MX - Non Gaussian Merton Black Scholes Theory

Advanced Series on
Statistical Science &
Applied Probability Vol.9
NON-GAUSSIAN
MERTON-BLACK-SCHOLES
THEORY
Svetlana I. Boyarchenko
Sergei Z . Levendorskii
World Scientific
NON-GAUSSIAN
THEORY
ADVANCED SERIES ON STATISTICAL SCIENCE &
APPLIED PROBABILITY
Editor: Ole E. Barndorff-Nielsen
Published
Vol. 1: Random Walks of Infinitely Many Particles
by P. Revesz
Vol. 2: Ruin Probabilities
by S. Asmussen
Vol. 3: Essentials of Stochastic Finance: Facts, Models, Theory
by Albert N. Shiryaev
Vol. 4: Principles of Statistical Inference from a Neo-Fisherian Perspective
by L. Pace and A. Salvan
Vol. 5: Local Stereology
by Eva B. Vedel Jensen
Vol. 6: Elementary Stochastic Calculus — With Finance in View
by T. Mikosch
Vol. 7: Stochastic Methods in Hydrology: Rain, Landforms and Floods
eds. O. E. Barndorff-Nielsen et al.
Vol. 8: Statistical Experiments and Decisions: Asymptotic Theory
by A. N. Shiryaev and V. G. Spokoiny
Vol. 9: Non-Gaussian Merton-Black-Scholes Theory
by S. I. Boyarchenko and S. 2. Levendorskii
Advanced Series on
Statistical Science &
Applied Probability
NON-GAUSSIAN
THEORY
Svetlana I. Boyarchenko
University of Texas at Austin, USA
Sergei Z . Levendorskii
Rostov State University of Economics, Russia
V f e World Scientific
wb New Jersey * London • Singapore • Hong Kong
Jersey-London-Sine
Published by
World Scientific Publishing Co. Pte. Ltd.
P O Box 128, Farrer Road, Singapore 912805
USA office: Suite IB, 1060 Main Street, River Edge, NJ 07661
UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
British Library Cataloguing-in-Publication Data

A catalogue record for this book is available from the British Library.
NON-GAUSSIAN MERTON-BLACK-SCHOLES THEORY

Copyright © 2002 by World Scientific Publishing Co. Pte. Ltd.
All rights reserved. This book or parts thereof, may not be reproduced in any form or by any means,
electronic or mechanical, including photocopying, recording or any information storage and retrieval
system now known or to be invented, without written permission from the Publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to
photocopy is not required from the publisher.
ISBN 981-02-4944-6
Printed in Singapore by Uto-Print

to our parents
This page is intentionally left blank
Preface
By now, the drawbacks of the Gaussian modelling in Financial Markets

and Investment under Uncertainty are well-known. In particular, Gaussian
models cannot produce so-called fat tails of observed probability densities,
which leads to under-pricing of financial risks. One can hardly make a
mistake by saying that the under-pricing of the risk was the main reason
for the Long Term Capital Management disaster or recent failures of rating
agencies to warn investors of a series of the defaults of the investment-
graded firms.
The purpose of the book is to introduce an analytically tractable and
computationally effective class of non-Gaussian models for shocks (Regular
Levy Processes of Exponential type (RLPE)), and related analytical meth-
ods similar to the initial Merton-Black-Scholes approach, which we call the
Merton-Black-Scholes theory (MBS-theory). The potential range of appli-
cations of the non-Gaussian variant of the MBS-theory is huge, and the list
of results we have obtained so far does not exhaust all the possibilities.
As applications to Financial Mathematics, we solve pricing problems for
several types of perpetual American options, barrier options, touch-and-out
options and some other options, provide analogues of several approximate
methods for pricing of American options in the finite horizon case, and
deduce explicit analytical formulas for the locally risk-minimizing hedging.
We suggest fast computational procedures for pricing of European options;
they can be used for hedging and pricing of American and barrier options
as well.
As applications to Corporate Governance, we consider problems of en-
dogenous default, pricing of bonds of corporations, yield spreads of junk
vn
Vlll Preface
bonds, optimal leverage, optimal timing of investment under uncertainty

and optimal choice of the installed capital, and the capital accumulation
under non-Gaussian uncertainty. In particular, the correct form of the
Marshallian law is suggested, and numerical results are produced to show
how large the difference between prescriptions of Gaussian models and non-
Gaussian ones can be. We also consider discrete time analogues of perpet-
ual American options and the problem of the optimal choice of capital, and
outline several possible directions in which the methods of the book can be
developed further.
We tried to choose applications interesting for Financial engineers, spe-
cialists in Financial Economics, Real Options, and Partial Differential Equa-
tions (PDE) (especially in Pseudodifferential Operators (PDO)), and we
hope that specialists in Stochastic Processes will benefit from the usage
of PDO-technique in non-Gaussian situations. We also tried to make the
book accessible for graduate students in relevant areas and mathematicians
without prior knowledge of Finance and Economics.
As in the MBS-theory, we reduce problems of pricing of options, other
derivative securities and corporate bonds, optimal timing of investment
etc. to boundary value problems for the (generalized) Black-Scholes equa-
tion. In the MBS-theory, when stochastic processes are assumed Gaussian,
the Black-Scholes equation is a differential one, while in the non-Gaussian
case, it is a non-local pseudodifferential (or integro-differential) equation,
the difference of properties being the same as the one between the infinites-
imal generators of Gaussian Feller processes and non-Gaussian ones. We
use Dynkin's formula, and basic results from the potential theory for the
Markov processes, which is simpler than the modern sophisticated approach
based on the theory of semimartingales. We hope that the reader will find
refreshing a book on Mathematical Finance and Investment under Uncer-
tainty, which makes no use of the Ito formula at all.
After a boundary value problem for the generalized Black-Scholes equa-
tion is developed, we use the PDO-technique to find the solution. The ana-
lytical part is unavoidably more difficult than the PDE-part in the Gaussian
MBS-theory but we manage to use the most simple part of the theory of
PDO by restricting ourselves to the case of RLPE: the generators of RLPE
are PDO with constant symbols having fairly favourable properties. This
is the reason why we solve the boundary value problems for the generalized
Black-Scholes equation in the case when the shocks are modelled as RLPE
but not as more general Feller processes. We may say that RLPE is the
Preface IX
simplest class of processes with stationary independent increments if the

Brownian motion (BM) is not available. The PDO-technique used in the
book can be applied for a wide class of strong Feller processes generalizing
the class RLPE in the same spirit as Gaussian diffusions generalize BM. To
illustrate this point, we construct a class of Levy-like Feller processes, and
show how to price European options under these processes.
Certainly, it was difficult to aim simultaneously at economists and math-
ematicians from different fields: almost each part of the book may seem
trivial to one of the groups of readers whereas the other groups may find
this very part illuminating. Probably, it would be better to write a separate
book for each group of readers, and in fact, we had already started to write
a book on Pseudodifferential Operators with Applications to Finance and
Economics, when during a discussion with Ole E. Barndorff-Nielsen it was
decided that a book aimed at a wider audience should be written first.
Having in mind a diverse audience, we tried to write a book so that it
was simple in the beginning and more technical in further chapters. The
main part of the book, especially chapters on Investment under Uncertainty
and Endogenous Default, are written (almost) as an economic publication
though economists may find some parts of the exposition too terse. We
wrote a detailed Introduction in order to explain informally the main in-
gredients of our approach, so that the reader can read any part of the book
she is interested in conjunction with the Introduction and Chapter 3 only;
for the reader who is interested in all technical details, in Chapter 2 we list
main definitions and results of the theory of Levy processes, which we use in
the book, and provide a scheme for the reduction of the pricing problem to
a boundary value problem for the generalized Black-Scholes equation; and
we finish the book with rigorous proofs of the most technical statements
and with a systematic list of the results of the theory of PDO, which are
used in the book.
During our work on various topics covered in the book, we benefited
from illuminating comments and suggestions by I. Bouchouev, J. Cvitanic,
A. Dixit, D. Duffie, J. M. Harrison, I. Karatzas, G. Peskir, A. N. Shiryaev,
and Ken-Iti Sato; long discussions with E. Eberlein and O. E. Barndorff-
Nielsen were especially useful.
We are thankful to our son, Dmitriy Boyarchenko, who read the manu-
script, finding errors, suggesting improvements in exposition, and correcting
our grammar.
X Preface
0.0.1 General notation

We denote the real (complex) n-dimensional space by R n (C n ); R™ denotes
the cone of real vectors with non-negative components, and R " + stands for
the cone of vectors with positive components. For x,y € R n , x > y means
x - y € R™, and x > y means x — y € R++- Z+ stands for the set of
non-negative integers.
Unless otherwise stated, (x, y) denotes the standard scalar product of
vectors x,y € R™, and | • | denotes the standard norm in R n (or in C",
depending on the context). The Lebesgue measure on R n is denoted by
dx, and the notation for partial derivatives used in the book is dj = gf-,
Dj = — idj, where i = y/—\.
For x £ R n and a multi-index a = (ai,...,a„) € (Z+) n , set \a\ =
Qi + . . . + a „ ,
xa =xtl • • • < " , da =8?1---d*~, Da = Df1---D%".

If a is a function on R™, and a is a multi-index, then a^ = daa; and if a
is a function on R " x R ? , then a^j(a;,^) = dfD^.a{x,^).
For U C R", 1[/ denotes the indicator function of U: lu(x) = 1 if
x € [/ and 0 otherwise.
For real a, b, set a A b :— min{a, b}, a V b := max{a, 6}, and a+ — a V 0,
a_ = a — a+.
If 5 is viewed as a subset of a set U, then Bc := U \B denotes the
complement of B in U.
For a subset B of a topological space, B denotes the closure of B, and
B° the interior.
<S(Rn) denotes the space of infinitely smooth functions vanishing at the
infinity faster any power of |x|, together with all derivatives, and Co(R")
is the space of continuous functions vanishing at the infinity.
If B\,B2 are normed spaces, and A : B\ —> B% is a bounded linear
operator then II^He^Sa denotes the operator norm of A.
For a positive integer m, C™(Rn) (resp., C m ( R " ) ) denotes the space
of m times continuously differentiable functions vanishing at infinity (resp.,
with each derivative up to order m uniformly bounded); the notation Lp(R n ),
p € [1, +oo], is also standard.
Depending on the context, T denotes the Fourier transform or a filtra-
tion on the filtered probability space fi. The Fourier transform of a function
/ is denoted by / .
Contents
Preface vii
0.0.1 General notation x
Chapter 1 Introduction 1
1.1 The Gaussian Merton-Black-Scholes theory 1
1.1.1 Strong points 1
1.1.2 Drawbacks 2
1.1.2.1 Fat tails and anomalous skewness and kurtosis 3
1.1.2.2 The Black-Scholes market and formula . . . . 4
1.1.2.3 Volatility smile and volatility surface 5
1.1.3 Remedies for the MBS-theory 6
1.1.3.1 General remarks 6
1.1.3.2 Jump-diffusion models and stochastic volatil-
ity models 6
1.1.3.3 Levy processes 7
1.2 Regular Levy Processes of Exponential type 8
1.2.1 Characteristic function and exponent, Levy measure and
Levy-Khintchine formula 9
1.2.2 Definition of RLPE in ID 10
1.2.3 Infinitesimal generators of RLPE as PDO 11
1.3 Pricing of contingent claims 13
1.3.1 Discrete time models with a discrete space of states: No-
arbitrage and equivalent martingale measures 13
xi
XII Contents
1.3.2 Discrete time models with a discrete space of states:

Completeness of the market, and pricing of derivative
securities 17
1.3.3 Absence of arbitrage, EMM and completeness in the
Gaussian Black-Scholes model market 19
1.3.4 Sufficient condition for no-arbitrage in a Levy market
and incompleteness of a Levy market. The pricing for-
mula for contingent claims of European type and the
problem of a choice of EMM 19
1.3.5 On pricing based on the utility maximization 22
1.4 The Generalized Black-Scholes equation 23
1.4.1 The informal derivation 23
1.4.2 An outline of the reduction of the pricing of contingent
claims to boundary value problems for the generalized
Black-Scholes equation: barrier options 24
1.4.3 The case of interest bearing securities 25
1.4.4 The generalized Merton-Black-Scholes theory 26
1.4.5 Optimal stopping problems and the smooth pasting con-
dition 27
1.4.6 The case of a dividend-paying stock 27
1.5 Analytical methods used in the book 28
1.5.1 The Fourier transform, and Complex Analysis 28
1.5.2 The Wiener-Hopf factorization and the Wiener-Hopf equa-
tion 28
1.5.3 The case of the non-stationary Black-Scholes equation
and the constant barrier 30
1.5.4 The case of the non-constant barrier and multi-asset con-
tracts 31
1.5.5 Pseudodifferential operators 31
1.6 An overview of the results covered in the book 31
1.6.1 Elements of the theory of Levy processes 31
1.6.2 Option pricing 31
1.6.3 Investment under uncertainty and capital accumulation 33
1.6.4 Endogenous default and pricing of the corporate debt . 34
1.6.5 Numerical methods 35
1.6.6 Extensions 35
1.6.7 Basics of PDO theory 36
1.7 Commentary 36
Contents xiii
Chapter 2 Levy processes 39

2.1 Basic notation and definitions 39
2.1.1 Random variables 39
2.1.2 Conditional expectation 41
2.1.3 The Fourier transform and the Laplace transform.
Characteristic functions 42
2.2 Levy processes: general definitions 45
2.2.1 Levy processes and infinitely divisible distributions . . . 45
2.2.2 The Levy-Khintchine formula and generating triplet . . 47
2.2.3 Constructions of Levy processes 48
2.2.3.1 Explicit constructions 48
2.2.3.2 Subordinators and subordinated processes . . 49
2.2.3.3 Linear transformations of Levy processes . . . 50
2.2.4 The Wiener-Hopf factorization 50
2.3 Levy processes as Markov processes 52
2.3.1 Markov property and Feller property 52
2.3.2 Stopping times and the strong Markov property . . . . 53
2.3.3 Resolvent operator and infinitesimal generator 55
2.3.4 Dynkin's formula 57
2.3.5 Duality 58
2.3.6 Absolutely continuous resolvents 58
2.3.7 Operators PqB 59
2.4 Boundary value problems for the Black-Scholes-type equation . 62
2.4.1 Boundary problems for the stationary Black-Scholes type
equation 63
2.4.2 Boundary problems for the non-stationary Black-
Scholes type equation 64
2.5 Commentary 66
Chapter 3 Regular Levy Processes of Exponential t y p e in

ID 67
3.1 Model Classes 67
3.1.1 An overview 67
3.1.2 KoBoL family: direct construction via the Levy-
Khintchine formula 68
3.1.2.1 Definition in terms of the Levy measure . . . 68
3.1.2.2 The characteristic exponent 69
3.1.2.3 Moments of a KoBoL process 71
Contents
3.1.2.4 Graphs of characteristic exponents and proba-

bility densities of different KoBoL processes . 72
3.1.2.5 Proofs of technical lemmas 73
3.1.3 Normal Inverse Gaussian processes and Normal Tem-
pered Stable Levy Processes: construction via subordi-
nation 77
3.1.4 Variance Gamma Processes 79
3.1.5 Hyperbolic Processes and Generalized Hyperbolic Pro-
cesses 79
3.1.6 Comparison of the tail behaviour of probability densities
for different model classes of processes 81
3.2 Two definitions of Regular Levy Processes of Exponential type 82
3.2.1 Definition in terms of the Levy measure 82
3.2.2 Definition in terms of the characteristic exponent . . . . 84
3.3 Properties of the characteristic exponents and probability den-
sities of RLPE 85
3.4 Properties of the infinitesimal generators 87
3.5 A "naive approach" to the construction of RLPE or why they
are natural from the point of view of the theory of PDO . . . . 87
3.5.1 Modified Stable Processes and RLPE 88
3.6 The Wiener-Hopf factorization 89
3.6.1 The case of Levy processes of exponential type . . . . . . 89
3.6.2 The Wiener-Hopf factorization for RLPE 91
3.6.3 Approximate formulas for the factors in the case of NIG,
HP, KoBoL and NTS Levy processes 94
Chapter 4 Pricing and hedging of contingent claims of Euro-

pean type 97
4.1 Equivalent Martingale Measures in a Levy market 97
4.1.1 Esscher transform 98
4.1.2 General case 101
4.2 Pricing of European options and the generalized Black-Scholes
formula 104
4.2.1 Pricing formulas: convolution with the pricing kernel and
the Fourier-inversion formula 104
4.2.2 Call and put options 106
4.2.3 Numerical examples: the comparison with the Black-
Scholes formula, and the shapes of the smile 107
Contents xv
4.2.4 The problem of the model fitting and evaluation . . . . 110

4.3 Generalized Black-Scholes equation and its properties for differ-
ent RLPE and different choices of EMM, and implications for
parameter fitting Ill
4.4 Other European options 113
4.4.1 Power options 113
4.4.2 Digital options 113
4.4.3 Combinations 114
4.4.4 General case: the condition on the rate of growth of the
payoff at the infinity and the origin 114
4.5 Hedging 115
4.5.1 General discussion 115
4.5.2 Locally risk-minimizing hedging ratio 116
4.5.3 Continuity of the locally risk-minimizing hedging ratio
till the expiry 119
4.5.4 Numerical Examples 119
4.6 Commentary 120
Chapter 5 Perpetual American options 121

5.1 The reduction to a free boundary problem for the stationary
generalized Black-Scholes equation 121
5.1.1 General discussion 121
5.1.2 Free boundary value problem for the price of the perpet-
ual American option 122
5.1.3 Main Lemma 123
5.2 Perpetual American put: the optimal exercise price and the
rational put price 124
5.2.1 Main Theorem 124
5.2.2 Proof of optimality in the class Mo 126
5.2.3 Proof of optimality in the class M 129
5.2.4 Failure of the smooth pasting principle for some
RLPE's and its substitute 132
5.2.5 Approximate formulas for the case of model RLPE . . . 134
5.2.6 Proof of Theorem 5.2 134
5.3 Perpetual American call 139
5.3.1 Main results 139
5.3.2 Proof of optimality in the case of unbounded payoffs . . 140
5.4 Put-like and call-like options: the case of more general payoffs . 143
xvi Contents
5.4.1 Put-like options 143

5.4.2 Call-like options 146
5.5 Commentary 149
Chapter 6 American options: finite time horizon 151

6.1 General discussion 151
6.1.1 The free boundary problem 151
6.1.2 The role of dividends and non-zero interest rate 152
6.2 Approximations of the American put price 153
6.2.1 Geske-Johnson approximation 153
6.2.2 Analytic method of lines 154
6.2.3 Carr's randomization 158
6.3 American put near expiry 159
6.3.1 Behaviour of the early exercise boundary near the ma-
turity date 159
6.3.2 A general equation for /i(A) in the case of NTS pro-
cesses 160
6.3.3 The case of processes of order v € [1,2) 162
Chapter 7 First-touch digitals 165

7.1 An overview 165
7.2 Exact pricing formulas for first-touch digitals 166
7.3 The Wiener-Hopf factorization with a parameter 169
7.3.1 Formulas for the factors revisited 169
7.3.2 Analytic continuation of the factors w.r.t. A 172
7.4 Price near the barrier 177
7.4.1 Asymptotics a s i - > +0, and r > 0 is fixed 177
7.4.1.1 Main Theorem 177
7.4.1.2 Numerical example 178
7.4.1.3 Proof of Theorem 7.4 179
7.4.2 Asymptotics as x —> +0 and r —> +oo 182
7.5 Asymptotics as r —> +oo 183
Chapter 8 Barrier options 185

8.1 Types of barrier options 185
8.1.1 Standard barrier options 185
8.1.2 Options with a rebate 186
8.2 Down-and-out call option without a rebate 187
Contents xvii
8.2.1 Reduction to the boundary problem for the Generalized

Black-Scholes equation 187
8.2.2 Reduction to the Cauchy problem for an ordinary differ-
ential operator with the operator-valued coefficient . . . 189
8.2.3 Construction of the resolvent, and the Wiener-Hopf fac-
torization with a parameter 192
8.2.4 Proof of Eq. (8.17) and continuity of the RHS in
Eq. (8.19) 196
8.3 Asymptotics of the option price near the barrier 197
8.4 Commentary 197
Chapter 9 Multi-asset contracts 199

9.1 Multi-dimensional Regular Levy Processes of Exponential type 199
9.1.1 Multi-dimensional KoBoL family 200
pered Stable Levy processes 201
9.1.3 Hyperbolic Processes, Generalized Hyperbolic Processes
and Variance Gamma Processes 202
9.1.4 Definition of multi-dimensional RLPE 202
9.2 European-style contracts 203
9.2.1 The Esscher transform 204
9.2.2 Pricing of power forwards 205
9.2.3 Pricing of European options on one asset 206
9.2.4 Pricing of exchange options and basket options 207
9.3 Locally risk-minimizing hedging with a portfolio of several
assets 209
9.3.1 The set-up 209
9.3.2 Weights of the hedging portfolio 210
9.3.3 Hedging of European claims 214
9.3.4 Locally risk-minimizing hedging of "power forwards" . . 215
9.3.5 Approximate locally risk-minimizing hedging:
general case 216
9.4 Weighted discretely sampled geometric average 216
Chapter 10 Investment under uncertainty and capital accu-

mulation 221
10.1 Irreversible investment and uncertainty 221
10.2 The investment threshold 223
xviii Contents
10.3 Capital accumulation under RLPE 225

10.4 Computational results 227
10.5 Approximate formulas and the comparative statics 230
Chapter 11 Endogenous default and pricing of the corporate

debt 231
11.1 An overview 231
11.2 Endogenous default 233
11.3 Equity of a firm near bankruptcy level and the yield spread for
junk bonds 239
11.4 The case of a solvent firm 242
11.5 Endogenous debt level and endogenous leverage 247
11.6 Conclusion 248
11.7 Auxiliary results 249
11.7.1 Proof of Eq. (11.36)-Eq. (11.38) 249
11.7.2 Proof of Eq. (11.45)-Eq. (11.46) 251
Chapter 12 Fast pricing of European options 255

12.1 Introduction 255
12.2 Transformation of the pricing formula for the European put . . 258
12.3 F F T and IAC 260
12.3.1 Fast Fourier Transform 260
12.3.2 Integration-along-cut method 261
12.3.2.1 Case a) 262
12.3.2.2 Case b) 263
12.4 Comparison of F F T and IAC 264
Chapter 13 Discrete time models 267

13.1 Bermudan options and discrete time models 267
13.2 A perpetual American put in a discrete time model 269
13.2.1 Process specification 269
13.2.2 The pricing problem for a perpetual American option as
an optimal stopping problem 270
13.3 The Wiener-Hopf factorization 272
13.3.1 General formulas 272
13.3.2 Some useful properties of the factors a± 276
13.3.3 The case of exponential polynomials 276
13.4 Optimal exercise boundary and rational price of the option . . 278
Contents xix
Chapter 14 Feller processes of normal inverse Gaussian

type 281
14.2 Constructions of NIG-like Feller process via pseudodifferential
operators 284
14.2.1 Naive construction: NIG-like generators with state-
dependent parameters 284
14.2.2 Constructions via infinitesimal generators: subordina-
tion of semigroups of operators 286
14.3 Applications for financial mathematics 289
14.3.1 Generalised Black-Scholes equation for contingent claims
under NIG-like Feller processes 289
14.3.2 Pricing of European options under NIG-like Feller pro-
cesses obtained via subordination 290
14.3.3 Pricing of European options under NIG-like Feller pro-
cesses obtained by the naive approach 290
14.4 Discussion and conclusions 294
Chapter 15 Pseudo-differential operators with constant sym-

bols 295
15.2 Classes of functions 297
15.2.1 General notation and remarks 297
n
15.2.2 The space 5 ( R ) 298
15.2.3 The Fourier transform 299
15.3 Space <S'(Rn) of generalized functions on R " 300
15.3.1 Generalized functions (distributions) 300
n
15.3.2 Operations in S'(R ) 301
15.3.3 The Fourier transform of generalized functions 302
15.3.4 The regularization of oscillatory integrals 303
15.3.5 Tensor product of generalized functions 304
15.3.6 Generalized functions on an open set. The support of a
generalized function 305
n
15.4 Pseudo-differential operators with constant symbols on R . . . 305
15.4.1 Definition of PDO with constant symbols 305
15.4.2 The Sobolev spaces Hs{Rn) 307
3 n
15.4.3 Action of PDO in the scale H (R ) 308
15.4.4 Elliptic equations 309
XX Contents
15.4.5 Restriction to a hyperplane 309

15.4.6 The Sobolev embedding theorem 310
15.4.7 Action in the Holder spaces 310
15.5 The action of PDO in the Sobolev spaces on R " ± 312
15.5.1 Spaces H(Rn±) and PDO in a half-space 312
15.5.2 Spaces Hs(Rn±) 312
15.5.3 The Cauchy-type integral and its decomposition . . . . 314
15.6 Parabolic equations 316
15.6.1 The Cauchy problem for the parabolic equation 316
15.6.2 Reduction to a family of ODE on R+ 317
15.6.3 Parabolic equation as ODE with the operator-valued co-
efficient 319
15.6.4 Parabolic equations and the Holder-Zygmund spaces . . 322
15.6.5 The inhomogeneous parabolic equation and the asymp-
totics of the solution as t —» +oo 323
15.7 The Wiener-Hopf equation on a half-line I 324
15.7.1 Statement of the problem 324
15.7.2 The Wiener-Hopf factorization 325
15.7.3 Main Theorems 332
15.7.4 The asymptotics of the solution near the boundary . . . 334
15.8 Parabolic equations on [0,T] x R + 338
15.8.1 The statement of the boundary problem 338
15.8.2 The Wiener-Hopf factorization with a parameter . . . . 339
15.8.3 Reduction to ODE with the operator-valued coefficient
and the solution 342
15.9 PDO in the Sobolev spaces with exponential weights, in ID . . 344
15.9.1 Generalized functions 344
15.9.2 The Sobolev spaces # S - T ( R ) 346
15.9.3 PDO with the symbols holomorphic in a strip, and their
action in the scale HS'~'(R) 351
15.9.4 Action of PDO in the Holder-Zygmund spaces with ex-
ponential weights 353
15.10The Sobolev spaces with exponential weights and PDO on a
half-line . . . . .a 354
15.10.1 Spaces fl"*'(R±) 354
15.10.2Spaces Hs<-f(R±) 355
15.10.3 The Cauchy-type integral and its decomposition . . . . 357
Contents xxi
15.11Parabolic equations in spaces with exponential weights 358

15.12The Wiener-Hopf equation on a half-line II 358
15.12.1 Statement of the problem 358
15.12.2 The Wiener-Hopf factorization of symbols holomorphic
in a strip 359
15.12.3Main Theorems 361
15.12.4 The asymptotics of solutions near the boundary and the
free boundary problems 362
15.12.5 The asymptotics of solutions at the infinity 362
15.13Parabolic equations on R x R + with exponentially growing data 363
Chapter 16 Elements of calculus of pseudodifferential oper-

ators 365
16.1 Basics of the theory of PDO with symbols of the class S ^ R " x
Rn) 366
16.1.1 Asymptotic summation and the composition theorem . 366
16.1.2 Parametrix and approximate square root 369
16.1.3 Boundedness theorem 372
16.1.4 The inverse operator and the square root 373
16.2 Operators depending on parameters 373
16.3 Operators with symbols holomorphic in a tube domain 377
16.4 Proofs of auxiliary technical results 379
16.5 Change of variables and pricing of multi-asset contracts . . . . 381
16.6 Pricing of barrier options under Levy-like Feller processes . . . 382
Bibliography 385
Index 393
Chapter 1
Introduction
In this chapter, we list both strong and weak points of the Gaussian MBS-
theory, explain why it needs modification, describe informally the technique
used in the book and present briefly the main results of the book.
1.1 The Gaussian Merton-Black-Scholes theory
1.1.1 Strong points

Ideally, an applied mathematical theory in Finance and Economics should
possess the following merits:
(i) the theory has a wide range of applications;

(ii) in many situations of interest, the theory provides analytical an-
swers in the form suitable for the analysis of the dependence of the
endogenous variables on the exogenous parameters (comparative
statics);
(iii) Mathematics used in the theory is accessible for the wide audience,
in simplest situations at least;
(iv) moreover, it is strongly advisable that models of the theory can be
explained and applied at the level of "the rule of the thumb";
(v) when the analytical solution is not known, effective numerical meth-
ods are available;
(vi) the basic principles of the theory and the answers it provides are
in good agreement with the reality.
(If a model satisfies (ii), or better, (ii)-(iv), economists call it "tractable").
l
2 Introduction
The Merton-Black-Scholes theory is so popular among practitioners and

researches in Finance and Economics just because it satisfies criteria (i)-
(v) almost to perfection. It has been successfully applied to most of the
problems in Mathematical Finance, Financial Economics, theory of Real
Options and in many other fields of Economics, essentially to any problem
where the uncertainty matters, and in model situations the answers have
been obtained in the form of fairly simple analytical expressions convenient
for further analysis. Mathematics used in the MBS-theory is the theory of
Partial Differential Equations (PDE), and in many cases of interest, only
the knowledge of Ordinary Differential Equations (ODE) suffices. Being at
the undergraduate level, ODE-technique is widely known, and its usage in
the MBS-theory can be explained informally, as Dixit and Pindyck (1994)
demonstrated. Finally, for PDE and ODE, a host of effective numerical
methods is available.
1.1.2 Drawbacks
In the case of the MBS-theory, the motto: "Our vices are continuation of
our virtues" 1 should be rephrased in the reverse order: "Our virtues are
continuation of our vices". Both the simplicity and success of the MBS-
theory stem from the possibility of reduction to boundary value problems
for PDE or ODE, and this reduction becomes possible due to the choice of
the simplest class of stochastic processes - Gaussian processes - to model
the evolution of prices in financial and other markets. (The rival modern
martingale approach, which tries to remain in the realm of Probability
theory and avoid the usage of PDE and PDO whenever it can help it,
also capitalizes on the extraordinary nice properties of Gaussian processes).
In the basic Black-Scholes model, the price of a stock (or index, like the
Standard & Poor 500 Index) follows the Geometric Brownian motion: St =
expX t , where Xt is the Brownian motion. Then the probability density
PAt (%) of the increments Xt+At — Xt is given by
( \ 1 ( (x-^t)2\
PAt{x) exp (L1)
= 7^Tt {-Âr)'
where p, and a2 are called the drift and volatility. As Eq. (1.1) shows, the
tails of the probability density decay faster than an exponenial function as
Translated from Russian

The Gaussian Merton-Black-Scholes theory 3
DAX 30 FTSE 100
Fig. 1.1 Log-densities for major indices
X -» ± 0 0 .
1.1.2.1 Fat tails and anomalous skewness and kurtosis

Unfortunately for the MBS-theory, empirical studies strongly reject the
Gaussian model. Figure 1.1, kindly provided by Ole E. Barndorff-Nielsen
and Neil Shephard, shows non-parametric estimators and fitted non-Gaussian
Generalized Hyperbolic model for several major world indices. It is clearly
seen that the tails of the log-densities decay approximately linearly; if the
Gaussian model fitted well, the empirical log-densities would have been
parabolas. Thus, BM fits poorly. In particular, the tails decay much slower
than the BM model suggests: "the tails are fat". One may also notice
that the central part is fitted well by the Levy stable distribution (we will
comment on the last observation in the next subsections).
The deviation from the Gaussianity can also be easily inferred from
anomalous values of skewness and kurtosis of empirical pAt- Recall that the
mean, m; variance, a1; skewness, A3; and kurtosis, K = A4, of a probability
distribution p(x)dx are defined as follows:
+00
xp(x)dx; a2 := ((x — m) 2 );
/ -00
4 Introduction
{{x-mf) _ ((x - m) 4 )
A3:- -3 ; «:- -4 .
For a Gaussian p, A3 = 0 and K = 3, whereas for empirical p(x)dx =

PAt(x)dx, the skewness AA*,3 is non-zero, and the kurtosis K^t can be much
larger than 3, especially for small At.
There exists another important feature of the financial markets which
is not captured by the classical model of financial markets: the quasi long
range dependencies. Whereas the estimated autocorrelation functions based
on log price differences on stocks or currencies are generally closely consis-
tent with an assumption of zero autocorrelation, the empirical autocorrela-
tion functions of the absolute values or the squares of the returns may stay
positive for many lags.
Thus, the model for the price process in the MBS-theory is unrealistic
but it might not have been a very serious problem if the answers provided by
the theory were in good agreement with reality. Much thinner tails of prob-
ability density of BM mean that under the Gaussian modelling, the extreme
events such as large drops of a price are assigned negligible probabilities.
This leads to under-pricing of financial risks with severe consequences like
the Long Term Capital Management disaster. Even if one does not require
that the theory takes the extreme events into account properly, one should
expect that in relatively calm periods of the market activity the theory
performs well, but the MBS-theory noticeably fails in this respect, too.
The most celebrated result of the theory, the Black-Scholes option pric-
ing formula, seemed to produce fairly good approximations to observed
option prices during the first several years after its invention, but later
persistent deviations have been noticed. After October 1987 crash, the sys-
tematic errors increased, and in fact, for a long time practitioners have been
using the Black-Scholes formula simply as a coding machine to re-express
and interpret the observed option prices. To explain how the machine
works, we describe briefly the Black-Scholes market and formula.
1.1.2.2 The Black-Scholes market and formula

Consider the market of a riskless bond yielding the constant rate of return
r > 0, and an asset, whose price follows the Geometric Brownian motion:
St — exp Xt, where Xt is the Brownian motion with the drift and diffusion
coefficients /J, and a2; in Finance, a2 is called the volatility.
Consider a contract which gives to its holder the right, but not the obli-
gation, to buy the underlying asset for the specified price (the strike price),
K, at the specified expiry date, T. This contract is called the European call
option. Denote its price at time t < T, conditioned on the current price
(spot price), St, of the underlying asset by Fca,\\(St,t). The Black-Scholes
formula reads
FcMSt,t) = StNfr) - Ke~rTN{d2), (1.2)
where r = T — t is the time to expiry,

HSt/K) + (r + a2/2)r
di = •= , d2 = «i - O\/T,
and
N{x) = -J= / S e't2'2dt

is the distribution function of the standard normal random variable. It
is seen that for fixed r, T, t and St, Eq. (1.2) establishes a one-to-one
correspondence between the option price and the volatility a2. Given the
observed option price, we can use Eq. (1.2) to calculate the volatility which
leads to this price. The volatility inferred in this way is called the implied
volatility, as opposed to the historic volatility inferred from the observations
of the dynamics of the price of the underlying asset. If the Black-Scholes
formula were correct, the implied volatility, (Timp, would have been indepen-
dent of both the time to the expiry, r, and the strike price, K, and equal
to the historic volatility, a. In reality, this is not the case.
1.1.2.3 Volatility smile and volatility surface

Figure 1.2 is a stylised graph of the implied volatility as a function of the
moneyness S/K, T being fixed. The reader can easily understand why it
is called the volatility smile. If the Black-Scholes formula were correct, the
graph would have been a straight line shown on the picture. Smiles of
similar shapes were typically observed before October 1987 crash.
Later, the shape of the smile changed; sometimes, the smile is more
pronounced, in other cases, it is downward sloping or upward sloping. It
can even be a frown. In all cases, the shape of the smile depends on the
time to the expiry, and becomes more pronounced for small r. The graph of
the implied volatility as a function of (K, r) is called the volatility surface.
The reader can find several examples of volatility surfaces in Chapter 4.
6 Introduction
0.194 ~
0.192 -
0.19 -
0.188 -
0.178 -
0.176 -
0.174^-
0.6 0.7 O.8 0.9 1 1.1 1.2 1.3 1.4
S/K
Fig. 1.2 Volatility smile
1.1.3 Remedies for the MBS-theory

1.1.3.1 General remarks
The volatility smile clearly demonstrates that the Black-Scholes model is
not adequate. Some researches try to deduce the smile from various frictions
and distortions in the market while keeping the Gaussian assumption intact,
whereas others invent heuristic models to describe dynamics of the volatility
surface. We believe that it is more reasonable to start with more realistic
models for prices of the underlying securities rather than trying to mend
the MBS-theory while remaining in the Gaussian world.
1.1.3.2 Jump-diffusion models and stochastic volatility models

The MBS-theory becomes more flexible if additional stochastic factors are
introduced, as in the jump-diffusion models, where Xt is modelled as a
mixture of independent BM and Poisson processes, or the diffusion coeffi-
cient itself is modelled as a stochastic process (stochastic volatility models).
There is a good rationale in introducing jumps since jumps in prices are
easy to notice, but jump-diffusion models either do not fit well to data, if
the number of jumps is small, or they are not very tractable, if many jumps
are allowed. Stochastic volatility models (SV-models) have the advantage
of using an additional observable, the volatility, which increases the accu-
racy of the pricing, and they provide a modelling of the important quasi
long range dependence typically observed in the financial markets, which

is impossible with Levy processes, but there is a price to pay: the usage of
SV-models leads to PDE in the space of higher dimensions, and these PDE
have variable coefficients. As a result, analytical formulas are usually not
available, though there are some exceptional tractable cases for European
options - see the commentary in the end of the chapter; for contracts with
early exercise features, analytical formulas are unknown, and one has to
resort to numerical methods from the very beginning.
There is another objection to the usage of jump-diffusion models, of
a more fundamental nature: the empirical studies of high-frequency data
suggest that the processes typically observed in Financial markets do not
have a diffusion component at all. Notice that there are important situa-
tions, when the pricing under processes with a Gaussian component differs
from the one under purely discontinuous processes not only quantitatively
but qualitatively as well, the typical example being the pricing of a barrier
option or touch-and-out option near the barrier.
1.1.3.3 Levy processes

As a process with the independent stationary increments, the Brownian
motion is a member of a wide family of Levy processes. As early as in
1963, Mandelbrot suggested to use the Levy stable distributions to model
the returns in the Financial markets. Mandelbrot's idea is corroborated
to some extent by the shape of the observed probability densities in the
central part, where they are fitted well by Levy stable distributions, but
the tails of the Levy stable distributions are too fat: polynomially decay-
ing, whereas many empirical studies suggest that the tails decay exponen-
tially; even worse, non-Gaussian stable distributions have infinite second
moments, which contradicts the observed convergence of empirical proba-
bility distributions to Gaussian ones over longer time scale 2 . Finally, if X
is the stable Levy process under the measure chosen by the market, then
the expectation of the stock price E[ex'] = oo, which makes the model
unsuitable for consistent pricing.
From the beginning of the 90th, several families of Levy processes with
probability densities having semi-heavy, that is, exponentially decaying
2
T h e last observation explains why the Gaussian MBS-theory performs much better for
options far from the expiry
8 Introduction
tails 3 have been used to model stock returns and price options: Variance
Gamma Processes (VGP), used by D. Madan with co-authors; Normal In-
verse Gaussian Processes (NIG), used by O.E. BarndorfF-Nielsen's group;
Hyperbolic Processes and Generalized Hyperbolic Processes (HP and GHP)
used by E. Eberlein's group; Truncated Levy Processes (TLP, constructed
by Koponen 4 , used by J.-P. Bouchaud and his group, and extended by Bo-
yarchenko and Levendorskii; and Normal Tempered Stable Levy Processes
(NTS Levy processes); we delegate the detailed discussion and references
to Chapter 3. As A.N. Shiryaev remarked, the name TLP was misleading,
and so we replace it with the name "KoBoL processes".
Processes of all the families listed above have been shown to fit better to
the dynamics of historic prices, and pricing formulas for European options,
based on these processes, also perform better than earlier models. Until
recently, almost no effective analytical formulas in more difficult situations
than the pricing of the European options and even simpler forwards and
futures have been known, though by now, there are many papers and a cou-
ple of books devoted to various aspects of the general theory of modelling,
pricing and hedging under Levy processes. The main goal of the book is to
partially fill in this gap, while remaining as close as possible to the initial
Merton-Black-Scholes framework.
1.2 Regular Levy Processes of Exponential type
When working with an empirical set of data, one specifies the type of the
process, that is, chooses a parameterized family of processes, and then
fits the parameters of the process to data. For a general theory like a
non-Gaussian analogue of the MBS-theory, too specific information is not
needed; in fact, unnecessary specification of the process makes the theory
harder to understand. There are two ways to describe the properties of the
process, which are sufficient but by no means necessary for our approach to
work, and the classes of Levy processes with exponentially decaying tails
used in empirical studies of Financial markets and described in Subsec-
tion 1.1.3 enjoy these properties. From the point of view of Probability
3
By semi-heavy or exponentially decaying tails we mean that the probability density
behaves, for x -> ±oo, as const • \x\p± e x p ( - c r ± | i | ) , for some p± € R and o± > 0.
4
Non-infinitely divisible truncated Levy distributions had been constructed earlier by
R.N. Mantegna
Regular Levy Processes of Exponential type 9
Theory, the most natural description comes in terms of the Levy density,
which can be visualized as the density of jumps of the process, but from the
point of view of PDO-theory and Analysis in general, it is more convenient
to work with the characteristic exponent of the Levy process.
1.2.1 Characteristic function and exponent, Levy measure

and Levy-Khintchine formula
Let X — {X t } t >o be a one-dimensional Levy process on a probability triple
(Q,T,P), and let E = Ep be the expectation operator under P 5 . The
reader can view fi as a set of all trajectories of the process, and the a-
algebra T of subsets of fi as a collection of all possible events, (fi, T, P )
is a filtered space, in the sense that T = Ut>o3~t, Fs C Tt for s < t, and
all Tt are a-subalgebras of T. Each Tt is generated by Xs, 0 < s < t,
that is, Tt is the smallest cr-subalgebra such that for any 0 < s < t and
- c o < a < b < +co, the preimage X s " 1 ((a,6)) is in Tt- Thus, Tt can be
viewed as the information revealed at time t (we know the values, which
the process X assumed up to the moment t).
Let X be a Levy process on R. Then the characteristic function of
the distribution of the random variable Xt can be represented in the form
E[el^Xt] = exp(-tV>(£))- The function ip is called the characteristic expo-
nent of X. By the Levy-Khintchine formula, xp admits the representation
2 r+oo
</-(0 = y£ 2 -n£ + y_ (l-e^+itxl^^Mdx), (1.3)
where a > 0 and 7 S R are constants, and II is a measure on R \ {0}

with /_^min{l,x 2 }II(da;) < +00. The parameters a2 and II, appearing
in Eq. (1.3), are called the Gaussian coefficient, and the Levy measure, and
the triple (<r2,7, II) is called the generating triplet. The density of II is called
the Levy density. If II = 0, the process is Gaussian, and if a2 = 0, the Levy
process is a pure non-Gaussian process, without the diffusion component.
Example 1.1 Let v € (0,2), and t/>(£) = c|£|". This is the characteristic
exponent of the stable Levy process of order v, and the Levy measure is
n(da?) = cV{-v)'l\x\-J-ldx.
5
For the list of basic definitions of Probability Theory, see Chapter 2
10 Introduction
For v = 2, xp(£) = |£|" is the characteristic exponent of the standard

Brownian motion, and II = 0.
1.2.2 Definition of RLPE in ID

Loosely speaking, a Levy process X is called a Regular Levy Process of
Exponential type (RLPE) if its Levy density has a polynomial singularity
at the origin and decays exponentially at the infinity. An almost equivalent
loose definition is: the characteristic exponent is holomorphic in a strip
Q^ € (A_, A + ), continuous up to the boundary of the strip, and admits the
representation
*KZ) = -irt + <KZ),

where </>(£) stabilizes to a positively homogeneous function at the infinity:
0(£)~c|f|", as £ - > o o , in the strip 3£ € (A_,A+). (1.4)
"Almost" means that though processes of BM, NIG, HP, GHP, KoBoL and
NTS Levy families satisfy conditions of both definitions, VGP satisfies the
conditions of the first definition but not the second one, since the character-
istic exponent behaves like constln |£|, as £ —¥ oo. For pricing of contingent
claims of European type, the additional property Eq. (1.4) is not essential,
but it is needed to obtain effective explicit formulas for the factors in the
Wiener-Hopf factorization formula, which we need in the study of perpet-
ual American options and barrier options. This is the reason why we will
mainly use the second definition. The adjective "exponential" needs no
explanation, and "regular" indicates that from the analytical point of view,
RLPE is the most tractable subclass of Levy processes, if the Brownian
motion is not available (notice that BM is an RLPE). We will call v the or-
der of the process, A_ and A+ the steepness parameters, and c the intensity
parameter of the process. The A_ (resp., A+) characterizes the rate of the
exponential decay of the right (resp., left) tail of the probability densities,
and c plays the part similar to the variance of the Brownian motion.
E x a m p l e 1.2 For v € (0,2], 5 > 0,a > |/?|, and /z e R,
HO = -il* + S[(a2 - (0 + i02T/2 - (a 2 - P2)v/% (1-5)

is the characteristic exponent of an RLPE of order v, with the steepness
parameters A_ = —a + /? and A+ = a + (3. With v = 2, we obtain
Regular Levy Processes of Exponential type 11
the characteristic exponent of the Brownian motion, and with v = 1, the

characteristic exponent of the model NIG. When v ^ 1,2, Eq. (1.5) gives
the characteristic exponent of an NTS Levy process.
Notice that if /x = 0, then in the limit a -> 0, Eq. (1.5) defines the charac-
teristic exponent of the stable Levy process.
1.2.3 Infinitesimal generators of RLPE as PDO

Let / belong to the space CQ (R) of twice continuously differentiable func-
tions vanishing at the infinity. Then for each x 6 R, there exists a limit
( ^ ^ M L M , (i.6)
and Lf is in Co(R), the space of continuous functions vanishing at the
infinity. The map / t-t Lf is called the infinitesimal generator of the
process X. The infinitesimal generator admits an explicit representation in
terms of the generating triplet:
Lf(x) = ^f"(x) + 1f(x)+ [+0°(f(x+y)-f(x)-lm<1}(y)f'(x))Tl(dy).

^ J — oo
(1.7)
By using the Levy-Khintchine formula, we derive from Eq. (1.7) the follow-
ing formula for the action of L on oscillating exponents: for £ € R,
L e «C _ _^(^) e ««_ ( 1-8 )
If / is sufficiently regular, for instance, / belongs to <S(R), the space of

functions of the class C°°(R), rapidly decreasing at the infinity with all
their derivatives, then we can decompose / into the Fourier integral
+oo
i x €
e /m, (1-9)
/
-oo
where / is the Fourier transform of / :
+oo
e-^f{x)dx.
/ -oo
By applying Eq. (1.8) to Eq. (1.9), we obtain

+oo
e"€(-tf(0)/(0#- (1-10)
/
-oo
12 Introduction
An operator of the form

+oo
efa{a(s,fl/(flde (1.11)
/
-oo
is called a pseudodifferential operator (PDO) with the symbol a. Thus, L is
a PDO with the symbol —ip, and therefore one can use the well-developed
machinery of the theory of PDO 6 .
In the theory of PDO, the properties of symbols of operators are crucial:
the more regular symbols are, the more explicit results can be obtained, and
here is the list of important properties, which characteristic exponents of
RLPE enjoy.
• From the point of view of the theory of PDO, Levy processes are
convenient since the symbols of the infinitesimal generators are in-
dependent of the state variable, x (in the PDO-language, these are
constant symbols), and therefore, many results can be obtained by
using the simplest tools of Complex Analysis.
• To study boundary value problems, explicit formulas for the factors
in the Wiener-Hopf factorization are needed, and Eq. (1.4) allows
one to derive relatively simple formulas. This explains why the
regularity condition in the definition of RLPE is important 7 .
• The derivatives of characteristic exponents of processes of model
classes of RLPE grow at the infinity slower than the characteristic
exponents themselves (cf. Eq. (1.5)); this property allows one to
generalize the class RLPE and construct a general class of Levy-like
Feller processes with good properties (see Chapter 14).
• Moreover, in empirical studies of Financial markets, the charac-
teristic exponents are characterized by a large parameter, like the
parameter a in Eq. (1.5), and so it is possible to develop approxi-
mate effective formulas. We will pursue this possibility in Chapters
7-8 and 10-11, where we study first-touch digital options, barrier
options, investment under uncertainty and endogenous default.
So, the characteristic exponents of RLPE's enjoy almost all desirable prop-
erties but one: the so-called transmission property, which ensures that a
solution of a regular boundary value problem, e.g., the Dirichlet problem,
6
Of course, all these definitions and constructions generalize to the n-dimensional case
7
Eq. (1.4) can be relaxed: see Chapter 15
Pricing of contingent claims 13
with smooth data is smooth up to the boundary; elliptic differential bound-

ary value problems satisfy the transmission property, the stationary Black-
Scholes equation being an example. If RLPE is not a Brownian motion,
the transmission property fails, and typically, a solution behaves near the
boundary point xo like ~ const\x — xo\K, where K € (0,1) is determined by
properties of the symbol. This observation indicates that near the bound-
ary (e.g., near the barrier, for a barrier option), the difference between the
answer the Gaussian model provides and the one, which an RLPE-model
gives, can be very large indeed. It also allows us to obtain asymptotic for-
mulas for a solution near the boundary. We use this possibility in Chapters
7-8 and 11, where we study the behaviour of prices of first-touch digitals
and barrier options near the barrier, and junk bonds.
1.3 Pricing of contingent claims
The aim of this section is to introduce the basic notions of Mathematical

Finance, namely, no-arbitrage assumption, equivalent martingale measure,
complete and incomplete markets, redundant securities and replication, and
hedging. First, we explain these notions in simple discrete-time models
with a discrete space of states, next describe briefly the situation in the
continuous time Gaussian Black-Scholes model, and after that discuss the
pricing and hedging in the Levy market.

No-arbitrage and equivalent martingale measures
Consider a two-period model of a financial market with n securities. Let
M be the number of possible states of the market tomorrow labelled by
OJI,. .. ,LJm, and the probability of the state uij is anticipated to be pj.
Thus, J^liLi Pj = 1> 0 < Pj < 1- The n securities are given b y n x m matrix
D, with Djk denoting the number of units of account paid by security j in
state k. Thus, j-row Dj is the vector of payoffs of security j . Let Sj be
the price of security j . Consider a portfolio 6 — (0\,..., 9n) € R", where
8j denoted the number of shares of security j . If 9j > 0 (resp., 9j < 0),
the investor is said to have a long position (resp., short position) in security
j ; both types of positions are allowed. Notice that 9j < 0 means that the
investor must deliver 9j shares of security j tomorrow. The portfolio 9
14 Introduction
has the market value S • 6 = Ylî Sj@j an^ payoff DT6. We say that a
portfolio 6 is an arbitrage portfolio if S • 6 < 0 and DT6 > 0, or 5 • 0 < 0 and
DT9 > 0. If an arbitrage portfolio exists, there exists an opportunity of
free lunches. In real financial markets, arbitrage opportunities may appear
but they are promptly eliminated due to the activity of arbitrageurs, who
make money by looking for those opportunities. Thus, the assumption of
no free lunches is sufficiently realistic, and as we will see, it can be used as
a cornerstone of the pricing theory.
Introduce an augmented payoff matrix
-S
TZ
DT
The no-arbitrage assumption implies that there does not exist a portfolio
6 such that 1Z6 > 0 and one of the components of 116 is positive. By the
separating hyperplane theorem, there exists a row vector A € R + + 1 such
that
XR. = 0. (1.12)
Since A is defined up to a positive scalar multiple, we may normalize to one

its first component. Define by A € R.++ the vector of the last m components
of A. Then the definition of 1Z and Eq. (1.12) imply together that
S = \DT. (1.13)
The A is called the vector of state prices. Notice that Eq. (1.13) deter-
mines prices but makes no use of the probabilities pj of the states of the
market tomorrow. In other words, for the no-arbitrage pricing, only the
information about possible future events but not their probabilities mat-
ter. Another interpretation is: if investors agree on the set of future events
(possible values of securities prices tomorrow) and there are no arbitrage
opportunities, they may disagree on probabilities of those events.
Suppose that our two-state model describes an investor who buys a
portfolio of securities at time 0, and liquidates (that is, sells) it at time
1, with no dividends paid in between. Then Djk admits a natural inter-
pretation as the price of the security Sj in state fc, at time 1. Denote
Sj(0) = Sj, and define a random variable Sj(l, •) on the probability space
fi = {wk | k — 1 , . . . ,m} by Sj(l,Uk) — Dkj- Assume further that one of
the securities, say, S„, is the riskless bond (usually denoted by B) yielding
a riskless return r; thus, B(l, CJ) — (1 + r)B(0). In this case, the last of the
equations in the system Eq. (1.13) is
m
J3(0) = ^ A f c ( l + r ) B ( 0 )
k=i
(notice that S in Eq. (1.13) is a row vector, so the equations are written
in a row), and therefore, vector q = (1 + r)X satisfies the following two
conditions:
m
0<qk<l, allfc; ^ 9 f c = l. (1.14)
fc=i
Thus, {<&}£!=! c a n be viewed as new probabilities of the states {uik}; let Q

denote the new probability measure on Q, and E^ the expectation operator
under Q. Introduce the discounted prices SJ(1,OJ) = (l + r ) - 1 S j ( l , u ; ) , and
rewrite Eq. (1.13) as the expectation under the new measure:
m
S(0) = £ > S * ( l , W f c ) := E Q [S*(1)]. (1.15)
k=l
The measure Q is called an equivalent martingale measure (EMM). In ex-

amples below, the n-th security is the riskless bond, and hence, Dn =
[l+r---l + r].
Example 1.3 Let n = m = 2. If the first security is risky, then d n ^ d\2,
hence D\ and Di are linearly independent, and the system Eq. (1.13) has
the unique solution: the EMM exists and it is unique.
Example 1.4 Let n = 2,m = 3. If the first security is risky, then D\
and D2 are linearly independent, therefore solutions to Eq. (1.13) exist;
since m > n, there are infinitely many of them. In this example, there are
infinitely many EMM.
Example 1.5 Let n = 3, m = 3. Suppose that £>i = aD2 (a £ R \ {0}).
Then generically, Eq. (1.13) has no solutions. In this example, there is
no EMM generically, and hence, it is possible to construct an arbitrage
portfolio. We leave the construction as an exercise for the reader.
In the case of a similar (T+l)-period model of n stocks paying no dividends,

the probability space tt consists of all trajectories of the n-dimensional pro-
cess 5 = {5(i)}o<(<T- For a fixed t, and a collection of subsets {-R(s)}o<s<i,
16 Introduction
R(s) e R " + , denote by A = A(R{0),..., R(t)) the set of trajectories, which

pass via R(s) at time s = 0 , . . . , t. Let Tt denote the subalgebra of the al-
gebra of subsets of fi generated by these sets of trajectories. Tt can be
interpreted as the algebra of events corresponding to the information avail-
able at time t. We have Ts C Tt, s < t; the collection T := {Tt} is called
the filtration, and fi is called a filtered space. Suppose that a probability
measure P is introduced on the measure space (fi, TT)- We assume that
any set in TQ has the probability 0 or 1, that is, the prices today are known
for sure.
An investor chooses an initial portfolio 6(0), and adjusts the portfolio
9(t) at each moment t = 1 , . . . , T - 1 so that 0(t - 1) • S(t) = 6{t) • S(t)
(we assume that she has no additional source of income). In making her
decisions, the investor takes into account the information available at time
t; hence, 0(t) is an (fi, JFt)-random variable; one says that a trading strategy
0 = {0(t)}o<t<T-i is adapted to the filtration T.
An adapted trading strategy 0 provides an arbitrage opportunity if and
only if one of the following two conditions is satisfied:
1) 5(0) • 6>(0) < 0, and S(T) • 0(T - 1) > 0 always, and
Ep[S(T)-0(T-l)] >0;
2) 5(0) • 0(0) < 0 and EP[S{T) • 6(T - 1)] > 0.
Assume that one of the securities is a riskless bond with the dynamics
B(t) = (1 + r ) * S ( 0 ) . Introduce the discounted price process
S*(t) = (l + r)-tS(t).
The same sort of argument as in the two-period model above shows that
if there is no arbitrage opportunity, then there exists a new probability
measure Q on the same measure space (U,TT) such that for all 0 < s <
t <T,
S*{s)=E<*[S*(t)\Ts}, (1.16)
where S Q [ 5 | Ts] denotes the conditional expectation of a random variable

5 given Ts. Since the set of events has not changed, the measures P and
Q are equivalent, and Eq. (1.16) means that the discounted price processes
are martingales under Q. We see that if there is no arbitrage, there exists
an EMM. It can be shown that the converse is also true: if there exists an
EMM, there is no arbitrage.
Example 1.6 In the multi-period setting, analogues of Examples 1.3-

1.4 are binomial model and trinomial model, respectively. In the binomial
model, at each time step, S(t+1)/S(t) assumes the value u with probability
p G (0,1), and the value d with probability 1 — p; in the trinomial model,
S(t + l)/S{t) can assume three values, with probabilities pj £ (0,1), j =
1,2,3.
In the binomial model, the EMM is unique, and in the trinomial, there
are infinitely many EMM. Clearly, the latter is more flexible in the sense
that it is much easier to adjust its parameters to data.

Completeness of the market, and pricing of derivative
securities
For simplicity, consider the two-period model. A contingent claim is a con-
tract with the specified payoff F(UJ) for each state of the market tomorrow.
The market is called complete if any contingent claim can be replicated,
that is, if lmDT = R m . In other words, for any contingent claim, one can
construct a portfolio (replicating portfolio) 6 such that FT = DT6. An
equivalent condition is: the payoff vectors of basic securities (that is, the
rows of the payoff matrix D) span R " \ which implies n > m. In other
words, the number of spanning securities is not less than the number of
future states of the world. In a complete market, Eq. (1.13) has a unique
solution, and hence, there exists a unique EMM. It is clear that EMM is
unique if and only if n > m and rankZ? = m is maximal. If the market is
complete, and n > m, we can choose a basis of R m from the rows of the
payoff matrix D. Suppose that the first m rows constitute a basis. Then
the first m securities can be used to span the others; thus, the latter are
redundant.
Example 1.7 In the two securities-two states model, suppose that the
payoff on the risky security in the first state of the world tomorrow is greater
than the one in the second state, that is, du > du- Introduce a call option
on the risky security, with the strike price K, du < K < du. An option
owner will exercise the option and buy 1-security tomorrow if the first of
the possible states of the world materializes, and will not otherwise. Hence,
the payoff row F ( l ) is [du — K 0], and the price F(0) of the call option
today is
F(0) = A • F(1)T = Ai (du -K) + A20 = A x (d n - K), (1.17)

18 Introduction
where A are determined from Eq. (1.13).
In a complete financial market, it is possible to perfectly hedge. To hedge

means to reduce risk against market fluctuations by making appropriate
transactions. In a complete market, the risk can be completely eliminated:
let F be a contingent claim, and 6 a replicating portfolio. Then the portfolio
(F, —6) is riskless: in each possible state of the world tomorrow, the payoff
of this portfolio is 0, and —8 is the perfect hedge for F.
If the market is incomplete, then rankD < n, and one can introduce
additional securities in order to obtain a complete market. The creation of
derivative securities, e.g., options of different kind, in real financial markets
may be seen as attempts to make the market more complete and increase
the possibility to hedge.
Example 1.8 Consider the two securities-three states model, which is

incomplete, and hence it is impossible to hedge against some financial risks
in the market: if FT 0 ImDT, then there is no hedge for F. Suppose that
îi > di2 > di3, and introduce the call option on the risky security, with
the strike price K, d i 3 < K < d 12 . Now we have the market with three
securities and the payoff matrix
dn d\2 ^13
D dn - K d\2 - K 0
l +r l+r l+r
The reader can easily verify that rankD = 3 if and only if d\\ ^ dw Thus,
in the case d n = d\2 the option is redundant, and in the case d n > di2 its
introduction makes the market complete.
Had the real financial markets been complete, there would have been no
need in the creation of derivative securities. This observation implies that
models of incomplete financial markets are more realistic than models of
complete markets.
In multi-period models, the situation is similar, only the role of a portfo-
lio is played by a trading strategy (another name: dynamic portfolio). For
details, see references in the review of literature at the end of the chapter.
1.3.3 Absence of arbitrage, EMM and completeness in the

Gaussian Black-Scholes model market
Under certain regularity assumptions, all of the properties listed above—the
no-arbitrage, completeness, in particular, possibility of replication of op-
tions and perfect hedge of options, and the existence of the unique EMM—
hold in the Black-Scholes market. The no-arbitrage is equivalent to the
existence of EMM. Moreover, it is possible to derive the formula for EMM
by looking for the perfect hedging strategy. So, the Black-Scholes model
has all the nice features one can imagine but it implies that there is no need
to introduce options at all: all of them are redundant.
The reader can learn the Gaussian theory of financial markets from
many excellent books - see the review of literature.
1.3.4 Sufficient condition for no-arbitrage in a Levy market

and incompleteness of a Levy market. The pricing
formula for contingent claims of European type and
the problem of a choice of EMM
Consider a continuous-time model market of a riskless bond, the riskless
rate of return being r > 0, and a risky stock. Suppose that the price of the
stock evolves as S(t) = expX(t), where X is a Levy process under the his-
toric measure8 P . From general results due to Delbaen and Schachermayer
(1994), it follows that the existence of EMM Q, which is absolutely contin-
uous with respect to the historic measure, is equivalent to the no-arbitrage
condition. So, as in the situations above, we can calculate prices by using
an EMM, but if X is neither the Brownian motion nor the Poisson process,
an EMM is not unique, and the market is incomplete. Moreover, typically
there are infinitely many different EMM; we discuss the restrictions on the
choice of EMM from a given class in Chapter 4. The first restriction is
quite universal; we will refer to it as the EMM-condition.
In the continuous-time models, the discounted price process is given by
S*(t) = e~rtS(t), therefore, by applying Eq. (1.16) with s = 0 to the riskless
bond and to the stock with the price dynamics B(t) — B(0)ert and S(t) =
S(0)expX(t), respectively, and using the definition of the characteristic
8
That is, the measure inferred from the observations of returns
20 Introduction
exponent, we obtain, for each t > 0:
5(0) = £ ( 0 ) e - ^ Q ( 0 ) , and 5(0) = S(0)e-t{r+^-^,
hence ^ ( 0 ) = 0, which is satisfied for any process without killing, and
r + ijjQ(-i)=0. (1.18)
We call Eq. (1.18) the EMM-condition. There are more subtle restrictions;
in particular, parameters c and v in Eq. (1.4) must be the same for the
historic measure and an EMM. Still, for any model class of RLPE, free
parameters remain, and one can introduce additional degrees of freedom
by considering mixtures of models processes. In Section 4, we produce
numerical results to show how one can change the price of an option and
the shape of the smile by playing with parameters of KoBoL.
Notice that if X is assumed to be Gaussian both under the historic
measure and an EMM, then the condition: c in Eq. (1.4) is fixed means
that the variance of the process does not change under the change of the
measure, and the EMM-requirement Eq. (1.18) fixes the drift (j, of an EMM:
r - / i - y = 0 ,
that is,
^Q(0 = y e 2 - i ( r - y ) € . (1-19)
The oldest variant of EMM is the Esscher transform, which have been
used in Actuarial Science for several decades, and in Financial Mathemat-
ics, from the beginning of the last decade: in terms of the characteristic
exponent, one looks for V Q m the form
^Q(0 = V'p«-^)-V'pH<?)) (1.20)

where 6 is real, and the EMM-requirement Eq. (1.18) leads to the equation
for 9
r + ipp(-i(6 + l))-ipp(-id) = 0. (1.21)
For a chosen EMM Q and an European option with the expiry date T and
the terminal payoff g(Xr), the pricing formula Eq. (1.16) can be written as
F(St,t) = (27T)-1 [exp[ixt ~ <r + tfQ(£))]$(« (1.22)

where x = \nSt, T = T — t, with the integration over an appropriate line

9f£ = a; a 6 (A_, A + ) is determined by the type of the growth of the payoff
g at the infinity. To deduce Eq. (1.22), it suffices to decompose g into the
Fourier integral
+oo
e-**tg{x)dx, (1.23)
/
-oo
substitute into Eq. (1.16) and use the definition of the characteristic expo-
nent.
Example 1.9 Consider a European call option with the strike price K
and the expiry date T. The terminal payoff is g(X{T)) = (e x ( T ) - K)+,
and the integral in Eq. (1.23) is well-defined for £ in the half-plane 3£ < — 1.
Take any a < - 1 , and calculate, for 9f = a:
r+oo
g(£) = / ( c (-«+i)* _ Ke-*')dx (1.24)
»€ - 1 %
_ Ke~^lnK
{% - m
~ (* + 0£ '
By substituting into Eq. (1-22), we obtain
2?r J_00+i<7 (t, + i)£
The reader may be tempted to decompose the integral in the RHS of

Eq. (1.22), as one does in the Gaussian case but she should not do that:
it makes any numerical integration procedure to perform worse, since the
integrand vanishes at infinity more slowly after the decomposition has been
made.
The integral in Eq. (1.22), Eq. (1.25) being an example, can be calculated
by using the Fast Fourier Transform (FFT) or its modifications; there are
cases when FFT performs poorly, and other methods must be used. In
some cases, the probability densities of probability distributions of a process
can be calculated explicitly, and then the pricing formula can be written
22 Introduction
essentially in the same spirit as in the Gaussian case: only instead of the
tabulated normal distribution another distribution is used.
When fitting the model to real data, the observed prices of the option are
compared with the ones calculated from Eq. (1.22). There are many papers
devoted to the derivation of the formula for EMM from the formula for the
historic measure by using different heuristic and economic arguments (see
a discussion in Chapter 4). We believe that the parameters of EMM should
be inferred from data on both the stock and options, and not derived by
some formal reasoning.
The perfect hedging in the Levy market is impossible, and an investor
can only try to minimize the risk, after some measure of the risk is chosen.
1.3.5 On pricing baaed on the utility maximization

The reader acquainted with Mathematical Finance has noticed that we
never mentioned the utility maximization of an investor. Apart from the
desire to make the book as short as possible, there is a more serious reason
for this omission. Though the concept of the utility of an economic agent
is very important for Economics and can be used to obtain many impor-
tant qualitative conclusions, it is hardly suitable for the realistic quantita-
tive analysis. In Economic literature, there is a host of utility functions
with very different properties, some criticism of the concept of utility op-
timization per se,- and the disagreement about time horizon of economic
agents, but even if we assume that the class of the utility functions for the
investor is known and the investor optimizes her life-time utility as it is
usually assumed (because it greatly simplifies the problem, not that it is
very realistic), one may notice that the utility is not directly observable,
and it is impossible to fit the parameters of the utility function to the data
with any reasonable degree of accuracy. The no-arbitrage assumption and
parameter fitting for EMM from a chosen class seems to be a much more
reasonable procedure, and traders are known to dislike the utility optimiza-
tion approach as well. On the other hand, one needs utility optimization
in models of financial markets with frictions, for instance, for option pric-
ing under transaction costs. The study of such situations goes beyond the
scope of this book, anyway.
The Generalized Black-Scholes equation 23
1.4 The Generalized Black-Scholes equation
1.4.1 The informal derivation

Let Q be an EMM chosen by the market. Consider a contingent claim
whose life span is a deterministic time interval [0, T], with terminal payoff
g. Let At be small, and t < T, t + At < T. We can apply Eq. (1.22) with
t + At instead of T and f(X(t + At), t + At) instead of g(XT):
+oo
e x p K " M(r + V> Q (0)]/«> * + A*)de, (1.26)
/
-oo
where / is the Fourier transform of / w.r.t. the first argument. By expand-
ing in the power series in At, dividing by At and passing to the limit as
At —¥ +0, we obtain
+oo
eixH-(r + rpQ(O)+dt)f(U)dt; = 0. (1.27)
/
-oo
By using the definition of PDO, we can write Eq. (1.27) as
dtf(x,t)-(r + rP^(Dx))f(x,t)=Q, t<T. (1.28)
The same result can be obtained in a more transparent form, by writing
Eq. (1.26) as
f(x, t) = exp[-At(r + ^(D))]f(x, t + At), (1.29)
and expanding into the power series in Eq. (1.29) rather than in Eq. (1.26).
If we add to Eq. (1.28) the terminal condition
f(x,T)=g(x), i£R, (1.30)

and specify the type of the behaviour at the infinity (say, the price f(x,t) of
the call option is bounded by ex, due to the evident no-arbitrage argument),
we obtain the well-posed problem Eq. (1.28), Eq. (1.30). By solving it, we
can recover Eq. (1.22) we started with. Certainly, there is not much sense
in this procedure, but for more complicated contingent claims, no explicit
pricing formula is known in advance, and so the reduction to an appropri-
ate boundary problem with subsequent solution by analytical methods is
reasonable.
Of course, to justify the formal derivation of Eq. (1.28), some regularity
conditions must be imposed. We will not discuss them here, because the
24 Introduction
analytical derivation, which we have outlined, is inferior to the derivation

based on Dynkin's formula or potential theory for Levy processes: the latter
can be applied to complex situations, where the life-time of a contingent
claim is stochastic, whereas the naive derivation is applicable only in the
case of the deterministic life-time.
Eq. (1.28) is the natural generalization of the Black-Scholes equation. To
see this, recall that Dx = —idx, and substitute the characteristic exponent
Eq. (1.19) into Eq. (1.28):
dtf(x,t) -(r-(r- y ) 3 a - y d 2 ) f(x,t) = 0,
or
2 / 2\
dxf
dtf + \dlf +[r-Y) ~rf = °- C 1 - 31 )
This is the Black-Scholes equation. In terms of the variable S = ex,
Eq. (1.31) assumes the standard form
2
ft/+yS2a|/ + r S 0 s / - r / = O. (1.32)
1.4.2 An outline of the reduction of the pricing of contin-

gent claims to boundary value problems for the gen-
eralized Black-Scholes equation: barrier options
For the first illustration, consider the down-and-out call option; this is an
example of a barrier option. Unlike a European option, the life-time of
a barrier option is random: should the price of the stock fall below the
specified value H (the barrier) before the expiry date T, the option expires
worthless or the option owner is entitled to some rebate gr(Xt,t). In the
Gaussian case, the trajectories of the process are continuous, and hence the
rebate must be specified at the barrier only but in the case of a Levy process
with discontinuous trajectories, the rebate must be specified for all values
of Xt = In St below the barrier h :— In H. We conclude that the price of the
down-and-out call option is the solution to the following boundary value
problem:
dtf(x,t)-(r + ^(Dx))f(x,t) = 0, t<T, (1.33)

x
f(x,T) = (e -K)+, x>h; (1.34)
f{x,t) = gr(x,t), x<h,t<T; (1.35)

x
f(x,t) = e + 0(l), x -+ +00. (1.36)
The last condition is justified as follows: if there is no barrier, we have

the call option, for which Eq. (1.36) holds. The influence of the barrier
decreases as the distance from it grows, hence the introduction of the barrier
can change only by the 0(1) term in Eq. (1.36).
So far, the derivation of the problem Eq. (1.33)-Eq. (1.36) was formal.
The rigorous derivation is based on one of the fundamental results of the
general theory of Markov processes: Dynkin's formula. First, we introduce
the early expiration region B := {(x,t) \ x < h}, next, we denote by T(B)
the hitting time of B by the two-dimensional process Xt = {Xt,t) (for the
definitions, see Chapter 2), and then we express the price of the option as
the sum of two expectations, the first one corresponding to the payoff in
the case of the early expiration, and the second one to the case when the
price of the underlying stock never crosses the barrier till the expiry date:
f(x,0) = E [e-""Wg r (X(T(B)) | X(0) = x, r{B) < T] (1.37)
+ E [ e - r T ( e x ( T ) - K)+ | X(0) = x, T{B) > T ] .
Let L = dt + L = dt — ipîDx) be the infinitesimal generator of the two-

dimensional process X. Dynkin's formula reads, for any stopping time r:
/(x,0) = E ( e-rs{r-L)f{X{s),s)ds\X(0) = x (1.38)

Jo
+ E[e~rTf(X(T),T)\X{0) = x].
Apply Eq. (1.38) with r = T(B)AT, and compare with Eq. (1.37). In view
of Eq. (1.34) and Eq. (1.35), we conclude that Eq. (1.33) must hold (this
requires justification: see Chapter 2). Thus, the naive way of writing the
boundary value problem for the price of an option is correct.
In many cases, for RLPE in particular, the justification is simplified
significantly by using the potential theory of Levy processes.
1.4.3 The case of interest bearing securities

Now suppose that during the life-time prior to expiry date r (random or
non-random), the owner of the security is entitled to a stream of revenues
26 Introduction
with the density g°(X(t),t). The typical examples are bonds of corpora-
tions: a typical bond pays some interest during its life-time; when the bond
matures, it pays the principal, and in the (random) event of the default,
the bond expires worthless or some amount is paid. In this case, the price
of the security is the sum of the stochastic integral which expressed the
present value of the stream of revenues, and the expectation of the payoff
ge(X(r),T) on the expiration date:
/(x,0) = Elf e-rsg°(X(s),s)ds \ X(0) = x (1.39)
+ £[e-rV(*(r),r) \X{0)=x].
If T is the hitting time of a closed region B c R x [0, +00), then by com-

paring Eq. (1.39) with Eq. (1.38), we obtain the boundary problem
dtf(x,t)-(r + ^(Dx))f(x,t) = g°(x,t), (x,t) e Bc; (1.40)

e
f(x,t) = g (x,t), (x,t)eB. (1.41)
1.4.4 The generalized Merton-Black-Scholes theory

The problem Eq. (1.40)-Eq. (1.41) is the general form of most of the prob-
lems which will be considered in this book, be they option pricing problems,
investment problems in Real Options theory (the problem of the capital ac-
cumulation is similar but more complicated) or pricing of defaultable bonds.
In each case, we know what the stream of revenues during the life-time is,
and what is the payoff at the expiry. The revenue density appears in the
RHS of the generalized Black-Scholes equation, and the payoff at the expiry
- in the RHS of the boundary condition. This procedure can be applied
formally though the rigorous treatment requires an additional justification.
As in the case of the barrier option below, usually the second expectation
in Eq. (1.39) admits a natural representation as a sum of two (or more)
expectations, and then the boundary condition Eq. (1.41) can be naturally
written as a system of boundary conditions. Usually, the first one comes
from the terminal condition (the payoff at the deterministic moment T),
and the second one represents the early exercise (or expiration) payoff.
Additional complications arise when the owner of the security can choose
the exercise time, that is, the region B (usually subject to some restric-
tions), the typical example being American options, but in such cases, we
also write down a problem of the form Eq. (1.40)-Eq. (1.41) and solve it
by analytical means. This is what we call the generalized Merton-Black-

Scholes theory.
1.4.5 Optimal stopping problems and the smooth pasting

condition
Consider the American put option, with the strike price K and the expiry
date T. If exercised, it gives the option owner the payoff g(x) = K — ex.
Options of the European type can be exercised only at the expiry date, but
the owner of an option of the American type has the right to exercise the
option at any time t € [0, T]. In other words, she chooses the optimal early
exercise boundary x = h(t) so that the option is exercised if and only if
at time t, Xt = lnS* reaches h(t) or falls below h(t), and in this case, she
receives K—ex. Introduce the exercise region B := {{x,t) | x < h(t),t < T}.
Thus, we have a free boundary value problem, when both the exercise region
B and the solution to the problem Eq. (1.40)-Eq. (1.41) must be found so
that the solution f{x,t) = f(B;x,t) be maximal. In the Gaussian theory,
one usually finds the candidate for the boundary of B by the so-called
smooth pasting condition (another name: the smooth fit principle). In the
case of processes with jumps, this principle may fail, and so we solve the
problem Eq. (1.40)-Eq. (1.41) for any candidate for the optimal exercise
boundary, and choose the right one by using the explicit formula for f(x, t).
Since the explicit formula is available only in relatively simple cases, we
manage to realize this program for perpetual American options, when the
time horizon is infinite, and for some approximate discretization procedures
for American options with the finite time horizon. Notice, however, that
even in the Gaussian case the explicit solution for American options with
the finite time horizon is not available.
1.4.6 The case of a dividend-paying stock

Suppose that the stock pays dividends at the constant rate A > 0. Then
the no-arbitrage argument shows that the process { e - ' r + A ' ' 5 t } must be
a martingale under the measure chosen by the market. In this case, we
prefer to use the name a risk-free measure instead of EMM. The treatment
of all the problems remains essentially the same, although in some cases,
the results change. In particular, in the no-dividend case, it is non-optimal
to exercise the American call option prior to expiry whereas in the case of a
28 Introduction
dividend-paying stock, there exists the early exercise boundary S = H* (i)

such that the exercise is optimal when the spot price St reaches the level
H*(t) or crosses it.
1.5 Analytical methods used in the book
1.5.1 The Fourier transform, and Complex Analysis

The Fourier transform is needed to write an analytic expression for the
price of contingent claims of European type, in the form of the oscillating
integrals; to calculate these integrals explicitly, simplify them or obtain
approximate formulas, standard tools from Complex Analysis—the Cauchy
theorem and the Residue theorem—are needed. All these tools are used in
more complex situations as well.
1.5.2 The Wiener-Hopf factorization and the Wiener-Hopf

equation
If the price of a claim is independent of time, the examples being perpetual
American options and infinitely lived bonds, the generalized Black-Scholes
equation becomes the stationary Black-Scholes equation, and we have to
solve the boundary-value problems of the type:
(r+iPQ(D))f(x) = g°(x), x > h; (1.42)

f{x) = gr{x), x<h. (1.43)
By introducing a new unknown u — f — gr, we reduce to the problem
(r + xPQ{D))u{x) = G(x), x>h- (1.44)

u{x) = 0, x < h. (1.45)
The problem Eq. (1.44)-Eq. (1.45) is called the Wiener-Hopf equation, and
it can be solved by the Wiener-Hopf factorization method. The following
scheme can be realized under fairly weak regularity assumptions on the
data and the symbol a(£) := r + V' Q (0; m the case of an RLPE, the latter
is sufficiently regular.
Step 1. Factorize a(£), that is, represent it in the form
a(0=a+(0M0. ^ R , (1-46)
Analytical methods used in the book 29
where a+ (resp., a_) admits the analytic continuation into the half-plane
9£ > 0 (resp., 9£ < 0), and does not vanish there. If a± and l / a ±
are polynomially bounded in the corresponding half-plane, the factors are
uniquely denned up to scalar multiples, due to the Liouville theorem, so in
some cases, they can be guessed.
The factorization problem can be solved for fairly wide classes of sym-
bols, by using analytical methods, and these constructions admit general-
izations for the multidimensional case, when one considers the Wiener-Hopf
equation in the half-space x„ > 0, and the analytic continuation is made
with respect to the dual variable £ n .
When ip® is the characteristic exponent of a one-dimensional Levy pro-
cess, the Wiener-Hopf factorization has an important probabilistic interpre-
tation in terms of the supremum and infimum processes Mt = s u p 0 < s < t Xs
and Nt =mf0<s<tXs:
r(r + ^ Q ( 0 ) _ 1 = ^ ( 0 ^ ( 0 , (1-47)
where
<!>+(£) =rE\f °° e-rte*M'dt (1.48)

r r+oo
4>-{Z)=rE\l e-rteiiNtdt (1.49)
Certainly, Eq. (1.48)-Eq. (1.49) are not explicit, and some additional efforts
are to be made to produce analytical formulas.
Step 2. The solution to the problem Eq. (1.44)-Eq. (1.45) in a natural
function class exists (for instance, this is the case if the data are smooth and
of compact support, and the solution is sought in the class of continuous
bounded functions); the solution is unique and given by
u = 0r-(D)l(h>+oo)0+(D)r-1/G, (1.50)
where IG is any sufficiently regular continuation of G from (h, +00) on R.

Step 3. Write down the RHS in Eq. (1.50) explicitly, by using the
definition of PDO, and calculate it by using tools of Complex Analysis.
We would like to make several general comments.
(1) Though the Wiener-Hopf equation is not widely known in Eco-

nomics and Finance, Steps 1-2 are easy to memorize and apply.
30 Introduction
(2) The third step requires the knowledge of some technique of the
reader, and in the end the final answer has to be computed numer-
ically by using some integration procedure.
(3) Fortunately, for wide classes of processes and parameters values
observed in real financial markets, effective approximate formulas
are available, which we demonstrate in the main part of the book.
(4) The very form of Eq. (1.50) allows us to derive an algebraic equation
for the optimal exercise price of the American perpetual options for
fairly general class of payoffs, and for the investment threshold in
Real Option theory; this equation is new even in the Gaussian case.
(5) The same form allows us to formulate the correct form of the Mar-
shallian Law and naturally separate the two factors which influ-
ence the capital accumulation. These results are important for
Economics, and are new in the Gaussian case as well.
(6) Finally, if the reader is willing to work with non-Gaussian Levy
processes, there is no hope to use differential equations as in the
Black-Scholes theory anyway: Steps 1-3 are the simplest general
scheme available.
In the case of American options with the finite time horizon, the explicit
analytical methods are not available, but after the discretization of time,
the Wiener-Hopf method can be applied as well.
1.5.3 The case of the non-stationary Black-Scholes equa-

tion and the constant barrier
Here the model example is the problem Eq. (1.33)-Eq. (1.36), with the
possible inclusion of a non-zero term in the RHS of Eq. (1.33). If the data
are independent of t, one can make the Fourier transform w.r.t. t (or the
Laplace transform, depending on the taste of the reader) and reduce the
problem to the family of problems on the line, which can be solved by
the Wiener-Hopf method. By making the inverse transform, we obtain the
result. In the case of more general payoffs, we introduce a new unknown
satisfying the homogeneous boundary condition, and consider the resulting
problem as the Cauchy problem on (—oo, T) with the data at t = T, for the
ordinary differential operator of the first order, with the operator-valued
coefficient. Once again, the problem can be solved by inverting a family of
the Wiener-Hopf problems and two integration procedures.
An overview of the results covered in the book 31
1.5.4 The case of the non-constant barrier and multi-asset

contracts
In this book, we consider only some special cases, when the reduction to one-
dimensional problems on the line is possible. In the theory of PDO, there
exists a well-developed machinery for handling multi-dimensional problems
of this sort, approximate and numerical methods including.
1.5.5 Pseudodifferential operators

In all parts of the book but one, only PDO with constant symbols arise,
and the corresponding part of the theory of PDO is, essentially, a part
of Complex Analysis. However, in the study of NIG-like Feller processes,
whose infinitesimal generators are PDO with non-constant symbols, all the
main ingredients of the theory of PDO are needed.
1.6 An overview of the results covered in the book
1.6.1 Elements of the theory of Levy processes

In Chapter 2, we list necessary definitions and results from the general
theory of Levy processes, and discuss in more detail the reduction of pric-
ing problems to boundary value problems for the generalized Black-Scholes
equation. In Chapter 3, RLPE are introduced, their main properties are
derived, and model classes of RLPE are compared. The properties of the
infinitesimal generators are discussed, and it is explained how one natu-
rally comes to the definition of the class RLPE by using "naive" PDO-
considerations. Explicit formulas for the factors in the Wiener-Hopf factor-
ization formula are obtained.
1.6.2 Option pricing

In Chapter 4, we consider contingent claims of European type. We discuss
the properties of the generalized Black-Scholes and the dependence of the
properties on the choice of EMM. We calculate prices of several types of op-
tions, and produce numerical examples to show how the prices and volatility
smiles depend on the choices of parameters of the model for the process and
EMM. We derive an explicit formula for the locally risk-minimizing hedging
ratio, which can be viewed as an analytical realization of M. Schweizer's
32 Introduction
idea of locally risk minimizing hedging. We show that for RLPE of order
less than 2, the hedging ratio is Holder continuous until expiry, even at
the strike, whereas the Gaussian delta-hedge is discontinuous at the strike,
at the expiry. This makes non-Gaussian RLPE-hedging much more stable
than the Gaussian one. We produce numerical examples and compare the
hedging ratio for different processes.
In Chapter 5, we consider perpetual American options. We formulate
the optimal stopping problem, and the corresponding free boundary prob-
lem. We explicitly solve the problem for a fairly general payoff g, and show
that in the case of the put options and similar more general options, the
optimal exercise price H = eh is determined from the equation
(^(D)-19)(h) = 0,
where <f)~ is determined from Eq. (1.49) with q = r + A, where A > 0 is the
dividend rate. In particular, for the put, the equation reduces to
/•+oo
[ / e->t+Nidt
where Nt is the infimum process.

In the case of calls, the exercise is optimal only if A > 0 (which is the
well-known fact), and then the optimal exercise price is determined by
r+oo
H = K<t>+{-i)=KqE e-Qt+M'dt
Jo '
where Mt is the supremum process. For more general call-like options, the
optimal exercise price is determined from the equation
(<P+(D)-1g)(h)=0.
By using these explicit formulas, we show that in some cases, the smooth
pasting principle fails, and discuss its generalizations.
In Chapter 6, we consider the American options with the finite time
horizon. The explicit formulas being non-available even in the Gaussian
model, we consider RLPE-analogues of several approximate methods used
in the Gaussian case. It is shown that the behaviour of the RLPE- price
of the American put near expiry drastically differs from the one in the
Gaussian case.
In Chapters 7 and 8, we derive explicit pricing formulas for touch-and-
out options and barrier options, respectively. We consider the asymptotics
of the price near the barrier, and explain why this is the place, where the
Gaussian model differs most from non-Gaussian ones.
In Chapter 9, we consider simplest of the multi-asset contracts, and
multi-asset hedging. We show that in the case of highly correlated assets
in the portfolio, the difference between the Gaussian (seemingly) riskless
portfolio and the RLPE-locally risk-minimizing one can be quite substan-
tial. Notice that the main cause for the Long Term Capital Management
disaster was the mispricing of risks, and the hedging model we suggest is
not much more difficult to implement than the Gaussian hedging.
1.6.3 Investment under uncertainty and capital accumula-

tion
In Chapter 10, we consider a risk-neutral, competitive, and value maxi-
mizing firm under demand uncertainty. The firm chooses optimal invest-
ment strategies; the investment is irreversible. This is a typical problem of
Real Options theory (for the Gaussian variant of the theory, see Dixit and
Pindyck (1996)). The problems of irreversible entry (resp., exit) from the
market are quite similar to the option pricing problem for perpetual calls
(resp., puts), if the possibility of the further expansion is ruled out. In the
model of sequential capital accumulation, additional subtle points arise.
About a century ago, Marshall suggested a rule that a firm should invest
as long as the present value of expected marginal revenue is not less than the
marginal cost of investment. However, as Dixit and Pindyck (1996) pointed
out, this rule did not take into consideration option-like characteristics of
investment opportunities. The methods we introduce allow us to restate
the Marshallian law as follows. Starting with the original price process,
define a new process, called the infimum process for the price of the firm's
output: Nt — info< s <tPt. Then the correct investment rule is: in the
formula for the profit function, replace the price process with the infimum
process started at the current level of the price and invest as long as the
present value of expected marginal revenue is not less than the marginal
cost of investment. This rule is applicable when the price can move in both
directions.
We also write down an analytic formula for the expected level of the
capital stock in terms of the infimum and supremum processes.
34 Introduction
1.6.4 Endogenous default and pricing of the corporate debt

In Chapter 11, we extend the structural approach to credit risk modelling
for the case of non-Gaussian processes. The central feature of this approach
is an attempt to model explicitly the evolution of the assets of the firm. We
focus on the choice of optimal capital structure and bankruptcy level by
a firm which keeps a constant profile of debt, but we model the evolution
of the assets of the firm as a Levy process. Notice that models of credit
risk relying on diffusion process for dynamics of the firms' assets cannot
capture the basic features of credit risk observed in practice. In particular,
empirical studies show that the credit spreads on corporate bonds are too
high to be matched by diffusion approach. Also, under a diffusion process,
firms never default by surprise because a sudden drop in the firm's value
is impossible. Therefore, if a firm is not currently in financial distress, the
probability of its default on very short-term debt is zero. This fact implies
that the credit spreads tend to zero as the maturity of debt tends to zero,
which contradicts empirical evidence.
We suggest using RLPE for problems of credit risk modelling and en-
dogenous default as more realistic than diffusion processes and almost as
tractable analytically as the latter. For a benchmark model, where the
debt has infinite maturity, we present analytical solutions for the endoge-
nous variables. In more complicated situations, an asymptotic solution may
be possible.
We study the firm near the bankruptcy and suggest two types of approx-
imate formulas for all endogenous variables. The first set of approximate
formulas can be used very close to the default level, and the second set
of approximate formulas is valid for Levy processes with large truncation
parameters, i.e., for the processes with not very fat tails, and the firm not
very close to the bankruptcy level. These approximations do not neglect the
effects of non-Gaussianity on one hand, and on the other hand, they allow
to perform comparative statics analysis. Using the approximate solutions,
we derive the optimal leverage for the firm. We produce several series of
numerical examples which show that if the firm's value is not very far from
the default level, the difference between the Gaussian and non-Gaussian
models can be sizable, and very close to the bankruptcy level - very large
indeed.
1.6.5 Numerical methods

In the literature on the non-Gaussian option pricing, one sees the state-
ments like: "The Fast Fourier Transform (FFT) can be used". In Chapter
12, we discuss the limitations of FFT and suggest a very fast method for
calculation of prices of European options under model classes of RLPE of
order v E (0,1], that is, NIG, and KoBoL and NTS Levy processes of order
v £ (0,1). We have not studied the new method for HP and GHP different
from NIG. For calculation of individual option prices, the method is several
hundred times (for options out of the money or deep in the money) faster
than FFT. For parameter fitting purposes, it is necessary to calculate the
prices of options for many values of the stock price, and here FFT with
its ability to calculate the prices at many points at once may have some
advantage. However, if one is satisfied with fitting of option prices for a
hundred values of the stock prices or so, which is usually the case, then the
new method is faster than FFT for processes of order less than one. For
options near expiry, it is several times faster; the relative advantage of IAC
decreases as the time to the expiry increases, and increases as the order of
the process v \. 0. For KoBoL of order 0.2, say, IAC can be 10-20 times
faster than FFT, and by using additional devices, it is possible to enhance
IAC further. For NIG, IAC is faster for options near expiry, and slower
for options 3-5 days to maturity or more. The IAC-method can be applied
for option pricing under Variance Gamma Processes as well, and in this
situation, IAC is especially effective.
IAC has the following additional advantage: it allows for a fairly effective
error estimates, which is impossible with FFT.
1.6.6 Extensions
In Chapter 13, we consider a discrete-time model, corresponding to perpet-
ual American options, that is, Bermudan options. From the analytic point
of view, there is not much difference with the corresponding continuous
time models, but the model becomes (infinitely) more flexible.
In Chapter 14, we consider the simplest extension to the case of Feller
processes: NIG-like Feller processes constructed by Ole E. Barndorff-Nielsen
and Levendorskii (2001). This is the first step for developing non-Gaussian
analogues of interest rate models.
36 Introduction
1.6.7 Basics of PDO theory

In the last two chapters, we provide the essentials of the theory of PDO
used in the book. This is useful for a reader interested in the systematic
exposition of technical tools used throughout the book, though we tried to
explain the most essential things in the main body of the book. There is
another aim of the chapter: a specialist with the PDE-background tends
to regard the Black-Scholes equation as a primitive object of Mathematical
Finance, and the exposition in Wilmott et al. (1993) is a model example
of this approach. The basic theory presented in Chapter 15 allows the
reader to calculate prices of contingent claims formally, without mentioning
stochastic integrals, thereby providing the background for PDE-approach
in non-Gaussian situations.
1.7 Commentary
The systematic exposition of the Gaussian theory of Mathematical Finance,

at different levels, can be found in many monographs. Dufne (1996) is a very
good book on the level intermediate between Economics and Mathematics;
Hull (2000) is, probably, the best professional book on derivatives; Karatzas
and Shreve (1997) and Shiryaev (1999) are excellent rigorous mathematical
texts (in addition, Shiryaev (1999) discusses various fundamental aspects
of non-Gaussian Mathematical Finance and empirical facts). The last two
books and Musiela and Rutkowski (1997), which reviews hundreds of pa-
pers and results, focus mainly on martingale methods, whereas Wilmott et
al. (1993, 1995) and Kwok (1998) use the PDE-approach and discuss rele-
vant numerical methods. The monograph Dixit and Pindyck (1996) is an
excellent exposition of the Gaussian Real Options theory for Economists,
in the PDE-framework (mainly, ID-case).
Non-Gaussian models (stable Levy processes) have been introduced to
Finance by Mandelbrot (1963); see also Fama (1965) and the collection
of papers Mandelbrot (1997). On truncated Levy distributions and their
application to empirical studies of Financial markets, see Mantegna and
Stanley (2000) and Bouchaud and Potters (2000). The last book contains
theoretical results on non-Gaussian pricing and hedging of European op-
tions (under processes of Koponen's family as well), futures and forwards,
and heuristic approximate methods for portfolio optimization and some
other problems.
Commentary 37
For stochastic volatility models see Zhu (2000), BarndorfF-Nielsen and

Shephard (2001c, 2001c, 2002), review paper Barndorff-Nielsen et al. (2001)
and the bibliography there. Notice that there is an overlap of classes of
SV-nodels and Levy models; see the discussion in Barndorff-Nielsen and
Shephard (2001a, 2001b, 2002) and Barndorff-Nielsen et al. (2001). It is
remarkable that in op. cit. (see also the bibliography there), the volatility
is driven by a Levy process of Ornstein-Uhlenbeck type, and nevertheless,
in some cases explicit pricing formulas for European options have been
obtained (which is not typical for SV-models even in the Gaussian case; see
however Heston (1993) and Duffie et al. (2000)). SV-model based on the
hyperbolic motion, was constructed in Eberlein et al. (2001). For modelling
of the evolution of the volatility surface, see Cont and da Fonseca (2001).
Chapter 2
Levy processes
In the first three sections of this chapter, we basically follow Bertoin (1996)
and Sato (1999) ([B] and [S]). The account is by no means complete, and the
chapter is not meant to substitute for any standard text on Levy processes.
We only aim to provide the list of basic definitions and results, which are
necessary should the reader want to go into the details of our constructions.
In Section 2.4, we consider the stochastic integral of a Levy process
stopped at the exterior of an open set, and deduce a boundary value problem
for the generalized Black-Scholes equation, which the integral solves. This
reduction is the cornerstone of the modified MBS-theory, and the main tool
in Chapters 4-11 and 14. The reduction is based on the Dynkin formula
and the representation of the q-order harmonic measure of a set relative to
a point in terms of the q-potential measure. It admits a generalization to
the case of a strong Markov process with absolutely continuous q-potential
measure.
2.1 Basic notation and definitions
2.1.1 Random variables

Let Q, be a set. A collection T of subsets of Q is called a a-algebra if the
following three conditions are satisfied:
(1) fi£f, 0 G T (0 is the empty set);

(2) If An € T for n = 1, 2 , . . . , then U£°=1 An and n£°=1,4„ are in JF;
(3) If A e F, then Ac G T.
39
40 Levy processes
A pair (Cl,!F) is called a measurable space. A map ft : T —> R + =

[0, +00) is called a a-additive measure if
1) /x[0] = 0, and
2) if An e F, n — 1,2,..., and A+nAj = 0 for i ^ j , then
»[UAj} = J2rtAn).
The triple (fi, .F, /i) is called a roeasune space. A measure space (il, T, fi)
and the cr-algebra J- are said to be p- complete if .F contains any B C fi
of outer measure 0: inf/x[{A G F\ A D B}] — 0. If .7* is not complete,
the completion of F" is the smallest u-algebra containing all subsets of fi of
outer measure 0.
The measure fj, is called finite if fi(Q) < 00, and er-finite if there is a
sequence An of elements of T such that fi(An) < 00, and L)An = fl.
Let (fi, J7, P ) and (Q',.F',P') be measure spaces. We say that / : Q, —>
ST is a {IF/!F')-measurable function or measurable function for short if for
a n y y l ' G - F , . T 1 ^ ' ) <E-T7-
Theorem 2.1 (The Radon-Nikodym Theorem) Let fj, and v be two a-

finite measures on the same measurable space (£2, .F).
Then the following two statements are equivalent
(i) for 4 e f , v (A) = 0 implies fi(A) — 0;
(ii) [i — fv, for some non-negative measurable function f.
If either of conditions (i)-(ii) holds, then the measure ^ is called absolutely

continuous w.r.t. v. One writes / = ^ and calls / a version of the Radon-
Nikodym density of fj, relative to v. If v is absolutely continuous w.r.t. (1
as well, fi and v are called equivalent measures.
If P[S"2] = 1, the triple (£l,F, P ) is called a probability space and P is
called the probability measure. Any set A G T is called an event, and T?\A\
is called the probability of the event A.
The cr-algebra generated by the open sets in R n , that is, the smallest
cr-algebra that contains all open sets in R n , is denoted by B(Rn) and called
the Borel a-algebra. A real-valued function on R™ is called measurable if
it is (yB(R")/S(R))-measurable. If P is absolutely continuous w.r.t. the
Lebesgue measure dx, then p = ^ is called the density of the measure P .
A mapping X from Q into R n is an R"-valued random variable (or
random variable on R n ) if it is (.F/5(R))-measurable. We write P[X G B\
Basic notation and definitions 41
for P[{w | X(u) e B}}. The map
B(R n ) 3 B H P[X G S] e [0, +00)
defines a probability measure on <B(Rn), which is denoted by Px(-B) and

called the distribution (or /aw) of X . Probability measures on B(R n ) are
called distributions on R ra .
Let X be a real-valued random variable. If the integral Jn X(w)P(cL/)
exists, then it is called the expectation of X and denoted by £[X]. UU e J7,
we write
£[X;[/] := E[Xlu] := /" X(w)P(dw).
If 5 is a measurable function and the integral
E[g(X)) = [ g(x)Px(dx)
exists, then it is called the expectation of g(X).

If X is a function on a measure space, and X(w) satisfies a property A
outside a set of 0 measure, then X is said to satisfy the property A almost
everywhere (a. e.); if the measure space is a probability space, one says
almost surely (a. s.) instead of almost everywhere.
Unless otherwise stated, a probability space is fixed, and random vari-
ables are denned on it.
Let T be a set, and {Xt}teT be a family of random variables. Then
a{Xt :t£T) denotes the smallest CT-subalgebra such that all Xt are a(Xt :
t S T)-measurable. It is called the cr-algebra generated by {Xt}teT-
Let (£2,.F, P ) be a probability space. Sub-cr-algebras T\% F2, • • • of T
are called independent if, whenever Ajk £ Fjk, and all jk are distinct,
P[Ajln-.-nAjn\ = P[Ah]...-p[Ajn}.
Let Xj be a R 7 ^-random variable, j'• = 1 , . . . The family {X,} is called

independent if cr-algebras cr(Xi),... are independent.
2.1.2 Conditional expectation

Let P be a cr-subalgebra of T, and let A" be a random variable with
£[1X1] < 00. Then there exists a random variable Y such that
(i) Y is ^"'-measurable;
42 Levy processes
(ii) E[\Y\] < oo;

(iii) for any U € F, we have E[Y; U] = E[X; U\.
Moreover, if Y is another random variable with these properties then
Y = Y, a. s., that is, P[Y = Y] = 1 (for the proof of these statements, see,
e.g., Rogers and Williams (1994), 11.39).
A random variable Y with properties (1)—(iii) is called a version of the
conditional expectation ^[X|J 7 '] of X given F, and we write Y = E[X\F],
a. s.
2.1.3 The Fourier transform and the Laplace transform.

Characteristic functions
We use the following definition of the Fourier transform, f = J-f, of a
function / :
f(0 = / e~i^f{x)dx, (2.1)
which is standard in the Uterature on PDO. The corresponding definition

of the Fourier transform of a measure /x on R n is
A(0 = / e-'^nidx). (2.2)
The integral in Eq. (2.1) is well-defined if / € L i ( R " ) , and the integral

in Eq. (2.2) is if the measure fi is finite: fJ,(Rn) < oo. In particular, the
Fourier transform is defined for a probability measure. If P y is absolutely
continuous w.r.t. the Lebesgue measure, and p is the density of P y , then
Py(0=P(fl.
The definition of the Fourier transform of a measure //, which is com-
mon in the probabilistic literature, coincides with the RHS in Eq. (2.2)
modulo sign and is called the characteristic function. We denote it by fi to
distinguish it from the Fourier transform. Thus,
MO = A(-0-
The function
E[e^'Y>] = f e'WPyidx) (2.3)
is called the characteristic function of the distribution of a random variable

Y, and denoted by P y . We have Py(£) = P y ( - £ ) -
Basic notation and definitions 43
1
The inverse Fourier transform T is denned by
/ ( * ) = (2TT)-» / e'«'*>/(Ode. (2.4)
^" and J r _ 1 are mutual inverses on the space 5 ( R n ) of infinitely differen-

tiable functions vanishing faster than any rational function at the infinity,
together with all derivatives, and extend by continuity to bounded mutual
inverses on L2(R"). By duality, the action of !F and that of J r _ 1 extend on
the space <S'(Rn) of continuous linear functional on S(R"); its elements
are called distributions or generalized functions.
Certainly, one may change the definitions of the Fourier transform and
its inverse and share the factor (2w)~n between T and T~x differently. Ifwe
move (27r) -n into the RHS of Eq. (2.1), then Eq. (2.4), our definition of the
inverse Fourier transform, becomes the definition of the Fourier transform
common in the probabilistic literature. This discrepancy in the definition
of the Fourier transform in these two fields of Mathematics will haunt us
in many places below.
The convolution of two distributions HI and \i2 on R n is the distribution
M = Mi * M2 defined by
H(B)= [ [ lB(x + y)^1(dx)ti2(dy). (2.5)
The dual measure to (i, denoted ft,, is defined by jl(B) = fi(—B), where
-B = {x\ -x£B}.
The next theorem lists several properties of the characteristic functions,
or equivalently, the Fourier transform of measures. For the proofs and
additional properties, see Section 2 in [S].
Theorem 2.2 Let fi, fix, fi2 be distributions o n R " .
(1) For any £ e R n , W) = A ( ~ 0 -

(2) The characteristic function of ft is /x(^).
(3) (Bochner's theorem) We have fi(0) = 1, |£(£)| < 1 for any £ G R n ,
and ft is uniformly continuous and nonnegative-definite in the sense
that, for each n = 1,2,..., and £j,£k 6 R n , r\j,% £ C",
n n
EE^-6)^>0- (2-6)
i=\ fc=i
44 Levy processes
Conversely, if (f>: R " —* C is continuous atO, nonnegative-definite,

and(j>(0) = 1, then 4>(£) is the characteristic function of a distribu-
tion on R n .
(4) 7 / A i ( 0 =fa{£)for all f 6 R n , then m = fi2-
(5) Ifn = fj.i* ii2, then /2(f) = /2i(0#2(0-
f£j J / X i ararf X2 are independent random variables on R", then
(7) Let Xj be R n J -random variable, j = 1,2, . . . , m , and A" = {Xj}.

Then X\,..., Xm are independent if and only if
Px{t) = PxAti)---Pxm(U),
for f = ( f i , . . . , f m ) e R n , w/iere n = m H h nm.
f#) / / // ftas a ./irate absolute moment of order m € N , i/iai zs,
/ |x| m /i(da;) < 00,
then jl is a function of the class Cm, and for any a £ (Z+) n ,

satisfying \a\ < m, the moment
{(i)a := / xa^dx) (2.7)
can be found as
(/*>„ = (-l)" a l(I> a A)(0). (2.8)
f5) If fi is of the class Cm in the neighbourhood of the origin, where m

is a positive even integer, then /i has the finite absolute moment of
order m. In particular, if (1 £ C°°, then fj, has finite moments of
any order, and they can be calculated from Eq. (2.8).
(10) If (1 e L j ( R " ) , then /t is absolutely continuous w.r.t. the Lebesgue
measure, and its density, call it g, is given by
g(x) = (2TT)-» / e«*-«/i(Ode.

•/R"
The convolution of two functions / and g, denoted f * g, is defined by
(f*9)(x)= f f{x-y)g{y)dy. (2.9)

Le"vy processes: general definitions 45
It is well-defined if, for instance, both / and g belong to L i ( R n ) , and a

straightforward calculation shows that
Rgt£) = /(0s(0- (2-io)

If /ii and fj,2 arc absolutely continuous w.r.t. the Lebesgue measure dx, and
their densities pi,P2 € L i ( R n ) , then (i is absolutely continuous as well, and
its density p is the convolution of the densities pi and p2-
The Laplace transform of a distribution on [0, +oo) is denned by
L M («) = f e~uxii{dx), V » u > 0. (2.11)

J\0,+oo)
Theorem 2.3 Let Hi,fj,2 and M ê distributions on [0, +oo). Then
(i) J/L M l (tt) = L M2 (u), for all u>0, then î = /i2-

fii,) If n = HI* H2, then L^(u) = L / J l (u)L / i 2 (u), /or aZZ 3Ju > 0.
fmj If Lp(u) admits the analytic continuation into a wider half-plane,
and for some a < 0,
/•a+ioo
(2m)-1 / |LM(«)|du < +oo, (2.12)
tfien // is absolutely continuous w.r.t. the Lebesgue measure on

[0,+oo), and its density is given by
g(x) = (liti)'1 / exuL^(u)du < +oo. (2.13)

Jo—ioo
2.2 Levy processes: general definitions
2.2.1 Levy processes and infinitely divisible distributions

A family X — {Xt}t>o of random variables on R " with parameter t £ [0, oo)
defined on a common probability space (Cl, F, P ) is called a stochastic pro-
cess. Xt and X t (w) are sometimes denoted by X(t) and X(t,u), respec-
tively. For u € Q, a map R+ 3 t (-• Xt(w) G R n is called the trajectory (or
sample path) of the process X . The index £ is usually (and in this book)
taken for time.
For 0 <to <ti < ••• <tm, the map
B((Rn)m) B x^Bj » P[X(h) 6 B i , . . .,X(tm) € flm] 6 R +

46 Levy processes
defines a probability measure on <6((R n ) m ). The family of all such measures

over all choices of m and 0 < t0 < ti < • • • < tm is called the system of
finite-dimensional distributions oi X. A stochastic process Y is called a
modification of a stochastic process X if
P[Xt = Yt] = l for t e [0, +oo).
Two stochastic processes X and Y (not necessarily defined on a common

probability space) are identical in law if the system of finite-dimensional
distributions are identical.
Definition 2.1 An R n -valued process {Xt}t>o is called a Levy process
(or process with stationary independent increments) if it has the properties
(1) a trajectory of the process is right-continuous on [0, oo), with left
limits at all t > 0, a.s.;
(2) for 0 < to < t\ < ••• < tm, the random variables X ^ X ^ -
Xto ,.-•,
Xtm — Xtrn_l are independent;
(3) Xo = 0 a. s.;
(4) the distribution of Xt+a — Xt does not depend on s;
(5) it is stochastically continuous, that is, for every t > 0 and e > 0,
limP[|X.-Xt|>e]=0.
s—>t
A Levy process on R™ is called an n-dimensional Levy process. If the first
condition is not required, X is called a Levy process in law.
Denote by /z m the m-fold convolution of a probability measure with
itself:
fim = n * • • • * fi (m times).
Definition 2.2 A probability measure /z on R " is called infinitely divis-

ible if, for any positive integer m, there is a probability measure fj,m on R n
such that fi — /u£J.
The collection of all infinitely divisible distributions is in one-to-one cor-
respondence with the collection of all Levy processes, when two processes
identical in law are regarded as the same.
Theorem 2.4 ([S], Lemmas 7.6 and 7.9, and Theorems 7.10 and 11.5)
(i) If [i is infinitely divisible, then there exists a unique continuous func-
tion <f>: R n —• C such that (f)(0) — 0 and exp[(/>(£)] = /*(£)/ for every t > 0,
Levy processes: general definitions 47
/i* = F~l exp[£0] is well-defined and infinitely divisible, and there is a Levy
process in law, X, such that Pxt = A**-
(ii) If {-Xt}t>o is a Levy process in law on R n , then for any t>0, Pxt
is infinitely divisible and Pxt = /•**, where /x = P x i •
(Hi) If X and Y are Levy processes in law on R n such that Pxt = P ^ ,
then X and Y are identical in law.
(iv) Let X be a Levy process in law on R n . Then it has a modification
which is a Levy process.
2.2.2 The Levy-Khintchine formula and generating triplet

From the last theorem, it follows that if X is an n-dimensional Levy process,
the characteristic function of the distribution of Xt admits the representa-
tion
£[ e ««.*t>] = e - ' * « ) , £ e R n , t > 0. (2.14)
The function ip is called the characteristic exponent of X . It characterizes

the law of X in the sense that two Levy processes with the same charac-
teristic exponent have the same law.
The Levy-Khintchine formula below describes all possible characteristic
exponents, hence all Levy processes.
Denote D = {x | \x\ < 1}.
Theorem 2.5 (i) Let X be a Levy process on R". Then its characteristic
exponent admits the representation
1>(0 = \(M,Z)-i('Y,Z)- [ (ei^-l-i(x,OlD(x)Mdx), (2.15)
where A is a symmetric nonnegative-definite nxn matrix, 7 £ R n , and II

is a measure on R n satisfying
n({0})=0, f (|x| 2 A l)U{dx) < 00. (2.16)
(ii) The representation Eq. (2.15) is unique.

(Hi) Conversely, if A is a symmetric nonnegative-definite nxn matrix,
7 G R", and U is a measure on R n satisfying Eq. (2.16), then there exists
a Levy process X defined by Eq. (2.14) and Eq. (2.15).
48 Levy processes
The triple {A, II, 7) is called the generating triplet of X. The A and II
are called the Gaussian covariance matrix and Levy measure of X. When
II = 0, X is Gaussian, and if A = 0, X is called purely non-Gaussian.
Notice that, essentially, the term —i(x,£)l£>(x) in Eq. (2.15) is needed
in order to ensure the convergence of the integral, and hence other functions
can be (and are) used instead of c(x) := 1 D ( X ) , for instance, c(x) = 1/(1 +
|x| 2 ); the A and II are independent of the choice of c. If II satisfies the
condition
n({0}) = 0, / (\x\ A l)IL(dx) < 00, (2.17)
which is stronger than Eq. (2.16), then Eq. (2.15) can be simplified
iKO = \(M,t) ~ i(T0,t) ~ I (ei<Xl€> - l)n(di), (2.18)
where
7o=7- / xlD(x)Yl(dx).
70 is called the drift of X.

If the sample paths of a Levy process have bounded variation on every
compact time interval a. s., we say that the Levy process has bounded
variation. A Levy process has bounded variation if and only if A — 0 and
Eq. (2.17) holds (see, e.g., [B], p.15).
2.2.3 Constructions of Levy processes

2.2.3.1 Explicit constructions
Some important classes of Levy processes, e.g. Hyperbolic processes, are
obtained by constructing a probability distribution and showing that it is
infinitely divisible. In Chapter 3, we will construct extended Koponen's
family and its analogue on a half-line by taking appropriate Levy measures
and making explicit calculations in Eq. (2.15). Once characteristic expo-
nents of some Levy processes are constructed, one can extend the list by
using the following two general devices. The subordination is the way Vari-
ance Gamma Processes, Normal Inverse Gaussian processes and Normal
Tempered Stable Levy processes are constructed, and the linear transfor-
mation of Levy processes can be used as an easy tool of construction of new
processes from given ones.
2.2.3.2 Subordinators and subordinated processes

([B], Chapter 3; [S], Chapter 6)
A suhordinator is a Levy process taking values in [0, +oo), which implies
that its trajectories are increasing. It allows interpreting a suhordinator as
a "random time". Subordination is a transformation of a stochastic process
through random time change by a suhordinator independent of the original
process. The resulting process is called subordinate to the original one.
The idea of subordination is due to Bochner (1949); the subordination can
be used to transform Markov processes into Markov processes, and Levy
processes into Levy processes.
Since the non-decreasing paths have bounded variation, the Gaussian
covariance of a suhordinator is zero, and the Levy measure satisfies
Eq. (2.17). Hence, the characteristic exponent of a suhordinator is of the
form
iKO = -*7o£ + / (1 - e fa «)n(cfa). (2.19)

Jo
The Laplace transform of the law of a suhordinator Z can be expressed as
E[exp(-XZt)] = exp(-t*(A)), (2.20)

where <& : R + —> R + is called the Laplace exponent of Z. Thus,
r+oo
$(A) = rp(i\) = 70A + / (1 - e-Xx)U{dx). (2.21)
Jo
Clearly, $(A) can be analytically extended into the upper half-plane 3JA > 0.
Example 2.1 Let a € (0,1), and
/•+0O
$(A) = AQ = - r ( - a ) - 1 / (1 - e-Xx)x~l-adx. (2.22)
Jo
Then Z is called a stable suhordinator with index a.
Theorem 2.6 Let Z be a suhordinator with the Laplace exponent $, let
Y be a Levy process with the characteristic exponent K, and suppose that Z
and Y are independent.
Define Xt(u)) = Yzt(w)(w), t > 0. Then X is a Levy process with the
characteristic exponent ~4>(£) — $ ( K ; ( £ ) ) .
50 Levy processes
For the proof and formulas for the generating triplet of V> in terms of the
drift and Levy measure of Z and the generating triplet of Y, see [S], The-
orem 30.1.
Example 2.2 By applying the stable subordinator with index a to the

Brownian motion with zero drift, we obtain the stable process on R", of
index 2a, with the characteristic exponent of the form c|£| 2 a .
2.2.3.3 Linear transformations of Levy processes

T h e o r e m 2.7 ([S], Proposition 11.10) Let X be a Levy process on R "
with generating triplet (A, II, 7), and letUbemxn matrix. Then {UXt}
is a Levy process on R n with generating triplet {Au,Tluilu) given by
Au = UAU', Uv = IK/" 1 ,
1U = U1+ f Ux(lDrn(Ux) - lDn(x))U(dx),
where UU'1 is the measure on R m \ {0} defined by (IK/ _1 )(.B) = U({x \

Ux € B}), and Dn is the unit ball in R n .
2.2.4 The Wiener-Hopf factorization

Let {Cl,F, P ) be a probability space, on which a one-dimensional Levy
process X is defined, and let CIQ be a subset of fi such that for each ui € flo,
the trajectory X.(LJ) is right-continuous with left limits. Define, on £2o,
Mt = sup 0 < J ! < t Xs and Nt = info<s<t Xs. On fi \ Clo, both Mt and Nt are
set to be 0. M = {Mt} and N = {Nt} are called the supremum process
and the infimum process, respectively.
The Laplace transform (in t) of the distribution of Xt or to be more
precise, the function
r+00
qiq + m r ^ l e-«E[e*x']dt,
Jo
can be factorized by using the Laplace transforms (in t) of the distributions
of the supremum and infimum processes. Among many factorization iden-
tities, we will use only the simplest one (Rogers and Williams (1994), p.81;
[S], Theorems 45.2 and 45.5; for more detailed exposition, see [S], Section
45).
Theorem 2.8 (i) Let q > 0. There exists a unique pair of infinitely
divisible distributions p+ and p~ having drift 0 supported on (—oo, 0] and
[0, +oo), respectively, such that their Fourier transforms <j>+ and <j>~ satisfy
qiq + mV1 = #W7(0, £GR" (2-23)
(ii) The functions <f>+ and (j>~ admit the following representations
r+oo r+oo
<t>t(£.)=q e-qtE[e^Mt}dt =q e-gtE[ei(<-Xt-Nt)}dt, (2.24)
Jo Jo
f+oo r+oo
<t>q{€)=q\ e-Ê[e^N*\dt = qj e-Ê\e^Xt-Mt)\dt, (2 25)
Jo Jo
and
r+oo r+oo
<£+(£) = exp / t-xe~qtdt I (eixii - l)ii*{dx) (2.26)
Jo Jo
r+OO rO
$<, ( 0 = exp / t-le~qtdt / (eix« - l)n*(dx) (2.27)

Jo J-oo
Notice that <£+(£) (resp., <^"(£)) admits the analytic continuation into the
upper half-plane Q£ > 0 (resp., lower half-plane S£ < 0) and does not van-
ish there. Thus, Eq. (2.23) is a special case of the Wiener-Hopf factorization
introduced in solving integral equations by Wiener and Hopf (1931), and
widely used in the theory of boundary value problems for PDE and PDO.
The formulas Eq. (2.24)-Eq. (2.25) are by no means explicit though
very convenient for theoretical considerations, and Eq. (2.26)-Eq. (2.27)
are not effective, either. Simple analytical formulas can be obtained for
special cases only.
Example 2.3 Let X be a Brownian motion with the drift 7 and variance
a2. Then the characteristic exponent is
m -r - i-rt.
It is clear that for q > 0, the equation q +1/>(£) = 0 has two roots —i/?_
and -i/3+ in the upper and lower half-planes, respectively, and therefore,
52 LSvy processes
X
q(q + V>(£)) admits the factorization Eq. (2.23) with
Clearly, (/>" is the Fourier transform of the exponential distribution with

parameter — /?_, and <f>t is the Fourier transform of the dual to the expo-
nential distribution with parameter /?+.
Example 2.4 Let X be a spectrally negative Levy process, i.e., a Levy

process having no upward jumps. Then the law of {Mt} is exponential (see,
e.g., [B]), and therefore, for some ft = j3(q) > 0,
<t>t(0 = P/(P~iO- (2-29)
Since 4>^(Q has no zeroes in the lower half-plane 3?£ < 0, Eq. (2.23) implies
that — i/3 is the root of q + ip(£). After this root is found, which is a simple
numerical procedure, we can calculate 0+ from Eq. (2.29), and after that
4>~ from Eq. (2.23).
The reader can consider spectrally positive Levy processes by using the
symmetry argument.
In Chapter 5, we will obtain fairly simple analytic expressions for the

factors in Eq. (2.23), for Levy processes of exponential type, and additional
representation for regular Levy processes of exponential type. For model
processes, we will also derive approximate formulas with the leading terms
of the form Eq. (2.28).
2.3 Levy processes as Markov processes
2.3.1 Markov property and Feller property

Informally, a Markov process models the memoryless evolution of the price
bundle X(t) = (Xi(t),.. .,Xn(t)) of n stocks in the market, with the fu-
ture behaviour of the bundle depending on the past only through its current
value. The property of spatial homogeneity of Levy processes yields signif-
icant simplification of the theory.
To give a rigorous definition for the past at a given time, it is convenient
to introduce the filtration {Ft}te{o,oo] in the cr-algebra J7, that is, for t <
+co, Ft is the P-completed cr-subalgebra of T generated by (Xs, 0 < s < t),
Levy processes as Markov processes 53
and ^Foo stands for the P-completion of T. Thus, Ts C Tt if s < t- The

cr-subalgebra Tt models the information available at time t.
Prom the definition of a Levy process, it follows that for every s > 0, the
process X' defined by X't = Xt+a—Xa is independent of J-a. Moreover, the
finite-dimensional representations of X' are the same as that of X. These
two statements taken together are referred to as the simple Markov property
of Levy processes. Prom this property, one deduces that the filtration {.Ft}
is right-continuous, that is J-t — C\a>tFa, and the Blumenthal zero-one law
holds: for any A G F0, P[A] = 0 or 1.
For every x G R", denote by P x the law of X + x under P , that is, the
law of the Levy process started at x:
Px[Xtl G Bu... ,Xtm G Bm] = P[x + Xtl G Bu... ,x + Xtm e Bm),
for 0 < ti < ••• < tm and Bi,...,Bm G S ( R n ) . Thus, P ° = P . Set

x
E [f(Xt)] := E[f(Xt)\Xo = x], and introduce the family of convolution
operators on Loo(R"), indexed by t > 0:
Ptf(x) := E*[f(Xt)} = / f(x + y)PXt(dy).
We will use this notation in the case of any measurable non-negative / as

well; in this case, Ptf(x) = +oo is possible and admissible.
The family {-Pt}te[o,+oo) i s a Markov semigroup, in the sense that Po = I
(the identity operator), PsPt = Pt+S for every t, s > 0, and 0 < Ptf < 1
whenever 0 < / < 1.
Denote by Co(R") the space of continuous functions on R n vanishing
at the infinity. The semigroup {Pt}te[o,+oo) °f a Levy process has the Feller
property, that is, for every / G Co(R"),
(i) Ptf G C 0 ( R n ) , for any t > 0;
(ii) limtô Ptf = f (uniformly).
([S], Theorem 31.5). The same two statements hold if, instead of Co(R"),
we use Co(R n ), the subspace of L ^ R " ) consisting of functions vanishing
at the infinity ([B], Proposition 1.1).
2.3.2 Stopping times and the strong Markov property
The simple Markov property can be reinforced by allowing for certain ran-
dom times, that is, random variables with values in [0, +oo] instead of the
deterministic time t.
54 Levy processes
A random time T is called a stopping time if for every t > 0, the event
{T < t} belongs to the <r-algebra Tt- Since the filtration is right-continuous,
we have that T is a stopping time if and only if {T < t) g Tt for alH > 0.
For an arbitrary random time T, we denote by TT the P-completion of the
tr-algebra generated by the process X killed at time T, X o k r , 1 and the
variable 1{T<+<X>}XT-
It can be checked that when T is a stopping time, !FT consists of events
A such that {T < t} n A G Tt for all t > 0. (Intuitively, a random variable
T is a stopping time if one can decide whether T < t or not by making
observations of the evolution of X up to the moment t; A 6 TT if A is an
event expressible by observations up to time T).
T h e o r e m 2.9 (The Strong Markov property). Let T be a stopping ime

with P [ T < +oo] > 0. Then conditionally on {T < +oo}, the process
{Xr+t — Xr}t>o is independent of TT and has the law P .
By definition, a sample path of a Levy process has left limits at all

times, a. s. From the Feller property of the semigroup, one deduces a much
stronger property called quasi-left-continuity.
T h e o r e m 2.10 Let {T„} n6 N be an increasing sequence of stopping times

with limn-KxjTn = T a.s. Then limn-too-Xr„ = XT as. on {T < oo}. In
particular, ifTn < T a.s., then X is continuous at time T a.s. on {T < oo}.
Two important classes of examples of stopping times are provided by

the first entrance time into a set B C R", Tg := inf{t > 0 | Xt e B}, and
the first hitting time of B, T'B := inf{t > 0 | Xt € B}.
L e m m a 2.1 ([B], p.22)

a) Let B be open or closed. Then
(i) both TB and T'B are stopping times;
(ii) XTB € B a.s. on {TB < oo}, and XT> € B a.s. on {TB < oo};
b) TB = T'B Px-a.s. for every x € BCUB°; in particular, PX[TB = T'B\ = 1
for every x, if B is open.
Note that if B is closed, the last equality may fail for some x on the bound-
ary of B.
J
kt denotes the killing operator, ktw(s) = w{s) if s < t, and d otherwise, where d denotes
the cemetery point of the extended probability space fi U d (for details, see [B], Section
0)
2.3.3 Resolvent operator and infinitesimal generator

Let q > 0. The resolvent operator, or q-potential operator Uq, is defined for
every measurable non-negative function / by
/•oo r /-oo
Uqf(x) = / e-«*P t /(i)dt = Ex / e - * 7 ( X ,)dt
The probabilistic interpretation is as follows: if r = r(q) is the random time

having an exponential law with parameter q > 0, and r is independent of
X, then £ [ / ( * + XT)] = qUqf{x), for all x.
The family of finite measures (Uq(x,dy), x e R n ) associated to the
resolvent operator by
Uqf(x) = f f(y)U«(x,dy)
is called the resolvent kernel, or q-potential kernel. Clearly, for B € B(R n ),

Uq(x, B) = Vq(B — x), where Vq(B), the q-potential measure, is given by
Vq{B) = E / e-qt lB{Xt)dt = (unB)(o).

Jo
Hence, Uq is the convolution operator:
Uqf = Vq*f. (2.30)
Prom the semigroup property of {Pt}t>o, the resolvent equation follows:
Uq - UT + (q - r)UqUr = 0, Vq,r>0,
and the resolvent equation implies that the image of Co := Co(R-") under Uq
does not depend on q. We denote it by V. The Feller property implies that
for every function / € Co, limq-^+oo qUqf = / in the uniform topology.
Thus, V is dense in Co- By using the last observation and the resolvent
equation once again, one deduces that the map Uq : Co —> V is a bijection.
We define the infinitesimal generator L : T> —* Co as the inverse to Uq :
Co -+ V. Thus,
Uq(q-L)f = f,Vf€C0, (2.31)
and
(q - L)Uqf = f, V / € V. (2.32)
56 Levy processes
We have, for every t > 0, / € L i ( R n ) n Loo(R n ) and £ € R n ,
F(Ptf)(t) = E
VR"
= E f emxt-v).e>f(y)dy
.JRn
= E J{Xt,S)
JR"
Hence, by using the definition of the characteristic exponent, we conclude

that
HPtfW = e-**<«.F/(0, t > 0, (2.33)
and therefore,
(2.34)
which means that Uq is a PDO with the symbol (<Z + ^>(0)~ x :
Uy = {q + ^{D))-lf. (2.35)
We have shown that Eq. (2.35) holds for / € £ i ( R n ) n L M ( R n ) . In some

situations of interest, / is unbounded, but the following condition holds:
(9 + S t y ) - 7 e L i . (2.36)
Lemma 2.2 If f G L\ and Eq. (2.36) holds, then Eq. (2.35) is valid.
Proof. Since / e L\, we have
(U"f)(x) = / e-qt(Ptf)(x)dt
Jo
/•+oo p
= / e-«*(27r)- n / e- < < I ' € >-iK«/(0d£<ft.
JO JR»
Due to Eq. (2.36), the last integral computed in the reverse order dtd£
converges absolutely, and hence we can apply the Fubini theorem and obtain
Eq. (2.35). •
Finally, for / € V satisfying Lf e L i ( R n ) , we deduce
?(Lf)<£) = -V>(0.F/(O, (2.37)

and then
Lf(x) = (^(-iKOWX*) (2-38)

n i(x
= (2n)~ [ e ^(-V(6)/(0^,
which means that L — —ip(Dx) is a PDO with the symbol —if).

By using the same argument as in the one-dimensional case in Introduc-
tion, one can show that for / G <S(Rn), the definition Eq. (2.38) coincides
with the classical definition
d2f
Lf{x) = \T,aik. + <7,/'(*)) (2.39)
J
3,k=l
+ f (f(x + y)-f(x)-lD(y){y,f'(x)))U(dy).
2.3.4 Dynkin's formula

Let T be a stopping time. From the strong Markov property, one deduces
Dynkin's formula: for g G Co := Co(R n ), q > 0 and x G R",
U"g(x) = Ex g(Xt)dt + Ex [e-gTUqg(XT)] • (2.40)

Jo
The stochastic integral in Eq. (2.40) is understood as
r+oo
Ex / e-**l { t < r } («)$(X t )<ft
Jo
Eq. (2.40) is also valid for any non-negative universally measurable function
/ - see [S], Eq. (41.3).2 It follows that Eq. (2.40) holds for g G Li + C 0 .
Lemma 2.3 Let (q — L)f := (q + ip(D))f belong to the class Li + C0 and
satisfy Eq. (2.36). Then
f(x)=Ex f e-gt(q-L)f(Xt)dt + Ex[e-«Tf(XT)}. (2.41)

Jo
Proof. Apply Eq. (2.40) to g = (q - L)f. Due to Eq. (2.36), Eq. (2.35)
holds, and hence, U^(q - L)f = f. Thus, Eq. (2.40) becomes Eq. (2.41).•
2
In situations which we consider in this book, only Borel functions will arise, so there is
no need to give the definition of a universally measurable function here.
58 Levy processes
2.3.5 Duality
The process X — —X is called the dual process of X; it is a Levy process
for the same triple (CI, T, P ) . The sign~is used to denote the ^-resolvent Uq
and the infinitesimal generator L of X, etc. For x e R n , P 1 denotes the
law of x + X under P , that is, the law X under P _ x . It is clear that ifi is
the complex adjoint to tp: $(£) = VK-£)>£ € R n , and hence, L = —ip(-D).
The name "dual" stems from the following proposition.
Proposition 2.1 Let f and g be non-negative measurable functions. We

have for every t>0
/ Ptf(x)g(x)dx = / f(x)Ptg(x)dx,
JR" JRn
and for every q > 0,
/ Uqf(x)g(x)dx = / f(x)Uqg(x)dx.
JR" JR"
/ / / is in the domain of L, and g is in the domain of L, then
/ Lf(x)g{x)dx = I f{x)Lg{x)dx.
JR" ./R"
Proof. The first statements is obtained by a change of variables, see [B],
Proposition II. 1.1 and [S], Proposition 41.7, and the last two follow. •
2.3.6 Absolutely continuous resolvents

We start with a modification of the statements on p.25 in [B] and on p.p.288-
289 in [S].
Theorem 2.11 The following statements are equivalent:
(i) For some q > 0, the measure V is absolutely continuous w.r.t. the
Lebesgue measure;
(ii) For any q > 0, the measure Vq is absolutely continuous w.r.t. the
Lebesgue measure;
(Hi) For some q > 0, and some x € R n , the measure Ug(x,dx) is
absolutely continuous w.r.t. the Lebesgue measure;
(iv) For every q > 0, and every x 6 R n , the measure Uq{x,dx) is
absolutely continuous w.r.t. the Lebesgue measure;
(v) The resolvent operators have the strong Feller property, that is, for
every q > 0 and f G Loo(R n ), the function Uqf is continuous.
When the assertions of Theorem 2.11 hold, we say that the resolvent kernel
is absolutely continuous, or that the (ACP)-condition holds. The following
Lemma gives convenient sufficient conditions which are satisfied by Regular
Levy Processes of Exponential type.
Lemma 2.4 Let an n-dimensional Levy process X satisfy (ACT)-condi-
tion, that is, the transition measures Pxt are absolutely continuous for all
t > 0. Then (ACP)-condition holds for X and for (n + 1)-dimensional
process {(Xt,t)}.
Proof. Let pt be the density of PXt, and B x [a, b) e B(Rn x R + ) . Then
Vq(B x [a,6)) = / / e-qtpt{x)dxdt,

J JBx[a,b)
which proves (ACP)-condition for {(Xt,t)}. The proof for X is similar. •

Let (ACP)-condition hold. Then Eq. (2.30) can be written as
U"f{x)= f vq(y-x)f(y)dy = (vq*f)(x), (2.42)
where vq is the density of Vq, and vq(x) = vq(—x) is the density of Vq.
2.3.7 Operators P%
n
Let B C R be an F a -set (for the general definition, see [S], p.279; for our
purposes, it suffices to notice that Borel sets are Fo-sets), and let T'B be
the hitting time of B by X. For g € Loo(B), define
Plg{x):=Ex[e-qT'Bg{XTlB)]. (2.43)
The map g i-> Pgg(x) defines the measure which is called the g-order
harmonic measure of B relative to x. By gathering Proposition 42.13,
Theorem 42.5 and Definition 42.6 from [S], we obtain
Lemma 2.5 Let B be an Fa-set. There exists a a-finite measure dfis
supported on B such that
PqBl(x)= f vq(y-x)d»B(y).
60 Levy processes
Prom Lemma 2.5, we deduce for a closed B:
(P«Bl,(q-L)w)L2=0, VweC™(Bc), (2.44)
where Bc := R n \ B, L = —•ip(D) = —ip(-D) is the generator of the dual

process, and L 2 = ^2(R- n ;R-) is the real space.
We need a more general result:
L e m m a 2.6 Let B C R™ be closed, and g be a bounded measurable func-

tion. Then
(Py, (q - L)w)L2 = 0, V w € C%°(BC). (2.45)
Proof. We derive Eq. (2.45) from Eq. (2.44) by using an additional as-
sumption which holds for RLPE's: the characteristic exponent ift of X
admits the analytic continuation into a tube domain R n 4- iU, where U is
an open subset of R n , containing 0.
For 7 G R n , define a function u 7 by u7(x) = e^1,x^ . As the first step of
the proof, we show that Eq. (2.45) holds with g = u 7 provided q+ipî^) >
0. Change the probability measure: P 7 = e ^ ^ H ' ^ - n O p . Under the new
measure, X is the Levy process with the characteristic exponent ip7(£) —
VK£ ~~ *7) — iJ{~i'y) since
Epi [e*x<] = E* 3 tV(-i7)+»<e-»7,Xt) = exp[-t(ip($ - i-y) - ip(-i'y))}.
Let Ef, {-P7,t}t>o a n d L-y be the corresponding expectation operator, semi-

group and infinitesimal generator. We have
e~^'x\Ptung){x) = [ e-{^xh(l'x+y)g(x + y)P(Xt£dy)
= e-t+(-i-r) f g(x + y)e<-y^+t^-'i^P(Xtedy)

JR"
e-^(-^)(P7,t5)(x),
therefore
,-{q+H-i-l))T'B 9+V>(-t7) 1(2:).
e-b'*) ptuyix) = E% =:Pi 7,B
Since q + ^(-17) > 0, Eq. (2.44) applies to i ^ J * ( ~ ' 7 ) l and L 7 :
(P^{-hh, (q + ^(-»7) - I 7 H = o, v w € c0°°(5c),

or equivalently,
(u-'P^, (q + V-H7) - L^w') = 0, V w ' e C0°°(BC),
and finally,
1
( P | u 7 , (q - u ' ^ - V H 7 ) K K ) = 0, V «,' e C0°°(SC).
We have L 7 = —ip(D — i-y) + i>(—ij), therefore
iLy(L7 - V'C—ry))"^1 = -u^(D-ij)u~l
= -V>(#)
= i,
and hence -u~ 1 (L 7 — T/>(—i7))u7 = L. Eq. (2.45) with g — u7 has been
proved.
For the next step, consider a compact B and an arbitrary bounded mea-
surable g. We can approximate g in the L°°-norm by continuous functions,
and each continuous function by polynomials (the Stone-Weierstrass theo-
rem). Fix e > 0 such that the ball V(e) := {7 | ||7|| < e} C U. Since for
any multi-index a,
iimon7-^(e^-ir-n^,
3=1 3=1
uniformly in x G B, we can approximate g in the -L°°-norm by a sequence

of functions of the form
gN(x)= Yl cNi~/e{'y'x),
-7€iv
where TN C V(e) is finite. Since Eq. (2.45) is valid for g — u 7 , provided

—7 € U and q + ip(-i'y) > 0, it holds for g = gw provided e > 0 is
sufficiently small. By passing to the limit e —+ 0, N -* +00, we conclude
that Eq. (2.45) holds for any bounded measurable g and compact B.
It remains to drop the assumption that B is bounded. Notice that it
suffices to consider non-negative g. For R > 0, set B(R) — B n V(R). We
have
Ex e-*TB^g{X{TB{R)))]<E* [e-' r *ff(X(r B ))l+|| S ||ootf g 1TB\B[R)

62 Levy processes
and
Ex
'e-iT'Bg(X(TB))\ <EX \e-^*™g(X(rB{R)))\+\\g\\o0E' g 1TB\B(R)
For a Levy process, T'(B \ B{R)) -+ +00 as R -> +00, a.s., hence
-iT'g
Ex 0 as 72 —> +00,
and we conclude that for any x,
FB^R)^1) ~* ^JtSfc), as .R -> +00.

Eq. (2.45) holds with B(R) instead of B, hence
/ PqB(R)9(x){q-L)w{x)dx = Q, VweC™(Bc).
The integrand converges pointwise to PBg(x)(q - L)w(x) as i? —> +00,

and it is bounded by a function of the class Li(R"), uniformly in R, since
(q - L)w e L i ( R n ) , and PB^R~.g{x) is bounded uniformly in R and x. By
passing to the limit, we finish the proof of Eq. (2.45). •
2.4 Boundary value problems for the Black-Scholes-type

equation
This section is the focal point, where Probability and Analysis meet to
produce in the end analytical formulas for prices of contingent claims and
value functions of firms in Real Options theory. By definition, each of those
prices and value functions is a sum of the stochastic integral which represent
the expected discounted stream of payments (or revenues) during the life-
time of the contingent claim or the firm, and the expected terminal payoff in
the end. By using Dynkin's formula and the potential theory, we deduce the
pseudo-differential equation for the price in the region, where the contingent
claim remains alive, and the terminal payoff translates into the boundary
condition. Unlike in the Gaussian case, the boundary condition is non-local:
the unknown function must be specified not only at the boundary but the
other side of the boundary as well. On solving the boundary problem in an
appropriate function class and obtaining a formula for the unique solution
in this class, we obtain the price we have been looking for. In the last step,
we use standard technical tools from elements of the theory of generalized
functions, the theory of PDO and Complex Analysis.
Boundary value problems for the Black-Scholes-type equation 63
2.4.1 Boundary problems for the stationary Black-Scholes

type equation
Let X be an n-dimensional Levy process satisfying (ACP)-condition, let
B C R " be a closed set, let T'B be the hitting time of B by X, let g° 6
Loo{B) and gc € Loo(Bc) be non-negative functions, and consider
fix )=EX e-qtgc{Xt)dt + Ex[e-«T*g°(XT.B)].

\I! (2.46)
We have already mentioned in Introduction (and will see in Chapters IV-

IX) that prices of various derivative securities in the financial market are
naturally defined by expressions like Eq. (2.46).
Theorem 2.12 Let X be an n-dimensional Levy process satisfying (ACP)-

condition. Then the stochastic expression Eq. (2.4-6) is a bounded solution
to the following boundary value problem
(9-WW gc(x), x € Bc, (2.47)

/(*) g°(x), xeB, (2.48)
where Eq. (2.47) is understood in the sense of generalized functions:
U,{l-L)w)L2={gc,w)L„ (2.49)
for all we C£°(BC).

Proof. Set fc = U*gc. Then {q - L)fc = gc, therefore
</ c , (q - L)w)L2 = (9C,W)L2, V » £ C0°°(BC),
and it remains to show that / " : — / — fc satisfies Eq. (2.49) with gc = 0.

In view of Eq. (2.45), it suffices to show that f° = Pgh for some bounded
measurable h. But
r+oo
/"(*) = -Ex / e-^gc{Xt)dt + PqB9°(x)
JT'B
-E
-Bê-^'BE*™
r
Jo
e-^B[/V(Xri3)l+P|50(x)
" 9 * ncc
9 (Xt)dt + PqB9°{x)
P%h{x)
64 Levy processes
where h = —Uqgc + g° is measurable and bounded. •
Remark 2.1 a) In each concrete case, we find bounded solutions of the

problem Eq. (2.47)-Eq. (2.48) by using tools from Analysis; in many cases,
the bounded solution is unique, and in other cases, assuming that gc and
g° are continuous and bounded, we single out the solution we need as the
unique continuous bounded solution.
b) In many applications, the data g° and gc may be unbounded. Suppose
that g° and gc are non-negative, measurable and finite a.e. Then we use
the following scheme:
(i) construct a sequence of non-negative bounded measurable <?° (resp.,

g„) with compact support, which converge pointwise to g° (resp.,
gc), a.e.: g°n(x) | g°(x) and gcn(x) | gc(x);
(ii) find the unique bounded solution un of the problem Eq. (2.47)-
Eq. (2.48) with the data p° and g„ (or the unique continuous
bounded solution);
(iii) calculate the limit u := linin-^+oo un; by the Monotone Conver-
gence Theorem, this is the price of the derivative security or the
value of the firm.
c) If g° (and/or gc) changes sign, we represent it as the difference of the

non-negative function, and calculate separately the stochastic integrals for
the non-negative data.
d) If the process does not satisfy (ACP)-condition, we can solve the problem
Eq. (2.47)-Eq. (2.48), and after that check that Dynkin's formula is appli-
cable to un, that is, (q - L)un is non-negative and universally measurable.
Luckily, RLPE's do satisfy (ACP)-condition, as we will see in Chapter 3.
2.4.2 Boundary problems for the non-stationary Black-

Scholes type equation
Let T > 0 be the deterministic time, let D be a closed subset of R™ x
( - o o , T ] , a n d s e t £ > r = {x | (x,T) $ D}, Bc = {(x,t) £ D | t e ( - o o , T ) } .
Let gr e Loo(D), gT e Loa(DT) and gc € Loo(Bc) be non-negative. Let
T'D be the hitting time of D by an (n + l)-dimensional process {(Xt,t)},
Boundary value problems for the Black-Scholes-type equation 65
and for ( i , t ) e R " x [ 0 , T] consider

rT'n/\T
f(x,t) = £*•* e~'13gc(Xs,s)ds (2.50)

Jo
T
+E^[e-" ogr(XT>D,T'Dn{T^<T}}
+E*' [e-*TgT(XT)l{rD>T}].
t
We will see that the prices of contingent claims with finite time horizon,
e.g., barrier options, can be expressed in this form.
Theorem 2.13 Let X be an n-dimensional Levy process satisfying (ACP)-
condition. Then the stochastic integral Eq. (2.50) is a bounded solution to
the following boundary value problem
(q-8t-L)f{x,t) = gc(x,t), V(i,t)eF; (2.51)

T T
f(x,T) = g (x), Vx€Z) ; (2.52)
r
f(x,t) = g (x,t), W(x,t)eD, (2.53)
where Eq. (2.51) is understood in the sense of generalized functions:
(f,(q + dt-L)w)L2 = (gc,w)L2, (2.54)
for all w G <S(R™ x R ) s.t. supp w C Bc.

Proof. Introduce an (n + l)-dimensional Levy process {Xt} := {(Xt,t)},
denote its infinitesimal generator by L, and define the function g° on B :=
R n x R \ Bc by
gT(x), t = T and x € DT
{ gr(x,t),
0,
t < T a n d (x,t) eD
t > T and x £ R n .
We can write Eq. (2.50) as Eq. (2.46), with (x,t) and X in place of a; and
X:
f(x,t) = Ex>t
Jo
f e-qsgc(Xs)ds + Ex>t[e-qTBg0(XT')} (2.55)
From Lemma 2.4, X satisfies (ACP)-condition, therefore we can apply

Theorem 2.12 to Eq. (2.55), and obtain Eq. (2.51)-Eq. (2.53), with L in
Eq. (2.51) in place of dt + L. It remains to note that the two operators are
equal. •
66 Levy processes
Remark 2.2 More general payoffs, and processes which do not satisfy
(ACP)-condition, can also be considered.
Remark 2.3 Since the proof of Theorem 2.12 is based on the results from
the theory of Markov processes having absolutely continuous potential mea-
sures, the main results of this section admit the straightforward generaliza-
tions for Markov processes.
2.5 Commentary
For the systematic treatment of the general theory of Levy processes, see
[B], [S], and for a shorter exposition, Rogers and Williams (1994); for the
general theory of Markov processes, see Blumenthal and Getoor (1968),
Rogers and Williams (1994), Sharpe (1989) and the bibliography there.
Chapter 3
Regular Levy Processes of

Exponential type in I D
In this Chapter, we introduce families ofRLPE which are used in empirical

studies of financial markets not only by researchers, but by some practi-
tioners as well. Then we give two general definitions of RLPE and study
basic properties of characteristic exponents of RLPE.
3.1 Model Classes
3.1.1 An overview
D.B. Madan with corroborators were the first to use Levy processes differ-
ent from the stable ones, namely, Variance Gamma Processes (VGP)—see
Madan and Seneta (1990), Madan, Carr and Chung (1998)—in financial
modelling. A host of processes has been introduced, starting from the
family of Generalized Hyperbolic distributions constructed by Barndorff-
Nielsen (1977): Hyperbolic Processes (HP) and Normal Inverse Gaussian
(NIG) processes were constructed in Eberlein and Keller (1995) and Barndorff-
Nielsen (1998), respectively, and their generalizations: Generalized Hyper-
bolic Processes (GHP), and Normal Tempered Stable Levy Processes (NTS
Levy) - in Eberlein and Prause (2001), Prause (1998), Raible (2000), and
Barndorff-Nielsen and Levendorskii (2001), Bamdorff-Nielsen and Shep-
hard (2001c), respectively. By now the list of papers, where processes of
this cluster have been used for modelling behaviour of asset prices and
term structure of interest rates, pricing of derivative securities, and risk
management, is fairly long: Barndorff-Nielsen and Jiang (1998), Eberlein,
Keller and Prause (1998), Prause(1998), Eberlein and Raible (1999), Eber-
67
68 Regular Levy Processes of Exponential type in ID
lein et al. (2000a, 2000b), Eberlein (2001a, 2000b), Eberlein and Prause
(2001), Barndorff-Nielsen and Prause (2001), Barndorff-Nielsen and Shep-
hard (2001c). In Barndorff-Nielsen and Shephard (2001c), Normal Modified
Stable laws were constructed, thereby generalizing the class of Generalized
Hyperbolic distributions. For extensions to Feller processes and stochastic
volatility models, see Chapter 14 and Commentary to Chapter 1, respec-
tively.
(Non-infinitely divisible) Truncated Levy distributions were constructed and
used for financial modelling purposes by Mantegna and Stanley (1994,
1997); see also the monograph Mantegna and Stanley (2000). Novikov
(1994) and Koponen (1995) introduced infinitely divisible analogues of
truncated Levy distributions. The extension of Koponen's family was con-
structed in Boyarchenko and Levendorskii (1999a, 1999b, 2000); we will call
it KoBoL family. Later this family was used by Carr et al. (2001) under
the name CGMY-model. It was shown that processes of the KoBoL family
provide good fit for many equities.
Below, we provide technical details of the derivation of necessary formulas
mainly for KoBoL processes, since for these processes, all proofs can be
made in full by using very simple analytical tools, and Variance Gamma
Processes can be regarded as members of KoBoL family. The reader can
find technical details concerning processes of the other families in op. cit.
3.1.2 KoBoL family: direct construction via the Levy-

Khintchine formula
3.1.2.1 Definition in terms of the Levy measure
For v < 2 and A > 0, set
IL+(v,\;dx) =x+"-le-Xxdx, TI-(is,\;dx) = xZ"~leXxdx,
where x+ = max{i, 0}, x_ = x+ — x.
Definition 3.1 A Levy process X is called a KoBoL process of order

v < 2 if it is a purely discontinuous Levy process with the Levy measure of
the form
U(dx) = c+U+ {v,-\_;dx) + c-XT {v,\+;dx), (3.1)
where c± > 0, and A_ < 0 < A + .

Model Classes 69
A+ and A_ are called the steepness parameters of the process.
Clearly,
+0O
min{l,x 2 }n ± (i/,A;dx) < +00,
/ -00
therefore Eq. (3.1) defines a Levy measure. Since the condition

+00
min{l,|x|}n ± (i/,A;da;) < +c
-00
/ -00
is satisfied if and only if v < 1, we conclude that a K0B0L process is a finite

variation process if and only if v < 1.
3.1.2.2 The characteristic exponent

To compute the characteristic exponent of a K0B0L process, define
V>±(i/, A;^) by the Levy-Khintchine formula Eq. (1.3) with a = 7 = 0 and
R(dx) = TV^ (v, \; dx). The following lemma will be proved at the end of
this subsection.
Lemma 3.1 For u < 2 and A > 0, there exists C{v,X) such that
i/>± (u, A; 0 = ±iZC{v, A) + 4>(V, A; ± 0 , (3.2)
where
a) if v < 2, v ^ 0 , 1 , then
0(i/,A;O=r(-i/)[A"-(A-iO"]; (3-3)
b) if v = 0, then
0(O,A;O = l n ( A - i £ ) - l n A ; (3.4)
c) if v — 1, then
</>(l,A;£) = A l n A - ( A - * £ ) l n ( A - i O . (3.5)
From Lemma 3.1, we deduce

Lemma 3.2 The characteristic exponent of a K0B0L process of order v
is of one of the following forms:
70 Regular L4vy Processes of Exponential type in ID
(i) ifv<2, v ^ 0,1, then
(3.6)
(ii) if v — 0, then
^ ( 0 = -i/ie + c+[ln(-A_-iO-ln(-A_)]+c_[ln(A + +^)-lnA + ]; (3.7)
(Hi) if v — 1, then
iMO = - i / ^ + c+[(-A_)ln(-A_)-(-A_-*01n(-A_-i0]
+c_ [A+ In A+ - (A+ + ifl ln(A + + if J], (3.8)
where fi € R, c± > 0, and A_ < 0 < A+.

Notice that we can (and will) use Eq. (3.6)-Eq. (3.8) to define KoBoL
processes, without referring to the initial definition in terms of the Levy
measure.
The triple (v, c+,c-) essentially determines the shape of the probability
density in a neighbourhood of the origin. The smaller the v and the larger
the c±, the higher the peak in the centre of the probability distribution,
and in the case c+ ^ c_, the fall-off from the peak is asymmetric. The
order of the process coincides with the order (in the sense of the theory of
PDO) of the real part of the symbol —V"(0 of the generator L = —ip(D)
of the process unless v = 0 or v = 1, when the standard definition of the
order of a PDO does not apply. The steepness parameters determine the
rate of exponential decay in far parts of the tails of the probability density:
the smaller the A+ (resp., —A_), the heavier the left (resp., right) tail.
Typically, — A_ > A + , that is, the left tail is heavier.
Usually, a good fit to observed data can be achieved with the choice
c_|_ = c_. We call this constant the intensity of the process, and denote it
by c. The c plays the part similar to that of the variance of a Brownian
motion. To simplify the formulas, in the sequel we restrict ourselves to the
case c± = c. The formulas Eq. (3.6)-Eq. (3.8) simplify as follows:
if v < 2, v ^ 0,1, then
<K0 - -irf + cT(-i/)[(-A_)" - (-A_ - itf + A + " - (A+ + i O I ; (3.9)
if v = 0, then
<K0 = -irf + cpn(-A_ - iO - ln(-A_) + ln(A+ + t£) - In A+]; (3.10)

Model Classes 71
if v — 1, then
iMO = -t/* + c[(-A_)ln(-A_)-(-A_-tflln(-A_-iO]

+A+ In A+ - (A+ + iO ln(A+ + if)] • (3.11)
Prom Eq. (3.9)-Eq. (3.11), it is clearly seen that a mixture of KoBoL-

processes of the same order v and steepness parameters A_ and A+ but
different \i and c is a KoBoL process with the same v, A_ and A+. In other
words, a subclass of KoBoL distributions with fixed v, A_ and A+ is closed
under convolution, as the class of normal distributions is.
3.1.2.3 Moments of a KoBoL process

Lemma 3.3 For the first four central moments, we have
mi : = mi.-r/T =
#'(°); (3-12)
Tn2 : = m 2 , r / r = V"(0); (3.13)
(3)
m3 : = lim m 3 r / r = -n/> (0); (3.14)
T—*0 '
7714 : = l i m m 4 T / T = -V ( 4 ) (0). (3.15)

T—>0
The proof will be given at the end of the subsection. Of course, Lemma 3.3 is
valid not only for KoBoL processes, but we will use it mainly for them. For
a KoBoL process of order v ^ 1, by using the property T(l +a;) = xF(x) of
the Gamma function, we deduce from Eq. (3.9)-Eq. (3.10) and Eq. (3.12)-
Eq. (3.15)
mx = n + cT(l - «/)[(-A-)"" 1 - A+"- 1 ]; (3.16)

mk = cT(k-u)[(-\_)u-k + (-l)kX+,/-k}, k>2. (3.17)
We leave to the reader the computation of the derivatives of the character-

istic exponents of KoBoL processes, and their evaluation at 0, for KoBoL
processes of order 1.
Consider the typical case of the fatter left tail, that is, A+ < — A_.
Since T(x) > 0 for x > 0, we conclude that m^ < 0, hence, the probability
distribution is negatively skewed. Due to v < 2, m<i and rm increase as A +
and/or — A_ decrease, and for small r , the skewness
, m3,T _ i / 2 «i3 , 0 1CA
x
*.r = -im~T —* (3 18)
-
m m
2,r 2
' X V—2.0
v-1.5
x
• v-1.0 .
." —- v-O.5
x
•
' ". x X
_"
• •
X
* _*
". * .*
'. X
-'
s -«-
X x
-* •_
X
\ •- " x
x ."
.•
\ . x X .* /
" -• ' ^^^"^_

N
-^_ *- x
^•^^^ N . x
^\^.x
jp'
-300 -200 —100 10O 200 300
Fig. 3.1 The characteristic exponent and log-density of a KoBoL process: dependence
on the order of the process. Parameters: n = 0,X+ = —A_ = 24,rri2 = 0.18, r = 1 day,
v £ {0.5; 1.0; 1.5; 2}
and the kurtosis

7Tl4 . _l"l4
(3.19)
m 2,r
grow unboundedly as r -¥ +0, which agrees with empirical observations
in real financial markets. Notice that Eq. (3.18)- Eq. (3.19) hold for all
RLPE (and any Levy process with the sufficiently regular characteristic
exponent).
3.1.2.4 Graphs of characteristic exponents and probability densi-

ties of different KoBoL processes
To compare Gaussian and non-Gaussian processes, we choose the parame-
ters so that the first and second central moments, m i , r and m 2 , r , of proba-
bility densities of different processes coincide. It allows us to illustrate the
errors which arise when a non-Gaussian process is modelled as a Gaussian
one, the drift and variance being inferred from the first two moments of the
observed probability densities.
In Fig. 3.1, we take fi = 0,A + = -A_ = 24; then m,\ = 0. For each
v G {0.5; 1; 1.5}, c is chosen so that m 2 = 0.18 is independent of v. v = 2
denotes the case of the Brownian motion with zero drift and variance a2 =
0.18.
Model Classes 73
1 1 1 1 i
Fig. 3.2 Probability density of a KoBoL process: dependence on time and the order of
the process. Parameters: n — 0,A+ = —A_ = 24,nx% = 0.3, v G {0.5; 1.0; 1.5;2}. Left
panel: j — \ day, right panel: T = 5 days.
In Fig. 3.2, we illustrate how the probability density of a KoBoL process

evolves in time, and in Fig. 3.3, how it depends on the asymmetry parameter
0 = (A++A_)/2.
In Fig. 3.4, we fix v = 1.4, m.2 = 0.18, and see how the skewness and
kurtosis depend on A + and A_.
3.1.2.5 Proofs of technical lemmas

Proof of Lemma 3.1 The case of "-" sign easily reduces to the case of "+"
sign by a change of variables x H> —X, £ >-> — £, therefore it suffices to prove
Eq. (3.2)- Eq. (3.5) for "+" sign only,
a) If v < 0, we may write
r+oo r+oo
</>+(i/,A;0 = / x-"-le-Xxdx- x'"'1 e~{x-*)x dx + C{v, A),
Jo Jo
where
C(i/,A) =i£ [ x~ve~Xxdx.

Jo
Fig. 3.3 Probability density and log-density of a KoBoL process: dependence on /3 =

(A+ + A- )/2. Parameters: n = 0, v = 0.5, A± = 24 ± /?, m 2 = 0.3, T = 3 days.
Fig. 3.4 Skewness and kurtosis of the probability density of a KoBoL process: depen-
dence on the steepness parameters. Parameters: fi — 0, v = 1.4, mi = 0.18, T = 1
day.
By making changes of variables x — yj A in the first integral, and x =

y/{\ - i£) in the second one, we obtain Eq. (3.2) with <j>(v,\;£) given by
Eq. (3.3).
Model Classes 75
In the case v 6 (0,1), we integrate by part:
rl>+{v,\;0 = - - { f °° (e~Xx - e ' ^ - ^ dx~v
= _I|( e -^_ e -(A-iC)^ a ; -|+°°
+oo .
/
(\e~Xx - (A - i O c - ^ " ^ * ) x~vdx
+ f (\e~Xx - (A - i O e - ( A - i S ) l + if (Ax - l)e~ A x ) x~vdx\
= ~-l [+°° (e~Xx - e - ( * - * ) * ) x-"dx
+ ^ [ e " A + f (\x- l)e-Xxx-"dx]\
r(1
" V - (A - »0"] + i£C(i/, A),
where
C(i/,A) = [ - e - A + / (1 - \x)e-Xxx-"dx\lv = /" x - " e - A x d x .

Jo Jo
Since -i/r(-i>) = T(l - v), Eq. (3.2)-Eq. (3.3) have been proved.
Finally, if v E (1,2), we integrate by part twice, and arrive at the same
formulas Eq. (3.2)-Eq. (3.3), with
C(u, A) = -(v - lTx[e-x + \C(v - 1, A)].
b) If v — 0, we differentiate ip+(0,X,^) w.r.t. JJ, — - i f and get
dx
__1 l-e~A
~ /x + A A
Since ^ + ( 0 , A,0) = 0, we can obtain V + (0, A,f) by integration:
= lnV +
^~A—
c) If v = 1, we differentiate ^ + (1,A,£) with respect to A and get
—i(>+(0,\,£). Clearly, liniA-n-oo V^Clj ^>£) — 0, therefore we can find
V>+(1,A,£) by integration:
+oo
V»+(0,/i,OdM
/A
/"+oo
fo
° /_ e~^ -_ 1 \
= / Iln(/x-i0-ln^-^ J d/z
= (0* - *0 1Q(M - *£) - / i + i ^ - Z i l n / i + zi + ^ l n / i ) ! ^ 0 0
—z£ / d/x
J\ V
= AlnA-(A-i£)ln(A-i£)-i£ l+lnA+/ dfx
(here we have used lim^_,. +00 [(^ — i£)\n((i — i£) — filnfi — ifc(l+ln/i)] = 0),
and Eq. (3.2) and Eq. (3.5) follow.
Proof of Lemma 3.3. Since
+oo
e-faS-nKO^ (3 20)
/
-oo
we have
+oo
xpT(x)dx
/ -oo
p + OO /" + 00
-t-oo /-t-oo
/ xe- i x «- r , ^>d£d:r
/ -oo J —oo
/+oo /.-foo
/ e-^^de-^dx
-oo J—oo
-f-OO /- + 0 0
/
/ e-"^'(0e-r*(€)dfda:.
-oo J—oo
Model Classes 77
Since for / e «S(R),

r+OO
i-OO r+oo
r-too
/ e-**tf{Qdtdx, (3.21)
/
-oo J — oo
we conclude that Eq. (3.12) holds.
Further,
+oo
r+oo
m2 (x - mi,T)2pT(x)dx
/ -oo
+oo
-i x2pT (x)dx — m\T
-t-OO /-+O0
/ i2e-fa«-T*«)deda;-m?iT>
/ -00 J — OO
and integrating by part twice and using Eq. (3.21) and Eq. (3.12), we obtain
Eq. (3.13). Eq. (3.14) and Eq. (3.15) are derived similarly.

pered Stable Levy Processes: construction via subor-
dination
As early as in 1973, Clark modelled financial processes by subordination
of Brownian motion, in the discrete time model, and Madan and Seneta
(1990) were the first who introduced a model (Variance Gamma) which
used a Levy process as a subordinator. We start with a family of processes
which are the most convenient from the analytical point of view.
Let v 6 (0,2), 8 > 0, and let p(x;u,8) denote the probability density
function of the positive i//2-stable law S(v, 8) on R + with the Laplace
exponent 8(26)v/2. For 7 > 0, define the exponentially tempered version of
p(x; v, 8) by
p{x; v, 8,7) = e^"p(x; v, 8)e~^x. (3.22)
The distribution with density Eq. (3.22) {v G (0,2), 8 > 0,7 > 0) is referred
to as a tempered stable law, and we denote it by TS(v,8,j). TS(v,8,y) is
infinitely divisible, and its Levy density and characteristic exponent are
given by
V2V/2 X
u(x) - A ~ -l-u/2 -\i2x
$(9) = J[( 7 2 + 20)"/ 2 - 7 " ] . (3.23)
Let Y(B) be the Brownian motion with drift B, and let X be the subor-
dination of Y(B) by TS(u, 5,7). Then the characteristic exponent of X,
denote it (f>, is of the form
.2
and by using Eq. (3.23), we obtain
0(0 = 8[(a2 - (B + itf)"'2 - (a2 - fy'%

where a > \B\ is found from 7 = (a 2 — /3 2 ) 1 / 2 . By adding the drift, we
obtain the characteristic exponent of a Normal Tempered Stable (NTS) Levy
process
W O = -if* + S[(a2 -(8 + i02)"/2 - (a2 - B2)"'2}. (3.24)

The a — B (resp., a + B) describe the rate of decay of the right (resp., left)
tail, and <5 plays the part of the variance of a Brownian motion.
In the case v — 1, the characteristic exponents of Normal Inverse Gaus-
sian Processes (NIG) obtain:
W O = -irf + <5[("2 - (P + i02)1/2 ~ ( ^ - /? 2 ) 1/2 ]- (3.25)

From Eq. (3.25), it is clearly seen that a mixture of NIG-processes with
the same a and B but different \i and S is a NIG process with the same a
and B. In other words, a subclass of normal inverse gaussian distributions
with fixed a and B is closed under convolution, as the class of normal
distributions is. From Eq. (3.24), the same conclusions hold for NTS Levy
processes of a fixed order.
NIG have another important property: for any t > 0, the following
explicit analytical formula for the probability density is available:
a
Pt(X) = -eMAS^^f2 - M + 8X]K^.6{jf-^6)\ (3.26)
where (y) — (1 + I2/I2)1/2, and K\ denotes the modified Bessel function of

the third kind with index 1.
By using Eq. (3.12)-Eq. (3.15) and Eq. (3.24)-Eq. (3.25), one can eas-
ily calculate moments of probability distributions of NTS Levy and NIG
processes.
Model Classes 79
It can be shown that an NTS Levy process is a finite variation process

if and only if v < 1. In particular, NIG are not finite variation processes.
3.1.4 Variance Gamma Processes

As it was mentioned in Section 3.1.1, Variance Gamma Processes (VGP)
can be defined as purely discontinuous Levy processes with the Levy density
of the form c(x^}ex-x + xI 1 e A + a ; ). Then the characteristic exponents are
given by Eq. (3.10), but a more common form of the characteristic exponent
of a VGP is
<K0 = -if* + c[Ma2 - (0 + iO 2 ) - ln(a 2 - /?2)], (3.27)

where a > \/3\ > 0, c > 0 and // € R. Similarly to NIG, a VGP without the
drift is a subordination of a Brownian motion; this time, the subordinator
is a Gamma process with the Laplace exponent of the form
$(0) = c[ln(A + 2 0 ) - l n A ] ,
where A > 0. Similarly to KoBoL and NIG processes, a mixture of VGPs

with the same a and (3 but different /x and c is a VGP with the same a and
P-
3.1.5 Hyperbolic Processes and Generalized Hyperbolic Pro-

cesses
Generalized hyperbolic distributions were introduced by Barndorff-Nielsen
(1977) as variance-mean mixtures of normal distributions. The Lebesgue
density of a generalized hyperbolic distribution depends on five parameters:
xKx_1/2(a^d^ + (x-^)2) exp[/3(x - M )],
where Ky denotes the modified Bessel function of the third kind with index
A. An integral representation of Ky is given by
1 f°°
K^{z) = ^ J V1'1 exp[-0.52(y + y^jdy. (3.29)
The a,/3,n and 5 play the same parts as in the case of normal inverse
gaussian distributions, which obtain as a subclass of the class of generalized
hyperbolic distributions with A = —1/2. By changing A, one changes the

heaviness of the tails of the distribution. With A = 1, one obtains the
subclass of hyperbolic distributions, with heavier tails. The densities of the
hyperbolic distributions are expressed by the following simple formula
d{x) =
o Jfx rr—^ exp[-oy<52 + ( x _ M ) 2 + p{x _ rfi (3.30)
2aoKi{dy/a2 - /32)
which explains the name "hyperbolic": the graph of In d(x) is a hyperbola,

whereas in the case of normal distribution, we get a parabola.
The characteristic function of a generalized hyperbolic distribution is
given by
_ ,w I »' - ^ "i V2 * » ( V « 2 - ( / » + «) 2 )
(3 3I)
«o —UrcmtFJ w-w '
for |/3-Hf| < a. A generalized hyperbolic distribution is infinitely divisible,
and the density of the Levy measure is: if A > 0, then
0x f /-co e x p ( - \ / 2 y + a2|x|) \
and if A < 0, then
p0x /-oo exp (-\/2y + a 2 |x|J

9{X) = dy. (3.33)
^W\Jo y(J?Lx(6y/Zy) + Y*x(6y/Zy))
Here J\ and Y\ are the Bessel functions of the first and second order,
respectively.
By using the correspondence between infinitely divisible distributions
and Levy processes, one can define a Generalized Hyperbolic Process (GHP)
as the Levy process {Xt} with E[e^Xi] given by the RHS in Eq. (3.31).
We will need the asymptotics of g(x) as x -» +0.
Lemma 3.4 As x -> 0,
g(x) = -\x\-2 + 0(\x\-1). (3.34)
Proof. Prom the well-known asymptotic formulas for Bessel functions it

follows that the denominators of the integrands in Eq. (3.32) and Eq. (3.33)
Model Classes 81
stabilize to positive constants as y -> +0, and have the following asymp-
totics, as y -> +00:
+0(!rl) = (1+0to 2) ( 5)
K^ ) ^ ~" ' "
Hence, by making the change of variable y = \x\~2z2/2 in Eq. (3.32) and
Eq. (3.33) and using Eq. (3.35), we obtain Eq. (3.34). D
From Eq. (3.34), we conclude that all Generalized Hyperbolic Processes are
of infinite variation; in particular, NIG and HP are.
3.1.6 Comparison of the tail behaviour of probability den-

sities for different model classes of processes
For GHP and NTS Levy processes, NIG in particular, and for VGP, set
A_ — —a + /3, A + = a + j3. Then for all model classes, A_ is characterized
by the behaviour of any probability density pT{x) as x —> +00:
Vw>A_, limsupe~ w x p T (x) < + 0 0 , (3.36)
and
V UJ < A_, lim e'^prix) = +00. (3.37)

x—>+oo
For K0B0L, NTS Levy processes and VGP, Eq. (3.36)-Eq. (3.37) easily
follow from the formulas for the characteristic exponents; for GHP, the
verification of Eq. (3.36)-Eq. (3.37) requires more effort. It follows from
Eq. (3.36)-Eq. (3.37) that if two model processes are characterized by dif-
ferent A_(l) < A_ (2), then the right tails of probability densities of the
second process are fatter, independently of the classes the processes belong
to, and of the orders of the processes.
If A_(l) = A_(2), then the behaviour may differ: for instance, Eq. (3.36)
holds with u = A_ for K0B0L processes of order v > 0 and NTS Levy
processes of any order v e (0,2), whereas Eq. (3.37) holds with the same
UJ = A_ for VGP and HP. Thus, the right tails of probability densities of the
latter group of processes are fatter than those of the former group, but this
observation has hardly any practical significance: it is very difficult to infer
A_ from any empirical data set, and hence an error in A_ will dominate an
additional subtle difference in the tail behaviour.
It can be shown that of two KoBoL processes with the same A_, the
one of the lower order has the fatter right tail, and the same is true of NTS
Levy processes.
In the same manner, A+ is characterized by the behaviour of the left
tail.
3.2 Two definitions of Regular Levy Processes of Exponen-

tial type
3.2.1 Definition in terms of the Levy measure

Definition 3.2 Let A_ < 0 < A+. We call X a Levy process of exponen-
tial type [A_, A+] if its Levy measure satisfies
—1 /-+00
/
e- A + x n(da;) + / e-x-xU{dx) < oo. (3.38)
-oo Jl
Lemma 3.5 Let X be a Levy process of exponential type [A_, A+]. Then
(i) the characteristic exponent ip is holomorphic in the strip 9£ €
(A_, A+), and continuous up to the boundary of the strip;
(ii) there exist C and v > 0 such that for all £ in the strip 9£ G [A_, A+],
mO\<C(l + W; (3.39)
(Hi) for any q > 0, there exist S > 0 and er_ < 0 < cr+ such that for any
[w_,o;+] C ((T-,<T+) and all £ in the strip 9£ € [CJ_,CJ + ],
q + Uip(0 > 6, (3.40)
where 5 = 8(q,oj-,LJ+) > 0;

(iv) if a± satisfy
q + ip(i((T- + 0)) > 0, and q + i>{i{o+ - 0)) > 0, (3.41)
then Eq. (340) holds;

(v) for any q > 0, the equation
<Z + <MO=0 (3.42)
has at most one purely imaginary root in the lower (resp., upper)
half-plane, call it —i(i+ (resp., —ift-);
Two definitions of Regular Levy Processes of Exponential type 83
(vi) the root —ife exists if and only if
g + lK»(A±=FO))<0, (3.43)
and if it exists, it is a simple root.
Proof. (i) is immediate from Eq. (3.38), and (ii) can be easily deduced
from the Levy-Khintchine formula, by considering separately the integrals
over |x| < |£|- x and \x\ > | £ | - J .
(iii)-(iv) Set M\{a) — f*™ e~axfi1(dx), where ^(dx) is the probability
distribution of Xi. By differentiating twice, we conclude that Mi is convex,
and clearly, Mi(0) = 1 < eq. Hence, there exist w_ < 0 < u+ and 5 > 0
such that for all o e [w_,w+], M\{a) < eq~s.
Now, for any £ € R , and these a,
exp(-3ty(£ + ia)) = | exp(-</>(£ + ia))\ =
+oo I /-+oo
/
-oo I J— oo
therefore Eq. (3.40) holds with cr_ = infw_, a+ — supo;+, and Eq. (3.41)
implies Eq. (3.40).
(v)-(vi) Notice that by the proof of (iii), a <-* q + ip(ia) is concave and
equals to q > 0 at 0. •
For the sake of brevity, from now on we consider processes with Levy
measures (almost) symmetric in a neighbourhood of the origin.
Definition 3.3 Let A_ < 0 < A+ and v € [0,2). A purely non-Gaussian
Levy process is called a Regular Levy Process of Exponential type [A_, A+]
and order v if its Levy measure satisfies Eq. (3.38) and, in a neighbourhood
of zero, admits a representation II(dx) = f(x)dx, where / satisfies the
following condition:
there exist v' < i/, c > 0, and C > 0 such that
\f{x)-c\x\-v-1\<C\x\-v'-1, V |x| < 1. (3.44)
Notice that a regular Levy process of exponential type has bounded varia-
tion if and only if v < 1, since this is equivalent to f_ (\x\ Al)Il(dx) < +oo.
All model classes in Section 3.1 are RLPE in the sense of Definition 3.3:
• KoBoL processes with c+ = C-, of order u > 0, v ^ 1, are RLPE
of order v and exponential type [A_, A + ];
• KoBoL processes with c+ = c_, of order 1, are RLPE of order 1

and any exponential type [A'Â!,.] C (A_,A+);
• VGP are RLPE of order 0 and any exponential type [A'_, \'+] C
(A-,A + );
• NTS Levy processes of order v are RLPE of order v and exponential
type [-a + ft, a + /?]; in particular,
• NIG processes are RLPE of order 1 and exponential type [—a +
P, a + /?];
• all GHP are RLPE of order 1 and any exponential type [A_, A+] C
3.2.2 Definition in terms of the characteristic exponent

The same calculations as in the proof of Lemma 3.1 prove that an RLPE
of order v > 0 in the sense of Definition 3.3 is an RLPE in the sense of the
following definition.
Definition 3.4 Let A_ < 0 < A+ and v £ (0,2]. A Levy process is called
a Regular Levy Process of Exponential type [A_, A+] and order v > 0 if the
following two conditions are satisfied:
(i) the characteristic exponent admits a representation
lMO = - * ^ + # 0 . (3-45)
where <fi is holomorphic in the strip 5£ £ (A_,A + ), is continuous
up to the boundary of the strip, and admits a representation
0(0 = cier+Oder), (3.46)

as £ -> oo in the strip 9f£ 6 [A_, A+], where v\ < u;
(ii) there exist u2 < v and C such that the derivative of (f> in Eq. (3.45)
admits a bound
\4>'{t)\<C{l + \ti\r, 9£G[A_,A+]. (3.47)
One can easily generalize both definitions by using c± > 0 in Eq. (3.44) on
the half-axis ±x > 0, and in Eq. (3.46), as *ft£ -4 ±oo.
From now on, we exclude RLPE of order 0, VGP in particular, from our
consideration, since the logarithmic behaviour of the characteristic expo-
nent as £ —¥ +oo leads to technical complications at many places; in fact,
Properties of the characteristic exponents and probability densities of RLPE 85
only for pricing of options of the European type, VGP are as convenient as
the other RLPE's. Each time one has to consider barriers or early exercise
boundaries, one has to use the Wiener-Hopf factorization, which can be
done much simpler when the characteristic exponent grows polynomially
at the infinity. At the same time, we include Brownian motions: according
to Definition 3.4, they are RLPE of order 2 and any exponential type. No-
tice that a mixture of independent RLPE is an RLPE, the order being the
largest one of the orders of the components, and the exponential type being
the intersection of the exponential types of the components. In particular,
any RLPE with a non-trivial Gaussian component is of order 2.
3.3 Properties of the characteristic exponents and proba-

bility densities of RLPE
Clearly, an RLPE is a Levy process of exponential type, therefore the prop-

erties listed in Lemma 3.5 hold for any RLPE. In addition, from Eq. (3.40)
and Eq. (3.45)-Eq. (3.46), we deduce the following lemma.
Lemma 3.6 Let X be an RLPE of order v > 0 and exponential type
[A_,A+]. Then
(i) There exist 6 > 0 and C such that for all £ in the strip 9£ €
[A-.A+],
q + J t y ( 0 > c(l + \Z\y - C. (3.48)
(ii) For any q > 0, there exist 6 > 0 and a_ < 0 < cr+ such that for
any [w_,w+] C (cr_,a+) and all £ in the strip 3£ 6 [w_,w + ],
<z+3ty(o>s(i + Kir, (3-49)

where 6 = 6(q, w_,o;+) > 0.
Theorem 3.1 For any T > 0, the probability density pr of an RLPE is
infinitely smooth and exponentially decay at the infinity together with all its
derivatives.
In particular, an RLPE satisfies the (ACT)-condition.
Proof. Take u € (A_, A+). Then from Eq. (3.20) and Eq. (3.45)-Eq. (3.46),
+oo
/ -oo
-t-oo—iu
-co—iu
/
/+oo
i
e -««-'-*K+ -)df. (3.50)
-oo
The last equality obtains when we shift the line of integration. To justify
the shift, we use the Cauchy theorem and Eq. (3.48). Differentiate w.r.t. x
under the integral sign in Eq. (3.50) s = 0 , 1 , . . . times; from Eq. (3.48), we
conclude that the resulting integrand admits the estimate via
Cs(l + | £ | ) s e x p [ - < m
and therefore the integral converges absolutely and uniformly in x. Hence,

e~wxpT(x) and any of its derivatives are uniformly bounded w.r.t. x. •
We leave to the reader the formulation and proof of a similar statement

about pT(x) as a function of two variables, defined on the open half-plane
r >0.
The characteristic exponents from model classes enjoy additional im-
portant properties. The following lemma will be needed to derive simple
approximate formulas for the factors in the Wiener-Hopf factorization for-
mula. We managed to prove this lemma only for model classes of RLPE,
on the case-by-case basis. We conjecture that this lemma holds for a much
wider variety of RLPE if not for all RLPE.
Lemma 3.7 Let X be one of the model processes, of order v > 0 and
exponential type [A_,A + ].
Then
(i) the <f> in Eq. (3-45) admits the analytic continuation into the com-
plex plane with two cuts: (—zco,iA_] and [i\+,+ioo), and outside
any neighbourhood of iA_ and i\+ satisfies the following estimate:
W O I < C ( l + K|)"; (3-51)
(%%) all the roots in the plane with the cuts are purely imaginary.
Proof. See Boyarchenko and Levendorskii (2002a). •

Properties of the infinitesimal generators 87
3.4 Properties of the infinitesimal generators
Let either v > 1 or /x = 0. Then from Eq. (3.45)-Eq. (3.46), we conclude

that there exists C > 0 such that for all £ in the strip 3£ G [A_, A+],
\m\<c(i + \t\r. 0.52)

The bound Eq. (3.52) means that the infinitesimal generator L — —ip(D)
is a PDO of order v, and Eq. (3.49) means that q - L is a strongly elliptic
PDO of order v. Moreover, since both estimates are valid on any line 9£ —
ui 6 [UJ- , w + ], q — L is a strongly elliptic operator in spaces with exponential
weights (see Chapter 15). Thus, q — L enjoys very favourable properties
from the point of view of the general theory of operators on the line (in the
multi-dimensional case, in the whole space), but if a boundary is present
as it is the case with barrier options and options of the American type,
then for RLPE, which are not Brownian motion, the following important
property of elliptic differential operators fails: a solution to a sufficiently
regular boundary problem, e.g., the Dirichlet problem, with smooth data,
is smooth up to the boundary. This failure leads to the most significant
differences between the behaviour of prices of contingent claims near the
boundary in the Gaussian model, and the behaviour of the same prices in
non-Gaussian models.
The final remark is: if v 6 (0,1) and (x ^ 0, then q — L is no longer
strongly elliptic since the upper bound for the symbol q + ip(Q:
I? + T K O I < C ( I + KI)
is of a greater order than the lower bound Eq. (3.49). This leads to ad-
ditional problems with solutions near the boundary, in particular, to the
failure of the smooth pasting principle for optimal stopping problems.
3.5 A "naive approach" to the construction of R L P E or

why they are natural from the point of view of the
theory of P D O
If one wants to construct a non-Gaussian process with exponentially de-

caying tails of the density functions, whose generator looks as simple as
possible and possesses the most tractable properties from the point of view
of the theory of pseudo-differential operators, then one naturally guesses
88 Regular Livy Processes of Exponential type in ID
the formula for the infinitesimal operator of the NIG Levy process. Here
is a list of observations which naturally lead to NIG Levy:
2
(i) the generator of the Brownian motion is ^-A, but the tails of a
Gaussian distribution decay faster than an exponential function;
(ii) stable non-Gaussian Levy processes have generators of the form
L = -S\ — A|"/ 2 , where S > 0 and v 6 (0,2), and their symbols are
non-smooth at the origin: — <S|£|", which leads to polynomial decay
of the tails of the density functions;
(iii) the tails of the density functions observed in financial markets, in
turbulence and in many other fields of study usually have exponen-
tial decay; this means that the symbols of generators must be not
only smooth but holomorphic in a strip of the form 9£ G (A_, A + ),
where A_ < 0 < A+ (in the multi-dimensional case, in a tube do-
main 9£ € U, where U € R n is an open set containing 0).
The natural candidate for a generator with the property (iii) is
L =-8[{a2 - ^ 2
- a% (3.53)
where a > 0, S > 0, and if we want an asymmetric version, we simply shift

the strip:
L = -6[{a2 -(/3 + iD)2)v/2 ~ (a2 - 02)v/2], (3.54)
where a > 0, 6 > 0, a > |/?|, and D — —id is the standard notation
in the theory of PDO; it has the symbol £ (The last terms in Eq. (3.53)
and Eq. (3.54) are needed to ensure that L • 1 = 0, which is necessary for
a process without killing). Eq. (3.54) gives the infinitesimal generator of
a (multi-dimensional) NTS Levy process, and in the simplest case of the
square root, of a (multi-dimensional) NIG.
3.5.1 Modified Stable Processes and RLPE

NIG, HP and GHP have been constructed as follows. By using Generalized
Inverse Gaussian (GIG) distributions to mix normal laws, one constructs
the corresponding NIG, Hyperbolic and Generalized Hyperbolic distribu-
tions; since GIG are infinitely divisible, the resulting distributions also are,
and we can use the latter to construct Levy processes. Barndorff-Nielsen
and Shephard (2001c) introduced an extension of the family of GIG to a
The Wiener-Hopf factorization 89
class of distributions on R + + , the Modified Stable or MS laws. As GIG

laws, MS Laws are obtained from one of the positive K ( K E (0,1)) laws by
exponential and power tempering (or tilting). By using MS distributions as
mixing distributions for normal variance-mean mixtures, BarndorfF-Nielsen
and Shephard (2001c) introduced and study Normal Modified Stable laws.
They surmise (but have not proved) that MS laws are infinitely divisible,
and so this construction may lead to a new important class of Levy pro-
cesses. Our guess is that this class will be a subclass of RLPE, as NIG, HP
and GHP are.
3.6 The Wiener-Hopf factorization
We derive two representations for the factors in the Wiener-Hopf factoriza-

tion formula Eq. (2.23), and approximate formulas for the factors.
3.6.1 The case of Levy processes of exponential type

We fix a branch of In by the requirement In a € R for a > 0. We also fix
w_ < 0 < w+, for which Eq. (3.40) hold.
T h e o r e m 3.2 Let X be a Levy process of exponential type, let there exist

c, v > 0 such that
(3.55)
and let for UJ = u±,

+ 00 + 8U)
WW]
I dt] < +co. (3.56)
Then
a) (ptiO admits the analytic continuation into a half-plane Sr£ > w_ and
can be calculated as follows:
<£+(£) = exp (27T*)-1 J_ (3.57)

oo+iw_ <? + W ) V
1
exp (27U)- j _ dr) ; (3.58)
oo+iuj- lit - V)
b) <j>q (£) admits the analytic continuation into a half-plane 9£ < w+ and
can be calculated as follows:
<Pq (£) = e x p (3.59)

J-oo+iu+ Q + nV) V
'
= exp
/ —n—^—dri ; (3.60)
c) </\j~(£) 1 (resp., <j>q (£) 1) admits the analytic continuation into a wider
half-plane 3£ > A_ (resp., Off < X+) by
_ 1
<(0 = 9 _ 1 (9 + ^ ( 0 ) ^ 7 ( 0 . ^ e (A_,w_]; (3.61)
1
W = 9 _ 1 (g + ^ ( 0 X ( 0 , ^ 6 K,A+). (3.62)
Proof, a) Consider the expression under the exponent sign in Eq. (2.26):
r+oo p—qt r+oo
HO • = -J- (eixi ~ l)^{dx)dt
r
Jo Jo
r+oo „-qt r+oo r+oo
= / / (eixt - 1)(2TT)- 1 / e-^-^^drjdxdt.
JO * JO J-oo
On the strength of Eq. (3.55), we may apply the Cauchy theorem and shift
the line of integration:
r+oo „—qt r+oo f+oo+iui-
f(0 = —— (eix* - 1)(2TT)- 1 / e-i">-t*l'l)dndxdt.
Jo <• Jo J-oo+iui-
Now the inner double integral converges absolutely, hence we can apply the
Fubini theorem and integrate w.r.t. x first:
r+oo „—qt r+oo+iuj-
= —-(27TZ)-1 / e-'^ttrt - Z)-1 - ri-^dndt.
Jo * J-oo+iui-
Integrate by part:
r+oo -qt r+oo+iu- _ r
\nri—^txl)'(n)e-t^dndt
z
Jo J-oo+iu- V
/•+OO
r+oo r+ ™
/-t CO+iuJ-
, _ ,
= (27T2)-1 / / v-£.,.,,
ln-!—^xlj'{T])e^''+^^dr1dt.
Jo J-oo+iui- V
From Eq. (3.56), the integral above calculated in the reverse order dtdr) con-
verges absolutely. Hence we can apply the Fubini theorem once again and
obtain Eq. (3.57); integrating in Eq. (3.57) by part, we arrive at Eq. (3.58).
b) The dual process X is of exponential type [—A+, — A_], its character-
istic exponent is i/>, and [—CJ+, - w _ ] plays the part of [a;_,a;+] in Lemma
3.5. Write down the Wiener-Hopf factorization for X and apply the com-
plex conjugation; then the "+"-factor for X becomes the "-"-factor for X,
and Eq. (3.57) for X becomes Eq. (3.59) for X.
c) follows from Eq. (2.23) and Lemma 3.5, (i). •
Remark 5.1 If X is an RLPE in the sense of Definition 3.4, then Eq. (3.56)
and Eq. (3.55) hold; hence, Theorem 3.2 holds as well.
Lemma 3.8 Let CJ_ and w+ be as in Theorem 3.2.
Then there exists C > 0 such that in the half-plane ± 3 £ > ± w T , </>*
admits estimates
(l + K i r C < r # ( O I < ( l + KI)C- (3-63)

Proof. In Eq. (3.57) and Eq. (3.59), make change of variables rj i-> \£\t]
and use Eq. (3.39) to notice that the expressions under the exponential sign
admit an estimate via Cln(2 + |£|). •
Eq. (3.63) is insufficient for the proofs below. More information about
properties of the factors is obtained in the next subsection.
3.6.2 The Wiener-Hopf factorization for RLPE

Let <T_ < 0 < a+ be from Eq. (3.40). Fix A > max{—CT_,CT+}, and set
A±(£) s = (A=F*£)S — exp[sln(AT*£)]- Next, choose d > 0 and K _ , K + £ R
SO that
B(0 := d-lA+(£)-K+A-(0-K-(q + </>(£)) (3-64)
satisfies
lim 5 ( 0 = 1- (3.65)
Choices of d, /t+ and K_ depending on properties of t/>, hence on v, /x and c

in Eq. (3.45)-Eq. (3.46), we have to consider four cases.
(1) If v G (1,2) or v € (0,1] and fi = 0, we set d = c, K+ = K- = v/2.

(2) If v € (0,1) and /i > 0, we set d = /j, K+ = 1, K_ = 0.

(3) If v £ (0,1) and / / < 0, we set d = |/J|, K+ = 0, K_ = 1.
(4) If v = 1, we set d = (c2 + /J 2 ) 1 / 2 , K± = 1/2 ± 7r_1 arctan(/j/c).
In all cases, Eq. (3.65) follows from Eq. (3.45)-Eq. (3.46). In the first three
cases, Eq. (3.65) is immediate, and in the last case, the simplest way is to
check that lnf?(£) -t 0 as £ -> ±oo:
hm lnJB(0
KS = ± — K++ q = - ~ - / c _ + l n
? A±oo ' - 2 ^ 2 ' (C2 + M 2)l/2
+ lim ( - K + - K _ + 1) In |f|
f-+±oo
, . 7ri . u
= ± (vK ++ - « - ) —- T larctan— = 0
' 2 c
by our choice of K + and rc_.
The last factor in Eq. (3.64) assumes values in a half-plane Rz > 0 by
Eq. (3.40), and the same is true of the product of the first three factors,
since the first one is positive, A_(f) and A + (f) assume values in the half-
plane but in different quadrants, and 0 < K± < 1. Hence, for all ( £ R ,
—-K < argB(f) < 7r, and therefore, b — In B is well-defined on R. Fix
w_ < 0 < w+ such that <r_ < u>-, ui+ < a+, where cr± are from Eq. (3.40),
and notice that all the arguments above are valid on any line 3f = a 6
[w_,w + ]. Next, for r > <*;_, T\ € [w_,r) and real f, set
• r + OO+in b/ \
(by the Cauchy theorem, 6+(77 + i r ) is independent of a choice of T\), and

for r < oj+, TI € (r, w-f] and real f, set
It follows from Eq. (3.45), Eq. (3.46), Eq. (3.64) and Eq. (3.65) that there
exist C,Ci,p > 0 such that for any 77 in a strip ^sr] G [w_,ui+],
IBfa) - 1| < C(l + M)-',

and therefore,
\bm<ci(i + M)-p- (3.68)

Hence, the integrals in Eq. (3.66)-Eq. (3.67) converge, and 6+(£) (resp.,
&_(£)) is well-defined in a half-plane §£ > w_ (resp., Q£ < LJ+).
We set
o±(0=A±(0" ± expft ± (0-

T h e o r e m 3.3
a) a+ (resp., a-) is holomorphic in a half-plane 9£ > w_ (resp., 9£ < w+).

It admits the analytic continuation into a wider half-plane $?£ > A_
(resp., 9£ < \+), and the continuous extension up to the boundary, by
a+(0 = d-Hq + mya-it), ê[A_,u,_], (3.69)

1
a_(0 = cT (9 + V'(0)/o+(0, ^GK,A+]; (3.70)
b) on a strip 3f£ G [A_,A+],
9 + ^(0 =do+(Oa-(0; (3-71)
c^ f/iere exist C, c > 0 and pi > 0 swc/i that in a half-plane Sr£ > w_,
c(l + | ^ i r + < K ( O I < C ( l + | e i r - ; (3.72)
1
MO* - A+(0 ± K + 1 < C ( i + i e i ) ± , , + - p l ; (3.73)
and in a half-plane 9£ < OJ + ,
c(l + | € i r - < | a _ ( 0 | < C ( H - | C i r - , (3.74)
1
MO* - A - ( 0 ± K - I < C ( l + \Z\)±K—Pl; (3.75)
d) factors in Eq. (2.23) and Eq. (3.71) are related by
^(O-1=a±(O/a±(0). (3.76)
Proof. a) The first statement is straightforward by Eq. (3.68), and once

c) is proved, the second one follows since d_1(q + i/KO) is holomorphic
on a strip 9£ € (A_,A + ), and admits the continuous extension up to the
boundary of the strip.
b) By the residue theorem, we have for T\ € (w_, 9£) and TI £ (9£, w+)
2yr
\J-oo+iTi J-oo+iT2 ) S — »7
+0O+iT2
i /• 6(T/) ,
27T,/_ 0 0 . H r 2 ^ — 77
= 6(0-Mfl.
Hence, exp 6+ (£) exp 6_ (£) = £?(£), and Eq. (3.71) is immediate on a nar-
row strip a;_ < 9f£ < u>+; on a wider strip 9£ € [A_,A+], it holds by
construction: see Eq. (3.69) and Eq. (3.70).
c) By using Eq. (3.68), we obtain
|tf + ir- vr'KT,)] < c(i +1£ - tjD-î + N)- p .

By considering separately a region, where |£ - 771 > |£|/2, and its comple-
ment, it is easy to show that the RHS admits an upper bound via
C i ( i + |£|)-'»(i + h i ) " 1 - " 1 + d ( i + | ^ | ) - p l ( i + II - 1 ? ! ) - 1 - " 1 ,

where pi = min{l, p}/2 > 0. By integrating, we obtain for £ in a half-plane
S£ > UJ- (see Eq. (3.66)-Eq. (3.67))
IM0I<C 3 (1 + Kir pi , (3-77)

and Eq. (3.72)-Eq. (3.75) follow from Eq. (3.77) and the definition of a±.
d) Notice that a±, l / a ± , </>* and l/<f>f are bounded by a polynomial
in the half-plane ± 9 £ > ±w T , therefore, by comparing Eq. (2.23) and
Eq. (3.71), we conclude that a±4>^ is holomorphic, polynomially bounded
and non-vanishing on the complex plane. By the Liouville theorem, this is
constant, and taking into account that <t>^(0) = 1, we obtain Eq. (3.76). •
3.6.3 Approximate formulas for the factors in the case of

NIG, HP, K0B0L and NTS Levy processes
We can write these formulas down for both representations (in Eq. (2.23)
and Eq. (3.71)). In the case of the former, the argument and formulas
are shorter. We use Eq. (3.57) and Lemma 3.7 to transform the line of
integration into the contour along the banks of the cut (-ioo,iA_]. In
empirical studies of financial markets, A+ and - A _ are usually large, of
order 20-50, and then for typical values of other parameters, both roots
—i/3± in Lemma 3.5 exist. Therefore, in the process of transformation, the

contour crosses the simple pole at T/ = —i/3+. By the residue theorem, we
obtain, for £ in the upper half-plane,
-i0+
</\7(0 = ex
P In-
-i/?+ - £ + K(0
where
tp'jiz - 0) _ ip'(iz + 0) -z — i£
* + ( 0 = (27T)"1 / " In dz.
J —{ q + ip{iz - 0) q + tp(iz + 0) —z
(3.78)
Thus,
0+
#(0 : exp $+(£)•
(3.79)
Similarly, from Eq. (3.59), we deduce, for £ in the lower half-plane,
-P-
4>7(0 = exp*,(0, (3.80)
-P- + tf
where
r+oo i\>'{iz - 0) il>'{iz + 0)
* r ( 0 = (2 In dz.
Jx+ [q + ip(iz-0) q + ip(iz + 0)
(3.81)
If —A_ (resp., A+) is large, |$2"(f)| (resp., | $ J ( 0 I ) is small uniformly in
£ in the upper (resp., lower) half-plane, which can be easily seen from the
explicit formulas for the characteristic exponents and Eq. (3.78) (resp.,
Eq. (3.81)). Hence, we may calculate the integrals in Eq. (3.78) and
Eq. (3.81) with a large relative error and still obtain </>£(£) from Eq. (3.79)
and 0~(£) from Eq. (3.80) with good accuracy. This observation can be
used to develop effective numerical procedures. In fact, even the simple
approximations
P+ .-,» -0- (3.82)

<(fl 47(0
produce errors of several percent only, for many typical parameters values.
In Section 5.2, we demonstrate how Eq. (3.80)-Eq. (3.82) can be applied
to obtain simple approximate formulas for the perpetual American put.
Since the (approximate) factors in Eq. (3.82) look exactly as the factors in
the Gaussian case, the comparison of Eq. (2.28) and Eq. (3.82) provides
an analytical explanation why a simple adjustment of parameters of the

Gaussian model can give fairly good fit even in an apparently non-Gaussian
situation.
Chapter 4
Pricing and hedging of contingent

claims of European type
We start with several constructions of an equivalent martingale measure

(EMM). Then, assuming that an EMM is chosen, we derive the generaliza-
tions of the Black-Scholes pricing formula and the Black-Scholes equation
(GBE). We discuss properties of GBE, and study the dependence of GBE
on a choice of EMM. In the last section, we derive an explicit analytic
formula for the weights of the locally risk-minimizing portfolio.
4.1 Equivalent Martingale Measures in a Levy market
Consider a continuous-time model market of a riskless bond, the riskless

rate of return being r > 0, and a risky stock. Suppose that the price of the
stock evolves as S(t) = expX(t), where X is a Levy process under the his-
toric measure P . From general results due to Delbaen and Schachermayer
(1994), it follows that no-arbitrage pricing is possible under any EMM Q,
which is absolutely continuous with respect to the historic measure. If X
is a non-Gaussian process, then typically an EMM is not unique, and the
market is incomplete. Moreover, often there are infinitely many different
EMM, and the choice of an EMM is a non-trivial problem, especially be-
cause of the general result due to Eberlein and Jacod (1997), who showed
for wide class of processes, that by choosing an appropriate EMM, one can
obtain any price of a European contingent claim, which belongs to the ev-
ident no-arbitrage interval. For instance, they showed for a European call
option with strike K, the expiry date T and the constant rate r, that the
values of the option computed under different EMM, span the whole inter-
val from (5(0) — e~rT)+ to 5(0), which is the smallest interval in which all
97
98 Pricing and hedging of contingent claims of European type
prices must lie due to the no-arbitrage considerations.

Follmer and Schweizer (1991) used a hedging argument to advocate a
choice of the EMM which is closest to the historic measure in a relative
entropy sense / log(dQ/dP)dQ. Another approach is to use the optimisa-
tion argument see e.g. Naik and Lee (1990), Keller (1997), Kallsen (1998,
2000), Goll and Kallsen (2000) Goll and Riischendorf (2000). In particular,
in Naik and Lee (1990) and Keller (1997), it is shown that the Esscher trans-
form of the historic measure is optimal w.r.t. some optimisation criterion.
The Esscher transform was introduced by F. Esscher in 1932 in Actuarial
science; see the general discussion in Shiryaev (1999). Madan and Milne
(1991) were the first to use it in Finance, it was also used in Eberlein et al.
(1998) and many other empirical studies of financial markets.
We start with the Esscher transform of the historic measure P , and then
discuss more general constructions. The reader may safely skip the general
case for the first reading, since all the results in the following sections and
chapters are formulated for any EMM under assumption that the choice has
already been made (by using observed prices on the stock and derivatives,
say). All the numerical examples but one are produced for the Esscher
transform.
4.1.1 Esscher transform

Let X be a Levy process on a filtered probability space (fi, T, (Ft), P ) . We
look for an EMM Q satisfying
^|*=exp[0X(t)-d(M)], (4-1)
where d(6, t) is a constant. Thus, the density process of the Esscher trans-
form enjoys the following special property: it depends on the current value
X(u), t) but not on the whole prehistory of the process, that is, on X(u, s), s 6
[0, r]; for a general density process, this property may fail. The discounted
price process of the stock S*(t) — e~rtS(t) must be a martingale under Q,
therefore for t > 0, we have
5(0) = E Q [5*(t)] = S(0)Ep[e^+6)x{t)-d{e't)-rt],
where E® (resp., Ep) is the expectation operator under Q (resp., P ) . By

using the definition of the characteristic exponent of a Levy process, we
Equivalent Martingale Measures in a Levy market 99
conclude that for any t > 0
-tTpp(-i(l + 6)) - d(8, t)-rt = 0. (4.2)
By applying the same consideration to the riskless bond with the dynamics
B(t) = B{0)ert, we obtain
- # P H 0 ) - d ( 0 , t ) = O. (4.3)
p
Prom Eq. (4.3), we find d(8, t) = —ttp (—id), and by substituting into
Eq. (4.2) and dividing by t, we obtain the equation for 8:
-r-xpp(-i(l + e))+ipp{-i6)=0. (4.4)
Let a solution to Eq. (4.4) exist. Then the Esscher transform exists, and
Eq. (4.1) can be written as
^\rt=exp[eX(t) + WP(-i6)}.
We conclude that the characteristic exponent of X under Q is given by
^Q(0 = ^ P ( e - ^ ) - ^ P ( - ^ ) - (4.5)
E x a m p l e 4.1 Let W(t) be the standard Brownian Motion, and let X(t) =
fit + aW(t) be a Brownian Motion with a drift.
Then </>p(£) = ^ £ 2 - */•»£> and Eq. (4.4) turns into
r = - ( - t ( l + 8)a)2/2 + n{l + 9) + (-i6a)2/2 - fi$. (4.6)
Eq. (4.6) has a unique solution 6 — -(/x + CT2/2 - r)/a2, and Eq. (4.5)
gives the well-known formula for the characteristic exponent of a Brownian
motion under the equivalent martingale measure:
tfQ(fl = y £ a - t f ( r - Y ) . (4-7)
Eq. (4.7) can also be obtained from Girsanov's theorem.
For an arbitrary Levy process, Eq. (4.4) may have no real solutions,
and hence, the Esscher transform may not exist. The following lemma
shows that if it exists, it is unique, and provides a necessary and sufficient
condition for the existence of the Esscher transform.
L e m m a 4.1 Let X be of exponential type [A_,A+] with X+ — A_ > 1,

under the historic measure P .
Then
(i) the function f(9) = -ipp (-i(9+1))+ipp (-i9) is strictly increasing

on [-A+,—A_ — 1];
(ii) Eq. (4-4) has at most one root on [—A+,— A_ — 1];
(Hi) Eq. (4-4) has a root on (—A+, —A_ — 1) if and only if
lira f(9) <r< lim f(9). (4.8)

0->-A++O 0->—A--1-0
Proof. Parts (ii) and (iii) follow from (i), and to prove (i), it suffices to
show that f'(9) > 0, Vffe (-A+, - A _ - 1). Let pi(x)dx be the law of Xi.
Set
-t-oo
x 6 e OI pi(x)da;,
/ -OO
F(a) = In 1(0, a),
and represent / in a form f(a) = F(a + 1) — F(a). Clearly, it suffices to

show that F' increases, or even simpler:
F"(6)>0, V0€(-A+,-A_-l). (4-9)
We have
+oo p+oo
F"(6) =
/ oo x2e0x (x)dx
Pl / e8xPl(x)dx
-a
=
+oo
-oo
xe0xpi(x)dx
J—oo
x
r /.+oo
/
U -oo
(I(2,a)I(0,a)-I(l,a)2)/I(0.a)2,
e0xpi(x)dx
and by the Cauchy-Schwarz inequality, I(2,a)I(0,a) > 1(1,a)2. Since x

and 1 are linearly independent, the last inequality is strict, and Eq. (4.9)
and Lemma obtain. D
Notice that condition Eq. (4.8) is easy to verify once an explicit formula for
ipp is chosen. For processes in the empirical studies of financial markets,
Eq. (4.8) is usually satisfied.
Equivalent Martingale Measures in a L€vy market 101
4.1.2 General case

We assume that X is a purely discontinuous Levy process under P , with
the generating triplet (0,b,U(dx)). For a similar result for processes with
the non-trivial Gaussian component, see Proposition 2.19 in Raible (2000).
Following Eberlein and Jacod (1997), we write Mr for the (possibly
empty) class of measures locally equivalent to P , under which e~TtS(t) is
a martingale, and M!T for a subclass of MT under which X is again a Levy
process; 34,r(n(efa:)) denotes a class of functions y : R ->• (0,+00) such
that
+00 /•
/
(y/yjx) - l) 2 II(da;) + / (ex - l)y(x)U{dx) < +00, (4.10)
-00 J{x>l]
and
+00
/ ((e* - l)j,(ar) - xl{_1A](x))U(dx). (4.11)

-OO
Theorem 4.1 a) If ybtT{Tl(dx)) = 0, then MT = M'T = 0.

b) If 34,r(H(dx)) 7^ 0, then both M.T and M'r are non-empty, and for
each y € ^y&,r(II(<ia;)), there is a measure Q € M.'r under which X is again
a Levy process with the generating triplet (0,b',ll'(dx)), where
b' = b+ f x(y(x) - l)n(dar), (4.12)
and
II'(A) = / y(x)U(dx); (4.13)

JA
conversely, if Q € M'r is the measure under which X is a Levy process,
then its generating triplet is (0,b',Tl'(dx)), where b' and TV (dx) are given
by Eg. (4.12) and Eg. (4.13).
c) Let y and Q be as in b). Then the characteristic exponents of X
under P and Q are related by
+00
(l-e^)(y(x)-l)U(dx), (4.14)
/
-00
and Eq. (4-12) can be written as
r + ipQ(-i)=0. (4.15)
Proof, a) and the first part of b) are Proposition 1 in Eberlein and Jacod
(1997); the converse part of b) is from Proposition 2.19 in Raible (2000).
Notice that the integral in Eq. (4.12) converges due to the Cauchy-Schwarz
inequality:
(J x(y(x) - l)H(dx)) <j {y/y~(x)-\)2Il{dx) J {y/^{x)+lfx2TL{dx).
The first factor in the RHS is finite due to Eq. (4.10), and the second factor
is bounded since both II(dx) and y{x)Xi{dx) are Levy measures,
c) We use Eq. (4.12) and Eq. (4.13)
+oo
(l-e»«+^l[.1,1](i)Mz)n(.fa)
/ -oo
/+oo
(1 - eixi + iz£l[-i,i](a;))n(dx)
-OO
+oo
(1 - e*t)(y(x) - l)Tl(dx)
/ -OO
+oo
(1 - e^)(y(x) - l)U(dx),
/ -OO
and Eq. (4.14) follows. Eq. (4.15) is immediate. •
It is straightforward to deduce from Theorem 4.1 the following description

of a wide class of EMM for an RLPE.
T h e o r e m 4.2 Let X be an RLPE in the sense of the first definition, of

order v > 0 and exponential type [A_, A+], where A_ < — 1 < 0 < A + .
Let 7 > v/2, fi- < — 1 < 0 < )U+, and let y be a positive measurable
function which satisfies Eq. (4-11) and the following conditions:
\y{x) - 1| < C\xP, \x\ < 1; (4.16)

( A++M+)x
y{x) < Ce - , x < _i; (4.17)
y(x) < Ce(-A"+^-)x, x > 1. (4.18)
Then
(i) there exists a measure Q £ M'r under which X is an RLPE of

order v and exponential type [fi-,/j,+], with the generating triplet
(0,6MT(dx));
Equivalent Martingale Measures in a Levy market 103
(it) the characteristic exponent of X under Q is given by Eq. (4-14)-

It is holomorphic in the strip 3£ G (/i_,/x + ) and continuous up to
the boundary of the strip, and admits a representation
^ Q ( O = - « X + 0Q(O, (4-19)
where fi is the same as \i in Eq. (3-45) for V»p, and <jfi enjoys the
property Eq. (3.46):
(Hi) there exists v-^<v such that as £ —> oo in the strip 9£ G [fj,-,fi+],
0Q(O=cK|"+O(|Sr), (4.20).
where c is the same as c in Eq. (3.46) for if)p.
Proof. From Eq. (4.16)-Eq. (4.18) and the conditions on the rate of the
decay of the Levy density II(dx), we see that Eq. (4.10) holds, and (i)
follows from Theorem 4.1. In view of Eq. (3.45)-Eq. (3.46) and Eq. (4.14),
it suffices to show that the RHS in Eq. (4.14) is holomorphic in the strip
9£ € [/*_,/*+], and satisfies an estimate
+oo
(I - eix^(y(x) - l)U(dx) o(ier2). (4.2i)
!/
as £ -> oo in the strip. The analyticity in the strip has been discusses
already, and Eq. (4.21) is proved by using Eq. (4.16) in the same way as
the conditions of the second definition of RLPE were deduced from the
conditions of the first one. (See Chapter 3). D
It is easy to construct many functions y which satisfy the conditions of

Theorem 4.2, thereby constructing many examples of EMM.
Example 4.2 Let X be a KoBoL of order v > 0, with steepness param-

eters A_ < - 1 < 0 < A+. Take a,d > 0, 7 G (v/2,v), and to give more
weight to negative jumps, choose /i+ G (0, X+), and set
y(x) = eax + dx!e ( M +- A +^.
(The second term makes the left tail of the Levy density fatter). Then
Eq. (4.10) holds, and direct computations (essentially the same as in the
calculation of a K0B0L process in Section 3.1) show that
V>Q(0 =rpP(Z- to) - V< P Ha) + cdT(7 - v)(,?-i - ( M + + %)"-''). (4.22)

If
r + i>p(-i - ia) - V- P (-m) + cdT(j - v){^ - (JJL+ + l ) " " 7 ) = 0, (4.23)
then Eq. (4.11) holds, and Q is an EMM.

If Eq. (4.23) has a solution, then generically, the manifold of solutions
is two-dimensional, and hence, we obtain a two-parameter family of EMM.
By taking more terms in the definition of y, we can obtain EMM depending
on more parameters.
4.2 Pricing of European options and the generalized Black-

Scholes formula
Prom now on, we assume that an EMM Q is chosen, and that X is an

RLPE under EMM, of order v > 0 and exponential type [A_, A+]. We will
assume that A_ < — 1 < 0 < A + , which is close to the necessary conditions:
A_ < - 1 is necessary for the stock to be priced, and A+ > 0 holds for
any exponential Levy process. The case A_ = — 1,A+ = 0 can also be
considered but at the expense of more technical and lengthy constructions.
To simplify the notation, we denote the expectation operator and char-
acteristic exponent under Q by E and ip, respectively, the only exception
being the section on the locally risk-minimizing hedging, where we will use
both the historic measure and an EMM.
4.2.1 Pricing formulas: convolution with the pricing kernel

and the Fourier-inversion formula
Let g(Xr) be the (non-negative) terminal payoff for a contingent claim with
the expiry date T. Then at time t <T, the no-arbitrage price of the claim
can be found from
e~rtF{Su t) = E[e-rTg(XT) | Xt = x], (4.24)
where x = In St- Let an explicit formula for the probability density of

X under Q be known as it is the case with NIG. Then we deduce from
Eq. (4.24)
+oo
Pr(y)g(x + y)dy,
/ -oo
Pricing of European options and the generalized Black-Scholes formula 105
where r = T—t is the time to the expiry; by the change of variable y i-» y—x,
+oo
Pr(y-x)g(y)dy. (4.25)
/
-oo
In the general case, pT can be expressed in terms of the characteristic ex-

ponent, by using the Fourier transform
+oo
e-«€-r*(Odf| (4.26)
/
-oo
and Eq. (4.25) assumes the form

+O0 /-(-OO
/ J(x-»K-T<>r++Wg{y)dZdy. (4.27)
/
-oo J — oo
Let g be measurable, and let there exist a € (A_, A+) such that the function
ga defined by ga{x) := e<rxg(x) belongs to the class Z-i(R). Then we can
define
+oo
e-^g(y)dy
/ -oo
on the line $J£ = a. On the strength of Eq. (3.48), we can apply the
Cauchy theorem and shift the line of the integration in the inner integral
in Eq. (4.27):
+oo p+oo+itr
/ -oo J— oo-\-i(T
Now the integral converges absolutely, and we can change the order of
integration and obtain
-f-00+2<7
e ^ ~ r ( r + ^ ) ) m d t ( 4 2 8 )
/
-oo+i<T
x
Set f(x,t) = F(e ,t). Then by using the notation of PDO, we can write
Eq. (4.28) as
f{x, t) = e x p [ - r ( r + il>(Dx))]g{x). (4.29)

4.2.2 Call and put options
For the call option with the strike price K, the terminal payoff is g(Xx) =
(eXr - K)+, and hence ga e Li(R) for any a < - 1 . Since A_ < - 1 , we
can choose a £ (A_, - 1 ) , and apply Eq. (4.28). The direct calculation gives
r+oo
e - » V - K)+dx
•oo
f+OO
= / ( e ^(i-'«) _ Ke-*x)dx
UnK
e(l-i£)\nK e(l-i£)\nK
Ke-iSlnK
(4.30)
and by substituting into Eq. (4.28), we obtain the pricing formula for the
European call option:
FcMSut) - - ^ L ^ ^ - 5 *, (4-31)
where a € ( A _ , - l ) is arbitrary. This is the generalization of the Black-

Scholes formula for calls.
Similarly, for the put option with the strike price K, the terminal payoff
is g(XT) - (K-eXr)+, and hence ga € Li(R) for any a > 0. Since A+ > 0,
we can choose a £ (0,A + ), and apply Eq. (4.28). The result is
*Put(i>ut) --— df, (4.32)

27T J_00+ia2 f(f + 0
where a2 G (0, A + ) is arbitrary.
The integrals in Eq. (4.31)-Eq. (4.32) can be calculated by using the
Fast Fourier Transform (FFT), its modifications, and other methods. We
consider these methods in Chapter 12.
Since the integrand in Eq. (4.31)-Eq. (4.32) is meromorphic in the strip
3£ G [CT, a2] with two poles at £ = —i and £ = 0 and decays as £ -» oo in
the strip, we may shift the line of the integration in Eq. (4.31) up. The line
of integration crosses the two poles, and the residue theorem gives
K
F (c ts _ ,êMi(-i) HSt/K) - r(r + y>(-»))]
^c3.n{Jt,t) = - — (2717) :
Z7T —I
K exp[»01n(,S t /JiO - r{r + V>(0)))
+Fput{St,t).
By taking into account that V>(0) = 0 and r + rp(—i) — 0, since we price the
options under an EMM, we obtain the so-called put-call parity relation
FcMSut) = St- Ke~rT + Fput(St,t). (4.33)
Of course, the put-call parity relation can be deduced in a much simpler

way, and it is valid for no-arbitrage pricing under any process, for which
the option prices are finite: it suffices to notice that the difference between
the payoff of the call and the payoff of the put, with the same strike price,
is equal to ST — K, and apply Eq. (4.24).
4.2.3 Numerical examples: the comparison with the Black-

Scholes formula, and the shapes of the smile
First, we give several series of examples for put pricing under KoBoL pro-
cesses, when an EMM is the Esscher transform of the historic measure. The
characteristic exponent of the process under the historic measure is of the
form
HO = -irf + cT(-u)[X+^ - (A+ + iOv + ( - A - ) " - (-A_ - i£)%
where fi € R, v £ (0,2],i/ ^ l,c > 0, and A+ > 0 > A_ (in the case
v = 1, the formula for the characteristic exponent is different: see Eq. (3.8)).
When v = 2, we obtain a Brownian motion. Parameters are chosen so that
the Esscher transform exists. In each series, we vary the strike-price ratio
S/K, and one of the parameters, the remaining ones being fixed; the spot
price of the stock is 100. To facilitate the comparison with the Black-Scholes
formula, in each series we fix mi = iip'(Q) and m 2 = i>"{0). On the left
panel, we plot the difference between the Black-Scholes price and the KoBoL
price of the put, and on the right panel, we plot the implied volatility.
Were the Black-Scholes model correct, the implied volatility would have
been equal to m 2 , and the implied volatility surface flat. The reader is
advised to compare the pictures below with similar pictures in Eberlein
et al. (1998) and Eberlein and Prause (2001) for option pricing under
Hyperbolic Processes, Normal Inverse Gaussian processes and Generalized
Hyperbolic processes.
N_-__ ^ v v 1 , 2 v. ,. c , 1.1
T(daya) ° OS °-9 T (days) ° o.e °9
S/K S/K
Fig. 4.1 Black-Scholes price vs. KoBoL price: dependence on the moneyness and time
to expiry. Parameters: r = 0.05, mi = 0.12, m 2 = 0.18,1/ = 1.4, A_ = -24, A+ = 20.
Fig. 4.2 Black-Scholes price vs. KoBoL price: dependence on the moneyness and
the left steepness parameter, A+. Parameters: r = 0.05,mi = 0.12,7712 = 0.18,v —
1.4, \- = -24, T= 1 (day).
In Fig. 4.1, it is easily seen that the volatility smile becomes more
pronounced as r -> 0. Fig. 4.2 demonstrates that in the region where
the steepness parameters are large, the dependence of the price on these
parameters is weak. This is the good news since the steepness parameters
Fig. 4.3 Black-Scholes price vs. KoBoL price: dependence on the moneyness and the
order of the process, v. Parameters: r = 0.05, m i = 0.12,7712 = 0.18, A_ = — 24, A+ =
20, T = 1 (day).
are especially difficult to infer from the data with a reasonable degree of
accuracy. In the region where the steepness parameters are small, the
dependence of the price on these parameters can be quite sizable. Since A+
and — A_ do not differ much, the smile is almost symmetric. By playing
with steepness parameters, it is not difficult to obtain smiles of different
shapes.
Fig. 4.3 shows that the order of the process, v, is the most important
parameter. As v —> 2 and/or min{A+,— A_} —• +00 so that mi and 7712
remain fixed, the KoBoL price converges to the Black-Scholes price; the
difference is especially large for small v and/or A+ and - A _ .
In the last series of numerical examples, we consider the process which
is characterized by the same parameters as in Fig. 4.1 (in particular, u —
1.4, A+ = 20, A_ = —24), but instead of the Esscher transform, we use the
family of EMM constructed in Example 4.2, with fi+ = 0.5,7 = li a n d
a = 9,6 + 0.05, ...,0 + 0.3. For each value of a, we find d so that the EMM-
condition r + ^(—i) = 0 holds. Thus, a = 6 gives the Esscher transform.
The larger the a, the more weight the EMM places on extreme events.
By comparing Fig. 4.1-4.3 with Fig. 4.4, we see that it is easier to
obtain asymmetric smiles with EMM different from the Esscher transform.
Fig. 4.4 Black-Scholes price vs. KoBoL price: EMM from Example 4.2. Parameters:
r - 0.05,mi = 0.12, m 2 = 0.18,1/ = 1.4, A_ = - 2 4 , A+ = 2 0 , ^ + = 0.5,7 = 1,T = 1
(day).
4.2.4 The problem of the model fitting and evaluation

Once a model for the observed process (i.e., for the historic measure and for
an EMM are chosen) one has to fit the parameters of the historic measure
and EMM to data. Different methods can be used:
• all the parameters can be obtained by the maximum likelihood
method or the method of least squares;
• some of the parameters can be determined separately by using other
considerations, and the rest are determined by the maximum like-
lihood method, etc.
Clearly, it is easier to fit the parameters of the historic measure to data, and
the parameters of the historic measure can be used to find some parameters
of an EMM, due to Theorem 4.1, which implies some relationship between
sets of the parameters of P and Q. In particular, if X is an RLPE under
P , of order v 6 (0,2), then it is an RLPE under Q, of the same order, and
with the same constant c > 0 in Eq. (3.46). Should one decide to look for
P and Q in the same model class of RLPE, then after the parameters of P
are fitted to the historic data and the EMM-condition r + ip^(—i) = 0 is
taken into account, not many free parameters remain. For instance, in the
case of KoBoL family, with c and v fixed, the EMM condition determine
the line of admissible points in the (/x, A_, A+)-plane. In the case of GHP,
Generalized Black-Scholes equation 111
not only the order 1 and the parameter S(= c in the general definition of
RLPE) are the same for P and Q, but the drift parameter \i as well (see
Raible (2000), Proposition 2.20). Since the class GHP is a 5-parametric
family, the EMM-condition leaves 2 free parameters. If one wishes to work
with a subclass of NIG, only one free parameter remains.
Finally, we would like to notice that by using the asymptotic formulas
for the prices of the up-and-out and down-and-out options near the barrier,
which we obtain in Chapter 7 (or similar formulas for barrier options),
and the observed prices of these options near the barrier, we can infer
whether there is a diffusion component or not. Further, if there is none and
hence, the process is of order v 6 (0,2), then one can obtain the following
information about the EMM chosen by the market without using the data
on the stock itself: first, whether v belongs to (1,2), (0,1) or {1}, and then
(i) if v e (1,2), then to find v;
(ii) if v — 1, then to obtain a simple relation between the parameters c
and (j, in Eq. (3.45)-Eq. (3.46);
(iii) if v € (0,1), then decide whether the drift /x = 0 o r not, and
• if fi = 0, then to find v\
• if n ^ 0, then to find the sign of fi.
Thus, the data on the touch-and-out options and barrier options near the
barrier can be used to decide whether the chosen class of processes is suit-
able for pricing of options on a given stock or not.
For more details, see the next section.
4.3 Generalized Black-Scholes equation and its properties

for different R L P E and different choices of E M M , and
implications for parameter fitting
Differentiate f(x,t) = F(ex,t) w.r.t. t by using Eq. (4.28) in the domain

t < T, x € R; due to Eq. (3.48), we may differentiate under the integral
sign and obtain, by using the definition of PDO:
%(x,t) = (27T)-1 / (r-nK0)ete£-r(r+*(<»gm

+oo+io
e «€-r(r-HM£))- ( £ )d £
/
= (r + 1>(Dx))f(x,t).
We conclude that the price / of a contingent claim of the European type
satisfies the following equation
{dt-{r + iPQ{Dx)))f(x,t)=0, t<T,xeK. (4.34)
This is the generalization of the Black-Scholes equation. It can be deduced

formally, by differentiating Eq. (4.29).
In the Introduction, we have demonstrated that if X is a Brownian
motion then Eq. (4.34) becomes the Black-Scholes equation. Notice that
we can write Eq. (4.34) in the form
(r - LQ)f(x, t) = 0, t<T, xeR, (4.35)
where L Q — dt + £ Q , and L Q = —rpCi(Dx) is the infinitesimal generator of

X under Q.
In Eq. (4.34), we have explicitly indicated an EMM Q in order to dis-
cuss the dependence of properties of the operator in the generalized Black-
Scholes equation on a choice of an EMM Q. Prom Eq. (4.19)-Eq. (4.20) and
Eq. (3.45)-Eq. (3.46), we conclude that the difference V Q (£) - V^CO grows
slower than 5fr/>p(£), as £ -+ oo in the common strip of the analyticity. In
the language of the theory of PDO, this means that the replacement of ipp
with V' is a weak perturbation, and the local properties of the operator
remain essentially the same. In particular, we will see that the orders of
the factors in the Wiener-Hopf factorization formula are independent of the
choice of EMM, and derive from this observation that the leading term of
the price of a barrier option (or a touch-and-out option) near the barrier is
independent of a choice of EMM. This property can be used to infer some
relations for parameters (v, c) (and in the case v < 1, for /x as well) of an
RLPE from observed prices of barrier options and touch-and-out options.
General properties of the Generalized Black-Scholes equation depend on
the order of the process as follows: if v > 1 or v < 1 but the drift \x = 0,
then the stationary part of ^ is strongly elliptic, and it is not if v < 1 and
H ^ 0. In the latter case, the price of a barrier option (and a touch-and-out
option) may be discontinuous at the barrier, whereas in the former case,
these prices are always continuous up to the boundary. Thus, by using the
empirical data, it is possible to decide whether processes of order v < 1
with the non-zero drift are suitable for modelling purposes at all. From
the point of view of the theory of PDO, strongly elliptic operators enjoy
Other European options 113
much more favourable properties than non-strongly-elliptic ones, and so it

is advisable to use processes of order v > 1 or processes of order v < 1 with
zero drift.
4.4 Other European options
We consider several examples; the reader can easily calculate the prices of
other options.
4.4.1 Power options

Consider a call-like option with the payoff (ST — K)+, where /? > 0 is
typically equal to 2. In order that the price be finite, we need to assume
that /? < — A_. Assume for simplicity that the inequality is strict; then the
calculations are essentially the same as for the call option above, and the
result is
F(Sttt) = - - J ^ ^ - ^ de, (4.36)
for any a e (A_, —0).
4.4.2 Digital options

For the digital call option with the strike price K, the terminal payoff is 1,
if ST > K, and 0 otherwise. Hence, g(x) = 1{K^+00)(X), and for any a < 0,
and 9£ = a,
+oo
e-^g(x)dx
/ -OO
r+oo
= / e~ixtdx
J\nK
_ exp[—i£,lnK]
We choose a € (A_,0), and substitute into Eq. (4.28):

+oo+iff
/
Similarly, for the digital put option with the strike price K, the terminal
payoff is 1, if ST < K, and 0 otherwise, and we obtain
+00+t<72
eiSMSr/K)^-!^ (438)
/
-00+l<72
for any er2 € (0, A+).
4.4.3 Combinations
There are several types of options whose payoffs are linear combinations
of various calls, puts, and constants. For instance, a straddle is the option
with the payoff of the form
g(ST) = \ST-K\-C = (K- ST)+ + (ST - K)+ - C,
therefore the price of the straddle is
Straddle(S t ,t) = FcaU(St,t) + Fput(St,t) - e~rTC, (4.39)
and an explicit expression can be obtained by substituting Eq. (4.31)-

Eq. (4.32) into Eq. (4.39).
Similarly, the reader can calculate prices of other combinations. Typ-
ically, the payoffs are piece-wise linear in S, though power analogues can
also be considered.
4.4.4 General case: the condition on the rate of growth of

the payoff at the infinity and the origin
Generically, the payoff may be non-zero in neighbourhoods of both ST — 0
and ST — +oo, and there may be no a G (A_, A + ) such that pCT(-) = e"'g(-)
is of the class L i ( R ) . In this case, we assume that g is measurable (one
is tempted to say "continuous" but there are options with discontinuous
payoffs), and that there exist w_,w+ € (A_, A + ) such that
e^-lR±(-M-)€Li(R).
For any XQ, decompose the terminal payoff g in the sum of the following
two functions
g+{x0;x) = l[x0i+oo)(x)g{x), g~(x0;x) = l ( _ 00ia!o] (a;)g(a;),

Hedging 115
and calculate the option price
F(g;Sut) = F(g+;St,t) + F(g-;St,t) (4.40)

+oo+iu-
/
-OO+tUI-
•oo+iu+
In order that the integrals converge better, it is advisable to choose xo to

be one of the zeroes of g, if g is continuous; if it is not, a point of the
discontinuity is a reasonable choice.
4.5 Hedging
4.5.1 General discussion

In the Black-Scholes model of the complete market, without frictions and
constraints, the hedging problem is solved easily, and the risk can be com-
pletely hedged away. If the market is incomplete then the problem of the
risk minimization, hence, the problem of the choice of an appropriate mea-
sure of risk arise. We derive the explicit formula for the weights of the locally
risk-minimizing portfolio, and we do not require that the portfolio be self-
financed; the idea of the locally risk-minimization is due to M. Schweizer
(see Follmer and Schweizer (1991)). This type of hedging is by no means
the only one. The mean-variance approach to hedging is due to Follmer
and Sondermann (1986). They considered the optimisation problem
mmE[(H-Vr'6)2}
where the contingent claim H is an .Fr-measurable square integrable ran-

dom variable, V£' is the terminal value of a self-financed portfolio of
the riskless bond and a risky asset with initial capital x and quantity 6
invested in the risky asset. Follmer and Sondermann (1986) solved the
mean-variance problem in the martingale case, by the direct application of
the Galtchouk-Kunita-Watanabe projection theorem. In the semimartin-
gale case, this quadratic optimisation problem is solved by Gourieroux
et al. (1996). For further references and the extensive discussion of the
mean-variance hedging, see Musiela and Rutkowski (1997) and Pham et
al. (1998), and many other approaches to hedging, with variations, can be
found in the references at the end of the chapter. For hedging under pro-
cesses of Koponen's family, see Bouchaud and Potters (2000) and Matacz
(2000).
We have chosen the locally-risk minimizing hedging in order to demon-
strate in the simplest situation how the PDO-technique allows one to obtain
analytical results.
4.5.2 Locally risk-minimizing hedging ratio

Suppose that an investor can go short in the money account, and consider
the following problem: given the investor's wealth W(t) at time t < T,
where T is the expiration date of an option, construct a portfolio short
one option, and long 6(t) shares of the stock, with the minimal expected
variance the next infinitesimal moment t + At < T; the residual
wQ(t) := W[t) + F(S(t),t) - d(t)S(t) (4.41)
is invested into the money account.

The price of the portfolio at time t + At is
W(t + At) = -F(S(t + At), t + At) + 6(t)S(t + At) + ertw0(t);
we calculate the variance of wealth under P conditioned on the information

available at the date t. Let Ef be the corresponding expectation operator.
Then
Ef[{W{t + At)-Ef[W{t + At)\)2)

= Ef[{F{S{t + At),t + At) - Ef[F(S{t + At),t + At)})2]
-20(t)E?[(F(S{t + At),t + At) - Ef[F(S{t + At),t + At)])
x(S(t + At) - E?[S(t + At)])]
+6(t) E?[(S(t + At) - E?[S(t + At)])2].
2
Minimizing w.r.t. 0(t), we obtain the following first order condition:
where
SS(t,At) = Ef[S(t + At)2} - Ef[S(t + At)}2,

Hedging 117
and
FS{t,At) = Ep[F{S{t + At),t + At)S{t + At)]

-Ef[F(S(t + At),t + At)]Ep[S(t + At)}.
Suppose that under P , X is of exponential type [A_,A + ] with A_ < - 2 ;

then SS(t,At) < +00, and an explicit formula for SS(t, At) is immediate
from the definition of the characteristic exponent
SS(t, At) = ( e - ^ P ( ~ 2 i ) _ e -2AtV P (-i)) 5(t)2
By expanding into the Taylor series around At = 0, we obtain
SS(t, At) = (-i/> p (-2i) + 2^p{-i))S(t)2At + 0(At2). (4.43)
To obtain a similar representation for FS(t,At), we need an additional

restriction on [A_, A+], the exponential type of X under P , and a similar
restriction on [/u_, fi+], the exponential type of X under Q. Both depend on
the rate of growth of the terminal payoff g(Xr) of the option at the infinity.
We assume that for some a £ (/x_, n+), g„ G Li(R), and A_ < a — 1 < A + ;
then the justification of the following calculations is an easy exercise.
First, notice an evident formula for the infinitesimal generator L of the
two-dimensional process {Ht} — {(Xt,t)}, where X is regarded as a Levy
process under P :
L = dt + Lp =dt-ipP(Dx). (4.44)
By using Eq. (4.44), we obtain
Ef[F(S(t + At), t + At)S{t + At)] - F{S(t), t)S(t) (4.45)
= (dt-i>P(Dx))f(x,t)exAt + o(At),
and
Ef[F{S(t + At), t + At)} - F(S(t),t) (4.46)
= (dt-ipp(Dx))f(x,t)At + o(At).
Since
rl>p(D,)exv(x) = exiPP(Dx - i)v{x),

and
Ep[S{t + At)]-S(t) = (exp[-Atipp(-i)]-l)S(t)

= Atipp(-i)ex+o(At),
we derive from Eq. (4.45)-Eq. (4.46) the representation
FS(t,At) = exAt[-^p(Dx-i) + ipp{Dx) (4.47)
+ 4>p{-i)]F(ex,t) + o{At).
Now we can substitute the result and Eq. (4.43) into Eq. (4.42), and obtain
d(t, x; At) = e-xBp(Dx)F(ex, t) + O(At),
where x = In S(t), and
B (4 48)
«) _^(_M)+2*P(-0 • -
In the limit At -» +0, we obtain
6(t, x) = e-xBp(Dx)F(ex,t). (4.49)
Example 4.3 Let Xt be a Brownian motion. Then
<KO = - ^ + <72£2/2,
and therefore,
o P ^ _ - « - « - ) 2 + (fl2 + (-0 2 _ 2ji _

" ^ - _(_ 2 j)2 + 2(-i)2 - 2 *•
Hence,
5 P (£> X ) =*(-*£*) = dx,
and
9(t,x) = (dxf)-e-x=dsF(S(t),t).
Thus, we recover the standard Black-Scholes delta-hedging result.

Hedging 119
4.5.3 Continuity of the locally risk-minimizing hedging ra-

tio till the expiry
In the Black-Scholes model, for calls and puts, 9(t, x) is discontinuous for
options at the money, near expiry, i.e., at (t,x) = (T,q), where x = I n 5 ,
q = In K, and S and K are the spot and strike prices, respectively.
Let X be a mixture of independent RLPE of order v < 2, each satisfying
Eq. (3.45)-Eq. (3.46) with vx < 1. Then from Eq. (3.45)-Eq. (3.46), it
follows that
\Bp(£)\<C(l + \S\y, (4.50)
where p = min{i/ — l , ^ i } < 1. By using Eq. (4.31), we conclude that

in the case of a European call option, say, 9 = 9(t,x) is defined for all
(*, x) e [0, T) x R by
K [+00+ia exp[iZMS/K) - r(r + ^Q(Q)}BP(Q Jf

eit x) de
> = ~^sL+i„ WTT) '
where a G (min{/u_,A_ - 1 } , - 1 ) . From Eq. (4.50), it follows that the
integrand is bounded (in modulus) by
Ciu + Kir1-',
where e > 0, and C\ can be chosen the same for each t <T and S/K on any
compact. Since the integrand depends on t continuously, we conclude that
9(t,x) is continuous (in fact, Holder continuous) on (—oo,+oo) x (—oo,T].
This makes the hedging under purely discontinuous RLPE models more
stable than the Gaussian hedging, near expiry; and at the strike, much
more stable indeed.
4.5.4 Numerical Examples

In Fig. 4.5, we plot the KoBoL 9 and the difference between 0's in the
Gaussian and KoBoL models. We assume that the EMM is the Esscher
transform of the historic measure, and we vary the time to expiry, r, from
6 hours to 1 hour. We see that the KoBoL 9 remains essentially stable
whereas the difference with the Black-Scholes 9 tangibly increases.
Fig. 4.5 Black-Scholes 6 vs. KoBoL 8: dependence on the moneyness and time to expiry.
Parameters: r = 0.05, m i = 0.12,m 2 = 0.18,1/= 1.4, A_ = - 2 4 , X+ = 20, K = 100.
4.6 Commentary
The results of this chapter are obtained in Boyarchenko and Levendorskii

(1999a, 2000, 2001a). For the detailed study and empirical results on option
pricing under Hyperbolic Processes and Generalized Hyperbolic Processes,
see Eberlein et al. (1998), Keller (1997), Prause (1998), Raible (2000),
Eberlein and Prause (2001). For different analytical formulas for option
pricing and hedging under non-Gaussian processes, see Bouchaud and Pot-
ters (2000) and Matacz (2000).
Chapter 5
Perpetual American options
In this chapter, we consider several types of perpetual American options.

We start with the American puts and calh, and then consider more general
payoffs. We show that in some cases, the smooth pasting principle fails,
and discuss generalizations of this principle
5.1 The reduction to a free boundary problem for the sta-

tionary generalized Black-Scholes equation
5.1.1 General discussion

Options of the European type can be exercised only at the expiry date, T,
but the owner of the option of the American type has the right to exercise
the option at any time t e [0, T]. If the expiry date is far away, one can use
the idealization T — oo. An American option with the infinite time horizon
is called a perpetual option.
Let the riskless rate r > 0 and the dividend rate qo > 0 of the stock
be fixed, let St = exp Xt be the price process of the stock under a chosen
EMM; we assume that X is an RLPE. Let g(Xt) be the payoff function
of a perpetual American option on the stock. For instance, for a put,
g(x) = K — ex, and for a call, g(x) — ex — K, where K is the strike price;
for the formulation of our results, it is more convenient to use g(x) rather
than m&x{g(x),0}. Clearly, this departure from the conventional notation
is quite admissible since it is non-optimal to exercise the option in the
region g(x) < 0. The option owner chooses the optimal strategy, that is,
the stopping time r, in order to maximize the expected discounted payoff
121
122 Perpetual American options
Ex{e~qTg(Xr)}, where x = In S is the logarithm of the spot price, q — r+q0,

and the expectation is taken under the EMM (see Shiryaev (1999), XVIII,
2). We consider stopping times satisfying 0 < r = r(w) < oo, u £ fi; the
class of such stopping times will be denoted M. Thus, V*{x), the rational
price of the perpetual American option, is given by
K(x)=sup^[e-«rfl(XT)]. (5.1)
M
5.1.2 Free boundary value problem for the price of the per-
petual American option
Suppose that the optimal stopping time is the hitting time of the exterior
of an open set C C R:
r , = inf{t > 0 | Xt £ C}. (5.2)
On the strength of Theorem 3.1 and Lemma 2.4, an RLPE satisfies (ACP)-
condition, hence, in the case of g € Loo(B), where B = R \ C , we can apply
Theorem 2.12 and conclude that V, is a bounded solution to the following
boundary value problem:
{q-L)u{x) = 0, xeC, (5.3)

u(x) = g(x), x & C. (5.4)
where L = —rp{D) is the infinitesimal generator of X (In the case of a non-

bounded g, the approximation of g by bounded functions and the limiting
procedure are needed to justify Eq. (5.3)-Eq. (5.4) - see Remark 2.1.) Notice
that not only the function u is unknown but the boundary dC as well. Thus,
we have a free boundary value problem. In the pure diffusion case, one
finds a candidate for the optimal stopping time Eq. (5.2) or equivalently
a boundary of C by using the smooth fit principle, which says that the
optimal choice of the boundary ensures that u € C 1 , whereas with an
non-optimal choice, u' is discontinuous at the boundary. Notice that this
heuristic principle requires certain regularity of g, and after the boundary
and solution are found, one has to verify that the solution is indeed optimal.
We will see that in the case of RLPE, the smooth pasting principle may
fail, and so we will use other heuristic considerations to find the boundary.
In cases when C is a semi-bounded interval with the boundary h (we denote
the corresponding class of stopping times by .Mo), we obtain an explicit
The reduction to a free boundary problem 123
formula for u = u(h,x), and by analysing the properties of the latter, find
the unique candidate for the optimal h. The verification of the sufficient
optimality conditions in the class MQ is relatively simple: we will be able to
directly compare V — V(h, x) for different h, and check that the candidate
for the optimal boundary is indeed optimal.
5.1.3 Main Lemma

The verification of the optimality in the class M is much more tricky.
Consider the following free boundary value problem.
Given a non-negative continuous function g, find an open set C and a
function V such that Eq. (5.3)-Eq. (5.4) hold, and
V{x) > g(x), xeC; (5.5)

(q-L)V{x) > 0, x<£C. (5.6)
Lemma 5.1 Let (C,V) be a solution to Eq. (5.3)-Eq. (5.6) such that V
is continuous, let r* be the hitting time of C, and let
W : = (q — L)V be universally measurable; (5-7)

U*W = V. (5.8)
Then r* and V* = V solve the optimization problem Eq. (5.1).

In all applications in this book, W will be Borel, so there is no need to give
the definition of a universally measurable function here.
Proof. Due to Eq. (5.3) and Eq. (5.6), W is non-negative, and by Eq. (5.7),
it is universally measurable, therefore for any stopping time T, Eq. (2.40)
holds:
VW(x) = E* f e-qtW(Xt )dt + EX e-

qT
UqW(XT) (5.9)
Jo
By substituting Eq. (5.8) into Eq. (5.9), we obtain
V(ax) = Ex fT e-q\q - L)V(Xt)dt + Ex \e-qTV(XT) (5.10)
From Eq. (5.10), Eq. (5.6) and Eq. (5.3), we conclude that for any stopping
time r,
qT
V(x) > Ex e' V(XT)] , (5.11)
a.e. (Eq. (5.6) and Eq. (5.3) say nothing about W at dC but this is
unnecessary: dC has the Lebesgue measure 0, and therefore for a.e. x,
U«{x,dC) = 0 - see [S], Proposition 41.9). Prom Eq. (5.10) and Eq. (5.3),
for a chosen stopping time r»,
V(x) = E* \e-qT' V(XT.)] . (5.12)
By using Eq. (5.5) and Eq. (5.4), we deduce from Eq. (5.11) and Eq. (5.12)
V(x) > Ex[e-"rg(XT)},

V{x) = E*[e-*T-g(XT,)},
a. e. Set V« = V. Since V is continuous, we can omit "a.e." and conclude

that a pair (r«, V*) is the optimal stopping time and the rational price. •
5.2 Perpetual American put: the optimal exercise price and

the rational put price
5.2.1 Main Theorem

In this section, we consider the perpetual American put with the strike
price K and the payoff g(x) = K — ex. The riskless interest rate is r > 0,
and qo > 0 is the dividend rate. We set q = r + qo, and we assume that X
is an RLPE of exponential type [A_, A + ], where A_ < — 1 < 0 < A + , under
the EMM Q chosen by the market. In the statement of the next theorem,
we use <7+ > 0 introduced in Lemma 3.5.
Theorem 5.1 The optimal exercise region for the perpetual American
put is S < H* or equivalently x < h*, where
H*=eh'=Kcj>;(-i), (5.13)
and the rational put price is given by
v
*(x) = ~ « r / etc.* — d
^ (5-14)
27r J_00+iu+ £(£ +1)
for any u>+ € (0, a+).
Example 5.1 Let X be a Brownian motion, let the volatility be a2, and
let there be no dividends, that is, q = r. Let /?_ = —2r/cr2 be the negative
Perpetual American put: the optimal exercise price and the rational put price 125
root of the characteristic equation
y/32 + ( r - y ) / 3 - r = 0
of the stationary Black-Scholes equation
?lv"
2
+K(r-—)V'-rV
2'
= 0.
Then from Eq. (2.28), </>-(£) = ( - £ _ ) / ( - / ? _ + i£), and Eq. (5.13) and
Eq. (5.14) assume the form
H, = eh- = Kf3-/(f3- - 1), (5.15)
and
K
V M - /+0O+iu,+ exp[i(x-/i,)Cl(-/3-),,
respectively. The condition w+ G (0, o+) implies that the line Q£ = u+ lies
below the pole £ = — i/3- of the integrand, therefore by pushing the line
of the integration up to the infinity, we cross this only pole in the upper
half-plane. By applying the residue theorem, we derive from Eq. (5.16)
Kexp\p-(x-h.)}
V.{x) = j - ^ - ,
or
K(x) =
G3c)' (-^-r^
yr^j:) • ( 5 - 17 )
Eq. (5.15) and Eq. (5.17) are due to Merton (1973), who derived them by
using the ODE technique and smooth pasting principle.
Remark 5.2 On comparing Eq. (5.13) and Eq. (2.26), one is tempted to
write Eq. (5.13) in the form
H. = eh* = KE[eNT\, (5.18)
where {Nt} is the infimum process, and T is an exponential random variable

of mean q~1, independent of X. In fact, this interpretation is a bit mis-
leading: from the proof, which is valid for more general payoffs, it follows
that the right interpretation of Eq. (5.13) is:
the optimal /i» is the solution to the equation
(Uf4(q-L)g)(x) = Ot (5.19)
where U^ is the resolvent of the supremum process.
We will also see that the general form for Eq. (5.14) is
Vt{x) = 9C/ftl(-oo, h .]0&(? - L)g. (5.20)

In the following subsections, we
(1) show how to guess the unique candidate for the optimal stopping
time in the class Mo of hitting times -r(a) of segments (—oo, a],
(2) show how to prove, relatively simply, for any RLPE, that this is
indeed the optimal stopping time in the class Mo, and then
(3) produce more involved proof of optimality in the class M, for any
mixture of independent model RLPE,
(4) discuss the smooth pasting condition, and
(5) obtain simple approximate formulas.
After Theorem 5.1 was obtained, Mordecki (2000) proved it (in the form
Eq. (5.18)) for any Levy process by using the different technique. Notice
that in Section 5.5 we manage to prove the optimality of Eq. (5.19) in
the class Mo for wider class of payoffs, and the more practically oriented
reader may be satisfied with a straightforward proof of the optimality in
the intuitively natural class Mo.
5.2.2 Proof of optimality in the class Aio

For h and x € R, set
V(h,x):=Ex[e-«rWg(XrW)]. (5.21)
Lemma 5.2 Let there exist /i, with the following properties:
a) if h < ht, then there exists x such that
V(h,x)<g(x); (5.22)
b) for any x > h*,
V(ht,x) >g{x); (5.23)

c) if h > ht, then for any x > h,
V{K,x)>V(h,x). (5.24)
Then r(ht) is an optimal stopping time in the class M.Q.
Proof. Clearly, the rational price of the option must satisfy Eq. (5.23),
hence Eq. (5.22) excludes h < h*. Due to Eq. (5.23), h* is an admissible
choice, and Eq. (5.23)-Eq. (5.24) ensure that a choice h > h* is no better
than h*. •
To apply Lemma 5.2, we need an explicit formula for V(h, x). We derive it
by applying the Wiener-Hopf method to the problem Eq. (5.3)-Eq. (5.4),
which now assumes the form
(q-L)V(x) = 0, x>h, (5.25)

V(x) = g(x), x<h. (5.26)
Set
w(x) = ((/>-(D)-lg)(x).
Theorem 5.2 For any h G R, a solution to the problem Eq. (5.3)-
Eq. (5.4) in the class of measurable functions bounded on [h, +00) exists,
and
(i) if K_ = 1, then a continuous bounded solution is unique. It is given

by
V = <j>-{D)l(_ooM<l>-{D)-1g; (5.27)
(ii) if K- € (0,1), then a bounded solution is unique. It is given by

Eq. (5.27), and it is continuous;
(Hi) if K_ = 0, then a bounded solution is unique. It is given by
Eq. (5.27), and it is continuous if and only ifw(h) — 0;
(iv) if K_ € (0,1], then V'(h - 0) = V'(h + 0) if and only if w(h) = 0.
We postpone the proof till Subsection 5.3.6.

To guess the unique candidate /i* for the boundary of the optimal exercise
region, first, notice that for /? < a+,
4>-q{D)±le^ = 4>-{-iH)±l^x (5.28)

(for the sign " + " , it is immediate from the definition of <£~, and the formula
for the sign "-" follows), and <^"(0) = 1. Second, write Eq. (5.27) for x > h,
as
V{h,x)= J p-(x-y)w{y)dy, (5.29)
and then as
r+oo
V(h, x) = g{x) - p~{x- y)w(y)dy. (5.30)
Jh
Recall that p~ — !F~l<j)~ is the probability density supported on [0, +oo),

and suppose that w(h) > 0; then w(x) > 0 in a right neighbourhood of h,
and from Eq. (5.30), we conclude that there exists x s.t. V(h,x) < g(x).
Hence, such an h is non-optimal. Define h as the solution to the equation
w(x) = 0, that is,
h = ln(K<p-(-i)). (5.31)
If w(h) < 0, then w{x) < w(h) < 0 for each x > h, and therefore from
Eq. (5.30), we conclude that for each x > h,
V(h,x) >g{x)-w(h).
Since for x < h, V(h,x) = g(x), we see that V(h, x) is discontinuous at h;

moreover, it jumps down as x crosses h from above. One cannot expect
such strange behaviour of the rational put price, therefore we guess that the
optimal /i» must be h. Note that Eq. (5.31) is another form of Eq. (5.13),
and that with this choice of h, Eq. (5.27) assumes the form Eq. (5.14). To
see this, set v(h,x) = l(_ 00> /,)(x)u;(x), calculate, for S£ = w+ > 0,
v(h,$) = fh e-ixt(K-ex<f>;(-i))dx
J—OO
= Ke-î-it)-1 - c^7(-i)(l - »0_1

- Ke-^K-itri-il-it)-1]
(the last equality follows from Eq. (5.31)), and then apply the definition of
PDO to
V(h,x) = ct>-(D)v(h,x).
To verify that h is the optimal solution in the class .Mo, it suffices to

show that for each h > h and x > h,
V(h,x)<V(h,x). (5.32)
But for y > h, w(y) < 0, and therefore from Eq. (5.29), we conclude that
V(h,x)-V{h,x)= f p-{x-y)w(y)dy<0.
Jh
5.2.3 Proof of optimality in the class Ai

Set h = h,C = (/i,+oo), define V by Eq. (5.27), and verify conditions of
Lemma 5.1. Eq. (5.3)-Eq. (5.4) hold by construction, and Eq. (5.5) follows
from Eq. (5.30) and the inequality w(y) < 0 for y > h. To check Eq. (5.6)-
Eq. (5.8), we set v(h, •) = 1/ „, M O ^ O J and calculate
W: - {q-L)V
= qiq-'iq + iPmîDWh,-)
X
= q^{D)- v{h,.).
Under the choice Eq. (5.31), v(h,£) decays as | £ | - 2 , as ^ —> oo, therefore
in the case K+ < 1, the Fourier transform of W admits a bound via C(l +
K l ) - 1 - 6 , where e > 0. Hence, W € C 0 (R), and Eq. (5.7)-Eq. (5.8) hold.
If K+ = 1, then from Eq. (3.73)-Eq. (3.76), we deduce the representation
^(Dyâ+iOrHl -iD) + f(D),

where /(£) admits a bound via C(l + |£|)'' with some p < 1, in the half-plane
3£ > w_, where w_ < 0. The argument above shows that f(D)v(h,-) €
Co(R), and evidently, (l-iD)v(h, •) is bounded, decays at the infinity and
is smooth except for the point h, where it jumps. Hence, Eq. (5.7)-Eq. (5.8)
hold in the case K+ = 1 as well.
It remains to verify Eq. (5.6). Since
^(0 _ 1 =9- 1 (9 + # ( 0 W 0 .
we have W(x + h) = KW(x), where
W{x) _ w - . J+~*» e x p ^ y (Q^ (5.33)

Thus, we need to show that
W(x) > 0, x < 0. (5.34)
We formulate sufficient conditions for Eq. (5.34) to hold; they are satisfied
by any model RLPE, and any mixture of independent model RLPE. Of the
process, we require
(i) the function (f> in Eq. (3.46) admits the analytic continuation into
the lower half-plane with the cut (—ioo, iA_], and admits the bound
Eq. (3.51) in this half-plane, outside a neighbourhood of iA_; and
if v = 2, there exist c and u\ < 2 such that <j>{tz) — c£2 satisfy
Eq. (3.51) with v\ instead of v\
(ii) in a neighbourhood of zA_, <j> may have a weak singularity:
\m\<C\Z-i\-\-a, (5.35)
for some a < 1;
(iii) for any z € (—oo,A_), the limit
^-(z) = i[ip(iz—0)—ip(iz+0)] exists and is non-positive. (5.36)
Lemma 5.3 Let X be a mixture of independent BM, HP, NIG, NTS

Levy and KoBoL. Then (i)-(iii) hold.
Proof. Notice that if (i)-(iii) hold for two independent Levy processes,
then (i)-(iii) hold for the mixture of these processes, and therefore, it suffices
to verify (i)-(iii) for each model class separately. The verification of (i) for
NIG (and more generally, NTS Levy) and KoBoL is trivial due to the
simplicity of the analytic expressions in Eq. (3.23) and Eq. (3.6). In both
cases, the characteristic exponents are continuous at the ends of the cuts,
and there is no singularity mentioned in (ii).
Verification of (iii) for NTS Levy processes: here A_ = —a + (3, and for
z<-a + @,
$_(*) = i6[{a2-(P + i(iz-0))2y/2-(a2-((3 + i{iz + 0))2y/2}

= iS[(a2 -{13-z- iG)2)v'2 -(ai-(j3-z + iQ)2f'2)
= iS[(a2 - (J3 - z)2 + iO)"'2 - (a 2 - (/? - z)2 - iO)^2}

= i8((z - /3)2 - a2)"'2{J™'2 - e-™'2]
= -S((z - (J)2 - a2y'22 sin[7ri//2] < 0.
Verification for KoBoL: for z < A_,
*_(z) = icT(~u)[-(-X.-i(iz-0)y + (-\_-i(iz + O)Y]

= -iCr(-^)[(-A_ + z + ioy - (-X- + z- ioy]
= -icT{-u){-z + X-yie™ - e'™]
= c r ( - ^ ) ( - z + A_)I/2sin(7ri/) < 0,
since r(—V)SXO.{-KV) < 0.

The formula for the characteristic exponent of HP being more involved,
the verification of (i)-(iii) for HP is rather long, and we omit it here. •
The verification of optimality for model classes RLPE is finished with the
next lemma.
L e m m a 5.4 Let X be an RLPE satisfying (i)-(iii). Then Eq. (5.34)
hold.
Proof. By using (i)-(iii), we can transform the contour of integration in
Eq. (5.33), and reduce to the integral over the banks of the cut (—ioo, iA_].
In the process of the transformation, the contour crosses two simple poles
at £ = 0 and £ = — i (if — i is a root of q + V"(£), as it is in the case of puts
on a non-dividend paying stock, there is no second pole, but there is no
need to consider this case separately: the corresponding term below will be
automatically 0), which gives the first two terms; in the integral over the
banks of the cut, we make the change of the variable £ = iz. The result is
x
W(x) = q-(q + 1p(-i))ct>-(-i)e (5.37)
X
, , „ s-i [ - ^-{z)4>-(iz)exp[-zx]
+
^ L—fe+i)—dz-
For z < 0, cj)~ (iz) > 0, and since A_ < —1, the denominator of the integrand
is positive (the integral is over the interval z < A_(< —1), hence both
factors in the denominator are negative, and the product positive). Prom
Eq. (5.36), the integrand is negative, and for any —z > 0, the function
exp[—zx] is an increasing function of x. Hence, the integrand is a decreasing
function of x on (—oo,0). By definition of cr_, (q + ip(—i<?)) > 0 on (<7_,0),

and since 0 < 1 < — <r_, (q+ip(—i))<t>q (—i) > 0. It follows that W decreases
on (—oo,0), and hence it suffices to show that W(+0) > 0. If K+ < 1, the
integrand is absolutely integrable uniformly in x e (—oo,0], and therefore
W(+0) = W(0)
= q - (q + n-i))4>q (-») + (2TT) / * dz.
By transforming the contour of integration back, and taking into account

that (q + ip(Z))4>- ( 0 = g<#(0 - 1 > w e arrive a t
-)-ooi-ia di
/ (0(-t^)(i-i0
•oo+iff *Pq
The integrand is holomorphic in the upper half-plane 3?£ > 0 and admits
an estimate via C(l + \£\)~2+K+, for 9£ > a > 0. Hence, we can push the
line of integration up, and in the limit a —• +oo obtain zero. This finishes
the proof in the case K+ < 1.
If K+ = 1, we can represent (q + V>(£))<?V(£) = # J ( £ ) _ 1 m the form
(q + tf(0 W 0 = 9 a + ( 0 ) - x ( - i 0 + X (€),
where a+(0) > 0, and \ e n J°y s a ll the properties of (q + i/>($))<t>q~ (€) in

the case K+ < 1, which have been used above. Hence, if we use x instead
of (q + ip(^))<j>q (£) in the constructions above, we obtain a non-negative
function. To finish the proof, it remains to notice that
J —oo+io -
•
5.2.4 Failure of the smooth pasting principle for some
RLPE's and its substitute
Theorem 5.3 Let (j>Q satisfy
9
+oo+icr
107(0(1-*0_1|de<+oo, (5.38)
/. oo+ia
for some a € (0, <r+). Then the price of the perpetual American put satisfies
the smooth pasting principle.
Proof. By differentiating under the integral sign in Eq. (5.14), we obtain

VI = v, where
** J-oo+io
Hence, V* is smooth if and only if v is continuous. Under condition Eq. (5.38),

the Fourier transform of v is of the class Li(R), therefore v is continuous,
and the smooth pasting principle holds. •
Theorem 5.4 Let <j>~ admit a representation <^C(£) = c + x{£)> where

c ^ 0, and \ satisfies
+oo+i<7
I ix(o(i-*o_1|de<+oo.
—oo+icr
Then the smooth pasting principle fails.
Proof. This time we obtain that v = t>i + u 2 , where v\ is continuous, and
cff f+00+i<T exp[i(x - h . ) f l # ^ , v x-fc.
which is discontinuous. •
Notice that for an RLPE, Eq. (5.38) fails if and only if [i > 0 and v € (0,1);
this is the case of a process of bounded variation, with positive drift.
As we already saw, the natural candidate for the optimal exercise price is
determined from the equation
w(x):=<l>-(D)-1g{x)=0,
and this candidate can be singled out formally in one of the following forms:
I. If there is a unique h such that V(h; •) is continuous, this h is the

candidate; if V(h; •) is continuous for all h, the candidate is chosen by
the standard smooth pasting principle.
II. If there is h such that V£(h;h ± 0) are finite, then h is the optimal
boundary.
The second principle works for purely non-Gaussian RLPE, i.e., for RLPE
of order v < 2.
In all cases, one may say that the optimal choice of h makes V(h, •) "more
regular" at h than generically.
5.2.5 Approximate formulas for the case of model RLPE

Let the steepness parameter A+ be large, and let — i(3~ be the root of the
"characteristic equation" q+i/>(£) = 0 in the upper half-plane. By using the
crude approximation Eq. (3.82), we deduce from Eq. (5.13) and Eq. (5.14)
the approximations
Hv = eh*>=K/3-/(l3--l),
and
V M - -JL f+°°+iU+ «xp[i(s-ftap)flH3-),,
respectively, which look exactly like Eq. (5.15)-Eq. (5.16) in the Brownian
motion case. However, if one fits an RLPE by a Brownian motion, one
obtains the different /3_, and the different optimal exercise price and ra-
tional put price. Generically, the larger the A + , the smaller the differences
between the exact solution (ft#,K) and the approximation (/i a p,K P ), and
between (/i*, Vr) and the Gaussian result.
5.2.6 Proof of Theorem 5.2

By the change of the variable x H-> h + x, we reduce the proof of Theorem
5.2 to the case h = 0. Next, take any 7 G (0,<7+), and for a function / ,
introduce the notation / 7 (x) = e'yxf(x). Denote
a(D):=q + i>(D)=q-L,
substitute V(x) = e'^V^x) into Eq. (5.25), after that multiply Eq. (5.25)
and Eq. (5.26) by e 7 *, and use the equality
e<xa{D)e-^x = a(D + 17). (5.39)
We obtain
a(D + il)Vy(x) = 0, x>0; (5.40)

Vy(x) = gy(x), x < 0. (5.41)
Notice that g-y decays exponentially as x —* — oo together with all deriva-

tives: on (—oo,0],
\g\a\x)\<Cse-^, (5.42)
for each s = 0 , 1 , 2 , . . . (for the next steps of the proof, Eq. (5.42) for s < 2
suffices). Construct G 7 which coincides with g 7 on R_ and admits a bound
Eq. (5.42) on R, for s < 2, and set u 7 = Vy - G 7 , F1 = -a(D + ry)G 7 .
Then u 7 solves the problem
a(D + ry)u 7 (x) = F 7 (x), x > 0; (5.43)

w7(x) = 0, a: < 0. (5.44)
Eq. (5.42) implies that G1 e H2(R), and from Eq. (3.45)-Eq. (3.46) we
conclude that F 7 € i 7 2 _ " ( R ) , where P = i/, if i/ > 1 or /i = 0, and P =
maxfi/, 1} otherwise. Recall that we are looking for V which is measurable
and bounded on (0, +oo). Hence, we are looking for « 7 which is measurable
and admits a bound via Celx. We want to reduce the problem to the case
of an unknown function of the class L2CR+). Since cr_ < 0, we can choose
7' 6 (<T_ — 7, —7). Set
u 7 i 7 '(x) = e 7 xUj(x), Fini(x) = e 7 x F 7 (x),
substitute u^(x) — e - 7 x u 7 i 7 '(x) into Eq. (5.43) and Eq. (5.44), and after
that multiply Eq. (5.43) by e 7 x. By using Eq. (5.39), we obtain
a(£> + i(7 + 7'))u 7 , 7 '(:r) = Fltl,{x), x > 0; (5.45)

u 7 | 7 ,(x) = 0, x<0. (5.46)
Nowu 7 ) y G I<2(R+), and on the strength of Eq. (3.45)-Eq. (3.46), Theorem

15.4 gives
a(£> + z( 7 + 7 ' ) K , 7 ' e . f r ^ R ) .
Hence, the expressions on both sides of Eq. (5.45) belong to i J - < / ( R ) , and
the support of their difference, call it F-, is a subset of (—00,0]. This
o
means that F- eH _ 1 / ( R - ) , and we can write the Wiener-Hopf equation
Eq. (5.45) in the form
a(D +1(7 + 7 ' ) K , 7 ' = -F7,7' + F-. (5.47)
Multiply by q~l, and then apply <f>^{D + 2(7 + 7')). Since in the strip
3£e(A_,A + )(D(<7_,<r + )D(<7_,0))
?-1a(0 = ^(0-V7(0- 1
and by our choice, 7 + 7' G (cr_, 0), we obtain
^ - ( D + i(7 + 7 / ))- 1 u 7 ,7' = ^ + ^ - . (5-48)

where
K: = q-^+iD + iij + y^F^

= q-^+iD + t ( 7 + i))ei'*{-a{D + »7))G?7
- - 0 7 (I> + i(7 + 7 ' ) ) - V ' * G 7 ,
and
^ - g - V + P + i d + T'))^-
By construction, G 7 € i / 2 ( R ) , and
u 7 , y e L 2 (R+) = t f ° ( R + ) . F_ € f f - p ( R _ ) .
Prom Theorem 3.3, we know that for any a e (CT_,CT + ),
c(l + |ei) K± < 1^(01 < C(l + |ei) K± , ± 3 £ > a,

where C, c > 0 depend on a, therefore
< ^ - ( £ > + i ( 7 + 7 ' ) r V , 7 ' d H - ( R + ) , if- €H-K~(R-), K € tf-"-(R+).

Notice that rc_ € [0,1], and consider two cases: a) K_ € [0,0.5); b) K_ £
[0.5,1].
o
In case a), H~K~ (R) is the direct sum of the subspaces H~K~ (R±), the
projections being 6±, the closures of the-multiplication-by-lR ± -operators
defined on a dense subset Z*2(R) C H~K-(R). Hence, we deduce from
Eq. (5.48):
<£-(£>+ i ( 7 + 7'))- 1 u 7 ,7' =-e+4>-(D + i(j + 7'))-1e-y'xG^. (5.49)
Next, we multiply Eq. (5.49) by <i>q(D + i(j 4- 7')), which establishes an

o
isomorphism between H _ r e ~(R+) and L 2 (R+):
u 7 , r = -4>~(D + i ( 7 + i))9+cj>-{D + i ( 7 + 7 ' ) r X e y x G 7 . (5.50)

Then we multiply Eq. (5.50) by e~i'x and use Eq. (5.39):
Ul = -4>-{D + n)9+cl>-(D + i 7 ) _ 1 G 7 .
After that, we return to
= 0-(£> + t 7 )0_0-(L> + » 7 ) - 1 a T .
Since G 7 coincides with g~, on R_, Eq. (3.74) and Eq. (3.76) ensure that
supp <j>-{D + J 7 ) - 1 ( G 7 - 9-y) C [0, +oo). Thus,
9_ct>-(D + h)-1Gy = 9-<l>-(D + h)~lg1,
and in the formula for Vy, we may replace G 7 with g 7 . By using Eq. (5.39),
we finally arrive at
V = 4,~{D)9-<l>-{D)-1g, (5.51)
which is Eq. (5.27). This proves the first two statements in (ii)-(iii), in
case a). To study the continuity of V, consider w := 4>~{D)~xg. Clearly,
w € H2~K~ (R), and since 2 — K_ > 1/2, we can apply Theorem 15.15, and
obtain
9^-(D)-lg = w(0)(l-iD)-16 + (l-iDr19.{l-iD)^(D)-1g, (5.52)

where 5 is the Dirac delta-function. Prom Eq. (5.51)-Eq. (5.52),
V = w(0)V1+V2, (5.53)
where
^^-(DXl-iD)-1*,
and
V2 = 4>"(D)(1 - iD)-l9.{\ - iD)4>-{D)-lg.
Notice that for any e > 0, 8 £ H'^-^R), and 0_(1 ~ iD)^ {D)~l g &
H°{R). Hence, V2 6 H\R). If K_ > 0, we obtain Vx e H 1 / 2 + P ( R ) J for
any p G (0, K_), and therefore, the same is true of V. But for s > 1/2,
i P ( R ) C Co(R), and therefore, V is continuous. This finishes the proof of
(ii) for the case K_ 6 (0,0.5).
If K_ = 0, we have from Eq. (3.75) and Eq. (3.76)
4>-(D)=a-(0) + T(D), (5.54)
where T(£) admits an estimate
1T(01<C(l + |£1)-"\ 9£<u>+,
with m < 0, therefore from Eq. (5.53), we conclude that
V = a_(0)w{0)(l-iD)-18 + V3, (5.55)
where V3 e Co(R). It is straightforward to check that the Fourier transform

of l(_ OO]0 ](x)e I is (1 - i £ ) - 1 , therefore (1 - i£>)-1(S = l(_ OOi0 ]e x , and we
conclude from Eq. (5.55), that V is continuous if and only if w(0) - 0. Part
(iii) has been proved.
Now we prove part (ii) in the case K_ e [0.5,1) and part (i). We notice
that for s e (—3/2, —1/2), the decomposition of HS(H) into the sum of the
o
subspaces # " ( R ± ) is not direct, the intersection of the latter couple being
an lD-subspace C • 8, where 8 is the Dirac delta-function. It follows that in
Eq. (5.49), an additional term C8 may appear, and in Eq. (5.51), the term
C(j)~ (D)8, where C is a constant. If K_ £ (0,1), we can show with the help
of Eq. (3.75) and Eq. (3.76), that
+oo
/ -oo
is unbounded as x —> +0. Further, for K_ > 0, V in Eq. (5.51) belongs

to H3(R) for some s > 112, and hence, is continuous. We conclude that
C(j>~{D)8 + V is bounded only in the case C = 0, and so we are left with
the same Eq. (5.51). This finishes the proof of (ii)-(iii).
If K_ = 1, then the argument above shows that V in Eq. (5.51) is
continuous, and the same argument shows that 4>~{D)8(x) is discontinuous
at 0. Hence, in order to get a continuous solution, we need C — 0. This
finishes the proof of part (i).
The last part (iv) can be proved by the same argument, after differen-
tiating in Eq. (5.51). Theorem 5.2 has been proved.
Perpetual American call 139
5.3 Perpetual American call
5.3.1 Main results

The terminal payoff for the call is unbounded: g{x) — ex — K, which leads to
some complications. The standard general argument, due to Merton (1973),
which can be found in many books (see, e.g., Shiryaev (1999)), shows that
if the stock pays no dividends, then in the case of the finite time horizon,
it is optimal to exercise the American call only at the expiration date; for
the limiting case of the perpetual call, this translates into the theorem
that for any stopping rule there is a better one. Hence, we consider the
case of a dividend-paying stock, when q = r + qo > r. Notice that under
condition q > r, q + ^(-i) > 0, therefore the proof of Eq. (3.40) shows
that there exist CT_ < - 1 < 0 < CT+ such that Eq. (3.40) holds for any
[w_,w+] c {o-,<y+).
Theorem 5.5 Let the dividend rate be positive, and let X be a regular
Levy process of exponential type [A_, A+], where A_ < —1 < 0 < A + . Then
the optimal stopping time is the hitting time of the interval [h*(K), +oo),
where
eh'(K) . = H*(Kj ._ K4>+(-i), (5.56)
and the rational call price is given by
where w_ e (cr-,—1) is arbitrary.
Remark 5.3 a) As in the case of the put, our method proves the optimality
in the class .Mo, for any RLPE, and in the class M, for any mixture of
independent model RLPE.
b) After Theorem 5.5 was obtained, Mordecki (2000) proved that it holds
for any Levy process satisfying E[eXt] < +oo, in the form
H*{K) = eh'(K) = KE[eMT], (5.58)
where M is the supremum process, and T is an exponential random variable

with the mean q~l.
c) Our proof shows that the natural general form for Eq. (5.56) is
E[({q-L)g)(NT)) = 0, (5.59)
where g(x) = ex — K, and N is the infimum process, and for Eq. (5.57),
the general form is
V*(x) = 0+(I?)l [ h . i + o o ) tf+(£>)-V (5.60)
In the next Section, we will obtain Eq. (5.59)-Eq. (5.60) for more general
payoffs.
d) Prom the point of view of the analytical technique, which we use, the
principal difference between the case q > r and the case q = r is that in the
former case, the condition Eq. (3.40) holds with u>_ < — 1 < 0 < u+, and
in the latter case, due to the EMM-condition r + i/)Q(—i) = 0, it holds with
u;_ > — 1 only. x A less important issue is the unboundedness of the payoff,
which forces one to approximate the payoff by a sequence of bounded payoffs
{gn} in order that Theorem 2.12 be applicable (cf. Remark 2.1). Modulo
this approximation, the proofs and statements for the put can easily be
transformed into their analogs for the call by changing the direction on
the real axis and the reflection of the complex plane w.r.t. the origin. In
particular, the essentially new elements appear only in the proof of the
analog of Eq. (5.27).
5.3.2 Proof of optimality in the case of unbounded payoffs

Let h, x € R. Denote by r(h) the hitting time of the segment [h, +oo), and
define V(h,x) by Eq. (5.21):
V(h,x) := V(g;h,x) := E*[e-^g(XT{h))}.
Similarly to Lemma 5.2, we have
Lemma 5.5 Let there exist h* with the following properties:
a) if h> h*, then there exists x such that
V(h,x)<g(x); (5.61)
1
To be more specific, in the former case, the real part of the operator q — L = q + t/>Q (D)
is positive-definite on the L2-space of function which grow as ex as x —• +00, whereas
in the latter case, e x is an eigenfunction of q — L.
Perpetual American call 141
b) for any x < h*,
V(h*,x)>g(x); (5.62)
c) if h < h*, then for any x < h,
V{h*,x)>V(h,x). (5.63)
Then r(h*) is an optimal stopping time in the class Mo-

To derive an explicit formula for V(h, x) — V(g; h, x), we use Theorem 2.12
but unlike in the case of puts, we cannot do it directly since this time g
in unbounded. Construct a pointwise increasing sequence {Xn} C C°°(R)
converging pointwise to 1, s.t. for each n, Xn{x) is non-negative, equals
1 on (—oo, h] and vanishes for sufficiently large x. Then set gn = gxn-
Clearly, we may consider only h such that g(h) > 0, and then on [h, +oo)
gn(x) | 5(x) pointwise. Hence, by the dominant convergence theorem, for
any x,
V(gn;h,x) |%/i,i), n -> oo. (5.64)
Since gn is bounded, we can apply Theorem 2.12 and conclude that V(gn; ft, •)
is a bounded solution to the following problem
(q-L)V(x) = 0 , x<h, (5.65)

V(x) = gn(x), x>h. (5.66)
Further, by replacing all the signs in the statement and the proof of Theo-
rem 5.2, we find that the unique bounded measurable solution (and in the
case K + = 1, the unique continuous bounded solution) is given by
V(gn; h, •) = ^ ( l O l ^ + ^ + t D ) - V - (5.67)
By using Eq. (3.72) and Eq. (3.76), it is easy to obtain
'i-(h,+oo)(t>q{D)~1(gn ~ l(/i,+oo)5n) = 0.
Hence, we can write Eq. (5.67) as
V(gn; h, •) - <A+(D)l ( / ,, + o o ) <(D)- 1 5 n (/i; •), (5.68)
whereg n (h;-) := l(h,+oo)9n- Take7 G (er_,-l),set0„ i 7 (/i;:r) = e~<xgn{h;x),

multiply Eq. (5.68) by e 7X , and use the equality
e"<l>UD)kle-'1x = <t>UD + i-y)*1. (5.69)

The result is
V(gn; h, •) = e-*x4>t{D + t 7 ) l ( h i + o o ) # ( 0 + i 7 ) _ W 1
; •)• ( 5 -™)
Due to the choice of 7, a sequence {gn^(h; •)} consists of bounded func-

tions exponentially decaying at the infinity, and for each x > h, gn^{h\ x) f
gy(h; x) (which is also bounded and exponentially decays at infinity). Hence,
gn,-y{h; •) —> g-f(h; •) in £2(^)1 and therefore, in the sense of generalized
functions,
V(gn;h,-)->W(g;h,-), (5.71)
where
W(g; h, •) = e~^(D + t 7 ) l ( h , + o o ) # ( 0 + ij)-%{h; •).
By using Eq. (3.72) and Eq. (3.76), we can show that
W(g; h, •) = e-i*4$(D + ij)l(ht+oo)<f>+(D + t 7 ) " 1 ^ ,
and by using Eq. (5.69), we obtain
W(g; h, •) = <t>+{D)\(h}+oo)4>${D)-lg.
Compare Eq. (5.64) with Eq. (5.71). Since the limit in the sense of gen-
eralized functions is unique, we have W(g; h, •) = V(g;h,-) in the sense of
the generalized functions. By explicitly computing W(g; h, •):
+00+17
^•(-*)_1e(1-^ Ke-*h
e f a C #(0 d£
/ -oo+it 1-tf -^
where 7 € (cr_,—1), it is easy to check that W(g;h, •) is continuous on

(h, +00) but if two continuous functions define the same generalized func-
tion, they coincide.
To sum up, we have shown that
V(g; h, •) = <t>+(D)l[K+oo)<t>+{D)-lg. (5.72)
We set w(x) = <j>+{D)~lg{x), and prove that the optimal h* in Lemma 5.5
is the solution to the equation w(x) = 0, that is, <f>%(—i)~1eh' — K = 0.
This and the rest of the proof can be done in essentially the same manner
as in the case of puts.
Put-like and call-like options: the case of more general payoffs 143
5.4 Put-like and call-like options: the case of more general

payoffs
5.4.1 Put-like options

In the set up of Subsection 5.3, we consider payoff functions g(x) more
general than K — ex. Let <r+ > 0 be from Lemma 3.5, let h £ R, and
consider the problem Eq. (5.25)-Eq. (5.26).
T h e o r e m 5.6 Let g and its derivatives g^ = D"g, s = 0,1,2, be mea-

surable, and satisfy the following estimates
|5 ( s ) (x)| < Ce-"'+x, x<h, (5.73)

(s) x
|ff (x)| < Ce-"'- , x>h, (5.74)
where w'_ < LJ'+ < a+. Then all the statements of Theorem 5.2 hold.
Proof. Eq. (5.73)-Eq. (5.74) are the very conditions we used in the proof
of Theorem 5.2, with only one exception: if tv'+ > 0, then in the proof, we
must take 7 S (LJ'+,O-+). •
Remark 5.4 Notice that Eq. (5.74) is not a real restriction: if g fails to
satisfy this condition, we can replace g with any Ig, which coincides with g
on (—00, h] and satisfies Eq. (5.74). In particular, for any such Ig,
V = rq{D)h-ooM^{D)-Hg, (5.75)
which is understood as
+O0 + tUJ+
eix%{0mdi, (5-76)
/
-OO-HCJ-L
1
where v = l(_ oo/l ](/) q (D) lg, and w + S (u'+,a+) is arbitrary.
1
Setw(x)=(4>g(D)- g){x).
T h e o r e m 5.7 Let Eq. (5.73)-Eq. (5.74) hold with LJ'_ < u'+ < a+ and
some h, and let there exist h\ < Ji2 such that the following conditions are
satisfied:
w(x) > 0, VxKhi, (5.77)

w(x) = 0, V hi < x < h2\ (5.78)
w{x) < 0, V x > h2. (5.79)
Then for any h G [hi, hi], T(/I) is an optimal stopping time in the class
Mo, o,nd the rational option price is given by
e fa V7(0«(Ode, (5-80)
where o>+ £ (OJ'+, A+), and v = 1,^ j^w.

Proof. We repeat the proof of optimality in the class Me for the put, by
making use of Lemma 5.2 and Theorem 5.6 instead of Theorem 5.2. •
Notice that Eq. (5.78) can be written in the form Eq. (5.19).
Consider payoffs of the form
m
9i ) =Ylci eMljx),
x
(5-81)
where -o+ < 71 < 72 < • • •• Then Eq. (5.74) and Eq. (5.73) hold, and in
concrete cases, conditions Eq. (5.77)-Eq. (5.79) are easy to verify since
w{x) = 5 > ^ - ( - i 7 j ) - V ' * . (5.82)

3=1
In particular, if they are satisfied then hi = /12; call it ht. Notice that
the necessary condition for Eq. (5.77) and Eq. (5.79) is ci > 0 and c; < 0,
respectively.
Example 5.2 Suppose, the option owner has the right to sell a share of
the stock for K + ay/S, where S is the spot price. Then
g(x) = K + aex'2 - ex,
w(x) = tf + a < £ - H / 2 ) - V / 2 - 4 > - H ) - V ,
and the optimal exercise boundary is h* = y/Y, where Y is the unique
positive root of the equation
K + a 0 - ( - i / 2 ) " 1 r - ^(-iy'Y2 = 0.
When /i* is found, one can easily calculate v = l(_0o,/i»]u; a n d *K£)> f° r a n v
£ in the strip 3£ e (0,o+):
0(0 = / e-^[tf + a<£-(-V2)-V/2-</>-(-0-V]Cb

J—00
Ke-^*Z J0.5-iZ)h, M-iOh.
and after substituting into Eq. (5.80), obtain the rational option price:
-t-0O-|-JU>+
e i(z-h.)^-(0
/ -OO-f
-OO-f 1W+
ILJ+
K ae05h- eh'
X I-TT
-i£ + < ^ H / 2 ) ( 0 . 5 - i £ ) ^-(-i)(l-^)J
where u>+ G (0,0+) is arbitrary. By using the approximation Eq. (3.82) for
<^>~(£), we obtain the approximation to the rational option price:
f+oo+iu+ _a
e<(x-M€.
•oo+iw+
„„0.5/u ph. 1
+
-t£ ^(-i/2)(0.5-iO ^(-0(1 ~^)J
The integrand has the only pole £ = — i/3_ hanging above the line of inte-
gration, and this pole is simple. By pushing the line of integration up and
using the residue theorem, we obtain the following approximation:
K P (x) = De?-(*-h'\ x>K,
where
-/?_ae°- 5/l - ~(3-eh'
D =K +
&-(-x/2)(0.5-/3_) ^-(-t)(l-/J_)
The next theorem shows that for model RLPE, the stopping time obtained
in Example 5.2 is optimal not only in the class Mo but in the class M as
well.
T h e o r e m 5.8 Assume the following conditions:
a) X is a mixture of independent model RLPE;

b) g can be represented in the form
i i
x c ex
9( ) =Y^ t vllk } ~^2ck exphk*]'
x
( 5 - 83 )
fe=i fc=i
where cjj: are positive, 7 ^ > — cr+, k = l,...,l, are not necessarily
different, as well as 7^" > — o~+, k = 1 , . . . , /, and satisfy
7^<7fc", Vfc; (5.84)
c) for all k, the solution to the equation
1 h 1 1 1
ct^{-iltr ^ =^{-ilkT ^ (5-85)
is the same; call it h„.
Then r(/i*) is the optimal stopping time in the class M, and the rational
option price is given by Eq. (5.80) with any w+ € (maxj(—7,), cr+), h — /i»,
and
m^t^W^^f + 7w7- *v (5 86)

-
^t (7/ - »0(7> - *0
Proof. Recall that in the case of the put option, we managed to check
all the conditions of optimality but Eq. (5.6) for any RLPE, and the same
argument is valid here, even for much more general payoffs than Eq. (5.81).
Thus, the additional conditions b) and c) are needed to verify Eq. (5.6).
Eq. (5.83) allows us to represent g as a sum of payoffs of the form
gj(x) = c+e>t*-cje>7*- (5.87)
and Eq. (5.85) implies that h, = h*(gj) is the same for all j . But when /i*
is the same for all payoffs gj, Eq. (5.6) is evidently "additive" w.r.t.
g in the sense that if it holds for all V*(gj,x) defined by gj, then it holds
for V*(g,x). Thus, it remains to consider the payoff Eq. (5.87), and we
can repeat the proof of Eq. (5.6) for the put with the straightforward
changes. D
5.4.2 Call-like options

In the set up of Subsection 5.4, we consider payoff functions g{x) more
general than e x - K. The statements and the proofs are obtained from the
ones by reflection of the real axis w.r.t. the origin, with the corresponding
changes of signs and signs of inequalities. We start with an auxiliary result
about the Wiener-Hopf equation on (—00, h).
Let cr_ < 0 be from Lemma 3.5, and consider the problem
{q - L)V(x) = 0, x<h, (5.88)

V(x) = g(x), x>h. (5.89)
Theorem 5.9 Let g and its derivatives g(3' = D3g, s = 0,1,2, be measur-
able, and satisfy the estimates Eq. (5.73)-Eq. (5.74) with o~- < w'_ < UJ'+ .
Then for any h € R, o solution to the problem Eq. (5.88)-Eq. (5.89) in the
class of functions bounded on (—oo,/i] exists, and
(i) if K+ = 1, then a continuous bounded solution is unique. It is given

by
V = tfiD^+^+iDr'g; (5.90)
(ii) if K+ € (0,1), then a bounded solution is unique. It is given by

Eq. (5.90), and it is continuous;
(Hi) if K+ = 0, then a bounded solution is unique. It is given by
Eq. (5.90), and it is continuous if and only if (<j>+(D)~lg)(h) = 0;
(iv) if K+ £ (0,1], then V'(h - 0) = V'(h + 0) if and only if
(<f>+(D)-ig)(h) = 0.
Remark 5.5 Notice that Eq. (5.73) is not a real restriction: if g fails to
satisfy this condition, we can replace g with any Ig which coincides with g
on [h, +oo) and satisfies Eq. (5.73). In particular, for any such Ig,
V(x) = <t>+(D)l[h,+oo)<f>+(D)-Hg, (5.91)
which is understood as
+oo+iw_
e^+TOSK, (5.92)
/
•oo+iui_
where v — l[h^+00)<j)+ (D)~xlg, and w_ € (cr-,uj'_) is arbitrary.
Set
w(x) = (<^+(Z))- 1 g )(x).
Theorem 5.10 Let Eq. (5.7S)-Eq. (5.74) hold with a~ <u'_ < u'+ and
some h, and let there exist hi < hi such that the following conditions are
satisfied:
w(x) < 0, Vz</ii; (5.93)

w(x) = 0, V hi < x < h2; (5.94)

w(x) > 0, Vi>ii2. (5.95)
Then for any h € [^1,^2], T(JI) is an optimal stopping time in the class
Aio, and the rational option price is given by
e 1 ( l - f c ) c #(0*(Ode. (5-96)
where u>_ £ (o--,u)'_), and v = 1,^ +00\U>-
Notice that Eq. (5.94) can be written in the form Eq. (5.59).
Consider payoffs of the form Eq. (5.81), where 71 < 72 < • • • < —<r_. Then
Eq. (5.74) and Eq. (5.73) hold, and in concrete cases, conditions Eq. (5.93)-
Eq. (5.95) are easy to verify since
l
w{x)=YJCj<t>t(-ilirl^jX- (5-97)
In particular, if they are satisfied then h\ — J12; call it h*. Notice that
the necessary condition for Eq. (5.93) and Eq. (5.95) is ci < 0 and c; > 0,
respectively.
Theorem 5.11 Assume the following conditions:
a) X is a mixture of independent model RLPE;

b) g can be represented in the form
1 1
9( ) = ~ 5Z fc P[Tfc" ] + J2 cfc exPb% *L
x c ex x
(5-98)
fc=i fe=i
where cj~ are positive, 7^" < — <J_, k = l,...,l, are not necessarily
different, as well as 7^ < — c _ , k — 1 , . . . , /, and satisfy
7+<7fc-, Vfc; (5.99)
c) for all k, the solution to the equation
cUti-ht)'1^^ = c* #(-i7fc-^1e'Y;^', (5-100)

is the same; call it h*.
Commentary 149
Then r(/i*) is the optimal stopping time in the class M, and the rational
option price is given by Eq. (5.96) with any <J+ G (<r_,minj(—jj)), h = h*,
and
7
HO = -i24<t>i(-nf)-^t-w + ' "7£ (5.101)
5.5 Commentary
The results of this chapter are obtained in Boyarchenko and Levendorskii

(2002a) but several of key points of the proofs in op. cit. are essentially
simplified here, and the exposition is somewhat different. In the Brownian
motion case, the price of the perpetual call (under the historic measure)
was calculated by McKean (1965); the pricing problem for the perpetual
put was solved by Merton (1973). For diffusions with jumps, the prices of
the perpetual puts and calls were computed in Mordecki (1998, 1999). For
RLPE, the price of perpetual American put was calculated in Boyarchenko
and Levendorskii (1998,1999b, 2000), in terms of the factors in the Wiener-
Hopf factorization formula, and in the earlier version of Boyarchenko and
Levendorskii (2002a), which was circulated in (2000), the calls were con-
sidered, too. In op. cit., it was surmised that the results in terms of the
factors in the Wiener-Hopf factorization formula should hold for any Levy
process. Later, Mordecki (2000) announced the pricing formulas for puts
and calls in terms of the infimum and supremum processes, respectively,
for any Levy process. Mordecki used the reduction to the similar result in
Darling et al.(1972) for puts, in the discrete time (for the formulation, see
Chapter 13), and a variant of this result for calls which he derived.
For applications of the Wiener-Hopf method to the Queuing theory and
Insurance, see Borovkov (1972) and Prabhu (1980).
Chapter 6
American options: finite time horizon
In the finite horizon case, there are no explicit analytical pricing formulas
even in the Brownian motion case. We formulate the free boundary value
problem for the price of an American option, and consider several analogues
of approximate methods used in the Gaussian option pricing theory. We
also derive an approximate formula for the early exercise boundary near the
maturity and show that its behaviour in the RLPE-case drastically differs
from the behaviour in the Gaussian case.
6.1 General discussion
6.1.1 The free boundary problem

Consider an American option with the expiry date T and the payoff function
g(Xt,t); since the option owner has the right to exercise the option any
instant t before or on the expiry date, g may depend not only on the log-
spot price Xt but on t as well. We assume that the riskless rate r > 0 is
constant, and X is an RLPE under the risk-neutral measure Q chosen by
the market. The option owner chooses her optimal exercise policy in order
to maximize the expected discounted payoff
K(x,t) = supE%-rrg(XT,r) | X0 = x],

T
where supremum is taken over the set of stopping times satisfying T(U/) <
T.Vwefl.
Let L = L® be the infinitesimal generator of X under Q. Assume that
the optimal stopping time is of the form r'B AT, where r'B is the hitting time
151
152 American options: finite time horizon
of a closed set B C R x (—oo, T] by a two-dimensional process Xt = (Xt,t).

Set C = R x [0, T) \ B (this is the region where the option remains alive),
and consider the following boundary value problem
(dt + L-r)V(x,t) = 0, (x,t)€C; (6.1)

V(x,t) = g(x,t), (x,t) eBoi t = T; (6.2)
V(x,t) > g(x,t), t<T,xeR; (6.3)
(dt + L-r)V(x,t) < 0, t<T, (x,t)$C. (6.4)
Denote by L = dt + L the infinitesimal generator of the two-dimensional

process X, and by Ur its resolvent.
Theorem 6.1 Let the following conditions hold:
a) the r-potential measure of dB is zero: UrldB = 0, a. e.;

b) V* is a continuous solution to the problem Eq. (6.1)-Eq. (6.4);
c) W* :— (r — L)V* is universally measurable, and
d) UrW*=V*.
Then B is the optimal early exercise region, and V* is the rational option
price.
Proof. Repeat the proof of Lemma 5.1 with the process X instead of X,
taking into account that a sample path of X cannot cross the line t = T by
a jump. •
For the detailed study of the free boundary problem for the American
put in the Gaussian case, see monographs Musiela and Rutkowski (1997),
Karatzas and Shreve (1998), Shiryaev (1999) and the bibliography in op.
cit.
6.1.2 The role of dividends and non-zero interest rate

rt
In the case of a non-dividend paying stock, the process {e~ St} is a mar-
tingale under Q, which translates into the EMM condition
r + y>QH) = 0 (6.5)
for the characteristic exponent ip** of the process X, under Q. If the stock
pays dividends at the constant rate A > 0, then {e~(r~x)tSt} is a martingale
Approximations of the American put price 153
under Q, and the EMM condition assumes the form
r - A + V>Q(-i)=0. (6.6)
Eq. (6.6) implies that
r + ^Q(_j)>0; (6.7)
and from the technical point of view, this is the difference between Eq. (6.5)
and Eq. (6.7) which leads to Merton's (1973) result on the no-early-exercise
of the American call on a non-dividend-paying stock. Merton's no-arbitrage
argument is universal; in particular, it is valid not only in the Gaussian case,
but also for any Levy process.
Similarly, if the interest rate is zero: r ~ 0, then it is non-optimal to
exercise the American put before maturity.
6.2 Approximations of the American put price
We consider RLPE-analogues of three methods of approximate pricing of

the American put, popular in Gaussian Mathematical Finance; the Amer-
ican call on a dividend-paying stock and other types of American options
can be treated similarly. The problem of pricing of American contingent
claims being of great practical interest, there exist a host of other meth-
ods (see, e.g., Hull (2000), Kwok (1998), Musiela and Rutkowski (1997),
Wilmott et al. (1993), and the bibliography there).
6.2.1 Geske-Johnson approximation

The approximation proposed by Geske and Johnson (1984) uses the dis-
cretization of the time parameter and the backward induction, as in any
other standard discrete time approach. Assume that the option can be ex-
ercised only on the deterministic dates ti < t 2 < • • • < tn = T. An option
which can be exercised before expiry but only on dates ti < ti < • • • < tn
fixed in advance is called a Bermudan option; thus, the Geske-Johnson
approximation is an approximation of an American option by the corre-
sponding Bermudan option. In the Gaussian case, the explicit analytical
formula for the joint distribution of the stock prices (Stl,..., Stn) is known,
and so the application of the following procedure is straightforward.
Let Vj(x) = V(ex,tj) be the value of the option at time t = tj. For
j = n, when tn = T, vn is the terminal payoff: vn(x) = max{.K" — e x , 0 } .
At time tj, j = n — 1,n — 2 , . . . , 1, the option owner chooses the optimal
exercise boundary hj as the solution to the equation
K-e° = e-rr>E°\vi+l{Xti)\, (6.8)
where Tj = tj+i — tj. In Eq. (6.8), the LHS is the payoff at the current
level of the log-price, x, and the RHS is the discounted expected value
of keeping the option alive. If an analytical formula for the probability
density pt of the process Xt under Q is known, as is the case with NIG, or
tabulated, which can be done for any model RLPE, then Eq. (6.8) can be
solved numerically, and after the optimal exercise boundary hj is found, Vj
is calculated as
v(Ax\ = l e'rTjx / - » PTi O/K+i (x + y)dy' x

> hi'
' \ K-e , x<hj.
In the Brownian motion case, it is known that Eq. (6.8) has a unique
solution, and the sequence {VJ} converges to the option exact price when
A :— maxTj —* 0; one should expect that the same holds in the RLPE-case
as well. To estimate the limit, one may use Richardson's approximation
scheme (see, e.g., Musiela and Rutkowski (1997)).
6.2.2 Analytic method of lines

Carr and Faguet (1994) suggested to use the time discretization in the
Black-Scholes equation. They studied the Brownian motion case; in our
non-Gaussian situation, we discretize the derivative dt in the generalized
Black-Scholes equation Eq. (6.1). In the Gaussian case, it is possible to use
the smooth pasting principle to replace the optimality conditions Eq. (6.3)-
Eq. (6.4). Since we have shown in the case of perpetual American put that
the smooth pasting principle may fail in the non-Gaussian case, we have to
find the optimal discretized boundary by other means similar to the ones
used in the perpetual option case.
Divide [0, T] into n subperiods by points tj = jA, j = 0 , 1 , . . . , n, where
A = T/n, and denote by Vj(x) the approximation to V(x, tj). Then vn(x) =
max{.ftT—ex, 0}, and by discretizing the derivative dt in Eq. (6.1), we obtain,
for j = n — 1, n — 2 , . . . , 0,
A _1 («j- + i - Uj) - (r + V>(I>x)H(a:) = 0 , a: > fy,
or
(1 + A(r + ip(Dx)))vj(x) = vj+1(x), x>hj. (6.9)
Eq. (6.2) assumes the form
Vj{x)=g(x), x<hj, (6.10)
where g(x) = K — ex. Set q = A - 1 4- r,
a A (0 = 1 + A(r + V(0) - A(g + V(£)),

factorize a A:
a A (0 = (l + rA)^+(0-V7(0 _1 .
where <j>^ is the Fourier transform of a probability distribution P^ (dx) sup-
ported on R T (see Chapter 5), and solve the problem Eq. (6.9)-Eq. (6.10)
by using the Wiener-Hopf factorization method.
Denote by u,i the solution to the problem Eq. (6.9)-Eq. (6.10) with
Vj+\ = 0. Theorem 5.2 gives
v
ji =<t>q{D)l(-00,hi}<l>q{D)~1g,
and reasoning the same way as in the proof of Theorem 5.2, we can show
that the solution to the problem Eq. (6.9)-Eq. (6.10) with g = 0, call it Vj2,
is given by
vj2 = (1 + rA)-1<j>-(D)l[hjt+oo)<l>+(D)vj+1.
Clearly, Vj = Vj\ + Vj2, hence
Vj = ^•(D)l(_OOifcil07(D)-1«;
+{l + T^)-1d>-{D)llhii+eo)^+(D)vj+1
+(l+rA)-1<t>-(D)l[hjt+oo)cl>+(D)vj+1
= (K - ex) + (1 +rA)-1<t>q-(D)l[h.<+oo)wj,
where
Wj = 4>+(D)vj+1-(l + rA)<f,-(D)-\K-e')
= <l>+(D)vj+1 _ ( l + r A ) ( ^ - ^ ( - i ) - V )
= 4>+(D)[vi+1 + e'-(l + rA)K].
To derive the last equality, we have used the EMM-condition r+ip(—i) = 0,

and the Wiener-Hopf factorization at £ = —i:
^-(- i )-v?(- i )-= 1+A 1 ( ; + r r i)) =(i +r A)-'

Hence,
Vj = (K - ex) + (1 + rA)-1cj>;(D)l[hjt+oo)4>Z(D)zj, (6.11)
where
Zj(x) = vj+1(x) + ex - (1 + r-A)iC.
Lemma 6.1 For j = n — 1 , . . . , 0, tf»e function Wj — <j>+(D)zj is increas-

ing, and it changes the sign.
Proof. We have vn = (K — ex)+, whence z n _i(ar) = (e x — K)+ — rAK

is increasing, and hence so is
wn-i(x) = (^+(D)z n _i)(x)

+oo
zn-i{x + y)P+(dy).
/ •oo
Now, by a similar argument, from Eq. (6.11) we deduce that
zn_2 = ex -K + vn~i - rKA

l
= (1 + rA)- cj>;(D)l[hn_u+oo)^(D)zn_1 - rKA
is increasing, and so is wn-2- By iterating, we deduce that Wj is increasing

for all j .
To show that Wj changes sign, notice that
0 < Vj(x) <K, Vx (6.12)

(Eq. (6.12) will be proved at the end of the proof of Lemma). By using
Eq. (6.12), we derive
limsuptUj(x) = hms\ip{(j>^(D)vj+i + <f>+(-i)ex - (1 + rA)K]
< 4>+(D)K-K<K-K = 0,
and
lim Wj(x)= lim [<t>+(D)vj+1+ct>+(-i)ex-(l+r&)K} = +M;

a:—>+oo x—»+oo ^ ^
therefore, Wj changes sign.
It remains to prove Eq. (6.12). By Theorem 5.2, for j < n, Vj is the
unique bounded (or unique bounded continuous) solution to the problem
Eq. (6.9)-Eq. (6.10), hence from Theorem 2.12 we conclude that Vj can be
represented in the form of the stochastic integral
Vj{x) = Ex
f
Jo
e-Â^Vj+^X^dt + EX ,-QTi
(K ex^)
where r, is the hitting time of (—oo, hj}. Since vn(x) — TOSX.{K — e x ,0} G
[0, K), we find that vn_i is non-negative, and it admits an upper bound
via
Ex{K(A-1q-\l - e~qTi) + e ^ ) ] < K,
that is, vn-i satisfies Eq. (6.12). By iterating, we derive Eq. (6.12) for
j=n-2,...,0. •
Prom Lemma 6.1, Wj has the only zero, call it hj^.
Theorem 6.2 The optimal exercise boundary at time tj is hj^, j =

n — l , n — 2 , . . . ,0.
Proof. To stress the dependence on the exercise boundary hj, we write

Vj(hj\x) instead of Vj(x). Clearly, Wj is negative to the left of hj^, and
positive to the right of it. Hence, with a choice hj < /iJ:», l[hj]+00)U>j is
negative in a right neighbourhood of hj, and since cf>~ (D) is the convolution
operator with the kernel supported on the positive axis, we conclude that
VJ (hj ,x) = K -ex + <f>- {D)l[hj ,+00)Wj (x)

x
< K-e
at some points, which is impossible if the choice of hj is optimal. If hj >

hjit, then
l[hj,+oc)Wj(x) < l[fc J > > + o o)U>j0c)
on [hjit,hj), and both sides coincide on the complement to this interval.

We conclude that
Vj(hj,x) < Vj(hj^,x), V x
(and for some x, the inequality is strict), which finishes the proof of Theo-
rem. D
To sum up: the algorithm for the discrete time approximation to the opti-
mal exercise boundary and the rational put price is as follows:
1. Set hn<* = h\K, vn(x) = max{K — e T ,0};
2. For j = n — 1, n — 2 , . . . , define
Wj = <j>+(D)[vj+1 + ex - (1 + rA)K],
next find hj,* as the unique solution to the equation
Wj (x) = 0,
and set
Vj(x) =K-ex + (l+ rA)~V7(I?)l[fcil.,+«x,)«'j(a:)-
6.2.3 Carr's randomization

Carr (1998) suggested the following procedure in the Gaussian case. The
first step of the procedure remains the same in the non-Gaussian case,
and we end up with the same algorithm as in the analytical method of
lines. First, we divide the period [0, T] into n sub-periods of the length
A = T/n, next, set tj = jA, and find approximations hj and Vj to the
optimal exercise boundary by using the backwards induction. We have
vn(x) — max{p(x), 0}, where g(x) = K — ex; for j = n — 1, n — 2 , . . . , 0, the
approximation u, is defined as
Vj(x) = ma,xV(vj+i,hj-,x),
hj
where V(v, h; x) is the value of a barrier option with the terminal payoff v,
the down-and-out barrier h, and the rebate g (we may assume that only
American put near expiry 159
barriers h s.t. g(x) > 0 for all x < h are considered); the maturity date, r, is
a random variable which is exponentially distributed with scale parameter
A = 1/A. Hence,
r+oo
Vj(x) = maxA / e~xtU(vj+i,hj-,x,t)dt,
h
i Jo
where U(VJ+I, hf,x, t) is the value (at time 0) of a barrier option with the
maturity date t, the terminal payoff t>j+i, down-and-out barrier hj, and
rebate g. Prom Theorem 2.13, we conclude that U(vj+\,hj;x,t) is equal
to W(x,0) (the arguments Vj+\,hj,t are suppressed), where W(x,s) is a
bounded solution to to the following boundary value problem
dsW(x,s)-(r + ip(Dx))W(x,s) = 0, x>hj,s<t;

W(x,t) — Vj+i(x), x > hj;
x
W(x,s) = K-e , x<hj, s<t.
By making the change of variable s = t — r , and applying the Laplace-

Carson transform 1 w.r.t. r, we obtain the family of problems
Xvj(x) + (r + ip(Dx))vj(x) — \vj+i(x), x>hj, (6.13)

x
vj(x) = K-e , x<hj. (6.14)
Divide Eq. (6.13) by A, and take into account that A - 1 = A; we see that the
problem Eq. (6.13)-Eq. (6.14) turns into the problem Eq. (6.9)-Eq. (6.10),
and as in the preceding subsection, the free boundary hj must to be chosen
so as to ensure that the solution is optimal. The result is the procedure
described at the end of Subsection 6.2.2
6.3 American put near expiry
6.3.1 Behaviour of the early exercise boundary near the

maturity date
To simplify the formulas, we normalize the strike price if to 1. In the
Brownian motion case, the following asymptotics of the optimal exercise
' T h e Laplace-Carson transform differs from the standard Laplace transform by an ad-
ditional A-factor.
boundary x — h(T — t) is known
h(T-1) ~ -<r^{t-T)\n{T-t), (6.15)
as t —> T—0 (see, e.g., Barles et al. (1995) and Lamberton (1995)). We have
been unable to study the behaviour of the optimal early exercise boundary
near the expiry in the proper continuous time setting; instead, we study the
asymptotics of the boundary point /i„_i = h(A) at the last moment before
expiry, as A —» 0, in the discretization of the continuous time model. We
believe that the asymptotics of h(A) can be used as a good proxy for the
asymptotics of the boundary near expiry.
We study NTS Levy processes, NIG in particular; similarly, KoBoL and
Hyperbolic processes can be treated. The main result is that in many cases,
in particular, for processes of order v G [1,2), the limit of h(A) is negative,
and not zero, as in the Gaussian case. Notice the opposite effect for the
boundary far from expiry: the numerical results for the optimal exercise
price h» of the perpetual American put show that typically, the Gaussian
model gives smaller /i*, and we may conclude that in the non-Gaussian
case, the optimal exercise boundary is more flat than in the Gaussian case.
This observation shows that there is some compensation for the apparent
difficulties with the design of appropriate numerical procedures for pricing
of the American put in the non-Gaussian case.
For processes of order between 0 and 1, the behaviour of h(A) of various
types is possible, depending on all the parameters of the process; we will
not list the corresponding formulas here.
6.3.2 A general equation for h(A) in the case of NTS pro-

cesses
Consider the equation for fo„_i = h(A):
4>+(D)[(l-e°)++e*-l-rA}=0.
Rewrite it as
< # ( Z ? ) [ ( e * - l ) + - r A ] = 0,
and then, by using the equality
0 + ( 0 ) . l = # ( O ) - l = l,
and the definition of PDO, as

+00+IUJ- ix
r+oo+iW_ e « <£+(£)
\^q^d$ = 7-A, (6.16)
/ -t|(i-tO'
-oo+iw_
where w_ e (A_, — 1) is arbitrary. Since
^(0 =
(9 + ^(O)0,-(O'
we continue the transformation of the equation for a: = h(A):
rA. (6.17)
oo+iW_ (9 + V » ( 0 ) ^ ( 0 ( - ^ ) ( l - i O
In the NTS Levy case, A± = ±a + 3, and the formula for the characteristic
exponent of an NTS Levy process is
tf(0 = - ^ + S[(a2 -(0 + it)2)"'2 - (a2 - 32Y'2\.

By transforming the line of the integration in Eq. (6.17) into the integral
over the cut in the lower half-plane, and multiplying both sides by q, we
obtain
xz 2
•a+0 e- q [ip(iz + 0)-tlj(iz-0)}idz
(27T)" 1 / = rqA.
J —I 4>q{iz)z(l + z){q + ip(iz + 0))(q + ip(iz - 0))
(6.18)
Notice that
rqA = r ( A _ 1 + r)A -> r as A -> 0, (6.19)
and
i>(iz±0) = nz - 8(a2 - 32)v/2 + 6(a2 - (3 - z ±iO)2)"/2
= Hz- S(a2 - 32Y'2 + 6((3 - z)2 - a2y/2e^v'2.
We conclude that
2
[^(iz + 0) - V(w - 0)]* = -a2),//2(e-i'r^2-ei7rl//2)z
<5((/3-z)
= 2Jsin^((^-z)2-a2r/2, (6.20)
and
(q + iP(iz + 0))(q + i>(iz-0)) = (9 + / i 2 - t f ( a a - / 3 2 ) " / 2 ) 2 (6.21)
2 2 2 v
+<5 ((/3 - z ) - a )
+25cos^((/3-z)2-a2y/2
2 2
.(q + ^ z - S i a -^ ).
Lemma 6.2 a) On the half-axis z < 0, <f>^(iz) admits a bound
ci(l - zq-V")"- < 4>~{iz) < ca(l - zq-l'v)K-, (6.22)
where c\,c-x > 0 are independent of q > 1 and z < 0.

b)IfO< -zq~1/v and -zq~1/v -» 0, as g -+ +oo, tfien
4>-(iz) ~ 1. (6.23)
Proof, a) Set A = — ig, and let a_(A,f) be the factor defined in Chapter
7. Then <f>-{£) = a_(A,0)/a_(A,O. and Eq. (6.22) follows from Eq. (7.37).
b) In Eq. (3.60), make the change of variable n — iuj+ + qxlvr( and use
Eq. (3.45)-Eq. (3.47) to show that the integral under the exponential sign
vanishes as q —> +oo. •
Now we are ready to study Eq. (6.18). It is convenient to consider separately

the following cases:
a) i / € (1,2);
b) u = 1;
c) v G (0,1) and \x = 0;
d) v G (0,1) and ^ < 0;
e) v G (0,1) and /* > 0.
6.3.3 The case of processes of order v G [1,2)

We confine ourselves to this case; the processes of order less than 1 can be
considered in the similar fashion but possible types of behaviour of /i(A)
are numerous.
Theorem 6.3 Let v G [1,2). Then for sufficiently small A > 0, the
solution x — /i(A) to Eq. (6.18) exists; it is unique, and as A —» 0, h(A) —*
h, where h is the unique negative solution to the equation
(6 24)
/ z(z + l) ^ - J s i n f "
Before proving Theorem 6.3, we make two remarks. First, the solution to
Eq. (6.24) exists, and it is unique since the LHS decreases monotonically
from +00 to 0 as x goes from 0 down (when x = 0, the integral diverges

due to the condition v > 1).
Second, as r —• 0, we have h —> —00, which makes the RLPE model
much more reasonable than its Gaussian counterpart. Recall that the early
exercise in non-optimal in the case r = 0 (see, e.g., Musiela and Rutkowski
(1997)) but nevertheless, in the Gaussian model, for arbitrary small r > 0,
the early exercise boundary satisfies /i(A) —• 1(= K) as A —• 0.
Proof. First we consider the case u e (1,2). Suppose that the following
two assumptions hold:
a) h = /i(A) ->• 0 as q = A - 1 + r -> +00;
b) — h(A) > Aq~xlv, where A > 0 is independent of q.
Then on (2h~1,h~1), we have e - 1 2 < e - 2 , q+ip(iz±0) x q, and <f>~ (iz) x 1
due to Eq. (6.22) 2 Therefore, the LHS in Eq. (6.18) is bounded away from
zero by
,i/h
c {-h)2-vdz=C]_{-h)l-v,
J2/h
where c, c\ > 0 are independent of q. Since 1 — v < 0, we conclude that the

LHS in Eq. (6.18) is unbounded as q —* +00 whereas the RHS stabilizes to
r. Hence, the solution with properties a) and b) does not exist.
Now assume that a) holds but instead of b), -hq1/" = 0(1). Then on
(-2ql'v, -q1/v), we have e~xz = O(l), q + i>{iz ± 0 ) x ? , and (j>~{iz) x 1.
Hence, the LHS in Eq. (6.18) is bounded away from zero by C2<^"~1)/V —•
+00 as q —* 00, and hence for large q, Eq. (6.18) is impossible.
We are left with the only possibility: h = h(A) < 0 is bounded away
from zero uniformly in q > C, for sufficiently large C. But when x — h
satisfies this property, the leading term of the asymptotics of the integral
in the LHS of Eq. (6.18) is determined by the integral over any interval of
the form (—qe, —a + /3), where e > 0 is an arbitrary constant. By using
Eq. (6.23), we see that the leading term of the LHS in Eq. (6.18) is equal
to the LHS in Eq. (6.24). This finishes the proof of Theorem 6.3 in the case
i/G(l,2).
2
For two functions / and g we write / x g if / = O(g) and vice versa.
Now, let v = 1. Then the LHS in Eq. (6.18) admits a bound from below
via
c3 I e-xz{-z)-xdz = c 3 f" exzz~ldz.
If x —• 0 and — xq is bounded, then the last integral is unbounded as a

function of q, and hence, Eq. (6.18) is impossible. Suppose that - x = A/q,
where A is bounded away from zero as q —* oo, and change the variable
z = (—x)~xz'; we obtain
fq _ fA dz
I exzz~1dz = / —z r - —• +oo as x —• - 0 ,
7i 7-1 e z'
and the proof finishes as in the case v G (1,2). •
Chapter 7
First-touch digitals
In this chapter, we consider simplest options, which are reducible to bound-

ary value problems with both the terminal condition and a fixed early ex-
ercise boundary. This makes the problem analytically tractable in contrast
to free boundary value problems for American options with the finite time
horizon. For first-touch digitals, the terminal payoff is zero, which makes
the problem technically simpler than similar problems for barrier options.
We also show that in a neighbourhood of the barrier, the non-Gaussian
RLPE-prices differ significantly from the Gaussian prices, not only quanti-
tatively but qualitatively as well.
7.1 A n overview
A first-touch digital is a digital contract which pays $ 1 when and if a specific

event occurs. Consider the first-touch digital (another name is a touch-and-
out option) which pays $ 1 the first time the stock price S hits or crosses
the level H from above. Possibly, it would be better say "hits (0, i7]". For
simplicity, in the sequel we say "crosses the level H". If the stock price never
crosses the level H before time T, the claim expires worthless. Denote by
Vd (H, T; S, t) the no-arbitrage price of such a contract. It can be interpreted
as an American put-like option with the digital payoff. Since the payoff is
the same for all levels of the stock price below the barrier, it is optimal
to exercise the option the first time the level H is crossed. The formulas
for the value Vu(H,T;S,t) of the similar contract, which pays $ 1 the first
time the stock price crosses the level H from below, easily follow by the
reflection of the real axis w.r.t. the origin. For explicit pricing formulas for
165
166 First-touch digitals
the case when the dynamics of the stock price is modelled as the geometric
Brownian motion, see, e.g., Ingersoll (2000); general formulas for the RLPE-
case were obtained in Boyarchenko and Levendorskii (2001a, 2002b), where
more general rebates than here were considered, namely, first-touch power
options, which pay not $ 1 but S&; and first-touch contracts which pay a
non-zero amount iff the first barrier has been crossed but the second one
(situated farther) has not. Notice that the contracts of this sort do not
make sense in the Gaussian theory since a continuous trajectory stops at
the barrier.
In addition to exact albeit complicated formulas, we obtain several sim-
pler approximate formulas. The first formula is valid near the boundary
H, and the next two are valid for options far from the terminal date. From
the asymptotic formula for the price of the first-touch digital as the price of
the underlying approaches the barrier we conclude that if there is no Gaus-
sian component and the process is purely discontinuous then the difference
between the option prices in the Gaussian model and in the RLPE model
is especially large near the boundary. Moreover, different RLPE models
usually produce essentially different types of behaviour, which can be used
to decide, which type of processes fits the data better.
7.2 Exact pricing formulas for first-touch digitals
The riskless rate r > 0 is constant, and an EMM Q is chosen so that under
Q, {Xt} = {InSt} is an RLPE of order v € (0,2] and exponential type
[A_, A+], where A_ < — 1 < 0 < A + ; V denotes the characteristic exponent
of X under the risk-neutral measure chosen by the market.
Set x = ln(S/H), u(x,t) = Vd(H,T;S,t). Then for t < T and x € R,
u(x,t) = E[e-rT'lT><T | Xt = x], (7.1)
where T" is the hitting time of (—oo,0] by X. By using Theorem 2.13, we

obtain
Theorem 7.1 The u is a solution to the following problem:
(dt-(r + iP(Dx)))u(x,t) = 0, x>0, t<T, (7.2)

u(x,t) = 1, x < 0 , t <T, (7.3)
u(x,T) = 0, x>0, (7.4)
Exact pricing formulas for first-touch digitals 167
in the class of bounded measurable functions.
Set T = T - t,v(x,T) = u(x,T - r ) , and rewrite Eq. (7.2)-Eq. (7.4) as

follows:
(dT+r + il>(Dx))v(x,T) = 0, x > 0, r > 0, (7.5)

V(X,T) = 1, x<0, r>0, (7.6)
v(x,0) = 0, x>0. (7.7)
To solve the problem Eq. (7.5)-Eq. (7.7), make the Fourier transform w.r.t.
T; since the terminal condition Eq. (7.7) is homogeneous, we obtain the
following family of the problems on R, parameterised by A with 3A < 0:
(i\ + r + il>(Dx))v{x,\) = 0, x>0, (7.8)

v{x,X) = (iA)~\ x<0. (7.9)
Denote by 9± the indicator function of R ± . We are going to apply the

Wiener-Hopf method in the form Theorem 5.6, to the problem Eq. (7.8)-
Eq. (7.9).
When iX is real, and q :— iX + r > 0, we can use the formulas obtained
in Chapter 3. To stress the dependence on A, we change the notation of
Chapters 2 and 3, and write a(A,£) = iX + r + V'(O) ^±(^> 0 a n ^ a±(X, £)
instead of a(£), 4>f(Q and a±(£). Notice that K±, the orders of the factors
a±(A, •), are independent of A.
The following consideration shows that we can use the factorization
identities obtained in Chapter 3, for A in the half-plane 9A < <7o provided
(T0 > 0 is sufficiently small. Formulas for 0 ± remain the same but in order
to get convenient estimates for a±, uniformly in A in an appropriate subset
of the complex plane, the construction has to be adjusted.
Recall that in the case of an RLPE, a(X, £) admits the analytic contin-
uation w.r.t. £ into a strip containing the real axis, and moreover, for each
q = iX + r > 0, there exist <r_ < 0 < cr+ such that
» o ( A , 0 > c0, V 3£ G (<7-,<r+). (7.10)
(In Chapters 3 and 5, we used Eq. (7.10) with c0 = 0; clearly, we can choose
-a- > 0 and o~+ > 0 so that Eq. (7.10) holds with c 0 > 0). It follows that if
0o > 0 is sufficiently small, then Eq. (7.10) holds uniformly in A in the half-
space 9A <CTO,and therefore we can repeat the constructions of Chapter

5 for these A.
By applying Theorem 5.6, we find
v(-,X) = ( M A ^ X i A r 1 ^ - . (7.11)
(Cf. Eq. (5.27)). If K_ < 1, the solution is unique in the class of bounded
functions, and if K_ = 1, it is unique in the class of bounded continuous
functions. Notice that the condition K_ = 1 holds (or fails) for all A simul-
taneously. The equality K_ = 1 implies that either the process is of order
2 and hence has a diffusion component, or it is a process of order u 6 (0,1)
and fi < 0, that is, a process of bounded variation, with negative drift. In
both cases, by using Eq. (7.1), we obtain that u is continuous. Explicitly,
Eq. (7.11) is
+oo+ia; +
e»^_(A,0(A0- 1 «,
/ -oo-f-iu;-)-
where w+ € (0,CT+) is arbitrary, and v is obtained by the Fourier inversion
/ eÂ+a:«)^_(A,0(AO-1^dA, (7.12)
-oo+ia J— oo+iu>+
where a < 0. We integrate by part w.r.t. £ and A in the open quadrant

(0, +oo) 2 , and by using the explicit formula for <^>_, obtain the integrand,
which (locally w.r.t. x and r ) admits a bound via
C U + K D - ^ I + IAI)-1-',
for some C,e > 0. We conclude that v is continuous on (0, +oo) 2 , therefore
Eq. (7.12) gives the price we are looking for. By returning to the initial
variables, we obtain
Theorem 7.2 For S > H and t < T,
+oo+i<r /•+oo+iw+
/ (7-13)
/
-oo-\-i<7 J — oo+iu;+
.e*((T-*)A+ln(S/ff)C)^_(Ai^(A^rl^dAj
for a < 0 and OJ+ € (0,CT+) sufficiently close to 0.
By making the inversion of the x-axis w.r.t. the origin, which leads to the
change X H. X, S i-» -S, H >-> -H, £ H-> - £ , and (f>-{\,0 >-» <MA, - 0 .
The Wiener-Hopf factorization with a parameter 169
and then making the change of variable £ i-» — £ in the integral, we obtain
the following theorem.
Theorem 7.3 For S < H and t<T,

+oo+icr /•+oo+iu>_
/ (7-14)
/
.e<((T-t)A+ln(S/if)O0+(A>O(AO-ldedAj
for a < 0 and u>- £ (<r_, 0).
The reader may notice that Eq. (7.13) and Eq. (7.14) are not computa-
tionally effective, though relatively short. Below, we derive fairly effective
formulas; some of them are long, however.
7.3 The Wiener-Hopf factorization with a parameter
The reader interested in the main formulas but not in the technical details
of the proofs can skip the next section, where we study the analytic con-
tinuation of the factors in the Wiener-Hopf factorization formula w.r.t. an
additional parameter A, and obtain necessary estimates.
7.3.1 Formulas for the factors revisited

To derive asymptotic formulas, we need both Eq. (2.23) and Eq. (3.71), but
we need them for A in a domain of the complex plane, which contains a set
of the form
E(e,0) = t e + E(0),
where
E(0) = {A| a r g A e ( - 7 r / 2 - 0 , - 7 r / 2 + 0)},
with e, 9 > 0, and the best situation is when a choice 6 > n/2 is possible.
Moreover, we need to have certain uniform estimates for the factors not
only w.r.t. £ in appropriate half-planes, but w.r.t. A g £(e,0) as well.
In such situations, the standard terminology in the theory of PDO is "the
factorization with a parameter17.
For the convenience of the reader, we rewrite Eq. (3.71) and Eq. (3.57)-
Eq. (3.60):
r + i\
= <MA,£)MA,0; (7.15)
r + i\ + il>(£)
the formulas for $ + (A,£), Q£> a-, are
P+oo+v_ £i n (iA + r + ^(77))

<MA,£) = exp drj (7.16)
2ni J_-oo+ip-
(
+oo+ip-
exp -F
^nij-oo+ip. »A + r + V(»?)
In^Zidr,
t]
(7.17)
where p- € (cr_, 9f), and the formulas for <^>_(A,f), 3£ < a+, are
<MA,£) = exp •drj (7.18)
+oo+ip+
= exp In v-t dr\ ,(7-19)
-oo+ip+ i\ + r + ip(r)) r)
where p+ € (9^, cr+). Recall that in Chapter 3, the formulas have been
proved for i\ + r > 0. Clearly, each of the expressions in Eq. (7.15) and
Eq. (7.16)-Eq. (7.19) admits the analytic continuation w.r.t. A into some
domain, but to study the estimates for the factors in the Wiener-Hopf
factorization formula, the second representation of the factors is more ap-
propriate.
Eq. (3.71) assumes the form
o(A,0 = d a + ( A , 0 a - ( A , 0 , (7.20)
and what is important, to obtain the answer in a simple form, we have to

modify the constructions in Chapter 3 in order that a±(A,£) be holomor-
phic w.r.t. A in a neighbourhood of some set S(e, 6) and satisfy uniform
estimates there. To this end, we need an additional restriction on the order
of an RLPE. We have to consider separately two cases. The first case is
v € (1,2]. Under this assumption, it possible to obtain good uniform esti-
mates for A E £(e,#) with 6 > ir/2. The second case is v < 1, and here we
impose the additional condition:
if v <E (0,1) then fi = 0. (7.21)

In the second case, a good estimate for the factors a± can be obtained on
A € £(e,0) with 9 < 7r/2, which is not sufficient for certain purposes. The
choice 9 > n/2 is still possible but only when a±(A, £) are constructed as
functions which are not holomorphic w.r.t. A (see Chapter 8).
Let c,n,u be as in Eq. (3.45)-Eq. (3.46). If v € (1,2] or v G (0,1)
and fi = 0, set d = c, K± — u/2; if v = 1, set d = (p? + c 2 ) 1 / 2 , n± =
0.5 ± 7T_1 arctan(/j/c). Thus, d and K± are the same as in Chapters 3 and
5. For s G R, set
A±(A,O s = ((iA + r ) 1 / " T ^ ) S ,
and fix 9 G (0,7ri//2). If u > 1, we can (and will) choose 9 > ir/2.
If p- < 0, and e > 0, are sufficiently small, then A+(A, £) s is well-defined
in the region A € £(e,0), 9£ > p _ , and satisfies there an estimate
|A+(A,0*l > c(l + IAI1/" + Ifl)*, (7.22)

where c > 0. Similarly, if p+ > 0, and e > 0 are sufficiently small, then
A_(A,£) s is well-defined in the region A € £(e,0),9£ < p+, and satisfies
there an estimate
| A _ ( A , 0 ' | > c ( l + |A|V" + |£|)«, (7.23)
where c > 0. Further, define
B(A (7 24)
'^-dA+(A,0«+A_(A,0«-- '
Due to Eq. (7.10), b(A,£) := ln.B(A,f) is well-defined for £ in a sufficiently
narrow strip 3£ € [p-,p+], where p_ < 0 < p+, and A € S(e,9), provided
e > 0 is sufficiently small. For £ in the half-plane 3 ^ > p_ and A e S(e, 9),
set
+
MA , f) i r "-^,, (7.25)
2?ri J-00+ip_ J] - £
a+(A,£) = dA+(\,t)K+ exp[6+(A,0], (7-26)
and for £ in the half-plane 3£ < p+ and A € £(e,#), set
MA,0 = --M Û, (7.27)

a_(A,0 = A_(A,0 K " exp[6_(A,0]- (7.28)

The functions Eq. (7.26), Eq. (7.28) satisfy Eq. (7.20), and the factors in
Eq. (7.20)-Eq. (7.15) are related by
(
*-™-$$- ™>
7.3.2 Analytic continuation of the factors w.r.t. A
If v > 1, fix 9 G (7r/2,7ri//2); if v G (0,1], fix 9 G (0,TTI//2), and assume
that Eq. (7.21) holds.
Lemma 7.1 Let X be an RLPE. Then there exist C,C\ > 0, p- < 0 < p+
and e > 0 such that for all £ in the strip 3£ € [p_,p+] and all A G £(e,0)
t/ie following three estimates hold:
|iA + r + tf(OI>ci(l + |A| + | £ r ) ; (7.31)
- 7 r < a r g B ( A , 0 < TT; (7.32)
and
C-1 <|B(A,TJ)|<C. (7.33)
Proof. It follows from Eq. (3.45)-Eq. (3.46) and Eq. (7.21), that as A ->
oo, and £ -¥ oo in the strip 9£ G [A_, A+],
I i\-inZ + c\£\, v-\.

Now, if C, i? > 0 are large enough, then Eq. (7.31)-Eq. (7.33) hold for
all (£,A) outside the ball |A|2 + |£| 2 < R2, which satisfy 9 f G [A_,A+],
arg A G [—7r/2 - 9, -n/2 + 9]. This can be shown as follows. If v < 1, the
estimates are evident due to the assumption 0 < 9 < iri//2 < n/2, and if
v > 1, then it suffices to notice that
\i\ + c | £ | f = |3?A|2 + | - 3A + c | £ | f
dominates the RHS in Eq. (7.31).

Fix these R and C. It remains to find C, c\ > 0, p_ < 0 < p+ and

e > 0 such that Eq. (7.31)-Eq. (7.33) hold for all (£, A) inside the ball of
radius R, such that 9£ € [p-,p+] and A G £(e,0). On the strength of
Eq. (3.45)-Eq. (3.46), there exist Co > 0, -p- > 0 and p+ > 0 such that
r + Sty(f) > 2CQ. Set e = Co/iJ. Then for A € E(e, 0) satisfying |A| < fl, we
have R(iA) > - c o , and Eq. (7.31)-Eq. (7.33) follow. D
From now on, we assume that the conditions of Lemma 7.1 hold, and use
e, p~, p+ constructed in this lemma. We also fix arbitrary w_ G {p~, 0) and
uj+e{0,p+).
L e m m a 7.2 a) There exist C,C\,c\ > 0 such that for all £ in the half-
plane Sf > w_ and aZi A € Y,(e,8),
|fc+(A,OI<C, (7.34)
and
d ( l + IAI1/" + |£|)»+ < |a + (A,OI < C i ( l + IAI1/" + |£|)*+; (7.35)
ftj There exists C,C\,Ci > 0 suc/i i/iai for all f in i/ie half-plane Sf < w+
and all A G £(e,0),
l&-(A,OI<C (7-36)
and
Cl(l + IAI1/" + K|)»- < |a_(A,OI < C i ( l + IAI1/" + | f | ) " - . (7.37)
Proof. We prove Eq. (7.34)-Eq. (7.35); Eq. (7.36)-Eq. (7.37) are proved
similarly. It suffices to prove Eq. (7.34); Eq. (7.35) is an evident corollary.
By using Eq. (3.45)-Eq. (3.47) and Eq. (7.33), we easily obtain the
following estimates
\B(X,n) - 1| < d ( l + |A| + M " ' ) / ( l + |A| + \ v n (7-38)
^^(A.^/BCA.II)! < C 2 (l + IAI1/" + M)"'- 1 ', (7.39)
where C\ and C?, are independent of A € £(e,#) and n in a strip 3 ^ €

[p-,p+], as well as all constants below, and v' < v.
By making the shift in the f-space, we may assume that in Eq. (7.25),
77 G R, and a := 3 f > 0. Change the variable f -> f + io, set K =
(IAI + 1) 1 /", and for each pair (A, £), introduce intervals Jj C R:
Ji = {fl\\ri-S\<K},
h = {ri\\n-Z\>K, \v\<K},
J3 = fo||»7-fl>M, \V\>K},
J4 - {fl\K<\r,-i\<\rt\, \r,\>K}.
By using the mean value theorem and Eq. (7.39), we obtain
b(X,V) b(X,0
(7.40)
£ + ia — T] £ + ia — r\
where
l^(A^,r7,a)|<C3(l + | A | ) ^ - ^ . (7.41)
Since
f
J-K
drj
-AT ™ - 7J
In-
-K-ia
K-ia
<2?r,
we deduce from Eq. (7.40)-Eq. (7.41) and Eq. (7.33)
b{X,j])dr]
\[ *{X\v)dTI <2nlnC + C3 [ (1 + \X\)^'-^"df, = C4. (7.42)
To prove the following estimate, only Eq. (7.33) is needed:
b(\,T))dT)
\[P±^<C5[ (l + |A|)-1/"c67 = C6. (7.43)
Further, we infer from Eq. (7.38) that b admits an estimate of the same
form as B — 1, and using this estimate on J3, we obtain
b(X,r])dT]
/ £ + ia - 7]
<C- f ~
1+W+ W dr). (7.44)
JJ3 J\n\>K \v\0-
By changing the variable 77 = Kr)', we see that the RHS in Eq. (7.44) is
bounded uniformly in A £ E(e, 0), £ £ R, a > 0. Since v' £ [0, v), a function
/ ( S ) = (l + |A| + S " ) / ( l + |A| + 0

is decreasing on [0, +00), and therefore, we deduce from Eq. (7.38) an esti-
mate, for 7] G J4,
|6(A,i,)| < C 8 (l + |A| + |£ - r,|" ) / ( l + |A| + |£ - 7,1"). (7.45)
From Eq. (7.45),
bMdT l + |A| + |g-q|"

[ > <cs[ di,, (7.46)
and the change of variable 77 = £ + Jf T/ shows that the RHS in Eq. (7.46)
is bounded uniformly in A G £(e, 0).
By gathering bounds Eq. (7.42)-Eq. (7.44) and Eq. (7.46), we obtain
Eq. (7.34). •
Lemma 7.3 a) For any N, there exist C, p > 0 such that if 9£ > w_,
A G £(e,0), and |A| < iVln(|f| + 1), then
IMA,0I<C(1 + |£|)-", (7.47)
|a + (A,0-dA + (A,0 K + |<C(l + |£|r +-p- (7.48)
6j For any N, there exist C, p > 0 suc/i tftai i/ 3£ < LJ+, A G £(e,#), and
|A| < JVln(|f| + l ) , i/ien
\b-(\,0\<C(l + \t\)-', (7.49)
and
ia_(A,o-A_(A,oK-i<^(i + ieir- (7.50)
Proof, a) and b) are proved similarly, and Eq. (7.48) is immediate from
Eq. (7.47). To prove the latter, take e G (0,1), and consider the integrand in
Eq. (7.25) in the regions |ri| < (f) e , and \n\ > (£)£, where {£) = (l + l ^ 2 ) 1 / 2 .
For |r?| < (£) e , we have
K- r/T1 < c ^ r 1
and for any a > 0,
i\ + r + il>{n)
In < CAv)°
dA+{\,r])K+A_(\,n)K-
Hence, for a positive c-i and 9£ > c-i + p-, we obtain an estimate
r+°°+*p- b(\,T))
w
L -oo+ip- £ — *?
M<(Oe^drl <c(0~ 1+€(1+<7) -
If \v\ > (0% e use Eq. (3.45)-Eq. (3.46) to obtain that
1+ U l W
dA+(A,n)«+A_(A,r?)«- ''
where pi = v — v\ > 0, therefore
b(\,n) = 0((r,)-").
Now, to finish the proof of Eq. (7.47), it remains to notice that
+oo dn
< c(0~pl/2
/ -oo WHS-V) n
Lemma 7.4 a) There exist C, p > 0 such that for all £ in the half-plane
3f >w_ and all X 6 £(e,0),
\dxb+(X,0\< 0(1 + 1X1^ + ^)-", (7.51)
and
|9 A (a + (A,a ±1 )l < C(l + IAI1/" + | ^ | ) ± K + - p ; (7.52)
fc) There exists C, p > 0 suc/i t/iat /or aH £ in ihe half-plane S£ < w+ and
aHAeE(e,0),
|aA6_(A,0l < C7(l + IAI1/" + |ei)- p , (7.53)

and
^(a-CA.O* 1 )! < C(l + lAI1/1' + \m±K-p- (7-54)
Proof. It suffices to prove Eq. (7.51). We have, with some Ci, C2,pi > 0,
|9A5(A,n)| < C1(l + \X\ + \v\rpi,

\B(X,r,)\ < C2,
and therefore,
l^6(A,n)| < C 3 (l + l A i r ' ^ l + Ir?!)-"1/2. (7.55)
Price near the barrier 177
From Eq. (7.55), we see that to finish the proof, it suffices to apply the
estimate
+°° dr]
I 2
(\)Kt WHZ-Tl)
<C(A>-pi/2(£)~pl/4
D
7.4 Price near the barrier
We consider V(X,T) = Va(H,T;S,t), the reader can easily reformulate the

results for Vd(H,T; S, t).
7.4.1 Asymptotics as x —> + 0 , and r > 0 is fixed

In this Subsection, we fix r > 0, and compute the leading term of the
asymptotics of v(x, r ) as x —> +0.
7.4.1.1 Main Theorem

We start with the case v>\. Let 6 € (7r/2, TTV/2) and e > 0 be sufficiently
small so that all the estimates of the Section 7.3 are valid on the set E(e, 0)
defined at the beginning of Section 7.3. Choose a symmetric contour £ C
£(e,#), which passes below the origin and coincides with the line of the
form A(s) = —Co + io~- + (1 — ie)s in the neighbourhood of s = — oo,
where e > 0. Of course, the symmetry condition is imposed to simplify the
formulation only.
Set
V(T) = (27a)- 1 f e iAr a_(A,0)A" 1 dA. (7.56)
(Here the normalization
a_(A,0/(l + tO"" "*1 as£->±oo

is important).
Theorem 7.4 Let v > 1. Then there exists 8 > 0 such that as x -> +0,
v(x, T) = 1 - T(K_ + l ) - 1 © ^ ) * " - + 0(xK'+s). (7.57)

Remark 7.1
a) The asymptotics Eq. (7.57) holds not only for processes of order u > 1
but for processes of order 1, and for processes of order v G (0,1), if /x — 0,
as well; only the factors in the Wiener-Hopf factorization formula have
to be defined differently, as in Chapter 8. The proof remains the same.
b) The asymptotics Eq. (7.57) is uniform on each segment [TO, +OO) of the
positive half-axis in the sense that for each To > 0, the constant in the
O-term can be chosen the same for all T >TQ.
c) If there is a Gaussian component, then K_ = 1, and from Eq. (7.57)
we conclude that the solution is of the class C1 up to the boundary. If
there is no Gaussian component and Eq. (7.21) holds, then «;_ 6 (0,1),
and the price of the option behaves not as in the Gaussian case.
d) Still, the relative error of the Gaussian approximation is not large here
since the leading term is 1 in the both models. Similar effect is more
pronounced for barrier options without a rebate and junk bonds, since
in these cases, the ratio of the RLPE-price to the Gaussian price may
be very large indeed.
7.4.1.2 Numerical example

We take the riskless rate r = 0.05, and assume that under the EMM chosen
by the market, X is a NIG with parameters S = 7.2, a = 40, /? = 8, and
H, fixed by the EMM requirement r + i/> Q (-i) = 0. Then K_ = 0.5682,
and therefore, near the barrier, for r > 0 fixed (and not too small) the
difference 1 — v(x, r) behaves approximately like const • yfx. The numerical
calculations show that this is really the case (see Fig. 7.1). Recall that in
the Gaussian model, the price is smooth up to the boundary, and therefore,
the difference 1 — v(x, T) is approximately linear as a function of x, near the
barrier. Thus, one should expect the significant difference between NIG and
Gaussian prices, the latter being calculated for the volatility a2 = (V>Q)"(0);
this difference is apparent on Fig. 7.2.
Further, from Eq. (7.57), we conclude that for a fixed T > 0, the function
W(X,T) := (1-V(X,T))/XK-
is approximately constant near the barrier, and Fig. 7.3 demonstrates this
effect.
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14

In (S/H)
Fig. 7.1 The NIG-price of the down-and-out option, at T = 3,5,10 and 15 days to
expiry. Parameters: r — 0.05,5 = 7.2, a = 40, /3 — 8.
7.4.1.3 Proof of Theorem 7.4

Proof. By shifting the line of the inner integration in Eq. (7.12) and
using the residue theorem, we obtain in the region x > 0, r > 0:
v = 1 + /,
where f(x,r) is given by
+00+MT p + CO+iul-
/ e i < r * + *«0_(A,O(AO- x d£dA, (7-58)
/
-oo+itr J — oo+iu>-
a < 0 and w_ E (<r_,0).

Integrate by part w.r.t. A in Eq. (7.58). In view of Eq. (7.54), we obtain
absolutely converging integral. Moreover, we can justify the deformation of
the line 9A = a into the contour C. After that we integrate by part once
0.16
• x=3
x T=5
- - T=10
x=15
xxx*xx**x><
0.02 0.04 0.06 0.08 0.12 0.14

In (S/H)
Fig. 7.2 The Gaussian model vs. NIG model: the price of the down-and-out option, at
T = 3,5,10 and 15 days to expiry. Parameters: r = 0.05, <S = 7.2, a — 40,0 = 8.
again and obtain
+ QO+IU1-
f(x,T) = (2n)-2 jj_ ri(rx+xe) a-(A,0) d£d\. (7.59)
'CJ-oo+iw- a~(A,f)A£
(Here we have used Eq. (7.30)). Notice that there exist C, c0 > 0 such that
9A>co|A|, V A e £ : |A| > C.
Hence, the integrand in Eq. (7.59) admits an estimate via
C e -cor|A| ( 1 + |A|)"-/"(1 + |A|X/" + I f l ) - " - ^ - 1 ! ^ - 1 ,
and therefore the Fubini theorem applies:
r+oo+iu)-
f(x,r) = (27T)'2 / + 0 O + " " f e**+*> a
~n'l!l*W. (7.60)
-oo+iw-
x x x x x x x x x x x
KXXXX X X X X X x x x x x x x
• t=3
X T=5
- - t=10
T=15
10
In (S/H)
Fig. 7.3 Graph of W(x,r) near the barrier. Parameters: r — 0.05,5 = 7.2, a = 40,
8, at T = 3,5,10 and 15 days to expiry.
In addition, there exist N,C such that if |A| > JVln(|£| + 1), then
JXr O-(A,0)
<Cexp[-r|A|co/2]($)- K — 2 , V^<w_,
and the same estimate holds for
JXr O-(A,0)
{l + iOK-+1X'
From Eq. (7.50), on the set |A| < iVln(|£| + 1), 9£ < w_ we have the
estimate
iXrf a_(A,0) a_(A,0)

^ C a e x p h c o r l A O I A-rl1/^C)\ - K - - l - p
a_(A,0Â (1+*£)«-+%
for some p > 0, C2 > 0. By taking the last three estimates together, we
derive the representation
a
I ^
Jc
n)vtxdX
o-(A,f)*f A
= ^l + i^~K"1 I eiXT^^-dX
Jc A
+ «&(0, (7-61)
where w(£) is holomorphic in the half-plane 3£ < w_ and admits an esti-

mate
\w(0\<C(0-K-1-p, 9£<w-, (7.62)
with some p > 0. It follows from Eq. (7.62), that for any e > 0, w belongs
to the Sobolev space HK-+1/2+p~e(R) and vanishes on (—oo, 0]. But for
any s > 1/2 and e 6 (0, s - 1/2), HS(R) is embedded into the Holder space
C*-i/2-e(H,) j therefore w € C K - + " _ € ( R ) , V e > 0. Since w vanishes on
(—oo,0], we conclude that
w(x) = 0(xK-+p-e), Ve>0. (7.63)
Further,
+oo-|-iu»_
eixi{l + i O - " - - 1 ^ = T(K_ + l j ^ i ^ - e - * , (7.64)
/ -oo-j-iu'-
and by combining Eq. (7.61), Eq. (7.63) and Eq. (7.64), we arrive at
f(x, T) = - r ( / s _ + 1)~1V{T)XK+- + 0(xK-+"-6). (7.65)
Since v = 1 + / , Eq. (7.57) follows from Eq. (7.65). D
7.4.2 Asymptotics as x —> + 0 and r —»• + o o

Transform the contour C into a contour £ i lying above the line 9A = a+ >
0; if a+ is not large, we remain in the domain of the analyticity of a_ (A, 0).
By using the residue theorem, we obtain
V{T) = a _ ( 0 , 0 ) + (2 7 ri)- 1 f e iAr a_(A,0)A~ 1 dA. (7.66)
The second term in Eq. (7.66) exponentially vanishes as r —> +oo, and we
obtain from Eq. (7.57):
Asymptotics as r —y +00 183
Theorem 7.5 There exist 5 > 0 and o+ > 0 suc/i £/ia£ as x -» +0 and
r -> +00,
U(X,T) =
* " rfc^+V"" + ^ " " ^ + °( e ~' + r )- (7-67)
7.5 Asymptotics as r —)• + 0 0
The result of this section is intuitively clear: when the terminal date is far
away, the price is essentially independent of time, the generalized Black-
Scholes equation reduces to the stationary one, and therefore, the price
is approximately equal to the solution of the stationary boundary value
problem on the half-axis. The formal proof is as follows.
In Eq. (7.58), transform the line of the integration 9 A — a into the
contour C, next change the order of the integration as in the beginning of
the proof of Theorem 7.4, and then transform C into C\. Then, similarly
to Theorem 7.5, we obtain
Theorem 7.6 There exists a+ > 0 such as r —* +00, uniformly in
x € [0, + 0 0 ) ,
+oo+iu>+
e ^ < / » _ ( 0 , 0 ( - ^ ) - 1 ^ + O( e - f f + r ), (7.68)
/ -oo+iw+
where w+ > 0 is sufficiently small.
Note that in the proofs of Theorem 7.5 and Theorem 7.6, an additional
assumption on v is used but they can be proved for any 1/, in the same way
as the pricing formula for the perpetual American options was derived in
Chapter 5.
Chapter 8
Barrier options
Barrier options are reducibJe to boundary value problems with a non-trivial

terminal condition and a fixed early expiration boundary. In the event of
early expiration, the option owner may be entitled to a rebate, that is,
the boundary condition can be both homogeneous and inhomogeneous. In
contrast to the Gaussian case, a rebate (if any) must be specified not at
the barrier only but everywhere on the other side of the barrier as well.
8.1 Types of barrier options
8.1.1 Standard barrier options

Consider a contract which pays the specified amount at the terminal date,
provided during the life-time of the contract, the price of the underlying
asset does not cross a specified barrier S = H(t) from above (down-and-
out barrier options) or from below (up-and-out barrier options). When
the barrier is crossed, the option expires worthless or the option owner is
entitled to some rebate. For simplicity, we study constant barriers only,
though non-constant barriers can also be examined.
Consider barrier options on a stock, without a rebate. The standard
variety of these options comprises four types of down-and-out and up-and-
out options, and four types of down-and-in and up-and-in options. The
"out" options with the barrier H are
(1) the down-and-out call option with the terminal payoff (ST — K)+\
we denote by V&0tCa\\(K, H, T; S, t) the price of this contract at the
185
186 Barrier options
time t and spot-price S;

(2) the down-and-out put option;
(3) the up-and-out put option;
(4) the up-and-out call option.
An "in" option becomes the European option when the specified barrier
is crossed (before or on the terminal date); otherwise it expires worth-
less. For instance, the up-and-in call option becomes the European call
option when the barrier is crossed from below; we denote its price by
Vui;ca.u(K, H,T;S,t). Similarly, one defines the up-and-in put option, and
down-and-in puts and calls. In all states of nature, a European option pays
the same amount as the portfolio of the down-and-out option and up-and-
in option with the same barrier (of course, this is pertinent to a portfolio
of the up-and-out option and down-and-in option with the same barrier
as well), therefore the standard no-arbitrage consideration shows that the
pricing problem for an "in" option reduces to the pricing problem of the
corresponding "out" option and a European option, for instance,
VdiiCall(K, H, T; S, t) = VcM(K, T; S, t) - VUO)Call(tf, H, T; S, t), (8.1)
where Vca\\(K, T; S, t) is the price of the European call with the same strike
K and expiry date T. The equalities similar to Eq. (8.1) hold for other
pairs of barrier options, therefore it suffices to write the explicit formulas
for "out" options only.
8.1.2 Options with a rebate

Suppose that when the barrier is crossed from above, a European option
expires but the option owner is entitled to some rebate. If the rebate is a
constant equal to R, then the standard no-arbitrage consideration shows
that the price of the barrier option with a rebate is equal to the price of
the portfolio, which consists of the same type of the barrier option but
without a rebate, and the first-touch digital with the payoff R. The same
observation holds for "up" options, and similarly, options with more general
rebates can be considered.
The first-touch digitals having been studied in Chapter 7, we conclude
that it suffices to consider barrier options without a rebate. In this case,
from the mathematical point of view, the pricing problem is a backward
parabolic equation on the half-line, and so the answer can be deduced from
Down-and-out call option without a rebate 187
the general theory of Chapter 15. We illustrate the usage of the general
theory by explicitly solving the pricing problem for the European down-
and-out call V<i0;ca,ii(K,H,T;S,t) in the case K < H. Similarly, other
barrier options can be considered. For the pricing formulas of other barrier
options, more general payoffs and rebates including, see Boyarchenko and
Levendorskii (2002b), where simpler and more direct approach is used. No-
tice, however, that the general theory of Chapter 15 admit a straightforward
generalization for the case of option pricing under Levy-like Feller processes,
when only approximate analytical formulas can be obtained (see the end
of Chapter 16), whereas the approach in Boyarchenko and Levendorskii
(2002b) relies on the explicit analytical representation of the solution, and
therefore, it is more difficult to generalize the latter approach.
8.2 Down-and-out call option without a rebate
8.2.1 Reduction to the boundary problem for the General-

ized Black-Scholes equation
We assume that the riskless rate r > 0 is constant, and under a risk-neutral
measure chosen by the market, X is an RLPE, of order v and exponential
type [A_,A + ], where A_ < —1 < 0 < X+.
Set x = In 5, denote by <?(x) = (e x — AT)l[ 0i+oo )(x) the terminal payoff
of a down-and-out call option with the barrier H = 1, without a rebate.
Let v{x, t) = Vdo(g; 1,T;S, t) be the price of the option at the time t < T
and spot price S > 1. Then
v(x,t) = E^[e^T^g(XT)lT0>T
where To is the hitting time of (—oo, 0] by the process X. According to the
general scheme described in Section 2.4, we approximate g by a sequence
of non-negative functions gn with compact support, which converges to g
pointwise: <7«(x) T g{x) f° r an Y x >Q. The standard construction of such
a sequence is as follows: take a non-increasing function \ S C°°(R) such
that x(x) = 1 for x < 1, x(x) = 0 for x > 2, and 0 < x(x) < 1 for x S [1,2],
and set gn{x) = g{x)x{x/n). Set
vn(x,t) - £(*•*> [e^-T)rgn(XT)lT0>T] .

On the strength of the dominated convergence theorem, we conclude that
188 Barrier options
for any (x,t) € (0,+co) x [0,T),
vn{x,t)]v{x,t). (8.2)
By applying Theorem 2.13 to vn(x, t), we find that vn is a bounded mea-

surable solution to the following problem
(dt - (r + 1>(Dx)))vn(x, t) = 0, x > 0, t < T; (8.3)

vn(x,T) = 5n(x), x>0. (8.4)
vn(x,t) = 0, x<0, t<T. (8.5)
Set T = T — t, and to n (x,T) = vn(x,t); we obtain the Cauchy problem

for a pseudodifferential parabolic operator on R+, with the zero boundary
condition:
(dT + (r + V(£> I )))^„(a;, t) = 0, x > 0, T > 0; (8.6)

w„(x,0) = flf„(x), x>0. (8.7)
tu n (x,T) ^ 0 , x<0, r>0. (8.8)
o*'
Here the initial data gn eH (R-+), for any s' < 1/2, since gn is smooth
up to the boundary x = 0, and has the compact support, and the unknown
wn is bounded and measurable. Take 7 G (A_,—1), and define wnn(x) —
eyxwn(x). Similarly, define 5,1,7,57, and w7. The choice of 7 is explained
as follows: <77(x) must decay as x —• +00 so that <77 S L2(R+). Thus, if we
consider a payoff with the different rate of growth at the infinity, another
choice of 7 may be needed; and for up-and-out put option, say, when the
problem on R_ is to be considered, and the payoff stabilizes to a constant
as x —> —00, one should choose 7 € (0, A + ).
Substitute wn(x) — e~yxwnn and gn(x) — eT",xgn~1 into Eq. (8.6)-
yx
Eq. (8.8), next multiply by e , and use the equality
e^V(£>x)e- 7X = V(£* + i7);
the result is
(dr + (r + il>(Dx + iy)))wn„(x,t) = 0, x > 0, r > 0; (8.9)

Wnn(x,0) = 9n,y(x), X > 0, (8.10)
wn^(x,r) = 0, x < 0, T> 0. (8.11)
Recall that we understand the problem Eq. (8.9)-Eq. (8.11) in the sense
of generalized functions, and we look for a solution which is bounded and
measurable on the strip [0, +oo) x [0,T]. The standard argument shows
that such a solution is unique, and we will explicitly construct it by using
the representation theorem for analytical semigroups.
8.2.2 Reduction to the Cauchy problem for an ordinary dif-

ferential operator with the operator-valued coefficient
Set a(£) = r + ip(£), wnn(T) :— wn^(-,r). Let p+ be the restriction-to-R +
operator; consider
• p+a(Dx + ij) as an unbounded operator A-, in Z,2(R-+) with the

domain Z>(Ay) ={u= (A 7 + CI)-xv \ v € L2(R)}, and
• the problem Eq. (8.9)-Eq. (8.11) as the Cauchy problem for an
ordinary differential equation with the operator coefficient:
w'nn(r) + A^wn^r) = 0, r>0, (8.12)

w„, 7 (0) = gn,f, (8.13)
and the unknown wnn is regarded as a continuous (vector)-function

on [0, +oo), assuming values in P ( A 7 ) .
Here / denotes the identity operator, and C is large enough so that on the
line 3£ = 7,
8B(£ + i7) + C > o ( l + K i r , (8.14)
with some CQ > 0; the existence of such C and Co follows from Eq. (3.45)-
Eq. (3.46).
The inverse (Ay 4- CI)-1 is constructed as follows. Under condition
Eq. (8.14), the symbol a(£ + ij) + C admits a factorization
a(£ + ij)+C = a 7 ,c ; + (£) a 7,C;-(£),
and we define
(A 7 + CI)~lv = o7,C;-(I')-1e+a7,c;+(£')-1^,
where Iv € L,2(R) is an arbitrary extension of v £ Lg(R+). Since a 7 ,C;±(£)~

are bounded in the upper and lower half-planes, respectively, we have:
190 Barrier options
1
1) on the strength of Theorem 15.13, 0+a7tC;+{D) lv is independent
of the choice of Iv e Z<2(R), and the map
£ 2 ( R + ) 9 » H 6+a7^+(D)-Hv 6 L 2 (R+)
is bounded;
2) on the strength of Theorem 15.12, the map
L 2 (R+) 9 / H+ ^ . . ( Z ? ) - 1 / € L 2 (R+)
is bounded.
Hence, (A 7 + CI)-1 is a bounded operator in L 2 ( R + ) , and a direct
calculation shows that it is the inverse to
Ay + C J : 2 > ( A y ) - * L 2 ( R + ) .
(see Theorem 15.22). On the formal level, the system Eq. (8.12)-Eq. (8.13)
is easy to solve:
*n,7( r ) = exP[-TAi]gn,y, (8.15)
but in order to justify the answer, and to ensure the calculation of the
limit Eq. (8.2), some additional efforts are to be made. In order that
the representation theorem for analytical semigroups (Theorem 15.18) be
applicable, we have to impose an additional restriction on the order v and
fi in Eq. (3.45):
if i/ S (0,1), t h e n / i = 0. (8.16)
Under this condition, the elliptic part of the Generalized Black-Scholes

equation is strongly elliptic, which makes the study of the corresponding
(backward) parabolic equation especially simple, and allows one to obtain
an answer in the form of the absolutely converging integral, which is impor-
tant for numerical implementation. The advantage of the general approach
outlined in Chapter 15 for PDO with constant symbols, and at the end
of Chapter 16, for PDO with variable symbols, is that the perturbation
arguments can be used relatively easily; and these arguments are necessary
for any approximate solution to be validated, hence, for any generalization
to the case of Levy-like Feller processes to be made (cf. Chapter 14, where
the approximate pricing under Levy-like Feller processes is studied).
The rigorous sense can be assigned to Eq. (8.15) as follows. Suppose,

we have shown that the resolvent (A + A 7 ) _ 1 satisfies an estimate
IKl + IAIXA + ^ m s d , (8.17)

where C\ is independent of A on the contour of the form E ^ c := {A | |A| >
C, argA € (-9,9)}, where C > 0 and 9 G (ir/2,ir). Then we can apply
Theorem 15.18, and obtain
wn„(r) = (27TI)-1 f erA(A + A^g^dX, (8.18)

J C-e,c
where £e,c is the boundary of Ee,c- Moreover, for any r > 0, the RHS
defines an operator, which is bounded in L2(R+). From our construction
of the sequence {gn,-y}n>i, it follows that it converges in Z,2(R+) to <?7;
hence, by passing to the limit in Eq. (8.18), and multiplying by e _ 7 X , we
obtain
lim wn(;T) = (2m)-1 f eTXe-^(X + AJ)-1e''gdX. (8.19)

n—*oo I r
When we construct the resolvent, we will be able to show that the RHS in
Eq. (8.19) is a continuous function; hence, the limit on the LHS is the price
of the down-and-out call option we have been looking for:
W(-,T) =(2iri)-1 f e TA e-T(A + A 7 )- 1 e' 1 '- 3 (iA. (8.20)

J Co,c
As it is to be expected, the resolvent will appear to be of the form
(A + A^)-1 = a_(A, D + i 7 )- 1 0+a + (A, D + i 7 ) - \ (8.21)
where a±(A,£) are properly constructed factors in the Wiener-Hopf factor-

ization formula. In the case v > 1, we can use the factors a± constructed
in Chapter 7 (notice that the "A" here corresponds to "iA" in Chapter 7),
but in the case v < 1, this construction fails. An appropriate construction
will be given in the next subsection, and now we finish the calculation of
the price.
By using Eq. (8.21) and the equality
e- 7X a±(A, D + i 7 ) ~ V x
= a±(A, D)'1,
192 Barrier options
we can rewrite Eq. (8.20):
w(;T) = (2ni)-1 f e T A a_(A,Z))- 1 6' + a + (A,£>)-VA- (8.22)

J Ce,c
In the simplest case g(x) = (e x — K)l[oi+ao-)(x), which we consider here,
d+a+^Dy'g = 8+a+{\,D)-l{ex-K)
= ^+(a+(A,-i)-1e--JPfa+(A,0)-1),
and since the Fourier transform of the function on the RHS is equal to
o+(A)-i)-1(^-l)-1-/fo+(A>0)-1(tO-1,
we arrive at the pricing formula
W{X,T) = (271-i)-1 / / eTA+il?a_(A,£)_1 (8.23)

•>£e,c J—oo+ia
•[a+(A,-i)-1(^-l)-1-ifa+(A,0)-1(ie)_1]Â.
We have proved
Theorem 8.1 The rational price of the down-and-out call option with
the barrier H = 1 and the strike K < H is given by
1 /• r+oo+ia
T t A+i
Vdo;cM(K,l,T;S,t) = ^- / e( - ) « l n S a _ ( A , 0 - 1 (8-24)
2m
JC»,C J-oo+ia
•[a+(A, - i ) ( * £ - I ) - 1 ~ ^ a + ( A , 0 ) - 1 ( i O " " 1 ] ^ A .
_1
8.2.3 Construction of the resolvent, and the Wiener-Hopf

factorization with a parameter
Parameters 6 and C can be found as follows.
Lemma 8.1 Let Eq. (8.16) hold. Then there exist 9 € (n/2,n), C > 0
and Co > 0 such that for any A € S»,c and £ in the strip 9 ^ S [A_, A + ],
|A + a ( O I > c o ( l + |A| + K r ) . (8.25)
Proof. From Eq. (3.45)-Eq. (3.46), we conclude that there exist R, c\ > 0
such that for all f in the strip $£ G [A_, A+], outside the ball of radius R,
centered at 0,
K a ( £ ) > 2 c 1 ( l + |£|) I/ . (8.26)

Consider the subset
U(C,Cl) := {(A,0 | A € C, !% € [A_, A+], C < |A| < Cl(l.+ \Z\T).
If C is sufficiently large then on U(C, c\), we have |£| > R, hence Eq. (8.26)
holds, and therefore, Eq. (8.25) holds as well, with some Co > 0. If C\ > 0
is sufficiently large, and |A| > Cj(l + |£|) y , then from Eq. (3.45)-Eq. (3.46),
we conclude that Eq. (8.25) holds, with possibly another CQ > 0. Finally,
let ci and C\ be chosen as above. Then on a set V(ci, C\, C), defined by
c l (i+Kir<iAi<ci(i+Kir, IAI>C,
we have, as C —> +oo:
A + a ( 0 ~ A + c|er, Hu^l, (8.27)
and
A + a ( 0 ~ A - i / i £ + c|£|, if v = 1. (8.28)
Define W{ci,CuC,6) as a subset of V(cuCuC), where argA e [-0,0]. It

is evident from Eq. (8.27)-Eq. (8.28) that if O — TT/2 is small, and C is large,
then on W{cud,C,6), Eq. (8.25) hold.
This finishes the proof of Lemma. •
Notice the importance of the condition Eq. (8.16): if v G (0,1) and /x ^ 1,
then it is impossible to obtain the estimate Eq. (8.25).
Now, by using Eq. (8.25), we will construct the Wiener-Hopf factoriza-
tion of a(A, £) := A + a(£). The reader may notice that we have changed
the notation of Chapter 7 in order to use the standard form of Eq. (8.18).
If v > 1, then we can use the factorization constructed in Chapter 7, with
the factors analytic w.r.t. A € T,gtc- In other cases, this nice factorization
is impossible, and the construction must be changed.
Fix a branch of In by the requirement that In y is real for y > 0, set
A±(A,£) s = (1 + IAI17" =p i£)s = exp[sln(l + \\\VV =F t£)], and choose d > 0
and K _ , K + s R a s i n the case A = 0 in Chapter 3:
1) if v G (0,2], v ^ 1, we set d — c, K_ = K+ = i//2;
2) if v = 1, we set d = (c2 4- / i 2 ) ^ 2 , K± = 1/2 ± -K~X arctan(/i/c).
The next steps are analogues of the corresponding steps in the construc-
tions of the factors in the Wiener-Hopf factorization formula in Chapter 3,
but two additional subtle points are to be taken into account:
194 Barrier options
(i) for each A 6 £e,c, the Wiener-Hopf factorization formula must be

proved for £ not only in a sufficiently narrow strip containing the real
axis, as in Chapter 3, but for all £ in the strip (A_,A + ), and
(ii) the estimates must be uniform w.r.t. A G Tig,c, and £ from any strip
of the form S£ € [p-,p+], where A_ < p_ < p+ < A + .
These two conditions allow us to derive the pricing formula for any con-
tinuous payoff which grows as S13 as S —• +oo, for any /? 6 (—A+, —A_) J
Notice that due to the EMM-condition r + ip{-i) = 0, (ii) fails if C — 0
and p- = — 1. This observation highlights the importance of the flexibility
of the contour Ee,c- In particular, with the choice C = 0, it is impossible
to use the pricing formula for the down-and-out call.
With our choice of K±, function
5 ( A , 0 :=d-1A+(A,0-'e+A_(A,fl-'t-(A + a(0) (8.29)
satisfies for all A € T>e,c, £ € R and a € [A_, A+]
lim B(\,£ + ia) = l, (8.30)

£-»±oo
and 6(A, £ + io") = In B(X, £ + ia) is well-defined for these A, £, a\ the proof
is the same as in Chapter 3.
Lemma 8.2 For any X € £e,c and a £ [A_,A + ], the winding number
around the origin of the curve {B(X, £ + ia) | — oo < £ < +00} is zero:
,.£=+00
(27!-)- 1 / daxgB(X,^ + ur)=0. (8.31)
Proof. Prom Eq. (8.25), B(A,£) ^ 0 V A £ E f l , c and £ in the strip

S£ G [A_,A + ]. Taking Eq. (8.30) into account, we conclude that the LHS
in Eq. (8.31) is an integer. Moreover, this integer is independent of A e
Y,Qtc and a € [A_,A + ]. If C > 0 is large enough, and X — C, then from
Eq. (3.45)-Eq. (3.46), the last factor in Eq. (8.29) assumes values in a half-
plane dtz > 0. The same holds for the product of the first three factors,
since the first one is positive, A_(A, £) and A+(A,£) assume values in the
same half-plane but in different quadrants, and 0 < K± < 1. Hence, for all
£ in the strip 3?£ € [A_,A + ], we have — IT < arg.B(C,£) < n, and therefore,
Eq. (8.31) holds. •
1
In fact, at the expense of more technical argument, (i)-(ii) can be established for all £
in the strip £$£ 6 [A_, A+], and then /? = —A_ can be allowed as well.
Under condition Eq. (8.31), 6(A,£) :— lni?(A,£) is well-defined on E ^ c x

{£ | 9£ S [A_, A + ]}. Next, for real £, a > A_ and C\ € (A_,<r), we set
b+(\,S + ia) = -(27TZ)"1 / + 0 ° + I < r i , y , , ? )

di;, (8.32)
and for real £, a < A + and <72 G (a, A+), we set

+oo+i<r2 k(\ „\
dr
° . >- (8-33)
/ .oo+iat Z + i<r-r)
By the Cauchy theorem, 6±(A, £ + ia) are independent of choices of a\ and
02-
It follows from Eq. (3.45), Eq. (3.46), Eq. (8.30) and Eq. (8.31), that
there exist C,p> 0 such that for all 77 in the strip §77 € [A_, A + ],
|6(A,77)|< (7(1 + 1771)-",

where C depends on A but not on 77 (and p > 0 is independent of both
A and 77). Hence, the integrals in Eq. (8.32) and Eq. (8.33) converge, and
6±(A,£) are well-defined and holomorphic in a half-plane ± 3 £ > ±A T .
The next lemma can be proved as the similar statement in Chapter 7.
L e m m a 8.3 For any [LJ'_,<JJ'+] C (A_,A+), there exists C\ > 0 such that
\b+{\,£)\<Cu VAeE9,c, 9£>wl, (8.34)
and
w
|6_(A,0l<Ci, VAe£e,c ^ + - (8.35)
By the residue theorem, for A_ < a\ < a < ui < A+,
&(A, 77)
- /
a +oo+io-i /•+oo+ia'2\
£ + ia - 77
= b(\,£ + ia).
Hence, B± = expb± satisfy B = B+B- on E ^ c x {^ | S^ £ (w_,w + )},
and if we set
o_(A,0 = A_(A,0 K --B_(A,0,

a+(A,0 = dA+(A,0"+B+(A,0,
196 Barrier options
then for A G ~Ee,c, 9£ € (u>_,u>+),
A + a ( 0 = a+(A,Oa-(A,0- (8-36)
Lemma 8.4
aj For any A e Se,c, a+(A, £) is holomorphic in the half-plane $>£ > A_,
and satisfies an estimate
c(l + \M1/U + | ^ i r < | a + ( A , 0 | < C(l + lAp/" + |£|)«+, (8.37)
where C, c > 0 are independent of X £ Ee,c and £ * n *ê half-plane

9 ^ > a>_ (Tmt may depend on u>_ > A_^;
6) For any A G Ee.c, fl-(A,£) is holomorphic in the half-plane 3£ < A+,
and satisfies an estimate
c(l + IAI1/" + |e|) K - < |a_(A,0| < C{\ + IAI1/" + |£|)«-, (8.38)
where C,c > 0 are independent of A e £fl,c a n ^ £ ^ ^ e half-plane
S£ < w+ (7m£ may depend on w+ < A_J;
c) /or a// A € Se,c flnd £ in the strip A_ < S£ < A + , Eq. (8.36) holds;
d) factors in Eq. (8.36) are uniquely defined by properties a) and b), up to
scalar multiples, depending on A.
Proof. Clearly, A±(A, £) K± satisfy a) and b), and since b± are holomor-
phic and bounded on the same sets due to Eq. (8.34)-Eq. (8.35), a) and b)
are proved; Eq. (8.36) has already been proved.
To prove d), fix A, and suppose, A + a(£) = a+(A,£) a -(A,£) is an-
other factorization with the same properties. Then a'+(A, £ ) / a + (A, 0 (resp.,
a'_(A,£)/a_(A,£)) is holomorphic in the upper half-plane 3£ > A_ (resp.,
the lower half-plane S£ < A + ). Both functions are bounded and non-zero,
and coincide on R. Hence, the analytic continuation of any of them is a
bounded holomorphic function on C. By the Liouville theorem, it must be
constant. •
8.2.4 Proof of Eq. (8.17) and continuity of the RHS in

Eq. (8.19)
The resolvent is defined by
(A, + A)- 1 ^ = a_(A, D + i 7 ) - 1 e + a + ( A , D + i-y)-Hg, (8.39)

A symptotics of the option price near the barrier 197
where Ig is an arbitrary extension of g. Represent the operator on the LHS

of the estimate Eq. (8.17) in the form
(1 +mAy + X)-1 g = A'yi_A'^+g,

where
A; ) _ = (l + | A | r - a _ ( A , D + i 7 r 1 ,
and
A ; i + = M l + |A|) K+ a + (A, D + n)-H.
Prom Eq. (8.37) and Eq. (8.38), we conclude that both operators are bounded
in £2(R+) (see Theorem 15.12 and Theorem 15.13), uniformly in A, and
hence, Eq. (8.17) has been proved.
To prove the continuity, we notice that gy G/f(R+), for any s < 0.5 since
g is the restriction to [0, +00) of a smooth function, which exponentially
decays at the infinity together with its derivative. Take 5 < 0.5 such that
s + K- > 0.5. By using Eq. (8.37) and Eq. (8.38) once again, we can show
Q 3 Q S-TK —
that the resolvent is bounded as an operator from i f ( R + ) to H (R+)i

-1
uniformly in A e S^,c- Hence, (A 7 + A) <?7 £H (R+)> uniformly in
A S Efl.C) and therefore, for each r, W7(-,T) £H (R-+)- I* remains to
notice that if s + K_ > 0.5, then H (R+) C C o ([0, +00)).
The continuity of the RHS in Eq. (8.19) has thus been proved.
8.3 Asymptotics of the option price near the barrier
Consider the price of a barrier option without a rebate near the barrier.
It can be shown that if the time to the expiry is fixed then its behavior is
qualitatively the same as of W{X,T) := 1 — V(x,r), where V(X,T) is the
price of the corresponding first-touch digital. See Fig.7.1-7.3, which clearly
demonstrate how significant the difference from the Gaussian case can be.
8.4 Commentary
Various aspects of pricing of barrier options have been considered in a num-

ber of papers and books, see e.g. Rubinstein and Reiner (1991), Wilmott
198 Barrier options
et al. (1995), Musiela and Rutkowski (1997), Hull (2000) and the bibliog-
raphy there, but to the best of our knowledge only Gaussian processes have
been allowed.
For the double barrier options with both upper and lower barriers (also
in the Gaussian case) see Geman and Yor (1996), Pelsser (2000), and the
bibliography there.
The results of this chapter, in a more general set-up, and by using a bit
different technique, are obtained in Boyarchenko and Levendorskii (2002b).
Chapter 9
Multi-asset contracts
The variety of derivative products on several assets is huge - see the ref-
erences at the end of the chapter, where fairly comprehensive overviews
for the Gaussian case can be found; in many cases, it is possible to obtain
RLPE-analogues of the results of the Gaussian theory.
We construct multi-dimensional analogues of RLPE, and calculate prices
of (power) forwards and several types of European options in the RLPE
market of several assets. We calculate weights of a locally risk-minimizing
portfolio of several underlying assets and a contingent claim of the Euro-
pean type, and the prices of the Asian options of the simplest type.
9.1 Multi-dimensional Regular Levy Processes of Exponen-

tial type
In the Gaussian case, the evolution of the tuple of several assets can be
modelled as
n
1=1
where bj, j = 1 , . . . , n, are constants, E := \<rij\ is a symmetric matrix, and
W(t) = (W\(t),..., Wn(t)) is an R n -valued standard Brownian motion pro-
cess. In other words, each component of a multi-dimensional Brownian mo-
tion is represented as a mixture of independent Brownian one-dimensional
motions. For a generic non-Gaussian Levy process, the similar reduction
to finite mixtures of one-dimensional processes is impossible.
The definition of multi-dimensional KoBoL uses the decomposition into
199
200 Multi-asset contracts
an integral of one-dimensional subordinators (Levy processes on a ray; in

ID-case, two subordinators are mixed: on R + and R _ ) . Multi-dimensional
VGP, NIG, HP, GHP and NTS Levy processes are obtained via subordina-
tion of multi-dimensional Brownian motions, by using the same subordina-
tors as in ID-case.
9.1.1 Multi-dimensional KoBoL family

For the sake of brevity, we construct multi-dimensional KoBoL processes
of order v 6 (0,2), v ^ 1, and leave to the reader the construction of the
multi-dimensional analogue of the KoBoL process of order 1.
Denote by (p, <f>) the polar coordinates in R™, and by (•, •) the standard
scalar product in R n . Let H'(d(f>) and A be a finite measure and positive
continuous function, respectively, on the unit sphere S n _i, and let p, € R n .
Set
Il(dx) = p-"-1 exp(-\(<t>)p)dpn.'(d<f>).
By repeating the proof of Lemma 3.1, which treats the ID-case, we conclude
that there exists 7 € R n such that the generating triplet (0,7, H(dx))
defines the characteristic exponent
1K0 = -*<//,£> + r(-i/) f [\(<t>y - (A(0) -i<^))"]n'(d0).
By applying Theorem 2.7, we can obtain a more general class of character-

istic exponents.
Lemma 9.1 Let £ be a positive-definite n x n matrix, let U.'(d(f>) and A

be a finite measure and positive continuous function, respectively, on the
unit sphere 5 n _ 1 , and let p € R™. Then the following formula defines a
characteristic exponent of a Levy process:
VKO = -i<M, 0 + T(~u) f [X^r - (A(0) - i<E£, $ ) " ] I I ' ( # ) . (9.1)
Notice that if p = 0, then as maxA(^) —> 0, the characteristic exponent

Eq. (9.1) converges to the characteristic exponent of a stable Levy process.
Definition 9.1 Let v € (0,2), v ^ 1. The Levy process with the charac-
teristic exponent Eq. (9.1) is called a KoBoL process of order v.
Multi-dimensional Regular Levy Processes of Exponential type 201
The class of characteristic exponents Eq. (9.1) is too large for parameters
fitting purposes since the function A and the measure II' belong to infinite-
dimensional spaces. The following subclass of the KoBoL family depends
on the finite number of parameters, and still seems to be sufficiently rich
for practical applications. Another possibility is to use a measure U'(d(f>)
supported on a finite subset of the unit sphere.
E x a m p l e 9.1 Let A > 0 and /? € R n satisfy \\(3\\ < A; then A(<£) :=

A + (/3, <f>) is positive on S„_i. Let c > 0, /i e R n , and W(d(j)) = cd<j>, where
d<f> is the standard measure on S n _i- Then the conditions of Definition 9.1
are satisfied, and Eq. (9.1) reduces to
0(0 = -i(M,0 (9.2)
+cT(-v) [ [(A + </?, 4>))v - (A + </3, <j>) - i<££, 4>)Y}dct>.
We call a Levy process with the characteristic exponent Eq. (9.2) a KoBoL
process of order v, with the intensity parameter c, steepness parameter A,
asymmetry parameter /?, pseudo-covariance matrix S, and pseudo-drift /i.
Direct calculations show that
*fr(0) = Hj+cTO--") [ (A-M/?,*))"-1^)^, (9.3)
dhp_
(0) = cT(2 -u)f (A + (/?, «A)) ,/ - 2 (E«A) j (E^) fc ^. (9.4)
If a KoBoL with the characteristic exponent Eq. (9.1) fits well to empirical
data, but one tries to fit the Gaussian process to the same data, the drift
and covariance matrix of the Gaussian process will be given by the RHS in
Eq. (9.3) and Eq. (9.4), respectively. This remark can be used when one
compares results in KoBoL and Gaussian cases.

pered Stable Levy processes
Let v & (0,2), 5 > 0,7 > 0, and let TS(v,5,j) be infinitely divisible
distribution introduced in Subsection 3.1.3; its Laplace exponent is given
by Eq. (3.23):
$(0) = 6[tf + 2O)''/2-'f]. (9.5)

Let Y(/3°) be the n-dimensional Brownian motion with the drift /3° 6 R™
and (non-degenerate) variance-covariance matrix A, and let X be the sub-
ordination of Y(/3°) by TS(u, 6,7). Then the characteristic exponent of X,
denote it (j>, is of the form
and by using Eq. (9.5) and letting
/3 = A-1lP, a 2 = 7 2 + (A/3,/3),
we obtain
<f>(0 = 6{[a2 - (A(0 + ifl,/? + i^]v/2 - [a2 - (A/3,/?>]"/2}.
Here (•,•) denotes the standard scalar product both in R n and in C n re-
garded as R 2 n with the basis {ei, • • •, en; ie\, • • •, ien}. With this identifica-
tion, A acts in R 2 n as a block-diagonal matrix, each of the diagonal blocks
being equal to the n x n matrix A.
By adding the drift, we obtain the characteristic exponent of a Normal
Tempered Stable (NTS) Levy process
+6{[a2 - (A{(3 - H O , / * + tf>]v/2 - [« 2 - W,P)Y/2}- (9-6)

With v = 1, this is a characteristic exponent of a NIG process.
9.1.3 Hyperbolic Processes, Generalized Hyperbolic Processes

and Variance Gamma Processes
These processes can be constructed as in ID-case in Subsection 3.1.4 and
Subsection 3.1.5, with the n-dimensional Brownian motion in place of the
one-dimensional one. For instance, a multi-dimensional VGP is defined by
9.1.4 Definition of multi-dimensional RLPE

Consider the characteristic exponent of a K0B0L process. Since A is con-
tinuous and positive, A := minA(^) > 0, and therefore, ij) satisfies the
European-style contracts 203
following conditions:
(a) there exists a bounded open set U £ R n containing the origin such that
ip admits the analytic continuation into the tube domain Hn+iU C C n ,
and the continuous extension up to the boundary of the tube domain;
(b)
tf(0 = -<</*. 0+lM0> (9-8)

where ipi has the following asymptotics, as £ = T] + icr —•> oo in the tube
domain R n + ill:
M^) = ôo(v/\v\)\v\v + 0(\vre), (9-9)

where e = 1, and if>oo is a positive positively homogeneous degree zero
function:
ipoc(sri) = lM»j) > 0, Vs>0,T7GRn\0. (9.10)
Definition 9.2 We say that a Levy process X is a Regular Levy Pro-

cess of Exponential type (RLPE), of order v £ (0,2], if its characteristic
exponent satisfies conditions a) and b) with some e > 0.
KoBoL processes provide examples of RLPE of order v € (0,2), HP and

NIG are RLPE's of order 1, and Gaussian processes and its independent
mixtures with other RLPE's are RLPE's of order 2. To incorporate VGP
into the class of RLPE-processes, the condition Eq. (9.9) must be replaced
with the following one:
V-i(0 = 1>oo(v/\v\) In \v\ + 0(|77|- £ ), (9.11)

where e > 0, and Voo is a positive positively homogeneous degree zero
function. If Eq. (9.11) holds, we call X an RLPE of order 0.
9.2 European-style contracts
We consider forward contracts, European puts and calls, and exchange and
basket options. The market consists of n > 1 stocks with the price processes
Sj(t) = expX,(£), j = 1 , . . . ,n, and a riskless bond yielding the constant
rate of return r > 0. We assume that X = (Xi,..., Xn) is an RLPE under
the historic measure P .
9.2.1 The Esscher transform

In this subsection, we show that all the classes of processes described in
Section 9.1 are closed under the Esscher transform, provided certain natural
mild restrictions are met. We look for an EMM satisfying
dQ = e{K'x^-ctdP,
where K € R n and c € R are to be found. Let F(S(t),t) - f(Xt,t) be

the underlying security in the market or contingent claims priced in the
market; then its discounted price under the EMM Q is
EQ'x[e-rtf(Xt, t)] = EV'x[e(K'Xl)'(c+r)tf{Xu *)]. (9-12)
The EMM-condition means, in effect, that for the price of the riskless bond
and the price for any of the underlying stocks, we must have
E^x[e-rtf(Xut)]=f(x,0), t>0. (9.13)
By applying Eq. (9.12) and Eq. (9.13) to the riskless bond, whose price is
deterministic: f(Xt,t) = ert, and using the definition of the characteristic
exponent, we obtain the following equation
P
e-t(V (-i*)-c) _ 1
whence
c = iPP{-in); (9.14)
and by applying Eq. (9.12) and Eq. (9.13) to any of the underlying stocks,
that is, with f(x, t) = eXj, j = 1 , . . . , n, and taking Eq. (9.14) into account,
we derive
P
exj-t(V- ( - i e i - i K ) - i / ' P ( - i K ) + 7 - ) _ gXj j = I ... n.
Here {ej}" = 1 is the standard basis in R n . Thus, we have the following

system of equations to determine K:
ip (—iej—iK) — ip(—iK)+r = 0, j = l,...,n, (9.15)
and if a solution exists, we can define the Esscher transform of the historic
measure P by
From Eq. (9.16), we derive the following formula for the characteristic ex-
ponent rfrQ of the process X under Q:
V>Q(£)=VP(£-«)-V>PHK)- (9-17)
We will use Eq. (9.17) in numerical examples.

In the Brownian motion case, V P is a second-order polynomial, hence
Eq. (9.15) is a linear system of n equations with n unknowns, therefore if
the Brownian motion is non-degenerate, the system has the only solution
K. Notice that for typical parameters' values, || K || is of order 1. In general,
the system Eq. (9.15) may have no solutions, one or many, but if X is a
KoBoL with the large a := A or NIG, NTS Levy, HP, GHP and VGP with
large a (large w.r.t. the other parameters of the process, which is usually
the case as the empirical studies of the financial markets reveal), then the
characteristic exponent ip^ admits a representation
V>P(0 = -i(b,0 + i(Hess £,£) + fl(a,0, (9-18)
where b = dtpp(0), Hess is the Hessian of ijjp, evaluated at 0, and R(a,£)

is small (together with its derivatives) in a region where || £ ||< 2, provided
a is large enough. If the inverse of Hess is bounded uniformly w.r.t. the
large parameter a, then by inserting Eq. (9.18) into Eq. (9.15) and using
the standard perturbation argument, it can easily be shown that Eq. (9.15)
has the unique solution in the ball || K ||< 2. This solution can be found
both in the form of a series in negative powers of the large parameter a,
and numerically, by using any standard iteration method with «o = 0 as
the initial approximation.
9.2.2 Pricing of power forwards

We assume that the parameters of the historic measure P and the EMM
Q chosen by the market have been fitted to the data, and therefore, the
analytical formulas for the characteristic exponents i/' P (0 a n d V ^ ( 0 a r e
known. Let F(S(t),t) = f(X(t),t) be the no-arbitrage price of the contract
with the expiry date T and the terminal payoff G(S(T)) = g(X(T)). Then
for r := T - t > 0,
f{x, t) = e x p [ - r ( r + ^(Dx))]g(x). (9.19)

For a = {au...,an) G -U, set S{T)a := Sy{T)a^ • • • Sn{T)a-, and con-

sider the power forward with the delivery price K, and the terminal payoff
G(S{T)) = S(T)a — K. Due to the condition —a € U, this forward contract
is priced in the market, and its price F(S, t) = f(x, t) at time t is
f{x, t) = e x p [ - r ( r + ^{Dx))]{e(a'x) - K).

From the equality
V>Q(£)e<Q'X> = ^ Q H " ) e ( a , I > , (9.20)

we obtain
F(S(t), t) = e(t-T)W>(-ia))sâ _ e(t-T)rR (g 2 1 )
Notice that at the initiation, the price of the forward contract must be 0,
and therefore, the delivery price is defined at time 0 by
9.2.3 Pricing of European options on one asset

Let Fca\i;n(S(t),t) be a European call on the last stock, Sn, with the expiry
date T and the strike price K. Set f{x,t) = Fca\\]n(eXl,... ,eXn;t), and
Xn
g{x) := gn(xn) = (e ~ K)+- Further, divide the dual variables into two
groups £ = (£',£n), where £' = ( £ i , . . . , £ n - i ) , and notice that the Fourier
transform of the identity is the Dirac delta function; hence, for r = T—t > 0
and any a < — 1 such that aen G U,
f{x,t) = (27r)- n / / ei(x,ti)-T(r+^(t))g^d£
= (2TT)-" / / e i < x -«>- T ('- + * Q «))(J £ , = o5„K„)d?

1
^R"- J-oo+ia
/• /• + 00 + ICT
= (2TT)-" / / ei<I'«>-T<r+*9<°'«"»<V=o5n(£n)de
+oo+icr
+oo-
ei*"t"-TW*«>*"»gn{£n)dSni
/ -oo+
or
r+oo+ia
-t-oo-t-ia
ei£nlnS.-r(r+^(0,Jn))5n($n)dSni (Q 2 2 )
/ -oo+to-
In other words, the pricing formula is the same as in the case of one stock
S(t) = expX(t), where X is an ID Levy process with the characteristic
exponent ip^(0, •). We leave to the reader the verification of the fact that
1/^(0, •) is the characteristic exponent of a Levy process in ID.
The other types of European options on one asset or on weighted geo-
metric averages of several assets are priced in the same fashion.
9.2.4 Pricing of exchange options and basket options

For simplicity, we consider the market of two stocks. Let F(S(t),t) be a
European option which gives its owner the right to exchange the first asset
S\ for the second asset S2 a t the expiry date T, for the price %K. The payoff
function is max{G(5(T)),0}, where G(ex) = g(x) = eX2 - eXl - K, and
its Fourier transform cannot be calculated in a simple closed from. This
undermines the effectiveness of the direct application of the pricing formula
Eq. (9.19). Instead, we suggest the following approximate procedure, which
works reasonably well if the time to expiry is not too large, and the steepness
parameter (A in the K0B0L model, and a in HP, NIG, and NTS Levy
model) is large. The same procedure applies to a basket option with the
payoff (e*2 + eXl — K)+, and to other types of exchange and basket options.
Set Z0 = {x \ g{x) = 0}, and Z± = {x | ± g(x) > 0}, and let X(0) be
the vector of log-spot prices. Consider the following three cases:
a) X(0) £ Z- is far from ZQ;

b) X(Q) € Z+ is far from Z0;
c) X(0) is near ZQ.
Certainly, "far" and "near" do not have the exact meaning; presumably, "at
the distance more than 0.25" and "at the distance less than 0.25" can be
used, and if the results do not agree with the desired error at the boundary,
where "far" becomes "near", one can change the definitions of "far" and
"near".
In the case a), we set f(X(0),t) — 0.
In the case b), we expand G(S) in the Taylor series around S(0):
G(S)~£ C a (S(0))(S-S(0)) Q ,
a
pick up terms up to a fixed order m (we believe that m = 4 suffices)
G(S)~ £ ca(S(0))(S-S(0)r,
\a\<m
and after multiplying out, obtain an approximation for g(x) = G(S) of the
form
g(x)~ £ da(S(0))e<a'x\ (9.23)
Now, by substituting the approximation Eq. (9.23) into Eq. (9.19), and
using Eq. (9.20), we obtain
F(5(0),0)~ ^T da(S(0))e-r(r+*Q(-te»S(0)a. (9.24)

|a|<m
Notice that the construction in the case b) presupposes that ZQ, where
max{<7(x), 0} loses smoothness, is sufficiently far so that the influence of Z0
can be neglected.
In the case c), we cannot neglect it, and the procedure becomes more
complicated. Let Xoo £ ZQ be the closest point to X(0). In a neighbour-
hood U of Xoo, which contains X(0) but is not too large (probably, for
typical parameters' values, the neighbourhood of radius 0.4-0.5 will do),
we introduce the new coordinate system as follows. Let x — x(x'l) be the
parameterisation of ZQ in the neighbourhood of Xoo, s.t. Xoo = ^(0), and
set x'2 = \n(g(x)+Ki), where K\ > 0 is sufficiently large so that g(x) + Ki
is positive and bounded away from zero in the chosen neighbourhood U.
Then in U, g{x) — ex* — K\, and the payoff becomes the payoff of a Eu-
ropean call: max{ex2 — Ki,0}. By making the change of variables in the
Cauchy problem for the generalized Black-Scholes equation
(dt-(r + ^(Dx)))f(x,t) = 0, t<T,

f(x,T) = max{ 5 (:r),0},
we obtain
(dt-(r + ^(x',Dxl)))f(x',t) = 0 , t<T, (9.25)

f(x',T) = max{ex* - Ki,0}, (9.26)
where ^(x', Dxi) is the PDO ^ ( A r ) written in the new coordinate sys-
tem. The calculus of PDO allows one to write down the asymptotic expan-
Locally risk-minimizing hedging with a portfolio of several assets 209
sion of the symbol, and after that, obtain the approximate solution to the
problem Eq. (9.25)-Eq. (9.26). For details, see Section 16.5.
9.3 Locally risk—minimizing hedging with a portfolio of sev-

eral assets
9.3.1 The set-up

Assume that the investor can go short in the money account, and consider
the following problem:
given a basket of (derivative) securities F, which are priced

in the market, and the investor's wealth at time t, W(t),
construct a portfolio
(F(S(t),t); e1(t)S1(t),...,0n(t)Sn(t))
with the minimal variance the next infinitesimally close mo-

ment t + At, assuming that the residual
n
w0(t) := W(t) -F(S(t),t) -J20j(t)Sj(t) (9-27)
3= 1
is invested into the money account.
The problem and construction of the hedging portfolio can easily be mod-
ified for the case when F is a basket containing not only derivatives but
some of the basic securities Sjt, I = 1 , . . . , m; in this case, the corresponding
dj, (t) are assumed to be zero. The modification for the case of the restricted
borrowing is also possible.
The locally risk-minimizing formulas for the weights of the hedging port-
folio, constructed in the paper, admit a natural modification for the case
of Levy-like Feller processes constructed by Barndorff-Nielsen and Leven-
dorskii (see Chapter 14). This modification can be applied to portfolios
containing bonds and stocks.
As in the one-asset-case in Chapter 4, we show that the weights of the
hedging portfolio for a European option are continuous up to the expiry
date even at the strike price, where they are discontinuous in the Gaussian
case.
We derive simple formulas for "power forward" contracts, and then we

use these formulas to obtain effective approximate hedging formulas for
other claims, e.g., exchange options and basket options.
Numerical examples for locally risk-minimizing hedging under KoBoL
processes show that even not near the expiry, the weights of the hedging
portfolio in our model can significantly differ from the ones in the Gaussian
model with the same covariance matrix. The effect of fat tails is especially
strong in the case of highly correlated stocks.
9.3.2 Weights of the hedging portfolio

Let E^ be the expectation operator under Q conditioned on the informa-
tion available at time t. In order for the formulas which we derive below to
make sense, it is necessary for the following two conditions to hold
(i) F and Sj, j = 1 , . . . , n , are priced under Q, that is,
E?[F(S(t + At),t + At)} < oo,
and
E?[Sj(t +At)} <oo, V j = l , . . . , n ;
(ii) F, Sj, FSj, and SjSk, j,k = l,...,n, are priced under P .
Let X be an RLPE both under the historic measure P and the EMM Q
chosen by the market, and let Up and U^ be the corresponding open sets
in Definition 9.2. Let {ej}J=1 be the standard basis in R n . We will use the
identity
^ Q ( D ) e ^ = ex^Q(D - iej), (9.28)
which is valid if — ej is in U^. It can easily be shown by using Eq. (9.20)

and Eq. (9.28), that if F is the price of a European option with the terminal
payoff G(S(T)) = g(X(T)) admitting the estimate
G(ST) < Cmax{l, S i ( T ) , . . . , 5 n ( T ) } , (9.29)
and
-ej £ UQ, -ej, -ej -ekeUp (9.30)

for all j , k = 1,..., n, then conditions (i) and (ii) are satisfied. Notice that
Eq. (9.29) holds for European calls and puts.
Fix t, and denote Xj = \nSj(t), f(x,t) = F(eXl,... ,ex";t).
The price of the portfolio at time t + At is
n
W(t + At) = F{S{t + A),t + At) + ^BjflSjit + At) + erAtw0(t),
3= 1
where wo{t) is the residual in Eq. (9.27), and we assume that the investor
chooses 6(t) to minimize the variance of W (t + At) under the historic mea-
sure P, conditioned on the information available at the date t. We calculate
Ef[(W(t + At) - Ef[W(t + At)})2)
= Ef[(F(S(t + At),t + At) - Ef[F(S(t + At),t + At)})2]
n
+2 J2 Oj(t)E?[(F(S(t + At),t + At) - E?[F(S(t + At),t + At)})
3=1
x (£,-(* + At) - Ef[Sj(t + At)})}
+ £ OAWkmfKSjit + AV-EflSjit + At)})

j,fc=i
x(Sk(t + At)-Ef[Sk(t + At)})}.
Minimizing w.r.t. 6(t), we obtain the following first order conditions: for
j = l,...,n,
71
J2 OkWEfHSjit + At) - Ef[Si(t + At)})(Sk(t + At) - E?[Sk(t + At)})}

fc=i
= -Ef[(F(S(t + At),t + At) - Ef[F{S{t + At),t + At)])(S,-(t + At)
-EflSjit + At)})}
or, equivalently,
(9.31)
n
£ 0k(t) (Ef{Sj(t + At)Sk(t + At)} - Ef[(Sk(t + At)}Ef[Sk(t + At)})
k=i
= -E?[{F(S(t + At),t + to)Sj(t + At)} + Ef[F(S(t + At),t + At)}
xE?[S3{t + At)}.
The explicit formula for the terms on the LHS is immediate from the defi-
nition of the characteristic exponent (and from Eq. (9.20)):
Ef[Sj(t + At)Sk(t + At)} = e x p ( - A # p ( - i ( e j + e fc )))Sj(t)S fc (t), (9-32)

and
EP[Sj(t+At)}Ep[Sk(t+At)} = exp(-At^p(-iej)+iJp(-iek)))Sj(t)Sk(t).
(9.33)
By applying the Taylor formula to Eq. (9.32) and Eq. (9.33) around At = 0,
we see that the LHS admits the representation
n
^ ^ ( t ) 5 , ( i ) 5 f c ( t ) [ - V P H f e +ek)) + 1>p(-iej) + ^p(-iek)} At+o(At).
fc=i
(9.34)
Notice an evident formula for the infinitesimal generator L of the (n + 1)-
dimensional process {Xt} — {{Xt, t)}, when X is regarded as a Levy process
under P :
L = dt + Lp = dt-rpp(Dx). (9.35)
By using Eq. (9.35), we obtain for the terms in the RHS of Eq. (9.31)
Ep[F(S(t + At),t + At)Sj(t + At)} - F(S(t),t)Sj(t) (9.36)
= (dt - i>p(Dx))f(x, t)ex*At + o(At),
and
Ep[F(S(t + At),t + At)} - F(S(t),t) (9.37)
= (dt-^p{Dx))f(x,t)At + o(At).
From Eq. (9.28), we can rewrite Eq. (9.36) as
Ep[F(S(t + At),t + At)Sj{t + At)} - F(S(t),t)Sj(t) (9.38)
= £,•(«)($ - 1>P(DX - %ej))f{x, t)At + o(At),

and by gathering Eq. (9.37)-Eq. (9.38) and using Eq. (9.33), we see that
the RHS in Eq. (9.31) admits the representation
Sj(t)[i>p(Dx -iej) - i>p{Dx) -^(-ie^fixÂt + oiAt). (9.39)

By inserting Eq. (9.34) and Eq. (9.39) into Eq. (9.31), dividing by At and
Sj(t) and passing to the limit At —> +0, we obtain, for j = 1 , . . . , n:
n
Jêk(t)Sk(t) [-4>p(-i(ej +ek)) + iPP(~iej) + VPHefc)] (9-40)
fc=i
= -[-rpp(Dx - ie,) + rpp(Dx) + ^p(-iej)]f(x, t).
For a,(3e -Up + i R n , such that a + j3 e -Up + i R n , set
A(a,f3) = -^p{-i{a + p)) + ^p{-ia)+i>p{-ij3),

A(iD,f3) = -Tpp(D-i{3) + i>p(D) + *ljp(-if3).
Further, introduce a matrix
COV:=COV(X):=[A(ei,ek)]lk=l,
a row vector with operator-valued components
A(iD) :=A(X;iD) := [A{iDtej)]j=l ni
and a row vector
w{t) := w{S(t),t) = & (*)£,-(t)W,...,„
of fractions of wealth invested in the underlying stocks. Then we can rewrite

Eq. (9.40) as
COVw(t)T = -A(iD)Tf{x,t). (9.41)
If
COV ^ 0, (9.42)
we can solve Eq. (9.41)

T
w(t) = -COV'1 A(iD)Tf(x, t). (9.43)
Notice that if X is a Brownian motion, then

The direct calculation shows that COV(X) = [djk] is the covariance matrix,
and
n
A(iD,ej) = y^gjfcdj.
k=i
Thus, in the Gaussian case, Eq. (9.41) reduces to the standard delta-hedging
result, which gives the riskless portfolio. In the non-Gaussian case, the
locally risk-minimizing formula Eq. (9.43) does not determine a riskless
portfolio, the latter being non-existent.
If Eq. (9.42) holds, and an explicit formula for F(S, t) = f(x, t) is avail-
able, Eq. (9.43) can be used to calculate w = w(t). Notice that Eq. (9.43)
gives locally risk-minimizing weights for any contingent claim, not neces-
sarily of the European type.
9.3.3 Hedging of European claims

Let F be a European claim with the payoff g(X(T)) at the expiry date T,
and T = T — t is the time to the expiry, then
+oo+ia% f+oo+i<Tn
/ Q
... / e *<*.€>-'-('-+* (fl)$(£)(£1.. . # „ ,
-oo+i<7i J — oo+i<rn
(9.44)
where g(£) is the Fourier transform of g, and the choice of real dj, j =
1,2,..., n, is determined by the type of the contingent claim (for instance,
for the put on the last stock, one takes <rn > 0 such that —anen e C/Q, and
<7i = • • • = <7n_! = 0). Therefore in Eq. (9.43),
... (9.45)
-OO-HiTi J —00+i(Tn
exp [t(x, 0 - r(r + V>Q(0)] Aft, e ^ S t f K i • • • # « •
If the terminal payoff depends on the price of the last stock only: g(Xx) =
g n ( X r ( n ) ) , for instance, if F is the price of a European put on the last
stock, then g(£) = <?„(£„) ® %=o, where £' = ( £ i , . . . ,£n-i), and hence
Eq. (9.45) reduces to a much simpler expression
+ 0 0 + t<7„
(9.46)
/ •oo+icrn
• exp [ixnn - r(r + ipQ(0, T)en))] A(ir)en, ej)gn(jf)dr}.
Theorem 9.1 Let X be an RLPE of order v G (0,2), satisfying
i/-e<l, (9.47)
where e > 0 is from Eq. (9.9), let Eq. (9.30) holds, and let F(S(t),t) be the
price of the European put or call.
Then the weights of the locally risk-minimizing hedging portfolio given
by Eq. (9.43) and Eq. (9.46) are continuous up to the expiry date, for all
strikes.
Proof. For European puts and calls, the Fourier transform of the termi-
nal payoff decays as |J7| -2 , as 7] —> 00, and hence, under condition Eq. (9.47),
the integrand in Eq. (9.46) is bounded (in modulus) by an integrable func-
tion C(l + |T?|) - 2 + l / - e . Since the integrand is a continuous function in
(X,T) 6 R x [0,+00), we conclude that the integral is continuous on the
same set, too. ' •
As the proof shows, the result is valid for straddles, strangles and essentially
any other European contingent claim used in real Financial markets, with
the payoff, which is continuous, piecewise smooth, and satisfies (i) and (ii)
of the previous section (but it fails for digitals, of course).
Since the Gaussian delta-hedge produces discontinuous weights at the
strike and expiry, our result Eq. (9.43)-Eq. (9.46) in the non-Gaussian
RLPE-case differs drastically from the Gaussian hedging, near expiry and
strike. The difference can be substantial in other regions of (x, £)-space as
well, especially in the market, where some of the assets are highly correlated
stocks.
9.3.4 Locally risk-minimizing hedging of "power forwards"

Consider the power forward with the terminal payoff of the form S(T)a :=
S!(T)ai • • • Sn(T)a~ - K and the price
F{S,t) = e-r(r+TfiQ(-ia))soc _ Ke-rrt
where T = T — t. By using an equality
A{iD, eJ)e</3'x> = A(0, e,-)e</,'1>, (9.48)
and taking into account that A(Q,ej) = 0, we derive from Eq. (9.43)
w(tf = -e-T('-+^Q(-ia!))COF-1A(a)T5Q, (9.49)

where A(a) — [A(a, e\)... A(a, e n )].
9.3.5 Approximate locally risk-minimizing hedging:

general case
Suppose that we know a reasonably good approximation for the price
F(S,t) in a neighbourhood of a fixed point (5°, t0), of the form
F(S,t)= £ da(S°,t0;t)(S-S°)a, (9.50)

|a|<m
where \a\ = a\ H + an. (We believe that for practical purposes, m = 4

should suffice.) If the intensity of large shocks is small (in the KoBoL model,
this means that A is large, and in the models with subordinated Gaussian
processes, a is large), then for the locally risk-minimizing hedging purposes,
these are the values in a relatively small neighbourhood of (S(to), to), that
matter most, and hence, the approximation Eq. (9.50) seems reasonable.
Multiply out in Eq. (9.50) and obtain a representation of the form
F(S,t)= Yl Pa(S°,to;t)Sa, (9.51)

|a|<m
and after that, use Eq. (9.48) with f3 € {a 1 ,. • • , a W } := {a I M < ™}-

The result is
w(S(t0),tof = -COV-1 £ Pa(S°,to;to)(S

0 a
) A(a)T.
\a\<m
9.4 Weighted discretely sampled geometric average
An Asian option is the name used to denote options whose terminal payoff
depend on the average asset values during some period of the option life
time. There are Asian options both of the European style and the American
style; the averaging can be in continuous time, e.g., with the payoff of the
form
G(S(T);A(T0,T)) = (A(T0,T) -K)+,
where
A(To,T) = (T-T0)-1 J S(u)du,

Weighted discretely sampled geometric average 217
or in the discrete time, when A(To,T) is replaced with an average price

AN(to, T) on sampling dates 0 < t0 < • • • < t^ = T:
N
AN(t0,T) = {N + l)-1^S(tk).
fc=0
The popular numerical approach to the valuation of Asian options is based

on the approximation of the arithmetic average by the geometric average.
This approximation is known to significantly underprice the option (see
the discussion in Musiela and Rutkowski (1997)) but is computationally
very simple. The averaging with different weights are also used, and the
"basket" variant of the weighted discretely sampled geometric average is
technically no more difficult.
Let 5m £ R " , m = 0 , . . . , N be vector of weights for stocks at the sam-
pling dates, and define the weighted discretely sampled geometric average
as
J{t)= £<<T\X(i m )).

tm<t
Consider an option with the terminal payoff depending on both X(T) and
J(T): g(X(T),J(T)).
In addition to the observable vector of log-spot prices at time £, X(t),
we have the observable J(£), and clearly, the price of the option at time t
depends on the triple (X(t),J(t),t); denote this price f(X(t),J(t);t). At
a sampling date tm,
f(x, J ; t m - 0) = f{x, J + (8m,x);tm), (9.52)
and by using the Taylor formula:
F ( J + x) = ^(sir'd'jF^x3 (9.53)
s>0
= exp[xdj}F(J),
we can rewrite Eq. (9.52) as
f(x, J; tm - 0) = exp [{5m, x)dj] f{x, J; tm). (9.54)

Set £_! = 0. On each interval [tm,tm+i), J(t) is constant, and therefore
the price f(x, J; t) can be calculated as the price of a European option with
the terminal payoff f(X(tm+i), J;tm+i — 0):
f(x, 7; t) = exp[(t - tm+1)(r + i>Q(Dx))}f(x, J; tm+1 - 0). (9.55)
Let t £ [tm,tm+i)- By using Eq. (9.54) and Eq. (9.55), we derive
/ ( * , J; t) = exp[(t - tm+1)(r + V> Q (£ X ))]exp {{6m+\x)dj}

N-l
H exp[(tk - tk+1)(r + ^(Dx))) exp [(6k+1,x)dj] g(x, J).
k~m+l
Denote the Fourier transform of / and g w.r.t. J by f and g, respectively,

and let rj be the variable dual to J. Then
f(x, ry; t) = exp[(t - tm+1)(r + ^(Dx))\ exp[(6m+1, x)irj\

N-l
11 exp[(tk - tk+1)(r + TpQ(Dx))} exp[(6k+1,x)ir,}g(x, r,).
k=m+l
Set 7fc = Yls=k ^ s - By using the commutation relation

exp[6(D x )]exp[z(7,:r)] = exp[i(7,x)]exp[6(D I + 7 ) ] ,
we can move all the exp[(5s, x)ir)]- factors to the left, and obtain
f(x,rj;t) = exp[( 7 m + 1 ,a;)i77]exp[(i-t m + 1 )(r + V Q ( i ) x + 7 m + 1 ^ ) ]

N-l
exp[ J2 {tk-tk+1){r + ^{Dx+1k+171))\g{x,r1)
exp[(t - T)r] exp[< 7 m+1 ,x)irj\ exp[$(D x , 77; t)]g(x, 77),
where
$(Dx,rj;t) = (t-tm+l)^Q(Dx+1m+1r1) (9.56)

N-l
+ J2 (tk-tk+1)^(Dx + ik+1r)).
fc=m-f 1
By using the Fourier transform, one easily computes
<f>{x, J; t) = exp[$ (Dx, Dy, t)]g{x, J ) , (9.57)
and then one finds
f(x, J; t) = e C - ^ O r , J + ( 7 m + 1 , x); t). (9.58)

Weighted discretely sampled geometric average 219
Here we have used the equality
JT-1 exp[( 7 m + 1 , x)irj\? = exp[( 7 m + 1 , x)dj]

and Eq. (9.53).
If g is independent of x, which is the most popular case, the answer
simplifies considerably:
M J; t) = exp[$(0, Dj; t)}g(J), (9.59)

and
f(x,J;t) = e^-T^cj>o(J+ (lm+1,x);t). (9.60)

We have proved the following theorem.
Theorem 9.2 At time t e [tm,tm+i), the price
F(S(t),J(t);t) = f(X(t),J(t);t)
of the Asian option with the payoff function g is given by Eq. (9.56)-
Eq. (9.58). If g is independent of x, then the simpler formulas Eq. (9.59)-
Eq. (9.60) hold.
Chapter 10
Investment under uncertainty and

capital accumulation
Here we consider a risk-neutral, price-taking, and value-maximizing firm

under demand uncertainty. The firm chooses optimal investment strategies;
the investment is irreversible. We suggest a Marshallian-like form for the
investment rule. It is applicable when the price can move in both directions,
and uses the infimum process of the price instead of the price process itself.
We also present an explicit formula for the expected level of the capital
stock in terms of the infimum and supremum processes. Both results are
new even for the Gaussian case. We derive simple approximate formulas,
which facilitate the comparative statics analysis. Other problems of Real
Options theory can be treated in the same fashion.
10.1 Irreversible investment and uncertainty
Here we analyse a benchmark model of irreversible investment under uncer-

tainty. The investor is a competitive, risk-neutral, and value-maximizing
firm. The firm's production function satisfies conventional properties such
as monotonicity, concavity, and the Inada conditions. The latter conditions
at zero ensure that it is always optimal to enter the market therefore we are
interested in the optimal investment strategies of a firm, which has been on
the market for some time. Time is continuous. For simplicity, we assume
that the only source of uncertainty is the unit price of the firm's output,
which follows a stochastic process. Even though there is empirical evidence
(see, for instance, Yang and Brorsen (1992) or Deaton and Laroque (1992))
that distributions of commodity prices exhibit significant skewness and fat
tails, traditionally the price process P = {Pt}t>o is modelled as a Gaussian
221
222 Investment under uncertainty and capital accumulation
process, for example, the geometric Brownian motion. We model the price,
Pt, as Pt = exp(-Xt), where X = {Xt}t>o is an RLPE. We demonstrate that
RLPE are not less tractable in models of investment under uncertainty than
Gaussian processes, the former being more realistic than the latter.
At each moment of time, the firm's manager has to decide whether to
increase or not the stock of capital given the current stock and the market
price. If the firm suffered an adverse demand shock in the past, its capital
stock may be too big for the current price level, but since investment is
irreversible, the firm has to stay put. Should the price increase, it may
become optimal to invest, and in this case, the manager needs to determine
the optimal amount of new capital. We assume that it is optimal to invest
when the price crosses a certain barrier, which is called the investment
threshold. As soon as the investment threshold is determined, the optimal
amount of investment follows automatically, as we will see later.
There are at least two different prescriptions how to choose the trigger
price of investment. One of them is known for almost a century, and is
called the net present value (NPV) rule, or the Marshallian law. This rule
(in marginal terms) says that investment is optimal as long as the present
value of the expected marginal revenue does not exceed the marginal cost of
investment. The second rule was spelled out by Dixit and Pindyck (1996),
and it is based on the option pricing approach to the problems of investment
under uncertainty. Dixit and Pindyck (1996) pointed out that the NPV
rule was incorrect since it did not take into consideration the option-like
nature of investment decisions. To correct the Marshallian prescription,
the marginal cost of investment has to be multiplied by a certain factor.
This correction factor comes from a solution to a so-called fundamental
quadratic equation and has no economic meaning.
We use the Real Options approach as well and solve the problem essen-
tially by the method of Chapter 5, because the choice of the investment
threshold is similar to the choice of the optimal exercise price of a per-
petual American call option. Our approach not only allows deriving an
analytical solution for the case of output price following an RLPE, but also
yields several new insights. The most operational one from the standpoint
of economic applications is a modified version of the Marshallian law.
We consider only such processes that the price can move both up-
wards and downwards with positive probability. If the price process is
non-decreasing almost surely, the irreversibility does not matter, and the
Marshallian law is correct. On the other hand, if the price process is non-
The investment threshold 223
increasing almost surely in t, then the firm will never increase the capital
stock further, because the state of the market can only deteriorate. When
the price can move in both directions with non-zero probability, starting
with the original price process, it is possible to introduce a non-trivial in-
fimum price process. N = {Nt = info<s<t Pt}t>o- The solution to the
problem under consideration allows us to resurrect the Marshallian law to
a certain extent. We state that it is optimal to invest as long as the present
value of expected marginal revenue, computed for the infimum price pro-
cess instead of the original price process, does not exceed the marginal
cost of investment. Notice that this investment rule reflects the "bad news
principle", which was first formulated by Bernanke (1983). The critical
price which triggers new investment depends on downward moves, because
the ability to avoid the consequences of "bad news" leads us to postpone
investment decisions. The main advantage of our investment rule is that
it obviates the need to introduce the correction factor like in Dixit and
Pindyck (1996). In other words, the firm's manager may remain in the
Marshallian world as long as she keeps in mind infimum processes instead
of real ones.
10.2 The investment threshold
Here we specify the objective of the firm and sketch the solution technique.
Details can be found in Boyarchenko (2001).
The firm's objective is to maximize the total expected discounted profit:
r+oo
r-roo
IL(K,x) = max Ex / e-rt(eXtG(Kt)-rKt)dt (10.1)
{Kt} Jo
where the meaning of the stochastic integral above is explained in Bo-
yarchenko (2001). Here we treat the spot log-price, x, and the level of
accumulated capital, K, as state variables, and Kt as a control variable.
Due to irreversibility of investment, in each state (K,x), Kt > K.
In order the stochastic integral in Eq. (10.1) were bounded, we impose
the following conditions:
a) resource constraint: there exists K < oo, s.t. Kt < K, V t ;

b) M = E[g°°e-rt+x'dt < oo.
Following the tradition in the literature, we are going to view the space
of state variables as a (disjoint) union of two regions: inaction and action

ones. For all pairs (K, x) belonging to the inaction region, it is optimal to
keep the capital stock unchanged. In the action region, investment becomes
optimal. Let the inaction region be closed. Denote by T the boundary of
the inaction region. Then the investment strategy defined by the choice of
the inaction region can be viewed as follows:
(i) do not invest as long as Xt < h(K), where h(K) is defined by

{K,h{K))eY-
(ii) invest when Xt > h(K), and increase the capital stock up to the level
4>(Xt) denned by (<f>{Xt),x) € T.
It follows from (i) and (ii), that h is non-decreasing and continuous
a. e. with each point of discontinuity is of the first kind, and h(K + 0) =
h(K), V K. Denote the set of such h by % and notice that every h £ H
uniquely defines the boundary of the inaction region. To argue this, we
notice that h 6 H defines the boundary of the inaction region by (<f>(x), x) €
T & 3 K : x e [h(K - 0),h(K)]; and T together with the rules (i) and
(ii) uniquely define the strategy. Thus, we can restate the original firm's
problem in an equivalent way: find the investment threshold x = h{K), h e
H, which separates the inaction and action regions. Let W(K, x; a) be the
NPV of the firm when a € H is chosen as the investment barrier, that is
W(K, a;h) = Ex\f °° e-rt(eXtG(Kt) - rKt)dt (10.2)
In Boyarchenko (2001), it is shown that the stochastic integral in Eq. (10.2)

is well-defined given conditions a) and b) and the properties of the produc-
tion function. The (optimal) investment threshold, h €H, satisfies
W(K, x; h) > W(K, x; a), V a e H, (10.3)

or equivalently, W(K,x; h) = U(K,x).
The conventional solution technique is to write the Bellman equation for
Eq. (10.3) in the inaction region; and apply the Ito lemma to obtain a second
order differential equation w.r.t. the value function V(K, P) := Tl(K, x) =
W(K, x; h). To specify the solution to this equation and to find the invest-
ment threshold P = H{K) = eh(K\ one uses boundedness of V(K,P) as
P -* 0; the value matching condition VK(K,H(K)) = 1, and the smooth-
pasting condition VKP{K,H(K)) = 0. Evidently, this technique is not
suitable for non-Gaussian modelling because the equations, which arise in
Capital accumulation under RLPE 225
such models, cannot be differential any more; they are pseudo-differential

equations. Besides, as we have seen it in Chapter 5, the smooth-pasting
condition does not always hold for non-Gaussian processes.
We solve the problem as follows. First, we show that the value func-
tion W(K, x; h) is differentiable w.r.t. its first argument and formulate the
sufficient conditions for the investment threshold in terms of this deriva-
tive. Next, we fix a prospective candidate for the investment threshold and
show that the derivative satisfies the Wiener-Hopf equation. By using the
Wiener-Hopf method, we derive an explicit formula for WK{K, X; h). From
this formula, it is possible to conjecture the investment threshold. Finally,
we check that the sufficient conditions hold for the obtained threshold. As
a result, we arrive at the following equation for the current level of price,
H(K), which triggers the new investment for the firm:
/•OO
E / e-rtNtG'(K)dt | No = H(K) C'(K).

Jo
This is the Marshallian law but with the infimum process started at the
current level of the price instead of the original price process. We would like
to stress that our argument provides the revised Marshallian law for more
general form of dependence on the stochastic factor of both the revenue
and cost (cf. Chapter 5 for similar results for perpetual American options).
10.3 Capital accumulation under R L P E
The next important problem, which can be solved in the framework of our
model is the accumulation of capital in the long run. As it was stressed by
Hubbard (1994), the benchmark models of investment under uncertainty
considered in Dixit and Pindyck (1996), do not suggest specific predictions
about the level of investment. Since the investment rule itself is not observ-
able, one has to use the data on investment and capital stock to evaluate
investment models; this shows the importance of theoretical results on the
long-run capital accumulation. Abel and Eberly (1999) examine the be-
haviour of the capital stock in the long run and calculate explicitly the
impacts of irreversibility and uncertainty on the expected long-run capital
stock. They assume that an exogenous demand shock follows a geometric
Brownian motion. Two types of effect are revealed: the user cost effect,
which tends to reduce the capital stock, and the hangover effect, which
arises because the irreversibility prevents a firm from selling capital when
its marginal revenue product is low. Neither of these effects dominates
globally, so the effect of increased uncertainty cannot be determined unam-
biguously.
We address a similar problem in the case when the output price follows
an RLPE. To this end, we specify the production function as a Cobb-
Douglas one, i.e., G{K) — dKe, 6 6 (0,1); the marginal cost is normalized
to unity. We consider a new born firm with the price process P starting at
PQ = 1. As before, our reasoning is valid when prices move both upwards
and downwards. We introduce a (non-trivial) supremum price process as
M = {MT = sup 0 < s < t P t }t>o and derive the following formula for the
expected value -£[1^] of the capital, Kt, at time t:
E[Kt] = {edW-fM-VW+it), (10.4)
where W_ and W+(£) are determined by the infimum process and supre-
mum one, respectively:
/•+oo
W_ = E / e-rtNtdt | No = 1 ,
Jo
W+(t) = E{MlK1~6) | Mo = 1].
The suggested form shows that in making investment decisions under

uncertainty, the anticipation of future negative shocks mostly matters. In
addition, it allows us to separate two effects of the increasing uncertainty on
the capital accumulation, working in the opposite directions: the increase
of the downward uncertainty, measured by the infimum process, increases
the investment threshold thereby decreasing the optimal level of the capital
stock, but the increase of the upward uncertainty, measured by an increas-
ing function of the supremum process, increases the expected capital stock
each instant the investment threshold is crossed. For a similar factorization
formula, in different terms, in a model of a firm which chooses both capital
and labour under the Gaussian process, see Abel and Eberly (1999). As
in Abel and Eberly (1999), we obtain explicit formulas for W_ and W+(t);
the formula for W+(t) is fairly complicated but we show that for large t a
good approximate formula is valid
W+{t) = HE{oo)eat + 0(eet),
where HE(oo) is given by a much simpler expression than W+(t) itself,

Computational results 111
e > 0 is arbitrary, and a = i _ 1 In E[Pt'( ~ ' | Po = 1]; a is assumed positive.

If the last condition fails, then in the model of the completely reversible
investment the expected level of the capital decreases with time. Following
Abel and Eberly (1999), we compare Eq. (10.4) with the corresponding
formula for the capital accumulation in the case of completely reversible
investment
E[KtR] = {edW^/^-^W^t),
where
r r+°o
/•-t-OO
W* = E / e-rtPtdt | Po = 1 ,
Jo
W*(t) = E[Pt1/{1~9) | Po = 1] = ea\
by considering the ratio of the expected levels of accumulated capital
K(t) = § ^ = UCHE(t),
where
uc = (W-/w*)1/(1-v
is the user-cost effect, and
mt) = E[M!Z~lIM° = 1]
- the hangover effect of the irreversibility, and study how both these effects
and the ratio n(t) itself depend on parameters of the process. Clearly,
UC < 1 (unless the process is non-decreasing), and HE(t) > 1 (unless
the process is non-increasing), so that both effects work in the opposite
directions, and the joint effect on the capital accumulation is ambiguous:
n(t) can be larger or smaller than 1, depending on parameters of the process.
10.4 Computational results
Dixit and Pindyck (1996) show that the Marshallian law prescribes very
low trigger price of investment. A natural question which arises in our
model is how the investment threshold changes if one considers a regular
Levy process instead of a Gaussian process with the same first and second
moments î,At and H2,At of observed probability density pAt(x) of the

process X, for a chosen small time interval At; for small At, Hi,At/At
and H2,At/At are good proxies for the first two coefficients in the Taylor
expansion of the characteristic exponent i>(k) at k = 0; they can be used to
infer the parameters of a RLPE process (in the Gaussian modelling, they
are the drift and variance).
To answer this question, we calculated the investment threshold for
different regular Levy processes of exponential type with the same "drift"
and "variance"
lim Hi At/At, and lim \ii At/At, (10.5)
and we found that usually the threshold decreases (insignificantly) as we

replace the Gaussian process with a regular Levy processes of exponential
type with the same "drift" and "variance"; only for "very non-Gaussian"
processes of order close to zero we found sets of parameters for which the
threshold increased. This result can be explained as follows. When one fits
the Gaussian curve to a real probability density, one is bound to disregard
the extreme events, and so, in fact, the "drift" and "variance" for a Gaus-
sian process inferred from the real data do not coincide with the ones in
Eq. (10.5) which one obtains by using a larger portion of rare events than
in the Gaussian modelling: Eq. (10.5) based on the larger data set, gives
larger value of the variance and hence it is incorrect to compare Gaussian
processes and non-Gaussian ones by using the first two moments - or even
the first 3 or 4 moments, since all of them are suitable tools of describing
the behaviour of probability density near zero, being the coefficients of the
Taylor series at zero.
Hence, a much more natural question is: what can be an effect of taking
into account more and more data on rare events. To answer this question,
we consider a series of KoBoL processes with the characteristic exponent
Eq. (3.6); we fix r = 0.06,/x = -0.06, c = c_ = c+ = 0.09, v = 1.6, and
see how the moments limAtjoHj^t/At, j = 2,3,4 ("variance", "skewness"
and "kurtosis") of the process and the investment threshold 1 change with
steepness parameters A+ and A_, which describe the fatness of the left and
the right tail, respectively.
The choice v = 1.6 means that near zero, the probability density does
x
The latter depends on the production function, of course; we plot the factor 4>7{—*),
which characterizes the process in the formula for the investment threshold.
Computational results 229
not differ too much from a Gaussian one, and the decrease of A + (respec-
tively, —A_) can be interpreted as the taking into account the more and
more of negative (respectively, positive) large jumps. We believe that the
result of this computational experiment is instructive. We allow A_ to vary
from —30 to —22, and A+ from —A_ to — A_—8, and we see that though the
variance does not change much, the threshold changes four-fold. The other
observation is that the threshold is much more sensitive to the increase in
skewness than in variance and kurtosis. Consider two cases: 1) A_ = —30
remains fixed and A + decreases from 30 to 22, and 2) both A+ and —A_
decrease from 30 to 22. In the first case, the variance and kurtosis increase
less than in the second case, and the threshold and modulus of skewness in-
creased more in the first case. So even this insignificant increase of variance:
by less than 7 % (and more significant increase of skewness and kurtosis),
which can easily be disregarded in practice since they come mostly from the
tails of the probability density, and hence, can be left over as too extreme
events, can have a dramatic impact on the investment threshold.
The results of the numerical example highlight the importance of ob-
taining the correct estimates for the parameters of an RLPE governing the
underlying stochastic variable. In principle, it is possible to infer these
parameters from the moments (up to the fourth order) of the correspond-
ing probability distribution. However, this will provide a good fit for the
central part of the distribution only, but not for the tails of the distribu-
tion. To obtain parameter estimates fitting the whole distribution, more
sophisticated methods should be used.
Now we consider the errors which may arise when one uses Gaussian
models in non-Gaussian situations, both in non-Gaussian and Gaussian
models. Consider the same firm and process as in the example for the
investment threshold above, with r = 0.06, 8 = 0.33 and d = r/6 in the
Cobb-Douglas function, c = 0.5, v = 1.1. We change A+ in a wider range
from 2 to 30; A_ changes in the range from -30 to -10. So, here we allow for
much fatter left tail. Following Abel and Eberly (1999), we determine the
last parameter, fi, from the requirement that the expected rate of growth
of the price remains constant. For each set of parameters, which define
a non-Gaussian KoBoL process Xt — In Pt, we take the Gaussian process
with the same variance as that of Xt.
The numerical results show (see Boyarchenko (2001) for details) that the
investment threshold in the non-Gaussian case decreases up to 15 percent,
the capital accumulated during 10 years by 13 percent. This means that

though the Gaussian model may be too pessimistic about the unobserved
investment threshold, it certainly can be too optimistic about observed
quantities such as capital accumulated.
The real-world processes may deviate from Gaussian ones even more (for
instance, the left tail can be even fatter, and the figures clearly show that
with the further decrease in A + , the capital accumulation will be depressed
much more). This example illustrates that the capital accumulation is
mainly depressed by the possibility of large downward jumps, which is not
taken into account properly by Gaussian models.
10.5 Approximate formulas and the comparative s t a t i c s
Let —A_ and A_ be large, which means that the rate of the exponential
decay of the tails of the probability density is large. Then for typical param-
eters' values for model classes of RLPE (Normal Inverse Gaussian processes,
Hyperbolic processes and KoBoL family), the equation r + if>(£) = 0 has
exactly two solutions; both of them are purely imaginary: —i[3±, where
/?_ < 0 < /3+. In this respect, model classes of RLPE are similar to
Brownian motion, but their characteristic exponents are not holomorphic
in the whole complex plane. Further, the approximate formulas Eq. (3.82)
are valid, which look like the exact formulas in the Brownian motion case.
It can be shown that for typical parameters' values, the relative error of
Eq. (3.82) is several percent or less. For details, see Chapter 5, where the
next correction term is given. Hence, these formulas can be used for com-
parative statics purposes, and in this approximate form, RLPE-modelling
is not more difficult than in the Brownian motion case - and more realistic
at the same time.
Chapter 11
Endogenous default and pricing of the

corporate debt
We explicitly solve the problem of optimal capital structure endogenous

default in Leland (1994a) setting, for a firm which assets follow a regular
Levy process of exponential type. This setting is highly stylised but nev-
ertheless highlights several important issues. We express the endogenous
default boundary in terms of the inRmum process for the process of the
assets of the firm, and therefore, the result makes sense for any Levy pro-
cess. Formulas for the value of the firm and values of equity and debt are
also suitable for any Levy process. For model classes RLPE, we suggest
simple approximate formulas, which are valid for the firm very close to the
bankruptcy level, and show that in this region, the equity's value can be
much less than the one given by the Gaussian model.
For RLPE with large steepness parameters we suggest simple approxi-
mations to solutions, which are valid not very close to the default boundary.
This allows one to study relatively easily the dependence of endogenous
variables on parameters of the process and other exogenous variables.
11.1 A n overview
Two approaches to credit risk modelling can mainly be divided into two
categories: intensity based models, where the primitive is a stochastic model
for the default event (see the monograph Bielecki and Rutkowski (2001),
and in the non-Gaussian setting, Eberlein et al. (2000a, 2000b), Eberlein
and Ozkan (2001), and the bibliography there), and the structural approach.
In this Chapter, we apply the structural approach to credit risk modelling
for the case of non-Gaussian processes. The central feature of this approach
231
232 Endogenous default and pricing of the corporate debt
is an attempt to model explicitly the evolution of the assets of the firm. The
structural approach was pioneered by Black and Scholes (1973) and Merton
(1974) and extended by Black and Cox (1976), Longstaff and Schwartz
(1995). In a series of papers, Leland (1994a, 1994b) and Leland and Toft
(1996) focus on the choice of optimal capital structure and bankruptcy level
by a firm which keeps a constant profile of debt.
All the aforementioned structural models of credit risk have relied on
diffusion process to model the evolution of the firm's assets. While the
diffusion approach is convenient, it cannot capture the basic features of
credit risk observed in practice. In particular, Jones, Mason and Rosenfeld
(1984) find that the credit spreads on corporate bonds are too high to be
matched by diffusion approach. Also, under a diffusion process, firms never
default by surprise because a sudden drop in the firm's value is impossible.
Therefore, if a firm is not currently in financial distress, the probability
of its default on very short-term debt is zero. This fact implies that the
credit spreads tend to zero as the maturity of debt tends to zero, which
contradicts empirical evidence (see, for example, Fons (1994) and Sarig
and Warga (1989)).
The first thing that must be said about realistic models of the assets
of the firm is that jumps of essentially any size have to be incorporated.
Moreover, both negative and positive jumps have to be admitted to cap-
ture the evolution of assets in the information economy. One of the ways to
divert from the diffusion approach is to model the underlying asset process
as a jump-diffusion process. Zhou (1997) uses a jump-diffusion model in
the framework of the structural approach to valuing default-risky securi-
ties. Under a jump-diffusion approach, a firm can default instantaneously
because of a sudden drop in its value. The model allows for matching the
size of credit spreads on corporate bond to empirical data. However, the
bankruptcy level is given exogenously in Zhou's model, and the author does
not study the optimal capital structure. It is also necessary to mention that
in order that jump-diffusion models could be fitted well to empirical data,
many jumps have to be incorporated, which makes the models not very
efficient from both the analytical and computational points of view.
Here we study the optimal capital structure and endogenous default by
modelling the evolution of the firm's assets as a Levy process. One of the
advantages of models based on Levy processes is that such models admit
any number of jumps (infinite number including). Realistic models can be
obtained within a family of regular Levy processes of exponential type. Our
Endogenous default 233
goal is to show that RLPE are analytically tractable for problems of credit
risk modelling and endogenous default. Roughly speaking, we are going to
demonstrate that RLPE are the "second-best" if the Brownian motion fails
to produce realistic results.
We solve the endogenous default problem by using essentially the same
approach as in Chapters 5 and 10. We take any candidate for the level
of bankruptcy and calculate the corresponding candidate for the value of
equity by applying the Wiener-Hopf method. After this, we show that there
is only one candidate for the level of bankruptcy for which the found equity
satisfies certain natural conditions. The optimality can be proved as in
Chapter 5.
The plan of the Chapter is as follows. In Section 11.2, we specify and
solve a simple endogenous default problem. In Sections 11.3 and 11.4, we
study the firm near the bankruptcy, and a solvent firm, respectively, and
suggest two types of approximate formulas for all endogenous variables. The
first set of approximate formulas can be used very close to the default level,
and the second set of approximate formulas is valid for Levy processes with
large truncation parameters, i.e., for the processes with not very fat tails,
and the firm not very close to the bankruptcy level. These approximations
do not neglect the effects of non-Gaussianity on one hand, and on the other
hand, they allow to perform comparative statics analysis. In Section 11.5,
using the approximate solutions, we derive the optimal leverage for the firm.
We produce several series of numerical examples which show that for the
firm not very far from the default level, the difference between the Gaussian
and non-Gaussian models can be sizable, and very near the bankruptcy level
- very large indeed. In Section 11.6 we discuss the results obtained in the
Chapter and address the comparative statics issue. Technical results are
proved in Section 11.7.
11.2 Endogenous default
Consider a firm whose assets evolve according to the law
Vt=expXt, (11.1)
where X = {Xt}t>o is a Levy process, started at x — Xo, under a given

EMM Q. The stochastic process V is assumed to be unaffected by the
financial structure of the firm. Following the tradition in the relevant lit-
erature (see, for example, Leland (1994a)) we assume that a riskless asset
exists and pays a constant interest rate r > 0.
The firm issues debt, which attracts coupon payment Cdt when the firm
is solvent. In the event of default at time t, a fraction a G (0,1) of the
firm's value is lost to bankruptcy costs, therefore the debt holders receive
(1 - a)Vt = (1 - a) expX t . Let VB = eh be the bankruptcy level, i.e., the
firm defaults should the value of its assets fall below VB • Let r(h) be the
hitting time of (—oo,h] by X. For D(Xt), the value of the debt, we have
rr{h)
D(x) = Ex 'Cdt + Ex \e-rT(h\\ - a)e x '< h >] . (11.2)
By applying Theorem 2.12, we obtain
(r - L)D{x) = C, x>h, (11.3)

x
D(x) = (l-o)e , x<h, (11.4)
where L — —ip(D) is the infinitesimal generator of X.

Eq. (11.3)-Eq. (11.4) are analogues of the stationary Black-Scholes
equation (3) and (6i) in Leland (1994a); condition (6ii) in the above paper
also holds but it suffices to assume that D is continuous and bounded.
Next, following the steps of Leland (1994a), we consider the value of the
bankruptcy costs BC(x). This time, we have no stream of payments, and
hence BC(x) satisfies the homogeneous generalized Black-Scholes equation:
(r - L)BC(x) = 0, x>h, (11.5)
and the boundary condition is
BC(x)=aex, x < h. (11.6)
Eq. (11.5) is the analogue of (8i) in Leland (1994a); the condition (8ii) in
that paper is valid here as well. In fact, it suffices to require that BC is
continuous and bounded; such a solution is unique.
Now we consider tax benefits, TB(x), associated with debt financing.
These benefits resemble a security that pays a constant coupon TC1 as
long as the firm is solvent, and pays nothing at the bankruptcy. Hence,
1
T here should not be confused with the stopping time r(/i) above; the latter will not
appear any more
TB(x) solves the problem
(r-L)TB{x) = TC, x > h, (11.7)

TB(x) = 0, x<h, (11.8)
and it is also continuous and bounded.

The total value of the firm is the sum of its assets and tax benefits less
the bankruptcy costs:
v(x) = ex + TB(x) - BC(x). (11.9)
The value of equity is the total value of the firm less the value of debt:
E(x) = v(x) - £>(x); (11.10)
it can be represented as
E(x) = ex + / ( x ) , (11.11)
where f(x) = TB(x) - BC(x) - D(x). Notice that problems Eq. (11.3)-
Eq. (11.4), Eq. (11.5)-Eq. (11.6) and Eq. (11.7)-Eq. (11.8) are boundary
problems for a linear operator on the same half-axis, each having a unique
solution. This means that / is a unique bounded solution to the boundary
problem of the same sort, only with the data being the sum of the data of
the above problems, i.e.,
(r-L)f(x) = (r-l)C, x > h,

x x
f(x) = -ae - (1 - a)e
x
= -e , x < h.
Now it is easy to see that f(x), D(x), TB(x),BC(x) are solutions to the
following boundary problem:
(r-L)u(x) = A0, x>h, (H-12)

w(x) = Aiex, x<h, (11.13)
where h, AQ,AI £ R are constants. Therefore we will first characterize the

solution to Eq. (11.12)-Eq. (11.13), and then find f(x), D(x), TB(x),
BC(x) substituting the corresponding values for A0 and A\. By the same
argument, it suffices to consider separately Eq. (11.12)-Eq. (11.13) in the
cases Ao = 0 and A\ = 0. In the first case, Theorem 5.2 is applicable, and

the answer is
^ - ( D J l f . o o ^ f P ) - 1 ^ = Aê' - e f c ^ ( - 0 _ 1 * i ( x - h)],
where
*1(x)=<t>;(D)9+(x)ex, (11.14)
and 6± is the indicator function of R ± . In the second case, the modification

of the part of the proof of Theorem 5.2 gives the unique bounded measurable
solution A 0 r _ 1 * o ( - — h), where
*o = (l>7(D)0+- (11.15)
We have proved
L e m m a 11.1 For any / i 6 R and AQ,AI € C, the problem Eq. (11.12)-

Eq. (11.13) has the unique solution in the class of bounded measurable func-
tions. For x > h, it is given by
u{x) = Aor - 1 *o(a: - h) + Ai[e' - eh(j>-{-i)-1î{x - h)). (11.16)
Now we can calculate E(x) — ex by using Eq. (11.15)-Eq. (11.16) with

A0 = C(T — 1) and A\ = — 1, and obtain for x > h,
E(x) ex + C{T—1) $o(x _K)_ex + efc^-(_t-)-i$1 (a; _ h)
C(r-l)
*o(x -h) + eh(f>-{-i)-1yl{x - h)
"2^r + eV r "H)- *o(x - h)
+eh<t>-{-i)-1[^l{x -h)- V0(x - h)};
simplifying, we obtain
E(x) = <f>;(D)gh(x), (11.17)
where
9h{x) = 0+(x- h) Ûe^H)- (11.18)

Recall that <pr is the Fourier image of the probability distribution pr ,

supported at [0, +00), hence, Eq. (11.17) can be written as the convolution
E{x) = (jp~ *gh){x) = Prix- y)w{y)dy, (11.19)

Jh
where
w(y) = r-xC{T - 1) + e ^ - ( - i ) - 1 .
The value of the firm's equity having being obtained, the next step should
be to determine the level of bankruptcy for the case, when the firm can
choose this level, i.e., in the absence of a covenant (for example, positive
net-worth requirement). Leland (1994a) computes the endogenous level
=
of bankruptcy by using a smooth pasting condition: dE/dV\v=vB 0-
However, as we already saw in Chapter 5, the smooth pasting fails for
some non-Gaussian processes. Therefore, we are going to find the optimal
bankruptcy level as in Chapter 5, without resorting to the smooth pasting
principle.
Set
d(h) = lim w(x) (11.20)

x—yh+o
= r-lC{T-l)+eh<t>-{-i)-1.
Notice that due to Eq. (11.20), d{h) > 0 iff w(x) > 0 V x > h. If d(h) < 0,
we deduce from Eq. (11.19)-Eq. (11.20) that E{x) is negative at some points
x > h, which is impossible. Hence, h> /i„, where /i* solves d(h) — 0, or
VB=eh' = C{l T)
~ <t>;{-i), (11-21)
r
and it remains to show that with any choice h > /i„, the value of the equity
E(h;x) = eh(f>r-(-i)-1^!1(x - h) - C ( 1 ~ T ) $ 0 ( S - h) (11.22)

is less than or equal to E(h*;x), for any x > h. By using Eq. (11.19), we
obtain
E(ht;x)-E(h;x)= [ p~[x - y)w(y)dy. (11.23)

Jh.
Since w(y) > 0 on [/i,,+oo), we conclude that E{h*;x) > E(h;x), and the
optimality 2 has been verified.
Thus, we have proved the following theorem.
T h e o r e m 11.1 The optimal default level VB = eh' is given by Eq. (11.21).
Rewrite Eq. (11.21) as
VB = C(l - T)E
f
Jo
a~rônt nt
e dt (11.24)
where {nt} is the infimum process. It follows from Eq. (11.24), that the
asset value, VB, at which bankruptcy occurs is proportional to the coupon,
C, independent of bankruptcy costs, a, and decreases as the corporate tax
rate, r, increases. Moreover, we can see that the bankruptcy level does not
depend on the current asset value, V, but it depends on the parameters of
the infimum process for the firm's assets.
In the case of the Brownian motion, the bankruptcy level Eq. (11.21) is
exactly the same as determined by the smooth pasting condition in Leland
(1994a); it can be shown exactly as in Chapter 5, where we have shown that
our general optimal exercise price for the perpetual American put coincides
with Merton's result if X is the Brownian motion.
Now we determine the values of the firm's equity, debt, tax benefits,
bankruptcy costs and the value of the firm itself, given the value of the
assets and the level of bankruptcy, the former being a linear function of
the coupon C. By substituting Eq. (11.21) into Eq. (11.22), we obtain the
value of equity under the optimal choice of the default level
E(x) = °(1~T)[*1(x - h.) - *o(x - /»•)]. (H-25)

r
and Eq. (11.16) assumes the form
u(x) = Aor-ôix - K) + Ax[ex - C(1

~ r ) $i(a: - h.)]. (11.26)
r
With A0 = C and A\ = 1 - a, we derive from Eq. (11.26) the value of debt
D(x) = (1 - a)ex + - ¥ „ ( * - ht) - C(1

~ r)(1
" Q)
^ ( x - h,), (11.27)
r r
2
In the restricted class of hitting times of semibounded segments (-co, h]; the verification
of the optimality in the wider class M can be made exactly as in Chapter 5.
Equity of a firm near bankruptcy level and the yield spread for junk bonds 239
hence the value of the firm is
v(x) = (1 - a)ex + — * 0 ( z - h.) + ^ — ^ * i ( x - ft»), (11.28)

r r
the tax benefits obtain with A0 = TC and A\ = 0,
TB{x) = — # 0 ( z - ft*), (n-29)

r
and the bankruptcy costs are
BC{x) = a[e* - ^ — ^ * i ( x - ft,)]. (11.30)

r
By using formulas Eq. (11.27)-Eq. (11.30), we can compute the rate of
return on the debt
R(x) = C/D(x), (11.31)
the «/ieW spread
YS{x) = R(x)-r, (11.32)
and the leverage
L(x) = D(x)/v(x). (11.33)
By using the explicit formulas for the factors <j)f, first, h» can be calcu-
lated from Eq. (11.21), then $ 0 ( z - M and $ i ( z - ft,) from Eq. (11.15)-
Eq. (11.14). Finally, the values in Eq. (11.25) and Eq. (11.27)-Eq. (11.33)
can be computed. However, since the formulas for the factors involve inte-
grals, it is difficult to use them for comparative statics purposes. Therefore,
we are going to derive approximate formulas for all the relevant values.
11.3 Equity of a firm near bankruptcy level and the yield

spread for junk bonds
Consider the firm near the bankruptcy, i.e., the case of V/VB near 1, or
equivalently, x :— \II(V/VB) > 0 small. Clearly, this is the situation when
the presence of jumps must be felt in the first place, and hence the Gaussian
models can be expected to produce serious errors, and we show that the
mistake in the valuation of the firm equity near the bankruptcy level may
differ by hundreds percent - even a thousand percent - between Gaussian

and non-Gaussian models.
We start with the following approximate formulas, which are valid as
x -> +0; in these formulas, the exponent K_ appears, which is opposite to
the order of the factor <j>~ (£) in the Wiener-Hopf factorization formula. To
be more specific, from Eq. (3.75)-Eq. (3.76), as £ -* ±oo,
# 7 ( 0 = a_(0)(l + it)~K- + 0(K|-"-'), (11-34)

where p G (0, u - v{) is arbitrary. For RLPE of order v € (1,2] or of order
v € (0,1] with the drift coefficient (i = 0, K_ = v/2; for processes of order
1, K_ can assume any value from (0,1), depending on / j , and for processes
of order less than 1, K_ = 0, if /x > 0, and K_ = 1, if fj, < 0. Finally, it is
possible to show that for a more general class of processes than RLPE any
K- £ (-0.5,1.5) can be realised, the model example being
1>(Q = -itf (11-35)

v
+c+r(-v)[\+» - (A+ + iO ] + c-T{-V)[{-\-r - (-A_ - iff]
with c+ ^ c_ (see Chapter 15).
In the last Section, we will derive from Eq. (11.34) the following ap-
proximate formulas, as a; —• +0:
q-(Q) xK- +0(xK-+l>), (11.36)

*o(x)
r(«_ +1)
a-(0) xK~ +0(XK-+P), (11.37)
*i(x)
r(«_ +1)
9i(x)-*0(x) M O ) xK-+1+0{xK-+1+p), (11.38)
T(K_ + 2)
for any p £ (0,v — v\). Notice that the order of the leading terms in
Eq. (11.34)-Eq. (11.38) depend only on «;_, which can be regarded as a
measure of the asymmetry of PDF in the central part; the O-terms in
Eq. (11.36) and Eq. (11.37) differ as follows from Eq. (11.38). It is possi-
ble to write down the next terms in the asymptotic formulas Eq. (11.34)-
Eq. (11.38), but they are more involved.
Notice that even the simple formula Eq. (11.38) allows one to deduce
the following approximation for the equity value near the bankruptcy level:
E = C (1-T),,(0W VA~»
rr(K_ + 2) \ VB)
Equity of a firm near bankruptcy level and the yield spread for junk bonds 241
In the Gaussian = 1, and for processes of order less than one,

K- = v/2 + K_,I, where «;__! £ (—0.5,0.5) is greater than 0, equal to 0
or less than 0 depending on the asymmetry of probability densities of the
process (with the characteristic exponent Eq. (11.35), say): if the upward
jumps prevail, K_,I < 0, and if the downward jumps do, then «_,! > 0.
Clearly, for V close to the bankruptcy level VB, when h = ln(y/Ve) is close
to 0, the ratio of the equity in the non-Gaussian model to the one in the
Gaussian model can be very large indeed: 10 or even more.
We may apply the approximate formulas Eq. (11.36)-Eq. (11.37) to the
formulas Eq. (11.27)-Eq. (11.33) for other endogenous variables, and derive,
for instance,
Since the error of the approximation Eq. (11.40) is of the same order as the
value of equity in Eq. (11.39) or even larger, the usage of these formulas
for the calculation of the leverage makes no sense: the only result will be:
the leverage is approximately one, which is evident. (With the next terms
in Eq. (11.34)- Eq. (11.38), more information can be obtained).
For the return on the debt and the yield spread more informative for-
mulas are available:
C C r
R= = (U 41)
D~ (l-a)Vb (l-a)(l-T)#r(-»)' '
c r[i-(i-tt)(i-T)^(-o]
S r (1L42)
~D ~ (i_a)(i_r)0-H) •
They give the limit values of the return on the debt and the yield spread of
the firm near the bankruptcy level. In Fig. 11.1 we plot the default level VB
and the limit values of the yield spread in the KoBoL model, and the ratios
of these quantities in the non-Gaussian model to the ones in the Gaussian
model with the same volatility, for the interest rate r = 0.06, corporate tax
level r = 0.35, and bankruptcy cost a = 0.5; the characteristic exponent
is given by Eq. (11.35) with c_ = c + = c, u = 1.2, and the steepness
parameters A_ and A + varying from -30 to -8 and from 30 to 8, respectively.
The volatility vol=0.35 is fixed, which defines c as a function of A_ and A+,
from
vol = cT{-v)v{v - 1)[A+"- 2 + ( - A _ ) " - 2 ] ;
the drift coefficient fi is determined from the EMM requirement
r + ip(-i) = 0 (11.43)
and Eq. (11.35), i.e.,
fi = r + cr(-*/)[A + " - (A + + 1)" -1- (-A_) v - (-A_ - 1)"]. (11.44)
We see that for the limit value of the yield spread, the difference between
the non-Gaussian case and the Gaussian one is much smaller, only several
percent.
11.4 The case of a solvent firm
If the firm is not near the bankruptcy level, i.e., h = ln(V/Ve) is not very
small, the approximate formulas above are not good or even not applicable
at all. Fortunately, there exists another set of approximate formulas, which
give good approximations, if h is not very small - and if it is large, these
approximations are very good indeed. Namely, as we noticed in Lemma 3.7
and used in Chapters 5 and 7, the models classes of RLPE: TLP, NIG, NTS
Levy and HP, have a very useful property, namely, a(£) = r + VK£) admits
the analytic continuation into the complex plane with cuts (—ioo,iA_] and
[i\+, +oo), has no zeroes outside the imaginary axis, and at most 1 zero in
(0, iA_|_) and at most one zero in (iA_,0).
Moreover, for parameter values typically observed in Financial Markets,
both zeroes exist, and the steepness parameters are large. Denote the zero
in the upper half-plane by i{3 (in Chapters 3, 5 and 7, it was denoted by
—i/?_). In the last Section, we will show that under these conditions, the
following approximate equalities hold:
*o(s) ~ l + p-XK(fi)e-px, (11.45)

x l 3x
*i(x) ~ e 4>;{-i) + {l + p)- K{fi)e-l , (11.46)
where
(11.47)
ii>'(if3)<t>+(ipy
The case of a solvent firm 243
1.04v.-
0.31s
Fig. 11.1 Left panels: the default level Vg and the limit value of the yield spread. Right
panels: KoBoL model vs. Gaussian model. Parameters: r = 0.06, T = 0.35, a — 0.5;
vol=0.35, v = 1.2
Below, we will see that in order to get the correct dependence (in terms of
curvature) of the value of the firm and the value of equity on asset value,
K(p) should be negative, which is the case here. To see this, notice first
that
r+oo
/•-t-oo
tfW) E / e-rte~'3rntdt\mt =0
Jo
where {mt = s u p 0 < s < t Xt} is the supremum process for Xt (see Eq. (2.24)),
and so it is positive. Hence, it suffices to prove that iip'{i[)) is negative.
But *(0) = r > 0, \P(/3) = 0, and $ is concave as it was shown in the

proof of Lemma 3.5, and we conclude that the function \J>(z) := r + ij)(iz)
is decreasing at z = /3.
For a BM, the approximate equalities Eq. (11.45)-Eq. (11.46) are exact
ones. Moreover, K(/3) = —/3 and
0 (e* - e-px).
*i(s)
0 +1
It is easy to verify that the following approximations for Eq. (11.27)-
Eq. (11.33), which we derive by substituting Eq. (11.45)-Eq. (11.46), are
similar to exact answers in Leland (1994a) (compare Eq. (7), (9) and (11)-
(13) there (it may be not so easily seen since the formulas in op. cit. are
written in different variables):
C ( l - r)
E = £7jz7) - 1 ~ tfW1 - C1 + P)-1)^^'-^
D
= V-
C(l-r)
(1 - a)ex + Cr~l{\ +
1 + K(fi)
m +1) (0
p~xK{p)e-0{-x-h^)
(11.48)
C(l-r)(l-a)(c,
+ (/? + ly^Kê-^-V)
C(l-T)
1 (l-r)(l-a) -0'
£ 1 + ^(0)
^ 0+ 1 (£) (11.49)
i> = F +
Cr (1 - T)a0 V_ -P (11.50)
0(0 + 1) 1 + 0 +
1 +
VB
KtJ$) (V
TB =
r 1+ 0
(0 (11.51)
BC =
C(l-r) Fr + K(fi) ( V \~ff
V
C(l-r) /3 + 1 UBJ
N _/3
0+ H0-
a C ( l - r ) AT(/3) / V
(11.52)
The case of a solvent firm 245
coupon
Fig. 11.2 Left panels: the value of the debt and the yield spread. Right panels: KoBoL
model vs. Gaussian model. Parameters: r = 0.06, T = 0.35, a = 0.5; vol=0.35, v = 1.2,
A_ = -30.
First, we apply these approximate formulas to study the dependence of the

value of the debt and the yield spread on the coupon and the steepness
parameter A + describing the left tail (see Fig. 11.2). We see that the de-
pendence is of the same type as in the Gaussian model in Leland (1994a);
in this series of examples, the difference between the Gaussian model and
non-Gaussian one is relatively small, up to 3.5 percent.
The next series of numerical examples shows that the difference between
non-Gaussian and Gaussian models can be significant not only very close
Fig. 11.3 Left panels: equity and debt value. Right panels: KoBoL model vs. Gaussian
model. Parameters: V/VB = 1.2, C = 10, r = 0.06, a = 0.5, r = 0.35, vol=0.35, u = 1.2.
to the bankruptcy level (see Fig.ll.3-11.4). We fix the ratio of the firm
assets to the bankruptcy level V/VB = 1.2, which makes the approximate
formulas Eq. (11.48)-Eq. (11.51) applicable, with a small error, if A+ and
—A_ are not small; and this is the case we consider here: both A+ and
—A_ vary from 30 to 8. We see that the non-Gaussian equity value and the
Gaussian one can differ by 23 percent. In this series of examples, the value
of the debt in non-Gaussian model can be larger by 6 percent than in the
Gaussian model, the same holds for the firm's value, and the yield spread;
the leverage differs insignificantly.
Endogenous debt level and endogenous leverage 247
11.5 Endogenous debt level and endogenous leverage
Now it is straightforward to verify that E(V) is a convex function of the

asset value V, which is due to the fact that K{fi) < 0. This reflects the
"option-like" nature of equity, even if debt has an infinite horizon. The
value of the firm, v, is strictly concave in V for positive C and r. Notice
that v is also strictly concave in C, so it is possible to find the optimal
coupon C* as a solution to
where Vjj is given by Eq. (11.21). Whence we derive

-1//S
Vr K{p)
(1 + /3+(IzzM) (11.53)
(l-r)#T(-i) P
It is interesting to notice that the optimal value of the firm and the optimal
value of debt are exactly the same functions of C* as in Leland (1994a),
i.e.,
v* = V +— - ^ - , (11.54)
r p+ 1
D* = — ^ r , (11-55)
r p +1
where
1
m = 1+
p + 1 + (1 - r ) a / 3 / r '
The optimal return on debt is equal to
r(/3 + 1)
R* = C*/D* =
mp
which is again the same expression as in the diffusion case, but for a different
P-
Numerical examples show that as far as —A_ and A+ are not small,
only the endogenous coupon and the default level can differ by 3 percent;
the values of other endogenous variables change less when we replace the
Gaussian model with the non-Gaussian one, of the same variance.
11.6 Conclusion
We applied the structural approach to modelling of the optimal capital

structure and endogenous default, when the evolution of underlying assets
follows a Levy process. We started with a simple benchmark model of
Leland (1994a) where the debt had infinite maturity. Formulas obtained
for endogenous variables in the model, are suitable for any Levy processes.
However, we suggested using a wide family of regular Levy processes of
exponential type as the "second best" as compared to diffusion processes
in terms of tractability. For these processes, analytical expressions for the
endogenous variables are available, and in more complicated situations, an
asymptotic solution can be possible.
We developed two types of approximate formulas; the first type is suit-
able for a firm very close to the bankruptcy level, and we used them to show
that near the bankruptcy level, the value of the equity in the non-Gaussian
model can differ several times from the Gaussian model; this striking dif-
ference is natural since the presence of jumps must be felt near the default
boundary in the first place.
If the firm is not very close to the default level, the approximation
above is not very good or not applicable at all, but in this region, another
approximate formulas are suggested. We exploited the special structure of
the characteristic exponent of RLPE with large steepness parameters (the
case supported by empirical evidence), in order to obtain a relatively simple
tractable approximations for the solutions of the non-Gaussian model. The
tractability of these approximate solutions can even amount to comparative
statics. The solutions look like those in the diffusion model, but this does
not mean that some Gaussian model can produce the same result. There
are many coefficients in the whole set of formulas, and it is impossible to fit
all of them by using just one parameter - the variance; for complex models,
even two-three parameters may be not enough. So, non-Gaussianity still
tells in the values of the above coefficients.
The explanation of the fact that seemingly Gaussian-like type of formu-
las are capable of capturing non-Gaussian effects stems from the following
observations. First, the approximation has small errors and for these errors,
the explicit bounds are available. On the other hand, the coefficients in the
formulas differ from the naive Gaussian ones by much larger margin.
These approximate formulas can be used for the comparative statics
purposes by following the next steps.
Auxiliary results 249
First, it is possible to study the dependence of the root /? on the param-

eters of the process. For instance, for a KoBoL process with — A_ = A+,
one can show that /? is a decreasing function in v, A+ and c _ 1 , and for a
model class of NIG, the root /? can be found as a positive solution to a
certain quadratic equation, with coefficients explicitly determined by the
characteristics of the process. In this case, the comparative statics for /? is
as simple as in the BM case.
Second, if the truncation parameters are not small, an approximate
formula (j>~ (—i) = P/(/3 +1) is valid, and this formula can be used for com-
parative statics purposes. The error in this formula is larger than the errors
in the approximate formulas for \&o and $ i , and so for computations, it is
better to calculate </>"(—i) by using more accurate numerical procedures.
Similarly, the approximation to K(/3) in terms of /? can be obtained.
Third, by using approximations for endogenous variables, one can study
comparative statics of those as functions of f3,4>~(—i),K(f3), and eventually,
as functions of the parameters of the process.
Notice, however, that the numerical results above show that for rela-
tively large values of the steepness parameters, the significant difference
with the Gaussian models is observed for a firm near the bankruptcy level,
hence the second set of approximate formulas can produce results essen-
tially different from the Gaussian ones only when the firm is not far from
the bankruptcy level.
The exact formulas can produce sizable difference with the Gaussian
model for a firm far from bankruptcy, but only if one of the truncation
parameters is small, i.e., the corresponding tail is fairly fat.
We believe that the most exciting effects of non-Gaussianity will be
revealed in a model with finite maturity of debt, and in the more detailed
study of junk bonds, which can be done by using the technique employed
in Chapters 7 and 8 for the study of barrier options and first-touch digitals.
11.7 Auxiliary results
11.7.1 Proof of Eq. (11.36)-Eq. (11.38)

The Fourier transform of 9j is well-defined in any half-plane of the form
S£ < cr, where a < —j, and is given by
*i(0 = ^ ( 0 ( ^ - j ) _ 1 -
Notice that as £ -r oo in the same half-plane,
(ic-jrâ+^r'+oder2),
therefore it follows from Eq. (11.34) that for j = 0,1,
# : (*)(»* - i ) " 1 = o-(0)(i + i 0 - " - ~ 1 + /,(£), (H-56)
where /(£) = 0 ( | £ | - K - _ 1 - P ) , as £ -> oo in the half-plane Q£ < a. Hence,

for any e > 0, Fj :— e~ixT~xfj belongs to the Sobolev space i J s ~ € ( R ) ,
where s = /c_ + l + p— 1/2; by the Sobolev embedding theorem, HS_£(R)
is embedded into the Holder space C s - £ - 1 / 2 ~ € 1 ( R ) , for any ei > 0. Since
p G (0, v — v\) is arbitrary, we conclude that Fj G C K _ + P l ( R ) , for any
pi £ ( 0 , i / - i / i ) .
A direct calculation shows that for any K > 0,
r+oo
lc
^(xr "*)(0 = / e-ix(-xK-xe~xdx (11.57)
= r(/c)(i + iO"K,
and from the definition of \&j, j = 0,1:
-oo+i<r
for any a < -j, and from Eq. (11.56)-Eq. (11.57) it follows that
9j(x) = a_(0)r(«_ + l ) _ 1 x + - e - x +ejxFj(x). (11.58)
We know that \Pj(x) = 0 for any x < 0, hence from Eq. (11.58), Fj also
vanishes on (-oo,0]. Since Fj G CK-+P(R), for any p G (0, v - i>i), we
conclude that a s i - 4 +0,
Fj(x) = 0(xK-+p). (11.59)
Eq. (11.58)-Eq. (11.57) prove Eq. (11.36)-Eq. (11.37), and Eq. (11.38) is
proved similarly, the equality
-oo+i<r
for any a < — 1, being used.

11.7.2 Proof of Eq. (11.45)-Eq. (11-46)

Take a < — 1. Then we can write
-00+t<7
V>(£) admits the analytic continuation into the complex plane with cuts
(—ioo-NA_], [i\+, -H'oo), and it has only one zero i/3 in the upper half-plane.
Therefore we can use the Wiener-Hopf factorization formula Eq. (2.23) to
define the meromorphic continuation of (j>~ (£) into the complex plane with
the cut [i\+, +ioo) and the only pole at k = i/3. Notice that as £ -*• i/3,
0r(O =
r
+
(r + WP) + i>'(il3)(Z ~ iP))&W) °(1)
T
= K{P){Z-iPrl+0{l). (11.60)
Transform the contour of integration 9 £ = a by pushing it up; first, it

crosses the pole £ = 0, and then £ = ifi, and in the end we obtain the
contour
£ = (-Moo - 0, i\+ - 0) U (i\+ + 0, +ioo + 0).
On the strength of the residue theorem, and from Eq. (11.60), we obtain
90(x) = 1 + K(/3)0-1e-0x + (2wi)-1 f eixl<l>T{€)Cldk, (11.61)
and similarly,
*1(x) = cl>;(-i)e* + K(p)(p + ir1e-ex + (2ni)-1 J eîOtf + ir'dZ.

(11.62)
By comparing Eq. (11.61)-Eq. (11.62) to Eq. (11.45)-Eq. (11.46), we see
that it remains to show that the integrals in Eq. (11.61)-Eq. (11.62) are
small. To see this, make the change of the variable £ = iz on the contour
£; in the result, we obtain, for a = 0,1,
-1 J eix^;
(27TJ)- 1 / e ^ - ^ + ia)-1^
r+oo
= (27r)-1/ e-"*{z){z + a)-ldz,
where
$(z) = i(4>~{iz - 0) - 4>~{iz + 0)).
For model classes of RLPE, $(z) grows at the infinity not faster than Cz"
- see, e.g., Eq. (11.35) - and hence, we conclude that the integrals in
Eq. (11.61)-Eq. (11.62) admit a bound via
+oo
r+oo r+oo
/-t-oo
zx l l V-\J0-ZX
l
C / e- z"- dz = -Cx~ j z"- de
+oo
= Cx-1A+"-1e-A+I+Cx-1(j/-l) f e-"z"-2dz.
JxA+
If A+ is large and x > 0 is not too small, then we conclude that the integrals
in Eq. (11.61)-Eq. (11.62) are small. For instance, if x = 0.2 (that is, we
are 20 percent from the default level), v is close to 1 and A + = 40, which
is quite often the case in the real financial markets, then the upper bound
for the integrals is less than 0.01.
Fig. 11.4 Left panels: firm's value, yield spread, and leverage. Right panels: KoBoL
model vs. Gaussian model. Parameters: V/VB = 1.2, C = 10, r = 0.06, a = 0.5,
T = 0.35, vol=0.35, v = 1.2.
Chapter 12
Fast pricing of European options
We consider the Fast Fourier Transform (FFT), and a new method of op-
tion pricing based on the transformation of the line of integration into the
integral over an appropriate cut in the complex plane (integration-along-cut
method - IAC). The method is applicable for option pricing under VGP,
NTS Levy and KoBoL of order less than 1, and NIG.
12.1 Introduction
Option pricing under non-Gaussian Levy processes has a serious disadvan-

tage from the computational point of view since an explicit formula for
the price kernel is usually not available, and even if the explicit formula is
known as it is the case with Normal Inverse Gaussian processes, the com-
putations are more time-consuming than in the Gaussian case. Madan and
Carr (1998) suggested to use an explicit formula for the characteristic ex-
ponent of a Levy process and the Fast Fourier Transform (FFT). However,
F F T can produce significant errors for options not at the money, close to
expiry, and in op. cit., for FFT, no general effective procedure for the choice
of the line of the integration and sufficient number of points for a required
upper bound of the computational error were suggested.
We consider a new method, which allows one to calculate option prices
fast, and extremely fast just prices of options out of the money and deep
in the money, even close to the expiry, with good accuracy. The method is
applicable to option pricing under several families of Levy processes used
in empirical and theoretical studies of financial markets, and hopefully, for
option pricing under some stochastic volatility models as well. The method
255
256 Fast pricing of European options
is based on the transformation of the line of integration into the integral

over the banks of an appropriate cut in the complex plane, which is related
to the characteristic exponent of the process. As a short-hand notation,
the name I AC {integration along a cut) method is suggested.
If one calculates the price of the European put or call for just one value
of the stock price out of the money or deep in the money, then for a given
upper bound of the computational error, IAC can be hundreds times faster
(or even more) than FFT. For options at the money, the relative advantage
of IAC as compared t o FFT, is not that significant. Nevertheless, even if
one calculates the prices of an option at a hundred points or so, which is
needed for the parameter fitting purposes, then for many processes, IAC
can be several times faster than F F T for options close to the expiry; for
processes of order close to zero, 10-20 times faster. For some processes, there
is a region in parameters' space where F F T with its ability to calculate the
option prices for many spot prices at once performs better. Typically, the
relative strength of IAC increases as the option approaches the maturity.
Notice that we can effectively control the computational error of IAC,
which is not possible with FFT. By using this property, it is possible to
show that F F T method with parameters' values used in Carr and Madan
(1998) may provide significant error, and compare the speed of IAC and
FFT for a given upper bound of the computational error (see Levendorskii
and Zherder (2001)).
The basic idea of IAC - to change the contour of integration in order
to obtain the integral with a better rate of convergence - is the same as
that of the saddle point method (SPM) (see, e.g., Fedoryuk (1988)), and
in principle, SPM is faster. However, the choice of the optimal contour for
SPM can be rather tricky, and for IAC, the choice is quite straightforward.
Another nice property of IAC is that the resulting integrand is a real-valued
function; this property enhances the speed of calculations.
After reduction to the cut is made, we have to calculate an integral of
the form
+oo
e-zuf(v,z)dz, (12.1)
/
where u > 0 is proportional to, roughly speaking, a linear function of the
distance between the log spot price and the log strike price, and v > 0
is proportional to the time to the expiry, T. The upper bound for / can
easily be obtained for model classes of RLPE. This allows one to estimate
Introduction 257
effectively the error of the replacement of the semi-infinite interval of the

integration with a finite one. One can easily calculate the derivatives of
/ w.r.t. z, obtain a bound for an appropriate derivative, and then find
the error of a chosen integration procedure. By using the upper bound for
the error, we find the number of points necessary to achieve the required
accuracy. If the option prices at many points are needed, and hence the
values of I(u, v) are to be calculated for (u =)uo < u\ < • • •, we find the
number of points, which is necessary to calculate J(uo, v) with the desired
computational error, calculate the array of values f(v,Zj) at these points,
and use the array to calculate I(ua,v) for all s.
Notice that there are many additional possibilities to enhance the speed
of the calculations. For instance, one can find the asymptotics of the inte-
grand as z —» oo and use this asymptotic expansion in the calculation of
the integral in a suitably chosen neighbourhood of the infinity, with a de-
sired error. In some cases, the whole contour lies in this neighbourhood of
the infinity; in cases when a part of the contour remains outside the neigh-
bourhood, one can use simple standard integration procedures to calculate
the integral over the remaining part of the contour. In many cases, the
numerical integration requires the calculation of values of the integrand in
several dozen of points or just several points, which makes the calculation
extremely fast.
Our starting point is the following observation made in Chapter 3 about
properties of model classes of RLPE. Namely, VGP, together with the other
model classes of RLPE listed above, enjoy the following property, which is
crucial for I AC: the characteristic exponent of X is holomorphic in the
complex plane with two cuts on the imaginary axis (one in the upper half-
plane, and the other in the lower one), and the same holds for mixtures of
independent processes of aforementioned types.
For payoffs of interest, like the ones for European calls and puts, and
for many others, their Fourier images are meromorphic with poles outside
the cuts, hence, one can fairly freely transform the contour of integration,
by using the Cauchy theorem and the residue theorem.
Using these observations, we can transform the line of integration and
obtain an integral with a better rate of convergence. The method suggested
here, IAC, uses the contour, which passes along the banks of an appropriate
cut in the complex plane. Notice that this reduction procedure can be
justified for an RLPE of order less than or equal to 1: for VGP, NIG, and
KoBoL and NTS Levy processes of order v < 1; for processes of order v > 1,
after reduction to the cut, the integrand grows at infinity, and hence, I AC
is not applicable. We have not checked whether IAC is applicable for all
GHP. Since the formula for the characteristic exponent of a GHP, which is
not a NIG, is rather complicated, and the relative advantage of IAC is small
for processes of order 1, we surmise that it is not rational to apply IAC to
the other GHP's (though this question deserves a more careful study).
The implementation of the method requires the detailed calculations in-
volving the characteristic exponent. We illustrate the usage of the method
by calculating the price of the European put option under the KoBoL pro-
cesses of order (0,1).
12.2 Transformation of the pricing formula for the Euro-

pean put
Consider the model market of a riskless bond and a stock. Let r > 0 be
a (constant) riskless rate, and let St = exp Xt be the price process of the
stock. We assume that X is a Levy process under an EMM Q chosen by
the market; ip is the characteristic exponent of X under Q. We assume
that the steepness parameters satisfy A _ < — 1 < 0 < A + .
Consider the European put with the strike price K and the terminal
date T. The payoff at the expiry equals max.{K — ST, 0}, and without loss
of generality, we may assume that K = 1. Then
+oo
e-fa€(i_e*)+dz
/ -OO
is well-defined in the half-plane 3£ > 0 and admits the meromorphic exten-

sion into the complex plane with two poles at £ = 0 and £ — —i. Assume
that X is a KoBoL, and denote by f(x, t) the price of the American put
at time t and spot price ex. Take a e (0, A + ), and express f(x,t) as (for
details, see Levendorskii and Zherder (2001))
fix t) = ^ /+°°+i<T eXpmX + TM) + T C

"((A+ +
* r + (
~ A ~ ~ *)V)] dC
2n J_00+ia -i£(-it + 1)
(12.2)
Transformation of the pricing formula for the European put 259
where cv = cT(—v), and
Ro = e x p [ - r r - TCV(X+V + (-A_)")]. (12.3)
Set
dv = A+" + (-A_)",
K = A+ + (-A_),
X+ = A+/K,
X- = -A-/K,
and notice that x+ + X- = 1- By changing the variable £ —» K£ , we obtain
ff rt = _ ^ _ / • + ° ° + i ' T / ' c e x p [ ^ ( x + rAt)]

AX
' ^ 2 7 r « y _ 0 0 + W K _ ^ ( - ^ + i/«)
• e x p K r C l / ( ( x - - if)" + (x+ + OTK-
Introduce new parameters:
£/ = K(X + T ^ ) ,
V = —K"TCU,
P = 1/K,
and denote:
a _ RQ ^M-rr + v{xv- + xu+)) (124)

1
27TK 27TK
We have U € R, V € R, p < l, ax € (0,x+), and
f(x,t) = Ri d£. (12.5)

./-oo+iff, -K(-*?+P)
In the next section, we describe two methods of computation of Eq. (12.5):
FFT and IAC.
12.3 F F T and IAC
12.3.1 Fast Fourier Transform

In this subsection we follow Carr and Madan (1998); for a detailed treat-
ment of FFT, see Elliot and Rao (1982). The Fast Fourier Transform is an
efficient algorithm which is used for computing sums of the following form:
iV-l
X(k) = ] T e-*frO-iXfc-Da.y), Jfc = l,..,JV, (12.6)
where N is a power of 2. The analogous algorithm for computing the sum

N-l
X{j) =£ e'W-W-VxW, J = 1,.-, N, (12.7)
fc=i
is called the inverse Fast Fourier Transform (iFFT).

Denote o~\ by a, and rewrite Eq. (12.5) as follows
f(x, t) [+°°+i° exp{iC/e - V[(X- - it,)" + (x+ + *)"]}

d£
Ri^1 J-oo+ia
J-oo+i -%(-% + p)
+°° expjixz + izrp, - V[(x- -iz + a)u + (x+ + iz- cr)v]}d_
r oe -i(z + ia)(-iz + a + p)
where A = e-^"^+rt>) m Introduce
R(x;a) = Rie-K^x+Tfi\
and
expjizTii - V[(X- -iz + a)" + (x+ +iz- <*)"]}
1p(Z:a) = ; ; r~; : r
V ;
(-iz + <r)(-iz + a + p)
then
+oo
eixzip(z; a)dz. (12.8)
/
-oo
For a chosen step of numerical integration r), we use the trapezoidal rule
for the integral Eq. (12.8) and obtain an approximation for f(x,t):
N
f{x, t) » R(x; a) £ e*"' ip(Zj; a)v, (12.9)
FFT and IAC 261
where Zj = r](j — 1). F F T returns N values of x with a regular spacing of

size A, so that values for x are
xk = -b + X{k-l), k = l,2,..,N,
where b — ^ . With x = Xk, equation Eq. (12.9) becomes
N
f(xk,t) » R(xk;<T) £iei*'<--b+x<k-1H(zi;<r)TI,
y
(12-10)
j=i
and substituting Zj = n(j - 1) into Eq. (12.10) yields:
N
f(xk,t) » Rixâ)^^-1™-1^-^^;^- (12-11)
j=i
To apply the FFT, one couples A and r] by
A*? = f . (12.12)
This restriction plays a role of some kind of "the uncertainty principle": if

we choose r] small in order to obtain a fine grid for the integration, then we
obtain relatively large spacing A on the "moneyness" line. To overcome this
difficulty we should increase the number of points N and use more accurate
method; for example, one can incorporate (as in Carr and Madan (1998))
Simpson's rule weightings in summation:
f(xk,t) « R(xk]a)Y/eiXn(k-1)(j-1)e-ibz^(zj;a)7l[3 + (-1)' - l ^ ] ,
(12.13)
where 5n is the Kronecker symbol.
The sum in Eq. (12.13) is of the form Eq. (12.6), hence it can be com-
puted by FFT. Notice that it is rather difficult to estimate the error of this
computational scheme.
12.3.2 Integration-along-cut method

Let v e (0,1). We consider the integral Eq. (12.5) in the following two
cases: a) U > 0; b) U < 0.
12.3.2.1 Case a)
If U > 0, we push the contour of integration to the cut in the upper half-
plane, the choice of a cut being defined by a natural requirement: iU£ is
negative at the cut. As a result, we obtain
r+<
/•-t-tc
°° exp[iUt] ,-V>(£+0) _ -VM-O)
f(x,t) i?, d£,
- * ( - « £ + p)
where <£(£) = ( x - - i O " + (x+ + *0"-
By changing variables £ —> iz, z S R, we obtain:
/•+0
+°° exp[-£/z] e-V^>(iz+0) _ e-V<t>(iz-0)\ dz_
/(x,t) = «!
*(* + P)
For z > x+, we have
</>(iz±0) = ( X _ + z T i0) v + (x+ - * ± «))"

- (x_+z)" + ( 2 - x + ) " e ± ^
= ( x - + z)" + (z - x+)"(cos(i/7r) ± isin(i/7r)).
Set
a„ = cos(j/7r), 6„ = sin(i>7r), (12.14)
then
exp[V^(iz + 0)) - exp(-K<?i>(iz - 0)]
= exp(-V(X- + z)u - V{z - x + ) " a „ ) [e-iV^z-x+^b" - em—x+)"b^
v
= - 2 t e x p ( - V ( X - + z) - V(z - x+)"a„) sin(V(z - X+YK),
and
+°° exp[-E/z] s i n ( y ^ ( z - X+Y)
f(x,t) 2i? x
x+ L z z
( + P)
exp ( - V [ ( x - + z)" + au{z - X+T]) dz.
By changing the variable z -> z + x+, we come to
(12.15)
+
" °° exp[-£/z] sm(V&„z")
2Rie-u*+ / exp ( - V [ ( l + z)v + auzv\) dz.
Jo (z + x+)(z + X+ + P)
FFT and I AC 263
Introduce the family of integrals

+00
. . f e zusm(vbvzv) . .
l{v,U,V,p1,p2)= / ; rz ; r eXp (-V (1 + Z) +avz])dz,
(12.16)
where u > 0, v > 0, pi > 0, pi > pi, v e (0,1). Thus, in the case U > 0,
the computation of f(x,t) reduces to the computation of the integrals in
Eq. (12.16):
f(x,t) = 2Rie-u*+l(v, U, V, X+, X++P)- (12.17)
12.3.2.2 Case b)
In the case U < 0, we push the contour of integration in Eq. (12.2) to
the cut in the lower half-plane. In the process of the transformation, the
contour crosses two simple poles £ — 0 and £ = — i. Using the residue
theorem, we obtain
f(x,t) — exp(—rr) — exp(x)

r>°° «PTO re-v*«-c» _ e -v*«+o)i dc
J-ix- -*£(-*£+ P) L J
Making the change of variables £ = — iz yields
f{x, t) = exp(—rr) — exp(x)

+Rl r ( _. } exp[£/z] r _v«_iz_0) _ e _v*<-«+o)l d^
z z + 1 J
Jx- ~ (- P)
For z > x-,
cf>(-iz±0) = {X--zTi0)l/ + (x+ + ^±i0)v

= {x++zY + {z-x-Ye^^
v
= (x+ + z) + (z - x_)"(cos(z/7r) qF isin(i/ir));
therefore, by using the notation of Eq. (12.14),
exp{-V<f>(-iz - 0)) - exp(-V<f)(-iz + 0)) =
= e x p ( - ^ ( x + + z)v - V(z - x_)"a„) (jn'-x-Yb* _ e -^(z-x_)"fc„\
= 2i exp(-V(X+ + z)v - V(z - X-)vav) sin(V(z - X-TK)-

Hence
f(x,t) = e x p ( - r r ) — exp(x)
r+ v
r °° exp\Uz] sm(VK(z - X-) )
Jy-
• exp (-V[(X+ + z)v + av(z- x _)"]) dz.
Change the variable z i-+ z + X-
f(x, t) = e x p ( - r r ) - exp(x) + (12.18)
f J. V j . vexp(-V[(l + z r + a ^ ] ) d z ,
Jo (z + X-)(-2 + X - - P )
and use the notation of Eq. (12.16):
/ ( x , t ) = e x p ( - r r - ) - e x p ( x ) + 2 i ? 1 e ^ - I ( z , , - f 7 , V, x_, X--p)- (12-19)
We conclude that any numerical procedure for the integral Eq. (12.16),
together with Eq. (12.19), can be used to compute f(x, t) in the case U < 0.
Example. In the case v = 0.5, we have av = 0, bv — 1, and the integral
becomes especially simple:
+oo
l(0.5,u,v,Pl,P2)= f * ' " s m ^ ) exp {-vy/T+7) dz. (12.20)
7 (^ + Pl)(^ + P2)
0
12.4 C o m p a r i s o n of F F T a n d I A C
When applied to the computation of a discrete approximation of an integral,

FFT produces an error, which depends on the value of the return parameter,
and becomes significant at the end of the interval of integration. To reach
the required accuracy, usually one should choose a large number of points
of integration.
Applying IAC, first, we choose a finite interval so that the integral
over the neighbourhood of the infinity is less than e/2, where e > 0 is the
desired error, and in the numerical integration over a finite interval, we
use the Simpson method, after an appropriate change of the variable has
been made; the number of points should be chosen so that the error of the
method is less than e/2, too.
Comparison of FFT and I AC 265
Numerical experiments produced in Levendorskii and Zherder (2001)

show that this scheme works well for large A_, A + , and, if A_, A+ are
fixed, for not large c„. The last property can be reformulated as follows:
IAC works better for small r, usually, less than 10 days to expiry. Also, IAC
is preferable for v € (0,0.5), or even for u from a wider interval, in certain
region of the space of other parameters. For instance, for v € (0.125,0.5)
and A+ and —A_ of order 40-50, IAC is 15 times faster than FFT, when
one calculates the prices of the put options at 100 points in the moneyness
interval (0.7, 1.05), 3 days to expiry.
On the contrary, F F T works better for c„ large and v of order 0.5 and
larger.
Finally, the speed of IAC can be enhanced by using additional devices
mentioned in the Introduction.
On the two figures below, we show results of numerical experiments,
which supplement the numerical results in Levendorskii and Zherder (2001).
We fix r and all the parameters of KoBoL but one (/x), which is determined
from the EMM-requirement r + ip(—i) = 0.
We study the dependence of the ratio of the computational time (IAC
vs. FFT) needed to calculate 100 option prices in the moneyness interval
(0.75, 1.05), with the given upper bound for the error e = 1 0 - 4 . Notice
that we were able to use relatively small number of points for F F T just
because we knew the result with the error less than e = 10~ 5 from the test
calculation made with the help of IAC; hence, in practice, FFT would have
required even more time.
In Fig. 12.1, the dependence on c and the order of the process is shown;
T = 3 (day) is fixed, and in Fig. 12.2, we show the dependence on the order
and the time to expiry.
(a)
0.9 - (b)
0.8
0.7 - / • • • • ' "
0.6 - / • • "
0.5 - / /
0.4 -
^ ^ ^ ^
0.3 - ^ — - ^ ^ ^ - - • • - • • • " " • "
0.2 - '"
0.1 -"
'
Fig. 12.1 T h e ratio t(IAC)/t(FFT). Parameters: e = 10~ 4 , r = 0.03, r = 3; A+ = 35,

A_ = - 5 5 ; a) c = 5; b) c = 3.
0.15 0.2 0.25 0.3 0.4 0.45
Fig. 12.2 T h e ratio t(IAC)/t(FFT). Parameters: e = 10~ 4 , r = 0.03; \+ = 35,

A_ = —55, c = 5.
Chapter 13
Discrete time models
We consider perpetual Bermudan options and more general perpetual Amer-

ican options in discrete time. For wide classes of processes and pay-offs,
we obtain exact analytical pricing formulas in terms of the factors in the
Wiener-Hopf factorization formulas. Simple formulas are derived in the
case when the original probability density is approximated by exponential
polynomials.
13.1 Bermudan options and discrete time models
In discrete time, one can make very weak assumptions on the underlying
stochastic process, and yet obtain analytical solutions; at the same time,
the discrete time approach makes it possible to solve the pricing problem
for non-standard American options known as Bermudan options. A Bermu-
dan option can be exercised on specified dates only, during the life of the
option. An example of a Bermudan option is a bond option whose exer-
cise is restricted to coupon payment dates. Consider a Bermudan option
which can be exercised at the prescribed dates tj = jAt only, and the cor-
responding perpetual American option in discrete time: the answer for the
latter option gives the optimal exercise price and the price of the Bermudan
option on the prescribed dates, the price of the Bermudan option at time
t € (tj,tj+i) can be calculated as the price of the European option with
the expiry date i,+i, and the price of the Bermudan option at time tj+i as
the terminal payoff.
The state space in our model is R, and not a discrete set as in multino-
mial models (for these models, see, for example, Duffle (1996) and Shiryaev
267
268 Discrete time models
(1999)). We assume that the increments are identically distributed and

independent but we do not suppose that the transition density is infinitely
divisible so that it is possible to pass to the continuous time limit.
The optimal stopping problem for a perpetual American call option
in discrete time was solved by Darling, Ligget, and Taylor (1972), and
Mordecki (2000) has obtained the pricing formula for a perpetual American
put by applying the technique of op. cit. The aforementioned results
are obtained with the probabilistic technique, without explicit analytical
formulas, and it is not clear how to apply this technique to other types of
perpetual American options.
In this Chapter, we include the results obtained in Boyarchenko and
Levendorskii (2001b). They are the discrete-time analogue of the results
of Chapter 5, and the technique is similar. After the general formula for
the optimal exercise price for a Bermudan option is obtained, we consider
an example, when the model for shocks makes it impossible to pass to the
continuous time limit.
The approach developed in this chapter makes it possible to solve a
discrete time version of the model examined in Chapter 10. Prom our point
of view, discrete time models are more relevant for Economics in general,
and Real options in particular. In the first place, such setting allows one
to make only week assumptions on the underlying stochastic process X.
Second, we believe this model to be especially useful in cases when the
time interval between observations is not very small, as it is the case in the
theory of Real Options. Notice that even though the Real Options literature
to date has provided many insights into capital budgeting decision-making,
theoretical models are not widely used by practitioners. Among one of
the primary reasons for that, Lander and Pinches (1999) point out that
many of the required modelling assumptions are often violated in practical
real options applications. In particular, this concerns the choice of the
stochastic process for the underlying variable. As we already mentioned
it in Chapter 10, in spite of the fact that the normality of the process
is rejected by empirical evidence, Gaussian processes are often used in the
investment literature. The second reason for theoretical models being rarely
used by practitioners is technical involvement of the models. The task of
fitting the parameters of a model process to the data by no means makes
life easier. In a discrete time model, the empirical probability density can
be approximated by exponential polynomials with desired accuracy and
A perpetual American put in a discrete time model 269
simplicity (although there is certainly a tradeoff between these two). After

that, to obtain the analytical expression for the investment threshold one
has only to find roots of two polynomials - an easy task for any practitioner.
13.2 A perpetual American put in a discrete time model
13.2.1 Process specification

Time is discrete: we consider only dates tk, k = 0 , 1 , . . . , when the option
can be exercised, and we assume them to be equally spaced: A := tj+i — tj
is independent of j . We normalize A = 1. Consider a market of a riskless
bond and a stock. We assume that the stock pays no dividend. Let the
riskless rate be fixed and let S = {Sn}n>o, Sn — expXn, be the price
process of the stock. Let Y, Yi,... Yn,... be independent and identically
distributed random variables on the probability space fi, and let P(dx) be
the probabiUty distribution of Y. We assume that under a chosen EMM,
the log-price Xn admits a decomposition into a sum of n independent and
identically distributed random variables Y\,...Yn: Xn — Y\ + Yi -\ (- Yn.
Of P(dx), we require
+oo
exP(dx) < oo. (13.1)
/
-oo
This condition is necessary in order that the stock be priced under the
chosen EMM, and the requirement that {Sn} is a martingale under the
chosen EMM assumes the form
+oo
exP{dx) = 0, (13.2)
/
•oo
where e~r < 1 is the discount factor per period.
The condition Eq. (13.1) implies that the Fourier transform of P(dx), p,
is holomorphic in the strip 3;£ 6 (0,1) and continuous up to the boundary
of the strip. To simplify the study of the pricing problem, we impose
additional conditions on p. We assume that p is holomorphic and bounded
in a wider strip:
|P(0I < C, 9£e(-A+>-A_), (13.3)
where A_ < — 1 < 0 < A + , and there exist positive constants e 0 ,Cb, and p
such that
\m)\ < 0)(1 + | £ i r p , ê(-€0,eo). (13.4)
Notice that Eq. (13.4) implies that P(dx) is absolutely continuous: P(dx) =
p(x)dx, and Eq. (13.4) is satisfied if p is continuous, differentiable a. e., and
+oo
e£°lxy(x)|dx < + o o .
/ •oo
Eq. (13.4) can be relaxed but we believe that this condition is weak enough
for practical purposes.
13.2.2 The pricing problem for a perpetual American op-

tion as an optimal stopping problem
Consider a perpetual Bermudan option on the stock, i.e., a perpetual option
that can be exercised on specified dates only. Let g(Xt) be the payoff
function for the option: for a put g(x) = K—ex, and for a call g(x) = ex—K,
where K is the strike price. Denote by V*(x) the rational price of the
perpetual Bermudan put; then (see e.g. Shiryaev (1999). XVIII, 2)
V,(x) = sup Ex[e-rTg(XT)}, (13.5)
where M denotes the set of all stopping times r = T(U;), satisfying r(w) <
o o , {J € Q.
As in Chapter 5, first we will find a solution to Eq. (13.5) in the class
Mo of hitting times r{a) of segments (—oo,a]. To this end, we consider
V(h; x), given by Eq. (5.21), which is the price of the put, when h is chosen
as the exercise boundary. The sufficient conditions for the optimal exercise
boundary are given by Lemma 5.2. Since this lemma is stated in terms of
the value function V(h,x), we need an explicit formula for it. To obtain
such a formula, one can proceed as follows.
Fix h, a prospective candidate for the optimal exercise boundary, then
the option pricing problem can be written as
V(x) = e~rE[V(x + Y)], if x > h; (13.6)

V(x) = g(x), ifx<h; (13.7)
A perpetual American put in a discrete time model 271
here (with some abuse of notation) V(x) := V(h;x). We can rewrite

Eq. (13.6) as
+oo
p(y)V(x + y)dy (13.8)
/
•oo
/+oo
p(y - x)V(y)dy, x > h.
•oo
By introducing the symbol a(£) = 1 - e _ r p ( - £ ) , and using an equality
+oo
p(y - x)u(y)dy = (f-lp(-OF(u))(x),
/ -OO
we may rewrite Eq. (13.8) as
(a(D)V)(x) = 0, x > h, (13.9)
Introduce a function 31, which coincides with g on (—00,h]; and it is

continuous on [h, +00) and vanishes faster than e~e°x as x —> +00. Set
u(x) = V(x + h)-gi(x + h), then we obtain from Eq. (13.9) and Eq. (13.7)
the following system of equations:
(a{D)u)(x) = -a(D)gi{x + h), x > 0; (13.10)

u(x) = 0, x<0. (13.11)
Thus, the original problem has been reduced to the Wiener-Hopf equa-
tion. As in the previous chapters, we can obtain the solution to Eq. (13.10)-
Eq. (13.11) using the Wiener-Hopf method. After the solution is obtained,
it is easy to guess the necessary condition for the exercise boundary. In
fact, under our standing assumption on the process, the optimal exercise
boundary is determined by the requirement: V(h; •) is continuous at h. We
call this requirement the continuous pasting condition as opposed to the
smooth pasting condition, which never holds in discrete time.
It can be shown (for details see Boyarchenko and Levendorskii (1998))
that the necessary condition is similar to Eq. (5.31), so that an explicit
formula for h obtains. After that, one has only to check that the sufficient
conditions hold for this h. This is done exactly as in Subsection 5.2.2. Fi-
nally, by using the following sufficient condition from Darling et al. (1972),
one can prove the optimality in the same fashion as it was done in op. cit.
for calls.
Lemma 13.1 Let Y, Yi, Y2,... be independent and identically distributed

random variables, let XQ — 0 and Xn = Y\ -\ (- Yn for n > 1. Let g and
f be non-negative functions and z a constant satisfying 0 < z < 1. If for
all x,
f(x) > g(x) (13.12)
and
f(x)>E[zf(x + Y)\, (13.13)
then
f(x)>E[zTg(x + XT)} (13.14)
for all x and all stopping times T.

Evidently, the last statement can be used to justify the solution in the same
way as Main Lemma in Chapter 5 was used.
Thus, we solve the pricing problem for perpetual options in discrete
time by essentially the same method as in continuous time. The only issue
remains to be addressed is the Wiener-Hopf factorization in discrete time
models. This is done in the next Section.
13.3 The Wiener-Hopf factorization
13.3.1 General formulas

Set Xn — maxo<j< n X,-, and X„ = mino<j<„X,'; their analogues in con-
tinuous time are called the supremum and infimum processes, respectively.
Let z € (0,1), and let T be a random variable independent of {Yj} and
taking values in { 0 , 1 , . . . } , with P(T = n) = (1 - z)zn.
Theorem 13.1 (Spitzer (1964))
E[e*x*] = (l-z)/[l-zp(-0] (13.15)

= E eaxT E ei£XT (13.16)
Moreover, we have the Spitzer identities

' °° yk roo
E ei<LXT
= exp ,**e _ 1V)dx P(Xk <x) (13.17)
and
°° ~k yO
(13.18)
Set
C«*T1 t (13.19)
(13.20)
and write Eq. (13.15)-Eq. (13.16) as
(1 - z)/[l - zp(-0] = M O M O - (13.21)
Lemma 13.2 0/" 0+ (resp., 4>-) is the Fourier transform of a distribution

P+(dx) supported on (—00,0] (resp., P-(dx) supported on [0,+00)/
b) 4>+(£) (resp., <£_(£),) is holomorphic in the half-plane S£ > 0 (resp.,
3£ < 0^ and continuous up to the boundary of the half-plane.
Proof. We have
0+(O = E[e*Xr]
/•+00
= Y V l - z)zn / eixidP(Xn < x)
= y + ° ° ete«d f ] T > - z)z n P(X n < 1) j
= / e-<*tp+(dx),
J — OO
where
P + (dx) = d ( £ ( 1 - *>n(-P(Xn < -X)U .
This proves a) for the sign "+"; the proof for the sign "-" is similar, and
b) follows from a). •
The Spitzer identities Eq. (13.17)-Eq. (13.18) are by no means computation-

ally effective. More convenient formulas can be obtained under additional
conditions on p.
Lemma 13.3 Let Eq. (13.4) hold. Then

a) there exist —e0 < o- < 0 < <r+ < e0 such that
»[1 - e " r p ( - 0 ] > 0, 9£e(ff_,<r+); (13.22)
6) <A+(£) is holomorphic in the half-plane S£ > <r_, and can 6e defined, for
any w_ > <r_ and S£ > u;_, 6j/
1
ex /-+°°+^- e - V ( - ^ ) .!?-£.
<M£) = P (13.23)
oo+iu,_ l - e - r p ( - n ) n
+0o+ia,_ gin[i_e-rp(-T7)]
= 6XP dn (13.24)
p/- «(£ - v)
«+(0) (13.25)
MO'
where
•1 +oo+tu;_
/-t-OO- lnfl-e-^-n)]
a+(£)=exp dr] (13.26)
2fft . / _ „ , +
c) $-{£,) is holomorphic in the half-plane Off < o~+, and can be defined, for
any w+ < cr+ and S£ < u>+, by
ex (13.27)
<M0 = P 27T*
m 7y__0 0 + j w + l_e-pp(-T7) n
+oo+iw+ glnfl-e-^-n)]
exp d77 (13.28)
2iriJ_ oo+iw+
MO) (13.29)
MO'
where
a_(f) = exp
1
2m f/•+°°+i"+ ln[l - e " p ( - " ) ]
-oo+iuq. "-£
dn (13.30)
Proof, a) For real cr, and n on the line Qn =CT,we have

r+oo
e- CTa: p(x)dx.
-oo
The Wiener-Hop} factorization 275
Clearly, M(0) — 1, and by continuity of M, there exists a neighbourhood

of zero, denote it (er_, <r+) such that for all a G (<r_, <r+), e~TM{o) < 1 and
Eq. (13.22) holds.
b) Set z = e~r. Since the Fourier transform of a sum of independent
random variables is the product of the Fourier transforms of summands,
and Yi, Y2,. • • are independent identically distributed random variables, we
have for £ G R:
«2L zk r°°
53 T J (eix*-l)dxP(Xk<x)
°° zk f°° f+°°
= E l / (e"C - 1 ) ( 2 f f ) " 1 / _ e-ixnP(-v)hdvdx
Take u>_ G (<r_, 0), and shift the line of integration:
= E T / (eil€ - 1)(27r)"1 / e-ix»p(-v)kdvdx

K
1 JO J—oo+iui-
Assume that Eq. (13.4) holds with p > 1. Then the integrand admits a
bound via the integrable function Ceu-X(l + |»/|) - ''. Hence, we may apply
the Fubini theorem and integrate w.r.t. x first:
(In the case p G (0,1], we have to resort to the theory of oscillatory integrals
in order to justify the change of order of the integration). As o;_ —> 0,
C(«_) := sup |p(-r?)| -» 1;
hence, for any z G (0,1), there exists o>_ G ( c - , 0 ) such that zC(wJ) <
1. This implies that ^2 J converges absolutely, and for such an u;_, we
can apply the Fubini theorem and change the order of the integration and
summation:
Inll-zpi-vMrj-t)-1-^1^
•oo+iw-
This proves Eq. (13.24)-Eq. (13.25), and Eq. (13.23) is deduced from
Eq. (13.24) by integration by part.
c) is proved as b). •
13.3.2 Some useful properties of the factors a±

Lemma 13.4 a) We have
a(0=o+(Oa_(0 (13.31)
b)For any w € (<x_,0), there exist C,pi > 0 such that for all £ in the
half-plane 9f£ > u>,
M O ^ - l l ^ C a + iel)-" 1 ; (13.32)
c) For any LJ G (0,cr+), there exist C,pi > 0 such that for all £ in the
half-plane 3£ < w,
\a-(t-)±l-l\<C{l + \Z\)-n. (13.33)

Proof, a) foUows from Eq. (13.21), Eq. (13.25), Eq. (13.29) and the equal-
ity a(0) = 1 - er.
b) It suffices to show that the integral in Eq. (13.26) admits an estimate
via C ( l + |f | ) _ P l for £ in the half-plane 9£ > u). Take w_ € («r_, w). From
Eq. (13.4), it follows that the integrand in Eq. (13.26) admits an upper
bound via
Ci(l + \T)\rp\r)-tr\
where C\ and p > 0 are constants. By considering separately cases \TJ\ <
If 1/2) \T}\ > 2|£| and |f |/2 < \rj\ < 2|f |, and using the triangle inequality, it
is easy to prove that there exists C 2 such that for all rj on the line Sty = w_
and all £ in the half-plane 3?£ > w,
Ci(i+M)-"|77-f|-1<c2(i+|£|)-"/2[(i+|^-f|)-1-''/2+(i+H)-1-''/2].
The RHS is integrable over the line QTJ = w_, and the result is C(l +
|f \)~p/2, with some constant C; Eq. (13.32) follows.
c) Eq. (13.33) is derived in the same fashion. •
13.3.3 The case of exponential polynomials
The discrete time model is more flexible than the continuous time model
since an empirical probability density can easily be approximated by an
appropriate model one so that the factors in the Wiener-Hopf factorization
formula are computed (relatively) easily. Here we consider the simplest

approximation by exponential polynomials. 1 Set p = p_ + p+, where p± =
1 R ± P - If P- is an exponential polynomial:
p.(x) =£c+eA+*|af'lR_(x), (13.34)

3= 1
+
where m is a non-negative integer, A+ > 0, and cj" are real.
Similarly, if p+ is an exponential polynomial:
p+(x) = ^ c J e ^ - V l ^ (x), (13.35)

3=1
where m~ is a non-negative integer, A_ < 0, and cj are real. We are going

to show that if both p± are exponential polynomials then p, <j>- and cj>+ are
rational functions. For rj in the half-plane 3ty < A+,
p_(-tj) = f eixr,p_(x)dx
J —OO
oo
r0 m+
ixr,+\+x(_xydx
•°° 3= 1
+
m
= 5 ^ c + r ( j + i)(A+ + tf 7 )-'- 1 ,
3=1
and for r) in the half-plane S77 > A_,
p+(-7,) = 2 c 7 r 0 ' + X )(- A - -iT?)"i_1-
Whence we see that
a(t) = l-e-rp(-t). (13.36)
is a rational function, i.e., function:
"M'WY <13 37>

'
1
More elaborate approximations are possible as well, see for example, Boyarchenko and
Levendorskil, (2001b).
where P and Q are polynomials, and therefore, the factor
1_ e-rp(-r))
in Eq. (13.27) is the rational function of the form R(T])/(P(T))Q(T])), where

R = P'Q — PQ' is a polynomial of degree deg R = deg P + deg Q — l. Since
for a fixed £,
l n ^ i = ln(l-^/r,) = 0(|r,r1)
as T] —> oo, we conclude that on any line S77 = w, which does not pass
through a root of P Q , the integrand in Eq. (13.27) admits an estimate via
C(l + |T?|) - 2 , and the same holds in the half-plane S77 > u>, if there are
no roots of PQ in this half-plane. Therefore, we can apply the residue
theorem and shift the line of the integration: w+ —* +00. Each zero of PQ
contributes to the resulting formula, and in the result, we obtain that 0_
is a rational function. Similarly, 4>+ is, and since <£_ (resp., <f>+) is bounded
and does not vanish in the half-plane 3£ < 0 (resp., S£ > 0), we conclude
that
* - < « = n r ^ n ^ . (13.38)
where {—z/3j} (resp., {—ijj}) are all the zeroes of P (resp., Q) in the upper
half-plane. Similarly, <£+(£) is defined by zeroes in the lower half-plane.
13.4 Optimal exercise boundary and rational price of t h e

option
In this Section, we present the analytical formulas for the optimal exercise
price and the rational price of the perpetual American put option in discrete
time.
The former can be derived as
fc. = ln(Jf^_(-t)), (13.39)
where <p-(-i) is given by analytical expression Eq. (13.28). Given the

Optimal exercise boundary and rational price of the option 279
optimal /i», we calculate the rational put price as
27ry_0o+ia,, (-if)(l-i£)
where u;+ £ (0, a + ) is arbitrary.

Notice, that to derive the solution in the class of hitting times Mo, we
used the following properties of the payoff function g:
(i) g is continuous on R;
(ii) g is bounded and positive in some neighbourhood of - c o ;
(iii) there exists /i* such that
for all x < h*, (4>-(D)~1g)(x) > 0, and
for all x > h,, ((A-(J5) _1 /)(x) < 0.
Hence the result can be generalized for any payoff functions satisfying the
properties above. These properties are most easily verified for the payoffs
given by Eq. (5.81). It is also possible to impose additional conditions on
the parameters jj in Eq. (5.81) to obtain the optimal solution in the class
M.\ the conditions and proof (based on the reduction to the case of the put)
are essentially the same as in Chapter 5, where the continuous time model
is studied.
Below we give the simplest example, when a discrete time process is not
supposed to correspond to the observation of a Levy process (in continuous
time).
Example 13.1 Let
p ± (*) = ^ l R ± ( x ) e A ^ ,
where A_ < - 1 < 0 < A + . Then

r0
1 H-oo \
and
a(r)) = 1 - e r p(-7?)
= 1-
2e r A_ — it]+ A+ + irj
2(1 - e r )A_A + + (1 - 2er)(A+ + \J)vt\ + 2er„2
r
rj
r
~ 2e (A + + iii){-\- - irf)
Denote by — i/3_ (resp., — i/3+) the root of the numerator in the upper
half-plane (resp., lower half-plane); then
i ( 2 e r - l ) ( A + + A_)±iy/]9
-ifc = 4e r
D := (2e -l) (A+ + A_)2-16er(er-l)A+A_,
r 2
and we have
(77 + »/3_)(ij + i{3+)
ofo) = 2er(r]-i\ )(T]-i\-)'
+
By using Eq. (13.38), we obtain
* /« (A + + i Q ( - / 3 - )
=
*-«> A+(-/3_+i0-
Insert into the formula for the optimal exercise price:
(A+ + l)(-/3_)
Hm = eh' = K
A + (-/3_ + l) '
and then into the formula for the price of the perpetual American put: for
x > /&*,
K ,+oo+iW+ e <(«-J».)€(A + +ifl(-/?_)
«.
'•w-5/. ,+iui+ A+HL+ifl(-iO(l-*)
Under condition Eq. (13.22), io+ hes below — i/3_, hence iu; + does as well,
and when we push the hne of the integration up, it crosses the pole — i/3_.
By the residue theorem, we obtain
K{\+ + p_)e^^-h')
K(x) =
A+(l-/3_)
Chapter 14
Feller processes of normal inverse

Gaussian type
In this Chapter, we describe main results of the paper Barndorff-Nieken

and Levendorskii (2001). We consider and analyse natural generalisations
of constructions of NIG processes and some related processes. The resulting
Feller processes are somewhat similar to the NIG Levy process but may, for
instance, possess mean-reverting features. In contrast to the RLPE case,
the infinitesimal generators are no longer PDO with constant symbols, and
hence simple explicit pricing formulas are not available. Nevertheless, effec-
tive approximations to solutions of the corresponding generahzation of the
Black-Scholes equation can be derived. This generalization is an analogue
of the Black-Scholes equation with variable coefficients in the Gaussian
modelling, which is used by some practitioners to reduce the smile.
14.1 Introduction
We construct a class of Feller processes with generators possessing the most

nice properties from the point of view of the theory of pseudodifferential
operators; this class naturally generalizes the class of NIG processes. Sim-
ilarly, generalizations of other classes of RLPE can be constructed, and
they enjoy similar nice properties. Namely, the corresponding PDO be-
long to the class of classical PDO (see Hormander (1985) and the definition
in Chapter 16). Notice that in the monograph of Jacob (1996), where a
pseudo-differential approach to the study of stochastic processes is strongly
advocated, a class of PDO with non-smooth symbols (model examples be-
ing generators of Levy stable processes) is considered; this leads to more
difficult situations having much less relevance to applications in Financial
281
282 Feller processes of normal inverse Gaussian type
Mathematics and Physics. 1

The aim of the Chapter is to consider generalizations of the construc-
tions of NIG Levy processes described in Chapter 3. In the probabilistic
language, we want to consider Feller processes such that for any fixed x,
the exponent ip in the equality Ex[el^x'~x^\ = e - ' ^ 1 ' ^ is a characteris-
tic exponent of NIG Levy type, but with one or more of the parameters
depending on x; in the language of the theory of PDO, we want to con-
sider processes whose generators are PDO with "non-constant" symbols,
i.e., symbols depending not only on the dual variable £ but on x as well.
A natural way from the probabilistic viewpoint is to consider processes ob-
tained by subordination from general diffusion processes; this leads to the
necessity of calculation of the symbol of the generator. Another, simpler,
variant is to consider processes whose generators are PDO's with symbols
—a(x, £) given by the same formula as the characteristic exponent of a NIG
process, with /z, S, a and /3 depending on x:
a(x,0 = - v ( x ) £ + S(x)[(a(x)2 - (/?(*) + i^) 2 ) 1 / 2 - (c*(x)2 - P^x) 2 ) 1 ' 2 ].

(14.1)
We consider and compare these constructions. One can try various types
of x-dependence, including fairly weird ones; we consider relatively simple
versions which may have useful applications.
Recall the roles played by the steepness parameter, a, and the asymme-
try parameter, f3:a — (3 describes the rate of exponential decay of the right
tail of the density function, and a + 0 describes the decay of the left tail.
The scale parameter, 5, plays essentially the same role as the variance in
Gaussian models. In other words, the larger the value of a — (3 the smaller
the probability of large positive jumps, and the larger the value of a + / 3 the
smaller the probability of large negative jumps. If we want to reproduce a
mean-reverting effect then at high levels of x, we should have probability
of large positive jumps smaller than that of negative jumps; at small levels
of x the probability of large positive jumps should be larger than that of
negative jumps. Thus, (a(x) — /3(x))/(a(x) + /3(x)) must be increasing.
If we want to have a less volatile behaviour for small levels of x, then we
1
There is one useful property which generators of NIG and other RLPE processes fail to
have: the transmission property, Boutet de Monvel (1971), or a weaker smoothness-in-
a-half-space property, Eskin (1973), which simplifies the treatment of boundary value
problems. This property ensures that a solution to a "sufficiently good" boundary value
problem, e.g., the Dirichlet problem, with smooth data is smooth up to the boundary.
Introduction 283
may either decrease the scale parameter 8 (which is in direct analogy with
a standard device in Gaussian modelling) or increase a. For instance, one
may expect that if for some c > 0, p > 1,
a(x) + 0{x) > cx~p, x>0,
then the trajectories of the process never reach 0 from above. Processes of
this type are suitable for interest rate modelling purposes. To treat such
processes, the calculus of PDO with degenerate symbols is needed; for such
a calculus, see Levendorskii (1993) and Levendorskii and Paneah (1994).
In this book, we consider the simpler case of bounded a, 0, 5, fj., assum-
ing that 6 and a±0 are bounded away from 0. In addition, we assume that
these parameters are smooth and their derivatives are small. The simplest
example which we have in mind, and which is capable of reproducing the
mean-reverting effect, is the case of constant 5 and a, and 0 given by
0{x) = — ^ arctan(e(x - x0)) + 0O, (14.2)

7T
where e > 0, x > 0, and \0o ± xl < a- The derivatives of 0 admit estimates
\0(s)(x)\ < Cse°, 5 = 0,1,..., (14.3)
where the Cs are independent of e, and if we assume that e is small, we can

obtain asymptotic solutions in terms (roughly speaking) of a power series
in e.
Another example obtains when e in Eq. (14.2) is not small but e\ = x/a
is small; in this case, ei plays the part of e above, and if both e and e\ are
small then an asymptotic solution in terms of a series in the still smaller
parameter eei is obtained.
By using a characterization theorem for certain operators satisfying the
positive maximum principle (Courrege (1966b)) and the pseudodifferential
operators technique, one can show that a general theorem due to Hille-
Yosida-Ray (see Courrege (1966a), Ethier and Kurtz (1986), Jacob (1996))
is applicable to PDO with the symbol Eq. (14.1) and a, 0,6 specified as
above (see Subsection 14.2.1 for details), and therefore the PDO — a(x,D)
is the infinitesimal generator of a Feller semigroup on Co, the space of
continuous functions vanishing at infinity.
If a small parameter is present then, as we will show, the two construc-
tions of NIG-like Feller processes give processes with approximately equal
generators.
14.2 Constructions of NIG-like Feller process via pseudod-

ifferential operators
14.2.1 Naive construction: NIG-like generators with state-

dependent parameters
Definition 14.1 Let U G R n be an open domain whose closure contains

the origin, and let m € R.
We write a G 5 m ( R n x ( R n + iU)) if the following two conditions are
satisfied2
(i) a G C°°(R" x R ? ) , together with all its derivatives, admits an ana-
lytic continuation w.r.t. £ into a tube domain S£ G U
(ii) for any multi-indices a, j3, the derivative aitl admits a continuous
extension up to the boundary of U, and satisfies an estimate
\al;](x,0\<Cap(Om-M, (14.4)
where the constants Cap are independent of (x, £) G R n x (R" + iU).
Since in many cases the holomorphicity of a is not required, we will use

this definition with U = 0 by letting U = {0}. For this U, we recover the
standard Hormander class S™0(Rn x R").
Definition 14.2 Let m G (0,2], a G 5 m ( R n x (R n + it/)) and, for each

fixed x G R n , suppose a(x, £) is the characteristic exponent of a Levy
process. Then we write a G FLSm(Rn x (Rn+it/)).
Let Eq. (14.4) be valid on R n x R n , i.e., a G 5 m ( R n x (R n + j{0})). Then

a PDO a(x, D) maps <S(R") into itself, continuously (see Theorem 16.2).
The next fact is a variation of many similar statements in the theory of
PDO; for the proof, see Chapter 16.
Theorem 14.1 Let Eq. (14-4) be valid on R n x R", and let there exist
c> 0 such that on R " x R n
9Ra(x,£) + l > c ( £ ) m . (14.5)
Then there exists A > 0 such that a(x, D) + A : <S(R") —> <S(Rn) is invert-
ible.
2
T h e more standard notation would be S J " 0 ( R n x ( R n + iU)); we have omitted t h e lower
indices to simplify the notation.
Constructions of NIG-like Feller process via pseudodifferential operators 285
Theorem 14.2 Let m € (0,2] and a e FLSm{Rn x ( R n + iU)). Then a

PDO a(x,D) has a closed extension (A,D(A)),D(A) C C 0 ( R n ) , such that
—A is the generator of a Feller semigroup on Co(R").
Proof. By the theorem due to Hille-Yosida-Ray (see, e.g., Courrege
(1966a), Ethier and Kurtz (1986), Jacob (1996)) it is necessary and suffi-
cient to verify the following three conditions:
(i) V{a{x, D)) := <S(Rn) is dense in C 0 ;
(ii) a(x,D) satisfies the positive maximum principle on T>(a(x, D));
(iii) for some A > 0, the range of a(x, D) + A is dense in Co(R n ).
Since <S(Rn) is dense in Co(R n ), (i) is fulfilled, and if we take Theorem
14.1 into account, we obtain (iii). Finally, (ii) is a special case of a theorem
due to Courrege (1966b); for a pseudo-differential version of this result, see
Jacob (1996). D
Notice that in Courrege (1966b) the smoothness of the symbol was not
required, and an analogue of Theorem 14.2 can be established for much nas-
tier symbols than here (see Jacob (1996)). We use the simplest conditions
suitable for our purposes.
Definition 14.3 Let the conditions of Theorem 14.2 hold.
Then we call a Feller process X with generator —A a Feller-Levy process
(on R n ) of order m and exponential type U and write X € FLPm(Rn; U).
If we do not want to specify the order m and/or U, we say that A" is a
regular Feller-Levy process of exponential type.
If the symbol a of A is independent of i , we call the process X a Levy
process (on R n ) of order m and exponential type U; if we do not specify
the order m and/or the exponential type U, we say that X is a regular
Levy process of exponential type.
In the sequel, we identify a(x, D) with its extension, A.
By choosing a particular Levy process as a starting point, we can obtain
special classes. In the following definition, we use NIG Levy as a model,
and we consider the simplest case possible; more involved versions can also
be considered.
Definition 14.4 {NIG-Xike Feller processes in ID) Let /i, 6,a,/3e C£°(R),
S and a be positive, /x and (3 real-valued, and let there exist C, c > 0 such
that for all x,
S(x) > c, a{x) - |/3(x)| > c, \n(x)\ < C. (14.6)
Let a be defined by Eq. (14.1).

Then we call a Feller process X with the generator — a(x, D) a NIG-like
Feller process.
Notice that it is a process of order 1 and exponential type [A_, A+] for
any A_ < A+ satisfying
s u p i - a ( x ) + p{x)} < A_ < 0 < A + < inf{o(x) + /?(*)}. (14.7)

X
X
Condition Eq. (14.7) can be satisfied due to Eq. (14.6).
Definition 14.5 (Multi-dimensional NIG-like Feller processes) Let /x, a,

/3 6 C £ ° ( R n ; R n ) , 5 € C ° ° ( R n ; R + + ) , and let there exist c> 0 and open
sets U, V C R " such that {0} c U C V; for all x, 8(x) > c; and
(a(x) - p{x) + i{,a(x)+P{x) -itfgR-, V (s,f) e R " x (Rn + iV).
Here (•, •) is a bilinear form in C n , which extends the standard scalar prod-
uct in R n .
Let a be defined by
a(x, 0 = -in(x)£ + 5(x)[(a(x) - /3(i) + if, a(x) + f3(x) - i f ) ^

-(a(x)-0(x),a(x)-P(z))1/2\-
Then we call a Feller process X with the generator —a(x, D) a 7V/G-like
Feller process.
This is a process of order 1 and exponential type U.
14.2.2 Constructions via infinitesimal generators: subordi-

nation of semigroups of operators
We will use the following theorem due to Phillips (see Theorem 32.1 in Sato
(1999)).
Theorem 14.3 Let {Zt : t > 0} be a subordinator with Levy measure

p and drift /3Q, and let Xt denote the law of Zt. Let {Pt : t > 0} be a
strongly continuous semigroup of linear operators on a Banach space B
with infinitesimal generator L. Define
Qtf = I PsfX'ids), feB. (14.8)

J[0,oo)
Constructions of NIG-like Feller process via pseudodifferential operators 287
Then {Qt : t > 0} is a strongly continuous contraction semigroup of linear

operators on B. Denote its infinitesimal generator by —A. Then T>(L) is
the core of A, and
-Af = /30Lf+ [ (Psf - f)p(ds). (14.9)

J(0,oo)
We apply Theorem 14.3 with Pt being the transition semigroup of a

diffusion process with the infinitesimal operator L,
Lf{x) = \(o-2(x)d, d)f(x) + (b(x), d)f(x) + c(x)f(x), (14.10)
where a2 G C£°(R n ;EndR n ) is a positive definite (uniformly in x) sym-

metric matrix, b G C£°(Rn;Rn), c G C£°(R n ), c non-positive; and with
the subordinator having Laplace exponent K(U) — (d2 + u)"l2 — d", where
v € (0,2), and d > 0. This means, in particular, that flo = 0, and the
Laplace exponent and the Levy measure, p, are related by
/•OO
K(W)=
/ (l-e^'Xtis). (14.11)
Jo
To show that A is a PDO and calculate its symbol, we notice that Eq. (14.9)
can be written as
/•OO
-Af = / (exp(sL) - l)p(ds)

Jo
= K(-L) = (d2 - L)v'2 - dv. (14.12)
Since d > 0 and &L is a non-positive elliptic operator, d? — SRL is a positive-

definite elliptic operator, and, therefore, fractional powers of d2—L are well-
defined. Moreover, in the theory of PDO there is a well-known result about
complex powers of an elliptic PDO, which states that these powers are in
turn PDO's; and an asymptotic expansion of the symbol is provided. For
PDO on compact manifolds, the result is due to Seeley (1967); an analogue
for rather general classes of PDO on R™ was obtained in Hyakawa and
Kumano-go (1971) (see also the monograph Grubb (1996)); and in the case
U = {0}, the result in Grubb (1996) suffices for our simple class; in the case
of a general U, a straightforward modification of constructions in Grubb
(1996) is needed.
Notice, however, that the 'W/G-generating" Laplace exponent (the case
v = 1) is so simple that no general result on the asymptotic expansion of
the symbol is required: one can easily calculate an asymptotic expansion of

the symbol, b = b(x, £), of the square root of an elliptic PDO d2 — L(x, D)
with positive-definite real part, by using the Ansatz
b{x,t) ~ 61 ( i , 0 + bo(x,Q + 6 _ i ( i , 0 + • • •, (14.13)
where bj € 5- 7 (R n x ( R n + iU)), and the composition theorem, one of the

basic tools of the theory of PDO. As the result, we obtain the following
theorem.
Theorem 14.4 Let U C R" be an open set whose closure contains 0 and
such that
d2 - 3JL(x, 0 > 0 V (x, 0 € R n x (R n + iU). (14.14)
Then
a) n(-L)+d is a PDO with the symbol of the class 5 1 ( R n x (Rn+iU)),
and the symbol admits an asymptotic expansion Eq. (14-13) in the sense
that for any N > 0,
N-l
b
rN = b-J2 i-i e S1-N(Rn x (Rn + iU)), (14.15)
j=0
where
h(x,0 = (d2-L(x,0)1/2; (14.16)
bo(x,0 = -î(ar,0_1 E ^Mh^x,*); (14.17)

M=i
b) if the coefficients of L depend on a small parameter e > 0 and for any
s derivatives of order s admit an estimate viaCses, where Cs is independent
of e, then uniformly in e > 0,
(t-Hj G Sj(Rn x ( R n + «?)), e~NrN e S1~N{Rn x (Rn + iU)); (14.18)
c) let the coefficients of L be of the class C^°(R n ) and d > 0 be large;

then Eq. (14-15)-Eq. (14-18) are valid with a small parameter d~l instead
ofe;
d) if the coefficients satisfy the condition in b) andd is a large parameter,
then Eq. (14-15)-Eq. (14-18) are valid with a small parameter e/d instead
ofe.
Applications for financial mathematics 289
For the proof and inductive formulae for bj, j < 0, see Chapter 16.
Theorem 14.4 means, in particular, that NIG-like Feller processes con-
structed by subordination from diffusions with the uniformly elliptic — L,
where L is the generator of the diffusion, are iVIG-like processes in the
sense of the "naive" definition. Moreover, if we use the symbol a(x, £) =
(d 2 — L(x, ^)) 1 / ' 2 — d to construct a NIG-like process, and if the coeffi-
cients of L are slowly varying and/or d is large, we obtain a process whose
generator differs insignificantly from -K(-L), the generator of the process
obtained via subordination.
14.3 Applications for financial mathematics
14.3.1 Generalised Black-Scholes equation for contingent

claims under NIG-like Feller processes
Let r > 0 be the riskless rate, let St = expX t be the price process of a
stock, where Xt is a NIG-like Feller process with the generator — a(x,D),
and let f(t,Xt) be the price of a contingent claim. In Chapter 2, it is
explained how to reduce the problem of the calculation of f(t,Xt) to the
corresponding boundary problem for a generalised Black-Scholes equation
(dt + L-r))f(t,x)=0,
where L is the infinitesimal generator of a Levy process. It was mentioned

that the reduction is valid for any Markov process having absolutely con-
tinuous potential kernels.
Lemma 14.1 For any t > 0, exp[—ta(x, D)} is an integral operator with
the kernel of the class C°°.
The proof will be given in Chapter 16. Lemma 14.1 implies that we can
apply the reduction of Chapter 2 to NIG-like Feller processes, and deduce
the generalised Black-Scholes equation
(dt-(r + a{x,Dx))f{t,x)=0. (14.19)

14.3.2 Pricing of European options under NIG-like Feller

processes obtained via subordination
Let T be the expiry date and g(Xr) the terminal payoff. Let Qt be a
strongly continuous semigroup constructed as in Theorem 14.3. Then by
solving equation Eq. (14.19) subject to the terminal condition f(T,x) =
g(x), we obtain
f(t,x) = e-Tr(QTg)(x),
for r = T - t > 0. By applying Eq. (14.8), we find
f(t,x) = e-Tr f Psg{x)XT(ds). (14.20)

J[0,oo)
Thus, if P3 is the generator of a diffusion, we can obtain a pricing formula

for the contingent claim by finding first Psg, all s > 0, which differs by a
factor esr from the price </>(T — s, x) of the contingent claim with the expiry
date T and the same payoff g:
Psg(x) = esr<l>(T-s,x),
under a diffusion process, and then integrating e~rTPsg(x) w.r.t. the mea-
sure XT(ds) to obtain
f(t,x)= f e^-T>4>{T-s,x)\T{ds).
7(0,00)
Recall that one should choose the parameters of the model so that the
EMM-requirement
ex = e - r r J pae*v(ds), vr > 0 (14.21)

J[0,oo)
be met.
14.3.3 Pricing of European options under NIG-like Feller

processes obtained by the naive approach
The pricing formula is evident
f(t,x) = e-TrQTg(x), (14.22)
where QT is the transition semigroup of a NIG-like Feller process with the

generator — a(x, D); but in order to apply it, we need to obtain an explicit
formula for QT. That can be derived from the representation theorem for
analytic semigroups Theorem 15.18 similarly to Chapter 8, where the tran-
sition semigroup was computed for the case of an operator of a boundary
problem for a generator of NIG and some other Levy processes, in order
to compute the rational prices of barrier options.
In order that the RHS in Eq. (14.20) and Eq. (14.22) be finite, some
regularity conditions on g are needed, and these conditions are related to
the domain U in the definition of the generator of a NIG-like Feller process;
for instance, in ID when U is an interval (A_, A + ), it suffices to assume that
g is piecewise continuous and satisfies estimates
g{x) < Ce~"*x, V ± x > 0, (14.23)
with A_ < w_ < w+ < A + . In the case of European puts and calls, it
amounts to the restriction on A_,A + : A_ < — 1 < 0 < A + . Due to the
put-call parity, it suffices to calculate the price of the European put, for
which g(x) = (K — e x ) + , where K is the strike price. Thus, below, we
discuss the pricing of the European put.
We consider a NIG-like Feller process in ID with the generator — a(x, D),
where a is defined by Eq. (14.1), and n,5,a,(3 satisfy Eq. (14.6) and
Eq. (14.7). The EMM-requirement reduces to
r + a(x,-i)=0. (14.24)
If 5,a,f3 are already chosen, one can use Eq. (14.24) and Eq. (14.1) to find
the drift fi.
The construction of the transition semigroup starts with
Lemma 14.2 Let B be either the Sobolev space H"(R), where s € R, or

the Holder space C3(R), where s is a positive non-integer.
Then there exist C0 > 0 and 9 £ (7r/2,7r) such that if argX £ [—9,0],
then
|| A(A + r + a(x, D))'1 \\B^B< Co- (14-25)
Proof. Fix large C and consider first the region Hc,e — {^ I \M > C, arg A G
[-6,0]}. Since
a(x,£) ifj£ + S\£\, as£->±oo,

for C fixed, we can easily find 6 G (7r/2,7r) and C\ such that for all A 6 Sc,e
and all (ar.f) G R x (R +1{0}),
lACA + r + a f o O r 1 ! ^ ! . (14.26)
By using Eq. (14.26), and arguing as in the proof of Theorem 14.1 (cf. also
Theorem 16.12), we obtain that Eq. (14.25) holds for A G £c,e, w i t h some
Co independent of these A.
To prove that for C fixed, there exists 6 G (TT/2,TV) such that Eq. (14.25)
holds for A satisfying |A| < C and argA G [—6,9} it suffices to notice that
since — a(x, D) is a generator of the Feller process, r + Ua(x, D) is positive-
definite.
Remark 14.1. Since the boundedness theorem is valid for much more
general spaces, the estimate Eq. (14.25) and the formulas and estimates
below are also valid in these spaces.
The estimate Eq. (14.25) means that the representation theorem for the
semigroup with the generator —r — a(x, D) is applicable, and the following
formula is valid:
e x p [ - r ( r + a(x, D))]g = e~TTQTg
= (2iri)-1 f eXT(X + r + a(x,D))-1gd\. (14.27)
Here £# is the contour A — A(CT), —OO < a < +oo, where arg X(a) = —6, a <
0, argA(o-) = 0,er > 0.
By using an asymptotic expansion of the symbol R\{x,£) of the resol-
vent (A + r + a ( x , D ) ) - 1 :
Rx(x,0 ~ (A + r + a(x,0)"1 + ••• (14-28)
(for the proof and the full asymptotic expansion, see Chapter 16), we can
compute the RHS in Eq. (14.27) in the form of a series. Since
(27ri)" 1 / eAT(A + r + a ( a ; , 0 ) _ ^ A - e x p [ - r ( r + a ( x , 0 ) ] ,
JCe
the leading term in the asymptotic expansion of the RHS in Eq. (14.27) is
-foo+itr
exp[ia£-T(r + a(x,0)]5(0#.
/ •OO+tlT
where a G (A_, A + ) is chosen so that g is well-defined on the line !3£ = a.

Consider the put with the strike price normalized to 1: K = 1. We
musl? take a £ (0, A + ), and compute
To sum up: a formula for the leading term for the price of the European
put with the strike price normalized to 1 is
In Chapter 16, we give formulas for the next terms of the asymptotic expan-
sion Eq. (14.29), and show that the first omitted term is already fairly small:
if the symbol a(x, £) depends on parameters (e, d) and satisfies estimates
lag}(i, 01 < C a/ ,eMd 1 -l°l- ml,, < 1 'l' J l> (14.30)
for any a, (3, where Cap is independent of (e, d), then the first omitted term
in Eq. (14.29) admits an estimate via Cr2e, where C is independent of
(T, e, d), and the next term is less still: of order r 2 e 2 / d .
Notice that the leading term looks exactly as in the case of the option
pricing under Levy processes, only when we calculate f(t,x), we use the
characteristic exponent depending on x. Thus, the leading term can be
calculated as easily as in the case of Levy processes.
Remark 14.2. Consider the NIG-hke Feller process defined by the symbol
Eq. (14.1) with constant 8, parameters fi,a,P being bounded and having
bounded derivatives, and satisfying
a 0 := mina(:r) » f3° := max \P(x)\. (14.31)
Then we have Eq. (14.30) with e = 1 and d = «o-

If the parameters a, [3 and fi depend on a small parameter e > 0 in
such a way that any of their derivatives of order s admits an estimate
via Cses, where Cs is independent of e, then we have Eq. (14.30), and the
effective small parameter is e/d, i.e., the second omitted term in Eq. (14.29)
is very small. The first omitted term in Eq. (14.29) (see Eq. (16.40)) is more
involved than the leading term but since it is small - near expiry, very small
indeed - it can be calculated with relatively large error, which simplifies
the task of the development of appropriate computational procedures.
14.4 Discussion and conclusions
We have used two approaches for construction of a class of Feller processes,

generalizing the class of Levy processes, in particular the Normal Inverse
Gaussian Levy processes: via subordination, and by describing a class of
PDO to which the generators belong; the processes themselves have been
constructed by using the representation theorem for analytic semigroups.
We have shown that the class of JV7G-like Feller processes obtained by
the first approach is a subclass of the NIG-hke Feller processes obtained
by the second approach, and have discussed applicability of both types of
processes to option pricing.
A model for shocks based on the subordination approach may seem
more natural from the probabilistic viewpoint, but it may be not so easy to
fit it to the data so that the EMM-requirement Eq. (14.21) is met. So, our
observation that symbols of generators obtained under the two approaches
differ little if the derivatives of the coefficients of a subordinated diffusion
are small can be used to justify the usage of the naive approach. The EMM-
requirement for the latter approach is easy to satisfy, and one can obtain
approximate analytic pricing formulae, which are computationally as easy
(or difficult) as the pricing formulae for the corresponding Levy process.
The symbols and the pricing formulae differ little even if derivatives are
not small but Eq. (14.31) holds. Notice that empirical studies in, inter alia,
Barndorff-Nielsen and Jiang (1998) allow one to expect that Eq. (14.31)
usually holds, and hence the pricing formula Eq. (14.29) can be used.
One may use another construction of NIG-like processes: suppose, that
there is a coordinate system in which the process under consideration is a
NIG Levy. In this case, the natural question is what the properties of the
generator written in an initial coordinate system are, in particular, what
its symbol looks like; this question can easily be answered by applying the
change-of-variables formula in the theory of PDO.
Finally, we note that results, similar to the ones in Section 14.2-Section
14.3, can be obtained for other Levy-like Feller processes obtained as mod-
ifications of other RLPE.
Chapter 15
Pseudo-differential operators with

constant symbols
This Chapter aims at the systematic development of basics of the PDO-

machinery and the theory of function spaces in the simplest situation, when
symbols are independent of the state variable. Together with the Calculus
of PDO with variable symbols, which is constructed in Chapter 16, this
Chapter can be used to obtain analogues of results of Chapters 7 and 8
for barrier and touch-and-out options under Levy-like Feller processes con-
structed in Chapter 14, the multi-dimensional case including. The exposi-
tion is deliberately purely analytical; we aim to show what is the natural
order for the things to appear, from the point of view of Analysis, and how
one arrives at general formulas, which can be used in various applications,
without using the notion of stochastic integrals.
We also demonstrate how the Wiener-Hopf factorization method works
in more general situations than in the main text, for option pricing under
more general classes of processes in particular.
15.1 Introduction
A specialist with the PDE-background tends to regard the Black-Scholes

equation as a primitive object of Mathematical Finance, and the exposition
in Wilmott et al. (1993) is a model example of this approach. The price
of a contingent claim is defined as a solution to the Black-Scholes equa-
tion, which is specified by appropriate boundary and regularity conditions;
among those, is the specification of the behaviour at infinity. Since the
courses in ODE and PDE are standard, many specialists know the tech-
nique quite well, but the knowledge of PDO seems to be not wide-spread.
295
296 Pseudo-differential operators with constant symbols
The aim of the Chapter is to give a systematic exposition of basic facts

of the theory of PDO, which are needed for the generalization of the PDE-
approach to Gaussian Mathematical Finance to the case of non-Gaussian
Mathematical Finance. The basic theory presented below allows one to cal-
culate prices of contingent claims formally, without mentioning stochastic
integrals.
Local regularity properties are naturally formulated in terms of appropriate
spaces, like Cm and the Sobolev spaces Hm. In Gaussian theory, the spaces
Cm of non-negative integer indexes m arise, and in the case of processes
without the diffusion component, the generic case is non-integer m. The
corresponding spaces are called Holder spaces. The spaces Cm are more
natural from the point of view of applications to Mathematical Finance
(one prefers to know the price at each point, not in the sense of generalized
functions) but the study in the scale {Hm} is more simple since the Fourier
transform can be applied much easier. After the regularity of the solution
in terms of the scale {Hm} is established, the regularity in the scale {Cm}
can be deduced by using the Sobolev embedding theorem. It is possible to
work with the scale {C m } from the beginning till the end but this requires
additional conditions on symbols of PDO.
When the region is unbounded, the behaviour of the solution at the infinity
must be specified, and the standard (very convenient) device is to consider
the Generalized Black-Scholes equation in spaces with appropriate weights,
which capture the behaviour of the data and solutions. In the variable
x = In 5, the solution may grow exponentially, the main example being
the European call option with the payoff (ex — K)+, therefore the natural
weights are exponential.
The theory of PDO and action of PDO in function spaces is a complicated
construction, and when one tries to learn how to handle it, one feels that
there are must be easier ways. In simplest situations, like in Chapters 7 and
8, this is really so. It is possible to avoid the usage of the systematic theory
of function spaces and action of PDO in these spaces, when the results
are obtained in the explicit form, and hence the verification of necessary
regularity conditions is straightforward. Usually, this is possible only in the
case of PDO with constant symbols and domains with flat boundaries, when
the usage of the Fourier transform greatly simplifies the study of boundary
value problems.
As we saw in Chapter 14, in the case of PDO with variable symbols, the
Classes of functions 297
explicit solutions are not available, and this is a situation, where a general
theory can establish the existence and uniqueness of the solution in an ap-
propriate function class, prove the regularity of the solution, and obtain
effective approximate formulas for the solution. In Chapter 8, we demon-
strated a realization of this programme for pricing of barrier options in the
simple case of RLPE; this Chapter contains additional potentially useful
constructions and treats more general classes of symbols and spaces. If one
repeats the same constructions in the case of Levy-like Feller processes and
boundary value problems for contingent claims (and uses necessary results
of Chapter 16 on PDO with variable symbols, like in Chapter 14), one
obtains approximate pricing formulas.
We give a collection of main definitions and results of the theory of PDO
with constant symbols, without the connection to Probability, since this
connection was illustrated in many places in the main text. The exposi-
tion is essentially self-contained; of the reader, only the knowledge of basic
Functional Analysis and Complex Analysis is required. All necessary re-
sults from the theory of generalized functions and function spaces are listed;
in many cases, proofs of the results of the theory of generalized functions
are given. We omit some long proofs if the technical tricks in the proofs are
not used later, and the proof can easily be found in many monographs. We
basically follow the exposition in the monograph Eskin (1973), with some
modifications.
15.2 Classes of functions
15.2.1 General notation and remarks

We will work with spaces of functions on a finite-dimensional space V, which
we identify with R n by fixing a basis, or on a subset of R". Elements
of V — R n are denoted by x,y,z,...; elements of the dual space V are
denoted by £, 77, £, The standard scalar product in R " will be usually
denoted by xy. By using the scalar product, we can identify V and V and
denote the pairing (x, £) by x£. When we wish to distinguish R n = V from
R n = V', we write R"^ and R n j , respectively. J R n f(x)dx denotes the
Lebesgue integral.
For x 6 R", set (x) = (1 4- |x| 2 ) 1 / / 2 , and for multi-indices a,/? and a
function a on R n x x R" € , set a ^ ( x , f ) = dfD%a(x,£).
Estimates will be uniform on certain sets, usually on the domain of a

function; constants in these estimates will be denoted by C,c,Ci,ci etc.
We adopt the following policy on constants: unless otherwise stated, they
can be chosen the same for all values of parameters they can depend on,
and they are positive.
Cj" (U) is a subspace of Cm(U), consisting of functions uniformly bounded
on U, together with all derivatives; C™(U) is the subspace of bounded func-
tions vanishing at the infinity with all the derivatives up to the order m.
Sometimes we will use spaces of vector-functions assuming values in
normed spaces. For instance, if B is a normed space and K C R", then
C™(K; B) is the space of functions with the finite norm
SU
\\f\\c»(K;B) = £ P ll/ (Q) (*)lle.
|a|<m
This is a natural generalization of the definition of the space C™(K). Sim-

ilarly, one modifies the definitions of other spaces.
15.2.2 The space 5 ( R " )

The starting point of the theory of PDO is the definition of the action of
PDO in an appropriate space of sufficiently regular functions (test func-
tions), and the choice depends on properties of a class of PDO one works
with. First of all, it is advisable that the space of test functions be invariant
under the action of PDO from a given class. For differential operators on
R", the natural choice is Co°(R"), since for u € C™(U) and A, a differen-
tial operator with smooth coefficients, Au e Co°(R"). Pseudo-differential
operators being non-local, Co°(R n ) is not a good candidate, and for many
classes of PDO, the best choice is <S(R"), the space of infinitly differen-
tiable functions vanishing at the infinity faster any negative power of (x),
together with all derivatives. The topology in 5 ( R " ) is defined by a system
of seminorms
\\u\\s.a,0 = snp\x'3u^(x)\, (15.1)

ieR
Classes of functions 299
where a, @ are multi-indices. S ( R n ) is the Frechet space. For k, v € <S(R"),

define the convolution k * v by
+oo
k(x - y)v(y)dy.
/ -oo
By using the Leibnitz rule and the triangle inequality, it is easy to show that
k*v £ <S(Rn), and that a fixed k € <S(Rn) defines a continuous operator
K : 5 ( R n ) 3 v H4 k * v € <S(Rn).
K is called the convolution operator with the kernel k. In many appli-

cations, this definition is used for wider spaces and classes of kernels, for
instance, with Za(R n ) instead of 5 ( R n ) .
15.2.3 The Fourier transform

For u € «S(Rn), define the Fourier transform Tu = u by
fi(£) = f e~ix!>u(x)dx; (15.2)
then the inverse Fourier transform T~1 is defined by
u(x) = (27r)- n f eixiu(Od£. (15.3)
In the literature, one can find many versions of the definition of the Fourier
transform. For instance, Eskin (1973) uses Eq. (15.2) with ix£ in the expo-
nent (hence, Eq. (15.3) with — ix£), and the factor {2ir)~n can be shared be-
tween T and T~1 differently; but in all cases, if dx and d£ are the Lebesgue
measures in R ^ and R n £, respectively, the product of the factors in front
of the integral signs in Eq. (15.2) and Eq. (15.3) must be (27r) _n .
The following properties of the Fourier transform are well-known.
Theorem 15.1 a) T is a topological isomorphism ofS(Rn), and J7-1 is

its inverse;
b) HDVm = Zaf(0, andf(x"f){0 = (-ljWDf'/tf);
c;^(u*«)(o=«(o«(0;
d) the Parceval indentity holds:
( M W « ) = (2^) n («,t;)L a (R-); (15.4)

in particular,
HfillL(R-) = (2*)n|Mli2(R")-
The Fourier transform Eq. (15.2) is well-defined on a wider space L i ( R n ) ,
and by using the Parceval identity Eq. (15.4), T can be extended from a
dense subset <S(Rn) C Z<2(Rn) to a topological isomorphism of L 2 (R")- In
the next subsection, a natural generalization of Eq. (15.4) is used to define
the Fourier transform in the space of generalized functions. This construc-
tion has a disadvantage of being indirect; in some important cases, one can
obtain analogues of the direct formulas Eq. (15.2)-Eq. (15.3); the corre-
sponding procedures are considered in Subsection 15.3.3-Subsection 15.3.4.
15.3 Space S'(R") of generalized functions on R"
15.3.1 Generalized functions (distributions)

n
A functional (j> on <S(R ) is called linear if for any a i , a 2 € C and / i , / 2 €
5(R»),
(<t>,otifi +0:2/2) = QT(0,/I) + 0 2 ( ^ / 2 )•
The space of continuous linear functionals on 5 ( R n ) is denoted by «S'(R n ).

Its elements are called distributions or generalized functions. The topology
in S' (R n ) is defined by
/«->/ ^=> (fn,u)^(f,u)VueS(Rn).

Example 15.1 Let / be a locally integrable function, which admits a
bound
L \f(x)\(x)-Ndx < +c». (15.5)
Then we can define a functional (j> on <S(Rn) by

+00
f(x)u(x)dx, u G S(Rn).
/ -00
n
Clearly, (j> G »S'(R ). Functionals of this form are called regular. In the
sequel, the regular functional defined by the function / will be denoted by
the same letter, / .
Space S'(R n ) of generalized functions on R" 301
The following statement easily follows from the definition.
L e m m a 15.1
(i) If two functions, f and f\, define the same regular functional and are
continuous on an open set Q, then f(x) = / i ( x ) , V i 6 f i .
(ii) If f and / „ , n = 1,2,... are measurable, fn(x) -4 f{x) a.e., f satisfies
Eq. (15.5), and |/„(x)| < f(x), Vn,x, then / „ - > • / in <S'(R n ).
E x a m p l e 15.2 For 7 £ R n , the Dirac delta-function (57 defined by

(5 7 ,u) = u(j) belongs to <S'(R n ). When 7 = 0, the index 0 is usually
omitted: S = 60.
15.3.2 Operations in S'(Rn)

The derivative Daf of a distribution / is defined by
(Daf,u) = (/,Dau), V u e <S(Rn).
Clearly, it is a linear functional, and since Da is a continuous operator in

<S(Rn), Daf is continuous. Hence, Daf € <S'(R n ), and we conclude that
generalized functions have derivatives of any order.
E x a m p l e 15.3 Let 6+ = l(o,+oo) D e the indicator function of R + : 0+(x) =

l,x > 0, 9+{x) = 0,x < 0. Then
/•+00
(D9+,u) = (9+,Du) = / Du(x)dx = - m ( 0 ) ,
Jo
i.e., d6l+/dx = J. Similarly, (DkS,u) = {-l)kDku{0),k > 1.
E x a m p l e 15.4 If / and its derivatives up to order m = |a| are continuous
functions growing not faster than a polynomial at the infinity, then the
functional Daf is a regular functional defined by a function Daf.
Let a G C°°(R n ) satisfy, for any a:
\a^(x)\<Ca(x)m°, (15.6)
where Ca and ma are independent of x. For / £ «S'(R"), define the func-

tional af by
(a/,u) = ( / , a u ) , Vue5(R").
Clearly, af € «S'(Rn), and the-multiplication-by-a operator, with a satisfy-

ing Eq. (15.6), is a continuous operator in <S'(Rn). We conclude that
A(x,D)= £ aa(x)Da,
\a\<m
differential operators with C°°(Rn)-coemcients satisfying Eq. (15.6), are

continuous in <S'(Rn).
15.3.3 The Fourier transform of generalized functions

Motivated by Eq. (15.4), one defines the Fourier transform / = Tf of a
generalized function / G <S'(Rn) by duality:
(Ff,Fu) = (2n)n(f,u), VuG<S(R n ). (15.7)
Example 15.5 The Fourier transform of the delta-function 5 = 6Q is

equal to 1 since with / = 6, the RHS in Eq. (15.7) equals (27r)nu(0), and
from Eq. (15.3), with Tf = 1, the LHS is equal to
/ J F ^ l K = (2 7 r) n (F- 1 u)(0) = (27r)n^0).
Theorem 15.2 a) T is a topological isomorphism of «S'(Rn);

b) F(DZf)(0=taf(0, andf(xaf)(0 = (-l)|a|^/(0;
Proof, a) The continuity of J"in»S'(Rn) is straightforward from Eq. (15.7).
Define the action of T~x in <S'(Rn) by
{T~lf,T-lu) = (2w)-n(f,u), VuG 5(R");

then it is a continuous operator in <S'(Rn), and J:~1Tf — f for any / .
b) is immediate from the property b) of the Fourier transform in The-
orem 15.1. D
Thus, we have analogues of the statements a) and b) of Theorem 15.1 but
not of part c), where the convolution is treated, and the reason is that the
convolution / * g is not defined for arbitrary pair of generalized functions.
However, if the Fourier transform / G C°°(Rn) satisfies Eq. (15.6), we may
define the convolution / * g by
f*g = F-1(fg),
Space S'(R n ) of generalized functions on R" 303
and then
Hf*g) = fg-
15.3.4 The regularization of oscillatory integrals

Suppose that u, the Fourier transform of a function u, is found, but the
integral in Eq. (15.3) does not converge absolutely. If the limit
( ^ _ 1 u ) ( x ) = (27T)-1 lim / e f a C «(0df. (15.8)

A,B-t+oo J_B
exists, we can regard Eq. (15.8) as a regularization of the integral Eq. (15.3),
and use Eq. (15.8) to calculate u — T~x f (similar procedure can be applied
to the Fourier transform Eq. (15.2), of course). This regularization is still
within the scope of the basic Calculus. The next two regularizations of the
integrals Eq. (15.2) and Eq. (15.3) (understood as the mere notation for the
Fourier transform and the inverse Fourier transform, respectively) use the
technique of the theory of generalized functions. In both cases, we assume
that u is measurable and satisfies an estimate
|fi(OI<C(0', (15-9)
where C, s are independent of f. This is exactly the situation, which we
encounter in the next Section, where the solutions to pseudo-differential
equations will be found by means of the Fourier transform and its inverse.
Regularization 1. The generalized function u, the inverse Fourier trans-
form of it, can be defined by
(u.i;) = (2ir)-n{u,v)
= (2TT)-"/ fi(0«m
= (27r) _n f d£u{0 f dxeixtv(x)
= (2rr)- n f d£ f dxe-ixû(0vW),
JR" JR"
where the integral is understood as the iterated one. Take a positive integer
N, notice that
(0-2N(l-&x)Ne-ixS=e-ixZ,
where A is the Laplacian: —A = 5Z™=1 D$> a n ^ integrate by part. The

result is
(«,«) = (2TT)- B I dt[ dxe-ixtu(0(0-2N(l-Ax)N:^xj.

n
JR JR»
If 2JV > s + n, the last integral converges absolutely, and we may write
(u,v) = (2n)-n f f e-"«ti(0(0~2JV(l - A , ) " ^ ) ^ . (15.10)
The formula Eq. (15.10) allows one to explicitly calculate the pairing of
the generalized function u, given by Eq. (15.3), and any test function v €
<S(Rn). This defines u. It is easy to see that Eq. (15.10) with 2N > s + n
defines a continuous linear functional over <S(Rn).
Regularization 2. Take a sequence of uniformly bounded measurable
functions tpn, converging pointwise to 1. Then consider integrals
/„ = (27r)-" f f e-**il>n(du{CMx)dZdx.
JRn JRn
They converge absolutely, and by integrating by parts, we obtain
In = (27r)-» f [ e- f a tyn(0fi(fltf>" 2 J V (l - Ax)N^(xjdtdx.

JR" JR"
Take 2N > s + n; then from Eq. (15.9) we conclude that the integrand is
bounded by the same integrable function for all n, and by construction, it
converges pointwise. By applying the Lebesgue theorem, we conclude that
the limit limn_>oo In exists and is equal to the RHS in Eq. (15.10). Notice
that the RHS in Eq. (15.10) is independent of the choice 2N > s + n.
15.3.5 Tensor product of generalized functions

Set x' = ( x i , . . . ,xn-i) and x = (x',xn). Denote by S(R™,-1) ® <S(R In ) a
linear span of functions u = ui <g> u2 of the form u(x',xn) = u\{x')u2{xn),
where ux € 5 ( R n - x ) and u2 € S ( R ) . Clearly, -1
^(R", ) <8><S(RXJ is a
subset of <S(R").
For / i € «S'(R n _ 1 ) and f2 € S'(R), the tensor product / i ®f2 is defined
on monomials u = ui <g> it2 by
(/i®/2,w) = (fi,ui)(f2,u2),
and extended by linearity to 5(R", _ 1 ) ®<S(Ra;n).
Pseudo-differential operators with constant symbols on R n 305
It can be shown that 5(R", _ 1 ) ® ^ ( R z J is a dense subset in <S(Rn), and

/ i ® h admits a unique extension to a continuous linear functional / e
«S'(R n ). We will denote it by the same symbol / i ® fa.
Let v € 5 ' ( R n _ 1 ) and v be a regular functional. Then the Fourier transform
of w — v <S> S is equal to w(£) = £(£')•
15.3.6 Generalized functions on an open set. The support

of a generalized function
If / is a regular functional defined by a locally integrable function f(x), then
the support of / (denoted by supp / ) is the complement to the maximal
open set on which f(x) = 0 a. e. Let U C R n be an open set. S(U) denotes
the closure of C™(U) in <S(Rn). The space of continuous linear functionals
on S{U) is denoted by S'{U). Let / € S'{U). The functional fu £ 5'(£7)
is called the restriction of / on U if (/, u) = (/t/,u) for all u € S(U). The
restriction operator will be denoted pu, hence, fu = Puf- Since S(U) is a
closed subspace of <S(R"), any / e 5'(17) admits an extension If 6 S ' ( R n ) ,
by the Hahn-Banach theorem (clearly, an extension is non-unique).
We say that / equals zero on [7, if fu = 0. Since CQ°(U) is dense in
S(U), fu = 0 iff (/, u) = 0 for all u e C%°(U). Let U be the maximal open
set on which / = 0. Then the complement to U is denoted by supp / and
called the support of / . If / is a regular functional, this definition coincides
with the one given earlier.
15.4 Pseudo-differential operators with constant symbols

on R"
15.4.1 Definition of PDO with constant symbols

Let a be a measurable function satisfying
KOI<<7<0"V (15.11)
where C,m are independent of £. Then we write a € S' m (R n ). The-
multiplication-by-a operator is continuous in <S'(R n ), therefore the compo-
sition Jr~1aJr is a continuous operator in 5 ' ( R n ) . It is denoted by a(D),
and A = a(D) := !F~1aJr is called a pseudodifferential operator (PDO)
with the symbol a, of order m. The action of A = a(D) on <S(Rn) can be
defined by
Au(x) = {2n)-n f eixia(Z)u(Od£, (15.12)
or equivalently,
Au{x) = (27r)~n / df / dye«*-v)*a{t)u(y), (15.13)

JR" JR"
where the integral is understood as the iterated one.
Lemma 15.2 A is a PDO with the symbol a € Sm iff A is a continuous

operator in <S'(R"), which acts on the oscillating exponents as follows:
Aeixnf = a(-y)eixn>. (15.14)
Proof. Since the Fourier transform of elxy is equal to (27r)n<57, and a<57 =
a(7)(57, we have
T-xaTeix~< = T-Xa(j)(2v)n5y
= a(7)JP-1(27r)n<57
= a(i)eix"<.
This proves that A = a(D) acts on oscillating exponents by Eq. (15.14).

Now suppose that A is a continuous operator in <S'(R n ), which acts on
oscillating exponents e7(a;) = e%xl by formula Eq. (15.14), where 0(7) are
some constants. The integral Eq. (15.3):
u(x) = (27r) _n / e € (x)6(f)df

JR"
converges in the topology of <S(R"), therefore
Au{x) = (27r) _n / Ae 4 (x)u(£)d£,

JR"
and applying Eq. (15.14), we obtain Eq. (15.12). •
By using Eq. (15.14), it is straightforward to show that
a) Let A — ^2>.<maaDa,aa £ C, be a differential operator. Then A is a

PDO with the"symbol a ( 0 = E i a | < m ° « £ a -
Pseudo-differential operators with constant symbols on R™ 307
b) Let the Fourier transform of k € S'(Rn) satisfy Eq. (15.11). Then the
convolution operator K with the generalized kernel k is a PDO with the
symbol k.
15.4.2 The Sobolev spaces H"(Rn)

For s e R, denote by H"(Rn) the space of distributions / € S ' ( R n ) with
the finite norm | | / | | s defined by
11/112=/ 1/(0 W « . (i5-i5)

This definition implies that / is locally integrable.
Example 15.6 Since the Fourier transform of the delta-function is 1,

6eHs(Rn) iff s< - n / 2 .
For s e R, introduce the symbol A s (£) = (£)".
Lemma 15.3 For m,s 6 R, A S (D) : S(Rn) -> <S(Rn), and A S (D) :
<S'(Rn) -» S'(R n ) ore topological automorphisms, and A S (D) : i f m ( R n ) ->
i f m _ s ( R n ) is an isometry, with inverses defined by \~S(D).
Proof. Evident. •
In the following theorem, the main properties of the scale of the Sobolev
spaces are collected; the proofs can be found in, e.g., Triebel (1978).
Theorem 15.3 Let s < s1. Then

a) H°(Rn) = L2(Rn);
b) Hs(R.n) has the topology of the Hilbert space, with the scalar product
(u,v)s = (2n)-n [ (Z)2su{0W)dZ; (15.16)

JR."
c) C£°(R n ) C <S(Rn) C Hs(Rn) C <S'(R") densely, and these embeddings
are continuous;
d) Hs (R n ) C H"(Iln), densely and continuously;
e) H~s(Rn) can be naturally identified with the dual space (if*(R n ))*, by
(u,v) = {2«yn f u(0WW. (15.17)

Lemma 15.4 Let m > 0 be an integer. Then u G Hm{Rn) iff Dau G

L2(R n ), for all \a\ < m, and the norm \\-\\'s, defined by
(Ni;)2 = E IIÎILCR-). (15-18)

\a\<m
is equivalent to the norm Eq. (15.15).

Proof. Clearly, there exist C, c > 0 such that for all £ G R n ,
|a|<m
and by multiplying by |u(£)|2 and integrating over R n , we obtain
c\\u\\2s < E H£ a "lli 2 (R») < C\\u\\l (15.19)

|a|<m
Since all the terms are non-negative, we conclude that ||u|| s < oo iff
||Z) a u||| ,Rn* < oo, \a\ < m, i.e., the first statement of Lemma holds.
The second statement is immediate from Eq. (15.19). •
15.4.3 Action of PDO in the scale if"(R")
Theorem 15.4 Let m, s € R and a e S"n(Rn). Then
A = a{D) : H"(Rn) -»• H3-m(Kn) (15.20)

is bounded.
Proof. By using Eq. (15.12), Eq. (15.11) and Eq. (15.14),
\\Au\\2s.m = f (0 a ( - m ) |a(0| 2 |fi(0| a de
< c 2 f (£) 2 ( s _ m ) (0 2 m "(£)l 2 ^

= C2\\u\t D
Theorem 15.5 Let m, s € R, and let a G 5 m (R") satisfy

ntr1\<c(Q-m, (15.21)
Pseudo-differential operators with constant symbols on R" 309
where C is independent o/£ € R n . Then the operator Eq. (15.20), and the
operator
A:S'(Rn)^S'(Rn) (15.22)
are invertible, the inverse A~* being a PDO with the symbol a ( £ ) - 1 .
Proof. By construction, AA~l and A-1 A act as the identity, and from
Eq. (15.21) and Theorem 15.4, A " 1 : H"-m(Rn) -> H°{Rn) is bounded.
Hence, it is the bounded inverse of the operator Eq. (15.20). Since both
A, A'1 are continuous in S ' ( R n ) , the operator Eq. (15.22) is invertible as
well. D
A symbol a € S m ( R n ) satisfying Eq. (15.21) and the operator A = a{D)
will be called elliptic. Usually one calls a and A = a(D) elliptic if Eq. (15.21)
is satisfied outside some compact.
15.4.4 Elliptic equations

Consider an equation
Au = f,
where A = a(D) is a PDO with the elliptic symbol a G 5 m ( R n ) , and

/ 6 f f s - m ( R n ) . By applying A~x, we obtain a unique solution in Hs(Rn):
u = a(£>)- 1 /,
or
u(x) = (2ir)-» / ete«a(0-1/(0de.
If the integral does not converge absolutely, it needs to be regularized, e.g.,

as it was explained in Subsection 15.3.4.
15.4.5 Restriction to a hyperplane
Theorem 15.6 Let s > 1/2 and n = 1. Then the restriction-to-{0}-

operator
p': C£°(R) 9 u H-u(0) G C

admits a unique bounded extension p' : i f s ( R ) —• C.
The proof is essentially the same as the proof of the following theorem in
Eskin (1973), Theorem 4.2.
Theorem 15.7 Let s > 1/2 and n > 1. Then the restriction-to-the
hyperplane -xn = 0- operator
p' : C%°(Rn) 9 « H u(-,0) € ^ ( R " " 1 )
admits a unique bounded extension p' : HS(R) -> fl""-1/2(Rn-1).
15.4.6 The Sobolev embedding theorem
Theorem 15.8 (Eskin (1973), Theorem 4.3) Let s > r + n/2, where
r > 0 is an integer. Then Hs(Rn) C Co(R n ), densely and continuously.
15.4.7 Action in the Holder spaces

For a positive non-integer s, denote by C s ( R n ) the closure of 5 ( R n ) in
<S'(R n ), in the norm defined by
for s <E (0,1):
Hullo- = IM|L„ + sup sup |/ir s |u(a: + h) - u(x)\;

x€R" |/i|<l
for s = m + s', where m > 0 is an integer and s' € (0,1),
IMIc= E H^Hc-
\a\<m
These spaces are called Holder spaces. The Holder spaces constitute a part
of the scale of the Holder-Zygmund spaces {C s (R n )} s 6 R (for the definition,
see, e.g., Triebel (1978)), namely, for non-integer positive s, C s (R n ) =
C s ( R n ) 1 . The scale of the Holder-Zygmund spaces enjoys the standard
property C s (R n ) C C ( R n ) , if s > r, and the Sobolev embedding theorem is
valid for the Holder-Zygmund spaces (see, e.g., Triebel (1978)):
Theorem 15.9 Let s > n / 2 . Then for any s' < s - n / 2 , H9(Rn) C
Cs ( R n ) , densely and continuously.
^ o t e that for integer s > 0, neither C s ( R n ) nor C^(R n ) coincide with C"(R n )
Pseudo-differential operators with constant symbols on R n 311
To prove the boundedness of a PDO in the scale of the Holder-Zygmund

spaces, one has to impose additional conditions on the symbol. We formu-
late the boundedness theorem for the Holder-Zygmund spaces, under the
following fairly restrictive condition, which holds in many cases of interest.
Theorem 15.10 Let s,m e R, and let a € C°°(R n ) satisfy, for any
multi-index a,
|a (a) (0l<Ca<0 m " W . (15-23)
where Ca is independent of £.
Then A = a{D) : C s (R n ) -> C s _ m ( R n ) is bounded, with the norm
admitting an estimate
\\a(D)\\ < C V sup ( 0 - m + H | a ( a ) ( 0 l , (15-24)
|a|<JV«eR"
where C and N depend on n,s,m but not on a.
Proof. This is a special case of Theorem 16.7. •

Theorem 15.11 Let s, m € R, and let a satisfy Eq. (15.21) and Eq. (15.23)
Then A = a(D) : C*(R n ) -+ C s _ m ( R n ) is invertible.
Proof. By Theorem 15.5, A is invertible in S' (R"), with the inverse given
by A'1 = a ( D ) - 1 . Prom the Leibnitz rule, and Eq. (15.23) and Eq. (15.21),
it follows that a ( £ ) - 1 satisfies Eq. (15.23) with — m instead of m. By
Theorem 15.10, A : C S (R") -+ C s - m ( R n ) and A'1 : C 8 - m ( R n ) -> C s (R n )
are bounded, hence, they are mutual bounded inverses. •
It is straightforward to show That if Eq. (15.23) holds then a(D) maps

<S(Rn) into itself, continuously, and this is a topological automorphism, if
both Eq. (15.23) and Eq. (15.21) hold.
R e m a r k 15.1 a) Since Theorem 15.10 and Theorem 16.7 are valid for PDO
in the Besov spaces, Theorem 15.11 is valid for PDO in the Besov spaces
as well, and the same holds for theorems of Section 15.6 on PDO acting in
the scale of the Holder-Zygmund spaces.
b) To obtain the corresponding statements for PDO acting in the Holder
spaces, one must impose an additional condition: s, s — m > 0.
15.5 The action of P D O in the Sobolev spaces on R"±
o *
15.5.1 Spaces i f ( R n ± ) and PDO in a half-space
Denote by H ( R n ± ) a subspace of H"(Rn) consisting of distributions /

with supp / C R n ± .
Lemma 15.5 (Eskin (1973), Lemmas 4-2 and 4-3). H(R™±) is equal to
the closure ofC$°(Rn±) in H"(Rn).
Theorem 15.12 Let a £ S m ( R n ) admit the analytic continuation w.r.t.
£n into the lower half-plane r = 9 £ n < 0 (resp. r > 0) and satisfy an
estimate
M?,tn + t r ) | < C{\ + |£| + \r\)m (15.25)
for all £ € R n , and r < 0 (resp., r > 0 /

o s 0 s—m 0 s 0 s—m
Then a{D) :H ( R n + ) - > # ( R n + ) (resp., a(D) :H ( R n - ) - + i /
n
(R _),) is bounded.
Proof. Consider the case "+". By Theorem 15.4, a(D) : Hs(Rn) ->
m
f p - ( R " ) is bounded. By using Eq. (15.25) and the Cauchy Theorem,
one shows that for u € C£°(R n +), and v € C£°(R n _), (a(D)u,v) = 0.
Hence, supp a(D)u C R n + , and it remains to apply Lemma 15.5. •
For m G R, set
± W
I ((OT^n) m , ifn>l.
Clearly, A™ is holomorphic w.r.t. £„ + i r in the half-space ± r > 0, and
satisfies Eq. (15.25), hence
o * o *—m
Am(D) :ff(R" T ) - > # (R%)
is bounded. The same holds with —m instead of m, and A™A±m = 1;
therefore, it is an isomorphism.
15.5.2 Spaces H"(Rn±)

Let U C R n be an open set. One writes / G HS(U) if / is a distribution on
U, which admits an extension If S Hs(Rn). The norm in HS(U) is defined
The action of PDO in the Sobolev spaces on R " ± 313
by
||/lh.=inf||//||.,
where infimum is taken over all extensions If G H s ( R n ) . Set p± = PR.»±.
It maps # s ( R n ) onto H s ( R n ± ) .
Theorem 15.13 Let a G 5 m ( R n ) admit the analytic continuation w.r.t.
£n into the upper half-plane r = 9£„ > 0 (resp. r < 0) and satisfy an
estimate Eq. (15.25) for all £ G R n , and r > 0 (resp., r < 0).
Then the operator
Hs(Kn+) 3 u 4 p+a(D)lu € if s _ m ( R n + ) (15.26)
(resp., the operator
Hs(Rn_) 3 u H-> p_a(D)lu € tfs-m(Rn-))
is well-defined and bounded.
Proof. Consider the case "+". Let I'u 6 Hs(Rn) be another extension
o «
of u. Then supp (lu-l'u) C R " _ , hence lu-l'u G H ( R " - ) , and a(D)(lu-
o s-m
I'u) 6ff ( R n - ) , by Theorem 15.12. Hence, p+a(D)(lu - I'u) = 0, and
the operator Eq. (15.26) is well-defined. To show the boundedness, take lu
with the norm ||Zu||s < 2 | | U | | R " + ; S ; then from Theorem 15.4,
||p + a(D)iu|| R n + ; 8 _ m < ||o(D)Ju||,_ m < C\\lu\\B < 2C||u|| R « + ; .. Q
Denote by 5 ( R n ± ) the space of restrictions on R " ± of functions from

S ( R n ) . For any extension If € <S'(R") of / G «S(R"±), P+Dalf can be
identified with the classical derivative Daf G <S(R n ±). This justifies the
usage of the notation Daf instead of p+Dalf, for / G i f s ( R n + ) . Similarly,
for / G # S ( R " _ ) , Daf is used instead oip-Dalf.
Lemma 15.6 Let s > 0 be an integer. Then an equivalent norm in
Hs(Rn±) can be defined by
iMiil»±>, = E iP a «iiL(R- ± )- (15-27)

|a|<s
Proof. For any integer s, there exists an explicit construction of a linear

continuous extension operator Is : Hs(Rn+) -> Hs(Rn) satisfying
II''«II'.<C|MIR..
(see, e.g., Subsection 2.9.3 in Triebel (1978) or Lemma 4.5 in Eskin (1973)).
Hence,
iMiR«±,.<r«n;<ciMiR. ±.s-
The opposite inequality is valid for any extension operator: by Lemma 15.4,
the norms || • || s and || • \\'s in i ? s ( R n ) are equivalent, therefore
N i k * . . < lli*«ll'. < CtUi'uU. < c2\\u\\K±tS. ^
Denote by pa the-restriction-to-hyperplane-{x | xn — a) operator, and

by using the Lebesgue Theorem, one shows that if s > 1/2 and / €
Hs(Rn), then paf is a continuous vector-function of a taking values in
Hs-1/2(Rn~1). It follows that the operator p', denned on «S(R n ±) by
p'f(x') = lima±o Pa f{x')> admits a unique bounded extension p' : Hs(Rn+) -*•
ff»-1/2(R»-i).
15.5.3 The Cauchy-type integral and its decomposition
Denote by 0+ (resp., 0-) the-multiplication-by-l( 0) +oo)( x ) (resp., l(_ OO)0 )(x))

operator. Clearly, 9± : Z/2(Rn) -> L2(R n ±) is bounded.
Take r > 0; then
T{9+{xn)e-TX")(£n) = 1 / ( T + i&,),
and since F(uv) = (27r) -n u * v, we obtain for / 6 «S(R n ),
• F ( M * n ) e - " » / ) « U „ ) = &™)-1 [+°° /^Vn) dVn. (15. 28)
By using the Lebesgue theorem, we see that the limit of the LHS in Eq. (15.28)
is the Fourier transform of 6+f, and therefore,
F{0+f) = n + / , (15.29)
where
I I + / ( 0 := (2W)- 1 f ^ /^' V n )
drjn.
Similarly,
F(0-f) = U-f, (15.30)

The action of PDO in the Sobolev spaces on R n ± 315
where
+o
° ke,vn)
n-/(e = - M - / J&iO-n,
•dr]n.
By using Eq. (15.29) and Eq. (15.30), and a series of rather technical esti-
mates, Eskin (1973) deduces a series of important results.
Theorem 15.14 (Eskin (1973), Theorem 5.1) For \s\ < 1/2, 0± admits
a unique continuous extension 8± : HS(R") —¥H(R-n±)-
Lemma 15.7 (Eskin (1973), Lemma 5.4) For \s\ < 1/2, any function
o *
/ G Hs(Rn) admits a unique representation / = / + + / _ , where f± EH
(R"±), and f± = 6±f.
Define the extension-by-zero operator e± : <S(Rn±) —>• I/2(R n ) by
e±f{x)
= { 0, ±xn < 0.
Prom Theorem 15.14 and Lemma 15.7, we deduce
Lemma 15.8 Let \s\ < 1/2. Then e± admits a unique bounded extension
s n
e± : H (R ±) -> Hs(Rn), and Hs(Rn±) can be identified with # ( R n ± ) .
Theorem 15.15 (Eskin (1973), Lemma 5.5) Let f <E Hm+s(Rn), m > 0
integer, \s\ < 1/2. Then
m
9+f = £ AI*(D) [ ( p ' A ^ 1 (£>)/) ® So] + Al m (D)0 + A™(£>)/, (15.31)
fc=i
and
m
A
9-f = Y, +k(D) [ ( p ' A ^ 1 (D)f) ® So] + A+m(D)6_A™{D)f, (15.32)
jt=i
where p' is the-restriction-to-the hyperplane xn = 0 operator.

Theorem 15.16 (Eskin (1973), Theorem 5.2). Let s = - m + s', where
m > 0 is an integer and \s'\ < 1/2.
s
o * o
Then any f € # ( R n + ) n # ( R n _ ) admits a decomposition
m
f = Y,fk®Dk~ls°> (15-33)
where fk G ff*+*-1/2(Rn-1).
15.6 Parabolic equations
15.6.1 The Cauchy problem for the parabolic equation

We call a symbol a of order m > 0 strongly elliptic, if a is measurable and
satisfies the following two conditions:
KOI < C(0m, £eR n ; (15.34)

9M0 > co(Om-ô, £eR n ; (15.35)
where C, CQ > 0, ao are independent of £.
Consider an equation
(dt + a(Dx))f(x, t) = g(x, t), t> 0, (15.36)
subject to the initial condition
f(x,+0)=go(x). (15.37)
The (initial) boundary problem Eq. (15.36)-Eq. (15.37) can be solved by

various methods. For essentially each of these methods, it is convenient
(and for some constructions and/or steps of the proofs, even crucial) to
work with a satisfying Eq. (15.35) with non-negativeCTO-SO the first natural
step is reduction to the case ao = 0, that is, to the case when a satisfies
Eq. (15.34), and a condition
Ba(fl > Q)<0 m , (15-38)

where CQ > 0 is independent of £ € R n . The reduction is as follows: change
the unknown
f{x,t)=e«>tr°{x,t)
and the RHS in Eq. (15.36)
g(x,t)=e°°tg''°(x,t)
Then from Eq. (15.36)-Eq. (15.37), f"> is a solution to the problem
{at + aao{Dx))f0{x,t) = g°°(x,t), t> 0;

/<r 0 (z,+0) = g0(x),
Parabolic equations 317
where aCTo(£) := a(£) + a0 satisfies Eq. (15.34) and Eq. (15.38). Having
in mind the possibility of such a change of the unknown and the data, we
assume in the sequel that a satisfies Eq. (15.34) and Eq. (15.38).
15.6.2 Reduction to a family of ODE on R+

By making the Fourier transform w.r.t. x, we obtain a family of the Cauchy
problems on the half-line for ordinary differential equations on R+, param-
eterised by £ e R":
(ft+o(0)/K,*)=9(€,«), *>0. (15-39)
/(£,0)=3o(£), (15.40)
(here / denoted the "partial" Fourier transform w.r.t x). By solving the
Cauchy problem Eq. (15.39)-Eq. (15.40):
f(U) = f exp[(s - t)a(0}g(S,s)ds + exp[-ta(0]&(0, (15-41)

Jo
and making the inverse Fourier transform, we obtain the solution to the
initial problem Eq. (15.36)-Eq. (15.37)
f(x, t)= I exp[(s - t)a{Dx)]g{x, s)ds + exp[-ta{Dx)]go{x). (15.42)

Jo
Due to Eq. (15.35), for any t > 0,
Pt(x) = (2-ir)-1 I exp[ix£ - ta(£)}d£

JR."
is a continuous uniformly bounded function in x (in fact, pt G C°°(R n )),

and hence, we deduce from Eq. (15.42)
f{t,x)= / pt-s{x-y)g{y,s)dyds+ pt{x - y)g0(y)dy.

JO JR." JR."
In the case g = 0, which we mainly consider in applications to Financial

Mathematics, the last equation reduces to
f{t,x)= f pt{x-y)g0(y)dy. (15.43)

JR."
In order that the RHS in Eq. (15.41)-Eq. (15.43) be well-defined, g and go

must satisfy certain conditions, and one should specify in which sense the
limit in Eq. (15.37) is understood.
If go € Li(Rn) then Eq. (15.43) defines a continuous function on R n x R + .
Notice that so far we have the definition of the Fourier transform for gener-
alized functions from <S' only, and hence we cannot allow for exponentially
growing data. This is a serious drawback since in some important applica-
tions, go(x) does grow exponentially as x —> oo; this forces us to introduce
spaces with exponential weights in Section 15.9. Below, we restrict our-
selves to the case of g0 € H* (R n ) and g € I/2(Rn * R+)-
Extend both / and g by zero for t < 0, and make the Fourier transform
w.r.t. t:
fit, V) = fa + o(O)" 1 (9it, V) + 9b(0)-

Due to Eq. (15.34) and Eq. (15.38), a(f, rj) — ir) + a(£) satisfies an estimate
c(i+\er + \v\) < m,v)\ < c{\ + icr + w), (15.44)

where C, c > 0 are independent of £ G R n and 77 in the lower half-plane
$ST] < 0. If m = 1, the estimate Eq. (15.44) suggests that we can consider
the operator of the problem Eq. (15.36)-Eq. (15.37) in the scale of Sobolev
spaces; in the general case, the natural scale is the scale of anisotropic
Sobolev spaces.
For s,r e R, define H^[' (R n x R) to be the space of generalized
functions with the finite norm
\H\2mMs,r) = 1 1 (M + (0m)2S(02rm,V)\2^dr].
JR" JR
o(«-r) of
Spaces Hm,i (R n x R+) a r e defined by the analogy with H(R-n x R + ) , and
similarly to Theorem 15.12, we deduce from Eq. (15.44), that for any s,r,
aiD^Dt)-1 :ff„~11'P)(Rn x R+) -+H^(Rn x R+) (15.45)
is bounded. By making the change of variable r\ = r]'(^)m, is easy to verify

that for any e > 0,
/ (hi + <0m)-1-e|flb(0l2d»? < C(t)-™\9o(0\2,

JR
therefore if go £ Hs ( R n ) , where s' > 0, then for any e > 0, St=o ® 9o £

•ffm^ 2 - '' 8 ' ( R n x R ) , and since supp go ® <5o C R n x R + , we have, for
any e > 0,
0(-l/2-e,S')
9o ® <5o € H m , i (Rn x R+). (15.46)
If we assume for simplicity that g € J^2(Rn x

R + ) J then from Eq. (15.46)
and Eq. (15.45) we conclude that
/(*,*) = (27T)-1 / / ^^+t^(ir, + a^))-1m,v)+9o^))^dfi
o U/2-e,0)
belongs to Hm\ ( R n x R+)- This statement can be made more exact,
but even this form suffices to prove that / is well-defined as a general-
ized function. The initial condition can also be understood in the sense of
generalized functions: take any v £ 5 ( R n ) , multiply Eq. (15.41) by 6(£),
and integrate w.r.t. f over R n . By using Eq. (15.34)-Eq. (15.38) and the
Lebesgue theorem, it is not difficult to show that
lim (/(*,-),«) = (27T)"/ goiOmdt
= (9o,v).
Under additional conditions on the data, it is possible to show that /(£, x) ->
go(x) as t —¥ +0, a. e.
To prove that Eq. (15.37) holds in even stronger sense, and make more
exact statements on the regularity of / , the following method is more con-
venient.
15.6.3 Parabolic equation as ODE with the operator-valued

coefficient
Let g0 £ Hs(Rn). Denote by As the unbounded operator in H3(Rn) with
s+1
the domain H ( R n ) ; the action of As is defined by
Asu = a(Dx)u.
We can interpret the problem Eq. (15.36)-Eq. (15.37) as the Cauchy prob-
lem for the ordinary differential equation on R + with the operator-valued
coefficient. For the sake of brevity, we consider and solve it in the case
9 = 0:
f'(t)+A*f(t) = 0, t>0, (15.47)
/ ( + 0 ) = go- (15.48)
Theorem 15.17 For any sell andg0 G Hs(Rn), the problem Eq. (15.47)-
Eq. (15.48) has a solution f € C°([0,+oo);Hs(R+)).
The solution is unique, and for any to > 0,r € R , j € Z+, and 7 <
Co, where CQ > 0 is the constant in Eq. (15.38), it satisfies the following
estimate:
sup||e77(j)(t)||r<+oo. (15.49)
t>t0
The proof is based on the following definitions and theorem.

For 6 € (0,TT) and C > 0, set S e , c = {A € C | |A| > C, argA € [-0,6]}.
Set Tig — Il$fi-
Definition 15.1 Let A be an operator in the Banach space B, and let
there exist 9 e (0,7r), C > 0, and Ci such that for A G T>e,c, A 4- A is
invertible, and the resolvent satisfies the estimate
IKA + ^ m ^ C M l + IAI)-1. (15.50)

Then A is called a weakly 9-positive operator. If C = 0, A is called 6-positive.
Let £9,0 = 9Se,c be a regular contour, with a parameterisation A = A(CT)
satisfying arg A (a) — ±6 for a in a neighbourhood of ±00. Set Co = Cgfl-
The following theorem is a special form of the representation theorem
for the analytic semigroups (see Section IX.10 in Yosida (1964)).
Theorem 15.18 Let A be an unbounded operator in the Banach space
B, and let there exist 9 € (7r/2, IT) and C such that A is weakly 9-positive.
Then
a) for any g0 £ B, and any t > 0, the following integral is well-defined
exp(-tA)g0 = (2ni)-1 f etx(\ + A)-1g0dX; (15.51)

J C-9,C
b) Eq. (15.51) defines a strongly continuous semigroup {Tt}t>o in B, i.e.,

• To = / ;
• Tt = exp(-L4), t > 0;
• Tt+r = TtTTfort,r>0;
• for any t > 0 and go € B, lim T _^ TTgo = Ttgo in the topology of B;
c) f(t) = exp(—tA)go is a solution to the problem Eq. (15.47)-Eq. (15.48)

with the initial datum go G B.
To deduce Theorem 15.17 from Theorem 15.18, we need two auxiliary re-
sults.
Lemma 15.9 There exist Co > 0 and 9 G (n/2,n) such that if (A,£) G
E# x R , then
KX + aiOr'l^CodXl + iO"1)-1- (15-52)
Proof. If c > 0 is sufficiently small and (0™ < c|A|, then from Eq. (15.34)
and the triangle inequality we deduce
|A + a ( 0 | > ( | A | + ( 0 m ) / 2 ,
and Eq. (15.52) holds for these (f,A). Fix c > 0; then from Eq. (15.38),
there exists 9 e (w/2,ir) such that if |A| < c - 1 ( £ ) m and A € £<?, then
SRA > — 5fta(£)/2, and hence for some ci,C2 > 0,
lA + a ^ l ^ c x ^ r ^ c a d A l + ^ n .
Thus, Eq. (15.52) holds in this case as well. •
From Theorem 15.4 and Lemma 15.9, we conclude that there exist C\ > 0
and 9 6 (TT/2, it) such that if A e Tig, then Eq. (15.50) hold. Hence, Theorem
15.18 is applicable, and the solution to the problem Eq. (15.47)-Eq. (15.48)
is given by Eq. (15.51) with A = As.
By making the Fourier transform w.r.t. x, then changing the variable
A = in, and deforming the contour {n € C | in G Cg} into the real line
$sn = 0, we obtain that Eq. (15.51) gives the solution Eq. (15.43) to the
problem Eq. (15.36)-Eq. (15.37) with g = 0. In particular, for any go G
Hs(Rn), f[t) = exp[-tAs]g0 is of the class C 0 ( [ 0 , + o o ) ; # s ( R " ) ) , and /
satisfies the initial condition Eq. (15.48), where the limit is understood in
the sense of the topology of i f s ( R n ) .
It remains to prove the estimate Eq. (15.49). First, notice that if
we make the change of the unknown f(t) = e _ 7 t / 7 ( t ) in Eq. (15.47)-
Eq. (15.48), we obtain
/;(*) + ( A s - 7 ) / 7 ( t ) = 0, (15.53)
A(0) = 5o- (15.54)
Due to Eq. (15.38) and the condition 7 < Co, the problem Eq. (15.53)—
Eq. (15.54) enjoys the same properties as the problem Eq. (15.47)-Eq. (15.48),
and estimates Eq. (15.49) for / are equivalent to estimates Eq. (15.49) with
7 = 0 for / T . Hence, it suffices to prove Eq. (15.49) with 7 = 0.
By differentiating under the integral sign in Eq. (15.51) and using Eq. (15.52),
we obtain
+00
(l + l a D ' V ^ d a l l s o l U .
/ -00
For any to > 0, the integral converges uniformly in t > to, since 8 >
7r/2, and we deduce that Eq. (15.49) holds with r = s. By applying to
Eq. (15.47) a(D)-1, which maps Hs(Rn) to Hs+m(Rn) due to Eq. (15.42)
and Theorem 15.4, we find that
f(t) = -a{D)-lf'(t)
satisfies Eq. (15.49) with r = s+m. By induction, we obtain that Eq. (15.49)
holds with any r, and Theorem 15.17 is proved.
15.6.4 Parabolic equations and the Holder-Zygmund spaces

The method in Subsection 15.6.2 uses the Fourier transform, and hence
it is difficult to use it in other scales of spaces, e.g., the Holder-Zygmund
spaces, whereas the method in Subsection 15.6.3 based on the general rep-
resentation theorem for analytic semigroups can be modified quite easily.
In fact, we only need to know that a(D) acts in a given scale of spaces, and
its resolvent enjoys the estimate Eq. (15.51).
Lemma 15.10 Let a satisfy the estimates Eq. (15.23) and Eq. (15.38).
Then there exist 8 € (ir/2,n) and Cj,j > 0, such that for all A G £#, and
fell"
£ \D^((\+aior'my <cj(i+ 1X1)^, j = o,i,.... (15.55)
Proof. For j = 0, this is Lemma 15.9, and Eq. (15.55) for j > 1 can be
easily proved by induction, the Leibnitz rule, Eq. (15.50) and Eq. (15.23)
being used. D
For s G R, denote by As an unbounded operator in C s ( R n ) , with the

domain
V{AS) = {ue C s (R n ) I A*u := a{D)u € C s ( R n ) } .
It is possible to show that V{AS) = C s + m ( R " ) .

From Lemma 15.10 and Theorem 15.10, we obtain
Lemma 15.11 Let s,m G R, and let a satisfy the estimates Eq. (15.23)
and Eq. (15.38). Then there exist 6 G (7r/2,7r) and C such that for all
AGE*,
||(A + A'r'Wc-ô < C(l + |A|)-\ (15.56)
and by replacing in the proof of Theorem 15.17 the scale of the Sobolev
spaces with that of the Holder-Zygmund spaces, we obtain the Holder-
Zygmund analogue of Theorem 15.17.
Theorem 15.19 Let a satisfy the estimates Eq. (15.23) and Eq. (15.38).
Then for anysGK and g0 G C s ( R n ) , the problem Eq. (15.47)-Eq. (1548)
has a solution f G C°([0,+oo);C s (R n )).
The solution is unique. It is given by the RHS of Eq. (15.51), and
for any to > 0, r, j G Z + , and 7 < CQ, where CQ > 0 is the constant in
Eq. (15.38), the solution satisfies the following estimate:
sup||e7t/(j)(*)ll^ <+oo. (15.57)

t>t0
Notice that due to the remark after Theorem 15.10, the results of this
subsection are valid for any other scale of Besov spaces.
15.6.5 The inhomogeneous parabolic equation and the asymp-

totics of the solution as t —> + 0 0
Consider the inhomogeneous variant of the problem Eq. (15.47)-Eq. (15.48):
f'(t) + Asf(t) = g(t), t>0, (15.58)

/(0) = go- (15.59)
Theorem 15.20 Let a satisfy the estimates Eq. (15.23) and Eq. (15.38).
Then for any s € R, and functions g 6 C°([0,+oo);C s (R n )) and go e
C*(R n ), the following statements hold:
a) the problem Eq. (15.58)-Eq. (15.59) has a unique solution of the class
C°([0,+oo);C s (R n )), which is given by
f(t) = f T t _ T 5 (r)dr + Tt9o; (15.60)

Jo
b) fe C 1 ((0,+oo);C s (R"))nC°((0,+oo);C s + m (R n ));
c) if g(t) = g is independent of t, then f admits the representation
f(t) = fi + h, (15.61)
where
f2 = exp[-tAs](g0 - A);
d) for any i 0 > 0, positive non-integer r, j 6 Z+, and 7 < CQ, where CQ > 0
is the constant in Eq. (15.38), fo satisfies the estimate Eq. (15.57).
Proof. a)-b) Direct verification; the details are left to the reader as an
exercise.
c) Set f2(t) = f{t) - A . The problem Eq. (15.58)-Eq. (15.59) becomes
the problem Eq. (15.47)-Eq. (15.48) for f2 with g0 - fx £ C s (R n ) on the
RHS of Eq. (15.48). After solving this problem and applying Eq. (15.57),
we obtain c). D
15.7 The Wiener-Hopf equation on a half-line I
15.7.1 Statement of the problem

Let m > 0 and a € 5 m ( R ) . Given (generalized) functions g on R + and gr
on R_, consider a problem: find u satisfying
a(D)u(x) = g°(x), x > 0, (15.62)

r
u{x) = g (x), x<0. (15.63)
The Wiener-Hopf equation on a half-line I 325
Take any extension lgr of gT on R, and introduce / = u — lgr, g = g° —

p±a(D)lgr. In terms of / and g, Eq. (15.62) and Eq. (15.63) become
a(D)f(x) = g(x), x > 0, (15.64)

/(x) = 0 , x<0. (15.65)
Assume that g € i P ~ m ( R + ) , and interpret the problem Eq. (15.64)-

Eq. (15.65) as follows: find u € J f ( R + ) satisfying
p+a{D)u = g. (15.66)
This is the Wiener-Hopf equation. In order that the problem Eq. (15.66)
be well-posed, the symbol a must satisfy some regularity conditions.
Definition 15.2 We say that a e 5 ^ g ( R ) and call it a regular symbol of

order m, if the following conditions hold:
(i) a e 5 m ( R ) ;
(ii) there exists C > 0 such that
a(O#0, V£€R; (15.67)
(iii) there exist non-zero complex constants d± and p\ £ (0,1] such that
as £ —> ±oo,
a(0 = d ± K r + O(|€r-pi); (15-68)
(iv) there exist C > 0 and /92 € (0,1] such that for all £ € R,
|a'(OI<C<0 B, "' ,a - (15-69)

Notice that Eq. (15.67)-Eq. (15.68) imply that a is elliptic, i.e., Eq. (15.21)
holds.
15.7.2 The Wiener-Hopf factorization

The Wiener-Hopf method starts with the construction of functions a+ (£)
and a-(£), which admit the analytic continuation into the upper half-plane
^£ > 0 and the lower half-plane Qf < 0, respectively, and satisfy
a(0 = a _ ( 0 o + ( 0 , £eR- (15.70)

Example 15.7 Let
j=0
be a polynomial of degree m, with no real zeroes. Clearly, conditions
Eq. (15.67)-Eq. (15.69) are satisfied, and if {fj}i<j<ro is the set of zeroes
of a, then we may set
M O = am I I (£ - &)>
a-(o = n «-&•)•
In the general case, the construction is rather lengthy, so the reader must
be patient.
Choose d € C \ {0} and K0 € C so that the function
B 0 (O := d - J ( l + i O " ° ( l - i O _ " ° a ( 0 ( l + C 2 ) _ m / 2
satisfies
lim B 0 ( 0 = 1- (15-71)
£-+±oo
By taking the logarithm of B o ( 0 and using Eq. (15.68) and Eq. (15.71),
we obtain that d and KQ solve the system
ln(d±/d) ± i7TKo = 0.
(We have chosen the branch of the logarithm so that In a € R for a positive
a.) We see that d must be equal to
d = (d+d_) 1 / 2 = exppn(d+d_)/2], (15.72)
and therefore
K0 = iln(d+/d_)/27r. (15.73)
Due to Eq. (15.67) and Eq. (15.69), Bo is a continuous function, which does
not vanish on R, hence from Eq. (15.71), we conclude that the winding
number of the curve {Bo(0 I £ £ ( _ 0 ° , +oo)} around zero
+oo
dargB0(0 (15.74)
/
-oo
is an integer. Set
K_ = m/2 - Ko-l, K+ = m / 2 + Ko + l, (15.75)
call K_ the index of the factorization of the symbol a, and introduce
B(fl = (£-0'(£ + 0-,£o(O;

then
B(0 = (-l)'(l + i O ' ( l - i O - 1 B o ( 0 (15-76)

1 +
= (-l)'d- (l + i 0 - " - ( l - t O - " a ( O , (15.77)
and from Eq. (15.74),

+oo
dargB(£) = 0. (15.78)
/
-oo
Due to Eq. (15.78), b = \nB is well-defined on R, and by using Eq. (15.68),
we conclude that
fl(0 = l + O ( K r p i ) . as£^±oo; (15.79)
hence
6(0 = Odd""" 1 ), as f -> ±oo. (15.80)
For £ € R and <r > 0, set

r+oo
&+(£ + *") = -(27TZ)- 1 / " . / ^ <&?, (15-81)
6-(£-ia) = (27TZ) _1
/" °° 7 - ^ dry. (15.82)
J-bo i - w -i)
Due to Eq. (15.80), the integrals in Eq. (15.81)-Eq. (15.82) converge. Set
o+(0 = d(-l)'(l-i0K+exp[b+(0], 3£>0, (15.83)

a_(0 = (l + i C - e x p I M O ] , ^ < 0 . (15.84)
Theorem 15.21 Functions a± enjoy t/ie following properties:
(a+) a + (£) is holomorphic in the upper half-plane ?s£ > 0, and admits
the continuous extension up to the boundary $x£ = 0;
(b+) in the closed half-plane, a+(£) admits an estimate
c{\ + \t\f*+ < \a+(S)\ < C{1 + |£|)®K+, (15.85)

where C, c > 0 are independent of £ in the half-plane 9£ > 0;

(c+) for any r > 0, e > 0 and j = 1,2,..., fftere exists CTtij such that
for all £ in the half-plane 9£ > r ,
1 ^ ( 0 1 < Cr,eJ(l + | ^ | ) S K + - ^ + £ ; (15.86)
(d+) a+ admits a representation
a + ( 0 ± 1 = ^ ( - l j ' U - ^ ) ± , c + + *£(0, (15-87)
where t+ admits the following estimate: for any e > 0, there exists Ce > 0
SMC/I i/iat /or a// £ in i/ie half-plane 9 £ > 0,
1*1(01 < C e ( l + | e | ) ± S K + - p + £ , (15-88)
w/iere /> = min{/9i,p 2 } G (0,1];

(a_) a-(£) is holomorphic in the lower half-plane 9?£ < 0, and admits
the continuous extension up to the boundary 9£ = 0;
(b_) in the closed half-plane, a_(£) admits an estimate
c(l + |£|)SK- < | o _ ( 0 | < C ( l + |e|) R K -, (15.89)
where C, c > 0 are independent of £ m t/ie half-plane S£ < 0;

(c_) /or ani/ r > 0, e > 0 and j = 1,2,..., there exists CT,ej such that
for all £ in the half-plane Sr£ < —r,
| o ^ ( O I < CTttJ(l + K|)*"->1+<; (15-90)
(oL) a_ admits a representation
o_(C) ±1 = ( l + i O ± K - + * - ( 0 , (15-91)
where tt admits the following estimate: for any e > 0, there exists C( > 0
SMC/I that for all £ in the half-plane 3£ < 0,
|£(0I <C£(l + |e|)±SK-p+e; (15.92)
fej on £/ie reaZ axis, i?g. (15.70) holds;

(f) conditions (a±) - (6±) define factors in Eq. (15.70) uniquely, up to
scalar multiples, i.e., if a'± satisfy (a±) — (b±) and Eq. (15.70), then there
exists a constant c such that
a'+=ca+, a'_=c~1a-. (15.93)

Proof. First, we prove (o+) - {d+); (o_) - (d_) are proved similarly.
Since the function d(—1)'(1 — i£)K+ satisfies (a+) — (d+), it follows from
Eq. (15.83), that it suffices to prove the corresponding properties of b+,
namely
b+ (£) is holomorphic in the upper half-plane 9£ > 0, admits the continuous
extension into the closed half-plane and satisfies estimates Eq. (15.87) and
Eq. (15.88) with K+ = 0.
By differentiating under the integral sign in Eq. (15.81) and using Eq. (15.80),
we obtain, for £ in the half-plane 9£ > T > 0 and j = 1,2,...:
-t-oo
/ (l + M)-' 1 K-'?r 1 - i *J- (15-94)
-oo
From the triangle inequality, for rj e R and the same £,
(i+iiir'î.ir'i^-i (15.95)
therefore for any UJ 6 [0, p^, the RHS in Eq. (15.94) admits an estimate via
+oo
le-ur1-^"^. (15.96)
/
-oo
We may take any ui < 1, u € [0, pi}\ with this choice, for j > 1, the integral
in Eq. (15.96) converges, and we obtain that b+(£) satisfies Eq. (15.87) with
K+ = 0 .
Now we prove Eq. (15.88) for b+ with K+ = 0, in the half-plane 3f > 0.
Fix such £, and introduce the following subsets of R:
MO = { » j e R | | e - f j | > i , lfl<M/2},
MO = {TjeR||£-»/|>l, \£\>2\r,\},
MO = {77 G R | K - rj| > 1, N / 2 < |^| < 2|TJ|},
MO = {77GR-I | ^ - r ? | < 1}-
Then for j - 1,2,3,4 and 3£ > 0, define b+tj(0 by Eq. (15.81) with the
integration over Jj(£). Clearly,
4
Similarly to Eq. (15.94), we obtain for j = 1,2,3
\b+A0\ <C f (1 + M)-'11£ - ul" 1 *,. (15.97)

Jut)
For any a G [0,1], there exists Ca such that for any r\ G <7i(0,
\i - nr1 < ca(i + |£|)-*(i + |*,|)- 1 + a , (15.98)
and from Eq. (15.97) and Eq. (15.98), we deduce for j = 1, with arbitrary
e>0,
KAO\ < ce(i + mrpi+e f (i+\v\rl-cdr, (15.99)

< Ci,<(l + | f | ) - " l + ' .
Further, there exist C, C\ > 0 such that for any e G (0, pi) and all 77 G J2(£)>
(l + H ) - p i k - ' 7 | - 1 < C(l + \r,\)-^(l + \i\)-1 (15.100)
1 £ +
< c 1 ( i + M)- - (i + ^ | ) - ' " %
and from Eq. (15.97) and Eq. (15.100) we deduce Eq. (15.99) with j = 2.
In the case j = 3, the proof of Eq. (15.99) is different:
l(*+,3(oi < c a + iei)-"1 f /

(I + K - T , ! ) - 1 ^
• 1<IC->?I<3|£|
< C(l + | £ | ) - " ( l + l n ( l + K|))
< C£(l + |^|)-" 1 + e . (15.101)
Now we consider b+^(£). Notice that b+ti(£) = 0 unless £ = £ + ia, where
( 6 R and 0 < a < 1, and use the Lagrange formula and Eq. (15.67),
Eq. (15.69); we obtain for 77 G Ji{0>
.6(1,) = 6(C) + 61«, vKv-Q, (15-102)

where 61(^,77) admits an estimate
|6i(C,i7)l<C(l + | f | ) - « . (15.103)
By substituting Eq. (15.102) in Eq. (15.81) with the integration over J4(£),
and using the estimate
IMC, l)(ri - C)/(C - r, + i(7)| < C(l + |^|)-" 2 (15.104)

which follows from Eq. (15.103), and the limit
lim / (e-i?)- 1 *? = lim f (C + 0_1dC
= taln^1)"^' j
<r->+0 _(a-2+1)1/2 +
= in, (15.105)
we conclude that b+ i4 (£) satisfies the estimate Eq. (15.99) with p = minjpi,
P2} instead of pi.
We have shown that b+j ( 0 , .7 = 1,2,3,4, satisfy the estimate Eq. (15.99)
with p instead of p\, hence &+(£) also does, and to finish the proof of
(a + ) - (d + ), it remains to show that b+ admits the continuous extension
up to the boundary 3£ = 0. We have proved the uniform boundedness of
b+j(£), as Qf£ —>• +0, 6 + ^ ( 0 being represented as a sum of two terras. Of
those, the first one has the limit
-{2-Ki)-1b(C)i^ = 6(0/2
as a —¥ +0; by the inspection of the proof, it is seen that the integrands in
the integrals defining the second term and 6+,j(f) for j = 1,2,3 have limits
as a —¥ +0 and £ remains fixed. Prom the Lebesgue theorem, all these
integrals also have limits as a ->• +0. This finishes the proof of (a + ) - (d+).
We see that Eq. (15.70) follows from Eq. (15.76) and the following equal-
ity:
M 0 + 6-(0 = &(0(=inB(0), £eR- (15-106)

Prom the proof of (a+) — (d+) above and from its implied counterpart for
(a_) — (d_), we infer that as a —>• +0,
6±(C ± u r ) - • 6(C)/2 ± 6 ( 0 ,
where b is the same for both signs ± , and Eq. (15.106) is proved.
It remains to prove part (f). Suppose, there exists another pair o'±
satisfying (o±) - (b±) and Eq. (15.70). Then on the real axis
<(O/MO = M0/a'-(0,
and the LHS (resp., the RHS) admits a bounded holomorphic extension
into the upper (resp., lower) half-plane. Thus, we have a bounded holo-
morphic function in the whole complex plane. By the Liouville theorem, it

is constant, and Eq. (15.93) is proved. •
15.7.3 Main Theorems

Take any extension Ig € # * - m ( R ) and write Eq. (15.66) as
a(D)u = lg + g-, (15.107)

0s—m
where g_ €H ( R - ) - On the strength of Eq. (15.85), we can multiply
Eq. (15.107) by a + (Z?)~ 1 ; from Eq. (15.70), the LHS becomes a^{D)u, and
we obtain
a_(£>)u = a + ( J D ) - 1 ^ 4- ffi,_, (15.108)
0 a—R«_ 0 a—S/t_
where on the strength of Theorem 15.12, g ^ - £H ( R - ) , a-(D)u eH
( R + ) , and by Theorem 15.4, a^D)'1^ € # • - * " - (R).
Consider the following 3 cases.
Case I. s - 3?/c_ E ( - 1 / 2 , 1 / 2 ) . In this case, we apply Lemma 15.7 and
deduce from Eq. (15.108)
0 8 — 3?K_
a-(D)u = 0+a+(D)-llg GH (R+)-
By applying a-(D)~1 and using Eq. (15.89) and Theorem 15.12, we obtain
u = a_(£>)- 1 (9 + a + (Z?)- 1 ^ €fl"(R+). (15.109)
Theorem 15.22 Let s - ftrc_ G ( - 1 / 2 , 1 / 2 ) . Then the operator
P+a(D) :H(R+) -> ff-m(R+)

is invertible, with the inverse
H"-m(R+) BgûE H"-m(R+),
where u is given by Eq. (15.109).
Proof. We have already shown that Eq. (15.109) defines a bounded
left inverse. By applying to the RHS of Eq. (15.109) p+a(D) and using
Eq. (15.93), we obtain p+a+(D)9+a+(D)~1lg. On the strength of Lemma
15.7, this expression can be represented as
p+a+(D)a+{D)-1lg - p+a+(D)e-a+(D)-1lg = g.
0 s—RK_
The second term on the LHS is equal to zero since 0_a + (.D) -1 Zg £H
O 5 — 771
(R_), and hence, from Theorem 15.12, a+(£>)0_a + (£))- 1 /g £H (R-).

Hence, the map defined by
g M- a-(D)-H+a+(D)-Hg
is the right inverse. •
Corollary 15.1 Let s - !RK_ 6 ( - 1 / 2 , 1 / 2 ) , and let g in Eq. (15.66)

belong to i ? r ( R + ) , where r + K+ = q + s1, q is a positive integer and
s'e ( - 1 / 2 , 1 / 2 ) .
Then u €i/(R-(-), the solution to Eq. (15.66), admits a representation
g
u = ^2cj{lg)AZi (D)a^(D)-160 + AZqa-(Dy1e+Aq_(D)a+(D)-1lg,
3=1
(15.110)
where
Cj(lg) = (AL-1{D)a+(D)-1lg)(0). (15.111)
Proof. Apply Eq. (15.31) to Eq. (15.109). •
Case II. s - 3?K_ > 1/2.

In this case, we may regard / as an element of a wider space Hs - m ( R + ) ,
s s
O 0
where \s' — JJ/s_| < 1/2, find a unique solution in H (R+) DH(R+), and
after that check whether the solution defined by Eq. (15.109) belongs to
H(R.+)- If it does, we have a unique solution of the Wiener-Hopf equation
Eq. (15.66), of the class ^ ( R + ) , and if it does not, Eq. (15.66) has no
solutions of this class. In the latter case, one may introduce into Eq. (15.66)
potential operators in order to obtain a solvable problem (see, e.g., Eskin
(1973), Section 8), but we will not need this construction here.
Case III. s — 3t/c_ = —q+s', where q is a positive integer and \s'\ < 1/2.
Due to Theorem 15.4, A.-(D)-qa+{D)-1lg e # S ' ( R ) , hence we can
apply Lemma 15.7:
K^{D)~ia+{D)-llg = ^ + A_( J D)-«a + (D)- 1 Zp + /5_A_(Z?)-«a + (£»)- 1 ^.

(15.112)
Multiply Eq. (15.112) by A_(D)«:
a+(D)-llg = A^Dye+A-iDya+iDy'lg (15.113)

1
+ A_( J D)^_A_(D)-«a+(D)- Z 9 .
By inserting Eq. (15.113) into Eq. (15.108) and rearranging terms, we ob-
tain
a_(D)u - A _ ( £ » ) ^ + A _ ( £ » ) - ? a + ( ^ ) - 1 ^ (15.114)
, 1
= A_(D)^_A_(D)-'a+(£ )" ^ + 9i,-
By Theorem 15.12, the LHS in Eq. (15.114) belongs to H (R+), and

0 s — S?/c_
the RHS belongs to H ( R - ) , hence by Theorem 15.16, there exist
constants Cj, 1 < j < q, such that
a_(D)u - A_(D)«6»+A_(D)-«a+(I»)- 1 Zg = ^ T c ^ - 1
^.
By multiplying with a_(£>) _ 1 and leaving w alone on the left, we obtain

Theorem 15.23 Let s — 3JK_ = — q + s', where q is a positive integer
and \s'\ < 1/2.
Then for any g € Hs~m(R+), a solution to the Wiener-Hopf equation
Eq. (15.66) exists but it is non-unique, and the general solution is given by
Q
u = a^(D)-1A-(D)"e+A^(D)-qa+(D)-llg + a^(D)'1^2cjDj-160.
j=i
(15.115)
One of standard ways to obtain a well-posed problem is to add to Eq. (15.66)
appropriate boundary conditions (see, e.g., Eskin (1973), Section 8); for
our purposes, a more appropriate way to choose a unique solution is a
specification of the behaviour of the solutions to the Wiener-Hopf equation
Eq. (15.66), near the boundary. Possible types of the behaviour are studied
in the next Subsection.
15.7.4 The asymptotics of the solution near the boundary

To simplify the exposition, we restrict ourselves to a subclass of 5™ (R),
which will arise in our applications to Financial Mathematics, though the
general case can be treated in the same manner.

Let m G (0,2], a G S™ g (R), and let K± defined from Eq. (15.75),
Eq. (15.73) and Eq. (15.68) satisfy: satisfy
K_ G ( - 1 / 2 , 3 / 2 ) ; pj+K-> 1. (15.116)
Let g, the RHS in Eq. (15.66), be sufficiently smooth; for simplicity, g G

i f r ( R + ) , where r > 1/2 - K+, though smaller r (depending on m and K±)
may suffice. We look for a solution to Eq. (15.66) in if ( R + ) , where s = 0;
when convenient, we may (and will) take smaller s.
Theorem 15.24 Let m G (0,2), a G S™ g (R), Eq. (15.116) hold, g G

i f r ( R + ) , where r > 1/2 — K+, and consider a solution u to the Wiener-
Hopf equation Eq. (15.66), of the class L 2 (R+).
Then
I. In the case 9t«;_ G (1/2,3/2), the following three statements hold:
a) for any c G C \ {0}, there exists the unique u, which admits the
following representation, as x —> +0:
u(x) = cxK--x + Oix^+P-1-'), (15.117)
where x+ = max{x,0}, p = min{pi,p2}> and e > 0 is arbitrary;

b) there exists the unique u, which admits the following representation,
as x —> + 0 :
u{x) = T ( K _ + l)-1c1(lg)xK- + 0{x*K-+p-e), (15.118)
where p — min{p 1 ,p 2 } ) e> 0 is arbitrary, and
Cl(lg) = (a+(D)- 1 / f l )(0); (15.119)
c) any solution to the Wiener-Hopf equation Eq. (15.66) of the class

L 2 (R+) admits either the representation Eq. (15.117) or Eq. (15.118).
II. In the case 3?K_ G (—1/2,1/2], the solution u G L 2 (R+) exists; it is
unique and admits the representation Eq. (15.118).
Proof. I. Prom Eq. (15.115),
u - A-{D)a^.(D)'1e+AZ1(D)a+(D)-1lg + ca-iD)-1^, (15.120)
where c is an arbitrary constant. By choosing Ig € i f r ( R ) and using the

assumption r+K+ > 1/2, we see that the equation Eq. (15.31) (with m = 1)
is applicable to / = AZ1 (D)a+(D)~1lg e ffr+B++1(R), and hence we can

rewrite Eq. (15.120) as
u = ca^(D)-1S0+c1a-(D)-1AZ1(D)S0 (15.121)
1 1 1
+a-{D)- AZ (D)9+A-(D)a+(D)- lg,
where c is an arbitrary constant, and c\ = (a+{D)~1lg){0). Suppose, c ^ 0.

0-l/2-«
By using Eq. (15.91) and taking into account that 60 £H (R+)> V e >
0, we deduce from Eq. (15.121)
u = c(l + iD)~K-60 + V! + v2 + v3, (15.122)
where
O »K_+p-l/2-£
vi = ctZ(D)S0eH (R+), V e > 0 ;
0»K_+l/2-£
v2 = cia-{D)-1AZ1{D)S0eH (R+), V e > 0;
v3 = a-(D)-1AZ1(D)6+A-(D)a+(D)-1lgeH K +12
~ (R+).
If 3?i/ > 0, we have
^(a;+~le"x)(0 = / <Tixi~xxv-xdx = r(i/)(l + i 0 ~ " , (15.123)

./o
therefore
(1 + » D ) - % = r H - ' ^ - ' e - * . (15.124)
By using Theorem 15.9 and Eq. (15.116), we conclude that for any e > 0,
vi € C H K - + p _ 1 _ e ( R ) , and v2,v3 e C R K - _ £ ( R ) . Since they are supported
on [0, +oo) and p < 1, we conclude that Vj(x) = 0(xUK-+p~1"t), for
any e > 0. By applying Eq. (15.124) with v = K_ to the first term in
Eq. (15.122), we obtain Eq. (15.117).
This proves a), and to prove b), we set c = 0 in Eq. (15.121), and apply
Eq. (15.91) to the first remaining term on the RHS:
u = ci(J$)(l + iD)-K-+1S0 + vu + v3, (15.125)
where
0K-+p+l/2-e
vn = c1(lg)tZ(D)AZ1(D)80 £H (R+), V e > 0,
and v3 is the same as in the proof of a). The same argument as above shows
that
«ii(*) + v3{x) = 0 ( x S K - + p - e ) , x -> +0,
and by applying Eq. (15.124) with v = K _ + 1 to the first term in Eq. (15.125),
we obtain Eq. (15.118).
II. If 5R/c_ = 1/2, take s < 0 such that s - /c_ £ ( - 3 / 2 , - 1 / 2 ) ; then the
same argument as in the part I shows that the solution u £H(R+) admits
either the representation Eq. (15.117) or Eq. (15.118). Since x+ ' is not
square integrable in the neighbourhood of zero, Eq. (15.117) is impossible.
It remains to consider the case 3JK_ £ ( - 1 / 2 , 1 / 2 ) , when the solution is
unique and given by Eq. (15.109). By applying Eq. (15.31) to Eq. (15.109)
and arguing as in the part I, we arrive at Eq. (15.118). •
In the following corollary of Theorem 15.24, we list cases when the bounded
and continuous solution exists and it is unique.
Theorem 15.25 Let assumptions of Theorem 15.24 hold, and let u £

L 2 (R+) be a solution to the Wiener-Hopf equation Eq. (15.66).
Then a) if K_ £ (—1/2,0), and ci(lg) ^ 0, then any solution u is
unbounded as x -> + 0 ;
b) if K_ = 0, then a bounded solution u exists and it is unique. The u
is continuous outside the origin, and if c\ (Ig) ^ 0, it is discontinuous at 0;
ifc1{lg) = 0,u£Co(R);
c) if K- € (0,1), then a bounded u exists; it is unique and belongs to
Co(R);
d) if K- = 1, then a continuous bounded u exists; it is unique and
belongs to Co(R);
e) if K_ £ (1,3/2), then all u form a 1-dimensional lineal in Co(R).
In the next theorem, we consider cases when u £ Co(R), and study the
properties of the derivative u' near 0.
Theorem 15.26 Let assumptions of Theorem 15.24 hold, and let u £

L2(R+) be a solution to the Wiener-Hopf equation Eq. (15.66).
Then
a) if 0 < /t_ < 1, and c\{lg) 7^ 0, then u'{x) is unbounded as x —• + 0 ;
b) if K- — \, then u'(+0) = c\{lg) exists;
c) ifO < K- < 1 andci{lg) — 0, thenu'(+0) - 0 exists, andu £ C : ( R ) ;
d) if 1 < K_ < 3/2, there exists a unique u with bounded u'; it is given
by Eq. (15.118) and belongs to &(&).
Proof. Note that the behaviour of u' near the origin is the same as the
behaviour of (1 + iD)u = u + u', apply (1 + iD) to Eq. (15.121) and argue
exactly as in the proofs of Theorem 15.24 and Theorem 15.25. •
15.8 P a r a b o l i c e q u a t i o n s o n [0,T] x R +
15.8.1 The statement of the boundary problem

m
Let m > 0, and let a 6 S ( R ) satisfy Eq. (15.68) with Rd± > 0. Then a is
strongly elliptic, that is, Eq. (15.35) holds. Consider the following problem:
(dt+a(Dx))u{x,t) = g(t,x), x > 0, t > 0; (15.126)

u(0,x) = g0(x), x>0; (15.127)
r
u(t,x) = g (t,x), x<0,t<0. (15.128)
By changing the unknown
u{x,t) = eCT0'u<r°(x,f),
and similarly changing the data, we obtain the problem
{dt + aao(Dx))u°°{x,t) = /"(M), x>0,t>0; (15.129)

ffo
u (0,2;) = g0(x), x>0; (15.130)
ao r ao
u (t,x) = g > {t,x), x<0,t<0, (15.131)
where aCTo(£) := a(£) + ao satisfies Eq. (15.38). In particular, a„0 satisfies

Eq. (15.67), and if we assume that a satisfies Eq. (15.69), we can conclude
that aCT0 belongs to 5™g(R-) and satisfies Eq. (15.38). Having in mind
this possibility of changing the unknown and data, we assume that a itself
belongs to 5 r ^ g (R) and satisfies Eq. (15.38).
Suppose that gr is sufficiently regular, say, gT € C^fO,+oo);-fiT r (R_)),
where r > m (smaller r are also admissible), take any extension lgr €
C ° ( [ 0 , + o o ) ; F r ( R ) ) , and define
f(t,x) = u(t,x)-lgr(t,x), xGR, t>0;

r r
G(t,x) = g(t,x) + (lg )'t(x,t)+a(Dx)lg (t,x), x > 0, t > 0;
r
GQ{X) = g0{x) - lg (x), x > 0.
Parabolic equations on [0,T] x R+ 339
In terms of / , the problem Eq. (15.126)-Eq. (15.128) becomes
(dt + a(Dx))f(x,t) = G(t,x), x>0,t>0; (15.132)

f(0,x) = G0(x), x>0; (15.133)
f{t,x) = 0, x<0, i<0. (15.134)
As in Subsection 15.6.3, we consider the problem Eq. (15.132)-Eq. (15.134)

as ODE with the operator coefficient. This time, the latter is not an op-
erator in the whole space but the operator of the Wiener-Hopf equation,
and hence the resolvent must be studied by using the technique of Section
15.7. The technique needs some modification in order to obtain the uniform
estimate Eq. (15.50) for the resolvent.
15.8.2 The Wiener-Hopf factorization with a parameter

Let 6 be the same as in Lemma 15.9. We define d,l,Ko, and /c_ as in
Subsection 15.7.2, but change the definitions of Bo, B, b = In B, b± and a±
in order to be able to obtain estimates uniform in A € £#. We set
A^(A,0 = (l + | A | 1 / m ± i O s , (15.135)
B(X,0 := ( - l ) ' d - 1 A I K - ( A , 0 A ; K + (A,0(A + a(fl). (15.136)
Due to Eq. (15.52),
B(A,O?*0, V(A,0 6 S « x R , (15.137)
and clearly, for any A G S J ,
lim B(X,0= lim 5 ( 0 , 0 = 1, (15.138)
since 5(0,£) = B(£) satisfies Eq. (15.79). Therefore, the winding number
around zero of the curve {B(X, £) | - oo < £ < +00} is independent of
A e S 9 ; on the strength of Eq. (15.78),
+00
dargfl(A,0=0. (15.139)
/
-00
By using Eq. (15.137) and Eq. (15.139), we conclude that b = \nB is well-
defined on Hg x R.
It follows from Eq. (15.68), that uniformly in (A,£) 6 S j x R ,
l*(A,OI < ^(1 + \M + Kl m _ P l )/(l + \M + Klm)> (15-140)

and as £ -»• ±oo,
B(A,0 = l + O ( ( 0 - p i ) , (15.141)
with the constant in the O-term depending on A.

Set b = l n B , and for (A,£) e S j x R and a > 0, set
b+(X,£ + ia) = -(2m)-1 f °° * ( A , T ? ) dr,, (15.142)
M A , £ - »ff) = (27ri) _1 / °° / ( A , 7 ? ) dq. (15.143)
Due to Eq. (15.141), the integrals in Eq. (15.142)-Eq. (15.143) converge.

Set
o+(A,fl = d(-l)lAK++(X,Oexp[b+(X,0}, 5 ( > 0 , (15.144)

o_(A,0 = A^-(A,0exp[6_(A,0], ^ < 0. (15.145)
The following theorem is an analogue of Theorem 15.21; we omit parts

(c±) — (d±) since estimates in (d±), uniform in A, do not hold in general,
and we will not need the parameter-depending analogues of (c±) (and the
proof requires additional conditions on a).
Theorem 15.27 Functions a± enjoy the following properties:

(o+) V A 6 T,$, a+(A, £) is holomorphic in the upper half-plane 3£ > 0,
and admits the continuous extension up to the boundary G£ = 0;
(b+) there exist C, c > 0 such that for all A € £# and all £ in the upper
half-plane 3 £ > 0,
c((0 + \X\^mfK+ < M A , 0 1 < C((0 + lAI1/™)^; (15.146)
(a_) V A £ £#, a_(A,£) is holomorphic in the lower half-plane 9£ < 0,

and admits the continuous extension up to the boundary 9 £ = 0;
(6_) there exist C, c > 0 such that for all A € £# and all £ in the lower
half-plane 9 £ < 0,
c « 0 + \X\^mfK- < |a-(A,OI < C « 0 + lAI1/™)**-; (15.147)

Parabolic equations on [0,T] x R+ 341
( c ) / o r a / i ( A , 0 G £« x R,
A + o(0 = a+(A,Oo-(A,0; (15-148)
(d) for any A G £ 0 , conditions (a±),(6±),(c) de/ine a±(A, •) uniquely,

up to scalar multiples.
Proof. Parts (a±), (c) and (d) are immediate from parts (a±),(e) and
(/) of Theorem 15.21, and the proof of {b±) is a modification of the proof
of the same parts of Theorem 15.21. As in the proof of Theorem 15.21, it
suffices to show that 6± are uniformly bounded: there exists C such that
for all A G £# and all £ in an open half-plane ± 3 ^ > 0,
\b±(\,0\<C (15.149)
We prove Eq. (15.149) for the sign "+"; the proof for the sign "-" is essen-
tially the same. Let A e S« and ^ > 0. Set K(X) = (1 + \X\)1/m, and
introduce the following subsets of R:
•MA, £) = {n G R | |£ - 7,1 > K(X), |£| < 1,1/2},

-MA, 0 = {r, G R | K - 7,1 > K(X), |f | > 2|T,|},
M*, 0 - {l € R | |C - i7| > K(X), \rj\/2 < |£| < 2|i,|},
J4(A,0 - {r,€R\\t-ri\<K(\)}.
Then for j = 1,2,3,4 and Sf > 0, define b+J{X, f) by Eq. (15.142) with the
integration over Jj(X, Q. Prom Eq. (15.34), and Eq. (15.69) and Eq. (15.38),
we conclude that B(X, f) is uniformly bounded and bounded away from zero
on £9 x R:
C'1 <\B(X,0\<C, (15.150)
and 9,,&(A,rj) = dnB(X,rj)/B(X,rf) admits an estimate
\dvb(X, 77)| < C({0 + lAp/" 1 )-^, (15.151)
where C is independent of (A, f) G £9 x R . By using Eq. (15.140), Eq. (15.150)

Eq. (15.151) and repeating the proof of the boundedness of b+ in the
proof of Theorem 15.21, we obtain that 6j,+ are uniformly bounded, and
Eq. (15.149) follows.
15.8.3 Reduction to ODE with the operator-valued coeffi-

cient and the solution
We have not developed a theory of the Wiener-Hopf equation in the Holder
spaces, and the construction of such a theory is much more involved than the
construction of the theory for the Sobolev spaces. Besides, this construction
needs additional conditions even in the case of operators in the whole space
-see Eq. (15.23). This explains why here (unlike Subsection 15.6.3) we
work in the scale of Sobolev spaces. Notice however that after the solution
is found, and an explicit formula based on Eq. (15.51) is written, it is
possible to describe the properties of the solution in terms of the Holder
estimates like we did in Subsection 15.7.4.
Similarly to Eq. (15.58)-Eq. (15.59), we want to interpret the problem
Eq. (15.132)-Eq. (15.135) as
f'{t) + Asf{t) = G{t), t>0, (15.152)
/(0) = Go, (15.153)
where As is an unbounded operator in Hs-m(R+); the data Go € Hs-m(R+)

and G £ C°([0, +oo); / F - m ( R + ) ) , and the solution / is sought in the space
C°([0, +oo); Hs~m(R+)). An appropriate s can be chosen and A" can be
relatively easily defined if
3J/c±e[0,m]. (15.154)
In applications to Financial Mathematics, which we consider, this condition

is satisfied.
Take any s satisfying
s-me (-1/2,1/2), (15.155)
and for g 6 Hs-m(R+), set
(AT'g = a_(O,I»)- 1 0 + a + (O, J D)- 1 / 9 , (15.156)
where Ig € H s _ m ( R ) is an arbitrary extension of g. From Eq. (15.146)-

Eq. (15.147), and the assumption 9foc± > 0, it follows that a + ( 0 , D ) _ 1 is
_ 0s-m
bounded in H" m ( R ) , and a_(0, D) l is bounded in H (R-+); the latter
s m
space can be identified with H ~ {R+) in view of Eq. (15.155) and on the
strength of Lemma 15.8. Finally, 6+ can be regarded as a bounded operator
Parabolic equations on [0,T\ x R + 343
from H3~m(K) to Hs~m(R+), and therefore (A3)'1 can be regarded as a

s m
bounded operator in H ~ (R+). By applying p+a(D)e+ to (As)~lg, we
obtain
p+a+(D)0+a+(D)-1lg = p+lg = g
(for more details, see the proof of Theorem 15.22), hence the kernel of
{A")'1 is trivial, and we can introduce A" as the left inverse to ( A 8 ) - 1 .
Thus, A3 is the operator in H"~m(R+) with the domain
V(A°) = {(AaYxu |u€ tfs-m(R+)},
the action being defined by p+a(D)e+.
Lemma 15.12 There exists C such that for all A € E#,
||(A + A s ) - 1 | | ^ - m ( R + ) ^ ^ - " . ( R + ) < C ( l + IAD"1. (15.157)
Proof. The same argument as above shows that
(A + A s ) - 1 5 = a_(A,D)- 1 ^ + a + (A,Z?)- 1 Z 5 , (15.158)
where Ig € i f s _ m ( R ) is arbitrary extension of g € f f s - m ( R + ) . From

Eq. (15.146)-Eq. (15.147) and Eq. (15.154) it follows that there exists C > 0
such that
1
|(1 + l A l f ^ M A ^ r ! <C, V A € E*,3£ > 0,
and
|(1 + | A | ) H K - / m a _ ( A , 0 _ 1 | < C, V A € E , , 9 £ < 0,
and we conclude that
(1 + |A|) w * + / m a+(A,D)- 1 : F s ~ m ( R ) -»• Hs"m(R) (15.159)
and
0 s—m 0 s—m
(l + I A D ^ - ^ a - t A , ! ) ) - 1 ^ (R+)-^fl- (R+) (15.160)
are bounded uniformly in A 6 E#. Using the identification of spaces

s-,n
if (R+) and i f ( R + ) , the equality K+ + K_ = m and Eq. (15.158)-
Eq. (15.160), we obtain Eq. (15.157). D
With the estimate Eq. (15.157) at hand, we can use Theorem 15.18 with
B = Hs_m(R+), define Tt = exp[-L4 s ] by Eq. (15.51) and obtain analogues
of Theorem 15.19 and Theorem 15.20. We state the inhomogeneous version.
Theorem 15.28 Let a satisfy conditions Eq. (15.34)-Eq. (15.38),

Eq. (15.67)-Eq. (15.69) and Eq. (15.154), and let s satisfy condition
Eq. (15.155).
Then for any G G C°([0,+oo);i7*- m (R+)) and G0 € H°-m(R+), the
following statements hold:
a) the problem Eq. (15.152)-Eq. (15.153) has a unique solution f G
C°([0,+oo);Hs-m(R+)), which is given by
f(t)= f Tt.TG(r)dr + TtG0; (15.161)

Jo
b) if s- 5te_ < 1/2, then
f e ^ ( ( O , +oo); H°-m(R+)) n C°((0, +oo); tf (R+));
c) if s — 3te_ > 1/2, then for any e > 0,
/ G C1((0,+oo);tfs-m(R+)) nC°([0,+oo);^1/2+SK--€(R+));
d) for any to > 0, j £ Z + , and 7 < CQ, where CQ > 0 is the constant in
Eq. (15.38), fi satisfies the estimate
sup ||eT*/O)(0|| ff .-m(R +) < +00. (15.162)

t>t„
15.9 P D O in the Sobolev spaces with exponential weights,

in I D
In this Section, real A± and u± are assumed to satisfy A_ < 0 < A+ and
w_ < 0 < UJ+.
15.9.1 Generalized functions
Definition 15.3 «S(R; [A_, A+]) denotes the space of C°°-functions such
that for any 7 G [A_, A + ], elxu G <S(R). The topology in <S(R; [A_, A+]) is
PDO in the Sobolev spaces with exponential weights, in ID 345
defined by a system of seminorms
ll«l|5;[A_,A+]i.,JV = sup \x° DN (ex+x + ex-')u(x)\,

zeR.
where N, s > 0 are integers.
5'(R; [—A+, —A_]) denotes the space of continuous linear functionals in
S(R;[A_,A+]). Its elements are called distributions or generalized func-
tions.
Definition 15.4 For 7 € R and a regular function u, u 7 is defined by
u 7 (x) = e 7X u(x). For 7 € [ - A + , - A _ ] and u e 5 ' ( R ; [ - A + , - A _ ] ) , define
u7e<S'(R;[-A+-7,-A_-7])by
(u 7 ,u) = (u,u 7 ), VuG«S(R;[A_+7,A++7]),
and (ex~x + e A + x ) _ 1 u e «S'(R) is defined by
((ex~x + ex+x)~lu,v) = (u, (ex~x + e***)-^) , V v € 5(R).
Lemma 15.13 a)lf\^=0 = A+, then 5(R;[A_,A+]) = 5 ( R ) .

6J The-multiplication-by- ( e A + x + e A _ x ) operator establishes a topological
isomorphism between <S(R; [A_, A+]) and <S(R) (and a topological automor-
phism o/C£°(R)j;
cj Tfte-muftiph"caiion-&i/-(e A+x +e A - 1 ) -1 operator establishes a topolog-
ical isomorphism between <S'(R; [—A+, — A_]) and <S'(R);
d) For 7 € [A_, A+], the map
5 ( R ; [A_, A+]) 9 w »-> u 7 € <S(R; [A_ - 7, A+ - 7])
and for any 7 € R, the map
are topological isomorphisms;

e) For 7 € [—A+, — A_], the map
S'(R; [-A+, -A_]) 9 u K> u 7 6 S'(R; [-A+ - 7, - A _ - 7])
is a topological isomorphism;
f) Co°(Ii) is embedded into S(R; [A_, A+]), densely and continuously.
9) If X- < A_ < A+ < A'+, then <S(R; [A'_, A'+]) C S(R;[A_,A+]),
densely and continuously; in particular, S(R; [X'_, X'+]) C «S(R), densely
and continuously;
h) If\'_ < A_ < A+ < X'+, thenS'(R; [-X+, -A_]) C S'(R; [-X'+, -X'_]),
densely and continuously.
Proof, a), b) and d) are immediate from the definition, and c) and e)
follow from b) and d), respectively, by duality.
f) follows from b) and Theorem 15.3, part c).
g) Observe that
and use the definition of 5 ( R ; [A'_, X'+ ]).

h) The embedding follows from g), by duality, and the denseness follows
from f). •
From g), the Fourier transform u of u G «S(R; [A_,A + ]) C «S(R) is well-

defined, and it is straightforward to show that
a) u(£) admits the analytic continuation into the strip Q£ G [A_,A+], and
the continuous extension up to the boundary;
b) for each r e [A_,A + ], u{- + IT) 6 <S(R), and u 6 C°°([A_, A + ];5(R)).
Definition 15.5 For u G <S'(R; [—A+, — A_]) and 7 G [—A+, —A_], define
u(- + n) eS'(R) by
(u(- + ij), v(- - ij)) = 2TT(U, V), V v G 5 ( R ; [A_, A+]).
15.9.2 The Sobolev spaces HS^(R)
Definition 15.6 Let s and 7 € R. We say that a distribution u €

«S'(R; [-|7|, |7|]) belongs to if s - 7 (R) if and only if u 7 G i f ' ( R ) . The norm
in tf ^ ( R ) is defined by
IMI.,7 = IKII».
where || • || s is the norm in the Sobolev space HS(R).
Clearly, u(- + fy) = u^, therefore u € # S , 7 ( R ) if and only if the norm
is finite.
I*i7 a r +oo+i7
+00+17
-00+17
-00+17
(l + k | 2 ) s K 0 | 2 ^
\N 1X '/ *2
/
(15.163)
If u T G L i ( R ) , one can calculate u(£) on the line 9£ = 7 explicitly: for
-t-00
e _il "u 7 (a;)da;
/ -00
/+00
e- ia "'e' 1 ' x u(x)dx
-00
/+00
e - "*u(a;)dx,
-00
which is Eq. (15.2), the only difference being that in Section 15.2, £ is
assumed real.
Example 15.8 Let u{x) = 9+(x)e0x. Then the Fourier transform u(£)
is well-defined for 3 £ < -K/3:
f+co
u(£) = / e-êP'dx = - ( 0 - tO"1,
JO
and it follows that u G H S ' T (R), if and only if s < 1/2 and 7 < -3fy?.
For s 6 R, and A > 0, define \'X{D) as the PDO with the symbol (1 + A2 +
£2y/2
Lemma 15.14 Let m,s G R, p G R, 7 G [A_,A+] and A > max{|A_ —
p|> •*+-/>}•
Tften
A'x{D)eP* :5(R;[A_,A+]) -+<S(R; [A_ -p,X+ - p]), (15-164)
is a topological isomorphism.
Proof. On the strength of Lemma 15.13 d), it suffices to consider the

case p = 0; since A^(D) and A^ S (D) are mutual inverses as operators in
<S(R) D <S(R; [A_, A + ]), it remains to show that for any s, and any integers
fc,7V>0,
MA±;fc,jv;s := sup \xkDNex±xA°x(D)u(x)\

xGR
admits a bound via a linear combination of terms of the form Mx±]k',N';Q-

Since A > max{A+, — A_}, and A^(£) is holomorphic in a strip Q£ G
[A_, A+], we have
eT\°x(D) = Al(D + i1y*,
for any 7 € [A_,A+]. Hence,
Mx±;k,N;, = sup \xkDNA{{D + iA ± )e A±x u(a;)|. (15.165)

x€R.
By Lemma 15.3, Asx(D + iX±) is continuous in <5>(R), which implies that the
RHS in Eq. (15.165) admits an estimate via a linear combination of terms
of the form Mx±-k',N';o- D
Corollary 15.2 Let conditions of Lemma 15.14 hold, and let
epxAax(D) : <S'(R; [-A+ + p, - A _ + p)) -)• S'{R; [-A+, -A_]) (15.166)
be defined by duality: foru € «S'(R; [—A++p, —A_+p]) andv € <5(R; [A_, A+]),
(e"xAUD)u,v) = (u,A{(D)e"xv).
Then the operator Eq. (15.166) is a topological isomorphism.

L e m m a 15.15 Let m,s € R , 7 , 7 - p G [A_, A+] and A > |7 — p\.
Then
A\(D)epx : i? m - T (R) -+ ZP"- S ' 7 -"(R)
is a topological isomorphism, with the inverse defined by e~pxA^s(D).
Proof. By definition,
tfm'7(R) 3«H«7e HS(R)
is an isometry, hence it suffices to prove that
eh-p)*\sx(D)e(p-r)x = A« (D + i ( 7 - p)) : Hm(R) -> Hm-S(R) (15.167)
is a topological isomorphism. Notice that the equality in Eq. (15.167) holds

since A > I7 — p\, and on the strength of the same condition, there exist
C, c > 0 such that
c(0° < |AX(£ + i( 7 - P))| < C(0S, n e R,

and by applying Theorem 15.5, we conclude that the operator Eq. (15.167)
is a topological isomorphism. D
Theorem 15.29 Let s < s', and A_ < 7 < A + . Then

a) ff°'7(R) = L 2 (R.;e 27 da;), and HS'°{K) = H°{R);
b) H S ' 7 (R) has the topology of the Hilbert space, with the scalar product
-t-oo
/
(n)2su(n + »7)ii(fj + n)dn; (15.168)
-OO
S
c) C0°°(R) C <S(R;[A_,A+]) C H (R) C <S'(R; [-A+, -A_]) densely, and
these embeddings are continuous;
d) i f s ' 7 ( R ) C i f s ' 7 ( R ) , densely and continuously;
e) i ? - s ' ~ 7 ( R ) can be naturally identified with the dual space (if s ' 7 (R))*,
by
+00+I7
w(t)v(Z)<%- (15.169)
/
-00+17
Proof, a) Evident.
b) We can define (u,v)s,-y = (u 7 ,w 7 ) s , and then
+00
/
(Ti)2s^(T))€y (n)dTj
-00
+00
/
{T})2SU(TJ + 17)1/(77 + i~f)dn.
-00
c) By using topological isomorphisms defined by the multiplication-by-
e7X operator, we reduce to the case s = 0. After that we apply part c) of
Theorem 15.3 and parts f)-h) of Lemma 15.13.
d) Lemma 15.15 with s = 0 and part d) of Lemma 15.13 reduce to the
part d) of Theorem 15.3.
e) Set A = |7|, and introduce an equivalent norm in HS(R) by
H i ; = ||AX(D)«||0.
The corresponding norm in i ? s ' 7 (R) is
||u||;, 7 = \\Asx(D)u,\\0,
and the scalar product is
(u,v)' = {Als(D)uy,Vl)0.
Since # * ' 7 ( R ) is the Hilbert space, any <j> G (if s ' 7 (R))* can be denned by
some u G # S ' 7 ( R ) :
(</>, v) = (u, u) g>7

= (Aj'(D)u 7 ,t; 7 )o
= (e 7I Al; s (£>)e 7X u,t;)
= (AJftD + fryJe2^.*),
and by the same calculation, for a fixed u 6 i f s ' 7 ( R ) , the RHS defines a
continuous linear functional in i f s , 7 ( R ) .
Hence,
A\S(D + ry)e 2 7 1 : ffs'7(R) -4 (tf S - 7 (R))*
is an isomorphism, and the same argument as in the proof Lemma 15.15

shows that
A\S(D + i 7 ) e 2 7 1 : HS'7(R) -> # ~ S ' - 7 ( R )
is a topological isomorphism. This proves that (if s ' 7 (R))* is isomorphic

to tf-s--7(R). To finish the proof of e), take w € i f _ s ' _ 7 ( R ) and v €
i / s , 7 ( R ) , and consider
(w,v) = (iy_7,i;7)
+oo
/
-oo
+00-M7
/ w(S)H0dt.
-00+17 •
The following lemma and theorem are evident corollaries of Lemma 15.4
and Theorem 15.6, respectively.
Lemma 15.16 Let s > 0 be an integer, and 7 G R. Then an equivalent
norm in ff s ' 7 (R) can be defined by
i«7= £ II« 7 ^°«IIL(R)-
Theorem 15.30 Let s > 1/2 and 7 G R. Then the restriction-to-{0}-

operator
p':C0°°(R)9uH«(0)6C
admits a unique bounded extension p' : iJ s , 7 (R) —> C.
15.9.3 PDO with the symbols holomorphic in a strip, and

their action in the scale H ' , 7 ( R )
Definition 15.7 Let m and w_ < OJ+ be real, and 0 G [w_,a;+]. We

write a € 5 m ( R + i[w-, w+]) if the following conditions hold:
(i) a(£) is holomorphic in the strip 9£ G (u>_,w+);
(ii) a is continuous up to the boundary of the strip;
(iii) there exist C > 0 such that for all £ in the strip 9£ € [w_,u; + ],
K0I < C(0m- (15.170)

If Eq. (15.170) holds for £ in the half-plane 3£ < UJ+ (resp., 3£ > <j_), we
write a G 5 m ( R + i(-oo,o; + ]) (resp., a G S m ( R + i[u;_,+oo))).
For m , s G R, 7 € [W_,CJ + ], and a G S m ( R + i[w_,w+]), define
a(D) : # S ' 7 ( R ) -> tfs-m'7(R)
by
a(5)«(0 = a(Ofi(0, ^ = 7;
if a(- + ry)u(- + 17) G I a ( R ) , we can use the explicit formula
+oo+i7
eixâ(0u(0dx. (15.171)
/
-00+17
Theorem 15.31 Let m,s G R and 7 G [a;_,w+], and let a G S m ( R +
i[u-,uj+]). Then the operator
a(D) : # S ' 7 ( R ) -> # s - m ' 7 ( R ) (15.172)
is continuous, with the norm bounded by the constant C in Eq. (15.170).

Proof. Since i f s ' 7 ( R ) 9 u >-> u 7 G i f s ( R ) is a topological isomorphism,
it suffices to prove that
e 7 X a(Z?)e- 7 1 = a(D + ij) : HS(R) -* F s " m ( R )
is continuous, with the norm bounded by the constant C in Eq. (15.170).

But this is immediate from Theorem 15.4. •
Theorem 15.32 Let m,s G R, and 7 G [w_,u;+], and let a G 5 m ( R +

i[o;_,cj + ]) satisfy
|a(0-1|<C<fl-m, 9£ = 7, (15-173)
where C is independent of £ on the line 9£ = 7.

Tften £/ie operator Eq. (15.172) is invertible.
Proof. By definition, A A _ 1 and A - 1 A act as the identity, and from

Eq. (15.173) and Theorem 15.31, A'1 : H"-m^(B.) -¥ # S - 7 ( R ) is bounded.
Thus, it is the bounded inverse to the operator Eq. (15.172). •
Lemma 15.17 Let the following conditions hold

a)a€Sm(R + i[cj-,uj+]);
b) there exist non-zero limits
d±=lima(0|ei"m, (15.174)
as 9?£ —>• ±00, with £ remaining in the strip $£ G [LJ-,UJ+];

Then
(i) The set of zeroes of a in the strip 3£ € (LJ-,U)+) is either empty or
a bounded discrete set, the accumulation points being possible on the lines
Sf = A± only;
(ii) There exists a discrete set Z(a) G (A_, A+) with A± as the only pos-
sible accumulation points, such that for anj/7 G (A_, A+)\Z(a), Eq. (15.173)
holds.
Proof, (i) Condition a) implies that a is holomorphic in the strip (w_, w + )

and continuous up to the boundary of the strip. From the condition
b), it follows that there exists an open bounded subset U of the strip
9£ G ( w _ , « + ) such that a has no zeroes in the strip outside U. Since
a is holomorphic in U, its zeroes in U form a discrete set with possible
accumulation points at the boundary of the strip only.
(ii) Follows from (i) and Eq. (15.174). •
Corollary 15.3 Let a satisfy conditions of Lemma 15.17, and let Z(a)
be the set defined in Lemma 15.17.
Then for any 7 G (CJ_, oj+)\Z(a), the operator Eq. (15.172) is invertible.
15.9.4 Action of PDO in the Holder-Zygmund spaces with

exponential weights
Definition 15.8 Let s,j G R. We write u G C S,7 (R) if and only if

u 7 G C S (R).
For integer r > 0 and 7 € R, we write u € C Q ' 7 ( R ) if and only if
ti7 € CJ(R).
The reader can easily deduce from Theorem 15.8-Theorem 15.11 their
weighted analogues:
Theorem 15.33 Let s > r + 1/2, where r > 0 is an integer, and 7 G R.

Then i?*' 7 (R) C Cg' 7 (R), densely and continuously.
Theorem 15.34 Let s > 1/2, and 7 G R. Then for any s1 < s - 1/2,
i f s ' 7 ( R ) C C ' 7 ( R ) , densely and continuously.
Theorem 15.35 Let m G R, and let a G C°°(R + i[w_, w+]) satisfy, for
any multi-index a,
\a{a)(0\<Ca(Om~M, (15.175)
where Ca is independent of £ € Ti. + i[u!-,u)+].
Then for any s G R and 7 G [u-,ui+], A = a(D) : C ' T ( R ) -»
s m 7
C ~ ' (R) is bounded, with the norm admitting an estimate
SU
\\a(D)\\<C^2 P «>-m+|a,|o(o)(OI, (15-176)
|a|<JV«6R+i7
where C and N depend on n,s,m but not on a.
Theorem 15.36 Letm G R, and leta satisfy Eq. (15.173) and Eq. (15.176).
Then for any s,7 G R, a(D) : C S ' 7 (R) -¥ C s - m ' 7 ( R ) is invertible.
The following theorem is an evident analogue of Corollary 15.3 for the

Holder-Zygmund spaces.
Theorem 15.37 Let a satisfy Eq. (15.176) and conditions of Lemma

15.17, and let 2(a) be the set defined in Lemma 15.17.
Then for any 7 G (CJ-,CJ+) \ 2(a), an operator a(D) : C S ' 7 (R) —•
s m 7
C ~ ' (R) is invertible.
15.10 The Sobolev spaces w i t h exponential weights and

P D O o n a half-line
Essentially, here we reformulate definitions and main results of Section 15.5.

As in Subsection 15.9.2, all the proofs are reduced to the case of the Sobolev
spaces without the weight but one additional feature emerges: the embed-
ding of the Sobolev spaces with different weights. This feature makes it
possible to write down the asymptotics of solutions at the infinity.
15.10.1 Spaces H (R±)
Definition 15.9 Let s,7 G R. We write u €H (R±) iff u 7 €H{R±).
Lemma 15.18 H (R±) is the closure o/C£°(R±) in HS^(R).
Proof. The map u i-» u 7 defines an automorphism of CQ°(R±) and a

topological isomorphism between H (R±) and H (R±), hence it remains
to apply Lemma 15.5. •
Evidently, u € H°'7(R±) can be identified with a function from L2(R±; e2"/xdx),

extended by 0 on R T .
Theorem 15.38 Let a be holomorphic in the half-plane 3£ < 7 (resp.,

9£ > j), continuous up to the boundary 9£ = 7, and satisfy an estimate
Wfll < C ( 0 m (15-177)
in the closed half-plane.

o s,y o *-m,7 o«,7 o s-m,f
Then a{D) :H (R+) ->ff (R+) (resp., a{D) :H (R_) ^H
(R_)J is bounded.
Proof. Use the isomorphism
H ' ( R ± ) 3uûy €#(R±) (15.178)
with r = s and r.= s — m, and Theorem 15.12. •
For A > 0 and m € R, define
A£±(£) = (1 + A T i 0 m ( = exp[mln(l + A T iOD-

The Sobolev spaces with exponential weights and PDO on a half-line 355
Lemma 15.19 Let A > |7| — 1. Then for any s G R ,
K,±{D):H (R*)^H (RT) (15.179)
is a topological isomorphism.
Proof. Clearly, A™± satisfies the condition of Theorem 15.38, for any m,
hence the operator Eq. (15.179) and its counterpart with —m instead of
m are continuous. Since A™±(£>)A^™(D) = I and the same holds for the
product in the reverse order, Eq. (15.179) is the topological isomorphism^
Theorem 15.39 Let s' < s and 7' < 7. Then
o».7 o*'.7' o S>V o»'>7
H (R+) CH (R+) and H ( R - ) CH (R-),
Proof. With 7 = 7 ' , this is an evident corollary of Eq. (15.29) d), there-
fore it suffices to consider the case s = s'. Take A > I7I, and use Lemma
15.19 to reduce to the case s = 0, i.e., to the statements:
L 2 (R+;e 2 7 x da;) C L2(R+;e2'r'xdx), L2(R_;e2Vldz) c L2(R-;e2~<xdx),
densely and continuously, which are evident for 7' < 7. •
15.10.2 Spaces H*''y(B.±)
Definition 15.10 Let 3,7 € R. We write u £ HS'"/{K±) iff u 7 G

HS(R±). The norm in Hs<i(R±) is denoted by || • | | R ± ; S , 7 .
Clearly, p±, the-restriction-to-R-t-operator, maps i f s , 7 ( R n ) onto Hs'7(Rn±).
The following lemma is a straightforward generalization (and corollary) of
Lemma 15.6.
Lemma 15.20 Let s > 0 be an integer, and 7 € R . Then an equivalent
norm in HS''*(R,±) can be defined by
\Mn±;sn = £ ll e7I£?a "HL(R ± )- ( 15 - 18 °)

|a|<»
Theorem 15.40 Let a be holomorphic in the half-plane Q£ > 7 (resp.,
^£ < "f), continuous up to the boundary $j£ = 7, and satisfy an estimate
Eq. (15.177) for all 9 £ > 7 (resp., 9£ < j).
Then the operator
Hs'i(R+) 3«4p+a(D)lu G H"-m"(R+)
(resp., the operator
tf S ' 7 (R_) 3 u 4 p-a(D)lu G Hs-m,7(R_))
is well-defined and bounded.
Proof. Use the isomorphism
H ' (R±) 9 u >-> u 7 G # (R±) (15.181)
with r = s and r = s — m, and Theorem 15.13. •

L e m m a 15.21 Let A > |'y| — 1. Then for any s G R,
p±A%±(D)l : # S , 7 ( R ± ) -> # s - r o ' 7 ( R ± ) (15.182)
are topological isomorphisms.

Proof. Clearly, A™± satisfies the condition of Theorem 15.40, for any m,
hence the operator Eq. (15.182) and its counterpart with —m instead of m
are continuous. Since for any u G i ? s _ m , 7 ( R ± ) , and an extension operator
I, we have
P±Vx,±(D)l P±A^(D)lu = p±A?<±(D)AiX(D)lu =u

(cf. the proof of Theorem 15.13), and the same holds for the product in
the reverse order, with u G ZP' 7 (R±), the operators in Eq. (15.182) are
topological isomorphisms. •
T h e o r e m 15.41 Let s' <s and j ' < 7. Then
Hs'7 (R+) C Hs' -7' ( R + ) and F s ' 7 ' (R_) C Hs'-7 ( R _ ) ,

Proof. With 7 = 7', this is an evident corollary of Eq. (15.29) d), there-
fore it suffices to consider the case s = s'. Take A > |7|, and use Lemma
15.21 to reduce to the case s = 0, i.e., to the statements:
L 2 (R+;e 2 7 *dx) C L 2 ( R + ; e 2 7 ' x d x ) , L 2 (R_;e 2 7 ' x dx) c L 2 ( R _ ; e 2 7 I d x ) ,
densely and continuously, which are evident for 7' < 7. •

The Sobolev spaces with exponential weights and PDO on a half-line 357
15.10.3 The Cauchy-type integral and its decomposition

The results of this subsection are generalizations (and immediate corollar-
ies) of results in Subsection 15.5.3.
First, notice that operators 9± are well-defined on CQ°(R), and are
bounded operators from L2(R) to L 2 (R±)-
Theorem 15.42 (cf. Theorem 15.14) For 7 € R and \s\ < 1/2, 0±
o*>7
admits a unique continuous extension 9± : ff s ' T (R) —>H (R-±)-
Lemma 15.22 (cf. Lemma 15.7) Let 7 G R and \s\ < 1/2.
Then any function f G H* , 7 (R) admits a unique representation f = /+ +
/ _ , where f± £H '(R±), and f± = 8±f.
Define the extension-by-zero operator e± : p± (CQ°(R)) -4 £2(R) by
.±/w = {f)' ± x±I>0

<0.
Lemma 15.23 (cf. Lemma 15.8) Let 7 G R and \s\ < 1/2. Then
e± admits a unique bounded extension e± : i7 s ' 7 (R±) —> i ? s ' 7 ( R ) , and
o »>7
iJ s ' 7 (R-t) can be identified with H (R±)-
Theorem 15.43 (cf. Theorem 15.15) Let f G tfm+s'7(R), m > 0 inte-
ger, \s\ < 1/2, and 7 G R. T/ien for A > |-y| — 1,
0
+f = £ fi,-K!'-(D)S0 + A*™ (£>)*+A£_(D)/, (15.183)
k=i
where /,-,_ := p ' A ^ D ) / G C, and

m
B.f = ] T /;,+A^+(£>)<*„ + AiS(D)9-A%+(D)f, (15.184)

*=i
where fjt+~p'Akx^(D)f€C.
Theorem 15.44 (cf. Theorem 15.16) Let 7 G R and s = —m + s', where
m > 0 is an integer and \s'\ < 1/2.
0 ».7 o s.7
T/ien any f £H (R+)n H (R~) admits a decomposition
m
/ = £/tI>*-%, (15.185)
jfe=i
where fk € C .
15.11 Parabolic equations in spaces with exponential weights
The results on parabolic equations in R x [0, +oo), in weighted spaces, can

be easily obtained from results of Section 15.6 by adding the index 7 in the
notation of spaces, and assuming that 7 is chosen so that Eq. (15.34) and
Eq. (15.35) hold on the line 9£ = 7. Details are left to the reader.
15.12 The Wiener-Hopf equation on a half-line II
15.12.1 Statement of the problem

Let m > 0, A_ < 0 < A + , A_ < A+ and a € 5 m ( R ; [A_, A+]). Given
(generalized) functions g on R + and gr on R _ , consider a problem: find u
satisfying
a{D)u(x) = g°(x), x > 0, (15.186)

T
u{x) = g (x), x<0. (15.187)
Take any extension lgT of gr on R, and introduce / = u — lgr, g = g° —

p+a(D)lgr. In terms of / and g, Eq. (15.186) and Eq. (15.187) become
a(D)f(x) = g(x), x > 0, (15.188)

f(x) = 0, x<0. (15.189)
Assume that g 6 i f s ~ m ' 7 ( R + ) , where 7 € [A_,A+], and interpret the

problem Eq. (15.188)-Eq. (15.189) as follows: find u GH (R+) satisfying
p+a(D)u = g. (15.190)
This is the Wiener-Hopf equation. In order that the problem Eq. (15.190)
be well-posed, the symbol a must satisfy some regularity conditions.
Definition 15.11 We say that a 6 5™ g (R + i[A_,A + ]) and call it a
regular symbol of order m, if the following conditions hold:
(i) a e S m ( R + i[A_,A+]);
(ii) there exist non-zero complex constants d± and p\ e (0,1] such that
a(0=d±Kr + O(Kr-"), (15-191)

The Wiener-Hopf equation on a half-line II 359
as 9f££ -4 ±00, with £ remaining in the strip 9£ € [A_, A+];

(iii) there exist C > 0 and p^ € (0,1] such that for all f € R + i[A_, A+],
\a'(0\<C(0m'p2- (15.192)
Notice that a e S™ g (R + i[A_,A + ]) satisfy conditions of Lemma 15.17,

hence there exists a discrete set Z(a) € (A_,A + ) with X± as the only
possible accumulation points, such that for any 7 € (A_, A+) \ Z(a)
|a(0_1|<C7(0-"\ V £ € R + i7, (15-193)

where C 7 depends on 7 but not on £ G R + ij.
15.12.2 The Wiener-Hopf factorization of symbols holo-

morphic in a strip
Fix 7 G (A_, A+) \ Z(a). Exactly as in Section 15.7, we can construct the
factors a± = a 7 i ± in the Wiener-Hopf factorization formula
o(0=o7,_(0*,,+ (0. (15.194)

on the line 5£ = 7 instead of the real line in Section 15.7. Clearly, they
enjoy all the properties listed in Theorem 15.21, with the half-plane 9£ < 7
(resp., 3£ > 7) in the part of the lower half-plane 3£ < 0 (resp., 3£ > 0);
the orders K± — K ± ( 7 ) depend on 7 6 (A_, X+)\Z(a), too. However, ^±(7)
are constant on each connected component of (A_,A+) \Z(a): to see this,
notice that K ± ( 7 ) depend continuously on 7, and since the limits d± are
independent of 7, the analysis of the calculation of K± (7) shows that for
any 7',7" € (A_, A + ) \ Z(a), the difference K+(J') — K + ( 7 " ) is an integer
(and the same holds for K _ ' S ) .
Since a is holomorphic in the strip 3£ € (A_, A+), and a 7 ) + ( f ) - 1 (resp.,
a 7 ) _(£) - 1 ) is holomorphic in the half-plane 9£ > 7 (resp., 3£ < 7), and
each of these functions is continuous up to the boundary of the correspond-
ing domain, we can analytically extend a 7 i ± on wider half-planes by
a7,_(£)=a(£)/a7,+ (0, 3f£e[7,A+], (15.195)
a7i+(0=a(£)/a7,_(0, 9£€[A_>7]. (15.196)
Note that a 7i _ (resp., a 7 ,_) may have zeroes in the strip 9£ € (7>A+]
(resp., 9 { e L , 7 ) ) .
By using properties of a e S%g{R + i[A_, A+]) and Eq. (15.195)-

Eq. (15.196), we deduce from Theorem 15.21 the following theorem.
Theorem 15.45 Let a € S™g(R; [A_, A+]). Then for any 7 € (A_, A+) \
Z(a), there exist functions a 7 ) ± and constants K±(J) € C, with the follow-
ing properties:
(a+) a 7> +(£) is holomorphic in the half-plane 9£ > A_, admits the
continuous extension up to the boundary Q£ = A_, and satisfies an estimate
| a 7 , + ( 0 | < C(l + | £ | ) * K + ( 7 \ (15.197)
where C is independent of 9£ > A_ ;

(o + ) in the half-plane 9£ > 7, a7i-|_(£) admits an estimate
c(l + | € | ) W K + ( T ) < K + ( 0 l , (15-198)
where c > 0 is independent of £, 3£ > 7 ;

(c + ) for any r > A_, e > 0 and j = 1,2,..., there exists CT,t,j such that
for all £ in the half-plane 5s£ > T,
|oV?+«)l < C T ,«j(l + | e | ) * " + W - ^ % (15.199)

where p = min{pi,p2} € (0,1];
(d+) a 7 Y admits a representation
a 7 , + ( 0 ± 1 = dî-lYil + A - iO±K+M + <A!+(0, (15-200)
to/iere d is defined by Eq. (15.72), A = max{—A_,A+}, and t * A + admits

the following estimate: for any e > 0, there exists Ce > 0 SMC/I £na£ /or all
£ in the half-plane 9£ > 7,
l<Ai+(OI < ^ ( 1 + | e | ) ± R K + ( ^ - " + e ; (15.201)

(a_) a 7 i _(£) is holomorphic in the half-plane 9£ < A+, admits the
continuous extension up to the boundary Ss£ = A + , and satisfies an estimate
| a 7 , _ ( £ ) | < C ( l + |£|)* K -W, (15.202)
where C is independent of 9£ < A+ /

(6_) :n i/ie half-plane 3£ < 7, a 7i _(£) admits an estimate
c(l + | £ | ) * K - M < | a 7 , _ ( a i , (15.203)

w/iere c > 0 is independent of £_, 9£ < 7 ;
The Wiener-Hopf equation on a half-line II 361
(c_) for any r < A+, e > 0 and j = l,2,..., there exists CT,e,j such that
for all f in the half-plane Q£ < T,
|a??-(OI < CV,«j(l + |f|) R «-W-"+% (15.204)
where p = min{/9i,p 2 } G (0,1];

(d_) a^ 1 . admits a representation
a^(£)±l = (1 + A + i f ) ± K - ^ + < , A i _ ( 6 , (15-205)
tu/iere A = max{—A_,A+}, and t"tx_ admits the following estimate: for

any e > 0, there exists Ct > 0 such that for all f in the half-plane 9 f < 7,
l < A i - ( O I < Ce(l + |f | ) ± s « - ^ ) - " + e ; (15.206)
(ej £g. (15.194) holds in the strip Qf € [A_,A+];

f/^ conditions (a±) - (b±) define factors in Eq. (15.194) uniquely, up
to scalar multiples.
15.12.3 Main Theorems

The results below can be deduced from Theorem 15.45 exactly as their
analogues in Subsection 15.7.3 are deduced from Theorem 15.21.
We assume that a G S™g(R + i[A_, A + ]), and we fix 7 £ (A_, X+)\Z(a).
Theorem 15.46 Let s - Ste-fr) € ( - 1 / 2 , 1 / 2 ) .
Then the operator
p+a(D) •£'}*.+) -> H°-m"(R+)
is invertible, with the inverse given by
where
u = a 7 ,_(D)- 1 0+a 7 , + (Z})- 1 Zg €H''\R+), (15.207)
and Ig G Hs~m,'1('R) is an arbitrary extension of g.
Corollary 15.4 Let s - H/c-fr) G ( - 1 / 2 , 1 / 2 ) , and let g in Eq. (15.190)

belong to H r ' 7 ( R + ) , where r + K+(l) = q + s', q is a positive integer and
s'e ( - 1 / 2 , 1 / 2 ) .
O »iT
Then u €H (R-+), the solution to Eq. (15.190), admits a representation
« = £ cj(lg)AÂD)a1AD)-16o+A^_a^(D)-1e+Al_(D)a^+{D)-1lg,
2=1
(15.208)
where A = max{-A_,A+}, and
Cj(lg) = (A{7_1aT,+ (D)- 1 / 5 )(0). (15.209)
Theorem 15.47 Let s — 9 ? K _ ( 7 ) = — q + s', where q is a positive integer

and \s'\ < 1/2.
Then for any g € i f s - m ' 7 ( R + ) , a solution to the Wiener-Hopf equation
Eq. (15.190) exists but it is non-unique, and the general solution is given
by
1
u= a 7 ; _(D)- A A ,_(Z))^ + A A ,_(D)-«a 7 ) + (Z>)- 1 Z f f +a 7 ,_(Z3)- 1 ^ c , ^ - 1
^.
3=1
(15.210)
15.12.4 The asymptotics of solutions near the boundary and

the free boundary problems
We leave to the reader the straightforward reformulations of all the results
of Subsections 6.4 and 6.5.
15.12.5 The asymptotics of solutions at the infinity

We assume that a e S™g(R+i[\-, A+]), and we fix 71,72 € ( A _ , A + ) \ Z ( a ) ,
7i < 72- In the strip 3?£ € (71,72), there may exist finitely many zeroes
£ i , . . . , £q of the symbol a. Let kj be the multiplicity of a zero £,•.
Theorem 15.48 Let u EH (R+) be a solution to Eq. (15.190).

Then there exist u\ 6 i/*' 7 2 (R+) and constants Cji,j = l,...,q,I =
l,...,kj, such that u(x) has the following asymptotics, as x —> +00:
q kj
u(x) = E ^Cjix1'1 exp[i£ja;] + ui(x). (15.211)
3=11=1
growing data]
Parabolic equations o n R x R.+ with exponentially growing data 363
15.13 Parabolic equations on R X R + with exponentially

growing data
Assume that m > 0, a € S™g(R + i[A_, A+]), and Eq. (15.191) holds with
Kd± > 0, and consider the problem
{dt+a(Dx))f(x,t) = g(x,t), i>0,x>0; (15.212)

f(x,t) = 0, x<0,t>0; (15.213)
/(i,+0) = go{x), x>0. (15.214)
(For simplicity, we impose the homogeneous boundary condition Eq. (15.213),

which corresponds to the case of barrier options without a rebate; the gen-
eral case can easily be reduced to this one by the appropriate change of
the unknown function). Assume that the data grow exponentially, for in-
stance, so G # S ' 7 ( R + ) , and g 6 C ( [ 0 , + o o ) ; i 7 s - m ' 7 ( R + ) ) , where 7 e
(A_,A+) \ Z{a). From Eq. (15.191) and Eq. (15.193), we conclude that
Eq. (15.35) holds on the line $>£ = 7: there exist Co > 0 and a0 6 R such
that
R a ( 0 > co(O m - *o, 9£ = 7- (15.215)

Change the unknown
f(x,t) = e°°t-^f(ro„(x,t)
and the RHS in Eq. (15.212) and Eq. (15.214):
g(x,t) = e'lt-^g^yfat),
lx
go(x) = e~ gon(x).
In terms of /<r0l7 and <7<70)7, the problem Eq. (15.212)-Eq. (15.37) becomes
(dt+aao^{Dx))fao^{x,t) = g<TOtl(x,t), t>0,x>0; (15.216)

fnnfat) = °> x < 0 , i > 0 ; (15.217)
fw(x,+0) = 90n(x), x>0, (15.218)
where a<T0i7(^) := a(£ +17) 4- <T0 satisfies Eq. (15.38):
Ha«r0,-r(0 > co(Om, fGR, (15.219)

and the data do not exhibit the exponential growth: 30,7 € HS(R+), and
9<ro,7 € C([0,+oo);Hs-m(R+)). By using Eq. (15.219) and taking into
account that aVo^ e S^nreg(R), we can use the results of Section 15.8 to
solve the problem Eq. (15.216)-Eq. (15.218), and then obtain the solution
to the problem Eq. (15.212)-Eq. (15.214). The details are left to the reader.
Chapter 16
Elements of calculus of
pseudodifferential operators
In this Chapter, we Ust basic definitions and facts for simple classes ofPDO
in R n , which are necessary for the proofs of the results on NIG-like Feller
processes in Chapter 14, and calculation of approximate prices of European
options in Chapter 14 and basket options and exchange options in Chapter
9. The results of this Chapter can be used to calculate prices of barrier
options under Levy-hke Feller processes.
In contrast to the case ofPDO with constant symbols, where the Fourier
transform establishes the equivalence of the action of PDO and the mul-
tiplication operator, and so rather weak regularity conditions on symbols
suffice, here appropriate estimates on the derivatives of symbols must be
imposed. For simplicity, one usually assumes that symbols are infinitely
smooth, and imposes conditions on all derivatives, though weaker condi-
tions suffice to obtain the main results. In Section 16.1, we explain the
main ingredients and structure of the general theory of PDO for the sim-
plest class. We omit the most technical proofs like the composition theorem
and the theorem on the boundedness in appropriate scales ofBanach spaces
but we include several proofs, which demonstrate how the theory of PDO
works in relatively simple situations.
In Section 16.2, PDO depending on a parameter are considered; this
variant of the theory of PDO is needed in order to obtain approximate so-
lutions to pseudo-differential equations, when the exact analytical formulas
are non-avaiiable, and in Section 16.3, we consider PDO with symbols holo-
morphic in a tube domain; this variant of the theory of PDO is necessary to
treat problems with exponentially growing data. In Section 16.4, the proofs
of the technical results of Chapter 14 are presented. The transformation of
365
366 Elements of calculus of pseudodifferential operators
the symbol of a PDO under a change of variables is considered in the last

section.
16.1 Basics of the theory of P D O with symbols of t h e class

S™ 0 (R n X R")
Definition 16.1 Let m e R. We write a £ S™ 0 (R n x R " ) , if a is in-

finitely smooth, and for all multi-indices a, 0 the derivative o[^| := d"D%a
satisfies the following estimate
|ag(*,fll<C« l/S <O m - |a| , (16-!)

where constant Catp is independent of (x, £).
We set 5 - ° ° ( R " x R n ) := n m S f t ( R n x R n ) .
We start with this by now quite classical case since results for PDO with
symbols of the class 5 m ( R n x ( R n + iU)) can be deduced quite easily
from the results for S^0(Rn x R n ) . Most of the results below, with the
exception of the boundedness theorem for the Holder spaces, can be found
Grubb (1996) and Hormander (1985) (and local versions in Shubin (1978));
however, the form and proofs of some results, which we give below, do not
necessarily coincide with the proofs in these monographs.
16.1.1 Asymptotic summation and the composition theo-

rem
The following notation and theorem simplify many formulations in the the-
ory of PDO.
Definition 16.2 We say that a £ 5 ^ 0 ( R n x R n ) admits an asymptotic
expansion
+oo
a^âj, (16.2)
j=0
if for any N
N-l
rN:=a~Y, «j £ S^N{Rn x R"). (16.3)
o=o
Basics of the theory of PDO with symbols of the class S™0(R.n x R n ) 367
The celebrated subclass of 5J" 0 (R n x R n ) (historically, the first class of

PDO) is the class Sm of classical PDO of order m, with symbols admitting
the asymptotic expansion Eq. (16.2) having the following property: a,j(x,£)
are positively homogeneous w.r.t. f, of order m — j , outside a unit ball:
aj(x,t£)=tm-jaj(x,t),
for all |f | > 1 and t > 0.

Remark 16.1. In the theory of PDO, it is often convenient to use a weaker
form of Eq. (16.3): rN G S™0~M(Rn x R n ) , where M - • +co as N -» +oo,
but we will not need this form here.
For large |f|, the error term rw(x,f) is relatively small, and in the case
of symbols depending on a parameter, it is relatively small uniformly in
(x,f), so one can regard Eq. (16.2) as an approximate equality.
Theorem 16.1 (Asymptotic summation; Shubin (1978), Proposition 3.5).

Let aj € S%£j(Rn x R n ) , j = 0 , 1 , . . . .
Then
(i) there exists a £ 5j^ 0 (R n x R n ) such that Eq. (16.2) holds, and
(ii) if a1 £ Sj™0(Rn x R") is another symbol satisfying Eq. (16.2), then
a1-a<E5-°°(RnxRn).
On the space <S(Rn), the action of a PDO a(x,D) with symbol a e

S f t ( R n x R") is defined by
a(x,D)u(x) = (2Tr)-n f eix«a(x,0u(£R. (16.4)
T h e o r e m 16.2 Let a e Sm0(Rn x R n ) .

Then a PDO a(x, D) maps <S(Rn) into itself, continuously.
The proof is a straightforward exercise.

If a is a polynomial in £:
\a\<m
then Eq. (16.4) gives the definition of a differential operator
D)u{x)= J2 aa(x)(Dau)(x).
|a|<m
If we write the coefficients on the right, we obtain another operator, denote

it ar:
ar(x,D)u= ^2 (Da(aau))(x).
\a\<m
The corresponding general definition of a PDO ar(x,D) with the right

symbol1 a for a € Sm0(Rn x R n ) is
+ OO /.+OO
/ -OO J —CO
where the integral is understood as the iterated one. PDO with the right
symbols naturally arise when one considers the adjoint operators. The
following theorem ensures that the adjoint operator belongs to the same
class of PDO, and gives the formula for its symbol. Its proof and the proof
of the Composition Theorem are rather technical, and we refer the reader
to Shubin (1978), Hormander (1985) and Grubb (1996).
Theorem 16.3 Let a e ^ ( R " x R " ) .
Then the (complex) adjoint a(x,D)* defined by
(a(x,D)u,v) = (u,a(x,D)*v)
is a PDO whose (left) symbol a* belongs to 5]" 0 (R n x R n ) and admits an

asymptotic expansion
a*(x,S)~ X>!rXa£i(:r,0.
|o|>0
Theorem 16.3 and Theorem 16.2 allow one to define the action of a{x, D)
on <S'(Rn) by duality:
(a(x, D)<j>, f) = (</>, a(x, D)*f), V / 6 5(R").
Clearly, if <f> € <S(R") then the new definition of a(x, D)<f> coincides with the
old one. Further, a(x, D) is a continuous map from <S'(Rn) into itself. This
map can be defined as a unique continuous extension of the map defined by
Eq. (16.4) on 5 ( R n ) , and the standard agreement in the theory of PDO is
to denote this map by the same symbol a(x, D); the restriction of a(x, D)
on any subspace L C <S'(Rn) is also denoted by the same symbol.
1
In this context, Eq. (16.4) is the definition of the PDO a(x,D) with the left symbol a.
Basics of the theory of PDO with symbols of the class S J " 0 ( R " x R n ) 369
The next basic fact and the most important tool in the theory is the
Composition Theorem.
T h e o r e m 16.4 Let a G S^ 0 (R n x R n ) and b G S^(Rn x Rn).
Then C = a(x, D)b(x, D) is a PDO with the symbol c € S™ 0 +m '(R n x R " ) ,
which admits an asymptotic expansion
c(x,0~ £ (a!)"1^(x,06(Q)(x,0, ( 16 - 5 )
|a|>0
in the sense that for any integer N > 0,
rN=c- £ (a\)-lal%a) G S^m'~N(Rn x R"). (16.6)

0<|a|<Af-l
16.1.2 Parametrix and approximate square root
Definition 16.3 We write a G / 5 ^ 0 ( R n x R " ) iff a G Sm0(Rn x R n ) is

invertible and admits a bound
l a ^ r ^ C ^ r ™ . (16-7)
where C is independent of (x,£).
We call a an elliptic symbol, and A = a(x, D) an elliptic operator.
R e m a r k 16.2. The standard definition of an elliptic symbol requires
Eq. (16.7) outside a compact set in £-space. We use Eq. (16.7), since it
suffices for our purposes and admits a trivial generalization for the case of
operators depending on a parameter.
The following definition formalizes an intuitive notion of an approximate
inverse.
Definition 16.4 We say that B — b(x, D) is a parametrix of A — a(x, D)
if and only if
AB = I + ti(x,D), BA = I + t2{x,D), (16.8)

where t,- G 5 - ° ° ( R n x R " ) .
Theorem 16.5 (Parametrix construction) Let a G JS™ 0 (R' ! xR n ). Then
A = a(x,D) has a parametrix B = b{x,D), where b G 5^™(R n x R n )
admits an asymptotic expansion
b ~ 6_ m + 6_ m _! + • • •, (16.9)
where
6_ m = l / o G S^(Rn x R"), (16.10)
and 6_ m _ s G S f ™ _ s (R n x R n ) , s = 1,2,... are defined inductively by
fc-^^-a-1 ^ (a!)-1a^(6_m_j)(Q). (16.11)
Proof. Step 1. We construct 6 such that
i4A(x,D) = / + t ( i ) i ? ) , (16.12)
and similarly one can construct 6'(x, JD) such that
b'(x,D)A = I + t'(x,D), (16.13)
where t, t' G 5 _ 0 0 ( R n x R n ) . By mimicking the standard argument on the

equality of the left and right inverses, one deduces that Eq. (16.13) holds
with b(x, D) (and generally, with different t' G 5 _ ° ° ( R n x R n ) ) .
Step 2. From Eq. (16.7), we deduce that 6_ m := 1/a G S ^ R " x R n ) ,
and by substituting the asymptotic expansion Eq. (16.9) with yet un-
known b-m-., G Sj~™ _J (R n x R n ) , s = 1,2,... (for any sequence of such
6_ m _ s ,s > 1, 6 is defined by Theorem 16.1), and by using the composi-
tion theorem we obtain Eq. (16.12), where t G 5 ^ Q ( R " X R") admits an
asymptotic expansion
«~$3*., (16-14)
S>1
with ( J e 5 1 " J ( R " x R " ) , s > l , given by
j+\a{=s
= a6_m_s+ J^ (a!) _ 1 « ( a ) (^-m-,)(a). (16.15)

j<s,j+[a\=s
Define 6_ m _ s ,s > 1, by Eq. (16.11). Then by induction, we find 6_ m _ 3 G

5 1 "J l _ s (R n x R " ) , and the RHS in Eq. (16.15) is zero. Hence, t given by
Eq. (16.14) is of the class 5f^°(R n x R " ) , and b(x,D) is a parametrix of
a(x,D). ' •
Basics of the theory of PDO with symbols of the class S^ 0 (R n x R n ) 371
Theorem 16.6 (Asymptotic square root) Let a G ISm0(Kn x R n ) satisfy
a(x,0£#-, V(x,£)eRnxRn. (16.16)
Then there exists b G 5^ 0 / 2 (R n x R n ) such that
b(x,D)2 = a(x,D) + t(x,D), (16.17)
where t G 5~°°(R n x R n ) , and b admits an asymptotic expansion
where
bm/2 := a}'2 G S^2(Rn x R"), (16.19)

_S
and bm/2-a G S|"Q (R- n x
R " ) , s > 1, are defined inductively by
bm/2-s = ~\b-)2 £ (^r^^-k^Hbm/^la). (16-20)

j,k<s,\a\+k+j=s
Proof. Fix a branch of In by the requirement In a G R for a > 0, and set
&m/a(*,0 = a ( * , 0 1 / 2 : = exp[-Ino(a;,Pl-
under assumption Eq. (16.16), the symbol bm/2 is well-defined, and by using
Eq. (16.7), one can easily check that it belongs to S™Q2(Rn x R n ) . Fix any
sequence of 6 m / 2 _ s G S™ 0 /2_S (R" x R n ) , s > 1, define b G S™ 0 /2 (R n x R»)
by Eq. (16.18), by using Theorem 16.1, and calculate the symbol of b(x, D)2
by using the composition theorem. We obtain that it admits an asymptotic
expansion
&/2 + E*'' < 16 - 21 )

s>l
where
f = Yl («0- 1 (&m/2- fc ) (a) (& m/2 -,)(a)

\a\+k+j=s
= 2bm/2-jbm/2
+ £ («!)-1(bm/2-fc)(a)(fcm/2-J)(Q). (16.22)
j,k<3,\a\+k+j=s
If we define bm/2-s,s > 1, by Eq. (16.20), we obtain that the RHS in

Eq. (16.22) vanishes, and Eq. (16.21) gives Eq. (16.17). •
16.1.3 Boundedness theorem

One also needs a boundedness result for PDO in Banach spaces embedded
in <S'(R"). Such most popular and simple spaces are Z<2-based Sobolev
spaces i P ( R n ) , since in this case estimates for norms can easily be ob-
tained by means of the Fourier transform (see, e.g., Eskin (1973), Taylor
(1981), Hormander (1985) and Grubb (1996)). For the theory of stochas-
tic processes, Holder spaces C"(Rn) are more appropriate. Until relatively
recent times, only local Holder estimates were available (see, e.g., Taylor
(1981)); in Yamazaki (1986), global estimates are obtained for Besov spaces
Bpq, p,q e (0, oo], and results, which are necessary for us, follow as special
cases, since
for s £ R, BJ 2 is the Sobolev space H3(Rn);
for s > 0, B^x(Rn) is the Holder-Zygmund space C s ( R n ) ,
and finally,
for non-integer s > 0, C s ( R n ) is the Holder space, with the norm denned
by
for s £ (0,1):
\H\c- = N U o o + sup sup \h\-s\u(x + h)-u(x)\;

x€R" |/i|<l
for s = m + s', where m > 0 is an integer and s' £ (0,1),
I M I o = J2 \\Dau\\c...
\a\<m
Theorem 16.7 Let p,q £ (0,oo], m,s,£ R, and a £ 5 ^ 0 ( R " x R n ) .

Thena(x,D) : Bpq(Rn) —> Bp~m(Rn) is bounded, and its norm admits an
estimate
||o(x,£>)||<C sup sup <0H-m|a5)W)l, (16-23)
where the constants C and N depend on s, m, p, q and n but not on a £

5 x m 0 (R n x R " ) .
Operators depending on parameters 373
16.1.4 The inverse operator and the square root

The results of this subsection can be found in Grubb (1996).
Theorem 16.8 Let a G I S^0(Rn xRn), and let for some s,A = a(x,D) :
Hs{Rn) -» H3-m(Rn) be invertible.
Then the inverse is a PDO of the class S^™(R n x R n ) , and its symbol
admits the same asymptotic expansion Eq. (16.9) as the parametrix.
Corollary 16.1 Under the conditions of Theorem 16.8, A is an invertible

operator in <S(Rn), and for any r, A : Hr(Rn) -> Hr~m{Rn) is invertible.
Proof. <S(Rn) C Hr(Rn), for any r, hence both A and A'1 maps <S(Rn)
into itself. Therefore, they are mutual inverses as operators in 5 ( R n ) . Since
5 ( R n ) C Hr{Rn) densely, and A " 1 : Hr-m{Rn) -> Hr{Rn) is bounded,
we conclude that A' is the inverse to A : H (R ) -> F r _ m ( R n ) .
1 r n
•
Remark 16.3. Corollary 16.1 is valid not for the scale H"(Rn) only but
for many other scales of spaces as well.
Theorem 16.9 Let the conditions of Theorem 16.6 hold, and let $la(x, D)
be positive-definite, and%a(x,D)($la(x,D))~1 be bounded.
Then there exists a PDO B = b(x, D) such that B2 = A, and its symbol
b e 5ô ( R n x R " ) admits the same asymptotic expansion Eq. (16.18) as
an approximate square root.
16.2 Operators depending on parameters
Consider a PDO with the symbol depending on a parameter (or parame-

ters). If the dependence is nice, one can exploit it and obtain useful results,
e.g., prove the invertibility of the operator for large (or small, depending on
the situation) values of the parameter(s), and derive approximate formulae
for the inverse. As it turns out, the simplest way to realize this possibility
is to repeat the whole sequence of definitions and theorems from the very
beginning by explicitly specifying the dependence on the parameter(s). The
proofs themselves can be repeated word by word, without any change ex-
cept in the notation. For similar results about operators with a parameter,
see, e.g., Shubin (1978) and Grubb (1996).
We consider the case when there are two parameters - one, call it e,
describing the possible smallness of the derivatives w.r.t. x, and the other,
call it d, which describes the smallness of the derivatives w.r.t. £. The case
when only derivatives w.r.t. x (respectively £) are small obtains by fixing
d (respectively e).
So, let £ C (0,1] and V C [1, +oo) be the sets on which e and d live; we
assume that the statement e/d —• 0 makes sense on the set £ x T>, and we
regard a pair (e, d) as a parameter.
The starting point is a natural generalization of the class S™ 0 (R n x R n ) .
Definition 16.5 Let m G R. We write a e S^Q(£ x 2?;R n x R n ) if

a = a(e, d; x, £) satisfies estimates
| o g j ( c , d ; x , 0 l < Cat3eM(d+\Z\r-W, (16.24)
for all multi-indices a, 0, where the constants Cap are independent of e G

£,deV,(x,£) £ R " xR".
We set
S-°°((£ x V);Rn x R n ) := n m € R S f t ( ( £ x D ) ; R n x R n ) .
Sometimes we will use these definitions with
(£ x £>)£0 = {(e,d) G £ x V | e/d < e 0 }
instead of £ x P .
Definition 16.6 We say that o G S™0(5 x D ; R " x R " ) admits an asymp-

totic expansion Eq. (16.2) if for any N
N-l
rN:=a-Y, <*j G e ^ S J ^ f x XJ; R n x R " ) . (16.25)
j=o
T h e o r e m 16.10 Let e^'a.,- G 5^" 3 '(5 x D; R " x R n ) , j = 0 , 1 , . . .

T/ien t/iere exists a G S£ 0 (£ x X>;R" x R n ) such that Eq. (16.2) holds in
the sense Eq. (16.25).
T h e o r e m 16.11 Let a G S^0(£ x D ; R " x R n ) and b G S^Q(£ X5;R"X

n
R ).
Then C = a(x, D)b(x, D) is a PDO with the symbol c G S^m> {£ x V; R n x
R n ) , which admits an asymptotic expansion Eq. (16.5) in the sense that for
Operators depending on parameters 375
any integer N > 0,
rN = c- Y, («!) _ 1 a ( a ) t(a) € eNS™+m'-N(£ xV;Rnx Rn).

0<\a\<N-l
(16.26)
Definition 16.7 We write a £ IS%0(£ x D ; R n x R n ) if and only if

o 6 ^ ^ ( f x X>;Rn x R " ) , and there exists e0 > 0 such that if e/d < e0
then a(e, d; x, £) is invertible and admits a bound
|a(e,d,s,0-1|<C,(d2 + |e|2)-m/2.
where C is independent of (e, d, x, £) with e/d < eo-
We call A — a(e,d; x, D) an elliptic operator with a parameter.
In the sequel, we omit arguments e, d, and write a(x, D).
Theorem 16.12 Let a £ IS^0(£ xT>;R" x R n ) , and s,s-m > 0 be

positive non-integers.
There exists e\ > 0 such that if e/d < e\, then A = a(x,D) : C S (R") —>
C3~m(Rn) is invertible, and its inverse B is a PDO, whose symbol b £
Si™((£ X P ) £ l ; R " x R n ) admits the asymptotic expansion Eq. (16.9) in
the sense of Definition 16.6, where
6_ m = l / o € S^J*((f x 2J) £ l ;R n x R n ) ,
and
b _ m _ 3 € esS^~a{{£ x 2?) €l ;R n x R"), s > 1,
are determined by Eq. (16.11).
Proof. Define 6_ m = l / o , B° = 6_ m (x, £>). Since a e / £ £ „ ( £ x D; R " x

n
R ) , there exists e0 > 0 such that 6_ m £ S^((£ x V)eo;Rn x R " ) , and
by applying Theorem 16.11, we obtain
AB° = I + t1(x,D), B°A = I + t2(x,D), (16.27)
where t5 € eS^({£ x 2?) £o ;R n x R " ) . Hence, (d/e)tj £ 5 ? 0 ( R n x R n )

uniformly in (e,d) £ (£ xV)eo, and by the boundedness theorem, there
exists C such that for all (e, d) £ (£ x V)Co,
||(d/e)ti(a;,I>)||c.-m_c-» < C. ( 16 -28)

and
\\(d/e)t2(x,D)\\c--tc-<C. (16.29)
When e/d < ei := min{e 0 ,1/C}, we conclude from Eq. (16.27)-Eq. (16.29)
that AB° and B°A are invertible in C s - m ( R n ) and C s ( R n ) , respectively.
By the boundedness theorem, B° : C3~m(Rn) -+ C*(R n ) is bounded, and
hence, B°{1+ ti(x,D))~1 and (1 + t2{x,D))-1B° are bounded right and
left inverses to A : C"(R ) — C s - m ( R n ) .
n
An analogue of Theorem 16.8 for operators with parameter(s) is valid:

a parametrix B has the symbol b G S£™((£ x Z>) eo ;R n x R " ) , which can
be constructed exactly as in Theorem 16.5, and A~* is a PDO with the
symbol of the class S{^((£ x V)tl; R n x R n ) .
They satisfy Eq. (16.8) with tj G S-°°{{£ x £>)£o; R " x R n ) . By multi-
plying the left equality in Eq. (16.8) by A~l from the right and using Theo-
rem 16.11, we find A'1 -B = t(x, D), where t G S-°°{(£ x V)ei; R n x R " ) .
This proves that the asymptotic expansion for the symbol of the parametrix
is the asymptotic expansion for the symbol of the inverse. •
Remark 16.4. Theorem 16.12 is valid for other spaces as well; for instance,
in the scale of Sobolev spaces H3(Rn), it is valid without the restrictions
on s and m.
Theorem 16.13 Let Ka e IS^Q(£ x V; R n x R n ) be positive, and Qa €

S^0(£xV;RnxRn).
Then there exist ex > 0 and b € S™J?({£ x V)tl;Rn x Rn) such that
b(x,D)2 = a(x,D), (16.30)
andb admits the asymptotic expansion Eq. (16.18) in the sense of Definition
16.6, with the terms given by Eq. (16.19)-Eq. (16.20).
Proof. Exactly the same argument as in the proof of Theorem 16.6 allows
us to construct b G S^0((£ x V)eo\Rn x R"), which satisfies Eq. (16.17)
with t G S~°°{{£ x V)e0;Rn x R n ) ; Eq. (16.18) is understood in the sense
of Definition 16.6, and all the terms are given by Eq. (16.19)-Eq. (16.20).
By using a straightforward modification of the results on fractional powers
in Seeley (1967), Shubin (1978) and Grubb (1996), we obtain that if e0 > 0
is sufficiently small and (e,d) G (£ x V)eo, then B = a{x,D)1/2 is well-
defined, its symbol b belongs to S™£2((£ x V)£o;Rn x R n ) , and b - a1'2 G
Operators with symbols holomorphic in a tube domain 377
eS™ 0 /2_1 ((£ x X>)£o;R" x R " ) . It follows that
b = b + k, (16.31)
where k e eS^2'1^ x P ) £ o ; R n x R " ) . Suppose, that for some j > 1,
k e eiS?f-j((£ x X>)£0;Rn x R n ) , (16.32)
and show that then Eq. (16.32) holds with j +1 instead of j ; this will finish
the proof of the Theorem.
By calculating b(x, D)2 with the help of Theorem 16.11 and taking
Eq. (16.30)-Eq. (16.32) and Eq. (16.17) into account, we find that
T := b(x, D)k(x, D) + fc(x, D)b(x, D)
is a PDO with the symbol t £ e* + 1 S£ 0 ~ i - 1 ((£ x X>)£o;Rn x R n ) .

Apply the parametrix of b(x, D) to T = t(x, D), and use Theorem 16.11;
the result is that 2k G e^ + 1 5^ 0 / 2 _ j _ 1 ((5 x D ) e o ; R n x R n ) , and the verifi-
cation is complete. •
16.3 Operators with symbols holomorphic in a tube domain
Let U C R n be an open set such that its closure contains the origin:
U 3 {0}, and consider the classes 5 m ( R n ) x ( R n +iU)) = Sm0(Rn x (R n +
iU)) introduced in Definition 14.1. If we work with the same spaces as in
Subsection 16.1, we can repeat all the statements there word by word by
simply adding +iU in the notation of each class of symbols. This means that
we make no use of the fact that symbols admit holomorphic continuation
into the tube domain, and if we consider European options with bounded
payoffs, e.g., European puts, we may do it and lose no essential information.
When the payoff is exponentially growing at infinity, as is the case with
European calls, appropriate spaces are spaces with exponential weights,
and one can consider the action of PDO in such spaces if and only if their
symbols admit analytic continuation into an appropriate tube domain. For
PDO of the class S^0{Rn x (R n + iU)), the part of 5 ( R " ) is played by
5 ( R n ; U) = {u | e ^ u e <S(Rn), V y € U},
with a natural system of seminorms, and the following theorem can be

378 Elements of calcvJLus of pseudodifferential operators
trivially deduced from Theorem 16.2 by using the equality
e<x'v>a(x,D)u = (e<x'y)a(x,D)e-<XtV)")e<x'v)u
= a(x,D + iy)e<x'y)u.
Theorem 16.14 Let a e S™ 0 (R n x (R n + iU)). Then a(x,D) is a

continuous operator in <S(R"; U).
As in Subsection 16.1, we define by duality the action of a(x,D) in
<S'(Rn; U), the space of continuous linear functionals over <S(Rn; —0). The
next step, namely, the action in Banach subspaces of <S'(Rn; U), is not very
difficult but technically more involved.
To make the main idea clear, we restrict ourselves to the ID-case,
when U = (A_,A + ) is an interval, and A_ < — 1 < 0 < A + . The multi-
dimensional case can be treated similarly.
Definition 16.8 Let o;_ < w+ and s > 0 be non-integer. CS(R; [w_, w+])
denotes the space of functions with the finite norm
IMIc.(Ri[u,_,w+]) = ll(e"- x + e w + I H | c » ( R ) . (16.33)

Similarly, one defines Sobolev (and Besov) spaces with weight.
Theorem 16.15 Letae S f t ( R x ( R + t[A_, A+])), [w_,w+] C [A_,A + ],
and s,s — m be positive non-integers.
Then a(x,D) : C S (R; [o;_,w+]) —> C s _ m ( R ; [w_,w+]) is bounded, and its
norm admits an estimate
\\a(x,D)\\<C sup sup {âl-m\a[;Ux,0\, (16.34)

\<*\ + \P\<N ( x , € ) € R x (R+i[u<_ ,w+]) ^'
where constants C and N depend on s,m and ui-,uj+ but not on a S

S f o ( R x ( R + t[A_,A + ])).
Proof. For any u G C S (R; [w_, u+\)
\\a(x,D)u\\c>mu,_,„+]) < \\<r-'a{x,D)u\\c.<ji) + \\e"+xa(x,D)u\\c,{R)

= \\(<r-*a{x,D)e-u-x)<r-xu\\c.W
+\\(eu+*a(x,D)e-u+x)er+xu\\c.pi)
By using the equality
e u,±x o(x, D)e~u±x = a(x, D + iu±),

Proofs of auxiliary technical results 379
and Theorem 16.7, we continue
= ||a(x, D + zw_)e w - I u|| c »(R) + \\a(x, D + icv+)e"+xu\\c.(R)

< C(|| e w -*ti||c.(R) + | | e u + - u | | c . ( R ) ) ,
where C can be chosen in the form of the RHS in Eq. (16.34). The sum
in the brackets defines the norm equivalent to the norm Eq. (16.33), which
finishes the proof of the Theorem. •
We live to the reader as an exercise the proof for spaces defined by Eq. (16.33)
with the weight (eu-x + e ^ 1 ) - 1 instead of e w - x + e"+ x .
16.4 Proofs of auxiliary technical results
Proof of Theorem 14-1. If m < 0, then a(x, D) is bounded in L 2 ( R n ) , there-

fore for A large enough, A + a(x, D) is invertible in L2(R n ). By Corollary
16.1, it is invertible in 5(11").
Now let m > 0. Due to Eq. (14.5), for A large enough, b(\;x,f) =
(A + a ( x , £ ) ) - 1 is well-defined on R n x R n . By using the composition
theorem, it is straightforward to show that
6(A; x, D)(X + a(x, D)) = I + t(X; x, D), (16.35)
where t satisfies: for any pair of multi-indices (a, (3),
lim sup |tg))(A;x,OI(0H=0- (16-36)
It follows from the boundedness theorem and Eq. (16.36) that if A is suffi-
ciently large, the RHS in Eq. (16.35) is an invertible operator in I,2(R n ).
Hence, for these A, a(x, D) + A : Hs(Rn) —> L2(R n ) has a bounded left in-
verse (/ 4- t(A; x, D)) _1 6(A; x, D). Similarly, it has a bounded right inverse,
hence it is invertible. By Corollary 6.10, a(x, D) + A is invertible in <S(Rn).
Theorem 14.1 has thus been proved.
Proof of Theorem 14.4. a) We have a(x,£) := d2 - L(x,£) e 5j > 0 (R n x
( R n +iU)) for any U, and if U satisfies Eq. (14.14), we see that Eq.' (16.16)
holds for all (x,£) G R n x (R n + iU). Hence, the analogues of Theorem
16.9 and Theorem 16.6 for PDO with symbols holomorphic in a tube do-
main are valid, which gives not only Eq. (14.15)-Eq. (14.17) but the other
terms of the asymptotic expansion of the symbol as well (see Eq. (16.18)-
Eq. (16.20)).
b)-d) are special cases of the analogues of Theorems 6.11 and 6.7 for
PDO depending on a parameter.
Theorem 14.4 has thus been proved.
Proof of Lemma 14-1- By using Eq. (14.26) and repeating the proof of
Theorem 16.8, it is possible to show that the symbol of (\+r + ip(x, .D)) - 1 ,
call it R(\;x,£), satisfies estimates
|flg>(A;x,OI < Ca0(\X\ + K| + l ) - 1 ® " 1 " 1 , (16.37)

for all multi-indices a and /3, therefore for any t > 0, the symbol <ft(x,£) of
the operator
exp[-t(r + iP(x,D)] = (2ni)-1 [ eXtR(X;x,D)d\,
decays as £ —> oo faster than any power of (£), together with all derivatives.
But the kernel kt(x, y) of a PDO exp[—t(r+ip(x, D)] is expressed via symbol
as follows:
+oo
/
-oo
and by differentiating under the integral sign, we conclude that kt € C°°(R n x
Rn).
Calculation of the omitted terms of the asymptotic expansion Eq. (14-28)
We use Theorem 16.8 and Theorem 16.5, and Eq. (16.9)-Eq. (16.13) with
\ + r + a(x, £) instead of a(x, £). For instance, the RHS of Eq. (14.28) with
two terms (the only case which may have some relevance to practice) is
(A+r+a(x,0)-1 + (A+r+a(x,0)-3 £ ala\x,t)a[a){x,t) + --- (16.38)

|a| = l
By substituting Eq. (16.38) into Eq. (14.27) we can obtain the next terms
in Eq. (14.29). In particular, the first omitted term is equal to
+oo+i<r e X p AT + ix$ E u i - i a ( a ) («»£Ka) (*. 0

2 — ,,, J . f 7 a ; ' 1 } d£d\, (16.39)
4* JcJ-
where a € (0,A + ). The integral converges absolutely, hence we can apply
the Fubini theorem and change the order of integration. After that we can
Change of variables and pricing of multi-asset contracts 381
transform the contour C$ by pushing it to the left. There is only one pole
of order three, and when it is crossed, by the residue theorem the term
_if r+00+ia exp[ix£ - r(r + a(x,£))] £ | a | = 1 o<°)(x,g)a (a) (x,Q
(16.40)
appears. After that we can push the contour to infinity. Since the integrand
in Eq. (16.39) admits a bound via
we obtain 0 in the limit. If the derivatives of the symbol admit estimates

Eq. (14.30), the integrand in Eq. (16.40) admits an estimate via
CeexP(-r|£|)|£|-2.
Since for a ^ 0 fixed,
exp(-r|£|)|£r2^ < d ,
-oo-H<r
where Cx is independent of T, we conclude that the expression in Eq. (16.40),

which is the first omitted term in Eq. (14.29), is of order r 2 e. Similarly, the
next term comes from an expression Eq. (16.40) with XJ Q I=2> a n ( ^ hence
is of order T2e2/d , as claimed at the end of Section 14.3. All the next
terms can be calculated as well, but they are relatively small w.r.t. the
first omitted term, and hence of no practical significance.
16.5 Change of variables and pricing of multi-asset con-

tracts
Let K : R n —> R n be a diffeomorphism, which coincides with the identity

operator outside some compact. It induces the map
K* : S(Rn) B u H u o / c e <S(Rn),
which is clearly an isomorphism. Let A be a PDO with the symbol a G

5™ 0 (R n x R n ) , and define an operator
Ax : 5 ( R n ) 9 U H [A(u o «)] o K " 1 e <S(R").

Then (see, e.g., Theorem 4.2 in Shubin (1978)) the following theorem is
valid.
Theorem 16.16 Ai is a PDO with the symbol <zi € 5 ^ 0 ( R n x R " ) ,
which admits an asymptotic expansion
ai(y,r/)| y = K ( x ) ~ 5 ; ( a ! ) - ' a w ( x , f « ' ( ^ ) • D?eiK^«\z=x, (16.41)

a
where a a ( x , £) = d?a(x,£), and
K\(X, z) = K(Z) — K(X) — K'(X)(Z — x).
Notice that Theorem 16.16 is valid for the other classes of symbols, which
we consider.
By using Theorem 16.16 and approximate calculations, similar to the
ones made in the preceding section, we can obtain approximate formulas
for the solution to the problem Eq. (9.25)-Eq. (9.26), and hence, approxi-
mations to prices of basket options and exchange options.
16.6 Pricing of barrier options under Levy-like Feller pro-

cesses
As we have seen in Chapter 8, the pricing formula is straightforward if

an explicit formula for the resolvent (X + A)-1 is available, where A is
the operator of the appropriate boundary value problem for the stationary
Generalized Black-Scholes equation. In Chapter 8, we considered option
pricing under RLPE, when the generator was a PDO with constant symbol,
and so the explicit construction was possible.
At the end of Chapter 14, we have constructed an approximate resol-
vent for the Generalized Black-Scholes equation in the case of pricing of
European options under NIG-like Feller processes and used it to obtain
an approximate pricing formula. The first step of the construction of the
approximate resolvent was the inversion of the symbol; the PDO with the
symbol (A + a(x,£))-1 served as an initial approximation, and the correc-
tions could be calculated by means of the calculus of PDO.
The same program, with inevitable complications, can be realized in
order to construct an approximate resolvent (A + A)-1, where A is an
operator of a boundary problem on R + , say, and acts a s u n p+a(x, D)u,
on functions u, which vanish on R _ . We
Pricing of barrier options under Livy-like Feller processes 383
i) first, construct the Wiener-Hopf factors of the symbol A + a(x, £) for

each x (and A G E ^ c , for some 9 € (7r/2,7r) and C > 0):
\ + a{x,£) =za-(\;x,£)a+(\;x,£),
ii) then, assuming that a is a symbol from a good class (that is, its deriva-
tives are small in an appropriate sense) and using the explicit formulas
for the factors, obtained in Chapter 3, we show that the factors also be-
long to classes of symbols with sufficiently good properties, uniformly
in A € £e,c, for some 0 € {TT/2, IT) and C > 0;
iii) next, define the initial approximation to the resolvent by
i?°u = a_ (A; x, D)6+a+ (A; x, D)lu,
where lu is an extension on R of a function u defined on R + (say,

u e L2(R+), and I = e+), and by using the calculus of PDO and the
theory of action of PDO in appropriate function spaces on a half-line
(here the generalization of the theory of Chapter 15 for the case of
PDO with variable symbols is needed), show that
R°X(X + A) = I-TX,
where T\ is an operator with the small norm, uniformly in A G Ee,c,
for some 8 e (7r/2, n) and C > 0;
iv) finally, we can calculate the resolvent with any precision, by summing
the following series:
(\ + A)-1 = (I + Tx+T$ + ---)R°x. (16.42)
By using Eq. (16.42) and Theorem 15.18, we can find an approximate price
for a barrier option under a Levy-like Feller process.
Of course, the concrete realization of the program requires long calcu-
lations and additional constructions.
Bibliography
Abel, A.B. and Eberly, J.C. (1999) "The effects of irreversibility and uncertainty
on capital accumulation", Journal of Monetary Economics 44, 339-377.
Barles, G., Burdeau, J., Romano, M. and Samsoen, N. (1995) "Critical stock
price near expiration", Mathem. Finance 5, 77-95.
Barndorff-Nielsen, O.E. (1977) "Exponentially decreasing distributions for the
logarithm of particle size", Proc. Roy. Soc. London. Ser. A 3 5 3 , 401-419.
Barndorff-Nielsen, O.E. (1998) "Processes of normal inverse Gaussian type", Fi-
nance and Stochastics 2, 41-68.
Barndorff-Nielsen, O.E. and Jiang, W. (1998) "An initial analysis of some Ger-
man stock price series", Working Paper Series, 15, Aarhus: CAF Univ. of
Aarhus/Aarhus School of Business.
Barndorff-Nielsen, O.E. and Levendorskii, S.Z. (2001) "Feller processes of Normal
Inverse Gaussian type", Quantitative Finance 1, 318-331.
Barndorff-Nielsen, O.E., Nicolato, E. and Shephard, N. (2001) "Some recent de-
velopments in stochastic volatility modelling", Working paper
Barndorff-Nielsen, O.E. and Prause, K. (2001) "Apparent scaling", Finance and
Stochastics 5, 103-113.
Barndorff-Nielsen, O.E. and Shephard, N. (2001a) "Modelling by Levy processes
for financial econometrics", in Levy processes: Theory and Applications,
O.E. Barndorff-Nielsen, T. Mikosch and S. Resnik (Eds.), Birkhauser, 283-
318.
Barndorff-Nielsen, O.E. and Shephard, N. (2001b) "Non-Gaussian Ornstein-
Uhlenbeck-based models and some of their uses in financial economics (with
discussion)", J. Royal Stat. Soc. 6 3 . 167-241.
Barndorff-Nielsen, O.E. and Shephard, N. (2001c) "Normal modified stable pro-
cesses", Working paper
Barndorff-Nielsen O.E. and Shephard N. (2002) Levy based dynamics models for
financial economics, forthcoming.
Bernanke, B.S. (1983) "Irreversibility, uncertainty, and cyclical investment",
385
386 Bibliography
Quarterly Journal of Economics 98, 85-106.

Bertoin, J. (1996) Levy processes, Cambridge University Press, Cambridge.
Bielecki, T.R. and Rutkowski, M. (2001) Credit risk: modelling, valuation and
hedging, Springer-Verlag, Berlin Heidelberg New York
Black, F. and Cox, J.C. (1976) "Valuing corporate securities: Some effects of
bond indenture provisions", Journal of Finance 3 1 , 351-367.
Black, F. and Scholes, M. (1973) "The pricing of options and corporate liabilities",
J. of Political Economy 8 1 , 637-659.
Blumenthal, R.M. and Getoor, R.K. (1968) Markov processes and potential theory,
Academic Press, New York.
Borovkov, A.A. (1972) Stochastic processes in queuing theory, Nauka, Moscow,
1972 (Transl. Springer, New York, 1976).
Bouchaud, J.-P. and Potters, M. (2000) Theory of financial risk, Cambridge Uni-
versity Press, Cambridge.
Boutet de Monvel, L. (1971) "Boundary problems for pseudo-differential opera-
tors", Acta Math. 126, 11-51.
Boyarchenko, S.I. (2001) " Capital accumulation under non-Gaussian processes
and the Marshallian law", CARESS Working Paper #01-03, Univ. of Penn-
sylvania, Philadelphia.
Boyarchenko, S.I. and Levendorskii, S.Z. (1998) "On rational pricing of derivative
securities for a family of non-Gaussian processes", Preprint 98/7, Insititut
fur Mathematik, Universitat Potsdam, Potsdam.
Boyarchenko, S.I. and Levendorskii, S.Z. (1999a) "Generalizations of the Black-
Scholes equation for truncated Levy processes", unpublished manuscript
Boyarchenko, S.I. and Levendorskii, S.Z. (1999b) "Option pricing for truncated
Levy processes", In Applications of Physics in Financial Analysis (Eu-
rophysics conference abstracts, Dublin, July 15-17 (1999)), P. Alstrom,
K.B. Lauritsen (Eds.), European Physical Society, 1999, S5.
Boyarchenko, S.I. and Levendorskii, S.Z. (2000) "Option pricing for truncated
Levy processes", Intern. Journ. Theor. and Appl. Finance 3:3, 549-552.
Boyarchenko, S.I. and Levendorskii, S.Z. (2001a) " Option pricing and hedging
under regular Levy Processes of exponential type", in Trends in Mathemat-
ics. Mathematical Finance, M. Kohlman and S.'Tang (Eds.), 121-130.
Boyarchenko, S.I. and Levendorskii, S.Z. (2001b) " Pricing of perpetual Bermudan
options", submitted to Quantitative Finance
Boyarchenko, S.I. and Levendorskii, S.Z. (2002a) "Perpetual American options
under Levy processes", SI AM Journ. of Control and Optimization, forth-
coming.
Boyarchenko, S.I., and Levendorskii, S.Z. (2002b) "Barrier options and touch-
and-out options under regular Levy processes of exponential type", sub-
mitted to Annals of Applied Probability
Carr, P. (1998) "Randomization and the American put", Review of Financial
Studies 11:3, 597-626.
Carr, P. and Faguet, D. (1994) "Fast accurate valuation of American options",
Bibliography 387
Working paper, Cornell University.

Carr, P., Geman, H., Madan, D.B. and Yor, M. (2001) "The fine structure of asset
returns: An empirical investigation", Journ. of Business, forthcoming
Carr, P. and Madan, D.B. (1998) "Option valuation using the fast Fourier trans-
form", Journal of Computational Finance 2, 61-73.
Clark, P.K. (1973) "A subordinated stochastic process model with fixed variance
for speculative prices", Econometrica 4 1 , 135-156.
Cont, R. and da Fonseca, J. (2001) "Deformation of implied volatility surfaces:
an empirical analysis", R.I. No 466. Centre de Mathematiques Appliques,
Ecole Polytechnique, Paris.
Courrege, Ph. (1966a) "Sur la forme integro-differentielle des operateurs de C^
da C satisfaisant au principe du maximum", Sim. Theorie du Potentiel,
1965/66. Expose 2
Courrege, Ph. (1966b) "Sur la forme integro-differentielle du generateur in-
finitesimal d'un semi-group de Feller sur une variete", Sem. Theorie du
Potentiel, 1965/66. Expose 3
Darling, D.A., Liggett, T. and Taylor, H.M. (1972) "Optimal stopping for partial
sums", Annals of Mathem. Statistics 43, 1363-1368.
Deaton, A. and Laroque, G. (1992): "On the Behavior of Commodity Prices,"
Review of Economic Studies, 59:1, 1-23.
Delbaen, F. and Schachermayer, W. (1994) "A general version of the fundamental
theorem of asset pricing", Math. Ann. 300, 463-520.
Delbaen, F. and Schachermayer, W. (1996) "The variance optimal martingale
measure for continuous processes", Bernoulli 9, 81-105.
Dixit, A.K. and Pindyck, R.S. (1994) Investment under uncertainty, 2d ed.,
Princeton University Press, Princeton, New Jersey.
Duffie, D. (1996) Dynamic asset pricing theory. 2d ed., Princeton University
Press, Princeton, New Jersey.
Duffie, D., Pan, J. and Singleton, K. (2000) "Transform analysis and asset pricing
for affine jump-diffusions", Econometrica 68, 1343-1376.
Eberlein, E. (2001a) "Application of generalized hyperbolic Levy motions to
Finance", in Ldvy processes: Theory and Applications, O.E. Barndorff-
Nielsen, T. Mikosch and S. Resnik (Eds.), Birkhauser, 319-337.
Eberlein, E. (2001b) "Recent advances in more realistic risk management: the hy-
perbolic model", in Mastering Risk Volume 2: Applications, Carol Alexan-
der (Ed.), Prentice Hall-Financial Times, 56-72.
Eberlein, E., Breckling, J. and Kokic, P. (2000a) "A new framework for the
evaluation of market and credit risk", in Data mining and Computational
Finance, G. Bol, G. Nakhaezadeh, K.-H. Vollmer (Eds.), Physica-Verlag,
Wirtschaftwissenschaftliche Beitrage Bd. 174, 51-67.
Eberlein, E., Breckling, J. and Kokic, P. (2000b) "A tailored suit for risk man-
agement: the hyperbolic model", in Measuring risk in complex stochastic
systems, J. Franke, W. Hardle and G. Stahl (Eds.), Lecture Notes in Statis-
tics 157, 189-202.
388 Bibliography
Eberlein, E. and Jacod, J. (1997) "On the range of options prices", Finance and
Eberlein, E., Kallsen, J. and Kirsten, J. (2001) "Risk management based on
stochastic volatility", FDM Preprint 72, University of Freiburg.
Eberlein, E. and Keller, U. (1995) "Hyperbolic distributions in Finance",
Bernoulli 1, 281-299.
Eberlein, E., Keller, U. and Prause, K. (1998)
"New insights into smile, mispricing and value at risk: The hyperbolic
model", Journ. of Business 7 1 , 371-406.
Eberlein, E. and Ozkan, F. (2001) "The defaultible Levy term structure: ratings
and restructuring", FDM Preprint 72, University of Freiburg.
Eberlein, E. and Prause, K. (2001) "The generalized hyperbolic model: finan-
cial derivatives and risk measures", in Mathematical Finance- Bachelier
Congress 2000, H. Geman, D. Madan, S. Pliska, T. Vorst (Eds.), Springer-
Verlag
Eberlein, E. and Raible, S. (1999) "Term structure models driven by general Levy
processes", Math. Finance 9, 31-53.
Eberlein, E. and Raible, S. (2001) "Some analytic facts on the generalized hyper-
bolic model", in Proceedings of the 3rd European Meetings of Mathemati-
cians, Birkhauser
Elliot, D.F. and Rao, K.R. (1982) Fast transforms: Algorithms, analyses, appli-
cations, Academic Press, New York
Eskin, G.I. (1973) Boundary problems for elliptic pseudo-differential equations,
Nauka, Moscow (Transl. of Mathematical Monographs, 52, Providence,
Rhode Island: Amer. Math. Soc, 1980).
Ethier, St. N. and Kurtz, Th. G. (1986) Markov processes, Wiley Series in Prob-
ability and Mathematical Statistics, John Wiley & Sons, New York
Fama, E.F. (1965) "The behavior of stock market prices", Journ. of Business 38,
34-105.
Fedoryuk, M.V. (1988) The saddle point method, Nauka, Moscow (In Russian)
Follmer, H. and Leukert, P. (1999) "Quantile hedging", Finance and Stochastics
3, 251-273.
Follmer, H. and Schweizer, M. (1991) "Hedging of contingent claims under in-
complete information", in Applied Stochastic Analysis, M.H.A. Davies and
R.J .Elliott (Eds.), New York, Gordon and Bleach, 389-414.
Follmer, H. and Sondermann, D. (1986) "Hedging of non-redundant contin-
gent claims", in Contributions to Mathematical Economics, A. Mas-Collel,
W. Hildebrand (Eds.), North-Holland, Amsterdam, 205-223.
Fons, J.S. (1994) "Using default rates to model the term structure of credit risk",
Financial Analysts Journal, September/October, 25-32.
Geman, H. and Yor, M. (1996) "Pricing and hedging double-barrier options: A
probabilistic approach", Mathem. Finance 6:4, 365-378.
Gerber, H. and Shiu, E. (1994) "Option pricing by Esscher transforms (with
discussion)", Trans. Soc. Actuaries 46, 99-191.
Bibliography 389
Geske, R. and Johnson, H.E. (1984) "The American put option valued analyti-
cally", Journ. of Financed 1511-1524.
Goll, T. and Kallsen, J. (2000) "Optimal portfolios for logarithmic utility",
Stochastic Processes and their Applications 89, 31-48.
Goll, Th. and Riischendorf, L. (2001) "Minimax and minimal distance martingale
measures and their relationship to portfolio optimization", Finance and
Stochastics 5:4, 557-581.
Gourieroux, C., Laurent, J.P. and Pham, P. "Mean-variance hedging and
numeraire", Mathem. Finance
Grubb, G. (1996) Functional calculus of pseudo-differential boundary problems,
Birkhauser, Boston Basel Berlin
Heston, S.L. (1993) "A closed-form solution for options with stochastic volatility,
with applications to bond and currency options", Rev. Finan. Stud. 6, 327-
343.
Hormander, L. (1985) The analysis of linear partial differential operators III,
Springer-Verlag, Berlin
Hubbard, R.G. (1994) "Investment under uncertainty: keeping one's options
open", Journal of Economic Literature 32, 1816-1831.
Hull, J.C. (2000) Options, futures & other derivatives. 4th ed., Prentice Hall, New
Jersey.
Hyakawa, K. and Kumano-go, H. (1971) "Complex powers of a system of pseudo-
differential operators", Proc. Japan Acad. 47, 359-364.
IngersoU, J.E., Jr. (2000) "Digital contracts: Simple tools for pricing complex
derivatives", Journal of Business 73:1, 67-88.
Jacob, N. (1996) Pseudo-differential operators and Markov processes, Academie
Verlag, Berlin
Jones, E.P., Mason, S.P. and Rosenfeld, E. (1984) "Contingent claims analysis
of corporate capital structures: An empirical investigation", Journal of
Finance 39, 611-627.
Kallsen, J. (1998) "Semimartingale modelling in Finance. Dissertation", Mathe-
matische Fakultat, Universitat Freiburg im Breisgau.
Kallsen, J. (2000) "Optimal portfolios for exponential Levy processes", Mathe-
matical Methods of Operations Research 51:3, 357-374.
Karatzas, I. and Shreve, S.E. (1998) Methods of mathematical Finance, Springer-
Verlag, Berlin Heidelberg New York.
Keller, U. (1997) "Realistic modelling of financial derivatives. Dissertation",
Mathematische Fakultat, Universitat Freiburg im Breisgau.
Koponen, I. (1995) "Analytic approach to the problem of convergence of truncated
Levy flights towards the Gaussian stochastic process", Physics Review E 52,
1197-1199.
Kramkov, D. and Schachermayer, W. (1999) "The asymptotic elasticity of util-
ity functions and optimal investment in incomplete markets", Ann. Appl.
Probab. 9, 904-950.
Kwok, Y.-K. (1998) Mathematical models of financial derivatives, Springer-
390 Bibliography
Verlag, Berlin Heidelberg New York.

Lamberton, D. (1995) "Critical price for an American option near matu-
rity", in Seminar on Stochastic Analysis, Random Fields and Applications,
E. Bolthausen, M. Dozzi, F. Russo (Eds.), Birkhaiiser, Boston Basel Berlin,
353-358.
Lander, D.M. and Pinches, G.E. (1998) "Challenges to the practical implemen-
tation of modelling and valuing real options", The Quarterly Review of
Economics and Finance 38, Special Issue, 537-567.
Leland, H.E. (1994a) "Corporate debt value, bond covenants, and optimal capital
structure", Journal of Finance 49, 1213-1252.
Leland, H.E. (1994b) "Bond prices, yield spreads, and optimal capital struc-
ture with default risk", Working paper 240, IBER, University of California,
Berkeley (1994b).
Leland, H.E. and Toft, K.B. (1996) "Optimal capital structure, endogenous
bankruptcy, and the term structure of credit spreads", Journal of Finance
5 1 , 987-1019.
Levendorskii, S.Z. (1993) Degenerate elliptic equations, Kluwer Academic Pub-
lishers, Dordrecht
Levendorskii, S.Z. and Paneah, B. (1994) "Degenerate elliptic equations and
boundary problems", in Encyclopaedia of Mathematical Sciences v . 6 3 ,
Yu.V. Egorov and M.A. Shubin (Eds.), Springer-Verlag, Berlin Heidelberg
New York, p. 131-202.
Levendorskii, S.Z. and Zherder, V.M. (2001) "Fast option pricing under regular
Levy processes of exponential type", submitted to Journal of Computational
Finance.
Longstaff, F.A. and Schwartz, E.S. (1995) "A simple approach to valuing risky
and floating rate debt", Journal of Finance 50, 789-819.
Madan, D.B., Carr, P. and Chang, E.C. (1998) "The variance Gamma process
and option pricing", European Finance Review 2, 79-105.
Madan, D.B. and Milne, F. (1991) "Option pricing with VG martingale compo-
nents", Mathem. Finance 1, 39-55.
Madan, D.B. and Seneta, E. (1990) "The VG model for share market returns",
Journ. of Business 6 3 , 511-524.
Mandelbrot, B.B. (1963) "The variation of certain speculative prices", Journ. of
Business 36, 394-419.
Mandelbrot, B.B. (1997) Fractals and scaling in Finance: Discontinuity, concen-
tration, risk, Springer-Verlag, New York.
Mantegna, R.N. and Stanley, H.E. (1994) "Stochastic process with ultraslow con-
vergence to a gaussian: the truncated Levy flight", Phys. Rev. Lett. 7 3 ,
2946-2949.
Mantegna, R.N. and Stanley, H.E. (1997) "Physics investigation of financial mar-
kets", Proceedings of the International School of Physics "Enrico Fermi",
Course CXXXIV, F. Mallamace and H.E. Stanley (Eds.), IOS Press, Ams-
terdam, 473-489.
Bibliography 391
Mantegna, R.N. and Stanley, H.E. (2000) An introduction to Econophysics. Corre-

lation and complexity in Finance, Cambridge University Press, Cambridge.
Matacz, A. (2000) "Financial modelling and option theory with the truncated
Levy process", Intern. Journ. Theor. and Appl. Finance 3 : 1 , 143-160.
McKean, H-P. (1965) "A free boundary problem for the heat equation arising from
a problem of mathematical economics", Industrial Management Review 6,
32-39.
Merton, R.C. (1973) "Theory of rational option pricing", Bell Journal of Eco-
nomics and Management Science # 4 (Spring), 141.
Merton, R.C. (1990) Continuous time Finance, Blackwell, Cambridge,
MA/Oxford.
Mordecki, E. (2000), "Optimal stopping and perpetual options for Levy pro-
cesses" , Prepublicaciones de Matematica de la Universidad de la Republica,
2000/35, (2000).
Mordecki, E. (1998) "Optimal stopping for a compound Poisson process with
exponential jumps", Publicaciones Matemdticas del Uruguay 7, 55-66.
Mordecki, E. (1999) "Optimal stopping for a diffusion with jumps", Finance and
Mordecki, E. (2000) "Optimal stopping and perpetual options for Levy pro-
cesses" , Talk presented at the 1 World Congress of the Bachelier Finance
Society, June 2000.
Musiela, M. and Rutkowski, M. (1997) Martingale methods in financial modelling,
Springer-Verlag, Berlin Heidelberg New York.
Novikov, E.A. (1994) "Infinitely divisible distributions in turbulence", Physics
Review E 50, R3303-3305.
Pelsser, A. (2000) "Pricing double barrier options using Laplace transform", Fi-
nance Stochastics 4 : 1 , 95-104.
Pham, H. Rheinlander, T. and Schweizer, M. (1998) "Mean-variance hedging for
continuous processes: New Proofs and Examples", Finance and Stochastics
2, 173-198.
Prabhu, N.U. (1980) Stochastic storage processes. Queues, insurance risks and
dams, Springer, New York.
Prause, K. (1998) "The generalized hyperbolic models: estimation, financial
derivatives and risk measurement", PhD thesis. Mathematics Faculty,
Freiburg
Raible, R. (2000) "Levy processes in Finance: Theory, Numerics, and Empiri-
cal Facts. Dissertation", Mathematische Fakultat, Universitat Freiburg im
Breisgau.
Revuz, D. (1970) "Mesures associees aux fonctionnelles additives de Markov. I",
Trans, of AMS 148, 501-531.
Rogers, L.C.G. and Williams, D. (1994) Diffusions, Markov processes, and mar-
tingales. V.l, 2d ed., Wiley, New York [1st ed. by D. Williams, 1979].
Rubinstein, M. and Reiner, M. (1991) "Breaking down the barriers", Risk 4,
28-35.
392 Bibliography
Sarig, O. and A. Warga, A. (1989) "Some empirical estimates of the risk structure
of interest rates", Journal of Finance 44, 1351-1360.
Sato, K. (1999) Levy processes and infinitely divisible distributions, Cambridge
University Press, Cambridge.
Schweizer, M. (1996) "Approximation pricing and the variance-optimal martin-
gale measure", Ann. Probab. 24:1, 206-236.
Schweizer, M. (1999) "A minimality property of the minimal martingale mea-
sure", Stat. Probab. Lett. 42, 27-31.
Seeley, R. (1967) "Complex powers of an elliptic operator", Amer.Math.Soc. Proc.
Symp. Pure Math. 10 288-307.
Sharpe, M.J. (1989) General theory of Markov processes, Academic Press, New
York.
Shiryaev, A.N. (1999) Essentials of stochastic Finance. Facts, models, theory,
World Scientific, Singapore New Jersey London Hong Kong.
Shubin, M.A. (1978) Spectral theory and pseudo-differential operators, Nauka,
Moskow (Transl. 1980)
Spitzer, F. (1964) Principles of random walk, Van Nostrand, Princeton.
Taylor, M. (1981) Pseudo-differential operators, Princeton University Press,
Princeton NJ
Triebel, H. (1978) Interpolation theory, function spaces, differential operators.
VEB Deutscher Verlag der Wissenschaften, Berlin.
Wiener, N. and Hopf, E. (1931) "Uber eine classe singularer Integralgleichungen",
Sitzber. Deutsch. Akad. Wiss. Berlin, Kl. Math. Phys. Tech., 696-706.
Wilmott P., Dewynne J.N. and Howison S. (1993) Option pricing: Mathematical
models and computation, Oxford Financial Press, Oxford.
Wilmott, P., Howison, S. and Dewynne, J. (1995) The mathematics of finan-
cial derivatives. A student introduction, Cambridge University Press, Cam-
bridge.
Yamazaki, M. (1986) "A quasi-homogeneous version of paradifferential operators,
I. Boundedness on spaces of Besov type", J.Fac.Sci.Univ. Tokyo Sect. I A
Math. 33, 131-174.
Yang, S.R. and Brorsen, B.W. (1992): "Nonlinear Dynamics of Daily Cash
Prices," American Journal of Agricultural Economics, 74:3, 706-715.
Yosida, K. (1964) Functional analysis, Springer-Verlag, Berlin, Heidelberg, New-
York
Zhou, C. (1997) "A jump-diffusion approach to modelling credit risk and valuing
defaultable securities", Board of Governors of the Federal Reserve System,
Finance and Economics Discussion Papers Series: 1997/27.
Zhu, J. (2000) Modular pricing of options (Lecture Notes in Economics and Math-
ematical Systems. Vol. 493), Springer-Verlag, Berlin Heidelberg New York.
Index
general notation p', 309

«[$(*. 0.x 0 + , 0 _ , 3 O l , 314
a<Q>,x u 7 , 135, 345
(x,y), x
abbreviations
0j, x
Dj, x a. e., 41
(x),x AF, 59
(ACP)-condition, 59
lorl, x
\a\, x (ACT)-condition, 59
3c, Da, X a. s., 41
X X
EMM, 15
It/, x EMM-condition, 19
a A 6, x PAC, 60
oV6,x PAN, 60
a±, x FFT, 255
Bc, x IAC, 256
B°, x GHP, 8, 67, 80
S, x HP, 8, 67
L„(R"), x KoBoL, 8, 68
.F, x NTS Levy, 8, 67
U, X NIG, 8, 67
6, <50, <J7, 301 NPV, 222
supp, 305 ODE, 2
®, 304 MBS-theory, 2
II • ||., 307, 346 PDO, 2, 12
l|-||R±;..7,355 PDE, 2
II-He, 310 RLPE, 10
H'll.,7,346 TLP, 8, 68
p + , p _ , 189,277 VGP, 8, 67
393
394 Index
spaces S-°°((£xV);Rn xRn),374

C'(Rn), 310 Sft>((£xP)eo;R"xRn),375
C S ' 7 (R), 353 S-°°((£ x V)eo;Rn x R n ) , 375
n
C 0 ( R n ) , x, 53 IS^0{SxV;R xRn),375
<?om(Rn), x,
C 5 ( R n ) , 310 absolutely continuous, 40
C m ( t / ) , 298 adapted, 16
C^{U), 298 adjoint operator, 368
CT(I0, 298 cr-algebra, 39
Cr(K;B), 298 almost everywhere, 41
CS°(R n ), 298 almost surely, 41
C s (R;[w_,w + ]), 378 arbitrage
i f s ( R n ) , 307 opportunity, 16
F ( R n ± ) , 312 portfolio, 14
# s ( R n ± ) , 312 arbitrageur, 14
H3(U), 312 asymptotic expansion, 366
^ ; r > ( R " x R ) , 318 asymptotic square root, 371
asymptotic summation, 367
n
HZI (R x R+), 318
tfs'7(R), 346 bad news principle, 223
H ( R ± ) , 354 bankruptcy, 233
i F ' ^ R i ) , 350 cost, 234
level, 234
S ( R n ) , 298 barrier, 24
S'(Rn), 300 Black-Scholes formula, 5
«S(R;[A_,A + ]), 344 Black-Scholes equation, 24
S ' ( R ; [ - A + , - A - ] ) , 345 Blumenthal zero-one law, 53
S(Rn;U), 377 Borel cr-algebra, 40
S ' ( R n ; t / ) , 378 bounded variation, 48
classes of symbols capital

F L S m ( R n x ( R n + iU)), 284 accumulation, 225
S m ( R n ) , 305 structure 232
5™ g (R), 325 characteristic exponent, 9, 47
S m ( R + i[w_,w + ]), 351 GHP, 80
S m ( R + i(-oo,u> + ]), 351 KoBoL, 84
Sm(R + i[w-,+oo)), 351 NIG, 78
S £ g ( R + i[A_,A+]),358 NTS, 78
Sm(Rnx(Rn+i£/)),284 RLPE, 83
S r ; o ( R n x ( R n + i t / ) ) , 284 VGP, 79
5 ^ 0 ( R n x R n ) , 284 characteristic function, 42
5 J o o ( R n x R n ) , 370 complete market, 17
I S f t ( R n x R " ) , 369 completion, 40
5To(f x P ; R " x R n ) , 374 Composition Theorem, 368
Index 395
contingent claim, 17 exponential type, 82, 83, 285

replicated, 17 extension, 305
continuous pasting condition, 271
convolution, 43, 44, 299 factorization with a parameter, 169
convolution operator, 307 Fast Fourier Transform, 260
coupon payment, 243 Feller property, 53
credit filtered space, 9, 16
risk, 231 filtration, 16, 52
spread, 232 forward contract, 203
functional
debt, 234 additive, 59
density, 40 continuous, 60
Levy, 9 integrable, 60
probability, 2 (7-integrable, 60
derivative security, 18 natural, 60
diffusion coefficient, 4 Fourier transform, 11, 42, 299
Dirac delta-function, 301 inverse, 43
distribution, 41 free boundary value problem, 122
generalized hyperbolic, 79 frown, 5
hyperbolic, 80
infinitely divisible, 46 Gaussian coefficient, 9
normal inverse gaussian, 79 Gaussian covariance matrix, 48
Levy stable, 7 generalized Black-Scholes equation,
truncated Levy, 8, 68 23
drift, 2, 4, 48 generalized kernel, 307
dual measure, 43 generalized function, 43
Dynkin's formula, 57 generating triplet, 9, 48
Geske-Johnson approximation, 153
early exercise boundary, 27, 151
effect hedge, 18
hangover, 225 hedging
user-cost, 225 delta-, 118, 214
elliptic locally risk-minimizing, 116, 209
equation, 309 mean-variance, 115
operator, 309, 369 ratio, 116
symbol, 369 historic measure, 19
endogenous default, 232 hitting time, 122
equity, 235 Holder space, 310
equivalent martingale measure, 15 Holder-Zygmund space, 310
equivalent measures, 40
Esscher transform, 20 identical in law, 46
event, 40 Inada conditions, 221
expectation, 41 independent random variables, 44
expiry date, 5 index of the factorization, 327
396 Index
infimum process, 29, 50, 223 modification, 46

infinitesimal generator, 11, 55 Modified Stable laws, 89
integration-along-cut method, 256 multi-index, x
intensity, 70
based model, 231 net present value, 222
parameter, 10 no-arbitrage, 14
investment no-early-exercise, 153
strategies, 221 Normal Modified Stable laws, 68
threshold, 222
irreversible investment, 33, 221 option
American, 121
jump-diffusion, 6 call-like, 146
junk bond, 239 perpetual, 27
perpetual call, 121
killing operator, 53 perpetual put, 124
KoBoL process, 68 put, 27
of order u, 68, 69 put-like, 143
kurtosis, 3 Asian, 216
barrier, 24, 185
Laplace transform, 45 down-and-out, 185
Laplace exponent, 49 up-and-out, 185
law, 41 basket, 203
tempered stable, 77 Bermudan, 267
Levy-Khintchine formula, 9 digital
Levy process, 46 call, 113
in law, 46 put, 114
spectrally negative, 52 down-and-out call, 24
leverage, 233 European
call, 5
Marshallian law, 33, 222 put, 27
Markov exchange, 203
semigroup, 53 first-touch-digital, 165
simple property, 53 perpetual, 121
storong property, 54 power, 113
mean, 3 straddle, 114, 215
measurable function, 40 strangle, 215
measurable space, 40 touch-and-out, 165
measure space, 40 order of process, 10, 84, 285
/^-complete, 40 order of PDO, 305
measure, 40
cr-additive, 40 parametrix, 369
finite, 40 payoff, 13
a-finite, 40 portfolio, 13
Levy, 9, 48 locally risk-minimizing, 115
Index 397
weight, 115, 210 random time, 49

position, 13 rate
long, 13 dividend, 124
short, 13 riskless, 97
^-positive operator, 320 rational price, 122
potential real option, 222
kernel, 55 rebate, 24, 185
measure, 55 redundant security, 17
operator, 55 region
probability, 40 action, 224
density, 2 optimal exercise, 124
measure, 40 inaction, 224
infinitely divisible, 46 regular functional, 300
space, 40 regular symbol, 325, 358
process replicating portfolio, 17
dual, 58 resolvent, 191
Feller-Levy, 285 equation, 55
Levy of exponential type, 82 kernel, 55
Generalized Hyperbolic, 8, 67, 80 operator, 55
Hyperbolic, 8, 67 restriction, 305
KoBoL, 8, 68 risk-free measure, 27
Levy, 8
NIG-like Feller, 286 sample path, 45
Normal Inverse Gaussian, 8, 67, 78 short position, 13
Normal Tempered Stable Levy, 8, skewness, 3
67, 78, 223 smooth fit principle, 27
regular Levy of exponential type, smooth pasting condition, 27
10, 83 Sobolev space, 307
stable Levy, 9 Sobolev embedding theorem, 310
Truncated Levy, 8, 68 Spitzer identities, 272
Variance Gamma, 8, 67, 79 spot price, 4
production function, 221 stable Levy process, 9
pseudodifferential operator, 12, 305 steepness parameter, 10, 68
stochastic ^tocesa, 4^
put-call parity, 107 stochastic volatility models, 6
stopping time, 54
quasi long range dependence, 4 strongly elliptic, 316
quasi-left-continuity, 54 strike price, 5
^-potential kernel, 55 structural approach 231
q-potential measure, 55 subordination, 49
q-potential operator, 55 subordinator, 49
support, 305
Radon-Nikodym theorem, 40 supremum process, 29, 50, 226
random variable, 40 SV-models, 6
398 Index
symbol, 12, 305, 367

constant, 12
right, 368
left, 368
system of finite-dimensional
distributions, 46
tax benefit, 234

tensor product, 304
terminal payoff, 20
test function, 298
time to expiry, 5
time discretization, 154
trading strategy, 16
trajectory, 45
translation operator, 59
transmission property, 12
trigger price, 222
universally measurable function, 57
variance, 3
volatility, 2, 4
implied, 5
historic, 5
smile, 5
surface, 5
weakly ^-positive, 320

winding number, 194
Wiener-Hopf
equation, 28
factorization, 51, 89
factorization method, 28
yield spread, 239

Advanced Series on Statistical Science and Applied Probability - Vol. 9
NON-GAUSSIAN
MERTON-BLACK-SCHOLES THEORY
This book introduces an analytically tractable and computationally
effective class of non-Gaussian models for shocks (regular Levy
processes of the exponential type) and related analytical methods
similar to the initial Merton-Black-Scholes approach, which the
authors call the Merton-Black-Scholes theory.
The authors have chosen applications interesting

for financial engineers and specialists in financial
'economics, real options, and partial differential equations
(especially pseudodifferential operators); specialists in
stochastic processes will benefit from the use of the pseudo
differential operators technique in non-Gaussian situations. The authors
also consider discrete time analogues of perpetual American options and
the problem of the optimal choice of capital, and oudine several possible
directions in which the methods of the book can be developed further.
Taking account of a diverse audience, the book has been written In such
a way that it is simple at the beginning and more technical in further
chapters, so that it is accessible to graduate students in relevant areas
and mathematicians without prior knowledge of finance or economics.
I I I M H HI 4*44 I
www.worldscientific.com
4955 he J!

MX - Non Gaussian Merton Black Scholes Theory

Uploaded by

Copyright:

Available Formats

MX - Non Gaussian Merton Black Scholes Theory

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

MX - Non Gaussian Merton Black Scholes Theory

Uploaded by

Copyright:

Available Formats

Advanced Series on

Statistical Science &

Applied Probability Vol.9

Editor: Ole E. Barndorff-Nielsen

Statistical Science &

British Library Cataloguing-in-Publication Data

NON-GAUSSIAN MERTON-BLACK-SCHOLES THEORY

Printed in Singapore by Uto-Print

By now, the drawbacks of the Gaussian modelling in Financial Markets

bonds, optimal leverage, optimal timing of investment under uncertainty

simplest class of processes with stationary independent increments if the

0.0.1 General notation

xa =xtl • • • < " , da =8?1---d*~, Da = Df1---D%".

1.3.2 Discrete time models with a discrete space of states:

Chapter 2 Levy processes 39

Chapter 3 Regular Levy Processes of Exponential t y p e in

3.1.2.4 Graphs of characteristic exponents and proba-

Chapter 4 Pricing and hedging of contingent claims of Euro-

4.2.4 The problem of the model fitting and evaluation . . . . 110

Chapter 5 Perpetual American options 121

5.4.1 Put-like options 143

Chapter 6 American options: finite time horizon 151

Chapter 7 First-touch digitals 165

Chapter 8 Barrier options 185

8.2.1 Reduction to the boundary problem for the Generalized

Chapter 9 Multi-asset contracts 199

Chapter 10 Investment under uncertainty and capital accu-

10.3 Capital accumulation under RLPE 225

Chapter 11 Endogenous default and pricing of the corporate

Chapter 12 Fast pricing of European options 255

Chapter 13 Discrete time models 267

Chapter 14 Feller processes of normal inverse Gaussian

Chapter 15 Pseudo-differential operators with constant sym-

15.4.5 Restriction to a hyperplane 309

15.11Parabolic equations in spaces with exponential weights 358

Chapter 16 Elements of calculus of pseudodifferential oper-

1.1 The Gaussian Merton-Black-Scholes theory

1.1.1 Strong points

(i) the theory has a wide range of applications;

(If a model satisfies (ii), or better, (ii)-(iv), economists call it "tractable").

The Merton-Black-Scholes theory is so popular among practitioners and

Translated from Russian

DAX 30 FTSE 100

Fig. 1.1 Log-densities for major indices

1.1.2.1 Fat tails and anomalous skewness and kurtosis

For a Gaussian p, A3 = 0 and K = 3, whereas for empirical p(x)dx =

1.1.2.2 The Black-Scholes market and formula

FcMSt,t) = StNfr) - Ke~rTN{d2), (1.2)

where r = T — t is the time to expiry,

N{x) = -J= / S e't2'2dt

1.1.2.3 Volatility smile and volatility surface

Fig. 1.2 Volatility smile

1.1.3 Remedies for the MBS-theory

1.1.3.2 Jump-diffusion models and stochastic volatility models

long range dependence typically observed in the financial markets, which

1.1.3.3 Levy processes

1.2 Regular Levy Processes of Exponential type

1.2.1 Characteristic function and exponent, Levy measure

where a > 0 and 7 S R are constants, and II is a measure on R \ {0}