. A little later we shall establish (see Corollary 1.6.11) that any fBm B H can be presented in the form (1.3.4) with a suitable Brownian motion W . H H Remark 1.3.4. It is easy to see that the domain D(M− ) of the operator M− has a form ⎧ 1 Lp (R), H ∈ ( 12 , 1), α = H − 12 , ⎨∪ 1≤p< α H D(M− )= I −α (Lp (R)), H ∈ (0, 12 ), ⎩ p≥1 ± all measurable functions, H = 12 .

1.4 Fractional Brownian Motion with H ∈ ( 12 , 1) on the White Noise Space Consider the probability space of the white noise. Namely, recall that S(R) denotes the Schwartz space of rapidly decreasing infinitely differentiable realvalued functions, and let S (R) be the dual space of S(R), i.e., the space of tempered distributions with weak∗ topology. We consider S (R) as a probability space Ω with the σ-algebra F of Borel sets. According to the Bochner– Minlos theorem, there exists the probability measure P on (Ω, F), such that for any function f ∈ S(R) with the norm f L2 (R) , it holds that 1 2 E exp(if, ω) = exp − f L2 (R) , (1.4.1) 2 where ·, · denotes the dual operation. Note that from (1.4.1), we obtain that

1.4 Fractional Brownian Motion with H ∈ ( 12 , 1) on the White Noise Space

Ef, ω = 0, Ef, ω2 = f 2L2 (R) ,

11

(1.4.2)

where f ∈ S(R), and the duality f, ω can be extended by isometry to f ∈ L2 (R). Note that from (1.4.1)–(1.4.2), it follows that the process Wt := 1[0,t] , ω is a standard Brownian motion. H 1 (R), f1 ∈ L 1−H Now, let H ∈ [ 12 , 1), f1 ∈ L2 (R) and f2 ∈ L H1 (R). Then M+ H M− f2 ∈ L2 (R), therefore, we can consider on L2 (R) the inner product of the form H H f2 )L2 (R) = f1 (x)(M− f2 )(x)dx. (f1 , M− R

By (1.1.1) and (1.3.3), it holds that H H (f, M− f2 )L2 (R) = (M+ f1 , f2 )L2 (R) .

According to (SKM93), denote the spaces Φ(R) = {φ |φ ∈ S(R), φk (0) = 0, k ≥ 0} = {φ |φ ∈ S(R), (φ, tk )L2 (R) = 0, k ≥ 0}. H It was proved in (SKM93) that M± (Φ(R)) ⊂ Φ(R) and that the space Φ(R) is closed in S(R). Now, define two stochastic processes H H (t)(ω) := M± 1(0,t) , ω, B±

t ∈ R.

H H H (t) are Gaussian, EB+ (t) = EB− (t) = 0. For the Then the processes B± covariance function, it holds that H H H H (t)B± (s) = (M± 1(0,t) )(x)(M± 1(0,s) )(x)dx. (1.4.3) EB± R

By considering the sign “−”, we obtain from (1.3.4) that the right-hand side of (1.4.3) coincides with H H EBtH BsH = (M− 1(0,t) )(x)(M− 1(0,s) )(x)dx R

1 = (|t|2H + |s|2H − |t − s|2H ). 2 One obtains the same result if one considers the sign “+”. Therefore, each of H the processes B± (t) has a modification that is a normalized fBm. The process H B− (t) is called a “backward” fBm. It coincides with usual Mandelbrot–van Ness representation andH depends only on the past, i.e. on {Ws , s ∈ (−∞, t)}. H (t) = R (M− 1(0,t) )(s)dWs , where Wt (ω) = 1(0,t) , ω. The process Indeed, B− H (t) is called a “forward” fBm; it admits the representation B+ ∞ (3) H α H B+ (t) = CH (uα − (u − t) )dW = (M+ 1(0,t) )(s)dWs , u + + t

R

12


and depends on future values of W , i.e. on {Ws , s ∈ (t, +∞)}. The case H ∈(0,1/2) can be considered similarly. Also, it is possible to conHk and of fractional Brownian sider the linear combinations of the operators M± motions with different Hurst indices (in what follows we consider only the case Hk ∈ [1/2, 1)): m Hk σ k M± f (x), σk > 0 M± f (x) := k=1

and M B± (t) =

m

Hk σk B± (t) = M± 1(0,t) , ω.

(1.4.4)

k=1 H Clearly, the operators M± are mutually adjoint in the same way as M± . Indeed, (f1 , M− f2 )L2 (R) = (M+ f1 , f2 )L2 (R)

for appropriate functions f1 , f2 .

1.5 Fractional Noise on White Noise Space Let N0 = N ∪ {0} and I be the set of all finite multiindices α = (α1 , . . . , αn ) with αi ∈ N0 . Denote |α| = α1 + · · · + αn , α! := α1 ! · · · αn !. (Of course, in this and similar situations α as a multiindex differs from our α = H − 1/2 but it will not lead to misunderstanding.) Define the Hermite polynomials by 2

hn (x) := (−1)n ex

dn −x2 (e ) dxn

and Hermite functions 2 hn (x) := π −1/4 (n!)−1/2 2−n/2 hn (x)e−x /2 ,

n ≥ 0.

It is well-known that the functions { hn , n ≥ 1} form an orthonormal basis in L2 (R) with Fourier transform eiλx hn (x)dx = (2π)1/2 in hn (x), n ≥ 1. R

Define Hα (ω) :=

n

hαi ( hi , ω),

i=1

the product of Hermite polynomials and consider a random variable F = F (ω) ∈ L2 (Ω) := L2 (S (R), F, P ). Then, according to (HOUZ96, Theorem 2.2.4), F (ω) admits the representation

1.5 Fractional Noise on White Noise Space

F (ω) =

cα Hα (ω),

13

(1.5.1)

α∈I

and 2

F L2 (Ω) =

α! c2α < ∞.

α∈I

Next, we introduce the following dual spaces. (i) F ∈ S if the coefficients from expansion (1.5.1) satisfy 2 α!c2α (2N)kα < ∞ F k = α∈I

m

for any k ≥ 1, where (2N)γ = j=1 (2j)γj , γ = (γ1 , . . . , γm ∈ I). (ii) F ∈ S ∗ if F admits the formal expansion (1.5.1) with finite negative norm 2 α! c2α (2N)−qα < ∞ F −q = α∈I

for at least one we say that F ∈ S−q ). q ∈ N (in this case For F = α cα Hα ∈ S, G = α dα Hα ∈ S ∗ , we define α! cα dα . F, G := α∈I

Taking into account the Parceval identity, we can also define ± f ∈ L2 (R)}, L2M± (R) = {f : M± f ∈ L2 (R)} = {f : M where, according to our notations, g(λ) = R eiλy g(y)dy is the Fourier transform of the function g. The inner product in L2M± (R) is defined by (f, g)M± :=

R

M± f (x)M± g(x)dx = (M± f, M± g)L2 (R) .

−1 Also, define an inverse operator M± in terms of the Fourier transform. −1 For g(x) = M± f (x) ∈ L2 (R), it holds that f (x) = M± g(x), and, according to Theorem 1.1.5, we have the equalities

f(λ) = g(λ)

∞

σk CHk |λ|−αk , (λ)

k=1 (λ)

where CHk = exp

αk πi 2

(3)

sign λ CHk and αk = Hk − 1/2. Hence,

−1 (M ± f )(λ) =

m k=1

σk CHk |λ|−αk (λ)

−1

f(λ).

14


−1 Lemma 1.5.1. The functions e± k := M± hk , k ≥ 1, exist and form an orthonormal basis in L2M± (R). H and σ1 = σ. Consider, Proof. Let, for simplicity, m = 1, so that M± = M± for example, the sign “– ”. Then it holds that

√ − −1 |λ|α hk (λ) = (σCH (λ))−1 ik 2π|λ|α hk (λ), α = H − 1/2. e k (λ) = (σCH (λ)) Therefore, e− k exists and belongs to S(R). The second assertion is evident.

M Now we want to present the linear combination B± (t) of fBms in terms of hk , k ≥ 1.

Lemma 1.5.2. It holds that ∞ t M B± (t) = M∓ hk , ω, hk (x)dx k=1

t ∈ R,

ω ∈ S (R),

(1.5.2)

0

and the series converges in L2 (Ω). Proof. Let ω ∈ S(R). Then, from equality (1.4.4) it follows that M B± (t) = M± 1(0,t) , ω = 1(0,t) , M∓ ω,

and M∓ ω ∈ S(R). Since 1(0,t) ∈ L2M± (R), it admits the expansion 1(0,t) =

∞

± 1(0,t) , e± k M± ek ,

k=1

where the series converges in L2M± (R). Then, 1(0,t) , M∓ ω =

∞

± 1(0,t) , e± k M± ek , M∓ ω,

k=1

and the series converges in L2 (Ω). Furthermore, ∞

± 1(0,t) , e± k M± ek , M∓ ω =

k=1

=

∞

k=1

R

∞ k=1

R

± M± 1(0,t) (x)M± e± k (x)dxM± ek , ω

1(0,t) (x)M∓ hk , ω = hk (x)dx

∞ k=1

t

M∓ hk , ω, hk (x)dx

0

i.e. we obtain (1.5.2) for ω ∈ S(R). Moreover, we can extend (1.5.2) on S (R) since S(R) is dense in S (R) in weak* topology, and this topology generates the weak convergence. Since hk , ω = Hεk (ω),

1.5 Fractional Noise on White Noise Space

15

where εk = (0, . . . , 1, . . . , 0), where 1 is in kth place, we have that 2 2 ! ! ∞ t ∞ t 2 M (x)dx (ε !) = M (x)dx h h = !1(0,t) !L2 ∓ k k ∓ k k=1 0 k=1 0 M∓ m ≤ 2m−1 k=1 σk2 t2Hk < ∞.

Now, we introduce the fractional noise B˙ H as the formal expansion B˙ xH (ω) =

∞

H M+ hk , ω, hk (x)

k=1

and the linear combination of fractional noises as ∞ B˙ xM (ω) = M+ hk , ω. hk (x) k=1

that here we consider only H ∈ [1/2, 1) and that B˙ x (ω) = ∞Recall, ˜ k , ω) is white noise. ˜ k (x)(h h k=1 Lemma 1.5.3. The fractional noise B˙ xH and the linear combination B˙ xM of such noises belong to S ∗ for any x ∈ R. Proof. It is sufficient to consider B˙ H . By using the Fourier transform and Theorem 1.1.5, we obtain that H −ixt −α hk (t)(it) dt ≤ CH,k + CH,k , M+ hk (x) = CH,k e |t|≤1 |t|>1 R where √ CH,k denotes suitable constants. We have that hk (λ) = Ck hk (λ), Ck = k i 2π, and √ −1/12 Ck for |λ| ≤ 2 √k |hk (λ)| ≤ 2 Ce−γλ for |λ| > 2 k where C > 0 and γ > 0 do not depend on λ and k. Therefore, H k −1/12 |t|−α dt M+ hk (x) ≤ C |t|≤1 −1/12 −α −α −γt2 |t| dt + e dt + √ k √ |t|

(1.5.3)

|t|>2 k

1 7/3 − H. So, for q > 7/3, it holds that ||B˙ xH ||2−q < ∞ for any x ∈ R. This completes the proof.

16


1.6 Wiener Integration with Respect to fBm Now we return to an arbitrary complete probability space (Ω, F, P ), and continue the considerations of Sections 1.1–1.3. H Consider the space LH 2 (R) := {f : M− f ∈ L2 (R)} equipped with the H = ||M f || . norm ||f ||LH L2 (R) − 2 (R) Definition 1.6.1. Let f ∈ LH 2 (R). Then the Wiener integral w.r.t. fBm is defined as H f (s)dBsH := (M− f )(s)dWs . (1.6.1) IH (f ) := R

R

BsH

and Ws are connected as in (1.3.4). As a particular case, consider Here, the step function f : R → R given by f (t) =

n

ak 1[tk−1 ,tk ) (t),

k=1

where t0 < t1 < · · · < tn ∈ R and ak ∈ R, 1 ≤ k ≤ n. Then, from the linearity H , we have that of the operator M− IH (f ) =

n

ak

k=1

R

H M− 1[tk−1 ,tk ) (s)dWs =

n

ak (BtHk − BtHk−1 ),

(1.6.2)

k=1

and the latter sum coincides with the usual Riemann–Stieltjes sum. A question arises: in which sense can we consider formula (1.6.1) as the extension of the sum (1.6.2)? Note, that for a step function, it holds that 2

IH (f )L2 (Ω) =

n

ai ak

i,k=1

! H !2 f !L = !M−

2 (R)

R

H H M− 1[tk−1 ,tk ) (x)M− 1[ti−1 ,ti ) (x)dx

(1.6.3) 2α−1

= 2αH R2

f (u)f (v) |u − v|

du dv,

where the last equality holds for H ∈ (1/2, 1) but not for H ∈ (0, 1/2). Nevertheless, for any 0 < H < 1 we have the following: Lemma 1.6.2 ((Ben03a)). For 0 < H < 1, it holds that the linear span of H the set {M− 1(u,v) , u, v ∈ R} is dense in L2 (R). Proof. (i) Let H ∈ (1/2, 1) (for H = 1/2 the assertion is evident). Since (b + x)−α − x−α ∼ Cx−1/2−H as x → ∞, we have that the function (b − −α x)−α + − (−x)+ ∈ L1/H (R). Therefore, for any a < b it holds that g(x) := 1−H α α 1(a,b) (x) ∈ L1/H (R). Therefore, 1(a,b) = M− g ∈ I− (L1/H (R)), and this M− is true also for step functions. Since the class of step functions is dense in α α (L1/H (R)) is dense in L2 (R). Let h ∈ I− (L1/H (R)), L2 (R), it follows that I− H h = M− g, g ∈ L1/H (R). Then there exists the sequence of step functions

1.6 Wiener Integration with Respect to fBm

17

gn → g in L1/H (R). From the Hardy–Littlewood theorem (Theorem 1.1.1) it follows that H gn − h||L2 (R) ≤ C||gn − g||L1/H (R) → 0, n → ∞. ||M− H α 1(u,v) , u, v ∈ R} is dense in I− (L1/H (R)), and So, the linear span of {M− therefore it is dense in L2 (R). (ii) Let H ∈ (0, 1/2). Due to the Parceval identity, it is sufficient to prove H1 that the linear span of the functions M is dense in L (R). According

to Theorem 1.1.5, we have that

−

(a,b)

2

(3) −α H M , − 1(a,b) (x) = CH CH (x)1(a,b) (x)|x|

where CH (x) = exp{iπ sign xα/2}. According to Lemma 1.1.8, for any ϕ ∈ L2 (R) there exists a sequence of step functions ϕn such that (3) (CH )−1 |CH (−x)ϕ(x) −ϕ n (x)|x|−α |2 dx → 0, n → ∞, R

= g(x) for some g ∈ L2 (R). Then, we obtain because (CH )−1 CH (−x)ϕ(x) that (3) |ϕ(x) − CH CH (x) ϕn (x)|x|−α |2 dx R (3) |(CH )−1 CH (−x)ϕ(x) −ϕ n (x)|x|−α |2 dx → 0, n → ∞. = (3)

R

H Remark 1.6.3. Let H ∈ (0, 1/2). Then the operator M− defines an isometric −α H isomorphism from L2 (R) to L2 (R). Indeed, the operator I− is bounded from 2 L2 (R) to L1/H (R), according to Theorem 1.1.1. Let fn be a Cauchy sequence H in LH 2 (R) and ϕn = M− fn . Then

= ϕn − ϕm L2 (R) → 0, m, n → ∞, fn − fm LH 2 (R) H −1 H −1 ) ϕn → (M− ) ϕ =: f in whence ϕn → ϕ ∈ L2 (R), and fn = (M− L1/H (R). We have that

= ϕL2 (R) < ∞, f LH 2 (R) and = ϕn − ϕL2 (R) → 0. fn − f LH 2 (R) It means that LH 2 (R) is complete, i.e., it is a Hilbert space, and equals the closure of the step functions under LH 2 -norm. By (1.6.3), there exists a unique continuous extension of fractional Wiener integrals for the step functions to

18


H the space LH 2 (R). For any f ∈ L2 (R) and the approximating sequence of step functions fn f (s)dBsH = lim fn (s)dBsH in L2 (R). (1.6.4) n→∞

R

R

H Remark 1.6.4. Now, let H ∈ (1/2, 1). Then, the domain of the operator M− coincides with −α α α ) = D(D− ) = ∪p≥1 I− (Lp (R)), D(I−

and, according to Theorem 1.1.1 we can take here only 1 ≤ p < α−1 since −α H LH 2 (R) = {f ∈ D(I− ) : M− f ∈ L2 (R)}.

Note, that α L2 (R) = ∪1≤p 3π | sin u| > 12 for u ∈ (πk + π4 , πk + 3π 4 ), (u + 2 ) 4 , α−1 α−1 α−1 π 3π (u + 2 − x)+ > (2u − x)+ for x > π, u > 4 and (u − x)α−1 > (2u − x) + + for x > 0. Consider ∞ ∞ 2 2 |f (u)|(u − x)α−1 du dx = |u|−p | sin u|(u − x)α−1 du dx R

x

≥ ≥

R

1 2

1 ≥ 4 +

1 4

0

∞

R

∞

x∨ π 4

|u|−p | sin u|(u − x)α−1 du

∞ R

∞

k=0

∞ k=1

πk+ 3π 4

πk+ π 4

∞

0

x

k=0 πk+ π 4

πk− π 4

u−p (u − x)α−1 + du

πk+ 3π 4

πk+ π 4

u+

2 dx 2

u−p (u − x)α−1 + du

dx 2 dx

2 π −p π (u + − x)α−1 + du dx 2 2

1.6 Wiener Integration with Respect to fBm

1 4

≥

+

π

2−2p 4

≥

∞

∞

k=0 ∞

+

πk+ π 4

∞

π

2−2p 8

πk+ 3π 4

k=1

∞

πk− π 4

k=0

∞ πk+ π 4 k=1

πk+ π 4

∞

π

πk− π 4

u−p (2u − x)α−1 + du

2

u−p (2u − x)α−1 + du

πk+ 3π 4

πk+ π 4

23

dx 2 dx

u−p (2u − x)α−1 + du

u−p (2u − x)α−1 + du

2 dx

2 2−2p ∞ ∞ −p = u (2u − x)α−1 + du dx π 8 π 4 ∞ ∞ 2 −2p 2 2α−2p v −p (2v − 1)α−1 dx = + du x π 8 π 4x 2 2−2p ∞ 2α−2p ∞ −p x dx v (2v − 1)α−1 du = ∞ ≥ 1 8 π 2

for H > p.

Now we consider the representation of the Wiener process via fBm, i.e., the relation which is inverse to the relation (1.6.1). 1−H Lemma 1.6.10. Let 0 < H < 1. Then M− 1(0,t) ∈ LH 2 (R) for all t ∈ R, and the underlying Wiener process W admits the representation " Wt = CH M 1−H 1(0,t) (s)dBsH , R

" where C H =

(3) (3) (CH C1−H )−1 .

1−H 1(0,t) ∈ LH Proof. We must check that M− 2 (R). Indeed, (3)

(3)

H− 12

1−H H M− · M− 1(0,t) = CH C1−H I−

1

(I−2

−H

−1 " 1(0,t) ) = (C 1(0,t) ∈ L2 (R). H)

Furthermore, according to Definition 1.6.1, it holds that 1−H 1−H H H " " C (M 1 )(s)dB = C (M− M− 1(0,t) )(s)dWs H H (0,t) s − R R = 1(0,t) (s)dWs = Wt . (1.6.17) R

Corollary 1.6.11. Any fBm B H admits a Mandelbrot–van Ness representation with respect to the Wiener process W from representation (1.6.17).

24


1.7 The Space of Gaussian Variables Generated by fBm. Denote BH = span{BtH , t ∈ R}, where the closure is taken in L2 (Ω). We are interested in the following question: which classes of integrands in the definition of the Wiener integral w.r.t. fBm are isometric to BH or to some of its subspaces? The following theorem from (PT00b) gives the general answer to this question. Theorem 1.7.1. Let I be some class of integrands and let Is ⊂ I be the class of step functions. Under the assumptions (i) I is a space with inner product (f, g)I , f, g ∈ I, (ii) for f, g ∈ Is (f, g)I = EI(f )I(g), (iii) the set Is is dense in I, we have the following: (a) there is an isometry between the space I and a linear subspace of BH which is an extension of the map f → I(f ) for f ∈ Is ; (b) I is isometric to BH if and only if I is complete. Proof. (a) Let f ∈ I. By (iii), there exists fn ∈ Is , such that {fn , n ≥ 1} is a Cauchy sequence in I with norm · I = (·, ·)I . According to (ii), I(fn ) is a Cauchy sequence in L2 (Ω), hence it converges to some r.v. ξ ∈ L2 (Ω). We set I(f ) := ξ. Since I(fn ) ∈ BH and BH is a closed subspace of L2 (Ω), we obtain that I(f ) ∈ BH . So, we can define the map I: I → BH . For any f, g ∈ I it holds that (f, g)I = lim (fn , gn )I = lim EI(fn )I(gn ) = EI(f )I(g). n→∞

n→∞

Moreover, ξ does not depend on the choice of the sequence fn → f in I. Since the map I is linear, we get an isometry between I and some subspace of BH . (b) Since BH is complete as a closed subspace of the complete space L2 (Ω), it follows that I is complete if I is an isometry between I and BH . Conversely, let I be complete. Then, for any η ∈ BH , it holds that η = lim ηn , ηn = I(fn ) ∈ span{BtH , t ∈ R}, fn ∈ Is . So, I(fn ) → η in L2 (Ω). Therefore, from (ii) it follows that fn is a Cauchy sequence in I, and from completeness, fn → f in I, η = I(f ).

Corollary 1.7.2. From Lemma 1.6.2, Remark 1.6.3 and Theorem 1.6.5, we 1 obtain the following: the space I = LH 2 (R) is complete for H ∈ (0, 2 ) and 1 H incomplete for H ∈ ( 2 , 1). Step functions are dense in L2 (R) for any H ∈ 1 (0, 1). Therefore, LH 2 (R) is isometric to BH for H ∈ (0, 2 ) and isometric to 1 a subspace of BH for H ∈ ( 2 , 1).

1.7 The Space of Gaussian Variables Generated by fBm.

25

Theorem 1.7.3. The space (|RH |, ·LH ) is incomplete for H ∈ ( 12 , 1), the 2 (R) space(FH, ·FH ) is incomplete unless H = 12 , and the space (|RH |, ·|RH |,2 ), H ∈ 12 , 1 , is complete. Proof. (i) Consider the space (|RH |, · |RH |,1 ), H ∈ ( 12 , 1). Evidently, if some space is dense in an incomplete space, then it is also incomplete. From Lemma 1.6.9, it follows that |RH | ⊂ LH 2 (R), and from Theorem 1.6.5, we have that (R) is incomplete. So, it is enough to establish that |RH | is dense in LH LH 2 2 (R). H H f ∈ L2 (R). Therefore, there exists a If the function f ∈ L2 (R), then g := M− sequence of step functions {gn , n ≥ 1} ⊂ L2 (R) such that gn − gL2 (R) → 0. H ϕn , where ϕn Evidently, any step function gn can be expressed as gn = M− 1−H is a linear combination of functions M− 1(a,b) , −∞ < a < b < ∞, and ϕn can be determined via Lemma 1.1.3. Note that H H f − ϕn LH = M− f − M− ϕn L2 (R) → 0, 2 (R)

n → ∞, so it is enough to prove that ϕn ∈ |RH |. As will be established in Corollary 1.9.3, there exists some constant C such that ϕn |RH |,2 ≤ Cϕn L 1 (R) , and as mentioned in the proof of Lemma 1.6.2, we have that H

1−H 1(a,b) ∈ L H1 (R) for all −∞ < a < b < ∞. Therefore, (|RH |, · |RH |,1 ) M− is dense in LH 2 (R), and hence incomplete. (ii) Consider the space FH , H = 12 . Let 0 < H < 12 , and let {fn , n ≥ 1} be the sequence of functions −p 1{ n1 0 the kernel lH (t, ·) ∈ BV [0, t] ∩ C[0, t], if H ∈ (0, 12 ). Therefore,

t t lH (t, s)dBsH = BsH dlH (t, s) = BsH (lH )s (t, s)ds 0 0 0 t (5) BsH s−α−1 (t − s)−α−1 (t − 2s)ds, = −αCH t

ItH (lH ) =

0

and this integral is obviously a Gaussian random variable. By using the fact that lH vanishes at the endpoints, we can easily show that EItH (lH )ItH (lH ) = t1−2α for any 0 < t < t :

30


EItH (lH )ItH (lH ) 1 t t 2H (u + s2H − |u − s|2H )(lH )s (t, s)(lH )u (t , u)du ds = 2 0 0 1 t t =− |u − s|2H (lH )s (t, s)(lH )u (t , u)du ds 2 0 0 s 1 t 2H (lH )s (t, s) (s − u) (lH )u (t , u)du ds =− 2 0 0

t 1 t (lH )s (t, s) (u − s)2H (lH )u (t , u)du ds − 2 0 s s t 2α = −H lH (t, s) (s − u) (lH )u (t , u)du ds 0 0

t

t

lH (t, s)

+H 0

(5)

= −αHCH ×

t

(1.8.11)

(u − s)2α (lH )u (t , u)du ds

s t

lH (t, s) 0

|u − s|2α sign(u − s)u−α−1 (t − u)−α−1 (t − 2u)du ds.

0

From (NVV99, Proposition 2.1), we can obtain that

t

|u − s|2α sign(u − s)u−α−1 (t − u)−α−1 (t − 2u)du

0

= −2/α · Γ (1 − α)Γ (1 + α). Therefore, EItH (lH )ItH (lH ) = t1−2α . We can conclude, similarly to part (i), that ItH (lH ) is a martingale w.r.t. its natural filtration, and

t

ItH (lH ) = α

s−α dBs

(1.8.12)

0

for some Wiener process B. Thus, we have proved the following result. Theorem 1.8.1. Let B H be an fBm with H ∈ (0, 1), and let MtH := ItH (lH ) =

t

lH (t, s)dBsH .

(1.8.13)

0

Then there exists a Wiener process B such that (1.8.12) holds. Moreover, σ{BsH , 0 ≤ s ≤ t} = σ{Bs , 0 ≤ s ≤ t}. The inverse relation can be obtained for any H ∈ (0, 1) in the following t way: evidently, for any t > 0 the random variable Yt := 0 s−α dBsH is well

1.8 Representation of fBm via the Wiener Process on a Finite Interval

31

defined. It can be proved similarly (but more easily) as the existence of ItH (lH ). Furthermore, f(s) := s−α ∈ BV [0, t] ∩ C[0, t] for any t > 0 and H ∈ (0, 12 ). Therefore, it holds that t −α H Yt = t Bt + α BsH s−α−1 ds. (1.8.14) 0

Now, let H ∈ ( 12 , 1), f ∈ BV [0, t] ∩ C[0, t] and

t

f (s)dBsH .

It (f ) := 0

Then t

2

E |It (f )| = 2Hα

t

f (u)f (s)|u − s|2α−1 du ds < ∞, 0

0

and it is easy to see, similarly to (1.8.10) that It (f ) =

BtH f (t)

−

t

BsH df (s). 0

Let fε (s) = f(s)1{ε<s 0. Then

t

t

s−α dBsH t H −α H −α = Bt t − Bε ε − α BsH s−α−1 ds. fε (s)dBsH

=

0

ε

0

Note that the trajectories of B H belong to C H−ρ [0, T ] for any 0 < ρ < H, (see Section 1.16). Therefore BεH ε−α → 0, ε → 0 a.s. By similar reasoning, t H −α t Bs s ds → 0 BsH s−α ds, ε → 0 a.s. ε ε 2 for H ∈ Evidently, E 0 f (s)dBsH → 0, ε → 0, and we obtain (1.8.14) ( 12 , 1). But (1.8.14) is an integral equation with respect to BsH , 0 ≤ s ≤ t and its solution has the form t t sα−1 Ys ds = sα dYs . (1.8.15) BtH = tα Yt − α 0

0

Let MtH := ItH (lH ) be the Molchan martingale. Then, for H ∈ (0, 12 ), integration by parts leads to the equality t t (5) (5) MtH = CH (t − s)−α s−α dBsH = −αCH (t − s)−α−1 Ys ds, 0

whence

0

32


t

(t − u)

α

MuH du

0

=

=

(5) −αCH

(5) −αCH B

Ys 0

t

(t − u)α (u − s)−1−α du ds

s

t

(α + 1, −α)

and Yt =

t

(6) CH α

Ys ds, 0

t

(t − u)α dMuH .

(1.8.16)

0

Therefore, t (t − u)α dMuH tα 0 t s t −α sα−1 (s − u)α dMuH ds = mH (t, s)dBs . (1.8.17) (6)

BtH = α CH

0

0

0

Let H ∈ ( 12 , 1). Then, by using Theorem 1.8.1, we obtain that t t (t − u)α dMuH = α (t − u)α−1 MuH du 0 0 t u (5) α−1 (t − u) (u − s)−α s−α dBsH du = CH α 0 0 t t (5) α−1 −α (t − u) (u − s) du s−α dBsH = CH α =

0 s (5) CH αB (α, 1 −

(1.8.18)

α) Yt = (CH )−1 α Yt , (6)

i.e. we have (1.8.16) and obtain (1.8.17). In this case the kernel mH (t, s) can t (6) be simplified to mH (t, s) = αCH s−α s uα (u − s)α−1 du. Remark 1.8.2. It easily follows from (1.8.17) and (1.8.18) that the process B H satisfying (1.8.17) is an fBm. Indeed, it is a Gaussian process with zero mean and covariance t∧s 1 2H H H . mH (t, u)mH (s, u)du = t2H + s2H − |t − s| E Bt Bs = 2 0 Now we state a result of Le Breton (LeB98), see also (KLeBR00), demon t strating how the Wiener integral 0 f (s)dBsH can be presented as an integral with respect to fundamental martingale M H : 1 Theorem 1.8.3. Let f ∈ LH 2 (R) vanish outside [0, T ], where H ∈ ( 2 , 1). Furthermore, let t (7) f KH (t, s) := CH f (u)uα (u − s)α−1 du, s

1.8 Representation of fBm via the Wiener Process on a Finite Interval

(7)

where CH = Then,

2Hα (1−2α)B(1−2α,α)

1/2 .

t

f (s)dBsH 0

33

= 0

t f KH (t, s)dMsH .

Proof. Note that t t (7) f f (KH (t, s))2 dM H s = (CH )2 (1 − 2α) (KH (t, s))2 s−2α ds 0 0 t t (7) f (u)f (v)uα v α = (CH )2 (1 − 2α) 0 0 u∧v × (u − s)α−1 (v − s)α−1 s−2α ds du dv.

(1.8.19)

(1.8.20)

0

Further, Lemma 2.2 from (NVV99) states that

1

tµ−1 (1 − t)ν−1 (c − t)−µ−ν dt = c−ν (c − 1)−µ B(µ, ν)

0

for µ, ν > 0, c > 1. Hence for u < v, µ = 1 − 2α, ν = α we have that u (u − s)α−1 (v − s)α−1 s−2α ds 0

v −α v 2α−1 −1 B 1 − 2α, α = u−1 u u = B 1 − 2α, α (uv)−α (v − u)2α−1 . Moreover, for v < u it holds that v (v − s)α−1 (u − s)α−1 s−2α ds = B 1 − 2α, α (uv)−α (u − v)2α−1 . 0

By substituting these equalities into (1.8.20), we obtain for the integral on the right-hand side that t t (7) 2 (CH ) (1 − 2α)B 1 − 2α, α f (u)f (v)|u − v|2α−1 du dv 0 0 t t f (u)f (v)|u − v|2α−1 du dv = E|IH (f )|2 < ∞. = 2Hα 0

0

Moreover, the system (Is (f ), MsH , 0 ≤ s ≤ T ) is Gaussian and M H is Gaussian martingale. Therefore it follows from Theorem 7.16 (LS01) that t It (f ) = 0

d H E(M I (f )) dMuH , t ∈ [0, T ]. u t dM H u

34


For u ≤ t, we have that E(MuH It (f )) t t (5) = CH 2Hα f (v)s−α (u − s)−α 1{s 0, ν > 0,

0

the second integral equals, for µ = 1 − α and ν = 1 − α, to t (5) CH 2Hα(1 − α)B 1 − α, 1 − α f (v)v α

u

z 1−2H (v − z)α−1 dz dv.

0

u

Therefore, the derivative in u of the right-hand side of (1.8.21) equals C(H)B α, 1 − α f (u) − C(H)(1 − 2α)B 1 − α, 1 − α f (u)B 1 − 2α, α t + C(H)(1 − α)B 1 − α, 1 − α u−2α f (v)v α (v − u)α−1 dv, u (5)

where C(H) = 2HαCH . It is easy to check that (1 − 2α)B 1 − α, 1 − α B 1 − 2α, α = B α, 1 − α . Therefore, dE(MuH It (f )) = C(H)B 1 − α, 1 − α · (1 − 2α)u−2α du t × f (v)v α · (v − u)α−1 dv u t (7) −2α f (v)v α · (v − u)α−1 dv. = CH (1 − 2α)u u

Hence

dE(MuH It (f )) d M H u

f = KH (t, u), and the theorem is proved.

1.9 The Inequalities for the Moments of Wiener Integrals

35

1.9 The Inequalities for the Moments of the Wiener Integrals with Respect to fBm These inequalities were originated with paper (MMV01). Indeed, the Hardy– Littlewood theorem has an immediate consequence, namely, the estimates for the moments of the Wiener integrals with respect to fBm. 1 (R) and there exTheorem 1.9.1. (i) Let H ∈ (0, 12 ). Then LH 2 (R) ⊂ L H H ists a constant CH > 0 such that for any f ∈ L2 (R), it holds that

f L 1 (R) ≤ CH f LH . 2 (R)

(1.9.1)

H

(ii) Let H ∈ ( 12 , 1). Then L H1 (R) ⊂ LH 2 (R) and there exists a constant CH > 0 such that for any f ∈ L H1 (R) it holds that f LH ≤ CH f L 1 (R) . 2 (R)

(1.9.2)

H

−α H Proof. (i) Let f ∈ LH 2 (R). This means that M− f = CH D− f ∈ L2 (R). −α −α Evidently, f = I− D− f and from the Hardy–Littlewood theorem (Theo1 , p = 2 and α = 12 − H), it follows that rem 1.1.1 with q = H (3)

−α −α −α f L 1 (R) = I− D− f L 1 (R) ≤ C2, H1 ,−α D− f L2 (R) = CH f LH . 2 (R) H

H

(ii) We directly apply the Hardy–Littlewood theorem with 1 , α = H − 12 and q = 2: p= H H f LH = M− f L2 (R) ≤ CH f L 1 (R) . 2 (R) H

H Corollary 1.9.2. Let f ∈ LH 2 (R). Then there exists I(f ) = R f (s)dBs and 1 E|I(f )|2 = f 2LH (R) . Therefore, we have for H ∈ (0, 2 ) that E|I(f )|2 ≥ 2

−2 2 CH f 2L 1 (R) and, for H ∈ ( 12 , 1), it holds that E|I(f )|2 ≤ CH f 2L 1 (R) . H

H

Since I(f ) is a Gaussian random variable, we obtain the following inequalities for the moments of the Wiener integrals with respect to fBm: for any r > 0, there exists a constant C(H, r), such that for H ∈ ( 12 , 1) E|I(f )|r ≤ C(H, r)f rL 1 (R) H

and such that for H ∈ (0, 12 ), we have that f rL 1 (R) ≤ C(H, r)E|I(f )|r . H

36


Corollary 1.9.3. Let H ∈ ( 12 , 1) and f ∈ L H1 (R). Then it follows from Theorem 1.9.1, (ii), (1.6.7) and (1.6.14), that f |RH |,2 ≤ Cf L 1 (R) . H

Corollary 1.9.4. Let f ∈ L H1 [a, b] and f = 0 outside (a, b). Then we obtain the following estimates: for any r > 0, there exists a constant C(H, r), such that for H ∈ ( 12 , 1), it holds that r b E f (s)dBsH ≤ C(H, r)f rL 1 [a,b] a H and r b b f (s)dBsH g(s)dBsH ≤ C(H, r)f rL 1 [a,b] grL 1 [a,b] . E a H H a Furthermore, for H ∈ (0, 12 ) the opposite inequality holds: r b f rL 1 [a,b] ≤ C(H, r)E f (s)dBsH H a Remark 1.9.5. Let H ∈ ( 12 , 1) and f ∈ |RH |. Then, from H¨ older inequality, we obtain the estimate 2 2α−1 f |RH |,2 = |f (s)| |f (u)||s − u| du ds R

≤

1 H

R

R

H

|f (s)| ds

1 1−H 1−H

ds

R

|f (u)||s − u|

2α−1

R

du

.

Further, from the Hardy–Littlewood theorem with α = 2H − 1, q = and p = H1 , we obtain that

1 1−H 1−H

ds

R

|f (u)||s − u|

2α−1

R

du

1 1−H

≤ CH f L 1 (R) . H

Therefore, f |RH |,2 ≤ CH f L 1 (R) . H

Remark 1.9.6. Next, we show that the lower inequality in the case H ∈ ( 12 , 1) fails. Indeed, let f (u) = sign u · |u|−p sin u with 12 < p < H. Then according to the proof of Lemma 1.6.9, it holds that f ∈ LH 2 (R). Nevertheless,


∞

1

|f (u)| H du =

0

∞

0

1

| sin u| H du = ∞, p |u| H

since

37

p < 1. H

Therefore, the inclusion L H1 (R) ⊂ LH 2 (R) is proper. Moreover, consider ε−H the function fε (u) = u , 0 ≤ u ≤ 1, 0 < ε < H. Then 2H H C0 1 Γ (1 − H + ε)Γ (2α) ∼ , ε → 0, fε 21 = , fε 2LH (R) = 2 H ε ε Γ (H − ε) ε where C0 = B(1 − H, 2α). Since

1 ε 1

= ε2α and we can let ε tend to 0, it

ε2H

follows that the inequality

f LH ≥ CH f L 1 (R) 2 (R) H

is impossible for H ∈ Remark 1.9.7. It is very easy to check that the function f (u) = u−H ∈ / |RH | for any H ∈ ( 12 , 1). Indeed, ( 12 , 1).

0

T

T

u−H s−H |u − s|2α−1 duds =

0

1

0

1

u−H s−H |u − s|2α−1 du ds,

0

for any T > 0, and this is possible only in the case when these integrals are infinite. Now, let H ∈ (0, 12 ). As mentioned in (SKM93), the domain of the operator −α D− does not coincide with any space Lr (R), 1 ≤ r ≤ +∞. Therefore, the ε−H with ε > α inclusion L H1 (R) ⊂ LH 2 (R) is strict. Moreover, let f (u) = u (note that ε can be negative). By direct computations, we get 1

f L2 (R) = (2ε − 2α)− 2 and

! −α ! !I− f ! = Kε,H (2ε − 2α)−H , L 1 (R) H

where Kε,H = Therefore,

Γ (ε − α) 1 H 1 (2H) , α = H − 2 . Γ (ε − 2α + 2 )

f ! −α L!2 (R) ↑ +∞, ε ↓ (α). !I− f ! L 1 (R) H

−α α Set g = I− f , f = D− g, then f L2 (R) = gLH and 2 (R)

gLH 2 (R)

gL 1 (R) H

↑ +∞,

ε ↓ α.

38


So, we cannot obtain the inverse inequality to (1.9.1). Consider now the upper bound for the moments of I(f ) with H ∈ (0, 12 ). As always, α = H − 12 . Let W22 (R) be the standard Sobolev space W22 (R) = {f : R → R | f L2 (R) + f L2 (R) < ∞}. , Theorem 1.9.8. Let f ∈ C 1 (R) W22 (R) and |f (x)| + |f (x)| ≤ C0 |x|α−1−ε for some ε > 0, as |x| → ∞. Then f ∈ LH 2 (R), and there exists a constant C(H) depending only on H, such that f LH ≤ C(H)f W22 (R) . 2 (R) Proof. Now, we have that 1/2 1/2 (3) H 2 α 2 f LH = |(M f )(t)| dt = C |(I f )(t)| dt (R) − − H 2 R

(3)

R

= CH

1

R

−H

|(D−2

f )(t)|2 dt

1/2

2 d 0 1/2 α = f (x − u)(−u) du dx dx −∞ R 2 ∞ 1/2 (2) α dx = CH f (x + u)u du R

0 2 1/2 √ (2) x+1 α dx ≤ 2CH f (u)(u − x) du x R 2 ∞ 1/2 α dx + f (u)(u − x) du . (2) CH

R

x+1

Further, it holds that 2 x+1 1/2 α dx f (u)(u − x) du x R x+1 1/2 2 ≤ (2H)−1/2 |f (u)| du dx = (2H)−1/2 f L2 (R) , R

and

R

√

2f L2 (R) +

(1.9.4)

x

2 ∞ 1/2 α f (u)(u − x) du dx x+1 R 2 ∞ 1/2 α−1 dx f (x + 1) − α = f (u)(u − x) du ≤

(1.9.3)

x+1

√

2|α| R

∞

x+1

2 1/2 f (u)(u − x)α−1 du dx .

(1.9.5)


39

From the generalized Minkowsky inequality, we obtain that 2 2 ∞ ∞ 1/2 1/2 α−1 α−1 f (u)(u − x) du dx = f (u + x)u du dx x+1 1 R R ∞ 1/2 ≤ uα−1 du |f (u + x)|2 dx ≤ −1/αf 2L2 (R) . R

1

(1.9.6)

The claim follows now immediately from (1.9.3)–(1.9.6).

Now we turn to the case when f = 0 outside some interval [0, T ]. In this case the conditions on f can be much less restrictive. Indeed, then ⎧ 0, x ≥ T, ⎪ ⎨ T α d 1 α (1.9.7) I− f (x) = − Γ (H+ 12 ) dx x f (t)(t − x) dt, x ∈ (0, T ), T ⎪ ⎩− α α−1 f (t)(t − x) dt, x ≤ 0. 1 Γ (H+ ) 0 2

−α Consider some partial cases. Let f ∈ I− (Lp [0, T ]), for some p > 1, i.e. we T 1 can present f as a fractional integral f (x) = Γ (−α) ϕ(t)(t − x)−1−α dt, x ϕ ∈ Lp [0, T ]. Then, according to (SKM93), for any x ∈ (0, T ) it holds that

−

d dx

T

T

f (t)(t − x)α dt = f (x)(T − x)α + α x

(f (x) − f (t))(t − x)α−1 dt. x

(1.9.8) The same equality holds for f ∈ C β [0, T ] for α + β > 0. −α (Lp [0, T ]), From (1.9.7) and (1.9.8) it follows immediately that for f ∈ I− β in particular, for f ∈ C [0, T ] with α + β > 0 we have that E

T

0 (2)

2 T α−1 =α f (t)(t − x) dt dx −∞ 0 T 2 (f (t) − f (x))(t − x)α−1 dt dx. f (x)(T − x)α + α 2

f (t)dBtH

+ (CH )2

T

2

(2) (CH )2

0

0

(1.9.9)

x

Introduce now some classes of functions vanishing outside [0, T ]: H 2 LH [0, T ] = f : [0, T ] → R| | (M f )(x)| dx < ∞ , 2 − R

and DH [0, T ] :=

f : [0, T ] → R

f 2DH [0,T ]

T

2

T

|f (x) − f (t)|(t − x)

α−1

:= 0

x

dt

dx < ∞ .

40


Theorem 1.9.9. (i) The following inclusion holds: for any p > that −α (Lp [0, T ]) DH [0, T ] ⊂ LH EpH [0, T ] := I− 2 [0, T ].

1 H

it holds

Moreover, there exists a constant C(H, p), such that for any f ∈ EpH [0, T ] we have that T 2 1/2 1 H f (t)dBt ≤ C(H, p) f Lp [0,T ] T H− p + f DH [0,T ] . E 0

(1.9.10) (ii) C β [0, T ] ⊂ LH [0, T ] and there exists a constant C(H, β) such that for any 2 f ∈ C β [0, T ] E

2 1/2

T

f (t)dBtH

≤ C(H, β)f C β [0,T ] T H + T H+β .

(1.9.11)

0

2 T Proof. (i) Let f ∈ EpH [0, T ]. Then E 0 f (t)dBtH equals the right-hand side of (1.9.9) and also f ∈ Lp [0, T ]. We have the following estimate:

T

T

If :=

|f (t)| |f (s)| 0

0

T

0

−∞

T

|f (t)| |f (s)|

≤ 0

0

1 ≤ 2(1 − H) ≤

((t − x)(s − x)α−1 )dx ds dt

0

−∞

√ ( st − x)2α−2 dx ds dt 2

T

|f (t)| t

H−1

dt

0

p−1 1 2(1 − H) Hp − 1

2(p−1) p

2

2

f Lp [0,T ] T 2H− p .

(1.9.12)

Therefore, for f ∈ Lp [0, T ] it holds that If < ∞. Then the Fubini theorem implies that the first term on the right-hand side of (1.9.9) equals, up to a constant T

0

T

f (t)f (s) 0

0

−∞

(t − x)(s − x)α−1 dx ds dt,

and can be estimated by the right-hand side of (1.9.12). Moreover, the H¨ older inequality implies that 0

T

2

2

(p − 2)

2

|f (x)| (T − x)2α dx ≤ f Lp [0,T ] T 2H− p

p−2 p

(2αp + p − 2)

From (1.9.12) and (1.9.13) we obtain (1.9.10) with

p−2 p

.

(1.9.13)

1.10 Maximal Inequalities for Wiener Integrals w.r.t. fBm (2)

C(H, p) = CH

(2(1 − H))−1/2 α

41

p − 1 (p−1) p Hp − 1 √ p−2 √ p−2 2p ∨ 2α . + 2 2αp + p − 2

(ii) In this case, 1 If ≤ 2(1 − H)

T

2

T

f (t)t

H−1

≤

dt

0 2

|f (x)| (T − x)2α dx ≤ 0

and T

T 2H 2 f C β [0,T ] T 2H , 2H

2

T

|f (x) − f (t)|(t − x)α−1 dt 0

1 2 f C β [0,T ] T 2H , (1.9.14) 2(1 − H)H 2

dx ≤

x

(1.9.15)

T 2H+2β 2 f C β [0,T ] . 2(α + β)2 (H + β)

Thus, we obtain (1.9.11) with 1 1 1 CH = + ∨ . H 2 (1 − H) 2H 2(α + β)2 (H + β)

1.10 Maximal Inequalities for the Moments of Wiener Integrals with Respect to fBm For any fixed T > 0, denote ζT∗ = sup0≤t≤T |ζt |, where ζt is any function on [0, T ]. If B H = {BtH , t ≥ 0} is a fractional Brownian motion, then from its self-similar properties we obtain that E((B H )∗T )p = C(H, p)T pH , where C(H, p) = E((B H )∗1 )p . (It is an interesting and open problem how to compute this maximal moment.) Now, let f ∈ LH 2 (R). We try to find possible bounds t for the process It = It (f ) := 0 f (s)dBsH both on random and nonrandom intervals. Denote IT∗ p := (E(IT∗ )p )1/p . (i) Upper bound on nonrandom interval, H ∈ ( 12 , 1). Note that the process It (f ) is Gaussian, therefore it admits entropy maximal estimates. In this context, suppose that f ∈ |RH | and consider on [0, T ] the semi-metric ρI generated by the process I, i.e. ρ2I (s, t)

2 t H := E(It − Is ) = E f (u)dBu . 2

s

For any ε > 0 denote by N ([0, T ], ε) the metric ε-capacity of ([0, T ], ρ), or the minimal number of points in the ε-net of the interval [0, T ] in the semimetric ρI , i.e. the minimal number of centers of closed ε-balls covering [0, T ].

42


Also, let H([0, T ], ε) := log N ([0, T ], ε) be the of this interval εmetric ε-entropy 1 2 in the semi-metric ρI , and let D(T, ε) = 0 H([0, T ], u) du be the Dudley integral. Lemma 1.10.1. Let ρ(s, t) be some semi-metric on [0, T ] and let ϕ(x), x > 0, be a continuous increasing function, such that ϕ(0) = 0. Also, let g be a function with g(v) ≥ 0, g ∈ L1 [0, T ], such that for any 0 ≤ s < t ≤ T , it holds t that ϕ(ρ(s, t)) ≤ s g(v)dv. Then T N ([0, T ], u) ≤ 1 +

0

g(v)dv . ϕ(2u)

Proof. Consider 0 = s0 < s1 < . . . < sM < T , where |sk+1 −sk | = 2u, 0 ≤ k ≤ M − 1, |T − SM | ≤ 2u. Such a partition exists, s because our condition ensures the continuity of ρ(s, t). Evidently, ϕ(2u) ≤ skk+1 g(v)dv, 0 ≤ k ≤ M − 1, and N ([0, T ], u) ≤ M + 1. So, M ϕ(2u) ≤

M −1 sk+1

T 0

sM

g(v)dv =

sk

k=0

i.e. M ≤

g(u)du ≤

T

g(v)dv, 0

a

g(v)dv · (ϕ(2u))−1 .

Lemma 1.10.2. The Dudley integral admits the estimate D(T, ε) ≤

ε

* log(1 + u

0

1 −H

H C

T

+ 12 |f (v)| dv) du, 1 H

0

H is some constant. where C Proof. According to (1.9.2) and Corollary 1.9.2, it holds that 2 t f (u)dBuH ≤ C(H, 2)f 2L 1 [s,t] . E s

H

t 1 1 If we choose ϕ(u) = u H and g(v) = |f (v)| H , then ϕ(ρI (s, t)) ≤ s g(v)dv. We obtain from Lemma 1.10.1, that for any u > 0 the metric u-entropy of the in 1 1 1 T 1 terval [0, T ] does not exceed log 1 + u− H (C(H, 2)) 2H · 2− H 0 |f (v)| H dv . 1 H = 2− H1 (C(H, 2)) 2H From here the claim follows with C .

Theorem 1.10.3. For any p > 0, there exists a constant Cp (H) such that IT∗ p ≤ Cp (H)f L 1 [0,T ] . H

1.10 Maximal Inequalities for Wiener Integrals w.r.t. fBm

43

Proof. Denote σ 2 := sup EIt2 . 0≤t≤T

Then √ according to (Lif95, Theorem 1, p. 141) and its corollary, for any r > 4 2D(T, σ2 ), we have the inequality

√ σ r − 4 2D(T, ) 2 P {IT∗ > r} ≤ 2 1 − Φ , (1.10.1) σ where Φ(x) =

x

√1 2π

−∞

e

−y 2 2

dy. Since

E(IT∗ )p ≤ p

∞

xp−1 (1 − F (x))dx,

0

where F (x) = P {IT∗ < x}, we obtain from (1.10.1) that for D = D(T, σ2 ) it holds that 4√2D ∗ p xp−1 (1 − F (x))dx E(IT ) ≤ p 0 ∞ √ + p √ xp−1 (1 − F (x))dx ≤ (4 2D)p 4 2D ∞ x √ dx + 2p (x + 4 2D)p−1 1 − Φ (1.10.2) σ 0 ∞ x √ dx ≤ (4 2D)p + p2p xp−1 1 − Φ σ 0 ∞ x √ 1−Φ dx + p2p (4 2D)p−1 σ 0 √ √ ≤ (4 2D)p + p2p σ p C1 (p) + 2p p(4 2D)p−1 σC1 (1), ∞ where C1 (p) = 0 xp−1 (1 − Φ(x)) dx. Now we estimate D = D(T, σ2 ). From Lemma 1.10.2 and Corollary 1.9.4,

σ 2

D≤

.

log 1 + u

0

H )H ≤ H(C

H )H H C H = (C

T

H C

T

1 H

/ 12

|f (v)| dv

du

0 ∞

1

z2 log 2

H H = C where C

1 −H

exp zdz , (exp z − 1)H+1

(1.10.3)

1

|f (v)| H dv. Therefore, D ≤ C H f L 1 [0,T ] , where

0 ∞ log 2

1

H

1

zdz z 2 (expexp . Evidently, σ ≤ (C(H, 2)) 2 f L 1 [0,T ] . z−1)H+1

By substituting these two estimates into (1.10.2), we obtain the proof.

H

44


(ii) Lower bound on nonrandom interval, H ∈ ( 12 , 1). According to Remark 1.9.6, the reverse inequality of (1.9.2) fails. Therefore, we obtain the lower bound under stronger assumptions. We suppose here that f = f (s) > 0 on 1 , gT∗ = ess sup0≤s≤T g(s) and assume that gT∗ < ∞. [0,T]. Denote g(t) = f (t) Theorem 1.10.4. For any p > 0, we have an estimate IT∗ p ≥ cp (H)T H (gT∗ )−1 . Proof. According to the lower bound obtained by Sudakov (Lif95, Theorem 5, p. 152), for any ε > 0 it holds that p

E(IT∗ )p ≥ (EIT∗ ) ≥ Cp H([0, T ], ε) 2 εp , p

1

where H([0, T ], ε) = log N ([0, T ], ε). Evidently, N ([0, T ], ε) ≥ 1 ∨ T (2gT∗ ε)− H . Therefore, 1 H([0, T ], ε) ≥ log 1 ∨ T (2gT∗ ε)− H . Indeed, take an arbitrary partition π = {0 = s0 < s1 < · · · < sn = T } such 2 12 sk H that E sk−1 f (s)dBs ≤ 2ε. Then 2 ! !2 sk H E sk−1 f (s)dBsH = !M− (1(sk−1 ,sk ) f )!L (R) ≥ (gT∗ )−2 E|Bsk − Bsk−1 |2 2 = (gT∗ )−2 (sk − sk−1 )2H , 1

1

so, (gT∗ )− H (sk − sk−1 ) ≤ (2ε) H . Hence N ([0, T ], ε) ≥ 1 ∨ T (2gT∗ ε)−1/H . 1 −1 2 · ε, with ε > 0, it holds For the function ϕ(ε) = log (1 ∨ T (2gT∗ ε)) H that 1 1 max∗ ϕ(ε) = e− 2 T H (2gT∗ )−1 , 2 ε 0, there exists a constant C(H, p) such that IT∗ p ≥ C(H, p) f L 1 [0,T ] . H

(iv) Upper bound on nonrandom interval, H ∈ (0, 12 ).


45

Theorem 1.10.6. Let f : [0, T ] → R, f ∈ Lp [0, T ] ∩ DpH [0, T ] for some T T 1 , where DpH [0, T ] = {f : [0, T ] → R | 0 ( x ϕ(x, t)dt)p dx < ∞} and p> H (t)−f (x)| ϕ(x, t) = |f(t−x) 1−α · 1{0<x −α. Then f Lp [0,T ] ≤ f C β [0,T ] T 1/p , T T p p1 1 ϕ(x, t)dt dx ≤ C f Cβ[0,T ] T α+β+ p , 0

x

and IT∗ p ≤ C2 (H, p) f C β [0,T ] (T H + T H + β), where It∗ := sup0≤s≤t |Is |. 2. Similarly to Theorem 1.10.6, we can suppose that f ∈ Lp [0, T ]∩DpH [0, T ] and obtain the estimate for IT∗ r , r > 0. Indeed, the estimate (1.10.5) holds for any r > 0, and we obtain from (1.10.6) and (1.10.7) that √ E(IT∗ )r ≤ (4 2Cp p−1 G1p (0, T, f ))r + r2r (C(H, p)G1p (0, T, f ))r C1 (r) + r2r (Cp p−1 G1p (0, T, f ))r−1 · C1 (1) · C(H, p)G1p (0, T, f ) ≤ (C(H, p, r))r (G1p (0, T, f ))r .

√ From here IT∗ r ≤ C(H, p, r)G1p (0, T, f ), where C(H, p, r) ≤ 4 2Cp · p−1 + r−1

1/2−1/r

1/r + r1/r 2 · Cp r · p− r (C1 (1)C(H, p))1/r . 2C(H, p) · 2 π1/2 r · (Γ ( r+1 2 )) 1/r Evidently, C(H, p, r) can be estimated as C(H, p, r) ≤ C(H, p)(Γ ( r+1 for 2 )) some constant C(H, p) depending only on H and p. r−1

We continue now with random intervals. Let F = {Ft , t ≥ 0} be the natural filtration generated by the fBm B H and let τ be any stopping time with respect to this filtration, i.e., the event {τ ≤ t} ∈ Ft for any t ≥ 0. (v) Upper bound on random interval, H ∈ ( 12 , 1). Let f be a measurable positive function on R, α = H − 12 . Theorem 1.10.8. Let the function sα f (s) be nondecreasing on R. Then, for any p > 0, there exists a constant C(H, p) such that for any stopping time τ we have that pH

2α

Iτ∗ p ≤ C(H, p)(E((f (τ )) 2α τ pH )) pH (E(τ pH ))

1−H pH

.

Remark 1.10.9. For a bounded positive function f with f (x) ≤ f ∗ < ∞, x ∈ R, we obtain that Iτ∗ p ≤ C(p, H)f ∗ (Eτ pH )1/p . In particular, for f (s) ≡ 1, we obtain the upper bound from (NV98, Theorem 1.2).


t

t

47

t

Proof. Denote Yt = 0 s−α dBsH . Then BtH = 0 sα dYs and It = 0 sα f (s)dYs . Integration by parts gives the following upper bound for It∗ : t It∗ = sup |Is | = sup tα f (t)Yt − Ys d(sα f (s)) ≤ 2f (t)tα Yt∗ . 0≤s≤t

0≤s≤t

0

Now we use the representation (1.8.16) for Yt , t t α H Yt = CH (t − s) dMs = αCH (t − s)α−1 MsH ds, 0

(1.10.8)

0

H tα (M H )∗ . Here C H = C (6) α whence Yt∗ ≤ C = (1 − α)−1/2 . t H , α From these two estimates, we obtain for any t > 0 that H t2α f (t)(MtH )∗ , and for the random stopping time τ it holds that It∗ ≤ 2C H τ 2α f (τ )(MτH )∗ . Iτ∗ ≤ 2C Therefore, for any p > 0 H )p E(τ 2αp (f (τ ))p ((M H )∗ )p ). E(Iτ∗ )p ≤ (2C τ

(1.10.9)

From the H¨ older inequality it follows that 1

1

E(τ 2αp (f (τ ))p ((MτH )∗ )p ) ≤ (E(τ 2αpq (f (τ ))pq ) q (E((MτH )∗ )pr ) r , (1.10.10) H H where q = 2α > 1 and r = 1−H . From the Burkholder–Davis–Gundy inequalities for martingales, it follows that for any p > 0 there exist constants cp , Cp > 0, such that p

p

cp EM H τ2 ≤ E((MτH )∗ )p ≤ Cp EM H τ2 . But M H t = t1−2α , and E((MτH )∗ )p ≤ Cp Eτ p(1−H) . Therefore,

E((MτH )∗ )pr ≤ Cpr Eτ pH ,

(1.10.11)

and the proof follows from (1.10.8)–(1.10.11) with 1

p H )p Cpr , r= C(H, p) = (2C

H . 1−H

(vi) Lower bound on random interval, H ∈ ( 12 , 1). Let f be, as before, a pos1 and gT∗ = sup0≤s≤T g(s). itive measurable function, T > 0 be fixed, g(t) = f (t) In order to proceed, we need the following auxiliary result from (NV98). Denote ξt := t2α |MtH |.

48


Lemma 1.10.10. For any p > 0 there exists a constant cp > 0, such that for any stopping time τ , it holds that E(ξτ∗ )p ≥ cp Eτ pH .

(1.10.12)

t Proof. Let p = 2. From the Itˆ o formula we obtain that ξt2 = 0 (s2α + t 4αs4α−1 (MsH )2 )ds + 2 0 s4α MsH dMsH . Therefore, for any bounded stopping time τ , it holds that τ 2 s2α ds = (2H)−1 Eτ 2H . (1.10.13) Eξτ ≥ E 0

For arbitrary stopping time τ , we obtain by applying (1.10.13) to bounded stopping time τ ∧ n, that Eξτ2∧n ≥ (2H)−1 E(τ ∧ n)2H , and the Fatou lemma gives (1.10.12) with p = 2. Let p < 2. Inequality (1.10.12) with p = 2 means that continuous and hence predictable process (ξt∗ )2 dominates the (nonrandom) process ϕ(t) = t2H . Then, from the Lenglart inequality, for p < 2, we obtain that E(ξτ∗ )p ≥ cp Eτ pH −p

with cp = (2H)2−p(4−p) (DM82, VI, p. 113). Finally, let p > 2. Set k > 0, δ > 0 and define a process with positive values by ηt = δ + kt2H + ξt2 . Then, from the Itˆ o formula, for p > 2, we obtain that t p p p p 2 −1 2 2 ηs ((1 + 2kH)s2α + 4αs2H−3 (MsH )2 ) ηt = δ + 2 0 t p p 1 −2 pη 2 −1 s4α MsH dMsH . + p(p − 2)ηs2 s6α (MsH )2 ds + 2 0 Therefore, for any bounded stopping time τ τ p p p −1 Eητ2 ≥ E ηs2 (1 + 2kH)s2α ds 2 0 τ p p p k 2 −1 s2H( 2 −1) s2α ds · (1 + 2kH) ≥ E 2 0

(1.10.14)

p

≥

k 2 −1 (1 + 2kH)Eτ pH . 2H

From the Fatou lemma, applied, for any stopping time τ , to τ ∧ n, we obtain (1.10.14) for τ ∧ n and for δ = 0.


So,

49

p

p

E(kτ 2H + ξτ2 ) 2 ≥

k 2 −1 (1 + 2kH) pH Eτ . 2H

From the inequality p

p

p

(kτ 2H + ξτ2 ) 2 ≤ 2 2 −1 (k 2 τ pH + ξτp ), we obtain that p −1 p k2 p (1 + 2kH) − k 2 Eτ pH . Eξτp ≥ 21− 2 2H

This means that (1.10.12) holds with 1 + 2H) p p ( cp = k 2 21− 2 k −1 >0 2H

for k <

1 . p H(2 2 −2)

Now we are in a position to establish the lower bound on a random interval for H ∈ ( 12 , 1). Theorem 1.10.11. Let, for any t ∈ [0, T ], the function ϕ(s) := s−α (t−s)−α g(s) be nondecreasing on [0, t]. Then, for any p > 0, there exists a constant c(H, p) > 0, such that for any stopping time τ ≤ T it holds that Iτ∗ p ≥ c(H, p)(gT∗ )−1 (Eτ pH )1/p . Remark 1.10.12. Either of the following conditions (a) and (b) is sufficient for Theorem 1.10.11: (a) g ∈ C 1 [0, T ] and for any s ∈ (0, T ), it holds that g (s) ≥ g(s)( αs − T α−s ). (b) The function g(s)s−α is nondecreasing on [0, T ] (or the function f (s)sα is nonincreasing on [0, T]; compare with the condition of Theorem 1.10.8). Remark 1.10.13. The class of functions satisfying the condition of Theorem 1.10.11 is nonempty. For example, f (s) = s−γ e−βs with γ ≥ α and β ≥ 0 belongs to this class. (In this case assumption (b) is satisfied.) Proof. Let 0 < a < b < 1. Then the martingale MtH can be represented as MtH =

at

0

bt

lH (t, s)dBsH +

t

lH (t, s)dBsH + at

lH (t, s)dBsH bt

bt

lH (t, s)g(s)dIs + MtH (1 − b).

:= MtH (a) +

(1.10.15)

at

The middle term can be integrated by parts, and we obtain from the condition of the theorem that

50


bt lH (t, s)g(s)dIs at bt = lH (t, bt)g(bt)Ibt − lH (t, at)g(at)Iat − Is d(lH (t, s)g(s)) at (5) ≤ CH It∗ gt∗ t−2α (1 − b)−α b−α + (1 − a)−α a−α .

(1.10.16)

Therefore, the process ξt = t2α |MtH | can be estimated as ξt ≤ t2α |MtH (a)+ (5) MtH (1 − b)| + CH It∗ gt∗ , where CH = 2CH (((1 − b)−α )b−α + (1 − a)−α a−α ). Now we use Lemma 1.10.10 and obtain Cp Eτ pH ≤ E(ξτ )p p−1 ≤ 2p−1 Eτ 2pα |MτH (a) + MτH (1 − b)|p + 2p−1 CH E(Iτ∗ )p (gτ∗ )p .

(1.10.17)

Further, from (1.10.8) we have that |MtH (a)| ≤ CH |(t(1 − a))−α Yat − (5)

|MτH (a)| ≤ CH

Ys d(t − s)−α |

0

(5) ∗ (5) ≤ CH Yat · 2(t(1 − a))−α ≤ 2CH C H

Hence

at

aα H ∗ (Mat ) . (1 − a)α

aα (MτH )∗ , (1 − a)α

(1.10.18)

(5) where CH = 2CH C H. In order to estimate MtH (1 − b), note at first that for fixed t, the process H sH := BtH − Bt−s B , 0 ≤ s ≤ t, is a fractional Brownian motion with Hurst index H. Therefore, (5)

MtH (1 − b) = CH

t

tb

(t − s)−α s−α dBsH = CH

(5)

t(1−b)

H , u−α (t − u)−α dB u

0

and similarly as in the above estimates (1.10.15), we obtain that α 1−b H |Mτ (1 − b)| ≤ CH (MτH )∗ , b

(1.10.19)

where M H is the Molchan martingale the symmetry of the kernel tfor BH . But t H H = 0 lH (t, s)dBsH = MtH . lH (t, s) leads to the equality Mt = 0 lH (t, s)dBt−s Hence, 1−b α ) (MτH )∗ . |MτH (1 − b)| ≤ CH ( (1.10.20) b From (1.10.7), (1.10.18), (1.10.20), (1.10.10) and (1.10.11) with f ≡ 1 we obtain that

1.10 Maximal Inequalities for Wiener Integrals w.r.t. fBm p−1 2p−1 CH E(Iτ∗ )p · (gτ∗ )p

≥ Cp Eτ pH − 2p−1 CH Eτ 2pα · (Mτ∗ )p

51

α aα (1 − b) pH + · cp . Eτ (1 − a)α bα

By choosing a sufficiently small and b close to 1, we obtain that E(Iτ∗ )p ≥ (gT∗ )−p Eτ pH · Cp,H , where 1−p

Cp,H = 2

1−p CH

* p−1 Cp − 2 CH cp

aα (1 − b)α + (1 − a)α b

+ > 0.

(vii) Upper and lower bounds for power functions and H ∈ ( 12 , 1). The function f (s) ≡ 1 does not satisfy the condition of Theorem 1.10.11. To cover this case, we consider the power functions f (s) = sγ , γ > −2α, and obtain a better result than in Theorems 1.10.11 and 1.10.6: Theorem 1.10.14. Let f (s) = sγ with γ > −2α. Then, for any p > 0, there exist constants cp,H and Cp,H , such that for any stopping time τ it holds that cp,H (Eτ p(H+γ) )1/p ≤ Iτ∗ p ≤ Cp,H (Eτ p(H+γ) )1/p . Proof. Consider the upper bound. Now inequality (1.10.4) has the form E(Iτ∗ )p ≤ (2CH )p E(τ (2α+γ)p Mτ∗ )p . (6)

By applying H¨ older’s inequality with q = H+γ 1−H ,

1+2α+2γ 4α+2γ

> 1 and r =

1+2α+2γ 1−2α

=

and the Burkholder–Davis–Gundy inequalities, we obtain that E(Iτ∗ )p ≤ (2CH )p (Eτ (6)

1+2α+2γ 2

p

1

1

) q (E(Mτ∗ )pr ) r

1

(6)

1

≤ (2CH )p Cp,H (Eτ (H+γ)p ) q (Eτ (H+γ)p ) r = Cp,H Eτ (H+γ)p . Consider the lower bound. We use expansion (1.10.15) and estimate its middle term similarly to the first part of (1.10.16) with g(s) = s−γ : bt lH (t, s)g(s)dIs at bt = |lH (t, bt)g(bt)Ibt | + |lH (t, at)g(at)Iat | + Is d(PH (t, s)g(s)) at ≤ CH b−γ−α (1 − b)−α t−2α−γ It∗ + CH a−γ−α (1 − a)−α It∗ · t−2α−γ bt (5) + CH It∗ |d((t − s)−α s−α−γ )|. (5)

(5)

at

52


The function ϕ(s) := (t − s)−α s−α−γ has the following derivative on (at, bt) : ϕ (s) = s−α−γ−1 (t − s)−α−1 ((γ + 2α)s − (γ + α)t). For γ > −α, on the interval [0, t], the function ϕ(s) has an extremal point γ+α smax = ρt, where ρ = γ+2α , and for −2α < γ < −α, no extremal point exists. Therefore, the variation of ϕ(s) on the interval [at, bt] can be estimated as

bt

|d((t − s)−α s−α−γ )| ≤ t−2α−γ b−γ−α (1 − b)−α + 2|ρ|−γ−α (1 − ρ)−α + a−γ−α (1 − a)−α . at

From here,

bt −γ lH (t, s)s dIs ≤ C(a, b, H, γ)t−2α−γ It∗ , at

where (5) C(a, b, H, γ) = 2CH b−γ−α (1 − b)−α + a−γ−α (1 − a)−α + |ρ|−γ−α (1 − ρ)−α . Therefore, for the process ξt := t2α+γ |MtH |, we have that ξt ≤ t2α+γ |MtH (a) + MtH (1 − b)| + C(a, b, H, γ)It∗ , whence for any stopping time τ and p > 0, it holds that E(ξτ )p ≤ (C(a, b, H, γ))p E(Iτ∗ )p +E(τ 2α+γ |MτH (a)+MτH (1−b)|)p . (1.10.21) Similarly to Lemma 1.10.10, we can establish the following bound for ξτ : E(ξτ )p ≥ cp Eτ p(H+γ) . Further, we apply (1.10.11), the bounds (1.10.15) and (1.10.17), and H¨ older’s inequality with q = 1+2α+2γ > 1 and r = 1+2α+2γ > 1, where 4α+2γ 1−2α 1 1 2α+γ H + = 1, and obtain the bounds of the pth moment of τ M τ (a) and q r τ 2α+γ MτH (1 − b): aαp E((τ )2α+γ (MτH )∗ )p (1 − a)αp H (Eτ (2α+γ)pq ) q1 (E((MτH )∗ )pr ) r1 ≤ C H Eτ p(H+γ) , ≤C

E(τ 2α+γ MτH (a))p ≤ (CH )p

H = (CH )p where C

α 1−α

(1.10.22)

αp . Similarly,

H Eτ p(H+γ) , E(τ 2α+γ MτH (1 − b))p ≤ C

(1.10.23)


H = (CH )p C

1−b b

53

αp .

From (1.10.21)–(1.10.23) E(Iτ∗ )p ≥ cp,H Eτ p(H+γ) , where

⎛ cp,H = ⎝

cp − (CH )p

αp a 1−a

+

1−b αp ⎞ p1 b

⎠ >0

(C(a, b, H, γ))p

for sufficiently small a and 1 − b.

(viii) Lower bound on random interval, H ∈ (0, 12 ). Here, we consider only power functions f (s) = sγ , s > 0. According to t (1.8.6), the integral 0 sγ dBsH exists, if t

t γ γ

0

u s and

0

0

t

t

−∞

uγ−1 sγ−1 0

(u − x)α−1 (s − x)α−1 dx du ds < ∞

0

s∧u

(u − x)α (s − x)α dx du ds < ∞.

0

If we choose γ > −H, then both of these inequalities hold. Theorem 1.10.15. Let H ∈ (0, 12 ) and f (s) = sγ with γ ∈ (−H, −α). Then, for any p > 0, there exists a constant c(H, p) such that It∗ p ≥ c(H, p)(Eτ p(H+γ) )1/p . Proof. We estimate the Molchan martingale from above:

(MtH )∗ = CH

(5)

t

s−α−γ (t − s)−α dIs

∗

0

≤ CH It∗ (5)

t

|d(s−α−γ (t − s)−α )|.

0

The last integral exists when −α − γ > −1 or γ < 1 − α. As before, the derivative of the function ϕ(s) = s−α−γ (t − s)−α , s ∈ (0, t), equals ϕ (s) = s−α−γ−1 (t − s)−α−1 (γ + 2α)s − (γ + α)t . So, for γ ∈ (−H, −α), the function ϕ(s) has the unique extremal point t γ+α s = γ+2α t, and 0 |d(s−α−γ (t − s)−α )| ≤ Cα t−2α−γ , where Cα :=

α γ + 2α

−α

γ+α γ + 2α

−α−γ .

Hence, for any stopping time τ and any p > 0 it holds that

54


((MτH )∗ )p ≤ (CH Cα )p(Iτ∗ )pτ (−2α−γ)p. (5)

Further, from the Burkholder–Davis–Gundy inequalities we obtain that pEτ p(1−H) E(Mτ∗ )p ≥ C .

Hence,

q1 r1 pEτ p(1−H) C ≤ C(H, p)(E(Iτ∗ )pq ) · (Eτ (−2α−γ)pr ) ,

where C(H, p) = (CH Cα )p. 1−H 1−H Now, we choose r = 1−2H−γ > 1, q = H+γ > 1 and p = obtain for C p q c(H, p) = C(H, p) (5)

p(H+γ) 1−H ,

and

that c(H, p)Eτ p(H+γ) ≤ E(Iτ∗ )p .

1.11 The Conditions of Continuity of Wiener Integrals with Respect to fBm Consider the case H ∈ ( 12 , 1). Let f ∈ L H1 [0, t], t ∈ [0, T ]. Then in particular, t the integral It (f ) = 0 f (s)dBsH exists on [0, T ] and E(It (f ))2 = ||f ||2LH [0,t] ≤ 2 CH ||f ||2L 1 [0,t] . According to (Lif95), a sufficient condition for the continuity H

of separable of It (f ) on [0, T ] is the finiteness of the Dudley ε modification 1 integral 0 H([0, T ], u) 2 du. But in our case, from (1.10.3) with ε instead of σ 2 , it follows that

ε

1 2

H([0, T ], u) du ≤ 0

ε

log 1 + u 0

≤ 0

ε

1 −H

H C

T

1 H

|f (u)| du

12 du

0

t 12 1 1 H u− 2H du · C |f (u)| H du H1 . So, for such f , a separable modification of It (f ) is continuous on [0, T ]. ( 12 , 1).

1.12 The Estimates of Moments of Solutions of SDEs with fBm

55

1.12 The Estimates of Moments of the Solution of Simple Stochastic Differential Equations Involving fBm (i) Let H ∈ ( 12 , 1) and Ft = σ BsH , 0 ≤ s ≤ t . Consider a stochastic differential equation of the form dXt = b(t, Xt )dt + f (t)dBtH , t ≥ 0.

(1.12.1)

Here, X0 is F0 -measurable random variable, E|X0 |p0 < ∞ for some p0 > 1 and b(t, x) : R+ × R −→ R is a measurable Lipschitz function, i.e. |a(t, x) − a(t, y)| ≤ C|x − y|

(1.12.2)

with some constant C. Furthermore, b is of linear growth, meaning that |b(t, x)| ≤ C(1 + |x|)

(1.12.3)

f ∈ L H1 [0, T ].

(1.12.4)

and Theorem 1.12.1. Let b satisfy (1.12.2), (1.12.3) and f satisfy (1.12.4). Then equation (1.12.1) has a unique solution. Proof. We establish now that for any p ≤ p0 the map t (AX)t := X0 + b(s, Xs )ds + It (f ) 0

is a contraction in the space Sp := ξ(t, ω), t ∈ [0, Tp ] ξ(t, ·) is Ft -measurable, with the norm

sup E|ξt |p < ∞ , t∈[0,Tp ]

1

ξSp := sup (E|ξt |p ) p , t∈[0,Tp ]

where Tp is a number such that Tp < C −1 . Indeed, from (1.12.2)–(1.12.4) it follows that t E|(AX)t |p ≤ 3p E|X0 |p + E|It (f )|p + 2p tp−1 C 1 + E|Xs |p ds . 0

This means that AX ∈ Sp if X ∈ Sp . Further, for t ≤ Tp t E|(AX)t − (AY )t |p ≤ E| (b(s, Xs ) − b(s, Ys )ds|p 0 t t p ≤ C pE |Xs − Ys |ds ≤ C p Tpp−1 E |Xs − Ys |p ds, 0

0

56


i.e., AX − AY Sp ≤ L X − Y Sp , where L = C p Tpp < 1. Therefore, on the interval [0, Tp ] equation (1.12.1) has unique solution. If we obtain this solution Xt by the method of successive approximations, and the initial process is some (0) continuous process Xs ∈ Sp , then by the continuity of the process I(b) and t the equicontinuity of the integral 0 b(s, ·)ds, the solution Xt is continuous on [0, Tp ]. The proof of the theorem is obtained by extension of the solution from [0, kTp ] to [0, (k + 1)Tp ] via the relation t b(s, Xs )ds + (It − IkTp ), (1.12.5) Xt = XkTp + kTp

where k ∈ N and XkTp is the solution of the “previous” equation taken at the point t = kTp . Existence, uniqueness and continuity of the solution of (1.12.5) is established similarly to previous estimates.

Now we establish the upper bound for the solution of equation (1.12.1) on a random interval. Theorem 1.12.2. Let the functions b and f satisfy the conditions of Theorem 1.12.1, E|X0 |p < ∞ for any p > 0 and the function sα f (s) be nondecreasing on R. Then, (a) for any T > 0, p > 0 and stopping time τ ∈ [0, T ], we have the estimate E(Xτ∗ )p ≤ 4p e4

p

C p T p−1 pH

(E|X0 |p + C p Eτ p 2α

+ (C(H, p))p (E((f (τ )) 2α τ pH )) H (Eτ pH )

1−H H

),

where a constant C(H, p) appeared in Theorem 1.10.8. (b) If, in addition, the function f is bounded, i.e. |f (x)| ≤ f ∗ < ∞, then p p p−1 E|X0 |p + C p Eτ p + (C(H, p))p (b∗ )p Eτ pH . E(Xτ∗ )p ≤ 4p e4 C T Proof. Let τ ∈ [0, T ] and τn = τ ∧ inf {t > 0 : |Xt | ≥ n}. Then τn ∗ p Xs∗ ds + Iτ∗n (f ))p (Xτn ) ≤ (|X0 | + Cτn + C 0 τn ≤ 4p (|X0 |p + C p τnp + C p (Xs∗ )p ds · τnp−1 + (Iτ∗n (f ))p ). 0

Therefore, by Gronwall’s inequality, we obtain that (Xτ∗n )p ≤ 4p e4

p

p−1 C p τn

(|X0 |p + C p τnp + (Iτn∗ (f ))p ).

Hence, E(Xτ∗n )p ≤ 4p e4

p

C p T p−1

(E|X0 |p + C p Eτnp + E(Iτ∗n (f ))p ).

By applying Theorem 1.10.6, we obtain (a) and (b) for τ = τn , n ≥ 1. By taking n → ∞, we obtain the proof.

1.13 Stochastic Fubini Theorem for the Wiener Integrals w.r.t fBm

57

Remark 1.12.3. Exponential estimates for the solution of the more simple version of equation (1.12.1), were obtained in (TV03). We shall return to this problem in Section 3.5.

1.13 Stochastic Fubini Theorem for the Wiener Integrals w.r.t fBm We consider now only the case H ∈ (1/2, 1). Let PT = [0, T ]2 . Theorem 1.13.1. Let the measurable function f = f (t, s) : PT → R satisfy the conditions 2α−1 | f (t, u) | |f (t, s)| |s − u| ds du dt < ∞ (1.13.1) [0,T ]3

and

2α−1

[0,T ]4

|f (t1 , u)| |f (t2 , s)| |s − u|

ds du dt1 dt2 < ∞.

(1.13.2)

T T Then both the repeated integrals I1 := 0 ( 0 f (t, s)dt)dBsH and T T I2 := 0 ( 0 f (t, s)dBsH )dt exist and I1 = I2 with probability 1. Proof. The existence of the integral I1 is evident, due to (1.13.2). As to I2 , T f (t, s)dBsH exists a.e. (mod λ), where λ is the Lebesgue measure, and 0 according to (1.13.1), it holds that 2 T 1/2 T T E f (t, s)dBsH dt ≤ T 1/2 E f (t, s)dBsH dt 0 0 0 0 1/2 ≤ T 2αH |f (t, s)||f (t, u)||s − u|2α−1 du ds dt < ∞.

T

[0,T ]3

We consider at first only the measurable and bounded functions. Let f ∗ := sup(t,s)∈[0,T ]2 |f (t, s)| < ∞. Then there exists the sequence of simple and totally bounded functions fn = fn (t, s), such that fn → f uniformly on PT . The statement of the theorem is evident for fn . Further, denote gn (t, s) := f (t, s) − fn (t, s) and obtain the estimate T T T T H H gn (t, s)dt dBs + gn (t, s)dBs dt |I1 − I2 | ≤ 0 0 0 0 =: I1n + I2n . Furthermore,

58


E|I1n |2 = 2αH ≤ 2αHT 2 =T

2H+2

PT

sup (t,s)∈[0,T ]2

T gn (t1 , s)dt1 gn (t2 , u)dt2 |s − u|2α−1 ds du 0 0 |gn (t, s)|2 |s − u|2α−1 ds du T

PT

|gn (t, s)| → 0, 2

sup (t,s)∈PT

and E|I2n |2 ≤ T 0

T

E

T

2 gn (t, s)dBsH dt ≤

0

sup

|gn (t, s)|2 T 2H+2 → 0,

(t,s)∈PT

as n → ∞, and we obtain the proof for bounded f . Now, let f satisfy (1.13.1) and (1.13.2). For fn (t, s) := f (t, s)1{|f (t,s)|≤n} , n ≥ 1 the theorem is already proved. Define Cn := {(t, s, u) ∈ [0, T ]3 | |f (t, s)| ≥ n, |f (t, u)| ≥ n},

f n = f − fn .

Then for any n ≥ 1 we have that T T |I1 − I2 | ≤ f (t, s)1{|f (t,s)|>n} dt dBsH 0 0 T T H + f (t, s)1{|f (t,s)|>n} dBs dt =: I1n + I2n . 0 0 Furthermore, we have that T T 2 | = 2αH f n (t1 , s)dt1 f n (t2 , s)dt2 |s − u|2α−1 ds du E|I1n 0 0 [0,T ]2 ≤ 2αH |f n (t1 , s)||f n (t2 , s)||s − u|2α−1 ds du dt1 dt2 → 0, [0,T ]4

as n → ∞, according to (1.13.2), and 2 E|I2n | ≤ T 2αH |f n (t, s)||f n (t, u)||s − u|2α−1 ds du dt → 0, [0,T ]3

as n → ∞, according to (1.13.1).

1.14 Martingale Transforms and Girsanov Theorem for Long-memory Gaussian Processes According to Section 1.8, the process

1.14 Martingale Transforms and Girsanov Theorem (5)

MtH := CH

t

59

s−α (t − s)−α dBsH

0

t is a square integrable martingale, and Bt := α 0 sα dMsH is a Wiener process. (6) t In turn, BtH = CH 0 mH (t, s)dBs . Moreover, the process Yt =

(6) CH

t

(t − s)α s−α dBs

(1.14.1)

0

(5) t has the property that MtH = CH 0 (t − s)−α dYs is square-integrable martingale. All these processes are Gaussian. Therefore, in some sense, it is more convenient to consider the processes of a form similar to Yt and Mt , and to avoid fractional Brownian motion itself. In this section we consider t long-memory Gaussian processes that can be presented as integrals Vt = 0 h(t−s)ϕ(s)dWs with some Wiener process Wt and establish the conditions allowing us to transform these processes, similarly to Yt , into square-integrable martingales. Let {Wt , FtW , t ≥ 0} be the standard Wiener 0 process on a complete probability space (Ω, F, P ) with F = F∞ := t≥0 FtW . Define the convolution of two measurable integrable functions ϕ1 and ϕ2 : R+ → R by t (ϕ1 ∗ ϕ2 )(t) = 0 ϕ1 (t − s)ϕ2 (s)ds, t ∈ R+ . Let h and ϕ satisfy the assumption

(h2 ∗ ϕ2 )t < ∞, t > 0. (1.14.2) t Define the Gaussian process Vt = 0 h(t − s)ϕ(s)dWs . Evidently, EVt = 0. In the case when h(s) = sα , ϕ(s) = s−α and H ∈ (1/2, 1), the covariance function between distant increments of the process Vt vanishes at a power rate. More precisely, t EVt (Vt+k −Vk ) = (t − s)α ((t + k − s)α − (k − s)α )s−2α ds 0 t (t − s)α (t + k − s)α−1 s−2α ds ≥ αt ϕ ∈ L2 (0, t),

0

≥ αt2−α B(α + 1, α)k α−1 , ∞ and the series k=1 k α−1 diverges for H ∈ (1/2, 1). Due to this reason, according to the generally accepted terminology (CCM03; Ber94; WTT99), such processes are said to have a long memory. Compare this to the notion of longrange dependence from Section 1.2. Denote by Ruv = E Vu Vv the correlation function. Then we have that u∧v h(u − s)h(v − s)ϕ2 (s)ds. Ruv = 0

Let FtX = σ{Xs , 0 ≤ s ≤ t} and HtX = H{Xs , 0 ≤ s ≤ t} be, correspondingly, σ-fields and Gaussian subspaces, generated by the process X

60


on the interval (0, t], X = W, V . It follows from (CCM03, Proposition 15) that FtV = FtW , t ∈ R+ if and only if HtV = HtW . A necessary and sufficient condition for this coincidence can be formulated as the only function f such that ∀t ∈ R+ f ∈ L2 (0, t) and ((f · ϕ) ∗ h)t = 0 is the zero function.

(1.14.3)

V W Evidently, in this case F∞ = F∞ . We give one sufficient condition for the latter relation. Denote by ∞ Ff (λ) := e−λs f (s)ds, λ > 0 0

the Laplace transform of f . The following result is a direct consequence of (CCM03, Proposition 17). Lemma 1.14.1. Let the following condition hold 0 < |Fh (λ)| < ∞,

F|ϕ| (λ) < ∞,

Fϕ (λ) = 0

(1.14.4)

V W on some interval λ ∈ (a, b) ⊂ (0, ∞). Then F∞ = F∞ .

Now, let (1.14.3) hold. Denote by L2 (V ) = L2 (W ) = L2 (Ω, F∞ , P ) the space of F∞ -measurable ξ with Eξ 2 < ∞. Let H(V ) be the closed subspace of L2 (V ) consisting of linear functionals of V . Suppose that the function R : R2+ → R has a bounded variation |R|t := varPt R on any rectangle Pt , t ∈ R2+ , and consider the measurable function g : R+ → R such that |g(s − u)| |g(t − v)| d |R|uv < ∞, s, t ∈ R+ . (1.14.5) P(s,t)

As stated by (HC78), we have an isomorphism I between Λ2 (R) and H(V ). Here Λ2 (R) is the completion of the space Λ of step functions N f (t) = k=1 αk 1[tk+1 ,tk ) (t) in the norm generated by a scalar product f, g = Denote by I(f ) =

R

R

f (u) g(v) dRuv ,

I(f ) =

N

αk (Vtk+1 − Vtk ).

k=1

f dV ∈ H(V ) the image of f ∈ Λ2 (R) and let

t

g(t − u) dVu := I(˜ g ),

Mt := 0

where g˜(s) = g(t − s) 1{s≤t} , t ≥ 0. Then {Mt , FtW , t ≥ 0} is a Gaussian process and EMs Mt =

P(s,t)

g(s − u) g(t − v) dRuv .

1.14 Martingale Transforms and Girsanov Theorem

61

Moreover, under the condition: the double Riemann integral P(s,t)

g(s − u) g(t − v) dRuv

exists, (1.14.6)

the process Mt can be considered for any t ≥ 0 as a limit of Riemann sums in the mean-square sense. Note that the following condition is sufficient for (1.14.6): the derivative h (s), s > 0, exists, h(0) = 0, and Ruv admits a representation + * u1 ∧v1 h (u1 − z) h (v1 − z) ϕ2 (z) dz du1 dv1 (1.14.7) Ruv = P(u,v)

0

and P(s,t)

* |g(s − u)| |g(t − v)|

+ h (u − z) h (v − z) ϕ2 (z) dz du dv < ∞.

u∧v

0

Now we are in a position to study conditions on ϕ, h and g supplying martingale properties of Mt . Definition 1.14.2. Gaussian process V is called (g)-transformable if the process t g(t − s) dVs Mt := 0

is a martingale. Remark 1.14.3. Since Mt is a Gaussian process, it is a square-integrable martingale if V is (g)-transformable. Denote U = {f : R+ → R (f ∗ q)t = 0, t ∈ R+ , for such q : R+ → R that (|f | ∗ |q|)t < ∞, t ≥ 0, if and only if q = 0}, s t AC[0, t] = {f : R+ → R f (s) = 0 f (u) du; 0 ≤ s ≤ t with 0 |f (u)| du < ∞}. Theorems 1.14.4 and 1.14.5 contain two groups of sufficient conditions on the functions ϕ, h, g ensuring (g)-transformability of Vt (statements 1) and 3)). Statements 2) and 4) demonstrate that these conditions are, in some sense, necessary. Theorem 1.14.4. 1) Let ϕ, h, g satisfy conditions (1.14.2), (1.14.3), (1.14.7) and (|g| ∗ |h |)t < ∞, t > 0, (g ∗ h )t = C0 , t > 0 for some C0 ∈ R. Then Vt is (g)-transformable and M t = C02

t 0

ϕ2 (s) ds.

(1.14.8) (1.14.9)

62


2) Let ϕ, h, g satisfy conditions (1.14.2), (1.14.3), (1.14.7) and (1.14.8), h ∈ U , ϕ = 0 (mod λ) (λ is the Lebesgue measure), (g ∗ h )t ∈ C(0, ∞), Vt be (g)-transformable. Then (g ∗ h )t = C0 , t > 0, for some C0 ∈ R. Theorem 1.14.5. 3) Let ϕ and h satisfy (1.14.2) and (1.14.3), ϕ = 0 (mod λ), g satisfies (1.14.6) and g ∈ AC[0, t],

t ≥ 0,

g(0) = 0,

(|g | ∗ (h2 ∗ ϕ2 )1/2 )t < ∞, t > 0, (g ∗ h)t = C0 , t > 0 for some C0 ∈ R.

(1.14.10) (1.14.11) (1.14.12)

t Then Vt is (g)-transformable and M t = C02 0 ϕ2 (s) ds. 4) Let ϕ and h satisfy (1.14.2), (1.14.3), ϕ = 0 a.e. (mod λ), the process Vt is (g)-transformable with g satisfying (1.14.10), (1.14.11), (g ∗h)t ∈ C(0, ∞). Then (g ∗ h)t = C0 , t > 0, for some C0 ∈ R. Remark 1.14.6. Conditions (1.14.9) and (1.14.12) mean, in particular, that corresponding convolutions have jumps at zero, so at least one of the functions involved is singular at 0. Remark 1.14.7. Let h(s) = sα , ϕ(s) = s−α , g(s) = s−α . Then statement 1) holds for H ∈ (1/2, 1) and statement 3) holds for H ∈ (0, 1/2). Proof of Theorem 1.14.4. 1) It follows from (1.14.7) that ft (z) :=

t

1 g(t − v)

0

z∧v

2 h (v − r) h (z − r) ϕ2 (r) dr dv,

0≤z≤t

0

is defined for a.a. z ≤ t for any t ∈ R+ fixed. Condition (1.14.7) ensures the Fubini theorem for ft , and from (1.14.8)–(1.14.9) we obtain that v z 2 g(t − v) h (v − r) h (z − r) ϕ (r) dr dv ft (z) = 0 0 z t 2 + g(t − v) h (v − r) h (z − r) ϕ (r) dr dv z 0 t z = h (z − r) ϕ2 (r) h (v − r) g(t − v) dv dr 0 r z = C0 h (z − r) ϕ2 (r) dr, 0

i.e. ft does not depend on t ≥ z. Further, for any 0 ≤ s ≤ t we have that s E(Mt − Ms )Ms = g(s − u) ft (u) − fs (u) du = 0. 0


63

It means that the Gaussian process Mt with EMt = 0 has uncorrelated, thus independent, increments. Hence, Mt is a Gaussian martingale, and it holds that t t u∧v 2 g(t − u) g(t − v) h (u − r) h (v − r) ϕ (r) dr du M t = 0

0

t

0

g(t − u)

= C0 0

v

h (v − r) ϕ (r) dr dv = 2

C02

0

t

ϕ2 (r) dr. 0

t 2) Let Mt = 0 g(t − s) dVs be a square integrable martingale with g satisfying (1.14.7) and (1.14.8). Then E(Mt − Ms ) Vs = 0,

0 ≤ s < t,

or

0=

s

v

h (v − r) ϕ2 (r)

t

h (u − r) g(t − u) du 0 0 r s h (u − r) g(s − u) du dr dv = (h ∗ (ϕ2 · ζ))s , − r

where

t−r

s−r

h (u) g(t − r − u) du −

ζ(r) = 0

h (u) g(s − r − u) du.

0

Since h ∈ U , we obtain ϕ2 · ζ = 0, and, taking into account that ϕ = 0, we derive that ζ(r) = 0 (mod λ), r ≤ s ≤ t. Together with continuity of

h ∗ g ∈ C(0, ∞) it means that (h ∗ g)t = C0 , t > 0, for some C0 ∈ R. Proof of Theorem 1.14.5. 3) Under condition (1.14.6) the integral Mt is a mean-square limit of Riemann sums, and condition (1.14.10) permits us to transform the sum: Mt = l.i.m.

|λN |→0

= l.i.m.

|λN |→0

t

N −1

g(t − si ) (Vsi+1 − Vsi )

i=0 N −1

V (si+1 ) (g(si+1 ) − g(si ))

i=0

g (t − s) Vs ds =

= 0

t

g (t − s) 0

s

h(s − z) ϕ(z)dWz ds,

0

where |λN | = max0≤i≤N −1 |g(si+1 ) − g(si )|, and the last integral is the limit of Riemann sums in the mean-square sense. Further, condition (1.14.11), according to (Pro90, p. 160) or (Leb95), permits to apply to Mt the stochastic Fubini theorem, and we obtain from (1.14.12) that

64


Mt =

t

t

t

g (t − u) h(u − z) du dWs = C0

ϕ(z) 0

ϕ(z) dWz . (1.14.13) 0

z

4) If the process Mt is a square-integrable martingale, then from (1.14.13) it follows that for any 0 ≤ s ≤ t s W 0 = E(Mt − Ms /Fs ) = ϕ(z) η(z) dWz , 0

where

η(z) = (g ∗ h)t−z − (g ∗ h)s−z .

s Hence 0 ϕ2 (z) η 2 (z) dz = 0, and, arguing similarly to the completion of the proof of Theorem 1.14.4, part 2), we obtain that (g ∗ h)t = C0 for some

C0 ∈ R. Consider some examples of the functions ϕ, h satisfying conditions 1) or 3). (One example is contained in Remark 1.14.7.) Example 1.14.8. Let g(x) = x−1/2 cosh(ax1/2 ), x h (x) = sν/2 Iν (as1/2 ) (x − s)γ ds, 0

where −1 < ν
0, whence (g ∗ h )t = Γ (γ + 1)2−ν−1 πaν , t > 0, and condition (1.14.9) holds. (For the details of the theory of Bessel functions of the first kind and their Laplace transforms see (Wat95) and (GR80).) Condition (1.14.8) is fulfilled since |h (x)| ≤ Cxν+γ+1 on any interval (0, t), where C depends on t. Conditions (1.14.2) and (1.14.7) hold for any ϕ ∈ L2 (0, t), t > 0; condition (1.14.3), according to Lemma 1.14.1, holds for any ϕ such that F|ϕ| (λ) < ∞, Fϕ (λ) = 0 for λ ∈ (a, b) ⊂ (0, ∞). In this case Vt is (g)-transformable, according to part 1) of Theorem 1.14.4.


65

x

Example 1.14.9. Let g(x) = x−1/2 cosh(ax1/2 ), h(x) = 0 t−1/2 cos(at1/2 ) dt. Then Fg (λ) = (π/λ)1/2 exp(a2 /4λ), Fh (λ) = (π/λ)1/2 exp(−a2 /4λ), Fg (λ)Fh (λ) = π/λ, λ > 0, so (g ∗ h )t = π, t > 0. Since |h(x)| ≤ Cx1/2 , we can conclude as in Example 1.14.8. x Example 1.14.10. Let g (x) = 0 t−1/2 cosh(at1/2 ) (x − t)γ dt, h(x) = xν/2 Iν (ax1/2 ) with γ ∈ (−1, − 12 ), ν ∈ (−1, 0), γ + ν = − 32 . Then Fg (λ) = π 1/2 λ−γ−3/2 exp(a2 /4λ), Fh (λ) = λ−ν−1 exp(−a2 /4λ), Fg (λ)Fh (λ) = π 1/2 λ−1 . Conditions (1.14.2), (1.14.3) and (1.14.11) hold for ϕ ∈ L2 (0, t), t > 0, F|ϕ| (λ) < ∞, Fϕ (λ) = 0 for some interval (a, b) ⊂ (0, ∞), (1.14.10) is evident, (1.14.6) is fulfilled at least for ϕ ∈ C(R+ ). So, if ϕ > 0, ϕ ∈ C(R+ ) and F|ϕ| (λ) < ∞ we have part 3) of Theorem 1.14.5. Remark 1.14.11. According to Proposition 7 from (HC78), under the condition h ∈ L2 (0, t), t > 0, ϕ ≡ 1, Vt is a semimartingale. In this case we transform semimartingale into martingale by (g)-transformation. For example, let h(x) = xε , 1/2 < ε < 1, ϕ(x) = 1. Then

t

h(t − s)dWs = ε

Vt = 0

t 0

s

(s − u)ε−1 dWu ds

0

is a semimartingale, more precisely, a process of bounded variation. Put g(x) = t s x−ε . Then Mt = ε 0 (t − s)−ε ( 0 (s − u)ε−1 dWu )ds = εB(ε, 1 − ε)Wt , where B(·, ·) is the beta-function. t Now, let Vt be equal to Yt from (1.14.1). Recall that BtH = 0 sα dVs is an fBm with Hurst index H, and in this case BtH can be presented as t BtH = 0 mH (t, s)dBs , where B is a Wiener process and the kernel mH (t, s) is defined in Section 1.8. Consider general conditions on function ψ : R+ → R t for the process Nt := 0 ψs dVs to be presented in a similar way. Theorem 1.14.12. Let conditions (1.14.2), (1.14.3) hold and also ε lim ψ 2 (ε) h2 (ε − u) ϕ2 (u) du = 0; (1.14.14) ε↓0 0 ψ(u) ψ(v) dRuv exists, s, t > 0; (1.14.15) the Riemann integral [0,(s,t)]

there exists a derivative ψ (s), s > 0 and (h2 ∗ ϕ2 )1/2 ψ ∈ L1 (0, t), Then

t

ψ(s) dVs = 0

where

(|h| ∗ |ψ |)t < ∞,

t > 0.

t

m(t, s)ϕ(s) dWs , t > 0, a.s., 0

(1.14.16)

66


m(t, s) = ψ(t)h(t − s) −

t

h(u − s)ψ (u) du,

s

W is a Wiener process. If (1.14.16) is strengthened to (h2 ∗ ϕ2 )1/2 ψ ∈ L2 (0, t), t > 0,

(1.14.17)

t then E( 0 ψ(s) dVs )2 < ∞. t Proof. Under (1.14.14)–(1.14.16), we can consider the integral 0 ψ(u) dVu as a mean-square limit of Riemann sums, and integrating by parts, we obtain the following limits in the mean-square sense t t ψ(u) dVu = lim ψ(u) dVu ε↓0 ε 0 t = ψ(t)V (t) − lim ψ(ε)V (ε) − ψ (u)V (u) du ε↓0 0 u t ψ (u) h(u − s)ϕ(s)dWs du. = ψ(t)V (t) − 0

0

Due to (1.14.16), the stochastic Fubini theorem can be applied to the last integral, and we obtain t t t t ψ(u) dVu = ψ(t)h(t − s)ϕ(s)ds − ϕ(s) h(u − s)ψ (u) du dWs 0

0

0

s

t

=

m(t, s)ϕ(s)dWs . 0

The second statement is evident.

Now let P and P be two probability measures on (Ω, F). Denote by loc Pt (Pt ) the restriction of P (P) on Ft and suppose that P P (it means that Pt Pt , t ∈ R+ ). Consider the density process Zt = E(Xt ) := −Xs , X is a local martingale. exp Xt − 12 X c t 0≤s≤t (1+ Xs )e t As before, we consider the Gaussian process Vt = 0 h(t − s)ϕ(s) dWs and suppose that Vt is (g)-transformable by the function g; moreover, the condi t tions (1.14.8)–(1.14.9) or (1.14.10)–(1.14.12) hold. Let Mt = C0 0 ϕ(s) dWs with C0 depending on g. Since Mt has continuous modification, the process [M, X] has P -locally bounded variation (see (JS87, Lemma 3.14)). Denote by At := M, Xt the P -compensator of [M, X]. Suppose further that the function ψ satisfies conditions (1.14.14)–(1.14.16) of Theorem 1.14.12. t Lemma 1.14.13. The integral 0 m(t, s) dAs exists for any t > 0 P - and Pa.s.


67

t Proof. Since m(t, s) = ψ(t)h(t − s) − s h(u − s)ψ (u) du, we consider t t t h(t − s) dAs and 0 s h(u − s)ψ (u) du dAs individually. From Kunita’s 0 inequality and (1.14.2),

t

t

|h(t − s)| d|A|s ≤ 0

|h(t − s)|2 dM s · Xt 0

t

|h(t − s)| ϕ (s) dsXt 2

= C0

2

12

12

0, D0 = 0; t 4) ψ = 0, the integral 0 |g(t − s)| |ψ −1 (s)| d|D|s < ∞ a.s., t > 0, and we have a representation t t t g(t − s) ψ −1 (s) dDs = δs ds, where |δs | ds < ∞ a.s. 0 0 0 t 2 ϕ−2 t > 0; E s δs ds < ∞, 0

5) EE(Xt ) = 1, where

C0−1

Xt =

t

1 E(Xt ) = exp Xt − Xt ; 2

ϕ−1 s δs dWs ,

0

or: 6) conditions (1.14.2), (1.14.3), (1.14.6), (1.14.10)–(1.14.12); 7) conditions 3)–5); t 8) a process Et = 0 m(t, s) δs ds has bounded variation and

t

|g(t − s)| |ψ −1 (s)| d|E|s < ∞, a.s.,

t > 0;

0

9) g ∈ U . t = Bt − Dt is Gaussian and admits the representation Then the process B t P t = m(t, s)ϕ(s)dW s under the measure P loc B P such that ddP W = E(Xt ). 0 Ft

t

Proof. In our case At = M, Xt = 0 δs ds, therefore from Theorem 1.14.14 the “drift” equals C0−1 Et . It is enough to establish that Dt = C0−1 Et . If conditions 1)–5) hold, then

t

t

m(t, s) δs ds = 0

−

t

t −1 h(u − s) ψ (u)du d g(t − s) ψ (s) dDs

0 t

ψ(t) h (t − s)

= 0

ψ(t) h(t − s)

0

0 s

g(s − u) ψ 0

−1

t

t

h (v − s) ψ (v) dv

dDu ds − 0

s


s

×

0 t t

= ψ(t) 0

t t

t

− 0

0

g(s − u) ψ −1 (u) dDu ds h (t − s) g(s − u) ds ψ −1 (u) dDu

u

h (v − s) ψ (v) g(s − u) ψ −1 (u) I{u ≤ s ≤ v ≤ t}dv ds dDu

0

69

t

= C0 ψ(t)

t

ψ −1 (u) dDu − C0

0

0

t

= C0 ψ(t)

ψ

−1

ψ (v) ψ −1 (u) I{u ≤ v ≤ t}dv dDu

0

t

(u) dDu − C0

0

t

(ψ(t) − ψ(u)) ψ −1 (u) dDu = C0 Dt .

0

If conditions 6)–9) hold, then for any t > 0 t t −1 g (t − s) ψ −1 (s) + g(t − s) ψ (s) ψ −2 (s) g(t − s) ψ (s) dEs = 0

0

s

s

s

h(s − u) δu du −

× ψ(s) 0

0

h(v − u) ψ (v) dv δu du ds.

u

(1.14.18) The right-hand side of (1.14.18) contains four integrals. Consider them separately. From (1.14.12), t s t g (t − s) h(s − u) δu du ds = C0 δu du. 0

0

0

Further, s ψ (s) g(t − s) 2 h(s − u)δu du ψ(s) ψ (s) 0 0 s s h(v − u)ψ (v)dv δu du ds − 0 u s t z ψ (z) = g (t − s) h(z − u) δu du ψ(z) 2 0 0 ψ (z) 0 z z h(v − u) ψ (v) dv δu du dz ds. −

t

0

0

It is sufficient to prove that s z ψ (z) h(z − u)δu du σs := ψ(z) 2 0 ψ (z) 0 z z h(v − u)ψ (v)dv δu du dz − 0 u s s h(v − u) ψ (v) dv δu du =: σ ¯s , = ψ −1 (s) 0

u

(1.14.19)

70


and then it follows that the right-hand side of (1.14.18) equals C0 ¯0 , and the derivative But σ0 = σ s ψ (s) s σ ¯s = − 2 h(v − u) ψ (v) dv δu du ψ (s) 0 s u + ψ −1 (s) h(s − u) δu du · ψ (s) = σs .

t 0

δu du.

0

We obtain that t t −1 −1 −1 g(t − s) ψ (s)dDs = g(t − s) ψ (s) d C0 0

0

t

s

g (t − s) 0

ψ

−1

z(s, u) δu du ,

0

or

s

(u) dDu ds =

C0−1

0

t

s

g (t − s)

ψ

0

−1

(u) dEu ds.

0

s t If g ∈ U then 0 ψ −1 (u) d (D − E)u = 0, whence Dt = 0 ψs · ψs−1 dDs = t t s ψt · 0 ψs−1 dEs − 0 ψs · 0 ψu−1 dEu ds = Et .

P.

Theorem 1.14.16 permits us to calculate the Hellinger process for P and

Let P P and Yt = E(Xt ), Xt be a continuous square-integrable martingale. According to (JS87, Corollary 1.37) the Hellinger process in a narrow sense of order β equals ht (β) = 12 β (1 − β) Xt .

Theorem 1.14.17. Let one of conditions 1)–5) or 6)–9) hold, then β (1 − β) ht (β) = 2C02

t

2 ϕ−2 s δs ds

0

=

β (1 − β) 2C02

t

0

ϕ−2 s

d ds

s

g(t − u) ψ −1 (u) dDu

2 ds.

0

The proof follows immediately from Theorem 1.14.16.

t Remark 1.14.18. It is possible to study if the process Vt = 0 h(t − s)ϕ(s)dWs is itself a semimartingale. In the case when ϕ ≡ 1 this question is investigated in (CCM98). Theorem 1.14.19. Let the function h be differentiable on R+ , t t |h (u)|du < ∞, t ≥ 0, and 0 (h (t − u)ϕ(u))2 du < ∞, t ≥ 0. 0 Then the process {Vt , FtW , t ≥ 0} is a semimartingale. t Proof. We have the representation h(t) = h(0) + 0 h (u)du, which together with the Fubini theorem supplies the following transformations: t t t t−s Vt = h(t − s)c(s)dWs = h(0) c(s)dWs + h (u)duϕ(s) dWs 0

0

0

0

1.15 Nonsemimartingale Properties of fBm

t

t

= h(0)

t

ϕ(s)dWs + 0

0

t

v

71

t

ϕ(s)dWs 0

s

+ 0

h (v − s)ϕ(s)dv dWs = h(0)

h (v − s)ϕ(s)dWs dv.

0

1.15 Nonsemimartingale Properties of fBm; How to Approximate Them by Semimartingales A process {Xt , Ft , t ≥ 0} is called semimartingale, if it admits the representation Xt = X0 + Mt + At , where M is an Ft -local martingale with M0 = 0, A is a process of locally bounded variation, X0 is F0 -measurable. Evidently, any semimartingale has locally bounded quadratic variation; if X is continuous, then M and A are continuous. Let Xt = BtH with H ∈ (0, 1/2). Then its quadratic variation is infinite, therefore, it is not a semimartingale. If H ∈ (1/2, 1), then the quadratic variation of X is zero, and if we suppose that X is semimartingale, then the quadratic variation of Mt = Xt − X0 − At is zero, and M is zero. But Xt = At since X has unbounded variation. Therefore, Xt = BtH is not a semimartingale for any H = 1/2. (There are many another elegant proofs of this fact.) Nevertheless, there are many approaches to how to approximate fBm by a sequence of semimartingales. 1.15.1 Approximation of fBm by Continuous Processes of Bounded Variation We follow here the approach of and (And05) and (AM06). According to (1.8.5) and (1.8.18), we can represent {BtH , t ≥ 0} with Hurst index H ∈ (1/2, 1) as BtH =

t

sα dYs , 0

where (8)

Yt = CH

t

(t − s)α s−α dBs ,

0 (8)

(6)

{Bt , t ≥ 0} is a Wiener process, CH = CH α . We can rewrite Yt as t t (8) α−1 Yt = CH α (u − s) du s−α dBs . 0

s

(1.15.1)

72


If we formally apply the stochastic Fubini theorem to the right-hand side of (1.15.1), we obtain that t u (8) α−1 −α Yt = CH α (u − s) s dBs du. (1.15.2) 0

0

But the right-hand side of (1.15.2) does not exist, since the variance of interior integral is infinite, u (u − s)2α−2 s−2α ds = ∞. 0

Thereupon, we introduce the “truncated” process for β ∈ (0, 1), t

(8)

Ytβ = CH α

βs

0

and BtH,β

t α

=

s

dYsβ

=

0

(8) CH α

(s − u)α−1 u−α dBu ds,

0

t

βs

α−1 −2α

(s − u)

α

s 0

u

dBu ds

(1.15.3)

0

is a process of bounded variation which will serve as an approximation of BtH . Theorem 1.15.1. We have that E(BtH − BtH,β )2 ≤ c1 t2H (1 − β)2α , where c1 = c1 (H) is some constant, independent of t and β. Proof. First, we want to change the limits of the integration in (1.15.3) and consider the process βt

(8)

Ztβ := αCH =

(8) CH

t

0

(s − u)α−1 ds u−α dBu

u/β βt

α −α

(t − u) u 0

dBu −

1−β β

α Bβt

.

(1.15.4)

We cannot apply here the stochastic Fubini theorem (Pro90, Theorem IV.4.5), βt t because it is valid if the integral 0 u/β (s − u)2α−2 u−2α ds du is finite but it is infinite. Therefore, we must go an indirect way. We consider the integral t βs (8) Ytβ,ε = D ε βε (s − u)α−1 u−α dBu ds, where D = αCH , and the Fubini theorem ensures the equality

βt

Ytβ,ε = Ztβ,ε := D βε

(

t

u/β

(s − u)α−1 ds)u−α dBu .


73

Furthermore, E|Ytβ,ε − Ytβ | ≤ D t

ε

0 βε

+ ε

(s − u)2α−2 u−2α du

0

×u

−2α

βs

(s − u)2α−2 u−2α du

0

1/2

ds ≤ D

du + α (βε)

1/2−α

ε

u−1/2 du

0

t

1/2 ds

β

(1 − u)2α−2

0

(s − βε)α−1 ds → 0

ε

and

E|Ztβ,ε

−

Ztβ |2

≤D

2

βε

0

t

(s − u)α−1 ds

2

u−2α du ≤ CD2 βε1−2α → 0

u/β

as ε → 0, where C > 0 is some constant. This means that Ytβ = Ztβ a.s. for any t ∈ [0, T ]. Therefore, for 1/2 < β < 1 t 2 α 1−β (8) 2 β 2 α −α E(Yt − Yt ) = (CH ) E (t − u) u dBu + Bβt β βt 2α t 1−β (8) 2 (8) 2 2α −2α ≤ 2(CH ) (t − u) u du + 2(CH ) βt β βt 2α 1−β (8) (8) ≤ H −1 (CH )2 (βt)−2α t2H (1 − β)2H + 2(CH )2 βt β ≤ c2 t(1 − β)2α

with c2 = (CH )2 · 22α−1 (H −1 + 2). (1.15.5) (8)

Integration by parts gives us BtH − BtH,β = tα (Yt − Ytβ ) − α

t

(Ys − Ysβ )sα−1 ds 0

whence we obtain from (1.15.5) that E(BtH

−

BtH,β )2

≤ 2t E(Yt −

Ytβ )2

t

+ 2α t E(Ys − Ysβ )2 s2α−2 ds 0 t s2α−1 ds · c2 (1 − β)2α , ≤ 2c2 t2H (1 − β)2α + 2α2 t 2α

2

0

and we can put c1 = 2c2 (α + 1).

1.15.2 Convergence B H,β → B H in Besov Space W λ [a, b]. For λ ∈ (0, 1/2) define the Besov space W λ [a, b] as the space of measurable functions f : [a, b] → R such that b b s |f (s)| |f (s) − f (y)| ds + dy ds < ∞. f a,b,λ := λ (s − y)λ+1 a (s − a) a a

74


Theorem 1.15.2. For any λ ∈ (0, 1/2), H ∈ (1/2, 1) and any [a, b] ⊂ [0, T ] EB H − B H,β a,b,λ ≤ c1 (H, λ, T )(1 − β)α . H,β

:= BtH − BtH,β . We have

Proof. Denote B t EB

H,β

b

λ = E a

H,β

|B s | ds + E (s − a)λ

b

a

a

s

H,β

H,β

|B s − B y | dy ds. (s − y)λ+1

(1.15.6)

From Theorem 1.15.1, b b b H,β H,β |B s | (E(B s )2 )1/2 sH 1/2 α E ds ≤ ds ≤ c (1 − β) ds 1 λ λ (s − a)λ a (s − a) a a (s − a) ≤ c1 (H, λ, T )(1 − β)α , (1.15.7) with c1 (H, λ, T ) = c1 · T H−λ+1 · (H − λ + 1)−1 . Consider the second term in the right-hand side of (1.15.6). Rewrite the difference in the numerator as 1/2

H,β

Bs

H,β

− By

= (BsH − BsH,β ) − (ByH − ByH,β ) s s β = uα d(Yu − Yuβ ) = uα dY u , y

(1.15.8)

y

β

where Y u = Yu − Yuβ . Equality (1.15.8) and integration by parts give us the estimates H,β b s H,β |B s − B y | dy ds (s − y)λ+1 a a s b s β α −λ−1 α β α β (s − y) Y u u du dy ds = s Y s − y Y y + α a a y b s β β ≤ (s − y)−λ−1 sα Y s − Y y dy ds a

a

b

+ a

s

β

(s − y)−λ−1 (sα − y α )|Y y |dy ds

a b s

−λ−1

(s − y)

+α a

a

s

β |Y u |uα−1 du dy ds

y

=: I1 (β) + I2 (β) + αI3 (β). Now we estimate I2 (β): b EI2 (β) ≤ α a 1/2

s

a

b

β

y α−1 (s − y)−λ (E(Y y )2 )1/2 dy ds

≤ c2 α a

s

y α−1 (s − y)−λ y 1/2 dy ds · (1 − β)α

a

≤ c2 (H, λ, T )(1 − β)α ,

(1.15.9)


75

1/2

where c2 (H, λ, T ) = c2 αT 1−λ . Similarly,

b

s

E I3 (β) ≤ ≤

s

(s − y)

1/2 c2

−λ−1

a a b s

−λ−1

(E(Y y s

(s − y) a

dy ds

u

a

β 2 1/2 α−1 u du u) )

α−1/2

du dy ds · (1 − β)α

(1.15.10)

y

≤ c3 (H, λ, T )(1 − β)α , H−λ+1

1/2

T . Now we use the representation where c3 (H, λ, T ) = c2 H(H−λ)(H−λ+1) (1.15.4) to estimate I1 (β): s y β β (8) α −α α −α (s − u) u dBu |Y s − Y y | ≤ CH (s − u) u dBu −

(8)

+ CH

βs

1−β β

βy

α

|Bβs − Bβy |,

therefore

b

s

I1 (β) ≤ (s − y)−λ−1 sα a a s y α −α α −α × (s − u) u dBu − (y − u) u dBu dy ds (8) CH

βs

(8)

+ CH

1−β β

α

b

a

βy

s

(1.15.11)

sα (s − y)−λ−1 |Bβs − Bβy |dy ds

a

=: I1 (β) + I2 (β). Further, (8)

E I2 (β) ≤ CH

1−β β

α

b

a

s

sα (s − y)−λ−1/2 dy ds β 1/2

a

(1.15.12)

= c4 (H, λ, T )(1 − β) , α

H−λ+1

(8)

where c4 (H, λ, T ) = CH 2α · T1/2−λ . (Here we see that indeed λ must be less than 1/2.) Next, we decompose I1 (β) into two integrals I1 (β) =

(8) CH

a

b

a

(βs)∨a

(8) +CH

b

s

=: I3 (β) + I4 (β). a

(βs)∨a

76

1 Wiener Integration with Respect to Fractional Brownian Motion (8)

EI3 (β) ≤ CH

b

a

(βs)∨a

(s − y)−λ−1 sα

a

2 1/2 y s α −α α −α × E (s − u) u dBu − (y − u) u dBu dy ds βs

≤

√

(8)

2CH

s

×

a

≤2 H

(βs)∨a

(s − y)−λ−1 sα

a

(s − u)2α u−2α du +

βs α

βy

b

−1/2

(8) CH

b

a

y

(y − u)2α u−2α du

1/2 dy ds

βy (βs)∨a

(s − y)−λ−1 (s + y)1/2 sα dy ds · (1 − β)H

a H−λ

≤ c(H, λ, T )(1 − β)

(1.15.13) with c(H, λ, T ) = EI4 (β) ≤

(8) CH

2H T 1+H−λ . λ(1−λ)H 1/2

(8)

= CH

b

s

(βs)∨a

a

2 1(βy,y) (u) dBu

s

(s − y)−λ−1 sα

(βs)∨a

a

s E

−λ−1 α

(s − y)

α −α

− (y − u) u

b

Finally,

2 − (y − u)α 1(βy,y) (u) u−2α du

s

(s − u)α u−α 1(βs,s) (u)

0

1/2 dy ds

s

(s − u)α 1(βs,s) (u)

1/2

0

dy ds.

The interior integral equals s y α α 2 −2α ((s − u) − (y − u) ) u du + (s − u)2α u−2α du βs

y βs

+

(y − u)2α u−2α du =: I5 (β),

βy

and via some routine calculations can be estimated as I5 (β) ≤ CH (1 − β)2α (s − y), where CH = 1 + 22α + Therefore

α 1−2α .

(8)

EI4 (β) ≤ CH (CH )1/2 (1 − β)α ≤

(8) CH (CH )1/2 (1

b

s

(s − y)−λ−1/2 dy ds

sα

− β)

α

≤ C(H, λ, T )(1 − β)H−λ

(βs)∨a

a

b H−λ

s a

1

ds

(1 − y)−λ−1/2 dy

β

(1.15.14)


77

H−λ+1

(8)

T with C(H, λ, T ) = CH (CH )1/2 (H−λ+1)(1/2−λ) . Summarizing (1.15.9), (1.15.10), (1.15.12)–(1.15.14), we obtain the proof.

We obtain another approximation, considering the “truncated” process of the form t (s−β)+ (8) β α−1 −α Yt := CH α (s − u) u dBu ds 0

and

BtH,β

0

t

t ≥ 0,

sα dYsβ ,

=

H ∈ (1/2, 1).

(1.15.15)

0

Evidently, we intend to obtain the approximation while β → 0. Theorem 1.15.3. The process B H,β satisfies the relations 2H t , t 0. Also, let f (t, u) : R2 → C, t ∈ R, be such a function that for all t ∈ R 3) f (t, ·) ∈ C 1 (R); 4) |f (t, u)| = 0(|u|−l ) as u → ∞ for some l > 1/2. Theorem 1.15.6 ((PT00b)). 1. Under conditions 1)–4) the following convergence (in the sense of convergence of finite-dimensional distributions) takes place:

80


ξaf (t) := a−1/2

n d (ξn +iηn ) → ξ f (t) := f t, f (t, u)(dW1 (u)+idW2 (u)), a R

n∈Z

as a → ∞, where W1 and W2 are two independent standard Brownian motions. 2. If, in addition, f (0, u) = 0, |f (t, u) − f (s, u)| ≤ c|f (t − s, u)|

for all

s, t, u ∈ R,

|f (t, u)| ≤ ctH |f (1, u + ln t)|

for some

0 0 ξa (t) converges weakly to ξ(t) in the space C[0, T ] endowed with the uniform topology. u Corollary 1.15.7. Let f(t, u) = (eie t − 1)e−Hu . Then the corresponding process ξaf (t) is called the normalized Weierstrass–Mandelbrot process and, according to Theorem 1.15.6, it converges weakly to the process u f ξ (t) := (eie t − 1)e−Hu (dW1 (u) + dW2 (u)).

R

tH have identical finite-dimensional disMoreover, the processes ξ f (t) and B tributions because they are both Gaussian, have zero mean and the same covariance functions.

Remark 1.15.8. The proof of Theorem 1.15.6 is based on the Functional Central Limit Theorem. (iii) Weak convergence of random walks to fBm in Besov spaces (in the scheme of series). Consider a random walk {Xn }n≥1 consisting of stationary Gaussian random variables with zero mean and correlations r(i−j) := EXi Xj . Recall that a positive function ϕ(x), x ≥ a for some a > 0 is said to be slowly varying at ∞ if for all t > 0 limx→∞ ϕ(tx)/ϕ(x) = 1. Denote by D = D[0, 1] the Skorohod space of right-continuous functions on the interval [0, 1] that have left-hand limits, and equip D with the metric d(x, y) := inf{ε > 0 : ∃λ ∈ Λ such that λ < ε and sup |x(t) − y(λ(t))| ≤ ε}. t

Here λ := sups=t | log(λ(t) − λ(s))/(t − s)| and Λ := {λ : [0, 1] → [0, 1], λ is strictly increasing and continuous mapping of [0, 1] into itself}.


81

Under this metric D is a separable and complete metric space, and we denote D by −→ the convergence in the Skorohod topology, which is the weak topology D induced by this metric. That is, X n −→ X if Eψ(X n ) → Eψ(X) as n → ∞ for any bounded and continuous ψ : D → R. We start with the following result of Taqqu: Lemma 1.15.9 ((Taq75)). Let {Xn }n≥1 be a stationary Gaussian sequence with mean 0 and correlations r(i − j) = EXi Xj . Assume that n

r(i − j) ∼ n2H ϕ(n)

as n → ∞,

(1.15.21)

i,j=1

with 0 < H < 1, ϕ slowly varying. [nt] D H H Then Zn −→ B , where Zn (t) = d−1 with dn ∼ n2H ϕ(n), B n i=1 Xi is an fBm with Hurst index H, not necessarily normalized. Remark 1.15.10. Condition (1.15.21) is satisfied for H ∈ (1/2, 1) when r(k) ∼ 2α−1 k 2α−1 ϕ(k), ϕ(k) as k → ∞ with ∞and for H ∈ (0, 1/2), when r(k) ∼ −k r(0) + 2 k=1 r(k) = 0. Further, define for a function f ∈ Lp [0, 1] the modulus of continuity in Lp [0, 1]: 1/p p ωp (f, t) := sup |f (x + h) − f (x)| dx , |h|≤t

Ih

where Ih := {x ∈ [0, 1], x + h ∈ [0, 1]}. Now, for 0 < γ < 1 and β > 0, we consider a real function ωβγ : (0, 1] → R of the form ωβγ (t) := tγ (1 + log 1/t)β , t ∈ (0, 1], and denote f p, ωβα := f Lp [0,1] + sup ωp (f, t)/ωβγ (t). 0 0 and p ≥ 1/H ∨ 1/β. The next result is proved in (BL01). Theorem 1.15.11. Let H ∈ (0, 1), β > 0, p > 1/H ∨ 1/β, and let {Xn }n≥1 be a stationary Gaussian sequence with mean 0 and correlations r(i − j) = E Xi Xj . Assume that n

r(i − j) ∼ Cn2H

as n → ∞,

where C > 0.

i,j=1

Then C −1/2 Zn → B H as n → ∞ weakly in the space lipp (H, β).

82


(iv) Convergence of martingale differences to fBm. We follow here Nieminen’s paper (Nie04), which generalizes the result from (Sot01). Consider the following scheme of series: let (Ω, F, P ) be a probability space, (Xi,n , Fi,n )n≥1 , 1 ≤ i ≤ n be a sequence of square integrable martingale-differences, i.e., 2 < ∞, E(Xi,n /Fi−1,n ) = 0, F0,n = (∅, Ω), Xi,n is Fi,n -adapted, EXi,n Fi,n ⊂ Fi+1,n ⊂ F. Consider the sequence of kernels for H ∈ (1/2, 1) s [nt] , u du Z (n) (t, s) = n mH n s−1/n for s ∈ [1/n, 1] and t ∈ [0, 1], where [x] = k for k ≤ x < k + 1, k ∈ Z. Define the processes Wtn :=

[nt]

Xi,n ,

t ∈ [0, 1],

i=1

and

t

Z (n) (t, s)dWsn =

Ztn := 0

[nt] n i=1

i/n

mH i−1/n

[nt] (n) , u du · ξi . n

Theorem 1.15.12 ((Nie04)). Let limn→∞ n(Xi,n )2 = 1 a.s., 1 ≤ i ≤ n and max1≤i≤n |Xi,n | ≤ Cn−1/2 a.s. for some C ≥ 1. D

Then Z n −→ B H , n → ∞, where the convergence is in D[0, 1]. (n)

In the case when ξi are i.i.d. random variables, the corresponding result is proved by Sottinen (Sot01) under weaker conditions. (n)

(n)

Theorem 1.15.13 ((Sot01)). Let ξi = 0, Dξi the Skorohod space D[0, T ] for any T > 0.

D

= 1. Then Zn −→ B H in

(v) Convergence of integral functionals. Using Theorem 1.15.13, we can prove the result, similar to limit theorems for integral functionals on random walks, established in (SS70) and (Yos78). For example, (Yos78) considers sufficient conditions for n−1 i Si ξi+1 D 1 √ −→ ,√ fn f (t, Wt )dWt , n n n 0 i=1 where ξi is a sequence of martingale differences, Si =

i

ξk , Wt is a Wiener

k=1

process. For technical simplicity, we consider i.i.d. random variables and the interval [0, 1]. Let {fn }, n ≥ 1, fn : R → R be the sequence of functions satisfying the conditions 1) fn , f ∈ C 1 (R) and ∀R > 0 ∃MR > 0 such that


83

sup sup (|fn (x)| + |fn (x)|) ≤ MR ;

n≥1 |x|≤R

(r)

(r)

2) fn ⇒ f uniformly on any [−R, R]. Let πr := {0 = t0 < t1 < · · · < (r) tpr = 1} be the sequence of partitions of [0, 1], |πr | → 0 asr → ∞. Denote (r) (r) and define − Zn ni , ∆Zn,j,r := Zn tj+1 − Zn tj ∆Zn ni := Zn i+1 n the sequence of integral sums Sn (πr ) :=

p r −1

(r) fn Zn (tj ) ∆Zn,j,r .

j=1

Lemma 1.15.14. Under the conditions of Theorem 1.15.13 2 p n−1 r −1 i 2 = P -lim lim (∆Zn,j,r ) = 0. P -lim ∆Zn n→∞ r→∞ n→∞ n i=1 j=1 Proof. We can prove even the convergence in L1 (P ). For this purpose, we can rewrite the difference Zn (t2 ) − Zn (t1 ) for any 0 ≤ t1 < t2 ≤ 1 in the form 1 ] nk √ [nt (n) mH [ntn2 ] , s − mH [ntn1 ] , s ds · ξk Zn (t2 ) − Zn (t1 ) = n k−1 √ + n

[nt 2 ] k=[nt1 ]+1

n

k=1

k n k−1 n

mH

[nt2 ] n ,s

(n)

ds · ξk .

m n mH nl , s ds, and Denote αn (m, l) := m−1 n αn (m, l2 ) − αn (m, l1 ), m ≤ l1 ≤ l2 , . Then βn (n, l1 , l2 ) := l1 ≤ m ≤ l2 . αn (m, l2 ), 2 ] √ [nt (n) Zn (t2 ) − Zn (t1 ) = n βn (k, [nt1 ], [nt2 ]) ξk , and k=1

E|Zn (t2 ) − Zn (t1 )|2 = n =n

[nt 1 ] k=1 [nt 2 ]

k n k−1 n

βn2 (k, [nt1 ], [nt2 ])

k=1

[nt2 ] n ,s

− mH

[nt1 ] n ,s

2 ds

2 mH [ntn2 ] , s ds k=[nt1 ]+1 2 2 [ntn1 ] [nt2 ] ≤ 0 ds + [ntn1 ] mH [ntn2 ] , s ds mH [ntn2 ] , s − mH [ntn1 ] , s n 2 [nt1 ] [ntn2 ] α (5) α−1 du ds = (CH )2 0 n s−2α [nt1 ] u (u − s) n 2 [nt2 ] [ntn2 ] α α−1 + [ntn1 ] u (u − s) du ds s n 2H = E|BH (t2 ) − BH (t1 )|2 = [ntn2 ] − [ntn1 ] ≤ (t2 − t1 )2H . +n

mH

[nt 2 ]

k n k−1 n

(1.15.22)

84


From (1.15.22) E E

p r −1

2

n−1

∆Zn

i=1 p r −1

|∆Zn,j,r | ≤

i=1

j=1

i 2 n−1 −2H ≤ n → 0, n

n→∞

and

i=1

(r)

(r)

(tj+1 − tj )2H → 0,

r → ∞.

Lemma 1.15.15. Under conditions 1) and 2)

n−1 i i ∆Zn lim lim sup P Sn (πr ) − f Zn >δ =0 r→∞ n→∞ n n i=1

for any δ > 0.

Proof. Let a function F : R → R be such that F (x) = f (x), x ∈ R. Then by the Taylor formula F (Zn (1)) − F (0) =

n−1 i=0

F Zn i+1 − F Zn ni n

n−1 2 = f Zn ni ∆Zn ni + 12 f (θi,n ) ∆Zn ni , i=0 i=0 p r −1 (r) (r) F Zn tj+1 − F Zn tj F (Zn (1)) − F (0) = n−1

=

p r −1 j=0

j=0

(r)

f Zn tj

∆Zn,j,r +

1 2

p r −1

f (θn,j,r ) (∆Zn,j,r ) , 2

j=0

i i+1 where the points θi,nare between n n and Zn n , and the points θn,j,r Z (r) (r) and Zn tj+1 . Therefore are between Zn tj n−1 Sn (πr ) − f Zn i ∆Zn i ≤ n n i=1 −1 p r 2 + 21 |f (θn,j,r ) | |∆Zn,j,r | ,

1 2

n−1 i=0

2 |f (θi,n )| ∆Zn ni

j=0

and for any δ > 0 n−1 i P Sn (πr ) − f Zn n ∆Zn ni > δ ≤ P sup0≤t≤1 |Zn (t)| ≥ R i=1 n−1 n−1 2 2 ∆Zn ni ≥ M2δR + P (∆Zn,j,r ) ≥ M2δR . +P i=0

i=0

(1.15.23) Note that Zn −→ B , and functionals sup and inf in the Sko are continuous rohod topology, whence P sup0≤t≤1 |Zn (t)| ≥ R → P sup0≤t≤1 |BtH | ≥ R , and the last probability tends to 0 as R → ∞, according to (Sin97). The proof follows now from Lemma 1.15.14 and (1.15.23).

D

H

Theorem 1.15.16. Under the conditions of Lemma 1.15.15


1 i i d fn Zn −→ f (BtH )dBtH , ∆Zn n n 0 i=1

n−1

85

n → ∞,

d

where −→ denotes here the convergence in distribution. 1 Remark 1.15.17. The existence of integral 0 f (BtH )dBtH for H ∈ (1/2, 1) and f ∈ C 1 (R) follows from (Zah98) (see also Section 2.1), and this integral is a limit a.s. of Riemann–Stieltjes sums. Proof. Consider the difference

1

f (BtH )dBtH −

∆n := 0

i i fn Zn ∆Zn n n i=1

n−1

4 1 (j) (1) ∆n,r , where ∆n,r = 0 f (BtH )dBtH − and write it in the form ∆n := j=1 p r −1 H H f B H(r) ∆Bj,r is independent of n, ∆Bj,r = B H(r) − B H(r) . j=1

tj

tj+1

∆(2) n,r

p r −1

=

j=1

∆(3) n,r

=

p r −1 j=1

∆(4) n,r =

p r −1

(n)

(n) f (Z (r) )∆Zj,r t

(n)

(r) tj

j=1

Z

(n)

fn Z

(n)

f

−

(n)

(r)

tj

j=1

)∆Zj,r −

(n)

∆Zj,r ,

(r)

tj

j=1 p r −1

j

fn (Z

p r −1

−

H f (BtH(r) )∆Bj,r j

p r −1

tj

(n)

∆Zj,r ,

(n) (n) ∆Z i . fn Z i n

j=1

n

(1)

From the result of Z¨ ahle (Zah98) cited above, P -limr→∞ ∆n,r = 0. By (4) Lemma 1.15.15 P -limr→∞ ∆n,r = 0. (2) As to ∆n,r , we have from the weak convergence of Z (n) to B H that p r −1

f

Z

(n) (r)

tj

j=1

(n)

d

∆Zj,r −→

p r −1

H f (BtH(r) )∆Bj,r , j

j=1

(3)

as n → ∞, for any fixed r ≥ 1. We must estimate now ∆n,r . The technique here is similar to x 1.15.15. xthe proof of Lemma Let F (x) = 0 f (t)dt, Fn (x) = 0 fn (t)dt. Then p r −1

(n)

F (Z1 ) = =

p r −1 j=1

(n) (n) F Z (r) − F Z (r) tj+1

j=1

f (Z

(n)

(r) tj

(n)

)∆Zj,r +

1 2

p r −1 j=0

tj

f θj,r

(n)

(n)

∆Zj,r

2

(1.15.24) ,

86


and, similarly, (n)

Fn (Z1 ) =

p r −1

fn (Z

(n)

(r) tj

j=1

(n)

)∆Zj,r +

1 2

p r −1 j=0

2 (n) (n) ∆Zj,r fn θj,r ,

(1.15.25)

(n) (n) (n) (n) where θj,r and θj,r are between Z (r) and Z (r) . Now, tj

(n)

(n)

tj+1

(n)

|F (Z1 ) − Fn (Z1 )| ≤ |Z1 |

|fn (t) − f (t)|,

sup (n)

|t|≤|Z1

|

whence (n)

(n)

(n)

P {|F(Z1 ) − Fn (Z1 )| ≥ δ} ≤ P {|Z1 | ≥ R} + P sup|t|≤R |fn (t) − f (t)| ≥

δ R

(1.15.26)

(The last event is not random.) Since fn uniformly converges to f on [−R, R], the last term in (1.15.26) is zero for all sufficiently large n, and (n) limn→∞ P {|Z1 | ≥ R} = P {|B1H | ≥ R} ≤ R12 . Therefore, from (1.15.24)– (1.15.26) and Lemmas 1.15.14–1.15.15 P -lim lim ∆(3) n,r = 0, r→∞ n→∞

and the theorem is proved.

Remark 1.15.18. The paper (Wang03) contains a result on a weak convergence to fBm in the Brownian scenery. (vi) fBm as a weak limit of Poisson shot noise processes. Let for all n ∈ Z\{0} Xn be i.i.d.r.v. with EX1 = 0 and EX12 ∈ (0, ∞), g : R+ → R be a continuously differentiable function with g (u) = O(u−1/2−ε ), u → ∞ for some ε > 0. Consider the special model of multiplicative shots: Xi (u) = g(u)Xi , u ≥ 0, and a shot noise model, which is defined as

N (t)

S(t) =

i=1

Xi (t − Ti ) +

[Xi (t − Ti ) − Xi (−Ti )], t ≥ 0,

i≤−1

where N is a two-sided homogeneous Poisson process with the rate α > 0 and points · · · < T−2 < T−1 < 0 < T1 < T2 < · · · . For t = 0 we put S(0) = 0. According to (KK04), the multiplicative process with the above restrictions on g and Xi exists and has the following sample path properties. Lemma 1.15.19. The process S possesses a right-continuous version with left limits on R+ and has a finite variation on any [0, T ], T > 0. Therefore, it is a semimartingale with respect to its natural filtration.

1.16 H¨ older Properties of fBm and of Wiener Integrals

87

Now, suppose that limu→∞ ug (u)/g(u) = γ with γ ∈ (0, 1/2). Introduce the rescaled process S(x, t) =

S(xt) , x ∈ [0, ∞), t > 0, σ(t)

where σ 2 (t) = V ar(S(t)). Theorem 1.15.20. Under the above assumptions, S(·, t)−→B H ,

t→∞

when the convergence is in D[0, ∞) with the metric of uniform convergence on compacts, and H = 1/2 + γ.

1.16 H¨ older Properties of the Trajectories of fBm and of Wiener Integrals w.r.t. fBm Let {ξt , t ∈ [0, T ]} be a separable modification of Gaussian process, ρ2ξ (s, t) = E(ξs −ξt )2 , G = G(x) : R+ → R+ be a continuous increasing function, G(0) = ε 0, D(T, ε) = 0 H(T, u)1/2 du be the Dudley integral (see Section 1.10), ρ(s, t) be some semi-metric in [0, T ]. Definition 1.16.1. A function Θ = Θ(x) : R+ → R+ is called a modulus of continuity if Θ(0) = 0 and for any x1 , x2 ≥ 0 Θ(x1 ) ≤ Θ(x1 + x2 ) ≤ Θ(x1 ) + Θ(x2 ). Definition 1.16.2. Let g : [0, T ] → R be some function. The function ε → ∆ρ (g, ε) :=

sup |g(s) − g(t)| ρ(s, t) ≤ ε s, t ∈ [0, T ]

is called a modulus of uniform continuity of the function g with respect to the semi-metric ρ. Definition 1.16.3. A modulus Θ(·) is called a uniform modulus of a Gaussian process ξ with respect to the semi-metric ρ if for a.a. ω ∈ Ω lim sup ∆ρ (ξ. (ω), ε)/Θ(ε) < ∞. ε→0

The next result is formulated in the book (Lif95). Theorem 1.16.4. 1. Let for any s, t ∈ [0, T ] ρξ (s, t) ≤ G(ρ(s, t)).

(1.16.1)

Then the function Θ(ε) := D(T, G(ε)) is a uniform modulus of the Gaussian process ξ with respect to the semi-metric ρ.

88


2. Under assumption (1.16.1) with ρ(s, t) = |s − t|, the function ε Θ(ε) = | log r|1/2 dG(r) 0

is a uniform modulus of the Gaussian process ξ with respect to ρ. Definition 1.16.5. We say that the function f : [0, T ] → R belongs to the space C β− [0, T ] if f ∈ C γ [0, T ] for any γ < β. Let ξt = BtH be an fBm with Hurst index H ∈ (0, 1). Then, evidently, we can take G(x) = xH , so from the second statement of previous theorem, the function Θ(ε) ∼ εH | log ε|1/2 will be a uniform modulus of B H on any [0, T ]. In particular, |BtH − BsH | ≤ c(ω)|t − s|H−β for any 0 < β < H, i.e. t B H ∈ C H− [0, T ] for a.a. ω and any T > 0. Now, let ξt = It (f ) = 0 f (s)dBsH with f ∈ LH 2 [0, t] for any 0 ≤ t ≤ T , H ∈ (1/2, 1). We can take ρ(s, t) = t 1 H H du, G(x) = C |f (u)| Hx , s ∆ρ (I, ε) =

sup t s

0≤s 0 the random variable that is the right-sided Riemann– Liouville fractional derivative of order β (in Weyl representation) of fBm B H , where 1 − H < β < 1/2 and H ∈ (1/2, 1): Gt :=

1 sup |D1−β B H (s)|, t ∈ [0, T ]. Γ (β) 0≤s 0 EGpt < ∞. Proof. By the Garsia–Rodemich–Rumsey inequality (GRR71), for any p ≥ 1 and ρ > p−1 there exists a constant Cρ,p > 0 such that for any continuous function f on [0, T ] and for all s < z ≤ t ∈ [0, T ] z z |f (x) − f (y)|p |f (z) − f (s)|p ≤ Cρ,p |z − s|ρp−1 dx dy. |x − y|ρp+1 0 0 Choose ε < β − (1 − H) and put ρ = H − 2ε , p =

2 ε

and f (t) = BtH :

|BzH − BsH | ≤ CH,ε |z − s|H−ε ξt,ε , where

t ξt,ε = 0

0

t

2ε

2

|BxH − ByH | ε |x − y|

2H ε

dx dy

, 0 < ε < H.

(1.17.1)

Since BxH − ByH is a Gaussian random variable, and E|BxH − ByH |2 = |x − y|2H , we have that for the random variable ξt,ε for any q > 1 q 2ε t t |BxH −ByH | 2ε q E|ξt,ε | = E 0 0 dx dy 2H |x−y| ε H H q T T E|Bx −By | ≤ Cq,H,T 0 0 dx dy ≤ Cq,H,T , |x−y|Hq which means that all moments of ξt,ε are finite. Further, for ε < β − (1 − H) z |BzH −BsH | Gt ≤ Cβ sup0≤s 1 from the H¨ older and Burkholder inequalities

t b q E|ξt,δ | =E 0

t ≤ Cq,t 0

0

t

E|

t

|

y

0

y x

b dWu |2/δ x u dx dy |x − y|1/δ

bu dWu |q dx dy ≤ Cq,t |x − y|q/2

t 0

0

t

|

y x

qδ/2

b2u du|q/2 dx dy ≤ Cq,t , |x − y|q/2

b ξt,δ

is continuous and strictly increasing, so, our Wiener Note that the process integral with respect to the Wiener process is dominated by a strictly increasing process with all moments bounded on [0, T ].

1.18 Power Variations of fBm and of Wiener Integrals w.r.t. fBm We start here with the simple result obtained by Rogers in (Rog97). Consider for fBm {BtH , t ≥ 0} with H ∈ (0, 1) and for p > 0 the sums 2n p H n(pH−1) Sn,p (t) = , B jtn − B H (j−1)t · 2 j=1

2

and Sñ,p (t) = 2−n

(1.18.1)

2n

2n p H H Bjt − B(j−1)t . j=1

Then Law(Sn,p (t)) = Law(Sñ,p (t)) (i.e., these sums have identical disH , t > 0) = tribution), due to the self-similarity property of B H : (Law(Bct H H Law(c Bt , t > 0)). H The sequence (BkH − Bk−1 )k∈N is stationary. Therefore, from the ergodic theorem Sñ,p (t) → E|BtH |p =: Cp tpH as n → ∞

1.18 Power Variations of fBm and of Wiener Integrals w.r.t. fBm

91

with probability 1 and in L1 (P ), whence d

→ Cp tpH , n → ∞, Sn,p (t) −

(1.18.2)

P

→ Cp tpH , n → ∞. so Sn,p (t) − From (1.18.1)–(1.18.2)

⎧ ⎪ 2n p ⎨0, H P H → +∞, B jtn − B (j−1)t − 2 ⎪ 2n ⎩ H 1/H j=1 E Bt ,

p> p
0 (recall that a function is regularly varying if ψ(xt) ψ(t) → ρ(x) as t → ∞ and in this case β ρ(x) = x for some β ≥ 0). Then ψ(σ(t)) = t1+Hε and all the assumptions of Theorem 1.18.1 are satisfied. So, limδ→0 supπ∈Π(δ) S(|x|p , π, B H ) = const for 1 1 any p > H . Evidently, this constant is zero since for any p > p > H

S(xp , π, B H ) ≤

sup

0≤t 0, any small t and 0 ≤ h ≤ t. S(σ −1 (x),π,X) Then lim inf k→∞ supπ∈Π(k) ≥ 1 with probability 1. ˜ Φ( 1 ) k

H

BtH . 2

Then conditions (a), (b), (c) and (d) hold. Put σ(t) = t , Xt = Moreover, for H ∈ (0, 12 ) σ (t) − σ 2 (t − h) = t2α+1 − (t − h)2α+1 ≤ h2α+1 for all 0 ≤ h ≤ t ≤ 1. The function Φ(t) now has the form Φ(t) = 1

(2 log log 1t ) H , whence limk→∞ supπ∈Π(k) ˜

limk→∞ supπ∈Π(k) ˜

tj ∈π

|BtH −BtH j

j−1 1

1 |H

(2 log log k) 2H

1

S(|x| H ,π,B H ) 1

(2 log log k) 2H

= 1 or, in other words,

= 1.

For H ∈ (1/2, 1) we have no assumption (e), so, give only upper bounds. Namely, from the first statement of Theorem 1.18.2, we can deduce that 1 H H H tj ∈π |Btj − Btj−1 | lim supk→∞ sup ≤ 1. 1 (2 log log k) 2H ˜ π∈Π(k) Moreover, the following result holds. Theorem 1.18.3. Under assumptions (a)–(c) lim sup S(ψ(x), π, X) ≤ 1,

δ→0 π∈Π(δ)

3 with probability 1, where ψ(x) is the inverse function to σ(t) 2 log log the origin. In our case it means that lim sup

δ→0 π∈Π(δ)

tj ∈π

ψ(|BtHj − BtHj−1 |) ≤ 1,

1 t

near

1.18 Power Variations of fBm and of Wiener Integrals w.r.t. fBm

93

3

where ψ(t) is the inverse function to tH 2 log log 1t . Let, as before, Π be the set of all partitions of the interval [0, 1]. Definition 1.18.4. For any p > 0 define p-variation of the function f on the interval [a, b] as vp (f ) = sup S(|x|p , π, f ). π∈Π

Also, let p-variation index of the function f be v(f ) := inf(p : vp (f ) < ∞). The last relations mean that v(BH ) = H1 with probability 1, and, moreover, 1 1 and = ∞ for p < . vp (BH ) < ∞ for p > H H This result was obtained in (Nrv99) from another point of view. Let {Xt , t ≥ 0} be a Gaussian process with stationary increments and E|Xt+s − sγ = 0} and γ ∗ := sup{γ > 0 : Xt |2 = σ 2 (s). Let γ∗ := inf{γ > 0 : lims↓0 σ(s) γ

s lims↓0 σ(s) = ∞}. Then 0 ≤ γ ∗ ≤ γ∗ ≤ +∞. If γ ∗ = γ∗ then we say that the process Xt has the Orey index γ(X) = γ ∗ = γ∗ . Let Xt have the Orey index γ(X) ∈ (0, 1); then it follows from the results of Berman (Ber69) and also 1 . Evidently, from (JM83) that the p-variation index of Xt equals v(X) = γ(X) the Orey index of the fBm equals its Hurst index and equals H. t Now consider briefly the Gaussian process Xt = It (f ) = 0 f (s)dBsH . 1 Let H ∈ ( 2 , 1) and the function f is essentially bounded on [0, 1], ess sup0≤t≤1 |f (t)| = f ∗ . Then, according to Theorem 1.10.3, E|Xt − Xs |2 ≤ σ 2 (|t − s|), where σ 2 (t) = CH (f ∗ )2 t2α+1 , therefore from Theorem 1.18.1 limδ→0 supπ∈Π(δ) S(|x|p , π, I) = 0 for any p > H1 and from Theorems 1.18.2 and 1.18.3 1 S(|x| H , π, I) lim supk→∞ sup ≤ 1 P -a.s., (1.18.4) Φ( k1 ) ˜ π∈Π(k)

sup S(ψ(x), π, I) ≤ 1 P -a.s.

lim

δ→∞

(1.18.5)

˜ π∈Π(δ)

3 1/2 where ψ(x) is the inverse to CH f ∗ tH 2 log log Let f∗ := ess inf 0≤t≤1 f (t) > 0. Then t E|It − Is |2 = CH s P

s

1 t

near the origin.

t

f (u)f (v)|u − v|2α−1 du dv ≥ CH f∗2 |t − s|2α+1 ,

whence S(|x|p , π, I) − → ∞ as |π| → 0 and p < rem 1.18.1 it means that

1 H,

lim sup S(|x|p , π, I) = ∞ P -a.s.,

δ→0 π∈Π(δ)

and together with Theo-

p
0 we can immediately conclude from Theorem 1.9.1 that E|It − Is |2 ≥ CH f 2L 1 [s,t] ≥ CH f∗2 |t − s|2α+1 , H

P

whence S(|x|p , π, I) − → ∞ as |π| → 0 and p < can deduce from Remark 1.10.7 that

1 H.

Let f ∈ C β [0, 1]. Then we

E|It − Is |2 ≤ CH f C β ([0,1]) ((t − s)2α+1 + (t − s)2H+2β ), whence (1.18.4)–(1.18.5) follow for H ∈ (0, 12 ).

t Remark 1.18.5. In the paper (CNW06) the process of the form 0 us dBsH is considered where us is a stochastic process with paths of finite q-variation and the integral is pathwise Riemann–Stieltjes integral (construction of such integrals is described in Section 2.1). The convergence in probability of the normalized power variations of these integrals is established and their deviations are considered. Remark 1.18.6. Modern results on power variation of the integrals and other processes related to fBm are established in (GuNu05), (Nrv99), (CNW06), (DN99).

1.19 L´ evy Theorem for fBm The idea of this problem belongs to E. Valkeila. The results are published in (MV06). We start with the classical Lévy theorem: Theorem 1.19.1. Let {µ(t), t ≥ 0} be a continuous local martingale with the angle bracket µt = t. Then µt is the Wiener process. The natural question is: how can the fBm be characterized in a similar way or by some other properties? Let {Ω, F, {Ft }t≥0 , P } be some stochastic basis, {Xt , t ≥ 0} be a stochastic process (not necessarily adapted, as for beginning). For any t > 0, denote tk := t nk , 1 ≤ k ≤ n. The main result of this section is: Theorem 1.19.2. Let the process Xt satisfy the following conditions: (a) trajectories of X are H¨ older of any order 0 < β < H, where 0 < H < 1; n (b) n2α k=1 (Xtk − Xtk−1 )2 → t2α+1 for any t > 0 in the space L1 (P ), as n → ∞. t (c) the process Mt := 0 s−α (t − s)−α dXs is an Ft -adapted continuous squareintegrable martingale, where α = H − 1/2. Then Xt is an Ft -adapted fBm with Hurst index H.

1.19 Lévy Theorem for fBm

95

Proof. We shall divide the proof into several steps. First, t consider the case H ∈ ( 12 , 1). Let the square-integrable martingale Wt := 0 sα dMs , t ∈ [0, T ], t T > 0 and the process Yt := 0 s−α dXs . For convenience we put T = 1. We can establish the existence in the pathwise sense of the latter integral using H¨ older properties of X and integration by parts. Evidently, t

Mt =

(t − s)−α dYs .

(1.19.1)

0

Lemma 1.19.3. The process Xt admits the representation + t * t 1 Xt = sα (s − u)α−1 ds u−α dWu , CH 0 u where CH = B(α, 1 − α). Proof. Equation (1.19.1) is a generalized Abel integral equation and has the formal solution t 1 Yt = (t − s)α−1 Ms ds. (1.19.2) CH 0 It is very easy to check that (1.19.1) becomes an identity, if we substitute (1.19.2) into (1.19.1), rewritten as Mt = t

−α

t

Yt + α

(t − s)−1−α (Yt − Ys )ds.

(1.19.3)

0

Moreover, the corresponding homogeneous equation t 0 = t−α Yt + α (t − s)−1−α (Yt − Ys )ds, 0

has only a zero solution, whence Yt admits the representation (1.19.2). Further,

t

0

=

tα CH

t

sα dYs = tα Yt − α

Xt =

0

t

(t − s)α−1 Ms ds − 0

1 = CH

sα−1 Ys ds

t * 0

t

α CH

t

0

s

(s − u)α−1 Mu du ds

sα−1 0

+ sα (s − u)α−1 ds dMu .

u

Remark 1.19.4. From Lemma 1.19.3, for any 1 ≤ k ≤ n, it follows that tk tk 1 α α−1 Xtk − Xtk−1 = u (u − s) du dMs CH 0 s

96


tk−1

− 0

1 = CH

tk−1

s

α−1

du dMs

uα (u − s)α du dMs

tk

tk−1

tk

+ tk−1

u (u − s)

tk−1

0

tk

α

uα (u − s)α−1 du dMs

.

(1.19.4)

s

Denote

tk

ϕtk (s) :=

uα (u − s)α−1 du,

tk−1

and

ψkt (s) :=

tk

uα (u − s)α−1 du.

s

Then Xtk := Xtk − Xtk−1

1 = CH

Now, let 0 < s < t, and let

s t

tk−1

ϕtk (s)dMs

tk

ψkt (s)dMs

+

0

.

tk−1

be a rational number, such that

s t

(1.19.5) ∈ Q.

Lemma 1.19.5. Let n ∈ N be an increasing sequence, such that n st ∈ N, n. tk := tk/ n L1 (P ) Then n 2α k=n s +1 (Xtk − Xtk−1 )2 −−−−→ t2α (t − s), n → ∞. t

Proof. Evidently, n t s

n

2α

∆Xtk

2

k=1

2α 2 n t s 2α t 2α t = n · → s2α+1 · = st2α . ∆X sks n t s s t s

k=1

2 n We know from condition (b) that n 2α k=1 ∆Xtk → t2α+1 , whence the claim follows.

2 n Now we want to estimate n 2α k=n s ∆Xtk in terms of the angle t bracket M , by using representations (1.19.4)) and (1.19.5). In order to do this, rewrite the increment of the process X in the form

∆Xt k

n

1 = CH

tk−2

ϕtk (s)dMs

0

+

ϕtk (s)dMs

tk−2

=: Evidently,

tk−1

1 (I k + I2k + I3k ). CH 1

tk

+ tk−1

ψkt (s)dMs


α−1

ϕtk (s) ≤

tα k (tk−1 − s)

and

∧

n−α α

97

(1.19.6)

2 ∆Xt k

n

k= n st +1

⎛ n n 2α ⎝ = 2 CH s

t n

n

n 2α

·

⎞

k

(I1k )2 + (I2k )2 + (I3k )2 + 2I1k · I2k + 2I1k · I3k + 2I2k · I3 ⎠ .

k= n t +1

(1.19.7) Now we shall estimate the terms on the right-hand side of (1.19.7). Lemma 1.19.6. There exist two constants C1 > 0, C2 > 0 such that C1 t

n

t

u dM u ≤ P -lim( n

2α

2α

2α

n →∞

s

(I1k )2 ) ≤ C2 t4α (M t − M s ).

k= ns t +1

Proof. For simplicity, we shall omit ∼, and consider only such n that n st ∈ N. From the Itˆ o formula for square-integrable martingales, it follows that 2 tk−2 tk−2 (I1k )2 = ϕtk (u)dMu = (ϕtk (u))2 dM u 0

0

tk−2

u

ϕtk (v)dMv · ϕtk (u)dMu .

+2 0

0

First, we estimate n

S1n := n2α

k=n st +2

tk−2

(ϕtk (u))2 dM u .

0

From (1.19.6), we obtain that

tk−2

(ϕtk (u))2 dM u

≤t

0

2α

2α 2 tk−2 k t 2α−2 (tk−1 − u) dM u . n n2 0

So, the estimate of S1n from above has the form S1n ≤ n2α−2 t2α+2

n k=n st +2

2α tk−2 k 2α−2 (tk−1 − u) dM u . n 0

Now, we rewrite the sum in (1.19.8) for 0 < s < t and 2 ≤ n st ≤ n − 3:

(1.19.8)

98


n

n S11 :=

=⎝

s

nt

n

i=1

k=n st +2

+

n−2

+

n−2 i=n st +1

n

i=n st +1

ti

ti−1

2α−2

(tk−1 − u)

dM u

0

k=n st +2

⎛

tk−2

⎞ ⎠

ti

2α−2

(tk−1 − u)

dM u

ti−1

k=i+2

n

2α−2

(tk−1 − u)

dM u .

(1.19.9)

k=i+2

Evidently, 1 n

n

(tk−1 − u)

2α−2

s+t−u

≤

x2α−2 dx · s−u

k=n st +2

1 1 ≤ (s − u)2α−1 · (1 − 2α)−1 , t t

and n 1 1 1 2α−2 2α−2 2α−1 (ti+1 − u) (tk−1 − u) ≤ (ti+1 − u) + . n n (1 − 2α)t k=i+2

We substitute these estimates into (1.19.9): n st t i 1 n n S11 ≤ (s − u)2α−1 dM u 1 − 2α i=1 ti−1 t * ti n 2α−2 2α−1 + 1 1 ti+1 − u ti+1 − u +n + dM u n (1 − 2α)t i=n st +1 ti−1 1 n s (M t − M s ). ≤ (s − u)2α−1 dM u + t2α−2 n2−2α 1 + t 0 1 − 2α

We return to (1.19.8) and obtain that s (s − u)2α−1 dM u + t4α S1n ≤ n2α−1 t2α+1 0

1 + 1 (M t − M s ). 1 − 2α

Note that the martingale M is Hölder continuous up to order 12 , so W is H¨ older continuous up to order 12 , W is Hölder continuous up to 1, and the integral s

s

(s − u)2α−1 dM u = 0

0

(s − u)2α−1 u−2α dW u

s exists. Therefore, n2α−1 0 (s − u)2α−1 dM u → 0, n → ∞. We obtain that limn→∞ S1n ≤ C2 t4α (M t − M s ). Now we estimate S1n from below: first,


99

2 t2 ϕtk (u) ≥ (tk−1 )2α (tk − u)2α−2 · 2 . n

Then,

n

S1n ≥ n2α−2 t2 nt s

=n

t

i=1

2α−2 2

+n

t

n−2 i=n st +1

ti

ti−1

ti

2α

(tk−1 )

k=n st +2

2α−2 2

tk−2

2α−2

(tk − u)

dM u

0

(tk−1 )2α (tk − u)2α−2 dM u

n k=n st +2

n

ti−1

(tk−1 )2α (tk − u)2α−2 dM u .

(1.19.10)

k=i+2

Consider the interior sum of the second term: n 1 2α 1 t−u 1 2α 2α−2 (tk−1 ) (tk − u) ≥ x2α−2 x + u − dx n t ti+2 −u n k=i+2

t−u 1 2α x2α−2 dx ≥ (ti+1 ) t ti+2 −u 2α 2α−1 i+1 − (t − u)2α−1 (ti+2 − u) . · ≥ t2α−1 n 1 − 2α So, S1n

t2α+1 n2α−1 ≥ 1 − 2α

n−3 i=n st +1

i+1 n

2α

1

ti

2α−1

(ti+2 − u)

ti−1

2 − (t − u)2α−1 dM u . 2α−1

Consider the function f (u) := (ti+2 − u) [ti−1 , ti ]: f (u) ≥ (ti+2 − ti−1 )

2α−1

− (t − ti−1 )

2α−1

− (t − u)2α−1 on the interval

=

32α−1 − 42α−1 2α−1 t . n2α−1

Therefore, S1n

t2α+1 n2α−1 ≥ 1 − 2α

n−3 i=n st +1

ti

ti−1

i+1 n

2α

t2α−1 dM u n2α−1

100


n−3

×(32α−1 − 42α−1 ) ≥ C1 t2α

lim

n→∞

≥ C1 t

S1n

2t 2α dM u , n

u+

i=n st +1

and

ti

ti−1

t

u2α dM u ,

2α s

or, in terms of W , lim S1n ≥ C1 t2α (W t − W s ) .

n→∞

Now, we try to prove that S2n → 0 in probability, where

n

S2n = n2α

tk−2

0

k=n st +2

u

ϕtk (s)dMs ϕtk (u)dMu .

0

Evidently, it is sufficient to consider the sums of the form S3n = n2α

n k=2

tk−2

0

u

ϕtk (s)dMs ϕtk (u)dMu ,

0

because the sums n st +2 t k−2 k=2

0

u

ϕtk (s)dMs ϕtk (u)dMu

0

can be considered in a similar way. We use a very weak version of the Lenglart inequality: if N is a locally square integrable martingale on R, then for any ε > 0, A > 0 and T > 0 we have that P { sup |N (t)| ≥ ε} ≤ 0≤t≤T

A + P {N T ≥ A}. ε2

(1.19.11)

Rewrite S3n as S3n = n2α

n−2 ti i=1

n

ti−1

where ψuM =

k=i+2

ϕtk (s)dMs dMu = n2α

0

k=i+2 n

u

ϕtk (u)

* ϕtk (s)dMs ,

0

ψuM dMu ,

0

u

ϕtk (u)

2 1− n

u∈

i−1 i , n n

.

Since the martingale M is continuous (and square integrable), we can localize it: let for some L > 1 τL = inf{t > 0 : |Mt | ∨ M t ≥ L},


101

M t = Mt∧τL , M t = M t∧τL , ψ u = ψuM , τL = ∞ if |Mt | ∨ M ∞ < L for all t > 0. By (1.19.11), it is sufficient to prove that for any L > 0, n4α

2 t(1− n )

0

= n4α

2

ψ u dM u

n−2 ti ti−1

i=1

n

2

u

ϕtk (u)

P

dM u → 0,

ϕtk (s)dM s

n → ∞.

0

k=i+2

(1.19.12) u t n t First, we estimate the function ψu := k=i+2 ϕk (u) 0 ϕk (s)dM s = n u t t t k=i+2 ϕk (u)(ϕk (u)M u − 0 M s (ϕk (s))s ds). Evidently,

ϕtk (u) u

= (1 − α)

tk

v α (v − u)α−2 dv.

tk−1

Therefore, |ψu | ≤ L

n

(ϕtk (u))2 + L(1 − α)

k=i+2

n

u

ϕtk (u) 0

k=i+2

tk

v α (v − s)α−2 dv ds.

tk−1

Estimate the terms separately: ϕtk (u) ≤

tα+1 α−1 (tk−1 − u) , n

whence, n

(ϕtk (u))2 k=i+2 +

t2+2α n

n t2+2α t2+2α 2α−2 ≤ (t − u) ≤ (ti+1 − u)2α−2 k−1 n2 n2 k=i+2

1

(tx − u)2α−2 dx =

i+1 n

t4α t2α+1 (ti+1 − u)2α−1 ≤ Cn−2α , + 2α n n 1 − 2α

and n k=i+2

u

ϕtk (u) 0

tk

v α (v − s)α−2 dvds ≤ C

tk−1

n

ϕtk (u)

k=i+2

≤C

n

tk

v α (v − u)α−1 dv

tk−1

(ϕtk (u))2 ≤ Cn−2α .

k=i+2 2

From these estimates, it follows that ψ u n4α ≤ C. Therefore, there exists the bounded dominant. In order to establish (1.19.12), it is sufficient to prove P that ψ u n2α → 0, 0 < u < 1. We have that

102


n

E(ψ u n ) = n E 2α 2

4α

u

0

2

n

=n E

ϕtk (s)dM s 0

k=i+2

4α

2

u

ϕtk (u)

dM s .

ϕtk (u)ϕtk (s)

k=i+2

Similarly to previous estimates, we obtain that

2

n

4α

n

≤ Cn

4α

ϕtk (u)ϕtk (s)

k=i+2

n k=i+2

2 α−1

× (tk−1 − s)

≤ Cn

4α−2

1 α−1 (tk−1 − u) n2

n 1 2α−2 (tk−1 − u) n

2

k=i+2

≤ Cn4α−2

n2−2α + n1−2α n

2 ≤ C, for some C > 0.

This means that the bounded dominant exists. Moreover,

n2α

n

n

ϕtk (u)ϕtk (s) ≤ Cn2α

k=i+2

1 (u − s)α−1 n

ϕtk (u) ·

k=i+2

≤ Cn2α ·

1 n

1 i+1 n

v α (v − u)α−1 du · (u − s)α−1 → 0

P

for any s < u. This means that S3n → 0, and the lemma is proved. Lemma 1.19.7. There exists a constant C3 > 0, such that P -lim n2α n→∞

n

(I2k )2 ≤ C3 t4α (M t − M s ).

k=n st +2

Proof. We apply the Itˆ o formula to (I2k )2 and obtain that (I2k )2 =

tk−1

(ϕtk (s))2 dM s +

tk−2

tk−1

tk−2

s

ϕtk (u)dMu ϕkt (s)dMs .

tk−2

From (1.19.6) it follows that n k=n st +2

tk−1

tk−2

(ϕtk (s))2 dM s · n2α ≤ t4α C(M t − M s ).


103

Similarly to the estimates from Lemma 1.19.6, we obtain that for any A > 0 and ε > 0 tk−1 s n

P n2α ϕtk (u)dM u ϕtk (s)dM s ≥ ε tk−2

k=n st +2

≤

A + P n4α ε2

n

tk−2

tk−1

2

ϕtk (u)dM u )2 (ϕtk (s) dM s ≥ A .

tk−2

tk−2

k=n st +2

s

So, it is sufficient to prove that

n

4α

n

k=n st +2

tk−1

tk−2

s

ϕtk (u)dM u

2

P

(ϕtk (v))2 dM v → 0.

tk−2

The existence of the bounded dominant is established by the estimates: 2 s 4α n · ϕtk (u)dM u (ϕtk (s))2

tk−2

≤ n4α ϕtk (s)M s − ϕtk (tk−2 ) · M tk−2 − ≤ CL n

2 4α

ϕtk (s)

+

ϕtk (tk−2 )

s

(ϕtk (u))u M u du

2

tk−2 tk

+ tk−2 2 4α

s

v α (v − u)α−2 dv du

2

· (ϕtk (s))2

· (ϕtk (s))2

tk−1

≤ 9CL2 n4α (1/n)4α · t4α ≤ CL t .

Therefore, we must prove, that for any s < v < t s P 2α n ϕtk (u)dM u (ϕtk (v)) → 0,

n → ∞.

tk−2 4α

Here (ϕtk (v))2 ≤ nt 2α . Taking into account that M is bounded and continuous, and by using the relation s 4α t 2 n (ϕk (v)) E (ϕtk (u))2 dM u dM u ≤ CE(M s − M tk−2 ) → 0, tk−2

for s < tk−1 , we obtain the necessary estimates, whence the proof follows. Lemma 1.19.8. There exists a constant C4 > 0 such that P -lim n2α n→∞

n

(I3k )2 ≤ C4 (M t − M s ) · t4α .

k= ns t +1

The proof is similar to Lemma 1.19.7.

104


Lemma 1.19.9. We have that lim n2α

n

n→∞

Iik Ijk = 0

k

in probability. Proof. Consider, for example, n2α of M . But in this case, n4α E

n

I1k I2k

n

k k k=1 I1 I2 ,

2

= n4α E

k=1

where we substitute M instead

n

(I1k )2 (I2k )2 ,

k=1

tk−1 t (ϕk (s))2 dM s , since I1k , I2k , I3k are pairwise orthogonal. where I2k = tk−2 Moreover, from inequality (1.19.11), it follows that we must only prove the relation n P n4α (I1k )2 (I2k )2 → 0. k=1

According to Lemma 1.19.6, we have that P -lim n2α n→∞

n

(I1k )2 ≤ C2 t4α M t

k=1

and

tk−1

n2α max

1≤k≤n

(ϕtk (s))2 dM s ≤ α−2 max (M tk−1 − M tk−2 ) → 0. P

1≤k≤n

tk−2

All other terms can be estimated similarly, whence the claim follows.

By using our estimates, we can conclude that for rational s, consequently for any s < t, the following claims hold: (a) there exist two constants, C1 > 0 and C2 > 0 such that

t

u2α dM u ≤ (t − s) ≤ C2 t2α (M t − M s ).

C1 s

This estimate can be rewritten in terms of W and W :

t

C1 (W t − W s ) ≤ (t − s) ≤ C2 t2α

u−2α dW u .

s

(b) P -lim n2α n→∞

n k=n st +1

(Xtk )2 = P -lim n→∞

t

ϕns dM s , s


105

where ϕns is a positive, bounded, nonrandom function, separated from 0 by some constant. From the left-hand side of (a), it follows that W t is absolutely continu t ous w.r.t. the Lebesgue measure, so W t = 0 θs ds, where θs is a bounded, possibly, random variable. From the right-hand side of (a), it follows that t t 1 1−2α −2α −2α 1−2α 1−2α u θu du ≥ (t − st ) ≥ C3 (t −s ) = C3 u−2α du. C2 s s This means that

t

u−2α (θu − C3 )du ≥ 0.

s

Evidently, for any set A ∈ F t A

u−2α (θu − C3 )du dP ≥ 0.

s

Now, let the set D ∈ σ{F × B[δ, 1]}, and let δ > 0 be fixed. Then µ(D) < ∞, where µ = P × λ, λ is the Lebesgue measure on [0, 1]. By the theorem of approximation of measurable sets, for any ε > 0 there exists a collection of the sets {Di = Bi × [si , ti ], Bi ∈ F, [si , ti ] ∈ B[δ, 1]}, such that

k k 4 4 4 µ D\ Di ) ( Di \D < ε. i=1

Therefore, since u

−2α

i=1

(θu − C3 ) is bounded on D, u−2α (θu − C3 )dµ ≥ 0.

(1.19.13)

D

Now, set D = {(ω, u) : θu − C3 < 0, and u ≥ δ} and we immediately obtain that µ(D) = 0. From here we conclude that W is t 1 equivalent to the Lebesgue measure, and Wt = 0 θs2 dVs , where {Vs , Fs , s ≥ 0} is some Wiener process. Now, if we do all the same calculations as before, but for “true” fractional Brownian motion BtH , we obtain that t n 2α H 2 P -lim n (Btk ) = P -lim ϕns s−2α ds n→∞

k=n st +1

n→∞

s

= P -lim n2α n→∞

n k=n st +1

(BtHk )2 .

106


t (It is sufficient to take s = 0.) Therefore, P -limn→∞ s ψun du = 0, where ψun = u−2α ϕnu (θu − 1). Consider any set D ∈ σ{F × B[δ, 1]}, repeat all the previous reasonings and obtain that θu ≡ 1 (otherwise, put D = {(ω, u): θu > 1 + α, or θu < 1 − α}). We proved Theorem 1.19.1 for H ∈ (1/2, 1). Now we consider the case H ∈ (0, 1/2). Similarly to Lemma 1.19.3, we can present the process Xt as Xt =

t

t

sα dMs ,

z(t, s)dWs , where Wt = 0

0

and z(t, s) :=

(6) (CH )−1 mH (t, s)

α t t = (t − s)α − αs−α uα−1 (u − s)α du. s s

Therefore, Xtk − Xtk−1 = −α

tk−2

tk−1

tk−1 tk

0

−α

tk−2 tk

+ tk−1 tk

−α

s tk

s −α

tk

u s −α u

tk−1 −α

(u − s)α−1 du dWs (u − s)α−1 du dWs α

(tk − s) dWs

−α

tk

s tk−1

uα−1 (u − s)α du dWs

s

= J1k + J2k + J3k + J4k . For H ∈ (0, 1/2) it is more convenient to deal with Wt , not Mt . Evidently, ⎛ n n k 2 2 2α 2α ⎝ J1 lim n (∆Xtk ) = lim n n→∞

+

n

n→∞

k=n st +2

k 2

J2k + J3k + J4

k=n st +2

k=n st +2

n

+

k

lim n2α

n→∞

n k=n st +2

from below and from above. As before,

⎞

J1 J2k + J3k + J4 ⎠ .

k=n st +2

First, estimate

k

J1k

2

1.19 Lévy Theorem for fBm n

lim n2α

n→∞

107

2

(∆Xtk ) → t2α (t − s).

k=n st +2

First, we obtain upper bound for the sum tk−2 n t 2 θk (s) dW s , S4n := n2α 0

k=n st +2

where θkt (s) =

tk s −α u

tk−1

(u − s)α−1 du, s ≤ tk−1 . Evidently, for s ≤ tk−2

θkt (s)

t n

α−1

≤

(tk−1 − s)

∧

1 −α

α t . n

(1.19.14)

Therefore, for such n, that n st ∈ N we have that

⎛ 2α

=n

⎝

k−2 ti

n

S4n = n2α s

n

n−2

+

n

⎞ ⎠

i=n st +1 k=i+2

i=1 k=n st +2

s

≤ n2α−1 t

t −u n

s+ 0

s

+ n2α−2 t2

s+ 0

t + 1 − 2α

2 θkt (u) dW u

ti−1

k=n st +2 i=1 nt

ti

2 θkt (u) dW u

ti−1

2α−1

t −u n

dW u 2α−2 dW u

2α−1 n−2 ti t n2α−1 dW u n ti−1 s

+ t2 n2α−2

i=n t +1

n−2 i=n st +1

ti

ti−1

dW u

2α−2 t . n

(1.19.15)

2α−1 s The integral 0 s + nt − u dW u , according to Lemma 2.1 (NVV99), can be estimated as 2α−1 2α−1+β s t t dW u ≤ C(ω) s + − s , s+ −u 0 n n for some random variable 0 < C(ω) < ∞, where β is Hölder index of W u . Evidently, β > 0, and it holds that 2α−1 2α−1+β s 1 t dW u · n2α−1 ∼ n2α−1 → 0. s+ −u n n 0

108


The same is true for

s

0

2α−2 t dW u · n2α−2 . s+ −u n

The last two integrals from (1.19.14) admit the estimate: n−2

i=n st +1

n−2

+

ti

t dW u 1 − 2α ti−1

2α−2 t t2 n2α−2 ≤ t2α C2 (W t − W s ) . n

dW u

ti−1

i=n st +1

2α−1 t n2α−1 n

ti

Now we obtain the lower bound for S4n . Return to M instead of W .

n

S4n = n2α

n

(tk )

t −1 n s

≥t n

n

n−2

= Ct

n

ti

ti−1

i=n st +1

Note that n

i=n st +1

t−t+

2 n

ti

ti

(tk )

tk − u

2α−2

≥ Ct2α+1 n2α−1

dM u

ti−1

2α−1+β · n2α−1 → 0,

n→∞

n

n−2

n−2 i=n st +1

J1k

n → ∞.

2

k=n st +1

i=n st +1

n

2α

(t − u)2α−1 dM u

lim n2α

≥ Ct

dM u

ti−1

Therefore,

2α+1 2α−1

2α−2

1 2α−1 (ti+2 − u) − (t − u)2α−1 dM u . t

n−2

2α−1

∼

n

i=n st +1 k=i+2

i=1 k=n st +2 2α+2 2α−1

(tk − u)

0 n−2

+

2 ϕkt (u) dM u

tk−2

2α

k=n st 2 2α−2

0

k=n st +2

≥ t2 n2α−2

tk−2

ti

2α−1

(ti+2 − u)

dM u

ti−1

(ti+2 − ti−1 )

2α−1

ti

ti−1

dM u .


The “remainder” term for

J1k

n

Rn := n2α

2

equals

tk−2

z

θkt (v)dWv θkt (u)dWu .

0

0

k=n st +2

109

For technical simplicity, it is enough to consider the stopped process W t , nr instead of Wt , and k=3 for any r ∈ N, instead of n st +1 n n k=n s +2 = − k=3 . We obtain that k=3 + t

E(Rn ) = n E 2

4α

nr k−2

=n E

nr−2 nr

=n

nr−2

E

i=1

Let us estimate

u

0

ti

t i−1 n

nr

u

·

θkt (u)dW u

0

2

u

θkt (v)dW v 0

k=i+3

θkt (v)dW v

t i−1 n

i=1 k=i+3 4α

ti

2

u

θkt (v)dW v

ti−1

k=3 i=1

4α

ti

·

θkt (u)dW u 2

θkt (v)dW v

·

dW u .

θkt (u)

0

u t t = θk (u)W u − W v θk (v) v dv 0 u t t ≤ L θk (u) + L θk (v) v dv . 0

It follows from (1.19.14), that u α t t t t θk (v) v dv = θk (u) − θk (0) ≤ C n

for some

C > 0.

0

Moreover,

2α

n

2

nr

2

tr

≤n

(v − u)

2α

θkt (u)

α−1

dv

ti+1

k=i+3

α 2

= Cn2α [−(tr − u)α + (ti+1 − u) ] ≤ C, and the integrand

4α

n

nr k=i+2

2

u

θkt (v)dW v

·

θkt (u)

≤ C,

0

i.e. there exists the integrable dominant. Therefore, it is sufficient to establish that for any u

110


n2α

nr

u

P

θkt (v)dW v · θkt (u) → 0.

0

k=i+3

We take the mathematical expectation and obtain that 2 u nr θkt (v)θkt (u) dW v . n4α E 0

k=i+3

The bounded dominant exists. Indeed, 2 2

nr

nr 4α t t 2α t θk (v)θk (u) ≤n θk (v) ≤ C, n k=i+3

k=i+2

as before. Further, we must prove that n2α

nr

θkt (v)θkt (u) → 0

k=i+3

for all fixed 0 < v < u. We have that n2α

nr k=i+2

×

tk

nr

θkt (v)θkt (u) ≤ n2α

tk−1

nr

(s − u)α−1 ds

tk−1

k=i+3

(s − v)α−1 ds ≤ n2α

tk

α−1

(tk−1 − u)

k=i+3

α−1

1 n

tk

(s − v)α−1 ds

tk−1

tr

≤ n2α−1 (ti+2 − u)

(s − v)α−1 ds ti+2

≤ Cnα−1 (u − v)α−1 → 0,

n→∞

for any

0 < v < u. P

From all these estimates, the remainder term Rn → 0, n → ∞, and we have established that C1 t4α (M t − M s ) ≤ lim n2α n→∞

n

J1k

2

≤ C2 t2α (W t − W s ) .

k=n st +2

(Note, that for H ∈ (1/2, 1), we obtained opposite estimates.) Note also k 2 that we cannot estimate J , i > 1, from above. Indeed, the inteα i α l that admits the estimate < n1 → 0 for grand of the form t n − u H ∈ (1/2, 1), now, for H ∈ (0, 1/2), tends to ∞. So, we mention that k n k k k k 2 ≥ 0, prove that J1 J2 + J3k + J4k → 0, and k=n st +2 J2 + J3 + J4 obtain the estimate from above: C1 t2α (M t − M s ) ≤ (t − s).


111

In the sequel, we realize this plan. It is sufficient to estimate nthe sums from k = 2 till k = n. By applying the Lenglart inequality to n2α k=2 J1k J2k , we obtain that it is sufficient to prove that

2 n tk−2 tk −α s (u − s)α−1 du dW s n4α u 0 t k−1 k=2 ⎞ ⎛ 2 tk−1 tk −α s ×⎝ (u − s)α−1 du dW s ⎠ tk−2 tk−1 u ≤ Cn4α

n

tk−2

2 θkt (s)dW s

0

k=2

tk−1

2α

(tk−1 − s)

P

dW s → 0.

tk−2

Integrate the last integral by parts: tk−1 2α 2α W tk−1 − W tk−2 (tk−1 − s) dW s = (tk−1 − tk−2 ) tk−2

− 2α

tk−1

2α−1

(tk−1 − s)

tk−2 −2α

≤ Cn

∆W tk−1 + C

Now recall that

tk−1

=

2α−1

tk−2

W tk−1 − W s ds

(tk−1 − s)

W tk−1 − W s ds.

2

tk−2

θkt (s)dW s

0 tk−2

2 θkt (s) dW s + 2

0

tk−2

s

θkt (v)dW v θkt (s)dW s . 0

0

It was proved that σ1n := n2α

n k=2

tk−2

2 θkt (s) dW s

0

is bounded in probability, and σ2n

2α

:= n

n k=2

tk−2

s

P

θkt (v)dW v θkt (s)dW s → 0,

n → ∞.

0

0

Therefore, 4α

n

n k=2

0

tk−2

2 θkt (s)dW s

· Cn−2α ∆W tk−1

112

1 Wiener Integration with Respect to Fractional Brownian Motion P

≤ Cσ1n · max ∆W tk−1 + Cσ2n · max ∆W tk−1 → 0, k

n → ∞.

k

Also, n

4α

n

tk−2

θkt (s)dW s

2 ·

0

k=2

tk−1

× W tk−1 − W s ds ≤ C(ω) (σ1n + σ2n ) n2α 2H−ε

≤ C(ω) (σ1n + σ2n ) n2α (tk−1 − tk−2 ) n

Consider n2α n2α

k=2

n

tk−1

(tk−1 − s)

ds

tk−2

1−ε 1 → 0, n

∼

2α−ε

n → ∞.

J1k J3k :

tk−2

tk

θkt (s)dWs ·

0

k=1

2α−1

(tk−1 − s)

tk−2

tk−1

−α

s tk

α

(tk − s) dWs .

As before, it is sufficient to prove that n4α

n

tk−2

θkt (s)dW s

2 ·

0

k=1

tk

tk−1

s tk

−2α

2α

(tk − s)

P

dW s → 0, n → ∞,

or, equivalently,

tk

n2α max k

2α

(tk − s)

P

dW s · (σ1n + σ2n ) → 0.

(1.19.16)

tk−1

Note that by (NVV99, Lemma 2.1) and due to H¨ older properties of W ,

tk

2α

(tk − s)

dW s ≤ C(ω) (tk − tk−1 )

2α+1−ε

tk−1

∼

2α+1−ε 1 , n

whence we obtain (1.19.16). Now, consider n2α J1k J4k ; other sums can be estimated similarly. After some transformations, n4α

n k=1

tk−2

2 ·

θkt (v)dW u

0

tk

tk−1

≤ n2α max k

tk

tk−1

s

tk

s−2α

tk

2 uα−1 (u − s)α du

dW s

s

2 uα−1 (u − s)α du

dW s · (σ1n + σ2n )


≤ n2α max k

tk

tk−1

tk

s

(u − s)2α du dW s · (σ1n + σ2n )

s

tk

≤ Cn2α max k

≤ Cn ·

tk

u2α−2 du ·

113

2α+1

s2α−1 (tk − s)

dW s · (σ1n + σ2n )

tk−1

1 max n k

tk

(tk − s)

2α

dW s · (σ1n + σ2n )

tk−1 2α+1−ε

≤ C max (tk − tk−1 ) k

· (σ1n + σ2n ) → 0,

n → ∞.

Due to all these estimates we have proved that n

t2α (t − s) = lim n2α n→∞

2

(∆Xtk ) ≥ C1 t4α (M t − M s ) ,

k=n st +2

i.e. M t − M s ≤ C2 t−2α (t − s) = C2 t1−2α − st−2α ≤ C2 t1−2α − s1−2α , or

t

u−2α dW u ≤ C2

s

t

u−2α du.

s

As before, it follows that W t is absolutely continuous w.r.t. Lebesgue measure, t θs ds, (1.19.17) W t = 0

0 ≤ θs ≤ C, C is some constant, θs possibly is random. Taking this into account, we can continue estimates from above: for example, if we take for simplicity the sums over k = 2 till k = n, then n2α

n

k 2

J2

=σ 1n + σ 2n := Cn2α

k=1

n k=1

× dW s + Cn

2α

n k=1

where

θkt (s)

tk

= tk−1

tk−1

u

tk−2

tk−2

s −α

tk−1

tk

tk−1

s −α u

θkt (v)dWv

θkt (u)dWu ,

tk−2

α

(u − s)α−1 du ≤ (tk−1 − s) C.

n k=1

(u − s)α−1 du

u

Therefore, σ 1n ≤ Cn2α

2

tk−1

tk−2

2α

(tk−1 − s)

dW s .

114


Direct estimates give nothing (because of singularity at tk−1 ). So, we go by an indirect way: for some A > 0,

tk−1

2α

(tk−1 − s)

dW s ≤

tk−2

t tk−1 − nA

tk−1

+ t tk−1 − nA

tk−2

2α t tk−1 − tk−1 − · ∆W tk nA tk−1 2α + (thanks to (1.19.17)) C (tk−1 − s) ds ≤

≤

t nA

t tk−1 − nA

2α

∆W tk + C

t nA

2α+1 .

Taking the sum, we obtain: 2α 2α+1 n t t n 2α 2α σ 1 ≤ Cn ∆W tk + Cn n nA nA k=1

1 t2α+1 . A2α+1 If we estimate the sum from k = n st + 1 to k = n, then ≤ CA−2α t2α W t + C

σ 1n ≤ CA−2α t2α (W t − W s ) + C = CA−2α t2α (W t − W s ) + C

1 A2α+1 1 A2α+1

s t2α+1 1 − t t2α (t − s).

Now we want to prove that

n tk−1 u P 2α t n θk (v)dWv θkt (u)dWu → 0, k=1

n → ∞.

tk−2

tk−2

As usual, it is enough to establish that

2 n tk−1 u t 2 P 4α t θk (u) dW u → 0. θk (v)dW v n k=1

tk−2

tk−2

But we can bound W u by Cdu, so, it is enough to prove that 4α

n

n k=1

tk−1

tk−2

2

u

θkt (v)dW v

2 P θkt (u) du → 0.

tk−2

By taking the mathematical expectation, we see that it is sufficient to establish that

1.19 Lévy Theorem for fBm n

n4α

tk−1

2 2 P θkt (v) dW v θkt (u) du → 0.

tk−2

tk−2

k=1

u

115

By substituting Cdv instead of dW v , we see that it is enough to establish that

n tk−1 u 2 2 θkt (v) dv θkt (u) du → 0. σ3n := n4α k=1

tk−2

tk−2

2

We have that (θkt (u)) ≤ Cn−2α , and

n tk−1 u 1 n σ3 ≤ dv du ≤ C → 0, n tk−2 tk−2

n → ∞.

k=1

Finally, n

n2α

J2k

2

≤ CA−2α t2α (W t − W s ) + C

k=n st +2

1 A2α+1

t2α (t − s).

Now, proceed with J3k : 2α

n

n

2 J3k

2α

=n

k=1

2α

+n

n

tk

u

tk−1

tk

tk−1

k=1

tk−1

k=1

n

(tk − s)

−α

s tk

2

−α

s tk α

(tk − s) dWs

u tk

α

dW s

−α

α

(tk − u) dWu .

The first term can be estimated as 2α n tk t C 2α n2α (tk − s) dW s ≤ C (W t − W s ) + 2α+1 t2α (t − s), A A tk−1 k=1

as before. And with the bound dW s ≤ Cds, the second term can be estimated as n tk u 4α 2α 2α n4α k=1 tk−1 (t − s) ds · (tk − u) du ≤ nCn 4α+2 → 0. Therefore, for tk−1 k k 2 k 2 J3 we have the same estimate as for J2 . Finally, estimate n2α

n

k 2

J4

= Cn2α

k=1

n

2α

= Cn

k=1

+ Cn2α

k=1

−2α

tk

tk−1

tk

2 uα−1 (u − s)α du dWs

s

2

tk

s

tk−1

n

s−α

tk−1

k=1 n tk

tk

u

α−1

(u − s) du α

dW s

s

u

tk−1

s−α

s

tk

v α−1 (v − s)α dv dWs

116


×u

−α

tk

v α−1 (v − u)α dvdWu .

u

The first term can be estimated with the help of (1.19.17) as n2α t−2α

n

tk

tk−1

k=2

tk

2

n → ∞.

s 1 p

If k = 1, then for

1 q

= 1, p, q > 1

t/n

+

2α −2α

u 0

t/n

2

t/n

n t

≤ n2α t−2α

dW s ≤ Cn−2H → 0

uα−1 (u − s)α du

α−1

(u − s) du α

2/p

t/n

(u − s)αq du

s

2α −2α

2/q

t/n

up(α−1) du 0

ds

s

ds

s

t/n

≤n t

3p 2 s(pH− 2 +1) p

0

(Hq− q2 +1) q2 t −s ds n 4α+1 t 2α −2α ds ∼ n t → 0, n

2α+ q2 t −s s =n t n 0 n 2 i.e. the “main term” of n2α k=1 J4k tends to 0. For the remainder term n 2 of n2α k=1 J4k it is sufficient to prove that for any ε > 0 2α −2α

σ4n

t/n

4α

:= n

n

−2α

tk

u

−α

2

tk

v

s

tk−1

k= nε t

×u

2 2α−2+ p

α−1

(v − s) dv α

ds

s

tk−1

2

tk

v

α−1

(v − u) dv

du → 0

α

n → ∞.

u

But σ4n ×

≤n

4α

n k= nε t

v

α−1

(v − u) dv α

u

2

tk

v

tk−1

2

tk

tk

tk−1

du ≤ n−6

u

α−1

(v − s) dv α

ds

s n

(tk−1 )

−4

∼ n−2 → 0,

n → ∞.

k= nε t

After all estimates, for s > 0 lim n2α

n→∞

n

(∆Xtk ) ≤ C2 A−2α t2α (W t − W s ) + C2 2

k=n st +2

We have the opposite estimate,

1 t2α (t − s). A2α+1

1.20 Multi-parameter Fractional Brownian Motion

C1 t2α (t − s) ≤ lim n2α n→∞

n

(∆Xtk )

117

2

k=n st +2

≤ C2 A−2α t2α (W t − W s ) + C2

1 A2α+1

t2α (t − s).

1 So, for A sufficiently large, C3 := C1 − C2 A2α+1 > 0, and we obtain that

C3 t2α (t − s) ≤ C2 A−2α t2α (W t − W s ) , whence W t − W s ≥

C3 2α C2 A (t − s), and t write W t = 0 θs ds,

constants do not depend on s and

t. Therefore, if we then ε1 ≤ θs ≤ ε2 , εi > 0, and t 1/2 Wt = 0 θs dVs with some Wiener process V . Then we can conclude the proof of the theorem by the same arguments as for H ∈ (1/2, 1).

1.20 Multi-parameter Fractional Brownian Motion 1.20.1 The Main Definition There can be at least two approaches to the definition of multi-parameter fBm. We consider the process which has a “fractional Brownian” property in each coordinate, but also it is possible to consider this property, for example, along any ray with its origin at zero (MY67). For technical simplicity we consider two-parameter fBm (fBm-field) {BtH , t ∈ R2+ }, where t = (t1 , t2 ). We suppose that s ≤ t if s = (s1 , s2 ), t = (t1 , t2 ) and si ≤ ti , i = 1, 2. Definition 1.20.1. The two-parameter process {BtH , t ∈ R2+ } is called a (normalized) two-parameter fBm with Hurst index H = (H1 , H2 ) ∈ (0, 1)2 , if it satisfies the assumptions (a) B H is a Gaussian field, Bt = 0 for t ∈ ∂R2+ ; 2Hi i (ti + s2H − |ti − si |2Hi ). (b) E BtH = 0, E BtH BsH = 14 i i=1,2

Evidently, such a process has the modification with continuous trajectories, and we will always consider such a modification. Moreover, consider “twoparameter” increments: ∆s BtH := BtH − BsH1 t2 − BtH1 s2 + BsH for s ≤ t. Then H will be they are stationary. Note, that for any fixed ti > 0 the process B(t i ,·) the fBm with Hurst index Hj , i = 1, 2, j = 3 − i, evidently, nonnormalized. 1.20.2 H¨ older Properties of Two-parameter fBm Denote PT := [0, T1 ] × [0, T2 ].

118


Definition 1.20.2. The function f : R2+ → R belongs to the class C λ1 ,λ2 (PT ) older of orders λ1 and λ2 on PT ), if there exists a constant for 0 < λi ≤ 1 (f is H¨ C > 0, such that for all s ≤ t, s, t ∈ PT |∆s ft | ≤ C (ti − si )λi , (1.20.1) i=1,2

|f (t) − f (s1 , t2 )| ≤ C|t1 − s1 | , |f (t) − f (t1 , s2 )| ≤ C|t2 − s2 |λ2 . λ1

(1.20.2)

The norm in the space C λ1 ,λ2 (PT ) is denoted as

|f (t) − f (s1 , t2 )| (t1 − s1 )λ1 0≤s 0 there exists the random variable 0 < c(ω) < ∞ P -a.s. such that 1/2 1 |∆s BtH | ≤ c(ω) (ti − si )Hi 1 + log ti −s . i i=1,2

1.20.3 Fractional Integrals and Fractional Derivatives of Two-parameter Functions 1 Γ (α1 )Γ (α2 ) .

For α = (α1 , α2 ) denote Γ (α) =

Definition 1.20.3. (SKM93) Let f ∈ P := [a, b] :=

[ai , bi ], a = (a1 , a2 ),

i=1,2

b = (b1 , b2 ). Forward and backward Riemann–Liouville fractional integrals of orders 0 < αi < 1 are defined as f (u) α1 α2 du, (Ia+ f )(x) := Γ (α) ϕ(x, u, 1 − α) [a,x] and

f (u) du, [x,b] ϕ(x, u, 1 − α) correspondingly, where [a, x] = [ai , xi ], [x, b] = [xi , bi ], du = du1 du2 , α1 α2 (Ib− f )(x) := Γ (α)

α1

ϕ(u, x, α) = |u1 − x1 |

i=1,2 α2

|u2 − x2 |

, u, x ∈ [a, b].

i=1,2

1.20 Multi-parameter Fractional Brownian Motion

119

Definition 1.20.4. Forward and backward fractional Liouville derivatives of orders 0 < αi < 1 are defined as f (u) ∂2 α1 α2 (Da+ f )(x) := Γ (1 − α) du, ∂x1 ∂x2 [a,x] ϕ(x, u, α) and α1 α2 f )(x) := Γ (1 − α) (Db−

∂2 ∂x1 ∂x2

[x,b]

f (u) du, ϕ(x, u, α)

x ∈ [a, b].

Definition 1.20.5. Forward fractional Marchaud derivatives of orders 0 < αi < 1 are defined as

∆u f (x)du f (x) α α 1 2 + α (D f )(x) := Γ (1 − α) α 1 2 a+ ϕ(x, u, α) [a,x] ϕ(x, u, 1 + α) ⎞ xi αi f (x) − f (ui , xj ) αj dui ⎠ , + (x − a ) (xi − ui )1+αi j j ai i=1,2,j=3−i

and the backward derivatives can be defined in a similar way. α1 α2 α1 α2 Let 1 ≤ p ≤ ∞, the classes I+ (Lp (P)) := {f | f = Ia+ ϕ, ϕ ∈ α1 α2 α1 α2 Lp (P)}, I− (Lp (P)) := {f | f = Ib− ϕ, ϕ ∈ Lp (P)}. Similarly to Theorem 13.1 (SKM93), the following result can be proved.

Theorem 1.20.6. Liouville and Marchaud derivatives coincide on the classes α1 α2 I± (Lp (P)). −(α α )

α1 α2 =: Ia+ 1 2 . Of course, we can introduce the Further we denote Da+ notions of fractional integrals and fractional derivatives on R2+ . For example, the Riemann–Liouville fractional integrals and derivatives on R2+ are defined f (t) α1 α2 f )(x) := Γ (α) ( −∞,x] ϕ(x,u,α) dt, by the formulas (I+ f (t) α1 α2 (I− f )(x) := Γ (α) [x,∞ ) ϕ(x,u,α) dt, 2 −(α α ) f (t) α1 α2 f )(x) := Γ (1 − α) ∂x∂1 ∂x2 (−∞,x] ϕ(x,t,α) dt, and (I+ 1 2 f )(x) = (D+ −(α1 α2 ) f (t) α1 α2 ∂2 f )(x) = (D− f )(x) := Γ (1 − α) ∂x1 ∂x2 [x,∞ ) ϕ(x,t,α) dt, (I− 0 < αi < 1. Evidently, all these operators can be expanded into the product α1 α2 α1 α2 = I+ ⊗ I+ , and so on. In what follows we shall consider of the form I+ only the case Hi ∈ (1/2, 1). Define the operator (3) H1 H2 α1 α2 f := CHi I± f. M± i=1,2

Definition 1.20.7. A random field {Xt , t ∈ R2+ } is a field with independent increments if its increments {∆si Xti , i = 1, n} for any family of disjoint rectangles { ( si , ti ] , i = 1, n} are independent.

120


Definition 1.20.8. The random field {Wt , t ∈ R2+ } is called the Wiener field if W = 0 on ∂R2+ , W is the field with the independent increments and E(∆s Wt )2 = area( ( s, t ]) =

(ti − si ).

i=1,2

Let we have a probability space (Ω, F, P ) with two-parameter filtration {Ft , t ∈ R2+ } on it. It means that Fs ⊂ Ft ⊂ F for s ≤ t. Denote Fs∗ := σ{Fu , s ≮ u}. Definition 1.20.9. An adapted random field {Xt , Ft , t ∈ R2+ } is a strong martingale if X vanishes on ∂R2+ , E|Xt | < ∞ for all t ∈ R2+ and for any s ≤ t E(∆s Xt | Fs∗ ) = 0. Evidently, any random field with constant expectation and independent increments is a strong martingale, in particular, the Wiener field is a strong martingale. It is not difficult to prove the following fact. Lemma 1.20.10. Let {Wt , t ∈ R2+ } be a Wiener field. Then the field H1 H2 BtH1 H2 := (M− 1(0,t) )(x)dWx (1.20.3) R2

is two-parameter fBm (not necessarily normalized). Similarly to the one-parameter case, it is easy to show that any twoparameter fBm can be represented by (1.20.3) via underlying random field W. Introduce the notion of Wiener integral w.r.t. two-parameter fBm. Definition 1.20.11. Let

H1 H2 2 2 1 H2 f ∈ LH := f : R → R : ((M f )(t)) dt < ∞ . − 2 R2

Then we denote R2 f (t)dBtH1 H2 as Wiener process W .

R2

H1 H2 (M− f )(t)dWt for the underlying

The following facts are proved similarly to the one-parameter case. (2) Theorem 1.20.12. Let the kernel lH (t, s) = lHi (ti , si ) · 1{0<s 0 < α < 1. Then for any a < c < d < b b 1−α α (Da+ 1[c,d) )(x)(Db− gb− )(x)dx = g(d) − g(c).

1 1−α

and

(2.1.2)

a

Proof. We have that

α (Da+ 1[c,d) )(x)

=

⎧ ⎪ ⎨ 0, (x−c)−α

x ≤ c, c < x ≤ d,

Γ (1−α) ,

⎪ ⎩ (x−c)−α −(x−d)−α Γ (1−α)

, d ≤ x ≤ b.

Therefore, by using (2.1.1), we obtain for αp < 1, or q >

1 1−α ,

that

b

b 1−α 1−α 1 α (Da+ 1[c,d) )(x)(Db− gb− )(x)dx = Γ (1−α) (x − c)−α (Db− gb− )(x)dx c b 1−α 1−α 1−α 1 −α − Γ (1−α) d (x − d) (Db− gb− )(x)dx = Ib− (Db− gb− )(c) 1−α 1−α − Ib− (Db− gb− )(d) = g(d) − g(c). a

Corollary 2.1.4. Let the function g ∈ C λ [a, b] for some λ ≤ 1, then 1−α (Lp [a, b]) for any p ≥ 1 and 1 − α < λ. So, we can put p > 2/λ, gb− ∈ Ib− α = 1 − λ/2 and obtain for g (2.1.2). Corollary 2.1.5. For any step function fπ (x) =

n−1 k=0

ck 1[ xk ,xk+1 ) (x) with

a = x0 < · · · < xn = b and g satisfying the conditions of Lemma 2.1.3, we n−1 b have that a f (x)dg(x) = ck (g(xk+1 ) − g(xk )). k=0

2.1 Pathwise Stochastic Integration

125

Further we suppose that g(b−) = g(b) and g(a+) = g(a). Denote by BV [a, b] the class of functions of bounded variation on [a, b]. 1−α α Lemma 2.1.6. Let the functions fa+ ∈ Ia+ (Lp [a, b]), gb− ∈ Ib− (Lq [a, b]) ∩ BV [a, b] with p ≥ 1, q ≥ 1, 1/p + 1/q ≤ 1 and b α α Ia+ (|(Da+ f )|)(x)|g|(dx) < ∞. (2.1.3) a

Then

b

b

f (x)dg(x) = (L-S) a

f (x)dg(x). a

Proof. We have that b b α α (L-S) a f (x)dg(x) = (L-S) a Ia+ (Da+ f )(x)dg(x) b x 1 α−1 α = Γ (1−α) (L-S) a ( a (x − y) (Da+ f )(y)dy)dg(x).

(2.1.4)

Condition (2.1.3) together with Fubini theorem permits us to change the order of integration: b x α (L-S) a ( a (x − y)α−1 (Da+ f )(y)dy)dg(x) b b α α−1 dg(x))dy = a (Da+ f )(y)( y (x − y) (2.1.5) b α b ∞ α−2 = (α − 1) a (Da+ f )(y)( y ( x (z − y) dz)dg(x))dy. Further, if y ∈ (a, b) is the point of continuity of function g, then b ∞ b z ( x (z − y)α−2 dz)dg(x) = y ( y dg(x))(z − y)α−2 dz y ∞ b b + b ( y dg(x))(z − y)α−2 dz = y g(z)−g(y) (z−y)2−α dz +

g(b)−g(y) (α−1)(b−y)α−1

=

(2.1.6)

Γ (α) 1−α α−1 (Db− gb− )(y).

Since set of discontinuity points of g is at most countable , and taking (2.1.4)– (2.1.6) together, we obtain the proof.

Now we consider the case of Hölder functions f and g. The existence of b (R-S) a f dg for f ∈ C λ [a, b], g ∈ C µ [a, b] with λ + µ > 1 was established by b Kondurar (Kon37). Moreover, this integral coincides with a f dg , as the next theorem states. Let f ∈ C λ [a, b] for some 0 < λ ≤ 1 and |f (x) − f (y)| ≤ c(λ)|x − y|λ , x, y ∈ [a, b]. Consider the following step function: fπ (x) =

n−1

f (xk )1[ xk ,xk+1 ) (x),

k=0

where the partition π = {a = x0 < x1 < · · · < xn = b}. Evidently, lim|π|→0 supπ fπ − f L∞ [a,b] = 0.

126

2 Stochastic Integration with Respect to fBm and Related Topics

Theorem 2.1.7. 1) For any 0 < α < λ ! ! α α lim sup !(Da+ fπ ) − (Da+ f )!L

1 [a,b]

|π|→0 π

= 0.

2) Let f ∈ C λ ([a, b]), g ∈ C µ [a, b] with λ + µ > 1, then (R-S) and b b f dg = (R-S) f dg. a

b a

f dg exists

a

b

(x)−f (x)| Proof. 1) It is sufficient to prove that a |fπ(x−a) dx → 0 and α b x −α−1 (x − y) |fπ (x) − f (x) − fπ (y) + f (y)|dy dx → 0 as |π| → 0. But a a |fπ (x) − f (x)| ≤ |f (xk ) − f (x)| ≤ c(λ)|π|λ for x ∈ [ xk , xk+1 ), therefore b |fπ (x)−f (x)| 1−α dx ≤ c(λ)|π|λ (b−a) → 0 as |π| → 0. Also, for x ∈ [ xk , xk+1 ) (x−a)α 1−α a

x A(x) := a (x − y)−α−1 |fπ (x) − f (x) − fπ (y) + f (y)|dy k−1 xi+1 (x − y)−α−1 |f (xk ) − f (x) − f (xi ) + f (y)|dy = xi i=0

k−1 xi+1 (x − y)−α−1 |f (y) − f (x)|dy ≤ 2c(λ) (x − y)−α−1 dy · |π|λ xi i=0 x −α λ−α k) k) + c(λ) xk (x − y)λ−α−1 dy ≤ 2c(λ)|π|λ (x−x + c(λ) (x−x 1−α λ−α

+

x

xk

≤ 3c(λ) |π| λ−α , λ−α

b which means that a A(x)dx → 0 as |π| → 0. α f (x) and 2) We take 1 − µ < α < λ, then the fractional derivatives Da+ 1−α (Db− g)b− (x) exist, and, moreover, b |g(y)−g(x)| |g(b)−g(x)| 1−α 1 |(Db− g)b− (x)| ≤ Γ (1−α) 1−α + (1 − α) x (y−x)2−α dy (b−x) 1 1−α ≤C · c(λ)(b − x)µ+α−1 1 + µ+α−1 ≤ Γ (1−α) for some constant C. Therefore, according to part 1) of the proof, b b b 1−α α α | a fπ dg − a f dg| ≤ a |(Da+ fπ )(x) − (Da+ f )(x)||(Db− g)b− (x)|dx b α α ≤ C a |(Da+ fπ )(x) − (Da+ f )(x)|dx → 0, (2.1.7) as |π| → 0. Furthermore, according to Corollary 2.1.5, b a

fπ dg =

n−1

f (xk )(g(xk+1 ) − g(xk )) → (R-S)

b

k=0

a

f dg,

and from (2.1.7)–(2.1.8) we obtain the desired equality. Now we establish the properties of generalized integral tion of upper and lower boundaries.

(2.1.8)

t s

f dg as the func-

2.1 Pathwise Stochastic Integration

127

Lemma 2.1.8 ((Zah98)). 1) Let a ≤ s < t ≤ b and the functions f and g satisfy the assumptions 1−α α (Lp [a, b]), gb− ∈ I− (Lq [a, b]) for some 0 < α < 1, (i) (f · 1(s,t) ) ∈ I+ p ≥ 1, q ≥ 1, 1/p + 1/q ≤ 1, 1−α α (Lp [s, t]), gt− ∈ I− (Lq [s, t]) for some 0 < α < 1, (ii) fs+ ∈ I+ p ≥ 1, q ≥ 1, 1/p + 1/q ≤ 1. Then

b

1(s,t) f dg = a

2) The equality

f dg. s

t

t

f dg + s

u

u

f dg = t

f dg s

holds for a ≤ s < t < u ≤ b, if all the integrals exist as generalized Lebesgue– Stieltjes integrals. Proof. 1) Let {ϕn (x), x ∈ R} be a sequence of smooth kernels, i.e. 0 ϕn ∈ C ∞ (R), ϕn ≥ 0, ϕn = 0 outside [−1/n, 0] and −1/n ϕn (x)dx = 1. More exactly, let ϕn (x) = nϕ(nx) for ϕ ∈ C ∞ (R), ϕ = 0 outside of [−1, 0]. Then we can approximate the function gb− by smooth functions gn := gb− ∗ ϕn , and the following properties hold: gn (b−) = n [x−b,x−a]∩[−1/n,0] (g(b−) − g(x − t))ϕ(nt)dt |x=b− = 0; 1−α 1−α gn )(x) = Db− ( R gb− (x − t)ϕn (t)dt) (Db− = 1(a,b) (x)(Γ (1 − α))−1 R gb− (x − t)ϕn (t)dt(b − x)α−1 b (2.1.9) + α x (y − x)2−α ( R (gb− (x − t) − gb− (y − t))ϕn (t)dt)dy 1 (x) b gb− (x−t)−gb− (y−t) gb− (x−t) dy dt = Γ(a,b) (1−α) R ϕn (t) (b−x)1−α + α x (y−x)2−α 1−α = 1(a,b) (x)((Db− gb− ) ∗ ϕn )(x);

! 1−α !q 1−α !(D ! b− gn ) − (Db− gb− ) Lq [a,b] !q ! 1−α 1−α ! !(D b− gb− ) ∗ ϕn − (Db− gb− ) Lq [a,b] b 0 1−α 1−α t = a | −1 ((Db− gb− )(x − n ) − (Db− gb− )(x))ϕ(t)dt|q dx b 0 1−α 1−α ≤ C a −1 |(Db− gb− )(· − nt ) − (Db− gb− )(·)|q dt dx → 0,

n → ∞. (2.1.10) Therefore, from this Lq -convergence, from Lemma 2.1.2 and the properties of convolutions, b α b 1−α 1 f dg = a (Da+ 1(s,t) f )(u)(Db− gb− )(u)du a (s,t) b α 1−α = limn→∞ a (Da+ 1(s,t) f )(u)(Db− gn )(u)du b t = limn→∞ a (1(s,t) f )(u)gn (u)du = limn→∞ s f (u)(gb− ∗ ϕn )(u)du. Further, for any c > 0 (c ∗ ϕn )(u) = 0, therefore

128


t

ϕn )(u)du

f (u)(gb− ∗ = s t f (u)(gt− ∗ ϕn )(u)du, =

t

f (u)(g ∗ ϕn )(u)du

s

(2.1.11)

s

and t t limn→∞ s f (u)(gb− ∗ ϕn )(u)du = limn→∞ s f (u)(gt− ∗ ϕn )(u)du t (2.1.12) = limn→∞ s f (u)(gt− ∗ ϕn ) (u)du. Thanks to Lemma 2.1.2, assumption (ii), (2.1.9) and (2.1.10), applied to t instead of b, t limn→∞ s f (u)(gt− ∗ ϕn ) (u)du t α 1−α = limn→∞ s (Ds+ f )(u)(Dt− (gt− ∗ ϕn ))(u)du t α s+ (2.1.13) 1−α = limn→∞ s (Ds+ fs+ )(u)((Dt− gt− ) ∗ ϕn )(u)du t α t 1−α = s (Ds+ fs+ )(u)(Dt− gt− )(u)du = s f dg, and we obtain the first statement. The second one we obtain by using some of the equalities from (2.1.11): t u f dg + t f dg = limn→∞ s f (r)(g ∗ ϕn )(r)dr u u + lim u n→∞ t f (r)(g ∗ ϕn )(r)dr = limn→∞ s f (r)(g ∗ ϕn )(r)dr = s f dg.

t s

2.1.2 Pathwise Stochastic Integration in Fractional Besov-type Spaces In this subsection we consider the approach to pathwise stochastic integration in fractional Besov-type spaces, introduced by Nualart and Rˇ a¸scanu (NR00) (see also (CKR93) and (NO03a)). Consider the following functional spaces. Let for 0 < β < 1 t ϕβf (t) := |f (t)|+ 0 |f (t)−f (s)|(t−s)−β−1 ds, and W0β = W0β [0T ] be the space of real-valued measurable functions f : [0, T ] → R such that f 0,β := sup ϕβf (t) < ∞. t∈[0,T ]

Furthermore, let W1β = W1β [0, T ] be the space of real-valued measurable functions f : [0, T ] → R such that t |f (t) − f (s)| |f (u) − f (s)| f 1,β := sup + du < ∞ (t − s)β (u − s)1+β 0≤s 0 and any 0 < β < H belong to W1β [0, T ]. Let f ∈ W1β [0, T ]. Then its restriction to [0, t] ⊂ [0, T ] belongs to β I− (L∞ [0, t]) and Λβ (f ) :=

sup 0≤s 1. Then Lemma 2.2.1. 1. Let f ∈ I± β1 β2 limβ1 →0,β2 →0 Da+(b−) f (x) = f (x), where the limit is in Lp (P). 2. Let, in addition, the function f be twice continuously differentiable in the neighbor2 β1 β2 f hood of the point x. Then limβ1 →1,β2 →1 Da+(b−) f (x) = ∂x∂1 ∂x (x). So, we can 2 00 11 put Da+(b−) f := f , Da+(b−) f := f .

Theorem 2.2.2. Let 0 < βi < 1 and 1 < p < β1−1 ∨ β2−1 . Then the operator β1 β2 is bounded from Lp (P) into Lq (P), where Ia+ 1 < q < p((1 − β1 p)−1 ∧ (1 − β2 p)−1 ). Proof. Denote r := p((1 − β1 p)−1 ∨ (1 − β2 p)−1 ). Since r > p, it is sufficient to consider q ∈ (p, r). Then for p1 + p1 = 1, p1 + 1r = 1 − βi , from the generalized i H¨ older inequality, it holds that q1 β1 β2 |(Ia+ f )(x)| ≤ C |f (u)|p (xi − ui )(βi −1)γq du [a,x]

×

|f (u)|p du

p1 − q1

[a,x] 1− p

q ≤ C f Lp (P)

i=1,2

(xi − ui )(βi −1)(1−γ)p dui

p1

[a,x] i=1,2

|f (u)|p

[a,x]

(xi − ui )(βi −1)γq du

q1

.

i=1,2

Here we choose γ satisfying the inequalities (1 − βi )γq < 1 and (1 − βi )(1 − γ)p < 1, which is equivalent to 1 − (p (1 − βi ))−1 < γ < (q(1 − βi ))−1 . Such a choice is possible, since the inequality 1 − (p (1 − βi ))−1 < (q(1 − βi ))−1 is equivalent to q < p(1 − βi p)−1 , and this is evident under our suppositions. By integration over P we obtain that ! ! ! β1 β2 ! !Ia+ f !

Lq (P)

≤C

1− p q f Lp (P)

|f (u)| du · p

P

P i=1,2

(βi −1)γq

(xi − ui )

q1 dx

≤ C f Lp (P) .

132


β1 β2 Corollary 2.2.3. Let f ∈ Lp (P), g ∈ Lq (P), Ia+ g ∈ Lr (P) for 1/p + 1/r = 1 and r < q((1 − β1 q)−1 ∧ (1 − β2 q)−1 ), i.e. 1/p + 1/q < 1 + β1 ∧ β2 . Then β1 β2 β1 β2 f (u)Ia+ g(u)du = g(u)Ib− f (u)du. P

P

Evidently, ρ1 ρ2 β1 β2 ρ1 +β1 ρ2 +β2 I± = I± I±

for f ∈

ρ1 +β1 ρ2 +β2 (L1 (P)), I±

on

L1 (P);

ρi , βi ≥ 0, ρi + βi ≤ 1

ρ1 ρ2 β1 β2 Da+(b−) Da+(b−) f

ρ1 +β1 ρ2 +β2 = Da+(b−) f;

ρ1 ρ2 ρ1 ρ2 for f ∈ Ia+(b−) (Lp (P)), g ∈ Ib− (Lq (P)), p, q > 1, 1/p + 1/q < 1 + ρ1 ∧ ρ2 ρ1 ρ2 ρ1 ρ2 Da+ f (u)g(u)du = f (u)Db− g(u)du. P

P

2.2.2 Generalized Two-parameter Lebesgue–Stieltjes Integrals We suppose that all the functions, considered on some rectangle P = [a, b], belong to the space D(P), i.e. they have the limits in all the quadrants, Q++ (x) = {s ∈ P|s ≥ x}, Q+− (x) = {s ∈ P|s1 ≥ x1 , s2 < x2 }, −+ Q (x) = {s ∈ P|s1 < x1 , s2 ≥ x2 }, Q−− (x) = {s ∈ P|s < x}, f (x) = lims→x,s≥x f (s), and on the sides of rectangle the limits that can be defined are supposed to exist and denoted as f (x1 , b2 −), f (b1 −, x2 ), f (b−). Denote fa+ (x) = ∆a f (x), x ∈ P, and fb− (x) := f (x) − f (x1 , b2 −) − f (b1 −, x2 ) + f (b−). Definition 2.2.4. Let f, g : P → R. The generalized two-parameter Lebesgue–Stieltjes integral of f w.r.t. g is defined by β1 β2 1−β1 1−β2 f dg := (Da+ fa+ )(u)(Db− gb− )(u)du P

+

i=1,2

P bi

ai

i (Daβii+ fai + )(u, ai )(Db1−β )(gbi − (u, bi −) − gbi − (u, bi −))du i−

+ f (a)∆a g(b),

(2.2.1)

under the assumption that all the integrals on the right-hand side exist. A more convenient formula for P f dg has a form β1 β2 1−β1 1−β2 f dg = (Da+ f )(u)(Db− gb− )(u)du. P

P

(We do not specify here the conditions ensuring the latter equality but it is very easy to do it, similarly to the one-parameter case.) The next results also can be proved similarly to the one-parameter case ((SKM93) and (Zah98)).

2.2 Pathwise Stochastic Integration w.r.t. Multi-parameter fBm

133

Theorem 2.2.5. Definition 2.2.4 is correct, i.e. the right-hand side of (2.2.1) does not depend on the choice of βi , i = 1, 2. Theorem 2.2.6. Let f : P → R, f ∈ C λ1 λ2 (P) and λi + βi < 1, i = 1, 2, β1 β2 0 < βi < 1. Then Ia+(b−) (fa+(b−) ) ∈ C λ1 +β1 λ2 +β2 (P). Theorem 2.2.7. Let the function f ∈ C λ1 λ2 (P). Then for any p ≥ 1 and 0 < εi < λi , i = 1, 2 ε1 ε2 (Lp (P)) fa+(b−) ∈ I± and

ε1 ε2 Da+(b−) fa+(b−) ∈ C λ1 −ε1 λ2 −ε2 (P).

β1 β2 (Lp (P)), gb− ∈ Theorem 2.2.8. Let f ∈ C(P), g ∈ BV (P), f ∈ I+ 1−β1 1−β2 1 1 I− (Lq (P)), i = 1, 2, j = 3 − i, p + q ≤ 1, 0 ≤ βi ≤ 1, i = 1, 2. Then the generalized two-parameter Lebesgue–Stieltjes integral P f dg equals the Riemann–Stieltjes integral P f (x)dg(x).

Theorem 2.2.9. 1. Let g ∈ C λ1 λ2 (P) for some 0 < λi ≤ 1, i = 1, 2. Then for any P1 = [c, d) ⊂ P 1P1 dg = ∆c g(d). P

(P) and let the partition π = π 1 × π 2 , where π i = {ai = xi0 < 2. Let g ∈ C i · · · < xni = bi } be the partition of [ai , bi ]. i −1 n Also, let fπ (x) = fj1 j2 1Pj1 j2 (x), where Pj1 j2 = i=1,2 [xiji , xiji +1 ). λ1 λ2

Then

i=1,2 ji =0

P

fπ dg =

i −1 n

i=1,2 ji =0

fj1 j2 ∆xj g(xj+1 ), where xj = (x1j1 , x2j2 ).

Now, let πn be the sequence of partitions of rectangle P, πn ⊂ πn+1 i,n and |πn | = maxi=1,2 max0≤ji ≤ni,n −1 (xi,n ji +1 − xji ). Let f : P → R, fj1 j2 = (n)

f (xnji +1 ). We say that the partitions πn are uniform, if n1 bi −ai xi,n (n) , i = 1, 2. ji =

(n)

= n2

and xi,n ji +1 −

n1

Theorem 2.2.10. 1. Let f ∈ C λ1 λ2 (P) for some 0 < λi ≤ 1, i = 1, 2. Then lim sup fπn − f L∞ (P) = 0,

n→∞ πn

where supπn is taken!over all the sequences of partitions mentioned above. ! ! β1 β2 ! β1 β2 = 0, 2. limn→∞ supπn !Da+ (fπn )a+ − Da+ fa+ ! L1 (P)

for any β1 ∨ β2 < λ1 ∧ λ2 and all the sequences of uniform partitions of P.

134


Proof. The first statement is a direct consequence of uniform continuity f on P. Further, let gn (x) = fπn (x) − f (x). For the second statement it is sufficient to prove that any of the following functions Gn1 (x) := gn (x)(x − a1 )−β1 (y − a2 )−β2 , x1 Gn2 (x) := (x2 − a2 )−β2 gn (x) − gn (s1 , x2 ) (x1 − s1 )−1−β1 ds1 , a 1x2 n −β1 gn (x) − gn (x1 , s2 ) (x2 − s2 )−1−β2 ds2 , G3 (x) := (x1 − a1 ) a2 n ∆s gn (x) (xi − si )−1−βi ds G4 (x) := [a,x]

i=1,2

≤ C(|πn |λ1 + |πn |λ2 ), whence tends to zero in L1 (P). First, note that |gn (x)| n λ1 λ2 1−βi (bi − ai ) → 0, n → ∞. Further, let G1 L1 (P) ≤ C(|πn | + |πn | ) i=1,2 i,n n n the point x ∈ Pjn := i=1,2 [xi,n ji , xji+1 ) =: [xj , xj+1 ). Then it holds that Gn2 (x)

−β2

= (x2 − a2 )

j −1 1

x1,n k+1

x1,n k

k=0

(x1 − s1 )−1−β1 ds1

x1

+

gn (x, xnj , s1 )(x1

−1−β1

− s1 )

ds1

,

xj1 2,n where gn (x, xnj , s1 ) = f (xnj ) − f (x) − f (x1,n k , xj2 ) + f (s1 , x2 ). Therefore,

|Gn2 (x)|I{x ∈ Pjn } ⎡ λi |xi,n − x | ≤ C(x2 − a2 )−β2 ⎣ i ji

+

λ1 −1−β1

(x1 − s1 ) xj1

⎡

≤ C(x2 − a2 )−β2 ⎣

x1,n k+1

x1,n k

/

x

+

(x1 − s1 )−1−β1 ds1

1,n λ1 2,n λ2 (x1,n + (x2,n j2 +1 − xj2 ) k+1 − xk )

k=0

1

a1

i=1,2 j 1 −1

x1,n j

(x1 − s1 )−1−β1 ds1

ds1

1,n −β1 λi (x1 − x1,n j1 ) (x1 − xj1 )

i=1,2

+

j 1 −1 k=0

((x1,n k+1

−

λ1 x1,n k )

2 λ1 −β1 + (x1 − x1,n ) , j1

+

(x2,n j2 +1

−

λ2 x2,n j2 ) )

x1,n k+1 x1,n k

(x1 − s1 )−1−β1 ds1


135

and

Gn1 L1 (P) ≤

Gn1 L1 (P n ) ≤ C

j

j1 ,j2

Pjn

j1 ,j2

λ1 −β1 (x2 − a2 )−β2 (x1 − x1,n j1 )

1,n −β1 λ2 + (x2 − a2 )−β2 (x2 − x2,n j2 ) (x1 − xj1 ) x1,n j 1 −1 k+1 1,n 1,n λ1 −β2 + (xk+1 − xk ) (x2 − a2 ) (x1 − s1 )−1−β1 ds1 x1,n k

k=0 −β2

+ (x2 − a2 )

(x2,n j2 +1

λ1 x2,n j2 )

−

1,n j 1 −1 x k+1

x1,n k

k=0 −β2

+ (x2 − a2 )

λ1 −β1 x1,n j1 )

(x1 − ⎛

≤ C(b2 − a2 )

1−β2

(x1 − s1 )−1−β1 ds1

dx (n)

⎝|πn |

λ1 −β1

n1

+ |πn |

λ2

1,n 1−β1 (x1,n j1 +1 − xj1 )

j1 =1

(n) n1 −1

+

1,n λ1 (x1,n k+1 − xk )

x1,n k

j1 =0 n1 −1 x1,n k+1 (n)

+ |πn |λ2

x1,n k

j1 =0

x1,n k+1

x1,n k

(

b1

x1,n k+1

(x1 − s1 )−1−β1 dx1 )ds1 ⎞

(x1 − s1 )−1−β1 ds1 dx1 + |πn |λ1 −β1 ⎠ .

a1

(2.2.2) The first, third and fifth terms on the right-hand side of (2.2.2) are bounded from above by C|πn |λ1 −β1 → 0, n → ∞, and it is true for any πn . The second and fourth terms can be effectively estimated when πn = πn is uniform. In this case (n)

|πn |λ2

n1

1,n 1−β1 (x1,n ≤ j1 +1 − xj1 )

j1 =1

and

(n)

|πn |λ2

n1−1 j1 =0

x1,n k+1

x1,n k

x1,n k a1

C (n) (n1 )λ2 −β1

→ 0,

n → ∞,

(x1 − s1 )−1−β1 ds1 dx1

(n)

≤ |πn |λ2

n 1

j1 =1

1,n 1−β1 (x1,n → 0, j1 +1 − xj1 )

Gn3 and Gn4 can be estimated in a similar way.

n → ∞.

Definition 2.2.11. We say that the two-parameter left Riemann–Stieltjes integral l- P f dg exists if the sums Sn have the limit for all sequences of uniform partitions of P with vanishing diameter.

136


Theorem 2.2.12. Let f ∈ C λ1 λ2 (P), g ∈ C µ1 µ2 (P) and λi + µi > 1, i = 1, 2. Then the generalized two-parameter Lebesgue–Stieltjes integrals P f dg and l- P f dg exist and coincide. Proof. It is sufficient to prove that Sn → P f dg. But the sums Sn equal Sn = P fπn dg. Denote f (n) := fπn . Then β1 β2 (n) 1−β1 1−β2 (n) f dg = Da+ f (x)Db− gb− (x)dx P

P

β1 β2 (n) β1 β2 for any 1−µi < βi < λi . According to previous theorem, Da+ f → Da+ f

in L1 (P), whence the proof follows.

Remark 2.2.13. We can use the Hölder properties of f in order to establish that P f dg = lim Sn , where Sn =

2,n 1,n 2,n n n n (f (x1,n j1 , ξj2 ) + f (ξj1 , xj2 ) − f (ξj ))∆xj g(xj+1 )

j1 j2

and ξjn is any point of Pjn . 2.2.3 Generalized Integrals of Two-parameter fBm in the Case of the Integrand Depending on fBm Since the trajectories of two-parameter fBm B H1 H2 a.s. belong to C H1 −ε1 H2 −ε2 (P) for any rectangle P ⊂ R2+ and any 0 < εi < Hi , the next result is a direct consequence of Theorem 2.2.12. Theorem 2.2.14. Let B H1 H2 be a two-parameter fBm with Hi ∈ (1/2, 1), and the function F : R+ × R → R, F ∈ C 1 (R+ × R). Then there exists the generalized two-parameter Lebesgue–Stieltjes integral P F (·, B H1 H2 )dB H1 H2 which coincides with the left Riemann–Stieltjes integral l- P F (·, B H1 H2 )dB H1 H2 . Remark 2.2.15. Theorem 2.2.14 holds if we replace F (·, B H1 H2 ) with any H¨ older field f ∈ C λ1 λ2 (P), such that λi + Hi > 1. It means that for such an f , we can consider the integral P f dB H1 H2 for any ω ∈ Ω , P (Ω ) = 1 as the limit of corresponding integral sums. 2.2.4 Pathwise Integration in Two-parameter Besov Spaces According to the form of two-parameter forward and backward fractional Marchaud derivatives (Definition 1.20.8), the Besov type spaces in this case receive the following form. Let Pt := [0, t] = i=1,2 [0, ti ], t ϕβ1 1 (f )(t) := 0 1 |f (t) − f (s1 , t2 )|(t1 − s1 )−β1 −1 ds1 , t ϕβ2 2 (f )(t) := 0 2 |f (t) − f (t1 , s2 )|(t2 − s2 )−β2 −1 ds2 ,


|∆s f (t)|(ϕ(t, s, 1 + β))−1 ds, 0 < βi < 1, βi := |f (t)| + ϕi (f )(t) + ϕβ3 1 β2 (f )(t).

ϕβ3 1 β2 (f )(t) := and

ϕβf 1 β2 (t)

137

Denote by R, such that

Pt

i=1,2 β1 ,β2 (PT ) the W0

Banach space of measurable functions f : PT →

f 0,β1 ,β2 := sup ϕβf 1 β2 (t) < ∞, t∈PT

W1β1 ,β2 (PT )

the Banach space of measurable functions f : PT → R, such that f 1,β1 ,β2 := sup0<s≤t 12 , λi + µ1 + µ2 > 2, i = 1, 2. Then there exists This limit will be called the integral of the second kind of f limn→∞ Sn =: S. w.r.t. g and denoted as S = P f d1 gd2 g. Proof. Let, for technical simplicity, T1 = T2 = 1. Also, let m > n. Consider the difference Sn − Sm = Sn − Smn + Smn − Sm , where n 2 −1

Smn =

f (r2−m , j2 2−n )(g((r + 1)2−m , j2 2−n ) − g(r2−m , j2 2−n ))

j1 ,j2 =0 r∈Aj1

× (g(r2−m , (j2 + 1)2−n ) − g(r2−m , j2 2−n )), Aj1 = {r : j1 2m−n ≤ r < (j1 + 1)2m−n }. It is sufficient to estimate only Sn − Smn , because Smn − Sm can be estimated similarly. We have that |Sn − Smn | ≤ |∆1mn | + |∆2mn |, where ∆1mn =

2n −1

f (tnj )∆jr g∆1j2 r g, ∆2mn =

2n −1

j1 ,j2 =0 r∈Aj1 j1 ,j2 =0 r∈Aj1 ∆jr g = ∆tnj g(r2−m , (j2 + 1)2−n ), ∆1j2 r g = ∆1(r2−m ,j2 2−n ) g((r + 1)2−m , (j2 + 1)2−n ), ∆1jr f = ∆1tnj f (r2−m , j2 2−n ), (j2 + 1)2−n ), ∆2j2 r g = ∆2(r2−m ,j2 2−n ) g(r2−m , (j2 + 1)2−n ).

∆1jr f ∆1j2 r g∆2j2 r g,

Transform ∆1mn into the sum ∆1mn

=

n 2 −1

f (tnj )∆j2 r g∆1jr g,

j1 ,j2 =0 r∈Aj1

where ∆j2 r g = ∆(r2−m ,j2 2−n ) (g((r + 1)2−m (j2 + 1)2−n )), and ∆1jr g = ∆1(r2−m ,j2 2−n ) g(tnj1 +1j2 ). The increments ∆j2 r g correspond to the


139

rectangles ∆j2 r = ( r2−m , (r + 1)2−m ] × ( j2 2−n , (j2 + 1)2−n ], that do not 2 intersect, and ∪∆j2 r = ( 0, 1 ] . Therefore the sum ∆1n,m can be presented as a two-parameter generalized Lebesgue–Stieltjes integral P fmn dg, where fmn (s) = f (tnj )∆1jr g · 1{s∈∆j2 r } . In turn,

P

fmn dg =

P

β1 β2 1−β1 1−β2 (D0+ g1− )(s)ds, fmn )(s)(D1−

where 1 = (1, 1), 0 = (0, 0), 1 − µi < βi < λi , i = 1, 2. With such a choice 1−β1 1−β2 of βi D1− g1− ∈ C µ1 +β1 −1µ2 +β2 −1 (P), in particular, there exists such 1−β1 1−β2 g1− )(s)| ≤ C, s ∈ P. Therefore, it is sufficient to a C > 0 that |(D1− β1 β2 β1 β2 fmn )(s)|ds → 0, n, m → ∞. Since D0+ prove that P |(D0+ fmn consists of four terms, we must consider them separately. Estimate only P |ϕmni (s)|ds, where s1 2 (fmn (s) − fmn (u1 , s2 ))(s1 − u1 )−1−β1 du1 , ϕmn1 (s) = s−β 2 0 and (si − ui )−1−βi du1 ; ϕmn2 (s) = [0,s] ∆u fmn (s) i=1,2

the other two terms can be considered similarly. Let s ∈ ∆j2 r . Then, taking into account that |f (s)| ≤ C for some C > 0, we obtain that j 2−n r2−m 2 ( 01 + j1 2−n )|fmn (s) − fmn (u1 , s2 )|(s1 − u1 )−1−β1 du1 |ϕmn1 (s)| ≤ s−β 2 −n j 2 1 2 ≤ s−β (|fmn (s)| + |fmn (u1 , s2 )|)(s1 − u1 )−1−β1 du1 2 0 −m −β2 r2 2 + Cs2 |f (tnj )|(s1 − u1 + 2−m )µ1 (s1 − u1 )−1−β1 du1 ≤ Cs−β 2 j1 2−n −nµ1 −n −β1 −m µ1 −β1 −mµ1 −m −β1 ×(2 (s1 − j1 2 ) + (s1 − r2 ) +2 (s1 − r2 ) ), whence

|ϕmn1 (s)|ds ≤ C P

+

∆ j2

2n −1

j1 ,j2 =0 r∈Aj1

2−nµ1

∆ j2 r

2 s−β (s1 − r2−m )µ1 −β1 ds + 2−mµ1 2 r

2 s−β (s1 − j1 2−n )ds 2

∆ j2

≤ C(1 − β2 )−1 (2n(β1 −µ1 ) + 2m(β1 −µ1 ) ) → 0,

2 s−β (s1 − r2−m )−β1 ds 2 r

m, n → ∞.

Further, from H¨ older properties of f and g, it follows that for u ≤ (j1 2−n , j2 2−n ) we have the estimate |∆u fmn (s)| ≤ 2(s2 − u2 + 2−n )λ2 2−nµ1 + C(s2 − u2 + 2−n )µ2 (s1 − u1 )−nµ1 , for u ∈ (j1 2−n , r2−m ) × (0, j2 2−n ) the estimate is |∆u fmn (s)| ≤ 2(s2 − u2 + 2−n )λ2 (s1 − u1 + 2−m )µ1 + C2−mµ1 (s2 − u2 + 2−n )µ2 , and ∆u fmn (s) = 0 otherwise. Hence, |ϕmn2 (s)| ≤ C2−nµ1 (s1 − j2 2−n )−β1 (s2 − (j1 − 1)2−n )λ2 ∧µ2 −β2 + C(s1 + j2 2−n + 2−m )µ1 −β1 (s2 − j2 2−m + 2−n )µ2 ∧µ2 −β2 ,

140


and P |ϕmn2 (s)|ds ≤ C2n(β1 +β2 −µ1 −µ2 ∧λ2 ) → 0, m, n → ∞. So, |∆1mn | → 0, m, n → ∞. Now we want to prove that |∆2mn | → 0, m, n → ∞. We can present ∆2mn as n 2 −1 2 2 ∆2,j ∆mn = mn , j2 =0

where 2 ∆2,j mn =

n 2 −1

∆1jr f ∆1j2 r g∆2j2 r g.

j1 =0 r∈Aj1 2 Moreover, ∆2,j can be presented as one-parameter generalized mn 1 ψj2 (u)d1 g(u, j2 2−n ), where ψj2 (u) = Lebesgue–Stieltjes integral 0 1 1 2 −n ∆ f ∆ g1{r2−m ≤u 2 we can choose 12 > β > 1 − µ1 in such a way that 1 + β − λ1 − µ2 < 0. Finally, for j1 2−n ≤ z ≤ u ≤ (r + 1)2−m |ψj2 (u) − ψj2 (z)| ≤ 2−nµ2 (u − z + 2−m )λ1 , and

2n −1 j2 =0

1 0

|

r2−m j1 r −n

(ψj2 (u) − ψj2 (z))(u − z)−1−β dz|du

≤ C2m(1+β1 −λ1 −µ2 ) → 0, m → ∞.

Remark 2.2.18. For f (s) = C ∆2mn = 0, and it is easy to see from the bounds of ∆1mn that the theorem will hold under the assumption λi , µi > 12 , i = 1, 2. Remark 2.2.19. Multiple stochastic fractional integral with Hurst parameter less than 1/2 was considered in (BJ06).

2.3 Wick Integration with Respect to fBm with H ∈ [1/2, 1) as S ∗-integration 2.3.1 Wick Products and S ∗ -integration Recall (see Sections 1.4–1.5), that the random variable F on the probability space S (R) belongs to S ∗ if F admits the formal expansion (1.5.1) with finite negative norm 2 F −q = α! c2α (2N)−qα < ∞ α∈I

for at least one q ∈ N. Introduce the following notations: Let the function Z : R → S ∗ , and for any F ∈ S we have that Z(t), F ∈ L1 (R) as a function of t ∈ R. (ii) In this case, define R Z(t)dt as the unique element of S ∗ such that 77 88 Z(t)dt, F = Z(t), F dt, (i)

R

R

∗

and say that Z is integrable in S . (iii) Define the Wick products: for F (ω) = α cα Hα (ω), and G(ω) = β dβ Hβ (ω), put (F ♦ G)(ω) = α,β cα dβ Hα+β (ω). According to the (HOUZ96), for F, G, H ∈ S it holds that (iv) F ♦ G = G ♦ F ; (v) (F ♦ G) ♦ H = F ♦(G ♦ H); (vi) H ♦(F + G) = H ♦ F + H ♦ G; (vii) F ♦ G ∈ S if F, G ∈ S; F ♦ G ∈ S ∗ if F, G ∈ S ∗ . In this section we consider only the case H ∈ [1/2, 1).

142


∗ Theorem 2.3.1. Let the process Y (t) ∈ S and admit an expansion Y (t) = α cα (t)Hα (ω), t ∈ R, with the coefficients, satisfying the inequality

K := sup{α! cα L1 (R) (2N)−qα } < ∞ 2

α

for some q > 0. Then the Wick product Y (t) ♦ B˙ tM is S ∗ -integrable, and, moreover, Y (t) ♦ B˙ tM dt = cα (t)M+ (2.3.1) hk (t)dt · Hα+εk (ω). R

R

α,k

Proof. Consider only B˙ tH , and for arbitrary B˙ tM the proof is the same. Since hk , ω = Hεk (ω), we have that the Wick product Y (t) ♦ B˙ tH ∈ S ∗ and H equals α,k cα (t)M+ hk (t)Hα+εk (ω). According to (HOUZ96, Lemmas 2.5.6 ∗ and 2.5.7), the S -integrability of Y (t) ♦ B˙ H follows from the inequality t

! !2 ! ! ! ! H ! β! ! cα (t)M+ hk (t)! ! !α,k:α+εk =β ! β∈I

(2N)−pβ < ∞

L1 (R)

for some p > 0. According to estimate (1.5.3), H M+ hk (t) ≤ Ck 2/3−H/2 < Ck5/12 for any k ≥ 1 and some C > 0. Therefore, H hk (t) dt ≤ Ck 5/12 cα L1 (R) , cα (t)M+ R

and ! !2 ! ! ! ! H ! cα (t)M+ hk (t)! ! ! !α,k:α+εk =β !

⎛

≤⎝

! ! ! ! H hk (t)! !cα (t)M+

α,k:α+εk =β

L1 (R)

⎛

≤C⎝

k 5/12 cα L1 (R) ⎠ .

Consider the sum S :=

β∈I

≤

β∈I

⎛ β! ⎝

⎞2

k 5/12 cα L1 (R) ⎠ (2N)−pβ

α,k:α+εk =β

⎛

β!(l(β))5/6 ⎝

α,k:α+εk =β

⎠ L1 (R)

⎞2

α,k:α+εk =β

⎞2

⎞2 cα L1 (R) ⎠ (2N)−pβ ,

2.3 Wick Integration with Respect to fBm

143

where l(β) equals the number of the last nonzero element in the index β (the length of the index β). Further, for any α, β there exists no more than one k, such that α + εk = β. Therefore, ⎛

⎞2

⎝

cα L1 (R) ⎠ ≤ l2 (β)

α,k:α+εk =β

2

cα L1 (R) .

α,k:α+εk =β

It means that S≤

(α + εk )!(l(α + εk ))17/6 cα L1 (R) (2N)−pα−pεk 2

α,k

≤K

(α + εk )! α!

α,k

≤K

(l(α + εk ))3 (2N)−(p−q)α−pεk

(|α| + 1)4 2−|α|(p−q) k −p < ∞,

α,k ∗ ˙H for p > q + 1, and we have established the S -integrability of Y (t) ♦ Bt . Now, for any F = β,k dβ,k Hβ+εk (ω) ∈ S, we have from the definition of the S ∗ -integral and of Wick product, that :: 9 88 9 7 7 H Y (t) ♦ B˙ tH dt, F = cα (t)M+ hk (t)Hα+ε (ω), F dt R

R

=

k

α,k

(α +

R α,k

(2.3.2) H εk )!cα (t)dα,k M+ hk (t)(ω)dt.

Note that

2

(α + εk )! |dα,k | (2N)2q(α+εk ) =: Cq < ∞

α,k

for any q ∈ N. Therefore α,k

R

H (α + εk )! |cα (t)| |dα,k | M+ (α + εk )! |dα,k | k 5/12 cα L1 (R) hk (t) dt ≤ α,k

⎛ ≤⎝

2

βk ! |dα,k | (2N)2qβk

α,k

⎞1/2 k 5/6 cα L1 (R) βk !(2N)−2q(α+εk ) ⎠ 2

α,k

⎛ ≤ ⎝Cq K

α,k

⎞1/2 β ! k (2N)−q|α| k −2q ⎠ k 5/6 11/12, βk = α + εk , because α βα!k ! (2N)−q|α| ≤ α (|α| + 1)2−q|α| < ∞. So, we can change the signs of sum and integral in (2.3.2) and obtain 88 7 7 H Y (t) ♦ B˙ tH dt, F = (α + εk )!dα,k cα (t)M+ hk (t)(ω)dt R

R

α,k

=

99 R

α,k

::

H cα (t)M+ hk (t)(ω)dt, F

,

whence (2.3.1) follows.

Corollary 2.3.2. Let Y (t) = α cα (t)Hα (ω) ∈ S ∗ be a process such that T T T EY 2 (t)dt < ∞ for some T > 0. Then α α! 0 c2α (t)dt = 0 EY 2 (t)dt < 0 2 ∞, whence K := supα {α! cα L1 (R) (2N)−qα } < ∞ for any q > 0 (hereafter we put cα (t) := cα (t)1[0,T ] (t)). So, we can use Theorem 2.3.1 and conclude that Y (t) ♦ B˙ tM is S -integrable, and, moreover, equality (2.3.1) holds. ∗

Corollary 2.3.3. Let Y (t) ≡ 1. Then the previous corollary holds with c0 (t) = 1, cα (t) = 0 for α = 0,whence

T

B˙ tM dt = 0

T

0

k

M+ hk (t)dt · Hεk (ω) = BTM .

In this connection, we can say that the fractional noise is the S ∗ -derivative of fBm. As a consequence, we can define R Yt ♦ dBtM := R Yt ♦ B˙ tM dt for the process Yt , satisfying the conditions of Theorem 2.3.1. Now, let Y ∈ L2 [0, T ] be some nonrandom function, H ∈ (1/2, 1). Then cα (t) = Y (t) = cα (t), for α = 0 and cα ≡ 0 for other α, so, by using Theorem 2.3.1, we obtain that

T

Y (t) ♦ B˙ tH dt = 0

T 0

T

H Y (t)M+ hk , ω. hk (t)dt ·

0

k

H Further, even for Y ∈ L1 [0, T ] we can replace the operator M+ and obtain T H H Y (t)M+ hk (t)dt = 0 M− Y (t)hk (t)dt, whence

T

Y (t) ♦ B˙ tH dt = 0

k

=

R

k

R

H M− Y (t) hk (t)dt · hk , ω

(2.3.3) H M− Y

(t) hk (t)dt · Hεk (ω),


145

where Y (t) = Y (t)1[0,T ] (t). The right-hand side of (2.3.3) corresponds to T H Y (t) ♦ B˙ t dt, where (HOUZ96, representation (2.5.22)) of the integral 0 M− 1/2 ˙ ˙ Bt = Bt is a white noise: 0

T H M− Y (t) ♦ B˙ t dt =

α,k

0

T

cα (t) hk (t)dt · Hα+εk (ω).

Therefore, for Y ∈ LH 2 [0, T ]

T

Y (t) ♦ B˙ tM dt = 0

R

M− Y (t) ♦ B˙ t dt =

R

M− Y (t) · B˙ t dt.

(2.3.4)

2.3.2 Comparison of Wick and Pathwise Integrals for “Markov” Integrands In this subsection we can, without losing generality, consider instead of S (R) the probability space Ω = C0 (R+ , R) of real-valued continuous functions on R+ with the initial value zero and the topology of local uniform convergence. There exists a probability measure P on (Ω, F), where F is the Borel σ-field, such that on the probability space (Ω, F, P ) the coordinate process B : Ω → R defined as, Bt (ω) = ω(t), ω ∈ Ω is the Wiener process. (i)

(ii)

Recall the notion of a stochastic derivative. Let F be a square-integrable random variable, and suppose that the limit . 1 lim h(s)ds) − F (ω.) exists in L2 (P ) F (ω. + β β→0 β 0 for any h ∈ L2 (R). Then this limit is called the directional derivative Dh F . If the directional derivative Dh F , h ∈ L2 (R), is absolutely continuous w.r.t. the measure h(x)dx, i.e. dDh (F ) (x) · h(x)dx, Dh (F ) = dh R

and (dDh (F ))/(dh) does not depend on h, then the Radon–Nikodym derivative (dDh (F ))/(dh) is called the stochastic derivative of F and is denoted by Dx F . (iii) We have a chain rule for the stochastic derivative: if Dx F exists and ϕ ∈ C 1 (R), then Dx ϕ(F ) has the stochastic derivative Dx ϕ(F ) = ϕ (F )Dx F.

146


(iv) Let u ∈ L2 (R) be a nonrandom function. Then it follows from (NP95, Proposition 5.5), that Dx us dBs = ux a.e. R

Recall the notion of the class D1,2 . This is the Banach space, obtained as a completion of the set P0 of smooth ! functionals ! F = f (Bt1 , . . . , Bti ), w.r.t. the norm F 1,2 := F L2 (P ) + ! Dx F HS !L (P ) , where F ∈ P0 , 1 and ·HS denotes the Hilbert–Schmidt norm. 2 Denote LM 2 (R) = {f : R → R : R |M− f (x)| dx < ∞}. (v)

Lemma 2.3.4. Let F ∈ D1,2 , f ∈ LM 2 (R). Suppose that the integrals (M− f )(s) · Ds F ds and F · (M− f )(s)dBs = F · f (s)dBsM R

R

R

belong to L2 (P ). Then F ♦ R f (s)dBsM exists and F ♦ f (s)dBsM = (F · M− f )(s)δBs R

R

=F ·

R

f (s)dBsM −

R

(M− f )(s) · Ds F ds.

(2.3.5)

Proof. By using (HOUZ96, Corollary 2.5.12) and (NP95, Theorem 3.2), we obtain for nonrandom f that F ♦ f (s)dBsM = F ♦ (M− f )(s)dBs R R = (F ♦ M− f )(s)δBs = (F · M− f )(s)δBs R R = F · (M− f )(s)dBs − (M− f )(s) · Ds F ds R R M = F · f (s)dBs − (M− f )(s) · Ds F ds. R

R

(Note that according to (NP95, Theorem 3.2), the Skorohod integral F · (M f )(s)δB exists if and only if the difference F · (M− f )(s)dBs − s R R − R (M− f )(s) · Ds F ds belongs to L2 (P )).

Using this result, we can compare the Wick integral and the pathwise integral w.r.t. fBm BtH , H ∈ (1/2, 1)(the latter integral coincides with H . Stratonovich integral). Therefore, now M± = M±


147

Lemma 2.3.5. Let ϕ ∈ C 1 (R), Ft = ϕ(BtH ), f (s) = 1[t,t+h] (s), t, h > 0. If H − BtH ) belong to L2 (P ), then ϕ (BtH ) and Ft · (Bt+h H H Ft ♦(Bt+h − BtH ) = F · (Bt+h − BtH )

− Hϕ (BtH )t2α h + c(ω)(t2α−1 h2 + h2H ), where c(ω) is a.s. finite and independent of t and h. Proof. According to equation (2.3.5), we can rewrite formally the left-hand side of the previous equality: H H Ft ♦(Bt+h − BtH ) = Ft · (Bt+h − BtH ) H − M− 1[t,t+h] (s)Ds ϕ(BtH )ds.

(2.3.6)

R

Further, according to the chain rule (iii), it holds that Ds ϕ(BtH ) = ϕ (BtH )Ds BtH ,

and Ds BtH

= Ds

R

H H M− 1[0,t] (u)dBu = (M− 1[0,t] )(s).

Therefore, H H Ft ♦(Bt+h − BtH ) = Ft · (Bt+h − BtH ) H H H M− 1[t,t+h] (s) M− − ϕ (Bt ) 1[0,t] (s)ds, R

and under the conditions of the lemma the right-hand side of equation (2.3.6) is well-defined. Finally, R

H H H (M− 1[t,t+h] )(s)(M− 1[0,t] )(s)ds = E(Bt+h − BtH )BtH

=

1 ((t + h)2H − t2H − h2H ) = Ht2α h + 2Hαθ2α−1 h2 − h2H , 2

where θ ∈ (t, t + h). The lemma is proved.

2+ε Remark 2.3.6. Evidently, the assumption E ϕ(BtH ) < ∞ for some ε > 0 H − BtH ) to belong to L2 (P ). is sufficient for Ft (Bt+h

Now, fix some T > 0 and consider the sequence πn = {0 = tn0 < · · · < = T } of partitions of [0, T ], such that πn ⊂ πn+1 and |πn | → 0 as n → ∞. Suppose that tnn

ϕ (BtH ) ∈ L2 (P ), ϕ(BtH ) ∈ L2+ε (P ), t ∈ [0, T ]

(2.3.7)

148


for some ε > 0. According to Lemma 2.3.5, we can write n

H ϕ(BtHni−1 ) ♦ ∆Bi,n

=

n

i=1

H ϕ(BtHni−1 )∆Bi,n

i=1

−H

n

ϕ (BtHni−1 )(tni−1 )2α ∆ti,n + Rn (T ),

i=1 H where ∆ti,n = tni − tni−1 , ∆Bi,n = BtHni − BtHni−1 . Here Rn (T ) is a remainder term and Rn (T ) → 0 a.s. as n → ∞. Furthermore, the process Ct := ϕ(BtH ) is Hö lder continuous up to order H. Also, by Theorem 2.1.7, part 2), the n H converges a.s. as n → ∞ to the pathwise integral sum i=1 ϕ(BtHni−1 )∆Bi,n T H H ϕ(Bs )dBs . Clearly, 0 n

ϕ (BtHni−1 )(tni−1 )2α ∆ti,n →

i=1

T

ϕ (BsH )s2α ds a.s.

0

Therefore, lim

n→∞

n

T

ϕ(BsH )dBsH − H

H ϕ(BtHni−1 ) ♦ ∆Bi,n =

0

i=1

T

ϕ (BsH )s2α ds a.s.

0

Moreover, under assumption (2.3.7) and

T

E

2 ϕ(BsH ) ds < ∞,

(2.3.8)

0

there exists the Wick integral prove that

T 0

ϕ(BsH ) ♦ dBsH . Now we are in a position to

T

ϕ(BsH ) ♦ dBsH = lim

n→∞

0

n

H ϕ(BtHni−1 ) ♦ ∆Bi,n .

(2.3.9)

i=1

Theorem 2.3.7. Under conditions (2.3.7) and 2 E sup ϕ(BsH ) + E sup (ϕ (BsH ))2 < ∞ s≤T

s≤T

equality (2.3.8) and (2.3.9), consequently, the equality

T

ϕ(BsH ) ♦ dBsH 0

holds a.s.

T

ϕ(BsH )dBsH

= 0

−H 0

T

ϕ (BsH )s2α ds

(2.3.10)


149

Proof. Let the random variables F, G ∈ D1,2 . According to equality (2.3.5) and (NP95, Theorem 3.2), for i ≤ k < ; H H · G ♦ ∆Bk,n E F ♦ ∆Bi,n * + H H F M− 1[tni−1 ,tni ] (s)δBs · GM− 1[tnk−1 ,tnk ] (s)δBs =E R *R + H H =E F GM− 1[tni−1 ,tni ] (s)M− 1[tnk−1 ,tnk ] (s)ds R * + (2.3.11) H H +E Dt F Ds GM− 1[tni−1 ,tni ] (t)M− 1[tnk−1 ,tnk ] (s)ds dt R×R

1 = E[F Grik ] 2 * +E R

H Dt F M− 1[tni−1 ,tni ] (t)dt ·

R

+ H Ds GM− 1[tnk−1 ,tnk ] (s)ds ,

where 2H rik = tnk−1 − tni + (tnk − tni−1 )2H − (tnk − tni )2H − (tnk−1 − tni−1 )2H . Put in (2.3.11) F = ϕ(BtHni−1 ), G = ϕ(BtHnk−1 ) and take the sum over 1 ≤ i ≤ k ≤ n. We obtain that

n 2 H E ϕ(BtHni−1 ) ♦ ∆Bi,n = S1n + S2n , i=1

where S1n =

Eϕ(BtHni−1 )ϕ(BtHnk−1 )rik ,

1≤i≤k≤n

and

S2n =

E

1≤i≤k≤n

× =

1 4

R

R

H H ϕ (BtHni−1 )M− 1[tni−1 ,tni ] (t)M− 1[0,tni−1 ] (t)dt

H H ϕ (BtHnk−1 )M− 1[tnk−1 ,tnk ] (s)M− 1[0,tnk−1 ] (s)ds

Eϕ (BtHni−1 )ϕ (BtHnk−1 ) (tnk )2H − (tnk−1 )2H − (∆tnk )2H

1≤i≤k≤n

× (tni )2H − (tni−1 )2H − (∆tni )2H . Evidently,

|S2n | ≤ H 2 E

n H 2α ϕ (Btni−1 ) ti · ∆tni i=1

2 .

(2.3.12)

150

2 Stochastic Integration with Respect to fBm and Related Topics iT n

If the partition πn is uniform, i.e. tni = S1n

≤2

1≤i≤n

, then for some CH > 0

2 iT 2H H E ϕ(Btni−1 ) n

2H T + CH n

H H ϕ(Btni−1 )ϕ(Btnk−1 ) ·

i

i−1

1≤i≤k≤n

k

(u − v)2α−1 du dv. k−1

(2.3.13) Now it is very easy to conclude from (2.3.10)–(2.3.13), that the sums Sn :=

n

H ϕ(BtHnk ) ♦ ∆Bk,n

k=1

form a Cauchy sequence in L2 (P ), at least, for uniform πn . From the estimate |F, g| ≤ F L2 (P ) gL2 (P ) , F ∈ L2 (P ), g ∈ S, we obtain that Sn − Sm , g → 0, n, m → ∞ for any g ∈ S. This means that {Sn } is a Cauchy sequence in the weak sense. If we establish the weak T convergence Sn → S := 0 ϕ(BsH ) ♦ dBsH , then the theorem will be proved, since the convergence will be in L2 (P ), as well. According to (2.3.1) and Corollary 2.3.2, we have that S =

T

ϕ(BtH ) ♦ B˙ tH dt =

0

ϕn (t) ♦ B˙ tH dt =

Sn =

0

α,k

T

0

T

0

α,k

T

H cα (t)M+ hk (t)dt · Hα+εk (ω),

H cnα (t)M+ hk (t)dt · Hα+εk (ω),

where ϕn (t) =

n

ϕ(BtHni−1 )1[tni−1 ,tni ) (t),

i=1

ϕ(BtH ) =

cα (t)Hα (ω), cnα (t) =

α

n

cα (tni−1 )1[tni−1 ,tni ) (t).

i=1

Denote dnα := cα − cnα . Then S − Sn =

β α,k:α+εk =β

Furthermore, for any g =

β

T

H dnα (t)M+ hk (t)dt · Hβ (ω).

0

gβ Hβ (ω) ∈ S and any q > 0


151

T == >> n H β! gβ dα (t)M+ hk (t)dt S − Sn , g ≤ α,k:α+εk =β 0 β ⎛ ⎞1/2 ≤⎝ β!(gβ )2 (2N)βq ⎠ β

⎞1/2

⎛

!2 ! ! ! ! n H ! ⎜ ×⎝ β! ! dα M+ hk ! ! ! ! !α,k:α+εk =β β

⎟ (2N)−βq ⎠

.

L1 [0,T ]

We estimate only the second multiplicand. According to (1.5.3), for H H ∈ (1/2, 1) M+ hk (t) ≤ Ck 5/12 with constant C independent of t, k. So, !2 ! ! ! ! n H ! ! d α M+ h k ! ! ! ! !α,k:α+εk =β

⎛

⎞2

≤C⎝

k 5/12 dnα L1 [0,T ] ⎠

α,k:α+εk =β

L1 [0,T ]

⎛

≤ C(l(β))5/6 ⎝

⎞2 dnα L1 [0,T ] ⎠ ,

α,k:α+εk =β

where l(β) equals the number of nonzero entries in β. Further, !2 ! ! ! ! n H ! β! (2N)−βq ! dα M+ hk ! ! ! ! ! β α,k:α+εk =β L1 [0,T ] ⎛ ⎞2 ≤ β! (2N)−βq l(β)5/6 ⎝ dnα L1 [0,T ] ⎠

β

≤

β

≤

17/6

β! l(β)

α,k:α+εk =β

dnα L1 [0,T ] (2N)−βq

α:∃k,α+εk =β

(α + εk )!(l(α + εk ))17/6 dnα L1 [0,T ] (2N)−q(α+εk ) 2

α,k

(α + ε )! k 2 (l(α + εk ))17/6 (2N)−qα (2N)−qεk ≤ sup α! dnα L1 [0,T ] α! α α,k

2 ≤ sup α! dnα L1 [0,T ] (|α| + 1)23/6 2−|α|q k −q . α

α,k

The last series converges for q > 1, and it follows from the continuity of ϕ and condition (2.3.10), that

152


2 sup α! dnα L1 [0,T ] ≤ α! dnα L2 [0,T ] · T α

α

! ! = T !ϕ(B·H ) − ϕn (·)!L

2 [0,T ]

→ 0, n → ∞.

Theorem 2.3.7 can be generalized to the processes of the form BtM :=

m

σk BtHk .

k=1

Suppose that H1 =

1 2

and Hk ∈ (1/2, 1), 2 ≤ k ≤ m.

Theorem 2.3.8. Assume that conditions (2.3.7), (2.3.8) and (2.3.10) hold with BtH replaced by BtM . Then T T M M ϕ(Bt ) ♦ dBt = ϕ(BtM )dBtM 0

−

n

0

H H (Hi + Hk ) σi σk C i k

i,k=1

T 0

1 ϕ (BsM )sHi +Hk −1 ds + σ12 2

T

ϕ (BsM )ds,

0

where

⎧ (3) (3) ⎪ CHi CHk B(Hi − 1/2, 2 − Hi − Hk ) ⎪ ⎪ , ⎪ ⎪ ⎪ ⎪ (Hi + Hk )(Hi + Hk − 1)Γ (Hi − 1/2)Γ (Hk − 1/2) ⎪ ⎪ ⎪ ⎪ Hi , Hk ∈ (1/2, 1), ⎪ ⎪ ⎨ (3) H H = CHk C i k , Hi = 1/2, Hk ∈ (1/2, 1), ⎪ ⎪ ⎪ Γ (Hk + 3/2) ⎪ ⎪ ⎪ ⎪ ⎪ 0, Hi ∈ (1/2, 1), Hk = 1/2, ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 1 , Hi = Hk = 1/2. 2 Proof. We start with (2.3.5) and conclude that M M − BtM ) = ϕ(BtM ) · (Bt+h − BtM ) ϕ(BtM ) ♦(Bt+h m Hk Hi − ϕ (BtM ) σi σk M− 1[t,t+h] (s)M− 1[0,t] (s)ds. i,k=1

R

Hk i Further, for f ∈ LH 2 (R), g ∈ L2 (R), Hi , Hk ∈ (1/2, 1) ∞ (1) Hk Hi M− f (s)M− g(s)ds = Ci,k,H (x − s)Hi −3/2 f (x)dx R R s ∞ (1) × (y − s)Hk −3/2 g(y)dy ds = Ci,k,H f (x)g(y)dx dy s R2 x∧y × (x − s)Hi −3/2 (y − s)Hk −3/2 ds, −∞

2.3 Wick Integration with Respect to fBm (3)

(1)

where Ci,k,H =

x∧y

−∞

153

(3)

CH CH i k Γ (Hi −1/2)Γ (Hk −1/2) .

Evidently,

(x − s)Hi −3/2 (y − s)Hk −3/2 ds (2) (2) H +H −2 Ci,k,H 1{y > x} + Ck,i,H 1{y ≤ x} . = |y − x| i k

∞ (2) with Ci,k,H = 0 z Hi −3/2 (1 + z)Hk −3/2 dz = B(Hi − 1/2, 2 − Hi − Hk ). Therefore, (1) H +H −2 Hk Hi M− f (s)M− g(s)ds = Ci,k,H f (x) |y − x| i k R R (2) (2) · Ci,k,H 1{x < y} + Ck,i,H 1{y < x} dx dy. Let f (x) = 1[t,t+h] (x), g(y) = 1[0,t] (y). Then Hk Hi M− 1[t,t+h] (s)M− 1[0,t] (s)ds R

=

(j) Ck,i,H

j=1,2

=

t 0

t+h

(y − x)Hi +Hk −2 dy dx

t

Ck,i,H ((Hi + Hk )(Hi + Hk − 1))−1 (j)

j=1,2

< ; × (t + h)Hi +Hk − tHi +Hk − hHi +Hk ; < H H (t + h)Hi +Hk − tHi +Hk − hHi +Hk =: C i k ; H H (Hi + Hk )tHi +Hk −1 h + (Hi + Hk )(Hi + Hk − 1)θHi +Hk −1 h2 =C i k < − hHi +Hk , θ ∈ (t, t + h). (2.3.14) 1/2 For Hi = 1/2 and Hk ∈ (1/2, 1) we have that M− = I is identity operator, and ∞ (3) CHk 1/2 Hk M− f (s)M− g(s)ds = f (s) g(y)(y − s)Hk −3/2 dy ds. Γ (Hk − 1/2) R s R For f and g as above, the last integral equals t t+h (3) CHk (y − s)Hk −3/2 dy ds Γ (Hk − 1/2) 0 t (3) 2 1 CHk = (t + h)Hk +1/2 − tHk +1/2 − hHk +1/2 Γ (Hk + 3/2) 1 1 H (Hk + 1/2)tHk −1/2 h =: C k 2

2 + (Hk + 1/2)(Hk − 1/2)tHk −2 h2 − hHk +1/2 .

(2.3.15)

154


At last, for Hi = Hk = 1/2 1/2 1/2 M− 1[0,t] (s)M− 1[t,t+h] (s)ds = 0.

(2.3.16)

R

1 Now we can proceed as in Lemma 2.3.5 and Theorem 2.3.7, put C 2 take into account (2.3.14)–(2.3.16) and obtain the proof.

1 2

:= 12 ,

2.3.3 Comparison of Wick and Stratonovich Integrals for “General” Integrands Now we consider the general process Ft instead of ϕ(BtM ). Suppose that fBm {BtH , t ≥ 0} is “one-sided”, H ∈ ( 12 , 1). Theorem 2.3.9. Let {Ft , Ft , t ∈ [0, T ]} be the stochastic process satisfying the conditions (i)

(ii)

2+ε

< ∞ for any t ∈ [0, T ] and Ft ∈ D1,2 for any t ∈ [0, T ], E |Ft | some ε > 0, sups,t∈[0,T ] |Ds Ft | is bounded in probability; limh↓0 supt∈[0,T ] |Dt Fs − Dt Fs+h | = 0 in probability;

older continuous of order α > 1 − H (this condition implies (iii) Ft is a.s. H¨ T the existence of the Stratonovich integral 0 Ft dBtH , H ∈ (1/2, 1)); T (iv) E 0 Ft2 dt < ∞ (this condition implies the existence of the Wick T integral 0 Ft ♦ dBtH , according to Corollary 2.3.2); (v)

there exists a sequence of partitions {πn , n ≥ 1} with |πn | → 0 as n → ∞ T n H converge to 0 Ft ♦ dBtH such that the integral sums k=1 Ftnk−1 ♦ ∆Bk,n in probability.

Then

T

Fs ♦ dBsH = 0

0

T

(3)

Fs dBsH − CH

T

0

s

(s − t)α−1 Ds Ft dt ds.

0

Proof. Consider for any 0 ≤ t < t + h ≤ T the function f (u) = 1[t,t+h] (u). Then we take into account that Ds Ft = 0 for s > t and s < 0 (since Ft is Ft (3) t t+h H adapted) and obtain that R M− f Ds Ft ds = CH 0 t (u−s)α−1 duDs Ft ds, t+h α where t (u − s)α−1 du ≤ hα . Hence,

E R

2

H M− f Ds Ft ds

Further, Ft ·

≤

α2

R

(3)

CH

2

t

2

|Ds Ft | ds < ∞.

h2α tE 0

H

H H M− f dBs = Ft · Bt+h − Bt , and, according to (i),


155

ε 2(2+ε) 2+ε H − BtH ε < ∞. E Bt+h

2 2 H 2+ε 2+ε E Ft · Bt+h − BtH ≤ E |Ft |

H H Therefore, R M− f · Ds Ft ds and Ft · R M− f dB sn belong to L2 (PH) and it exist. follows from Lemma 2.3.4 that the integral sums k=1 Ftnk−1 ♦ ∆Bk,n Moreover, H H H n n M− 1[tnk−1 ,tnk ] (s)Ds Ftnk−1 ds Ftk−1 ♦ ∆Bk,n = Ftk−1 · ∆Bk,n − R H H − 1[tnk−1 ,tnk ] (s) M+ (D. Ftnk−1 ) (s)ds = Ftnk−1 · ∆Bk,n R

=F

tn k−1

·

H ∆Bk,n

−

(3) CH

tn k

tn k−1

0

tn k−1

(s − u)α−1 Du Ftnk−1 du ds. (2.3.17)

Consider the difference, n n n tk tk−1 (s − u)α−1 Du Ftnk−1 du ds n k=1 tk−1 0 T s α−1 − (s − u) Du Ftnk−1 1[tnk−1 ,tnk ) (s)du ds 0 0 ≤C·

α

sup 0≤u≤t≤T

|Du Ft | · |πn | · T → 0,

(2.3.18)

as n → ∞ in probability, according to (i). Further, according to (i) and (ii), T s (s − u)α−1 Du Ftnk−1 1[tnk−1 ,tnk ) (s)du ds 0 0 T s α−1 − (s − u) Du Fs du ds → 0 (2.3.19) 0 0

in probability. Now, the proof follows from (v) and (2.3.17)–(2.3.19).

Now consider one sufficient condition for (v) (condition (v) seems to be the most artificial among other conditions (i)–(iv)). To this end, consider the middle part n of (2.3.11), from which we obtain nthat for any step processes Fn (t) = k=1 Fk,n 1[tnk−1 ,tnk−1 ) (t) and Gn (t) = k=1 Gk,n 1[tnk−1 ,tnk−1 ) (t) . E

n

Fn (t) ♦ dBtH

k=1

=E R

·

n

/ Gn (t) ♦ dBtH

k=1 H H M− Fn (t)M− Gn (t)dt + E

R2

H H M− Ds Fn (t)M− Dt Gn (s)ds dt.

(2.3.20)

156


The next result was motivated by (Ben03a, Theorem 2.2.8). Theorem 2.3.10. Let the stochastic process {Ft , Ft , t ∈ [0, T ]} satisfy the assumptions (i)–(iv) and T (vi) E 0 Ft2 dt < ∞; (vii) the operator Ft : [0, T ] → D1,2 is continuous in L2 ([0, T ] × P ). T n H exist, the integral 0 Fs ♦ dBsH Then the integral sums k=1 Ftnk−1 ♦ ∆Bk,n exists and T n H Fs ♦ dBsH = lim Ftnk−1 ♦ ∆Bk,n in L2 (P ) n→∞

0

k=1

for any sequence of increasing partitions πn with |πn | → 0 as n → ∞. n H Proof. Under condition (vi), the existence of sums k=1 Ftnk−1 ♦ ∆Bk,n and T H the integral 0 Fs ♦ dBs was established in Theorem 2.3.9. Further, using (2.3.20) and (vii), we obtain that 2 n T H H Ft ♦ dBt − Ftnk−1 ♦ ∆Bk,n E 0 k=1 ; H 1/2, or ∗ 1/H < 2) L2 [0, T ] ⊂ LH for nonrandom 2 [0, T ]. We obtain that the S -integral T functions from L2 [0, T ] coincides with the Wiener integral 0 Y (t)dBtH from Definition 1.6.1. Another approach to Skorohod integration w.r.t. fBm was developed in the papers (AN02), (Nua03), (Nua06). The main idea is to use the basic tools of a stochastic calculus of variations (Malliavin calculus) with respect to B H . Recall some of these notions for H ∈ (1/2, 1). (For H ∈ (0, 1/2) see, for example, (AMN00).) Let S be a family of smooth random variables of the form F = f (BtH1 , . . . , BtHn ) with f ∈ Cb∞ (Rn ) and ti ∈ [0, T ], 1 ≤ i ≤ n. Let H be a closure of the linear space of step functions defined on [0, T ] with respect to the scalar product t

s

1[0,t] , 1[0,s] H := 2αH

|r − u|2α−1 du dr. 0

0

Then the derivative operator D : S → Lp (Ω, H) for p ≥ 1 is defined as n ∂f DH F = (BtH1 , BtH2 , . . . , BtHn )1[0,ti ] . ∂x i i=1

Let Dk,p (H) be the Sobolev space, the closure of S with respect to the norm F pk,p = E(|F |p ) +

k j=1

E(Dν F pH⊗j ),

2.4 Skorohod, Forward, Backward and Symmetric Integration

161

where Dj is the jth iteration of D. The Skorohod integral (divergence operator) δH is defined as the adjoint of DH : D1,2 (H) ⊂ L2 (Ω) → L2 (Ω, H), defined by the means of the duality relationship E(GδH (u)) = EDH G, uH , u ∈ L2 (Ω, H), G ∈ S. Its domain is denoted by Dom(δH ). Introduce the Banach space |H| ⊗ |H| as the class of all the measurable functions ϕ : [0, T ]2 → R such that ϕ2|H|⊗|H|

:= (2αH)2 [0,T ]4

|ϕu,v ||ϕs,t ||s − u|2α−1 |t − v|2α−1 du dv ds dt < ∞,

and denote |H| := |RH | with the norm · |RH |,2 (see (1.6.7)). Denote also n S|H| the family of |H|-valued random variables of the form F = Fi h i , where Fi ∈ S and hi ∈ |H|. Put D F := k

n

i=1

D Fi ⊗ hi , and define the space k

i=1

Dk,p (|H|) as the completion of S|H| with respect to the norm F pk,p,|H| = E(F p|H| ) +

k

E(Di F pH⊗i ⊗|H| ).

i=1

Then D1,2 (|H|) ⊂ Dom(δH ). The basic property of the divergence operator is that for every u ∈ D1,2 (|H|) we have 2

E(|δ(u)|2 ) ≤ uD1,2 (|H|) . Consider the forward integral w.r.t. fBm ((AN02), (LT02)). It is defined as t t H us dBsH,− := P − lim ε−1 us (B(s+ε)∧t − BsH )ds. (2.4.1) 0

ε→0

0

(Note that in a similar way the symmetric Stratonovich integral can be de t t H H − B(s−ε)∧t )ds, and fined: 0 us dBsH,− := P − limε→0 (2ε)−1 0 us (B(s+ε)∧t also backward integral can be defined.) In (LT02) the ucp-limit is considered instead of the P -limit, where ucp-convergence is uniform convergence in probability on [0, T ]. Moreover, it is mentioned in (AN02) that forward, backward and symmetric integrals with integrand u and w.r.t. fBm coincide with each other under the following suppositions: u ∈ D1,2 (|H|) with t t |Ds ur ||r − s|2α−1 ds dr < ∞ a.s.). Also, it was proved that for processes 0 0 t t u ∈ D1,2 (|H|) with 0 0 |Ds ur ||r − s|2α−1 ds dr < ∞ a.s. we have the equality t t t us dBsH,s = δH (u) + 2αH |Ds ur ||r − s|2α−1 drds. (2.4.2) 0

0

0

162


Evidently, for u ∈ C β [0, T ] with β + H > 1 all the integrals, symmetric, forward, backward, and pathwise, coincide. We use this fact in order to establish the conditions of coincidence of Skorohod integrals introduced in (Ben03a) and in (AN02). Theorem 2.4.6. Fix a time interval [0, T ]. Let φ ∈ C 1 (R) and satisfy, together with its derivative φ , the growth condition |φ(x)| ≤ C exp(λxb ) for t some λ > 0 and 0 < b < 2. Then the integrals δH (φ(B H )) and 0 φ(BsH )δBsH coincide on [0, T ] a.s. Proof. According to Proposition 3.3 (Nua06), under the condition of the theorem (even under the less restrictive condition |φ(x)| ≤ C exp(λx2 ) for λ < (4T 2H )−1 ), the divergence operator δH (φ(B H )) exists on [0, T ] and satisfies the relation T T H H H φ(Bs )dBs − H φ (BsH )s2α ds a.s., δH (φ(B )) = 0

0

T

where 0 φ(BsH )dBsH is the pathwise integral. According to Theorem 2.3.7, under conditions (2.3.10), which evidently hold now, the same equal T ity is valid for the integral 0 φ(BsH ) ♦ dBsH . Therefore, δH (φ(B H )) and T φ(BsH ) ♦ dBsH coincide a.s. on [0, T ]. Further, the conditions of Theo0 H |X|)(t))2 dt can be rem 2.4.4 also hold now. Indeed, for example, E R ((M− T T H 2 bounded in our case by C 0 |φ(Bs )| ds. Therefore, 0 φ(BtH )δBtH exists T and equals 0 φ(BtH ) ♦ B˙ tH dt. Finally, we use Theorem 2.3.1 and Corollary 2.3.2 and obtain the proof.

Remark 2.4.7. A general S-transform approach to the stochastic fractional integration is presented in (Ben03b); see also (CC00) and (Cou07).

2.5 Isometric Approach to Stochastic Integration with Respect to fBm 2.5.1 The Basic Idea Some special approach to stochastic integration w.r.t. fBm was considered in (MV00). We will work with a continuous stochastic process {Xt , 0 ≤ t ≤ T } defined on a complete probability space (Ω, F, P ). Let Ft := FtX be the sigmafield generated by X on [0, t]. We assume that X0 = 0. Given a partition πn := {ti : 0 = t0 < t1 < · · · < tn = T } and X a stochastic process, define n. Assume first that the integrand ∆Xi by ∆Xi := Xti − Xti−1 for 1 ≤ i ≤ fi 1[ti−1 ,ti ) (t), where the random f is a simple predictable process: ft = i

variables fi are assumed to be Fti−1 measurable and ti ∈ πn ; denote the

2.5 Isometric Approach to Stochastic Integration with Respect to fBm

163

class of simple predictable processes by Ls . With such an f ∈ Ls and any (continuous) process X, define the stochastic integral of f with respect to X by fi ∆Xi . (f, X) := i

Assume now that |πn | → 0 as n → ∞. If the process X is the standard Brownian motion B, f := L2 (P ⊗ λ)- lim f n , where λ is the Lebesgue measure on [0, T ], one can define the integral (f, B) as the L2 -limit of the simple stochastic integrals (f (n) , B) using the classical Itô isometry E(f

(n)

T

2

(fs(n) )2 ds.

, B) = E

(2.5.1)

0

Assume now that the process X is any continuous stochastic process and f is a simple predictable process. Define now a semi-norm for (f, X) using (2.5.1). Note that such a semi-norm does not depend on the process X. It is the main feature of this approach. If the process X is the standard Brownian motion, then the semi-norm is a norm and the integrals of simple function converge to the classical stochastic integral defined by Itˆ o. For an arbitrary integrator X, even if the semi-norm is a norm, it may happen that the integrals of simple functions of processes have no limit. However, they have a limit in the completion of the space integral sums with respect to this norm. In this sense we generalize the Itô construction of stochastic integrals. In particular, we show that if X is a fractional Brownian motion B H , then we can define a norm by putting T 1/2 ! ! 2 !(f, B H )! := E f ds s G 0

in the space G of random variables of the form {g ∈ G : G = (f, B H ), f ∈ Ls }. Even more turns out to be true: for any k ≥ 2 define random variables (f, X (k) ) by the formula (f, X (k) ) := fi (∆Xi )k i

and define again a semi-norm for such random variables by putting ! ! T 1/2 ! (k) ! X ) := E fs2 ds . !(f, ! k G

0

Again, if the Brownian motion B H , then ! process X is a fractional ! pr H (k) !(f, (B ) )! k is a norm. Denote by L (P ⊗ λ) the space of predictable 2 G T process f with the property E 0 fs2 ds < ∞. Now, let f ∈ Lpr 2 (P ⊗ λ) be a predictable process and f (n) a sequence of simple predictable processes such that

164


! ! ! (n) ! !f − f !

L2 (P ⊗λ)

→0

as n → ∞. Define the higher-order generalized integral (f, (B H )(k) ) as a limit in the Banach space (J k , ·Gk ), which is the space of some kind of extended random variables g, which are limits of the sequences of the form (f, (B H )(k) ) with respect the norm ·Gk . 2.5.2 First- and Higher-order Integrals with Respect to X Wiener Integrals Further, if (Y, ·Y ) is a complete metric space, then the Y - lim stands for the limit on the space Y with respect m to the norm ·Y . Assume that f is a simple deterministic process, ft = i=1 fi 1[ti−1 ,ti ) (t). Then ·G is a norm if and only if (f, X) =

m

fi ∆Xi = 0 ⇐⇒ fi = 0, 1 ≤ i ≤ m.

(2.5.2)

i=1

Let X = (Xt )t∈[0,T ] be a square integrable process with EXt = 0, X0 = 0, and write R(t, s) for the covariance function, R(t, s) = E Xt Xs . Consider the quadratic forms Bm = E((f, X))2 where f ∈ Ls has deterministic coefficients fi , 1 ≤ i ≤ m. Then condition (2.5.2) is equivalent to the following: The quadratic form Bm is positive definite for each m ≥ 1.

(2.5.3)

We can write Bm in terms of the correlation function R: Bm =

m i=1

+2

[fi2 (R(ti , ti ) − 2R(ti−1 , ti ) + R(ti−1 , ti−1 ))]

fi fj [R(ti , tj ) − R(ti−1 , tj ) − R(ti , tj−1 ) + R(ti−1 , tj−1 )].

i=j,i,j≤m

(2.5.4) Put δii := R(ti , ti ) − 2R(ti−1 , ti ) + R(ti−1 , ti−1 ) and δij := R(ti , tj ) − R(ti−1 , tj ) − R(ti , tj−1 ) + R(ti−1 , tj−1 ). Then condition (2.5.3) is equivalent to the property that the matrix (δij )i,j≤m is positive definite for each m ≥ 1. Assume that condition (2.5.2) is valid for the process X and assume that f ∈ L2 [0, T ]. Then there exists f n ∈ Ls such that f n − f L2 [0,T ] → 0 as n → ∞. Moreover, the sequence (f n , X)


165

is a Cauchy sequence in the space (E s , ·E s ), where E s is the subspace of Ls consisting of deterministic simple functions f . Complete E s with respect the norm ·E s and denote this Banach space by E. Define now the integral T T fs dXs as the limit of (f n , X) in the space E. We say that 0 fs dXs is the 0 generalized Wiener integral with respect to process X. Note that Ls in dense in L2 [0, T ] and hence also E s is dense in E, by using the isometry. We clarify the connection between random variables and Wiener integrals defined above. Let ζ n be a sequence of random variables of the form ζ n := (f n , X) with some f n ∈ Ls . Assume now that ζ = P -limn ζ n and f − f n L2 [0,T ] → 0, n → ∞. We show later that it may happen that P {|ζ| < ∞} < 1 or even P {|ζ| < ∞} = 0. But even in the above situation the limit T fs dXs = E- lim(f n , X) n

0

defines the generalized Wiener integral. In this kind of situation we say that T the random variable ζ is one of the representatives of 0 fs dXs in the space T of random variables and 0 fs dXs is one of the representatives of the random T variable ζ in the space E: write this as ζ ↔ 0 fs dXs . It is easy to check that if X is a process with non-correlated increments and with the property E Xt2 > E Xs2

(2.5.5)

where s < t, then condition (2.5.2) is satisfied. Note first that condition (2.5.5) is equivalent to the condition E(Xt − Xs )2 > 0 for s < t. Since the process X has non-correlated increments, we have that E

m i=1

2 fi ∆Xi

=

m

fi2 E(∆Xi )2 = 0

i=1

if and only if fi = 0, i ≤ m. Note that if X is a square integrable martingale and EXt2 > EXs2 , s < t, then (2.5.2) is satisfied. Similarly, if X is a stationary process with so-called orthogonal vector measure ϕ(dλ) such that the spectral measure F (dλ) := E|ϕ(dλ)|2 is equivalent to the Lebesgue measure, then condition (2.5.2) is satisfied. If the process X is the standard Brownian motion B, then (f, B)E s = E(f, B)2 = f L2 [0,T ] and then the limits of simple integrals (f (n) , B) in the space E and in L2 (P ) are the same. Similarly, if the process X is a continuous square integrable t martingale M with the angle bracket M t = 0 as ds, where 1/K ≤ Eas ≤ K, the limits in the space E and L2 (P ) are the same.

166


First-order Stochastic Integrals with Respect to X Let F := {Ft , t ∈ [0, T ]} be a filtration on (Ω, F, P ) satisfying the usual conditions of right continuity and completeness. The notation X ∈ F means that Xt is Ft measurable. So, let X ∈ F be a process and introduce the space Gs of random variables ξ: ξ=

m

fi ∆Xi

i=1

where fi ∈ Fti−1 and fi ∈ L2 (P ), 1 ≤ i ≤ m, m ≥ 1. Let f be as above, i.e., f ∈ Ls and the coefficients fi , 1 ≤ i ≤ m satisfy fi ∈ Fti−1 and fi ∈ L2 (P ). Then we can define a surjection I from Ls → Gs by I(f ) := (f, X) =

m

fi ∆Xi .

i=1

Introduce the following semi-norm on Gs : m 1/2 (f, X)Gs := E fi2 (ti − ti−1 ) .

(2.5.6)

i=1

It is easy to check that the condition (f, X) = 0 P -a.s. if and only if fi = 0 P -a.s. for 1 ≤ i ≤ m

(2.5.7)

is a necessary and a sufficient condition for I to be a bijection and ·Gs to be a norm. Let X be a square integrable process, which satisfies (2.5.7). Now let f be a T predictable process with E 0 fs2 ds < ∞. Then there exist processes f n ∈ Ls such that T

(fs − fsn )2 ds → 0

E 0

as n → ∞. Now Ls is the space of elementary “predictable” processes g, m where gt := i=1 fi 1[ti−1 ,ti ) (t), and fi ∈ Fti−1 , 1 ≤ i ≤ m. Complete again T the space Gs with respect to the norm ·Gs . The integral 0 fs dXs =: I(f ) is defined using the extension of the isometry I on the completed Banach space G. The sequence f n is a Cauchy sequence with respect the norm ·G T and the integral 0 fs dXs is the limit of the elementary integrals (f n , X) in T the space (G, ·G ). We say that the integral 0 fs dXs defined for predictable pr f ∈ L2 (P ⊗ λ) is the first order generalized stochastic integral with respect T (1) to the process X. Later we will use the notation 0 fs dXs for this integral. n If ζ be a sequence of random variables of the form ζ n := (f n , X)


167

with some f n ∈ Ls and assume that ζ = P -limn ζ n and f − f n Lpr → 2 (P ⊗λ) 0, n → ∞. Hence also T fs dXs = G- lim(f n , X). n

0

It may happen that P {|ζ| < ∞} < 1 or even P {|ζ| < ∞} = 0. Again the T (1) random variable ζ is one of the representatives of the integral 0 fs dXs in T (1) the space of random variables and 0 fs dXs is one of the representatives of T (1) the random variable ζ in the space G: write this again as ζ ↔ 0 fs dXs . The first-order integral is linear: (af + bg, X) = a(f, X) + b(g, X). Higher-order Stochastic Integrals with Respect to X Let (X, F) be again a stochastic process defined on (Ω, F, P ). Introduce the space Gs,k of the random variables ξ: ξ :=

m

fi (∆Xi )k

i=1

where k > 1, fi ∈ Fti−1 , fi ∈ L2 (P ), 1 ≤ i ≤ m. If f ∈ Ls is a predictable step function, define a surjection I k from Ls to Gs,k by putting I k (f ) := (f, X (k) ) :=

m

fi (∆Xi )k .

i=1

We suppose that any simple function has different values on the adjoining segments of the partition. With this assumption only one partition corresponds to a simple function, we have only one zero function and I k is a surjection. Introduce the following semi-norm on Gs,k : ! ! ! ! !(f, X (k) )!

Gs,k

m 1/2 := E fi2 (ti − ti−1 ) = f L2 (P ⊗λ) . i=1

Let f and g be simple predictable processes, defined with respect to different partitions πf and πg . Consider f + g on the partition π := πf ∪ πg , put (f, X (k) ) + (g, X (k) ) := (f + g, X (k) ) and see that ! ! ! ! ! ! ! ! ! ! ! ! (2.5.8) !(f, X (k) ) + (g, X (k) )! s,k ≤ !(f, X (k) )! s,k + !(g, X (k) )! s,k . G

G

G

Again it is easy to check that the condition (f, X (k) ) = 0 P -a.s. if and only if fi = 0 for 1 ≤ i ≤ m, m when f ∈ Ls , f = fi 1[ti−1 ,ti ) (·) i=1

(2.5.9)

168


is a necessary and sufficient condition for I k to be a bijection, for Gs,k to be a linear space and for ·Gs,k to be a norm. n ∈ Ls such that If f is a predictable process from Lpr 2 (P ⊗ λ), take f f − f n L2 (P ⊗λ) → 0. Assume that property (2.5.9) holds for the process T (k) X with some k > 1. Define the integral 0 fs dXs := I k (f ) as the limit k

k

of (f n , X (k) ) in the completed Banach space (G , ·Gk ), where G is the completion of Gs,k with respect to norm ·Gs,k . We say that such an integral T (k) fs dXs is the kth order generalized stochastic integral of f with respect 0 to the process X. Assume now that property (2.5.9) holds for all k ≤ N . Define the Banach space GN by 1 2 N GN := G × G × · · · × G and define the norm in GN by ·GN :=

N

·Gk .

k=1

In view of (2.5.8), ·GN satisfies the triangle inequality and hence it is really a norm. N The elements g ∈ G have the form g=

N k=1

T

fk (s)dXs(k)

0

where fk is a predictable process from L2 (P ⊗ λ). Note also that there is N pr a bijection between such a g from G and (f1 , . . . , fN ) ∈ ⊗N k=1 L2 (P ⊗ λ) N equipped with the norm fk L2 (P ⊗λ) . k=1

The following examples clarify the definition of the generalized integrals of higher order. We assume that the process X satisfies property (2.5.9) for each 1 ≤ k ≤ N below. Processes with bounded variation. Assume that the process X is a continuous process with bounded variation and consider the random variables XTm , where N m XTm := (∆Xk )l . l=1 k=1 P

When |π| → 0 we have that XTm → XT and the right-hand side converges in N the space G towards the element N l=1

0

T

dXs(l) .


Here the random variable XT is a representative of the integral N T (l) zero is a representative of the sum dXs . 0

T 0

169 (1)

dXs

and

l=2

Standard Brownian Motion. Assume that X is a standard Brownian motion, X = B. Define again the random variable XTm by XTm :=

N m

(∆Bk )l .

l=1 k=1 P

Now, when |π| → 0, XTm → BT + T , so the constant T is a representative of N T T (2) (l) the integral 0 dBs and zero is a representative of the sum dBs . 0 l=3

2.5.3 Generalized Integrals with Respect to fBm Fractional Brownian Motion and Property (2.5.7) Theorem 2.5.1. Property (2.5.7) holds for fBm B H , H ∈ (0, 1). fi ∆BiH = 0 almost surely. Assume that m0 is the Proof. Assume that i≤m

largest index for which P {fm0 = 0} > 0. Then from presentations (1.8.17)– (1.8.18) we have H ∆Bm = 0

tm0

tm0 −1

mH (tm0 , s)dWs +

tm0 −1

0

(mH (tm0 , s) − mH (tm0 −1 , s))dWs

= Am0 + Bm0 , For the term Bm0 we have Bm0 ∈ Ftm0 −1 . Put Ωc := {ω : |fi | ≤ c, i ≤ m0 }. Then Ωc ∈ Ftm0 −1 and m0

1Ωc fi ∆BiH =

i=1

m

1Ωc fi ∆BiH = 0.

i=1

Hence we can conclude the following: 0=E

m0

1Ωc fi ∆BiH

2

i=1

=E

1Ωc fi ∆BiH

2

+ fm0 1Ωc Bm0 −1 + fm0 Am0

i≤m0 −1

The right-hand side of (2.5.10) is equal to

(2.5.10) .

170


E

(fi ∆BiH 1Ωc ) + fm0 1Ωc Bm0 −1

i≤m0 −1

+E

2 fm 1 0 Ωc

tm0

tm0 −1

2

(B H (tm0 , s))2 ds .

Hence, from (2.5.10), since

tm0

tm0 −1

(B H (tm0 , s))2 ds > 0

we have that fm0 1Ωc = 0 almost surely for any c > 0 and so fm0 = 0 P -a.s.

This shows that condition (2.5.7) is fulfilled. Hence fi = 0 for all i ≤ m. Fractional Brownian Motions and Property (2.5.9) Theorem 2.5.2. Property (2.5.9) holds for fBm B H , H ∈ (0, 1). Proof. We know from Theorem 2.5.1 that the claim holds for k = 1. Assume now that k > 1 and let m0 , Am0 , Bm0 and W be as in the proof of Theorem 2.5.1. Put fic := 1Ωc fi . Note that fic ∈ Ftm0 −1 for i ≤ m0 . Denote by χ the random variable m 0 −1 χ := fic (∆BiH )k . i=1

For the random variable χ we have that χ ∈ Ftm0 −1 , and this fact is used below. Assume that fi (∆BiH )k = 0. With the above notation we have i≤m

from this assumption that also χ+

c fm 0

k k r=0

r

(Bm0 )k−r (Am0 )r = 0.

Write the expression in (2.5.11) as ⎛ ⎞ k c ⎝χ + fm (Bm0 )k−r (Am0 )r ⎠ 0 r 0≤r≤k, r even ⎛ ⎞ k c + ⎝fm (Bm0 )k−r (Am0 )r ⎠ =: χ1 + χ2 . 0 r

(2.5.11)

(2.5.12)

0≤r≤k, r odd

The random variable Am0 is a Gaussian random variable with zero expectation and hence for odd r E(Am0 )r = 0 and by conditioning on Ftm0 −1 in (2.5.12) it is easy to see that E(χ1 χ2 ) = 0. So from this we can conclude that Eχ22 = 0, using also (2.5.11) and (2.5.12). But


171

2 χ22 = fm (γ1 + γ2 ) 0

with

γ1 :=

0≤r≤k, r odd

and γ2 :=

r=q, r,q odd

2 k k−r r (Am0 ) (Bm0 ) r

k k (Bm0 )2k−r−q (Am0 )r+q . r q

(2.5.13)

(2.5.14)

All the terms in (2.5.13) are nonnegative and since r + q is even, the same holds for the expression (2.5.14), too. Note also that if r = 1, then k 2 (Bm0 )2k−2 (Am0 )2 > 0 2 (γ1 + γ2 )) = 0. Hence fm0 = 0 almost surely. But at the same time E(fm 0 almost surely. From this follows that fi = 0 almost surely for all i ≤ m. We

have shown that fBm B H satisfies property (2.5.9) for all k ≥ 1.

Some Properties of the Generalized Integrals In this subsection we discuss some of the properties of the generalized integrals. At this stage we have results mostly on Wiener integrals. Assume that B H is again an fBm with index H. Take fsn := nγ 1(T /2−1/2n,T /2+1/2n] (s). 2

Then f n L2 [0,T ] = n2γ−1 . If H ∈ (1/2, 1), 1/2 < γ < H, then f n L2 [0,T ] → ∞ and the generalized integral does not exist, but E((f n , B H ))2 = n2γ−2H → 0, and the limit exists in L2 (P ). If H < γ < 1/2, then E((f n , B H ))2 → ∞, but f n L2 [0,T ] → 0. Hence the integral exists in G and it is = 0, but the limit P

does not exist in L2 (P ). Note also that here we have that |(f n , B H )| → ∞. L2 -integrals and Wiener integrals, H ∈ (1/2, 1). If B H is an fBm with Hurst index H ∈ (1/2, 1), then according to (1.9.2) we have the following estimate for L2 -integral, valid for any p > 0: p T p E fs dBsH ≤ cH,p f L 1 [0,T ] . (2.5.15) 0 H Hence, if (f (n) , B H ) converges in G, it also converges in L2 (P ). L2 -integrals and Wiener integrals, H ∈ (0, 1/2). Before the continuation, we prove the following theorem, which is the opposite to (2.5.15). Theorem 2.5.3. Let f ∈ Ls and B H is an fBm with Hurst index H ∈ (0, 1/2). Then 2 T 2 H E fs dBs ≥ C f L2 [0,T ] . (2.5.16) 0

172


Proof. If f ∈ Ls and (f, B H ) =

i

E(f, B H )2 =

fi ∆BiH , then

(fi2 E∆BiH )2 +

i

fi fk E(∆BiH ∆BkH ).

(2.5.17)

i=k

But E(∆BiH ∆BkH ) < 0 and hence fi fk E(∆BiH ∆BkH ) ≥ |fi ||fk |E(∆BiH ∆BkH ). Use this in (2.5.17) to obtain the inequality

2 H 2 H E(f, B ) ≥ E |fi |∆Bi . i

Hence we can assume that fi ≥ 0 for all i ≤ n in proving (2.5.16). Denote by D(R) the space of functions f with the two properties: f ∈ C ∞ (R) and f has compact support. Let φ ∈ D(R). Then the Fourier transform φ of φ belongs to S(R) ⊂ FH ⊂ (R) (see Lemma 1.6.8), and moreover, LH 2 2 2 H H −2α E φt dBt = E φ (t)Bt dt = cH |φ(λ)||λ| dλ, (2.5.18) R

R

R

where cH is some constant. We want to prove that there exists a sequence (φn )n≥1 , φn ∈ D(R) such that L2 (P ) (φn ) (t)BtH dt → (f, B H ). (2.5.19) R

To prove (2.5.19) it is sufficient to prove it for f ∈ Ls , fu = a1[s,t) (u), s < t ≤ T and a > 0. Take φn ∈ D(R) such that supp(φn ) ⊂ [s − 1/n, t + 1/n] and φn = a on [s + 1/n, t − 1/n]. Then

n

(φ ) R

(u)BuH du

t+1/n

=

n

(φ ) t−1/n

(u)BuH du

s+1/n

+

(φn ) (u)BuH du

s−1/n

and, for example, t+1/n t+1/n H H (φn ) (u)BuH du ≤ (φn ) (u)(Bt+1/n − BuH )du aBt+1/n − t−1/n t−1/n ≤a

sup u∈[t−1/n,t+1/n]

H |Bt+1/n − BuH |.

From self-similarity of B H and Remark 1.10.7 with f = 1, T = 2/n sup u∈[t−1/n,t+1/n]

L2 (P )

H |Bt+1/n − BuH | −−−−→ 0


and so E(f, B H )2 = lim n

R

173

|φn (λ)||λ|−2α dλ.

Since for any λ ∈ R f(λ) = limn→∞ φn (λ), we have, using the Fatou lemma and relation (2.5.18), 2 ∞ ∞ 2 −2α n 2 −2α H |f(λ)| |λ| dλ ≤ lim inf |φ (λ)| |λ| dλ = E fi ∆Bi . n→∞ −∞ −∞ i

We have that ∞

|f(λ)|2 |λ|−2α dλ −2α 2 |f (λ)| dλ + ≥ε −∞

|λ|>ε

|λ|≤ε

|f(λ)|2 |λ|−2α dλ.

(2.5.20)

Put ρ(λ) := |λ|−α 1[−ε,ε] (λ). Since H ∈ (0, 1/2), we have that ρ ∈ L1 (R). Also, ∞ ε itλ ρ(t) := e ρ(λ)dλ = cos(tλ)|λ|−α dλ. −∞

−ε

This integral is finite and hence ρ(·) is the Fourier transform of ρ(·). Use the Parceval identity to obtain |f(λ)|2 |λ|−2α dλ |λ| 0 we can choose πn in such a way that kn b n n (2.6.3) f (t)dg(t) − f (tk )∆g(tk ) < δ. a k=1

176


Further, according to (FdP01, Corollary 20), kn b ε n n fm (t)gm (t)dt − fm (tk )∆gm (tk ) ≤ C |πn | · fm C α [a,b] · gm C β [a,b] , a k=1

(2.6.4) where 0 < α < α, 0 < β < β, and α + β = 1 + ε. If fn − f C α [a,b] → 0, m → ∞, then fm − f C α [a,b] → 0, m → ∞ for 0 < α < α, and fm C α [a,b] ≤ C1 , where C1 does not depend on m ≥ 1. Similarly, gm C β [a,b] ≤ C2 . From these bounds and from (2.6.4) we obtain that kn b ε n n fm (t)gm (t)dt − fm (tk )∆gm (tk ) ≤ C3 |πn | . a

(2.6.5)

k=1

ε

Choose such n that (2.6.3) holds and also C3 |πn | < δ; then for such fixed n we can choose such m that k kn n (2.6.6) f (tnk )∆g(tnk ) − fm (tnk )∆gm (tnk ) < δ. k=1

k=1

It is possible since supt∈[a,b] |gm (t) − g(t)| ≤ gm − gC β [a,b] → 0, and the same is true for fm . The proof of the first statement follows now from (2.6.3)–(2.6.6). The third statement follows from 1) and (FdP01, Lemma 19), which states (1) that the bound (2.6.2) holds for any f ∈ C0 [a, b] (it means that f ∈ C (1) [a, b] (1) and f (a) = 0) and g ∈ C [a, b]. The second statement follows from 1) and (FdP01, Theorem 22). Indeed, according to 3) b f (t) − f (0) dg(t) ≤ C f C α [a,b] · gC β [a,b] · (b − a)1+ε , a whence b f (t)dg(t) ≤ C f C α [a,b] · gC β [a,b] · ((b − a)1+ε ∨ (b − a)β ). a

β [a, b] with β > 1 − H. In this Further we consider H ∈ ( 12 , 1). Let f ∈ Cpw N bi H case the sum i=1 ai f (t)dBt exists. The next result means that this sum can be represented as a unique integral.

Lemma 2.6.4. Let f be piecewise H¨ older of order β > 1 − H on the interval [a, b]. Then there exists the Riemann–Stieltjes integral

2.6 Stochastic Fubini Theorem for Stochastic Integrals w.r.t. fBm

b

f (u)dBuH a

=

N

bi

177

f (u)dBuH

ai

i=1

and for an arbitrary sequence πn of partitions of [a, b] it can be represented as a limit b kn f (u)dBuH = lim f (unk )∆BuHnk . |πn |→0

a

N

k=1

= [a, b), [ai , bi ) are disjoint and f ∈ C α [ai , bi )). Proof. Put πni := [ai , bi ) ∩ πn . Evidently, πni ≤ |πn |. It follows from boundedness of f and continuity of B H that (We suppose that

i=1 [ai , bi )

i j:un j ∈πn

f (unj )∆BuHnj

→

bi

f (u)dBuH ,

ai

even in the case when πni does not contain ai or(and) bi . N Therefore, k:un ∈πn f (unk )∆BuHnk = i=1 k:un ∈πi f (unk )∆BuHnk n k k b N b → i=1 aii f (u)dBuH = a f (u)dBuH , as |πn | → 0.

Let 0 < T1 < T2 , Φ = Φ(t, u, ω) : PT := [T1 , T2 ]2 × Ω → R be the random function measurable in all the variables. Theorem 2.6.5. Let there exist the set Ω ⊂ Ω such that P (Ω ) = 1 and let for any ω ∈ Ω the function Φ(s, u, ω) satisfy the conditions: older of order β > 1 − H in u ∈ 1) ∀s ∈ (T1 , T2 ) Φ(t, ·, ω) is piecewise H¨ [T1 , T2 ], and there exists C = C(ω) > 0 such that Φ(t, ·, ω)Cpw β [T1 ,T2 ] ≤ C; T2 H 2) the function T1 Φ(t, u, ω)dBu is Riemann integrable in the interval [T1 , T2 ]. Then there exist the repeated integrals

T2 T2

I1 := T1

Φ(t, u, ω)dBuH

T1

T2 T2

dt and I2 := T1

Φ(t, u, ω)dt dBuH ,

T1

and I1 = I2 P -a.s. Proof. We fix ω ∈ Ω and omit ω throughout the proof. The integral T2 Φ(t, u)dBuH exists according to Lemma 2.6.4 and condition 1); the reT1 peated integral I1 exists according to condition 2). Since Φ(t, ·) is piecewise T H¨ older, then from the evident bound T12 |Φ(t, u1 ) − Φ(t, u2 )| ds ≤ C(T2 − T α T1 ) |u1 − u2 | we obtain that T12 Φ(t, u)ds is piecewise Hölder of order α in u ∈ [T1 , T2 ]. Further, since B H is Hölder up to order H > 12 and α+H > 1, the integral I2 also exists. The integral I1 can be presented as a limit of integral sums,

178


I1 = lim

|πn |→0

k n −1 T2 k=0

Φ(tnk , u)dBuH ∆tnk .

(2.6.7)

T1

For any point tnk ∈ πn , according to condition 1), there exists a finite number of points {u1,k < u2,k < · · · < ul(k),k } such that Φ(·, u) is Hölder between them. Denote {T1 = u0 < u1 < u2 < · · · < uL(n) = T2 } :=

kn 4

{u1,k < u2,k < · · · < ul(k),k } ∪ {T1 , T2 }.

k=1

For any interval [ui , ui+1 ] we consider the sequence of partitions πi,r , r ≥ 1 of the form (0)

(1)

(m )

πi,r := {ui = ui,r < ui,r < · · · < ui,r r = ui+1 }, |πi,r | → 0, r → ∞. L(n)−1 (0) (N ) Then π ˜r := i=0 πi,r ∪ {T1 , T2 } := {T1 = ur < · · · < ur r = T2 } πr | = is a partition of interval [T1 , T2 ] w.r.t. argument u, its diameter |˜ πr | → 0, r → ∞. max1≤i≤L(n)−1 |π|i,r , and |˜ Estimate the difference |I1 − I2 |: k n −1 N r −1 Φ(tnk , ur(j) )∆BuH(j) ∆tnk |I1 − I2 | ≤ I1 − r k=0 j=0 N r −1 k n −1 n,r + I2 − Φ(tnk , ur(j) )∆tnk ∆BuH(j) =: ∆n,r 1 + ∆2 . (2.6.8) r j=0 k=0 Further, ∆n,r 1

k n −1 T2 n H n ≤ I1 − Φ(tk , u)dBu · ∆tk k=0 T1 k N n −1 T2 r −1 n H n (j) H + Φ(tk , u)dBu − Φ(tk , ur )∆Bu(j) ∆tnk . r j=0 k=0 T1

Since Φ is piecewise Hölder, then, according to Lemma 2.6.4, T2 N r −1 n H n (j) H Φ(t , u)dB − Φ(t , u )∆B (j) → 0, r → ∞. k u k r ur T1 j=0 kn −1 T2 n H n According to (2.6.7), I1 − k=0 Φ(t , u)dB · ∆t u k k → 0, n → ∞. T1 Therefore,


lim lim ∆n,r = 0. 1

179

(2.6.9)

n→∞ r→∞

Further, ∆n,r 2

N r −1 T2 (j) H ≤ I2 − Φ(t, ur )dt · ∆Bu(j) r T1 j=0 (2.6.10) Nr −1 kn −1 tn k+1 (j) n (j) H + Φ(t, ur ) − Φ(tk , ur ) dt · ∆Bu(j) . r j=0 k=0 tnk

The second term can be expanded as k n r −1 n −1 tk+1 N (j) n (j) H (2.6.11) Φ(t, ur ) − Φ(tk , ur ) ∆Bu(j) dt r n k=0 tk j=0 k −1 L(N )−1 n n tk+1 (j) n (j) H = Φ(t, ur ) − Φ(tk , ur ) ∆Bu(j) dt . r n tk (j) k=0 i=0 ur ∈πi,r Since the function Φ(s, u) − Φ(tnk , u) is Hölder on any interval [ui , ui+1 ), we have that Φ(t, ur(j) ) − Φ(tnk , ur(j) ) ∆BuH(j) lim |πi,r |→0

r

(j)

ur ∈πi,r

ui+1

=

Φ(t, u) − Φ(tnk , u) dBuH . (2.6.12)

ui

(j) Moreover, ∀ 0 ≤ i ≤ L(n) − 1 the sequence fir (t, tnk ) := ur(j) ∈πi,r Φ(t, ur ) (j) − Φ(tnk , ur ) ∆B H(j) has the integrable dominant. Indeed, we can use the ur bounds from (FdP01, Corollary 20), Lemma 2.6.3, and the boundedness of H¨ older norms, and obtain that (j) ur+1 r r n n n H |fi (t, tk | ≤ fi (t, tk ) − Φ(t, u) − Φ(tk , u) dBu (j) ur (j) ur+1 + Φ(t, u) − Φ(tnk , u) dBuH ur(j) ! ! ε ≤ C |πi,r | · Φ(t, ·) − Φ(tnk , ·) (j) (j) β · !B H ! (j) (j) H C[ur ,ur+1 ]

C[ur ,ur+1 ]

180


(j) ur+1 Φ(t, u) − Φ(tnk , u) dBuH + ur(j) (j) ur+1 n H ≤C + Φ(t, u) − Φ(tk , u) dBu , ur(j)

(2.6.13)

where β < β, H < H and β + H > 1. Using the second statement of Lemma 2.6.3 and condition 1) of this theorem, we obtain the bound (j) ur+1 n H (j) Φ(t, u) − Φ(tk , u) dBu ur ! ! ≤ C Φ(t, ·) − Φ(tnk , ·)C α [T1 ,T2 ] · !B H !C H [T ,T ] ≤ C. (2.6.14) 1

pw

2

Estimates (2.6.13) and (2.6.14) mean that we can use the Lebesgue dominant convergence theorem and obtain that tnk+1 ui+1 tnk+1 lim Φ(t, u) − Φ(tnk , u) dBuH dt, fir (t, tnk )dt = r→∞

tn k

where the integrand in t. Therefore, lim

ui

tn k

i=0

k=0

= T2

= T1

T1

Φ(t, ur(j) ) − Φ(tnk , ur(j) ) ∆BuH(j) dt r

(j)

ur ∈ πi,r

n k n −1 tk+1

k=0 T2

tn k

ui

Φ(t, u)−Φ(tnk , u) dBuH is measurable and bounded

k n −1 L(n)−1 tnk+1

r→∞

ui+1

tn k

Φ(t, u)dBuH

T2

Φ(t, u) − Φ(tnk , u) dBuH dt

T1

dt −

k n −1 T2 k=0

Φ(tnk , u)dBuH ∆tnk .

(2.6.15)

T1

T According to condition 2) of this theorem, the integral T12 Φ(t, u)dBuH is Riemann integrable in t, therefore T2 T2 k n −1 T2 (2.6.16) lim Φ(tnk , u)dBuH ∆tnk = Φ(t, u)dBuH dt. n→∞

k=0

T1

T1

T1

From Lemma 2.6.4, L(n)−1 T2 (r) H I2 − Φ(t, uj )dt · ∆Bu(r) → 0, as n → ∞. j T1 r=0 Now the proof follows from (2.6.8)–(2.6.17).

(2.6.17)


181

t

Let I(t) = 0 f (s)dBsH for some stochastic process trajectories f with 1 β from C [0, T ] with β + H > 1. Consider the integral H ∈ ( 2 , 1) J1 (t) = t l (t, s)I(s)ds that will appear in connection with the Girsanov theorem 0 H and stochastic differential equations in subsections 2.8.2 and 3.2.3, and also, t t H let J2 (t) = 0 f (u) u lH (t, s)ds dBu . Lemma 2.6.6. Both the integrals, J1 and J2 , exist and J1 = J2 P -a.s. Proof. It follows from (FdP01) that the trajectories of I(t), t ∈ [0, T ] are H¨ older of order H − ε for any 0 < ε < H, whence the existence of J1 (t) follows. Further, elementary calculations + * u2 u2 u2 1 (t − s)−α s−α ds ≤ (t − s)−2α ds + s−2α ds ≤ (u2 − u1 )1−2α 2 u1 u1 u1 t demonstrate that the function f (u)· u lH (t, s)ds is Hölder up to order β ∧(1− 2α) > 1 − H, and J2 (t) exists. We can present these integrals in the following way: t t t t H Φ(s, u)dBu ds, J2 = Φ(s, u)ds dBuH , J1 = 0

0

0

0

where Φ(s, u) = lH (t, s)f (u)1{0≤u≤s} . The function Φ will satisfy both the conditions of Theorem 2.6.5, if we put T1 = δ and T2 = t − δ for any 0 < δ < 2t . In particular, Φ(s, ·) is piecewise H¨ older of order β on [δ, t − δ] with one point u = s of H¨ older discontinuity for any s ∈ [δ, t − δ]. Therefore, the following equality holds a.s.: s t−δ t−δ t−δ lH (t, s) f (u)dBuH ds = f (u) lH (t, s)dsdBuH . δ

δ

δ

u

The last equality can be rewritten as J1 − R1 = J2 − R2 , where

(2.6.18)

t−δ s δ lH (t, s) f (u)dBuH ds + lH (t, s) f (u)dBuH ds 0 0 δ 0 t s + lH (t, s) f (u)dBuH ds =: R11 + R12 + R13 ;

R1 =

δ

0

t−δ

R1 =

δ

t H f (u) lH (t, s)ds dBu +

0

u

t

+ t−δ

δ

t−δ

f (u)

t

t−δ

lH (t, s)ds dBuH

t f (u) lH (t, s)ds dBuH =: R21 + R22 + R23 . u

182


s According to (FdP01, Theorem 22), there exists C > 0 such that f (u)dBuH ≤ CsH−ε for any fixed 0 < ε < 1 . Therefore, 2 0 δ 1 1 |R11 | ≤ C s 2 −ε (t − s)−α ds ≤ Ct1−α (1 − α)−1 δ 2 −ε → 0 as δ → 0. 0

Similarly, 1 |R12 | ≤ C1 δ H−ε · δ −α · δ 1−α → 0 and |R13 | ≤ C2 t 2 −ε δ 1−α → 0 as δ → 0, where C1 and C2 are some constants, possibly depending on ω. t As mentioned above, the process f (u) · u lH (t, s)ds is Hölder of order β∧(1−2α) > 1−H. Therefore, by using again (FdP01, Theorem 22), we obtain the bounds |R21 | ≤ Cδ H−ε , |R22 | ≤ C1 (t − 2δ)H−ε , and |R23 | ≤ Cδ H−ε with some constants C, C1 , depending on ω. Taking in (2.6.18) a limit as δ → 0,

we obtain from all these estimates that J1 = J2 a.s.

2.7 The Itˆ o Formula for Fractional Brownian Motion 2.7.1 The Simplest Version First, we present a very elegant proof of the Itˆ o formula involving fBm from (Shi01). Lemma 2.7.1. Let B H be an fBm with H ∈ (1/2, 1), F ∈ C 2 (R). Then for any t > 0 t F (BuH )dBuH . F (BtH ) = F (0) + 0

Proof. The Taylor formula with the reminder term in the integral form gives us x F (u)(x − u)du. F (x) = F (y) + F (y)(x − y) + y

Let the sequence of partitions πn = {0 = < tn1 < · · · < tnkn = t}, |πn | → 0, k n ; < F (tnk ) − F (tnk−1 ) n → ∞. Then F (BtH ) − F (0) = tn0

=

kn

k=1

F

(BtHnk−1 )(BtHnk − BtHnk−1 ) + Rtn ,

where Rtn =

kn B Hn t k BtHn

F (u)(BtHnk − u)du.

k=1 k=1 k−1 Further, sup F (BuH ) < ∞ a.s. and for H ∈ (1/2, 1), and 0ut

P -lim n→∞

Therefore |Rtn |

1 2

kn 2 H Btnk − BtHnk−1 = 0. k=1

2 kn H P sup F (BuH ) Btnk − BtHnk−1 −→ 0. Even if we do

0ut

k=1

not know that the limit of integral sums

kn k=1

F (BtHnk−1 )(BtHnk − BtHnk−1 ) exists

2.7 The Itˆ o Formula for Fractional Brownian Motion

183

(but we know it from Theorem 2.1.7), we can obtain this existence now and, moreover, t F (BuH )dBuH . F (BtH ) − F (0) = 0

2.7.2 Itˆ o Formula for Linear Combination of Fractional Brownian Motions with Hi ∈ [1/2, 1) in Terms of Pathwise Integrals and Itˆ o Integral , Denote C β− [a, b] = 0 1. Further calculations are obvious: we use the Taylor formula and pass to the limit, as usual, taking into account that kn Hi P Hj Hj Hi for any 1 i m and 2 j m k=1 Btnk − Btnk−1 Btnk − Btnk−1 −→ 0 as n → ∞.

Now, consider the process Yt =

m

σi BtHi , where Hi ∈ (1/2, 1) for any

i=1

1 i m. We can forecast that in this case the class C 1 (R) of functions can be used. m σi BtHi , where Hi ∈ (1/2, 1) for any 1 i m. Theorem 2.7.3. Let Yt = i=1

Let F ∈ C 1 (R), and F ∈ C β [0, t] with (β + 1) min Hi > 1 for any t > 0. Then for any t > 0 t m σi F (Ys )dBsHi . (2.7.1) F (Yt ) − F (0) = i=1

0

Proof. Clearly, condition (β + 1) min Hi > 1 ensures the existence of t F (Ys )dBsHi as the limit of Riemann sums for any i > 1. Consider convo0 lutions Fn = F ∗ ϕn with ϕn from Lemma 2.1.8. Then Fn ∈ C ∞ (R), formula (2.7.1) holds for any Fn and for any 1 − min Hi < γ < β · min Hi we have that

184


γ γ D0+ Fn → D0+ F in L1 [a, b] as n → ∞ for any a, b ∈ R, which can be proved similarly to (2.1.10). Therefore, t (F (Ys ) − Fn (Ys ))dBsHi 0

1−γ Hi sup Dt− Bt− (s) 0st

sup |Ys |

0st

γ D Fn (s) − Dγ F (s) ds → 0, 0+ 0+

− sup |Ys | 0st

whence the proof follows.

Remark 2.7.4. Theorems 2.7.2 and 2.7.3 can be extended to the functions F of several variables, o formula has the following t depending also on t. The Itˆ form: let Yti = 0 fi (s)dBsHi , where H1 = 1/2, Hi ∈ (1/2, 1), 2 i m − 1, t t Ytm = 0 g(s)ds, 0 f12 (s)ds < ∞ a.s., fi ∈ C βi [0, t] a.s. for βi + Hi > 1, t |g(s)| ds < ∞ a.s., F = F (t, x) : R+ × Rn → R, F ∈ C 1 (R+ ) × C 2 (R) 0 2 t ∂F t (Z )f (s) ds < ∞, 0 ∂F × C 1 (Rn−1 ), the integrals 0 ∂x s 1 ∂t (Zs ) ds < ∞, 1 t ∂2F t ∂F ∂F (Zs ) f12 (s)ds < ∞, and 0 ∂x (Zs ) |g(s)| ds < ∞ a.s, ∂x (Zs )fi 0 ∂x2 1 i 1

∈ C γ [0, t] a.s. for γ + Hi > 1 and any t > 0, where Zs = (s, Ys1 , . . . , Ysm ). Then

m−1 t ∂F ∂F (Zs )ds + = F (0) + (Zs )fi (s)dBsHi ∂x i 0 ∂t 0 i=1 t ∂F 1 t ∂2F + (Zs )g(s)ds + (Zs )f12 (s)ds. (2.7.2) 2 0 ∂x21 0 ∂xm t t In particular, for the process Yt = 0 a(s)dBsH + 0 b(s)ds we have that t

F (t, Yt1 , . . . , Ytm )

F (t, Yt ) = F (0, Y0 ) + 0

t

t Ft (s, Ys )ds + Fx (s, Ys )b(s)ds 0 t Fx (s, Ys )a(s)dBsH , H ∈ (1/2, 1). +

(2.7.3)

0

2.7.3 The Itˆ o Formula in Terms of Wick Integrals The next result is a direct consequence of Theorems 2.3.8 and 2.7.3. Theorem 2.7.5. Let the function F = F (t, x) : R+ × R → R be continuously differentiable in t and twice continuously differentiable in x. Let 2+ε ) < ∞, t > 0 for some ε > 0, Yt be as* in Theorem 2.7.2, E ∂F t ∂x (t, Y 2 + ∂F 2 ∂ 2 F E sup + ∂x2 (s, Ys ) < ∞, t > 0. Then ∂x (s, Ys ) 0st


t

F (t, Yt ) − F (0, 0) = 0 m

+

∂F (s, Ys )ds + ∂t

σi σk C˜Hi ,Hk (Hi + Hk )

i,k=1

t 0

t

0

185

∂F (s, Ys ) ♦ dYs ∂x

∂2F (s, Ys )sHi +Hk −1 ds. ∂x2

(2.7.4)

2.7.4 The Itˆ o Formula for H ∈ (0, 1/2) We use the integral representation of fBm via the underlying Wiener process B on the finite interval [0, t] : t H Bt = mH (t, s)dBs 0 s t t (6) α (6) −α α α−1 −α −α = CH t u (t − u) dBu − CH α s u (s − u )dBu ds. 0

0

0

Let the function F ∈ C (R) and we want to expand F (BtH ). Note that z (6) H H BtH = Bt,t , where for 0 < z < t Bz,t = CH z α 0 u−α (t − u)α dBu z s (6) − CH α 0 sα−1 0 u−α (s − u)−α dBu ds. Therefore 3

F (BtH )

1 (6) 2 t H = F (0) + F + (CH ) F (Bz,t )(t − z)2α dz 2 0 0 t z (6) H F (Bz,t )z α−1 u−α (t − u)α dBu dz = F (0) + αCH 0 0 t (6) H F (Bz,t )(t − z)α dBz + CH 0 t z (6) H − αCH F (Bz,t )z α−1 u−α (t − u−α )dBu dz 0 0 t 1 (6) H + (CH )2 F (Bz,t )(t − z)2α dz. (2.7.5) 2 0 t

H H (Bz,t )dz Bz,t

Further, (6)

H = BzH + αCH z α Bz,t

z

0

= BzH

u−α

t

(v − u)α−1 dv dBu t z (6) + αCH z α u−α (v − u)α−1 dBu dv, (2.7.6) z

z

0

whence t r z (6) H ) = F (BzH ) + F BzH + αCH z α u−α (v − u)α−1 dBu dv F (Bz,t z z 0 z (6) α −α α−1 × αCH z u (r − u) dBu dr =: F (BzH ) + φ(F , z, t), (2.7.7) 0

186


H and similar relation holds for F (Bz,t ). But r z 1 z −α u−α (v − u)α−1 dBu dv = u [(r − u)α − (z − u)α ] dBu . (2.7.8) α z 0 0

Substituting (2.7.6)–(2.7.8) into (2.7.5), we obtain the following result. Theorem 2.7.6. Let H ∈ (0, 1/2), B H be an fBm with Hurst index H, t (6) z represented as BtH = 0 mH (t, s)dBs . Denote Yr,z := CH 0 u−α (r −u)α dBu , 0 z r, Yz := Yz,z . Then

t

t

= F (0) + F + F (BzH )(t − z)α dBz 0 0 t 1 (6) 2 t H H α−1 −α F (Bz )z Yt,z dz + (CH ) F (Bz )(t − z)2α dz + Rt , 2 0 0

F (BtH )

(BzH )αz α−1 Yt,z dz

(6) CH

where

t (6) φ(F , z, t)αz α−1 Yt,z dz + CH φ(F , z, t)(t − z)α dBz 0 0 t t 1 (6) −α φ(F , z, t)z α−1 Yt,z dz + (CH )2 φ(F , z, t)(t − z)2α dz. 2 0 0 t

Rt = α

Remark 2.7.7. The different approaches to the Itˆ o formula for fBm with H ∈ (1/2, 1) are contained in the papers (Lin95), (DH96), (DU99), (AN02), (DHP00), (BO04), (CCM03), (FdP01). An elegant version of the Itˆ o formula for F (BtH ) for any H ∈ (0, 1) was obtained by C. Bender in (Ben03a) and (Ben03c), but in terms of distributions. If the distribution F is of function type, continuous at 0 and of polynomial growth, the form of such an Itˆ o formula coincides with (2.7.4) for m = 1. For the other forms of the Itˆ o formula for fBm with H ∈ (0, 1/2) see also (Nua03), (GRV03), (ALN01), (AMN00), (CN05). 2.7.5 Itˆ o Formula for Fractional Brownian Fields First, we prove one auxiliary result for H¨ older two-parameter functions. Let the function F : R → R, F ∈ C 3 (R), F is the Lipschitz function, f (t) := F (g(t)), g ∈ C µ1 µ2 (R2+ ) with µi > 1/2, i = 1, 2. (2.7.9)

i,n , < · · · t = t Let the rectangle Pt = [0, t] ⊂ R2+ be fixed, πni := 0 = ti,n n i 0 2 where ti,n k =

kti 2n ,

1 kt2 fik = f ( it 2n , 2n ),

∆1ik f = fi+1k − fik , ∆2ik f = fik+1 − fik , ∆ik f = ∆1ik+1 f − ∆1ik f.


187

Lemma 2.7.8. Under assumption (2.7.9) lim Ijn = 0, 1 j 7, where I1n I4n I7n

= = =

2n −1

∆1ik f ∆ik g,

i,k=0 2n −1

i,k=0 2n −1

I2n

=

i,k=0

fik ∆ik g∆2ik g,

F

i,k=0

2n −1

I5n

n→∞

∆2ik f ∆ik g,

=

2n −1 i,k=0

I3n =

2n −1 i,k=0

∆1ik f (∆2ik g)2 ,

fik ∆ik g∆1ik g,

I6n =

2n −1

(∆1ik f )2 ∆2ik g,

i,k=0

(gi,k )∆1ik g(∆2ik g)2 .

Proof. Consider I1n (I2n is similar). We can rewrite I1n = Pt fñ dg, where 1 1 1 2 × ktn2 , (k+1)t . Further, fñ = ∆1 f for s ∈ ∆n := itn1 , (i+1)t n n ik

ik

2

2

2

2

Pt

fñ dg =

Pt

α1 α2 ˜ 1−α1 1−α2 (D0+ g1− )(s)ds, fn )(s)(D1−

1−α 1−α 1 2 where 1−µ1 < αi < µi , i = 1, 2. Since (D g1− )(s) C for some C > 1− α1 α2 0, it is sufficient to prove that lim Pt (D0+ fñ )(s) ds = 0, and in turn, for n→∞ this purpose it is sufficient to prove that Pt |φn,i (s)| ds → 0, 1 i 4, where −α2 s1 ˜ −α ˜ (fn (s) − fñ (u, s2 ))(s1 − u)−1−α1 du, φn,1 (s) = s−α 1 s2 fn (s), φn,2 (s) = s2 0 −α1 s2 ˜ (f (s) − fñ (s1 , v))(s2 − v)−1−α2 dv, φn,3 (s) = s1 0 n φn,4 (s) = [0,s] ∆u,v fñ (s)(s1 − u)−1−α1 (s2 − v)−1−α2 du dv. The relation (i+1)t1 1 |φn,1 (s)| ds → 0 is evident. Further, if it 2n s < 2n , then Pt −n −α2 i2 −1−α1 −nµ1 |φn,2 (s)| Cs2 (s1 − u1 ) du · 2 , whence 0 t2 −α2 n(α1 −µ1 ) |φ (s)| ds C 0 s2 ds2 · 2 → 0, n → ∞. Similarly, Pt n,2 −nµ1 |φ (s)| ds → 0, n → ∞. Finally, |φ n,3 n,4 (s)| ds C2 Pt Pt n 2 −1 × (s − u)−1−α1 (s2 − v + 2−n )µ2 −α2 −1 du dv ds1 ds2 ∆n [0,ti,n ] 1 ik k i,k=0 n(α1 +α2 −µ1 −µ2 )

= C2 → 0, n → ∞. Of course, similar estimates hold for I3n n n n and I4 . As to I5 , I6 and I7n , their estimates resemble each other, so, we consider only I5n . Note that lim Sn := lim

n→∞

n→∞

n −1 2

f (tni2n )(∆1i2n gi+12n )2 lim C · 2n · 2−2nµ1 = 0.

i=0

n→∞

Now, present the sum Sn as Sn =

n 2 −1

(fik (∆ik g)2 + 2fik ∆ik g∆1ik g + ∆2ik f (∆1ik g)2 + ∆2ik f (∆ik g)2

i,k=0

+ 2∆2ik f ∆1ik g∆ik g) =:

1i5

Sn,i ,

188


where Sn,1 C · 2−2n(µ1 +µ2 −1) → 0, n → ∞, similarly, Sn,4 → 0, Sn,5 → 0, n → ∞. According to previous estimates lim Sn,2 = lim I3n = 0. Therefore, n→∞ n→∞

lim I5n = lim Sn,3 = 0. n→∞

n→∞

Remark 2.7.9. Let F : R → R, F ∈ C 3 (R) and F is the Lipschitz function, the field g(t) is a linear combination of the fractional Brownian fields, g(t) =

m

H1i H2i

with Hji >

σi Bt

i=1

1 , j = 1, 2, 1 i m. 2

Clearly, the previous lemma holds for such g(t) and f (t) = F (g(t)). Theorem 2.7.10. For any t ∈ R2+ F (g(t)) = F (0) + F (g)dg + Pt

Pt

F (g)d1 g d2 g.

Proof. According to the one-parameter Itˆ o formula (Theorem 2.7.3)

t1

F (g(t)) = F (0) +

F (g(s1 , t2 ))d1 g(s1 , t2 )

0 n

= F (0) + lim

2

n→∞

f (tni,2n )∆1i,2n gi+1,2n a.s.

i=0

The prelimit sum can be presented as n 2 −1

F

(g(tnik ))∆ik g +

i,k=0

n 2 −1

F

(g(tnik ))∆1ik g∆2ik g +

i,k=0

+

1 2

n 2 −1

n 2 −1

2 F (g(sik n ))∆ik g∆ik g

i,k=0

n F (g(θik ))(∆2ik g)2 ∆1ik g +

i,k=0

1 2

n 2 −1

n F (g(θik ))(∆2ik g)2 ∆ik g,

i,k=0

(2.7.10) 2 −1 n where θik ∈ ∆nik . According to Theorem 2.2.9, F (g(tnik ))∆ik g → i,k=0 F (g)dg a.s. Furthermore, according to Theorem 2.2.17 and Lemma 2.7.8, Pt n 2 −1 2n −1 2 F (g(tnik ))∆1ik g∆2ik g → Pt F (g)d1 g d2 g, F (g(sik n ))∆ik g∆ik g → 0, n

i,k=0 −1 2n 1 F (g(tnik ))(∆2ik g)2 ∆1ik g 2 i,k=0

i,k=0

F (g(tnik ))(∆2ik g)2 ∆ik g → 0, i,k=0 −1 2n n F , 12 F (g(θik ))(∆2ik g)2 ∆1ik g → i,k=0

→ 0,

due to the Lipschitz properties of

−1 2n

1 2

and 0,

2.7 The Itˆ o Formula for Fractional Brownian Motion 1 2

−1 2n i,k=0

189

n F (g(θik ))(∆2ik g)2 ∆ik g → 0, n → ∞, a.s., and the assertion of the

theorem is proved.

Remark 2.7.11. The theorem holds even for F ∈ C 2 (R), such that F is the Lipschitz function. To prove this, we must rewrite the sum of second and 2n −1 n F (g(θik ))∆1ik g∆2ik g. fourth term on the right-hand side of (2.7.10) as i,k=0 Then we can prove that this sum has a limit Pt F (g)d1 g d2 g, similarly to Theorem 2.2.17. Also, the sum of third and fifth terms can be rewritten as 2n −1 n F (g(θik ))∆ik g∆2ik g, and we can prove that its limit is zero.

i,k=0

2.7.6 The Itˆ o Formula for H ∈ (0, 1) in Terms of Isometric Integrals, and Its Applications Definitions If f ∈ L2 (P ⊗ λ), f is predictable, π is a partition, then fπ is the step function f (ti−1 )1[ti−1 ,ti ) (t). fπ = i → − Define the class of functions Φ as follows: f ∈ Φ if the following conditions are satisfied: → − i i 2 i !(i)i !f := (f : i ≥ 1), where f ∈ L (P ⊗ λ), f is predictable and !f ! 2 < ∞. L (P ⊗λ) i → − (ii) f is uniformly tight: P {supt≤T supi |f i (t)| > C} → 0 as C → ∞. (iii) The random variable u defined by u := (fπi , (B H )(i) ) (for the noi

tations see Section 2.5.2) does not depend on the partition π, and the series → − converges absolutely with probability one, when f ∈ Φ. − − → i → − Write ( f , B H ) for the sum (fπ , (B H )(i) ), and put U := {u : u = i → −−→ − − → ( f , B H ), f ∈ Φ}. Let Φp be the projection of Φ to the first p coordinates. The following example shows that U is nonempty. Example 2.7.12. Assume that f ∈ Cb∞ (R): then f (BTH ) − f (0) =

n

∆f (BtHi )

i=1

and if f k := (1/k!)f (k) , k ≥ 1, then → −−→ − f (BTH ) − f (0) = ( f , B H ), → − f (BTH ) − f (0) ∈ U and f ∈ Φ, (f 1 , . . . , fp ) ∈ Φp for any p ≥ 1.

190


→ −−→ − Lemma 2.7.13. If u ∈ U, u = ( f , B H ) with u = 0, then f i = 0, i ≥ 1. Proof. Since u does not depend on the partition, take first the partition {0, T }. The random variable u has a representation f0i (BTH )i , (2.7.11) u= i

where f0i are real numbers, since F0 is the trivial σ-algebra. But u=0 isince f0 y i = 0 and from (2.7.11) it follows that for almost all y ∈ R we have that hence f0i = 0 for all i ≥ 1. Next, consider the partition {0, t, T }. We have that u= f0i (BtH )i + fti (BTH − BtH )i = 0. i

i

i

From the above we get that f0i = 0 for all i ≥ 1 and hence also fti = 0 for all i ≥ 1.

The Itˆ o Formula for Isometric Integrals The following is an analogue of the Itˆ o formula in this context. Theorem 2.7.14. Assume that the Hurst index H satisfies H ∈ (0, 1/2). There exists one-to-one correspondence between U and the set [1/H]

(f i , (B H )(i) ) . V := v : v := i=1

Proof. We must show that there exists one-to-one correspondence between U → and Φ[1/H] . Assume that f ∈ Φ[1/H] . Then there exists a vector − g ∈ Φ such → − i i that f = g for i ≤ [1/H]. Assume that h is another element from Φ such −−→ → −−→ − → that f i = hi for i ≤ [1/H]. Put u := (− g , B H ) and v := ( h , B H ). Then u−v =

∞

(g i − hi , (B H )(i) ).

i=1/H+1

On one hand, since u and v are independent of the partition π, we can take a partition π such that |π| < 1. Then for any ε > 0 we have that P {|u − v| > ε} ≤ P (D) + P {|u − v| > ε, Ω \ D} and D is the set D := {supt≤T supi |fti − gti | ≥ C}. But P {|u − v| > ε, Ω \ D} ≤

C E |∆BkH |i ε i>1/H

k

(2.7.12)

2.8 The Girsanov Theorem for fBm and Its Applications

and since E

191

|∆BkH |i ≤ CT (|π|)Hi−1

k

we have that P {|u − v| > ε, Ω \ D} → 0 as |π| → 0. By property (iii) of Φ we can choose C such that P (D) < δ for any δ > 0. Use these estimates in (2.7.12) to conclude that u = v. On the other → −−→ − → −−→ − → − − → hand, if u = ( f , B H ) = ( h , B H ) we have from Lemma 2.7.13 that f = h . To finish, note that from Example 2.7.12 it follows that the random variable [ T H(i) 1/H](1/i!) 0 f (i) (xs )dBs .

f (BTH ) − f (0) is a representative of i=1

Example 2.7.15 (Fractional Doleans exponent). Assume that [1/H] = 2p, where p ∈ N. Then the random variable yt = exp(BtH − t/(2p)!) − 1 is a representative of 2p−1 1 t ys d(BsH )(i) . i! 0 i=1 We say that y is the Doleans exponent of B H .

2.8 The Girsanov Theorem for fBm and Its Applications 2.8.1 The Girsanov Theorem for fBm −α −α Consider the kernel lH (t, s) = CH s (t − s) , 0 < s < t. Let H Ft = σ Bs , 0 s t = σ {Bs , 0 s t}, where B is underlying Wiener process in the representation t t lH (t, s)dBsH , Bt = α sα dMsH . MtH = (5)

0

0

Assume that the random process {φt , t 0} is adapted to filtration Ft and satisfies t lH (t, s) |φs | ds < ∞, t > 0, P -a.s. (2.8.1) 0

Assume also that we have the representation t t lH (t, s)φs ds = α δs ds, t > 0, 0

(2.8.2)

0

with some Ft -adapted process δ satisfying t |δs | ds < ∞, P -a.s., t > 0, 0

(2.8.3)

192


and

t

s2α δs2 ds < ∞, t > 0.

E 0

Define a square-integrable martingale L by Lt :=

t 0

(2.8.4) sα δs dBs .

Theorem 2.8.1. Assume that we have (2.8.1)–(2.8.4) and the martingale L satisfies E exp {Lt − 1/2Lt } = 1, t > 0. tH := BtH − t φs ds is an fBm with respect to measure Then the process B 0 Q, where the measure Q is defined by 1 dQ L = exp L − t t . dP Ft 2 Proof. Note first that the integral t t t H H H Mt := lH (t, s)dBs = lH (t, s)dBs − lH (t, s)φs ds 0

0

(2.8.5)

0

exists, since both integrals exist as pathwise integrals (the first integral was studied in Section 1.8 and (2.8.2) ensures the existence of the second integral). Moreover, from (2.8.2) it follows that t t t H H −α Mt = Mt − α δs ds = α s dBs − δs ds . 1

Evidently, M H

0

2 := P -lim t

t

1−2α

|π|→0 ti ∈π

0

0

2 1 (MtHi − MtHi−1 )2 exists and equals M H = t

tH := α . Therefore, for any θ ∈ R we have for M MtH that

t t θ2 1 H 2 1 θ2 t1−2α M + Lt − Lt = θ s−α dBs − θ δs ds − 2 2 2 1 − 2α t 0 0 t t t 1 + sα δs dBs − s2α δs2 ds = (θs−α + sα δs )dBs 2 0 0 0 1 t 2 −2α 1 − (θ s − 2δs θ + δs2 s2α )ds =: Rt − Rt , (2.8.6) 2 0 2 t where R is a square-integrable martingale given by Rt := 0 (θs−α +sα δs )dBs . But (2.8.6) means that the process θ2 1 H 2 1 H M + Lt − Lt Kt := exp θMt − 2 2 t tH − θM

is a local P -martingale. This implies, in turn, that the process 1 2

θ 2 H H is a local Q-martingale. From (Ell82, Theorem exp θMt − 2 M t


193

H is a local Q-martingale with the angle 13.22), we can conclude that M t −2α t H s , where B is a standard ds and so Mt = α 0 s−α dB bracket M t = 0 s Brownian motion with respect to Q (and is obtained from B by subtracting a drift). This means that t t H s . lH (t, s)dBs = α s−α dB (2.8.7) 0

0

H , (2.8.5) and (2.8.7), we can obtain Now, using two representations for B H and then conclude from Remark 1.8.2 that it is the fBm with (1.8.17) for B respect to the measure Q.

2.8.2 When the Conditions of the Girsanov Theorem Are Fulfilled? Differentiability of the Fractional Integrals If we analyze the conditions of the Girsanov theorem, we see that condition (2.8.2) is a principal concern. Now we shall establish that in one particular but t important case this condition holds. Let the process I(t) := 0 lH (t, s)φ(s)ds t with φ(t) = 0 a(s, ω)dBsH , where the integrand a = a(s, ω) : R × Ω → R is measurable in its variables and for a.a. ω ∈ Ω is Hölder in s with some index β ∈ (1/2, 1). According to Theorem 2.1.7, the integral φ(t) exists as a pathwise integral for ω ∈ Ω , P (Ω ) = 1. Moreover, according to Lemma t t 2.6.6, there exists a repeated integral J(t) := 0 a(u, ω) u lH (t, s)ds dBuH and the equality I(t) = J(t) holds for ω ∈ Ω . Lemma 2.8.2. Let a ∈ C ρ [0, t] for any t > 0 and for any ω ∈ Ω , P (Ω ) = 1, ρ ∈ (1/2, 1). Then for any t > 0 I(t) admits the representation t (5) δs ds, I(t) = CH t1−2α 2α−2

where δs = s ω ∈ Ω.

s 0

0

u

1−α

−α

(s − u)

a(u, ω)dBuH , and δ ∈ L1 [0, t] for any t > 0,

Proof. Further we suppose everywhere that ω ∈ Ω and argument ω will be omitted. We rewrite J(t) as t

1

J(t) = t1−2α

a(u)lH (1, s)ds dBuH 0

=

(5) CH t1−2α

0

u/t t t u

s2α−2 (s − u)−α u1−α a(u)ds dBuH =: CH t1−2α M (t). (5)

Consider now the function s t s2α−2 (s − u)−α u1−α a(u)dBuH ds. N (t) := 0

0

194


The following results ensure its existence: (i) According to (NVV99, Lemma 2.1), for the function g ∈ C β [0, T ] with t 0 < γ + β < 1, f (0) = 0 the integral 0 (t − u)γ dg(u) exists and equals

t

0

(t − u)γ dg(u) = lim (εγ (g(t − ε) − g(t)) ε→0 t−ε (g(u) − g(t))(t − u)γ−1 du). + tγ g(t) + γ

(2.8.8)

0

(ii) According to Lemma 2.6.3, for f ∈ C γ [a, b], g ∈ C β [a, b], γ + β > 1, 0 < ε < γ + β − 1 b f (t)dg(t) C f C γ [a,b] gC β [a,b] ((b − a)1+ε ∨ (b − a)β ), (2.8.9) a where C does not depend of f and g. Using (2.8.8)–(2.8.9), we obtain the following estimates for 0 < s1 < s2 < t: s2 s2 −α H a(z)(s2 − z) dBz = lim −ε a(v)dBvH s1

−α

ε→0

s2

s2 −ε s2 −ε

−1−α

s2

+ (s2 − s1 ) +α (s2 − z) s1 s1 z ! H! 1−α+ε lim C aC ρ [0,t] !B !C H [0,t] (s2 − s1 ) ∨ (s2 − s1 )−α+H ε→0 s2 −ε −1−α 1+ε H (s2 − z) dz , (2.8.10) +α (s2 − z) ∨ (s2 − z) a(z)dBzH

a(v)dBvH dz

s1

where H is any constant not exceeding H and 0 < ε < ρ + H − 1. Evidently, the right-hand side of (2.8.10) by CK1 (t)(s2 − s1 )−α+H , ! H ! can be estimated where K1 (t) aC ρ [0,t] !B !C H [0,t] (t ∨ 1)1+ε−H , C does not depend on ρ, B H , t. Further, u s2 s2 −α 1−α H 1−α −α H (s2 − u) u a(u)dBu = u d (s2 − z) a(z)dBz s1 s1 s s2 1 u s2 = s1−α (s2 − z)−α a(z)dBzH − (1 − α) u−α (s2 − z)−α a(z)dBzH du 2 s1

s1

s1

=: L(s1 , s2 ). The estimate

|L(s1 , s2 )| Cs1−α K1 (t)(s2 − s1 )−α+H 2 s2 + C(1 − α)K1 (t) u−α (u − s1 )−α+H du s1 1−α CK1 (t) s2 (s2 − s1 )−α+H + (s2 − s1 )1−2α+H

(2.8.11)


195

means that |L(0, s)| CK1 (t)s1−2α+H . Now it is clear that t s2α−2 s1−2α+H ds CK1 (t)tH < ∞. |Nt | CK1 (t) 0

Consider the function t Nε (t) := s2α−2 1{s∈[ε,t]} 0

s−ε

u1−α (s − u)−α a(u)dBuH ds.

0

Evidently, for any ε > 0 the function φε (s, u) := 1{s∈[ε,t],0us−ε} s2α−2 u1−α (s − u)−α a(u) is piecewise-Hölder in u with index ρ ∧ (1 − α) > 1/2 (u = s − ε is the point of H¨ older discontinuity), and the function ψε (s) := 0

t

φε (s, u)dBuH = s2α−2 1{s∈[ε,t]}

s−ε

(s − u)−α u1−α a(u)dBuH

0

is Riemann integrable on [0, t]. Therefore, φε (s, u) satisfies the conditions of the stochastic Fubini Theorem 2.6.5, whence Nε (t) exists and equals

t−ε

t

u1−α a(u)

Mε (t) := 0

s2α−2 (s − u)−α ds dBuH .

u+ε

Further, t s |N (t) − Nε (t)| s2α−2 u1−α (s − u)−α a(u)dBuH ds s−ε εε s + s2α−2 u1−α (s − u)−α a(u)dBuH ds 0 0 t s2α−2 CK1 (t)(s1−α ε−α+H + ε1−2α+H )ds ε ε s2α−2 CK1 (t)s1−2α+H ds + 0

≤ CK1 (t)(ε−α+H + εH ) → 0, ε → 0. For M (t) − M ε (t) we use one of the integral transformations from (NVV99, c Lemma 2.2): for µ ∈ R, ν > −1, c > 1 the integral 1 tµ (t − 1)ν dt 1−1/c ν s (1 − s)−µ−ν−2 ds, and as a result obtain the bound = 0

196


t−ε u+ε 1−α 2α−2 −α H |M (t) − Mε (t)| C a(u)u s (s − u) ds dBu 0 u t t 1−α 2α−2 −α H +C a(u)u s (s − u) ds dBu u t−ε ε u+ε t−ε −α −α H =C a(u) s (1 − s) ds dBu 0 0 1− ut t +C a(u) s−α (1 − s)−α ds dBuH =: A1 (ε) + A2 (ε). t−ε 0 According to the stochastic Fubini theorem 2.6.5,

ε/t

A1 (ε) = C

s−α (1 − s)−α

t−ε

a(u)dBuH ds

0

0

t

+C

−α

s

−α

(1 − s)

ε/t

A2 (ε) = C

s−α (1 − s)−α

0

ε(1−s) s

a(u)dBuH ds

0

ε/t

and

t(1−s)

a(u)dBuH ds. t−ε

Therefore, |A1 (ε)| C

ε −α ε 1−α t t 0 H 1 ε(1 − s) −α α + CK1 (t) s (1 − s) ds → 0, ε → 0, s ε/t t−ε

a(u)sBuH

1−

and

ε/t

|A2 (ε)| CK1 (t)

s−α (1 − s)−α (ε − ts)H ds → 0, ε → 0.

0

Therefore, N (t) = M (t), and our lemma is proved.

3 Stochastic Differential Equations Involving Fractional Brownian Motion

3.1 Stochastic Differential Equations Driven by Fractional Brownian Motion with Pathwise Integrals 3.1.1 Existence and Uniqueness of Solutions: the Results of Nualart and Rˇ a¸scanu Consider the function σ = σ(t, x) : [0, T ] × R → R satisfying the assumptions: σ is differentiable in x, there exist M > 0, 0 < γ, κ ≤ 1 and for any R > 0 there exists MR > 0 such that (i) σ is Lipschitz continuous in x: |σ(t, x) − σ(t, y)| ≤ M |x − y|,

∀t ∈ [0, T ], x, y ∈ R;

(ii) x-derivative of σ is local Hölder continuous in x: |σx (t, x) − σx (t, y)| ≤ MR |x − y|κ ,

∀|x|, |y| ≤ R, t ∈ [0, T ];

(iii) σ is Hölder continuous in time: |σ(t, x) − σ(s, x)| + |σx (t, x) − σx (s, x)| ≤ M |t − s|γ ,

∀x ∈ R, t, s ∈ [0, T ].

Let 0 < β < 1/2, f ∈ W0β [0, T ], g ∈ W11−β [0, T ]. We need some preliminary estimates, in addition to Lemmas 2.1.9 and 2.1.10. Consider on W0β [0, T ] the norm, equivalent to · 0,β : f 0,β,λ := sup e−λt ϕβf (t). t∈[0,T ]

Lemma 3.1.1 ((NR00)). Let assumptions (i)–(iii) hold with γ > β. Then the following statements hold. t 1. There exists the integral G(σ) (f )(t) := 0 σ(·, f (·))dg, t ∈ [0, T ]. 2. G(σ) (f ) ∈ C 1−β [0, T ] ⊂ W0β [0, T ].

198


3. G(σ) (f )1−β ≤ C1 Λ1−β (g)(1 + f 0,β ). 4. G(σ) (f )0,β,λ ≤ C2 Λ1−β (g)λ2β−1 (1 + f 0,β,λ ), λ ≥ 1, where C1 and C2 depend only on M, β, γ, T and |σ(0, 0)|. 5. For any f, h ∈ W0β [0, T ] such that fT∗ ∨ h∗T ≤ R G(σ) (f ) − G(σ) (h)0,β,λ ≤ C3 λ2β−1 Λ1−β (g)(1 + Cf + Ch )f − h0,β,λ , where Cf := supr∈[0,T ]

r

|fr −fs |κ ds, 0 (r−s)β+1

C3 depends only on M, β, γ, R, MR , T.

Proof. We prove only statement 5; the others can be proved in a similar, but more simple way. It is easy to check via the Taylor formula in the integral form that the function σ satisfying (i)–(iii) admits the following bound: for any R > 0, ti ∈ [0, T ], i = 1, 2, and |xi | ≤ R, 1 ≤ i ≤ 4 |σ(t1 , x1 ) − σ(t2 , x2 ) − σ(t1 , x3 ) + σ(t2 , x4 )| ≤ M |x1 − x2 − x3 + x4 | + M |x1 − x3 ||t2 − t1 |γ + MR |x1 − x3 |(|x1 − x2 |κ + |x3 − x4 |κ ).

(3.1.1)

Therefore, from Lemma 2.1.9, part 1, G(σ) (f ) − G(σ) (h)0,β,λ t 1 ≤ Cβ,T Λ1−β (g) supt∈[0,T ] e−λt 0 ((t − r)−2β + r−β )ϕσ(·,f (·))−σ(·,h(·)) (r)dr 1 ≤ Cβ,T Λ1−β (g) supr∈[0,T ] (e−λr ϕσ(·,f (·))−σ(·,h(·)) (r)) t −λ(t−r) ((t − r)−2β + r−β )dr. × 0e (3.1.2) The last integral in (3.1.2) can be estimated by ∞

t e−λu u−2β du + 0 e−λu (t − u)−β du λt ∞ = λ2β−1 0 e−u u−2β du + λβ−1 0 e−u (λt − u)−β du ≤ λ2β−1 C1,β + λβ−1 C2,β 0

z ∞ with C1,β = 0 e−u u−2β du, C2,β = supz≥0 0 e−u (z − u)−β du. Evidently, for λ ≥ 1 t e−λ(t−r) ((t − r)−2β + r−β )dr ≤ λ2β−1 (C1,β + C2,β ).

(3.1.3)

(3.1.4)

0

Further, from the Lipschitz property (i) and (3.1.1), it follows that r ϕσ(·,f (·))−σ(·,h(·)) (r) ≤ M |f (r) − h(r)| + M 0 |f (r) − f (s) − h(r) M + h(s)|(r − s)−β−1 ds + γ−β |f (r) − h(r)|rγ−β r κ r |f (r)−f (s)|κ + MR |f (r) − h(r)| 0 |r−s|β+1 ds + 0 |h(r)−h(s)| ds . β+1 |r−s| The proof follows now from (3.1.2)–(3.1.5), with C3 = (C1,β directly T γ−β + C2,β )(M + MR ) 1 + γ−β .

(3.1.5)

3.1 SDEs Driven by fBm with Pathwise Integrals

199

The next lemma describes the situation with the Lebesgue integrals. Let the function b = b(t, x) : [0, T ] × R → R satisfy the assumptions (iv) for any R ≥ 0 there exists LR > 0 such that |b(t, x) − b(t, y)| ≤ LR |x − y|,

∀|x|, |y| ≤ R, ∀t ∈ [0, T ];

(v) there exists the function b0 ∈ Lp [0, T ] and L > 0 such that |b(t, x)| ≤ L|x| + b0 (t),

∀(t, x) ∈ [0, T ] × R.

Lemma 3.1.2. Let 0 < β < 1/2, assumptions (iv) and (v) hold with p = β −1 , f ∈ W0β [0, T ]. Then the following statements hold. t 1. There exists the Lebesgue integral F (b) (f )(t) := 0 b(s, f (s))ds, t ∈ [0, T ]. 2. F (b) (f ) ∈ C 1−β [0, T ]. 3. F (b) (f )1−β ≤ C4 (1 + fT∗ ) ≤ C4 (1 + f 0,β ). 4. F (b) (f )0,β,λ ≤ C5 λ2β−1 (1 + f 0,β,λ ), where λ ≥ 1, C4 and C5 depend only on β, T, L and b0 Lp [0,T ] . 5. Let f, h ∈ W0β [0, T ] with fT∗ ∨ h∗T ≤ R. Then F (b) (f ) − F (b) (h)0,β,λ ≤ C6 λβ−1 f − h0,β,λ ,

λ ≥ 1,

where C6 depends on β, R, T, LR . Proof. We prove only statement 4. Indeed, from Lemma 2.1.10, t |b(s,f (s))| t (L|f (s)|+b0 (s)) 3 3 ds ≤ Cβ,T ds ϕβF (b) (f ) (t) ≤ Cβ,T 0 (t−s)β 0 (t−s)β 1−β β t t |f (s)| − 1−β 3 ≤ Cβ,T L 0 (t−s)β ds + 0 (t − s) ds b0 L1/β [0,T ] t |f (s)| 3 1−2β L 0 (t−s) ≤ Cβ,T B0,β , β ds + cβ t where cβ = Hence

1−β 1−2β

1−β

(3.1.6)

3 , B0,β = b0 Lp [0,T ] , Cβ,T = T β + 1/β.

t |f (s)| 3 · L · supt∈[0,T ] e−λt 0 (t−s) F (b) (f )0,β,λ ≤ Cβ,T β ds 3 −λt 1−2β + Cβ,T cβ B0,β supt∈[0,T ] e t t 3 ≤ Cβ,T · L · sups∈[0,T ] e−λs |f (s)| 0 e−λu u−β ds 3 + Cβ,T cβ B0,β λ2β−1 supz≥0 e−z z 1−2β ≤ C5 λ2β−1 (1 + f 0,β,λ ), 3 (L · Γ (1 − β) + cβ B0,β supz≥0 e−z z 1−2β ). where C5 = Cβ,T

Now, let 0 < β < 1 be fixed, g ∈ W11−β [0, T ]. Consider the (deterministic) differential equation t t b(s, Xs )ds + σ(s, Xs )dgs , t ∈ [0, T ], (3.1.7) Xt = X0 + 0

0

200


where X0 ∈ R, and the coefficients σ, b : [0, T ] × R → R are measurable functions satisfying (i)–(v) with p = 1/β, 0 < γ, κ ≤ 1 and κ . 0 < β < β0 = 12 ∧ γ ∧ 1+κ Theorem 3.1.3. Equation (3.1.7) has the unique solution X ∈ W0β [0, T ]. This solution belongs also to the space C 1−β [0, T ]. Proof. Let the function f ∈ W0β [0, T ]. Then, according to statements 3 of Lemmas 3.1.1 and 3.1.2, G(σ) (f ) ∈ C 1−β [0, T ] and F (b) (f ) ∈ C 1−β [0, T ]. So, if X is the solution of (3.1.7) and X ∈ W0β [0, T ], then X = X0 + F (b) (X)(t) + G(σ) (X)(t) ∈ C 1−β [0, T ]. Now we prove the uniqueness. Let X and Y be two solutions from C 1−β [0, T ] and XC 1−β [0,T ] ∨ Y C 1−β [0,T ] ≤ R. Then from statements 5 of Lemmas 3.1.1 and 3.1.2, for β < γ X − Y 0,β,λ ≤ F (b) (X) − F (b) (Y )0,β,λ + G(σ) (X) − G(σ) (Y )0,β,λ ≤ (C3 Λ1−β (g)λ2β−1 (1 + CX + CY ) + C6 λβ−1 )X − Y 0,β,λ , λ ≥ 1. Note, that for β
1 − H, κ > H −1 − 1 (the constants M, MR , R, LR and the function b0 can depend on ω). Then there exists the unique solution {Xt , t ∈ [0, T ]} of equation (3.1.11), X ∈ L0 (Ω, F, P, W01−H+ε [0, T ]) with a.a. trajectories from C H−ε [0, T ].

202


Remark 3.1.5. Theorem 3.1.4 admits evident generalization to the multidimensional case. Consider the equation on Rd Xti

=

X0i

+

t

bi (s, Xs )ds + 0

m j=1

t

σji (s, Xs )dBsHj ,

1 ≤ i ≤ d, t ∈ [0, T ],

0

(3.1.12) where the processes B Hj are fBms with Hurst index Hj ∈ (1/2, 1), 1 ≤ j ≤ the matrix of “diffusions” and b = (bi )di=1 the m. Denote by σ = (σji )d,m i,j=1 2 1/2 “drift” vector, |σ| := ( i,j |σji | ) , |b| := ( i (bi )2 )1/2 , and suppose that assumptions (i)–(v) hold with these notations, H = min1≤j≤m Hj , p = (1 − H + ε)−1 , γ > 1 − H, κ > H −1 − 1. Then there exists the unique vector solution Xt of equation (3.1.12) on [0, T ] in L0 (Ω, F, P, W01−H+ε [0, T ]) with a.a. trajectories from C H−ε [0, T ]. 3.1.2 Norm and Moment Estimates of Solution We consider equation (3.1.7), suppose that the assumptions of Theorem 3.1.3 hold and, in addition, the coefficient σ satisfies the following growth condition: (v’) |σ(t, x)| ≤ M (1 + |x|µ ) for some 0 ≤ µ ≤ 1. Lemma 3.1.6. The solution of (3.1.7) satisfies the estimate X0,β ≤ C0 exp(C1 (Λ1−β (g))κ ), where 0 < β < β0 = 1/2 ∧ γ ∧

κ 1+κ ,

⎧ −1 ⎪ ⎨ (1 − 2β) , if µ = 1, 1−2β −1 κ = (1 − β) , if 0 ≤ µ < 1−β , ⎪ 1−2β ⎩> µ , if 1−β ≤ µ < 1, 1−2β

(3.1.13)

and the constants C0 and C1 depend on T, β, µ and on the constants from conditions (i)–(v). Proof. Evidently, ϕβX (t) ≤ |X0 | + ϕβF (b) (X) (t) + ϕβG(σ) (X) (t).

(3.1.14)

t |Xu | 3 1−2β ϕβF (b) (X) (t) ≤ Cβ,T (L 0 (t−u) B0,β ) β du + cβ t t |Xu | 3 4 ≤ LCβ,T 0 (t−u)β du + Cβ,T ,

(3.1.15)

From (3.1.6)


203

t r t s )| r )−σ(u,Xu )| |G(σ) (X)| ≤ Λ1−β (g) 0 |σ(s,X ds + β 0 0 |σ(r,X du dr β s (r−u)β+1 µ t |Xs | t r |Xr −Xu | ≤ Λ1−β (g) M 0 sβ ds + M 0 0 (r−u)β+1 du dr 1−β tγ−β+1 + M t1−β + M (γ−β)(γ−β+1) t r |Xr −Xu | µ t , ≤ Cβ,γ,T Λ1−β (g) + M Λ1−β (g) 0 |Xssβ| ds + 0 0 (r−u) β+1 du dr (3.1.16) and, similarly to (2.1.15)–(2.1.16) t 0

+

t (X)−G(σ) s (X)| ds ≤ M Λ1−β (g)(Cβ,γ,T + 0 (t−s)β+1 u − u)−β 0 |Xu − Xv |(u − v)−β−1 dv du).

(σ)

|Gt

t 0

(t

Let us estimate the “worst” integral t

t 0

|Xu |µ (t − u)−2β du (3.1.17)

|Xu |µ (t − u)−2β du: 1/p

1/q ds , (2β−ρ)q 0 0 0 (t − s) (3.1.18) where we must choose µp = 1, (2β − ρ)q < 1, whence ρ > 2β + µ − 1, and estimate (3.1.18) takes the form µ t t |Xu | µ −2β |X | (t − u) du ≤ C du u β,µ,T 0 0 (t−u)ρ/µ (3.1.19) t |Xu | ≤ Cβ,µ,T 1 + 0 (t−u)ν du ,

t

|Xu |µ (t − u)−2β du ≤

where ν =

ρ µ

>

2β+µ−1 µ

|Xu |µ (t − u)ρ

p

t

du

(for µ = 1 we put ν = 2β).

From (3.1.14)–(3.1.19) we obtain that ϕβX (t) admits an estimate ϕβX (t) ≤ K1 (1 + Λ1−β (g)) + K2 (1 + Λ1−β (g)) ·

0

t

ϕβX (u)((t − u)−ν + u−β )du

with constants K1 and K2 depending on on T, β, µ and on the constants from conditions (i)–(v). Evidently, (t − u)−ν + u−β = For µ > ν=β>

1−2β 1−β we 2β+µ−1 . µ

uβ + (t − u)ν ≤ (tβ + tν )u−β (t − u)−ν . uβ (t − u)ν

have that ν > β; for 0 < µ ≤

1−2β 1−β

we can put

In any case

ϕβX (t) ≤ K1 (1 + Λ1−β (g)) + K2 (1 + Λ1−β (g))tν

0

t

ϕβX (u)u−ν (t − u)−ν du.

(3.1.20) In (NR00) the following version of the Gronwall lemma was proved: if 0 ≤ c < 1, a, b ≥ 0, x : R+ → R+ is a continuous function such that for each t ∈ [0, T ]

204


xt ≤ a + bt

c

t

(t − s)−c s−c ds,

(3.1.21)

0

then

1

xt ≤ C3 exp{C4 tb 1−c },

(3.1.22)

where C3 and C4 depend only on a, b, c. The proof follows from (3.1.20)– (3.1.22).

In the case of equation (3.1.11) g(t) = B H (t, ω) and instead of Λ1−β (g) 1−β H 1 sup0≤s 3/4 for µ = 1 and β < 12 − µ4 if 1−2β 1−β ≤ µ < 1), then EXq0,β < ∞ for any q > 0. 3.1.3 Some Other Results on Existence and Uniqueness of Solution of SDE Involving Processes Related to fBm with (H ∈ (1/2, 1)) It follows from the results of Subsection 3.1.1, that it is possible to consider an SDE involving fBm with H ∈ (1/2, 1) as an ordinary differential equation for any ω ∈ Ω , P (Ω ) = 1. Therefore, the results for the ordinary differential equations with the H¨ older continuous forcing can be applied. One of these results belongs to Ruzmaikina (Ruz00). Another approach was developed in the papers (CQ00), (GA98), (GA99a), (GA99b), (Jum93), (IT99), (KKA98a), (Kli98), (KZ99), (MN00), (Mis03), (Zah99) (Zah01) and (Zah05). For example, in the papers (Zah99) and (Zah05) the author considers SDEs of the form dXti =

m

σji (t, Xt )dZtj,− + bi (t, Xt )dt,

j=0

(3.1.23)

t ∈ [0, T ], Xt0 = X0 , 0 ≤ t0 < T, under the following assumptions: (vi) σji ∈ C 1 (Rd ×[0, T ], Rd ) and all partial derivatives are locally Lipschitz in x ∈ Rd ; (vii) bi ∈ C(Rd × [0, T ], Rd ) is locally Lipschitz in x ∈ Rd (with probability 1 in the random case). Here 1 ≤ i ≤ d. Also, X0 is an arbitrary vector random variable. The integrals w.r.t. the processes Ztj are the generalized stochastic forward integrals. What are they and what processes can we consider here? (Recall that the forward (not generalized) integrals were introduced in the Section 2.4.)


205

Suppose that Y is a stochastic càgl` ad (left continuous with right limits) process and Z a stochastic càdl` ag (right continuous with left limits) process on [0, T ]. Then the generalized stochastic forward integral is defined as 1 t t Zt− (s + u) − Zt− (s) ds du, (3.1.24) Y dZ − := lim ε uε−1 Ys ε→0 u 0 0 0 whenever the right-hand side is determined, where lim stands t for uniform 1 on [0, T ] convergence in probability (ucp-convergence), and 0 = limδ↓0 δ a.s. We use the same notation as for the forward integral in the Section 2.4. Similarly, the generalized quadratic variation process (bracket) is defined as 1 t 1 ε−1 (Zt− (s+u)−Zt− (s))2 ds du+(Zt −Zt− )2 (3.1.25) u , [Z]t := lim ε ε→0 0 0 u whenever the convergence holds uniformly in probability. If Z is a semimartingale and Y an adapted c` agl` ad process then integral (3.1.24) agrees with the t− usual Itˆ o integral 0+ Y dZ, and notion of the generalized bracket coincides with the classical one. If Z is a continuous process with the generalized bracket [Z], and the function F = F (t, x) : R × [0, T ] → R, F ∈ C 1 ([0, T ]) × C 1 (R), then the simple Itô formula holds for 0 ≤ s < t ≤ T : t F (t, Zt ) = F (s, Zs ) + s ∂F (u, Zu )dZu− ∂x t ∂F 1 t ∂2F + s ∂t (u, Zu )du + 2 s ∂x2 (u, Zu )d[Z]u . Now we suppose that the paths of Z j , 1 ≤ j ≤ m from equation (3.1.23) , H W3α , H ∈ (1/2, 1), belong to the Sobolev–Slobodeckij space W3 − := β

1 H

− 1.

Existence of Pathwise Solution for Bounded Coefficients Now we relax the conditions on coefficients to obtain the existence (not uniqueness) of the solution of equation (3.1.28). Theorem 3.1.10. Let the coefficient b be bounded and continuous, coefficient σ be bounded and H¨ older of order 1 > ρ > 1/H − 1. Then equation (3.1.28) has a pathwise solution. Proof. We consider the sequence {ψn (x), x ∈ Rd , n ≥ 1} of smooth kernels, such that ψn ≥ 0; ψn = 0, |x| ≥ n1 ; ψn ∈ C ∞ (Rd ); Rd ψn (x)dx = 1. Introduce the functions b(y)ψn (x − y)dy, σn (x) = σ(y)ψn (x − y)dy. bn (x) = Rd

Rd

Then for any x ∈ Rd and any 1 ≤ i ≤ d |∂xi bn (x)| =

Rd

b(y)∂xi ψn (x − y)dy ≤ b∞

Rd

|∂xi ψn (x − y)| dy ≤ Cn b∞ ,

where b∞ := supx∈Rd |b(x)|. The same estimate is true for σn , and it means that bn and σn are Lipschitz continuous, with the depending on n. Further, for any constants possibly x ∈ Rd , |bn (x)| = Rd b(y)ψn (x − y)dy ≤ b∞ i.e. bn are bounded functions. The same is true for σn . Finally, for any N > 0, x, y ∈ Rd , |x| ≤ N , |y| ≤ N and 1 ≤ i ≤ d

208


|∂xi σn (x) − ∂xi σn (y)| ≤ σ∞ ≤ σ∞

Rd

|∂xi ψn (x − z) − ∂xi ψn (y − z)| dz 2 sup ∂xi ,xj ψn (z) (N + 1)d |x − y|,

1 |z|≤ n ,1≤i,j≤d

i.e. ∂xi σn satisfy the local Lipschitz conditions. These estimates demonstrate that bn and σn satisfy conditions (i’)–(iii’) that in turn ensure the pathwise existence (and uniqueness) of the solution of the equation Xtn = X0 +

t

bn (Xsn )ds + 0

t

σn (Xsn )dBsH .

(3.1.29)

0

We fix some ω ∈ Ω and denote by C different constants even if they depend on Ω. According to Theorem 3.1.4 and Remark 3.1.5, the solution Xtn is Hölder continuous of order H −δ for any δ > 0; from H¨ older continuity of σ we obtain that ρ |σ(x − z) − σ(y − z)| |ψn (z)| dz ≤ C |x − y| , |σn (x) − σn (y)| ≤ Rd

(3.1.30) therefore, σn (Xsn ) belongs to the space W2β [0, T ] for any β < Hρ. By using estimate (2.1.14) and the boundedness of σn for any 0 ≤ s ≤ t, we obtain for each 1 − H < β < ρH (this is possible since ρ > 1/H − 1) the estimate t n H σ (X )dB n u u s

t u t σn (Xun ) − σn (Xyn ) |σn (Xun )| ≤G du + dy du β (u − y)β+1 s (u − s) s s ρ t u n Xu − Xyn (t − s)1−β ≤ G σ∞ + CG dy du. 1−β (u − y)β+1 s s Here GT := Λ1−β (B H ) (see Section 1.17) and EGpT < ∞ for any p > 0 (Lemma 1.17.1 and Remark 1.17.2). Finally, we can estimate t t |Xtn − Xsn | ≤ bn (Xun )du + σn (Xun )dBuH ≤ b∞ (t − s) s s ρ t u n Xu − Xyn σ∞ 1−α (t − s) + GT + CGT dy du. 1−α (u − y)β+1 s s Consider any fixed interval [0, T ] and denote X1−β,T :=

sup 0≤s 0. Indeed, X n are Hölder of order H − ε for any ε > 0 with constant, possibly depending on n (Theorem 3.1.4 and Remark 3.1.5), therefore X n 1−β,T < ∞ a.s. Now, from (3.1.31), σ∞ 1−β ρ t u n Xu − Xyn

X n 1−β,T ≤ b∞ T β + GT + CGT

sup 0≤s 1−β , which is possible for some β > 1 − H, since ρ > 1/H − 1. Note that for 0 < ρ < 1 the equality P = C(1 + P ρ ) has the unique root P0 > 0 and the inequality P ≤ C(1 + P ρ ) holds for P ≤ P0 . Therefore, X n 1−β,T ≤ P0 (T, ρ), where P0 (T, ρ) depends only on T and ρ, not on n. This means that

|Xtn − Xsn | ≤ P0 (T, ρ)(t − s)1−β ,

(3.1.32)

which according to the Arcela criterion means that the sequence {Xtn , t ∈ [0, T ]}, n ≥ 1 is tight for any ω ∈ Ω in the space C[0, T ]. Evidently, we can conclude that there exists {Xtnk , t ∈ [0, T ]}, nk ≥ 1, such that Xtnk → Xt in the space C[0, T ]. We can suppose that Xtn → Xt in C[0, T ]. Now it is sufficient to prove that X(t) is a solution of (3.1.28). Let consider some auxiliary estimates. First, t t t n (bn (Xsn ) − b(Xs ))ds ≤ |b (X ) − b (X )| ds+ |bn (Xs ) − b(Xs )| ds. n n s s 0

0

0

(3.1.33) Further, for any x, y ∈ Rd |bn (x) − bn (y)| = (b(x − u) − b(y − u))ψn (u)du Rd ≤ |b(x − u) − b(y − u)| ψn (u)du ≤ sup |b(x − u) − b(y − u)| . 1 |u|≤ n

1 |u|≤ n

(3.1.34) The process {Xt , t ∈ [0, T ]} is continuous on [0,T], so bounded for any ω ∈ Ω. Let C(T, ω) = sup0≤s≤T |Xs |. For any ε > 0 there exists η > 0 such that sup |s−z| 0 and any ω ∈ Ω there exists such n0 ∈ N that |Xsn − Xs | < η, n ≥ n0 , s ∈ [0, T ]. From these estimates, |bn (Xsn ) − bn (Xs )| ≤ sup |bn (Xsn − u) − bn (Xs − u)| 1 |u|≤ n

≤

sup |b(s) − b(z)| < β,

sup

1 |s−z| 0 is arbitrary, we obtain that a.s. t |bn (Xsn ) − bn (Xs )| ds → 0,

n ≥ n0 . (3.1.36)

n → ∞.

(3.1.37)

0

The second term on the right-hand side of (3.1.33) can be estimated in such a way: |bn (Xs ) − b(Xs )| ≤

Rd

|b(Xs − u) − b(Xs )| ψn (u)du

≤

sup

sup |b(z − u) − b(z)| → 0,

1 |z|≤C(T,ω) |u|≤ n

n → ∞,

since b is a continuous function. Moreover, bn are bounded, which implies t the convergence 0 |bn (Xs ) − b(Xs )| ds → 0, n → ∞ a.s. We obtain that t t b (Xsn )ds → 0 b(Xs )ds a.s., t > 0. Furthermore, 0 n t t (σn (Xsn ) − σ(Xs ))dBsH ≤ (σn (Xsn ) − σn (Xs ))dBsH 0 0 t H + (σn (Xs ) − σ(Xs ))dBs . (3.1.38) 0

Now we can estimate the first term of (3.1.38) for any 1 − H < β < 12 : t t |σn (Xsn ) − σn (Xs )| (σn (Xsn ) − σn (Xs ))dBsH ≤ G ds sβ 0 0 t u |σn (Xun ) − σn (Xrn ) + σn (Xr ) − σn (X)| +G dr du. (u − r)1+β 0 0

(3.1.39)

Similarly to estimates (3.1.34)–(3.1.37), we obtain that a.s. t sups≤T |σn (Xsn ) − σn (Xs )| → 0 and 0 |σn (Xsn ) − σn (Xs )| s−β ds → 0, while n → ∞. Further, recall the estimate (3.1.30). For any sufficiently small t u t u ε > 0, present 0 0 on the right-hand side of (3.1.39) as 0 0 t u−ε ε u t u = ε 0 + 0 0 + ε u−ε , and the integrals on the right-hand side can be estimated as


t

t

u−ε

u−ε

≤ 2 sup |σn (Xsn ) − σn (Xs )| ε

s≤T

0

ε

0 −α

≤ Cε

211

dr du (u − r)1+β

· sup |σn (Xsn ) − σn (Xs )| → 0 s≤T

a.s. for any fixed ε > 0. Further, from (3.1.32), t

t

u

≤C ε

u−ε

ρ

ρ

|Xun − Xrn | + |Xr − Xu | dr du (u − r)1+β ε u−ε t u (u − r)ρ(1−β)−1−β dr du = Cερ(1−β)−β ≤C u

0

(3.1.40)

u−ε

is small for small ε > 0, and moreover, C does not depend on n. The integral ε u ≤ Cερ(1−β)−β+1 . (3.1.41) 0

0

t u Therefore, since ε > 0 is arbitrary, 0 0 → 0 a.s. while n → ∞. The second term of (3.1.38) can be estimated as t t |σn (Xs ) − σ(Xs )| (σn (Xs ) − σ(Xs ))dBsH ≤ G ds sβ 0 0 t u |σn (Xu ) − σn (Xr ) − σ(Xu ) + σ(Xr )| +G dr du. (u − r)1+β 0 0 Evidently, |σn (Xs ) − σ(Xs )| ≤

Rd

|σ(Xs − u) − σ(Xs )| ψn (u)du ρ

≤ sup |u| ≤ 1 |u|≤ n

1 → 0, nρ

n→∞

t and |σ (X ) − σ(X )|s−β ds → 0 a.s., n → ∞. Now, as before, t u 0 n t su−ε εs u t u t u−ε = ε 0 + 0 0 + ε u−ε and ε 0 ≤ 2 n1ρ · ε−β → 0 for any 0 0 fixed ε > 0 and other integrals can be estimated as in (3.1.40)–(3.1.41). So, t t σ (X n )dBsH → 0 σ(Xs )dBsH a.s. while n → ∞ and 0 n s t t

Xt = 0 σ(Xs )dBsH + 0 b(Xs )ds. The theorem is proved. Remark 3.1.11. By similar, but even more simple arguments we can prove the existence of the solution of the equation t t Xt = X0 + b(Xs )ds + f (s)dBsH , 0

0

where b is bounded and continuous, f ∈ C 1−H [0, T ], X0 is a real-valued random variable.

212


Differentiability and Local Differentiability of the Solution Here we shall use the elements of Malliavin calculus with respect to fBm B H , contained in Section 2.4. Suppose that we consider some subspace Ω1 ⊂ Ω and restrict F and P to Ω1 . Denote the mathematical expectation w.r.t. restricted measure P1 as E1 . Definition 3.1.12. Random variable F belongs locally to the space ∞ D1,p (H) on [0, T ] if there exists a sequence Ω1 ⊂ Ω2 ⊂ . . . ⊂ Ω such that n=1 Ωn = Ω p p and F 1,p,n := En (|F | ) + En (DF )pH < ∞. In this case we say that F is locally differentiable, F ∈ D1,p,loc . According to Lemma 1.5.4 (Nua95), see also (Nua98), we can formulate the sufficient conditions of local differentiability. Let {Fr , r ≥ 1} be a sequence of r.v. from D1,p,loc satisfying the conditions (viii) Fr → F in any Lp (Ωn ), n ≥ 1, (ix) supr Fr 1,p,n < ∞ for any n ≥ 1. Then F belongs to D1,p,loc . Remark 3.1.13. Suppose that there exists a localizing sequence (Ωn , n ≥ 1), such that F ∈ D1,p,loc for any p > 1. Then we say that F ∈ D1,∞,loc . Consider equation (3.1.28) and suppose that its coefficients X0 , b and σ satisfy conditions (i’)–(ii’) and (x) b ∈ C 1 (Rd ); d (xi) |∂xi σ(x) − ∂xi σ(y)| , ≤ M |x − y|, ∀x, y ∈ R ; (xii) X0 ∈ D1,∞ := p≥1 D1,p (H), X0 is a bounded F0 -adapted random variable. Theorem 3.1.14. 1. Let conditions (i’)–(ii’) and (x)–(xii) hold. Then the unique solution Xt of equation (3.1.28) is locally differentiable in the sense that Xti ∈ D1,∞,loc for any 1 ≤ i ≤ d with the same localizing sequence. 2. Let equation (3.1.28) be semilinear, i.e. σ(x) = σx, conditions (i’), (ii’), (x), (xii) hold for b and X0 and H > 3/4. Then Xti ∈ D1,∞ for any 1 ≤ i ≤ d. Proof. 1. Let T > 0 be fixed. According to Theorem 3.1.4 and Remark 3.1.5, under conditions (i’)–(i”) and (x) equation (3.1.28) has the unique solution Xt on the interval [0, T ]. Moreover, it can be obtained by successive approx(n)

(0)

imations, Xt , n ≥ 0 , t ∈ [0, T ] where Xt = X0 ∈ D1,∞ . Further we consider the case d = 1, for technical simplicity; in the general case they are (k) similar. We use induction. Suppose that Xt ∈ D1,∞ , 1 ≤ k ≤ n, and the (k) older continuous of order derivatives Ds Xt , 0 ≤ s ≤ t ≤ T , 1 ≤ k ≤ n are H¨ (n) older 1 − β for some 1 − H < β < 1/2. Since the approximations Xt are H¨ continuous of any order not exceeding H, and from conditions (i’) and (x) σ (n) (n) and b are bounded, σ (Xr )Ds Xr is Hölder continuous in r of order 1 − β.


213

t t (n) (n) (n) (n) Therefore, the integrals s σ (Xr )Ds Xr dBrH and s b (Xr )Ds Xr dr exist, 0 ≤ s ≤ t ≤ t ≤ T and t t (n+1) = σ(Xs(n) ) + σ (Xr(n) )Ds Xr(n) dBrH + b (Xr(n) )Ds Xr(n) dr. Ds Xt s

Hence for any 1 − H < β < (n+1)

|Ds Xt

s 1 2,

from (2.1.14),

| ≤ |σ(Xs(n) )| + M

t

s

t +G s

r

t

|Ds Xr(n) |dr + M G

|Ds Xr(n) |(r − s)−β dr

s

|σ (Xr(n) )Ds Xr(n) − σ (Xu(n) )Ds Xu(n) |(r − u)−1−β du dr, (3.1.42)

s

where G = 1/Γ (β)

sup 0<s 1−2β , if 1−β ≤ µ < 1; ⎪ ⎩ 1 , if 0 ≤ µ < 1−2β 1−β 1−β . Finally, |σ(Xsn )| ≤ C exp{CGκ }, and from (3.1.42) it follows that t Ds Xr(n) t (n+1) dr Ds Xt ≤ C exp{CGκ } + M Ds Xr(n) dr + M G β s s (r − s) t r Xr(n) − Xu(n) Ds Xr(n) du dr + MG (r − u)1+β s s t r Ds Xr(n) − Ds Xu(n) du dr. (3.1.43) + CG (r − u)1+β s s It follows from Lemmas 3.1.1, 3.1.2 and 3.1.6 that |Xsn − Xrn | ≤ exp{CGκ } |s − r|

1−β

for any 1 − H < β < 12 . In this case 1 − 2β > 0 and from (3.1.43)

(3.1.44)

214


t (n+1) κ Ds Xt ≤ C exp{CG } + M Ds Xr(n) dr s t Ds Xr(n) t κ (n) + MG dr + M G exp{CG } X D (r − s)1−2β dr s r β s (r − s) s t r Ds Xr(n) − Ds Xu(n) du dr, + CG (r − u)1+β s s or, briefly, (n+1) ≤ C exp{CGκ } Ds Xt t Ds Xr(n) t r Ds Xr(n) − Ds Xu(n) + CG exp{CGκ } dr + CG du dr. β (r − u)1+β s (r − s) s s Further, for any 0 ≤ u < r ≤ T r (n+1) (n+1) (n) (n) H Ds Xr − Ds Xu = σ (Xv )Ds Xv dBv + u

r

b (Xv(n) )Ds Xv(n) dv,

u

and we obtain by similar estimates that r Ds Xu(n) dv Ds Xr(n+1) − Ds Xu(n+1) ≤ CG exp{CGκ } β u (v − u) r v Ds Xv(n) − Ds Xz(n) dz dv, + CG (v − z)1+β u u whence r Ds Xr(n+1) − Ds Xu(n+1) s

(r − u)1+β

(3.1.45)

du

r Ds Xv(n) 1 dv du ≤ CG exp{CGκ } 1+β β s (r − u) u (v − u) r r v Ds Xv(n) − Ds Xz(n) 1 dz dv. + CG 1+β (v − z)1+β s (r − u) u u

r

t Ds Xt(n) −Ds Xu(n) (n) 2 := Ds Xt , ϕn (t) := s du, ϕ10 (t) := Ds X0 , Denote (t−u)1+β ϕ20 (t) := 0, C˜1 (ω) := C exp{CGκ }, C˜2 (ω) := CG exp{CGκ }, C˜3 (ω) := CG. Then for ϕn (t) := ϕ1n (t) + ϕ2n (t)

ϕ1n (t)


t

215

t

ϕ1n (r)(r − s)−β dr + C˜3 (ω) ϕ2n (r)dr s s t ϕn (r)(r − s)−β dr; ≤ C˜1 (ω) + C˜2 (ω) + C˜3 (ω)T β

ϕ1n+1 (t) ≤ C˜1 (ω) + C˜2 (ω)

s

ϕ2n+1 (t)

t 1 ϕ1n (v) dv du 1+β β s (t − u) u (v − u) t t 1 ϕ2n (v)dv du + C˜3 (ω) 1+β s (t − u) u t v du ϕ1n (v) dv = C˜2 (ω) β 1+β s s (v − u) (t − u) t v ϕ2n (v) + C˜3 (ω) t

≤ C˜2 (ω)

s

s

du dv. (t − u)1+β

v

Since s (v − u)−β (t − u)−1−β du ≤ C(t − v)−2β , with ∞ C = 0 u−β (1 + u)−1−β du, we have that ϕ2n+1 (t)

t ˜ − v) dv + C3 (ω)C ϕ2n (v)(t − v)−β dv s t ϕn (v)(t − v)−2β dv. ≤ C C˜2 (ω) + C˜3 (ω)T β

t

≤ C˜2 (ω)C

−2β

ϕ1n (v)(t s

s

Finally,

t

ϕn+1 (t) ≤ C(ω) 1 + 0

−2β

−2β

dv ϕn (v) (t − v) +v t 2β −2β −2β ≤ C(ω) 1 + t ϕn (v)(t − v) v dv , 0

where C(ω) = C C˜1 (ω) ∨ C˜2 (ω) + C˜3 (ω)T β ∨ 1 for some C > 0. It is very easy to check by induction, similarly to (3.1.20)–(3.1.21), that 1

ϕn (t) ≤ C(ω)C1 exp{C2 t (C(ω)) 1−2β } =: ψ(t), where C1 and C2 depend only on β. In particular, ϕ1n ≤ ψ(t) and ϕ2n ≤ ψ(t). ˜ < ∞ a.s., and from (3.1.45) it follows that Evidently, sups≤t≤T ψ(t) =: C(ω) (r − u)1−β (n+1) (n+1) ˜ + (r − u) , − Ds Xu Ds Xr ≤ C(ω)C(ω) 1−β (n+1)

i.e. Ds Xr is H¨ older continuous of index 1−β (it is necessary for induction). ˜ Denote Ωk := {ω : C(ω) ≤ k}. Then

216


Ek

p (n) Ds Xt ≤ k p < ∞.

sup 0≤s≤t≤T

(3.1.46)

Moreover, t t (n+1) (n) (n) H ≤ b(X X ds + − X ) − b(X ) (σ(X ) − σ(X ))dB t t s s s s s 0 0 t t Xs(n) − Xs (n) ≤M ds Xs − Xs ds + C(ω) sβ 0 0 t r σ(Xr ) − σ(Xr(n) ) − σ(Xn ) + σ(Xu(n) ) du dr. + C(ω) (r − u)1+β 0 0 From Lemma 7.1 (NR00), conditions (i’), (x) and from (3.1.44), it follows that σ(Xr ) − σ(Xr(n) ) − σ(Xu ) + σ(Xu(n) ) ≤ C Xr − Xu − Xr(n) + Xu(n) + C Xr(n) − Xr |Xr − Xu | + Xr(n) − Xu(n) 1−β + C Xr − Xu − Xr(n) + Xu(n) . ≤ C Xr(n) − Xr C exp{CGκ } |u − r| Then t (n+1) − Xt ≤ C(ω) Xt 0

(n) (n) X r − X u − Xr + Xu

r

du dr (r − u)1+β t (n) + Xs − Xs s−β ds .

0

0

By similar estimates we obtain t X (n+1) − Xu(n+1) − Xt − Xu t 0

(t − u)1+β t ≤ C(ω) 0

du

du (t − u)1+β

(n) Xs − Xs s−β ds

u

t

r

+ u

t

u

(n) (n) X r − X v − Xr + Xv (r − v)1+β

t (n) (n) (n) Denote ξn1 (t) := Xt − Xt , ξn2 (t) := 0 Xt − Xu − Xt + Xu × (t − u)−1−β du, then t t 1 (t) ≤ C(ω) ξn1 (s)s−β ds + ξn2 (s)ds , ξn+1 0

0

du dr .


t 2 1 −β ξn+1 (t) ≤ C(ω) ξn (s)s

t

s

ξn2 (s)

+ 0

0

217

s

du ds 1+β 0 0 (t − u) t du ds ≤ C(ω) s−β (t − s)−β (ξn1 (s) + ξn2 (s))ds. (t − u)1+β 0

t Let ξn (t) = ξn1 (t)+ξn2 (t), then ξn+1 (t) ≤ C(ω)t2β 0 s−2β (t − s)−2β ξn (s)ds. Denote C4 (ω) := sup |ξ0 (s)|. Then it is easy to obtain that 0≤s≤T

ξn (t) ≤ (C(ω))n C4 (ω)

Γ (1 − 2β)n+1 n(1−2β) t . Γ (n(1 − 2β))

(3.1.47)

p

Hence, Ek sup0≤t≤T |ξn (t)| ≤ Ck (n+1)p < ∞ for some C > 0 and any p > 0, where Ωk = {ω : C(ω) p ≤ k, C4 (ω) ≤ k}. Finally, we obtain from (3.1.47) that (n) Ek sup Xt − Xt → 0, n → ∞. Together with (3.1.46) it means that 0≤t≤T

X(t) ∈ D1,∞,loc . 2. Let equation (3.1.28) be semilinear, i.e. it has a form

t

Xt = X0 +

t

Xs dBsH ,

b(Xs )ds + σ 0

0

where b satisfies conditions (i’), (ii’), (x), X0 satisfies condition (xii). Then (n+1) ˜ Ds Xt ≤ C1 (ω) + C(ω)

|Ds Xrn | dr β s (r − s) t + C(ω) |σ| t

s

s

r

|Ds Xrn − Ds Xun | du dr, (r − u)1+β

Ds Xrn+1 − Ds Xun+1 r r z |Ds Xzn | |Ds Xzn − Ds Xvn | ≤ C(ω) dz + C(ω)|σ| dv dz, β (z − v)1+β u (z − u) u u or in terms of ϕ1n (t) and ϕ2n (t) from Part 1 of the proof, t t ϕ1n (t) ≤ C˜1 (ω) + C(ω) ϕ1n (r)(r − s)−β dr + C(ω) |σ| ϕ2n (r)dr, s s t t ϕ1n (v)(t − v)−2β dv + C(ω) ϕ2n (v)(t − v)−β dv. ϕ2n (t) ≤ C(ω) s

s

Repeating the same estimates as in Part 1 but with other constants, we obtain ϕ1n (t) ≤ C˜1 (ω) exp{CGκ (t − s)}.

218


p 1 Evidently, E ϕ1n (t) ≤ Cp < ∞ since now, of course, µ = 1, κ = 1−2β , 3 and for H > 4 the coefficient β > 1 − H can be chosen in such a way that 1 1−2β < 2. Moreover, in this, semilinear, case, (n+1) − Xt Xt

t

≤ C(ω) 0

t

0

|Xsn − Xs | ds + C(ω) sβ

t 0

r

(n) (n) Xr − Xr − Xu + Xu

0

(r − u)1+β

du dr,

(n+1) (n+1) − Xt − Xu + Xu Xt

du (t − u)1+β t t du (n) −β C(ω) − X ≤ X s ds s s 1+β (t − u) 0 u t r Xr − Xv − Xr(n) + Xv(n) dv dr + C(ω) 1+β (r − v) u u t (n) ≤ C(ω) Xs − Xs s−β (t − s)−β ds 0 t s Xs − Xv − Xs(n) + Xv(n) (t − s)−β ds, + C(ω) (s − v)1+β 0 0

or

t

ξn1 (s)s−β ds

t

≤ C(ω) + C(ω) ξn2 (s)ds, 0 0 t t 2 (t) ≤ C(ω) ξn1 (s)s−β (t − s)−β ds + C(ω) ξn2 (s)(t − s)−β ds, ξn+1 1 ξn+1 (t)

0

0

t Xt(n) +Xu(n) −Xt +Xu 2 (n) 1 where ξn (t) = Xt − Xt , ξn (t) = 0 du. (t−u)1+β Repeating the same estimates as in Part 1, but with other constants, we obtain: C n+1 Gn+1 C˜4 (ω) , sup ξn (t) ≤ Γ (n(1 − 2β)) 0≤t≤T

where ξn (t) = ξn1 (t) + ξn2 (t),

C˜4 (ω) = sup |ξ0 (t)| = sup 0≤t≤T

0≤t≤T

|X0 | + |Xt | + 0

t

|Xt − Xr | dr . (t − r)1+β

3 According to Corollary 3.1.7, E C˜4p (ω) < ∞ for any p ≥ p 1 if H > 4 . Clearly, (n) EGp < ∞ for any p ≥ 1. Therefore, E sup Xt − Xt ≤ Cp < ∞ and we 0≤t≤T

obtain the proof.


219

Remark 3.1.15. It is easy to see that under conditions (i’)–(ii’) and (x)–(xii) the derivative Ds Xt satisfies the equation t t σ (Xr )Ds Xr dBr + b (Xr )Ds Xr dr. (3.1.48) Ds Xt = σ(Xs ) + s

s

Remark 3.1.16. For differentiability and local differentiability of the solutions of SDE involving fBm see also (NS05) and (MS07b). Smoothness of the Functionals of the Solution We consider equation (3.1.28) and suppose that the coefficients b and σ satisfy the conditions of Theorem 3.1.10 and the condition (xiii) b, σ ∈ C 1 (R). Note that under these conditions equation (3.1.28) has a pathwise solution. Let Xt be any solution of (3.1.28) and the function F ∈ C 2 (R). Then for any fixed T T > 0 0 |F (Xs )b(Xs )| ds < ∞ a.s. Suppose that the process F (Xs )σ(Xs ) ∈ D1,2 (|H|) and a.s.

T

0

T

|Ds (F (Xu )σ(Xu ))| |u − s|

2α−1

du ds < ∞.

s

According to the Itˆ o formula (2.7.3) and equality (2.4.2), it holds that

t F (Xs )b(Xs )ds + F (Xs )σ(Xs )dBsH 0 0 t t = F (X0 ) + F (Xs )b(Xs )ds + F (Xs )σ(Xs )δBsH 0 0 t t 2α−1 + CH Ds (F (Xu )σ(Xu )) |u − s| du ds. t

F (Xt ) = F (X0 ) +

0

(3.1.49)

s

By using this equality, we can prove the following result. Denote t t t H b (Xu )du + σ (Xu )dBu , 0 ≤ s < t ≤ T. εs := exp s

s

Theorem 3.1.17. Let the conditions of Theorem 3.1.10, condition (xiii) and the following conditions hold: T (xiv) E 0 |F (Xt )b(Xt )| dt < ∞, the function f (s) := EF (Xs )b(Xs ) is continuous on [0, T ]; (xv) F (Xs )σ(Xs ) ∈ D1,2 (|H|) and

T

E 0

s

T

|Ds (F (Xu )σ(Xu ))| |u − s|

2α−1

du ds < ∞.

220


Then the function ϕ(t) := EF (Xt ) is differentiable in t and ϕ (t) = EF (Xt )b(Xt ) t 2α−1 + 2αHE (F (Xs )σ (Xs )) σ(Xs ) |s − t| (εs0 )−1 ds · εt0 . 0

Proof. From the Itˆ o formula (3.1.49) and conditions (xiv)–(xv) it follows that t ϕ(t) = EF (X0 ) + EF (Xs )b(Xs )ds 0 t t 2α−1 EDs (F (Xu )σ(Xu )) |u − s| du ds, + CH 0

CH = 2αH.

s

Note t that the mathematical expectation of the divergence operator E 0 F (Xs )σ(Xs )δBsH = 0. Therefore, under (xiv) and (xv) we can differentiate ϕ and obtain that t 2α−1 EDs (F (Xt )σ(Xt )) |t − s| ds. ϕ (t) = EF (Xt )b(Xt ) + 2αH 0

Further, from the chain rule, Theorem 3.1.14 and Remark 3.1.15 Ds (F (Xt )σ(Xt )) = (F (Xs )σ(Xs )) Ds Xt , the derivative Ds Xt exists and satisfies linear differential equation (3.1.48), whence Ds Xt = σ(Xs )εts 1 {s ≤ t} . Therefore ϕ (t) = EF (Xt )b(Xt ) t 2α−1 + 2αHE (F (Xs )σ (Xs )) σ(Xs ) |s − t| (εs0 )−1 ds · εt0 . 0

3.1.5 Semilinear Stochastic Differential Equations Involving Forward Integral w.r.t. fBm León and Tudor in their paper (LT02) established the existence of a global solution of a semilinear stochastic differential equation with forward integrals (for the definition and properties of forward integral see Section 2.4). Let p > 1 and γ ∈ (0, 1). A process u ∈ D1,p (|H|) belongs to L1,p γ if upL1,p := E(upL 1 [0,T ] ) + E(DupL 1 [0,T ]2 ) < ∞. γ

γ

γ

It follows from (AN02) that L1,p γ ⊂ Dom(δH ) for any 0 ≤ γ ≤ H.

(3.1.50)


221

The next statement from (LT02) establishes the relationship between the forward integral (understood in the sense of ucp-convergence) and the diver t gence operator that we denote here as 0 us δBsH (for P-convergence such a statement was proved by (AN02), see (2.4.1)). Here something like condition (3.1.50) is needed. Theorem 3.1.18 ((LT02)). 1)Let {ut , t ∈ [0, T ]} be a stochastic process, u ∈ Lγ1,2 for some 1/2 < γ < H and the trace condition holds, T t |Ds ur ||r − s|2α−1 dsdr < ∞ a.s. 0

0

t t Then both the integrals, 0 us dBsH,− and 0 us δBsH , exist for any t ∈ [0, T ] and t t T t us dBsH,− = us δBsH + 2αH Ds ur |r − s|2α−1 dsdr. 0

0

0

0

2) Now consider the semilinear stochastic differential equation t t Xt = X0 + b(s, Xs )ds + σs Xs dBsH,− , t ∈ [0, T ], 0

(3.1.51)

0

where the coefficients b : Ω × [0, T ] × R → R and σ : Ω × [0, T ] → R are measurable, X0 is random variable, b and σ satisfy the following assumptions: (xvi) For all ω ∈ Ω, t ∈ [0, T ] and x, y ∈ R |b(ω, t, x) − b(ω, t, y)| ≤ κ(ω)|x − y|, |b(ω, t, 0)| ≤ κ(ω) for some random variable κ(ω). (xvii) σ is forward integrable and there is ε0 > 0 such that r T r H limc→∞ sup0 0 r t r H limε→0 P sup0≤t≤T 0 ( 0 σs ε−1 (B(s+ε)∧T − BsH )ds − 0 σs dBsH,− )

H ×σr ε−1 (B(r+ε)∧T − BrH )dr > c = 0. Also, denote by A the class of all processes X such that (σX) is forward integrable and for any c > 0 and t ∈ [0, T ] s s H limε→0 limη→0 P 0 σs Xs exp{− 0 σr ε−1 (B(r+ε)∧T − BrH )dr}

H H ×(η −1 (B(s+η)∧T − BsH ) − ε−1 (B(s+ε)∧T − BsH ))ds > c = 0.

222


Theorem 3.1.19 ((LT02)). Under assumptions (xvi)–(xviii) equation (3.1.51) has a unique solution in the class A that is given by the unique solution of the equation t

t t H,− X0 + σs dBs exp σs dBsH,− b(u, Xu )du, t ∈ [0, T ]. Xt = exp 0

0

u

Here are some classes of coefficients satisfying assumptions (xvii) and (xviii). Example 3.1.20 ((LT02)). Assume that the stochastic process {σt , t ∈ [0, T ]} satisfies the following conditions: (xix) σ ∈ Lγ1,2 for some 1/2 < γ < H, and for some t0 ∈ [0, T ]

T

1

|Ds σt0 | γ ds

E 0

2γ

< ∞;

(xx) there exists β such that for all s, t ∈ [0, T ] E|σt − σs | ≤ c|s − t|β/2 and

E

T

1

|Du (σs − σt )| γ du

2γ

≤ c|s − t|β ;

0

T (xxi) E|σt | < ∞ and E( 0 |Dr σt ||t − r|2α−1 dr)2 < ∞, t ∈ [0, T ]; (xxii) there exists µ ∈ (0, H) such that (a) limc→∞ sup0 0 and 2

θε = ε

−1−µ+H

0

T

0

r

(s+εµ )∧T

|Du σs ||s − u|2α−1 du ds

2

1/2 dr

,

(s−εµ )∨0

P

(b) θε → 0 as ε → 0, (c) β > 2(1 − H + µ) and H − µ > 1/2 ∨ µ. Then σ satisfies assumptions (xvi) and (xvii). Example 3.1.21. Let {σt , t ∈ [0, T ]} be an absolutely continuous process of the form t σ˙ s ds, t ∈ [0, T ], σt = σ0 = 0

Lγ1,2

for some 1/2 < γ < H, and σ satisfies conditions (xxi) and with σ0 , σ˙ ∈ (xxii). Then σ satisfies assumptions (xvii) and (xviii).


223

3.1.6 Existence and Uniqueness of Solutions of SDE with Two-Parameter Fractional Brownian Field In this subsection we use the notations introduced in Subsection 2.2.4. We continue with the estimates of the two-parameter generalized Lebesgue–Stieltjes integrals (the first result in this direction was formulated in Lemma 2.2.16), and use these estimates to obtain the conditions of existence and uniqueness of solution of SDE involving the two-parameter fBm. The estimates of the norms of integrals on the whole duplicate the corresponding estimates from Lemmas 3.1.1–3.1.2 and Theorem 3.1.3, but are much more technical. Therefore we omit the proofs. For details, see (MisIl03). Another approach to SDEs involving a two-parameter fractional Brownian field was developed in (TT03). Denote PT = [0, T1 ] × [0, T2 ] ⊂ R2+ and introduce the following norm on the space W0β1 , β2 (PT ): ||f ||β1 ,β2 ,λ1 ,λ2 := sup e−λ1 t1 −λ2 t2 ϕβf 1 , β2 (t). t∈PT

Also, recall that ||f || = sup |f (t)|. t∈PT

Lemma 3.1.22. Let the function σ : PT × R → R and satisfy the following conditions: (xxiii) 1) σ ∈ C 3 (PT × R); 2) ∃ C > 0 such that |Dσ(t, x)| ≤ C, where the symbol D stands for any differentiation that is possible according to item 1) and (t, x) ∈ PT ×R; 3) |σ(r, 0)| ≤ C; Also, let f ∈ W0β1 , β2 (PT ), g ∈ W11−β1 ,1−β2 (PT ), for some 0 < βi < i = 1, 2. Then the following statements hold:

1 2,

1) ||σ(·, f (·))||0,β1 , β2 ≤ Cβ1 ,β2 ,T (1 + ||f ||)(1 + ||f ||0,β1 , β2 )2 ; 2) The generalized Lebesgue–Stieltjes integral (σ) Gt (f ) := σ(s, fs )dgs PT

exists, belongs to the spaces C 1−β1 ,1−β2 and W0β1 , β2 (PT ) and admits in these spaces the following estimates: (a) ||G(σ) (f )||1−β1 , 1−β2 ≤ Cβ1 ,β2 ,T Λ1−β1 ,1−β2 (g)(1 + ||f ||) ×(1 + ||f ||0,β1 , β2 )2 ; 1 2 λ−1+2β (b) ||G(σ) (f )||β1 , β2 , λ1 , λ2 ≤ Cβ1 ,β2 ,T Λ1−β1 ,1−β2 (g)λ−1+2β 1 2 (3.1.52) ×(1 + ||f ||2 ) 1 + ||f ||β1 ,β2 ,λ1 , λ2 + ||f ||2β ,β , λ1 , λ2 . 1

2

2

2

224


Here Cβ1 ,β2 ,T depends only on β1 , β2 and T .

Remark 3.1.23. All the estimates hold for fs = gs = BsH1 H2 with Hi ∈ 12 , 1 , i = 1, 2. Remark 3.1.24. Let the function σ be bounded, f1 (x) = f (x) + C0 , where C0 ∈ R be some constant. Then ||G(σ) (f1 )||β1 , β2 , λ1 , λ2 can be estimated by the right-hand side of (3.1.52), i.e. this estimate does not depend on C0 . Lemma 3.1.25. Let the function σ satisfy the condition (xxv) σ ∈ C 3 (PT ) × C 5 (R), and conditions (xxiii), 2) and 3) hold. Also, let f, h ∈ W0β1 , β2 and g ∈ W11−β1 ,1−β2 for some 0 < βi < 12 and i = 1, 2. Then ||G(σ) (f ) − G(σ) (h)||β1 , β2 , λ1 , λ2 ≤

Cβ1 ,β2 ,T Λ1−β1 ,1−β2 (g) 1 λ1−2β 1

2 λ1−2β 2

(1 + ||f || + ||h||)2

×(1 + ||f ||0,β1 , β2 + ||h||0,β1 , β2 )2 ||f − h||β1 ,β2 ,λ1 , λ2 + ||f − h||2β

λ1 1 ,β2 , 2

,

λ2 2

,

λi ≥ 1, i = 1, 2. Lemma 3.1.26. 1) Let the function b = b(t, x) : PT × R → R be of linear growth: |b(t, x)| ≤ C(1 + |x|). Also, let f ∈ W0β1 ,β2 (PT ). Then the integral (b) Ft (f ) := b(s, f (s))ds ∈ C 1 (PT ) for t ∈ PT and Pt

F (b) (f )β1 ,β2 ,λ1 ,λ2 ≤

Cβ1 ,β2 ,T 1−β1 1−β2 λ1 λ2

(1 + f β1 ,β2 ,λ1 ,λ2 ) .

2) If the function b is bounded, we have the same situation as described in Remark 3.1.24. 3) If f, h ∈ W0β1 ,β2 (PT ) and ||f || ≤ N , ||h|| ≤ N , then F (b) (f ) − F (b) (h)β1 ,β2 ,λ1 ,λ2 ≤

Cβ1 ,β2 ,T,N 1 1−β2 λ1−β λ2 1

||f − h||β1 ,β2 ,λ1 ,λ2 ,

λi ≥ 1, i = 1, 2, where Cβ1 ,β2 ,T,N depends on β1 , β2 , T and N . Consider now a stochastic differential equation on the plane, (b) (σ) Xt = X0 + b(s, Xs )ds + σ(s, Xs )dBsH1 ,H2 = X0 + Ft (X) + Gt (X), Pt

Pt

(3.1.53) where t ∈ PT ⊂ R2+ , B H1 ,H2 is the fractional Brownian field with the Hurst indices Hi ∈ ( 12 , 1), σ, b : PT × R → R are measurable bounded functions, σ satisfies conditions (xxv), (xxiii), 2), and the function b(s, x) is continuous in s and Lipschitz in x. The two-parameter process Xt : PT × Ω → R will be called a solution of (3.1.53) if it converts (3.1.53) into identity for a.a. ω ∈ Ω and any t ∈ PT , and

3.2 The Mixed SDE Involving Both the Wiener Process and fBm

225

(σ)

the integral Gt (X) exists for a.a. ω ∈ Ω as the two-parameter generalized Lebesgue–Stieltjes integral. The proof of the main result, stated in the next theorem, relies, in particular, on the boundedness of the coefficients and on Remark 3.1.24. Theorem 3.1.27. Under the conditions mentioned above, SDE (3.1.53) has a unique solution in the class W0β1 ,β2 (PT ) and for a.a. ω ∈ Ω, X ∈ C 1−β1 ,1−β2 for any 1 − Hi < βi < 12 , i = 1, 2.

3.2 The Mixed SDE Involving Both the Wiener Process and fBm Real objects varying in time (climate and weather derivative, prices on the stock market etc.) can have a component with a long memory (that is modeled by fBm with H ∈ (1/2, 1)) and also a component without memory (that is modeled by a Wiener process). Therefore, it is natural to consider stochastic differential equations involving both Brownian and fractional Brownian motions. We refer to such equations as mixed stochastic differential equations(and, correspondingly, to such models as mixed models). The conditions of existence of a local solution of the mixed SDE were formulated in Theorem 3.1.9. Of course, we would like to establish the conditions of the existence of a global solution. We start with the semilinear SDE. 3.2.1 The Existence and Uniqueness of the Solution of the Mixed Semilinear SDE Consider an SDE of the form t t t b(s, Xs )ds + σ1 Xs dWs + σ2 Xs dBsH , t ∈ [0, T ], (3.2.1) Xt = X0 + 0

0

0

where X0 is an F0 -measurable random variable, σ1 and σ2 are real numbers, {Wt , Ft , t ∈ [0, T ]} and {BtH , Ft , t ∈ [0, T ]} are a Wiener process and fBm, correspondingly, on the same probability space (Ω, F, Ft , t ∈ [0, T ]), without any suppositions on their dependence. Theorem 3.2.1. Let the function b satisfy Lipschitz and linear growth conditions in x: |b(t, x) − b(t, y)| ≤ L|x − y|, |b(t, x)| ≤ L(1 + |x|),

L > 0, x, y ∈ R,

and is continuous in both variables, b ∈ C([0, T ] × R). Then there exists the unique solution {Xt , t ∈ [0, T ]} of equation (3.2.1), and the trajectories of X a.s. belong to C 1/2− [0, T ].

226


Proof. First, we use Theorem 3.1.9 and construct a local solution. In this order we consider an auxiliary system of partial differential equations (3.1.27) that now acquires the following form: ∂h ∂Zj (Y, (Z1 , Z2 )) = σj h(Y, (Z1 , Z2 )), j = 1, 2, h(Y0 , 0, 0) = X0 . The solution of this system has the form h(Y, (Z1 , Z2 )) = (Y − Y0 + X0 ) exp{σ1 Z1 + σ2 Z2 },

(3.2.2)

where Z1 (t) = Wt , Z2 (t) = BtH . Now we try to construct the local solution Xt of equation (3.2.1) in the form of Xt = h(Yt , (Z1 (t), Z2 (t))), where the trajectories of Y a.s. belong to C 1 [0, T ], Y (0) = Y0 be some F0 -measurable random variable. Applying the Itˆ o formula (2.7.2) from Remark 2.7.4, we obtain that 2 ∂h (Yt , Z1 (t), Z2 (t))dZi (t) dXt = ∂Z i i=1

∂h 1 (Yt , Z1 (t), Z2 (t))Yt dt + σ12 h(Yt , Z1 (t), Z2 (t))dt. (3.2.3) ∂Y 2 Comparing (3.2.1) and (3.2.3), we get the ordinary differential equation for the process Y : Yt = (c1 (t))−1 b(t, (Yt − Y0 + X0 )c1 (t)) − 12 σ12 (Yt − Y0 + X0 ) =: f (t, Y ), Y (0) = Y0 , (3.2.4) where c1 (t) = exp{σ1 Z1 (t) + σ2 Z2 (t)}. Further we fix ω ∈ Ω and put for this ω L1 (T ) := max0≤t≤T (c1 (t))−1 > 0, L2 (T ) := max0≤t≤T c1 (t) > 0, D1 = LL1 (T ), D2 = L + 12 σ12 . Then for t ≤ a0 and |Yt − Y0 | ≤ b0 with some a0 , b0 > 0 we have that M := max0≤t≤T |f (t, Yt )| ≤ L(L1 (T ) + b0 + |X0 |) + 12 σ12 (b0 + |X0 |) = D1 + D2 (b0 + |X0 |) =: M0 , and by the Picard theorem, the solution of equation (3.2.4) exists and is unique on the interval [0, l(0) ], where l(0) := min(a0 , b0 /M ) ≥ min(a0 , b0 /M0 ) =: t0 ; consequently, the solution exists on [0, t0 ]. By using (3.2.2), the solution at the point t0 can be bounded by |h(Yt0 , Z1 (t0 ), Z2 (t0 ))| ≤ |Yt0 − Y0 + X0 |L2 (T ) ≤ (b0 + |X0 |)L2 (T ). Evidently, the trajectories of the solution belong to C 1/2− [0, t0 ], since Y is continuously differentiable (recall that b ∈ C([0, T ] × R)) and exp{σ1 Z1 (t) + σ2 Z2 (t)} = exp{σ1 Wt + σ2 BtH } ∈ C 1/2− [0, t0 ]. Now we want to extend the solution for [0, T ]. The value Xt0 will be the (1) new initial value X0 , and +

(1)

|X0 | ≤ (b0 + |X0 |)L2 (T ).


227

Now, for |t − t0 | ≤ a1 , |Yt − Yt0 | ≤ b1 , for some a1 and b1 > 0, the solution of (3.2.4) exists on the interval [t0 , t1 ], where t1 − t0 = min(a1 , b1 /M1 ) with M1 = D1 + D2 (b1 + (b0 + |X0 |)L2 (T )). In the nth step of this procedure of the extension of the solution we obtain tn − tn−1 = min(an , bn /Mn ) with n−1 (T ) + |X0 |Ln+1 (T )) and the solution Mn = D1 + D2 (bn + k=0 bn−1−k Lk+1 2 2 exists on [tn−1 , tn ]. Now we have two possibilities: if |X0 | ≤ 1 we can put bn = 1, 0 ≤ k ≤ n n (T )))−1 and bn /Mn = (D1 + D2 ( k=0 Lk2 (T ) + |X0 |Ln+1 2 Ln+2 (T )−1

≥ (D1 + D2 L2 2 (T )−1 )−1 =: Kn . If |X0 | > 1 then we put bk = |X0 |, 0 ≤ k ≤ n and in this case also bn /Mn ≥ Kn . For both the cases put an = Kn , n ak . n ≥ 0 and tn − tn−1 = an , tn = k=0 (a) Let L2 (T ) ≤ 1. Then the series an diverges, so, there exists only a n≥0

finite number of aforementioned steps and we obtain the existence of a solution on the whole interval [0, T ]. an converges, possibly, its sum (b) Let L2 (T ) > 1. Then the series n≥0

S ≤ T and we obtain the existence of a solution on [0, S). Therefore, we have established the existence of a finite solution on [0, S2 ]. By the same method we can extend it on [ S2 , S] with the same step S2 , since the size of step does not depend on the initial value X0 . So, we can extend the solution with the step S2 on the whole [0, T ]. The uniqueness of the solution follows from Theorem 3.1.9. It follows from its construction (see (3.2.2) and (3.2.4)), that the trajectories

of solution belong to C 1/2− [0, T ]. 3.2.2 The Existence and Uniqueness of the Solution of the Mixed SDE for fBm with H ∈ (3/4, 1) Now we consider a mixed SDE without any semilinear restrictions but only for H ∈ (3/4, 1). Existence and Uniqueness of Solution of Mixed SDE for fBm with H ∈ (3/4, 1) and with Stabilizing Term We follow here the approach of (MP07). Consider the following mixed SDE: Xt = X 0 + 0

t

t a(s, Xs )ds + b(s, Xs )dWs 0 t t + c(s, Xs )dBsH + ε c(s, Xs )dVs , t ∈ [0, T ], 0

(3.2.5)

0

where a, b, c : [0, T ] × R → R are measurable functions, V, W are independent Wiener processes, ε > 0 and B H is independent of W and V fractional

228


Brownian motion with H ∈ (3/4, 1), X0 is independent of W, B H and V . The t integral ε 0 c(s, Xs )dVs will play the role of the stabilizing term. It permits us to establish the existence and uniqueness of the solution of (3.2.5), adapted to the filtration Ft , t ≥ 0, where Ft = σ{X0 , Ws , (εVs + BsH )|s ∈ [0, t]}.

(3.2.6)

The results are valid also for the case when b = 0. If ε = 0 and b = 0, we obtain equation (3.1.6) with g = B H , whose existence and uniqueness conditions were formulated in Theorem 3.1.4. As we shall see, the stabilizing term permits us to avoid the smoothness condition on c, for example, the existence and H¨ older properties of ∂x c(s, x). The main result that we use in the proof was stated by Cheridito (Che01b). For the completeness of exposition we shall present it here. Its proof originated in the papers (HH76) and (Hit68). Proposition 3.2.2. 1. Let {Wt , t ∈ [0, T ]} be a Wiener process, {BtH , t ∈ [0, T ]} be an independent fBm with H ∈ (3/4, 1), γ ∈ R \ {0}, MtH,γ := Bt + γBtH ,

t ∈ [0, T ],

H,γ

with its own filtration {FtM , t ∈ [0, T ]}. H,γ Then {MtH,γ , FtM , t ∈ [0, T ]} is equivalent to Brownian motion; consequently it is a semimartingale. 2. There exists a unique real-valued Volterra kernel h = hγ ∈ L2 [0, T ]2 such that t s H,γ h(s, u)dMuH,γ ds, t ∈ [0, T ] Bt := Mt − 0

0

is a Brownian motion. Furthermore, t s H,γ Mt = Bt − r(s, u)dBu ds, 0

t ∈ [0, T ],

(3.2.7)

0

where r = rγ ∈ L2 [0, T ]2 . H, 1

As a consequence, the process NtH,ε := BtH + εVt = ε(Vt + 1ε BtH ) = εMt ε can be represented as t s NtH,ε = εVt + εrε (s, u)dVu ds, (3.2.8) 0

0

where V is some Wiener process with respect to filtration Ft := σ{εVs + BsH , s ∈ [0, t]} and, from the independence of V, W, B H and X0 , it is a Wiener process w.r.t. {Ft , t ∈ [0, T ]}. Using (3.2.26), we can rewrite the equation (3.2.4) in the form


Xt = X0 +

t

t

a(s, Xs )ds + b(s, Xs )dWs 0 0 t t +ε c(s, Xs )dVs + c(s, Xs ) 0

229

0

s

εrε (s, u)dVu ds.

(3.2.9)

0

The drift coefficient of equation (3.2.9) equals a(s, x) + c(s, x, ω), s where c(s, x, ω) = c(s, x) 0 εrε (s, u)dVu . Evidently, the random variable s εrε (s, u)dVu is not bounded, but we can consider the sequence of stopping 0 t s times τ M = inf{t ∈ [0, T ] : 0 ( 0 εrε (s, u)dVu )2 ds > M } ∧ T, and consider the sequence of corresponding stopped equations. The existence and uniqueness of the solutions of these equations can be established by standard methods and then it is easy to pass to the limit when M → ∞. Finally, we obtain the following result (note that in this section we begin with the new numeration of the conditions). Theorem 3.2.3. Let the following conditions hold: (i) The functions |a(s, 0)| + |b(s, 0)| + |c(s, 0)| ≤ L, s ∈ [0, T ] and |a(s, x)| + |b(s, x)| + |c(s, x)| ≤ L(1 + |x|), for some constant L > 0; (ii) there exists an increasing function l(s) : [0, T ] → R such that ∀x, y ∈ R |a(s, x) − a(s, y)| + |b(s, x) − b(s, y)| + |c(s, x) − c(s, y)| ≤ l(s)|x − y|; (iii) the initial value X0 is square-integrable. Then equation (3.2.9), and consequently equation (3.2.5), has on [0, T ] the unique Ft -adapted solution Xt . The Existence and Uniqueness of the Solution of the Mixed SDE Involving fBm with H ∈ (3/4, 1) as the Limit Result for the Equations with the Stabilizing Term Now we want to pass to the limit as ε → 0 in equation (3.2.5). Let ε = 1/N, N ≥ 1, and consider the sequence of the equations with the stabilizing term t t XtN = X0 + a(s, XsN )dt + b(s, XsN )dWt 0 0 (3.2.10) t 1 t N H N + c(s, Xs )dBs + c(s, Xs )dVs , t ∈ [0, T ]. N 0 0 Let the coefficients a, b, c and X0 satisfy conditions (i), (ii) and (iii). Then, according to Theorem 3.2.3, equation (3.2.10) has a unique strong solution, say {XtN , t ∈ [0, T ]}. Evidently, the solutions are adapted to different filtrations FtN = σ{X0 , Ws , (N −1 Vs + BsH ), s ∈ [0, t]}. The aim of this section is to establish the conditions of existence and uniqueness of the solution of the limit mixed equation

230


Xt = X0 +

t

a(s, Xs )ds +

t

0

0

t

t ∈ [0, T ].

c(s, Xs )dBsH ,

b(s, Xs )dWs + 0

(3.2.11) Let the coefficients of equation (3.2.11) satisfy assumption (iii) and the following ones: there exist such constants B, L, M > 0, γ ∈ (1 − H, 1) and κ ∈ (3/2 − H, 1) that (iv) all the coefficients are bounded: |a(s, x)| + |b(s, x)| + |c(s, x)| ≤ L, ∀s ∈ [0, T ], ∀x ∈ R; (v) all the coefficients are Lipschitz in x: |a(t, x) − a(t, y)| + |(b(t, x) − b(t, y)| + |c(t, x) − c(t, y)| ≤ L|x − y|, ∀t ∈ [0, T ], ∀x, y ∈ R, (vi) the x-derivative of the function c exists and is Hölder continuous in t: ∀s, t ∈ [0, T ], ∀x ∈ R |c(s, x) − c(t, x)| + |∂x c(s, x) − ∂x c(t, x)| ≤ L|s − t|γ . (vii) the x-derivative of the function c is Hölder continuous in x: |∂x c(t, x) − ∂x c(t, y)| ≤ L|x − y|κ , for ∀t ∈ [0, T ], ∀x, y ∈ R . Remark 3.2.4. Note that for H ∈ [3/4, 1) 3/2 − H > 1/H − 1, so condition (vii) is more restrictive than the corresponding condition (ii) used in Theorem 3.1.4. In general, this last group of conditions is evidently more strong than conditions (i)–(ii) of Theorem 3.2.3. Now consider for β < (1/2 ∧ γ ∧ κ/2 ∧ (κ − 12 )) some “stochastic analog” of the functional space of Besov type: W β [0, T ] := {Y = Yt (ω)|(t, ω) ∈ [0, T ] × Ω, ||Y ||β < ∞} with the norm

Y β := sup t∈[0,T ]

2

E(Yt ) + E 0

t

|Yt − Ys | ds (t − s)1+β

2 ,

and prove that the solution of SDE (3.2.10) belongs to this space for any N > 1. We shall denote different constants as C if they do not depend on N and it is unimportant to the stated results. First of all we prove the H¨ older continuity of the solution of equation (3.2.10), by using (1.17.1) and (1.17.2). Theorem 3.2.5. For any δ ∈ (0, 1/2) the solution of equation (3.2.10) is H¨ older continuous with parameter 1/2 − δ.


231

Proof. Consider |XrN − XzN | for 0 < z < r < T : r r 1 r |XrN − XzN | ≤ a(u, Xu )du + b(u, Xu )dWu + c(u, Xu )dVu N z z z r C c b + c(u, Xu )dBuH ≤ L(r − z) + Cξr,δ |r − z|1/2−δ + ξr,δ |r − z|1/2−δ N z r |c(u, XuN )|du H + Λ1−β (B ) uβ z r u N |c(u, Xu ) − c(v, XvN )| + Λ1−β (B H ) dv du (u − v)1+β z z r u |XuN − XvN | dv du, ≤ Cr (ω)(r − z)1/2−δ + Cr (ω) (u − v)1+β z z where

b c ∨ ξt,δ ∨ 1), Ct (ω) := C(Λ1−β (B H ) ∨ ξt,δ

Ct (ω)

and are defined by (1.17.2), ≤ ments of any order. Therefore, for δ < 1/2 − β we have that

b ξt,δ

c ξt,δ

φr,s := s

r

CT (ω)

and

CT (ω)

(3.2.12) has the mo-

r |XrN − XzN | dz ≤ C (ω) (r − z)−1/2−δ−β dz r (r − z)1+β s r u r 1 |XuN − XvN | dvdudz + 1+β (u − v)1+β s (r − z) z z r (r − u)−β φu,s du . ≤ Cr (ω) (r − u)1/2−β−δ + s

From the modified Gronwall inequality (Lemma 7.6 (NR00)) it follows that 1

φr,s ≤ Cr (ω)(r − s)1/2−β−δ exp{Cr (ω) 1−β }. Return to |XrN − XzN |: |XrN − XzN | ≤ Cr (ω)(r − z)1/2−δ r 1 r (ω)(r − z)1/2−δ , + Cr (ω) exp{Cr (ω) 1−β } (v − z)1/2−β−δ dv ≤ C z

r (ω) = C (ω) exp{C (ω) 1−β }, and the theorem is proved for where C r r 0 < δ < 1/2 − β, and consequently for 0 < δ < 1/2. 1

t (ω). It also has moments Introduce the random variable C(ω) := sup C 0≤t≤T

of any order. Now we want to prove that the solution of (3.2.10) belongs to the space {W β [0, T ], · β } for all N > 1.

232


Theorem 3.2.6. Under assumptions (iii)–(vi) the solution of equation (3.2.10) belongs to the space W β [0, T ] of Besov type with norm · β for all N > 1 and any β < (1/2 ∧ γ ∧ κ/2 ∧ κ − 12 ). Proof. In order to prove the statement of this theorem, we want to estimate AN 1 (t)

+

AN 2 (t)

:=

E(XtN )2

t

+E 0

|XtN − XsN | ds (t − s)1+β

2 .

First, for AN 1 (t) we have that E(XtN )2 ≤ 5E(X0 )2 + 5E

0

t

a(s, XsN )ds

2 + 5E

t

b(s, XsN )dWs

2

0

t 1 t + 5E( c(s, XsN )dBsH )2 + 5E( c(s, XsN )dVs )2 . (3.2.13) N 0 0 2 t Evidently, E 0 a(s, XsN )ds ≤ L2 T 2 , 2 2 2 t t E 0 b(s, XsN )dWs ≤ L2 T, E N1 0 c(s, XsN )dVs ≤ LN 2T ≤ L2 T . Further, for δ < 1/2 − β we have that E

2 2 t c(s, X N ) s c(s, XsN )dBsH ≤ E C (ω) ds β s 0 0 t 2 t s 2 |c(s, XsN ) − c(u, XuN )| L 2 ≤ CE C (ω) t du ds ds + 1+β 2β (s − u) s 0 0 0 ⎞ ⎞

2 t s L(s − u)γ + LC(ω)(s − u)1/2−δ + du ds ⎠⎠ (s − u)1+β 0 0 t

2 2 (ω)C 2 (ω))T 3−2β−2δ ) ≤ C(EC (ω)(L2 T 2−2β + L2 T 2(1−β+γ)) + L2 E(C

with C(ω) = Λ1−β (B H ). From all these estimates it follows that AN 1 (t) < ∞. (t). We have that Consider now AN 2

2 a(u, XuN )du| ≤ 4E ds (t − s)1+β 0

2

2 t t t t | s b(u, XuN )dWu | | s c(u, XuN )dVu | −2 + 4E ds + 4N E ds (t − s)1+β (t − s)1+β 0 0

2 t t | s c(u, XuN )dBuH | + 4E ds . (3.2.14) (t − s)1+β 0 AN 2 (t)

Evidently,

t

|

t s


t

E

|

t

a(u, Xu )du| ds (t − s)1+β

2 ≤ CL2 t2−2β .

s

0

233

Now, let ρ ∈ (β, 1/2), then we have the estimate

E

t

|

t

0

b(u, Xu )dWu | s ds (t − s)1+β

2

t

≤ Ct

1−2ρ

t

≤ Ct1−2ρ

t

0

E|

t

b(u, Xu )dWu |2 ds (t − s)2+2β−2ρ s

2

b (u, Xu )du ds ≤ CL2 t1−2β , (3.2.15) (t − s)2+2β−2ρ s

0

and similarly,

E

t

|

t s

0

Now we estimate E N := E

c(u, Xu )dVu | ds (t − s)1+β

t 0

|

t s

2 ≤ CL2 t1−2β .

c(u, Xu )dBuH |(t − s)−1−β ds

2 . Since

t t c(u, Xu )dBuH ≤ C(ω) |c(u, Xu )|(u − s)−β du s

s

t + s

u

N −1−β |c(u, xN ) − c(r, X )|(u − r) dr du ≤ C(ω) u r

s

t t × |c(u, Xu )|(u − s)−β du + s

s

s

u

L(u − r)γ + LC(ω)(u − r)1/2−δ dr du , 1+β (u − r)

we have that for δ < 1/2 − β E N can be bounded by

E

t

C(ω) 0

≤ C(L t

2 2−4β

L(t − s)1−β + L(t − s)1+γ−β + LC(ω)(t − s)3/2−δ−β ds 1+β (t − s)

2

2 (ω)). EC (ω) + L2 t2+2γ−4β EC (ω) + L2 t3−2δ−4β EC (ω)C (3.2.16) 2

2

2

Therefore, AN 2 (t) satisfies the inequality 2

2 2−2β AN + L2 T 1−2β + L2 T 2−4β EC (ω) 2 (t) ≤ C(L T

2 (ω)) < ∞. (3.2.17) + L2 T 2+2γ−4β EC (ω) + L2 T 3−2δ−4β EC (ω)C 2

2

Finally, the statement of our theorem follows from inequalities (3.2.13)– (3.2.17) with sufficiently small δ > 0.

234


Introduce for any R > 1 the stopping time τR by τR := inf{t : Ct (ω) ≥ R} ∧ T,

(3.2.18)

where Ct (ω) is defined by (3.2.12). Evidently, for any ω ∈ Ω there exists R(ω) such that τR = T for all R > R(ω). Define the processes {XτNR ∧t , N ≥ 1, t ∈ [0, T ]} as the solutions of equation (3.2.10) stopped at the moment τR , and prove that they are fundamental in the norm · β of the space W β [0, T ]. Theorem 3.2.7. Under assumptions (iii)–(vi) the sequence N {Xt∧τ , N ≥ 1, t ∈ [0, T ]} of solutions of equations (3.2.10) is fundamental in R the norm · β for any β < (1/2 ∧ γ ∧ κ/2 ∧ κ − 12 ). Proof. Consider N M AN,M (t) + AN,M (t) := E(Xt∧τ − Xt∧τ )2 1 2 R R t 2 N M N M |Xt∧τ − Xt∧τ − Xs∧τ + Xs∧τ | R R R R +E ds (t − s)1+β 0 τR ∧t 2 N M |Xt∧τ − Xt∧τ − XsN + XsM | N M 2 R R = E(Xt∧τ − X ) + E ds . t∧τR R (t − s)1+β 0

First, for AN,M (t) we have the estimate 1 (t) ≤ 4E AN,M 1

τR ∧t

0

2 (a(s, XsN ) − a(s, XsM ))ds

τR ∧t

+ 4E

τR ∧t

0

−

b(s, XsM ))dWs

(c(s, XsN )

−

c(s, XsM ))dBsH

0

+ 4E + 4E

2 (b(s, XsN )

τR ∧t

2

0

c(s, XsN ) c(s, XsM ) − N M

2

dVs

=: 4(I1 + I2 + I3 + I4 ).

t t N M N M Then I1 ≤ CT L2 0 E(Xs∧τ −Xs∧τ )2 ds, I2 ≤ CL2 0 E(Xs∧τ −Xs∧τ )2 ds, R R R R 2 −2 −2 I4 ≤ CL T (N + M ). Now we are in a position to estimate I3 : 2 τR ∧t I3 ≤ 2R E |c(s, XsN ) − c(s, XsM )|s−β ds 0 τR ∧t s +E |c(s, XsN ) − c(s, XsM ) − c(u, XuN ) + c(u, XuM )| 2

0

0

× (s − u)−1−β duds

2

= 2R2 (I4 + I5 ).


235

Further, I4 ≤ CL T 2

1−2β

τR ∧t

E

(XsN

−

XsM )2 ds

2

= CL T

t

AN,M (s)ds. 1

1−2β

0

0

By using Lemma 7.1 (NR00), we estimate I5 as

2 L|XsN − XsM − XuN + XuM | I5 ≤ 3E du ds (s − u)1+β 0 0 τR ∧t s 2 N 2 L |Xs − XsM |(s − u)γ + 3E du ds (s − u)1+β 0 0 τR ∧t s 2 L|XsN − XsM |(|XsN − XuN |κ + |XsM − XuM |κ ) + 3E du ds (s − u)1+β 0 0 τR ∧t

s

= 3(I6 + I7 + I8 ). Here

t

I6 ≤ CT L2

s∧τR

E 0

0

I7 ≤ CT L2 0

I8 ≤ E

τR ∧t

0

s

0

N M |Xs∧τ − Xs∧τ − XuN + XuM | R R du (s − u)1+β

2 ds,

t N M s2(γ−β) E|Xs∧τ − Xs∧τ |2 ds, R R

2 L|XsN − XsM |2R(s − u)κ(1/2−δ) du ds (s − u)1+β t N M ≤ CT L2 R2 sκ−2κδ−2β E|Xs∧τ − Xs∧τ |2 ds, R R 0

where we choose δ in such a way that κ − 2κδ − 2β > 0. It is possible since β < κ − 1/2 so κ − 2β > 1/2 − β > 0. Finally, I5 ≤ C

t

N,M 2(γ−β) 2 κ−2κδ−2β AN,M (s) + (s + CR s )A (s) ds, 2 1

0

and AN,M (t) ≤ CR2 1

t

AN,M (s)ds + CR2 1 0

t

AN,M (s)ds 2 0

+ C(N −2 + M −2 ). Return to AN,M (t). It admits the following estimate: 2

(3.2.19)

236


⎛ ⎝ AN,M (t) ≤ C E 2

τR ∧t

τR ∧t s

0

τR ∧t

τR ∧t

τR ∧t

τR ∧t

s

+E

0

(b(u, XuN ) − b(u, XuM ))dWu ds (t − s)1+β

2

2

2 (c(u, XuN ) − c(u, XuM ))dBuH ds (t − s)1+β 2 ⎞ τR ∧t c(u,XuN ) c(u,XuM ) ( − )dV u N M s ds ⎠ (t − s)1+β

s

+E 0

(a(u, XuN ) − a(u, XuM ))du ds (t − s)1+β

τR ∧t

+E 0

= C(I9 + I10 + I11 + I12 ). Further, for β < ρ < 1/2

I9 ≤ CT

τR ∧t

E

(t − s)

τR ∧t

L2 |XuN − XuM |2 du ds (t − s)2+2β−2ρ 0 t t 2 N M 1−2β E Xs∧τ − X ds ≤ CT AN,M (s)ds, ≤ CT 1−2β s∧τR 1 R 1−2ρ

0

0

I10 ≤ CT

1−2ρ

t

t s

0

0

N M E|Xu∧τ − Xu∧τ |2 du R R ds (t − s)2+2β−2ρ

≤ CT 1−2ρ 0

t

AN,M (s) 1 ds. 1+2β−2ρ (t − s)

For I12 we have I12 ≤ CT 1−2β (N −2 + M −2 ). Now consider I11 : I11 ≤ CR2 T 1−2ρ (I13 + I14 ), where I13 ≤ CE 0

τR ∧t

τR ∧t s

2 t N M Xu∧τ − Xu∧τ du s (u − s)−2β du r r ds (t − s)ν t AN,M (s)(t − s)−1+2ρ−4β ds, ≤C 1 0


τR ∧t

τR ∧t

237

2 L|XuN − XuM − XvN + XvM | dv du (u − v)1+β 0 s s τR ∧t u 2 + L|XuN − XuM |(u − v)ρ−1−β dv du s s τR ∧t u + L|XuN − XuM | |XuN − XvN |κ + |XuM − XvM |κ

I14 ≤ CE

s

u

s

× (u − v)−1−β dv du

2

(t − s)−ν ds =: C(I15 + I16 + I17 ),

where ν = 2 + 2β − 2ρ, ρ > β. In turn, I15 ≤ CT

t

2ρ−2β

N M |Xs∧τ − Xs∧τ − XuN + XuM | 2 R R du ds (s − u)1+β t AN,M (s)ds, = CT 2ρ−2β 2

s∧τR

E 0

0

0

I16 ≤ C

t

E

τR ∧t s

|XuN − XuM |(u − s)γ−β du

2

(t − s)ν

0

ds

t

AN,M (s)ds, 1

≤ CT 2ρ+2γ−4β 0

where β < γ, β < ρ. Furthermore, I17 ≤ CR2 E

τR ∧t

τR ∧t u s s

|XuN − XuM |(u − v)κ(1/2−δ)−1−β dv du (t − s)ν

0

2 ds,

where we chose 0 < δ < 1/2 − β/κ; note that β < κ − 1/2. Similarly to I16 , t 2 κ−2κδ+2ρ−4β AN,M (s)ds, I17 ≤ CR T 1 0

where κ − 2κδ + 2ρ − 4β > 0 for sufficiently small δ since ρ > β and κ > 2α. Therefore we have t 2 AN,M I14 ≤ CR (s) + AN,M (s) ds. 1 2 0

t

Hence I11 ≤ CR

4 0

Finally,

(s) AN,M N,M 1 + A2 (s) ds. (t − s)1+2β−2ρ

238


AN,M (t) 2

≤ CR

t

AN,M (s)(t 1

4

−1−2β+2ρ

− s)

0

t

AN,M (s)ds 2

ds + 0

+ C(N −2 + M −2 ).

(3.2.20)

From (3.2.19) and (3.2.20) we obtain that the sum AN,M (t) + AN,M (t) admits 1 2 N,M the same estimate as A2 (t), i.e. AN,M (t) 1

+

AN,M (t) 2

≤ CR

4

t

AN,M (s)(t − s)−1−2β+2ρ + AN,M (s) ds 1 2

0

+ C(N −2 + M −2 ); taking into account that ρ > β and using the modified Gronwall lemma (NR00), we obtain that AN,M (t) + AN,M (t) 1 2 ≤ CR4 (N −2 + M −2 ) exp{t(CR4 )1/(2ρ−2β) },

(3.2.21)

and we can put, for example, ρ := 1/4+β/2. When N, M → 0, we obtain that the right-hand side of (3.2.21) tends to zero, whence the proof follows.

Theorem 3.2.8. The SDE (3.2.11) has a solution on the interval [0, T ], and this solution is unique. Proof. Since the space {W β [0, T ], · β } is complete, from Theorem 3.2.6 we can define XτR ∧t := lim XτNR ∧t , N →∞

where the limit is taken in space Wβ [0, T ] (in particular, we have that the limit exists in L2 (Ω × [0, T ])). Using similar estimates and Theorem 3.2.6, we can prove that XτR ∧t is the unique solution of the original equation (3.2.11) on the interval [0, τR ]. From the definition (3.2.18) of τR we have τR1 ≤ τR2 for R1 ≤ R2 . So XτR1 and XτR2 coincide a.s. on the interval [0, τR1 ]. Where R → ∞ we obtain the existence and uniqueness of the solution of SDE (3.2.11) on the whole interval [0, T ].

3.2.3 The Girsanov Theorem and the Measure Transformation for the Mixed Semilinear SDE Consider equation (3.2.1) and suppose that W is underlying Wiener process for B H and that the coefficient b(t, x) satisfies the condition of Theorem 3.2.1 and can be presented as b(t, x) = e(t, x)x, where e ∈ Cb (R+ × R). Denote eˆ(t, x) := e(t, x)t−α , α = H − 12 , H ∈ ( 12 , 1). Now we try to change the measure P for another probability measure Q such that QT PT , where PT := P |FT ,


239

QT := Q|FT , and such that the drift e(t, Xt )Xt dt will be annihilated under satisfy the assumptions QT . First, let some probability measure Q T dQ 1 T 2 ϕs dWs − ϕ ds = exp dP 2 0 s 0 FT

and

T

E exp 0

1 ϕs dWs − 2

T

ϕ2s ds

=1

(3.2.22)

0

T with E 0 ϕ2s ds < ∞. t ˆ t will be a Then from the Girsanov theorem the process Wt − 0 ϕ2s ds =: W ¯ Wiener process under the measure QT . Also, let the measure Q be such that ¯ dQ = exp LT − 1 LT , dP FT 2 1 E exp LT − LT = 1, (3.2.23) 2 t t t 0 sα dMsH , where Lt = 0 sα δs dWs , MtH = 0 lH (t, s)dBsH , Wt = α t t t t l (t, s)ψs ds = α 0 δs ds, t > 0 with E 0 s2α δs2 ds < ∞, 0 |δs | ds < ∞, 0 H H := B H − t ψs ds is P -a.s., t > 0 (see Subsection 2.8.1). Then the process B t t 0 ¯ . Now we need in the equality Q = Q ¯ = an fBm w.r.t. to measure Q FT FT FT t Q|FT . Hence, in particular, Lt = 0 ϕs dWs , whence ϕs = sα δs . Therefore we want to find ϕ and ψ in such a way that common drift equals and

σ1 ϕt + σ2 ψt = −e(t, Xt ),

t ∈ [0, T ].

Now we apply the Abel rearrangement to the relation t t t lH (t, s)ψs ds = α δs ds = α s−α ϕs ds : 0 (5)

CH

t

0

(t − u)α−1 0

(u − s)−α s−α ψs ds du

t

=α

(t − u)α−1 0

(5)

0

0

or

u

whence after differentiation

s−α ϕs ds du,

0

B(α, 1 − α)CH

u

0

t

s−α ψs ds = α

0

t

(t − u)α −α u ϕu du, α

(3.2.24)

240


(6) (αCH )−1 t−α ψt

=α

t

(t − u)α−1 u−α ϕu du.

(3.2.25)

0

Substituting (3.2.25) into (3.2.24), we obtain that (9)

σ1 ϕt + σ2 CH tα

t

0

(t − u)α−1 u−α ϕu du = −e(t, Xt ), CH = αCH α . (3.2.26) (9)

(6)

Denote θt := t−α ϕt , then (9)

σ1 θt + σ2 CH

t

(t − u)α−1 θu du = −ˆ e(t, Xt ).

(3.2.27)

0

Equation (3.2.27) is a Volterra equation with weak singularity, and its unique solution has the form t ∞ (t − s)nα−1 eˆ(t, Xt ) 1 eˆ(s, Xs )ds, θt = − − ρn σ1 σ1 0 n=1 Γ (nα) (9)

where ρ = σ2 CH Γ (α). Now we must check conditions (3.2.22) (3.2.23).

and 1 T 2 Evidently, it is sufficient to check Novikov’s condition: E exp 2 0 ϕt dt < t) ∞ and E exp 12 LT < ∞. But ϕt = − e(t,X σ1 t ∞ nα−1 − σ11 tα 0 n=1 ρn (t−s) ˆ(s, Xs )ds and is bounded since e is bounded. FurΓ (nα) e ther, δs = α s−α ϕs , and Novikov’s condition evidently holds for the function L, too. So, we have proved the following result. Theorem 3.2.9. Under our suppositions equation (3.2.1) under measure Q obtains the differential form t + σ2 Xt dB H , dXt = σ1 Xt dW t

X(0) = X0 ,

and its solution has a form t + σ2 B H − 1/2σ 2 t}. Xt = X0 exp{σ1 W t 1

3.3 Stochastic Differential Equations with Fractional White Noise 3.3.1 The Lipschitz and the Growth Conditions on the Negative Norms of Coefficients Now we return to Wick integration with respect to fBm (see Sections 1.5 and 2.3). Consider the SDE of the form

3.3 Stochastic Differential Equations with Fractional White Noise

Xt = X0 +

t 0

b(s, Xs )ds +

m t j=1

0

241

H

σj (s, Xs ) ♦ B˙ s j ds,

(3.3.1)

t ∈ [0, T ], where all Hj ∈ [1/2, 1) are different, B˙ Hj are the fractional noises. The equation, similar to (3.3.1), but with white noise was studied by V˚ age (Vage96). Note that the proof of the existence and uniqueness result in (Vage96) is in fact based not on the structure of white noise, but on its inclusion into S ∗ , and this fact holds for fractional noise also, see Lemma 1.5.3. According to Theorem 1 (Vage96), the negative norm of the Wick products admits the following estimate: F ♦ G−r ≤ Cr,q F −r G−q for random variables F ∈ S−r , G ∈ S−q , r < q − 1. According to Lemma 1.5.3, B˙ tHk ∈ S−q for any q > 7/3, in particular, B˙ tHk ∈ S−3 and, moreover, supt≥0 B˙ tHk −q ≤ Cq for q > 7/3 and some Cq > 0. Therefore, for any r > 0 and F ∈ S−r F ♦ B˙ tHk −r ≤ CF −r . Suppose now that the coefficients b and σ and the initial value X0 of equation (3.3.1) satisfy the conditions: (i) for any 1 ≤ j ≤ m and some r > 0, b, σj : [0, T ]×S−r → S−r , X0 ∈ S−r , the functions b(t, Xt ) and σj (t, Xt ), 1 ≤ j ≤ m are strongly measurable on [0, T ] for any X ∈ C([0, T ], S−r ); (ii) for some r > 0 b(t, x) − b(t, y)−r + b(t, x)−r +

m j=1

m j=1

σj (t, x) − σj (t, y)−r ≤ Cx − y−r ,

σj (t, x)−r ≤ c(1 + x|−r ),

0 ≤ t ≤ T,

0 ≤ t ≤ T.

It follows from strong measurability of σj and Theorem 6 (Vage96) that H σj (t, x) ♦ B˙ t j is also strongly measurable. Further, condition (ii) ensures the t t H existence of 0 b(s, Xs )ds and 0 σj (s, x) ♦ B˙ s j ds, 0 ≤ t ≤ T, that can be considered as the Bochner integrals in S−r for X ∈ C([0, T ], S−r ). The next result can be proved with the help of the standard method of successive approximations (similar proof for white noise is contained in (Vage96)). Theorem 3.3.1. Under conditions (i) and (ii) equation (3.3.1) has on [0, T ] the unique solution X ∈ C([0, T ], S−r ). 3.3.2 Quasilinear SDE with Fractional Noise As mentioned in (Vage96), simultaneous fulfilment of the Lipschitz and growth conditions on the negative norms of coefficients is very restrictive. To avoid this, we consider the quasilinear equation of the form t m t Xt = X0 + b(s, Xs , ω)ds + σj (s)Xs ♦ B˙ sHj ds, (3.3.2) 0

j=1

0

242


where Hj ∈ [1/2, 1), the coefficients and the initial value X0 satisfy the following conditions: (iii) σj (s), 1 ≤ j ≤ m are nonrandom functions, σj ∈ L1/Hj [0, T ]; (iv) the function b(s, x, ω) : [0, T ] × R × S → R is measurable in all the arguments, |b(s, x, ω)| ≤ C(1 + |x|), ω ∈ S (R), s ∈ [0, T ], x ∈ R; |b(s, x, ω) − b(s, y, ω)| ≤ C|x − y|, x, y ∈ R, ω ∈ S (R), s ∈ [0, T ]; (v) X0 ∈ Lp (Ω) = Lp (S (R)) for some p > 0. Theorem 3.3.2. Under conditions (iii)–(v) the equation (3.3.2) has on [0, T ] the unique solution X ∈ Lp (Ω) for any p < p. Proof. Let, for simplicity, m = 1, H1 = H ∈ (1/2, 1). Consider the differential form of equation (3.3.2) dXt = b(t, Xt ) + σ(t)Xt ♦ B˙ tH , dt

X(0) = X0 .

(3.3.3)

Put σt (s) := σ(s)1[0,t] (s), and suppose that Jσ (t) is the Wick exponent of the t form Jσ (t) = exp♦ (− 0 σ(s)dBsH ), where the Wick exponent is defined as ∞ X♦ n exp♦ X := n! . Then, according to formula (1.6.1), n=0

♦

Jσ (t) = exp

−

R

H (M− σt )(s)dWs .

Denote also Zt := Jσ (t) ♦ Xt . By the rules of stochastic differentiation (see, for example, (HOUZ96)), dZt dJσ (t) dXt = Jσ (t) ♦ − σ(t) ♦ Xt ♦ B˙ tH , dt dt dt and we obtain from (3.3.3) that dZt dJσ (t) = ♦ b(t, Xt ). dt dt

(3.3.4)

Now we use the Gjessing lemma (Gje94), which states that dJσ (t) dJσ (t) H · b(t, T−(M−H σt ) Xt , ω + M− σt ), ♦ b(t, Xt , ω) = dt dt

(3.3.5)

where T is the shift operator, Tω0 F (ω) = F (ω + ω0 ) for any ω0 ∈ S (R). Similarly, Zt = Jσ (t)·T−(M−H σt ) Xt , and from (3.3.4)–(3.3.5) we obtain that Zt is the solution of the ordinary differential equation dZt dJσ (t) H = · b(t, Jσ−1 (t) · Zt , ω + M− σt ), Z0 = X0 , dt dt

(3.3.6)

3.4 The Rate of Convergence of Euler Approximations

243

for any ω ∈ S (R). Equation (3.3.6) differs from the corresponding equation (3.6.15) for the white noise (see the book (HOUZ96)) only with the function H σt instead of σt . However, it has the same structure, which means that M− conditions (iii)–(v) ensure the existence and the uniqueness of the solution of equation (3.3.2) for any ω ∈ S (R) on the interval [0, T ]. Now we estimate the moments of the solution Xt . First, from conditions (iii)–(iv) t H σs ))|ds |Zt | ≤ |X0 | + 0 Jσ (s)|a(s, Jσ−1 (s)Zs , ω − (M− t −1 ≤ |X0 | + C 0 Jσ (s)(1 + Jσ (s)|Zs |)ds t t ≤ |X0 | + C 0 Jσ (s)ds + C 0 |Zs |ds, and from the Gronwall inequality it follows that T |Zt | ≤ (|X0 | + C 0 Jσ (s)ds) exp{CT }, T E|Zt |p ≤ exp{pCT }2p (E|X0 |p + CE 0 |Jσ (s)|p ds). Since

(3.3.7)

H σt )(s)dWs } E|Jσ (s)|p + E exp♦ {−p R (M− 2 H 2 = exp{p M− σt L2 (R) },

H σt ) ∈ L2 (R), and condition (iii) and inequality (1.9.2) ensure that (M− p therefore we obtain from (3.3.7) that E|Zt | < ∞ for any p > 0. Further, T−(M−H σt ) Xt = Zt Jσ−1 (t), and E|Jσ−1 (t)|q < ∞ for any q > 0, therefore H σt ∈ L2 (R), we obtain from T−(M−H σt ) Xt ∈ Lp (Ω) for any p < p. Since M− Corollary 2.10.5 (HOUZ96) that X ∈ Lp (Ω) for any p < p.

3.4 The Rate of Convergence of Euler Approximations of Solutions of SDE Involving fBm The numerical solution of stochastic differential equations driven by Wiener process is essentially based on the method of time discretization and has a long history. We refer to the monograph (KP92), which contains an almost complete theory of the numerical solution of such SDEs with regular coefficients. The paper (KP94) is devoted to the Euler approximations for SDEs driven by semimartingales. Concerning the numerical solution of SDEs driven by fBm, we mention first the paper (GA98), where the equations with the modified fBm (which is a special semimartingale) are studied. The papers (Nou05; NN06) study Euler approximations for the homogeneous one-dimensional SDEs involving fBm and having bounded coefficients with bounded derivatives up to third order. It is proved that the error of the approximation is a.s. equivalent to δ 2α ξt , and the process ξt is given explicitly. These papers also discuss the Crank–Nicholson and the Milstein schemes for SDEs driven by fBm. Here we present the results on the rate of convergence of Euler approximations of solutions for SDE with nonstationary coefficients. Of course, our approach differs from those proposed in (Nou05),(NN06).

244


3.4.1 Approximation of Pathwise Equations Consider the multidimensional equation (3.1.12) with the coefficients satisfying the Rd -version of assumptions (i)–(v) of Subsection 3.1.1 with Hj = H, 1 ≤ j ≤ m, b0 (t) = L (see Remark 3.1.5 for additional notations). Under these assumptions, this equation has the unique solution {Xt , t ∈ [0, T ]} and for a.a. ω ∈ Ω the trajectories of the solution belong to C H− [0, T ]. T , τn = nT Now, let t ∈ [0, T ], δ = N N = nδ, n = 0, . . . , N . Consider the discrete Euler approximations of the solution of equation (3.1.12), Yτi,δ = Yτi,δ + bi (τn , Yτδn )δ + n+1 n

m

σji (τn , Yτδn )∆Bτj,H , n

Y0i,δ = X0i ,

j=1

and the corresponding continuous interpolations Yti,δ = Yτi,δ +bi (τn , Yτδn )(t−τn )+ n

m

σji (τn , Yτδn )(Btj,H −Bτj,H ), n

t ∈ [τn , τn+1 ].

j=1

(3.4.1) The continuous interpolations satisfy the equation Yti,δ

=

X0i

+ 0

t

bi (tu , Ytδu )du

+

m j=1

0

t

σji (tu , Ytδu )dBuj,H ,

(3.4.2)

where tu = τnu , nu = max{n : τn ≤ u}. For simplicity we denote the vector of solutions as Xt = (Xti )i=1,...,d , the vector of the continuous approximations as Ytδ = (Ytδ,i )i=1,...,d . Theorem 3.4.1. 1) Let the modification of conditions (i)–(v’) from Section 3.1 hold for the vector case, with γ > 1 − H, κ = µ = 1, LR = L, MR = M and b0 (t) = L. Then for any ε > 0 and 0 < ρ < H there exist δ0 > 0 and Ωε,δ0 ,ρ ⊂ Ω such that P (Ωε,δ0 ,ρ ) > 1 − ε and for any ω ∈ Ωε,δ0 ,ρ , δ < δ0 one has |Ytδ | ≤ C(ω), |Ysδ − Yrδ | ≤ C(ω)(ts − tr )H−ρ , 0 ≤ r < s ≤ T . 2) If, instead of (v) and (v’) we assume that b and σ are bounded functions, then |Ytδ | ≤ C(ω), |Ysδ − Yrδ | ≤ C(ω)(s − r)H−ρ , 0 ≤ r < s ≤ T . In both the cases C(ω) does not depend on δ. Proof. 1) We can always assume that δ ≤ 1. It follows immediately from (i) and (iii), Section 3.1.1 and (3.4.2) that for any β ∈ (1 − H, γ ∧ 1/2)


245

t m t i i,δ δ δ j,H bi (tu , Ytu ) du + |Yt | ≤ X0 + σji (tu , Ytu )dBu 0

≤ X0i + L

t 0

+ GT

m t

0

m t δ σji (tu , Ytδ ) u−β du 1 + Ytu du + GT u j=1

0

j=1

j=1

0

σji (tr , Ytδ ) − σji (tu , Ytδ ) (r − u)−β−1 du dr

r

r

0

u

i t δ −β T 1−β β Yt u du + LT + mM GT + LT ≤ X0 + mM GT u 1−β 0 t tr (tr − tu )γ + Ytδr − Yuδ + Yuδ − Ytδu + M GT 0

0

× (r − u)−β−1 du dr, (3.4.3) where GT := Λ1−β (B H ). (We use here the equality t = t for t ≤ u < r.) r u r T 1−α Denote C1 (ω) := |X0 | + mM GT 1−α + LT , C2 (ω) := mM GT + LT β . Further, note that tr − tu ≤ r − u + δ. Also, it follows from representations (3.4.1) and (3.4.2) that for any ρ ∈ (0, H) δ Yu − Ytδ ≤ L 1 + Ytδ (u − tu ) + M · C(ω, ρ) 1 + Ytδ (u − tu )H−ρ u u u δ ≤ C3 (ω) 1 + Ytu (u − tu )H−ρ , (3.4.4) where the value C(ω, ρ) appears in the relation |BtH − BsH | ≤ C(ω, ρ)|t − s|H−ρ , s, t ∈ [0, T ], C3 (ω) = LT 1−H+ρ + M · C(ω, ρ). Moreover, for γ > β t tr (tr − tu )γ (r − u)−β−1 du dr Pt := 0 0 t tr (r − u)γ + δ γ (r − u)−β−1 du dr ≤ 0 0 t t rγ−β dr + β −1 δ γ (r − tr )−β dr, ≤ (γ − β)−1 0

0

and for any k ≥ 0 and any power π > −1 τk+1 τk+1 π (r − tr ) dr = (r − τk )π dr = C1 δ π+1 with C1 = (π + 1)−1 , τk

τk

whence

t

(r − tr )−β dr ≤

0

Therefore

0

T

(r − tr )−β dr = C1 N δ 1−β = C1 δ −β .

(3.4.5)

246


Pt ≤ C1 T γ−β+1 + β −1 C1 δ γ−β ≤ C1 T γ−β+1 + β −1 C1 =: C2 . Estimate now

t

tr

Qt := 0

δ Yu − Ytδ (r − u)−β−1 du dr, u

0

using (3.4.4) and (3.4.5): t δ,∗ Qt ≤ C3 (ω) 1 + Yt

(3.4.6)

tr

(u − tu )H−ρ (r − u)−β−1 du dr t (r − tr )−β dr ≤ C4 (ω) 1 + Ytδ,∗ δ H−β−ρ ≤ C3 (ω) 1 + Ytδ,∗ δ H−ρ β −1 0

0

0

(3.4.7) δ with C4 (ω) = C3 (ω)β · C1 . Note that := sup0≤s≤t Ys < ∞ for any t ∈ [0, T ] a.s. Substituting (3.4.6) and (3.4.7) into (3.4.3), we obtain that t δ δ −β Yt ≤ C5 (ω) + mC2 (ω) Yt u du + mC4 (ω)Ytδ,∗ δ H−β−ρ u 0 (3.4.8) t tr + C6 (ω) ϕr,u du dr, −1

0

Ytδ,∗

0

where ϕr,u := − − u)−β−1 , for 0 < v < tu < T, 0 < β < 1 with C5 (ω) = mC1 (ω) + mGT C1 + mC1 GT C3 (ω) + C4 (ω), C6 (ω) = M GT . To simplify the notations, in what follows we remove subscripts from C(ω) and C, writing C(ω) for all constants depending on ω and C for all nonrandom constants. Summing up everything, we can write t t tr δ −β Yt u du + ϕ du dr . Ytδ,∗ ≤ C(ω) 1 + Ytδ,∗ δ H−β−ρ + r,u u |Ytδr

Yuδ |(r

0

In turn, we can estimate

ts 0

0

0

(3.4.9) ϕs,u du. First, similarly to the previous estimates,

ts 1 ts δ δ δ Yt (v − u)−β dv Yt − Yuδ ≤ C(ω) Y dv + 1 + 1 + t s v v u u ts tv 2 σ(tv , Ytδ ) − σ(tz , Ytδ ) (v − z)−β−1 dz dv + v z u u ts 1 δ Yt (v − u)−β dv ≤ C(ω) (ts − u)1−β + v u ts tv ts (v − tv )−β dv + ϕv,z dz dv + δγ u u u ts tv 2 δ Yz − Ytδ (v − z)−β−1 dz dv ; + z u

u

(3.4.10)


247

multiplying by (s − u)−β−1 and integrating over [0, ts ], we obtain that

5

s

ϕs,u du ≤ C(ω) 0

Qis ,

(3.4.11)

i=1

where Q1s

s

(ts − u)

1−β

:=

−β−1

(s − u)

du ≤

0

Q2s

ts

(s − u)−2β du ≤ C;

ts Ytδ (v − u)−β dv := (s − u) v 0 u ts δ v −β −β−1 Ytv (v − u) (s − u) du dv ≤ C = ts

−β−1

0

0

where C =

∞ 0

(3.4.12)

0

ts

δ Yt (s − v)−2β dv,

0

v

(3.4.13)

(1 + y)−β−1 y −β dy;

Q3s

ts

−β−1

ts

(s − u) (v − tv )−β dv du 0 u ts −1 γ (s − v)−β (v − tv )−β dv. ≤β δ

:= δ

γ

(3.4.14)

0

Let ts = nδ for some 0 < n ≤ N . The last integral can be estimated as

ts

I :=

(s − v)−β (v − tv )−β dv =

0

where (k+1)δ

n−2 (k+1)δ k=0

−β

(k+1)δ

≤ (s − (k + 1)δ) kδ

kδ

(n−1/2)δ

+

nδ

+ (n−1)δ

, (n−1/2)δ

(v − τv )−β dv ≤ C(s − (k + 1)δ)−β δ 1−β ,

kδ

and the last two integrals are bounded by Cδ 1−2β . Therefore, I ≤ Cδ −β . Further, using estimate (3.4.4), we can conclude that ts ts tv Q4s := (s − u)−β−1 ϕv,z dz dv du 0 u u ts tv z∧v ϕv,z (s − u)−β−1 du dz dv ≤ 0 0 0 tv ts (s − v)−β ϕv,z dz dv. ≤C 0

0

Finally, similarly to the previous estimates,

(3.4.15)

248


tv Yzδ − Ytδ (v − z)−β−1 dz dv du (s − u) z 0 u u ts tv ts (s − u)−β−1 (v − z)−β−1 dz dv du · δ H−ρ 1 + Ytδ,∗ ≤ C(ω) s 0 u u H−ρ−β ≤ C(ω) 1 + Ytδ,∗ . δ s (3.4.16) t Now, denote ψs := Ysδ,∗ + 0 s ϕs,u du. Then it follows from (3.4.9) and (3.4.11)– (3.4.16) that for any t ∈ [0, T ] (including t = kδ) t δ,∗ H−β−ρ (t − v)−2β + v −β ψv dv . + ψ(t) ≤ C(ω) 1 + Yt δ

Q5s

:=

ts

−β−1

ts

0

Let ε > 0 be fixed. Note that all constants C(ω) are finite a.s. and independent of δ. Thus, we can choose δ0 > 0, ρ small enough such that H − β − ρ > 0, and Ωε,δ0 ,ρ such that C(ω)δ0H−β−ρ ≤ 1/2 on Ωε,δ0 ,ρ and P (Ωε,δ0 ,ρ ) > 1 − ε. Then for any ω ∈ Ωε,δ0 ,ρ t 1 (t − v)−2β + v −β ψv dv, ψt ≤ C(ω) + ψt + C(ω) 2 0 whence

t 2β ψt ≤ C(ω) 1 + t (t − v)−2β v −2β ψv dv , 0

and it follows immediately from the last equation and (3.1.22)–(3.1.23) that ψt ≤ C(ω) whence, in particular, Ytδ ≤ C(ω), t ∈ [0, T ]. Moreover, from (3.4.10) with u = tr , r ≤ s, taking into account that ts (v − tv )−β dv = (1 − β)−1 δ −β (ts − tr ), we obtain the bound tr δ Yt − Ytδ ≤ C(ω) (ts − tr )1−β + δ γ−β (ts − tr ) + (ts − tr ) s r ts (v − tv )−β dv ≤ C(ω)(ts − tr )1−β , + δ H−ρ tr

and statement 1) is proved. 2) Let |b(t, x)| ≤ b, |σ(t, x)| ≤ σ. Then it is very easy to see that estimate (3.4.8) will take the form t tr δ Yt ≤ C(ω) 1 + ϕr,v du dr , 0

0

(3.4.10) will take the form ts δ Yt − Yuδ ≤ C(ω) (ts − u)1−β + (δ γ + δ H−ρ ) (v − tv )−β dv s u ts tv ϕv,z dz dv , + u

u


249

and instead of (3.4.11)–(3.4.16) we obtain

ts

ϕs,u du ≤ C(ω) 1 +

0

0

ts

−β

(s − v)

tv

ϕv,z dz dv ,

0


Remark 3.4.2. It is easy to see that we proved a little more than Theorem 3.2.3 states. Namely, we proved that the norm in Besov space, sup0≤s≤T ψs , is bounded by C(ω) on Ωε,δ0 ,ρ , with C(ω) not depending on δ. Now we establish the estimates of the rate of convergence of our approximations (3.4.2) for the solution of equation (3.1.12) with pathwise integral w.r.t. fBm. We establish even more, namely, the estimate of convergence rate for the norm of the difference Xt − Ytδ in some Besov space, similarly to the result of Theorem 3.4.1. Denote ∆u,s (X, Y δ ) := Xs − Ysδ − Xu + Yuδ . Theorem 3.4.3. Let the modification of conditions (i)–(v’) from Section 3.1 hold for the vector case, with γ > 1 − H, κ = µ = 1, LR = L, MR = M and b0 (t) = L, and suppose also that: 1)the coefficient b is H¨ older continuous in time: |b(t, x)−b(s, x)| ≤ C|t−s|θ , C > 0, 2α < θ ≤ 1, α = H − 1/2; 2) the exponent γ from condition (iii)(Section 3.1) satisfies γ > H. Then: 1. For any ε > 0, β ∈ (1 − H, 1/2) and any sufficiently small ρ > 0 there exists δ0 > 0 and Ωε,δ0 ,ρ such that P (Ωε,δ0 ,ρ ) > 1 − ε and for any ω ∈ Ωε,δ0 ,ρ , δ < δ0 ts ∆u,s (X, Y δ ) (s − u)−β−1 du Uδ := sup Xs − Ysδ + 0≤s≤T

≤ C(ω) · δ

0 2α−ρ

,

where C(ω) does not depend on δ and ε (but depends on ρ); 2. If, in addition, the coefficients b and σ are bounded, then for any ρ ∈ (0, 2α) there exists C(ω) < ∞ a.s. such that Uδ ≤ C(ω)δ 2α−ρ , C(ω) does not depend on δ. Proof. 1. Denote Ztδ := sup0≤s≤t Xs − Ysδ . Then

250


s

Ztδ ≤ sup

0≤s≤t

+ sup

0 m

|b(u, Xu ) − b(tu , Ytδu )|du

0≤s≤t i,j=1

s (σji (u, Xu ) − σji (tu , Ytδu ))dBuj,H 0

t

≤

|b(u, Xu ) − b(u, Yuδ )|du + 0 t

+ 0

t

|b(u, Yuδ ) − b(tu , Yuδ )|du 0

|b(tu , Yuδ ) − b(tu , Ytδu )|du

(3.4.17)

m s (σji (u, Xu ) − σji (u, Yuδ ))dBuj,H + sup 0≤s≤t i,j=1 m

+ sup

0≤s≤t i,j=1

0

s (σji (u, Yuδ ) − σji (tu , Yuδ ))dBuj,H 0

6 m s (σji (tu , Yuδ ) − σji (tu , Ytδu ))dBuj,H =: Ik .

+ sup

0≤s≤t i,j=1

0

k=1

Now we estimate separately all these terms. Evidently, t I1 ≤ L Zuδ du.

(3.4.18)

0

Condition 1) implies that for δ < 1 t θ |u − tu | du ≤ Cδ θ ≤ Cδ 2α . I2 ≤ C

(3.4.19)

0

It follows from Theorem 3.4.1 that for any ε > 0 and any ρ ∈ (0, H) there exists δ0 > 0 and Ωε,δ0 ,ρ ⊂ Ω such that P (Ωε,δ0 ,ρ ) > 1 − ε and C(ω) independent of ε and δ such that for for any ω ∈ Ωε,δ0 ,ρ it holds that δ Yt − Ysδ ≤ C(ω) |t − s|H−ρ . In what follows we assume that δ < δ0 < 1. Therefore I3 ≤ L · C(ω)δ H−ρ · t ≤ C(ω)δ H−ρ ,

ω ∈ Ωε,δ0 ,ρ .

(3.4.20)

Now we go on with I4 . For 1 − H < β < 1/2 m * t σji (u, Xu ) − σji (u, Ytδ ) u−β du I4 ≤ C(ω) u t + 0

0

i,j=1 r

0

σji (r, Xr ) − σji (u, Xu ) − σji (r, Yrδ ) + σji (u, Yuδ ) + × (r − u)−β−1 du dr =: I7 + I8 .

(3.4.21)


Evidently,

t

I7 ≤ C(ω)

Zuδ u−β du.

251

(3.4.22)

0

According to (3.1.1), under conditions (i)–(iii) |σ(t1 , x1 ) − σ(t2 , x2 ) − σ(t1 , x3 ) + σ(t2 , x4 )| ≤ M |x1 − x2 − x3 + x4 | γ κ κ + M |x1 − x3 | |t2 − t1 | + |x1 − x2 | + |x3 − x4 | . (3.4.23) 12 Therefore, I8 ≤ k=9 Ik , where t

r

Xr − Yrδ (r − u)γ−β−1 du dr,

r

Xr − Yrδ |Xr − Xu |κ (r − u)−β−1 du dr,

r

Xr − Yrδ Yrδ − Yuδ κ (r − u)−β−1 du dr,

I9 = C(ω) 0

0

t I10 = C(ω) 0

0

t I11 = C(ω) 0

0

t I12 = C(ω) 0

r

∆u,r (X, Y δ )(r − u)−β−1 du dr.

0

Taking into account that β > H > α, we obtain that t I9 ≤ C(ω) Zuδ du

(3.4.24)

0

It follows from Theorem 3.1.4 that under assumptions (i)–(v) for any 0 < ρ < H there exists a constant C(ω) such that sup |Xt | ≤ C(ω),

0≤t≤T

sup 0≤s≤t≤T

|Xt − Xs | ≤ C(ω) |t − s|

H−ρ

.

(3.4.25)

Moreover, we can choose ρ > 0 and β > 1 − H such that κ(H − ρ) > β and H − ρ > 2β, because κH > 1 − H. In this case t T r δ κ(H−ρ)−β−1 Zr (r − u) du dr ≤ C(ω) Zrδ dr. (3.4.26) I10 ≤ C(ω) 0

0

0

Evidently, on the corresponding set Ωε,δ0 ,ρ the same estimate holds for I11 . Now estimate I5 . t σ(u, Yuδ ) − σ(tu , Yuδ ) u−β du I5 ≤ C(ω) 0 t r σ(r, Yrδ ) − σ(tr , Yrδ ) − σ(u, Yuδ ) + σ(tu , Yuδ ) + C(ω) 0 0 −β−1

× (r − u)

du dr =: I13 + I14 .

252


Obviously,

I14

I13 ≤ C(ω)δ γ , t t tr t r ≤ C(ω) ≤ C(ω) + 0

t

r

+ 0

0

0

0

tr

(3.4.27) tr

δ γ (r − u)−β−1 du dr

0

(r − u)γ + (r − u)H−ρ (r − u)−β−1 du dr ≤ C(ω)(δ γ−β + δ H−ρ−β ).

tr

(3.4.28)

Similarly, t σ(tu , Yuδ ) − σ(tu , Ytδ ) u−β du I6 ≤ C(ω) u 0 t r σ(tr , Yrδ ) − σ(tr , Ytδ ) − σ(tu , Yuδ ) + σ(tu , Ytδ ) + C(ω) r u 0

(3.4.29)

0

× (r − u)−β−1 du dr =: I15 + I16 . Here

t

δ H−ρ u−β du ≤ C(ω)δ H−ρ ,

I15 ≤ C(ω)

(3.4.30)

0

t I16 ≤ C(ω) t

0 r

r

+

0

0

t ≤ C(ω) δ H−ρ 0

tr

tr

(r − u)−β−1 du dr

0

(r − u)H−ρ−β−1 du dr ≤ C(ω)δ H−ρ−β .

+ 0

t

tr

tr

(3.4.31) Substituting (3.4.18)–(3.4.31) into (3.4.17), we obtain that on Ωε,δ0 ,ρ t t Ztδ ≤ C(ω) Zrδ r−β dr + δ H−ρ−β + δ H−ρ + θr dr , (3.4.32) where θr =

r 0

0

0 −β−1

∆r,u (X, Y )(r − u) δ

du. Recall that H − ρ > 2α, therefore

t Ztδ ≤ C(ω) Zrδ r−α + θr dr + δ 2α−ρ . 0

Now we estimate θt . Evidently, for t > u t δ b(s, Xs ) − b(ts , Ytδ ) ds ∆t,u (X, Y ) ≤ s u

m t j,H δ + σji (s, Xs ) − σji (ts , Yts ) dBs . i,j=1

Therefore we obtain that θt ≤

u

9 k=1

Jk , where


t

b(s, Xs ) − b(s, Ysδ ) (t − u)−β−1 ds du,

t

J1 = 0

u

t

t

b(s, Ysδ ) − b(ts , Ysδ ) (t − u)−β−1 ds du,

J2 = 0

u

t

t

b(ts , Ysδ ) − b(ts , Ytδ ) (t − u)−β−1 ds du,

J3 = 0

σ(s, Xs ) − σ(s, Ysδ ) (s − u)−β (t − u)−β−1 ds du,

t

J4 = C(ω) 0

u

t

t

J5 = C(ω) 0

u

t

t

σ(s, Ysδ ) − σ(ts , Ysδ ) (s − u)−β (t − u)−β−1 ds du, σ(ts , Ysδ ) − σ(ts , Ytδ ) (s − u)−β (t − u)−β−1 ds du,

J6 = C(ω) 0

s

u

t

s

u

t t

r

J7 = C(ω) 0

253

u

σ(r, Xr ) − σ(r, Yrδ ) − σ(v, Xv ) + σ(v, Yvδ )

u

× (r − v)−β−1 (t − u)−β−1 dv dr du, t t

r

J8 = C(ω) 0

u

σ(r, Yrδ ) − σ(tr , Yrδ ) − σ(v, Yvδ ) + σ(tv , Yvδ )

u

× (r − v)−β−1 (t − u)−β−1 dv dr du, t t

r

J9 = C(ω) 0

u

σ(tr , Yrδ ) − σ(tr , Ytδ ) − σ(tv , Yvδ ) + σ(tv , Ytδ ) r

u

v

× (r − v)−β−1 (t − u)−β−1 dv dr du. s t It is clear that J1 ≤ C 0 Zsδ 0 (t − u)−β−1 du ds, J3 ≤ C(ω)δ H−ρ . Further, J4 ≤ C

Zsδ 0

The inner integral Therefore

s 0

t

s

J2 ≤ Cδ θ ,

(s − u)−β (t − u)−β−1 du ds.

0

(s−u)−β (t−u)−β−1 du ≤ (t−s)−2β

∞ 0

(1+y)−β−1 y −β dy.

254


t

J4 ≤ C

(t − s)−2β Zsδ ds.

0

Similarly to J2 , J5 ≤ C(ω)δ γ , and similarly to J3 , J6 ≤ C(ω)δ H−ρ . Estimating J7 , J8 and J9 is, of course, t t a rbit more complicated, but not dramatically. Obviously, J8 ≤ C(ω)δ γ 0 u u (r − v)−β−1 (t − u)−β−1 du t = C(ω)δ γ 0 (t − u)−2β dv dr du ≤ C(ω)δ γ ; similarly J9 ≤ C(ω)δ H−ρ . Now we apply to J7 inequality (3.4.23) and obtain the following estimate of the integrand: 1 σ(r, Xr ) − σ(r, Yrδ ) − σ(v, Xv ) + σ(v, Yvδ ) ≤ M ∆r,v (X, Y δ ) κ 2 κ + Xr − Yrδ (r − v)γ + Xr − Yrδ |Xr − Xv | + Xr − Yrδ Yrδ − Yvδ . (3.4.33) 13 According to this, we write J7 ≤ k=10 Jk , where, in turn, t t r ∆r,v (X, Y δ )(r − v)−β−1 (t − u)−β−1 dv dr du J10 = C(ω) 0 u u t r v (t − u)−β−1 ∆r,v (X, Y δ )(r − v)−β−1 dv dr du = C(ω) 0 0 0 t (t − r)−β θr dr; ≤ C(ω) 0

t t

r

J11 = C(ω)

0

u t

≤ C(ω)

u r

Zrδ

0

Xr − Yrδ (r − v)γ−β−1 dv dr(t − u)−β−1 du

(t − u)−β−1

0 t

r

(r − v)γ−β−1 dv du dr

u

(t − r)−β Zrδ dr,

≤ C(ω) 0

t t

r

J12 = C(ω) 0

u r

t

u

r

≤ C(ω)

0

0 t

≤ C(ω)

Xr − Yrδ |Xr − Xv |κ (r − v)−β−1 dv dr(t − u)−β−1 du Zrδ (r − v)κ(H−ρ)−β−1 (t − u)−β−1 dv dr du

u

Zrδ (t − r)−β dr,

0

t and J13 ≤ C(ω) 0 Zrδ (t−r)−β dr is obtained the same way. Summing up these t estimates, we obtain that J7 ≤ C(ω) 0 (t − r)−β Zrδ + θr dr, whence t θt ≤ C(ω) (t − r)−2β Zrδ + θr dr + δ H−ρ + δ θ . 0

(3.4.34)


255

Coupling together (3.4.32) and (3.4.34), and taking into account that H − ρ > 2α, θ > 2α, we obtain t δ 2α (t − r)−2β + r−β Zrδ + θr dr Zt + θt ≤ C(ω) δ + 0 (3.4.35) t δ 2α 2β −2β −2β ≤ C(ω) δ + t Zr + θr dr . (t − r) r 0

The proof now follows immediately from (3.4.35) and (3.1.22)–(3.1.23). Statement 2 is obvious.

3.4.2 Approximation of Quasilinear Skorohod-type Equations Now we proceed with the problem of the numerical solution of Skorohodtype equations driven by fractional white noise. From now on, we assume that our probability space is the white noise space, i.e. (Ω, F, P ) = (S (R), B(S (R)), µ), the symbol ♦ stands for the Wick product, Wt = ˙ is the white noise. (See also 1[0,t] , ω is the standard Brownian motion, W Sections 1.4, 1.5, 2.3 and Subsection 3.3.2.) Consider the quasilinear Skorohod-type equation driven by fractional white noise that is the one-dimensional analog of equation (3.3.2): Xt = X0 +

t

0

t

σ(s)Xs ♦ B˙ sH ds,

b(s, Xs , ω) ds +

(3.4.36)

0

with nonrandom initial condition X0 . Suppose that the coefficients b and σ satisfy conditions (iii)–(iv) of Theorem 3.3.2 (in this subsection we always refer to them as to conditions (iii)–(iv)), and (vi) “Smoothness” of b w.r.t. ω: for any t ∈ [0, T ] and for h ∈ L1 (R) |b(t, x, ω + h) − b(t, x, ω)| ≤ C(1 + |x|) |h(s)| ds. R

(vii) H¨ older continuity of b w.r.t. t or order H with constant that grows linearly in x: H

|b(t, x, ω) − b(s, x, ω)| ≤ C(1 + |x|) |t − s| ; (viii) H¨ older continuity of σ w.r.t. t or order H: H

|σ(t) − σ(s)| ≤ C |t − s| . Remark 3.4.4. Condition (vii) holds if, for example, the coefficient b has the stochastic derivative growing at most linearly in x. It is obviously true if b is nonrandom and H¨ older of order H.

256


Consider the fractional Wick exponent H σt (s)dWs Jσ (t) = exp − M− R

1 2 H = exp − M− σt (s)dWs − σt |RH |,1 2 R

It easily follows from Theorem 3.3.2 that for nonrandom X0 under conditions (iii)–(iv) the equation (3.4.36) has a unique solution that belongs to all Lp (Ω) and can be represented in the form Zt = Jσ (t) ♦ Xt ,

or

Xt = J−σ (t) ♦ Zt ,

where the process Zt solves (ordinary) differential equation Zt = X0 + 0

t

H Jσ (s)b(s, Jσ−1 (s)Zs , ω + M− σs ) ds.

(3.4.37)

This gives the following idea of construction of time-discrete approximations of the solution of (3.4.36). Take the uniform partition {τn = nδ, n = 1, . . . , N } of [0, T ] and define first the approximations of Z in a recursive way: 0 = X0 , Z τ τ + J(τ n )b(τn , J−1 (τn )Zτ , ω + M σ Z =Z n )δ, n+1 n n where

(3.4.38)

t !2 1! := exp − J(t) 1[0,t] !|R |,1 , σ (s)dBsH − !σ H 2 0 H n := σ 1[0,τn ] , M := M− . σ (s) := σ(ts ), σ

n are easily computable as finite sums of Note that both σn |RH |,1 and M σ elementary integrals. Further, we interpolate continuously by t s )b(ts , J−1 (ts )Zt , ω + M σ t = X0 + J(t Z ns ) ds, (3.4.39) s 0

where ns = max{n : τn ≤ s}, and set t = T−M (σ1 ) J−1 (t)Zt , X [0,t]

(3.4.40)

where for ω0 ∈ S (R) Tω0 is the shift operator, Tω0 F (ω) = F (ω + ω0 ). Lemma 3.4.5. Under the assumption (vi), the following estimate is true: α e 1 b(t, e−α1 x, ω) − eα2 b(t, e−α2 x, ω) ≤ C(1 + eα1 + eα2 + |x|) |α1 − α2 | .


257

Proof. Write α e 1 b(t, e−α1 x, ω) − eα2 b(t, e−α2 x, ω) ≤ eα1 b(t, e−α1 x, ω) − eα1 b(t, e−α2 x, ω) + eα1 b(t, e−α2 x, ω) − eα2 b(t, e−α2 x, ω)

and apply (vi).

Lemma 3.4.6. Let ξ1 and ξ2 be jointly Gaussian variables. Then for q ≥ 1 1 2q 2 0 such that sup0≤s≤T E exp{λvs2 } < ∞, is sufficient for the Novikov condition,

T if it has the form E exp 12 0 vs2 ds < ∞. Therefore, the proof follows immediately from (3.5.6), (3.5.7) and (3.5.9). Let H ∈ (1/2, 1). In this case δt is a fractional derivative of the form: t d (5) −α −α δt = (t − s) s h(s)ds CH dt 0 ⎛ ⎞ t (5) = CH ⎝t−2α h(t)+α (t−α h(t) − r−α h(r))(t − r)−α−1 dr)1(0,T ) (t)⎠, 0


Now we establish more convenient conditions for the existence of a weak solution in terms of g and b. Denote the function h(s, x) := g(s)b(s, x). Theorem 3.5.3. Let 0 < |f (t)| < f ∗ for any t ∈ [0, T ] and one of the following assumptions holds: (iii) H ∈ (0, 1/2) and h(t, x) is of linear growth: |h(t, x)| ≤ C(1 + |x|),

(t, x) ∈ [0, T ] × R

3.5 SDE with the Additive Wiener Integral w.r.t. Fractional Noise

265

(iv) H ∈ (1/2, 1), f is essentially bounded on [0, T ] and h(s, x) is H¨ older continuous: ρ

γ

|h(t, x) − h(s, y)| ≤ C (|x − y| + |t − s| ) , where 1 > rho > 1 −

1 2H

and 1 ≥ γ > α.

Then equation (3.5.1) has a weak solution. Proof. In both cases we must check the conditions of Theorem 3.5.2. 2 t Let H ∈ (0, 1/2). Then t2α 0 s−α (t − s)−α−1 h(s)ds ≤ Ct2α sup0≤s≤t |h(s)| t−4α ≤ CT −2α (1 + |X0 | + IT∗ (f )). Note that now α < 0). Furthermore, inequality (3.5.7) is transformed into 2

E exp{λ(IT∗ (f ))2 } < ∞

for some

λ > 0.

The last inequality follows from the Fernique theorem (Fer74) about exponential integrability of the square of the supremum norm of a Gaussian process (recall that the process It (f ) is Gaussian). For H ∈ (1/2, 1) ρ

ρ

|h(s)| ≤ |h(0, 0)| + C (sγ + |X0 | + |Is (f )| ) ,

s

sα 0

(3.5.10)

s s−α h(s) − r−α h(r) h(s) − h(r) dr = dr α+1 α+1 (s − r) 0 (s − r) s −α s −α (s − r−α )(h(r) − h(s)) s − r−α α + sα dr + s h(s) dr. α+1 (s − r)α+1 0 0 (s − r)

Further, |h(s) − h(r)| ≤ |h(s, X0 + Is (f )) − h(r, X0 + Ir (f ))| γ ρ ≤ C (|s − r| + |Is (f ) − Ir (f )| ) , therefore s s s ρ h(s) − h(r) |Is (f ) − Ir (f )| γ−α−1 ≤ C dr (s − r) dr + C dr α+1 (s − r)α+1 0 (s − r) 0 0 s ρ |Is (f ) − Ir (f )| dr. ≤C +C (s − r)α+1 0 Similarly to Lemma 1.17.1, it follows from the Garsia–Rodemich–Rumsey inequality that for any 0 < ε < H H−ε

|Is (f ) − Ir (f )| ≤ CH,ε |r − s| where

ξε ,

266


T

T

ξε = 0

0

|Ix (f ) − Iy (f )| |x − y|2H/ε

ε/2

2/ε

dx dy

.

Further, according to Corollary 1.9.4, it holds that T T f 2/ε L 1 [x,y] 2 H ≤ C H, dx dy 2H/ε ε |x − y| 0 0

Eξε2/ε

2 2/ε ≤ C H, (f ∗ ) T 2 . (3.5.11) ε

Therefore, 0

s

ρ

|Is (f ) − Ir (f )| ρ dr ≤ CH,ε ξερ (s − r)α+1

s

(s − r)ρ(H−ε)−α−1 dr ≤ Cξερ 0

for some constant C and such ε that ρ(H − ε) − α > 0, and s h(s) − h(r) ρ dr ≤ C (1 + ξε ) . α+1 (s − r) 0

(3.5.12)

The next term admits an estimate s −α s (s − r−α )(h(s) − h(r)) |h(s) − h(r)| −α sα r dr dr ≤ α+1 (s − r) (s − r) 0 0 s ρ |Is (f ) − Ir (f )| −α r dr ≤ C (1 + ξερ ) ; (3.5.13) ≤C +C (s − r) 0 the proof follows now from (3.5.10)–(3.5.13), because for ρ < 1

T 2ρ Eλ ≤ CE exp λ |Is (f )| s−2α ds + λT Cξε2ρ < ∞. 0

3.5.2 Existence of a Weak Solution for SDE with Discontinuous Drift Consider equation (3.5.1) for the case when f ≡ 1, b(s, x) = b(x) and b(x) is Hölder continuous of order ρ ∈ (1 − 1/2H, 1) except on a finite number of points, where there is a jump discontinuity (MN04). Theorem 3.5.4. Suppose that the function b(x) is H¨ older continuous of order ρ ∈ (1 − 1/2H, 1) in a finite number of intervals (−∞, a1 ), (a1 , a2 ), . . . ,(aN −1 , aN ), (aN , +∞) and there is a jump discontinuity in the points H ai , 1 ≤ i ≤ N , that is, b(ai −) = b(ai +) = b(ai ). Let Bt be an fBm with

Hurst parameter H ∈ weak solution.

√ 1 1+ 5 , 2 4

. Then equation (3.5.1) with f ≡ 1 has a


267

Remark 3.5.5. The case H ∈ (0, 1/2) is not specific now; for example, if b is discontinuous but bounded we have a weak solution. Proof. A function b(x) satisfying the conditions of Theorem 3.5.4 can be decomposed as follows: b(x) = d(x) +

N

ci sign(x − ai ),

i=1 1 where the function d is Hölder continuous of order ρ ∈ (1 − 2H , 1), and ci ∈ R. Then, in order to prove Theorem 3.5.4 it suffices to check that the function sign(x − ai ) satisfies condition (3.5.8) for all λ > 0. We have now that h(s) = b(X0 + BsH ) = sign(X0 + BsH ). Since T 1−2α sign(X0 + BsH )s−α 2 ds ≤ T , 1 − 2α 0 it suffices to consider the term s −α s sign(X0 + BsH ) − r−α sign(X0 + BrH ) α As = s dr. (s − r)α+1 0

We have

s

r

As =

(s − r)α+1

0

s

≤

sign(X0 + BsH ) − sign(X0 + BrH ) (s − r)α+1

0 s

1 − s α sign(X0 + BrH ) r

+ =

sign(X0 + BsH ) − s α sign(X0 + BrH )

0 A1s

(s − r)α+1

dr

dr

+ A2s .

The term A2s can be easily bounded: s s α −1 2 r dr = c, As ≤ α+1 0 (s − r) where

c= 0

1

z −α − 1 dz < ∞. (1 − z)α+1

For the term A1s we can write s A1s ≤ 2 1{X0 +BsH >0,X0 +BrH 0, x > 0 and for some fixed 0 < ε < H. Set T −2α Sε := 0 |BsH + x| H−ε ds. We can write, assuming ε < 13


2α H−ε

E exp λξε

2α ⎜ H−ε = E ⎝exp λξε Sε 1

Sε

⎛

⎜ + E ⎝exp λξε ≤ E exp

2α H−ε

2− ε λξε H−ε

1−3ε Sε −1 + ε, < ∞. By the H¨ older inequality, assuming H−ε we obtain that

2α (H−ε)(1−ε)

T

|BsH

Sε ≤ CT,ε

−1+ε

+ x|

ds

.

0

Hence, 2H−3ε 1−3ε

Sε

ρ

T

|BsH

≤ CT,ε

−1+ε

+ x|

ds

,

0 (2α)(2H−3ε) where ρ = (H−ε)(1−ε)(1−3ε) can be expressed as ρ = 4α + δ, where δ > 0 tends to zero as ε tends to zero. Therefore, it suffices to show that E exp λψε4α+δ < ∞, (3.5.16)

where

T

ψε = 0

1{|BsH +x| 1, 1/γ + 1/γ = 1 and γβ = 2. Then T

(y−x)2 T 1 β β − 2σt2 √ g(t, X ) dt = g(t, y) e dy dt t 0 0 2πσt R 1/γ 1/γ − γ (y−x)2 T T 2 −γ 1 γβ 2σ t ≤ √2π 0 R g(t, y) dy dt e σt dy dt 0 R 1/γ 1/γ T T −γ z 2 1−γ = √12π 0 R g(t, y)2 dy dt e σ dz dt t 0 R 2−β 1/γ β 2 T T − 2−β ≤ C 0 R g(t, y)2 dy dt σt dt . 0

ˆ E

Finally, put

β 2−β

= 1 − β1 , γ = 1 + 1r . T ≥ C(H)||f ||L 1 (0,t) , so 0 σt−r dt < ∞,

= r > 1, which means that β =

From inequality (1.9.1), ||f ||LH 2 (0,t) whence the proof follows.

2r 1 1+r , α H

Lemma 3.5.10. Let bn (t, x) = bn (t, x)1{|x| ≤ C1 } be a sequence of measurable functions, |bn (t, x)| ≤ C2 , limn→∞ bn (t, x) = b(t, x), for all (t, x) ∈ [0, T ] × R, and the conditions of Lemma 3.5.9 hold. Let also the cor(n) responding solutions Xt of the equations t (n) Xt = X 0 + bn (s, Xs(n) )ds + It (f ), t ∈ [0, T ], 0

converge a.s. to some process Xt for all t ∈ [0, T ]. Then the process X is a solution of equation (3.5.1). T (n) Proof. It is sufficient to prove that limn→∞ In := limn→∞ E 0 |bn (s, Xs ) (1) (2) − b(s, Xs )|ds = 0. But In ≤ In + In , where In(1)

T

|bn (s, Xs(n) ) − b(s, Xs(n) )|ds,

=E 0

T

|b(s, Xs(n) ) − b(s, Xs )|ds.

In(2) = E 0

T (1) Evidently, from (3.5.18) and finiteness of bn and b, In ≤ C( 0 R |bn (t, x) − (2) b(t, x)|2 dt dx)1/2 → 0, n → ∞, and also In → 0, n → ∞.

Theorem 3.5.11. Let both the functions h(t, x) and b(t, x) satisfy the linear growth condition. Then equation (3.5.1) has the unique strong solution.

274


Remark 3.5.12. The next condition is sufficient for both the functions h(t, x) and b(t, x) to be of linear growth: |b(s, x)| ≤ C(|f (s)| ∧ 1)(1 + |x|).

(3.5.19)

Proof. For any R > 0 denote b(R) (t, x) := b(t, x)1{|x|≤R} . Let ϕ be a smooth nonnegative function with compact support in R such that R ϕ(x)dx = 1. De fine bR,j (t, x) := j R b(R) (t, y)ϕ(j(x − y))dy. Let for n ≤ k bR,n,k = ∧kj=n bR,j , bR,n = ∧∞ bR,j . The functions bR,n,k are Lipschitz in x uniformly in t and j=n bR,n,k ↓ bR,n , k → ∞, bR,n ↑ b(R) , n → ∞, for a.a. x and any t. Equation R,n,k as an ordinary differential (3.5.1) with bR,n,k has the unique solution X equation with Lipschitz coefficient. By the comparison theorem for ODEs the R,n . The sequence X R,n R,n,k decreases in k, hence it has a limit X sequence X (R) increases in n, hence it has a limit X . Applying Lemma 3.5.10 we obtain (R) that {Xt , t ∈ [0, T ]} is a solution of (3.5.1) with drift b(R) (t, x). Then we (R) apply standard techniques: all Xt are bounded by (IT∗ (f ) + |x|)eCT , and (R) (3.5.1) has a unique solution on any [0, τR ], where τR = inf{t : |Xt | ≥ R}. It means that (3.5.1) has a unique solution on the whole interval [0, T ].

3.5.5 Existence of a Strong Solution for Discontinuous Drift Let Ω = C0 ([0, T ], R) be the Banach space of continuous functions, null at time 0, equipped with the supremum norm, and P be the unique probability measure on Ω such that the canonical process is an fBm with Hurst parameter H ∈ (1/2, 1). Assume also that the canonical filtration is augmented with the P -negligible sets. We consider the following partial case of equation (3.5.1): t b(Xs )ds + BtH (3.5.20) Xt = X0 + 0 √

with b(x) = sign x, H ∈ (1/2, H0 ), H0 = 1+4 5 . According to Theorem 3.5.4, equation (3.5.20) has a weak solution. Now we intend to prove the existence of its strong solution. For this purpose consider the following approximations of the function b(x) = sign x: ⎧ x ≤ 0; ⎪ ⎪ −1, ⎪ ⎪ 0 < x ≤ n12 ; ⎨ n3 x2 − 1, 1/n2 < x ≤ n1 − n12 ; bn (x) = 2nx − 1, ⎪ 1 2 3 ⎪ 1/n2 < x ≤ n1 ; ⎪ ⎪ 1 − n (x − n ) , 1/n − ⎩ 1 1, x ≥ n. Then

⎧ 0, x ≤ 0; ⎪ ⎪ ⎪ 3 ⎪ 2n x, 0 < x ≤ n12 ; ⎨ 1/n2 < x ≤ n1 − n12 ; bn (x) = 2n, ⎪ 1 3 ⎪ 2n (x − n ), 1/n − 1/n2 < x ≤ n1 ; ⎪ ⎪ ⎩ 0, x ≥ n1 .


275

Evidently, any bn ∈ C(R); moreover, it is Lipschitz: |bn (x1 ) − bn (x2 )| ≤ 2n3 |x1 − x2 |. Lemma 3.5.13. For any x ∈ R bn+1 (x) > bn (x), n ≥ 1. 1 ). Proof. It is sufficient to consider the interval (0, n+1 1 3 2 (a) For x ∈ (0, (n+1)2 ] bn+1 (x) = (n + 1) x − 1 > n3 x2 − 1 = bn (x). 1 1 bn (x) = n3 x2 − 1, bn+1 (x) = 2(n + 1)x − 1. But the (b) For x ∈ ( (n+1) 2 , n2 ]

inequality 2(n + 1)x − 1 > n3 x2 − 1 holds for x < 2(n+1) n3 , and it is our case. 1 1 − (n+1) bn+1 (x) = 2(n + 1)x − 1 > 2nx − 1 = bn (x). (c) For x ∈ ( n12 , n+1 2] 1 1 1 1 1 − (n+1) bn+1 (x) = 1 − (n + 1)3 (x − n+1 )2 , (d) For x ∈ ( n+1 2 , n − n2 ] 1 3 2 bn (x) = 2nx − 1. The function ϕ(x) := (n + 1) (x − n+1 ) + 2nx − 2 has 1 1 n ϕ (x) = 2(n + 1)3 (x − n+1 ) + 2n = 0 for x0 = n+1 − (n+1) 3 , it is the point 1 1 1 1 of local minimum and x0 ∈ ( n+1 − (n+1)2 , n − n2 ] for n > 2. So, we must 1 1 1 1 − (n+1) check the inequality ϕ(x) < 0 for x = n+1 2 and x = n − n2 , and it evidently holds. 1 ) the inequality bn+1 (x) = 1 − (n + 1)3 (x − (e) Finally, for x ∈ ( n1 − n12 , n+1 A 1 1 2 2 3 n(n + 1))x > 1 n+1 ) > 1 − n (x − n ) = bn (x) is equivalent to (2n + 1 + and it is sufficient to check it in the point x = n1 − n12 : (2n + 1 +

A

1 1 3n2 − 2n − 1 n−1 − 2 > (3n + 1) 2 = n(n + 1)) >1 n n n n2

for n ≥ 2. Therefore, bn+1 (x) ≥ bn (x), x ∈ R.

Consider the approximating equation t n Xt = x + bn (Xsn )ds + BtH .

(3.5.21)

0

The functions bn are Lipschitz, therefore equation (3.5.21) has a unique strong solution Xtn on [0, T ], and Xtn ≤ Xtn+1 for any t ∈ [0, T ] a.s. Moreover, for any 0 < ε < H |Xtn1 (ω) − Xtn2 (ω)| ≤ C(ω)|t2 − t1 |H−ε + |t2 − t1 |, so, the set {Xn (·, ω), n ≥ 1} is tight for any ω ∈ Ω , P (Ω ) = 1. We obtain that Xtn (ω) ↑ Xt (ω), ω ∈ Ω , where the limit process X is continuous in t. Further, t t n ≤ b (X )ds − b(X )ds n s s 0 0 t t n |bn (Xs ) − bn (Xs )|ds + |bn (Xs ) − b(Xs )|ds. 0

0

(3.5.22)

276


Note that |bn (Xsn ) − bn (Xs )| = bn (Xs ) − bn (Xsn ) ≤ 2. Consider all the cases of mutual values of Xs , Xsn . (a) For Xsn < 0, Xs ∈ (0, n1 ] bn (Xs ) − bn (Xsn ) ≤ 21{Xs ∈(0, n1 ]} . (b) For Xsn < 0, Xs > n1 bn (Xs ) − bn (Xsn ) ≤ 21{Xs >0,Xsn n1 |bn (Xs ) − bn (Xsn )| ≤ 21{Xs >0,Xsn ∈[0, n1 ]} → 0 a.s., n → ∞. Further, t

t

|bn (Xs ) − b(Xs )|ds ≤ 2 0

0

1{Xs ∈[0, n1 ]} ds.

(3.5.23)

We obtain from (3.5.22) – (3.5.23) and (a)–(d) that t t t n lim | bn (Xs )ds − b(Xs )ds| ≤ 6 lim 1{Xs ∈(0, n1 )} ds n→∞ n→∞ 0 0 0 t 1{Xs =0} ds. =6 0

Therefore, to prove the existence of a strong solution of (3.5.20) it is sufficient T to prove that E 0 1{Xs =0} ds = 0, and in turn it is sufficient to establish the existence of bounded density ps (x), x ∈ R, s > 0 of the process Xs . For this purpose, return to Xsn : since the functions bn are continuously differentiable, then Xsn has a stochastic derivative, and on our probability space t Ds Xun bn (Xun )du, Ds Xtn = 1 + s

t = exp{ s bn (Xun )du} ≥ 1, since bn ≥ 0. whence Now we use the result of (Nua95): let the random variable F ∈ D1,2 , h ∈ Dom δ. Then F has a continuous h ∈ H, DF, hH = 0 a.s. and DF,h H and bounded density h f (x) = E 1{F >x} δ . DF, hH Ds Xtn

Now we put F := Xtn , ht (s) := 1{0≤s≤t} . Then t t t

exp bn (Xun )du DF, hH = 2αH 0 0 s

t bn (Xun )du |v − s|2α−1 dv ds ≥ CH t2H > 0. × exp v ht (s) Consider the function θ(s) = DF,h = ht (s)ξ, where ξ is a bounded random t H −1 variable, ξ = DF, hH , Eξ 2 < ∞. To establish that θ ∈ Dom δ, it is sufficient to verify that


277

T

(Ds ξ)2 ds < ∞.

E

(3.5.24)

0

Indeed, t

t

Ds ξ = Ds

t

bn (Xun )du

exp 0

0

exp

z

t

t

bn (Xun )du

v t

t

n = DF, h−2 · exp b (X )du × |v − z|2α−1 dvdz n u H 0 0 z t

t bn (Xun )du |v − z|2α−1 bn (Xun )du dv dz, × exp −1

v

z

where |bn (x)| ≤ 2n3 (since |bn (x1 ) − bn (x2 )| ≤ 2n3 |x1 − x2 |). Therefore, −2 −2H |Ds ξ| ≤ CH t ·4n3 ·C(H, n, t) (note that |bn (x)| ≤ 2n), and (3.5.24) holds. We obtain that θ ∈ Dom δ, and the density pnt (x) := pXtn (x) equals h n n pt (x) = E 1{Xt >x} δ . DXtn , ht H Let ψ(y) := 1[a,b] (y). Then from Proposition 2.1.1 (Nua95) b pnt (x)dx P {a ≤ Xn (t) ≤ b} = a

h E 1{Xtn >x} δ dx DXtn , ht H a

n Xt h ψ(x)dx · δ =E DXtn , ht H −∞ y h n ψ(z)dz = ϕ(y) = = E ϕ(Xt )δ DXtn , ht H −∞ 7 8 h =E Dϕ(Xtn ), DXtn , ht H H 1 ≤ E (Dϕ(Xtn ), hH ) CH t2H b −2H = C1,H t E 1{Xtn >x} δ(h) dx

b

=

a

≤ C1,H t

−2H

b

E|δ(h)|

dx. a

Therefore, ptn (x) ≤ C2,H t−2H , and P {a ≤ Xt ≤ b} = limn→∞ P {a ≤ Xtn ≤ b} = C2,H t−2H (b − a) for any continuous points of distribution function of Xt . Choosing a ↑ 0, b ↓ 0, we obtain density pt (0) ≤ C2,H t−2H . So, we have proved the following result: Theorem 3.5.14. Equation (3.5.20) with b(x) = sign x has a strong solution.

278


3.5.6 Estimates of Moments of Solutions for Regular Case and H ∈ (0, 1/2) Now we consider the case, when H ∈ (0, 1/2) and condition (3.5.19) holds. Then equation (3.5.1) has a unique strong solution. Suppose, in addition, 1 . Then the integral It = It (f ) that f ∈ Lp [0, T ] ∩ DpH [0, T ] for some p > H is continuous on [0, T ] (see Section 1.11). Evidently, the solution Xt is also continuous on [0, T ]. Let τN = inf{t > 0 : |Xt | ≥ N } ∧ T. Then |Xt∧τN | ≤ N. The solution admits the evident estimate t (1 + |Xs∧τN |)ds, |Xt∧τN | ≤ |X0 | + |It∧τN | + C 0

and for any r > 1 t E|Xt∧τN |r ≤ 3r |X0 |r + C r E( 0 (1 + |Xs∧τN |)ds)r + E|It∧τN |r t ≤ 3r |X0 |r + (6C)r tr + (6C)r E 0 |Xs∧τN |r ds · tr−1 + 3r E|It∧τN |r t ≤ g(t) + (6C)r tr−1 0 |Xs∧τN |r ds. Here

g(t) = 3r |X0 |r + (6C)r tr + 3r E|It∗ |r .

(3.5.25)

(3.5.26)

From the Gronwall inequality we obtain that E|Xt∧τN |r ≤ g(t)(1 + C1 tr e

C1 tr r

),

where C1 = (6C)r . Let N → ∞, then it holds that E|Xt |r ≤ g(t)(1 + C1 tr e

C1 tr r

).

(3.5.27)

Now, it follows from Theorem 1.10.6 and the part 2 of Remark 1.10.7, that there exists a constant C(H, p) such that It∗ r

1/r r+1 ≤ C(H, p) Γ G1p (0, t, f ). 2

(3.5.28)

It follows from (3.5.25)–(3.5.28) that C1 tr E|Xt |r ≤ g(t) 1 + C1 tr e r

(3.5.29)

r 1 r where g(t) = 3r |X0 |r + (6C)r tr + 3r C(H, p)r Γ ( r+1 2 ) (Gp (0, t, f )) . Estimate r (3.5.29) means that E|Xt | < ∞, t ∈ [0, T ], and this permits us to reduce the value of the multiplier g(t). Indeed, if we know that E|Xt |r < ∞, we can write the following inequality instead of (3.5.25):


t

E|Xt |r ≤ E |X0 | + |It | + C ≤ g1 (t) + C1 t

(1 + |Xs |)ds 0

279

r (3.5.30)

t

E|Xs | ds,

r−1

r

0

where from (1.9.10) and (1.10.4) g1 (t) = (3|X0 |)r + C1 tr + 3r sup0≤s≤t E|Is |r ≤ (3|X0 |)r + C1 tr + Cr (C(H, p))r (G1p (0, t, f ))r , Cr = 3r Cr , Cr = Hence, from the Gronwall inequality it follows that E|Xt |r ≤ g1 (t)(1 + C1 tr e

C1 tr r

2r/2 Γ ( r+1 2 ). π 1/2

).

(3.5.31)

Let estimate similarly E|Xt − Xt |r , 0 ≤ t < t ≤ T. t t E|Xt − Xt |r ≤ (2C)r E( t (1 + |Xs |)ds)r + 2r E| t f (s)dBsH |r ≤ (4C)r (1 + g1 (T )(1 + C1 T r e

C1 T r r

(3.5.32)

))(t − t)r + 2r Cr (Gp (t, t , f ))r ,

where Gp (t, t , f ) = C(H, p)(f Lp (t,t ) (t − t)H−1/p + f DH (t,t ) ), t t f DH (t,t ) = ( t ( x |f (x) − f (t)|(t − x)α−1 dt)2 dx)1/2 . Let f ∈ C β [0, T ] with α + β > 0, 0 < β < 1. Then f Lp (t,t ) · (t − t)H−1/p ≤ f C β [0,T ] (t − t)H , 1 (t − t)H+β f DH (t,t ) ≤ f C β [0,T ] · CH,β 1 with CH,β = (H + β − 1/2)−1 (2H + 2β)−1/2 . Therefore

E|Xt − Xt |r ≤ (4C)r (1 + g1 (T )(1 + C1 T r e + 2r Cr (CH,β,T )r (t − t)

rH

C1 T r r

))(t − t)r

(3.5.33)

,

1 where CH,β,T,p = C(H, p)(1 + CH,β T β ) f C β [0,T ] . Estimates (3.5.31) and (3.5.33) can be strengthened by appropriate choice of partitions of [0, T ]. More exactly, take t0 := (6C)−1 . Then for 0 ≤ t ≤ t0 it follows from (3.5.30) that t E|Xt |r ≤ g1 (t) + 6C 0 E|Xs |r ds, and from the Gronwall inequality

E|Xt |r ≤ g1 · e6Ct ≤ e · g1 ,

0 ≤ t ≤ t0 ,

r

where g1 = (3|X0 |) + 1 + Cr (G1 (0, T, f ))r . Further, for t0 ≤ t1 ≤ 2t0 t E|Xt |r ≤ 3r |Xt0 |r + 3r E| f (s)dBsH |r + (6C)r (t − t0 )r

t0

t

t

|Xs |r ds ≤ 3r g1 e+ Cr (G1 (0, T, f ))r +1+6C

+(6C)r (t−t0 )r−1 E t0

|Xs |r ds, t0

280


whence E|Xt |r ≤ g2 e6C(t−t0 ) ≤ g2 e, where g2 = 3r g1 e + Cr (G1 (0, T, f ))r + 1. Further, by induction, for kt0 ≤ t ≤ (k+1)t0 we have that E|Xt |r ≤ gk+1 e, where gk+1 ≤ 3r gk e+Br ≤ · · · ≤ (3r e)k (g1 +Br ) for Br = Cr (G1 (0, T, f ))r +1. The 1 2number of such steps on the interval [0, T ] does not exceed k = tT0 + 1 ≤ 6CT + 1. It means that for any 0 ≤ t ≤ T E|Xt |r ≤ (3r e)6CT +1 (g1 + Br ) ≤ (3e)(6CT +1)r (3r |x|r + 2 + 2Cr (G1 (0, T, f ))r ), (3.5.34) and similarly to (3.5.34) we obtain that E|Xt − Xt |r ≤ (4C) Dr (t − t)r + 2r Cr (G2 (t, t , f ))r , r

where Dr = 1 + (3e)(6CT +1)r (g1 + Br ) = 1 + (3e)(6CT +1)r (2 + (3|X0 |)r + 2Cr (G1 (0, T, f ))r ).

(3.5.35)

For f ∈ C [0, T ] with 0 < β < 1, H + β > 1/2 we have that β

E|Xt − Xt |r ≤ (4C)r (1 + (3e)(6CT +1)r (g1 + Br ))(t − t)r + + 2r Cr (CH,β,T,p ) (t − t)Hr , (3.5.36) r

whence E|Xt − Xt |r ≤ (4C)r Dr (t − t)r + 2r Cr (CH,β,T,p ) (t − t)Hr . r

(3.5.37)

3.5.7 The Estimates of the Norms of the Solution in the Orlicz Spaces The results of Subsections 3.5.7– 3.5.9 were motivated by the papers (KM06) and (KM07). Let the function U (x) = exp{x2 } − 1, (Ω, F, P ) be some probability space. Definition 3.5.15. The Orlicz space LU (Ω) generated by the function U (x) is the space of random variables ξ on (Ω, F), such that for some constant Cξ > 0 EU ( Cξξ ) < ∞. The next result is proved in the monograph (BK00). Theorem 3.5.16. The Orlicz space LU (Ω) is the Banach space with respect to the Luxemburg norm 2 ξ ξU = inf{r > 0 : E exp ≤ 2}. r2


281

Let T be some set of parameters. Definition 3.5.17. The random process Y = {Yt , t ∈ T} belongs to the space LU (Ω), if for any t ∈ T the random variable Yt belongs to this space. Introduce the notations a := (3e)6CT +1 , b := 3|X0 |a, c := 3aG1 (0, T, f ), √ √ √ √ d2 c1 = c 2, d := max {c1 , a e, b e}, h := (3 + 2 2) exp{ 2c 2 }. Theorem 3.5.18. Let the conditions of the Theorem 3.5.11 hold and {Xt , t ∈ [0, T ]} be the solution of equation (3.5.1). Then for any ε > 0 ε2 P {|Xt | ≥ ε} ≤ h exp − 2 . (3.5.38) 2c Proof. The next inequality follows from (3.5.34): r+1 2cr E|Xt |r ≤ 2ar + br + √ 1 Γ . 2 π

(3.5.39)

Furthermore, from the Stirling formula Γ (u) =

√ 2πuu−1/2 e−u eθ(u)

with θ(u)
l e, and obtain the inequality

√ ( ε )2 1 ( ε )2 1 2 1 P {|Xt | ≥ ε} ≤ 2 aε l e + εb l e + 2 2 exp − εl 2e √ 2

2

2 = exp ln εb εl 1e + 2 exp ln aε εl 1e + 2 2 exp − εl √ Let ln aε ∨ ln εb ≤ − 12 , i.e. ε ≥ (a ∨ b) e.

1 2e

. (3.5.43)

Then

√ √ ε2 ε2 P {|Xt | ≥ ε} ≤ (3 + 2 2) · exp − 2 = (3 + 2 2) · exp − 2 . (3.5.44) 2el 2c 2

d ≥ 1, so it follows from Evidently, (3.5.44) holds for ε ≥ d. But exp 2c 2

(3.5.44) that inequality (3.5.38) holds for any ε > 0.

Theorem 3.5.19. Let the conditions of Theorem 3.5.11 hold and {Xt , t ∈ [0, T ]} be the solution of equation (3.5.1). Then the random variable Xt belongs to the Orlicz space LU (Ω), and its norm in this space admits an estimate √ Xt U ≤ 2(1 + h)c. Proof. The statement of this theorem follows from Theorem 3.5.18 and the next lemma, which is the partial case of Theorem 2.3.4 (BK00).

Lemma 3.5.20. Letξ be a random variable such that for any ε > 0 ε2 P {|ξ| ≥ ε} ≤ C1 exp − 2C for some Ci > 0, i = 1, 2. Then ξ ∈ LU (Ω) and 2 2 √ ξU ≤ 2(1 + C1 )C2 . Now √ introduce the notations B1 := 2( e)−1/2 CH,β,T,p , B2 := 4C √ce T 1−H , B3 := 4C(1 + 2a + b)T 1−H ,

√ √ 3 ∨B4 , B6 := B4 e, B4 := B1 + B2 , B5 := (2 2 + 1) exp B2B 2 4 √ B7 := 2(1 + B5 )B6 . Theorem 3.5.21. Let {Xt , t ∈ [0, T ]} be the solution of equation (3.5.1), the conditions of Theorem 3.5.11 hold and the function f ∈ C β [0, T ] with H + β > 1/2. Then for any ε > 0 and 0 ≤ t < t ≤ T ε2 P {|Xt − Xt | ≥ ε} ≤ B5 exp − 2 (3.5.45) 2B6 (t − t)2H and

Xt − Xt U ≤ B7 (t − t)H .

(3.5.46)


283

Proof. Inequality (3.5.46) follows from (3.5.45) and Theorem 3.5.19. So we prove only (3.5.45). It follows from inequalities (3.5.37) and (3.5.40) that √ √ E|Xt − Xt |r ≤ 2B1r rr/2 + 2 2B2r rr/2 + B3r (t − t)rH . So, for any ε > 0 √ r √ r r/2 B r P {|Xt − Xt | ≥ ε} ≤ r (t − t)rH 2 Bε1 + 2 2 Bε2 + ε3 √ r r ≤ 2 2 Bε4 rr/2 + Bε3 (t − t)rH . Now we substitute r =

1 e

2

ε

and obtain for r ≥ 1, i.e. for

(t −t)H B

ε > (t − t)H B6 , that for q(ε) :=

4 2

ε 2(t −t)2H B62

√ P {|Xt − Xt | ≥ ε} ≤ 2 2 exp {−q} + exp ln Be3 (t − t)H · q . √ H Also, let ln Bε3 (t − t)H ≤ − 12 , i.e. ε ≥ e(t √ − t) BH3 . Then for ε ≥ ε0 , where ε0 := (B3 ∨ B4 ) e(t − t) we have an inequality √ P {|Xt − Xt | ≥ ε} ≤ (2 2 + 1) exp {−q(ε)} ≤ B5 exp {−q(ε)} . If 0 < ε < ε0 , then √ P {|Xt − Xt | ≥ ε} ≤ (2 2 + 1) exp {q(ε0 )} exp {−q(ε)} = B5 exp {−q(ε)} .

Corollary 3.5.22. Let {Xt , t ∈ [0, T ]} be a solution of equation (3.5.1) for which the conditions of Theorem 3.5.11 hold and the function f ∈ C β [0, T ] with H + β > 1/2. Then for any λ ∈ R 2 λ 2 2H B (t − t) E exp {λ|Xt − Xt |} ≤ 2 exp . 4 7 This statement follows directly from (3.5.46) and the following lemma, which is a partial case of Lemma 2.3.4 (BK00). Lemma 3.5.23. If the random variable ξ belongs to the space LU (Ω) , where U (x) = = exp x2 − 1, then for any λ ∈ R E exp {λ|ξ|} ≤ 2 exp

λ2 ξ2U 4

.

284


3.5.8 The Distribution of the Supremum of the Process X on [0, T ] First we present some facts from the theory of stochastic processes that belong to the Orlicz spaces. Let T be some infinite set of parameters, Y = {Yt , t ∈ T} be some real-valued process from the space LU (Ω), where U (x) = exp{x2 } − 1, supt∈T Yt U < ∞, ρY (t, s) = Yt − Ys U be a semi–metric on T. Let the space (T, ρY ) be separable and the process Yt be a separable process on (T, ρY ). Also, let N (ε) = N (T, ε) be the metric capacity of (T, ρY ), i.e. the minimal number of closed balls of radius ε that cover (T, ρY ). Note that N (ε) → ∞ as ε → 0. (See also the beginning of Section 1.10, where similar questions are discussed for Gaussian processes.) The next theorem is a partial case of Theorem 3.3.4 (BK00). Theorem 3.5.24. Let the following assumption holds: ε0 (ln(1 + N (ε)))1/2 dε < ∞, 0

where ε0 := sup ρY (t, s). Then the random variable sup |Yt | belongs to the t,s∈T

t∈T

space LU (Ω) and sup |Yt |U ≤ K := inf Yt U + t∈T

t∈T

e2 θ(1 − θ)

θε0

(ln(1 + N (ε)))1/2 dε < ∞,

0

(3.5.47) where 0 < θ < 1 and N (θε0 ) > e2 − 1. Remark 3.5.25. The statement of the theorem remains true if we replace N (ε) by any function N1 (ε) ≥ N (ε). Remark 3.5.26. Under the assumption of Theorem 3.5.24 for any ε > 0 we have that ε2 P {sup |Yt | ≥ ε} ≤ 2 exp − 2 , (3.5.48) K t∈T where K was defined in (3.5.47). Inequality (3.5.48) is implied by the following one: if ξ ∈ LU (Ω), then for any ε > 0 ε2 . (3.5.49) P {|ξ| ≥ ε} ≤ 2 exp − ξU In turn, inequality (3.5.49) is a partial case of Theorem 3.3.4 (BK00). Theorem 3.5.27. Let {Yt , t ∈ T = [a, b]} be the separable process from the space LU (Ω), and let there exist σ = σ(h) : [0, b − a] → R+ , increasing and continuous in h, and such that σ(0) = 0 . Also, let sup Yt − Ys U ≤ σ(h),

|t−s|≤h

(3.5.50)


and

ε0

0

285

1/2 3(b − a) du < ∞, ln 1 + (−1) 2σ (u)

(u) is the inverse function to σ(u), and ε0 = σ(b − a). where σ Then sup |Yt | ∈ LU (Ω) and the following estimate holds: (−1)

t∈[a,b]

! ! ! ! ! ! ! sup |Yt |! ≤ K1 := inf Yt U t∈T !t∈[a,b] ! U

+

e2 θ(1 − θ)

θ ε0

0

1/2 3 b−a du. ln 1 + 2 σ (−1) (u)

Here θ is any number from the interval ⎞ ⎛ 3(b−a) σ 2(e 2 −1) ⎝0, 1 ∧ ⎠. σ(b − a)

(3.5.51)

(3.5.52)

Moreover, for any ε > 0, we have the estimate ε2 P { sup |Yt | ≥ ε} ≤ 2 exp − 2 . K1 t∈[a,b]

(3.5.53)

Proof. The claim follows from Theorem 3.5.24 with T = [a, b]. Indeed, according to (3.5.50), the process Y is separable in the space ([a, b], ρY ), where b−a + 1, and for ρY (t, s) = Yt − Ys U . It is easy to see that N (u) ≤ 2σ(−1) (u) 0 < u ≤ ε0 , i.e. for

θ ε0

≥ 1, we have that N (u) ≤

b−a σ (−1) (u)

1/2

(ln (1 + N (u)))

du ≤

0

0

θ ε0

b−a 3 2 σ (−1) (u) .

Therefore

1/2 3 b−a du. ln 1 + 2 σ (−1) (u)

The inequality N (θ ε0 ) > e2 − 1 can be reduced, according to Remark 3.5.25, 3(b−a) to the inequality 2σ(−1) (θε0 ) > e2 −1, i.e. to (3.5.52). Inequality (3.5.53) follows now from (3.5.48).

Theorem 3.5.28. Let the condition of Theorem 3.5.11 hold, {Xt , t ∈ T = [0, T ]} be the solution of equation (3.5.1) and 0 ≤ t1 < t2 ≤ T . Then the random variable sup |Xt | ∈ LU (Ω), and t1 ≤t≤t2

! ! H ! ! γ (t2 − t1 ) ! sup |Xt |! ≤ (h + 1)c1 + e2 CH,γ θ− 2H =: L, !t ≤t≤t ! 1−θ 1 2 U where 0 < θ
0 2 ε P { sup |Xt | ≥ ε} ≤ 2 exp − 2 . L t1 ≤t≤t2

(3.5.55)

Proof. We use Theorem 3.5.27 with [a, b] = [t1 , t2 ]. The process Xt is continuous with probability 1, hence is separable. It follows from (3.5.46) that σ(h) = B7 hH . It is easy to see that in this case ε0 = σ(t2 − t1 ) and 1/2 θε t2 −t1 I(θ ε0 ) := 0 0 ln 1 + 32 σ(−1) du (u) σ(−1) (θε0 ) 1/2 −t t = HB7 0 ln 1 + 32 2 v 1 v H−1 dv.

(3.5.56)

Since for 0 < γ ≤ 1 and x > 0 ln(1 + x) =

1 1 xγ ln((1 + x)γ ) ≤ ln(1 + xγ ) ≤ , γ γ γ

we obtain from (3.5.56) the following estimate for any 0 < γ < 2H: γ σ(−1) (θε0 ) H−1− γ γ 2 dv · (t − t ) 2 I(θ ε0 ) ≤ 32 2 HB7 · γ1 0 v 2 1 γ γ = CH,γ (t2 − t1 ) 2 (σ (−1) (θ ε0 ))H− 2 . Evidently,

1

1

ε0 ) = θ H σ (−1) ( ε0 ) = θ H (t2 − t1 ). σ (−1) (θ Therefore γ

I(θ ε0 ) ≤ CH,γ θ1− 2H (t2 − t1 )H .

(3.5.57)

Now the proof follows from (3.5.56)–(3.5.57) and Theorems 3.5.19 and 3.5.27.

Remark 3.5.29. Estimate (3.5.54) demonstrates that up to constants the estimates for distribution of the supremum of the process X are of the same form as similar estimates for the Gaussian process (see (Fer74), for example). Corollary 3.5.30. Let {Xt , t ∈ [0, T ]} be a solution of equation (3.5.1) under the conditions of Theorem 3.5.11 and 0 ≤ t1 < t2 ≤ T. Then for any p ≥ 1 we have an estimate p p1 ≤ Cp · L, (3.5.58) E sup |Xt | t1 ≤t≤t2

1

where L is defined in (3.5.54) and Cp = 2 p

√ p 2 .

Proof. This statement follows from Theorem 3.5.28. Indeed, it was established in Lemma 2.33 (BK00), that for the random variable ξ ∈ LU (Ω) , U (x) = 1 exp x2 −1 and p ≥ 1 (E |ξ|p ) p ≤ Cp ξU . Now (3.5.58) follows from (3.5.54).


287

Corollary 3.5.31. Let {Xt , t ∈ [0, T ]} be the solution of equation (3.5.1), 0 ≤ t1 < t2 ≤ T. Then for any λ ∈ R 2 2 λ L E exp λ sup |Xt | ≤ 2 exp . 4 t1 ≤t≤t2 This estimate follows from Theorem 3.5.27 and Lemma 3.5.23. 3.5.9 Modulus of Continuity of Solution of Equation Involving Fractional Brownian Motion Definition 3.5.32. We say that the C-function U (x) (C-function is continuous, even, convex function that increases in x > 0 and is zero at the zero point) satisfies the ∆2 -condition if there exist such constants x0 > 0 and L0 > 1, that U 2 (x) ≤ U (L0 x) for x ≥ x0 . Example 3.5.33. The function U (x) = exp x2 −1 satisfies ∆2 -condition with √ x0 := 0 and L0 := 2. Theorem 3.5.34. Let {Yt , t ∈ T} be a stochastic process from the Orlicz space LU (Ω), where the function U (x) satisfies the ∆2 -condition with constants x0 , L0 , and let Z0 := x0 ∨ L0 . Let ρY (t, s) = Yt − Ys U , t, s ∈ T be a semi-metric generated by Y. Also, let (T, ρY ) be the separable space and the process Y be the separable process in the space (T, ρY ) . Put ε0 := sup ρY (t, s), let N (u) be t,s∈T

the minimal number of closed u-balls covering (T, ρY ) , N1 (u) ≥ N (u), u > 0 and let N1 (u) increase in u. If for any ε > 0 ε U (−1) (N1 (u)) du < ∞,

q(ε) :=

(3.5.59)

0

then for any ε ∈ (0, ε0 ) such that N1 (ε0 ) ≥ U (Z0 ), and for any x ≥ Z0 √ 3+ 2 |Yt − Ys | ≥x ≤ . (3.5.60) P sup U (x) 0 0, 0 < γ ≤ 1 it is easy to

12 1 −1 v H +2 dv 2

0

H

γ

1 + 2γ/2 v − 2H 1

γ2

0 1

+2

ln

≤ γ/2

H

1 γ γx ,

γ− 2 γ 1 − 2H

δ t2 − t1

1

dv ≤ γ − 2

H− γ2

δ t2 − t1

≤ CH,γ

H

δ t2 − t1

H− γ2

and the proof immediately follows from (3.5.64) and (3.5.65).

, (3.5.67)

4 Filtering in Systems with Fractional Brownian Noise

4.1 Optimal Filtering of a Mixed Brownian–Fractional-Brownian Model with Fractional Brownian Observation Noise Consider the real-valued signal process Xt and the observation process Yt defined by the following system of equations: ⎧ t N t ⎪ ⎪ ⎪ Xt = η + a(s, X )ds + bi (s, Xs ) dWsi s ⎪ ⎪ ⎪ 0 0 ⎪ i=1 ⎪ ⎨ M t (4.1.1) cj (s)dBsHj , t ∈ [0, T ] , ⎪+ ⎪ 0 ⎪ j=1 ⎪ ⎪ t t ⎪ ⎪ ⎪ ⎩ Yt = ξ + A (s, Xs ) ds + C(s)dBsH , 0

0

where {W i , 1 ≤ i ≤ N } are independent Wiener processes, {B Hj , 1 ≤ j ≤ M } are independent fractional Brownian motions with Hurst indices Hj ∈ ( 12 , 1), B H is an fBm with Hurst index H ∈ ( 12 , 1), all the processes are mutually independent, random initial conditions (η, ξ) are independent of each other and independent of all the processes (W i , B Hj , B H ), the functions a, b, A : [0, T ] × R → R, cj , C : [0, T ] → R are measurable in their variables and satisfy the conditions that are sufficient for the existence of pathwise integrals w.r.t. corresponding fBms. The problem is to construct the optimal filter of the signal X according to the observation Y , which will be expressed in terms of the conditional expectation πt (X) := E(Xt /FtY ), where FtY := σ{Ys , 0 ≤ s ≤ t}. Note that the partial cases of this problem were considered in (KLeBR99), (KLeBR00), where N = 1, cj = 0 (see also (KKA98b), (LeB98)), and in (Pos05), where bi = 0. Suppose that the following condition holds: (i) the function C ∈ LH 2 (R), does not vanish and 1/C(s) is bounded on [0, T ], cj ∈ LH (R). 2

292


Here we use the approach to the solution of optimal filtering problem developed in (KLeBR00) but simplify it and modify it in accordance with our model (4.1.1). Introduce the following processes, connected with fBm B H : t t Zt∗ := lH (t, s)C −1 (s)dYs = lH (t, s)D(s, Xs )ds 0

0

t

lH (t, s)dBsH = Jt (D) + MtH ,

+

(4.1.2)

0

t where Jt (D) = 0 lH (t, s)D(s, Xs )ds, MtH is the Molchan martingale, introduced in (1.8.5), D(s, Xs ) = A(s, Xs )/C(s). Recall that 1 (5) lH (t, s) = CH s−α (t − s)−α 1{0<s 0. Then the integral t (5) Jt (D) = Γ (1 − α)CH E(s, X)s−2α ds 0

has a.s. bounded variation, so, it follows from (4.1.2) that Zt∗ is the semiH martingale w.r.t. the σ-field Ft := σ{ η, ξ, Xs , Wsi , 1 ≤ i ≤ N, Bs j , 1 ≤ j ≤ M, Ys , 0 ≤ s ≤ t } and admits the representation t (5) Zt∗ = MtH + CH E(s, X)s−2α ds, CH = Γ (1 − α)CH , (4.1.3) 0

and, in addition, Zt∗ is FtY -adapted. Further, let

4.1 Optimal Filtering of a Mixed Model in Fractional Brownian Observation

νt := Zt∗ − CH

t

πs (E(s, X))s−2α ds,

293

(4.1.4)

0

where πs (Q) := E(Q/FsY ), It follows from (4.1.4) and (4.1.3) that t νt = CH

E(s, X) − πs (E(s, X)) s−2α ds + MtH .

0

Moreover, for 0 ≤ s < t ≤ T E(νt − νs /FsY ) = E(MtH − MsH /Fs /FsY ) = 0, t and the integral CH 0 E(s, X) − πs (E(s, X)) s−2α ds is continuous and has a bounded variation. Hence νt = t1−2α , where νt is calculated w.r.t. the filtration {FtY , 0 ≤ t ≤ T }. So, ν is a continuous Gaussian martingale w.r.t. this filtration. (Evidently, ν is adapted to this filtration since Zt∗ is adapted.) Further we need the following evident result. Lemma 4.1.1. Any square-integrable martingale {Mt , FtY , t ∈ [0, T ]} with M0 = 0 admits the representation t Mt = ϕs dνs , t ∈ [0, T ], 0

where the process ϕt is FtY -adapted and E

t 0

ϕ2s s−2α ds < ∞.

The next statement is proved similarly to Theorem 18.11 (Ell82); see also t := E(Xt /FtY ). Theorem 3 (KLeBR00). For any integrable process X, let X Theorem 4.1.2. Let {St , Ft , t ∈ [0, T ]} be the semimartingale of the form

t

αs ds + mt , t ∈ [0, T ],

St = S0 + 0

T where ES02 < ∞, E 0 αs2 ds < ∞ and {mt , Ft , t ∈ [0, T ]} be a square inte t grable martingale with mutual bracket m, M H t = 0 λs s−2α ds. Then the process {St , t ∈ [0, T ]} satisfies the following stochastic differential equation: t t X)−Ss πs (E(s, X)) dνs , t ∈ [0, T ]. s +CH Ss E(s, S t = S0 + α s ds+ λ 0

0

Proof. If we define the FtY -adapted process Mt := St − S0 −

t

α s ds, 0

(4.1.5)

294


t then for s ≤ t E(Mt − Ms /FsY ) = E(St − Ss /FsY ) − s α u du t t Y = E( s αu du/Fs ) − s α u du = 0. Therefore, Mt is a FtY -square-integrable martingale. By Lemma 4.1.1, Mt admits the representation t Mt = ϕs dνs , t ∈ [0, T ], (4.1.6) 0

whence St = S0 +

t

α s ds + 0

t

ϕs dνs . 0

Now we use the same reasonings as in Theorem 18.11 (Ell82). On the one hand, with the help of (4.1.3) the product St Zt∗ can be decomposed by the Itˆ o formula as t t ∗ H −2α Ss (dMs + CH E(s, X)s ds) + Zs∗ (αs ds + dms ) St Zt = 0

0

t

+

λs s−2α ds,

0

whence

∗ St Zt∗ = S t Zt =

t

X)s−2α (CH Ss E(s,

0

s s−2α )ds + N 1 , +α s Zs∗ + λ t

(4.1.7)

where Nt1 is continuous FtY -martingale. On the other hand, using (4.1.4) and (4.1.5)–(4.1.6) we obtain the following decomposition for St Zt∗ : t St Zt∗ = (Zs∗ α s + CH Ss πs (E(s, X))s−2α + ϕs s−2α )ds + Nt2 , (4.1.8) 0

where Nt2 is a continuous FtY - martingale. It follows from (4.1.7)–(4.1.8) that X) = CH Ss πs (E(s, X)) + ϕs , whence the proof s + CH Ss E(s, N 1 = N 2 and λ follows.

Now we can establish the form of the optimal filter in our model. In this t H order we rewrite all the integrals 0 cj (s)dBs j , 1 ≤ j ≤ M, by using Theorem 1.8.3, in the form t t c Hj cj (s)dBs = KHjj (t, s)dMsHj , 0

0

where (7)

C KH (t, s) = CH

t

C(u)uα (u − s)α−1 du. s

4.2 Optimal Filtering in Conditionally Gaussian Linear Systems

295

Further, consider for any t ∈ [0, T ] the process Xut

u

:= η +

a(s, Xs )ds + 0

+

i=1

M j=1

N

0

u

u

bi (s, Xs )dWsi

0

c

KHjj (t, s)dMsHj , 0 ≤ u ≤ t,

(4.1.9)

so that Xtt = Xt from (4.1.1). Evidently, {Xut , 0 ≤ u ≤ t} is the semimartingale with respect to the filtration {Ft , 0 ≤ t ≤ T }. Therefore we can use Theorem 4.1.2 to establish the following result. Theorem 4.1.3. Let φ ∈ Cb2 (R), πt (φ) = E(φ(Xt )/FtY ), Lts φy (x) = a(s, y)φ (x) +

N M 1 2 c bi (s, y)φ (x) + βj (KHjj (t, s))2 s−2αj φ (x), 2 i=1 j=1

0 ≤ s ≤ t ≤ T, βj = 1 − 2αj , and the conditions (i)–(iii) hold. Then the equation for the optimal filter πt has the form: t t πt (φ) = π0 (φ) + πs (Lts φXs (Xst ))ds + CH (πst (φE) − πst (φ)πs (E))dνs , 0

0

where πst (φE) = E(φ(Xst )E(s, X)/FsY ), πst (φ) = E(φ(Xst )/FsY ), πs (E) = πs (E(s, X)). Proof. It follows from (4.1.9) that Xt is a “boundary” value of the semimartingale Xut , 0 ≤ u ≤ t. u N 1 t 2 Since φ(Xut ) = φ(η) + (φ (Xst )a(s, Xs ) + φ (Xs )bi (s, Xs ) 2 i=1 0 M N u c (1 − 2αj )(KHjj (t, s))2 s−2αj φ (Xst ))ds + φ (Xst )bi (s, Xs )dWsi + +

j=1 M u j=1

0

i=1

0

s = 0 in our case, the proof immediately φ (Xst )KHjj (t, s)dMsHj and λ c

follows from Theorem 4.1.2.

4.2 Optimal Filtering in Conditionally Gaussian Linear Systems with Mixed Signal and Fractional Brownian Observation Noise Now we suppose that the real-valued signal process {Xt , t ∈ [0, T ]} and the observation process {Yt , t ∈ [0, T ]} satisfy the following system of equations:

296


⎧ t N t ⎪ ⎪ ⎪ Xt = η + a(s)Xs ds + bi (s)dWsi ⎪ ⎪ ⎪ 0 0 ⎪ i=1 ⎪ ⎨ M t + cj (s)dBsHj , ⎪ ⎪ ⎪ j=1 0 ⎪ ⎪ t t ⎪ ⎪ ⎪ ⎩ Yt = ξ + A(s)Xs ds + C(s)dBsH , t ∈ [0, T ] 0

(4.2.1)

0

where {W i , 1 ≤ i ≤ N } are independent Wiener processes, {B Hj , 1 ≤ j ≤ M } are independent fBms with Hurst indices Hj ∈ ( 12 , 1), B H is an fBm with Hurst index H ∈ ( 12 , 1), W i , B Hj , B H are mutually independent, random initial conditions (η, ξ) are independent of all the processes (W i , B Hj , B H ), a, bi , cj , A, C : [0, T ] → R, are bounded measurable functions which satisfy the conditions sufficient for the existence of Lebesgue integrals, corresponding pathwise integrals w.r.t. fBms and Itˆ o integrals w.r.t. Wiener processes. As before we suppose that C(s) does not vanish and 1/C(s) is a bounded function on [0, T ]. Suppose also that the conditional distribution π0 := E(η/ξ) is Gaussian. Under these assumptions the mutual distribution of the pair (X, Y ) is well-defined, and this pair is conditionally Gaussian pair, i.e., for any 0 ≤ t1 ≤ t2 ≤ · · · ≤ tn ≤ t ≤ T the joint conditional distribution of (Xt1 , . . . , Xtn ) given FtY is Gaussian. The same is obviously true for the system ((X, E(·, X)), Y ). Then for any t ∈ [0, T ] the optimal filter πt has a Gaussian distribution, which can be completely characterized t := E(Xt /F Y ) and conditional variance by its conditional mean value X t 2 2 Y σ t := E((Xt − Xt ) /Ft ), t ∈ [0, T ]. Denote D(s) := A(s)/C(s) and note that now E(s, xs ) = D(s)xs for any x ∈ C 1/2− [0, T ]. Suppose that the following version of the condition (iii) is now fulfilled: α (L1 [0, T ]) for any x ∈ C 1/2− [0, T ]. Evidently, the (iii”) D(s)xs s−α ∈ I0+ set of such D(s) is nonempty. Consider for any t ∈ [0, T ] the semimartingale that is similar to (4.1.9): Xut

u

:= η +

a(s)Xs ds + 0

N i=1

u

bi (s)dWsi

+

0

M j=1

0

u

c

KHjj (t, s)dMsHj ,

0 ≤ u ≤ t, so that Xtt = Xt from (4.2.1). Denote σ 02 := E(η 2 /ξ) − (π0 )2 . Lemma 4.2.1. For all t ∈ [0, T ] u t = π0 + s ds + CH X a(s)X 0

=

σ 02

D(s) σs2 dνs ,

0

σ t2

u

t

a(s)( σst )2 ds

+2 0

+

N i=1

0

t

b2i (s)ds

(4.2.2)

4.2 Optimal Filtering in Conditionally Gaussian Linear Systems

+

M

t

(1 − 2αj ) 0

j=1

t

+ CH

c (KHjj (t, s))2 s−2α ds

− (1 −

t

2 2α)CH

297

D(s) σs2 s−2α ds

0

t 2 t 2 t ) dνs , D(s) (X σs2 (X s ) Xs − (Xs ) Xs − 2 s

0

(4.2.3)

st )(Xs − X s )/FsY . where ( σst )2 := E (Xst − X Proof. By using Theorem 4.1.2 and independence of {W i , M Hj } of M H we obtain that u u t t Y s πs (E))dνs Xu := E(Xu /Fu ) = π0 + a(s)Xs ds + CH (Xs E(s, X) − X 0

u

= π0 +

0

s ds + CH a(s)X

u

D(s) σs2 dνs ,

0

0

whence (4.2.2) follows. Now we apply the Itˆ o formula to the semimartingale t , 0 ≤ u ≤ t} : {X u u u t 2 2 t st dνs 2a(s)(Xs )Xs ds + 2CH D(s) σs2 X (Xu ) = (π0 ) + 0

0

u

2 + CH (1 − 2α)

D(s) σs2 s−2α ds, t ∈ [0, T ].

(4.2.4)

0

On the other hand,

u

(Xut )2 = η 2 +

2a(s)(Xst )Xs ds 0

+

N

b2i (s)ds +

0

i=0

+

u

M

0

j=1

N u

2bi (s)Xst dWsi +

0

i=1

u

(1 − 2αj ) M 0

j=1

u

(KHjj (t, s))2 s−2αj ds c

c

2KHjj (t, s)Xst dMsHj ,

whence 2 t 2 (X u ) = E(η /ξ) +

u

t 2a(s)(X s )Xs ds +

0

+ + CH

M j=1

u

N i=1

u

(1 − 2αj ) 0

u

b2i (s)ds

0

(KHjj (t, s))2 s−2αj ds c

t 2 t 2 D(s)((X s ) Xs ) − (Xs ) Xs )dνs , t ∈ [0, T ].

(4.2.5)

0

Subtracting (4.2.4) from (4.2.5) for u = t, we obtain (4.2.3).

298


4.3 Optimal Filtering in Systems with Polynomial Fractional Brownian Noise In all previous filtering models the noises were presented as the integrals w.r.t. a Wiener process or w.r.t. an fBm, but everywhere with nonrandom integrands. In this section we consider the simple case of a random integrand. Let the signal process {Xt , t ∈ [0, T ]} and the observation process {Yt , t ∈ [0, T ]} are defined by the following system of equations: Xt = η +

t

a(s, Xs )ds + 0

Yt = ξ +

N

bn (BtH1 )n ,

n=1

t

t

C(s)dBsH2 , t ∈ [0, T ],

A(s, Xs )ds + 0

0

where (BtH1 , BtH2 , t ∈ [0, T ]) are fBms with Hurst indexes Hi ∈ (1/2, 1), a, A : [0, T ] × R → R are measurable functions and bn , 1 ≤ n ≤ N are real numbers. Suppose that the pair (η, ξ) does not depend on (B H1 , B H2 ), condition (i) holds for the function C, and condition (ii) holds for E(s, xs ) with any x ∈ C H1 − [0, T ]. First we try to present the power term (BtH1 )n in the form t t (BtH1 )n = Mn (t, s)dBs + Kn (t, s)ds, 0

0

where B is the underlying Wiener process (it means that BtH1 = t mH1 (t, s)dBs with the kernel mH1 (t, s) from Section 1.8), Mn (t, s) and 0 Kn (t, s) are some Fs -adapted random functions. Evidently, for n = 1 t BtH1 = mH1 (t, s)dBs . 0

Therefore M1 (t, s) = mH1 (t, s), K1 (t, s) = 0. For arbitrary n ≥ 2 (BtH1 )n = t t n(BsH1 )n−1 mH1 (t, s)dBs + 0 n(n−1) (BsH1 )n−2 (mH1 (t, s))2 ds. 2 0 So, the signal process can be presented as

t

Xt = η +

a(s, Xs )ds + 0

N

bn 0

=η+

t

a(s, Xs )ds + 0

t

M (t, s)dBs + 0

Kn (t, s)ds 0

t

K(t, s)ds, 0

where Mn (t, s) = n(BsH1 )n−1 mH1 (t, s), Kn (t, s) =

t

Mn (t, s)dBs +

n=1

t

n(n − 1) H1 n−2 (Bs ) (mH1 (t, s))2 , 2

4.3 Optimal Filtering with Polynomial Fractional Noise

M (t, s) =

N n=1

bn Mn (t, s), K(t, s) =

Suppose that B, M H2 t =

N

bn Kn (t, s).

n=1 t

t

λs s−2α ds, where MtH2 =

l (t, s)dBsH2 . 0 H2

0

Consider the family of semimartingales s s Xst := η + a(u, Xu )du + Mn (t, u)dBu + 0

299

0

s

Kn (t, u)du,

0

s ∈ [0, t], t ∈ [0, T ]. Let the function φ ∈ C 2 (R). Then the process φ(Xst ), s ∈ [0, t] is a semimartingale with the representation s s φ(Xst ) = φ(η) + φ (Xut )M (t, u)dBu + Ltu (φ(·))du, 0

0

where Ltu (φ(·)) = (a(u, Xu ) + K(t, u)) φ (·) + 12 φ (·)(M (t, u))2 . So, Theorem 4.1.3 gives the following representation for the optimal filter πs (φ(Xst )) : s s πu (φ (Xut )M (t, u)λu ) πs (φ(Xst )) = π(φ(η)) + πu (Ltu (φ(Xut )))du + 0

0

+ CH (πu (φ(Xut )E(u, X)) − πu (φ(Xut ))πu (E)) dνu . If we put s = t then the equation for the optimal filter πt (φ(Xt )) receives the form: t t t t πu (φ (Xut )M (t, u)λu ) πu (Lu (φ(Xu )))du + πt (φ(Xt )) = π(φ(η)) + 0

0

+ CH (πu (φ(Xut )E(u, X)) − πu (φ(Xut ))πu (E)) dνu .

5 Financial Applications of Fractional Brownian Motion

5.1 Discussion of the Arbitrage Problem 5.1.1 Long-range Dependence in Economics and Finance As mentioned in the paper (WTT99), long-range dependence in economics and finance has a long history and is an area of active research (e.g., see (Lo91), (CKW95)). The importance of long-range dependent processes as stochastic models lies in the fact that they provide an explanation and interpretation of an empirical law that is commonly referred to as the Hurst law or Hurst effect. In short, for a given set of observations n Xi , n ≥ 1, and sample variance {Xi , i ≥ 1} with partial sum Y (n) = S 2 (n) = n−1

n

i=1

(Xi − n−1 Y (n))2 , n ≥ 1, the rescaled adjusted range statistic

i=1

or R/S-statistic is defined by 1 R t t (n) = , n ≥ 1. max Y (t) − Y (n) − min Y (t) − Y (n) 0≤t≤n S S(n) 0≤t≤n n n Hurst in (Hur51) found that many naturally occurring empirical records appear to be well represented by the relation E ((R/S)(n)) ∼ c1 nH as n → ∞, with typical values of the Hurst parameter H ∈ (1/2, 1), and c1 a finite positive constant not depending on n. But in the case when the observations come from a short-range dependent model, then E (R/S(n)) ∼ c2 n1/2 as n → ∞, where c2 does not depend on n. The discrepancy between these two relations is called the Hurst effect or Hurst phenomenon. The analysis of the R/S-statistic, provided in (WTT99), (TTW95) and (TT97), leads to the recommendation to use a diverse portfolio of time-domain-based and frequency-domain-based graphics and statistical methods, including the graphical R/S-method, the modified R/S-statistic (Lo91) and Whittle’s approach. Also, another (possibly,

302


surprising) recommendation is: in the case when statistical analysis cannot be expected to provide a definitive answer concerning the presence or absence of long-range dependence in asset price returns, a more revealing and also much more challenging approach to tackle this problem consists of providing a mathematically rigorous physical “explanation” for the presence or absence of the long-range dependence phenomenon in stock returns. 5.1.2 Arbitrage in “Pure” Fractional Brownian Model. The Original Rogers Approach Suppose that we establish that the existence of long-range dependence on the financial market in which we operate, and we must model a share price process using long-range dependence of returns. So, we can try to replace the clasical log-Brownian model (BlSc73) dSt = St (µdt + σdWt ), t ≥ 0 involving some Brownian motion W by the model involving fBm B H : dSt = St (µdt + dBtH ), t ≥ 0,

(5.1.1)

where H ∈ (1/2, 1). Three main problems arise immediately: what will be the class of financial strategies, what will be the kind of stochastic integral w.r.t. the fBm used in the model and is such a model arbitrage-free or not? There has been wide discussion on these topics and we will present here the main (in our opinion) results and conclusions. It seems that the first attempt to construct arbitrage on the financial market that is modeled with fBm, was made by Rogers (Rog97). He did not use geometric fBm like (5.1.1) but fBm itself, and exploits its stationary properties, obtains an arbitrage possibility and immediately concludes that fBm is an absurd model for finance markets (as we shall see later, the situation is not so dramatic). The notion of arbitrage that will be used (only in this subsection) is the following: we say that an arbitrage exists on the interval [a, b] if there is some trading strategy whose gains process {ηt , a ≤ t ≤ b} satisfies the following conditions: (a) ηa = 0; (b) ηt ≥ −β for all a ≤ t ≤ b and some β > 0; (c) P {ηb > 0} > 0. The brief description of the Rogers construction is the following. Suppose that (Ω, F, {Ft , t ∈ R+ }, P ) is a filtered probability space and {Xt , t ∈ R+ } is a continuous integrable adapted process. For any a > 0 and 0 ≤ t < b define τ (t, b, a) := inf{u > t : Xu − Xt ∈ / [−a, a]} ∧ b. Lemma 5.1.1. Let, for any rational a, b, t with t < b, E Xτ (t,b,a) − Xt /Ft = 0 a.s. Then X is a local martingale.

(5.1.2)

5.1 Discussion of the Arbitrage Problem

303

Proof. For any stopping time T ≤ c equality (5.1.2) can be extended to (5.1.3) E Xτ (T,b,a) − XT /FT = 0 a.s. Indeed, we can approximate T by a sequence of stopping times T (n) = 2−n ([2n T ] + 1), taking discretely many rational values and decreasing to T. Now fix N ∈ N, define τ := τ (0, N, N ), fix ε > 0 and define the stopping times σ0ε = 0, ε σn+1 := inf{u > σnε : Xu − Xσnε ∈ / (−ε, ε)} ∧ τ,

n ≥ 0. Evidently, σnε ↑ τ as n → ∞. From (5.1.3) it follows that ε E Xσn+1 /Fσnε = Xσnε . Since |Xσnε | ≤ N + |X0 |, we have that for any n ≥ 0 Xσnε = E Xτ /Fσnε , and as ε → ∞ we obtain that for any t < N Xt∧τ = E(Xτ /Ft ), which means that Xt∧τ is a martingale, and this is sufficient.

Now, as we have seen in Section 1.15, fBm B H is not a semimartingale (in particular, it is not a local martingale) unless H = 1/2. As a conclusion, we obtain from Lemma 5.1.1 that for fBm {BtH , t ∈ R} the following is true: if we define for any n ∈ N the process H H nH Yn (t) := (Bt·2 , −n −21−n − B−21−n )2 H

B 0 ≤ t ≤ 1, H ∈ (1/2, 1), and Yn := F−2 −n , then there exist a > 0 and ε > 0, such that P {E(Yn (τn )/Yn−1 ) ≥ ε} ≥ ε

where τn = inf{t > 0 : Yn (t) ∈ [−a, a]} ∧ 1. Note that by the scaling properties of B H the sequence {Yn , n ∈,Z} of C[0, 1]-valued random variables is stationary and even ergodic since n σ{Yk , k ≤ −n} is trivial. The ergodic theorem guarantees that P {E(Yn (τn )/Yn−1 ) ≥ ε for infinitely many n ≥ 0} = 1. Consider the period (−21−n , −2−n ] and call this period “promising” if E(Y (τn )/Yn−1 ) ≥ ε. There will be infinitely many “promising” periods. The investment strategy is the following one: we invest a unit amount in each “promising” period but immediately sell our holding and wait until the end of the period if Yn goes out of [−a, a] during the promising period. So, the gain ζn made during a “promising” period satisfies the relations −a ≤ ζn ≤ a,

304


E(ζn /Yn−1 ) ≥ ε, and for the “nonpromising” period ζn = 0. Denote accumulated gain by ηn = ζk . Then we can take λ > 0 sufficiently small such k≤n

that

E(e−ληn /Yn−1 ) ≤ e−ληn−1 .

Therefore, e−ληn is a nonnegative supermartingale convergent a.s. to 0. If we stop ηn at the first time ν when ηn < −a, then P {ν < ∞} ≤ exp{−λa} < 1 and on the event {ν = +∞} ηn → +∞. Finally, the arbitrage strategy can be described as follows: invest a unit amount in Y (which is the same as investing an amount 2nH in B H during period n ) in each “promising” period until either ηn has risen to 1 or falls to below −a. The former happens at least with probability 1 − exp{−λa}, and the resulting gain is 1, and if the latter happens we lose at most 2a. If the latter happens we invest 1/2 in each “promising” period until either ηn has risen to 1 or has fallen below 5a 2 . If the latter happens we lose at most 3a, and invest 1/4 in each “promising” period until either η has risen to 1 or has fallen below 13a 4 and so on. To continue in this way, successively halving the stake when things go badly, we shall eventually be successful and make a net gain of at least 1, and the worst that can happen is that our wealth meantime could fall to 4α, so we have arbitrage in our definition. 5.1.3 Arbitrage in the “Pure” Fractional Model. Results of Shiryaev and Dasgupta Consider a (B(r), S(r))-market with Bt (r) = ert , H St (r) = eµt+σBt , t ≥ 0,

(5.1.4)

H ∈ (1/2, 1). Let for simplicity µ = r, σ = 1. We construct a portfolio H π = (β, γ) with βt = 1 − e2Bt , γt = 2(BtH − 1). For such a portfolio we have that the corresponding capital Xtπ equals H 2 Xtπ = βt Bt (r) + γt St (r) = ert eBt − 1 . From the Itˆ o formula (2.7.5) for a pathwise integral w.r.t. fBm, Xtπ = 0

t

t H H 2 H rers eBs − 1 ds + 2 ers+Bs eBs − 1 dBsH t t0 = βs dBs (r) + γs dSs (r), 0

0

(5.1.5)


305

and (5.1.5) exactly means that the strategy π is self-financing strategy in usual sense. So, for this portfolio X0π = 0 and Xtπ > 0 a.s. for any t > 0, and everyone understands that it is an arbitrage possibility (in any appropriate definition). This is Shiryaev’s example (Shi01). A very close result was obtained by Dasgupta (Das98). He considered a one-dimensional portfolio πt , 0 ≤ t ≤ 1, the same model as in (5.1.4), defined discounted gain as t Gt = π(s)Bs−1 (r) σdBsH + (µ − r)ds , 0

and determined arbitrage as the following possibility: (a) there exists α ∈ R such that P {Gt ≥ α, 0 ≤ t ≤ 1} = 1; (b) P {Gt ≥ 0} = 1, (c) P {G1 > 0} > 0. Now, consider the particular case µ = r and the particular portfolio H H (5.1.6) πt = 2ert+σBt eσBt − 1 . With portfolio (5.1.6) the gain process equals t t H H H 2eσBs eσBs − 1 σdBsH = e2σBs 2σdBsH Gt = 0

t

−2

H

eσBs

0

2 H H H 2σdBsH = e2σBt − 1 − 2eσBt + 2 = eσBt − 1 .

0

Of course, we obtain arbitrage possibility. As a conclusion, we see that the “pure” continuous-time model based on fBm is not arbitrage-free, if the arbitrage possibility is defined in any appropriate terms. The same fact is emphasized in PhD thesis of Cheridito (Che01b), the paper of Salopek (Sal98); see also an early discussion on arbitrage with fBm in finance in (MS93). Now we can discuss discrete-time models and “mixed” models (the latter ones are much more promising). 5.1.4 Mixed Brownian–Fractional-Brownian Model: Absence of Arbitrage and Related Topics Let {Wt , t ≥ 0} be a standard Wiener process and {BtH , t ≥ 0} be an fBm with the Hurst index H ∈ (1/2, 1) , both defined on a filtered probability space Ω, F, {Ft , t ≥ 0}, P . a mixed version of the Black–Merton–Scholes model, i.e. a Consider B, S -market with a bond B and a stock S, where Bt = ert ,

H

St = eaWt +bBt

+ct

,

r, a, b, c ∈ R, t ∈ R+ .

(5.1.7)

For a given strategy (or a portfolio) π = {βt , γt , t ≥ 0} the capital {Xt , t ≥ 0} corresponding to this portfolio equals

306


Xt = Bt · βt + St · γt . We make the following assumptions about the strategy π: 1) π is a self-financing strategy, i.e. t t βs dBs + γs dSs ; Xt = X0 + 0

(5.1.8)

(5.1.9)

0

2) π is a Markov-type strategy, i.e. βt = β St , t , γt = γ St , t .

(5.1.10)

One needs to be accurate with condition (5.1.9), for it to reflect the real economic concept of “self-financing”. This entails that the meaning of the second integral in (5.1.9) should be specified clearly. We understand it now in the pathwise sense, i.e. as the following limit with probability 1: t n−1 γs dSs = lim γsk (Ssk+1 − Ssk ). 0

max|sk+1 −sk |→0

k=0

n−1

Here, the sum k=0 γsk (Ssk+1 − Ssk ) is an obvious formula for the capital, earned on the price variation of S with a piecewise buy-and-hold strategy t {˜ γt , t ∈ R+ } = {γsk , sk ≤ t < sk+1 , t ≥ 0}. Hence, the integral 0 γs dSs , as the capital earned on S with the continuous strategy {γt , t ∈ R+ }, agrees with the “fundamental moral” in the definition of self-financing conditions (for discussion on this topic see Section 5.2.2). We say that the strategy π has an arbitrage opportunity if there exists T > 0 such that X0 = 0,

XT ≥ 0 (P − a.s.),

P (XT > 0) > 0.

In the mixed model (5.1.7) with a = 0 and b = 0, some results in this direction have been obtained in the papers of (Ku99), (Che01b), (MV02), (Zah02a). More exactly, Kuznetsov (Ku99) established the absence of arbitrage under the condition of independence of processes W and B H . As we mentioned in Subsection 3.4.2, Cheridito (Che01b) proved that, for H ∈ (3/4, 1), the mixed model with independent W and B H is equivalent to the one with Brownian motion and hence it is arbitrage-free. Z¨ ahle (Zah02a) proved the absence of arbitrage in the general mixed model with independent Wiener process and the process of zero quadratic variation (Dirichlet processes, see, for example, (Fol81b)). In the mixed model, studied in the paper (MV02), there is no requirement of independence. Conversely, the absence of arbitrage is demonstrated under the condition that the process B H is connected with the process W as in formula (1.8.17). The main result of this subsection is that the mixed market is arbitragefree without any conditions on the dependence of W and B H , if we restrict ourselves to the self-financing Markov-type strategies with smooth β and γ.


307

Conditions of Self-Financing and Their Consequences Note that in the case of the Markov-type strategy (5.1.10), the process of capital Xt can be written as a function of price of the stock S at the moment t: (5.1.11) Xt = Φ St , t , where Φ(x, t) = ert · β(x, t) + x · γ(x, t).

(5.1.12)

We prove in this section that the self-financing assumption strongly restricts the class of possible functions Φ in (5.1.11). t In the case of γt = γ(St , t) with smooth γ(·, ·), the integral 0 γs dSs exists and it can be presented in the form t t t t a2 H γs dSs = a γs Ss dWs + b γs Ss dBs + c+ γs Ss ds, (5.1.13) 2 0 0 0 0 where the first integral on the right-hand side is the Itˆ o integral, the second integral is the pathwise Riemann–Stieltjes integral and the third one is the Riemann integral. Formula (5.1.13) gives the Itˆ o formula for an exponent of the mixed process. In addition, we shall refer in this subsection to the Itˆ o formula for processes with generalized quadratic variation (see Subsection 2.7.2). The Itô integral in (5.1.13) appears due to the choice of the left endpoint sk in the expression under the summation sign in (5.1.12). Such a choice is crucial for condition (5.1.9) to have the economic sense of self-financing. The t second integral 0 b γs Ss dBsH does not depend on the choice of inner points of the intervals. Theorem 5.1.2. Let the B, S -market be given by (5.1.7) with a = 0. Suppose also that for all t > 0 the support of the distribution of St coincides with supp(St ) = [0, +∞).

(5.1.14)

Then in the class of Markov-type strategies (5.1.10) with β(x, t), γ(x, t) ⊂ C 2 (0, +∞) × C 1 [0, +∞) the condition of self-financing (5.1.9) is equivalent to the following one: (i) There exists a function φ(x, t) ∈ C 2 (0, +∞) × C 1 [0, +∞) , which satisfies the equation φt (x, t) +

a2 2 x φxx (x, t) + r x φx (x, t) − r φ(x, t) = 0, 2

and the strategy (β, γ) can be expressed in terms of φ:

(5.1.15)

308


β(x, t) = e−rt φ(x, t) − x · φx (x, t) ; γ(x, t) = φx (x, t).

(5.1.16)

Remark 5.1.3. Condition (5.1.14) holds, for example, in the case when the processes W and B H are jointly Gaussian, and, hence, log(St ) = aWt + bBtH + ct, t ≥ 0 is a Gaussian process. Remark 5.1.4. Under condition (i) we have the identity Φ(x, t) = φ(x, t). Proof of Theorem 5.1.2. Below we use the Itô formula for processes with generalized quadratic variation; see (3.1.25), (3.1.26). Firstly, the Itˆ o formula holds for continuous processes with generalized bracket. Secondly, if the process Z has the usual bracket, then it has the same generalized bracket. Let consider the process St and prove that it has usual bracket. Indeed, n

∆Stk

2

=

k=0

n

k=0

+

+c tk+1

H

− eaWtk +bBtk +c tk

2

H 2 H bB +c tk+1 e2aWtk+1 e tk+1 − ebBtk +c tk

n k=0

+2

k+1

k=0 n

=

n

aWtk+1 +bBtH

e

bBtH +c tk k

eaWtk+1 e

eaWtk+1 − eaWtk

2

eaWtk+1 − eaWtk

H

e2bBtk +2c tk

bBtH

e

k+1

+c tk+1

H

− ebBtk +c tk

k=0

=: I1n + I2n + I3n . t Evidently, I2n → 0 Su2 a2 du, a.s. and in L2 (P ). Further, H H H bBtk+1 +c tk+1 − ebBtk +c tk ≤ ebBtk +c tk b ∆BtHk + c ∆tk e and the trajectories of B H belong to the class C H− [0, T ] with H > 1/2. Therefore I1n → 0 a.s., and the same is true for I3n . It means that the bracket of S has the form t

a2 Su2 du.

[S]t = 0

(5.1.17)

Let o formula (2.7.8) to the processes Bt β St , t and us apply the Itˆ St γ St , t from (5.1.8). We obtain the equalities t Bt β St , t − β 1, 0 = d Bu β Su , u 0 t t t β Su , u dBu + Bu βt Su , u du + Bu βx Su , u dSu (5.1.18) = 0 0 0 t 1 Su , u d[S]u , Bu βxx + 2 0


309

and t d Su γ Su , u St γ St , t − γ 1, 0 = 0 t t t γ Su , u dSu + Su γt Su , u du + Su γx Su , u dSu = 0 0 0 t 1 2γx Su , u + Su γxx Su , u d[S]u . + 2 0

(5.1.19)

Combining equations (5.1.18) and (5.1.19), we obtain: t t Xt − X 0 − β Su , u dBu − γ Su , u dSu 0

t

Bu βt Su , u +Su γt Su , u

= 0

t du +

0

Bu βx Su , u +Su γx Su , u dSu

0

+

1 t 2

Bu βxx Su , u + 2γx Su , u + Su γxx Su , u d[S]u .

(5.1.20)

0

Comparing equations (5.1.20) and (5.1.9), we conclude that the condition of self-financing of the strategy π = {βt , γt , t ∈ R+ } is equivalent to the equation t t Bu βt Su , u +Su γt Su , u du + Bu βx Su , u +Su γx Su , u dSu 0 0 1 t Bu βxx Su , u + 2γx Su , u + Su γxx Su , u d[S]u = 0, t > 0. + 2 0 (5.1.21) From the same Itô formula and definition of the process S, we obtain that t t 1 2 H S t = S0 + a Su du, Su d aWu + bBu + cu + 0 0 2 t where the integral 0 Su dWu exists as the usual Itô integral, and the integral t S dBuH exists as the limit of the Riemann–Stieltjes sums, because 0 u S ∈ C 1/2− [0, T ], B H ∈ C H− [0, T ], and 1/2 + H > 1. Substituting equation (5.1.17) into equation (5.1.21), we obtain that equation (5.1.21) can be rewritten as t

Bu βt Su , u + Su γt Su , u du

0

t + 0

Bu βx Su , u + Su γx Su , u Su d aWu + bBuH + c + a2 /2 u

310

+


a2 2

t

Bu βxx Su , u + 2γx Su , u + Su γxx Su , u Su2 du = 0. (5.1.22)

0

Let us take the quadratic variation of the both sides of (5.1.22). Evidently, the usual bracket of all Lebesgue integrals in (5.1.22) vanishes, and the bracket of the Itˆ o integral equals * · + Bu βx Su , u + Su γx Su , u Su d aWu = 0

= a2

t

Bu βx Su , u + Su γx Su , u

2

t

Su2 du.

0

t Establish now that the usual bracket of the process 0 Bu βx Su , u + Su γx Su , u Su d bBuH a.s. equals 0. In this order denote fu := b Bu βx Su , u + Su γx Su , u . Evidently, the trajectories of this process belong to the class C 1/2− [0, T ]. Further, from the estimate in Proposition 22 (FdP99), it follows that tk+1 ! ! 1/2+H−2δ H H fu dBu − ftk ∆Btk ≤ C f C 1/2−δ !B H !C H−δ ∆tk , tk

with constant C not depending on f and B H and such δ that 1/2+H −2δ > 1, i.e. δ < α/2. Therefore, n

+2

k=0 n

tk+1

fu dBuH

2

≤2

tk

ftk

n k=0

2

∆BtHk

2

tk+1

tk

fu dBuH − ftk ∆BtHk

2

n ! !2 1+2H−4δ 2 ∆tk ≤ 2 C 2 f C 1/2−δ !B H !C H−δ

k=0

k=0

+2

n

ftk

2

∆BtHk

2

→ 0 a.s.

k=0

From all these estimations and (5.1.22) we obtain 2

t

a

2 2 Su du = 0. Bu βx Su , u + Su γx Su , u

(5.1.23)

0

Since (5.1.23) holds for all t > 0, we easily deduce that Bu βx Su , u + Su γx Su , u = 0 for all u > 0 and almost all (a.a.) ω ∈ Ω.

(5.1.24)


311

Substituting (5.1.24) into (5.1.22) we obtain another equation for all t > 0: t a2 t Bu βt Su , u + Su γt Su , u du + Bu βxx Su , u + 2γx Su , u 2 0 0 Su , u Su2 du = 0. + Su γxx This means that the equality

Bu βt Su , u + Su γt Su , u (5.1.25) 2 a Su , u + 2γx Su , u + Su γxx Su , u Su2 = 0 Bu βxx + 2

holds for all u > 0 and a.a. ω ∈ Ω. Condition (5.1.14) of the theorem ensures that equations (5.1.24) and (5.1.25) may hold if and only if (5.1.26) Bt βx x, t +x γx x, t = 0; 2 a Bt βt x, t +x γt x, t + Bt βxx x, t +2γx x, t +x γxx x, t x2 = 0, 2 (5.1.27) for all t > 0, x > 0. The last relations mean that the strategy β(St , t), γ(St , t) is selffinancing if and only if the pair β(x, t), γ(x, t) satisfies equations (5.1.26), (5.1.27). Now assume that condition (i) of the theorem holds. Substituting β and γ from (5.1.16) into (5.1.26) and (5.1.27) we obtain an identity 0 = 0 in the first equation and identity (5.1.15) in the second one. Conversely, if (5.1.26) and (5.1.27) hold, we set φ(x, t) := Bt · β(x, t) + x · γ(x, t). For such function φ we obtain from (5.1.26) that φx (x, t) = Bt · βx (x, t) + γ(x, t) + x · γx (x, t) = γ(x, t), β(x, t) = Bt−1 φ(x, t) − x · γ(x, t) = e−rt φ(x, t) − x · φx (x, t) , i.e. we come to (5.1.16). Substituting β and γ from (5.1.16) into identity (5.1.27), we obtain that φ(x, t) satisfies equation (5.1.15).

Remark 5.1.5. Let the process {Zt , t ≥ 0} be defined on Ω, F, {Ft , t ≥ 0}, P with Z0 = 0 and [Z] ≡ 0, where [Z] stands for usual bracket, i.e. quadratic variation. Then it is not hard to see that Theorem 5.1.2 is valid for the B, S˜ -market with Bt = ert ,

S˜t = eaWt +Zt +ct ,

˜ if only condition (5.1.14) holds for the process S.

312


Absence of Arbitrage Theorem 5.1.6. Let the B, S -market be given by (5.1.7) with a = 0. Let the support of the distribution of St coincides with supp(St ) = [0, +∞)

(5.1.28)

for all t > 0. Then there is no arbitrage strategy in the class of self-financing Markovtype strategies (5.1.10) with β(x, t), γ(x, t) ⊂ C 2 (0, +∞) × C 1 [0, +∞) . Proof. Theorem 5.1.2 states that for any strategy in the class, described in the theorem, the process of capital Xt is given by Xt = φ St , t , where φ satisfies the equation φt (x, t) +

a2 2 x φxx (x, t) + r x φx (x, t) − r φ(x, t) = 0. 2

(5.1.29)

Suppose that an arbitrage strategy exists. So, there exists T > 0 such that X0 = 0,

XT ≥ 0 (P − a.s.).

(5.1.30)

Together with (5.1.28) conditions (5.1.30) are equivalent to the following ones: φ(1, 0) = 0,

φ(x, T ) ≥ 0

∀ x > 0.

(5.1.31)

We are going to prove that φ ≡ 0 is the only function that satisfies (5.1.29) and (5.1.31) simultaneously. Hence, it would mean that there is no arbitrage strategies in the given class. Let us use the standard approach in solving equation (5.1.29). Suppose the function φ satisfies equation (5.1.29) with boundary conditions (5.1.31). Then a new function η(z, t), given by η(z, t) = θ(az, T − t),

z ∈ R, t ∈ [0, T ],

where θ(z, t) = e−(α z+β t) φ (ez , t) ,

α=

r 1 − , 2 a2

β=−

r2 a2 + 2, 8 2a

satisfies a heat equation ηt (z, t) =

1 η (z, t) 2 zz

(5.1.32)


313

with additional conditions ∀z ∈ R

η(z, 0) ≥ 0,

η(0, T ) = 0.

Here, an inverse change is given by

1 r φ(x, t) = x( 2 − a2 ) · e

2

2

r − a8 + 2a 2

·η

(5.1.33)

ln(x) ,T − t . a

The continuous solution of equation (5.1.32) is well known and has the form (z − ξ)2 − 12 η(z, t) = η(ξ, 0) · (2πt) · exp − dξ, 2t R which together with boundary conditions (5.1.33) gives η ≡ 0 and, therefore, φ ≡ 0.

Convergence of Lebesgue–Stieltjes Integrals to the Integral w.r.t. fBm In this subsection, we use Theorem 1.15.3 and prove a theorem which establishes the convergence in probability of integrals with respect to B H, β from (1.15.15) to the integral with respect to fBm. Theorem 5.1.7. Let the process f be such that for some ε > 0 and for a.a. ω∈Ω f (·, ω) ∈ C 2(1−H)+ε [0, T ]. (5.1.34) Then

T

P f (u) dBuH, β − →

0

T

f (u) dBuH

as

β → 0+,

0

P

→ denotes the convergence in probability. where − Proof. For any N > 0 we introduce the step process of the form fN (u) =

N

f (uk−1 )1[uk−1 ,

uk ) (u),

u ∈ [0, T ),

fN (T ) = f (uN ),

k=1

where

kT , 0 ≤ k ≤ N. N Then the following obvious inequality holds: T T H, β H f (u) dBu − f (u) dBu 0 0 T H, β T H, β H ≤ f (u) − fN (u) dBu + fN (u) d Bu − Bu 0 0 T + fN (u) − f (u) dBuH =: I1 (N, β) + I2 (N, β) + I3 (N ). 0 uk =

314


1 We shall establish that for the subsequence Nβ such that Nβ = following convergence holds: P → 0, I1 Nβ , β −

P I2 Nβ , β − → 0,

P I3 Nβ − → 0 as

T

2

β 1/2

the

β → 0+.

Condition (5.1.34) is equivalent to the relation: there exists a finite random variable K = K(ω) such that P -a.s. ∀ 0 ≤ x < y ≤ T we have λ

|f (x) − f (y)| ≤ K |x − y|

(5.1.35)

with λ = 2(1 −H) + ε. Consider I1 Nβ , β . We use (5.1.33), (5.1.34), (5.1.35) to obtain: T H, β f (u) − fNβ (u) dBu I1 Nβ , β = 0

(u−β)+ N uk H− 12 α−1 21 −H ˜ =C f (u) − f (uk−1 ) u (u − y) y dWy du u 0 k−1 k=1 N (u−α)+ uk 1 1 ˜ y du ≤CK (uk − uk−1 )λ uH− 2 (u − y)α−1 y 2 −H dW 0 uk−1 k=1

=: C K ζ1 (N, β). ˜ is now the underlying Wiener process (before it was denoted B, where W but now B is bond process). From now on C means a constant, the value of which is not interesting for us. Without loss of generality we may assume that β < T /2. Let estimate the mathematical expectation of ζ1 Nβ , β :

Eζ1 Nβ , β ≤ β

λ 2

N

≤β

N

≤β

λ 2

uk

u

(u−β)+ α−1 21 −H ˜ E (u − y) y dWy du 0 1/2

(u−β)+

α

(u − y)2H−3 y 1−2H dy 1/2

1−β/T

(1 − y)

2H−3 1−2H

y

dy

0

≤Cβ

λ 2

N k=1

1/2

(1 − y)

2H−3 1−2H

0

du

0

uk−1

k=1

u

H− 12

uk−1

k=1 λ 2

uk

y

uk

uH−1 du

uk−1

1/2

1−β/T

(1 − y)

2α

2H−3

dy + 2

dy

1/2

12 λ ≤ C β 2 1 + β 2α−1 , Substituting λ = 2(1 − H) + ε in (5.1.36) we obtain

(5.1.36)


12 Eζ1 Nβ , β ≤ C α1−H+ε/2 1 + β 2α−1 = O β ε/2 → 0,

315

β → 0+.

P

Hence, I1 (Nβ , β) − → 0 as β → 0+. Let consider I2 Nβ , β . N β I2 Nβ , β = f (uk−1 ) BuH,k β − BuHk − BuH,k−1 − BuHk−1 k=1

N f (uk ) − f (uk−1 ) · BuH, β − BuH + f (T ) B H, β − BTH ≤ T k k k=1

≤K

N

uk − uk−1

λ H, β · Buk − BuHk + f (T ) BTH, β − BTH .

k=1

P P The term f (T ) BTH, β − BTH − → 0 because BTH, β − → BTH as β → 0+. Denote λ H, β N Bu − BuH . With the help of Theorem ζ2 (Nβ , β) := k=1 uk − uk−1 k k 1.15.3, the mathematical expectation of BtH, β − BtH can be estimated in the following way: ⎧ tH , t 0. For N =

T

β 1/2

2

, ρ = ε/2 and λ = 2(1 − H) + ε we obtain

from (5.1.37) that N λ λ Eζ2 Nβ , β ≤ β 2 E BuH,k β − BuHk ≤ β 2 [Nβ ] + 1 o β α−ρ =

=o β

k=1 2(1−H)+ε−1 + 2

(α− 2ε ) = o(1) → 0,

β → 0+.

P

→ 0 as β → 0+. Hence, I2 (Nβ , β) − Finally, it follows from Theorem 2.1.7 that T T fNβ (u) dBuH − f (u) dBuH → 0 I3 Nβ = 0 0 a.s., and hence in probability, as β → 0+.

316


The Capital Process as a Limit of Semimartingales Let the B, S -market be given by (5.1.7) and a Markov-type strategy ˜ γ˜ ) be self-financing for this market. Then the capital, based on this strat(β, egy, is given by t t γ˜ Ss , s dSs . β˜ Ss , s dBs + Xt = X 0 + 0

0

For β > 0 and the given β(·, ·), γ(·, ·) consider the processes H, β

Stβ = eaWt +bBt and

t

Xtβ = X0 +

β˜ Ssβ , s dBs +

0

t

+ct

γ˜ Ssβ , s dSsβ .

(5.1.38)

0

The Itô formula and definition of B H, β imply that the process X β can be rewritten as Xtβ = X0 +

t

r Bs β˜ Ssβ , s + b (BsH, β )s + c Ssβ γ Ssβ , s ds

0

t

+a

Ssβ γ Ssβ , s dWs

(5.1.39)

0

with (BsH, β )s = CH αsα (6)

(s−β)+

˜ u, (s − u)α−1 u−α dW

0

which means that X β is a semimartingale at least if the following condition holds: T 2 E Ssβ γ˜ Ssβ , s ds < ∞. (5.1.40) 0

˜ ·), γ˜ (·, ·) satisfy the Theorem 5.1.8. Let H ∈ (3/4, 1) and the pair β(·, assumptions: (ii)

∀t ≥ 0

˜ t), γ˜ (·, t) ∈ C 1 (R) β(·,

(iii) ∀ T, L > 0 there exists K = K(T, L) > 0 such that 1 ˜ ˜ s) + |˜ γ (x, t) − γ˜ (x, s)| ≤ K |t − s| 2 , ∀ |x| ≤ L, t, s ∈ [0, T ]. β(x, t) − β(x, (iv)

∀ T > 0 there exist M = M (T ) > 0 and N = N (T ) > 0 such that ˜ N γx (x, t)| ≤ M (1 + |x| ), ∀ t ∈ [0, T ]. βx (x, t) + |˜ P

Then Xtβ − → Xt as β → 0+ for any t ∈ [0, T ].


317

Remark 5.1.9. Evidently, conditions (ii)–(iv) imply (5.1.40) and the pair B, S β can be regarded as a new stock market with a price of the stock P

→ St as being a semimartingale. It follows from Theorem 1.15.2 that Stβ − β → 0+ at any moment t ≥ 0. If, additionally, condition (5.1.14) holds for ˜ γ˜ ∈ (C 2 × C 1 )(R+ ), then the strategy β(S ˜ β , s), γ˜ (S β , s) is selfS β and β, s s financing and the market B, S β is arbitrage-free. In this case the process X β is a process of capital in this market. Proof of Theorem 5.1.8. Using (5.1.38), (5.1.39) and (5.1.9), we may write Xtβ − Xt t t t = γ Ssβ , s dSsβ − γ Ss , s dSs β˜ Ssβ , s − β˜ Ss , s dBs + 0 0 0 t t β β f (s) − f (s) ds + a g (s) − g(s) dWs =r 0 0 t t t β a2 g (s) − g(s) ds, +b g β (s) dBsH, β − g(s) dBsH + c + 2 0 0 0 where

f β (s) = ers β˜ Ssβ , s , g β (s) = Ssβ γ˜ Ssβ , s ,

f (s) = ers β˜ Ss , s , g(s) = Ss γ˜ Ss , s .

To prove that Xtβ → Xt , it is enough to establish that

t

P f β (s) − f (s) ds − → 0;

(5.1.41)

0

t

P g β (s) − g(s) dWs − → 0; 0 t t P g β (s) dBsH, β − → g(s) dBsH ; 0 0 t β P g (s) − g(s) ds − → 0, as β → 0+.

(5.1.42) (5.1.43) (5.1.44)

0

t The convergence in (5.1.41), (5.1.42) and (5.1.44) holds if 0 f β (s) 2 2 t P P → 0 and 0 g β (s) − g(s) ds − → 0 as β → 0+, which, in turn, − f (s) ds − follows immediately from the relations 2 E f β (s) − f (s) ≤ C β 2α , 2 E g β (s) − g(s) ≤ C β 2α , which will be proved in Lemma 5.1.10. Let us prove (5.1.43). Obviously, the following inequality holds:

(5.1.45) (5.1.46)

318


t ≤ g(s) dBsH, β 0

t t β H, β H g (s) dBs − g(s) dBs 0 0 t t β H, β H g (s) − g(s) dBs . − g(s) dBs + 0

(5.1.47)

0

The trajectories of the process η(t) = aWt + bBtH + ct a.s. belong to the 1 space C 2 − [0, T ]. It means that for any ρ > 0 there exists K1 (δ, ω) > 0 such that 1

|η(t) − η(s)| ≤ K1 (δ, ω) |t − s| 2

−ρ

∀ t, s ∈ [0, T ]. (5.1.48) 1 Let us prove that the process g(s) =: ψ η(s), s also belongs to C 2 − [0, T ] P -a.s. Indeed, it follows from (iii) that ∀ L > 0 there exists K2 (L) > 0 such that 1

|ψ(x, t) − ψ(x, s)| ≤ K2 (L) |t − s| 2 ,

,

∀ |x| ≤ L, t, s ∈ [0, T ].

(5.1.49)

˜, N ˜ >0 It follows from the definition of ψ(x, s) and (iv) that ∃ M ˜ exp{N ˜ |x|}, |ψx (x, s)| ≤ M

∀ s ∈ [0, T ].

(5.1.50)

Now we use (5.1.48)–(5.1.50) to obtain ψ η(t), t − ψ η(s), s ≤ ψ η(t), t − ψ η(s), t + ψ η(s), t − ψ η(s), s ≤ sup |ψx (x, t)| · |η(t) − η(s)| + ψ η(s), t − ψ η(s), s |x|≤|η(s)| ∨|η(t)|

˜ sup |η(t)| K1 (δ, ω) |t − s| ˜ exp N ≤M

1 2 −δ

+ K2

t∈[0,T ]

1

sup |η(t)| |t − s| 2 t∈[0,T ]

≤ K3 (δ, ω) |t − s|

1 2 −δ

,

where

˜ sup |η(t)| K1 (δ, ω) + T δ K2 ˜ exp N K3 (δ, ω) = M t∈[0,T ]

sup |η(t)| . t∈[0,T ]

For any H ∈ (3/4, 1) it is possible to find ε = ε(H) > 0 such that C [0, T ] ⊂ C 2(1−H)+ε [0, T ]. So, we can apply Theorem 5.1.7 to the first term on the right-hand side of (5.1.47) and obtain its convergence to 0 in probability. Consider the second term on the right-hand side of (5.1.47). Using (5.1.46) we obtain, as in (5.1.36), that 1 2−


319

t β H, β g (s) − g(s) dBs E 0 t (s−β)+ β ˜ y ds g (s) − g(s) · C sα ≤E (s − y)α−1 y −α dW 0 0 12 t (s−β)+ 2 ≤C sα E g β (s) − g(s) · (s − y)2α−2 y −2α dy ds 0

0

≤Cβ

α

(1−β/T )+

2α−2 −2α

(1 − y)

y

12

t

sH−1 ds

dy 0

0

12 1 ≤ C β α 1 + β 2α−1 = O β 2α− 2 , β → 0+, t which means that E 0 f β (s) − f (s) dBsH, β → 0 if H ∈ 34 , 1 .

Lemma 5.1.10. Inequalities (5.1.45) and (5.1.46) are true for every s ∈ [0, T ] and β ∈ (0, 1) with a constant C that does not depend on β and s. Proof. We prove only inequality (5.1.46) since (5.1.45) can be established similarly. Denote a function ψ(x, s) : = exp{x} · γ exp{x}, s . Then the processes g β (s) and g(s) are given by g β (s) = ψ aWs + bBsH, β + cs, s , g(s) = ψ aWs + bBsH + cs, s . We obtain from the H¨ older inequality that 2

∂ψ(x, s) H, β β 2 H E g (s) − g(s) ≤ E sup ∂x · b Bs − Bs

≤ b2

x∈I1 (s, β, ω)

1 ∂ψ(x, s) 2p p H, β 1 H 2q q E B − B , E sup s s ∂x x∈I1 (s, β, ω)

(5.1.51)

where p, q > 1, 1/p + 1/q = 1 and I1 (s, β, ω)

= x : aWs + min(bBsH, β , bBsH ) < x − cs < aWs + max(bBsH, β , bBsH ) . 1 In the case when 2q < 1−H (which is equivalent to the inequality p > we can use Theorem 1.15.2 and derive the following estimation: 2qH q1 1 s , s 0 and all β ∈ (0, 1). Summarizing (5.1.51), (5.1.52) and (5.1.56) 1 we obtain that for any p > 2α 2 E g β (s) − g(s) ≤ C β 2α−1/p ,

s ∈ [0, T ], β ∈ (0, 1),

(5.1.57)

where constant C does not depend on p or β. Since p is arbitrary, inequality (5.1.46) follows from (5.1.57).


321

5.1.5 Equilibrium of Financial Market. The Fractional Burgers Equation Definition 5.1.11. The financial market described by equation (5.1.7) is in are the equilibrium on [0, T ] if both the kernel ϕt and likelihood ratio dQ dp Ft functions of t and Wt , twice differential in both the variables, and do not depend on the path of {Ws , 0 ≤ s < t} (for the corresponding notations see Subsection 3.2.3). This definition generalizes the usual definition of equilibrium of the financial market involving only the Wiener process (see (HC93)), where the path’s independence of dQ dp Ft is declared, and the kernel ϕt equals simply e(t, Wt ), up to a constant multiplier. Theorem 5.1.12. If the financial market is in equilibrium, then ϕt satisfies the Burgers equation 1 −ϕ(s, x)ϕx (s, x) = ϕt (s, x) + ϕxx (s, x). 2 t t Proof. Let ϕt = g(t, Wt ), and 0 ϕs dWs − 12 0 ϕ2s ds = G(t, Wt ), where g, G ∈ C 2 (R+ × R). Then t 1 t 2 g(s, Ws )dWs − g (s, Ws )ds = G(t, Wt ), t ∈ [0.T ]. 2 0 0

From the Itˆ o formula, t t 1 (Gt (s, Ws ) + Gxx (s, Ws ))ds + Gx (s, Ws )dWs . G(t, Wt ) = 2 0 0

From here g(s, Ws ) = Gx (s, Ws ), − 12 g 2 (s, Ws ) = Gt (s, Ws ) + 12 Gxx (s, Ws ), or, simply, g(s, x) = Gx (s, x), − 12 g 2 (s, x) = Gt (s, x) + 12 Gxx (s, x). Further, g2 (s, x) = G22 (s, x), − 12 g 2 (s, x) = Gt (s, x) + 12 gx (s, x). Therefore, 1 (s, x), gt (s, x) = Gtx (s, x), −g(s, x)gx (s, x) = Gtx (s, x) + gxx 2 whence the proof follows.

Remark 5.1.13. It is easy to see that the “principal” kernel θt = ϕt t−α satisfies the equation 1 sα+1 θ(s, x)θx (s, x) = αθ(s, x) + sθt (s, x) + s θxx (s, x), 2 s > 0, x ∈ R, and α = H − 1/2, which can be called, in this connection, the fractional analog of the Burgers equation. (Recall that the usual Burgers equation has the form ut = uxx + uux .)

322


5.2 The Different Forms of the Black–Scholes Equation on the Fractional Market 5.2.1 The Black–Scholes Equation for the Mixed Brownian–Fractional-Brownian Model Consider a mixed version of the Black–Merton–Scholes model (5.1.7) with the value process Xt , described by (5.1.8), and self-financing strategies, defined by (5.1.9)–(5.1.10). Consider C(t, St ), the price of a European call option with striking price K at time t ∈ [0, T ]. Suppose that C ∈ C 1 [0, T ] × C 2 (R), then S(t)) := C(T − t, S(t)) according to the Itˆ we can present the function C(t, o formula from Theorem 2.7.2 as t 2 (u, Su ) + cC (u, Su )Su + C a Su C C(t, S(t)) = C(0, x) + t S S 2 0 +

2 a Su2 du Css 2

t

+a 0

S (u, Su )Su dWu + b C

t

CS (u, Su )Su dBuH .

(5.2.1)

0

Now, let the portfolio on value process consist of one option and an amount of −δ of underlying assets. The number −δ will be specified later. The value − δS. of this portfolio equals X = C The jump in the value of this portfolio in one-step time equals 2 SdWu + bC SdB H S 2 du + aC − δdS = C + cC + a C dX = dC t S S S u 2 SS a2 S du + cSdu . − δ aSdWu + bSdBuH + 2 If we choose δ =

(5.2.2)

∂C ∂S

to eliminate the stochastic noise, then 2 · S 2 du. + a C dX = C t SS 2 The return of an amount X invested in bank account equals r Xdt at time dt. For absence of arbitrage, these values must be the same. Hence we obtain the traditional Black–Scholes equation 2 + 1 a2 S 2 ∂ C − r C + r SC = 0, C t S 2 ∂S 2 or, in terms of C(t, St ),

∂2C 1 −Ct + a2 S 2 2 − r C + r SCS = 0. 2 ∂S Remark 5.2.1. The same equation was obtained by Z¨ ahle (Zah02a) for the process Zt instead of aWt + bBtH , where Zt = aWt + bZt , and Z is continuous process with vanishing generalized quadratic variation.

5.2 The Different Forms of the Black–Scholes Equation

323

5.2.2 Discussion of the Place of Wick Products and Wick–Itˆ o–Skorohod Integral in the Problems of Arbitrage and Replication in the Fractional Black–Scholes Pricing Model This section appears as a result of the interesting discussion of the related problems contained in the papers (SV03) and (BH05). The fact of the existence of arbitrage in the “pure” fractional Brownian model is, to some degree, the consequence of the fact that the mathematical expectation of the stochastic integral w.r.t. fBm defined in the pathwise sense is nonzero (and you immediately obtain such an integral as a limit of the portfolio value created by step buy-and-hold strategies; we discussed this topic in Subsection 5.1.4). Note, however, that the arbitrage opportunity constructed by Rogers (Rog97) does not depend on any particular notion of integration. The same is true for the pre-limit arbitrage of the fractional Black–Scholes model considered in (Sot01). Nevertheless, many efforts were made to create the “pure” fractional model which will be “free of arbitrage”, with the help of the stochastic integral constructed by Wick products. We mention in this connection the papers (HO03), (EvH03), (Ben03), (BO03), (BHOS02), (Mis04). Now we present the corresponding list of propositions for alternative definitions of portfolio values and self-financial conditions: (i) the price of risky asset S is modeled by a geometric fBm and is the solution of the equation dSt = St ♦ dBtH ,

S0 = s0 ,

(5.2.3)

where H ∈ (1/2, 1) everywhere. In this case

1 St = s0 exp♦ (BtH ) = s0 exp BtH − t2H 2

(5.2.4)

(see Section 2.3.1 for the definition of the Wick integral and recall that ∞ exp♦ (X) = n=0 X ♦ n ). Such an approach was developed in (EvH03) and (HO03). The portfolio value is defined in (EvH03). The standard way is Vt = ft Bt + gt St , where f and g are the respective numbers of units of the riskless and the risky asset held in the portfolio. However, in (HO03) the portfolio value is defined as Vt = ft Bt + gt ♦ St . The standard Itˆ o-type self-financing condition dVt = gt dSt is replaced by dVt = gt St ♦ dBtH in (EvH03) and by dVt = gt ♦ dSt in (HO03). The paper (BH05) claims that the definition of Vt as Vt = ft Bt + gt St together with dVt = gt St ♦ dBtH (where we put Bt ≡ 1) has no economic interpretation as a self-financing condition. Here are the brief arguments. Consider a buy-and-hold portfolio. It must satisfy Vt − Vu = gu (St − Su ),

(5.2.5)

324


t from intuitive point of view. However, in our case Vt − Vu = u gu Sz ♦ dBzH , t where the last integral, in general, does not coincide with gu u Sz ♦ dBzH and does not coincide with the right-hand side of (5.2.5). To be precise with this statement, consider the following example from (BH05): let the initial capital x > 0; at time t = 0 we put our money into the bank account and wait until t = 1. Since Bt ≡ 1 we receive x at time t = 1. At this moment we put our money into the risky asset, i.e., buy x/S1 shares at the price S1 and hold this position until t = 2. The value of this portfolio at time t = 2 is V2 = Sx1 S2 . Evidently, such a strategy must be considered as self-similar since 2 nothing was added or subtracted. Nevertheless, Sx1 S2 = x + 0 gu Su ♦ dBuH 2 with gu = Sx1 1(1,2] (u). Indeed, E(x + 0 gu Su ♦ dBuH ) exists and equals x, but

1 S2 1 1 (1 − 22H ) = xE exp B2H − B1H − 22H + 12H = x exp S1 2 2 2

1

1 (1−22H ) exp ·(2−1)2H = x exp{1−22α }, ×E exp{B2H −B1H } = x exp 2 2 which is not x unless H = 1/2. There are some other objections concerning this model, see (BH05). As to the model with dVt = gt ♦ dSt , simple buy-and-hold strategies will be self-financing in this case. However, the objection in this case is that such a definition of portfolio Vt = ft dBt + gt ♦ dSt is hard to motivate from the economic point of view. The reasoning in (BH05) is more moral and practical than mathematical: indeed, to calculate the value of portfolio in this case one needs to know Wick calculus and it is hard to instruct the broker how to do it. But there are also some mathematical reasonings against this model, because it can be proved that there exists a portfolio f = 0, g1 > 0 such that g1 ♦ S1 < 0 with positive probability (index 1 stands for the moment of time here). It is sufficient to put Ω = {ω ∈ Ω|B1H (ω) ∈ (1/2, 3/2)}, g1 = S1 − 1, where S1 = exp{B1H − 1/2}. Then g1 > 0 on Ω , P (Ω ) > 0, g1 ♦ S1 = S1 ♦ S1 − S1 = exp{2B1H − 2} − exp{B1H − 12 } < 0 on Ω . In spite of all this criticism, we can say some positive words about Wick (and Skorohod) models with fBm in finances. For other interesting facts and approaches to these topics see, for example, (AOPU00),(Oks07). First, we mention that geometric fBm can be written in two forms: xE

(1)

St

H

= S0 eµt+σBt or

(2)

St

H

= S0 eµt+σBt

2

− σ2 t2H

.

(5.2.6)

The first form is very simple to understand but the second one is similar to 1 2 (2) usual geometrical Brownian model St = S0 eµt+σBt − 2 σ t , because ESt = S0 (2) for µ = 0. (In Section 6.1 we shall consider the null hypothesis H : S = St (1) against A : S = St , but in a more complex form, see below.) As mentioned in (SV03), if we consider it in the Riemann–Stieltjes sense, (2) the geometric fBm St with µ = 0 is the solution of the equation

5.2 The Different Forms of the Black–Scholes Equation (2)

dSt

(2)

= St (dBtH − Ht2α dt), (2)

325

(5.2.7)

(2)

(2)

(2)

and in the Wick–Skorohod sense δSt = St δBtH or dSt = St ♦ dBtH , i.e. we obtain the model (5.2.3). Nevertheless, due to the Riemann–Stieltjes interpretation, we can consider self-financing condition as t gs Ss(2) d(BsH − Hs2α ds), Vt = V0 + 0

and it has a clear economic meaning. Indeed, one can consider the Riemann– Stieltjes integral as an almost sure limit of simple predictable trading strategies. (2) Now we use the Itô formula (Theorem 2.7.6) for m = 1, St := St , Yt = 2 σBtH + µt − σ2 t2H , H ∈ (1/2, 1) and F(t, x) = F (t, S0 ex ), take (5.2.7) into account and obtain t ∂F (u, Su )du F(t, Yt ) := F (t, St ) = F (0, S0 ) + 0 ∂t t t ∂F ∂F 2 2α + (u, Su )Su (µ − Hσ u )du + σ (u, Su )d(BuH − Hu2α du) ∂x 0 0 ∂x t ∂2F ∂F + Hσ 2 (u, Su )Su du u2α (u, Su )Su2 + 2 ∂x ∂x 0 t t ∂F ∂F = F (0, S0 ) + (u, Su )du + µ (u, Su )Su du 0 ∂t 0 ∂x t t ∂F ∂2F +σ (u, Su )d(BuH − Hu2α du) + Hσ 2 u2α 2 (u, Su )Su2 du. ∂x 0 ∂x 0 Consider the assumption E sup 0≤s≤t

∂F ∂x

2 (s, Ss )Ss

+ E sup 0≤s≤t

∂2F ∂x2

(s, Ss )Ss2

2

< ∞.

(5.2.8)

St ) := C(T − t, St ), where C(t, x) is the price of some Let F (t, St ) := C(t, European option with C(T, x) = c(x), and S satisfying assumption (5.2.8). in differential form as Then, similarly to (5.2.2), we can present dC t = σ dC

∂C · S(dBtH − Ht2α dt) ∂S ∂C ∂C ∂2C + + σ 2 Ht2α 2 S 2 dt. + µS ∂S ∂t ∂S

Now, if the portfolio of value process V consists of one option and an − δ · S, the jump in amount of −δ of underlying assets, then the value V = C the value of this portfolio in one time step equals

326


t − δ · dSt dVt = dC ∂C · St (dBtH − Ht2α dt) − δ(σSt (dBtH − Ht2α dt)) =σ ∂S ∂C ∂C ∂2C + + σ 2 Ht2α 2 St2 − µSt δ dt. + µSt ∂S ∂t ∂S If we choose δ :=

∂C ∂S

to eliminate the stochastic noise, then dV =

∂C ∂t

+ σ 2 Ht2α

∂2C S 2 dt. ∂S 2

The return on an amount Vt invested in the bank account equals rV dt at time dt. For absence of arbitrage they must be equal, whence we obtain the fractional Black–Scholes equation (“Wick” version): ∂2C ∂C ∂C = 0. + σ 2 Ht2α 2 S 2 + rS − rC ∂t ∂S ∂S We can solve this equation on the segment [0, T ] with boundary condition c(x) = (x − K)+ , where K > 0 is strike price, and obtain ln S + r(T − t) + (T 2H − t2H ) σ2 K 2 √ C(t, S) = C(T − t, S) = SΦ σ T 2H − t2H ln S + r(T − t) − (T 2H − t2H ) σ2 K 2 √ − Ke−r(T −t) Φ , σ T 2H − t2H where Φ( · ) is a function of standard normal distribution. Note that it coincides with the solution of usual Black–Scholes equation for H = 1/2.

6 Statistical Inference with Fractional Brownian Motion

6.1 Testing Problems for the Density Process for fBm with Different Drifts As we have seen in Subsection 5.2.2, the form of geometric fBm (5.2.6) depends on the kind of integral that is used in its calculations: if we use the Riemann– Stieltjes integral, 1 t (1) (1) (1) Ss ds + σ Ss(1) dBsH , then St = S0 + µ 0 (1)

0

(1)

St = S0 exp{µt + σBtH }, and if the behavior of geometric process is guided by the Wick integral, 1 t (2) (2) (2) Ss ds + σ Ss(2) ♦ dBsH , then St = S0 + µ 0 (2)

0

(2)

St = S0 exp{µt + σBtH − 12 σ 2 t2H }. So, the natural question arises: what trend actually has geometric fBm? This question was considered in the paper (KMV05), and here we present a solution of this problem. In what follows P the notation Xn = oP (1) means that Xn −→ 0, Xn = OP (1) means that lim lim sup P {|Xn | ≥ C} = 0 . Assume that H ∈ (1/2, 1). For a fixed µ ∈ R

C→∞

n

let Pµ,σ,σ be the distribution of the process Xt := σBtH + µt −

σ 2 2H t , 0≤t≤T 2

(6.1.1)

in the space C[0,T ] of continuous functions. Similarly, Pµ,σ is the distribution of the process (6.1.2) Xt := σBtH + µt, 0 ≤ t ≤ T in the space C[0,T ] .

328


Suppose now that we observe a trajectory of the process {Xt , 0 ≤ t ≤ T } in the space C[0,T ] . Denote by PX the law of X. We want to test the following complex hypothesis: H : PX ∈ {Pµ,σ,σ : µ ∈ R, σ ∈ R+ } against the complex alternative A : PX ∈ {Pµ,σ : µ ∈ R, σ ∈ R+ }. From the point of view of the general theory, models of observation (6.1.1) and (6.1.2) are equivalent to the classical model t = X

t

lH (t, s)dXs = σMtH + µB1 t1−2α − σ 2 HB2 t

(6.1.3)

0

and

t = σM H + µB1 t1−2α , X t

where t = X

t

lH (t, s)dXs ,

MtH

t

lH (t, s)dBsH ,

=

0

(6.1.4)

0 (5)

the kernel lH is defined in Section 1.8, B1 := CH B(1 − α, 1 − α), (5) B2 := CH B(1 + α, 1 + α). Introduce the following density processes (Radon–Nikodym derivatives) based on the observed trajectory of X: f1 (X : µ, σ, σ) := and f2 (X : µ, σ) :=

dPµ,σ,σ (X) dP0,σ

(6.1.5)

dPµ,σ (X). dP0,σ

(6.1.6)

Theorem 6.1.1. Assume we observe X on the interval [0, T ]. We have

µ 2 T − bX T1 − c µ T 1−2α + dµT − kσ 2 T 2H (6.1.7) f1 (X : µ, σ, σ) = exp a 2 X σ σ2 and µ

2 t − c µ T 1−2α , f2 (X : µ, σ) = exp a 2 X σ σ2 where

t1 = X c=

t

s , a = B1 , s2α dX

0 1 2 B 2 1,

d = B1 B2 H,

HB2 2(1−H) , HB22 8(1−H) .

b=

k=

(6.1.8)

(6.1.9)

Proof. Follows immediately from (6.1.1) to (6.1.6) and the classical Girsanov theorem.

6.1 Testing Problems for the Density Process for fBm with Different Drifts

329

6.1.1 Observations Based on the Whole Trajectory with σ and H Known In this section we demonstrate how to test the hypothesis H against the alternative A, when σ is known and the whole trajectory {Xt : t ∈ [0, T ]} is observed. We can use the likelihood ratio to test this (for the likelihood ratio see (Bor84), p. 319). In our problem the likelihood ratio l(X· ) = l(X· |σ) has the form supµ∈R f1 (X; µ, σ, σ) . (6.1.10) l(X· |σ) := supµ∈R f2 (X; µ, σ) Note that in (6.1.10) both upper bounds are attained, since the densities f1 and f2 are the quadratic functions of µ. More precisely, we have 2 t )2 T 2α−1 − 2αbX t · T 2α supµ∈R f1 (X; µ, σ, σ) = exp 4σa 2 c (X

(6.1.11) T s ds − 4α2 kσ 2 T 2H , + 2αb 0 s2α−1 X and the value of µ giving the maximal value in (6.1.11) is µ H :=

t + dT σ 2 aX . 2cT 1−2α

(6.1.12)

Similarly, for the denominator in (6.1.10) we have sup f2 (X; µ, σ) = exp

aX 2 T 2α−1

T

4σ 2 C

µ∈R

(6.1.13)

an the maximum in (6.1.13) is achieved by µ A :=

T T 2α−1 aX . 2c

(6.1.14)

We obtain the following theorem as a direct consequence of (6.1.10) – (6.1.14): Theorem 6.1.2. The likelihood l(X· |σ) from (6.1.10) admits the representation T

2α s ds − 4α2 kσ 2 T 2H . s2α−1 X l(X· |σ) = exp − 2αbXT T + 2αb 0

Remark 6.1.3. Note that in the case when H = 12 we have l(X. |σ) = 1. It means that our method does not work in this case, because the drift (−σ 2 2t ) has the same order in t as µt, and we cannot distinguish them. Therefore our method works worse if H is close to 12 .

330


Next we describe the testing procedure. Given a confidence level 1 − ρ, ρ ∈ (0, 1/2), consider the critical areas defined by K1 := {X· : l(X· |σ) ≥ Kρ } and K2 := {X· : l(X· |σ) < kρ }. The critical values 0 < kρ ≤ Kρ are chosen in such a way that we have sup Pµ,σ (K1 ) ≤ ρ, sup Pµ,σ,σ (K2 ) ≤ ρ.

µ∈R

(6.1.15)

µ∈R

The test is now clear: if X· ∈ K1 we accept H, if X· ∈ K2 we accept A. If l(X· |σ) ∈ [kρ , Kρ ) then no hypothesis is accepted. Inequalities (6.1.15) show that the probabilities of so-called errors of the first and of the second kind will not exceed the level ρ. Next we compute the critical values Kρ , kρ . To compute kρ recall that under A the process X has the same distribution as the process σZt + µt. Similarly, to compute Kρ we use the fact that under H the process X has the 2 same distribution as the process σZt + µt − σ2 t2H . We have that T

l(σZ· +µ·|σ) = exp −2αbσMTH ·T 2α +2αbσ s2α−1 MsH ds −4α2 kσ 2 T 2H 0

and σ2 · |σ = exp l(σZ· + µ · |σ) exp 8α2 kσ 2 T 2H }. l σZ· + µ · − 2 Hence, we have that Pµ,σ (K1 ) = P − 2αbσMTH · T 2α

T + 2αbσ 0 s2α−1 MsH ds ≥ log Kρ + 4α2 kσ 2 T 2H

(6.1.16)

The random variable in the above expression is Gaussian with zero mean and variance α2 HB22 σ 2 T 2H . v2 = 1−H Therefore, by (6.1.15) log K v ρ Pµ,σ (K1 ) = 1 − Φ + , v 2

(6.1.17)

where Φ is the distribution function of standard normal distribution. If ξρ is 2 such that 1 − Φ(ξρ ) = ρ, then Kρ ≥ exp{vξρ − v2 }. Similarly, 1 v 1 log Pµ,σ,σ (K2 ) = 1 − Φ + , (6.1.18) v kρ 2 that is kρ ≤ exp{−vξρ + v2 kρ = Kρ −1 . 2 }),

v2 2 }.

Finally, we can choose Kρ = max(1, exp{vξρ −


331

6.1.2 Discretely Observed Trajectory and σ Unknown Assume now that we observe the process X discretely and the intensity σ of the fractional noise is unknown. We replace the parameter σ in l(X· |σ) with a consistent estimate σ n , where n is the number of time points, and instead of the stochastic integrals w.r.t. X we will use sums in terms of the increments of X. We obtain a quasi-likelihood ratio, which is constructed from the observations. The critical values will be computed uniformly w.r.t. all possible values of µ and σ. We will give an asymptotic description of the critical levels. First, choose the critical values independently of the parameter σ. For Kρ ≥ 1 we have that E A 1 1 v log Kρ + ≥ 2 log Kρ = 2 log Kρ , v 2 2 and from (6.1.17) Take Kρ∗ := e

2 ξρ 2

A Pµ,σ (K1 ) ≤ 1 − Φ( 2 log Kρ ).

and put K1∗ := {X· : l(X· ) ≥ Kρ∗ }. Then we have sup Pµ,σ (K1∗ ) ≤ ρ.

(6.1.19)

µ,σ>0

2 ξρ

Similarly, using (6.1.18) and taking kρ∗ = e− 2 and if K2∗ := {X· : l(X· ) ≤ kρ∗ } we will have (6.1.20) sup Pµ,σ,σ (K2∗ ) ≤ ρ. µ,σ>0

Put

K0∗ := {X· : kρ∗ < l(X· ) < Kρ∗ };

note that K0∗ is (a conservative variant of) the region, where neither the √ HB2 σ hypothesis H nor the hypothesis A is accepted. Let C1 := α√1−H . √ Theorem 6.1.4. Assume that T > sup Pµ,σ (K0∗ ) ≤ µ,σ

and sup Pµ,σ,σ (K0∗ ) ≤

µ,σ,σ

2

C1 ξρ

1/H . Then we have that

C 2 T 2H

4 −H T exp − 1 C1 32 C 2 T 2H

4 −H T exp − 1 . C1 32

(6.1.21)

(6.1.22)

Proof. We have that v 1 + log kρ∗ . Pµ,σ (K0∗ ) ≤ Pµ,σ ({X· : l(X· ) > kρ∗ }) = 1 − Φ 2 v

(6.1.23)

332


We have the following inequality for x > 0: ∞ u2 x2 1 1 1 − Φ(x) = √ e− 2 . e− 2 du ≤ √ 2π x 2πx C1 T H 2

Apply this to (6.1.22) with x =

−

ξρ2 1 , C1 T H 2

√ and if T >

2ξρ C1

1/4 we

obtain (6.1.21). The estimate (6.1.22) is obtained similarly. Corollary 6.1.5.

lim sup Pµ,σ {K0∗ } = 0

T →∞ µ

and

lim sup Pµ,σ,σ {K0∗ } = 0.

T →∞ µ

Assume that we observe the process X at points 0 ≤ tn,1 < · · · tn,n ≤ T , where tn,k ∈ π n . Put ∆n = max{tn,1 , |π n |, T − tn,n } and assume that lim ∆n = 0.

(6.1.24)

n→∞

We will introduce a discrete version of the functional l(X· ). Put sk = tn,k , n2 is some ∆sk = sk+1 − sk , xk = Xtn,k and ∆xk = xk+1 − xk . Assume that σ 2 consistent estimator of σ . Put

n−1 (5) −α ln (x1 , . . . , xn ) = exp − 2αbT 2α CH s−α ∆xk k+1 (T − sk ) k=0

+

(5) 2αbCH

n k=1

s2α−1 k+1

k−1

s−α i+1

−α

(sk − si )

∆xi

∆sk − 4α

2

k σn2 T 2H

.

i=0

With the help of constants Kρ∗ and kρ∗ from (6.1.19) and (6.1.20) define the critical domains ∗ K1n := {(xn,1 , . . . , xn,n ) ∈ Rn |ln (xn,1 , . . . , xn,n ) ≥ Kρ∗ }

and

∗ K2n := {(xn,1 , . . . , xn,n ) ∈ Rn |ln (xn,1 , . . . , xn,n ) < kρ∗ }.

∗ If the observations belong to K1n then H is accepted and if the observations ∗ belong to K2n then A is accepted.

Theorem 6.1.6. Assume that we have (6.1.24) as n → ∞. Then for any µ ∈ R, σ > 0 we have that Pµ,σ,σ

ln (xn,1 , . . . , xn,n ) −→ l(X· |σ)

(6.1.25)

and Pµ,σ

ln (xn,1 , . . . , xn,n ) −→ l(X· |σ).

(6.1.26)


333

Proof. We prove the claim (6.1.25) (the claim (6.1.26) is proved similarly). · |σ) the random variable l(X· |σ), when the process Xt is reDenote by l(X 2 2H xn,1 , . . . , x n,n ) placed by the process σZt + µt − σ t2 , 0 ≤ t ≤ T , and by ln ( n ∆Zn,k + µ∆tk − the variable where we replace ∆Xn,k by σ for any ε > 0, C > 0 we have that

σ 2 (∆tk )2H . 2

Then

· |σ)| > ε} Pµ,σ,σ {|ln (xn,1 , . . . , xn,n ) − l(X · |σ)| ≥ C} ≤ Pµ,σ,σ {|ln (xn,1 , . . . , xn,n )| ≥ C} + Pµ,σ,σ {|l(X · |σ)| > εe−C }. + Pµ,σ,σ {| log ln ( xn,1 , . . . , x n,n ) − log l(X (6.1.27) The first two probabilities can be chosen sufficiently small for large C > 0. · |σ) and ln ( The structure of the functionals l(X xn,1 , . . . , x n,n ), the facts that Pµ,σ,σ t (5) n−1 −α −α σ n −→ σ, CH ∆sk → 0 lH (t, s)ds and k=1 sk+1 (T − sk ) n

(5) CH

s2α−1 k

k−1

s−α i+1

(sk − si )

−α

∆si ∆sk →

s

2α−1

s 0

i=1

k=1

T

lH (s, u)du ds, 0

supply that it is sufficient to prove that n

(5)

CH

Pµ,σ,σ

−α s−α ∆BkH −→ MTH , k+1 (T − sk )

(6.1.28)

k=1

where ∆Yk := Y(k+1)T /n − YkT /n for any process Y, and that (5) CH

n

s2α−1 k+1

−→

−α

s−α i+1 (sk − si )

∆BiH

∆sk

i=1

k=1 Pµ,σ,σ

k−1

T

s2α−1 MsH ds.

(6.1.29)

0 −α To prove (6.1.28) consider for fn,T (s) = CH s−α 1{s∈[sk ,sk+1 )} k+1 (T − sk ) (5)

E

MTH

T

−

(5) CH

n−1

2 s−α k+1

(T − sk )

−α

∆BkH

k=1 T

(lH (T, s) − fn,T (s)) (lH (T, u) − fn,T (u)) |u − s|2α−2 du ds.

= 2Hα 0

0

T T We have that fn,T (s) ↑ lH (T, s) for s ∈ (0, T ), and 0 0 lH (T, s)lH (T, u) × |u − s|2α−1 du ds < ∞. Therefore, by monotone convergence,

T

T

(lH (T, s) − fn,T (s)) (lH (T, u) − fn,T (u)) |u − s|2α−1 du ds → 0 0

0

334


as n → ∞ and (6.1.28) follows. To finish, we prove (6.1.29). Denote g(s) := s2α−1 MsH and (5)

gn (s) := CH s2α−1 k+1

k−1

−α

s−α i+1 (sk − si )

∆BiH 1{s∈[sk ,sk+1 )} .

i=1

Then, for any s ∈ (0, T ], H 2α−1 E|g(s) − gn (s)| ≤ |s2α−1 − s2α−1 k+1 | E|Ms | + 2Hαs 1/2 s s × E 0 0 (lH (s, u) − fn,s (u))(lH (s, r) − fn,s (r))|u − r|2α−1 du dr , (6.1.30) and as in previous inequalities, the second term on the right-hand side of (6.1.30) goes to zero, moreover the left-hand side can be dominated, according to Remark 1.9.5, by

(1) s2α−1 T 1−H + C 2 ||lH (s, ·)||L [0,s] ≤ C (3) T 1−H , C H H H 1/H are some constants, i = 1, 2, 3. From here where C T H E| 0 (g(s) − gn (s))ds| → 0, as n → ∞ and we obtain (6.1.29). (i)

Corollary 6.1.7. Assume that (6.1.24) holds. Then ∗ ∗ lim sup Pµ,σ (K1n ) ≤ ρ, lim sup Pµ,σ,σ (K2n )=0 n

and

n

∗ lim lim sup(Pµ,σ + Pµ,σ,σ )(K0n ) = 0, T

where

∗ K0n

n

:= {(xn,1 , . . . , xn,n ) : kρ∗ < log ln (xn,1 , . . . , xn,n ) < Kρ∗ }.

Proof. By Theorem 6.1.6 we have, as n → ∞: ∗ ∗ ) → Pµ,σ (K1∗ ), Pµ,σ,σ (K2n ) → Pµ,σ (K2∗ ) Pµ,σ (K1n

and

∗ (Pµ,σ + Pµ,σ,σ )(K0n ) → (Pµ,σ + Pµ,σ,σ )(K0∗ ).

Hence the statements of the corollary follow from Theorem 6.1.6.

Note that according to Corollary 6.1.7 the proposed test procedure has asymptotically the level of errors less than or equal to ρ for both kinds of errors. Note also that the probability not to make a decision goes to zero as T → ∞. It is also easy to see from the proof of Theorem 6.1.6 and Corollary 6.1.7 that this convergence is uniform for all µ and all σ ≥ σ0 > 0, where σ0 is fixed.

6.2 Goodness-of-fit Test

335

6.2 Goodness-of-fit Test 6.2.1 Introduction Suppose that H was tested against A, and we conclude that, e.g. A is true. Consider a certain functional depending on the trajectory of the observed process {Xt , 0 ≤ t ≤ T }. If the distribution of this functional under A is known we can construct the corresponding goodness-of-fit test. For a given confidence level we either reject A or do not reject A. If we reject A it means that the observed trajectory does not fit the model described by A, and we conclude finally in this case that both A and H are wrong. If the parameters in the models are unknown we propose an asymptotical test which provides a given confidence level as T → +∞. 6.2.2 The Whole Trajectory Is Observed and the Parameters µ and σ Are Known Introduce a functional which depends on the whole observed trajectory {x(t), t ∈ [0, T ]}, in a linear way: T Z(T, s)dXs , QT := 0

where Z(T, s) = s1/4−H (T − s)3/4−H. We choose here the exponents 14 − H and 34 − H different from 12 − H in order to obtain the functional which is essentially different from MTH . The reason for that will be clear from Theorem 6.2.3. The integral exists in both 2 cases when Xt = σBtH + µt and Xt = σBtH + µt − σ2 t2H . Denote 5 7 1 7 − H, − H , B4 = B H + , − H . B3 = B 4 4 4 4 Theorem 6.2.1. Let the parameters µ and σ be known. (i)

Assume that we have H: Xt = σBtH + µt −

σ 2 2H . 2 t

Then

RTH := T H−1 QT − µB3 · T 1−H + σ 2 H · B4 · T H ∼ N (0, C2 σ 2 ); (ii)

Assume that we have A: Xt = σBtH + µt. Then RTA := T H−1 QT − µB3 · T 1−H ∼ N (0, C2 σ 2 ), where

1

C2 = 2Hα 0

0

1

1

3

(us) 4 −H ((1 − u)(1 − s)) 4 −H · |u − s|2α−1 du ds.

336


Proof. Assume H. Then we have T Z(T, s)dBsH + µT 1−2α B3 − σ 2 HB4 T QT = σ

(6.2.1)

0

and so RTH = T H−1 σ

T

Z(T, s)dBsH . 0

Obviously, RTH is normally distributed with mean zero and with variance T T 1 3 (us) 4 −H ((T −u)(T −s)) 4 −H |s−u|2α−1 du ds, E(RTH )2 = σ 2 T 2α−1 2αH 0

0

i.e. E(RTH )2 = σ 2 C2 and the first claim now follows. Assume A. Then we can write QT as T Z(T, s)dBsH + µT 1−2α B3 , QT = σ

(6.2.2)

0

and the second claim follows from (6.2.2) as above. The goodness-of-fit tests are based on the statistics H

RT :=

RTH σ(C2 )

1 2

,

RTA

A

RT :=

1

σ(C2 ) 2

.

Fix a confidence level 1 − ρ, ρ ∈ (0, 12 ), and let ξ ρ2 be a standard normal law, i.e P {N (0, 1) ≥ ξ ρ2 } = and reject A if

A |RT |

ρ 2.

We reject H

ρ 2 -fractile of a H if |RT | > ξ ρ2 ,

>ξ . ρ 2

A Pµ,σ,σ

Note that under H, RT −−−−→ −∞, T → +∞, therefore the inequality A RT < −ξ ρ2 is an additional argument in favor of H. H Pµ,σ

Also, if A is true, then RT −−−→ +∞, T → +∞, therefore the inequality A RT > ξ ρ2 is an additional argument in favor of A. Remark 6.2.2. Suppose that in reality we have the model Xt = σBtH1 + µt, H1 > H, not σBtH + µt. Denote the law of X in this case by P . Then 3 T H−1 T 1 −H H RT = s4 (T − s) 4 −H dBsH1 , 1 (C2 ) 2 0 H

H

P

and E(RT )2 has the order T 2(H1 −H) for large T , thus RT − → ∞, T → ∞, and H

A

RT = RT +

σHB4 T H (C2 )

1 2

= T H1 −H OP (1) +

σHB4 T H (C2 )

1 2

Therefore our statistics can distinguish this case, too.

P

− → +∞,

T → ∞.


337

6.2.3 Goodness-of-fit Tests with Discrete Observations Asymptotic Behavior of Discrete Statistics for µ Unknown and σ Known Suppose for simplicity that we observe the values X kT , k = 0, 1, . . . , n. We n substitute in RTA , RTH a discretization of QT , T := Q

n−1 k=0

(k + 1)T 14 −H kT 34 −H Xk . T− n n

Instead of µ we substitute the estimates (6.1.12) and (6.1.14), respectively. Thus we define T − µ H := T H−1 Q R A B3 T 1−H + σ 2 HB4 T H T and

TA := T H−1 Q T − µ R H B3 T 1−H .

Under hypothesis H we have TH = σT −H R

n−1 k=0

+ µT 1−H

n−1 k=0

k + 1 14 −H k 34 −H 1− B H kT n n n

(6.2.3)

n−1 k + 1 14 −H k 34 −H 1 σ 2 H k + 1 14 −H 1− T · − n n n 2 n k=0

k 34 −H k + 1 2H k 2H −µ A B3 T 1−H + σ 2 HB4 T H , × 1− − n n n and under hypothesis A TA = σT −H R

n−1 k=0

+ µT 1−H

n−1 k=0

k + 1 14 −H k 34 −H 1− B H kT n n n

(6.2.4)

k + 1 14 −H k 34 −H 1 1− H B3 T 1−H . · −µ n n n

To begin we find the rate of convergence of the integral sums in (6.2.3) and (6.2.4) to the corresponding integrals. A by Define R T A := R T

σ T 1−H

T

s1/4−H (T − s)3/4−H dBsH + B3 T 1−H (µ − µ A ) 0

H similarly, with µ H replacing µ A . and R T A − R A and R H − R H . Put We study the differences R T T T T

338


qn (T, s) :=

n−1 k=0

(k + 1)T 14 −H kT 34 −H · 1[ kT , (k+1)T ) (s), T− n n n n

1

sδ (1 − s)β ds,

I(δ, β) := B(δ + 1, β + 1) = 0

and In (δ, β) =

n−1 k=0

k+1 n

We have that A − R A = T H−1 R T T

δ 1−

k n

β

1 . n

T

(qn (T, s) − q (T, s)) dBsH 0

− T 1−H µ (In (1/4 − H, 3/4 − H) − I (1/4 − H, 3/4 − H))

(6.2.5)

and 2 TH − R TH = R TA − R TA − σ T H 2H In (H − 3/4, 3/4 − H) R

n−1 2 1/4−H 3/4−H σ2 k+1 k TH (6.2.6) − I (H − 3/4, 3/4 − H) − 1− 2 n n k=0

2H 2H H−3/4 3/4−H k+1 k k+1 1 k . × − − 2H 1− n n n n n Using self-similarity, we obtain that

2 T E T H−1 (qn (T, s) − q (T, s)) dBsH 0

2

1

(qn (1, s) − q (1, s)) dBsH

=E

.

(6.2.7)

0

According to Remark 1.9.5 we have 1 2 E (qn (1, s) − q (1, s)) dBsH ≤ cH ||qn (1, s) − q(1, s)||2L1/H [0,1] . (6.2.8) 0

Now we use these preliminary calculations to prove the next result. Let n = n(T ) be the number of approximation points. Theorem 6.2.3. Assume (iii) For

1 2

< H ≤ 34 , Tβ → 0, n(T )

T → ∞, with β =

H . H + 14


(iv) For

3 4

339

< H < 1, Tβ → 0, T → ∞, with β = H. n(T )

Then under H and under A

TH − R TH = oP (1), R

T →∞

(6.2.9)

TA − R TA = oP (1), R

T → ∞.

(6.2.10)

A ∼ N (0, r2 ), where H ∼ N (0, r2 ), and under A R Moreover, under H R T T

1

1

2α−1

ϕ(s)ϕ(u) · |u − s|

r2 := 2σ 2 αH 0

du ds,

0

with 1

3

ϕ(s) := s 4 −H (1 − s) 4 −H −

B3 −α s (1 − s)−α . B1

Proof. To prove the claims note first that using Lemmas B.0.1 and B.0.2 from H = oP (1) under H and R A − R A = oP (1) H − R Appendix B we have that R T T T T H under A. Next, we substitute (6.1.12) into RT and obtain H = R T

σ T 1−H

T

1

3

s 4 −H (T − s) 4 −H −

0

B3 −α s (T − s)−α dBsH . B1

H ∼ N (0, r2 ). Similarly, one shows that under A This implies that under H R T A 2 ∼ N (0, r ). R

T Remark 6.2.4. For the kernel lH (t, s) instead of Z(t, s) we obtain the degenH and R A . This is the reason why we take the kernel erate distribution of R T T Z(t, s). Goodness-of-fit Test Based on Theorem 6.2.3, we construct the goodness-of-fit test similarly to H ρ the one from Subsection 6.2.2. Choose ξ 2 as there. We reject H if RT > rξ ρ2 , A ρ and we reject A if R T > rξ 2 . The test is applicable for large T only, contrary to the test from Subsection 6.2.2, because for the probability pH (T ) that H is rejected when H is true, we have now lim pH (T ) = ρ

T →∞

and similarly for A and pA (T ).

340


6.2.4 On Volatility Estimation In this subsection we construct an estimator for the parameter σ. We end this subsection by giving the goodness-of-fit test for the case where both µ and σ are unknown. Introductory Computations for Volatility Estimation 2 Assume H. Then the background process is Xt = σBtH +µt− σ2 t2H , t ≥ 0. We make observations at time points tk = kT n , k = 0, 1, . . . , n. Put, as before, ∆Xk = X k+1 T − X k T , k = 0, . . . , n − 1. Then we have, with obvious notation, n n that ∆Xk = σ∆BkH + µ∆tk − k = 0, . . . , n − 1. Consider now

∆Xk TH

σ2 ∆(t2H )k , 2

and write this as

∆Xk 1 µ∆tk σ2 = σ ε + − ∆(t2H )k . k TH nH TH 2T H

(6.2.11)

∆B H nH

In (6.2.11) we used the notation εk = TkH . By self-similarity the distribution of the vector (ε0 , . . . , εn−1 ) is the same as of the vector H BH 1 − B0 n

1 nH

B1H − B H (n−1) n

,...,

1 nH

d

H = B1H − B0H , B2H − B1H , . . . , BnH − Bn−1 ,

where we again used self-similarity. Simple computation gives Eεk = 0, Eε2k = 1 and Eεk εl = If k > l ≥ 1 and gives

1 2

1 |k − l + 1|2H − 2|k − l|2H + |k − l − 1|2H . 2 ≤ H < 1, then, applying the mean value theorem twice 0 ≤ Eεk εl ≤ 2Hα(k − l)2α−1 .

Denote µ1 :=

nH µ∆t , TH

yt :=

nH ∆Xt TH

(6.2.12)

and rewrite (6.2.11):

yt = σεt + µ1 −

σ 2 −H H 2H T n ∆t . 2

To simplify the notation put yk := σεk + µ1 −

σ 2 H H 2H T n ∆τk , k = 0, 1, . . . , n − 1, 2

(6.2.13)

2H − ( nk )2H . We use a sample variance to estimate σ : where τk2H = ( k+1 n )

σ n2 :=

n y1 + · · · + yn (yn2 − y 2n ) with y n := . n−1 n

(6.2.14)


Let zk = σεk −

σ 2 H H 2H T n ∆τk , 2

k = 0, 1, . . . , n − 1. Then

z¯n = σ ε¯n − and σ 2 =

341

σ 2 H H−1 T n 2

(6.2.15)

n nσ 2 (zn2 − z 2n ) = ε2 − σT H nH εn ∆τn2H n−1 n−1 n σ 2 2H 2H T n (∆τn2H )2 − ε2n + 4 σ 2 2H 2H H H 2 2H 2H + σT n εn ∆τn − T n (∆τn ) . 4

(6.2.16)

Again we have a problem with the rate of the discretization with respect to the observation interval. We start with one lemma: Lemma 6.2.5. Assume that X, Y are two standard normal random variables: EX = EY = 0 and V ar(X) = V ar(Y ) = 1. Assume that EXY = q. Then E (X 2 − 1)(Y 2 − 1) = 2q 2 .

(6.2.17)

Lemma 6.2.6. With the notation above: If H < 34 , then E|ε2n − 1| ≤ C √1n . 3 (vi) If H = 34 , then E|ε2n − 1| ≤ C logn n . (v)

(vii) If

3 4

< H < 1, then E|ε2n − 1| ≤ Cn2α−1 .

Proof. We have that ε2n − 1 =

n−1 1 2 (ε − 1). n i=0 i

From Lemma 6.2.5 and (6.2.6): n−1 2 1 2 E(ε2i − 1)2 + 2 E ε2n − 1 = 2 n i=0 n

0≤j 0. Differential equation (6.3.1) can be rewritten in the integral form t t a(s, Xs )ds + b(s, Xs )dBsH , t ∈ [0, T ]. Xt = X0 + θ 0

(6.3.2)

0

Here we use pathwise construction of the integral w.r.t. fBm. Suppose that equation (6.3.2) has unique pathwise solution. (Sufficient conditions of existence and uniqueness of the solution on the interval [0, T ] are presented in Theorem 3.1.4.) Now, let T > 0 be fixed. We are in a position to find the likelihood ratio dPθ (t) dP0 (t) for the probability measure Pθ (t) corresponding to our model and the probability measure P0 (t) corresponding to the model with zero drift. Suppose that the following assumption holds: t) (i) b(t, Xt ) = 0, t ∈ [0, T ] and a(t,X b(t,Xt ) is a.s. Lebesgue integrable on [0, T ]. Denote ϕt :=

a(t,Xt ) b(t,Xt )

and introduce the new process H := B H + θ B t t

t

ϕs ds.

(6.3.3)

0

= (1 − Let also the following conditions hold (recall that α = (1 − 2α)1/2 , α −1/2 ): 2α) t (ii) lH (t, s)|ϕ(s)|ds < ∞, t ∈ [0, T ]

0

t

lH (t, s)ϕ(s)ds = α

(iii) θ 0

t

δs ds, t ∈ [0, T ] 0

and

t

s2α δs2 ds < ∞, t ∈ [0, T ].

(iv) E 0

6.3 Consistency of Drift Parameter Estimates

345

t

s is a square-integrable martingale for the Wiener Then Lt = 0 sα δs dB t sH process B w.r.t. the measure P0 (t) such that 0 lH (t, s) dB t −α s . According to the Girsanov theorem for fBm (Theorem 2.8.1), =α 0 s dB under the assumptions (i)–(iv) and 1 (v) E exp Lt − Lt = 1, 2 H is an fBm on [0, T ] w.r.t. the measure Q defined via the the process B t relation 1 dPθ (t) = exp Lt − Lt , t ∈ [0, T ]. (6.3.4) dP0 (t) 2 Remark 6.3.1. We can try to present the likelihood ratio (6.3.4) as a function of the observed process Xt , according to statistical tradition. Toward this end recall that t t t sH = α s = lH (t, s) dB s−α dB lH (t, s)b−1 (s, Xs )dXs . (6.3.5) 0

0

0

t

Suppose that the process Jt := 0 lH (t, s)b−1 (s, Xs )dXs admits a differential of the form dJt = F (t, Xt )dXt ; then, evidently, Lt = 0

t

s = α sα δs dB

t

s2α δs F (s, Xs )dXs , 0

and δs is a functional of the process X under the conditions of Lemma 6.3.2 (see below). In turn, the existence of the differential dJt can be es t tablished separately for 0 lH (t, s)ϕs ds (it is realized in Lemma 6.3.2) and for t l (t, s)dBsH = MtH , but the last problem is of the same complexity as the 0 H original one. Another possibility is to establish, similarly to Lemma 6.3.2, the conditions of the existence of the derivative (sα δs ) , in general, this problem is solvable; then we can rewrite t α s (sα δs ) ds, B Lt = t δt Bt − 0

is an adapted functional of X. Indeed, we can present B and, of course, B H (see (6.3.5)), relation (6.3.3) and the equality via X with the help of B t t BtH = 0 b−1 (s, Xs )dXs − 0 ϕs ds. Consistency of the Drift Parameter Estimates In order to find the maximum likelihood estimate of the parameter θ, we use likelihood ratio (6.3.4), which can be rewritten as t t dPθ (t) s − 1 = exp sα δs dB s2α δs2 ds , dP0 (t) 2 0 0

346


where δs is defined according to the integral representation (iii). First we establish sufficient conditions ensuring the existence of representation (iii). t Denote ψ(t, x) = a(t,x) b(t,x) , so that ψ(t, Xt ) = ϕ(t), I(t) := 0 lH (t, s)ϕ(s)ds. Lemma 6.3.2. Let ψ(t, x) ∈ C 1 [0, T ] ∩ C 2 (R). Then for t > 0 I (t) = C(H)ψ(0, 0)t−2α +

t

lH (t, s) (ψt (s, Xs ) + θψx (s, Xs )a(s, Xs )) ds

0 (5)

− αCH + (1 −

t

s−1−α (t − s)−α

0

(5) 2α)CH t−2α

+

s

t

s

2α−2

s 0

u1−α (s − u)−α ψx (u, Xu )b(u, Xu )dBuH ds

0

(5) CH t−1

(ψt (u, Xu ) + θψx (u, Xu )a(u, Xu )) du ds

0

t

u1−α (t − u)−α ψx (u, Xu )b(u, Xu )dBuH ,

0 (5)

where C(H) = (1 − 2α)B(1 − α, 1 − α)CH . Proof. According to the Itˆ o formula (2.7.3), s ϕs = φ(0, 0) + (ψt (u, Xu ) + ψx (u, Xu )θa(u, Xu ))du 0

+

s

ψx (u, Xu )b(u, Xu )dBuH .

0

Substituting (6.3.6) into the integral I(t) = I(t) = C(H, 1)ψ(0, 0)t1−2α +

t

l (t, s)ϕs ds, 0 H

t

t

lH (t, s)

+θ 0

we obtain

ψt (u, Xu )du ds

0

ψx (u, Xu )a(u, Xu )du ds

0

t

s

lH (t, s)

+

s

s

lH (t, s) 0

(6.3.6)

0

ψx (u, Xu )b(u, Xu )dBuH ds,

(6.3.7)

0

(5)

C(H, 1) = CH B(1 − α, 1 − α) and now our aim is to differentiate I(t). The first term on the right-hand side of (6.3.7) is obviously differentiable, i.e. can t be presented as C(H)ψ(0, 0) 0 s−2α ds. The second and the third terms can be transformed using integration by parts: s u s s s−α ψt (u, Xu )du = u−α ψt (u, Xu )du − α u−1−α ψt (v, Xv )dvdu, 0

and

0

0

0


s−α

s

ψx (u, Xu )a(u, Xu )du =

0

s

u−1−α

−α

0

s

347

u−α ψx (u, Xu )a(u, Xu )du

0 u

ψx (u, Xu )a(v, Xv )dv du.

(6.3.8)

0

According to representation (6.3.8), there exist a.s. the fractional derivatives of order α, i.e. the derivatives of fractional integrals: s t d t lH (t, s) ψt (u, Xu )duds = lH (t, s)ψt (s, Xs )ds dt 0 0 0 (5)

−αCH d dt

t

s

t

s−1−α (t − s)−α

0

0

0

−

s

ψt (u, Xu )du ds.

t

(6.3.9)

lH (t, s)ψx (s, Xs )a(s, Xs )ds

0

(5) αCH

ψx (u, Xu )a(u, Xu )du ds =

lH (t, s) 0

t

−1−α

s

−α

(t − s)

0

s

ψx (u, Xu )a(u, Xu )du ds.

(6.3.10)

0

Further, it follows from Lemma 2.8.2 that t s lH (t, s) ψx (u, Xu )b(u, Xu )dBuH ds 0

=

(5) CH t1−2α

0

t 2α−2

s 0

s

u1−α (s − u)−α ψx (u, Xu )b(u, Xu )dBuH ds. (6.3.11)

0

The proof follows immediately from relations (6.3.9)–(6.3.11).

Now, we can rewrite (6.3.6) as T 2 θ T α dPθ (t) θ 2α 2 s − = exp s I (s)dB s (I (s)) ds . (6.3.12) dP0 (t) α 0 2(1 − 2α) 0 It follows from (6.3.12) that the maximum likelihood estimate is achieved under the condition T θ T 2α α s s I (s)dB − s (I (s))2 ds = 0, α 0 0 whence

t s α 0 sα I (s)dB . θt = t s2α (I (s))2 ds 0

Using Lemma 6.3.2, we obtain

(6.3.13)

348


Bt + θα

t

t , sα I (s)ds = B

(6.3.14)

0

t is a Wiener process under measure Q. Substituting (6.3.14) where B into (6.3.13) we obtain t α 0 sα I (s)dBs θt = θ + t . s2α (I (s))2 ds 0

(6.3.15)

t Recall that under condition (iv) 0 sα I (s)dBs is the square-integrable t Pθ -martingale with angle bracket 0 s2α (I (s))2 ds. Theorem 6.3.3. Let the conditions of Theorem 3.1.4 and (i)–(v) hold for any T > 0 and, moreover, ∞ (vi) s2α (I (s))2 ds = ∞ a.s. 0

Then the maximum likelihood estimate θT is strongly consistent as T → ∞. Proof follows immediately from representation (6.3.15) and from Theorem Xt 6.10 (LS86). This theorem establishes that X → 0 a.s. if Xt is a squaret integrable martingale and X∞ → ∞ a.s. In other words, t

sα I (s)dBs → 0, t → ∞, t0 s2α (I (s))2 ds 0 with Pθ -probability 1. Example 6.3.4. Consider the linear model of the form dXt = θXt dt + Xt dBtH . In this case ϕt = 1, so δt = C(H)t−2α . Hence

Since α

t 0

t 0

δs ds =

t

l (t, s)ds 0 H

= C(H, 1)t1−2α ,

t α 0 s−α dBs . θt = θ + C(H)t1−2α

s−α dBs is the square-integral martingale with the angle bracket

t1−2α → ∞ when t → ∞, then, according to Theorem 6.3.3, a.s. as t → ∞ . So, the estimate θt is consistent with probability 1.

α 0t s−α dBs C(H)t1−2α

→ 0,


349

6.3.2 Consistency of the Drift Parameter Estimates in the Mixed Brownian–fractional-Brownian Diffusion Model with “Linearly” Dependent Wt and BtH Now we consider the linear mixed Brownian–fractional-Brownian diffusion model represented by the stochastic differential equation of the form dXt = θXt dt + σ1 Xt dBt + σ2 Xt dBtH ,

(6.3.16)

Xt=0 = X0 ∈ R, 0 ≤ t ≤ T, T > 0, {θ, σ1 , σ2 } ⊂ R, σ1 σ2 < 0, θ is a parameter that we need to estimate. We suppose that the Wiener process B and the fBm B H in (6.3.16) are connected via the relations (1.8.3), (1.8.5). The integral form of equation (6.3.16) is t t t Xt = X0 + θ Xs ds + σ1 Xs dBs + σ2 Xs dBsH , 0 ≤ t ≤ T. (6.3.17) 0

0

0

The existence and the uniqueness of the solution of the equation (6.3.17) was established in Theorem 3.2.1. The Girsanov Theorem for the Mixed Fractional Diffusion Model First we try to change the probability measure Pθ for the another measure P0 , Pθ (T ) ∼ P0 (T ) in order to exclude the drift θXt dt from equations (6.3.16) and (6.3.17). We introduce probability measures P0,i , i = 1, 2 and Pθ,i , i = 1, 2 as follows. The probability measures P0,1 (t) and Pθ,1 (t) are determined by the following condition: t dPθ,1 (t) 1 t 2 = exp ψs dBs(1) − ψs ds dP0,1 (t) 2 0 0 for a nonrandom function ψs such that 0 (1)

t

ψs dBs(1) −

E exp Here the process Bt

t 0

ψs2 ds < ∞ and

1 2

t

ψs2 ds

= 1.

0

is created according to the Girsanov theorem, t (1) ψs ds, (6.3.18) Bt := Bt + 0

(1)

and Bt is a standard Wiener process with respect to the probability measure P0,1 (t). The probability measures P0,2 and Pθ,2 (t) satisfy the relation t dPθ,2 (t) 1 t 2α 2 = exp sα δs dBs(2) − s δs ds , dP0,2 (t) 2 0 0

350


t where δs satisfies the relation 0 lH (t, s)|δs | ds < ∞, t ∈ [0, T ] and admits the following integral representation: t t lH (t, s)ϕs ds = α δs ds, (6.3.19) 0 (2)

the Wiener process Bt

0

is defined from the equation

t

t

lH (t, s) dBsH,2 = α 0

s−α dBs(2) .

0

Moreover, the process

t

BtH,2 := BtH +

ϕs ds

(6.3.20)

0

is a fractional Brownian motion on [0, T ] with respect to the measure P0,2 (t). So, the total drift coefficient equals t t σ1 ψs ds + σ2 ϕs ds = θt, 0

0

and if we suppose that the functions ψ and ϕ are continuous, we obtain that σ1 ψt + σ2 ϕt = θ. Obviously, from (6.3.18)–(6.3.20) and since the likelihood ratios must coincide, we obtain that t H H M t = Mt + α s−α ψs ds

(6.3.21) dPθ,i (t) dP0,i (t)

0

and H = M H + α M t t

t

δs ds, 0

whence tα δt = ψt , t ∈ [0, T ]. Moreover,

t

t

lH (t, s)ϕs ds = α 0

s−α ψs ds.

0

Multiplying by (t − s)α−1 and integrating, we obtain t s (5) α−1 CH (t − s) u−α (s − u)−α ϕu du ds 0

0

t

(t − s)α−1

=α 0

s

δu du ds, 0

(6.3.22)


351

and the Fubini theorem applied to both sides of (6.3.22) gives t α t C(H, 2) u−α ϕu du = (t − u)α δu du, α 0 0 whence ϕt =

1 tα C(H, 3)

t

(t − u)α−1 u−α ψu du.

(6.3.23)

0

α . Substituting (6.3.23) Here C(H, 2) = CH B(α, 1 − α), C(H, 3) = C(H,2) into (6.3.21), we obtain a Volterra equation of the second kind, with weak singularity, of the form t σ2 α t σ 1 ψt + (t − u)α−1 u−α ψu du = θ, C(H, 3) 0 (5)

or 1 σ2 ρt + σ1 C(H, 3)

t

(t − u)α−1 ρu du = 0

et , σ1

(6.3.24)

where ρt = t−α ψt , et = θt−α . We solve (6.3.24) using successive approximations t 1 σ2 et (n+1) ρt + (t − u)α−1 ρu(n) du = . (6.3.25) σ1 C(H, 3) 0 σ1 (0)

(1)

σ2 and start with ρt = 0, ρt Denote for simplicity C := σ1 C(H,3) we obtain from (6.3.25) that C t et (2) ρt = (−1) (t − u)α−1 eu du + . σ1 0 σ1

=

et σ1 .

Then

It is very simple now to prove by induction that for n > 1 (n)

ρt

=

t n−1 et Γ k (α) 1 ds + , (−C)k es (t − s)kα−1 σ1 Γ (kα) σ 1 0 k=1

(n)

and the solution ρt = limn→∞ ρt ρt =

∞ 1 (−C)n σ1 n=1

evidently can be represented as a series

t

es (t − s)nα−1 0

et Γ n (α) ds + . Γ (nα) σ1

Hence ψt = tα ρt =

∞ Γ n (α) t −α tα θ θ (−C)n s (t − s)nα−1 ds + σ1 n=1 Γ (nα) 0 σ1

∞ Γ n (α) θ θ tnα + . = Γ (1 − α) (−C)n σ1 Γ ((n − 1)α + 1) σ 1 n=1

(6.3.26)

352


The series on the right-hand side of (6.3.26) can be expressed in terms of the ∞ zn (see, for example, (Po99)), Mittag–Leffler function Eρ (z) := n=0 Γ (n/ρ+1) At = −Ctα π(sin πα)−1 E1/α (−CΓ (α)tα ), and in these terms ψt =

θ (At + 1). σ1

Therefore, the likelihood ratio for the mixed fractional Brownian model equals t 1 t 2 dPθ,1 (t) = exp ψs dBs(1) − ψs ds dP0,1 (t) 2 0 0 t θ 1 θ2 t 2 (As + 1)dBs(1) − (A + 1) ds , = exp s σ1 0 2 σ12 0 whence the maximum likelihood estimate for θ equals T (1) (A(s) + 1)dBs θT1 = σ1 0 t (A(s) + 1)2 ds 0 T = σ1

0

T t (A(s) + 1)dBs + σθ1 0 (A(s) + 1)2 ds (A(s) + 1)dBs . = θ + σ1 0T t 2 (A(s) + 1) ds (A(s) + 1)2 ds 0 0

For the demonstration of the consistency of the estimate θT1 with proba t bility 1, it is sufficient to prove the divergence of the integral 0 (A(s) + 1)2 ds when t → ∞. Note that C < 0 since σσ12 < 0, and ∞

(−CΓ (α))n

n=1

=

∞ n=1

∞ tnα tnα > = (−CΓ (α))n Γ ((n − 1)α + 1) n=1 Γ (n + 1)

(−CΓ (α))n

tnα = exp {−CΓ (α)tα } → ∞, n!

when t → ∞ because α > 0 and −CΓ (α) > 0. Note that δt = t−α ψt satisfies conditions (ii)–(vi). So we have proved the following result. Theorem 6.3.5. The drift parameter maximum likelihood estimate of the linear Brownian–fractional-Brownian model (6.3.16) is consistent with probability 1. The Asymptotic Normality of the Maximum Likelihood Estimates First, consider one of the limit theorems for the stochastic integrals w.r.t. the Wiener process {Wt , Ft , t > 0}. Let {h(s), s ≥ 0} be an Fs -adapted pre t dictable function such that E 0 h2 (s)ds is finite for any t > 0 and


353

nt

Fn (t) = σ{h(s), W (s), s ≤ nt}. Consider the sequence Yn (t) := 0 h(s)dWs . Evidently, Yn (t) are Fn (t)-square-integrable martingales, t ∈ [0, T ], and their nt angle brackets equal Yn (t) = 0 h2 (s)ds. Suppose that the following conditions hold: (vii) there exists an increasing real-valued sequence {An , n ≥ 1} such that An ↑ ∞, n → ∞ and for some constant c0 > 0 we have that n P h2 (s)ds · A−2 n → c0 , 0

Consider the sequence of normalized square-integrable martingales nt n −2 P Xn (t) := A−1 n · 0 h(s)dWs . Then Xn (1) = 0 h2 (s)ds · An → c0 , therefore Xn satisfy conditions of Theorem 4.1 (LS86), if we consider the set of convergence points consisting of one point t = 1. By using this theorem we obtain the following result: Lemma 6.3.6. Let condition (vii) holds. Then the random variable

n

h(s)dWs ·

Zn := 0

−1/2

n 2

h (s)ds 0 −1/2

weakly converges to the random variable c0

N (0, 1).

Proof. From the Theorem 4.1 (LS86) and the condition (vii) we obtain that Xn (1) weakly converges to the value Z(1) of the Gaussian martingale Z with independent increments such that Z (t) = c0 t. Evidently, 1/2 Z(1) ∼ c0 N (0, 1). Moreover, from the same condition, the weak convergence holds: Zn = n 0

An h2 (s)ds

−1/2

1/2 · Xn (1) → c0

N (0, 1).

Consider the estimate θn1 satisfying relation (6.3.15). We see that for the pure fractional diffusion model h(s) = A(s) + 1 and is nonrandom. Therefore we obtain from Lemma 6.3.6 that 1/2 n (A(s) + 1)2 ds (θn1 − θ) → N (0, 1). 0

Moreover, under the assumption (viii) there exists an increasing real-valued sequence {An , n ≥ 1} such that An ↑ ∞, n → ∞ and n P s2α (Is )2 ds · A−2 n → ϕ0 , n → ∞, 0

354


we have a weak convergence ϕ0 An (θn − θ) → N (0, 1). 1/2

In this sense we say that the estimates θn and θn1 are asymptotically normal. 6.3.3 The Properties of Maximum Likelihood Estimates in Diffusion Brownian–Fractional-Brownian Models with Independent Components Now we consider an “opposite” situation when the components of the diffusion model are independent, more exactly, the processes B H and B are independent, where B H is a fBm and B is a Wiener process. The Estimates of the Drift Parameter in the Mixed Brownian– Fractional-Brownian Diffusion Model Where Bt and BtH are Independent Let the diffusion equation contain stochastic differentials with respect to fBm and the Wiener process, dXt = θXt dt + σ1 Xt dBt + σ2 Xt dBtH , Xt=0 = X0 ∈ R, 0 ≤ t ≤ T , T > 0, {θ, σ1 , σ2 } ⊂ R \ {0}, where the processes Bt and BtH are independent. Evidently, we can rewrite the solution of our simple linear equation as Xt = X0 exp{θt + σ1 Bt + σ2 BtH − 1/2σ12 t}. It was mentioned by B.L.S. Prakasa Rao in the private conversation that we cannot prove the equivalence of the observation of the whole process Xt and the observation of its two independent components, Bt and BtH , i.e., we cannot separate these components (note that the measures corresponding to these processes are singular). So, we suppose that we observe both the components. Let, as before, θ be the parameter to be estimated. We shall try to represent the estimate of θ via the components Bt and BtH because it seems to be impossible to represent it via the whole process Xt . Let Pθ be the basic probability measure corresponding to the process X. We introduce probability measures P0,i , i = 1, 2 and Pθ,i , i = 1, 2 as follows. The probability measures P0,1 (t) and Pθ,1 (t) are determined by the following condition: t dPθ,1 (t) 1 t 2 = exp ψs dBs(1) − ψs ds dP0,1 (t) 2 0 0 for a nonrandom function ψs such that

t

t

ψs dBs(1) −

E exp 0

0

1 2

ψs2 ds < ∞ and

t

ψs2 ds 0

= 1.

6.3 Properties of Maximum Likelihood Estimates (1)

Here the process Bt

355

is created according to the Girsanov theorem, t (1) Bt := Bt + ψs ds (6.3.27) 0

(1)

and Bt is a standard Wiener process with respect to the probability measure P0,1 (t). The probability measures P0,2 and Pθ,2 (t) satisfy the relation t dPθ,2 (t) 1 t 2α 2 = exp sα δs dBs(2) − s δs ds , dP0,2 (t) 2 0 0 t where δs satisfies the relation 0 lH (t, s)|δs | ds < ∞, t ∈ [0, T ], admits the following integral representation: t t lH (t, s)ϕs ds = α δs ds, (6.3.28) 0

0

(2)

the Wiener process Bt is defined from the equation t t lH (t, s) dBsH,2 = α s−α dBs(2) , 0

0

and the process

BtH,2 := BtH +

t

ϕs ds 0

is a fractional Brownian motion on [0, T ] with respect to the measure P0,2 (t). So, the total drift coefficient equals t t σ1 ψs ds + σ2 ϕs ds = θt, 0

0

or, if we suppose that the functions ψ and ϕ are continuous, σ1 ψt + σ2 ϕt = θ.

(6.3.29)

Since Bt and BtH are independent, the final probability measure P0 (t) is the product of the measures P0,1 (t) and P0,2 (t). Thus the final likelihood ratio is * t dPθ (t) 1 t 2 (1) = exp ψs dBs − ψ ds dP0 (t) 2 0 s 0 t + 1 t 2α 2 × sα δs dBs(2) − s δs ds 2 0 0

t

0

t

sα δs dBs(2) −

ψs dBs(1) +

= exp

0

1 2

0

t

ψs2 + s2α δs2 ds .

(6.3.30)

356


Solving equations (6.3.28) and (6.3.29) with respect to the functions ψt and δt , respectively, we obtain 1 (θ − σ2 ϕt ), σ1 t δt = α lH (t, s)ϕs ds . ψt =

0

(6.3.31) (6.3.32)

t

Substituting equalities (6.3.31) and (6.3.32) into likelihood ratio (6.3.30), we get at the point t = T that s T T dPθ (T ) 1 (1) α = exp (θ−σ2 ϕs ) dBs + α s lH (s, u)ϕu du dBs(2) dP0 (T ) σ1 0 0 0 s

1 − 2

0

T

.

1 (θ − σ2 ϕs )2 + s2α α σ12

s

2 / ds . lH (s, u)ϕu du

0

(6.3.33)

s

If follows from (6.3.33) that the maximum likelihood estimate θ 1T of the parameter θ satisfies the equality 1 σ1

T

dBs(1) 0

1 − 2 σ1

T

(θ − σ2 ϕs ) ds = 0, 0

which can be rewritten as follows: (1)

σ1 BT + σ2

T

ϕs ds − θT = 0. 0

This gives us the following estimate of the parameter θ: (1) σ2 σ1 BT + θ 1T = T

T 0

ϕs ds . T

(6.3.34)

Now we solve equation (6.3.29) with respect to the function ϕt and substitute it into equation (6.3.34):

T σ 1 (1) BT − ψs ds . (6.3.35) θ 1T = θ + T 0 Substituting (6.3.27) into (6.3.35) yields BT θ 1T = θ + σ1 . T

(6.3.36)

It is evident that the estimate (6.3.36) of parameter θT1 is strongly consistent.

6.3 Properties of Maximum Likelihood Estimates

357

We can construct another estimate of the parameter θ.The function δtis t expressed via ϕt by equality (6.3.28). Denote also ζt := 0 lH (t, s)ψs ds . t Then t t 1 δt = α lH (t, s)ϕs ds = α lH (t, s)(θ − σ1 ψs ) ds σ2 0 0 t t

t θ σ1 =α lH (t, s) ds − ζt σ2 σ 2 0 t =α

θ σ1 C(H)t−2α − ζt , σ2 σ2

(5)

(6.3.37)

(5)

where C(H) = CH (1 − 2α)B1 CH , B1 = B (1 − α, 1 − α). Using equality (6.3.37) for likelihood ratio (6.3.30), taking the logarithms, differentiating with respect to θ, and equating the derivative to zero, we obtain at the point t=T T T θC(H) −2α σ1 −α (2) s dBs − α s − ζs ds = 0, σ2 σ2 0 0 or T C(H) 1−2α σ1 T s−α dBs(2) − α 3 θ T +α lH (T, s)ψs ds = 0. σ2 σ2 0 0 This implies another estimate for the parameter θ: θ 2T

=

σ2 α

T 0

s−α dBs + σ1 (2)

T

l (T, s)ψs 0 H (5) CH B1 T 1−2α

ds

.

(6.3.38)

Now we substitute the expression (6.3.31) for the function ψt into rela(5) tion (6.3.38) and obtain with C(H, 1) = CH B1 that . / T T σ2 2 −α (2) lH (T, s)ϕs ds − α s dBs . θT = θ− C(H, 1)T 1−2α 0 0 T T (2) Recall that α 0 s−α dBs = 0 lH (T, s) dBsH,2 . Further, T T H,2 lH (T, s)ϕs ds − lH (T, s) dBs = − 0

0

T

lH (T, s) dBsH . 0

So, the second estimate of the parameter θ is given by T σ2 θ 2T = θ + lH (T, s) dBsH , C(H, 1)T 1−2α 0 or

358


θ 2T = θ +

σ2 α C(H, 1)

T 0

s s−α dB , 1−2α T

(6.3.39)

s is some Wiener process. The strong consistency of the estimate θ 2 where B T is also clear. Now we compare the estimates θT1 and θT2 . First we compute the variances of the remainder terms in formulae (6.3.36) and (6.3.39) and compare σ12 T −1 and σ22 C(H, 1)−2 T 2α−1 . Since H ∈ 12 , 1 , it is obvious that there exists a number N such that σ12 T −1 < σ22 C(H, 1)−2 T 2α−1 for all T > N . It means that the variance of the deviation between the estimate θT1 and true value is smaller than that of the corresponding deviation between the estimate θT2 and the true value. It this sense, the estimate θT1 is better than θT2 . Local Asymptotic Normality and Asymptotic Efficiency of the Estimate of the Drift Parameter in a Linear Brownian Diffusion Model Consider (only for comparison with the fractional case, see below) a pure linear Brownian model + 1 1 ,1 . dXtθ = β θXt dt + cXt dBt , Xt=0 = X0 , c ∈ R \ {0}, t ∈ [0, T ], β ∈ T 2 Put Θ = (0, ∞) and let θ ∈ Θ. According to Definition 2.1 (IK81), a family of measures Pθ (t) has the property of local asymptotic normality (LAN) at the point θ ∈ Θ as t → ∞, if dPθ+A(t,θ)u (t) 1 = exp uξt,θ − u2 + ζt (u, θ) Zt,θ (u) := (6.3.40) dPθ (t) 2 for some function A(t, θ) and any number u ∈ R, where ξt,θ ⇒ N (0, 1) as Pθ (t)

t → ∞ with respect to the measure Pθ (t), and ζt (u, θ) → 0, t → ∞, for all numbers u ∈ R. We say in this case that the LAN property holds for the family of measures Pθ (t) as t → ∞ at the point θ. Theorem 6.3.7. The LAN property holds for the family of measures Pθ (t) as t → ∞ at any point θ ∈ Θ. Proof. We change the probability measure Pθ (t), which corresponds to the process Xtθ for the measure P0 (t). Then the drift θXt dt disappears and we obtain t t , X 0 = X0 + c X 0 dB t

s

0

t = Bt + tθ/(cT β ) is a Wiener process w.r.t. the measure P 0 (t). where B Consider the likelihood ratio corresponding to this change of measure with ϕs = θ/(cT β ):


t θ θ2 s − 1 d B ds β 2 0 (cT β )2 0 cT θ 1 θ2 B = exp − t . t cT β 2 (cT β )2

dPθ (t) = exp dP0 (t)

t

359

Now we consider the linear model with a parameter θ shifted by A(t)u. The likelihood ratio for such a change of measure is of the form Pθ+A(t)u (t) 1 1 2 = exp (θ + A(t)u)Bt − (θ + A(t)u) t dP0 (t) cT β 2(cT β )2 and

dPθ+A(t,θ)u (t) dPθ+A(t,θ)u (t) = · dPθ (t) dP0 (t)

dPθ (t) dP0 (t)

−1

1 1 θ 1 θ2 2 t − B = exp (θ + A(t)u) B (θ + A(t)u) t − − t t cT β 2(cT β )2 cT β 2 (cT β )2 uA(t) 1 2 A2 (t) A(t)uθ = exp Bt − u t− t . cT β 2 (cT β )2 (cT β )2 √ Set A(t) := cT β / t. Then √ t dPθ+A(t,θ)u (t) B 1 2 uθ t = exp u √ − u − . dPθ (t) cT β t 2

√ t / t ⇒ N (0, 1) under both the measures P0 (t) and Pθ (t) and, in Since B √ addition, uθ t/(cT β ) → 0 as t → ∞ for T ≥ t and α > 12 , the above definition implies the LAN property for the family Pθ (t) as t → ∞ and at any point θ ∈ Θ.

Consider now the asymptotic efficiency of the estimate of parameter θ. According to definition (11.3), introduced in the monograph (IK81), an estimate {θt , t > 0} of a parameter θ is asymptotically efficient under the LAN property for the cost function ω(A−1 (t, θ)x) at the point θ if lim lim sup EPθ (t) ω A−1 (t, θ)(θt − θ ) = Eω(N (0, 1)). δ→0 t→∞ |θ −θ| 0.

360


Further we consider the cost function ω A−1 (t, θ)x ∈ Wp , where Wp ⊂ W is the class of functions of W that have a dominant polynomial. Consider the maximum likelihood estimate of the parameter θ in a linear Brownian model cT β cT β 1 cT β Bt = Bt . θt = Bt + β θt = θ + t t cT t To prove the asymptotic efficiency of the estimate θt we use Theorem 1.3 of Chapter III from (IK81). According to this theorem, the estimate θt is asymptotically efficient in the sense mentioned above if the following conditions hold: limt→∞ A−1 (t, θ2 )A(t, θ1 ) = B(θ1 , θ2 ) exists, the convergence is uniform in θi ∈ Θ and B(θ1 , θ2 ) is continuous in θ1 ; (b) ζt (θ) := A−1 (t, θ)(θt − θ) ⇒ N (0, 1) uniformly in θi ∈ Θ as t → ∞ with respect to the measure Pθ (t); (c) for any N > 0 random variables |A−1 (t, θ)(θt −θ)|N , are Pθ (t)-integrable for any θ ∈ Θ uniformly in t > t0 (N ). (a)

β

√ does not depend on Condition (a) holds in our case because A(t) = cT t θ. Now we check condition (b): √ t cT β 1 −1 Bt = Bt √ ⇒ N (0, 1) ζt (θ) = A (t, θ) θt − θ = cT β t t

under both the measures P0 (t) and Pθ (t). Condition (c) now is evident. Thus the estimate θt is asymptotically efficient as t → ∞. Local Asymptotic Normality and Asymptotic Efficiency of the Estimate of the Drift Parameter in a Linear Fractional Brownian Diffusion Model Now consider a pure linear fractional Brownian model 1 θXt dt + Xt dBtH , Xt=0 = X0 , θ ∈ Θ, t ∈ [0, T ], β ∈ (1 − H, 1]. Tβ It will be clear later that in this model it is sufficient to consider β ∈ 1 − H, 12 . Now ϕt = θ/T β . Then t t θ θ δs ds = lH (t, s) β ds = β C(H, 1)t1−2α , δt = (θ/T β )C(H, 1)t−2α α . α T T 0 0 t s C(H, 1)−1 t2α−1 , where 0 s−α dB Therefore θt = T β α t t θ s = α α s−α dB s−α dBs + β C(H, 1)t1−2α . T 0 0 dXt =

In other words,

t 0 s−α dBs T βα . θt = θ + C(H, 1)t1−2α


361

Theorem 6.3.8. The LAN property holds for the family Pθ (t) as t → ∞ at any point θ ∈ Θ. Proof. We change the probability measure Pθ (t) for the measure P0 (t). As a result, the drift θXt dt disappears. The corresponding likelihood ratio is given by t dPθ (t) 1 t 2α 2 α = exp s δs dBs − s δs ds dP0 (t) 2 0 0 θC(H, 1) α t −α 1 2 1−2α = exp s dBs − (θC(H, 1)) t . Tβ 2T 2β 0 Now we consider the linear model with parameter θ shifted by A(t)u and denote for simplicity K = C(H). (θ + A(t)u)K t Pθ+A(t)u (t) s = exp s−α dB dP0 (t) Tβ 0 1−2α

1 2 t ((θ + A(t)u)K) . 2T 2β 1 − 2α The likelihood ratio for this model is of the form −1 dPθ+A(t,θ)u (t) dPθ+A(t,θ)u (t) dPθ (t) = · dPθ (t) dP0 (t) dP0 (t) t K K t1−2α K t1−2α 1 −α A(t)u − θ = exp A(t)u s d B − . s Tβ 2 T β 1 − 2α T β 1 − 2α 0

−

/Kt1−H . Then the likelihood ratio obtains the form Set A(t) := T β α t s dPθ+A(t,θ)u (t) s−α dB 1 2 uθKt1−H 0 = exp α u . − u − dPθ (t) t1−H 2 T βα Since

t α

0

s s−α dB ⇒ N (0, 1) 1−H t

and

uθKt1−H → 0 as t → ∞, T βα the LAN property holds for the family Pθ (t) as t → ∞ at any point θ ∈ Θ.

Now we check the asymptotic efficiency of the estimate θt . Consider conditions (a)-(c). Two of them, (a) and (c), are evident. To check (b) we use the following relations: t C(H)t1−H T β 0 s−α dBs ζt (θ) = A−1 (t, θ)(θt − θ) = T βα C(H, 1)t1−2α

362


=

t −α s dBs 0

t1−H

α

⇒ N (0, 1).

Therefore, the estimate θt of the parameter θ is asymptotically efficient as t → ∞. Remark 6.3.9. The maximum likelihood estimators for the drift coefficient in the stochastic differential equations involving fBm were considered also in the paper (TV03), the estimate of the diffusion coefficient for diffusion driven by fBm is contained in the paper (LL00).

A Mandelbrot–van Ness Representation: Some Related Calculations

Now we calculate the constant that appeared in the Mandelbrot–van Ness representation of fBm (see Section 1.3, Theorem 1.3.1). Lemma A.0.1. The following equalities hold: (2) CH

:=

1 ((1 + s) − s ) ds + 2H R+

−1

α 2

α

=

(2H sin πHΓ (2H)) Γ (1 + α)

1/2

(2)

Proof. Recall that the constant CH is chosen to normalize the fBm H (2) (2) α B t = CH kH (t, u)dWu = CH Γ (1 + α) (I− 1(0,t) )(x)dWx R

R

(see Lemma 1.1.3). Therefore, the first equality is evident, since

0

2

R

(kH (t, u)) du =

−∞

=t

2H 0

t

((t − x) − (−x) ) dx +

∞

(t − x)2α dx

α 2

α

0

1 . ((1 + s) − s ) ds + 2H α

α 2

We obtain the second equality if we note that 2 1 α 2 α 1(0,t) )(x) dx F(I (I− 1(0,t) )(x) dx = − 2π R R and according to Theorem 1.1.5

απi −α α 1(0,t) )(x)(λ) = 1 F(I sign λ (λ)|λ| exp (0,t) − 2 =

απi eitλ − 1 −α |λ| exp sign λ . iλ 2

.

364

A Mandelbrot–van Ness Representation: Some Related Calculations

Therefore, R

1 2π

R

|eitλ − 1|2 |λ|−2α−2 dλ

1 sin2 tλ|λ|−2α−2 dλ 2π R R ∞ ∞ 2 (1 − cos tλ) sin2 tλ 1 1 = dλ + dλ 2α+2 π 0 λ π 0 λ2α+2 ∞ 1 (1 − cos λ)2 1 ∞ sin2 λ t2H , = t2H dλ + dλ = 2α+2 2α+2 π 0 λ π 0 λ 2H sin πHΓ (2H) =

1 2π

α (I− 1(0,t) )(x)2 dx =

(1 − cos tλ)2 |λ|−2α−2 dλ +


B Approximation of Beta Integrals and Estimation of Kernels

These results were obtained by E.Valkeila (KMV05). Lemma B.0.1. Assume that −1 < δ < 0, β > −1 and n ≥ 2. Then for β ≥ 0 |I(δ, β) − In (δ, β)| ≤ C1 (δ, β)n−α−1 ,

(B.0.1)

and with −1 < β < 0 we have |I(α, β) − In (δ, β)| ≤ C2 (δ, β)n−α−β−1

(B.0.2)

(for the value of the constants, see the proof ). Proof. We start the proof with I(δ, β) − In (δ, β) =

1 n

sδ (1 − s)β ds − n−δ−1

k+1

0 n−2

+

k n

k=1 1

+

n

1 1− n

s (1 − s) − δ

β

k+1 n

δ β k ds 1− n

sδ (1 − s)β ds − n−β−1 .

We work first with the integral on (0, 1/n). We have

1 n

sδ (1 − s)β ds − n−δ−1 =

0

1 n

sδ − n−δ ds

0

1 n

+

β sδ (1 − s) − 1 ds;

0

here

0≤ 0

1 n

sδ − n−δ ds = −δ/(δ + 1)n−δ−1 ,

(B.0.3)

366


if β ≥ 0, then

β sδ (1 − s) − 1 ds| ≤

1 n

| 0

1 n

sδ ds

0

and if β < 0 and s ≤ 1/n, then 0 ≤ (1 − s) − 1 ≤ 2−β − 1. Use these estimates in (B.0.3) to obtain β

|

1 n

sδ (1 − s)β ds − n−δ−1 | ≤ C1 (δ, β)n−δ−1 .

(B.0.4)

0

Next, we work with the integral on (1 − 1/n, 1). We have

1

1 1− n

+

sδ (1 − s)β ds − n−β−1 =

1

β

1 1− n

(1 − s)

1

1 1− n

β (1 − s) − n−β ds

sδ − 1 ds,

and this gives 1 |β| −β−1 n sδ (1 − s)β ds − n−β−1 ≤ + 2−δ n−β−1 . 1 1 +β 1− n

(B.0.5)

We continue with the middle term. We have

δ β n−2 k+1 n k+1 1 k β δ s (1 − s) ds − 1− k n n n n k=1

=

+

n−2 k=1

n−2

k n

k=1

k+1 n k n

k+1 n

k+1 n

sδ −

k+1 n

δ β

(1 − s) ds

δ β k β (1 − s) − 1 − ds . n

(B.0.6)

The first term on the right-hand side of (B.0.6) is always positive, when δ < 0. We use the estimate δ

δ

δ

sδ − ((k + 1) /n) ≤ (k/n) − ((k + 1) /n) . If β ≥ 0, then (1 − s)β ≤ 1 and so for the first term on the right-hand side of (B.0.6) we obtain

n−2 (k+1)/n δ β δ s − ((k + 1) /n) (1 − s) ds 0≤ k=1

k/n


≤ n−δ−1

n−2

k δ − (k + 1)

δ

≤ n−δ−1 .

367

(B.0.7)

k=1

If β ≤ 0 then

(k+1)/n

sδ − ((k + 1)/n)

δ

β

(1 − s) ds ≤

k/n

− (k + 1)

δ

(n − k)

β+1

− (n − (k + 1))

β+1

1 n−δ−β−1 k δ 1+β

δ ≤ n−δ−β−1 k δ − (k + 1) ,

and this gives the estimate

n−2 (k+1)/n α β 0≤ (sα − ((k + 1) /n) ) (1 − s) ds ≤ n−α−β−1 . k=1

(B.0.8)

k/n

Finally, the second part of the middle term is

n−2 (k+1)/n δ β β ((k + 1) /n) (1 − s) − (1 − k/n) ds . Jn := k=1

k/n

If β ≥ 0, then with calculations similar to above |Jn | ≤ n−δ−1 ,

(B.0.9)

1 |Jn | ≤ − 2β n−α−β−1 . β

(B.0.10)

and if β < 0, then

Combining the bounds (B.0.3)–(B.0.7) and (B.0.9) we get C1 (δ, β), and com

bining the bounds (B.0.3)–(B.0.6), (B.0.8) and (B.0.10) we get C2 (δ, β). Lemma B.0.2. Put 3/4−H 2H 2H n−1 k + 1 1/4−H k k+1 k Hn := − 1− n n n n k=0 H−3/4 3/4−H k+1 1 k −2H 1− . n n n Then

1

|Hn | ≤ Cn− min(1, 4 +H) .

Proof. The proof of Lemma B.0.2 is similar to Lemma B.0.1. The proof of the following lemma is obvious.

(B.0.11)

368


Lemma B.0.3. Consider the expression

2H 2H 2 n−1 k k+1 1 u ¯n (H) := − . n n n k=0

Then |¯ un (H)| ≤

C . n2

(B.0.12)

References

[AOPU00] Aase, K., Øksendal, B., Privault, N., Ubøe, J.: White noise generalization of the Clark-Haussmann-Ocone theorem with applications to mathematical finance. Finance Stoch., 4, 465–496 (2000) [AS96] Abry, P., Sellan, F.: The wavelet-based synthesis for fractional Brownian motion proposed by F. Sellan and Y. Meyer: Remarks and fast implementation. Appl. Comp. Harmon. Analysis, 3, 377–383 (1996) [AS95] Adler, R.J.; Samorodnitsky, G.: Super fractional Brownian motion, fractional super Brownian motion and related self-similar (super) processes. Ann. Prob., 23, 743–766 (1995) [ALN01] Alòs, E., Le´ on, I.A., Nualart, D.: Stratonovich stochastic calculus with respect to fractional Brownian motion with Hurst parameter less than 1/2. Taiwanesse J. Math., 5, 609–632 (2001) [AMN00] Alòs, E., Mazet, O., Nualart, D.: Stochastic calculus with respect to fractional Brownian motion with Hurst parameter less than 1/2. Stoch. Proc. Appl., 86, 121–139 (2000) [AMN01] Alòs, E., Mazet, O., Nualart, D.: Stochastic calculus with respect to Gaussian processes. Ann. Prob., 29, 766–801 (2001) [AN02] Alòs, E., Nualart, D.: Stochastic integration with respect to the fractional Brownian motion. Stoch. Stoch. Rep., 75, 129–152 (2002) [And05] Androshchuk, T.: The approximation of stochastic integral w.r.t. fBm by the integrals w.r.t. absolutely continuous processes. Prob. Theory Math. Stat., 73, 11–20 (2005) [AM06] Androshchuk, T., Mishura Y.: Mixed Brownian–fractional Brownian model: absence of arbitrage and related topics. Stochastics: Intern. J. Prob. Stoch. Proc., 78, 281–300 (2006) [AG03] Anh, V., Grecksch, W.: A fractional stochastic evolution equation driven by fractional Brownian motion. Monte Carlo Methods Appl. 9, 189–199 (2003)

370

References

[AHL01] Anh, V., Heyde, C., Leonenko, N.: Dynamic models of long– memory processes driven by Lévy noise with applications to finance and macroeconomics. J. Appl. Probab. 39, 730–747 (2002) [AI04] Anh, V., Inoue, A.: Prediction of fractional Brownian motion with Hurst index less than 1/2. Bull. Aust. Math. Soc., 70, 321– 328 (2004) [AI05a] Anh, V., Inoue, A.: Financial markets with memory I: Dynamic models. Stoch. Anal. Appl., 23, 275–300 (2005) [AI05b] Anh, V., Inoue, A.: Financial markets with memory II: Innovation processes and expected utility maximization. Stoch. Anal. Appl., 23, 301–328 (2005) [ALN01] Anh, V., Leonenko, N., Nguyen, C.: Stochastic differential equations with fractional Riesz–Bessel input. Inform. Techn. Econom. Management, 1, 1–16 (2001) [ALP02a] Ayache, A., Leger, S., Pontier M.: Drap brownien fractionnaire. (Fractional Brownian sheet.) Potential Analysis, 17, 31– 43 (2002) [ALP02b] Ayache, A., Leger, S., Pontier, M.: Les ondelettes a la conquite du drap brownien fractionnaire. (Wavelets conquering the fractional Brownian field.) C. R. Acad. Sci. Paris Sér. I Math., 335, 1063–1068 (2002) [BLOPST03] Bardet, J.M., Lang, G., Oppenheim, G., Philippe, A., Stoev, S., Taqqu, M.S.: Semi-parametric estimation of the long-range dependence parameter: a survey. In: Doukhan, P., Oppenheim, G., Taqqu, M. (eds) Theory and Applications of Long-Range Dependence. Birkh¨ auser, Boston, 557–579 (2003) [BJ06] Bardina, X., Jolis, M.: Multiple fractional integral with Hurst parameter less than 1/2. Stoch. Proc. Appl., 116, 463–479 (2006) [BP88] Barton, R.J., Vincent Poor, H.: Signal detection in fractional Gaussian noise. IEEE Trans. Inform. Theory, 34, 943–959 (1988) [BaNu06] Baudoin, F., Nualart, D.: Notes on the two-dimensional fractional Brownian motion. Ann. Prob., 34, 159–180 (2006) [Ben03a] Bender, C.: Integration with respect to a fractional Brownian motion and related market models. PhD Thesis, Hartung-Gorre Verlag, Konstanz (2003) [Ben03b] Bender, C.: An S-transform approach to integration with respect to a fractional Brownian motion. Bernoulli, 9, 955–983 (2003) [Ben03c] Bender, C.: An Itˆ o formula for generalized functionals of a fractional Brownian motion with arbitrary Hurst parameter. Stoch. Proc. Appl., 104, 81–106 (2003) [BE03] Bender, C., Elliott, R.J.: On the Clark–Ocone theorem for fractional Brownian motions with Hurst parameter bigger than a half. Stoch. Stoch. Rep., 75, 391–405 (2003)

References

371

[BE04] Bender, C., Elliott, R.J.: Arbitrage in a discrete version of the Wick-fractional Black-Scholes market. Math. Oper. Res., 29, 935–945 (2004) [BSV06] Bender, C., Sottinen, T., Valkeila, E.: No-arbitrage pricing beyond semimartingales. Preprint, Weierstrass-Institut Ang. Anal. Stoch., No. 1110, Berlin (2006) [Ben03] Benth, F.E.: On arbitrage-free pricing of weather derivatives based on fractional Brownian motion. Appl. Math. Finance, 10, 303–324 (2003) [Ber94] Beran, J.: Statistics for Long-memory Processes. Chapman & Hall, New York (1994) [Ber69] Berman, S.M.: Harmonic analysis of local times and sample functions of Gaussian processes. Trans. Amer. Math. Soc., 143, 269– 281 (1969) [Ber70] Berman, S.M.: Gaussian processes with stationary increments: local times and sample function properties. Ann Math. Stat., 41, 1260–1272 (1970) [BGK06] Bhansali, R.J., Giraitis, L., Kokoszka, P.S.: Estimation of the memory parameter by fitting fractionally differenced autoregressive models. J. Multivariate Anal. 97, 2101–2130 (2006). [BHOS02] Biagini, F., Hu, Y., Øksendal, B., Sulem, A.: A stochastic maximum principle for processes driven by a fractional Brownian motion. Stoch. Proc. Appl., 100, 233–254 (2002) [BHOZ07] Biagini, F., Hu, Y., Øksendal, B., Zhang T.: Stochastic Calculus for Fractional Brownian Motion and Applications. Forthcoming. Probability and Its Applications, Springer (2007) [BO02] Biagini, F., Øksendal, B.: Minimal variance hedging for fractional Brownian motion. Methods Appl. Analysis, 10, 347–362 (2003) [BO04] Biagini, F., Øksendal, B.: Forward integrals and an Itˆ o formula for fractional Brownian motion. Dept. of Math., University of Oslo, No. 22 (2004) [Bi81] Bichteler, K.: Stochastic integration and Lp -theory of semimartingales. Ann. Prob., 9, 49–89 (1981) [BG96] Bisaglia, L., Guegan, D.: A review of techniques of estimation in long memory processes: application to intraday data. Comp. Stat. Data Analysis, 26, 61–81 (1997) [BG98] Bisaglia, L., Guegan, D.: A comparison of techniques of estimation in long-memory processes. Comp. Stat. Data Analysis, 27, 61–81 (1998) [BH05] Björk, T., Hult, H.: A note on Wick products and the fractional Black–Scholes model. Finance Stoch., 9, 197–209 (2005) [BlSc73] Black, F., Scholes, M.: The pricing of options and corporate liabilities. J. Political Econ., 81, 637–654 (1973)

372

References

[Bor84] Borovkov, A.A.: Mathematical Statistics. Estimates of Parameters. Testing Hypotheses. (Russian) Nauka, Moscow (1984) [BL01] Boufoussi, B., Lakhel, El H.: Weak convergence in Besov spaces to fractional Brownian motion. C. R. Acad. Sci. Paris Sér. I Math., 333, 39–44 (2001) [BO03] Boufoussi, B., Ouknine, Y.: On a SDE driven by a fractional Brownian motion and with monotone drift. Electronic Comm. Prob., 8, 122–134 (2003) [BK00] Buldygin, V.V., Kozachenko, Yu.V.: Metric Characterization of Random Variables and Random Processes. AMS, Providence, Rhode Island (2000) [CC98] Carmona, P., Coutin, L.: Fractional Brownian motion and the Markov property. Electronic Commun. Prob., 3, 95–107 (1998) [CC00] Carmona, P., Coutin, L.: Integrale stochastique pour le mouvement brownien fractionnaire. C. R. Acad. Sci. Paris Sér. I Math., 330, 231–236 (2000) [CCM98] Carmona, P., Coutin, L., Montseny, G.: Applications of a representation of long memory Gaussian processes. Preprint, Université de Toulouse. Laboratoire de Statistique et Probabilités (1998) [CCM03] Carmona, P., Coutin, L., Montseny, G.: Stochastic integration with respect to fractional Brownian motion. Ann. Inst. Henri Poincaré, Prob. Stat., 39, 27–68 (2003) [Che01a] Cheridito, P.: Mixed fractional Brownian motion. Bernoulli, 7, 913–934 (2001) [Che01b] Cheridito, P.: Regularizing fractional Brownian motion with a view towards stock price modelling. PhD thesis, Zurich (2001) [CKM03] Cheridito, P., Kawaguchi, H., Maejima, M.: Fractional Ornstein– Uhlenbeck processes. Electronic J. Prob., Paper no. 3, 1–14 (2003) [CN02] Cheridito, P., Nualart, D.: Symmetric integration with respect to fractional Brownian motion. Preprint, University of Barcelona (2002) [CN05] Cheridito, P., Nualart, D.: Stochastic integral of divergence type with respect to the fractional Brownian motion with Hurst parameter H < 1/2. Ann. Inst. Henri Poincaré, Prob. Stat. 41, 1049–1081 (2005) [CMS80] Chou, C.S., Meyer, P.A., Stricker, C.: Sur les intégrales stochastiques de processus prévisibles non bornées. Séminaire Prob. XIV, In: Lect. Notes Math. 784, 128–139 (1980) [ChW83] Chung, K.L., Williams, R.J.: Introduction to Stochastic Integration. Birkh¨ auser, Boston (1983) [CKR93] Ciesielski, Z., Kerkyacharian, G., Roynette, B.: Quelques espaces fouctionnels associes á des processus gaussiens. Studia Mat., 107, 171–204 (1993)

References

373

[CM95] Cioczek-Georges, R., Mandelbrot, B.: A class of micropulses and antipersistent fractional Brownian motion. Stoch. Proc. Appl., 60, 1–18 (1995) [CM96] Cioczek-Georges, R., Mandelbrot, B.: Alternative micropulses and fractional Brownian motion. Stochastic Processes Appl., 64, 143–152 (1996) [CNW06] Corcuera, J.M., Nualart, D., Woerner, J..C.: Power variation of some integral long-memory processes. Bernoulli, 12, 713–735 (2006). [Cou07] Coutin, L.: An introduction to (stochastic) calculus with respect to fractional Brownian motion. In: Donati-Martin, C. et al. (eds) Séminaire de Probabilités XL, Lecture Notes in Math. 1899, Springer (2007) [CD99] Coutin, L., Decreusefond, L.: Abstract nonlinear filtering theory in the presence of fractional Brownian motion. Ann. Appl. Prob., 9, 1058–1090 (1999) [CD01] Coutin, L., Decreusefond, L.: Stochastic Volterra equations with singular kernels. In: Cruzeiro, Ana Bela et al. (eds), Stochastic Analysis and Mathematical Physics. Birkh¨ auser, Boston. Prog. Probab. 50, 39–50 (2001) [CNT01] Coutin, L., Nualart, D., Tudor, C.A.: Tanaka formula for the fractional Brownian motion. Stoch. Proc. Appl., 94, 301–315 (2001) [CQ00] Coutin, L., Qian, Z.: Stochastic differential equations for fractional Brownian motions. C. R. Acad. Sci. Paris Sér. I Math., 331, 75–80 (2000) [CQ02] Coutin, L., Qian, Z.: Stochastic analysis, rough path analysis and fractional Brownian motion. Prob. Theory Rel. Fields, 122, 108–140 (2002) [CM96] Csörg¨ o, S., Mielniczuk, J.: Extreme values of derivatives of smoothed fractional Brownian motions. Prob. Math. Stat., 16, 211–219 (1996) [CKW95] Cutland, N.G., Kopp, P.E., Willinger, W.: Stock price returns and the Josef effect: a fractional version of the Black–Scholes model. In: Bolthausen, E., Dozzi, M., Russo, F. (eds) Sem. Stoch. Anal., Random Fields Appl. Birkh¨ auser, Boston, 327– 351 (1995) [DH96] Dai, W., Heyde, C.C.: Itˆ o’s formula with respect to fractional Brownian motion and its application. J. Appl. Math. Stoch. Analysis, 9, 439–448 (1996) [Das98] Dasgupta, A.: Fractional Brownian motion. Its properties and applications to stochastic integration. PhD Thesis, University of North Carolina (1998)

374

References

[Daye03] Daye, Z.J.: Introduction to Fractional Brownian Motion. Advanced FE, www.stat.purdue.edu/∼zdaye/Projects/fBM.doc (2003) [Dec01] Decreusefond, L.: A Skorokhod-Stratonovitch integral for the fractional Brownian motion. In: Stochastic Analysis and Related Topics VII. Proc. 7th Silivri Workshop. Birkh¨ auser, Boston, 177–198 (2001) [Dec03] Decreusefond, L.: Stochastic integration with respect to fractional Brownian motion. In: Doukhan, P., Oppenheim, G., Taqqu, M. (eds) Theory and Applications of Long-Range Dependence. Birkh¨ auser, Boston, 203–225 (2003) ¨ unel, A.S.: Application du calcul des varia[DU95] Decreusefond, L., Ust¨ tions stochastiques an mouvement brownien fractionnaire. C. R. Acad. Sci. Paris Sér. I Math., 321, 1605–1608 (1995) ¨ unel, A.S.: Fractional Brownian motion: [DU98] Decreusefond, L., Ust¨ Theory and applications. ESAIM Proc., 5, 75–86 (1998) ¨ unel, A.S.: Stochastic analysis of the frac[DU99] Decreusefond, L., Ust¨ tional Brownian motion. Potential Analysis, 10, 177–214 (1999) [DS06] Delbaen, F., Schachermayer, W.: The Mathematics of Arbitrage. Springer, Berlin (2006) [Del72] Dellacherie, C.: Capacités et Processus Stochastiques. Springer, Berlin (1972) [DM82] Dellacherie, C., Meyer, P.-A.: Probabilities and Potential. B: Theory of Martingales. North-Holland, Amsterdam (1982) [DH03] Deo, R.S., Hurvich, C.M.: Estimation of long-memory in volatility. In: Doukhan, P., Oppenheim, G., Taqqu, M. (eds) Theory and Applications of Long-Range Dependence. Birkh¨ auser, Boston, 313–325 (2003) [DDM70] Doléans-Dade, C., Meyer, P.A.: Intégrales stochastiques par rapport aux martingales locales. Séminaire Prob. IV. In: Lect. Notes Math. 124, 77–107 (1970) [Do03] Doukhan, P.: Models, inequalities and limit theorems for stationary sequences. In: Doukhan, P., Oppenheim, G., Taqqu, M. (eds) Theory and Applications of Long-Range Dependence. Birkh¨ auser, Boston, 43–100 (2003) [DKL03] Doukhan, P., Khezour, A., Lang,G.: Non-parametric estimation under long-range dependence. In: Doukhan, P., Oppenheim, G., Taqqu, M. (eds) Theory and Applications of Long-Range Dependence. Birkh¨ auser, Boston, 303–313 (2003) [DN99] Dudley, R.M., Norvai˘sa, R.: An introduction to p-variation and Young integrals. Concentrated advanced course. Maphysto, CMPhS University of Aarhus, Denmark (1999) [Dun00] Duncan, T.E.: Some applications of fractional Brownian motion to linear systems. In: System theory: modeling, analysis, and

References

[Dun01] [Dun04]

[Dun06] [DHP00]

[DPM02]

[DF02]

[DvZ04] [DvZ05]

[ElN93]

[ElN99] [Ell82] [EvH01]

[EvH03]

[ENO02]

[Fel36] [Fer74]

375

control. Kluwer Int. Ser. Eng. Comput. Sci, 518, Kluwer Academic Publishers, Boston, 97–105 (2000) Duncan, T.: Some aspects of fractional Brownian motion. Nonlinear Analysis Theory Methods Appl., 47, 4775–4782 (2001) Duncan, T.: Some processes associated with a fractional Brownian motion. In: Mathematics of Finance. Proc. AMS-IMS-SIAM Conf. Contemporary Mathematics, 351, AMS, Providence, RI, 93–101 (2004) Duncan, T.E.: Prediction for some processes related to a fractional Brownian motion. Stat. Prob. Lett., 76, 128–134 (2006) Duncan, T.E., Hu, Y., Pasik-Duncan, B.: Stochastic calculus for fractional Brownian motion, I. Theory. SIAM J. Control Optimization, 38, 582–612 (2000) Duncan, T., Pasik-Duncan, B., Maslowski, B.: Fractional Brownian motion and stochastic equations in Hilbert spaces. Stoch. Dyn., 2, 225–250 (2002) Dzhaparidze, K., Ferreira, J.A.: A frequency domain approach to some results on fractional Brownian motion. Stat. Prob. Lett., 60, 155–168 (2002) Dzhaparidze, K., van Zanten, H.: A series expansion of fractional Brownian motion. Prob. Theory Relat. Fields, 130, 39–55 (2004) Dzhaparidze, K., van Zanten, H.: Krein’s spectral theory and the Paley-Wiener expansion for fractional Brownian motion. Ann. Prob., 33, 620–644 (2005) El-Nouty, Ch.: A Hanson-Russo-type law of the iterated logarithm for fractional Brownian motion. Stat. Prob. Lett. 17, 27–34 (1993) El-Nouty, Ch.: On the large increments of fractional Brownian motion. Stat. Prob. Lett., 41, 169–178 (1999) Elliott R.J.: Stochastic Calculus and Applications. Applications of Mathematics, 18. Springer-Verlag, New York (1982) Elliott, R.J., van der Hoek, J.: Fractional Brownian motion and financial modelling. In: Kohlmann, M., Tang, S. (eds) Proc. Workshop Math. Finance, Konstanz, Birkh¨ auser, Basel, 140–150 (2001) Elliott, R.J., van der Hoek, J.: A general fractional white noise theory and applications to finance. Math. Finance, 13, 301–330 (2003) Erraoui, M., Nualart, D., Ouknine, Y.: Hyperbolic stochastic partial differential equations with additive fractional Brownian sheet. Preprint No. 307, Univ.de Barcelona (2002) Feller, W.: Zur Theorie der Stochastischen Prozesse (Existenzund Eindeutigkeitssatze). Math. Ann., 113, 113–160 (1936) Fernique, X.M.: Regularitè des trajectories de fonctions alèatories gaussiennes. In: Lect. Notes Math., 480, 2–95 (1974)

376

References

[FdP99] Feyel, D., de la Pradelle, A.: Fractional integrals and Brownian processes. Potential Analysis, 10, 273–288 (1999) [FdP01] Feyel, D., de la Pradelle, A.: The FBM Itˆ o’s formula through analytic continuation. Electronic J. Prob., paper 26, 1–22 (2001) [Fol81a] F¨ ollmer H.: Calcul d’Itˆ o sans probabilités. Séminaire Prob. XV, In: Lect. Notes Math. 850, 144–150 (1981) [Fol81b] F¨ ollmer, H.: Dirichlet processes. In: Stochastic integrals, Proc. LMS Durham Symp. 1980, Lect. Notes Math. 851, 476–478 (1981). [Gap04] Gapeev, P.: On arbitrage and Markovian short rates in fractional bond markets. Stat. Prob. Lett., 70, 211–222 (2004) [GRR71] Garsia, A., Rodemich, E., Rumsey, H.: A real variable lemma and the continuity of paths of some Gaussian processes. Indiana Univ. Math. Journal, 20, 565–578 (1970/71) [GR03a] Giraitis, L., Robinson, P.M.: Edgeworth expansions for semiparametric Whittle estimation of long memory. Ann. Stat. 31, 1325–1375 (2003) [GR03b] Giraitis, L., Robinson, P.M.: Parametric estimation under long– range dependence. In: Doukhan, P., Oppenheim, G., Taqqu M. (eds) Theory and Applications of Long-Range Dependence. Birkh¨ auser, Boston, 229–250 (2003) [Gje94] Gjessing, H.: Wick calculus with applications to anticipating stochastic differential equations. Manuscript, Univ. of Bergen (1994) [Gol84] Goldman, A.: Estimations analytiques concernant le mouvement brownien fractionnaire a` plusieurs paramètres. (Analytic estimations concerning a fractional Brownian motion.) (French) C. R. Acad. Sci. Paris Sér. I Math., 298, 91–93 (1984) [Gor77] Gorodeckii, V.V.: On convergence to semi-stable Gaussian processes. Theory Prob. Appl. 22, 498–508 (1978) [GRV03] Gradinaru, M., Russo, F., Vallois, P.: Generalized covariations, local time and Stratonovich Itˆ o’s formula for fractional Brownian motion with Hurst index H ≥ 1/4. Ann Prob., 31, 1772–1820 (2003) [GR80] Gradshteyn, I.S.; Ryzhik, I.M.: Table of Integrals, Series, and Products. Academic Press, New York (1980) [GA98] Grecksch, W., Anh, V.: Approximation of stochastic differential equations with modified fractional Brownian motion. Z. Anal. Anwend., 17, 715–727 (1998) [GA99a] Grecksch, W., Anh, V.: Approximation of stochastic Hammerstein integral equation with fractional Brownian motion input. Monte Carlo Methods Appl., 5, 311–323 (1999) [GA99b] Grecksch, W., Anh, V. A parabolic stochastic differential equation with fractional Brownian motion input. Stat. Prob. Lett., 41, 337–346 (1999)

References

377

[GN96] Grippenberg, G., Norros, I.: On the prediction of fractional Brownian motion. J. Appl. Prob., 33, 400–410 (1996) [GuNu05] Guerra, J., Nualart, D.: The 1/H-variation of the divergence integral with respect to fractional Brownian motion for H > 1/2 and fractional Bessel processes. Stoch. Proc. Appl., 115, 91–115 (2005) [GK05] Gushchin, A.A., K¨ uchler, U.: On recovery of a measure from its symmetrization. Theory Prob. Appl. 49, 323–333 (2005) [HM06a] Herbin, E., Merzbach, E.: A characterization of the set-indexed fractional Brownian motion by increasing paths. C. R., Math., Acad. Sci. Paris 343, 767–772 (2006) [HM06b] Herbin, E., Merzbach, E.: A set-indexed fractional Brownian motion. J. Theor. Probab. 19, 337–364 (2006) [HH76] Hida, T., Hitsuda, M.: Gaussian Processes. American Mathem. Society: Translations of Math. Monographs (1976) [Hit68] Hitsuda, M.: Representation of Gaussian processes equivalent to Wiener process. Osaka J. Mathematics, 5, 299–312 (1968) [Ho96] Ho, H.-C. On central and noncentral limit theorems in density estimation for sequences of long-range dependence. Stoch. Proc. Appl., 63, 153-174 (1996) [HH03] Ho, H.-C., Hsing, T.: A decomposition for generalized U -statistics of long memory linear processes. In: Doukhan, P., Oppenheim, G., Taqqu, M. (eds) Theory and Applications of Long-Range Dependence. Birkhäuser, Boston, 143–156 (2003) [HC93] Hodges, S., Carverhill, A.: Quasimean reversion in an efficient stock market: the characterization of economics equilibria which support Black-Scholes option pricing. The Economic J., 103, 395-405 (1993) [HOUZ96] Holden, H., Øksendal, B., Ubøe, J., Zhang, T.: Stochastic Partial Differential Equations. Birkh¨ auser, Boston (1996) [Hu05] Hu,Y.: Integral tranformations and anticipative calculus for fractional Brownian motions. Mem. Amer. Math. Soc., 175 (2005) [HN04] Hu,Y., Nualart, D.: Some processes associated with fractional Bessel processes. Preprint No. 347, Univ. de Barcelona (2004) [HN06] Hu,Y., Nualart, D.: Rough path analysis via fractional calculus. Manuscript. arXiv:math.PR/0602050 v1 2Feb (2006) [HO02] Hu, Y., Øksendal, B.: Chaos expansion of local time of fractional Brownian motions. Stoch. Anal. Appl., 20, 815–837 (2002) [HO03] Hu, Y., Øksendal, B.: Fractional white noise calculus and applications to finance. Infin. Dimens. Anal. Quantum Prob. Relat. Top., 6, 1–32 (2003) [HOS05] Hu, Y., Øksendal, B., Salopek, D., M.: Weighted local time for fractional Brownian motion and applications to finance. Stochastic Anal. Appl., 23, 15–30 (2005)

378

References

[HOS03] Hu, Y., Øksendal, B., Sulem, A.: Optimal portfolio in a fractional Black and Scholes market. Infin. Dimens. Anal. Quantum Probab. Relat. Top. 6, 519–536 (2003) [HC78] Huang, S.T., Cambanis, S.: Stochastic and multiple Wiener integrals for Gaussian processes. Ann. Prob. 6, 585–614 (1978) [Hur51] Hurst, H.E.: Long-term storage capacity in reservoirs. Trans. Amer Soc. Civil. Eng., 116, 400–410 (1951) [HBS65] Hurst, H.E., Black, R.P., Simaika, Y.M.: Long Term Storage in Reservoirs. An Experimental Study. Constable, London (1965) [HP04a] H¨ usler, J., Piterbarg, V.: Limit theorem for maximum of the storage process with fractional Brownian motion input. Stoch. Proc. Appl., 114, 231–250 (2004) [HP04b] H¨ usler, J., Piterbarg, V.: On the ruin probability for physical fractional Brownian motion. Stoch. Proc. Appl., 113, 315–332 (2004) [IK81] Ibragimov, I.A., Khas’minski, R.Z.: Statistical Estimation: Asymptotic Theory. Springer, Berlin (1981) [IT99] Iglói, E., Terdik, Gy.: Bilinear stochastic systems with fractional Brownian motion input. Ann Appl. Prob., 9, 46–77 (1999) [INA06] Inoue, A., Nakano, Y., Anh, V.: Linear filtering of systems with memory and application to finance. Journal of Appl. Math. and Stoch. Anal., 2006, 1–26 (2006) [IN07] Inoue, A., Nakano, Y.: Optimal long term investment model with memory. Appl. Math. Optimization, 55, 93–122 (2007). [INA07] Inoue, A., Nakano, Y., Anh, V.: Binary market models with memory. Stat. Prob. Lett., 77, 256–264 (2007). [Itô42] Itô, K.: Differential equations determining Markov processes. Zenkoku Shijo Sugaku Danwakai, 244, 1352–1400 (1942) [Itô44] Itô, K.: Stochastic integral. Proc. Imp. Acad. Tokyo, 20, 519– 524 (1944) [Itô51] Itô, K.: Stochastic differential equations. Memoirs of the Amer. Math. Soc. 4 (1951) [Jac79] Jacod, J.: Calcul stochastique et problème de martingales. Lect. Notes. Math., 714 Springer, Berlin (1979) [JS87] Jacod, J., Shiryaev, A.N.: Limit Theorems for Stochastic Processes. Springer, Berlin (1987) [JM83] Jain, N.C., Monrad, D.: Gaussian measures in Bp1 . Ann. Prob., 11, 46–57 (1983) [Jost06] Jost, C.: Transformation formulas for fractional Brownian motion. Stoch. proc. Appl., 116, 1341–1357 (2006) [Jum93] Jumarie, G.: Stochastic differential equations with fractional Brownian motion input. Int. J. Syst. Sci., 24, 1113–1131 (1993) [Kam96] Kamont, A.: On the fractional anisotropic Wiener field. Prob. Math. Stat., 16, 85–98 (1996)

References

379

[KS98] Karatzas, I, Shreve, S.E.: Methods of Mathematical Finance. Springer, New York (1998) [KK97] Kasahara, Y., Kosugi, N.: A limit theorem for occupation times of fractional Brownian motion. Stoch. Proc. Appl., 67, 161–175 (1997) [KM96] Kasahara, Y., Matsumoto, Y.: On Kallianpur-Robbins law for fractional Brownian motion. J. Math. Kyoto Univ., 36, 815–824 (1996) [KO99] Kasahara, Y., Ogawa, N.: A note of the local time of fractional brownian motion. J. Theor. Prob., 12, 207–216 (1999) [KK73] Kawada, T., Kˆ ono, N.: On the variation of Gaussian processes. In: Proc. 2nd Japan–USSR Symp. Prob. Theory, Lect. Notes Math., 330, 176–192 (1973) [KKA98a] Kleptsyna, M.L., Kloeden, P.E., Anh, V.V.: Existence and uniqueness theorems for stochastic differential equations with fractional Brownian motion. Probl. Inf. Transm. 34, 54–56 (1998) [KKA98b] Kleptsyna, M., Kloeden, P., Anh, V.: Nonlinear filtering with fractional Brownian motion. Probl. Inf. Transm., 34, 150–160 (1998) [KLeB99] Kleptsyna, M.L., Le Breton, A.: Extension of the Kalman–Bucy filter to elementary linear systems with fractional Brownian noises. Stat. Infer. Stoch. Proc., 5, 249–271 (2002) [KLeB02] Kleptsyna, M.L., Le Breton, A.: Statistical analysis of the fractional Ornstein–Uhlenbeck type process. Stat. Infer. Stoch. Proc., 5, 229-248 (2002) [KLeBR99] Kleptsyna, M.L., Le Breton, A., Roubaud, M.C.: An elementary approach to filtering in systems with fractional Brownian observation noise. In: Grigelionis B. (ed) Prob. Theory Math. Stat. (Proc. 7th Vilnius Conf.), TEV, Vilnius, 373–392 (1999) [KLeBR00] Kleptsyna, M.L., Le Breton, A., Roubaud, M.C.: General approach to filtering with fractional Brownian noises – application to linear systems. Stoch. Stoch. Rep., 71, 119–140 (2000) [Kli98] Klingenh¨ ofer, F.: Differential equations with fractal noise. PhD Thesis, University of Jena (1998) [KZ99] Klingenh¨ ofer, F., Z¨ ahle, M.: Ordinary differential equations with fractal noise. Proc. Amer. Math. Soc., 127, 1021–1028 (1999) [KP92] Kloeden, P.E., Platen, E.: Numerical Solution of Stochastic Differential Equations. Springer, Berlin (1992) [KK04] Kl¨ uppelberg, C., K¨ uhn, Ch.: Fractional Brownian motion as a weak limit of Poisson shot noise processes – with applications to finance. Stoch. Proc. Appl., 113, 333–351 (2004) [KP94] Kohatsu-Higa, A., Protter, P.: The Euler scheme for SDE’s driven by semimartingales. In: Stochastic Analysis

380

References

[Kol31]

[Kol40]

[Kon37] [Kono96]

[Kop84] [KS03]

[KK99]

[KM06]

[KM07]

[KM00]

[KMV05]

[Ku99] [LN03]

on Infinite-Dimensional Spaces, Pitman Res. Notes in Math. Ser. 310, 141–151 (1994) ¨ Kolmogoroff, A.N.: Uber die analytischen Methodenin der Wahrscheinlichkeitsrechnung. Math. Annalen, 104, 415–458 (1931) Kolmogorov, A.N.: The Wiener spiral and some other interesting curves in Hilbert space. Dokl. Akad. Nauk SSSR, 26, 115–118 (1940) Kondurar, V.: Sur l’intégrale de Stieltjes. Recueil Math., 2, 361– 366 (1937) Kˆ ono, N.: Kallianpur–Robbins law for fractional Brownian motion. Prob. Theory Math. Stat. Proc. 7th Japan-Russia Symp. World Scientific, Singapore, 229–236 (1996) Kopp, P.E.: Martingales and Stochastic Integrals. Cambridge, Cambridge Univ. (1984) Koul, H.I., Surgailis,D. : Robust estimation in regressian models with long-memory errors. In: Doukhan, P., Oppenheim, G., Taqqu, M. (eds) Theory and Applications of Long-Range Dependence. Birkh¨ auser, Boston, 339–355 (2003) Kozachenko, Y.V., Kurchenko, O.O.: An estimate for the multiparameter fractional Brownian motion. Theory Stoch. Proc., 5 (21), 113–119 (1999) Kozachenko, Y.V., Mishura Y.S.: Maximal upper bounds of moments of stochastic integrals and the solutions of stochastic differential equations involving fractional Brownian motion with Hurst index H < 1/2. I (Ukrain.) Prob. Theory Math. Stat., 75, 47–59 (2006) Kozachenko, Y.V., Mishura Y.S.: Maximal upper bounds of moments of stochastic integrals and the solutions of stochastic differential equations involving fractional Brownian motion with Hurst index H < 1/2. II (Ukrain.) Prob. Theory Math. Stat., 76, 53–69 (2007) Krvavych,Y., Mishura,Y.: The stochastic Fubini theorem for integrals containing random integrand and fractional Brownian motion as integrator. Theory Stoch. Proc., 6 (22), 79–89 (2000) Kukush, A., Mishura, Y., Valkeila, E. Statistical inference with fractional Brownian motion. Stat. Infer. Stoch. Proc., 8, 71–93 (2005) Kuznetsov, Yu.A.: The absence of arbitrage in a model with fractal Brownian motion. Russ. Math. Surv., 54, 847–848 (1999) Lanjri, Z.N., Nualart, D.: Smoothness of the law of the supremum of the fractional Brownian motion. Electron. Commun. Prob., 8, 102–111 (2003)

References

381

[Leb95] Lebedev, V.A.: Fubini theorems for depending on parameter stochastic integrals with respect to L0 -valued random measures. Theory Prob. Appl., 40, 313–323 (1995) [LeB98] Le Breton, A.: Filtering and parameter estimation in a simple linear system driven by a fractional Brownian motion. Stat. Prob. Lett., 38, 263–274 (1998) [LT02] León, J.A., Tudor, C.: Semilinear fractional stochastic differential equations. Bol. Soc. Mat. Mexicana, 8, 205–226 (2002) [LL00] León, J.R., Lude˜ na, C.: Estimating the diffusion coefficient for diffusions driven by fBm. Stat. Inference Stoch. Process., 3, 183– 192 (2000). [LP95] Leu, J.-S., Papamarcou, A.: On estimating the spectral exponent of fractional Brownian motion. IEEE Trans. Inf. Theory, 41, 233–244 (1995) [Le48] Lévy, P.: Processus stochastiques et mouvement Brownien. Gauthier–Villars, Paris (1948) Second edition (1965) [Lif95] Lifshitz, M.A.: Gaussian Random Functions (Russian). TViMS, Kiev (1995) [Lin95] Lin, S.J.: Stochastic analysis of fractional Brownian motions. Stoch. Stoch. Rep., 55, 121–140 (1995) [Lind93] Lindstrøm, T.: Fractional Brownian fields as integrals of white noise. Bull. London Math. Soc., 25, 83–88 (1993) [LS86] Liptzer, R.S., Shiryaev, A.N.: Theory of Martingales. (Russian). Nauka, Moscow (1986) [LS01] Liptser, R.S., Shiryaev, A.N.: Statistics of Random Processes. 1: General Theory. Applications of Mathematics. 5. Springer, Berlin (2001) [Lo91] Lo, A.W.: Long-term memory in stock market prices. Econometrica 59, 1279–1313 (1991) [Mac81a] Maccone, C.: Time rescaling and Gaussian properties of the fractional Brownian motions. Nuovo Cimento B (11), 65, 259–271 (1981) [Mac81b] Maccone, C.: Eigenfunction expansion for fractional Brownian motions. Nuovo Cimento B (11), 61, 229–248 (1981) [MS93] Maheswaran, S., Sims, C.A.: Empirical implications of arbitragefree asset markets. In: Philips, P.C.B. (eds) Models, Methods, and Applications of Econometrics. Basil Blackwell (1993) [Mae03] Maejima, M.: Limit theorems for infinite variance sequences. In: Doukhan, P., Oppenheim, G., Taqqu, M. (eds) Theory and Applications of Long-Range Dependence. Birkh¨ auser, Boston, 157–164 (2003) [Mal97] Malliavin, P.: Stochastic Analysis. Springer, Berlin (1997) [Man97] Mandelbrot, B.B.: Fractals and Scaling in Finance. Discontinuity, Concentration, Risk. Springer, New York (1997)

382

References

[MvN68] Mandelbrot, B.B., van Ness J.W.: Fractional Brownian motions, fractional noices and applications. SIAM Review, 10, 422–437 (1968) [MN03] Maslowski, B., Nualart, D.: Evolution equations driven by a fractional Brownian motion. J. Funct. Anal., 202, 277–305 (2003) [Mas93] Masry, E.: The wavelet transform of stochastic processes with stationary increments and its application to fractional Brownian motion. IEEE Trans. Inf. Theory, 39, 260–264 (1993) [Mas96] Masry, E.: Convergence properties of wavelet series expansions of fractional Brownian motion. Appl. Comput. Harmon. Analysis, 3, 239–253 (1996) [MS99] Massoulie, L., Simonian, A.: Large buffer asymptotics for the queue with FBM input. J. Appl. Probab. 36, 894–906 (1999). [MMV01] Memin, J., Mishura, Yu., Valkeila, E.: Inequalities for the moments of Wiener integrals with respect to fractional Brownian motions. Stat. Prob. Lett., 51, 197–206 (2001) [MP80] Métivier, M., Pellaumail, J.: Stochastic integration. Academic Press, New York (1980) [Me76] Meyer, P.A.: Un cours sur les intégrales stochastiques. Séminaire Prob. X, In: Lect. Notes Math., 511, 246–400 (1976) [MN00] Mikosch, T., Norvaiˇsa, R.: Stochastic integral equations without probability. Bernoulli, 6, 401–434 (2000) [Mis03] Mishura, Yu.: Quasilinear stochastic differential equations with fractional-Brownian component. Prob. Theory Math. Stat., 68, 95–106 (2003) [Mis04] Mishura, Yu.: Fractional stochastic integration and BlackScholes equation for fractional Brownian motion with stochastic volatility. Stoch. Stoch. Rep., 76, 363–381 (2004) [Mis05] Mishura, Yu.: An estimate of ruin probabilities for the long range dependence models. Prob. Theory Math. Stat., 72, 93– 100 (2005) [MisIl03] Mishura, Yu., Ilchenko S.: Itˆ o formula for fractional Brownian fields. Prob. Theory Math. Stat., 69, 141–153 (2003) [MisIl04] Mishura, Yu., Ilchenko S.: Generalized two-parameter Lebesgue– Stieltjes integrals and their applications to fractional Brownian fields. Ukr. Math. J.,56, 435–450 (2004) [MisIl06] Mishura, Yu., Ilchenko S.: Stochastic integrals and stochastic differential equations involving fractional Brownian field. Theory Prob. Math. Stat., 75, 85–101 (2006) [MN04] Mishura, Y., Nualart, D.: Weak solutions for stochastic differential equations with additive fractional noise. Stat. Prob. Lett., 70, 253–261 (2004) [MP07] Mishura, Yu., Posashkov S.: Existence and uniquenes of solution of mixed stochastic differential equation driven by fractional

References

[MR01]

[MS07a]

[MS07b]

[MV00] [MV02]

[MV06]

[Mol69]

[Mol03]

[MG69]

[MY67] [MS03]

[Mur92]

[Nar98]

[NN06]

383

Brownian motion and Wiener process. Theory Stoch. Proc., 13(29), 152–165 (2007) Mishura, Yu., Rudomino-Dusyatska, N.: Consistency of drift parameter estimates in fractional-Brownian diffusion models. Theory Stoch. Proc., 7(23), 103–112 (2001) Mishura, Yu., Shevchenko G.M.: The rate of convergence for Euler approximations of solutions of stochastic differential equations driven by fBm. http://arxiv.org/abs/0705.1773 Mishura, Yu., Shevchenko G.M.: On differentiability of solution to stochastic differential equation with fractional Brownian motion. Theory Stoch. Proc., 13(29), 243–250 (2007) Mishura, Yu., Valkeila, E.: An isometric approach to generalized stochastic integrals. J. Theor. Prob., 13, 673–693 (2000) Mishura, Yu. S., Valkeila E.: The absence of arbitrage in a mixed Brownian-fractional Brownian model. Proc. of the Steklov Institute of Math., 237, 215–224 (2002) Mishura, Yu. S., Valkeila E.: An extension of the Lévy characterization to fractional Brownian motion. Preprint, Helsinki Univ. Techn. A514 (2006) Molchan, G.: Gaussian processes with spectra which are asymptotically equivalent to a power of λ. Theory Prob. Appl. 14, 530–532 (1969) Molchan, G.: Historical comments related to fractional Brownian motion. In: Doukhan, P., Oppenheim, G., Taqqu, M. (eds) Theory and Applications of Long-Range Dependence. Birkh¨ auser, Boston, 39–42 (2003) Molchan, G., Golosov, J.: Gaussian stationary processes with asymptotic power spectrum. Soviet Math. Dokl. 10, 134–137 (1969) Monin, A., Yaglom, A.M.: Statistical Hydromechanics. Mechanics of Turbulence. I-II. (Russian) Nauka, Moscow (1967) Moulines, E., Soulier, P.: Semi-parametric spectral estimation for fractional processes. In: Doukhan, P., Oppenheim, G., Taqqu, M. (eds) Theory and Applications of Long-Range Dependence. Birkh¨ auser, Boston, 251–301 (2003) Muraoka, H.: A fractional Brownian motion and the canonical representations of its coefficient processes. Soochow J. Math., 18, 361–377 (1992) Narayan, O.: Exact asymptotic queue length distribution for fractional Brownian traffic. Adv. Perform. Analysis, 1, 39–63 (1998) Neunkirch, A., Nourdin, I.: Exact rate of convergence of some approximation schemes associated to SDEs driven by a fractional Brownian motion. Preprint, available online at http://arxiv.org/abs/math.PR/0601038 (2006)

384

References

[Nie04] Nieminen, A.: Fractional Brownian motion and martingaledifferences. Stat. Prob. Lett., 70, 1–10 (2004) [Nor95] Norros, I.: On the use of the fractional Brownian motion in the theory of connectionless networks. IEEE J. Sel. Areas Commun., 13, 953–962 (1995) [Nor97] Norros, I.: Four approaches to the fractional Brownian storage. In: Lévy Véhel, J., Lutton, E., Tricot, C. (eds) Fractals in Engineering. Springer, 154–169 (1997) [Nor99] Norros, I.: Busy periods of fractional Brownian storage: a large deviations approach. Adv. Perform. Analysis, 2, 1–19 (1999) [NVV99] Norros, I., Valkeila, E., Virtamo, J.: An elementary approach to a Girsanov formula and other analytical results on fractional Brownian motions. Bernoulli 5(4), 571–587 (1999) [Nrv99] Norvaiˇsa, R.: p-variation and integration of sample functions of stochastic proceses. In: Grigelionis B. (ed) Prob. Theory Math. Stat. (Proc. 7th Vilnius Conf.), TEV, Vilnius, 521–540 (1999) [Nou05] Nourdin, I.: Schémas d’approximation associés à une équation différentielle dirigée par une fonction h¨ oldérienne; cas du mouvement brownien fractionnaire. C. R. Acad. Sci. Paris Sér. I Math., 340, 611–614 (2005) [NV98] Novikov, A., Valkeila, E.: On some maximal inequalities for fractional Brownian motions. Preprint 186, University of Helsinki (1998) [Nua95] Nualart, D.: The Malliavin Calculus and Related Topics. Springer, Berlin (1995) [Nua98] Nualart, D.: Analysis on Wiener space and anticipating stochastic calculus. In: Lect. Notes Math., 1690, 123–227 (1998) [Nua03] Nualart, D.: Stochastic integration with respect to fractional Brownian motion and applications. In: Gonz´ alez-Barrios, J.M. et al. (eds), Stochastic Models. Contemp. Math., 336, AMS, Providence, RI, 3–39, (2003) [Nua06] Nualart, D.: Fractional Brownian motion: stochastic calculus and applications. In: Proceed. Intern. Congress Mathem. Madrid, Spain. European Mathem. Society (2006) [NO02] Nualart, D., Ouknine, Y.: Regularization of differential equations by fractional noise. Stoch. Proc. Appl., 102, 103–116 (2002) [NO03a] Nualart, D., Ouknine, Y.: Besov regularity of stochastic integrals with respect to the fractional Brownian motion with parameter H > 1/2. J. Theor. Prob., 16, 451–470 (2003) [NO03b] Nualart, D., Ouknine, Y.: Stochastic differential equations with additive fractional noise and locally unbounded drift. Progress in Probability, 56, 353–365 (2003) [NP95] Nualart, D., Pardoux, E.: Stochastic calculus with anticipating integrands. Prob. Theory Rel. Fields, 78, 535–581 (1995)

References

385

[NR00] Nualart, D. R˘ a¸scanu, A.: Differential equations driven by fractional Brownian motion. Collect. Math., 53, 55–81 (2000) [NS05] Nualart, D., Saussereau, B.: Malliavin calculus for stochastic differential equations driven by a fractional Brownian motion. Preprint No. 371, Univ. de Barcelona (2005) [Oks03] Øksendal, B.: Stochastic Differential Equations. An Introduction with Applications. Springer, Berlin (2003) [Oks07] Øksendal, B.: Fractional Brownian motion in finance. In: Jensen, B.S., Palokangas, T. (eds) Stochastic economic dynamics. CBS Press, Copenhagen (2007) [OZ01] Øksendal, B., Zhang, T.: Multiparameter fractional Brownian motion and quasi-linear stochastic partial differential equations, Stoch. Stoch. Rep. 71, 141–163 (2001) [PT02a] Pérez-Abreu, V., Tudor, C.: Multiple stochastic fractional integrals: a transfer principle for multiple stochastic fractional integrals. Bol. Soc. Mat. Mexicana, 3, 187–203 (2002) [PT00a] Pipiras, V., Taqqu, M.: Convergence of the WeierstrassMandelbrot process to fractional Brownian motion. Fractals, 8, 360–384 (2000) [PT00b] Pipiras, V., Taqqu, M.: Integration questions related to fractional Brownian motion. Prob. Theory Rel. Fields, 118, 251–291 (2000) [PT01] Pipiras, V., Taqqu, M.S.: Are classes of deterministic integrands for fractional Brownian motion on an interval complete? Bernoulli, 7, 873–897 (2001) [PT02b] Pipiras, V., Taqqu, M.S.: Deconvolution of fractional Brownian motion. J. of Time Series Analysis, 23, 487–501 (2002) [PT03] Pipiras, V., Taqqu M.S.: Fractional calculus and its connection to fractional Brownian motion. In: Doukhan, P., Oppenheim, G., Taqqu, M. (eds) Theory and Applications of Long-Range Dependence. Birkh¨ auser, Boston, 165–202 (2003) [Po99] Podlubny, I.: Fractional Differential Equations. An Introduction to Fractional Derivatives, Fractional Differential Equations, Some Methods of Their Solution and Some of Their Applications. Academic Press, San Diego (1999) [Pos05] Posashkov, S. Optimal filtering in systems with fractionalBrownian noises. Theory Prob. Math. Stat, 72, 120–128 (2005) [Pri98] Privault, N.: Skorokhod stochastic integration with respect to a non-adapted process on Wiener space. Stoch. Stoch. Rep., 65, 13–39 (1998) [Pro90] Protter, P.: Stochastic Integration and Differential Equations. A New Approach. Applications of Mathematics, 21. Springer, Berlin (1990)

386

References

[RZ91] Ramanathan, J., Zeitouni, O.: On the wavelet transform of fractional Brownian motion. IEEE Trans. Inform. Theory, 37, 1156– 1158 (1991) [RLT95] Reed, I.S., Lee, P.C., Truong, T.K.: Spectral representation of fractional Brownian motion in n dimensions and its properties. IEEE Trans. Inf. Theory, 41, 1439–1451 (1995) [Rog97] Rogers, L.C.G.: Arbitrage with fractional Brownian motion, Mathem. Finance, 7, 95–105 (1997) [Ros87] Rosen, J.: The intersection local time of fractional Brownian motion in the plane. J. Multivariate Analysis, 23, 37–46 (1987) [RV93] Russo, F., Vallois, P.: Forward, backward and symmetric stochastic integration. Prob. Theory Rel. Fields, 97, 403–421 (1993) [RV95a] Russo, F., Vallois, P.: Itˆ o formula for C 1 -functions of semimartingales. Prob. Theory Rel. Fields, 104, 27–41 (1995) [RV95b] Russo, F., Vallois, P.: The generalized covariation process and Itˆ o formula. Stoch. Proc. Appl., 59, 81–104 (1995) [RV98] Russo, F., Vallois, P.: Forward and backward calculus with respect to finite quadratic variation processes. Preprint, Institut Galilée, Mathématiques, Université Paris (1998) [RV00] Russo, F., Vallois, P.: Stochastic calculus with respect to continuous finite quadratic variation processes. Stoch. Stoch. Rep., 70, 1–40 (2000) [Ruz00] Ruzmaikina, A.A.: Stieltjes integrals of H¨ older continuous functions with applications to fractional Brownian motion. J. Stat. Phys., 100, 1049–1069 (2000) [Sai92] Sainty, Ph.: Construction of a complex-valued fractional Brownian motion of order N . J. Math. Phys., 33, 3128–3149 (1992) [Sal98] Salopek, D.M.: Tolerance to arbitrage. Stock. Proc. Appl. 76, 217–230 (1998) [SKM93] Samko, S.G., Kilbas, A.A., Marichev, O.I.: Fractional Integrals and Derivatives. Theory and Applications. Gordon and Breach Science Publishers, New York (1993) [ST94] Samorodnitsky, G., Taqqu, M.S.: Stable Non-Gaussian Random Processes. Chapman & Hall, New York (1994) [Sch99] Schoenmakers, J.G.M., Kloeden, P.: Robust option replication for a Black–Scholes model extended with nondeterministic trends. J. Appl. Math. Stoch. Anal.,12, 113–120 (1999) [Seb95] Sebastian, K. L.: Path integral representation for fractional Brownian motion. J. Phys. A, Math. Gen., 28, 4305–4311 (1995) [Sh96] Shieh, N.-R.: Limit theorems for local times of fractional Brownian motions and some other self-similar processes. J. Math. Kyoto Univ., 36, 641–652 (1996) [Shi99] Shiryaev, A.N.: Essentials of Stochastic Finance: Facts, Models and Theory. World Scientific, Singapore (1999)

References

387

[Shi01] Shiryaev A.N.: Arbitrage and replication for fractal models. Preprint, MaPhySto, Aarhus (2001) [Sin97] Sinaj, Ya.G.: Distribution of the maximum of a fractional Brownian motion. Russ. Math. Surv., 52, 359–378 (1997) [Sin94] Singer, P.: An integrated fractional Fourier transform. J. Comp. Appl. Math., 54, 221–237 (1994) [SL95] Sithi, V.M., Lim, S.C.: On the spectra of Riemann-Liouville fractional Brownian motion. J. Phys. A, Math. Gen., 28, 2995–3003 (1995) [Sko75] Skorokhod, A.V.: On a generalization of a stochastic integral. Theory Prob. Appl., 20, 219–233 (1975) [SS70] Skorohod, A.V., Slobodenyuk, N.P.: Limit Theorems for Random Walks (Russian). Naukova Dumka, Kiev (1970) [Sot01] Sottinen, T.: Fractional Brownian motion, random walks and binary market models. Finance Stoch., 5, 343–355 (2001) [SV03] Sottinen, T., Valkeila E.: On arbitrage and replication in the fractional Black–Scholes pricing model. Stat. Decisions, 21, 93– 107 (2003) [Sur03a] Surgailis, D.: CLTs for polynomials of linear sequences: diagram formula with illustrations. In: Doukhan, P., Oppenheim, G., Taqqu, M. (eds) Theory and Applications of Long-Range Dependence. Birkh¨ auser, Boston, 111–128 (2003) [Sur03b] Surgailis, D.: Non-CLTs: U-statistics, multinomial formula and approximations of multiple Wiener–Itˆ o integrals. In: Doukhan, P., Oppenheim, G., Taqqu, M. (eds) Theory and Applications of Long-Range Dependence. Birkhäuser, Boston, 129–142 (2003) [SW03] Surgailis, D., Woyczynski, W.A.: Limit theorems for the Burgers equation initialized by data with long-range dependence. In: Doukhan, P., Oppenheim, G., Taqqu, M. (eds) Theory and Applications of Long-Range Dependence. Birkh¨ auser, Boston, 507– 523 (2003) [Tal95] Talagrand, M.: Hausdorff measure of trajectories of multiparameter fractional Brownian motion. Ann. Prob., 23, 767–775 (1995) [Tal96] Talagrand, M.: Lower classes for fractional Brownian motion. J. Theor. Prob., 9, 191–213 (1996) [TX96] Talagrand, M., Xiao, Y.: Fractional Brownian motion and packing dimension. J. Theor. Prob., 9, 579–593 (1996) [Taq75] Taqqu, M.: Weak convergence to fractional Brownian motion and to the Rosenblatt process. Z. Wahr. verw. Gebiete, 31, 287– 302 (1975) [Taq77] Taqqu, M.: Law of the iterated logarithm for sums of non linear functions of Gaussian variables that exhibit long-range dependence. Z. Wahr. verw. Gebiete, 40, 203–238 (1977)

388

References

[Taq03] Taqqu, M.: Fractional Brownian motion and long-range dependence. In: Doukhan, P., Oppenheim, G., Taqqu, M. (eds) Theory and Applications of Long-Range Dependence. Birkh¨ auser, Boston, 5–38 (2003) [TTW95] Taqqu, M.S., Teverovsky, V., Willinger, W.: Estimators for long-range dependence: an empirical study. Fractals, 3, 785–798 (1995) [TT97] Taqqu, M.S., Teverovsky, V.: On estimating long-range dependence in finite and infinite variance series. In: Adler, R., Feldman, R., Taqqu, M.S. (eds) A Practical Guide to Heavy Tails: Statistical Techniques and Applications. Birkh¨ auser, Boston, 177–217 (1997) [Thao03] Thao, T.H.: A note on fractional Brownian motion. Vietnam J. Math., 31, 255–260 (2003). [TTV03] Tindel, S., Tudor, C., Viens, F.: Stochastic evolution equations with fractional Brownian motion. Prob. Theory Rel. Fields, 127, 186–204 (2003) [Tud03] Tudor, C.A.: Weak convergence to the fractional Brownian sheet in Besov spaces. Bull. Braz. Math. Soc. (N.S.), 34, 389–400 (2003) [TT03] Tudor, C., Tudor, M.: On the two-parameter fractional Brownian motion and Stieltjes integrals for H¨ older functions. J. Math. Anal. Appl., 286, 765–781 (2003) [TV03] Tudor, C.A., Viens, F.G.: Itˆ o formula and local time for the fractional Brownian sheet. Electronic J. Prob., 8, 1–31 (2003) [Vage96] V˚ age, G.: A general existence and uniqueness theorem for WickSDEs in (Y )n−1,k . Stoch. Stoch. Reports, 58, 259–284 (1996) [Wang03] Wang, W.: Weak convergence to fractional Brownian motion in Brownian scenery. Prob. Theory Rel. Fields, 126, 203–220 (2003) [WRL03] Wang, X.-T., Ren, F.-Y., Liang, X.-Q.: A fractional version of the Merton model. Chaos, Solut. Fractals, 15, 455–463 (2003) [Wat95] Watson, G.N.: A Treatise on the Theory of Bessel Functions. Cambridge Mathematical Library edition. Cambridge University Press, Cambridge (1995) [Wie23] Wiener N.: Differential space. J. Math. Phys. 2, 131–134 (1923) [WTT99] Willinger, W., Taqqu, M., Teverovsky, V.: Stock market process and long-range dependence. Finance Stoch., 3, 1–13 (1999) [Wyss00] Wyss, W.: The fractional Black–Scholes equation. Fraction. Calc. Appl. Anal., 3, 51–61 (2000) [Xiao91] Xiao, Y.: Multiple points of fractional Brownian motion and Hausdorff dimension. (Chinese) Chin. Ann. Math., Ser. A, 12, 612–618 (1991)

References

389

[Xiao96] Xiao, Y.: Packing measure of the sample paths of fractional Brownian motion. Trans. Amer. Math. Soc., 348, 3193–3213 (1996) [Xiao97a] Xiao, Y.: Packing dimension of the image of fractional Brownian motion. Stat. Prob. Lett., 33, 379–387 (1997) [Xiao97b] Xiao, Y.: Hausdorff measure of the graph of fractional Brownian motion. Math. Proc. Camb. Philos. Soc., 122, 565–576 (1997) [Yag57] Yaglom, A.M.: Certain types of random fields in n-dimentional space similar to stationary stochastic processes. Theory Prob. Appl., 2, 292–338 (1957) [Yag87] Yaglom, A.M.: Correlation Theory of Stationary and Related Random Functions I. Springer, New York (1987) [Yin96] Yin, Z.-M.: New methods for simulation of fractional Brownian motion. J. Comp. Phys., 127, 66–72 (1996) [Yor76] Yor, M.: Sur les intégrales stochastiques optionelles et une suite remarquable de formules exponentielles. (French) Séminaire Prob. X, In: Lect. Notes Math. 511, Springer, Berlin, 481–500 (1976) [Yor88] Yor, M.: Remarques sur certaines constructions des mouvements browniens frationnaires. (Remark on certain constructions of fractional Brownian motions.) (French) Séminaire Prob. XXII, In: Lect. Notes Math., 1321, Springer, Berlin, 217–224 (1988) [Yos78] Yoshihara, K.-I.: A weak convergence theorem for functionals of sums of martingale differences. Yokohama Math. J., 26, 101–107 (1978) [Zah98] Z¨ ahle, M.: Integration with respect to fractal functions and stochastic calculus. I. Prob. Theory Rel. Fields, 111, 333–374 (1998) [Zah99] Z¨ ahle, M.: On the link between fractional and stochastic calculcus. In: Grauel, H., Gundlach, M. (eds) Stochastic Dynamics, Springer, 305–325 (1999) [Zah01] Z¨ ahle, M.: Integraton with respect to fractal functions and stochastic calculus. II. Math. Nachr., 225, 145–183 (2001) [Zah02a] Z¨ ahle, M.: Long range dependence, no arbitrage and the BlackScholes formula. Stochastics and Dynamics, 2, 265–280 (2002) [Zah02b] Z¨ ahle, M.: Forward integrals and stochastic differential equations. In: Dalang, R.C., Dozzi, M. (eds) Sem. Stoch. Anal, Random Fields Appl. III, Progress in Prob., 52, Birkh¨ auser, 293–302 (2002) [Zah05] Z¨ ahle, M.: Stochastic differential equations with fractal noise. Math. Nachr., 278, 1097–1106 (2005) [Zha96] Zhang, L.: Some liminf results on increments of fractional Brownian motion. Acta Math. Hung., 71, 215–240 (1996) [Zha97] Zhang, L.: A note on liminfs for increments of a fractional Brownian motion. Acta Math. Hung., 76, 145–154 (1997)

Index

(g)-transformable, 61 C-function, 287 ∆2 -condition, 287 b-self-similar, 7 arbitrage, 302 arbitrage opportunity, 306 asymptotic efficiency, 359 backward integral, 161 Besov space, 73, 128 Bessel function of the first kind, 64 Black–Scholes equation, 322 fractional, 326 bracket, 205 Burkholder–Davis–Gundy inequalities, 47 capital, 305 chain rule for stochastic derivative, 145 complex alternative, 328 complex hypothesis, 328 composition formulas for fractional integrals, 2 conditionally Gaussian pair, 296 confidence level, 330 convolution, 59 critical areas, 330 critical values, 330 density process, 66 derivative operator, 160 directional derivative, 145 discounted gain, 305

divergence operator, 161 Dudley integral, 42 entropy maximal estimates, 41 errors of the first and of the second kind, 330 field with independent increments, 119 fractional analog of the Burgers equation, 321 fractional Brownian motion, 7 approximation, 71 backward, 11 forward, 11 geometric, 302 Mandelbrot–van Ness representation, 9, 23 multi-parameter, 117 fractional derivative Marchaud, 3 two-parameter, 119 Riemann–Liouville, 2 two-parameter, 119 Weyl representation, 3 fractional Doleans exponent, 191 fractional noise, 15 fractional Wick exponent, 256 fundamental martingale, 27 Gaussian subspaces, 59 generalized Lebesgue–Stieltjes integral, 123 generalized quadratic variation, 205 Girsanov theorem, 67

392

Index

goodness-of-fit test, 335 H¨ older continuous functions, 2 Hardy–Littlewood theorem, 1 Hellinger process, 70 Hermite functions, 12 Hermite polynomials, 12 Hurst effect, 301 Hurst law, 301 Hurst phenomenon, 301 integration-by-parts formula for fractional derivatives, 4 for fractional integrals, 2 Itˆ o formula, 182 for H ∈ (0, 1/2), 185 for fractional fields, 186 for Wick integrands, 184 Lévy theorem, 27, 94 Lenglart inequality, 48 likelihood ratio, 329 local asymptotic normality, 358 local martingale, 66 local solution, 205 long memory, 59 long-memory Gaussian processes, 59 long-range dependence, 8 Markov-type strategy, 306 martingale, 27 maximum likelihood estimate, 345 metric ε-capacity, 41 metric ε-entropy, 42 Mittag–Leffler function, 351 mixed version of the Black–Merton– Scholes model, 305 modulus of continuity, 87 modulus of uniform continuity, 87 Molchan martingale, 27 observation process, 291 observed trajectory, 328 optimal filter, 291 optimal filtering problem, 292 Orlicz space, 280 portfolio, 305 quasi-likelihood ratio, 331

random walk, 80 rate of convergence of Euler approximations, 243 regularly varying, 91 rescaled adjusted range statistic or R/S-statistic, 301 Riemann–Liouville fractional integral on (a, b), 1 on R, 1 two-parameter, 118 self-financing strategy, 305, 306 semi-metric, 41 semi-norm, 163 semimartingale, 65, 71 signal process, 291 Skorohod integral, 158, 161 Skorohod space, 80 Skorohod topology, 81 slowly varying at ∞, 80 smooth functionals, 146 Sobolev space, 38 Sobolev–Slobodeckij space, 205 spectral density function, 8 spectral representation, 8 stochastic derivative, 145 stochastic differential equation Euler approximations, 243 mixed, 225 moment estimates for solution, 56 pathwise, 197 quasilinear of Skorohod type, 255 weak solution, 263 with additive Wiener integral, 262 with fractional white noise, 241 stochastic Fubini theorem, 57, 174 stochastic integral generalized of first order, 166 generalized forward, 205 generalized of kth order, 168 Skorohod, 158, 161 Stratonovich, 146 symmetric Stratonovich, 161 w.r.t. fBm forward, 161 Wick, 144 Wiener, 16 stopping time, 46

Index Stratonovich integral, 146 symmetric, 161 strong martingale, 120 strong Molchan martingale, 121 testing procedure, 330 The Fourier transform, 5 two-parameter left Riemann–Stieltjes integral, 135 two-sided Wiener process, 9

uniform modulus, 87 weak solution, 263 Weierstrass–Mandelbrot process, 79 white noise, 10 Wick products, 141 Wiener field, 120 Wiener integral generalized, 165 w.r.t. fBm, 16 moment inequalities, 35

393

Lecture Notes in Mathematics For information about earlier volumes please contact your bookseller or Springer LNM Online archive: springerlink.com

Vol. 1735: D. Yafaev, Scattering Theory: Some Old and New Problems (2000) Vol. 1736: B. O. Turesson, Nonlinear Potential Theory and Weighted Sobolev Spaces (2000) Vol. 1737: S. Wakabayashi, Classical Microlocal Analysis in the Space of Hyperfunctions (2000) Vol. 1738: M. Émery, A. Nemirovski, D. Voiculescu, Lectures on Probability Theory and Statistics (2000) Vol. 1739: R. Burkard, P. Deuflhard, A. Jameson, J.-L. Lions, G. Strang, Computational Mathematics Driven by Industrial Problems. Martina Franca, 1999. Editors: V. Capasso, H. Engl, J. Periaux (2000) Vol. 1740: B. Kawohl, O. Pironneau, L. Tartar, J.-P. Zolesio, Optimal Shape Design. Tróia, Portugal 1999. Editors: A. Cellina, A. Ornelas (2000) Vol. 1741: E. Lombardi, Oscillatory Integrals and Phenomena Beyond all Algebraic Orders (2000) Vol. 1742: A. Unterberger, Quantization and Nonholomorphic Modular Forms (2000) Vol. 1743: L. Habermann, Riemannian Metrics of Constant Mass and Moduli Spaces of Conformal Structures (2000) Vol. 1744: M. Kunze, Non-Smooth Dynamical Systems (2000) Vol. 1745: V. D. Milman, G. Schechtman (Eds.), Geometric Aspects of Functional Analysis. Israel Seminar 19992000 (2000) Vol. 1746: A. Degtyarev, I. Itenberg, V. Kharlamov, Real Enriques Surfaces (2000) Vol. 1747: L. W. Christensen, Gorenstein Dimensions (2000) Vol. 1748: M. Ruzicka, Electrorheological Fluids: Modeling and Mathematical Theory (2001) Vol. 1749: M. Fuchs, G. Seregin, Variational Methods for Problems from Plasticity Theory and for Generalized Newtonian Fluids (2001) Vol. 1750: B. Conrad, Grothendieck Duality and Base Change (2001) Vol. 1751: N. J. Cutland, Loeb Measures in Practice: Recent Advances (2001) Vol. 1752: Y. V. Nesterenko, P. Philippon, Introduction to Algebraic Independence Theory (2001) Vol. 1753: A. I. Bobenko, U. Eitner, Painlevé Equations in the Differential Geometry of Surfaces (2001) Vol. 1754: W. Bertram, The Geometry of Jordan and Lie Structures (2001) Vol. 1755: J. Azéma, M. Émery, M. Ledoux, M. Yor (Eds.), Séminaire de Probabilités XXXV (2001) Vol. 1756: P. E. Zhidkov, Korteweg de Vries and Nonlinear Schrödinger Equations: Qualitative Theory (2001) Vol. 1757: R. R. Phelps, Lectures on Choquet’s Theorem (2001) Vol. 1758: N. Monod, Continuous Bounded Cohomology of Locally Compact Groups (2001)

Vol. 1759: Y. Abe, K. Kopfermann, Toroidal Groups (2001) Vol. 1760: D. Filipović, Consistency Problems for HeathJarrow-Morton Interest Rate Models (2001) Vol. 1761: C. Adelmann, The Decomposition of Primes in Torsion Point Fields (2001) Vol. 1762: S. Cerrai, Second Order PDE’s in Finite and Infinite Dimension (2001) Vol. 1763: J.-L. Loday, A. Frabetti, F. Chapoton, F. Goichot, Dialgebras and Related Operads (2001) Vol. 1764: A. Cannas da Silva, Lectures on Symplectic Geometry (2001) Vol. 1765: T. Kerler, V. V. Lyubashenko, Non-Semisimple Topological Quantum Field Theories for 3-Manifolds with Corners (2001) Vol. 1766: H. Hennion, L. Hervé, Limit Theorems for Markov Chains and Stochastic Properties of Dynamical Systems by Quasi-Compactness (2001) Vol. 1767: J. Xiao, Holomorphic Q Classes (2001) Vol. 1768: M. J. Pflaum, Analytic and Geometric Study of Stratified Spaces (2001) Vol. 1769: M. Alberich-Carramiñana, Geometry of the Plane Cremona Maps (2002) Vol. 1770: H. Gluesing-Luerssen, Linear DelayDifferential Systems with Commensurate Delays: An Algebraic Approach (2002) Vol. 1771: M. Émery, M. Yor (Eds.), Séminaire de Probabilités 1967-1980. A Selection in Martingale Theory (2002) Vol. 1772: F. Burstall, D. Ferus, K. Leschke, F. Pedit, U. Pinkall, Conformal Geometry of Surfaces in S4 (2002) Vol. 1773: Z. Arad, M. Muzychuk, Standard Integral Table Algebras Generated by a Non-real Element of Small Degree (2002) Vol. 1774: V. Runde, Lectures on Amenability (2002) Vol. 1775: W. H. Meeks, A. Ros, H. Rosenberg, The Global Theory of Minimal Surfaces in Flat Spaces. Martina Franca 1999. Editor: G. P. Pirola (2002) Vol. 1776: K. Behrend, C. Gomez, V. Tarasov, G. Tian, Quantum Comohology. Cetraro 1997. Editors: P. de Bartolomeis, B. Dubrovin, C. Reina (2002) Vol. 1777: E. García-Río, D. N. Kupeli, R. VázquezLorenzo, Osserman Manifolds in Semi-Riemannian Geometry (2002) Vol. 1778: H. Kiechle, Theory of K-Loops (2002) Vol. 1779: I. Chueshov, Monotone Random Systems (2002) Vol. 1780: J. H. Bruinier, Borcherds Products on O(2,1) and Chern Classes of Heegner Divisors (2002) Vol. 1781: E. Bolthausen, E. Perkins, A. van der Vaart, Lectures on Probability Theory and Statistics. Ecole d’ Eté de Probabilités de Saint-Flour XXIX-1999. Editor: P. Bernard (2002) Vol. 1782: C.-H. Chu, A. T.-M. Lau, Harmonic Functions on Groups and Fourier Algebras (2002)

Vol. 1783: L. Grüne, Asymptotic Behavior of Dynamical and Control Systems under Perturbation and Discretization (2002) Vol. 1784: L. H. Eliasson, S. B. Kuksin, S. Marmi, J.-C. Yoccoz, Dynamical Systems and Small Divisors. Cetraro, Italy 1998. Editors: S. Marmi, J.-C. Yoccoz (2002) Vol. 1785: J. Arias de Reyna, Pointwise Convergence of Fourier Series (2002) Vol. 1786: S. D. Cutkosky, Monomialization of Morphisms from 3-Folds to Surfaces (2002) Vol. 1787: S. Caenepeel, G. Militaru, S. Zhu, Frobenius and Separable Functors for Generalized Module Categories and Nonlinear Equations (2002) Vol. 1788: A. Vasil’ev, Moduli of Families of Curves for Conformal and Quasiconformal Mappings (2002) Vol. 1789: Y. Sommerhäuser, Yetter-Drinfel’d Hopf algebras over groups of prime order (2002) Vol. 1790: X. Zhan, Matrix Inequalities (2002) Vol. 1791: M. Knebusch, D. Zhang, Manis Valuations and Prüfer Extensions I: A new Chapter in Commutative Algebra (2002) Vol. 1792: D. D. Ang, R. Gorenflo, V. K. Le, D. D. Trong, Moment Theory and Some Inverse Problems in Potential Theory and Heat Conduction (2002) Vol. 1793: J. Cortés Monforte, Geometric, Control and Numerical Aspects of Nonholonomic Systems (2002) Vol. 1794: N. Pytheas Fogg, Substitution in Dynamics, Arithmetics and Combinatorics. Editors: V. Berthé, S. Ferenczi, C. Mauduit, A. Siegel (2002) Vol. 1795: H. Li, Filtered-Graded Transfer in Using Noncommutative Gröbner Bases (2002) Vol. 1796: J.M. Melenk, hp-Finite Element Methods for Singular Perturbations (2002) Vol. 1797: B. Schmidt, Characters and Cyclotomic Fields in Finite Geometry (2002) Vol. 1798: W.M. Oliva, Geometric Mechanics (2002) Vol. 1799: H. Pajot, Analytic Capacity, Rectifiability, Menger Curvature and the Cauchy Integral (2002) Vol. 1800: O. Gabber, L. Ramero, Almost Ring Theory (2003) Vol. 1801: J. Azéma, M. Émery, M. Ledoux, M. Yor (Eds.), Séminaire de Probabilités XXXVI (2003) Vol. 1802: V. Capasso, E. Merzbach, B. G. Ivanoff, M. Dozzi, R. Dalang, T. Mountford, Topics in Spatial Stochastic Processes. Martina Franca, Italy 2001. Editor: E. Merzbach (2003) Vol. 1803: G. Dolzmann, Variational Methods for Crystalline Microstructure – Analysis and Computation (2003) Vol. 1804: I. Cherednik, Ya. Markov, R. Howe, G. Lusztig, Iwahori-Hecke Algebras and their Representation Theory. Martina Franca, Italy 1999. Editors: V. Baldoni, D. Barbasch (2003) Vol. 1805: F. Cao, Geometric Curve Evolution and Image Processing (2003) Vol. 1806: H. Broer, I. Hoveijn. G. Lunther, G. Vegter, Bifurcations in Hamiltonian Systems. Computing Singularities by Gröbner Bases (2003) Vol. 1807: V. D. Milman, G. Schechtman (Eds.), Geometric Aspects of Functional Analysis. Israel Seminar 20002002 (2003) Vol. 1808: W. Schindler, Measures with Symmetry Properties (2003) Vol. 1809: O. Steinbach, Stability Estimates for Hybrid Coupled Domain Decomposition Methods (2003) Vol. 1810: J. Wengenroth, Derived Functors in Functional Analysis (2003)

Vol. 1811: J. Stevens, Deformations of Singularities (2003) Vol. 1812: L. Ambrosio, K. Deckelnick, G. Dziuk, M. Mimura, V. A. Solonnikov, H. M. Soner, Mathematical Aspects of Evolving Interfaces. Madeira, Funchal, Portugal 2000. Editors: P. Colli, J. F. Rodrigues (2003) Vol. 1813: L. Ambrosio, L. A. Caffarelli, Y. Brenier, G. Buttazzo, C. Villani, Optimal Transportation and its Applications. Martina Franca, Italy 2001. Editors: L. A. Caffarelli, S. Salsa (2003) Vol. 1814: P. Bank, F. Baudoin, H. Föllmer, L.C.G. Rogers, M. Soner, N. Touzi, Paris-Princeton Lectures on Mathematical Finance 2002 (2003) Vol. 1815: A. M. Vershik (Ed.), Asymptotic Combinatorics with Applications to Mathematical Physics. St. Petersburg, Russia 2001 (2003) Vol. 1816: S. Albeverio, W. Schachermayer, M. Talagrand, Lectures on Probability Theory and Statistics. Ecole d’Eté de Probabilités de Saint-Flour XXX-2000. Editor: P. Bernard (2003) Vol. 1817: E. Koelink, W. Van Assche (Eds.), Orthogonal Polynomials and Special Functions. Leuven 2002 (2003) Vol. 1818: M. Bildhauer, Convex Variational Problems with Linear, nearly Linear and/or Anisotropic Growth Conditions (2003) Vol. 1819: D. Masser, Yu. V. Nesterenko, H. P. Schlickewei, W. M. Schmidt, M. Waldschmidt, Diophantine Approximation. Cetraro, Italy 2000. Editors: F. Amoroso, U. Zannier (2003) Vol. 1820: F. Hiai, H. Kosaki, Means of Hilbert Space Operators (2003) Vol. 1821: S. Teufel, Adiabatic Perturbation Theory in Quantum Dynamics (2003) Vol. 1822: S.-N. Chow, R. Conti, R. Johnson, J. MalletParet, R. Nussbaum, Dynamical Systems. Cetraro, Italy 2000. Editors: J. W. Macki, P. Zecca (2003) Vol. 1823: A. M. Anile, W. Allegretto, C. Ringhofer, Mathematical Problems in Semiconductor Physics. Cetraro, Italy 1998. Editor: A. M. Anile (2003) Vol. 1824: J. A. Navarro González, J. B. Sancho de Salas, C ∞ – Differentiable Spaces (2003) Vol. 1825: J. H. Bramble, A. Cohen, W. Dahmen, Multiscale Problems and Methods in Numerical Simulations, Martina Franca, Italy 2001. Editor: C. Canuto (2003) Vol. 1826: K. Dohmen, Improved Bonferroni Inequalities via Abstract Tubes. Inequalities and Identities of Inclusion-Exclusion Type. VIII, 113 p, 2003. Vol. 1827: K. M. Pilgrim, Combinations of Complex Dynamical Systems. IX, 118 p, 2003. Vol. 1828: D. J. Green, Gröbner Bases and the Computation of Group Cohomology. XII, 138 p, 2003. Vol. 1829: E. Altman, B. Gaujal, A. Hordijk, DiscreteEvent Control of Stochastic Networks: Multimodularity and Regularity. XIV, 313 p, 2003. Vol. 1830: M. I. Gil’, Operator Functions and Localization of Spectra. XIV, 256 p, 2003. Vol. 1831: A. Connes, J. Cuntz, E. Guentner, N. Higson, J. E. Kaminker, Noncommutative Geometry, Martina Franca, Italy 2002. Editors: S. Doplicher, L. Longo (2004) Vol. 1832: J. Azéma, M. Émery, M. Ledoux, M. Yor (Eds.), Séminaire de Probabilités XXXVII (2003) Vol. 1833: D.-Q. Jiang, M. Qian, M.-P. Qian, Mathematical Theory of Nonequilibrium Steady States. On the Frontier of Probability and Dynamical Systems. IX, 280 p, 2004. Vol. 1834: Yo. Yomdin, G. Comte, Tame Geometry with Application in Smooth Analysis. VIII, 186 p, 2004.

Vol. 1835: O.T. Izhboldin, B. Kahn, N.A. Karpenko, A. Vishik, Geometric Methods in the Algebraic Theory of Quadratic Forms. Summer School, Lens, 2000. Editor: J.-P. Tignol (2004) Vol. 1836: C. Nˇastˇasescu, F. Van Oystaeyen, Methods of Graded Rings. XIII, 304 p, 2004. Vol. 1837: S. Tavaré, O. Zeitouni, Lectures on Probability Theory and Statistics. Ecole d’Eté de Probabilités de Saint-Flour XXXI-2001. Editor: J. Picard (2004) Vol. 1838: A.J. Ganesh, N.W. O’Connell, D.J. Wischik, Big Queues. XII, 254 p, 2004. Vol. 1839: R. Gohm, Noncommutative Stationary Processes. VIII, 170 p, 2004. Vol. 1840: B. Tsirelson, W. Werner, Lectures on Probability Theory and Statistics. Ecole d’Eté de Probabilités de Saint-Flour XXXII-2002. Editor: J. Picard (2004) Vol. 1841: W. Reichel, Uniqueness Theorems for Variational Problems by the Method of Transformation Groups (2004) Vol. 1842: T. Johnsen, A. L. Knutsen, K3 Projective Models in Scrolls (2004) Vol. 1843: B. Jefferies, Spectral Properties of Noncommuting Operators (2004) Vol. 1844: K.F. Siburg, The Principle of Least Action in Geometry and Dynamics (2004) Vol. 1845: Min Ho Lee, Mixed Automorphic Forms, Torus Bundles, and Jacobi Forms (2004) Vol. 1846: H. Ammari, H. Kang, Reconstruction of Small Inhomogeneities from Boundary Measurements (2004) Vol. 1847: T.R. Bielecki, T. Björk, M. Jeanblanc, M. Rutkowski, J.A. Scheinkman, W. Xiong, Paris-Princeton Lectures on Mathematical Finance 2003 (2004) Vol. 1848: M. Abate, J. E. Fornaess, X. Huang, J. P. Rosay, A. Tumanov, Real Methods in Complex and CR Geometry, Martina Franca, Italy 2002. Editors: D. Zaitsev, G. Zampieri (2004) Vol. 1849: Martin L. Brown, Heegner Modules and Elliptic Curves (2004) Vol. 1850: V. D. Milman, G. Schechtman (Eds.), Geometric Aspects of Functional Analysis. Israel Seminar 20022003 (2004) Vol. 1851: O. Catoni, Statistical Learning Theory and Stochastic Optimization (2004) Vol. 1852: A.S. Kechris, B.D. Miller, Topics in Orbit Equivalence (2004) Vol. 1853: Ch. Favre, M. Jonsson, The Valuative Tree (2004) Vol. 1854: O. Saeki, Topology of Singular Fibers of Differential Maps (2004) Vol. 1855: G. Da Prato, P.C. Kunstmann, I. Lasiecka, A. Lunardi, R. Schnaubelt, L. Weis, Functional Analytic Methods for Evolution Equations. Editors: M. Iannelli, R. Nagel, S. Piazzera (2004) Vol. 1856: K. Back, T.R. Bielecki, C. Hipp, S. Peng, W. Schachermayer, Stochastic Methods in Finance, Bressanone/Brixen, Italy, 2003. Editors: M. Fritelli, W. Runggaldier (2004) Vol. 1857: M. Émery, M. Ledoux, M. Yor (Eds.), Séminaire de Probabilités XXXVIII (2005) Vol. 1858: A.S. Cherny, H.-J. Engelbert, Singular Stochastic Differential Equations (2005) Vol. 1859: E. Letellier, Fourier Transforms of Invariant Functions on Finite Reductive Lie Algebras (2005) Vol. 1860: A. Borisyuk, G.B. Ermentrout, A. Friedman, D. Terman, Tutorials in Mathematical Biosciences I. Mathematical Neurosciences (2005)

Vol. 1861: G. Benettin, J. Henrard, S. Kuksin, Hamiltonian Dynamics – Theory and Applications, Cetraro, Italy, 1999. Editor: A. Giorgilli (2005) Vol. 1862: B. Helffer, F. Nier, Hypoelliptic Estimates and Spectral Theory for Fokker-Planck Operators and Witten Laplacians (2005) Vol. 1863: H. Führ, Abstract Harmonic Analysis of Continuous Wavelet Transforms (2005) Vol. 1864: K. Efstathiou, Metamorphoses of Hamiltonian Systems with Symmetries (2005) Vol. 1865: D. Applebaum, B.V. R. Bhat, J. Kustermans, J. M. Lindsay, Quantum Independent Increment Processes I. From Classical Probability to Quantum Stochastic Calculus. Editors: M. Schürmann, U. Franz (2005) Vol. 1866: O.E. Barndorff-Nielsen, U. Franz, R. Gohm, B. Kümmerer, S. Thorbjønsen, Quantum Independent Increment Processes II. Structure of Quantum Lévy Processes, Classical Probability, and Physics. Editors: M. Schürmann, U. Franz, (2005) Vol. 1867: J. Sneyd (Ed.), Tutorials in Mathematical Biosciences II. Mathematical Modeling of Calcium Dynamics and Signal Transduction. (2005) Vol. 1868: J. Jorgenson, S. Lang, Posn (R) and Eisenstein Series. (2005) Vol. 1869: A. Dembo, T. Funaki, Lectures on Probability Theory and Statistics. Ecole d’Eté de Probabilités de Saint-Flour XXXIII-2003. Editor: J. Picard (2005) Vol. 1870: V.I. Gurariy, W. Lusky, Geometry of Müntz Spaces and Related Questions. (2005) Vol. 1871: P. Constantin, G. Gallavotti, A.V. Kazhikhov, Y. Meyer, S. Ukai, Mathematical Foundation of Turbulent Viscous Flows, Martina Franca, Italy, 2003. Editors: M. Cannone, T. Miyakawa (2006) Vol. 1872: A. Friedman (Ed.), Tutorials in Mathematical Biosciences III. Cell Cycle, Proliferation, and Cancer (2006) Vol. 1873: R. Mansuy, M. Yor, Random Times and Enlargements of Filtrations in a Brownian Setting (2006) Vol. 1874: M. Yor, M. Émery (Eds.), In Memoriam PaulAndré Meyer - Séminaire de Probabilités XXXIX (2006) Vol. 1875: J. Pitman, Combinatorial Stochastic Processes. Ecole d’Eté de Probabilités de Saint-Flour XXXII-2002. Editor: J. Picard (2006) Vol. 1876: H. Herrlich, Axiom of Choice (2006) Vol. 1877: J. Steuding, Value Distributions of L-Functions (2007) Vol. 1878: R. Cerf, The Wulff Crystal in Ising and Percolation Models, Ecole d’Eté de Probabilités de Saint-Flour XXXIV-2004. Editor: Jean Picard (2006) Vol. 1879: G. Slade, The Lace Expansion and its Applications, Ecole d’Eté de Probabilités de Saint-Flour XXXIV2004. Editor: Jean Picard (2006) Vol. 1880: S. Attal, A. Joye, C.-A. Pillet, Open Quantum Systems I, The Hamiltonian Approach (2006) Vol. 1881: S. Attal, A. Joye, C.-A. Pillet, Open Quantum Systems II, The Markovian Approach (2006) Vol. 1882: S. Attal, A. Joye, C.-A. Pillet, Open Quantum Systems III, Recent Developments (2006) Vol. 1883: W. Van Assche, F. Marcellàn (Eds.), Orthogonal Polynomials and Special Functions, Computation and Application (2006) Vol. 1884: N. Hayashi, E.I. Kaikina, P.I. Naumkin, I.A. Shishmarev, Asymptotics for Dissipative Nonlinear Equations (2006) Vol. 1885: A. Telcs, The Art of Random Walks (2006) Vol. 1886: S. Takamura, Splitting Deformations of Degenerations of Complex Curves (2006)

Vol. 1887: K. Habermann, L. Habermann, Introduction to Symplectic Dirac Operators (2006) Vol. 1888: J. van der Hoeven, Transseries and Real Differential Algebra (2006) Vol. 1889: G. Osipenko, Dynamical Systems, Graphs, and Algorithms (2006) Vol. 1890: M. Bunge, J. Funk, Singular Coverings of Toposes (2006) Vol. 1891: J.B. Friedlander, D.R. Heath-Brown, H. Iwaniec, J. Kaczorowski, Analytic Number Theory, Cetraro, Italy, 2002. Editors: A. Perelli, C. Viola (2006) Vol. 1892: A. Baddeley, I. Bárány, R. Schneider, W. Weil, Stochastic Geometry, Martina Franca, Italy, 2004. Editor: W. Weil (2007) Vol. 1893: H. Hanßmann, Local and Semi-Local Bifurcations in Hamiltonian Dynamical Systems, Results and Examples (2007) Vol. 1894: C.W. Groetsch, Stable Approximate Evaluation of Unbounded Operators (2007) Vol. 1895: L. Molnár, Selected Preserver Problems on Algebraic Structures of Linear Operators and on Function Spaces (2007) Vol. 1896: P. Massart, Concentration Inequalities and Model Selection, Ecole d’Été de Probabilités de SaintFlour XXXIII-2003. Editor: J. Picard (2007) Vol. 1897: R. Doney, Fluctuation Theory for Lévy Processes, Ecole d’Été de Probabilités de Saint-Flour XXXV-2005. Editor: J. Picard (2007) Vol. 1898: H.R. Beyer, Beyond Partial Differential Equations, On linear and Quasi-Linear Abstract Hyperbolic Evolution Equations (2007) Vol. 1899: Séminaire de Probabilités XL. Editors: C. Donati-Martin, M. Émery, A. Rouault, C. Stricker (2007) Vol. 1900: E. Bolthausen, A. Bovier (Eds.), Spin Glasses (2007) Vol. 1901: O. Wittenberg, Intersections de deux quadriques et pinceaux de courbes de genre 1, Intersections of Two Quadrics and Pencils of Curves of Genus 1 (2007) Vol. 1902: A. Isaev, Lectures on the Automorphism Groups of Kobayashi-Hyperbolic Manifolds (2007) Vol. 1903: G. Kresin, V. Maz’ya, Sharp Real-Part Theorems (2007) Vol. 1904: P. Giesl, Construction of Global Lyapunov Functions Using Radial Basis Functions (2007) Vol. 1905: C. Prévôt, M. Röckner, A Concise Course on Stochastic Partial Differential Equations (2007) Vol. 1906: T. Schuster, The Method of Approximate Inverse: Theory and Applications (2007) Vol. 1907: M. Rasmussen, Attractivity and Bifurcation for Nonautonomous Dynamical Systems (2007) Vol. 1908: T.J. Lyons, M. Caruana, T. Lévy, Differential Equations Driven by Rough Paths, Ecole d’Été de Probabilités de Saint-Flour XXXIV-2004 (2007) Vol. 1909: H. Akiyoshi, M. Sakuma, M. Wada, Y. Yamashita, Punctured Torus Groups and 2-Bridge Knot Groups (I) (2007) Vol. 1910: V.D. Milman, G. Schechtman (Eds.), Geometric Aspects of Functional Analysis. Israel Seminar 2004-2005 (2007) Vol. 1911: A. Bressan, D. Serre, M. Williams, K. Zumbrun, Hyperbolic Systems of Balance Laws. Lectures given at the C.I.M.E. Summer School held in Cetraro, Italy, July 14–21, 2003. Editor: P. Marcati (2007) Vol. 1912: V. Berinde, Iterative Approximation of Fixed Points (2007)

Vol. 1913: J.E. Marsden, G. Misiołek, J.-P. Ortega, M. Perlmutter, T.S. Ratiu, Hamiltonian Reduction by Stages (2007) Vol. 1914: G. Kutyniok, Affine Density in Wavelet Analysis (2007) Vol. 1915: T. Bıyıkoˇglu, J. Leydold, P.F. Stadler, Laplacian Eigenvectors of Graphs. Perron-Frobenius and FaberKrahn Type Theorems (2007) Vol. 1916: C. Villani, F. Rezakhanlou, Entropy Methods for the Boltzmann Equation. Editors: F. Golse, S. Olla (2008) Vol. 1917: I. Veselić, Existence and Regularity Properties of the Integrated Density of States of Random Schrödinger (2008) Vol. 1918: B. Roberts, R. Schmidt, Local Newforms for GSp(4) (2007) Vol. 1919: R.A. Carmona, I. Ekeland, A. KohatsuHiga, J.-M. Lasry, P.-L. Lions, H. Pham, E. Taflin, Paris-Princeton Lectures on Mathematical Finance 2004. Editors: R.A. Carmona, E. Çinlar, I. Ekeland, E. Jouini, J.A. Scheinkman, N. Touzi (2007) Vol. 1920: S.N. Evans, Probability and Real Trees. Ecole d’Été de Probabilités de Saint-Flour XXXV-2005 (2008) Vol. 1921: J.P. Tian, Evolution Algebras and their Applications (2008) Vol. 1922: A. Friedman (Ed.), Tutorials in Mathematical BioSciences IV. Evolution and Ecology (2008) Vol. 1923: J.P.N. Bishwal, Parameter Estimation in Stochastic Differential Equations (2008) Vol. 1924: M. Wilson, Littlewood-Paley Theory and Exponential-Square Integrability (2008) Vol. 1925: M. du Sautoy, L. Woodward, Zeta Functions of Groups and Rings (2008) Vol. 1926: L. Barreira, V. Claudia, Stability of Nonautonomous Differential Equations (2008) Vol. 1927: L. Ambrosio, L. Caffarelli, M.G. Crandall, L.C. Evans, N. Fusco, Calculus of Variations and NonLinear Partial Differential Equations. Lectures given at the C.I.M.E. Summer School held in Cetraro, Italy, June 27– July 2, 2005. Editors: B. Dacorogna, P. Marcellini (2008) Vol. 1928: J. Jonsson, Simplicial Complexes of Graphs (2008) Vol. 1929: Y. Mishura, Stochastic Calculus for Fractional Brownian Motion and Related Processes (2008)

Recent Reprints and New Editions Vol. 1618: G. Pisier, Similarity Problems and Completely Bounded Maps. 1995 – 2nd exp. edition (2001) Vol. 1629: J.D. Moore, Lectures on Seiberg-Witten Invariants. 1997 – 2nd edition (2001) Vol. 1638: P. Vanhaecke, Integrable Systems in the realm of Algebraic Geometry. 1996 – 2nd edition (2001) Vol. 1702: J. Ma, J. Yong, Forward-Backward Stochastic Differential Equations and their Applications. 1999 – Corr. 3rd printing (2007) Vol. 830: J.A. Green, Polynomial Representations of GLn , with an Appendix on Schensted Correspondence and Littelmann Paths by K. Erdmann, J.A. Green and M. Schocker 1980 – 2nd corr. and augmented edition (2007)

Stochastic Calculus for Fractional Brownian Motion and Related Processes

Stochastic calculus for fractional Brownian motion and related processes

Brownian Motion and Stochastic Calculus

Brownian motion and stochastic calculus

Brownian Motion and Stochastic Calculus, 2nd Edition

Brownian Motion Calculus

Brownian Motion Calculus

Brownian Motion and Stochastic Calculus (Graduate Texts in Mathematics)

Brownian Motion and Stochastic Calculus (Graduate Texts in Mathematics)

Brownian Motion and Stochastic Flow Systems

Markov processes, Brownian motion, and time symmetry

Levy processes and stochastic calculus

Brownian motion

Brownian Motion

Lévy Processes and Stochastic Calculus

Levy processes and stochastic calculus

Brownian motion

Continuous Martingales and Brownian Motion

Continuous martingales and Brownian motion

Aspects of Brownian Motion

Aspects of Brownian motion

Malliavin Calculus for Processes with Jumps (Stochastic Monographs : Theory and Applications of Stochastic Processes 2)

Malliavin Calculus for Processes with Jumps (Stochastic Monographs : Theory and Applications of Stochastic Processes, Vol 2)

Malliavin Calculus for Processes with Jumps (Stochastic Monographs: Theory and Applications of Stochastic Processes, Vol 2)

Stochastic Mechanics and Stochastic Processes

Fractional Calculus and Its Applications

q-Fractional Calculus and Equations

Dynamical theories of Brownian motion

Dynamical theories of Brownian motion

Dynamical theories of Brownian motion

Aspects of Brownian Motion (Universitext)

Stochastic Calculus for Fractional Brownian Motion and Related Processes