Optimal Trading Strategies Under Arbitrage

Johannes Karl Dominik Ruf

Submitted in partial fulllment of the
Requirements for the degree
of Doctor of Philosophy
in the Graduate School of Arts and Sciences
COLUMBIA UNIVERSITY
2011
c _2011
Johannes Karl Dominik Ruf
All Rights Reserved
ABSTRACT
Optimal Trading Strategies Under Arbitrage
Johannes Karl Dominik Ruf
This thesis analyzes models of nancial markets that incorporate the possibility of
arbitrage opportunities. The rst part demonstrates how explicit formulas for opti-
mal trading strategies in terms of minimal required initial capital can be derived in
order to replicate a given terminal wealth in a continuous-time Markovian context.
Towards this end, only the existence of a square-integrable market price of risk
(rather than the existence of an equivalent local martingale measure) is assumed.
A new measure under which the dynamics of the stock price processes simplify is
constructed. It is shown that delta hedging does not depend on the no free lunch
with vanishing risk assumption. However, in the presence of arbitrage opportu-
nities, nding an optimal strategy is directly linked to the non-uniqueness of the
partial dierential equation corresponding to the Black-Scholes equation. In order
to apply these analytic tools, sucient conditions are derived for the necessary
dierentiability of expectations indexed over the initial market conguration. The
phenomenon of bubbles, which has been a popular topic in the recent academic
literature, appears as a special case of the setting in the rst part of this thesis.
Several examples at the end of the rst part illustrate the techniques contained
therein.
In the second part, a more general point of view is taken. The stock price
processes, which again allow for the possibility of arbitrage, are no longer assumed
to be Markovian, but rather only Ito processes. We then prove the Second Funda-
mental Theorem of Asset Pricing for these markets: A market is complete, meaning
that any bounded contingent claim is replicable, if and only if the stochastic dis-
count factor is unique. Conditions under which a contingent claim can be perfectly
replicated in an incomplete market are established. Then, precise conditions un-
der which relative arbitrage and strong relative arbitrage with respect to a given
trading strategy exist are explicated. In addition, it is shown that if the market
is quasi-complete, meaning that any bounded contingent claim measurable with
respect to the stock price ltration is replicable, relative arbitrage implies strong
relative arbitrage. It is further demonstrated that markets are quasi-complete, sub-
ject to the condition that the drift and diusion coecients are measurable with
respect to the stock price ltration.
Contents
Contents
Acknowledgments iii
Chapter 1: Outline of Thesis 1
Chapter 2: The Markovian Case 4
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Market model and market price of risk . . . . . . . . . . . . . . . . 9
2.3 Strategies, wealth processes and arbitrage opportunities . . . . . . . 17
2.4 Optimal strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4.1 Optimizing a given strategy . . . . . . . . . . . . . . . . . . 23
2.4.2 Hedging of contingent claims . . . . . . . . . . . . . . . . . . 31
2.4.3 Smoothness of hedging price . . . . . . . . . . . . . . . . . . 36
2.5 Change of measure . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.6 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.8 Condition that hedging price solves a PDE . . . . . . . . . . . . . . 66
Chapter 3: Completeness and Relative Arbitrage 70
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.2 Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.2.1 Market model . . . . . . . . . . . . . . . . . . . . . . . . . . 75
i
Contents
3.2.2 Market prices of risk and stochastic discount factors . . . . . 76
3.2.3 Trading strategies and claims . . . . . . . . . . . . . . . . . 77
3.3 Existence of (super-)replicating trading strategies . . . . . . . . . . 78
3.4 Completeness and Second Fundamental Theorem of Asset Pricing . 91
3.5 Relative arbitrage and strong relative arbitrage . . . . . . . . . . . 94
3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
Bibliography 100
ii
Acknowledgments
Acknowledgments
I am very grateful to Professor Ioannis Karatzas for serving as my advisor in my
doctoral studies. His scholarship and integrity have set an ideal to which I inspire.
Professor Karatzas is not only an excellent researcher with deep insights, but also
a great instructor. I thank him for always sharing his ideas, for his generosity in
oering fast, helpful, and constructive feedback, for introducing me to other junior
and senior researchers, for supporting me in all my academic projects, and, last but
not least, for his friendliness in our academic and non-academic interactions.
I am also deeply indebted to Professors Rama Cont, Dilip Madan, Philip
Protter, and Josef Teichmann for agreeing to serve on my dissertation committee.
I have had many enlightening discussions with each of them over the course of my
graduate studies, which sharpened my understanding of the subject matter of this
thesis.
Without Dr. Bob Fernholz at Intech, there would be no Stochastic Portfolio
Theory, which is the fundamental building block of this thesis. I therefore thank
Bob Fernholz for sharing his insights and for further creating the space and atmo-
sphere for our many meetings at Intech, where research is done at its nest: a free
ow and exchange of ideas.
Dr. Adrian Banner, Dr. Daniel Fernholz, Dr. Tomoyuki Ichiba, Dr. Phi Long
iii
Acknowledgments
Nguen-Tranh, Dr. Vassilios Papathanakos, Radka Pickova, Subhankar Sadhukhan,
and Dr. Phillip Whitman have regularly attended these meetings. Their feedback
on my presentations and the exchange of ideas with them have greatly improved
the results presented here.
Besides the previously mentioned colleagues, I have been discussing this the-
sis with friends and colleagues within the Department of Statistics at Columbia
University, mainly with Emilio Seijo and Dr. Li Song, and outside of Columbia
University, primarily with Professor Erhan Bayraktar, Professor Kostas Kardaras,
Dr. Sergio Pulido, Winslow Strong, Professor Johan Tysk, and Dr. Hao Xing. Dur-
ing a visit for the Minerva lectures at Columbia University, Professor Hans Follmer
spent valuable time reviewing an early draft of the rst part of this thesis. Two
referees and an associate editor of the Journal of Mathematical Finance also con-
tributed very helpful comments on a paper that serves as the foundation for the
rst part of this thesis. I am deeply indebted to all of them and to the audiences
in various conferences and seminars for giving me the opportunity to discuss this
research.
It would be harder to read this thesis if its English had not been proofread
by ve very patient friends. I thank Michael Agne, Rachel Schutt, Zach Shahn,
Ekaterina Vinkovskaya, and, in particular, Ashley Grith.
This thesis is not directly, but indirectly, inuenced by joint work with amaz-
ing co-authors. In particular, I thank Professor Peter Carr, Dr. Travis Fisher,
Professor Carlos Oyarzun, Professor Matthias Scherer, Amal Moussa, Tyler Mc-
Cormick, Professor Andrew Gelman, Professor Tian Zheng, and my other co-
authors for sharing their knowledge with me and for many good collaborations.
iv
Acknowledgments
I am grateful to my colleagues at Commerzbank, D-Fine, J.P. Morgan, and
Morgan Stanley for making my internships such a valuable experience. In partic-
ular, I express my gratitude to Dr. J urgen Hakala, Andreas Weber, Dr. Jochen
Alberty, Dr. Jorn Rank, Dr. Lei Fang, Dr. Xiaolan Zhang, Daniel Leiter, Michael
Jansen, Michael Sternberg, Dr. Len Umantz, and Tyler Ward.
Without seeing the passion of my mathematics and nance teachers, I would
never have started my doctoral studies. I am indebted to Inge Schneider, Siegfried
Ranz, Dr. Josef Krutel, Helmut Freischmidt, Herbert Sittenberger, Ottmar Fis-
cher, Elvira Rendle, and Professors Werner Balser, Werner L utkebohmert, Volker
Schmidt, Franz Schweiggert, Frank Stehling, Karsten Urban, R udiger Kiesel, Ul-
rich Stadtm uller, Amanda Adkisson, Kerry Back, Dante DeBlassie, and Jean-Luc
Guermond.
I am grateful to the Cusanuswerk, the Fulbright Commission, the National
Science Foundation (DMS Grant 09-05754), and Columbia University for both their
nancial and intellectual support.
This thesis ends my doctoral studies at the Department of Statistics at
Columbia University. I am grateful to Professors David Madigan and Richard
Davis for creating such a friendly and creative environment in the department.
I thank Dood Kalicharan, Faiza Bellounis, and Anthony Cruz for their always
helpful support in all personal and administrative matters. All faculty members
at Columbia University have always been very supportive; in particular, I would
like to thank Professors Jan Vecer and Zhiliang Ying, both of whom have served
on my oral committee, and Professors Jose Blanchet, Mark Brown, Rama Cont,
Michael Hogan, Ioannis Karatzas, Steve Kou, Martin Lindquist, Shaw-Hwa Lo,
v
Acknowledgments
Panagiota Daskalopoulos, Michael Sobel, Victor De la Pe na, Philip Protter, Daniel
Rabinowitz, Bodhisattva Sen, and Tian Zheng. I also would like to thank all of
the intelligent and passionate students whom I had the opportunity to teach or to
instruct in the capacity of their teaching assistant.
My friends at Columbia University have challenged me with interesting ques-
tions and with their achievements, and most importantly, have been excellent
friends. Each encounter has been very important to me. I thank all of them and
mention a necessarily very incomplete, but hopefully representative, list: Xiaodong
Li, Amal Moussa, Shawn Simpson, Li Song, Xiaoru Wu, Priya Dev, Stacey Han-
cock, Samantha Sartori, Maggie Wang, Tomoyuki Ichiba, Alexandra Chronopoulou,
Georgios Fellouris, Ragnheidur Haraldsdottir, Petr Novotny, Libor Pospisil, Rachel
Schutt, Tyler McCormick, Ivy Ng, Subhankar Sadhukhan, Emilio Seijo, Heng Liu,
Radka Pickova, Tony Sit, Ekaterina Vinkovskaya, Junyi Zhang, Michael Agne,
Stephanie Zhang, Zach Shahn, Gerado Hernandez, Aleks Jakulin, Siliva Ardila,
Juan Campos, Rodrigo Carrasco, Romain Deguest, Rouba Ibrahim, Xuedong He,
Tulia Herrera, Yu Hang Kan, Olga Kolesnikova, Arseniy Kukanov, Xianhua Peng,
Matthieu Plumettaz, Daniela Wachholz, Cecilia Zenteno, Yori Zwols, Avishek Ad-
hikari, Danielle Cohen, Amit Duvshani, Hugo Escobar, Mei-Ho Lee, Adam Shavit,
Joan Shavit, Casandra Silva Sibiln, Esther Quintero, Fabio Dominguez, Ashley
Grith, Olympia Hadjiliadis, Eva Jozsef, Gabor Jozsef, Helena Kauppila, Robert
Neel, Oya Okman, Aly Sanoh, Bjarni Torfason, Olger Twyner, Jin Wang, and Yun
Zhou.
Over the last several years, I have also been supported by many friends out-
side of Columbia University. I am especially grateful to Erhan Bayraktar, Matthias
vi
Acknowledgments
B uhlmeier, Jan Cameron, David Pereda Cubian, Tanja Djokic, Dan Freeman, Jo-
hannes Eberle, Selim Gokay, Sebastian Grimm, Philipp Illeditsch, Bernadette Kalz,
Ulrike Kiechle, Dieter Kiesenbauer, Samuel Knossow, Dimitrije Kostic, Raphael
Laguerre, Marco Maggis, Jan Mai, Eva Mair, Cassia Marchon, Christian Markus,
John McFall, Meike Meuser, Yuka Mitsuhashi, Nicola Nowak, Carlos Oyarzun,
Soumik Pal, Svetlana Poznanovik, Lavinia Ostafe, Sergio Pulido, Fabiola Rangel,
Simone Rechtsteiner, Klaus Reuter, Michael Salomon, Matthias Scherer, Markus
Schierle, Kai Schonle, Halina Tegetmeyer, Katharina Zaglauer, Hao Xing, and Gor-
dan

Zitkovc.
Most importantly, I thank my family for their support, especially my brothers
Benedikt, David, Emanuel, and Simon, my parents Margit and Franz Ruf, and my
grandparents Maria and Walter Vetter, Karl and Walburga Ruf. They have always
given me unconditional support in all of my endeavors and their faith in me has
always been a driving and inuential force in my life. They might not understand
everything, but nothing in this thesis would have been written without them.
vii
To Margit and Franz, Benedikt, David, Emanuel, and Simon
To my friends and colleagues
viii
Chapter 1. Outline of Thesis 1
Chapter 1
Outline of Thesis
This thesis consists of two main chapters. Chapter 2 treats Markovian market
models. We illustrate how optimal trading strategies can be computed using the
classical idea of delta hedging. This chapter generalizes the results in Fernholz and
Karatzas (2010) and contains the results of the paper
Ruf, J. (2011+). Hedging under arbitrage. Mathematical Finance, forth-
coming.
That paper focuses on replicating European-style contingent claims. Chapter 2
contains additional results related to optimizing a given trading strategy.
The models studied in Chapter 2 can be considered complete, meaning that
any contingent claim studied here can be replicated. It is also of interest, however, to
study conditions under which a contingent claim can be generated in an incomplete
model, which is the subject of Chapter 3. More specically, it is shown that the
question of completeness and the question of existence of arbitrage can be addressed
separately from one another. Chapter 3 is based on the paper
Ruf, J. (2011). Completeness and arbitrage.
Chapter 1. Outline of Thesis 2
Both chapters deal with markets that do not satisfy the no free lunch with
vanishing risk (NFLVR) assumption, which is typically a standard assumption
in the literature, and which is discussed in the introductions of the following two
chapters. The replicability of contingent claims is also studied in both chapters.
Chapter 2 restricts its analysis to Markovian models, for which explicit formulas
for the replication can be easily derived. Chapter 3 derives existence results for a
more general class of models.
The following two chapters are self-contained and therefore exhibit some
redundancies. Notational inconsistencies in the two chapters have been minimized,
but not entirely eliminated.
The work presented in this thesis was primarily motivated by a desire to
better understand the models studied in Stochastic Portfolio Theory (SPT). SPT
is not predicated upon the no-arbitrage assumption, but instead models nancial
markets and studies the existence of arbitrage opportunities that arise; see the
survey paper by Bob Fernholz, who developed the eld, and by Ioannis Karatzas,
a major contributor, for an overview of the recent developments in Fernholz and
Karatzas (2009).
In some sense, because it demonstrates that the concept of a price exists
even in markets studied by SPT, which may allow for the presence of arbitrage op-
portunities, this thesis unies SPT and the classical theory of Financial Mathemat-
ics. Furthermore, and as previously mentioned, the characterization of replicable
claims and the Second Fundamental Theorem of Asset Pricing, which connects the
replicability of any contingent claim in the economy with the uniqueness of some
pricing operator, can be proven without having to exclude arbitrage.
Thus, SPT has claried which assumptions are necessary and which assump-
tions are extraneous for relevant tasks in Mathematical Finance, such as the pricing
and replication of contingent claims. In particular, the assumption of NFLVR is of-
Chapter 1. Outline of Thesis 3
ten, despite its mathematical convenience, too strong an assumption. As discussed
in Chapter 2, excluding unbounded prots with bounded risks is often a sucient
assumption, and also leads to a pricing measure, but one which may no longer be
equivalent to the original one.
On the other hand, the many strong tools developed, in particular by Freddy
Delbaen and Walter Schachermayer in the 1990s, for the so-called classical no-
arbitrage theory, proved invaluable to this thesis development. Many of the proofs
within this thesis, especially those in Chapter 3, begin by transforming models
with possible arbitrage opportunities into the no-arbitrage framework, in order to
thereafter apply the powerful tools of the classical theory of Financial Mathematics.
The academic literature has documented that the no-arbitrage condition is
not necessary for the existence of well-dened option prices. Karatzas et al. (1991b)
were among the rst to characterize replicable claims via duality methods, which
they accomplish without relying on the existence of equivalent local martingale
measures. Eckhard Platen developed the Benchmark Approach to Mathematical
Finance, which establishes the real-world pricing formula, which also does not
require an arbitrage-free market, as one of its important concepts; see Platen (2006).
Furthermore, in Section 10 of the previously mentioned survey article by Fernholz
and Karatzas (2009), the martingale representation theorem is applied, which yields
the result that claims can be replicated under certain conditions. This thesis builds
on all of these results and generalizes them to less restrictive assumptions.
Chapter 2. The Markovian Case 4
Chapter 2
The Markovian Case
2.1 Introduction
In a nancial market, an investor usually has several trading strategies at her
disposal to obtain a given wealth at a specied point in time. For example, if
the investor wanted to cover a short-position in a given stock tomorrow at the
cheapest cost today, buying the stock today is generally not optimal, as there may
be a trading strategy requiring less initial capital that still replicates the exact
stock price tomorrow. In this chapter, we show that optimal trading strategies, in
the sense of minimal required initial capital, can be represented as delta hedges.
We generalize the results of Fernholz and Karatzas (2010)s paper On optimal
arbitrage, in which specically the market portfolio is examined, to a wide class
of terminal wealths that can be optimally replicated by delta hedges.
We shall not restrict ourselves only to markets satisfying the No free lunch
with vanishing risk (NFLVR) or, more precisely, the No arbitrage for general
admissible integrands (NA) condition.
1
Thus, we cannot rely on the existence
1
We refer the reader to Delbaen and Schachermayer (2006) for a thorough introduction to NA,
NFLVR and other notions of arbitrage. Since we shall assume the existence of a square-integrable
market price of risk, we implicitly impose the condition that NFLVR fails if and only if NA fails;
Chapter 2. The Markovian Case 5
of an equivalent local martingale measure, which we otherwise would have done.
However, we shall construct another probability measure to take the place of the
risk-neutral measure. We do not exclude arbitrage a priori for several reasons.
First, we cannot always assume the existence of a statistical test that relies upon
stock price observations to determine whether an arbitrage opportunity is present,
as illustrated in Example 3.7 of Karatzas and Kardaras (2007). In such a situation,
a typical agent, who needs to rely on a path-by-path analysis, would not be aware
of an arbitrage opportunity and could consequently not benet from it. Second,
examining possible arbitrage opportunities, rather than excluding them a priori,
is of interest in itself. Further arguments and empirical evidence supporting the
consideration of models without an equivalent local martingale measure are dis-
cussed in Section 0.1 of Kardaras (2008) and Section 1 of Platen and Hulley (2008).
A model of economic equilibrium for such models is provided in Loewenstein and
Willard (2000a). In the spirit of these papers, we shall impose some restrictions
on the arbitrage opportunities and exclude a priori models that imply unbounded
prot with bounded risk, which can be recognized by a typical agent.
This chapter is set in the framework of Stochastic Portfolio Theory. For an
overview of this eld, we recommend the reader consult the monograph by Fern-
holz (2002) and the survey paper by Fernholz and Karatzas (2009). This chapter
contributes to Stochastic Portfolio Theory a clearer understanding of pricing and
hedging and its relation to several other current research directions, such as the
Benchmark Approach, developed by Eckhard Platen and co-authors in a series
of papers. Indeed, we generalize some of the Benchmark Approachs results here
and provide tools to compute the so-called real-world prices of contingent claims
under that approach. The monograph by Platen and Heath (2006) provides an
excellent overview of the Benchmark Approach .
see Proposition 3.2 of Karatzas and Kardaras (2007).
Chapter 2. The Markovian Case 6
Stochastic Portfolio Theory is a suitable framework for studying so-called
relative arbitrage opportunities: Given a specic strategy, are there other strate-
gies that outperform the original one? A related important observation is made
in Fernholz et al. (2005): If one assumes the market to be diverse, that is, if
no company can take over the whole market, and to have a bounded volatility
structure, then a relative arbitrage opportunity with respect to the market port-
folio exists. The existence of relative arbitrage would not conict per se with the
NFLVR assumption as, for example, the existence of admissible suicide strategies
in arbitrage-free markets
2
shows. Another example is a stock price that is a strict
local martingale. Then there exists a relative arbitrage opportunity with respect
to this stock. From this point of view, it seems articial that one should exclude
relative arbitrage with respect to the money market and we shall also explicitly
study some models in which such arbitrage is possible. Here, our analysis extends
parts of the work done by Delbaen and Schachermayer (1995a) about the Bessel
process and its reciprocal. Depending on which process is chosen to model the
stock price, either there is arbitrage possible with respect to the money market or
there is arbitrage possible with respect to the stock. Both cases can be treated in
a uniform way provided that one abstains from making the NFLVR assumption.
There have been several recent papers treating the subject of bubbles;
a very incomplete list consists of the work by Loewenstein and Willard (2000b),
Cox and Hobson (2005), Heston et al. (2007), Jarrow et al. (2007; 2010), Pal and
Protter (2010), and Ekstrom and Tysk (2009). A bubble is usually dened within
a model that guarantees NFLVR as the dierence between the market price of a
tradeable asset and its smallest hedging price. A given asset has a bubble if and
only if there exists a relative arbitrage opportunity with respect to this asset. The
analysis here includes the case of bubbles, but is more general, as it also allows for
2
See Section 6.1 of Harrison and Pliska (1981) or Section 1.2 of Karatzas and Shreve (1998)
for an example.
Chapter 2. The Markovian Case 7
models that imply arbitrage: Models including bubbles rely on an equivalent local
martingale measure, which diers from models that allow for arbitrage, in which a
more complicated change of measure not relying on Girsanovs theorem is necessary.
To wit, while the bubbles literature concentrates on a single stock whose price
process is modeled as a strict local martingale, we consider markets with several
assets with the stochastic discount factor itself being represented by a (possibly
strict) local martingale. In the presence of an asset with a bubble, our contribution
is limited to the bubbles representation as a relative arbitrage opportunity and to
the explicit representation of the optimal replicating strategy. We also discuss the
reciprocal of the three-dimensional Bessel process as the standard example for a
bubble.
Two phenomena, in particular, have been repeatedly discussed in the bubbles
literature: The lack of a unique solution of the corresponding Black-Scholes PDE
for an asset and the failure of the classic put-call parity; see, for example, Cox and
Hobson (2005). Both these observations also hold for the more general arbitrage
situation. We characterize the hedging price as the minimal nonnegative solution
for the Black-Scholes PDE and suggest a modied put-call parity, which generalizes
to models with arbitrage opportunities.
We set up our analysis in a continuous-time Markovian context; to wit, we
focus on stock price processes whose mean rates of return and volatility coecients
only depend on time and on the current market conguration. Furthermore, we
concentrate, on (possibly time-inhomogeneous) strategies that depend only on the
current stock prices. This restriction to a Markovian model is certainly not the most
general one, but it provides us with a rich setup, which provides valuable insight
into the most interesting strategies. For such a model and a given Markovian
trading strategy, we nd an optimal strategy, by which we mean an investment
decision rule that uses minimal initial capital but that, nevertheless, leads to the
Chapter 2. The Markovian Case 8
identical terminal wealth as produced by the original strategy. Since we do not
rely on a martingale representation theorem, we can allow for a larger number of
driving Brownian motions than the number of stocks, which generalizes the ideas
of Fernholz and Karatzas (2010) to not only a larger set of strategies, but also to a
broader set of models for the specic case of the market portfolio.
Next, we prove that a classical delta hedge yields the the cheapest hedging
strategy for European contingent claims. This is of course well-known in the case
where an equivalent local martingale measure exists and is extended here to models
that allow for arbitrage opportunities and that are not necessarily complete. In
this context, we provide sucient conditions to ensure the dierentiability of the
hedging price, generalizing results by Heath and Schweizer (2000), Janson and Tysk
(2006), and Ekstrom and Tysk (2009). This set of conditions is also applicable
to models satisfying the NFLVR assumption. Because the computations for the
optimal trading strategy under the real-world measure are often too involved and
because we cannot always rely on an equivalent local martingale measure, we derive
a non-equivalent change of measure and formulas based thereon, as illustrated, for
instance, by a new generalized Bayes rule.
The next section introduces the market model, discusses dierent notions of
arbitrage, and contains an initial result concerning the independence of some price
candidates from the choice of the market price of risk. Section 2.3, after dening
strategies and their associated wealth processes, concludes the discussion of arbi-
trage. In Section 2.4, we present some of our rst main results, including (1) the
precise representation of an optimal strategy designed to either replicate a given
wealth process or hedge a non path-dependent European claim and (2) sucient
conditions for the dierentiability of the hedging price. A modied put-call parity
follows directly. In Section 2.5, we prove the next main result of this chapter, which
is a change to a non-equivalent probability measure that simplies computations.
Chapter 2. The Markovian Case 9
This section also contains several other corollaries such as a generalized Bayes rule
and a discussion of a change-of-numeraire technique. Section 2.6 then provides sev-
eral examples that illustrate various aspects of the chapters results and Section 2.7
draws the conclusions. Finally, Section 2.8 serves as an appendix to the chapter
and discusses a sucient condition for a technical assumption made in Section 2.4.
2.2 Market model and market price of risk
In this section, we introduce the market model, discuss the existence of a market
price of risk and dene the stochastic discount factor. We assume the perspective of
a small investor who takes positions in a frictionless nancial market with nite time
horizon T in order to accumulate wealth or hedge a nancial claim. By small we
mean that the investors trading activities have no impact on prices. Equivalently,
the investor is a price-taker and the stock prices are given exogenously.
We use the notation R
d
+
:= s = (s
1
, . . . , s
d
)
T
R
d
, s
i
> 0, for all i =
1, . . . , d and assume a market in which the stock price processes are modeled as
positive continuous Markovian semimartingales. That is, we consider a nancial
market S() = (S
1
(), . . . , S
d
())
T
of the form
dS
i
(t) =S
i
(t)
_

i
(t, S(t))dt +
K

k=1

i,k
(t, S(t))dW
k
(t)
_
(2.1)
for all i = 1, . . . , d and t [0, T] starting at S(0) R
d
+
and a money market B().
Here : [0, T] R
d
+
R
d
denotes the mean rate of return and : [0, T] R
d
+

R
dK
denotes the volatility. We assume that both functions are measurable.
For the sake of convenience we only consider discounted (forward) prices
and set the interest rate constant to zero; that is, B() 1. The ow of in-
formation is modeled as a right-continuous ltration F = T(t)
0tT
such that
W() = (W
1
(), . . . , W
K
())
T
is a K-dimensional Brownian motion with indepen-
dent components. In Section 2.5, we impose more conditions on the ltration F and
Chapter 2. The Markovian Case 10
the underlying probability space . For the moment, we assume that all stochastic
integrals that appear are measurable with respect to the ltration F. The under-
lying measure and its expectation shall be denoted by P and E, respectively. The
current state of the market S(0) should be clear from the context and so we shall
omit specifying S(0) as an index for measures and expectations in most cases.
We only consider those mean rates of return (, ) and volatilities (, ) that
imply the stock prices S
1
(), , S
d
() exist and are unique and strictly positive.
More precisely, denoting the covariance process of the stocks in the market by
a(, ) (, )
T
(, ), that is,
a
i,j
(t, S(t)) :=
K

k=1

i,k
(t, S(t))
j,k
(t, S(t))
for all i, j = 1, . . . , d and t [0, T], we impose the almost sure integrability condition
d

i=1
_
T
0
([
i
(t, S(t))[ + a
i,i
(t, S(t))) dt < .
Under this condition, the stock prices S
1
(), . . . , S
d
() can be expressed as
S
i
(t) =S
i
(0) exp
__
t
0
_

i
(u, S(u))
1
2
a
i,i
(u, S(u))
_
du+ (2.2)
+
K

k=1
_
t
0

i,k
(u, S(u))dW
k
(u)
_
> 0
for all i = 1, . . . , d and t [0, T]. Furthermore, we assume the existence of a market
price of risk, which generalizes the concept of the Sharpe ratio to several dimensions
by setting the risk factors W
k
() in relation to the mean rates of return
i
(, ).
Denition 1 (Market price of risk). A market price of risk is a progressively mea-
surable process (), which maps the volatility structure (, ) onto the mean rate
of return (, ). That is,
(t, S(t)) = (t, S(t))(t) (2.3)
holds almost surely for all t [0, T].
Chapter 2. The Markovian Case 11
Furthermore, we assume that () is square-integrable, to wit,
_
T
0
|(t)|
2
dt < (2.4)
almost surely. The existence of a market price of risk is a central assumption in
both the Benchmark Approach (see Chapter 10 of Platen and Heath 2006) and in
Stochastic Portfolio Theory (see Section 6 of Fernholz and Karatzas 2009). This
assumption enables us to discuss hedging prices, as we do throughout this thesis.
Similar assumptions have been discussed in the economic literature. For example,
in the terminology of Loewenstein and Willard (2000a), the existence of a square-
integrable market price of risk excludes cheap thrills but not necessarily free
snacks. Theorem 2 of Loewenstein and Willard (2000a) shows that a market with
a square-integrable market price of risk is consistent with an equilibrium where
agents prefer more to less. We discuss the connection between a market price
of risk and its square-integrability with various no-arbitrage notions in Remark 1
below.
Based on the market price of risk, we can now dene the stochastic discount
factor as
Z

(t) := exp
_

_
t
0

T
(u)dW(u)
1
2
_
t
0
|(u)|
2
du
_
(2.5)
with dynamics
dZ

(t) =
T
(t)Z

(t)dW(t) (2.6)
for all t [0, T]. In classical no-arbitrage theory, Z

() represents the Radon-

Nikodym derivative that translates the real-world measure into the generic risk-
neutral measure with the money market as the underlying. Since we do not want
to impose NFLVR a priori in this thesis, but are rather interested in situations in
which NFLVR does not necessarily hold, we shall not assume that the stochastic
Chapter 2. The Markovian Case 12
discount factor Z

() is a true martingale. Thus, we can only rely on a local mar-

tingale property of Z

(). Cases where Z

() is only a local martingale have, for

example, been discussed by Karatzas et al. (1991b), Schweizer (1992), in the Bench-
mark Approach starting with Platen (2002) and Heath and Platen (2002a;b) and in
Stochastic Portfolio Theory; see, for example, Fernholz et al. (2005) and especially,
Fernholz and Karatzas (2010). On the other hand, much eort has been made to
strengthen Novikov (1972)s condition to ensure that the stochastic discount factor
Z

() be a true martingale, for example by Wong and Heyde (2004), Hulley and
Platen (2009), Mijatovic and Urusov (2009), and the literature therein.
A market price of risk () does not have to be uniquely determined. Unique-
ness is intrinsically connected to completeness, as we shall see in Chapter 3, and we
need not assume it. In general, innitely many market prices of risk may exist. To
illustrate, think of a model with d = 1, K = 2, (, ) 0 and (, ) (1, 1). Then,
for any y R, the constant process () (y, y)
T
is a square-integrable market
price of risk. Another example of this non-uniqueness follows the next proposition.
We observe that the existence of a square-integrable market price of risk
implies the existence of a Markovian square-integrable market price of risk. To see
this, we dene (, ) :=
T
(, )((, )
T
(, ))

(, ), where denotes the Moore-

Penrose pseudo-inverse of a matrix. Given the existence of any market price of risk,
we know from the theory of least-squares estimation that (, ) is also a market
price of risk. Furthermore, we have |(t, S(t))|
2
|(t)|
2
for all t [0, T] almost
surely for any market price of risk (), which yields the square-integrability of
(, ).
The next proposition shows that any square-integrable Markovian market
price of risk maximizes the random variable that will later be a candidate for a
hedging price. We denote by T
S
() the augmented ltration generated by the
stock price process. We emphasize that the next result only holds so long as the
Chapter 2. The Markovian Case 13
terminal payo M is T
S
(T)-measurable. We generalize the following proposition
in Theorem 6 of Chapter 3.
Proposition 1 (Role of Markovian market price of risk). Let M 0 be a random
variable measurable with respect to T
S
(T) T(T). Let () denote any square-
integrable market price of risk and (, ) any square-integrable Markovian market
price of risk. Then, with
M

(t) := E
_
Z

(T)
Z

(t)
M

T(t)
_
and M

(t) := E
_
Z

(T)
Z

(t)
M

T(t)
_
for t [0, T], where we take the right-continuous modication
3
for each process,
we have M

() M

() almost surely. Furthermore, if both Z

() and Z

() are
T
S
(T)-measurable, then Z

(T) Z

(T) almost surely.

Proof. Due to the right-continuity of M

() and M

() it suces to show for all

t [0, T] that M

(t) M

(t) almost surely. We dene c() := () (, S()). For

the sequence of stopping times

n
:= T inf
_
t [0, T] :
_
t
0
c
2
(s)ds n
_
,
where n N, we set c
n
() := c()1
n
and observe that
Z

(T)
Z

(t)
=
Z
c
(T)
Z
c
(t)
exp
_

_
T
t

T
(u, S(u))(dW(u) + c(u)du)

1
2
_
T
t
|(u, S(u))|
2
du
_
= lim
n
Z
c
n
(T)
Z
c
n
(t)
exp
_

_
T
t

T
(u, S(u))(dW(u) + c
n
(u)du)
1
2
_
T
t
|(u, S(u))|
2
du
_
with Z
c
() and Z
c
n
() dened as in (2.5). The limit holds almost surely since both
v() and (, ) are square-integrable, which again yields the square-integrability of
3
See Theorem 1.3.13 of Karatzas and Shreve (1991).
Chapter 2. The Markovian Case 14
c(). Since
_
T
0
c
2
n
(t)dt n, Novikovs Condition (see Proposition 3.5.12 of Karatzas
and Shreve 1991) yields that Z
c
n
() is a martingale. Now, Fatous lemma, Girsanovs
theorem and Bayes rule (see Chapter 3.5 of Karatzas and Shreve 1991) yield
M

(t) liminf
n
E
Q
n
_
exp
_

_
T
t

T
(u, S(u))dW
n
(u)
1
2
_
T
t
|(u, S(u))|
2
du
_
M

T(t)
_
,
(2.7)
where dQ
n
() := Z
c
n
(T)dP() is a probability measure, E
Q
n
its expectation operator,
and W
n
() := W() +
_

0
c
n
(u)du a K-dimensional Q
n
-Brownian motion. Since
(, S())c
n
() 0 we can replace W() by W
n
() in (2.1). This yields that the
process S() has the same dynamics under Q
n
as under P. Furthermore, both
(, S()) and M have, as functionals of S(), the same distribution under Q
n
as
under P. Therefore, we can replace the expectation operator E
Q
n
by E in (2.7) and
obtain the rst part of the statement. The last inequality of the statement follows
from setting M = 1
Z

(T)>Z

(T)
and observing that M must equal zero almost
surely.
We remark that the inequality M

() M

() can be strict. As an example,

choose M = 1 and a market with one stock and two Brownian motions, to wit, d = 1
and K = 2. We set (, ) 0, (, ) (1, 0) and observe that (, S()) (0, 0)
T
is
a Markovian market price of risk. Another market price of risk () (
1
(),
2
())
T
is dened via
1
() 0, the stochastic dierential equation
d
2
(t) =
2
2
(t)dW
2
(t)
for all t [0, T] and
2
(0) = 1. That is,
2
() is the reciprocal of a three-dimensional
Bessel process starting at one. Itos formula yields Z

()
2
(), which is a
strict local martingale (see Exercise 3.3.36 of Karatzas and Shreve 1991), and thus
M

(0) = E[Z

(T)] < 1 = E[Z

(T)] = M

(0).
Chapter 2. The Markovian Case 15
Under the assumption that an equivalent local martingale measure exists,
Theorem 12 of Jacka (1992), Theorem 3.2 of Ansel and Stricker (1993) or Theo-
rem 16 Delbaen and Schachermayer (1995c) show that a contingent claim can be
hedged if and only if the supremum over all expectations of the terminal value of
the contingent claim under all equivalent local martingale measures is a maximum.
In our setup, we also observe that the supremum over all M

(0) in the last propo-
sition is a maximum, attained by any Markovian market price of risk. Indeed, we
shall prove in Theorem 2 that, under weak analytic assumptions, claims of the form
M = p(S(T)) can be hedged. The general theory lets us conjecture that all claims
measurable with respect to T
S
(T) can be hedged. Theorem 6 of Chapter 3 conrms
this conjecture.
As pointed out by Ioannis Karatzas in a personal communication (2010),
Proposition 1 might be related to the Markovian selection results, as in Krylov
(1973), Section 4.5 of Ethier and Kurtz (1986), and Chapter 12 of Stroock and
Varadhan (2006). There, the existence of a Markovian solution for a martingale
problem is studied. It is observed that a supremum over a set of expectations
indexed by a family of distributions is attained and the maximizing distribution is
a Markovian solution of the martingale problem. This potential connection needs
to be worked out in a future research project.
From this point forward, we shall always assume the market price of risk to
be Markovian. As we shall see, this choice will lead directly to the optimal trading
strategy.
Remark 1 (Market price of risk and NA, NUPBR, NIA). Proposition 3.6 of Delbaen
and Schachermayer (1994) shows (compare also Proposition 3.2 of Karatzas and
Kardaras 2007) that NFLVR holds, if and only if NA and no unbounded prot
with bounded risk (NUPBR) hold. NUPBR is also known as arbitrage of the
rst kind (compare Ingersoll 1987; Kardaras and Platen 2009) and as the BK
Chapter 2. The Markovian Case 16
property (compare Kabanov 1997; Kardaras 2010, Proposition 1.2). If NUPBR
holds, then, in particular, scalable arbitrage opportunities do not exist.
The existence of a square-integrable market price of risk guarantees the ex-
istence of a positive stochastic discount factor, which again ensures that NUPBR
holds as it is proven in Theorem 3.12 of Karatzas and Kardaras (2007). Moreover,
since it is shown in Lemma 3.1 of Delbaen and Schachermayer (1995b) that NA
holds, if and only if no immediate arbitrage (NIA) holds and the possibility to
make some prot using a credit line is excluded. However, since immediate arbi-
trage is again scalable we can also conclude that NUPBR implies NIA. Therefore, if
NUPBR holds, then NFLVR fails, if and only if the second component of NA fails,
to wit, if and only if it is possible to make some prot using a credit line. Indeed,
the application of this chapters results to the optimal hedging problem of a bond
serves to quantify exactly how much some prot is in a given model.
On the other hand, a careful analysis of Section 10 in Karatzas et al. (1991a)
or Theorems 3.5 and 3.6 in Delbaen and Schachermayer (1995b), using the fact that
the ranges of (, ) and a(, ) are identical, reveals that a necessary condition for
NIA is the existence of a market price of risk that satises an integrability condition
strictly weaker than the condition in (2.4). Furthermore, Theorem 1 of Levental and
Skorohod (1995) and Proposition 1.1 of Lyaso (2010) motivate the integrability
condition in (2.4) to prevent general scalable arbitrage opportunities.
A toy example for a market without a market price of risk P Lebesgue-
almost everywhere (and thus with scalable arbitrage) can be described by a drift
(, ) and a volatility structure (, ) such that the set / [0, T] R
d
+
dened
as / := (t, s) : (t, s) s.t. (t, s)(t, s) = (t, s) has positive measure, by which
we mean p
,
:= P(Lebesgue(t : (t, S(t)) /) > 0) > 0. We can decompose
(, ) uniquely into the sum of two vectors
1
(, ) in the range of (, ) and
2
(, )
orthogonal to its columns. Then, we have
2
(t, s) ,= 0 for all (t, s) / and
Chapter 2. The Markovian Case 17

2
T
(, )(, ) 0 always. Investing according to
2
(, ) would thus switch o the
risk factors and lead to nonnegative mean rate of return
2
T
(, )(, ) = |
2
(, )|
2
.
Investing according to such a strategy (see Section 2.3 for a precise denition) would
lead to a wealth process (as in (2.11) below) that is greater than one with probability
p
,
. This arbitrage opportunity could be leveraged arbitrarily by replacing the
strategy
2
(, ) with
2
(, ) multiplied by a constant, leading to an immediate
and unbounded prot. This line of thought, enriched with deep measure-theoretic
results, is the underlying idea for the proof of the existence of a market price of risk
under the NIA condition in Theorem 3.5 of Delbaen and Schachermayer (1995b).
2.3 Strategies, wealth processes and arbitrage op-
portunities
In this section, we introduce trading strategies, describe investors wealth processes
and dene relative arbitrage. We denote the proportion of the investors wealth
invested in the i
th
stock by
i
. The proportion of the wealth that is not invested
in stocks gets invested in the money market, which yields zero interest rate. The
next denition states this more precisely.
Denition 2 (Markovian trading strategy and associated wealth process). We call
a function : [0, T] R
d
+
R
d
a (Markovian trading) strategy and the process
V

() with dynamics
dV

(t) =
d

i=1

i
(t, S(t))V

(t)
dS
i
(t)
S
i
(t)
(2.8)
for all t [0, T] and with initial condition V

(0) = 1 its associated wealth process.

To ensure that V

() does not explode and to exclude doubling strategies we restrict

Chapter 2. The Markovian Case 18
ourselves to strategies that satisfy the integrability condition
d

i=1
_
T
0
(
i
(t, S(t))V

(t))
2
a
i,i
(t, S(t))dt < , (2.9)
and the nonnegativity condition V

(t) 0 for all t [0, T] almost surely. We shall

use for any v > 0 the notation V
v,
() vV

() and interpret v as the investors

initial capital.
For (2.8), we have used that the strategy is self-nancing; that is, no wealth
is consumed and no money is added to the wealth process from outside. To wit,
the wealth at any point of time is obtained by trading the initial wealth according
to the strategy (, ).
We have assumed that a strategy only depends on the current conguration
of the market and not on its past, in order to preserve the Markovian property
of the model. This has the economic interpretation that investment decisions are
based upon the current market environment only. It would be of interest to extend
the here presented results to a more general framework allowing for non-Markovian
stochastic processes and strategies that may depend on the past of the market,
perhaps relying on the Clark-Ocone formula (compare Karatzas and Ocone 1991).
However, we allow the strategies to be time-inhomogeneous. Denition 2 allows for
functionally generated portfolios (compare Remark 7 of Section 2.5) and hedging
strategies for non-path dependent options (European and American style). Dening
for all i = 1, . . . , d and t [0, T] the functions h
i
(t, ) := V
v,
(t)
i
(t, S(t)) as the
dollar value and
i
(t, ) := h
i
(t, )/S
i
(t) as the number of shares held, (2.8) can be
written in the more familiar forms
dV
v,
(t) =
d

i=1

i
(t, )dS
i
(t)
=
d

i=1
h
i
(t, )
dS
i
(t)
S
i
(t)
(2.10)
Chapter 2. The Markovian Case 19
=h
T
(t, )(t, S(t)) ((t, S(t))dt + dW(t))
for all t [0, T].
The conditions in (2.4) and (2.9) in conjunction with Holders inequality
yield that
d

i=1
_
T
0
[
i
(t, S(t))V

(t)
i
(t, S(t))[ dt <
almost surely, which guarantees the existence of a strong solution for V

(). If the
condition in (2.9) holds with (, )V

() replaced by (, ) then V

() stays strictly
positive. In this case, analog to (2.2), the solution of the stochastic dierential
equation in (2.8) is given as
V

(t) =exp
__
t
0

T
(u, S(u))(u, S(u))du +
_
t
0

T
(u, S(u))(u, S(u))dW(u)

1
2
_
t
0

T
(u, S(u))a(u, S(u))(u, S(u))du
_
(2.11)
for all t [0, T]. For example, the strategy
0
(, ) 0 invests only in the money
market and its associated wealth process satises V

0
() 1. Usually, trading
strategies do not lead to wealth processes that only depend on the current state of
the market, as the next remark discusses:
Remark 2 (Markovianness of wealth process and dependence on whole path). Ob-
viously, the wealth process of an investor jointly with the stock price process is
Markovian if the investor uses a Markovian trading strategy. Yet, at time t [0, T]
the wealth process does not only depend on the current stock prices S(t) but in most
cases also on past stock prices S(u), u t. Important exceptions from this rule
are the market portfolio
m
i
(t, s) := s
i
/

d
j=1
s
j
and investments in single stocks or
the money market only; that is,
j
i
(t, s) :=
j
(i) for some j 1, . . . , d +1 and for
all (t, s) [0, T] R
d
+
, where
j
represents Kroneckers delta function. However, as
we shall see in Theorem 1 of Section 2.3, the dependence of the associated wealth
Chapter 2. The Markovian Case 20
processes on the past does not represent a problem in our setup for nding optimal
strategies.
The change of numeraire, that is, the change of the denomination in which
the wealth process is quoted, is one of the most useful techniques in mathematical
nance; compare Geman et al. (1995) for a derivation and discussion of the change
of numeraire technique. It also plays a fundamental role in this chapter. For every
numeraire, a special market price of risk exists:
Denition 3 (-specic market price of risk). Let (, ) denote a strategy and
(, ) a market price of risk. Dene the corresponding -specic market price of
risk

(t, s) : [0, T] R
d
+
R
K
as

(t, s) := (t, s)
T
(t, s)(t, s). (2.12)
The following computations show that the -specic market price of risk
exactly translates the volatilities into the mean rates of return relative to the wealth
process of (, ). Let (, ) be any other strategy and V

() always strictly positive.

Then, we have from (2.11)
V

(t)
V

(t)
=exp
_
_
t
0
((u, S(u)) (u, S(u)))
T
(u, S(u))du
+
_
t
0
((u, S(u)) (u, S(u)))
T
(u, S(u))dW(u)

1
2
_
t
0
_

T
(u, S(u))a(u, S(u))(u, S(u))

T
(u, S(u))a(u, S(u))(u, S(u))
_
du
_
and thus after a short calculation,
d
_
V

(t)
V

(t)
_
=
V

(t)
V

(t)
_
(t, S(t)) (t, S(t))
_
T
_
_
(t, S(t)) a(t, S(t))(t, S(t))
_
dt
+ (t, S(t))dW(t)
_
Chapter 2. The Markovian Case 21
=
V

(t)
V

(t)
_
(t, S(t)) (t, S(t))
_
T
(t, S(t))dW

(t), (2.13)
where
W

(t) := W(t) +
_
t
0

(u, S(u))du (2.14)

for all t [0, T]. Another short computation yields
Z

(t)V
v,
(t) = v exp
_

_
t
0

T
(u, S(u))dW(u)
1
2
_
t
0
|

(u, S(u))|
2
du
_
(2.15)
for all t [0, T].
Remark 3 (Change of numeraire). The expression in (2.15) should be contrasted to
one in (2.5). The market price of risk (, ) is replaced by the -specic market price
of risk

(, ) when we multiply Z

() by a strictly positive wealth process V

v,
().
This is a well-known fact in the no-arbitrage theory of change of numeraire; compare
for example Chapter 9 of Shreve (2004). However, if Z

()V
v,
() is not a true
martingale, then Z

(T)V

(T) is not a Radon-Nikodym derivative and the process

W

() is not necessarily a Brownian motion under an equivalent local martingale

measure. Corollary 4 of Section 2.5 will extend the classical change of numeraire
to this case.
Arbitrage has been mentioned several times. We conclude this section by
discussing exactly what we mean by it. The next denition goes back to Section 3.3
of Fernholz (2002).
Denition 4 (Arbitrage). We call a strategy (, ) with P(V

(T) V

(T)) = 1
and P(V

(T) > V

(T)) > 0 a relative arbitrage opportunity with respect to the

strategy (, ). We call (, ) a classical arbitrage opportunity if (, ) invests fully
in the money market, that is, if (, ) 0.
For a detailed study of arbitrage, and in particular no-arbitrage conditions,
we refer the reader to the monograph by Delbaen and Schachermayer (2006). Jar-
row et al. (2007; 2010) discuss these conditions with respect to the existence of
Chapter 2. The Markovian Case 22
bubbles and suggest using the stronger condition of no dominance rst proposed
by Merton (1973). Here, we take the opposite approach. Instead of imposing a new
condition, the goal of this analysis is to investigate a general class of models and
study how much can be said in this more general framework without relying on the
tool of an equivalent local martingale measure.
For the sake of completeness and to put this work into perspective we remind
the reader how a bubble is frequently dened in the existing literature.
4
From
Theorem 1 below it follows then that the existence of a bubble implies a relative
arbitrage opportunity.
Denition 5 (Bubble). We say that a strategy (, ) contains a bubble if the stochas-
tic discount factor Z

() is a true martingale and if Z

()V

() is a strict local mar-

tingale, that is, not a martingale, under the equivalent local martingale measure.
In this context, it is important to remind ourselves that Z

() is a true mar-
tingale if and only if there exists an equivalent local martingale measure Q, under
which the stock price processes are local martingales. The question of whether Q
is a martingale measure or only a local martingale measure is not connected to
whether Z

() is a strict local or a true martingale.

2.4 Optimal strategies
In this section, we derive the representation of optimal strategies in terms of delta
hedges. In Subsection 2.4.1, we start from a given Markovian trading strategy
and nd an optimal strategy leading to the same terminal wealth. As Remark 2
discusses, this result can be interpreted as a hedging result for a certain class of
4
In the bubbles literature, an alternative denition appears, based upon the characterization
of the pricing operator as a charge, that is, an only nitely additive measure. However, it can
be shown that this characterization is equivalent to the one here, which relies on strict local
martingales; see Section 8 of Jarrow et al. (2010) for the proof and literature that relies on this
alternative characterization.
Chapter 2. The Markovian Case 23
possibly path-dependent payos, namely those which are strictly positive and for
which a (possibly suboptimal) Markovian trading strategy is known to replicate
them. Subsection 2.4.2 treats the hedging of non path-dependent European claims.
Finally, Subsection 2.4.3 provides sucient conditions under which the hedging
price in Subsection 2.4.2 is suciently dierentiable.
2.4.1 Optimizing a given strategy
Simple examples for strategies that we shall optimize are the market portfolio,
where the portfolio weights are chosen as the market weights for stocks, or a strategy
that invests the whole wealth in the money market. Given such a strategy, we look
for a new strategy whose associated wealth at time horizon T exactly replicates
the original value. We choose the new strategy to be optimal in the sense of
minimal required initial capital. This criterion of optimality is directly related
to the criterion of the shortest time to beat a portfolio by a given amount; see
Section 6.2 of Fernholz and Karatzas (2009).
If D 0 is a nonnegative T(T)-measurable random variable such that
E[D[T(t)] is a function of S(t) for some t [0, T], we use the Markovian structure
of S() to denote conditioning on the event S(t) = s by E
t,s
[D]. For the mo-
ment, we assume that the associated wealth process stays strictly positive to avoid
notational diculties. We start by dening the function U

: [0, T] R
d
+
[0, 1]
as
U

(t, s) :=E
Tt,s
_
Z

(T)V

(T)
Z

(T t)V

(T t)
_
(2.16)
=E
Tt,s
_
exp
_

_
T
Tt

T
(u, S(u))dW(u)
1
2
_
T
Tt
|

(u, S(u))|
2
du
__
.
(2.17)
The last equality follows directly from (2.15). As we show in Theorem 1, U

can
be interpreted as a hedging price. It obviously depends on the strategy (, ).
Chapter 2. The Markovian Case 24
Proposition 1 yields that U

does not depend on the choice of the (Markovian)

market price of risk (, ).
We shall assume throughout this section that U

solves the PDE

t
U

(t, s) =
1
2
d

i=1
d

j=1
s
i
s
j
a
i,j
(T t, s)D
2
i,j
U

(t, s)
+
d

i=1
d

j=1
s
i
a
i,j
(T t, s)
j
(T t, s)D
i
U

(t, s), (2.18)

where D
i
, D
2
i,j
denote the partial derivatives with respect to the variable s. Sec-
tion 2.8, which serves as an appendix, provides a sucient condition for this as-
sumption, and Remark 5 illustrates that smoothness of U

is sucient for U

to
solve the PDE.
The next theorem is the rst key result of this chapter. It shows that U

can be interpreted as a hedging price for the wealth process V

(): There exists a

strategy that costs U

(T, S(0)) and replicates the wealth at time T. Furthermore,

there is no other strategy that replicates the wealth for less initial capital. Platen
(2008) suggests calling this fact Law of the Minimal Price to contrast it to the
classical Law of the One Price, which appears if an equivalent martingale measure
exists.
Theorem 1 (Optimal strategy). Let (, ) denote any Markovian trading strategy
with a strictly positive associated wealth process V

() and let U

solve the PDE in

(2.18). Then, a new strategy (, ) exists such that the associated wealth process
V
v,
() with initial wealth v := U

(T, S(0)) 1 is always strictly positive and has

the same value as V

() at time T, that is,

V
v,
(T) = V

(T).
Thus, whenever Z

()V

() is a strict local martingale, there exists a relative arbi-

trage opportunity with respect to (, ). The strategy can be explicitly repre-
Chapter 2. The Markovian Case 25
sented as

i
(t, s) =s
i
D
i
log U

(T t, s) +
i
(t, s) (2.19)
for all (t, s) [0, T] R
d
+
. Furthermore, (, ) is optimal: There exists no strategy
(, ) such that
V
v,
(T) V

(T) = V
v,
(T) (2.20)
almost surely for some v < v.
Proof. Let us start by dening the martingale N

() as
N

(t) :=E[Z

(T)V

(T)[T(t)] = Z

(t)V

(t)U

(T t, S(t)) (2.21)
for all t [0, T] and denoting by L the innitesimal generator of S(), that is,
L =
d

i=1
s
i

i
(t, s)D
i
+
1
2
d

i=1
d

j=1
s
i
s
j
a
i,j
(t, s)D
2
i,j
. (2.22)
Since
dU

(T t, S(t)) =
_
LU

t
U

_
(T t, S(t)) dt
+
K

k=1
d

i=1
S
i
(t)
i,k
(t, S(t))D
i
U

(T t, S(t))W
k
(t)
holds for all t [0, T], the product rule of stochastic calculus and (2.15) yield
dN

(t)
Z

(t)V

(t)
=dU

(T t, S(t)) + U

(T t, S(t))
d(Z

(t)V

(t))
Z

(t)V

(t)

k=1

k
(t, S(t))
d

i=1
S
i
(t)
i,k
(t, S(t))D
i
U

(T t, S(t))dt.
We obtain the equality
dN

(t)
N

(t)
=
K

k=1
_
d

i=1
S
i
(t)
i,k
(t, S(t))
D
i
U

(T t, S(t))
U

(T t, S(t))

k
(t, S(t))
_
dW
k
(t)
Chapter 2. The Markovian Case 26
+ C

(t, S(t))dt,
where
C

(t, s) :=
_
LU

t
U

_
(T t, s)
U

(T t, s)

d

i=1
s
i
D
i
U

(T t, s)
U

(T t, s)
K

k=1

k
(t, s)
i,k
(t, s)
=
1
U

(T t, s)
_
1
2
d

i=1
d

j=1
s
i
s
j
a
i,j
(t, s)D
2
i,j
U

(T t, s) (2.23)
+
d

i=1
d

j=1
s
i
a
i,j
(t, s)
j
(t, s)D
i
U

(T t, s)

t
U

(T t, s)
_
=0,
since U

solves the PDE in (2.18). Thus, we can write

dN

(t)
N

(t)
=
K

k=1
_
d

i=1

i,k
(t, S(t))
_
S
i
(t)D
i
log (U

(T t, S(t))) +
i
(t, S(t))
_

k
(t, S(t))
_
dW
k
(t)
=
K

k=1

k
(t, S(t))dW
k
(t)
=
d(Z

(t)V
v,
(t))
Z

(t)V
v,
(t)
,
where the last equality follows from (2.15). Then, N

(0) = v = Z

(0)V
v,
(0) and
both processes N

() and Z

()V
v,
() have the same dynamics such that
Z

(T)V

(T) = N

(T) = Z

(T)V
v,
(T);
see Theorem 1.4.61 of Jacod and Shiryaev (2003). Since zero is an absorbing state
for any nonnegative supermartingale and since Z

(T)V

(T) > 0 almost surely, we

observe that V

() is a strictly positive process almost surely.
Optimality comes from the fact that for any strategy (, ) and for any
initial wealth v 0 the process Z

()V
v,
() is bounded from below by zero, further
Chapter 2. The Markovian Case 27
implying that it is a supermartingale. Assume we have some strategy (, ) such
that (2.20) is satised. Then, we obtain
v E[Z

(T)V
v,
(T)] E[Z

(T)V

(T)] = E[Z

(T)V
v,
(T)] = v, (2.24)
which concludes the proof.
We obtain from (2.21) and the last proof that
V
v,
(t) =
N

(t)
Z

(t)
= V

(t)U

(T t, S(t)), (2.25)
which we can rewrite as
U

(T t, S(t)) =
V
v,
(t)
V

(t)
.
Thus, U

(T t, S(t)) can be interpreted as the fraction of two dierent wealth

processes at time t that lead to the same terminal wealth, namely the wealth
processes associated with the optimal strategy (, ) and the original strategy (, ),
respectively.
We would like to emphasize that we have not shown that (, ) is unique.
Indeed, since we have not excluded the case that two stock prices have identical
dynamics this is not necessarily true. However, if the strategy (, ) is also optimal,
then (2.24) yields that Z

()V

() is a martingale, and thus V

() V

(); to wit,
the optimal wealth process is unique.
The next remarks discuss various assumptions of the last theorem:
Remark 4 (Completeness of the market). One remarkable feature of the last result
is that we have not required the market to be complete. In contrast to Fernholz and
Karatzas (2010), we do not rely on the martingale representation theorem but in-
stead directly derive a representation for the conditional expectation process of the
nal wealth V

(T) in the form of another wealth process (, ). This means that

given the existence of some Markovian trading strategy (, ) to achieve V

(T),
Chapter 2. The Markovian Case 28
an optimal strategy (, ) exists to achieve V

(T). The explanation for this phe-

nomenon is that all relevant sources of risk for exploiting the relative arbitrage are
completely captured by the tradeable stocks. However, we remind the reader that
we live here in a setting in which the mean rates of return and volatilities do not
depend on an extra stochastic factor. In a more incomplete model, with jumps
or additional risk factors in mean rates of return or volatilities, this result can no
longer be expected to hold. We revisit this discussion in Chapter 3.
Remark 5 (PDE in (2.18)). The essential assumption of this section is that U

solves the PDE in (2.18). Sucient conditions are existence and dierentiability
conditions on the function H

of (2.53) in conjunction with the condition in (2.57)

in Section 2.8. Another sucient condition is dierentiability of U

or, more pre-

cisely, that U

C
1,2
([0, T] R
d
+
). Then, the proof of Theorem 1 yields that U

automatically solves the PDE in (2.18), at least at all points (t, s) [0, T]R
d
+
that
can be attained by S() at time t. This can be seen from the fact that the process
N

() of (2.21) is a martingale; thus, its dt-term must disappear. This corresponds

exactly to the condition C

(, ) 0, where C

is dened in (2.23). Multiplying

this equation by U

(T t, s) we obtain the PDE in (2.18). Alternatively, Remark 3

of Fernholz and Karatzas (2010) briey discusses general but technical assumptions
for the necessary dierentiability of U

. Furthermore, it can be observed that it is

sucient that U

solves the PDE only on the subset of [0, T] R

d
+
where the stock
price lives. Example 1 of Section 2.6 illustrates this point.
The condition of dierentiability in time t can be slightly weakened to piece-
wise dierentiability. If there are m points 0 < t
1
< . . . < t
m
< T where U

is not
dierentiable, then we can nd an optimal strategy up to time t
1
, starting from t
1
to
t
2
and so on. This will neither change the optimal strategy (, ) nor the minimal
initial capital v in any way. This small modication allows us to include strate-
gies (, ) with structural breaks, by which we mean strategies whose arbitrage
Chapter 2. The Markovian Case 29
properties are changed at nitely many time steps. An example is a full investment
up to time t (0, T) in one strategy that can be arbitraged and afterwards a full
investment in another strategy that cannot be arbitraged.
Furthermore, as Example 5 of Section 2.6 illustrates, the dierentiability of
U

in the stock price dimension is only a sucient but not a necessary condition
for the existence of an optimal strategy.
The PDE in (2.18) always has the constant function as a solution. The next
result classies U

within the class of all PDE solutions as the minimal nonnegative

solution. This result generalizes Theorem 1 of Fernholz and Karatzas (2010).
Proposition 2 (Characterization of U

). The function U

is the smallest function

that solves the PDE in (2.18) and is nonnegative for all (t, s) [0, T] R
d
+
that
can be attained by S() at time t with initial condition U

(0, s) 1 for all s R

d
+
.
Furthermore, the PDE

t
U(t, s) =
1
2
d

i=1
d

j=1
s
i
s
j
a
i,j
(T t, s)D
2
i,j
U(t, s)
+
d

i=1
d

j=1
s
i
a
i,j
(T t, s)
j
(T t, s)D
i
U(t, s),
where we have exchanged (, ) by (, ), has U

(, ) 1 as its minimal nonnega-
tive solution.
Proof. Consider any suciently smooth function

U

: [0, T] R
d
+
R
+
that solves
the PDE in (2.18) and the initial condition

U

(0, s) 1 for all s R

d
+
. Dene, as
in (2.21), the process

N() as

(t) := Z

(t)V

(t)

(T t, S(t))
for all t [0, T], which is, as in the proof of Theorem 1, a positive supermartingale.
Thus, we have

(T t, S(t)) =

(t)
Z

(t)V

(t)
Chapter 2. The Markovian Case 30

E
t,S(t)
_

(T)
_
Z

(t)V

(t)
=
E
t,S(t)
_
Z

(T)V

(T)

(t)V

(t)
= U

(T t, S(t))
for all t [0, T]. The second statement of the proposition comes from the fact
that Z

()V

() is a martingale, which implies U

(, ) 1, and from the same
considerations as above.
The hedging price for the stock of Example 4 in Section 2.6, for instance,
is one of many solutions of polynomial growth of the corresponding Black-Scholes
type PDE. For example, consider h
1
(t, s) := s and h
1
(t, s) times the hedging price
of (2.52), that is, h
2
(t, s) := 2s(1/(s

t)) s < s, for all (t, s) [0, T] R

+
,
where denotes the cumulative normal distribution. Then, both h
1
and h
2
solve
the PDE

t
h(t, s) =
1
2
s
4
D
2
h(t, s)
with the identical boundary condition h(t, 0) = 0 and h(0, s) = s for all (t, s)
[0, T] R
+
.
The reason for non-uniqueness in this case is the fact that the second-order
coecient has super-quadratic growth preventing standard theory from being ap-
plied; see, for example, Section 5.7.B of Karatzas and Shreve (1991). Furthermore,
the boundary condition at innity is not specied precisely enough. Both solutions
grow polynomially, but clearly h
2
is always smaller than h
1
. In this specic ex-
ample, the corresponding process 1/S() is a three-dimensional Bessel process and
therefore stays away from the boundary. If the drift, however, is removed, it is a
Brownian motion, which can hit zero. Thus, boundary conditions need to be pre-
cisely specied for a PDE in 1/s at zero, which corresponds to the precise boundary
Chapter 2. The Markovian Case 31
condition at innity for the PDE above. Indeed, as the next section shows, the ex-
istence of an arbitrage opportunity is equivalent to the positive probability of some
process imploding to zero under some measure Q, which corresponds exactly to the
observation that 1/S() in the above example can hit zero.
For the special case (, ) 0 in one dimension, and under some assumptions
on the volatility parameter (, ), Ekstrom et al. (2009) suggest a numerical algo-
rithm that utilizes this characterization and nds the minimal nonnegative solution
of a Black-Scholes type PDE that does not have a unique solution.
2.4.2 Hedging of contingent claims
So far, we have started from a given Markovian trading strategy (, ) and then
optimized it. However, one might imagine situations in which one wants to hedge
a contingent claim but does not know a possibly suboptimal strategy (, ) a priori.
How can we nd, in such a situation, an optimal strategy? In the following we re-
solve this problem for Markovian claims. We shall also provide weak sucient con-
ditions for the corresponding hedging price to be dierentiable in Subsection 2.4.3.
We now explicitly allow the associated wealth processes to hit zero.
To simplify computations later on, we introduce some notation. As before,
the expectation operator corresponding to the event S(t) = s is written as E
t,s
.
Using the Markovian structure of our model, we denote, outside of the expectation
operator, by (S
t,s
(u))
u[t,T]
a stock price process with the dynamics of (2.1) and
S(t) = s, in particular, S
0,S(0)
() S(). We observe that Z

(u)/Z

(t) depends
for u (t, T] on T(t) only through S(t) and we write similarly (

Z
,t,s
(u))
u[t,T]
for
(Z

(u)/Z

(t))
u[t,T]
, with

Z
,t,s
(t) = 1 on the event S(t) = s. When we want to
stress the dependence of a process on the state we shall write, for example,
S(t, ).
We emphasize the standing assumptions made in Section 2.2, namely, that
Chapter 2. The Markovian Case 32
the stock price process S() with dynamics specied in (2.1) starting in S(0) R
d
+
is R
d
-valued, unique and stays in the positive orthant. Furthermore, a square-
integrable Markovian market price of risk exists almost surely.
For any measurable function p : R
d
+
[0, ), representing the payo of the
contingent claim, we dene a candidate h
p
: [0, T] R
d
+
[0, ) for the hedging
price of the corresponding European option, similar to the denition of U

in (2.16)
as
h
p
(t, s) := E
Tt,s
_

(T)p(S(T))
_
. (2.26)
The only dierence between h
p
and U

is that we do not normalize h

p
with a
wealth process. Since S() is Markovian, h
p
is well-dened. The equation in (2.26)
appears as the real-world pricing formula in the Benchmark Approach; compare
Equation (9.1.30) of Platen and Heath (2006).
Let us denote by supp(S()) the support of S(), that is, the smallest closed
set in [0, T] R
n
such that
P((t, S(t)) supp(S()) for all t [0, T]) = 1.
We call i-supp(S()) the union of (0, S(0)) and the interior of supp(S()) and assume
that
P((t, S(t)) i-supp(S()) for all t [0, T)) = 1.
This assumption is made to exclude degenerate cases, where S() can hit the bound-
ary of its support with positive probability.
Denition 6 (Point of support). We call any (t, s) i-supp(S()) a point of support
for S().
We remark that each such point (t, s) satises t < T. For example, if S() is
a one-dimensional geometric Brownian motion then the set of points of support for
S() is exactly (0, S(0)) (t, s) (0, T) R
+
.
Chapter 2. The Markovian Case 33
Applying Itos rule to (2.26) yields the following result, which in particular
provides a mechanism for pricing and hedging contingent claims under the Bench-
mark Approach. Its proof is similar to the one of Theorem 1. In order to avoid
introduction of extra notation and to be consistent with Theorem 1, we state the
optimal trading strategies in terms of proportions of the current wealth. This might
formally lead to a division by zero when the wealth process hits zero, but in that
case no investments will happen anyway. We refer the reader to Theorem 1 of Ruf
(2011), where the theorem and its proof are stated in terms of numbers of shares
held.
Theorem 2 (Markovian representation for non path-dependent European claims).
Assume that we have a contingent claim of the form p(S(T)) 0 and that the
function h
p
of (2.26) is suciently dierentiable or, more precisely, that we have
for all points of support (t, s) for S() that h
p
C
1,2
(|
Tt,s
) for some neighborhood
|
Tt,s
of (T t, s). Then, with

p
i
(t, s) := s
i
D
i
log (h
p
(T t, s)) (2.27)
for all i = 1, . . . , d and (t, s) [0, T] R
d
+
, and with v
p
:= h
p
(T, S(0)), we get
V
v
p
,
p
(T t) = h
p
(t, S(T t))
for all t [0, T]. Furthermore, the strategy
p
(, ) is optimal in the sense of Theo-
rem 1 and h
p
solves the PDE

t
h
p
(T t, s) =
1
2
d

i=1
d

j=1
s
i
s
j
a
i,j
(t, s)D
2
i,j
h
p
(T t, s) (2.28)
at all points of support (t, s) for S().
Proof. Let us start by dening the martingale N
p
() as
N
p
(t) :=E[Z

(T)p(S(T))[T(t)] = Z

(t)h
p
(T t, S(t))
Chapter 2. The Markovian Case 34
for all t [0, T]. Although h
p
is not assumed to be in C
1,2
((0, T]R
d
) but only to be
locally smooth, we can apply a localized version of Itos formula; see, for example,
Section IV.3 of Revuz and Yor (1999). Then, the product rule of stochastic calculus
can be used to obtain the dynamics of N
p
(). Since N
p
() is a martingale, the
corresponding dt-term must disappear. This observation, in connection with (2.3)
and the positivity of Z

(), yields the PDE in (2.28). Itos formula, now applied to

h
p
(T , S()), and the PDE in (2.28) imply
dh
p
(T t, S(t)) = h
p
(T t, S(t))
d

i=1

p
i
(t, S(t))
dS
i
(t)
S
i
(t)
for all t [0, T]. Then, both h
p
(T , S()) and V
v
p
,
p
() are stochastic exponentials
and solve the same stochastic dierential equation. Theorem 1.4.61 of Jacod and
Shiryaev (2003) yields h
p
(T , S()) V
v
p
,
p
().
The optimality of
p
(, ) follows exactly as in Theorem 1.
The last result generalizes Proposition 3 of Platen and Hulley (2008), where
the same result is derived for a one-dimensional, complete market with a time-
transformed squared Bessel process of dimension four modeling the stock price
process.
We remark that, as before, we neither assumed a complete market nor uti-
lized a representation theorem. In particular, at no point did we assume invertibility
or full rank of the volatility matrix (, ). Under these general assumptions, there
is no hope to be able to hedge all contingent claims on the Brownian motion W(T).
However, W(T) appears in this class of models only as a nuisance parameter and
it is of no economic interest to trade in it directly.
Remark 6 (Delta hedging). Writing (2.27) as

p
i
(t, S(t))
V
v
p
,
p
(t)
S
i
(t)
= D
i
h
p
(T t, S(t))
Chapter 2. The Markovian Case 35
and observing that the left-hand side is the number of shares invested in stock i
at time t shows that the optimal strategy is a delta hedge as in classical Financial
Mathematics, when one tries to hedge a contingent claim. Of course, (, ) of (2.19)
can be interpreted in a similar way: U

is the risk-adjusted expected nal wealth

relative to the current wealth. Since everything has been expressed with respect to
a wealth process V

() the associated strategy (, ) is added to obtain the optimal

strategy (, ).
Example 9.2.2 of Fernholz et al. (2005) illustrates that the classical put-call
parity can fail. Using the machinery of this section, we can directly show that a
modied version of the put-call parity holds. An equivalent version in the situation
of an equivalent local martingale measure with possible bubbles has already been
derived in Lemma 7 of Jarrow et al. (2007). The put-call parity is sometimes applied
incorrectly in the literature, see, for example, Emanuel and Macbeth (1982)
5
. In
this context, we refer the reader also to the discussion in Madan and Yor (2006).
Corollary 1 (Modied put-call parity). For any L R we have the modied put-
call parity for the call- and put-options (S
1
(T)L)
+
and (LS
1
(T))
+
, respectively,
with strike price L:
E
t,s
_

(T)(L S
1
(T))
+
_
+ s
1
U

1
(T t, s)
= E
t,s
_

(T)(S
1
(T) L)
+
_
+ LU

0
(T t, s), (2.29)
where
0
(, ) 0 denotes the strategy for holding a monetary unit and
1
(, )
(1, 0, . . . , 0)
T
the strategy for holding stock S
1
().
Proof. The statement follows from the linearity of expectation.
Due to Theorem 2, under weak dierentiability assumptions, optimal strate-
gies exist for the money market, the stock S
1
(T), the call and the put. Thus, the
5
We thank Peter Carr for pointing us to this reference.
Chapter 2. The Markovian Case 36
left-hand side of (2.29) corresponds to the sum of the hedging prices of a put and
the stock, and the right-hand side corresponds to the sum of the hedging prices of
a call and L monetary units. The dierence between this and the classical put-call
parity is that the current stock price and the strike L are replaced by their hedging
prices. Section 2.2 of Bayraktar et al. (2010b) have recently observed an another
version of the put-call parity. Instead of replacing the current stock price by its
hedging price, they replace the European call price by the American call price and
restore the put-call parity this way.
2.4.3 Smoothness of hedging price
Next, we shall provide sucient conditions under which the function h
p
of the last
subsection is suciently smooth. Towards this end, we need the following denition.
Denition 7 (Locally Lipschitz and locally bounded). We call a function f : [0, T]
R
d
+
R locally Lipschitz and locally bounded on R
d
+
if for all s R
d
+
the function
t f(t, s) is right-continuous with left limits and for all M > 0 there exists some
C(M) < such that
sup
1
M
|y|,|z|M
y,=z
[f(t, y) f(t, z)[
|y z|
+ sup
1
M
|y|M
[f(t, y)[ C(M)
for all t [0, T].
In particular, if f has continuous partial derivatives, it is locally Lipschitz
and locally bounded. We require several assumptions in order to show the neces-
sary dierentiability of h
p
in Theorem 3 below. It is subject to future research to
determine the precise conditions which yield the existence of a delta hedge, possibly
without requiring h
p
to be the classical solution of a PDE.
(A1) The functions
k
(, ) and
i,k
(, ) are for all i = 1, . . . , d and k = 1, . . . , K
locally Lipschitz and locally bounded.
Chapter 2. The Markovian Case 37
(A2) For all points of support (t, s) for S() there exist some C > 0 and some
neighborhood | of (t, s) such that
d

i=1
d

j=1
a
i,j
(u, y)
i

j
C||
2
(2.30)
for all R
d
and (u, y) |.
(A3) The payo function p is chosen so that for all points of support (t, s) for S()
there exist some C > 0 and some neighborhood | of (T t, s) such that
h
p
(u, y) C for all (u, y) |.
If h
p
is constant for

d d coordinates, say the last ones, Assumption (A2) can
be weakened to requesting the uniform ellipticity only in the remaining d

d 1
coordinates; that is, the sum in (2.30) goes only to d

d 1 and R
d

d1
.
Assumption (A3) holds in particular if p is of linear growth; that is, if p(s)
C

d
i=1
s
i
for some C > 0 and all s R
d
+
, since

Z
,t,s
()S
t,s
i
() is a nonnegative
supermartingale for all i = 1, . . . , d.
We emphasize that the conditions here are weaker than the ones used in
Section 9 of Fernholz and Karatzas (2010) for the case of the market portfolio,
which can be represented as p(s) =

d
i=1
s
i
. In particular, the stochastic integral
component in Z

() does not present any technical diculty in our approach.

We proceed in two steps. In the rst step, we use the theory of stochastic
ows to derive continuity of S
t,s
(T) and

Z
,t,s
(T) in t and s. This theory relies
on Kolmogorovs lemma, see, for example, Theorem IV.73 of Protter (2003), and
studies continuity of stochastic processes as functions of their initial conditions.
We refer the reader to Kunita (1984) and Chapter V of Protter (2003) for an
introduction to and further references for stochastic ows. We shall prove continuity
of S
t,s
() and

Z
,t,s
() at once and introduce for that the d +1-dimensional process
X
t,s,z
() := (S
t,s
T
(), z

Z
,t,s
())
T
.
Chapter 2. The Markovian Case 38
The following lemma modies Theorem V.37 of Protter (2003) for our con-
text. This result will be of use below.
Lemma 1 (Stochastic ow, globally Lipzschitz). Fix

d N. We consider a system
of

d stochastic dierential equations of the form
dY
t,y
i
(u) =
i
(u, Y
t,y
(u))dt +
K

k=1

i,k
(u, Y
t,y
(u))dW
k
(u), (2.31)
Y
t,y
i
(t) = y
i
for all u [t, T], for all (t, y) [0, T] R

d
, where Y
t,y
() = (Y
t,y
1
(), . . . , Y
t,y

d
())
T
denotes a

d-dimensional vector. The drift : [0, T] R

d
R

d
and volatility
coecient : [0, T] R

d
R

dK
are assumed to be measurable and to satisfy the
global Lipschitz condition

i=1
[
i
(u, y
1
)
i
(u, y
2
)[ +

i=1
K

k=1
[
i,k
(u, y
1
)
i,k
(u, y
2
)[ C|y
1
y
2
|
for all (u, y
1
, y
2
) [0, T] R

d
R

d
for some constant C > 0.
Then, the stochastic dierential equation in (2.31) has a unique solution
Y
t,s
(). It has a modication, which we again call Y
t,s
(), and which satises the
following continuity property: Fix any countable set of time indices T = t
i

[0, T]
iN
. Then, there exists a subset

with P(

) = 1 such that for all

, k N, t T, and for all y
1
, y
2
R

d
with |y
1
y
2
| 2
k2
we have
sup
u[t,T]
|Y
t,y
1
(u) Y
t,y
2
(u)| c
1
()2
c
2
()k
for some constants c
1
(), c
2
() R
+
. In particular, the constants c
1
() and c
2
()
can be chosen independently of t T.
Proof. The lemma basically states Theorem V.37 of Protter (2003). The explicit
continuity comes from an analysis of the arguments in the proof of Kolmogorovs
Lemma; compare Theorem IV.73 of Protter (2003). There, we use Chebyshevs
inequality simultaneously for all t T and then follow the proof line by line.
Chapter 2. The Markovian Case 39
We can now prove the continuity of the process X
t,s,1
() in t and s using a
localization technique:
Lemma 2 (Stochastic ow, locally Lipzschitz). We x a point (t, s) [0, T]R
d
+
so
that X
t,s,1
() is strictly positive and an R
d+1
+
-valued process. Then, under Assump-
tion (A1), we have for all sequences (t
k
, s
k
)
kN
[0, T] R
d
+
with lim
k
(t
k
, s
k
) =
(t, s) that
lim
k
sup
u[t,T]
|X
t
k
,s
k
,1
(u) X
t,s,1
(u)| = 0
almost surely, where we set X
t
k
,s
k
,1
(u) := (s
T
k
, 1)
T
for u t
k
. In particular, for
K() suciently large we have that X
t
k
,s
k
,1
(u, ) is strictly positive and R
d+1
+
-valued
for all k > K() and u [t, T].
Proof. Since the class of locally Lipschitz and locally bounded functions is closed
under summation and multiplication, Assumption (A1) yields that the drift and
diusion coecients of X
u,y,z
() are locally Lipschitz for all (u, y, z) [0, T] R
d
+

R
+
. We start by assuming t
k
t for all k N and obtain
sup
u[t,T]
|X
t
k
,s
k
,1
(u) X
t,s,1
(u)| sup
u[t,t
k
]
|(s
T
k
, 1)
T
X
t,s,1
(u)|
+ sup
u[t
k
,T]
|X
t
k
,s
k
,1
(u) X
t
k
,s,1
(u)|
+ sup
u[t
k
,T]
|X
t
k
,s,1
(u) X
t
k
,S
t,s
(t
k
),

Z
,t,s
(t
k
)
(u)|
for all k N. The rst term on the right-hand side of the last inequality goes to
zero as k increases by the continuity of the sample paths of X
t,s,1
(). The arguments
and the localization technique in the proof of Theorem V.38 in Protter (2003) in
conjunction with Lemma 1 yield that
lim
k
sup
u[

t,T]
|X

t,y
k
,z
k
(u) X

t,s,1
(u)| = 0
Chapter 2. The Markovian Case 40
for all

t t, t
1
, t
2
, . . . and any sequence ((y
T
k
, z
k
)
T
)
kN
R
d+1
+
with (y
T
k
, z
k
)
T

(s
T
, 1)
T
as k almost surely. The convergence is uniformly in

t t, t
1
, t
2
, . . ..
We now choose for (y
T
k
, z
k
)
T
the sequences (s
T
k
, 1)
T
and (S
t,s
T
(t
k
, ),

Z
,t,s
(t
k
, ))
T
for all . This proves the statement if t
k
t for all k N. In the case of the
reversed inequality t
k
t, we observe
sup
u[t,T]
|X
t
k
,s
k
,1
(u) X
t,s,1
(u)| sup
u[t,T]
|X
t
k
,s
k
,1
(u) X
t
k
,s,1
(u)|
+ sup
u[t,T]
|X
t
k
,s,1
(u) X
t,s,1
(u)|,
which again yields continuity, similar to above.
In the second step, we use a technique from the theory of PDEs to conclude
the necessary smoothness of h
p
. The following result has been used by Ekstrom,
Janson and Tysk. We present it here on its own to underscore the analytic compo-
nent of our argument:
Lemma 3 (Schauder estimates and smoothness). Fix a point (t, s) [0, T) R
d
+
and a neighborhood | of (t, s). Suppose Assumption (A1) holds along with inequality
in (2.30) for all R
d
and (u, y) | and some C > 0. Let (f
k
)
kN
denote
a sequence of solutions of the PDE in (2.28) on |, uniformly bounded under the
supremum norm on |. If lim
k
f
k
(t, s) = f(t, s) on | for some function f : |
R, then f solves the PDE in (2.28) on some neighborhood

| of (t, s). In particular,
f C
1,2
(

|).
Proof. We refer the reader to the arguments and references provided in Section 2 of
Janson and Tysk (2006) and Theorem 3.2 of Ekstrom and Tysk (2009). The central
idea is to use the interior Schauder estimates by Knerr (1980) in conjunction with
Arzel`a-Ascoli type of arguments to prove the existence of rst- and second-order
derivatives of f.
We can now prove the smoothness of the hedging price h
p
:
Chapter 2. The Markovian Case 41
Theorem 3. Under Assumptions (A1)-(A3) there exists for all points of support
(t, s) for S() some neighborhood | of (T t, s) such that the function h
p
dened
in (2.26) is in C
1,2
(|).
Proof. We dene p : R
d+1
+
R
+
by p(s
1
, . . . , s
d
, z) := zp(s
1
, . . . , s
d
) and p
M
:
R
d+1
+
R
+
by p
M
() := p()1
p()M
for some M > 0 and approximate p
M
by a
sequence of continuous functions p
M,m
(compare for example Appendix C.4 of Evans
1998) such that lim
m
p
M,m
= p
M
pointwise and p
M,m
2M for all m N. The
corresponding expectations are dened as

h
p,M
(u, y) := E
Tu,y
[ p
M
(S
1
(T), . . . , S
d
(T),

Z

(T))]
for all (u, y)

| for some neighborhood

| of (T t, s) and equivalently

h
p,M,m
.
We start by proving continuity of

h
p,M,m
for large m. For any sequence
(t
k
, s
k
)
kN
[0, T] R
d
+
with lim
k
(t
k
, s
k
) = (t, s), Lemma 2, in connection with
Assumption (A1), yields
lim
k
p
M,m
(S
t
k
,s
k
(T),

Z
,t
k
,s
k
(T)) = p
M,m
(S
t,s
(T),

Z
,t,s
(T)).
The continuity of

h
p,M,m
follows then from the bounded convergence theorem.
Now, Lemma 2.6 of Janson and Tysk (2006), in connection with Assump-
tion (A2), guarantees that

h
p,M,m
is a solution of the PDE in (2.28). Lemma 3 then
yields that rstly,

h
p,M
and secondly, in connection with Assumption (A3), h
p
also
solve the PDE in (2.28) on some neighborhood | of (T t, s). In particular, h
p
is
in C
1,2
(|).
The last theorem generalizes the results in Ekstrom and Tysk (2009) to
several dimensions and to non-continuous payo functions p. Chapters 6 and 15
of Friedman (1976) and Janson and Tysk (2006) have related results, but they
impose linear growth conditions on a(, ) so that the PDE in (2.28) has a unique
solution of polynomial growth. We are especially interested in the situation in
Chapter 2. The Markovian Case 42
which multiple solutions may exist. Heath and Schweizer (2000) present results in
the case when the process corresponding to the PDE in (2.28) does not leave the
positive orthant. As Fernholz and Karatzas (2010) observe, this condition does not
necessarily hold if there is no equivalent local martingale measure. In the case of
Z

() being a martingale, our assumptions are only weakly more general than the
ones in Heath and Schweizer (2000) by not requiring a(, ) to be continuous in the
time dimension. Further results are also obtained by Section III.7 of Kunita (1984),
but under strong continuity assumptions on a(, ). However, in all these research
articles, the authors show that the function h
p
indeed solves the PDE in (2.28) not
only locally but globally and satises the corresponding boundary conditions. We
have here abstained from imposing the stronger assumptions these papers rely on
and concentrate on the local properties of h
p
. For our application, it is sucient to
observe that h
p
(T t, S(t)) converges to p(S(T)) as t goes to T; compare the proof
of Theorem 2.
The next section provides an interpretation of our approach to prove the
dierentiability of h
p
; all problems on the spatial boundary, arising for example
from a discontinuity of a(, ) on the boundary of the positive orthant, have been
conditioned away, so that S() can get close to but never actually attains the
boundary.
2.5 Change of measure
We obtained in Theorem 1 a precise description of an optimal strategy (, ) to
replicate the wealth V

(T) at time T. However, in order to compute this strategy

we need to compute the deltas of the expectation U

of the risk-adjusted wealth

Z

(T)V

(T). In Theorem 4, we shall provide a useful representation of U

by
performing a change of measure. To be able to do then the computations, we
provide the dynamics of the stock price processes and a formula for conditional
Chapter 2. The Markovian Case 43
expectations under the new probability measure in Corollaries 2 and 3. We end
this section by proving a result concerning the change of numeraire in Corollary 4,
illustrating in Proposition 3 how a canonical probability space can be constructed to
satisfy the technical assumptions of this section, and in several remarks discussing
connections of this work to some literature.
Theorem 1.4 of Delbaen and Schachermayer (1995b) shows that NA implies
the existence of a local martingale measure which is absolutely continuous with
respect to P. On the other side, a consequence of this section is the existence of a
local martingale measure under NUPBR, such that P is absolutely continuous with
respect to it. Indeed, as discussed in Remark 1, NA and NUPBR together yield
NFLVR, which again yields an equivalent local martingale measure corresponding
exactly to the one discussed in this section. Another point of view, which we
do not take here, is the recent insight by Kardaras (2010) on the equivalence of
NUPBR and the existence of a nitely additive probability measure that is, in
some sense, weakly equivalent to P and under which S() has some notion of weak
local martingale property.
Our approach via a generalized change of measure is in the spirit of the
work by Follmer (1972; 1973), Meyer (1972), Section 2 of Delbaen and Schacher-
mayer (1995a), and Section 7 of Fernholz and Karatzas (2010). They show that for
the strictly positive P-local martingale Z

(), there exists a probability measure Q

such that P is absolutely continuous with respect to Q and dP/dQ = 1/Z

(T

),
where

is the rst hitting time of zero by the process 1/Z

(). Their analysis has

been built upon by several authors, for example by Section 2 of Pal and Protter
(2010). We complement this research direction by determining the dynamics of
the P-Brownian motion W() under the new measure Q. These dynamics do not
follow directly from an application of a Girsanov-type argument since Q need not
be absolutely continuous with respect to P. Similar results for the dynamics have
Chapter 2. The Markovian Case 44
been obtained in Lemma 4.2 of Sin (1998) and Section 2 of Delbaen and Shirakawa
(2002). However, they rely on additional assumptions on the existence of solutions
for some stochastic dierential equations. Wong and Heyde (2004) prove the ex-
istence of a measure

Q satisfying E
P
[Z

(T)] =

Q(

> T), where W() has the

same

Q-dynamics as we derive, but P is not necessarily absolutely continuous with
respect to

Q.
For the results in this section, we do not need that U

solves the PDE in

(2.18). However, we make the technical assumption
6
that the probability space
is the space of right-continuous paths : [0, T] R
m
for some m N
with left limits at t [0, T] if (t) ,= and with an absorbing cemetery point
. By that we mean that (t) = for some t [0, T] implies (u) = for
all u [t, T] and for all . This point will represent explosions of Z

(),
which do not occur under P, but may occur under a new probability measure Q
constructed below. We further assume that the ltration F is the right-continuous
modication of the ltration generated by the paths , or more precisely by the
projections
t
() := (t). Concerning the original probability measure we assume
that P( : (T) = ) = 0 and that for all t [0, T], is an absorbing state for
Z

(); that is, Z

(t) = implies Z

(u) = for all u [t, T]. This assumption,

consistent with the dynamics of (2.6) species Z

() only on a set of measure zero

and is made for notational convenience.
We emphasize that we have not assumed completeness of the ltration F.
Indeed, we shall construct a new probability measure Q that is not necessarily
equivalent to the original measure P and can assign positive probability to nullsets
of P. If we had assumed completeness of F, we could not guarantee that Q could
be consistently dened on all subsets of these nullsets, which had been included in
6
The results in this section hold for more general probability spaces and ltrations, basically
for these ltered spaces that allow for extension theorems. However, for the sake of clarity, we
restrict ourselves here to this special version of a probability space and ltration. Compare the
appendix of Follmer (1972) for details.
Chapter 2. The Markovian Case 45
F during the completion process. The fact that we need the cemetery point and
cannot restrict ourselves to the original canonical space is also not surprising. The
point represents events that have under P probability zero, but under Q have
positive probability. Follmer and Imkeller (1993) discuss another example where
a change of measure needs additional events, and thus extensions of the original
probability space.
All these assumptions are needed to prove the existence of a measure Q with
dP/dQ = 1/Z

(T

). After having ensured its existence, one then can take the
route suggested by Theorem 5 of Delbaen and Schachermayer (1995a) and start
from any probability space satisfying the usual conditions, construct a canonical
probability space satisfying the technical assumptions mentioned above, and then
perform all necessary computations on this space. We shall detail these technical
steps in Proposition 3.
For now, the goal is to construct a measure Q under which the computation
of U

simplies. For that, we dene the sequence of stopping times

i
:= inft [0, T] : Z

(t) i
with inf := and the sequence of -algebras T
i
:= T(

i
T) for all i N.
We observe that the denition of T
i
is independent of the probability measure and
dene the stopping time

:= lim
i

i
with corresponding -algebra T
,
:=
T(

T) generated by

i=1
T
i,
.
Within this framework, Meyer (1972) and Example 6.2.2 of Follmer (1972)
rely on an extension theorem (compare Chapter 5 of Parthasarathy 1967) to show
the existence of a measure Q on (, T(T)) satisfying
Q(A) = E
P
_
Z

i
T)1
A

(2.32)
for all A T
i,
, where we now write E
P
for the expectation under the original
measure. We summarize these insights in the following theorem:
Chapter 2. The Markovian Case 46
Theorem 4 (Generalized change of measure). There exists a measure Q such that
for all stopping times

T with

T T and for all T(

T)-measurable random variables

D 0 we have
E
P
_
Z

T
_
D
_
= E
Q
_
D1
1/Z

T)>0
_
, (2.33)
where E
Q
denotes the expectation with respect to the new measure Q. That is,
P is absolutely continuous with respect to Q. Under this measure Q, the process

W() =
_

W
1
(), . . .

W
K
()
_
T
with

W
k
(t

) := W
k
(t

) +
_
t

k
(u, S(u))du (2.34)
for all k = 1, . . . , K and t [0, T] is a K-dimensional Brownian motion stopped at
time

.
Proof. The existence of a measure Q satisfying (2.32) follows as in the discussion
above. Now, for any set A T(

T) we have
A =
_
A
_

T
__

_
i=1
_
A
_

i1
<

T

i
__
.
From the fact that

T holds, if and only if 1/Z

T) = 0 holds and from the

identity in (2.32) we obtain
Q
_
_
A
_
_
_
1
Z

T
_ > 0
_
_
_
_
_
=

i=1
Q
_
A
_

i1
<

T

i
__
=

i=1
E
P
_
Z

i
T)1
A

i1
<

_
=

i=1
E
P
_
Z

T
_
1
A

i1
<

_
=E
P
_
Z

T
_
1
A
_
,
where the last identity holds since P
_

T
_
= 0. This yields the representation of
(2.33). From Girsanovs theorem (compare Theorem 8.1.4 of Revuz and Yor 1999)
Chapter 2. The Markovian Case 47
we obtain that on T
i,
the process

W() is under Q a K-dimensional Brownian
motion stopped at

i
T. Since

i=1
T
i,
generates T
,
and forms a -system, we
get the dynamics of (2.34).
Thus, an equivalent local martingale measure exists, if and only if Q(1/Z

(T) >
0) = 1. On the other hand, if no equivalent local martingale measure exists, then
valuing a wealth process must include the barrier aspect 1/Z

(T) > 0. To wit, al-

lowing for arbitrage requires calculating the Q-probability of the reciprocal 1/Z

()
of the stochastic discount factor hitting zero. We emphasize that we need not know
(, ) to calculate the corresponding hedging price U

, but only its nal associated

wealth as a function of the stock prices S(). However, in this case we cannot obtain
the strategy (, ) from Theorem 1.
A further consequence of Theorem 4 is the fact that the dynamics of the
stock price process and the reciprocal of the stochastic discount factor simplify
under Q as the next corollary shows.
Corollary 2 (Evolution of important processes under Q). The stock price process
S() and the reciprocal 1/Z

() of the stochastic discount factor evolve until the

stopping time

under Q according to
dS
i
(t) = S
i
(t)
K

k=1

i,k
(t, S(t))d

W
k
(t)
and
d
_
1
Z

(t)
_
=
1
Z

(t)
d

k=1

k
(t, S(t))d

W
k
(t)
for all i = 1, . . . , d and t [0, T].
Proof. This is a direct consequence of the representation of

W() in (2.34) and
Denition 1 of the market price of risk.
Chapter 2. The Markovian Case 48
The results of the last corollary play an essential role when we do com-
putations, since the rst hitting time of the reciprocal of the stochastic discount
factor can in most cases be easily represented as a rst hitting time of the stock
price. This now usually follows some more tractable dynamics, as we shall see in
Section 2.6. Theorem 4 also holds for expectations conditioned on T(t): the next
corollary generalizes the well-known Bayes rule for classical changes of measures;
compare Lemma 3.5.3 of Karatzas and Shreve (1991). Similar computations appear
already in Proposition 4.2 of Follmer (1972).
Corollary 3 (Bayes rule, Q-martingale property of 1/Z

()). Let

T denote any
stopping time with

T T. For all T(

T)-measurable random variables D 0 the

representation
E
Q
_
D1
1/Z

T)>0

T(t)
_
= E
P
_
Z

T
_
D[T(t)
_
1
Z

_
t

T
_1
1/Z

(t

T)>0
(2.35)
holds Q-almost surely (and thus P-almost surely) for all t [0, T]. Furthermore,
for any process N(), N()1
1/Z

()>0
is a Q-martingale, if and only if N()Z

() is
a P-martingale. In particular, the process 1/Z

() is a Q-martingale.
Proof. We observe that for all T
_

i


T
_
-measurable random variables D 0,
(2.32) can be rewritten as
E
Q
[D] = E
P
_
Z

i


T
_
D
_
. (2.36)
due to the martingale property of the bounded process Z

i
) under P. We x
a time t [0, T]. For any A T(t), (2.33) with D replaced by D1

A
where

A =
A
_

T > t
_
T
_

T t
_
and the same techniques, as in the proof of Theorem 4,
yield
E
Q
_
D1
1/Z

T)>0
1

A
_
=E
P
_
Z

T
_
D1

A
_
Chapter 2. The Markovian Case 49
=E
P
_
_
E
P
_
Z

T
_
D

T(t)
_
1
Z

_
t

T
_1

A
Z

_
t

T
_
_
_
=

i=1
E
P
_
E
P
_
Z

T
_
D

T(t)
_
1
Z

_
t

T
_
1

i1
<t

A
Z

i
t

T
_
_
=

i=1
E
Q
_
_
E
P
_
Z

T
_
D

T(t)
_
1
Z

_
t

T
_1

i1
<t

A
_
_
=E
Q
_
_
E
P
_
Z

T
_
D

T(t)
_
1
Z

_
t

T
_1
1/Z

(t

T)>0
1

A
_
_
,
where the second-to-last equality relies on the identity of (2.36). This yields the
representation in (2.35). The other statements follow from choosing

T = T, D =
N(T) and D = 1/Z

(T).
For the case of strict local martingales the equivalence of the last corollary is
generally not true. Take as an example N() 1 and Z

() a strict local martingale

under P. Then, Z

()N() Z

() is a local P-martingale but N()1

1/Z

()>0

1
1/Z

()>0
is clearly not a local Q-martingale. The reason for this lack of symmetry
is that a sequence of stopping times that converges P-almost surely to T need not
necessarily converge Q-almost surely to T.
We have seen that Theorem 4 implies that 1/Z

() stopped at zero is a
martingale under the new measure. As Delbaen and Schachermayer (1995a) and
Pal and Protter (2010) have discussed, the other direction holds trivially true:
Let Q denote some measure; M() a Q-martingale started at some positive value
M(0) > 0; and T
0
the rst hitting of zero by M(). Then, under the new measure
d

P := M(TT
0
)dQ, the process 1/M() is again a local martingale due to Girsanovs
theorem and Itos formula. It is a martingale, if and only if M() does not hit zero
Chapter 2. The Markovian Case 50
under the original measure Q.
In order to simplify computations even more, the following change of numeraire
for strictly positive wealth processes can be useful.
Corollary 4 (Change of numeraire). Let (, ) and (, ) denote two strategies
such that V

() is strictly positive and

T a stopping time with

T T. There exists
a measure Q

such that P is absolutely continuous with respect to Q

and
E
P
_
Z

T
_
V

T
__
= E
Q

_
_
V

T
_
V

T
_1
1/(Z

T)V

(

T))>0
_
_
, (2.37)
where E
Q

denotes the expectation with respect to the new measure Q

. Under this
measure Q

, the process W

() = (W

1
(), . . . W

K
())
T
with
W

k
(t

) := W
k
(t

) +
_
t

k
(u, S(u))du (2.38)
for all k = 1, . . . , K and t [0, T] is a K-dimensional Brownian motion stopped
at time

:= lim
i
inft [0, T] : Z

(t)V

(t) i. The process

(, ) here
is exactly the -specic market price of risk from Denition 3. The equality in
(2.13) holds until the stopping time

and, in particular, the processes S() and

1/(Z

()V

()) evolve until

under Q

according to
dS
i
(t) = S
i
(t)
d

j=1
a
i,j
(t, S(t))
j
(t, S(t))dt +S
i
(t)
K

k=1

i,k
(t, S(t))dW

k
(t) (2.39)
and
d
_
1
Z

(t)V

(t)
_
=

(t, S(t))
1
Z

(t)V

(t)
dW

(t)
for all i = 1, . . . , d and t [0, T]. Furthermore, the statements of Corollary 3 hold
with Q replaced by Q

and Z

() replaced by Z

()V

(). This yields the represen-

tation
U

(T t, s) = Q

_
1
Z

(T)V

(T)
> 0

T(t)
_
. (2.40)
Chapter 2. The Markovian Case 51
Proof. The proof goes exactly along the lines of Theorem 4 and Corollaries 2 and 3
with the obvious modications.
We emphasize the similarity of the Q

-dynamics of S() in (2.39) and the

PDE in (2.18) for U

.
The next proposition demonstrates how one can construct a probability space
that satises the technical conditions of this section:
Proposition 3 (Canonical probability space). Let (, T, P) denote any probability
space, equipped with a ltration F = T() that satises the usual conditions. There
exists a probability space (

,

T,

P), equipped with a ltration

T(), which supports
a probability measure Q such that

P is absolutely continuous with respect to Q, and
such that (2.33) holds for any

T(

T)-measurable random variable D 0 for any

T()-stopping time

T.
Furthermore, (

,

T,

P) has the same distributional properties as (, T, P);
that is, it supports a K-dimensional Brownian motion W(), a vector-valued process
(, ), a vector-valued process (, ), a matrix-valued process (, ), a d-dimensional
progressively measurable stock price process S() that satises (2.1), and a process
Z

() that satises (2.6). The ltration

T() can be assumed to be completed under
Q but not necessarily under

P. However, any process () that is progressively
measurable with respect to the

P-completed ltration has a modication

() that is
progressively measurable with respect to

T() and that is indistinguishable from ()
under

P. Furthermore, the process

W() of (2.34) is a Brownian motion under Q,
at least up to some stopping time.
Proof. We map the probability space on the canonical space

= C([0, ), R
n

) of R
n
-valued functions which are absorbed in the cemetery point
and continuous before absorption. We use here n = 1 +K +d +K +d +dK. The
paths in R
n
are the images of Z

(), W(), S(), (, S()), (, S()), and (, S()).

Chapter 2. The Markovian Case 52
As underlying ltration

T() we choose the right-continuous version of the canonical
one, that is, the ltration generated by the paths but not completed by the null-sets.
The mapping from to

induces a measure

P.
Although the ltration

T() is not completed and stochastic integrals appear,
the relations in (2.1) and (2.6) hold since all processes appearing are progressively
measurable with respect to

T(). The limit in the denition of stochastic integrals
is therefore a-fortiori

T()-measurable.
We observe that this probability space satises the technical requirements
to introduce the measure Q; see Meyer (1972). We now apply Theorem 4.
We nally augment the ltration

T() with all Q-nullsets without changing
the dynamics of the underlying processes under Q, and therefore nor under

P (see
Theorem 2.7.9 of Karatzas and Shreve 1991). We refer the reader to the argument
explicated in the rst remark of Section 1 in Delbaen and Schachermayer (1995a)
for the existence of an indistinguishable modication

() of some process () as
in the statement of the proposition.
The next remarks relate our results to the existing literature:
Remark 7 (Portfolio-generating functions). Fernholz (1999) introduces and dis-
cusses strategies (, ) of the form

i
(t, s) =
_
D
i
log(R(
m
(t, s))) + 1
d

j=1

m
j
(t, s)D
j
log(R(
m
(t, s)))
_

m
i
(t, s)
for all i = 1, . . . , d and (t, s) [0, T] R
d
+
where
m
(, ) denotes the market port-
folio with
m
i
(t, s) := s
i
/

d
j=1
s
j
and R any positive twice dierentiable function
satisfying some weak boundedness conditions. Then, Fernholz (1999) shows that
the pathwise formula
log
_
V

(T)
V

m
(T)
_
= log
_
R(
m
(T, S(T)))
R(
m
(0, S(0)))
_
+
_
T
0
(t, S(t))dt
Chapter 2. The Markovian Case 53
holds where : [0, T] R
d
+
R is some function that can be written down
explicitly. This yields, in connection with Corollary 4, the formula
U

(T, s) =
1
R(
m
(0, S(0)))
E
Q

m
_
R(
m
(T, S(T))) exp
__
T
0
(t, S(t))dt
_
1
1/(Z

(T)V

m
(T))>0
_
,
which can be used to compute optimal trading strategies.
Remark 8 (Perfect balance and optimal growth). Kardaras (2008) discusses in the
case of the market portfolio
i
(t, s) =
m
i
(t, s) := s
i
/

d
j=1
s
j
for all (t, s) [0, T]
R
d
+
and i = 1, . . . , d the perfect balance condition (, ) = a(, )(, ), which
is exactly the mean rate of return appearing in the dynamics of (2.39). If the
perfect balance condition holds under the real-world measure P, then each
component of the market portfolio
m
(, ) is a martingale. If (, ) is not the market
portfolio then this martingale property usually does not hold for the components
of (, ). However, the condition still implies that the strategy (, ) is growth-
optimal in the sense of Problem 4.6 of Fernholz and Karatzas (2009). That means
that (, ) maximizes the mean rate of return
T
(, )(, ) 1/2
T
(, )a(, )(, )
of the logarithm of the associated wealth process over all strategies (, ). More
generally, if
(, ) = a(, )(, ) + (, )c(, ),
for some c : [0, T]R
d
+
R
K
such that the stochastic exponential of

(, ) c(, )
in (2.12) is a martingale, then there is no arbitrage possible with respect to (, ).
This follows directly from the fact that the martingale property implies that P and
Q

are equivalent and thus rstly, W

() of (2.38) is a true Brownian motion and

secondly, the fraction V

()/V

() of (2.13) is a supermartingale for any strategy

(, ) under Q

; compare Corollary 4. Formally, for c(, ) ,= 0 the perfect balance

condition of Kardaras (2008) is satised in the case of (, )
m
(, ); however,
Chapter 2. The Markovian Case 54
the interest rates are out of balance. In the literature, a growth-optimal portfolio
is often also called numeraire portfolio; compare Section 3.5 of Platen (2002).
Remark 9 (Connections to the work of Delbaen and Schachermayer (1995a)). Del-
baen and Schachermayer (1995a) show that under equivalent technical assumptions,
there exists for any strictly positive local martingale Z() a measure such that un-
der the new measure 1/Z() is a martingale that can hit zero. In their work, Z()
represents the reciprocal of the stock price while in this work we treat the case of a
stochastic discount factor. In both situations a positive probability of 1/Z() hitting
zero under the new measure leads to an arbitrage opportunity. In this work, we
can additionally compute a strategy that uses minimal initial capital to perform the
arbitrage. Furthermore, we do not look only at arbitrage with respect to the money
market but also at arbitrage with respect to a much broader class of strategies.
2.6 Examples
In this section, we discuss several examples for markets that imply arbitrage oppor-
tunities. Examples 1, 2 and 3 study dierent strategies for the three-dimensional
Bessel process with drift. Example 4 concentrates on the reciprocal of the three-
dimensional Bessel, a standard example in the bubbles literature. Finally, Exam-
ple 5 illustrates a process that leads to a hedging price that is not dierentiable
and not even continuous but where the delta hedge still works.
Example 1 (Three-dimensional Bessel process with drift - money market). One of
the best known examples for markets with arbitrage is the three-dimensional Bessel
process, as discussed in Section 3.3.C of Karatzas and Shreve (1991). A Bessel
process starting at some point x > 0 is in distribution identical to the Euclidean
norm of a three-dimensional Brownian motion with the rst component starting
at x and the other two components starting at zero. We study here a class of
Chapter 2. The Markovian Case 55
models that contain the Bessel process as special case and generalize the example
for arbitrage of A.V. Skorohod in Section 1.4 of Karatzas and Shreve (1998). For
that, we begin with dening an auxiliary stochastic process X() as a Bessel process
with drift c, that is,
dX(t) =
_
1
X(t)
c
_
dt + dW(t) (2.41)
for all t [0, T] with W() denoting a Brownian motion on its natural ltration
F = F
W
and c [0, ) a constant. The process X() is strictly positive, since
it is a Bessel process, thus strictly positive under the equivalent measure where
W(t) ct
0tT
is a Brownian motion. The stock price process is now dened via
the stochastic dierential equation
dS(t) =
1
X(t)
dt + dW(t) (2.42)
for all t [0, T]. Both processes X() and S() are assumed to start at the same
point S(0) > 0. From (2.41) and (2.42) we obtain directly the identity X(t) =
S(t) ct, which yields the stock price dynamics
dS(t) =S(t)
_
1
S
2
(t) S(t)ct
dt +
1
S(t)
dW(t)
_
for all t [0, T]. Furthermore, since c 0 holds, we have strictly positive stock
prices S(t) > ct 0 for all t [0, T]. Thus, for c > 0, the model allows for an
obvious arbitrage in the sense of Denition 1.2 in Guasoni et al. (2010). If c = 0
then S() X() and the stock price process is a Bessel process. Of course, the
market price of risk is exactly 1/X() or, more precisely, we have
(t, s) = 1/(s ct)
and
Z

(t) = exp
_

_
t
0
1
S(u) cu
dW(u)
1
2
_
t
0
1
(S(u) cu)
2
du
_
Chapter 2. The Markovian Case 56
for all (t, s) [0, T] R
+
with s > ct. Thus, the reciprocal 1/Z

() of the stochastic
discount factor hits zero exactly when S(t) hits ct. This follows directly from the
Q-dynamics of 1/Z

() derived in Corollary 2 and a strong law of large numbers as

in Lemma A.2 of Kardaras (2008).
Let us start by looking at a general, for the moment not-specied Markovian
trading strategy (, ) whose associated wealth at time T is a function of the stock
price, that is, V

(T) =: p(S(T)). For all (t, s) [0, T] R

+
with s > ct, by relying
on Corollary 3 and changing the measure Q to

Q under which
_

S(t)
_
0tT
:=
S(t) ct
0tT
is a Brownian motion, we obtain
U

(T t, s) =E
t,s
_
Z

(T)V

(T)
Z

(t)V

(t)
_
=
1
V

(t)
E
Q
_
p(S(T))1
min
tuT
S(u)cu>0

T(t)

S(t)=s
=
1
V

(t)
E

Q
_
exp
_
c
_

S(T)

S(t)
_

c
2
(T t)
2
_
p(S(T))1
min
tuT

S(u)>0

T(t)
_

S(t)=sct
=
1
V

(t)
_

0
exp
_
c(y s + ct)
c
2
(T t)
2
_
p(y + cT)

1
_
2(T t)
_
exp
_

(y s + ct)
2
2(T t)
_
exp
_

(y + s ct)
2
2(T t)
__
dy
(2.43)
=
1
V

(t)
_
_

cTs

Tt
1

2
exp
_

z
2
2
_
p(z

T t + s)dz
exp(2cs 2c
2
t)

_

cT2ct+s

Tt
1

2
exp
_

z
2
2
_
p(z

T t s + 2ct)dz
_
, (2.44)
where we have plugged in the density of a Brownian motion absorbed at zero (com-
pare Problem 2.8.6 of Karatzas and Shreve 1991) and made use of the substitution
z = (y s + cT)/

T t and z = (y + s + cT 2ct)/

T t, respectively.
Chapter 2. The Markovian Case 57
Let us consider the investment in the money market only, to wit, the strategy

0
(, ) 0 and V

0
(t) 1 p(s) for all (t, s) [0, T]R
+
. This yields the hedging
price of one monetary unit
U

0
(T t, s) =
_
s cT

T t
_
exp(2cs 2c
2
t)
_
s cT + 2ct

T t
_
, (2.45)
where denotes the cumulative standard normal distribution function. In the
special case c = 0 we have
U

0
(T t, s) = 2
_
s

T t
_
1. (2.46)
For the rst derivative we obtain

s
U

0
(T t, s) =2c exp(2cs 2c
2
t)
_
s cT + 2ct

T t
_
+

2
(T t)
exp
_

(cT s)
2
2(T t)
_
,
which simplies in the case of c = 0 to

s
U

0
(T t, s) =

2
(T t)
exp
_

s
2
2(T t)
_
. (2.47)
It can be easily checked that U

0
solves the PDE in (2.18) for all (t, s) [0, T] R
+
with s > ct. This is sucient to apply Theorem 1 (compare Remark 5) to nd the
optimal hedging strategy of one monetary unit:

0
(t, s) = s

s
log
_
U

0
(T t, s)
_
. (2.48)
For c = 0 we obtain thus from (2.46) and (2.47) the representation

0
(t, s) =
2
s

Tt

_
s

Tt
_
2
_
s

Tt
_
1
> 0, (2.49)
where denotes the standard normal density.
Chapter 2. The Markovian Case 58
It comes at no surprise that, in order to beat the money market, we have to
be long the stock. The strategy
0
(, ) has for c = 0 another interpretation. To
derive it, we observe that U

0
(T t, s) is the probability that a Brownian motion

W() starting at s does not hit zero before time T t. Using the density of the
hitting time T
0
:= inft 0 :

W(t) = 0, (compare for example Proposition 2.8.5
of Karatzas and Shreve 1991) yields
U

0
(T t, s) =Q(T
0
> T t) =
1

2
_

Tt
s
y
3
2
exp
_

s
2
2y
_
dy,
which gives

s
U

0
(T t, s) =
1

2
_

Tt
_
1
y
3
2

s
2
y
5
2
_
exp
_

s
2
2y
_
dy
and

0
(t, s) =1
1

2
_

Tt
s
3
y
5
2
exp
_

s
2
2y
_
dy
U

0
(T t, s)
.
This is exactly

0
(t, s) = 1 s
2
E
Q
_
1
T
0

min
0uTt

W(u) > 0
_
.
It is well-known that a Bessel process allows for arbitrage. Compare for example
Example 3.6 of Karatzas and Kardaras (2007) for an ad-hoc strategy that corre-
sponds to a hedging price of (1) for a monetary unit if S(0) = T = 1. We have
improved here the existing strategies and found the optimal one, which corresponds
in this setup to a hedging price of U

0
(1, 1) = 2(1) 1 < (1).
Remark 10 (Multiple solutions for the PDE in (2.18)). We observe that the hedging
price U

0
in (2.45) depends on the drift c. Also, U

0
is suciently dierentiable,
thus by Proposition 2 uniquely characterized as the minimal nonnegative solution
of the PDE in (2.18), which does not depend on the drift c. The uniqueness of
U

0
by Proposition 2 and the dependence of U

0
on c do not contradict each other,
Chapter 2. The Markovian Case 59
since the nonnegativity of U

0
has only to hold for these points (t, s) [0, T] R
+
that can be attained by S() at time t. For a given time t [0, T], these are
only the points s > ct. Thus, as c increases, the nonnegativity condition weakens
since it has to hold for fewer points, thus U

0
can become smaller and smaller.
Indeed, plugging in (2.45) the point s = ct yields U

0
(T t, ct) = 0. In summary,
while the PDE itself does only depend on the (more easily observable) volatility
structure of the stock price dynamics, the mean rate of return determines where
the PDE has to hold and thus, contributes to determining the exact amount of
possible arbitrage.
In the next example, we price and hedge a European call within the same
class of models as in the last example:
Example 2 (Three-dimensional Bessel process with drift - stock and European call).
Since we do not know a priori any (possibly suboptimal) strategy that leads to the
value (S(T) L)
+
at time T for some strike L 0 we cannot rely on Theorem 1
and have to tackle this question slightly dierently using the results of Theorem 2.
Plugging in p(y) = (y L)
+
in (2.44), dening
h
p
(T t, s) := E
t,s
_
Z

(T)
Z

(t)
(S(T) L)
+
_
for all (t, s) [0, T] R
d
+
with s > ct, and using the notation a b := maxa, b
we can simplify the expected risk-adjusted value as follows:
h
p
(T t, s) =
_

(cTL)s

Tt
1

2
exp
_

z
2
2
_
(z

T t + s L)dz exp(2cs 2c
2
t)

_

(cTL)2ct+s

Tt
1

2
exp
_

z
2
2
_
(z

T t s + 2ct L)dz
=
_
T t
2
exp
_

(s (cT L))
2
2(T t)
_
+ (s L)
_
s (cT L)

T t
_
exp(2cs 2c
2
t)
_
_
T t
2
exp
_

((cT L) 2ct + s)
2
2(T t)
_
Chapter 2. The Markovian Case 60
+ (2ct s L)
_
(cT L) + 2ct s

T t
__
. (2.50)
The modied put-call parity of Corollary 1 could now be applied to give us directly
the hedging price of a European put. If L cT, in particular if L = 0, the last
expression simplies to
h
p
(T t, s) =s
_
s cT

T t
_
+ exp(2cs 2c
2
t)
_
2ct s cT

T t
_
(s 2ct) LU

0
(T t, s),
where U

0
denotes the hedging price of one monetary unit given in (2.45). It is
simply the dierence between the hedging price of the stock and L monetary units
since L cT implies S(T) > cT L almost surely and the call is always exercised.
Using L = 0 we get the value of the stock.
For L = c = 0, the last equality yields h
p
(t, s) = s for all (t, s) [0, T] R
+
and the stock cannot be arbitraged. There are at least two other ways to see this
result right away. Simple computations show directly that Z

(T) = S(0)/S(T) if
c = 0 and thus, for the strategy
1
(, ) 1, which invests fully in the market,
we obtain U

1
(, ) 1. Alternatively, using the representation of U

1
implied by
(2.33) we see that the hedging price is just the expectation of a Brownian motion
stopped at zero, thus the expectation of a martingale started at one. Every method
on its own shows the lack of relative arbitrage with respect to the market if c = 0.
On the other hand, if c > 0, then relative arbitrage with respect to the
market is possible. In this case, the representation implied by (2.33) shows that as
soon as the Brownian motion is stopped, which is the rst time S(t) equals ct, the
value of the random variable of which the expectation is taken jumps to zero and
thus the stopped process is not a martingale any more. Obviously,
U

1
(T t, s) = E
t,s
_
Z

(T)S(T)
Z

(t)s
_
=
1
s
h
p
(T t, s)
Chapter 2. The Markovian Case 61
is suciently dierentiable and thus, Remark 5 and Theorem 1 yield the optimal
arbitrage opportunity

1
(t, s) =
2ct

Tt

_
scT

Tt
_
+
_
scT

Tt
_
+ exp(2cs 2c
2
t)
_
2ctscT

Tt
_
(2cs 4c
2
t + 1)
U

1
(T t, s)
.
We can now nd the corresponding strategy for the call price of (2.50). Assuming
for the sake of notation that L cT, Theorem 2 yields

p
(t, s) =
_
s
_
s L

T t
_
+ s exp(2cs 2c
2
t)
_
2ct s L

T t
_
(2cs + 2cK 4c
2
t + 1) 2c

T t
_
L + s 2ct

T t
__
/h
p
(T t, s)
as the optimal strategy. If c = 0 this simplies to

p
(t, s) =
s
_

_
sL

Tt
_
+ 1
_
s+L

Tt
__
h
p
(T t, s)
.
Two notable observations can be made. First, in this model both the money
market and the stock simultaneously have a hedging price cheaper than their current
price, as long as c > 0. Second, in contrast to the classical theory of Financial
Mathematics, the mean rate of return under the real-world measure does matter
in determining the hedging price of calls (or other derivatives) since it inuences
the possibilities of arbitrage.
We can now also nd a quantile hedge for strategies (, ) whose associated
wealth process V

(T) is only a function of the market S(T). To wit, for such (, )

and some given [0, 1) one can compute the value and optimal strategy for the
quantity

U
,
(T, s) := inf v > 0 : strategy such that P
s
(V
v,
(T) > V

(T)) 1 ;
(2.51)
that is,

U
,
(T, s) represents how much initial capital is needed to obtain the termi-
nal wealth V

(T) with a given probability 1. This question has been resolved in

Chapter 2. The Markovian Case 62
the case of the existence of an equivalent local martingale measure by Section 2.4
of Follmer and Leukert (1999) relying on the Neyman-Pearson lemma. The next
example illustrates a possible approach for markets with arbitrage in the case of
the market portfolio in the Bessel process setup. Recently, Bayraktar et al. (2010a)
solved the problem of nding a quantile hedge by means of formulating a stochastic
control problem.
Example 3 (Three-dimensional Bessel process - quantile hedging). In order to com-
pute the quantity in (2.51) it is clearly sucient to compute the optimal strat-
egy and the initial wealth for the contingent claim D = S(T)/S(0)1
S(T)
for
:= infz > 0 : P
S(0)
(S(T) z) . By similar considerations as in Examples 1
and 2, the probability P
S(0)
(S(T) z) equals the expectation of a Brownian motion
starting at S(0) and stopped at zero to be above z at time T. We obtain, similar
to (2.43),
P
S(0)
(S(T) z) =E
Q
S(0)
_
1
S(TT
0
)>z
S(T T
0
)
S(0)
_
=
1
S(0)
_

z
y

2T
_
exp
_

(y S(0))
2
2T
_
exp
_

(y + S(0))
2
2T
__
dy
=

T
S(0)

2
_
exp
_

(z S(0))
2
2T
_
exp
_

(z + S(0))
2
2T
__
+
_
S(0) z

T
_
+ 1
_
z + S(0)

T
_
,
where S() is a Brownian motion starting at S(0) under Q
S(0)
and T
0
the rst hitting
of zero by S(). The truncation and the optimal strategy can now be computed
as we have done for calls in Example 2. We omit the computations since they do
not contain any new insights.
Pal and Protter (2010) compute call prices for the reciprocal Bessel process
model. This process has appeared several times in the bubbles literature, often
Chapter 2. The Markovian Case 63
called the constant elasticity of variance (CEV) process; see, for example, Sec-
tion 2.2.2 of Cox and Hobson (2005) or Example 1.2 of Heston et al. (2007). We
discuss next how the results of the last examples relate to this model and illustrate
that even under the NFLVR condition relative arbitrage is possible.
Example 4 (Reciprocal of the three-dimensional Bessel process). Let the stock price

S() have the dynamics

d

S(t) =

S
2
(t)dW(t)
for all t [0, T] with W() denoting a Brownian motion on its natural ltration F =
F
W
. The process

S() is exactly the reciprocal of the process S() of Examples 1 and
2 with c = 0, thus strictly positive. We observe that there is no classical arbitrage
since the mean rate of return is zero and thus P is already a local martingale
measure. However, there is arbitrage possible with respect to the stock. To wit, if
one wants to hold the stock at time T, one should not buy the stock at time zero,
but use the strategy
1
(, ) below for a hedging price smaller than

S(0) along with
the suboptimal strategy
1
(, ) 1. That is, the strategy
1
(, ) contains a bubble
according to Denition 5.
We have already observed that

S(T) = 1/S(T), which is exactly the stochas-
tic discount factor in Example 1 for c = 0 multiplied by

S(t). Thus, as in (2.46)
the hedging price for the stock is
U

1
(T t, s) = 2
_
1
s

T t
_
1 < 1 (2.52)
along with the optimal strategy

1
(t, s) =
2
_
1
s

Tt
_
s

T t
_
2
_
1
s

Tt
_
1
_ + 1 < 1
for all (t, s) [0, T) R
+
similar to (2.49). By Corollary 4, the hedging price U

1
could also be calculated as one minus the probability of explosion (to ) of the
Chapter 2. The Markovian Case 64
process

S() before time T under Q

1
, where it has the dynamics
d

S(t) =

S
3
(t)dt

S
2
(t)dW

1
(t).
Alternatively, one could also calculate the probability of implosion (to zero)
of the process S() = 1/

S() with dynamics

dS(t) =dW

1
(t)
for all t [0, T] under Q

1
, which is again a Brownian motion as in Example 1 but
now starting at 1/S(0). For pricing calls, we observe
_

S(T) L
_
+
= L

S(T)
_
1
L

1

S(T)
_
+
=
L
S(t)

S(t)
S(T)
_
1
L
S(T)
_
+
for L > 0. Thus, the price at time t of a call with strike L in the reciprocal Bessel
model is the price of L

S(t) puts with strike 1/L in the Bessel model and can be
computed from Example 2 and Corollary 1. For S(0) = 1, simple computations
will lead directly to Equation (6) of Pal and Protter (2010). The optimal strategy
could now be derived with Theorem 2.
The next example
7
illustrates that U

need not be dierentiable or even

smooth in the stock price dimension in order to nd an optimal strategy.
Example 5 (U

not dierentiable). Let us consider a market with time horizon

T = 2 and one stock with dynamics
dS(t) =
_

_
1
S(t)>
1
2

dW(t) if t < 1,
1
S(t)>1
_
1
S(t)1
dt + dW(t)
_
if t 1.
Thus, up to time t = 1 the stock price is either constant or evolves as Brownian
motion stopped at 1/2. If at time t = 1 the stock price is less than or equal to
7
We developed this example after a helpful conversation with Daniel Fernholz.
Chapter 2. The Markovian Case 65
one, it stays constant and otherwise evolves as a three-dimensional Bessel process
shifted by one. From (2.46) the hedging price for the money market is
U

0
(t, s) =
_

_
1 if s 1,
2
_
s1

t
_
1 if s > 1
for t [0, 1]. Then, the hedging price U

0
is not continuous for s = 1, thus not
dierentiable. However, there always exists an optimal strategy
0
. For (t, s)
[0, 1] [0, 1] no arbitrage is possible, which implies that
0
(2 t, s) = 0 is optimal.
For (t, s) [0, 1] (1, ) we know that the stock price always stays above one
and the optimal strategy is the one given in (2.49) with s replaced by s 1 on
the right-hand side. For t (1, 2], the function U

0
(t, s) = E
2t,s
[U

0
(1, S(1))] is
easily shown to be suciently dierentiable. Therefore, we can apply Theorem 1 to
obtain
0
(2t, s). We have illustrated that, although U

0
(t, s) is not dierentiable
in the stock price dimension, namely for (t, s) [0, 1] 1 in this example, there
can nevertheless exist an optimal strategy
0
(, ).
2.7 Conclusion
It has been proven that, under weak technical assumptions, there is no equivalent
local martingale measure needed to nd an optimal hedging strategy based upon
the familiar delta hedge. To ensure its existence, weak sucient conditions have
been introduced that guarantee the dierentiability of an expectation parameterized
over time and over the original market conguration. The dynamics of stochastic
processes simplify after a non-equivalent change of measure and a generalized Bayes
rule has been derived. From an analytic point of view, results of Fernholz and
Karatzas (2010) concerning non-uniqueness of the Cauchy problem of (2.18) have
been generalized to a class of PDEs that allow for a larger set of drifts. With this
newly developed machinery, some optimal trading strategies have been computed
Chapter 2. The Markovian Case 66
addressing standard examples for which so far only ad-hoc and not necessarily
optimal strategies have been known.
2.8 Condition that hedging price solves a PDE
In this section, which serves as an appendix to this chapter, we provide a neces-
sary condition for U

of (2.16) solving the PDE in (2.18) and especially for being

suciently dierentiable.
One way to ensure smoothness of U

is to follow the arguments in Subsec-

tion 2.4.3. To start, we formulate an additional assumption:
(A1) The functions

k
(, ) are for all k = 1, . . . , K locally Lipschitz and locally
bounded.
In particular, this assumption restricts the possible strategies (, ); however, it
allows for the market portfolio
m
(, ), for example.
Then, Assumptions (A1), (A1), and (A2) guarantee the necessary smooth-
ness of U

. This can be seen directly from Theorem 3, when we replace Z

() by
Z

()V

() and set the payo function p() 1.

An alternative way to show smoothness goes along the lines of Section 9 in
Fernholz and Karatzas (2010): First, we remove the stochastic integral by assuming
that there exists a real-valued function H

C
1,2
([0, T] R
d
+
) such that
d

i=1

i,k
(t, s)s
i
D
i
H

(t, s) =

k
(t, s) (2.53)
for all k = 1, . . . , K and (t, s) [0, T] R
d
+
. That is, if the covariance process a(t, s)
has a multiplicative inverse a
1
(t, s) on [0, T] R
d
+
, then H

has partial derivatives

of the form
D
i
H

(t, s) =

d
j=1
a
1
i,j
(t, s)
j
(t, s)
i
(t, s)
s
i
. (2.54)
Chapter 2. The Markovian Case 67
This condition basically means that (, ) and (, ) are suciently smooth in time
and space and have an anti-derivative. As Remark 5 discusses, this assumption can
easily be slightly generalized. Applying Itos formula to H

yields
H

(T, S(T)) H

(t, S(t))
_
T
t
_
LH

(u, S(u))

t
H

(u, S(u))
_
du
=
_
T
t

T
(u, S(u))dW(u),
where L is the innitesimal generator of S() dened in (2.22). Collecting all
deterministic terms in a function k

: [0, T] R
d
+
R, we obtain
k

(t, s) :=LH

(t, s) +

t
H

(t, s)
1
2
|

(t, s)|
2
=
1
2
d

i=1
d

j=1
s
i
s
j
a
i,j
(t, s)D
2
i,j
H

(t, s) +
d

i=1
s
i

i
(s, t)D
i
H

(s, t)
+

t
H

(t, s)
1
2
d

i=1
d

j=1
s
i
s
j
a
i,j
(t, s)D
i
H

(t, s)D
j
H

(t, s)
=
1
2
d

i=1
d

j=1
s
i
s
j
a
i,j
(t, s)D
2
i,j
H

(t, s) +
d

i=1
d

j=1
s
i
a
i,j
(t, s)
j
(t, s)D
i
H

(t, s)
+

t
H

(t, s) +
1
2
d

i=1
d

j=1
s
i
s
j
a
i,j
D
i
H

(t, s)D
j
H

(t, s), (2.55)

where the last equality follows from the denitions of H

and

(, ) in (2.53) and
(2.12). Using that, (2.17) can now be written as
U

(t, s) =exp(H

(T t, s))E
Tt,s
_
exp(H

(T, S(T))) exp

__
T
Tt
k

(u, S(u))du
__
.
To proceed, we make the additional assumption that the deterministic function
G

: [0, T] R
d
+
R
+
dened as
G

(t, s) :=exp(H

(T t, s))U

(t, s) (2.56)
=E
Tt,s
_
exp(H

(T, S(T))) exp

__
T
Tt
k

(u, S(u))du
__
,
Chapter 2. The Markovian Case 68
which does not involve a stochastic integral any more, solves the time-inhomogeneous
Cauchy problem

t
G

(t, s) = LG

(t, s) + k

(T t, s)G

(t, s). (2.57)

To sum up, this second approach requires the existence of a smooth function H

satisfying (2.54), and G

being a solution of the Cauchy problem in (2.57). Chap-

ter 9 of Fernholz and Karatzas (2010) discusses general conditions under which the
later assumption is satised; however, we feel that these assumptions tend to be
more restrictive than assumptions (A1), (A1), and (A2). Nevertheless, if they hold
true, then the next lemma concludes the argument:
Lemma 4 (PDE for U

). If G

dened in (2.56) solves the PDE in (2.57) then

U

solves the PDE in (2.18).

Proof. Omitting arguments for the sake of clarity, we have
D
i
U

=U

D
i
H

+ exp(H

)D
i
G

,
D
2
i,j
U

=U

D
2
i,j
H

+ U

D
i
H

D
j
H

+ exp(H

)D
i
H

D
j
G

+ exp(H

)D
j
H

D
i
G

+ exp(H

)D
2
i,j
G

.
Therefore, collecting the U

terms and comparing them to the representation of k

in (2.55) and collecting the exp(H

) terms we obtain
1
2
d

i=1
d

j=1
s
i
s
j
a
i,j
D
2
i,j
U

+
d

i=1
d

j=1
s
i
a
i,j

j
D
i
U

= U

_
k

t
H

_
+exp(H

)
_
d

i=1
s
i
D
i
G

j=1
a
i,j
(
j
+ s
j
D
j
H

) +
1
2
d

i=1
d

j=1
s
i
s
j
a
i,j
(t, s)D
2
i,j
G

_
= exp(H

)(G

)
_
exp(H

)

t
H

_
G

+ exp(H

)LG

,
where the last equality follows from the identities
d

j=1

j,k
(
j
+ s
j
D
j
H

) =
k
Chapter 2. The Markovian Case 69
and
K

k=1

i,k

k
=
i
for all i, k = 1, . . . , d. That proves the statement since time goes in the reverse
direction.
Chapter 3. Completeness and Relative Arbitrage 70
Chapter 3
Completeness and Relative
Arbitrage
3.1 Introduction
This chapter examines conditions under which contingent claims can be replicated
by dynamic trading in the stock market. Let S() be a continuous, d-dimensional
Ito-process (the stock price process) with respect to a ltration T() and D 0
be an T(T)-measurable random variable (the claim). The question then is when
D can be represented as a stochastic integral of some progressively measurable
process (the trading strategy) with respect to S(). Replicable claims have been
completely characterized if S() satises the No Free Lunch with Vanishing Risk
(NFLVR) condition. This notion was introduced by Delbaen and Schachermayer
(1994) and is equivalent to the existence of an equivalent measure under which S()
is a local martingale. If the supremum over the expected values of D under all
equivalent local martingale measures (ELMMs) is attained, then the claim D can
be replicated.
We generalize this characterization for replicable claims to markets that do
Chapter 3. Completeness and Relative Arbitrage 71
not necessarily satisfy NFLVR, but allow for a stochastic discount factor; this cor-
responds to a weak structural restriction on the drift of the process. Stochastic
discount factors are local martingales and take the place of the Radon-Nikodym
derivatives that are used to change the original measure to an ELMM in the case of
NFLVR. If the supremum over the expected values of D multiplied by all stochastic
discount factors is attained, then the claim D can be replicated.
NFLVR is a mathematical concept introduced to characterize markets that
admit an ELMM, and thus exclude arbitrage opportunities. However, from an eco-
nomic perspective, it is reasonable to consider models which do not satisfy NFLVR.
As Loewenstein and Willard (2000a) and Hugonnier (2010) discuss, models with-
out NFLVR may nevertheless lead to an equilibrium where rational agents have
an optimum. Thus, although NFLVR is a convenient mathematical assumption, it
does not always accurately reect our economic intuition of arbitrage. In partic-
ular, the existence of a stochastic discount factor prevents arbitrage opportunities
from being scaled up, thus allowing for the existence of a numeraire portfolio and
of solutions to utility maximization problems; see Karatzas and Kardaras (2007).
In a similar vein, the theory of real world pricing in the Benchmark
Approach (see Platen and Heath 2006) acknowledges that no ELMM is needed
to have the concept of a price for contingent claims. Models that allow for some
kind of arbitrage are furthermore studied in the framework of Stochastic Portfolio
Theory (see Fernholz 2002; Fernholz and Karatzas 2009), which starts from the
premise of realistically describing the evolution of market weights over long time
horizons and provides simple testable conditions, such as diversity or sucient
intrinsic volatility, under which arbitrage does exist. These insights and ideas lead
to the conclusion that models which impose the existence of a stochastic discount
factor, but do not necessarily additionally assume NFLVR, are a natural class of
models to study. This chapter is therefore a step to close the gap in the theory
Chapter 3. Completeness and Relative Arbitrage 72
between the class of models with and without the assumption of NFLVR.
A nonnegative stock price process S() has been dubbed a bubble, if the set of
ELMMs is nonempty and S() is a strict local martingale under an ELMM. Such a
stock price models an asset that is overpriced compared with its intrinsic value, as
measured by its expectation under an ELMM. It behaves locally like a martingale,
but in the long run, behaves like a strict supermartingale. Academic literature has
recently devoted substantial attention to bubbles, given that they are able to model
seemingly overpriced stocks as in the Internet Bubble within the framework of
NFLVR. We suggest Cox and Hobson (2005), Heston et al. (2007), and Jarrow
et al. (2010) as some initial references to this literature. An asymmetry within
the class of admissible trading strategies is the reason that this phenomenon of
overpriced stock prices appears in models that satisfy NFLVR. Such models allow
for the bond to be sold, but usually do not allow for the stock to be sold in order
to prot from it being overpriced. For this subtle point, we refer the reader to
Yan (1998), where allowable strategies are introduced to avoid the asymmetry
introduced by admissibility constraints. If one is willing to accept the presence of
bubbles, then a natural next step is to allow for some kind of arbitrage, essentially
reecting a bubble in the money market. Such arbitrage arises, for example, after
a change of numeraire with an asset that has a bubble.
Having characterized the claims that can be perfectly replicated, it is a natu-
ral next step to identify the markets in which all claims can be perfectly replicated.
Such a market is then called complete. For markets without arbitrage opportu-
nities, the Second Fundamental Theorem of Asset Pricing (2nd FTAP) gives a
sucient and necessary condition, stating that a market is complete if and only
if the ELMM is unique. This insight regarding the equivalence of the existence of
a replicating strategy for any claim and the uniqueness of a pricing measure can
be traced back to the seminal papers by Harrison and Kreps (1979) and Harrison
Chapter 3. Completeness and Relative Arbitrage 73
and Pliska (1981). For a list of more recent results and references, we point the
reader to Section 1.8 of Karatzas and Shreve (1998). Recently, Lyaso (2010) has
studied completeness in markets where capital gains and additional information to
the investors are modeled separately.
In this chapter, we show that the 2nd FTAP can be extended to markets
that do not proscribe arbitrage. Its generalized version then states that a market
is complete if and only if the stochastic discount factor is unique. Clearly, in the
case of NFLVR, this condition reduces to the classical one since then any stochastic
discount factor generates an ELMM. We conclude that the question regarding the
existence of arbitrage and the question regarding the completeness of the market
can be addressed separately from one another; see also Jarrow et al. (1999) and
Section 10.1 of Fernholz and Karatzas (2009). The proof of the 2nd FTAP in
markets that do not proscribe the existence of arbitrage is simple. It relies on a
change-of-numeraire technique and an application of the classical 2nd FTAP.
Often, however, completeness is too strong a requirement. We instead intro-
duce the notion of quasi-completeness, which only takes into consideration claims
measurable with respect to the stock price ltration. We show that markets whose
drift and diusion components are measurable with respect to the stock price l-
tration are quasi-complete; this generalizes Proposition 1 in Chapter 2, where the
Markovian case is studied.
An important element in Stochastic Portfolio Theory is the concept of (strong)
relative arbitrage. One says that there exists a relative arbitrage opportunity with
respect to some trading strategy () if there exists another trading strategy ()
that outperforms (); that is, if trading according to () yields a higher terminal
wealth than trading according to (). The relative arbitrage is called strong if the
terminal wealth is strictly dominated almost surely. It has been unclear up until
this point whether a relative arbitrage opportunity necessarily implies a strong one.
Chapter 3. Completeness and Relative Arbitrage 74
Having now the characterization of perfect replication and completeness at hand,
we can resolve this question: The existence of relative arbitrage does usually not
imply that of strong relative arbitrage; however, it does if the market is quasi-
complete. We can further state very precise conditions for the existence of both
relative arbitrage and strong relative arbitrage opportunities.
We have included several examples of toy markets that illustrate various
subtle points of our results in the sections that follow. Example 6 demonstrates
that, although a given claim might be measurable with respect to the stock price
ltration, the trading strategy to replicate the claim does not necessarily have
to be measurable with respect to this ltration. Example 7 illustrates that the
drift is important for determining whether a market is quasi-complete or not. In
Example 8, we study a stock price with a bubble whose minimal replicating cost is
not below its current price. Changing this model slightly then yields Example 9,
which treats a model without an ELMM but in which the minimal replicating price
for $1 is again $1. Finally, Example 10 provides the dynamics of a stock price that
implies a strong but diminishing arbitrage opportunity.
We introduce the model and admissible trading strategies in Section 3.2. For
a given claim, we provide necessary and sucient conditions to decide whether it
can be replicated in Section 3.3. In Section 3.4, we state and prove a generalized
version of the 2nd FTAP and discuss the concept of quasi-completeness. We then
apply the tools developed in the previous sections to link relative arbitrage and
strong relative arbitrage in Section 3.5 and we conclude in Section 3.6.
3.2 Setup
This section introduces the probabilistic market model and the concepts of market
prices of risk, stochastic discount factors, trading strategies, and (contingent) claims.
Throughout the chapter, we shall assume that all equalities and inequalities only
Chapter 3. Completeness and Relative Arbitrage 75
hold in an almost-sure sense.
3.2.1 Market model
We x a canonical
1
probability space (, T, P). We assume that = C([0, ), R
K
),
that is, is the space of all continuous functions W() = (W
1
(), . . . , W
K
())
T
taking values in R
K
for some xed K N. Furthermore, we x P so that the
process W() has the law of a K-dimensional Brownian motion. We denote by
F = T(t)
t0
the ltration generated by the paths of W(), and assume it satises
the usual assumptions. We further assume for some xed d N the existence of
a vector of d continuous, adapted processes S() = (S
1
(), . . . , S
d
())
T
with values
in (0, )
d
, which represent the price processes of the risky assets in an economy.
We assume the existence of a K-dimensional, vector-valued process () and of a
dK-dimensional, matrix-valued process (), both progressively measurable with
respect to the underlying ltration F, such that the dynamics of S() can be written
as
dS
i
(t) =S
i
(t)
K

k=1

i,k
(t)
_

k
(t)dt + dW
k
(t)
_
(3.1)
for all t 0 and i = 1, . . . , d. The process () is not assumed to be of rank d or
K. Thus, we do not exclude a-priori stock price models with a non-tradable state
variable, such as stochastic volatility models.
The strict positivity of S() will be guaranteed by imposing S(0) (0, )
d
and the integrability condition
_
T
0
_
|(t)|
2
+
d

i=1
K

k=1

2
i,k
(t)
_
dt < (3.2)
for all T 0.
1
To generalize the results presented here to more general semimartingale models is subject to
future research.
Chapter 3. Completeness and Relative Arbitrage 76
We denote by T
S
() the with the null sets of T augmented, right-continuous
ltration generated by S(). More precisely, we dene T
S
(t) := (S(u), u t) for
all t 0. Since S() is a strong solution of (3.1), we have the inclusion T
S
(t) T(t)
for all t 0.
3.2.2 Market prices of risk and stochastic discount factors
The special structure of the drift is a standard assumption imposed in order to ex-
clude the possibility of an arbitrage opportunity that could otherwise get scaled un-
boundedly; see Section 10 of Karatzas et al. (1991a) and the proof of Theorem 3.5 in
Delbaen and Schachermayer (1995b). We call the process () = (
1
(), . . . ,
K
())
T
in (3.1), which maps the volatility on the drift, a market price of risk.
It is clear that this process is usually not uniquely determined; for example
if the number of rows of () is smaller than the number of columns, that is, if
d < K. In this case, a set of R
K
-valued, F-progressively measurable processes ()
exists such that (3.1) is satised with any () replacing (). To make this precise,
we dene
:=
_
: [0, ) R
K
progressively measurable

_
T
0
|(t)|
2
dt < for all T 0
_
,

t
:=() [ ()() ()() . (3.3)
We call
t
, which as a direct consequence of (3.2) contains (), the set of all
market prices of risk. If any ()
t
replaces () in (3.1), the dynamics of S()
are unchanged. We observe that the stochastic process
m
() dened as

m
() := ()

()(), (3.4)
where denotes the Moore-Penrose pseudo-inverse of a matrix, is again a market
price of risk and therefore is also an element of
t
; see Corollary 1 of Penrose (1955).
Chapter 3. Completeness and Relative Arbitrage 77
For any process () in , we dene Z

() as its stochastic exponential; that

is,
Z

(t) := exp
_

_
t
0

T
(u)dW(u)
1
2
_
t
0
|(u)|
2
du
_
(3.5)
for all t 0. For ()
t
, we call Z

() a stochastic discount factor.

3.2.3 Trading strategies and claims
We shall consider a small investor who can trade dynamically in both a risk-free
asset, which pays zero interest rate, and in d stocks with price processes given by
S(). The assumption that the risk-free interest rate equals zero is made here for
convenience. We shall assume that the investor can trade in the market without
any frictions. In particular, we assume that the investor faces no transaction costs,
is allowed to trade continuous fractions of shares, and does not inuence market
prices. However, the investor shall be restricted to always having nonnegative
wealth, as specied in the next paragraph.
We call any progressively measurable vector () = (
1
(), . . . ,
d
())
T
, where
each component of () species the number of shares held by the investor following
that trading strategy, an (admissible) trading strategy for initial capital p 0 if the
corresponding wealth process V
p,
() with V
p,
(0) = p and dynamics
dV
p,
(t) =
d

i=1

i
(t)dS
i
(t) (3.6)
for all t 0 stays nonnegative.
For any T > 0, we call any nonnegative T(T)-measurable random variable
D 0 a (contingent) claim. A claim represents a certain monetary payo at time
T. Even without the existence of a traded asset S
i
() with S
i
(T) = D for some
i = 1, . . . , d, there might still exist some trading strategy () and some p > 0 such
that V
p,
(T) = D (respectively, V
p,
(T) D), in which case the claim is said to
be replicated (respectively, superreplicated) by ().
Chapter 3. Completeness and Relative Arbitrage 78
Remark 11 (On the admissibility constraint). In the classical theory of Financial
Mathematics, one also has to introduce a notion of admissibility for trading strate-
gies in order to prevent the investor from following the notorious doubling strate-
gies; see the discussion in Section 6 of Harrison and Kreps (1979). Usually, a more
general condition than the nonnegativity of the corresponding wealth process is
assumed. However, any such condition implies that the risk-adjusted wealth pro-
cess Z

()V
p,
() is a supermartingale; see Strasser (2003). This is no longer true
when one abstains from imposing the no-arbitrage condition. For example, if Z

()
is a strict local martingale and the wealth process is only restricted to stay above
some constant < 0, then Z

()V
p,
() is usually no longer a supermartingale.
This motivates the admissibility constraint made here, which mandates that the
wealth process stay nonnegative. We observe that under NFLVR, due to the super-
martingale property, any wealth process of a (super-)replicating strategy for some
nonnegative claim D 0 is again nonnegative, independently from the admissibil-
ity constraint. This fact will be used in the proofs of Section 3.3, where we apply
results of the no-arbitrage theory to obtain a characterization for claims that can
be replicated in markets that do not proscribe arbitrage. We shall revisit these
observations in Remark 12.
3.3 Existence of (super-)replicating trading strate-
gies
Given a specic claim, it is of interest to specify conditions under which its pay-
o can be obtained by means of dynamic trading in the stocks. Theorem 8.5 of
Karatzas et al. (1991b) provides a sucient condition for the replicability of strictly
positive claims. The authors allow for markets that are incomplete, as well as for
markets that admit arbitrage opportunities. Here, we extend their result to more
Chapter 3. Completeness and Relative Arbitrage 79
general volatility matrices () and to claims that are only nonnegative, and fur-
ther provide a minimal superreplicating price for general nonnegative claims. By
relying on duality methods, the question regarding the existence of superreplicating
strategies has been answered in full generality for markets satisfying NFLVR. We
refer the reader to Jacka (1992), Ansel and Stricker (1993), El Karoui and Quenez
(1995), and Delbaen and Schachermayer (1995c) for more on this topic; see also
Kramkov (1996) and Follmer and Kabanov (1998) for a more general class of mod-
els. In the following, we show that these results also extend to markets without an
ELMM.
Throughout this section, we x an horizon T > 0 and an T (T)-measurable
random variable D 0, which represents the claim. We dene
p := D
0
:= sup
()

E[Z

(T)D] [0, ], (3.7)

with
t
as in (3.3). We shall see in Theorem 5 that p represents the minimal
superreplicating price for D. We now introduce the sequence of stopping times

0
:= 0,

n
:= T inf
_
t [0, T]

m
(t) n
_
for all n N, where
m
() has been dened in (3.4) and denotes the minimum.
For any stopping time
n
and any T(
n
)-measurable random variable

D 0, we
set
p
n
(

D) := sup
()

E
_
Z

(
n
)

D
_
. (3.8)
By analogy with p of (3.7), p
n
(

D) can be considered the minimal price for super-
replicating

D at time
n
, as we shall demonstrate below. Towards this end, we
dene
p
e
n
(

D) := sup
()
e
n
E
_
Z

(
n
)

D
_
, (3.9)
Chapter 3. Completeness and Relative Arbitrage 80
where we denote

e
n
:= () [ (
n
)(
n
) (
n
)(
n
), E[Z

(T)] = 1 , = .
for all n N. The next denition is in the spirit of Delbaen and Schachermayer
(1995c):
Denition 8 (Maximal trading strategy). We call a trading strategy () maximal
if the supremum is a maximum in (3.7) with D = V
p,
(T).
Theorem 8 (b) below will motivate the word maximal, since it shows that
no trading strategy which outperforms a maximal one exists. It is already now clear
that () is maximal if some ()
t
exists such that Z

()V
p,
() is a martingale
up to time T.
We can now resolve the question regarding the existence of a (super-)replicating
strategy for claims in models that do not proscribe arbitrage. As the next lemmas
clarify, our argument utilizes the fact that the existence of a square-integrable mar-
ket price of risk guarantees that the market is basically, up to a stopping time, free
of arbitrage. Thus, we shall be able to nd a sequence of time-consistent trading
strategies, which eventually lead, path-by-path, to a superreplicating strategy.
Lemma 5 (Localized (super-)replication). Assume

D 0 is T(
n
)-measurable for
some n N. Then, the equality p := p
n
(

D) = p
e
n
(

D) holds, and the supremum
in (3.8) is attained if and only if it is attained in (3.9). The supremum p is the
minimal superreplicating price for

D at time
n
. More precisely, if p < then an
admissible trading strategy () exists such that
V
p,
(
n
)

D.
If p < , then there exist an T(
n
)-measurable claim

D

D, a trading
strategy (), and a market price of risk ()
e
n
, such that Z

()V
p,
() is a
martingale up to time
n
and V
p,
(
n
) =

D.
Chapter 3. Completeness and Relative Arbitrage 81
Furthermore, no trading strategy () exists for which V
c,
(
n
)

D, for any
c [0, p
n
(

D)).
Proof. First, assume that there exist () and c [0, p
n
(

D)) such that V
c,
(
n
)

D. We observe that Z

()V
c,()
() is a supermartingale for any ()
t
. Thus, we
obtain
E
_
Z

(
n
)

D
_
E[Z

(
n
)V
c,
(
n
)] c < p
n
(

D)
for all ()
t
, which leads to a contradiction with the denition of p
n
(

D) as a
supremum in (3.8).
Now, observe that

Z() Z

m
(
n
) is a martingale since it is bounded
by n. In particular, it denes a new measure Q on T(T), which is equivalent to
P, by dQ/dP =

Z(T). We introduce a ctional market

S() by

S() S(
n
).
Then,

S
i
() is a Q-local martingale for all i 1, . . . d. Thus, NFLVR holds for
the new market. Since the probability space is the canonical one, any measure

Q
on T(T) under which

S() is a local martingale and which is equivalent to P has a
representation d

Q/dP = Z

(T) for some ()

e
n
. Then, by the classical theory
for arbitrage-free markets, a trading strategy () exists such that

V
p
e
n
(

D),
(
n
)

D, where

V
p
e
n
(

D),
() is dened as in (3.6) with S() replaced by

S(); see Theorem 9
of Delbaen and Schachermayer (1995c). However, we have S(
n
)

S(
n
),
and therefore V
p
e
n
(

D),
(
n
)

D.
Together with the rst part of the proof, where we have shown that any c 0
that satises V
c,
(
n
) 0 for some trading strategy () also satises c p
n
(

D),
this also yields p
e
n
(

D) p
n
(

D). The inequality in the other direction follows from
the fact that for any ()
e
n
there exists ()
t
with Z

(
n
) = Z

(
n
). To
see this, set (t) = (t)1
tn
+
m
(t)1
t>n
for all t 0.
Corollaries 10 and 14 of Delbaen and Schachermayer (1995c) yield the ex-
istence of a claim

D

D, a trading strategy (), and a market price of risk
Chapter 3. Completeness and Relative Arbitrage 82
()
e
n
, such that Z

()V
p,
() is a martingale up to time
n
and V
p,
(
n
) =

D.
Assume now that the supremum in (3.8) is attained, say by ()
t
, but that
the supremum in (3.9) is not. Then, again by Corollaries 14 and 10 of Delbaen
and Schachermayer (1995c), there exists a claim

D

D with P(

D >

D) > 0 and
a trading strategy (), such that V
p,
(
n
) =

D. However, the supermartingale
property of Z

()V
p,
() leads directly to a contradiction.
The next lemma will be of use in Theorem 5, when we need to prove time
consistency of strategies. It generalizes the equality p
n
(

D) = p
e
n
(

D) of Lemma 5.
The measurability of the essential suprema in the following lemma is guaranteed as
in Lemma 7 below.
Lemma 6 (Suciency of local martingale measures). Fix n N. Assume again
that

D 0 is T(
n
)-measurable for some n N. Then, the equality
ess sup
()

E
_
Z

(
n
)
Z

(
n1
)

T(
n1
)
_
= ess sup
()
e
n
E
_
Z

(
n
)
Z

(
n1
)

T(
n1
)
_
(3.10)
holds.
Proof. As in the proof of Lemma 5, each ()
e
n
corresponds to some ()
t
.
So, we need only show that the left-hand side is less than or equal to the right-
hand side in (3.10). Towards this end, we x n N and introduce the process
W
Q
() = (W
Q
1
(), . . . , W
Q
K
())
T
with
W
Q
k
() := W
k
() +
_

0

m
k
(u)du
for all k = 1, . . . , K. Analogously to (3.5), we dene
Z
,Q
() := exp
_

_

0

T
(u)dW
Q
(u)
1
2
_

0
|(u)|
2
du
_
and observe
Z

() Z

m
()Z

m
,Q
() (3.11)
Chapter 3. Completeness and Relative Arbitrage 83
for all () . Fix any ()
t
and denote by
i

iN
a sequence of stopping
times dened as

i
:=
n
inf
_
t
n1

m
,Q
(t) iZ

m
,Q
(
n1
)
_
for all i N. The equality in (3.11) and Fatous lemma yield
E
_
Z

(
n
)
Z

(
n1
)

T(
n1
)
_
= E
_
Z

m
(
n
)
Z

m
(
n1
)
Z

m
,Q
(
n
)
Z

m
,Q
(
n1
)

T(
n1
)
_
= E
_
Z

m
(
n
)
Z

m
(
n1
)
lim
i
Z

m
,Q
(
i
)
Z

m
,Q
(
n1
)

T(
n1
)
_
liminf
i
E
_
Z

m
(
n
)
Z

m
(
n1
)
Z

m
,Q
(
i
)
Z

m
,Q
(
n1
)

T(
n1
)
_
= liminf
i
E
_
Z

(i)
(
n
)
Z

(i)
(
n1
)

T(
n1
)
_
ess sup

()
e
n
E
_
Z

(
n
)
Z

(
n1
)

T(
n1
)
_
,
since

(i)
:=
_

m
() +1

i
t
_
()
m
()
__

e
n
for all i N; this proves the statement.
We continue by introducing the sequence of random variables
D
n
:= ess sup
()

E
_
Z

(T)
Z

(
n
)
D

T(
n
)
_
=: ess sup
()

n
0 (3.12)
for all n N. If p = D
0
< , then D
n
< for all n N. We discuss in the next
lemma the measurability of each D
n
; in particular, we show that D
n
represents a
claim:
Lemma 7 (Measurability of D
n
). For any n N, the essential supremum D
n
of
(3.12) is T(
n
)-measurable.
Chapter 3. Completeness and Relative Arbitrage 84
Proof. Fix n N. Theorem A.3 of Karatzas and Shreve (1998) yields that D
n
exists
and is T(
n
)-measurable. This is due to the observation that for any
(1)
(),
(2)
()

t
, there exists
(3)
()
t
, which is dened by
(3)
(
n
)
(1)
(
n
) and

(3)
(t) :=
(1)
(t)1
D

(1)
n
D

(2)
n

+
(2)
(t)1
D

(1)
n
<D

(2)
n

(3.13)
for all t >
n
and therefore satises the fork property
D

(3)
n
= D

(1)
n
D

(2)
n
, (3.14)
where denotes the maximum.
The next lemma proves a dynamic programming principle (DPP). It is es-
sential for the results that follow below.
Lemma 8 (Multiplicative DPP). The sequence of random variables (D
n
)
nN
sat-
ises the equalities
D
n1
= ess sup
()
e
n
E
_
Z

(
n
)
Z

(
n1
)
D
n

T(
n1
)
_
if D
n
< for all n N.
Proof. Fix n N such that D
n
< . In conjunction with the fork property of
(3.13) and (3.14), Theorem A.3 of Karatzas and Shreve (1998) yields the existence
of a sequence of random variables (
(i)
)
iN
such that D

(i)
n
D
n
as i . Fix
> 0 and dene i

as
i

:= min
_
i N

(i)
n
D
n

_
.
Then, D

(i

)
n
is T(
n
)-measurable and we obtain
ess sup
()
e
n
E
_
Z

(
n
)
Z

(
n1
)
D
n

T(
n1
)
_
ess sup
()
e
n
E
_
Z

(
n
)
Z

(
n1
)
D

(i

)
n

T(
n1
)
_
+
D
n1
+ .
Chapter 3. Completeness and Relative Arbitrage 85
Since the choice of was arbitrary, this yields one inequality; the other direction
follows from
D
n1
ess sup
()

E
_
Z

(
n
)
Z

(
n1
)
ess sup

()

E
_
Z

(T)
Z

(
n
)
D

T(
n
)
_

T(
n1
)
_
= ess sup
()

E
_
Z

(
n
)
Z

(
n1
)
D
n

T(
n1
)
_
and Lemma 6.
Next, we set p
n
:= p
n
(D
n
) for all n N, as in (3.8). The following time
consistency follows from the same argument as the DPP of Lemma 8: The sequence
(p
n
)
nN
satises
p
n
= p (3.15)
for all n N, with p as in (3.7). We can now state and prove the main result of
this section:
Theorem 5 ((Super-)replicating strategy). There exists no trading strategy ()
for which V
c,
(T) D, for any c [0, p). If p < , then a trading strategy ()
exists such that V
p,
(T) D. If the supremum in (3.7) is attained, then one can
choose () so that V
p,
(T) = D, that is, the claim D can be exactly replicated by
dynamic hedging. If an ELMM exists, then the supremum in (3.7) can be replaced
by the supremum over all ()
t
for which Z

() is a martingale.
Proof. The rst part of the statement follows as in Lemma 5. Assume in the
following that p < , thus D
n
< for all n N. Now, we inductively construct
for each n N trading strategies
(n)
() that satisfy
V
p,
(n)
(
n1
) = V
p,
(n1)
(
n1
) (3.16)
and V
p,
(n)
(
n
) D
n
. According to Lemma 5 and due to (3.15), there exist a
contingent claim

D
1
D
1
, a trading strategy
(1)
(), and a market price of risk
Chapter 3. Completeness and Relative Arbitrage 86

(1)
()
e

1
, such that V
p,
(1)
(
1
) =

D
1
and Z

(1)
()V
p,
(1)
() is a martingale up
to time
1
. Assume that we have determined
(n1)
(),
(n1)
(), and

D
n1
:=
V
p,
(n1)
(
n1
) D
n1
, such that Z

(n1)
()V
p,
(n1)
() is a martingale up to time

n1
. We observe that p
n
(D
n
+

D
n1
D
n1
) = p
n
= p, since by Lemma 5
p p
e
n
(D
n
+

D
n1
D
n1
)
sup
()
e
n
E
_
Z

(
n1
)
_
ess sup

()
e
n
E
_
Z

(
n
)
Z

(
n1
)
D
n

T(
n1
)
_
+

D
n1
D
n1
__
sup
()

E
_
Z

(
n1
)

D
n1
_
= p
due to the DPP of Lemma 8. By Lemma 5 again, there exist a contingent claim

D
n
D
n
+

D
n1
D
n1
D
n
, a trading strategy
(n)
(), and a market price of
risk
(n)
()
e
n
such that V
p,
(n)
(
n
) =

D
n
and Z

(n)
()V
p,
(n)
() is a martingale
up to time
n
.
Now, the DPP of Lemma 8 yields
V
p,
(n)
(
n1
) =E
_
Z
(n)
(
n
)
Z
(n)
(
n1
)
V
p,
(n)
(
n
)

T(
n1
)
_
ess sup
()
e
n
E
_
Z

(
n
)
Z

(
n1
)
(D
n
+

D
n1
D
n1
)

T(
n1
)
_
ess sup
()
e
n
_
E
_
Z

(
n
)
Z

(
n1
)
D
n

T(
n1
)
_
+ E
_
Z

(
n
)
Z

(
n1
)

T(
n1
)
_
(

D
n1
D
n1
)
_
=D
n1
+

D
n1
D
n1
=

D
n1
,
and thus V
p,
(n)
(
n1
) V
p,
(n1)
(
n1
). Assume that the event V
p,
(n)
(
n1
) >
V
p,
(n1)
(
n1
) has positive probability. Since Z

(n1)
()V
p,
(n1)
() is a martingale,
this implies that the event V
p,
(n)
(
n1
) < V
p,
(n1)
(
n1
) should also have pos-
Chapter 3. Completeness and Relative Arbitrage 87
itive probability, leading to a contradiction. Thus, this inductive procedure yields
trading strategies
(n)
() that satisfy (3.16) and V
p,
(n)
(
n
) D
n
for all n N.
We dene a new trading strategy () as
(t) =
(1)
(0)1
0
(t) +

n=1

(n)
(t)1
t(
n1
,n]
for all t 0. We observe that
V
p,
(
n
) =

D
n
D
n
holds for all n N. We now x any such that
n()
= T for some n().
Then, we obtain
V
p,
(T)() = V
p,
(
n()
)() D
n()
() = D()
with equality if the supremum in (3.7) is attained, due to the observation that

D
n
= D for all n N in that case. Since for almost all such an n() exists,
() (super-)replicates D.
If an ELMM exists, we are in the context of the classical theory of Finan-
cial Mathematics and then it is sucient to take the supremum in (3.7) over all
ELMMs to obtain the minimal superreplicating price; see Delbaen and Schacher-
mayer (1995c).
The previous theorem proves, in particular, a conjecture in Chapter 2. There,
the Markovian case is discussed and it is demonstrated that the supremum in (3.7)
is always attained, as long as D is measurable with respect to T
S
(), the ltration
generated by the stock price processes S(). For path-independent European-style
claims, an explicit trading strategy for the exact replication is constructed, but the
question of whether path-dependent claims could be hedged is not resolved. For a
more precise statement of these results, see Theorem 6 below.
We wish to draw the readers attention to a few subtle points concerning
the previous theorem. First of all, even if the supremum in (3.7) is not attained,
Chapter 3. Completeness and Relative Arbitrage 88
there might nevertheless exist a trading strategy () that replicates D, that is, a
trading strategy such that V
p,
(T) = D. The claim D = S(2) in Example 8 below
illustrates this point. However, such a trading strategy () is not maximal in the
sense of Denition 8. Second, the replicating price p in (3.7) depends strongly on
the admissibility constraint V
p,
() 0, as the next remark discusses:
Remark 12 (Relevance of the admissibility constraint). We have observed in Re-
mark 11 that the precise choice of the admissibility constraint is not relevant for
determining the costs of replicating a nonnegative claim in markets without arbi-
trage. This is no longer true in markets that do not proscribe arbitrage. Indeed,
if we allow for strategies () whose associated wealth process is only required to
stay above a constant < 0, then the minimal nonnegative price p

to (super-
)replicate a claim D can be computed as
p

:= sup
()

E[Z

(T)(D + )] p.
In particular, it is possible that p

< 0. The fact that p

is the minimal (super-

)replicating price can be seen as in Lemma 5. The strategy () that (super-
)replicates D under these weaker admissibility condition is exactly the same strategy
that (super-)replicates D + in Theorem 5.
A further subtle point that we want to emphasize is that the trading strategy
() which replicates some T
S
()-measurable claim D in Theorem 5 for the price p is,
in general, progressively measurable only with respect to T(), but not necessarily
with respect to T
S
(). The next example illustrates this point. To determine
sucient conditions that imply the measurability of the replicating trading strategy
with respect to T
S
() is a future research project.
Example 6 (Measurability of trading strategies). We set d = K = 2, S
1
(0) =
S
2
(0) = 2,
1,2
()
2,1
() 0, and
i
() 1/S
i
() for i = 1, 2. Furthermore, we
Chapter 3. Completeness and Relative Arbitrage 89
dene I
1
:= 1
W
1
(1)0
, I
2
:= 1
W
1
(1)<0
and

1,1
(t) =
1
S
1
(t)
1
[1,)
(t)
_
1 I
1
1
t>
1

_
,

2,2
(t) =
1
S
2
(t)
1
[1,)
(t)
_
1 I
2
1
t>
2

_
,
for all t 0, where we set
i
:= inf t 0 [ S
i
(t) 1 for i = 1, 2. Thus, up
to time t = 1, the market does not move. Then, one of the stock price processes
has the dynamics of the reciprocal of a three-dimensional Bessel process, while the
other one has the same dynamics only until it hits 1. The sign of W
1
(1) decides
which of the two processes has which dynamics.
We observe that I
1
is not measurable with respect to the stock price ltration
T
S
() up to the stopping time
1

2
> 1. More precisely, for any stopping time
<
1

2
, any event A T
S
( ) is independent of the event W
1
(1) 0; thus
W
1
(1) 0 / T
S
( ).
If ()
t
denotes any market price of risk, then
1
(t) = 1/S
1
(t) for t 1
(t [1,
1
]) if I
1
= 0 (I
1
= 1) and
2
(t) = 1/S
2
(t) for t 1 (t [1,
2
]) if I
2
= 0
(I
2
= 1). We thus obtain from Itos formula
Z

(t) =
S
1
(0)
S
1
(t)
S
2
(0)
S
2
(t)
Z

(1)
2

i=1
_
1 + I
i
_
c
i
(, t
i
, t) 1
_
_
for all t 1 with
c
i
(, t
0
, t
1
) := exp
_

_
t
1
t
0

i
(u)dW
i
(u)
1
2
_
t
1
t
0

2
i
(u)du
_
for all i = 1, 2 and t
0
, t
1
0.
We now x T = 2 and D = 1 and obtain
p = sup
()

E[Z

(2)]
= E
_
S
1
(0)S
2
(0)
S
1
(2)S
2
(2)
_
Chapter 3. Completeness and Relative Arbitrage 90
=
1
2
_
E
_
S
2
(0)
S
2
(2)

W
1
(1) 0
_
+E
_
S
1
(0)
S
1
(2)

W
1
(1) < 0
__
= 2(2) 1,
where denotes the cumulative standard normal distribution function and the last
equality is derived from the expectation of the reciprocal of a three-dimensional pro-
cess, starting at 2; see, for example, (2.46). Furthermore, the supremum is attained,
for example by (). Thus, there exists a trading strategy, () = (
1
(),
2
())
T
,
which exactly replicates D = 1 and which can be explicitly represented as

i
(t) =
2

2 t
(1 I
i
)1
[1,2]
(t)
_
S
i
(t)

2 t
_
for i = 1, 2, where denotes the standard normal density, and the corresponding
(unique) wealth process
V
p,
(t) = 1
[0,1)
(t)p + 1
[1,2]
(t)
2

i=1
(1 I
i
)
_
2
_
S
i
(t)

2 t
_
1
_
for all t [0, 2]; compare (2.49) and (2.46).
We observe that V
p,
( ) depends for all stopping times > 1 on I
1
, thus is
not measurable with respect to the stock price ltration T
S
( ) for all stopping times
(1,
1

2
). Therefore, there exists no trading strategy () that is measurable
with respect to the stock price ltration T
S
() and that replicates D = 1 for initial
costs p.
The last example can easily be adapted to an example for an arbitrage-free
market with a claim that is T
S
(T)-measurable for some T > 1 and that can be
replicated by a maximal T()-measurable trading strategy, but not by a maximal
T
S
()-measurable trading strategy. Towards this end, we introduce a new market
with two stocks

S
i
() := 1/S
i
() for i = 1, 2; both of them are now local martingales,
one of them stopped at
i
. Now, we consider the claim D = S
1
(2)I
2
+ S
2
(2)I
1
. In
order to ensure the measurability of D with respect to T
S
(2), we replace
i
by

i
1.5. Then, we proceed with the argument of the previous example.
Chapter 3. Completeness and Relative Arbitrage 91
3.4 Completeness and Second Fundamental The-
orem of Asset Pricing
In this section, we extend the Second Fundamental Theorem of Asset Pricing to
include markets that do not proscribe arbitrage opportunities. Furthermore, we
discuss two notions of completeness. To start, we formally introduce the concept
of a complete market in the next denition:
Denition 9 (Complete market). A market is called complete if for all T > 0 and
all bounded T(T)-measurable random variables D 0 there exist p > 0 and a
maximal trading strategy () that replicates D; that is, there exists a maximal
trading strategy () such that V
p,
(T) = D. Alternatively, if there exist some
T > 0 and some T(T)-measurable random variable D 0 for which no maximal
replication exists, then the market is called incomplete on [0, T].
In particular, by the martingale representation theorem, a market is complete
if d = K and (t) is invertible for all t > 0; see Section 10.1 of Fernholz and Karatzas
(2009). As previously noted, the notion of completeness is often too strong and we
therefore introduce the notion of quasi-completeness, a slight generalization:
Denition 10 (Quasi-complete market). We call a market quasi-complete if for every
T > 0 and every bounded T
S
(T)-measurable random variable D 0, there exists
a maximal trading strategy () that replicates D.
It follows directly from this denition that any complete market is necessarily
quasi-complete but not vice versa. We call a function g : R
+
C(R
+
, R
d
+
) R
non-anticipative functional if g(t, x()) = g(t, x( t)) for all t R
+
and all x()
C(R
+
, R
d
+
); that is g is non-anticipative if g(t, x()) depends on the path of x only
up to time t. We have the following result, which generalizes Pag`es (1987), Due
(1988)
2
, and Proposition 1 in Chapter 2:
2
We thank Martin Schweizer for pointing us to this reference.
Chapter 3. Completeness and Relative Arbitrage 92
Theorem 6 (Sucient conditions for quasi-completeness). If S() of (3.1) can be
represented as the unique solution of
dS
i
(t) =S
i
(t)
K

k=1

i,k
(t, S())
_

k
(t, S())dt + dW
k
(t)
_
,
where
i,k
and

k
are non-anticipative functionals for all i = 1, , d and k =
1, , K, then the market is quasi-complete. Furthermore, for any T > 0 and
T
S
(T)-measurable random variable D 0,
m
() maximizes the expression in (3.7).
Proof. The proof of Proposition 1 in Chapter 2 carries through with only minor
modications.
We emphasize that we have not assumed that the volatility matrix ()
has full rank in the previous theorem. As demonstrated in the next example,
an incomplete market is generally not quasi-complete if
m
() is not progressively
measurable with respect to T
S
():
Example 7 (Relevance of drift for quasi-completeness). We set K = 1, d = 1,
S(0) = 1, (t) = 0,
m
(t) = 0 for all t [0, 1], and (t) = 1/S(t),

m
(t) =
1
S(t)
t1
[W(1)[
for all t > 1. This market is a slight extension of Example 1 in Chapter 2. We con-
sider D = 1, which is T
S
(2)-measurable. For any ()
t
, where
t
is introduced
in (3.3), we obtain
E
_
Z

(2)
Z

(1)
D

T(1)
_
=
_
(1 c) exp(2c)(1 (1 + c))
_

c=
1
|W(1)|
,
where denotes the standard normal cumulative distribution function; see (2.45).
The last function is decreasing in c 0. We thus obtain
sup
()

E[Z

(2)D] = 2(1) 1;
however, the supremum is not attained. Thus, the model is not quasi-complete.
Chapter 3. Completeness and Relative Arbitrage 93
The stock price process in the previous example is sometimes called a bub-
ble, since under the ELMM its price tends to decrease in expectation due to its
strict local martingality. We refer the reader to Jarrow et al. (2010) for a deni-
tion, further references, and a thorough discussion regarding bubbles in incomplete
markets. The next lemma prepares the proof of the Second Fundamental Theorem
of Asset Pricing:
Lemma 9 (Rank of volatility matrix in complete market). If a market is complete,
then rank((t)) = K Lebesgue-almost everywhere. In particular, d K.
Proof. Fix some T > 0 and assume a complete market. We now show rank((t)) =
K Lebesgue-almost everywhere on [0, T]. We introduce a new, ctional market
with d + 1 stocks
_
S
1
()
V
p,
()
, . . . ,
S
d
()
V
p,
()
,
1
V
p,
()
_
,
where p is dened in (3.7) with D = 1 and () is the corresponding maximal trading
strategy, as for example determined in the proof of Theorem 5, such that V
p,
(T)
1. Then, Theorems 11 and 4 of Delbaen and Schachermayer (1995c) yield together
that NFLVR holds for the new market. If we denote the volatility matrix of the
new market by (), then a simple computation shows that rank(()) rank( ()),
and hence, that the new market is also complete. Although we have not assumed
d + 1 K, the argument of Theorem 1.6.6 in Karatzas and Shreve (1998) works
and proves the result.
We can now formulate and prove the Generalized Second Fundamental The-
orem of Asset Pricing:
Theorem 7 (Generalized Second Fundamental Theorem of Asset Pricing). A mar-
ket is complete if and only if any process ()
t
satises (t) =
m
(t) Lebesgue-
almost everywhere.
Chapter 3. Completeness and Relative Arbitrage 94
Proof. If the market is complete, then we have rank((t)) = K Lebesgue-almost
everywhere by Lemma 9. This is equivalent to the Lebesgue-almost everywhere
uniqueness of () in
t
. For the reverse direction, we observe that the supremum
in (3.7) is always taken over a singleton, and is thus trivially attained.
We remark that any complete market implies Z

() Z

m
() for all ()
t
.
Thus, in the no-arbitrage framework, this directly translates into the uniqueness
of the ELMM. However, it is important to note that the question regarding the
completeness of the market can be addressed separately from the question regarding
the existence of arbitrage; see also Jarrow et al. (1999).
3.5 Relative arbitrage and strong relative arbi-
trage
In this section, we analyze the interplay of relative arbitrage and strong relative
arbitrage opportunities. The concept of relative arbitrage traces back to Merton
(1973), where the term dominant portfolio is used. He writes:
Security (portfolio) A is dominant over security (portfolio) B, if on
some known date on the future, the return on A will exceed the return
on B for some possible states of the world, and will be at least as large
as on B, in all possible states of the world.
We also refer to Jarrow et al. (2007; 2010) for a thorough discussion of Mertons
no-dominance principle in connection with the existence of bubbles. Delbaen and
Schachermayer (1994; 1995c) coined the term maximal element for a terminal
wealth V
p,
(T) that cannot be dominated by another terminal wealth V
p,
(T). In
the following, we use the terminology of Stochastic Portfolio Theory; see Fernholz
Chapter 3. Completeness and Relative Arbitrage 95
and Karatzas (2009). This line of research does not focus on nding the right con-
ditions to exclude arbitrage opportunities, but instead studies these opportunities;
see, for example, Fernholz and Karatzas (2010), where relative arbitrage with re-
spect to the market portfolio is studied. We now provide the precise denition on
which we shall rely:
Denition 11 (Relative and classical arbitrage). We say that there exists relative
arbitrage with respect to a trading strategy () over the time horizon [0, T] if
there exists a trading strategy () such that P(V
p,
(T) V
p,
(T)) = 1 and
P(V
p,
(T) > V
p,
(T)) > 0. We say that () is a strong relative arbitrage if
P(V
p,
(T) > V
p,
(T)) = 1. If () 0, which corresponds to holding the risk-free
money market, then we sometimes substitute the word relative by classical.
Obviously, the existence of strong relative arbitrage necessarily implies that
of relative arbitrage. However, the converse is less obvious. Using the insights
developed in the previous sections, we shall discuss conditions under which the
existence of relative arbitrage implies that of strong relative arbitrage in Theorem 8.
We start by giving an example showing that this implication does not always hold:
Example 8 (Relative arbitrage without strong relative arbitrage). Let K = 1, d = 1,
S(0) = 2, () 0 and (t) = 0 for t [0, 1] [2, ). Set
(t) =
1
S(t)
1
W(1)0>t
1

2 t
for t (1, 2), where
:= inf
_
t 1 :
_
t
1
1

2 s
dW(s) = 1
_
. (3.17)
Then we have < 2, which yields S(2) = 2 on the event W(1) < 0, S(2) = 1 on
the event W(1) 0, and S() being a strictly positive, local martingale.
We consider the buy-and-hold trading strategy () 1, such that D :=
V
2,
(2) = S(2). Since NFLVR is satised here, it is sucient to take the supremum
Chapter 3. Completeness and Relative Arbitrage 96
in (3.7) over

:= ()
t
: E[Z

(2)] = 1 ,
to wit, the subset of
t
that generates the ELMMs. For any ()

we have
Q

(W(1) 0) > 0, where Q

is dened by dQ

/dP = Z

(2), such that S() is a

strict local martingale under any ELMM. However,
p = sup
()

E[Z

(2)S(2)] = 2 inf
()

(W(1) 0) = 2.
That is, the cheapest trading strategy to superreplicate one share S(2) at time
T = 2 costs p = 2. Fix any trading strategy (). Then, on the event W(1) < 0
we always have V
2,
(2) = 2 = S(2). This shows that no strong relative arbitrage
exists with respect to () over the time horizon [0, 2].
However, relative arbitrage exists. The trading strategy () 0 yields
V
2,
(2) = 2. Thus, P(V
2,
(2) > S(2)) = P(W(1) 0) = 1/2 > 0. To conclude,
although the cheapest superreplicating price of a given terminal wealth V
p,
(T)
might be p, the trading strategy () might nevertheless be dominated in the sense
of Merton (1973).
Delbaen and Schachermayer (1998) discuss a model in which the stock price
process is a strict local martingale under one measure, but actually a true martingale
under an equivalent measure. The previous example exhibits a stock price process
such that Z

()S() is a strict local martingale for all ()

t
, but where the
cheapest price to replicate the stock price is the current stock price itself. This
example can be easily modied to obtain a market that allows for arbitrage, but
where the cheapest superreplicating price, to pay at time 0, for $1 at time T > 0 is
again $1:
Example 9 (Free lunch with vanishing risk but without strong classical arbitrage).
We again set K = 1 and d = 1. We now consider the stock price process
Chapter 3. Completeness and Relative Arbitrage 97

S() := 1/S() with S() dened as in Example 8, which corresponds to a change of

numeraire. Itos formula yields the dynamics
d

S(t) =

S(t)(t)
_
(t)dt dW
t
_
with () as in Example 8. Corollary 15 of Delbaen and Schachermayer (1995c) di-
rectly yields that this market does not allow for an ELMM. Any stochastic discount
factor

Z

() in the new model can be written as

Z

() = Z

()S()/S(0), where Z

()
denotes a stochastic discount factor in the original model of Example 8.
We now set T = 2 and consider the claim D = 1, which corresponds to
holding exactly $1 at time 2. and obtain sup
()
E[

Z

(2)D] = 1. Thus, despite
the existence of arbitrage opportunities, the cheapest price to hold $1 is again $1 and
no strong classical arbitrage exists, due to reasoning similar to that in Example 8.
However, starting with $1, one can achieve a terminal wealth that is larger than $1
with positive probability by following the trading strategy () 1.
We can now state precise conditions for the existence of relative arbitrage
and strong relative arbitrage opportunities:
Theorem 8 (Conditions for relative arbitrage and strong relative arbitrage). Fix
T > 0 and a trading strategy () admissible for some initial capital p > 0.
(a) There exists a strong relative arbitrage opportunity with respect to () over
the time horizon [0, T] if
p := sup
()

E[Z

(T)V
p,
(T)] < p. (3.18)
The converse holds if a trading strategy () and a constant > 1 exist such
that V
p,
(T) V
p,
(T) ,= 0.
(b) There exists a relative arbitrage opportunity with respect to () over the time
horizon [0, T], if and only if
E[Z

(T)V
p,
(T)] < p (3.19)
Chapter 3. Completeness and Relative Arbitrage 98
for all ()
t
.
(c) In particular, the existence of relative arbitrage implies that of strong relative
arbitrage over the time horizon [0, T] if the market is quasi-complete and
V
p,
(T) is T
S
(T)-measurable.
Proof. We prove (a), (b), and (c) separately:
(a) Assume (3.18) holds. According to Theorem 5, a trading strategy () exists
such that
V
p,
(T) V
p,
(T) + p p > V
p,
(T),
which shows the existence of strong relative arbitrage.
We observe that for any ()
t
and for any trading strategy (), ad-
missible with respect to the initial capital p, the process Z

()V
p,
() is a
supermartingale. Thus, if a strong relative arbitrage () and some > 1 as
in the statement of the theorem exist, then
sup
()

E[Z

(T)V
p,
(T)] sup
()

E[Z

(T)V
p,
(T)]
1

< p,
which implies (3.18).
(b) In a similar vein, assume that a relative arbitrage () with respect to ()
exists. Then,
E[Z

(T)V
p,
(T)] < E[Z

(T)V
p,
(T)] p.
for all ()
t
, yielding (3.19). For the reverse direction, let us introduce,
as in Lemma 9, a ctional market with d + 1 stocks
_
S
1
()
V
p,
()
, . . . ,
S
d
()
V
p,
()
,
1
V
p,
()
_
.
Chapter 3. Completeness and Relative Arbitrage 99
Then, (3.19) yields that no ELMM exists for the new market. Thus, the
ctional market allows for classical arbitrage and Theorems 4 and 11 of Del-
baen and Schachermayer (1995c) yield the existence of a relative arbitrage
opportunity.
(c) If the market is quasi-complete, then the supremum in (3.18) is always a max-
imum, and consequently, relative arbitrage implies strong relative arbitrage
in the case of quasi-completeness.
The next example illustrates the fact that p = p in (3.18) does not necessarily
exclude a strong relative arbitrage opportunity:
Example 10 (Diminishing strong relative arbitrage). We use the same setting as in
Example 8 with () modied as follows:
(t) =
1
S(t)

i=1
1
i
1
[W(1)[[i1,i)>t
1

2 t
for t (1, 2), where the stopping is dened as in (3.17). This yields S(2) = 11/i
on the event [W(1)[ [i 1, i) for all i N. We obtain
p = sup
()

E[Z

(2)S(2)] = 2 inf
()

i=1
1
i
Q

([W(1)[ [i 1, i))
_
= 2,
where

and Q

are dened in Example 8. However, the trading strategy () 0

yields V
2,
(2) = 2 > S(2) and is thus a strong relative arbitrage. This example
shows that (3.18) is sucient, but not necessary for the existence of strong relative
arbitrage.
It is clear that we need to assume in part (c) of Theorem 8 that V
p,
(T) be
T
S
(T)-measurable. To see this, one could, for example, construct a wealth process
V
p,
() in a quasi-complete model that has exactly the same dynamics as S() in
Example 8 and allows for relative arbitrage but not strong relative arbitrage.
Chapter 3. Completeness and Relative Arbitrage 100
3.6 Conclusion
In this chapter, we have illustrated that in general the concepts of arbitrage and
completeness can be considered separately from each other. More precisely, we
have proven a version of the Second Fundamental Theorem of Asset Pricing for
markets that do not proscribe arbitrage. We have also provided necessary and
sucient conditions for claims in incomplete markets to be exactly replicable. We
have introduced the concept of quasi-complete markets to generalize the idea of
complete markets and have further shown that relative arbitrage implies strong
relative arbitrage in quasi-complete markets.
We have assumed that the agent can trade dynamically and without any
constraints in the market. It is an open question for markets that allow for the
presence of arbitrage opportunities, in which manner trading constraints interfere
with the replication of claims. In particular, it is not clear under which trading
constraints certain arbitrage opportunities disappear. This is subject to future
research. A good starting point is the theory for markets that satisfy NFLVR, as
developed in Cvitanic and Karatzas (1993) and Follmer and Kramkov (1997).
BIBLIOGRAPHY 101
Bibliography
Ansel, J.-P. and Stricker, C. (1993). Couverture des actifs contingents. Annales de
lInstitut Henri Poincare, 30(2):303315.
Bayraktar, E., Huang, Y.-J., and Song, Q. (2010a). Outperforming the market
portfolio with a given probability.
Bayraktar, E., Kardaras, C., and Xing, H. (2010b). Strict local martingale deators
and pricing American call-type options. Finance and Stochastics, forthcoming.
Cox, A. and Hobson, D. (2005). Local martingales, bubbles and option prices.
Finance and Stochastics, 9(4):477492.
Cvitanic, J. and Karatzas, I. (1993). Hedging contingent claims with constrained
portfolios. Annals of Applied Probability, 3(3):652681.
Delbaen, F. and Schachermayer, W. (1994). A general version of the Fundamental
Theorem of Asset Pricing. Mathematische Annalen, 300(3):463520.
Delbaen, F. and Schachermayer, W. (1995a). Arbitrage possibilities in Bessel pro-
cesses and their relations to local martingales. Probability Theory and Related
Fields, 102(3):357366.
Delbaen, F. and Schachermayer, W. (1995b). The existence of absolutely continuous
local martingale measures. Annals of Applied Probability, 5(4):926945.
Delbaen, F. and Schachermayer, W. (1995c). The no-arbitrage property under a
change of numeraire. Stochastics and Stochastic Reports, 53:213226.
Delbaen, F. and Schachermayer, W. (1998). A simple counterexample to several
problems in the theory of asset pricing. Mathematical Finance, 8(1):111.
Delbaen, F. and Schachermayer, W. (2006). The Mathematics of Arbitrage.
Springer.
Delbaen, F. and Shirakawa, H. (2002). No arbitrage condition for positive diusion
price processes. Asia-Pacic Financial Markets, 9:159168.
BIBLIOGRAPHY 102
Due, D. (1988). An extension of the Black-Scholes model of security valuation.
Journal of Economic Theory, 46:194204.
Ekstrom, E., Lotstedt, P., Sydow, L. V., and Tysk, J. (2009). Numerical option
pricing in the presence of bubbles.
Ekstrom, E. and Tysk, J. (2009). Bubbles, convexity and the Black-Scholes equa-
tion. Annals of Applied Probability, 19(4):13691384.
El Karoui, N. and Quenez, M.-C. (1995). Dynamic programming and pricing of
contingent claims in an incomplete market. SIAM Journal on Control and Opti-
mization, 33(1):2966.
Emanuel, D. C. and Macbeth, J. D. (1982). Further results on the constant elasticity
of variance call option pricing model. Journal of Financial and Quantitative
Analysis, 17(4):533554.
Ethier, S. N. and Kurtz, T. G. (1986). Markov Processes: Characterization and
Convergence. John Wiley & Sons.
Evans, L. C. (1998). Partial Dierential Equations. American Mathematical Soci-
ety.
Fernholz, D. and Karatzas, I. (2010). On optimal arbitrage. Annals of Applied
Probability, 20(4):11791204.
Fernholz, E. R. (1999). Portfolio generating functions. In Avellaneda, M., editor,
Quantitative Analysis in Financial Markets. World Scientic.
Fernholz, E. R. (2002). Stochastic Portfolio Theory. Springer.
Fernholz, E. R. and Karatzas, I. (2009). Stochastic Portfolio Theory: a survey. In
Bensoussan, A., editor, Handbook of Numerical Analysis, volume Mathematical
Modeling and Numerical Methods in Finance. Elsevier.
Fernholz, E. R., Karatzas, I., and Kardaras, C. (2005). Diversity and relative
arbitrage in equity markets. Finance and Stochastics, 9(1):127.
Follmer, H. (1972). The exit measure of a supermartingale. Zeitschrift f ur
Wahrscheinlichkeitstheorie und Verwandte Gebiete, 21:154166.
Follmer, H. (1973). On the representation of semimartingales. Annals of Probability,
1(4):580589.
BIBLIOGRAPHY 103
F ollmer, H. and Imkeller, P. (1993). Anticipation cancelled by a Girsanov trans-
formation: a paradox on Wiener space. Annales de lInstitut Henri Poincare,
29.
F ollmer, H. and Kabanov, Y. (1998). Optional decomposition and Lagrange mul-
tipliers. Finance and Stochastics, 2(1):6981.
Follmer, H. and Kramkov, D. (1997). Optional decompositions under constraints.
Probability Theory and Related Fields, 109:125.
Follmer, H. and Leukert, P. (1999). Quantile hedging. Finance and Stochastics,
3(3):251273.
Friedman, A. (1976). Stochastic Dierential Equations and Applications. Vols 1
and 2. Academic Press.
Geman, H., Karoui, N. E., and Rochet, J. (1995). Changes of numeraire, changes
of probability measure and option pricing. Journal of Applied Probability, 32.
Guasoni, P., Rsonyi, M., and Schachermayer, W. (2010). The Fundamental The-
orem of Asset Pricing for continuous processes under small transaction costs.
Annals of Finance, 6(2):157191.
Harrison, J. M. and Kreps, D. M. (1979). Martingales and arbitrage in multiperiod
securities markets. Journal of Economic Theory, 20(3):381408.
Harrison, J. M. and Pliska, S. (1981). Martingales and stochastic integrals in
the theory of continuous trading. Stochastic Processes and their Applications,
11(3):215260.
Heath, D. and Platen, E. (2002a). Consistent pricing and hedging for a modied
constant elasticity of variance model. Quantitative Finance, 2(6):459467.
Heath, D. and Platen, E. (2002b). Perfect hedging of index derivatives under a
minimal market model. International Journal of Theoretical and Applied Finance,
5(7):757774.
Heath, D. and Schweizer, M. (2000). Martingales versus PDEs in nance: an
equivalence result with examples. Journal of Applied Probability, 37:947 957.
Heston, S., Loewenstein, M., and Willard, G. (2007). Options and bubbles. Review
of Financial Studies, 20.
Hugonnier, J. (2010). Rational asset pricing bubbles and portfolio constraints.
BIBLIOGRAPHY 104
Hulley, H. and Platen, E. (2009). A visual criterion for identifying ito diusions as
martingales or strict local martingales.
Ingersoll, J. E. (1987). Theory of Financial Decision Making. Rowman & Littleeld
Publishers.
Jacka, S. (1992). A martingale representation result and an application to incom-
plete nancial markets. Mathematical Finance, 2:239250.
Jacod, J. and Shiryaev, A. N. (2003). Limit Theorems for Stochastic Processes.
Springer, 2nd edition.
Janson, S. and Tysk, J. (2006). Feynman-Kac formulas for Black-Scholes-type
operators. Bulletin of the London Mathematical Society, 38(2):269282.
Jarrow, R., Jin, X., and Madan, D. (1999). The Second Fundamental Theorem of
Asset Pricing. Mathematical Finance, 9(3):255273.
Jarrow, R., Protter, P., and Shimbo, K. (2007). Asset price bubbles in complete
markets. In Fu, M. C., Jarrow, R. A., Yen, J.-Y. J., and Elliott, R. J., editors,
Advances in Mathematical Finance, volume in honor of Dilip Madan, pages 97
121. Birkhauser.
Jarrow, R., Protter, P., and Shimbo, K. (2010). Asset price bubbles in incomplete
markets. Mathematical Finance, 20(2):145185.
Kabanov, Y. M. (1997). On the FTAP of Kreps-Delbaen-Schachermayer. Statistics
and Control of Stochastic Processes (Moscow, 1995/1996), pages 191203.
Karatzas, I. and Kardaras, C. (2007). The numeraire portfolio in semimartingale
nancial models. Finance and Stochastics, 11(4):447493.
Karatzas, I., Lehoczky, J., and Shreve, S. E. (1991a). Equilibrium models with
singular asset prices. Mathematical Finance, 1(3):1129.
Karatzas, I., Lehoczky, J., Shreve, S. E., and Xu, G. (1991b). Martingale and
duality methods for utility maximization in an incomplete market. SIAM Journal
on Control and Optimization, 29.
Karatzas, I. and Ocone, D. (1991). A generalized Clark representation formula,
with application to optimal portfolios. Stochastics and Stochastics Reports, 37.
Karatzas, I. and Shreve, S. E. (1991). Brownian Motion and Stochastic Calculus.
Springer, 2nd edition.
Karatzas, I. and Shreve, S. E. (1998). Methods of Mathematical Finance. Springer.
BIBLIOGRAPHY 105
Kardaras, C. (2008). Balance, growth and diversity of nancial markets. Annals of
Finance, 4(3):369397.
Kardaras, C. (2010). Finitely additive probabilities and the Fundamental Theo-
rem of Asset Pricing. In Chiarella, C. and Novikov, A., editors, Contemporary
Mathematical Finance. Springer.
Kardaras, C. and Platen, E. (2009). On the semimartingale property of discounted
asset-price processes.
Knerr, B. F. (1980). Parabolic interior Schauder estimates by the maximum prin-
ciple. Archive for Rational Mechanics and Analysis, 75(1):5185.
Kramkov, D. (1996). Optional decomposition of supermartingales and hedging
contingent claims in incomplete security markets. Probability Theory and Related
Fields, 105:459479.
Krylov, N. V. (1973). On the selection of a Markov process from a system of
processes and the construction of quasi-diusion processes. Mathematics of the
USSR-Izvestiya, 7:691709.
Kunita, H. (1984). Stochastic dierential equations and stochastic ows of dieo-
morphisms. In Lecture Notes in Mathematics, volume

Ecole d

Ete de Probabilites
de Saint-Flour XII-1982, pages 143303. Springer.
Levental, S. and Skorohod, A. (1995). A necessary and sucient condition for ab-
sence of arbitrage with tame portfolios. Annals of Applied Probability, 5(4):906
925.
Loewenstein, M. and Willard, G. A. (2000a). Local martingales, arbitrage, and
viability. Free snacks and cheap thrills. Economic Theory, 16(1):135161.
Loewenstein, M. and Willard, G. A. (2000b). Rational equilibrium asset-pricing
bubbles in continuous trading models. Journal of Economic Theory, 91(1):17
58.
Lyaso, A. (2010). The FTAP in the special case of Ito process nancial markets.
Madan, D. and Yor, M. (2006). Itos integrated formula for strict local martingales.
In Seminaire de Probabilites, XXXIX, pages 157170. Springer.
Merton, R. C. (1973). Theory of rational option pricing. Bell Journal of Economics,
4(1):141183.
Meyer, P. (1972). La mesure de H. Follmer en theorie de surmartingales. In
Seminaire de Probabilites, VI. Springer.
BIBLIOGRAPHY 106
Mijatovic, A. and Urusov, M. (2009). On the martingale property of certain local
martingales.
Novikov, A. (1972). On an identity for stochastic integrals. Theory of Probability
and its Applications, 17(4):717720.
Pag`es, H. (1987). Optimal consumption and portfolio policies when markets are
incomplete.
Pal, S. and Protter, P. E. (2010). Analysis of continuous strict local martingales
via h-transforms. Stochastic Processes and their Applications, 120(8):14241443.
Parthasarathy, K. (1967). Probability Measures on Metric Spaces. Academic Press.
Penrose, R. (1955). A generalized inverse for matrices. Proceedings of the Cambridge
Philosophical Society, 51:406413.
Platen, E. (2002). Arbitrage in continuous complete markets. Advances in Applied
Probability, 34(3):540558.
Platen, E. (2006). A benchmark approach to nance. Mathematical Finance,
16(1):131151.
Platen, E. (2008). The law of minimum price.
Platen, E. and Heath, D. (2006). A Benchmark Approach to Quantitative Finance.
Springer.
Platen, E. and Hulley, H. (2008). Hedging for the long run.
Protter, P. E. (2003). Stochastic Integration and Dierential Equations. Springer,
2nd edition.
Revuz, D. and Yor, M. (1999). Continuous Martingales and Brownian Motion.
Springer, 3rd edition.
Ruf, J. (2011+). Hedging under arbitrage. Mathematical Finance, forthcoming.
Schweizer, M. (1992). Martingale densities for general asset prices. Journal of
Mathematical Economics, 21:363378.
Shreve, S. E. (2004). Stochastic Calculus for Finance II. Continuous-Time Models.
Springer.
Sin, C. (1998). Complications with stochastic volatility models. Advances in Applied
Probability, 30(1):256268.
BIBLIOGRAPHY 107
Strasser, E. (2003). Necessary and sucient conditions for the supermartingale
property of a stochastic integral with respect to a local martingale. In Seminaire
de Probabilites, XXXVII, pages 385393. Springer.
Stroock, D. W. and Varadhan, S. R. S. (2006). Multidimensional Diusion Pro-
cesses. Springer, Berlin. Reprint of the 1997 edition.
Wong, B. and Heyde, C. C. (2004). On the martingale property of stochastic
exponentials. Journal of Applied Probability, 41(3):654664.
Yan, J.-A. (1998). A new look at the Fundamental Theorem of Asset Pricing.
Journal of the Korean Mathematical Society, 35(3):659673.