Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
THE RISK-SPREAD OPTION IN A POTENTIAL THEORETIC FRAMEWORK
By
MICHAEL C. SWEARINGEN
A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY
UNIVERSITY OF FLORIDA
2000
Copyright 2000
by
Michael C. Swearingen
To My Late Father Thank You for a Lifetime of Encouragement
Wish You Were Here
iv
PREFACE
Participants in fixed-income markets trade in various securities whose value
ultimately depends upon a particular rate of interest at which a financial entity is willing
to issue debt. The financial entities involved may be governments or corporations, which
determines whether the interest rate is riskless or risky, respectively. The risk involved in
this context is the credit risk associated with the possibility of default on a corporate
bond. Since there is no credit risk with government bonds, it is reasonable to assume that
the risky interest rate should always be greater than the riskless interest rate. But, what
happens when there is inflation? Does the difference between these interest rates, or risk
spread, remain the same as the government rate rises? This raises the issue of hedging
against the inflation risk associated with corporate bonds.
The main result of this dissertation is the development and pricing of an originally
designed interest rate derivative which shall be known as the risk-spread option. This
option may be used by investors to hedge away the risk associated with the difference
between the government riskless interest rate and that of a corporate bond. The potential
theoretic approach to this pricing problem is general enough to generate various
stochastic models of both the riskless and risky interest rates. Moreover, it provides a
model that is analogous to physical systems that employ potential theory; thus, the
physics of fixed-income finance is illuminated revealing a mathematical structure behind
economic intuitions.
v
The first section of Chapter 1 establishes the basic definitions and assumptions in
fixed-income finance and arbitrage pricing theory that will be used throughout this work.
Sections 2 and 3 lay down a general potential theoretic framework in which to evaluate
interest rate derivatives. As an example of the procedure developed in these sections,
Section 4 uses an Ornstein-Uhlenbeck process to derive models for the risky and riskless
interest rates as well as bond prices. In Section 5, the risk-spread option is introduced by
means of a discrete example.
Chapter 2 contains the main theoretical work necessary to represent the price of the
risk-spread option as the solution to a Cauchy problem. In Chapter 3, the Fourier and
Laplace transforms are used to represent the solution in a more tractable form. Also, it is
shown how to hedge the risk-spread option using a portfolio of riskless and risky bonds.
In the final section of Chapter 3, the graphs of the riskless and risky yield curves are
displayed for various parameters. A summary of the results together with some remarks
on advantages, disadvantages, and proposed future improvements is found in the
concluding chapter.
vi
TABLE OF CONTENTS
page
PREFACE .......................................................................................................................... iv
ABSTRACT.....................................................................................................................viii
CHAPTERS
1 A POTENTIAL THEORETIC FRAMEWORK FOR FIXED-INCOME MARKETS....1
1.1. The Essentials of Mathema tical Finance ................................................................ 1 1.1.1 The Fundamentals of Fixed-Income Finance................................................ 2
1.1.2 A Review of Arbitrage Pricing Theory......................................................... 6 1.2 Potential Approach I: Riskless Bonds and the Martingale Measure....................... 8 1.3 Potential Approach II: Risky Bonds and the Forward Martingale Measure......... 12
1.3.1 Risky Bonds ................................................................................................ 13 1.3.2 The Forward Martingale Measure............................................................... 21
1.4 A Simple Example of the Potential Theoretic Approach...................................... 26 1.5 The Risk-Spread Option........................................................................................ 28
2 A CAUCHY PROBLEM FOR THE RISK-SPREAD OPTION ....................................31
2.1 Derivation of the Cauchy Problem........................................................................ 31 2.2 The Potential Theoretic Parametrix Method......................................................... 37
2.2.1 The Gaussian Semigroup............................................................................ 37 2.2.2 The Fundamental Solution ......................................................................... 39
2.3 Preliminary Technical Results .............................................................................. 43 2.3.1 Differentiability of the Gaussian Semigroup.............................................. 43 2.3.2 Basic Potential Theory ............................................................................... 48
2.4 The Derivatives of the Gaussian Potential............................................................ 54 2.5 A Series Representation of the Fundamental Solution ......................................... 61
2.5.1 Convergence and Continuity...................................................................... 62 2.5.2 Hölder Continuity....................................................................................... 65
2.6 The Solution to the Cauchy Problem.................................................................... 71 3 NUMERICAL RESULTS AND APPLICATIONS .......................................................73
3.1 The Fourier and Laplace Transforms .................................................................... 73 3.2 Delta Hedging with the Risk-Spread Option ........................................................ 79 3.3 Numerical Properties of the Yield Curve.............................................................. 82
vii
4 SUMMARY AND CONCLUSIONS ............................................................................ 91 4.1 Summary of Results .............................................................................................. 91 4.2 Future Projects and Model Extensions ................................................................. 92
REFERENCES ..................................................................................................................94
BIOGRAPHICAL SKETCH .............................................................................................96
viii
Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy
THE RISK-SPREAD OPTION IN A POTENTIAL THEORETIC FRAMEWORK
By
Michael C. Swearingen
August 2000
Chairman: Joseph Glover Major Department: Mathematics
A fixed-income economy, which includes defaultable securities, is developed through
a potential theoretic approach to modeling the spot rate of interest. Under the assumption
of an arbitrage free market, the riskless and risky state-price densities are used as inputs
to generate the respective spot rates in a Markovian setting. The riskless state-price
density is simply the discounted conditional expectation of the derivative of the
martingale measure Q with respect to the reference probability P associated with the
underlying Markov process tX . The risky state-price density is an original modification
of its riskless counterpart. If the time to default is modeled as the first jump in a
generalized Poisson process with intensity ( )t tXλ = λ , then the risky state-price density
is defined as the discounted conditional expectation of the derivative of the forward
martingale measure F with respect to P. However, the discounting is done with respect to
the default intensity λ rather than the riskless spot rate. Furthermore, it is revealed
ix
through the resulting expression for the risky bond price, that the default intensity λ is the
risk spread between the riskless and risky spot rates.
The main example used to illustrate this procedure is the well-known Ornstein-
Uhlenbeck process from which a Cox-Ingersoll-Ross model of both spot rates is derived.
In addition to computing bond prices with this example, a Cauchy problem for an
originally designed option on the risk spread is derived through the Feynmann-Kac
Theorem. A series solution is then developed using a modern potential theoretic version
of the classical parametrix method for parabolic partial differential equations.
1
CHAPTER 1 A POTENTIAL THEORETIC FRAMEWORK FOR FIXED-INCOME MARKETS
1.1 The Essentials of Mathematical Finance
The two main prerequisites of mathematical finance that are imperative to an
understanding of this dissertation are fixed-income finance and arbitrage pricing theory.
This section begins by establishing the probabilistic setting in which these concepts
will be reviewed. According to Musiela and Rutkowski (1998), an economy is a family
of filtered probability space ( ) , , :Ω µ µ ∈F P , where the filtration t t [0,T]∈=F F
satisfies the usual conditions, and P is a collection of mutually equivalent probability
measures on the measurable space( )T,Ω F . We model the subjective market uncertainty
of each investor by associating to each investor a probability measure from P. Investors
with more risk tolerance will be represented by probability measures that weight
unfavorable events relatively lower, whereas conservative investors are characterized by
probability measures that weight unfavorable events relatively higher. Moreover, it is
assumed that investment information is revealed to each investor simultaneously as
events in the filtration F. Since the measures in P are mutually equivalent, the investors
agree on the events that have and have not occurred. It is convenient to further assume
that investors initially have no other information, i.e. 0F is trivial with respect to each
probability measure in P. This assumption asserts that the initial information available to
investors is objective.
2
1.1.1 The Fundamentals of Fixed-Income Finance The foundation of a working knowledge of fixed-income finance rests on an
understanding of the inherent relationship between the various interest rates and bonds.
Consider the economy ( ) , , :Ω µ µ ∈F P on the interval [ ]0,T and a Markov process
tX with ( )t sX :0 s t≡ σ ≤ ≤F . Implicit in this statement is the assumption that the state-
variable probability xP P≡ associated with tX belongs to P for some fixed element x of
the state space of tX . A zero-coupon bond, or discount bond, of maturity T is a security
that pays the holder one unit of currency at time T. The prices of government and
corporate discount bonds at time t ≤ T are denoted by B(t,T) and B(t,T)% , respectively.
The local expectations hypothesis (L-EH) relates the discount bond to the instantaneous
interest rate, or spot rate, for borrowing and lending over the time interval [ ]t , t dt+ .
Denote the riskless spot rate by ( )t tr r X= and assume that it is a nonnegative, adapted
process with almost all sample paths integrable on [0,T] with respect to Lebesgue
measure. The L-EH asserts that
( )( )T
P s ttB(t,T) E exp r X ds = −
∫ F (1.1)
According to Musiela and Rutkowski (1998), the economic interpretation of this
hypothesis is that “...the current bond price equals the expected value ... of the bond price
in the next (infinitesimal) period, discounted at the current short-term rate” (p. 283). This
statement is better understood in a discrete-time setting. In fact, using a left sum
approximation to the integral in (1.1) with the partition n
i i 0t
= of [ ]0,T yields
3
( )i 1
n
P t ii 1
B(0,T) E exp r X t−
=
= − ∆ ∑ (1.2)
( )( ) ( )0 i 1
n
P t 1 t ii 2
E exp r X t exp r X t−
=
= − ∆ − ∆ ∑
( )( ) ( )i 1 1
n
1 P P t i ti 2
exp r x t E E exp r X t−
=
= − − ∆ ∑ F
( )( ) ( )( )x1 P 1exp r x t E B t ,T= − .
Under the assumption of no arbitrage, it will be shown that (1.1) holds under the risk-
neutral measure in Section 2. Naturally, a similar relationship holds between the risky
bond B% and the risky spot rate tr% , which will be derived in Section 3.
The savings account tB is a process that represents the price of a riskless security that
continuously compounds at the spot rate. More precisely, it is the amount of cash at time
t that accumulates by investing one dollar initially, and continually rolling over a bond
with an infinitesimal time to maturity. Hence, we have
( )( )t
t s0B exp r X ds≡ ∫ . (1.3)
When a security tS is divided by the savings account, the resultant process is the price
process of the security discounted at the riskless rate.
Another bond of importance is known as the coupon bond, which pays the holder
fixed coupon payments 1 nc ,...,c at fixed times 1 nT,...,T with nT T= . The price of the
coupon bond is simply the present value of the sum of these cash flows. Denoting the
price of a riskless coupon bond at time t by cB (t,T) , we have
i
c i iT t
B (t,T) c B(t,T)>
= ∑ . (1.4)
A similar relationship holds for the risky coupon bond cB% .
4
In practice, the coupons are typically structured by setting ic c= for i 1,...,n 1= − ,
and nc N c= + , where N is the principal, or face value, and c is a fixed amount that is
generally quoted as a percentage of N called the coupon rate.
A problem that arises in comparing coupon bonds is that the uncertainty of the rate at
which the coupons will be reinvested causes uncertainty in the total return of the coupon
bond. Hence, coupon bonds of different coupon rates and payment dates are not directly
comparable. The continuously compounded riskless yield-to-maturity (YTM) Y(t,T) is
the unique solution to the equation
( )( )( )i
c i iT t
B (t,T) c exp Y t,T T t>
= − −∑ , (1.5)
and represents the total return on the coupon bond under the assumption that each of the
coupon payments occurring after t are reinvested at the rate Y(t,T) . The risky YTM
Y(t,T)% is defined in a similar fashion.
The interested reader should verify that there exists a unique, adapted, nonnegative
process 0 t TY(t,T)
≤ ≤ given the adapted coupon bond process, coupon payments, and
payment dates. In fact, this follows by noting that the LHS of (1.5) is a decreasing
function of Y, and that the price of a coupon bond will never exceed the sum of the
coupon payments.
The yield-to-maturity expectations hypothesis (YTM-EH) relates the riskless YTM
and the riskless spot rate. Musiela and Rutkowski (1998) state this hypothesis as the
assertion that “...the [continuously compounded] yield from holding any [discount] bond
is equal to the [continuously compounded] yield expected from rolling over a series of
single-period [discount] bonds” (p.284). To gain a better understanding of this statement,
5
we first observe that the YTM of a discount bond is simply the continuously compounded
interest rate. Hence, in a discrete-time setting with the partition n
i i 0t
= of [ ]t,T , we have
that the yield of a discount bond ( )i 1 iB t , t− is given by
( )i 1i 1 i tY(t , t ) r X−− = , (1.6)
from which we deduce that the bond price is given by
( ) ( )( )i 1i 1 i t iB t , t exp r X t−− = − ∆ . (1.7)
Since the YTM-EH asserts that the yield of ( )B t,T is equal to the yield expected
from rolling over a series of discount bonds ( )i 1 iB t , t− , it follows that
Y(t,T) ( )( )n
P i 1 i ti 1
1 1ln B t,T E ln B(t , t )
T t T t −=
≡ − = − − −
∏ F (1.8)
( )i 1
n
P t i ti 1
1E r X t
T t −=
= ∆ −
∑ F .
Taking the limit as the mesh of the partition tends to zero, we obtain the continuous-
time discount bond price and YTM under the YTM-EH:
( )T
P s ttB(t,T) exp E r X ds = − ∫ F (1.9)
and
( )T
P s tt
1Y(t,T) E r X ds
T t = − ∫ F . (1.10)
The last interest rate that we will consider is the instantaneous forward interest rate,
or forward rate for borrowing or lending over the time interval [ ]s,s ds+ as seen from
time t s≤ . This will be denoted by f(t, s) in the riskless case and f(t,s)% in the risky case.
If the dynamics of the process t s Tf(t,s)
≤ ≤ are specified, then the price of the
discount bond is defined by
6
( )T
tB(t,T) exp f(t,s)ds≡ −∫ . (1.11)
Alternatively, if the dynamics of the discount bond are known, then we have
f(t,T) lnB(t,T)T
∂≡ −
∂, (1.12)
provided that this derivative exists. By combining (1.9) and (1.12) we obtain
( )( )P T tf(t,T) E r X= F . (1.13)
Therefore, the YTM-EH asserts that the forward rate is an unbiased estimate of the
spot rate under the state-variable probability measure P. Under the assumption of no
arbitrage, it is shown in Section 3 that this holds under the forward martingale measure.
1.1.2 A Review of Arbitrage Pricing Theory
The terminology presented in this review may be found in Musiela and Rutkowski
(1998). Consider the economy ( ) , , :Ω µ µ ∈F P on the interval [ ]0,T . A trading
strategy, or portfolio, tφ is a vector of locally bounded, adapted processes of tradable
asset holdings. Moreover, it is assumed that every sample path is right continuous with
left limits. A trading strategy tφ is called self-financing, if the wealth process ( )tV φ of
the trading strategy neither receives nor pays out cash flows external to the assets that
comprise the strategy. More precisely, let iφ denote the holding of asset iS . Then, a
self-financing trading strategy ( )1 n,...,φ = φ φ is defined by asserting that ( )n
i it t t
i 1
V S=
φ ≡ φ∑
satisfies
( ) 1 1 n nt t t t tdV dS ... dSφ = φ + + φ . (1.14)
7
A self-financing strategy φ is called an arbitrage portfolio, if its associated wealth
process satisfies all of the following conditions for some (thus for all) P ∈ P:
• ( )0V 0φ = (Zero Investment)
• ( )( )TP V 0 1φ ≥ = (Zero Risk)
• ( )( )TP V 0 0φ > > (Possible Gain).
Hence, an investor taking advantage of an arbitrage opportunity may become infinitely
wealthy without risk. Under the assumption that arbitrage portfolios do not exist, it has
been shown that there exists a risk-neutral, or martingale measure, Q in our economy
under which the discounted asset process 1t t tZ B S−≡ follows a martingale. This result,
known as the Fundamental Theorem of Asset Pricing, is proven in a quite general setting
by Delbaen and Schachermayer (1994, 1998).
The next topic for review is the arbitrage pricing of financial derivatives. A self-
financing trading strategy φ is Q-admissible if the discounted wealth process
( ) ( )1t t tV B V−φ ≡ φ is a Q-martingale and uniformly bounded below with respect to
[ ]t 0,T∈ . The uniform boundedness condition is included to disallow trading strategies
in which the investor’s debt may become arbitrarily large. A contingent claim, or TF -
measurable random variable, C is Q-attainable if there exists a Q-admissible trading
strategy φ that replicates the value of C at time T (i.e. ( )TV Cφ = ). The market is
defined by M(Q) ≡ (S,Φ), where Φ consists of the Q-admissible trading strategies. A
market is said to be complete if every contingent claim is attainable.
Under the assumption of no arbitrage, an attainable claim C is uniquely replicated for
each martingale measure Q. In fact, we define the arbitrage price process ( )t C Qπ of C
8
to be the wealth process of the uniquely replicating trading strategy. Since this strategy is
Q-admissible, it follows that
( ) ( )1t t Q T tC Q B E B C−π = F . (1.15)
It is shown in Musiela and Rutkowski (1998) that ( ) ( )t 1 t 2C Q C Qπ = π for distinct
martingale measures 1Q and 2Q , if C is attainable with respect to both measures. Hence,
the definition of arbitrage price is independent of the choice of martingale measure and
will be denoted by ( )t Cπ . Therefore, if we assume that the market is complete, then the
pricing of contingent claims does not depend on the choice of martingale measure.
Alternatively, we may assume that the martingale measure is unique from which it
follows that the market is complete in the restricted sense that every contingent claim C
with ( )1T 1B C , ,Q− ∈ ΩL F is attainable (Björk, 1996). Either assumption will suffice for
the contingent claims considered in this dissertation. Furthermore, it will be shown in the
next section that the expression for the arbitrage price (1.15) can be rewritten with respect
to the state-variable probability P associated with the Markov process tX used to model
market uncertainty.
1.2 Potential Approach I: Riskless Bonds and the Martingale Measure
Equipped with the notions from our review of mathematical finance, we will now
present the potential approach to developing models of the riskless spot rate. The
fundamentals of this approach will carry over to the next section where a framework in
which credit derivatives such as risky bonds and the risk-spread option can be priced.
9
Consider the economy ( ) , , :Ω µ µ ∈G P on the interval [ ]0,T and a Markov
process tX with ( )t sX : 0 s t≡ σ ≤ ≤G . Combining the concepts of fixed-income finance
with those of arbitrage pricing theory, we see that a discount bond is simply a contingent
claim with the constant value one. Under the assumption of no arbitrage, it follows from
(1.15) that there exists a risk-neutral measure Q such that
( ) ( ) ( )T1
t t Q T t Q s ttB(t,T) 1 B E B E exp r(X )ds− = π = = −
∫G G . (2.1)
This proves the previously stated assertion that the L-EH given in (1.1) is satisfied under
the risk-neutral measure Q.
Rogers (1997) has shown that the expectation in (2.1) can be rewritten with respect to
the state-variable probability P using the state-price density
( )t 1t s P t t t0
dQexp r(X )ds E B N
dP−
ζ ≡ − =
∫ G , (2.2)
where t P t
dQN E
dP
≡
G . Before proving this result, we recall the following abstract
version of Bayes Rule:
Lemma 1 – Let Q and P be probabilities on the measurable space ( ),Ω J , H be a sub-
σ-algebra of J , ( )1f , ,Q∈ ΩL J , and dQ
NdP
≡ . Then
( ) ( )( )
PQ
P
E f NE f
E N=
HH H . (2.3)
proof : See pg. 458 of Musiela and Rutkowski (1998). <
10
Theorem 2 – For any contingent claim C, we have ( ) P T tt
t
E CC
ζ π =ζ
G. (2.4)
proof : From (1.15) and Lemma 1, it follows that
( ) ( ) ( )( )
1P T T t P T t1
t t Q T t ttP T t
E CB N E CC B E B C B
E N
−− ζ π = = =
ζ
G GG G . < (2.5)
An immediate consequence of Theorem 2 is the desired expression:
( ) P T t
t
EB t,T
ζ =ζ
G. (2.6)
This is the fundamental result of the potential approach. Since r is nonnegative, it is
easily verified that tζ is a positive supermartingale with respect to P. In fact, we have
( ) ( )( )t s
t s
P t s t P s t0
dQE E exp r X ds
dP+
+
+
ζ = −
∫
GG G (2.7)
( )( )t s
t
s P t0
dQexp r X ds E
dP+
≤ −
∫
GG
( )( )t
s P t t0
dQexp r X ds E
dP
= − = ζ
∫ G .
Furthermore, if we assume that ( )P tE 0ζ → as t → ∞, then tζ is a potential. From
(2.6) it follows that this assumption translates into the reasonable financial assumption
that the price of the riskless bond B(0,t) tends to zero as the time until maturity increases
to infinity.
The general potential approach to fixed-income finance outlined in Rogers (1997) is
to generate models of the spot rate through a judicious choice of the state-price density.
The only mathematical restriction is that tζ must be a positive supermartingale with
11
respect to P. A specific procedure is to choose a positive function g defined on the state
space of tX and use this to define tζ by
t tt
0
U g(X )e
U g(X )
α−α
αζ ≡ , (2.8)
where ( )0
U gα
α > is the α-potential of g defined by
( ) ( )( )x sP s0
U g x E e g X ds∞α −α= ∫ . (2.9)
Since this tζ is clearly nonnegative, we must only verify that it is a supermartingale
with respect to P. In fact, consider the martingale
( )st P s t0
M E e g X ds∞
−α = ∫ G . (2.10)
Applying the Markov property of tX , we deduce
( )tt t tM A e U g X−α α= + , (2.11)
where ( )t s
t s0A e g X ds−α≡ ∫ is an increasing process. It follows from (2.8) that
( )t t t0
1M A
U g(X )αζ = − (2.12)
from which we deduce that tζ is a supermartingale.
Given the above model for the state-price density, (2.6) may be employed to derive
the price of a riskless bond. We will now derive an expression for the riskless spot rate
by comparing (2.2) with (2.12). It follows from (2.2) that
1t t t t td B dN r(X ) dt−ζ = − ζ , (2.13)
On the other hand, we deduce from (2.12) that
( ) ( )( )t
t t t
0
1d dM e g X dt
U g X−α
αζ = − . (2.14)
12
Since tN is a (local) martingale, a comparison of (2.13) and (2.14) reveals the desired
expression for the riskless spot rate
( ) ( ) ( )tt t
t0 t t
e g X g Xr X
U g(X ) U g(X )
−α
α α= =ζ
, (2.15)
where the model of the state-price density (2.8) was used in the final equality.
Rather than specify g directly, it is convenient to model its α-potential by a
nonnegative function f, defined on the state space of tX , which lies in the domain of the
infinitesimal generator G of tX . With ( ) ( )t tf X U g Xα≡ we rewrite (2.8) as
( )( )
ttt
0
f Xe
f X−αζ = . (2.16)
Since Uα is the inverse of α – G, we have g = (α – G)f. Hence, the expression for the
spot rate (2.15) may be rewritten as
( ) ( )
( )t
tt
G f Xr
f X
α −= . (2.17)
1.3 Potential Approach II: Risky Bonds and the Forward Martingale Measure
The main goal of this section is to extend the results of the previous section to the
risky setting by developing a model in which the risky bond price is expressed in a
similar fashion to the riskless bond in (2.6). The main difference is that the riskless state-
price density tζ is replaced with the risky state-price density tϕ , which is expressed in
terms of the forward martingale measure instead of the risk-neutral measure.
Consider the economy ( ) , , :Ω µ µ ∈F P on the interval [ ]0,T with F defined
below. Let the Markov process tX denote the state-variable process, and ν be a random
13
time denoting the time of default. Following Lando (1998), we define F so as to make ν
a stopping time and adapted to tX as follows:
• tG ( )sX :0 s t≡ σ ≤ ≤
• tH [ ]( )s : 0 s t≡ σ ν > ≤ ≤
• tF t t≡ ∨G H . In the first of the following two sub-sections, an expression for the risky bond price is
derived under the assumption that ν may be represented by the first jump of a generalized
Poisson process. The forward martingale measure is introduced in the second sub-
section to derive the risky bond analog of (2.6).
1.3.1 Risky Bonds
A risky bond is a contingent claim that pays the holder one unit of currency at
maturity in the event that there is no default. Hence, under the assumption of no
arbitrage, it follows from (1.15) that there exists a risk-neutral measure Q such that
[ ]( ) [ ]( ) ( ) [ ]T
1t t Q T t Q s tT T Tt
B(t,T) 1 B E B 1 E exp r(X )ds 1−ν > ν > ν >
= π = = − ∫% F F . (3.1)
In applying the potential approach to risky bonds, it will be convenient to rewrite the
conditional expectation in (3.1) with respect to tG . Before this can be achieved, we will
recall the Monotone Class Theorem (MCT) and use it to prove some preliminary results.
Theorem 1 (The Monotone Class Theorem) – Let A and D be collections of subsets
of a set C . Then ( )σ ⊆A D if the following conditions are satisfied:
(i) A B∈∩ A for every A and B in A
(ii) ⊆A D (iii) ∈C D (iv) B \ A ∈D for every A and B in D with A ⊂ D
(v) ii 1
A∞
=
∈∪ D for every increasing sequence i i 1A
∞
= of sets from D .
14
A proof of this result may be found in Blumenthal and Getoor (1968) among other places.
The first result that we will prove using the MCT is
Lemma 2 – For every t T≤ and tA ∈ H , we have that either
[ ]( ) [ ]( )T TQ t A Q tν > = ν >∩ G G (3.2)
or
[ ]( )TQ t A 0ν > =∩ G . (3.3)
proof: It suffices to show that for every t T≤ , tA ∈ H , and TC∈G we have that either
[ ]( ) [ ]( )Q t A C Q t Aν > = ν >∩ ∩ ∩ (3.4)
or
[ ]( )Q t A C 0ν > =∩ ∩ . (3.5)
Let t T≤ . We will apply the MCT with ≡ ΩC ,
[ ] t : 0 s t≡ ν > ≤ ≤A , (3.6)
and
[ ]( ) [ ]( ) t TA :Q t A C Q t C or 0 for every C≡ ∈ ν > = ν > ∈∩ ∩ ∩D H G . (3.7)
Let 1 20 s s t≤ < ≤ . Since [ ] [ ] [ ]1 2 2s s sν > ν > = ν >∩ , it follows that A is closed
under intersections and ⊆A D . Hence, A satisfies the hypotheses of the MCT. We
proceed to verify the hypotheses on D .
Since it is clear that Ω ∈D , we begin by showing that D is closed under proper
differences. Let A ∈ D , B∈D , and TC∈G with A B⊂ . We have
[ ]( )( ) [ ]( ) [ ]( )( )Q B \ A t C Q B t C \ A t Cν > = ν > ν >∩ ∩ ∩ ∩ ∩ ∩ (3.8)
[ ]( ) [ ]( )Q B t C Q A t C= ν > − ν >∩ ∩ ∩ ∩ .
If [ ]( )Q B t C 0ν > =∩ ∩ , then the RHS of (3.8) is also equal to zero since A B⊂ .
So, assume that [ ]( )Q B t C 0ν > >∩ ∩ . Since B∈D we have that
15
[ ]( ) [ ]( )Q B t C Q t Cν > = ν >∩ ∩ ∩ . (3.9)
Hence, if [ ]( )Q A t C 0ν > =∩ ∩ , then the RHS of (3.8) is equal to [ ]( )Q t Cν > ∩ . On
the other hand, if [ ]( )Q A t C 0ν > >∩ ∩ , then
[ ]( ) [ ]( )Q A t C Q t Cν > = ν >∩ ∩ ∩ (3.10)
since A ∈ D . From (3.9) and (3.10) we deduce that the RHS of (3.8) is zero. It follows
that B \ A ∈D .
The final hypothesis of the MCT that must be verified is that D is closed under
increasing sequences. Let i i 1A
∞
= be an increasing sequence of sets from D and define
ii 1
A A∞
=
≡ ∪ . We will show that A ∈ D . We begin by defining the pairwise disjoint
sequence of sets i i 1B
∞
= by 1 1B A≡ and i 1 i 1 iB A \ A+ +≡ . Since D is closed under
proper differences, we have that iB ∈D for every i 1≥ . Furthermore, it is clear that
ii 1
A B∞
=
= ∪ . It follows that
[ ]( ) [ ] [ ]( )i ii 1i 1
1 Q A t C Q B t C Q B t C∞ ∞
==
≥ ν > = ν > = ν >
∑∩ ∩ ∩ ∩ ∩ ∩∪ . (3.11)
Since this series is finite, we have that
[ ]( )iQ B t C 0ν > =∩ ∩ (3.12)
for all but finitely many i. Define
[ ]( ) iI i 1 :Q B t C 0≡ ≥ ν > >∩ ∩ . (3.13)
Since iB ∈D for every i 1≥ , it follows that
[ ]( ) [ ]( ) [ ]i ii I i I
Q A t C Q B t C Q B t C∈ ∈
ν > = ν > = ν >
∑∩ ∩ ∩ ∩ ∩ ∩∪ , (3.14)
[ ]( ) [ ]( )maxiQ A t C Q t C≤ ν > ≤ ν >∩ ∩ ∩ ,
16
where maxi max i I= ∈ and the last inequality holds since maxiA ∈D . Furthermore,
since iB ∈D for every i I∈ , we also have that
[ ]( ) [ ]( ) [ ]( )ii I
Q A t C Q B t C nQ t C∈
ν > = ν > = ν >∑∩ ∩ ∩ ∩ ∩ , (3.15)
where n denotes the cardinality of I. Comparing (3.14) with (3.15), we see that n must be
zero or one from which we deduce that A ∈ D .
Therefore, we conclude from the MCT that ( )t = σ ⊆H A D . This implies that
either (3.2) or (3.3) holds for every t T≤ and tA ∈ H . <
Theorem 3 – For every t T≤ we have
( ) [ ]( )( )
TT t t
T
Q T |Q T | 1
Q t |ν>
ν >ν > ∨ =
ν >GG H G . (3.16)
proof : Define ( )t TY Q t≡ ν > G for every t T≤ and rewrite (3.16) as
[ ]( ) [ ]Q t T t TT tE 1 Y 1 Yν> ν >∨ =G H . (3.17)
In the following proof, be aware that tY is TG -measurable for each t T≤ and is not
necessarily tG -measurable.
Since the RHS of (3.17) is measurable with respect to T t∨G H , it suffices to show
that for every t T≤ and T tD ∈ ∨G H we have
[ ]( ) [ ]( )Q t Q TT D t DE 1 Y E 1 Yν> ν >=∩ ∩ . (3.18)
Let t T≤ . We will prove (3.18) using the MCT with ≡ ΩC ,
T tA B : A ,B≡ ∈ ∈∩A G H , (3.19)
and
[ ]( ) [ ]( ) T t Q t Q TT D t DD : E 1 Y E 1 Yν> ν>≡ ∈ ∨ =∩ ∩D G H . (3.20)
17
We begin by checking that A satisfies the hypotheses of the MCT. Clearly, A is
closed under intersections. Let A B∈∩ A for some TA ∈G and tB∈H . We will show
that ⊆A D by considering the following two cases:
Case 1 – [ ]( )TQ T B 0ν > >∩ G
Since t TB ∈ ⊆H H , it follows from Lemma 2 that
[ ]( ) ( )T TQ T B Q Tν > = ν >∩ G G . (3.21)
Furthermore, since [ ] [ ]T tν > ⊆ ν > we have that
[ ]( ) [ ]( )T TQ t B Q T B 0ν > ≥ ν > >∩ ∩G G . (3.22)
Thus, another application of Lemma 2 yields
[ ]( ) ( )T TQ t B Q tν > = ν >∩ G G . (3.23)
From (3.21) and (3.23) we deduce that
[ ] ( )( )Q tT A BE 1 Yν> ∩ ∩ [ ]( )( ) [ ]( )( )Q A t T Q A t TE 1 Y Q T B E 1 Y Q T= ν > = ν >∩ G G (3.24)
( ) ( )( )Q A t T Q A T TE 1 Y Y E 1 Y Q t= = ν > G
[ ]( )( ) [ ] ( )( )Q A T T Q Tt A BE 1 Y Q t B E 1 Yν>= ν > = ∩ ∩∩ G .
Hence, A B ∈∩ D which implies that ⊆A D in this case.
Case 2 – [ ]( )TQ T B 0ν > =∩ G
From the assumption of this case, we have that
[ ] ( )( ) [ ]( )( )Q t Q A t TT A BE 1 Y E 1 Y Q T B 0ν> = ν > =∩ ∩ ∩ G . (3.25)
So, it suffices to show that
[ ] ( )( )Q Tt A BE 1 Y 0ν> =∩ ∩ . (3.26)
18
Hence, we may assume without loss of generality that TY 0> . This implies that
[ ]( ) [ ]( )C CT TQ t B Q T Bν > ≥ ν >∩ ∩G G (3.27)
[ ]( ) [ ]( )CT T TQ T B Q T B Y 0= ν > + ν > = >∩ ∩G G .
It follows from Lemma 2 that
[ ]( ) ( )CT TQ t B Q tν > = ν >∩ G G (3.28)
since CtB ∈H . On the other hand, we have that
( ) [ ]( ) [ ]( )CT T TQ t Q t B Q t Bν > = ν > + ν >∩ ∩G G G . (3.29)
Combining (3.28) and (3.29) we obtain
[ ]( )TQ t B 0ν > =∩ G (3.30)
from which we deduce that
[ ] ( )( ) [ ]( )( )Q T Q A T Tt A BE 1 Y E 1 Y Q t B 0ν> = ν > =∩ ∩ ∩ G . (3.31)
Therefore, we deduce from (3.25) and (3.31) that ⊆A D in this case as well.
We proceed to verify that D satisfies the hypotheses of the MCT. Clearly, Ω ∈D .
In fact, we have that
[ ]( ) ( )( ) ( )Q t Q t T Q t TTE 1 Y E Y Q T E Y Yν> = ν > =G (3.32)
( )( ) [ ]( )Q T T Q TtE Y Q t E 1 Yν>= ν > =G .
The next step is to show that D is closed under proper differences. Let A and B be
sets in D with A B⊂ . We deduce that B \ A ∈D from the following result:
[ ] ( )( )Q tT B\AE 1 Yν> ∩ [ ]( ) [ ]( )( ) [ ]( )( ) [ ]( )( )Q t Q t Q tT B \ T A T B T AE 1 Y E 1 Y E 1 Y
ν> ν > ν> ν >= = −∩ ∩ ∩ ∩ (3.33)
[ ]( )( ) [ ]( )( )Q T Q Tt B t AE 1 Y E 1 Y
ν> ν>= −∩ ∩
[ ]( ) [ ]( )( ) [ ] ( )( )Q T Q Tt B\At B \ t AE 1 Y E 1 Yν>ν> ν>
= = ∩∩ ∩ .
19
The final hypothesis of the MCT that must be verified is that D is closed under
increasing sequences. Let i i 1A
∞
= be an increasing sequence of sets from D and define
ii 1
A A∞
=
≡ ∪ . We will show that A ∈ D . As in Lemma 2, we begin by defining the
pairwise disjoint sequence of sets i i 1B
∞
= by 1 1B A≡ and i 1 i 1 iB A \ A+ +≡ . Since D is
closed under proper differences, we have that iB ∈D for every i 1≥ . Furthermore, it is
clear that ii 1
A B∞
=
= ∪ . Thus, it follows that A ∈ D from the following equality:
[ ]( ) [ ]( ) [ ]( ) [ ]( )i iQ t Q t Q T Q TT A T B t B t A
i 1 i 1
E Y 1 E Y 1 E Y 1 E Y 1∞ ∞
ν> ν> ν> ν >= =
= = =∑ ∑∩ ∩ ∩ ∩ . (3.34)
Hence, we conclude from the MCT that ( )T t∨ = σ ⊆G H A D . Thus, (3.18) holds
for every t T≤ and T tA ∈ ∨G H from which we obtain the desired result (3.16). <
We now return to the problem of rewriting the conditional expectation in (3.1) with
respect to tG . Conditioning with respect to T t∨G H first yields
( )T
Q Q s [ T] T t ttB(t,T) E E exp r(X )ds 1 ν>
= − ∨ ∫% G H F (3.35)
( ) ( )T
Q s T t ttE exp r(X )ds Q T | = − ν > ∨ ∫ G H F
[ ] ( ) ( )( )
T TQ s tt t
T
Q T1 E exp r(X )ds
Q tν>
ν >= −
ν > ∫
G FG ,
where the last inequality follows from Theorem 3. This expression for the risky bond is
more attractive than (3.1) since the argument of the conditional expectation is now TG -
measurable. However, we are still unable to replace the conditioning on tF with
conditioning on tG . This will require an independence assumption and the following
lemma:
20
Lemma 4 – Let 1F , 2F , and 3F be sub σ-algebras of the σ-algebra F such that 1 2∨F F
is independent of 3F . Then for every integrable, 1F -measurable function Y we have
( ) ( )2 3 2E Y | E Y |∨ =F F F .
proof : See Section 9.2 of Chung (1974). <
Assuming that TG is independent of tH for each t T≤ , we may apply Lemma 4 with
1 T≡F G , 2 t≡F G , and 3 t≡F H to (3.35) to obtain
( ) [ ] ( ) ( )( )
T TQ s tt t
T
Q TB t,T 1 E exp r(X )ds
Q tν>
ν >= −
ν > ∫ % G GG . (3.36)
The financial interpretation of this independence assumption is that the riskless spot
rate up until the time horizon T is independent of the default status of the bond prior to
maturity. We will now present a model of the default time in which this independence
assumption is somewhat relaxed.
As in Lando (1998), we model the default time by t
s0inf t 0 : (X )dsν ≡ > λ ≥∫ E ,
where E is distributed under Q as a unit exponential random variable that is independent
of TG , and λ is a nonnegative, continuous function on the state space. The default time
may be regarded as the first jump in a generalized Poisson process with intensity
( )t
t s0X dsΛ ≡ λ∫ , and we deduce that ( ) ( )T tQ t | expν > = −ΛG . In fact, we clearly have
that tΛ is measurable with respect to tG . So, it suffices to show that for every tA ∈G we
have [ ]( ) ( )( )Q A tQ t A E 1 expν > = −Λ∩ . It follows that
[ ]( ) [ ]( ) ( ) ( )( )t
ut Q A Q A tQ t A Q A E 1 e du E 1 exp
∞ −
Λν > = Λ < = = −Λ∫∩ ∩E . (3.37)
21
Hence, (3.35) becomes
[ ] ( )( )T
Q s tt tB(t,T) 1 E exp r X dsν>
= − ∫% % F , (3.38)
where r r≡ + λ% . Now, since E is independent of TG , we may apply Lemma 4 with
1 T≡F G , 2 t≡F G , and ( )3 ≡ σF E to obtain
( )( ) ( ) ( )( )T T
Q s t Q s tt tE exp r X ds E exp r X ds − ∨ σ = − ∫ ∫% %G E G . (3.39)
Since ( )t t t t t⊆ ≡ ∨ ⊆ ∨ σG F G H G E , we may condition with respect to tF on both
sides of (3.39) to obtain
( )( ) ( )( )T T
Q s t Q s tt tE exp r X ds E exp r X ds − = − ∫ ∫% %F G . (3.40)
Combining this with (3.38) we deduce the desired expression for the risky bond:
[ ] ( )( )T
Q s tt tB(t,T) 1 E exp r X dsν>
= − ∫% % G . (3.41)
We conclude that the process tλ represents the risk spread between the riskless spot
rate tr and risky spot rate tr% . Intuitively, this makes sense since the probability of default
increases with tλ .
1.3.2 The Forward Martingale Measure
The next step before applying the potential approach to risky bonds is to develop an
understanding of the forward martingale measure (FMM) which is denoted by F. We
define this probability measure through the derivative
t
tt
dF B(t,T)Y
dQ B B(0,T)≡ ≡
G. (3.42)
22
It should be obvious from (2.1) that (3.42) defines a Q-martingale, since it is shown there
that the discounted bond price follows a martingale under Q.
The FMM arises when pricing forward contracts in a market without arbitrage.
Formally, a forward contract is an agreement established at time 0t T< to exchange an
asset for a prearranged delivery price at time T. More precisely, we have
Definition 3 – A forward contract written at time 0t T< on an attainable contingent
claim C for settlement at time T with delivery price K is an attainable contingent claim
H C K≡ − , where the delivery price is a fixed amount of cash determined at time 0t .
Since there is no initial exchange of money between the participants of a forward
contract, the delivery price must be set such that ( )0t H 0π = . In other words, the delivery
price K is set equal to the arbitrage price of C. If this is not the case, then it can be shown
that an arbitrage portfolio exists.
Example 4 – Let H be a forward contract as in the previous definition with 0t 0= , and
suppose that ( )0 H 0π > . Let Ctφ and H
tφ denote the replicating trading strategies of the
attainable claims C and H, respectively. Then an arbitrage portfolio can be constructed
by taking a long position in the forward contract, short selling Ctφ , and purchasing
( )( )0 C
B 0,T
π riskless bonds of maturity T. Denoting this portfolio by tψ we see that
( )
( )0 C H
t t t
C,0,...,0
B 0,T
πψ = − φ + φ
, (3.43)
where the first coordinate denotes riskless bond holdings. It follows that
( ) ( )( ) ( ) ( ) ( ) ( ) ( )0 C H C
0 0 0 0 0
CV B 0,T V V C V 0
B 0,T
πψ = − φ + φ = π − φ ≡ . (3.44)
This implies that the trading strategy ψ requires no initial investment. Furthermore,
23
( )TV ψ ( )
( ) ( ) ( ) ( ) ( )( ) ( )0 0C H
T T
C CB T,T V V C C K
B 0,T B 0,T
π π= − φ + φ = − + − (3.45)
( ) ( )( )
( )( )
0 0C KB 0,T H0
B 0,T B 0,T
π − π= = > a.s.
Hence, ψ also satisfies the zero risk and possible gain conditions of an arbitrage
portfolio. In fact, ψ represents extreme arbitrage in the sense that a positive profit will
almost surely be realized.
On the other hand, if ( )0 H 0π < then an arbitrage portfolio can be constructed by
taking a short position in the forward contract, short selling ( )
( )0 C
B 0,T
π riskless bonds of
maturity T, and purchasing Ctφ . That is, we construct an arbitrage portfolio by negating
the holdings of the previous case. Therefore, ( )0 H 0π = in an arbitrage free market. <
Although the forward contract has an initial price of zero, its arbitrage price may
fluctuate before it matures. We define the forward price of C at time t as the delivery
price for which ( )t H 0π = and form the adapted process ( ) 0
C t t TF t,T
≤ ≤. It follows from
(3.45) that the forward price is given by
( ) ( )( )t
C
CF t,T
B t,T
π= . (3.46)
Hence, the forward price C is simply the arbitrage price of C discounted by the
riskless bond price. This is not surprising since the FMM is a special case of a change of
numeraire with the riskless bond chosen as the new numeraire (Björk, 1996).
The next result justifies the name “forward martingale measure” for F.
Theorem 5 – Let C be an attainable contingent claim that is integrable with respect to F.
Then, the forward price process ( ) C 0 t TF t,T
≤ ≤ is a martingale under F.
24
proof : Since C is integrable with respect to F and ( )CF T,T C= , it suffices to show that
for every t ∈ [0, T] we have
( ) ( )C F tF t,T E C= G . (3.47)
We apply Bayes rule (Lemma 2.1) to obtain
( ) ( )( ) ( )Q T t 1
F t Q t T tQ T t
E Y CE C E Y Y C
E Y−= =
GG GG , (3.48)
where the martingale tY is defined in (3.42). From (3.42) and (3.46) we deduce
( ) ( )( )
( )( ) ( )
1t Q T t t
F t C
B E B C CE C F t,T
B t,T B t,T
−π
= = =G
G . < (3.49)
Theorem 6 – [ ] ( )( )T
F s tt tB(t,T) 1 B(t,T)E exp X dsν>
= − λ ∫% G . (3.50)
proof : An immediate consequence of (3.49) is
( ) ( ) ( )t F tC B t,T E Cπ = G . (3.51)
A comparison of (1.15) (with tF replaced by tG ) and (3.51) yields
( )( ) ( ) ( ) ( )T
Q s t t F ttE exp r X ds C C B t,T E C − = π =
∫ G G . (3.52)
Applying (3.52) to the attainable contingent claim ( )( )T
stC exp X ds= − λ∫ , we obtain the
desired result (3.50) from the risky bond equation (3.38). <
It follows from this theorem that the YTM-EH (1.13) holds under the forward
martingale measure. In fact, it follows from (3.52) with TC r= that
( ) ( ) ( )( )T
F T t Q T s tt
1E r E r exp r X ds
B t,T = − ∫G G (3.53)
( ) ( )( )T
Q s tt
1E exp r X ds
B t,T T
∂ = − − ∂ ∫ G
( ) ( ) ( ) ( )1
B t,T lnB t,T f t,TB t,T T T
∂ ∂= − = − =∂ ∂
.
25
We proceed to extend the potential approach to the risky bond. A comparison of the
expectation in (2.1) with that in (3.50) provides motivation for the following:
Definition 7 – The risky state-price density is given by ( )t
t t
dFexp
dPϕ ≡ −Λ
G.
It follows from Theorem 6 that we may proceed as in Theorem 2.2 to obtain the risky
analog of (2.6):
[ ] ( ) P T t
tt
EB(t,T) 1 B t,Tν >
ϕ =ϕ
% G. (3.54)
In fact, it follows from Bayes Rule (Lemma 2.1) that
( )( )( )( )T
P s ttT P T tF s tt
tP t
dFE exp X ds
EdPE exp X ds
dFE
dP
− λ ϕ − λ = = ϕ
∫∫
G GGG
. (3.55)
Hence, (3.54) follows from (3.50) and (3.55). Therefore, the risky bond price may be
determined by specifying the risky state-price density in a similar fashion to the
procedure outlined in Section 2.
We finish this section by noting that the risky state-price density may be expressed in
terms of its riskless counterpart and the risk spread. From the relation
( ) ( )
t t tt
tt
dF dF dQ B(t,T) dQ B(t,T)
dP dQ dP B B 0,T dP B 0,T= = = ζ
G G GG (3.56)
it follows that
( ) ( )t t t
B(t,T)= exp
B 0,Tϕ −Λ ζ . (3.57)
Inserting this into (3.54) yields
[ ]
( )( )T
P s T tt
tt
E exp X dsB(t,T) 1 ν >
− λ ζ =ζ
∫%G
. (3.58)
26
This expression for the risky bonds resembles (2.5); however, it does not follow
directly from Theorem 2.2 since [ ]t1 ν> is not tG -measurable. This result illustrates the
importance of the independence assumption used to derive (3.41).
1.4 A Simple Example of the Potential Theoretic Approach
Consider the economy ( ) , , :Ω µ µ ∈F P on the interval [ ]0,T with F defined as in
the beginning of the previous section. Let the state-variable process tX denote the
Gaussian diffusion with state space dR satisfying
t t tdX dW X dt= − θ (4.1) for some positive parameter θ, where tW is a d-dimensional Brownian motion. Hence,
tX is the well-known stationary Ornstein-Uhlenbeck process given by
( )tt st 0 s0
X e X e dW−θ θ= + ∫ . (4.2)
It follows that tX has distribution ( )t0 tN e X , V−θ , where ( )2 t
t
1V 1 e
2− θ≡ −
θ. Also, the
generator of this process is given by
d
2i
i 1 i
1 fGf f x
2 x=
∂≡ ∇ − θ
∂∑ (4.3)
for every f ∈ ( )2 dC R . Define the function df : +→R R by
2
f(x) exp xd
α =
(4.4)
for some positive parameter α. It follows from (2.17) that
( ) ( )
( )2t
t tt
G f X 1r X
f X 2
α −= = σ , (4.5)
where ˆ ˆ4 ( )σ ≡ α θ − α and ˆd
αα ≡ . Since the riskless spot rate must be positive, this
27
induces the condition ˆθ > α . We will show in Chapter 3 that this is the well-known
mean-reverting Cox-Ingersoll-Ross (CIR) process.
Next, the riskless state-price density given by (2.16) is
( )( ) ( )t
tt 0 t
0
e f(X )exp t r exp r
f(X )
−α
ζ = = − α + κ κ , (4.6)
where2
d
ακ ≡
σ. Hence, the price of the riskless bond may be calculated from (2.6):
( )B t,T ( ) ( )( ) ( )( )P T t
t P T tt
Eexp r E exp r
ζ= = − α τ + κ κ
ζ
G G (4.7)
( )( ) ( )( )tX
t Pexp r E exp rτ= − α τ + κ κ
( ) ( )( )d
2t
ˆˆ1 2 V exp V r−τ τ= − α − +ατ ,
where V
Vˆ1 2 Vτ
ττ
≡− α
and T tτ ≡ − . We deduce that the riskless YTM is given by
( ) ( ) ( )t
1 1 dˆ ˆY t,T lnB t,T V r ln 1 2 V2τ τ
≡ − = +ατ+ − α τ τ . (4.8)
Finally, we use (1.12) to derive an expression for the riskless forward rate:
( ) ( ) ( )( )
tlnB t,T exp 2 r ˆf t,T 2 V
ˆ ˆT 1 2 V 1 2 V ττ τ
∂ − θτ = − = − αα ∂ − α − α
. (4.9)
Similarly, in the risky setting we use a function dh : +→R R defined by
2
h(x) exp xd
β =
(4.10)
to model the risky state-price density, where β is a positive parameter. We deduce that
( ) ( )
( )2t
t tt
G h X 1X
h X 2
β −λ = = σ% , (4.11)
where ˆ ˆ4 ( )σ ≡ β θ − β% and ˆ
ˆd
ββ ≡ . Since the risk spread must be positive, this induces the
28
condition ˆθ > β . It follows that the risky state-price density is given by
( )( ) ( )t
tt 0 t
0
e h(X )exp t exp
h(X )
−β
ϕ = = − β + κλ κλ% % , (4.12)
where2
d
βκ ≡
σ% % . From (3.54), (4.11), and (4.12) we obtain the risky analog of (4.7):
B(t,T)% [ ] ( ) P T t
tt
E1 B t,Tν >
ϕ =ϕ
G (4.13)
[ ] ( )( ) ( )( )d2
ttˆ1 B t,T 1 2 V exp V
−
τ τν >= − β − λ +βτ% ,
where V
V ˆ1 2 Vτ
τ
τ
≡− β
% . We deduce that the risky YTM is given by
( ) ( ) [ ] ( ) ( )tt
1 1 d ˆY t,T lnB t,T 1 Y t,T V ln 1 2 V2
τ τν >
≡ − = + λ +βτ+ − β τ τ % % % . (4.14)
We conclude this example with an expression for the risky forward rate:
( ) ( )[ ] ( ) ( )
( )t
t
lnB t,T exp 2 ˆf t,T 1 f t,T 2 VˆˆT 1 2 V1 2 V
τν >ττ
∂ − θτ λ = − = + − ββ ∂ − β− β
%% . (4.15)
1.5 The Risk-Spread Option
Continuing with the economy of the previous section, we will conclude this chapter
by presenting the payoff of the risk-spread option. Consider an investor who purchases a
risky bond at t 0= that matures at time T. We would like to construct an option that will
guarantee that he would receive a minimum return of γ above the riskless spot rate. We
define the risk-spread option by its payoff of
( )( )T
s0C exp ds 1
+= γ − λ −∫ , (5.1)
and we refer to γ as the risk-spread insurance level.
29
We begin with a simple deterministic, discrete-time example under the assumption
that the risky bond does not default. Suppose that the riskless spot rate is given by
1 0t
2 0
p if t tr
p if t t
<= ≥
, (5.2)
and that the risk spread is given by
1 0t
2 0
q if t t
q if t t
<λ = ≥
, (5.3)
where 1 2q q> γ > . From (5.1) and (5.3) we have that
( )( )( )2 0C exp q T t 1= γ − − − . (5.4)
Since the risky rate t t tr r= + λ% is deterministic, it follows from (3.41), (5.2) and (5.3) that
( ) ( ) ( )( ) ( )( )( )T
s 0 1 1 0 2 20B 0,T exp r ds exp t p q exp T t p q= − = − + − − +∫% % . (5.5)
Because we are assuming that the risky bond does not default, the investor receives one
dollar at maturity. Hence, the total (continuously compounded) return the investor
receives from the option and the risky bond at maturity is
R( ) ( )0
1 1 Cln
T B 0,T C
+= + π % (5.6)
Assuming that ( )0 C 0π = , we have that
( )0
0 00 t
t T tR r r
T T
−= + + γ% .
It can be seen from this result that the investor is compensated when the risk spread
drops below the insurance level γ after time 0t . If 0t 0= , then the investor is guaranteed
30
to receive a return of γ above the riskless spot rate. On the other hand, if 0t T= , then the
option expires out-of-the-money and the investor receives the risky spot rate of return.
However, from (1.15) we see that ( ) 10 TC B C 0−π = > . This implies that the return on
the risk-spread option is the riskless spot rate. In fact, the rate of return of holding any
deterministic derivative is the riskless spot rate. Of course, there is no need for the risk-
spread option in a deterministic world.
In the general stochastic, continuous-time case we deduce from (5.4) and (5.6) that
R ( )( )
( ) ( )
T
s0
0
exp ds1ln
T B 0,T C
+ γ − λ = + π
∫% (5.7)
( ) ( ) ( )( )( )T
s 00
1ds ln B 0,T C
T
+= γ − λ − + π∫ % .
We would like to find 0γ >% such that
( )( )T
s s0
1R r ds
T
+= + γ − λ∫ % % . (5.8)
If there exists a constant γ% such that (5.7) and (5.8) are equal, we define the effective
risk-spread insurance level γ% as the solution to
( )( ) ( )( ) ( ) ( )( )P T P T T 0E E r ln B 0,T CT
+ + ∂γ − λ = γ − λ − − + π
∂%% . (5.9)
In the next chapter, the general stochastic, continuous-time case for the risk-spread
option is studied in detail. In particular, a representation for the arbitrage price of this
option is derived as the solution to a Cauchy problem.
31
CHAPTER 2 A CAUCHY PROBLEM FOR THE RISK-SPREAD OPTION
2.1 Derivation of the Cauchy Problem
Consider the economy ( ) , ,P : PΩ ∈F P on the interval [ ]0,T with F defined as in
Section 1.3. Let ( )T
t0C exp ( (X )) dt 1+= γ − λ −∫ denote the payout at time T of a risk-
spread option with fixed risk-spread insurance level γ, where t 0 t TX ≤ ≤ is the state-
variable process with natural filtration tG and probability P ∈P defined in Section 1.4.
According to (1.2.6), the price of this option at time t is given by
( )t Cπ ( )( )( )( )T
P s T t0P T t
t t
E exp X ds 1E C+ γ − λ − ζ ζ = =
ζ ζ
∫ GG (1.1)
( )T
P s T t0
t
E exp dsB(t,T)
Ψ ζ = −ζ
∫ G,
where
( ) ( )( )s s sX X+
Ψ ≡ Ψ ≡ γ − λ (1.2)
and ( )B t,T denotes the riskless bond price.
Using the model of the state-price density developed in the simple example of Section
1.4, it follows from (1.4.6) that
( )T
P s T t0E exp ds Ψ ζ ∫ G ( )( ) ( ) ( )T
0 P T s t0exp T r E exp r exp ds = − α + κ κ Ψ ∫ G (1.3)
( )( ) ( ) ( ) ( )tt X
0 s P s0 0exp T r exp ds E exp r exp ds
τ
τ = − α + κ Ψ κ Ψ ∫ ∫ .
32
Combining (1.1) and (1.3) yields
( ) ( )( ) ( ) ( ) ( )t
t t s t0C exp r exp ds u X , B t,Tπ = − α τ + κ Ψ τ −∫ . (1.4)
where
( ) ( ) ( ) ( ) ( )x xP s P s0 0
u x, E exp r exp ds E g X exp dsτ τ
τ τ τ ≡ κ Ψ = Ψ ∫ ∫ , (1.5)
( )2ˆg(x) exp x≡ α , d
αα ≡ , and T tτ ≡ − .
The goal of this section is to prove that (1.5) is the unique solution of class
( )2,1 du C [0,T]∈ ×R to the following Cauchy problem:
Lu(x, ) 0τ = (1.6)
for every ( ) dx, (0,T]τ ∈ ×x R , and
u(x,0) g(x)= (1.7)
for every dx ∈R , where
L G∂
≡ + Ψ −∂τ
(1.8)
and G is the generator of the state-variable process given by (1.4.3). This result is known
as the Feynman-Kac Theorem and is proven in (Karatzas & Shreve, 1991) under the
assumption that there are constants M 0> and 1µ ≥ such that for every dx ∈¡ we have
( )2
0 Tmax u(x, ) M 1 x µ
≤τ≤τ ≤ + . (1.9)
We will extend this result by replacing (1.9) with a less restrictive bound using the
following lemma, which appears as Problem 3.4.12 in (Karatzas & Shreve, 1991).
Lemma 1 – Let t t 0 tM ,
≤ < ∞G be a continuous, real-valued martingale such that 0M 0=
almost surely. Also, let tC be a continuous, real-valued process of bounded variation
such that t tC M+ ≤ ρ almost surely for some 0ρ > and every t 0≥ . Then, for every
n 2> ρ , the semimartingale t t tY C M≡ + satisfies
( ) ( )21
2t
0 t T
nP max Y n 3 2 exp
8−
≤ ≤
≥ ≤ πρ − ρ
. (1.10)
33
proof : We will first construct a time-changed Brownian Motion as in (Karatzas &
Shreve, 1991). For every t 0≥ define the stopping time
s
t
inf s 0 : M t if 0 t M
if t M
∞
∞
≥ > ≤ <τ ≡ ∞ ≥
. (1.11)
Without loss of generality1, assume that our probability space is rich enough to contain a
standard one-dimensional Brownian motion tt 0 t
B , τ ≤ < ∞% G that is independent of
1 See Remark 3.4.1 in (Karatzas & Shreve, 1991) for a technical justification.
t 0 tY
≤ < ∞ and define
tt t t MB B B M
∞τ∧≡ − +% % (1.12)
for every t 0≥ . It follows from Problem 4.7 of (Karatzas & Shreve, 1991) that
tt 0 t
B , τ ≤ < ∞G is a standard one-dimensional Brownian motion such that the filtration
t 0 tτ ≤ <∞
G satisfies the usual conditions and for every [ ]t 0,T∈ we have
t
t MM B= a.s. (1.13)
For every positive integer n, we define the stopping time
n t
nR inf t 0 : B
≡ ≥ ≥ 2 . (1.14)
Let n 2> ρ and note that
t t t t t
nM Y C Y Y
2≥ − ≥ − ρ ≥ − (1.15)
for every [ ]t 0,T∈ . Hence,
[ ]t t n n0 t T 0 t T
nmax Y n max M R RΤ≤ ≤ ≤ ≤
≥ ⊆ ≥ ⊆ Μ ≥ ⊆ ρ ≥ 2 . (1.16)
Using integration by parts, it can be verified that for every x 0> we have
2 2
x
u 1 xexp du exp
2 x 2
∞ − ≤ −
∫ . (1.17)
34
From (1.17), the symmetry of Brownian motion, and the reflection principle (Revuz &
Yor, 1991), it follows that
( )t0 t TP max Y n
≤ ≤≥ ( )n t t0 t 0 t
n nP R P max B 2P maxB
2 2≤ ≤ρ ≤ ≤ρ
≤ ≤ ρ = ≥ ≤ ≥
(1.18)
2
n2
n 4 u4P B exp du
2 22ρ
∞
ρ
= ≥ ≤ − π ∫
( )2 21
26 n n
exp 3 2 expn 2 8 8
− ρ≤ − ≤ πρ − π ρ ρ . <
We will use this lemma to prove a version of the Feynman-Kac Theorem for the
state-variable process tX defined in Section 1.4. Recall that
( )t t stt 0 s0
X e X e dW−θ −−θ= + ∫ , (1.19)
where tW is a d-dimensional Brownian motion and θ is a positive parameter. Define
tt t 0 tY e X X Mθ≡ = + , (1.20)
where the process t t 0 t
M ,≤ < ∞
G is the continuous, real-valued martingale defined by
t s
t s0M e dWθ≡ ∫ (1.21)
for every t 0≥ . Furthermore,
( ) ( )t 2 s 2 t 2 T
t 0
1 1M e ds e 1 e 1
2 2θ θ θ= = − ≤ −
θ θ∫ . (1.22)
Hence, for every t 0≥ we have
( )2 T0 0t
1X M X e 1
2θ+ ≤ ρ ≡ + −
θ. (1.23)
Therefore, it follows from Lemma 1 that
( ) ( )21
2t
0 t T
nP max Y n 3 2 exp
8−
≤ ≤
≥ ≤ πρ − ρ
. (1.24)
35
Theorem 2 (Feynman-Kac) – Let ( )2,1 du C [0,T]∈ ×R satisfy the Cauchy problem
given by
Lu(x, ) 0τ = (1.25)
for every ( ) dx, (0,T]τ ∈ ×x R , and
u(x,0) g(x)= (1.26)
for every dx ∈R , where
L G∂
≡ + Ψ −∂τ
(1.27)
and G is the generator of the state-variable process defined by
d
2i
i 1 i
1 fGf f x
2 x=
∂≡ ∇ − θ
∂∑ (1.28)
for every f ∈ ( )2 dC R . Furthermore, assume that there exists positive constants K and h,
with 1
h8 d
<ρ
and ρ defined by (1.23), such that for every dx ∈R we have
( ) ( )2
0 Tmax u x, Kexp h x
≤τ≤τ ≤ . (1.29)
Then, u is uniquely represented by
( ) ( ) ( )xP s0
u x, E g X exp dsτ
τ τ = Ψ ∫ . (1.30)
proof : Let ( ) dx, (0,T]τ ∈ ×x R with 0X x= almost surely. We will apply ˆIto's formula
to the process s 0 sV
≤ ≤ τ defined by
( )s s sV u X , s≡ τ − E . (1.31)
where
( )s
s q0exp dq≡ Ψ∫E (1.32)
It follows that for every [ ]s 0,∈ τ we have
( ) ( ) ( )i 2s 0 q q q q q q
i
d s s
0 0i 1
V V u X , q dX u X , q dqx q
=
∂ ∂− = τ − + ∇ + + Ψ τ − ∂ ∂ ∑∫ ∫E E . (1.33)
Recalling that tX satisfies t t tdX dW X dt= − θ , we see that the first term becomes
36
( ) ( ) ( ) ( )i iq q q q q q
i i
d s s
0 0i 1
u X , q dW X u X , q dqx x
=
∂ ∂τ − − θ τ − ∂ ∂ ∑ ∫ ∫E E . (1.34)
Combining (1.25), (1.27), (1.28), (1.33), and (1.34) we deduce that
( ) ( )is 0 q q q
i
d s
0i 1
V V u X , q dWx
=
∂− = τ −∂∑∫ E (1.35)
for every [ ]s 0,∈ τ . Moreover, the expectation of the RHS of (1.35) is zero, hence
( ) ( ) ( )( )x x0 P s P s su x, V E V E u X , sτ = = = τ − E . (1.36)
Let n tS inf t 0 : Y n d≡ ≥ ≥ for every n ≥ 1, and fix ( )m 0,∈ τ . Then,
( )u x,τ ( )( )n n
xP S m n S mE u X , S m∧ ∧= τ − ∧ E (1.37)
( ) [ ]( ) ( ) [ ]( )n nn n
x xP m m P S n SS m S mE u X , m 1 E u X , S 1> ≤= τ − + τ −E E .
As n → ∞ and m ↑ τ , we see that the first term on the RHS of (1.37) approaches
( )( ) ( ) ( )x xP P s0
E u X ,0 E g X exp dsτ
τ τ τ = Ψ ∫E . (1.38)
From (1.2), (1.24), and (1.29), we see that the second term is dominated by
( ) [ ]( )n n n
xP S n S S mE u X , S 1 ≤τ − E ( ) [ ]( )n n
xP S n S mE u X , S 1 ≤≤ γ τ − (1.39)
( ) [ ]( )n n
2xP S S TKE exp h X 1 ≤≤ γ ( ) ( )2 x
nKexp hdn P S T≤ γ ≤
( ) ( )( )di2 x
t0 t Ti 1
Kexp hdn P max Y n≤ ≤
=
≤ γ ≥∑
( ) ( )2d 1
2 2
i 1
nKexp hdn 3 2 exp
8−
=
≤ γ πρ − ρ
∑
( )1
221 8 hd
3 Kd 2 exp n8
− − ρ≤ γ πρ − ρ
,
which approaches zero as n → ∞ since 1 8h d 0− ρ > . Therefore, by combining (1.38)
and (1.39) with (1.37) we deduce the desired result (1.30). <
37
The main goal of this chapter is to construct a unique solution ( )2,1 du C [0,T]∈ ×R to
this Cauchy problem that satisfies (1.29). The next section introduces the basic
terminology inherent in the potential theoretic parametrix method. In Sections 3 and 4,
the relevant potentials are studied and technical differentiability results are provided. The
fifth section provides a series representation for the fundamental solution of the Cauchy
problem and establishes some continuity results. Finally, the chapter concludes by
constructing the solution to the Cauchy problem in Section 6.
2.2 The Potential Theoretic Parametrix Method
2.2.1 The Gaussian Semigroup The solution presented in this section updates the parametrix method (Friedman,
1964) by using the modern theories of potentials and semigroups. The canonical family
of semigroups associated with the differential operator L is the Gaussian family
Z : 0 T, 0δτ < τ ≤ δ > defined by
( )d 22 x
Z x exp2 2
δτ
δδ ≡ − πτ τ (2.1)
for every ( ) dx, (0,T]τ ∈ ×R and 0δ > .2 We will omit the subscript τ when we wish to
refer to the Gaussian semigroup as a function on d (0,T]×R .
For each (0,T]τ∈ , we define the operator [ ]Zδτ ⋅ by
[ ]( ) ( ) ( )dZ f x Z f (x) Z (x y)f y dyδ δ δ
τ τ τ≡ ∗ ≡ −∫R (2.2)
for every ( )df C∈ R for which this convolution exists. Because of the bound (1.29) , it is
important that the convolution is finite for the following subset of ( )dC R :
2 The superscript δ will be dropped for the standard Gaussian semigroup (i.e. δ = 1).
38
Definition 1 – For every positive δ, let
( ) ( )2df C : f Aexp h for some positive h and A 032 dδ
δ≡ ∈ ≤ ⋅ < >
ρ RA (2.3)
denote the set of δ-admissible functions and denote 1A by A , where
( )2 T0
T 1max , X e 1 .
4d 2θ ρ ≡ + − θ
Also, the bound parameter of a δ-admissible function
f is defined by
( )2
fh inf 0 h : f Aexp h for some A 032 d
δ≡ < < ≤ ⋅ >
ρ . (2.4)
It follows that the initial function g appearing in our Cauchy problem (1.26) is δ-
admissible for some δ depending on α . The next result gives an example of a class of
admissible functions for which the convolution in (2.2) is finite.
Lemma 2 – Let (0,T]τ∈ , 0 2< ε ≤ δ , and f ε∈A . Then [ ] 2Z fδτ ε∈A . More precisely,
for every dx ∈R we have that
[ ]( )Z f xδτ ( ) ( )2
fd
exp h Z x dδτ≤ −∫R
ξ ξ ξ (2.5)
( )d2 2 2f
ff f
hexp x Cexp h x
2 h 2 h
δ δ= ≤ δ − τ δ − τ % ,
where
d2
f
C2Th
δ≡ δ − and f
f ff
hh 2h
2Th
δ≡ <
δ −% .
proof: We first note that since Zδτ and f are continuous functions on dR , it follows
from (2.5) that [ ] ( )dZ f Cδτ ∈ R . Therefore, to prove that [ ] 2Z fδ
τ ε∈A , it suffices to
verify that (2.5) holds. After completing the square, the integrand becomes
d 22 2f f
f f
h ( 2 h )exp x exp x
2 2 h 2 2 h
δ δ δ − τ δ − − πτ δ − τ τ δ − τ ξ . (2.6)
Since f 2h δ∈A we have f2 h 0δ − τ > , thus we see that (2.5) holds for every dx ∈R . <
39
The final property of the family of Gaussian semigroups that we will review is
continuity in time. Clearly, for every dx ∈R and f ε∈A with 0 2< ε ≤ δ , the maps
( )Z xδττ a and [ ]( )Z f xδ
ττ a are continuous. Furthermore, we may continuously
extend the action of the Gaussian semigroup by defining [ ]( ) [ ]( )00
Z f x limZ f x f(x)δ δτ
τ→≡ = .
In fact, this follows from the weak convergence of the measure defined by the Gaussian
semigroup to the point mass measure as 0τ → .
2.2.2 The Fundamental Solution
We begin this section by noting that for a given function ( )dF C (0,T]∈ ×R , we will
refer to the family (0,T]
Fτ τ∈ as a continuous family of functions from ( )dC R , where
( ) ( )F x F x,τ ≡ τ for every ( ) dx, (0,T]τ ∈ ×R .
Definition 3 – A fundamental solution for L is a continuous family ( ]0,Tτ τ∈Γ of
functions from ( )dC R such that for every (0,T]τ∈ and f ∈A we have:
(i) [ ]( ) ( )L f L f 0τ τΓ ≡ Γ ∗ =
(ii) [ ] ( )00
f lim f fττ→
Γ ≡ Γ ∗ = .
It will be shown that the solution to the Cauchy problem given by (1.25) and (1.26)
may be represented by [ ]( )u(x, ) g xττ = Γ , (2.7)
where the fundamental solution τΓ shall be constructed using the following parametrix
method. Consider the following Cauchy problem for a given f ∈A :
pL u(x, ) 0τ = (2.8)
u(x,0) f(x)= , (2.9)
where the principal part pL of the differential operator L is defined by
2p
1L
2
∂≡ ∇ −
∂τ. (2.10)
40
It follows from the Feynman-Kac Theorem that the solution to this simple heat equation
may be represented by
( ) [ ]u x, Z fττ = . (2.11)
A fundamental solution for pL will be known as a parametrix associated with L. In
our case, it follows that the Gaussian semigroup Zτ is a parametrix for L. This result will
be used to motivate an expression for a fundamental solution for L that satisfies (2.7).
Definition 4 – The Gaussian potential is a family of operators 0 TU τ ≤τ≤
defined by
[ ] [ ]s0U f Z f ds
τ
τ τ−≡ ∫ (2.12)
for every ( )df C∈ R for which this integral is defined. For the continuous family
0 Tf fτ <τ≤
≡ of functions from ( )dC R , we define
[ ] [ ]s s0U f Z f ds
τ
τ τ−≡ ∫ . (2.13)
As with the Gaussian semigroup, we will omit the subscript τ when we wish to refer to
the Gaussian potential as a function on d (0,T]×R .
Recall that our state-variable process satisfies
t t tdX dW X dt= − θ . (2.14)
Hence, if θ and Ψ are identically zero, then it follows from (1.5) and (2.14) that
PL L= . This inspires us to look for a fundamental solution for L that satisfies (2.7) in
terms of the Gaussian potential of some continuous family 0 Tτ <τ≤ϕ ≡ ϕ of functions
from ( )dC R , as follows:
[ ]Z Uτ τ τΓ = + ϕ . (2.15)
The function ϕ will be determined by condition (i) of Definition 3. This condition
implies that for every (0,T]τ∈ and f ∈A we have
[ ]( ) [ ]( ) [ ][ ]( ) ( ) [ ]( )0 L f L Z f L U f L Z f L U fτ τ τ τ τ= Γ = + ϕ ≡ ∗ + ϕ ∗ . (2.16)
41
Consequently, we are led to an investigation of the differential properties of the
Gaussian semigroup and the Gaussian potential. In particular, we will show in Section 3
that the differential operator L may be passed through the convolution
[ ]Z fτ . More explicitly, for every ( ) dx, (0,T]τ ∈ ×R and f ∈A we will show that
[ ]( )( ) ( ) ( )dL Z f x LZ x f dτ τ= −∫R
ξ ξ ξ . (2.17)
In Section 4, we will derive sufficient conditions on ϕ so that the differential operator
L may be passed through the convolution [ ][ ]LU fτ ϕ . More precisely, it will be shown
that [ ]U τ ϕ is well-defined provided that there exists positive constants C and δ
with 1δ ≤ such that for every ( ) dx, (0,T]τ ∈ ×R we have
( ) ( )x CZ xδτ τϕ ≤ . (2.18)
In this case, we shall say that ϕ is Z -boundedδ . Furthermore, if ϕ is locally, uniformly
Hölder continuous in space, then we will prove that
[ ][ ]( ) [ ][ ]( ) [ ]s s0L U f L Z f ds f
τ
τ τ− τϕ = ϕ − ϕ∫ (2.19)
for every (0,T]τ∈ and f ∈A , where
[ ]f fτ τϕ ≡ ϕ ∗ (2.20)
and
[ ][ ]( ) [ ]( )s s s sL Z f L Z fτ− τ−ϕ ≡ ϕ ∗ . (2.21)
In Section 5, we will construct a series representation for a function ϕ satisfying
(2.16) and prove that it satisfies the conditions above, which have been asserted to be
sufficient for ϕ to satisfy (2.19). We will define ϕ in terms of the following potential:
42
Definition 5 – The L-potential is a family of operators 0 TVτ ≤τ≤
defined by
[ ] [ ]( )s0V f L Z f ds
τ
τ τ−≡ ∫ (2.22)
for every ( )df C∈ R for which this integral is defined. For the continuous family
0 Tf fτ <τ≤
≡ of functions from ( )dC R , we define
[ ] [ ]( )s s0V f L Z f ds
τ
τ τ−≡ ∫ . (2.23)
As with the Gaussian Potential, we will omit the subscript τ when we wish to refer to
the L-potential as a function on d (0,T]×R . It follows that (2.19) may be rewritten as
[ ][ ]( ) [ ][ ] [ ]L U f V f fτ τ τϕ = ϕ − ϕ , (2.24)
where [ ][ ] [ ]V f V fτ τϕ ≡ ϕ ∗ . (2.25)
Combining (2.24) with (2.16), we deduce that ϕ must satisfy the following Volterra
integral equation for every (0,T]τ∈ and f ∈A :
[ ] [ ]( ) [ ][ ]f L Z f V fτ τ τϕ = + ϕ . (2.26)
It is shown in Section 5 that the solution of (2.26) may be represented by the series
[ ]m
m 0
V LZ∞
τ τ=
ϕ = ∑ , (2.27)
where mVτ is an operator defined by
0V Iτ ≡ (2.28) and m 1 mV V V+
τ τ τ≡ o (2.29)
for every (0,T]τ∈ and m 0≥ .
43
2.3 Preliminary Technical Results 2.3.1 Differentiability of the Gaussian Semigroup In this sub-section we begin presenting the technical results that are required to
complete the construction of a fundamental solution as outlined in Section 2. The main
result is that for every ( ) dx, (0,T]τ ∈ ×R and f ∈A we have
[ ]( )( ) ( ) ( )dL Z f x LZ x f dτ τ= −∫R
ξ ξ ξ . (3.1)
This will require a few preliminary results.
Lemma 1 – Let ( ) dx, (0,T]τ ∈ ×R , and let n, m, ε, and δ be positive constants with
ε < δ . For every constant µ with 2m n j
02
− +≤ µ ≤ , there exists a positive constant A
such that for j = 0,1, and 2 we have
( )n j
2m n j 2m ji
x Z A(x) Z x
x x
δεττ− + − µµ
∂≤
τ ∂ τ (3.2)
Furthermore, for every constant µ with 2m n 2
02
− +≤ µ ≤ , there exists a positive
constant B such that
( )n
2m n 2 2m
x Z B(x) Z x
x
δεττ− + − µµ
∂≤
τ ∂τ τ. (3.3)
proof : Let n, m, δ, and ε be positive with ε < δ and define ( )2
xW x,
2τ ≡
τ. Then for
every µ with 2m n
02
−≤ µ ≤ there is a positive constant A such that
44
n
m
xZ (x)δ
ττ
d 22nm x
x exp2 2
− δδ = τ − πτ τ
(3.4)
( ) ( )( )
( )dm2
2m n 2
2W(x, ) exp W(x, )exp W(x, )
2x
−µ
− − µµ
τ − δ − ε τ δ = −ε τ πτ τ
( )2m n 2
AZ x
xετ− − µµ
≤τ
.
This establishes (3.2) for j 0= . For j 1= we deduce from (3.4) that
( )n
mi
x Zx
x
δτ∂
τ ∂ ( ) ( )
n n 1
im m 1
x xxZ x Z x
+δ δτ τ+
δ= ≤ δ
τ τ τ (3.5)
( )2m n 1 2
AZ x
xετ− + − µµ
≤τ
%
holds for every 2m n 1
02
− +≤ µ ≤ . Similarly,
( )n 2
m 2i
x Zx
x
δτ∂
τ ∂ ( ) ( )
n n 2
im mi
x x xZZ x x 1 Z x
x
δδ δττ τ
δ δ ∂ δ= + ≤ − τ τ ∂ τ τ τ (3.6)
( )n 2 n
m 2 m 1
x xZ x
+δτ+ +
≤ δ δ +
τ τ ( )2m n 2 2
AZ x
xετ− + − µµ
≤τ
holds for every 2m n 2
02
− +≤ µ ≤ , which implies (3.2) for j = 2. Finally, we obtain
(3.3) from the following inequality:
( )n
m
x Zx
δτ∂
τ ∂τ ( ) ( )
n 2 n 2 n
m 2 m 2 m 1
x x x d xdZ x Z x
2 2 2 2
+δ δτ τ+ +
δ δ= − ≤ +
τ τ τ τ τ (3.7)
( ) ( )n 2 n
d 2m n 2 2m 2 m 1
T x d x AZ x Z x
2 2 x
+δ ετ τ+ − + − µ+ + µ
δ≤ + ≤
τ τ τ . <
We will now apply this lemma to obtain a result on the differentiability of the
Gaussian semigroup.
45
Theorem 2 – For every 0δ > and f δ∈A we have [ ] ( )2,1 dZ f C (0,T]δ ∈ ×R . In fact,
[ ]( ) ( )j j
j ji i
dZ f x Z x f ( )d
x xδ δτ τ
∂ ∂= −
∂ ∂∫Rξ ξ ξ (3.8)
and
[ ]( ) ( )d
Z f x Z x f ( )dδ δτ τ
∂ ∂= −
∂τ ∂τ∫Rξ ξ ξ (3.9)
for every ( ) dx, (0,T]τ ∈ ×R and j 1,2∈ . Furthermore, [ ]j
ji
Z fx
δτ
∂∂
and [ ]Z fδτ
∂∂τ
are
functions in 2δA for every (0,T]τ∈ .
proof: Let 0δ > and f δ∈A . Clearly, the integrands above are continuous functions
of ( )x,τ on d (0,T]×R . Hence, to prove assertions (3.8) and (3.9) as well as show that
[ ] ( )2,1 dZ f C (0,T]δ ∈ ×R , it suffices to show that the integrals in these assertions are
locally, uniformly bounded with respect to ( ) dx, (0,T]τ ∈ ×R .
Let f δ∈A , j 1,2∈ , and dB : 1≡ ∈ <Rξ ξ . From (3.8) we obtain
( ) ( ) ( )C
j j j
j j ji i i
d B BZ x f ( ) d Z x f ( ) d Z x f ( ) d
x x xδ δ δτ τ τ
∂ ∂ ∂− ≤ − + −∂ ∂ ∂∫ ∫ ∫R
ξ ξ ξ ξ ξ ξ ξ ξ ξ . (3.10)
Clearly, the first integral on the RHS of (3.10) is bounded uniformly in
( ) dx, (0,T]τ ∈ ×R . For the second integral, consider a compact subset dM ⊆ R and let
( )x, M (0,T]τ ∈ × . We deduce from Lemma 1 with 0µ = that for every positive ε < δ
there exists a positive constant A such that
( )C
j
jiBZ x f ( )d
xδτ
∂ −∂∫ ξ ξ ξ ( ) ( )
C
j 2
fB
A x Z x exp h x d− ε
τ≤ − −∫ ξ ξ ξ (3.11)
( ) ( )C
2
fB
A exp h x Z x dετ≤ −∫ ξ ξ .
Since we may choose ε arbitrarily close to δ, we may assume that f ε∈A . It follows
from Lemma 2.2 that there exists positive constants C and f fh 2h<% such that
46
( ) ( ) ( ) ( )C
2 2 2f f f
Bexp h x Z x d Cexp h x Cexp h Nε
τ − ≤ ≤∫ % %ξ ξ , (3.12)
where ( )N diam M= .
By combining (3.10), (3.11), and (3.12), we deduce that the integral in (3.8) is locally,
uniformly bounded with respect to ( ) dx, (0,T]τ ∈ ×R . Moreover, this result may also be
established for the integral in (3.9) by repeating the proof using (3.3) to obtain the
analogue of (3.11). Consequently, we have shown that assertions (3.8) and (3.9) hold for
every ( ) dx, (0,T]τ ∈ ×R and have proven that [ ] ( )2,1 dZ f C (0,T]δ ∈ ×R . Finally, it
follows from (3.12) that [ ]j
ji
Z fx
δτ
∂∂
and [ ]Z fδτ
∂∂τ
are in 2δA for every (0,T]τ∈ . <
We proceed to prove the main result of this sub-section.
Theorem 3 – For every f ∈A we have [ ] ( )dLZ f C (0,T]∈ ×R . In fact,
[ ]( )( ) ( )dL Z f x LZ x f ( )dτ τ= −∫R
ξ ξ ξ (3.13)
for every ( ) dx, (0,T]τ ∈ ×R . Furthermore, [ ] 2LZ fτ ∈A for each (0,T]τ∈ .
Remark: The differential operator L acts in the variables ( ) dx, (0,T]τ ∈ ×R .
proof: Clearly, the integrand in (3.13) is continuous in ( ) dx, (0,T]τ ∈ ×R . Hence, to
prove (3.13) as well as show that [ ] ( )dZ f C (0,T]δ ∈ ×R , it suffices to show that the
integral in (3.13) is locally, uniformly bounded in ( ) dx, (0,T]τ ∈ ×R .
We begin by proving that for every compact dM ⊆ R there is a C 0>% such that
( ) ( )1
2LZ x C Z x−
δτ τ− ≤ τ −%ξ ξ (3.14)
for every ( )x, M (0,T]τ ∈ × and d∈Rξ . Let dM ⊆ R be compact with diameter N,
( )x, M (0,T]τ ∈ × , and d∈Rξ . It follows from Lemma 1 with 1
2µ = that there is a
C 0> such that for every positive 1δ < we have
47
LZ (x )τ − ξ ( )d
ii 1 i
x Z (x ) x Z (x )x
τ τ=
∂≤ θ − + Ψ −∂∑ ξ ξ (3.15)
d
i 1 i
N Z (x ) Z (x )x
τ τ=
∂≤ θ − + γ −∂∑ ξ ξ
( ) ( ) ( )1 1
2 2C Z x Z x , C Z x− −δ δ δ
τ τ τ≤ τ − + γ − τ ≤ τ −%ξ ξ ξ .
Now, let ( )x, M [q,T]τ ∈ × for some q 0> and 1δ < be a positive constant such that
fh32 d
δ<
ρ. It follows from (3.15) and Lemma 2.2 that there exists positive constants 1C ,
2C , and f fh 2h<% such that
( ) ( )d
LZ x f dτ −∫Rξ ξ ξ ( ) ( )
122
1 fd
C Z x exp h d−
δτ≤ τ −∫R
ξ ξ ξ (3.16)
( ) ( )1 1
2 22 22 f 2 fC exp h x C q exp h N
− −≤ τ ≤% % .
We deduce that the integral in (3.13) is locally, uniformly bounded with respect to
( ) dx, (0,T]τ ∈ ×R from which we conclude that [ ]( ) ( )dL Z f C (0,T]∈ ×R and (3.13)
holds for every ( ) dx, (0,T]τ ∈ ×R . Also, the bound (3.16) implies that [ ]( ) 2L Z fτ ∈A
for each (0,T]τ∈ . <
This result is the first step towards the construction of the fundamental solution of our
Cauchy problem. Recall that we wish to express the fundamental solution in the form
[ ]Z Uτ τ τΓ = + ϕ (3.17)
for some continuous family 0 Tτ <τ≤ϕ ≡ ϕ of functions from ( )dC R that satisfies
[ ]( ) [ ]( ) [ ][ ]( )0 L f L Z f L U fτ τ τ= Γ = + ϕ (3.18)
for every (0,T]τ∈ and f ∈A . Theorem 3 allows us to pass the differential operator L
through the convolution in the first term on the RHS of (3.18).
48
2.3.2 Basic Potential Theory Before turning our attention to the second term on the RHS of (3.18), we will
determine two classes of functions for which the Gaussian and L-potentials are defined.
Naturally, the class of admissible functions is a good place to start.
Definition 4 – For every 0δ > , the bound parameter of a family 0 Tf fτ <τ≤
≡ of
functions from δA is defined by f f0 T
h sup h32 dτ
<τ≤
δ≡ ≤
ρ. Furthermore, the family f is said
to be uniformly bounded if fh32 d
δ<
ρ.
Theorem 5 – Let 2ε ≤ and 0 Tf fτ <τ≤
≡ be a continuous, uniformly bounded family of
functions from εA . Then, [ ]U f and [ ]V f are functions in ( )dC (0,T]×R . Also, for
each (0,T]τ∈ we have that [ ]U fτ and [ ]V fτ are elements of 2εA .
proof: Recall from Definitions 2.4 and 2.5 that for every ( ) dx, (0,T]τ ∈ ×R we have
[ ]( ) [ ]( )s s0U f x Z f x ds
τ
τ τ−≡ ∫ (3.19)
and
[ ]( ) [ ]( )( )s s0V f x L Z f x ds
τ
τ τ−≡ ∫ . (3.20)
From Lemma 2.2 and Theorem 3 we see that the integrands in (3.19) and (3.20) are
continuous for every ( ) dx, ,s (0,T] (0, )τ ∈ × × τx R . Therefore, the desired assertions
follow from Lemma 2.2 and (3.16). In fact, we have that
[ ]( ) [ ]( ) ( ) ( )2 2
s s f f0 0U f x Z f x ds Cexp h x ds CTexp h x
τ τ
τ τ−≤ ≤ ≤∫ ∫ % % (3.21)
and
[ ]( )V f xτ [ ]( )( ) ( ) ( )1
22
s s f0 0L Z f x ds Cexp h x s ds
τ τ −τ−≤ ≤ τ −∫ ∫%% (3.22)
( )2
f2C Texp h x≤ %%
for every ( ) dx, (0,T]τ ∈ ×R , where C 0> , C 0>% , and f fh 2h<% . <
49
We will use this result in conjunction with the following class of functions:
Definition 6 – Let 0δ > . A family 0 Tτ <τ≤ψ ≡ ψ of functions from ( )dC R is called
Z -boundedδ , if there exists a positive constant C such that
( ) ( )x CZ xδτ τϕ ≤ (3.23)
for every ( ) dx, (0,T]τ ∈ ×R .
Consider a continuous family 0 Tτ <τ≤ψ ≡ ψ of Z - boundedδ functions and an
admissible function f δ∈A . It follows from Lemma 2.2 that
[ ]fτψ ( ) ( )d
x f dτ≤ ψ −∫Rξ ξ ξ (3.24)
( ) ( ) ( )2 2
f fd
C Z x exp h d Cexp h xδτ≤ − ≤∫R
%%ξ ξ ξ ,
where C 0> and f fh 2h<% . Hence, [ ] 0 T
fτ <τ≤ψ is a continuous, uniformly bounded
family of functions from 2δA . In particular, it follows from Theorem 5 that
[ ] ( )dU f C (0,T] ψ ∈ × R and [ ] 2U fτ δ ψ ∈ A for each (0,T]τ∈ and 1δ ≤ .
The next goal is to show that ψ is in the domain of the Gaussian potential and
combine our results to obtain
[ ] [ ][ ]U f U fτ τ ψ = ψ (3.25)
for every (0,T]τ∈ . The following lemma is the first step towards this goal and
essentially follows from the semigroup property of the Gaussian kernel.
Lemma 7 – For every a and b with a,b 1−∞ < < and positive δ we have
( ) ( ) ( ) ( )a b 1 a bs s
d0( s) s Z x Z d ds B 1 a,1 b Z x− − δ δ − − δ
τ− ττ − − = − − ττ
∫ ∫Rd ξ ξ ξ . (3.26)
where B denotes the beta function.
50
proof : We begin by substituting w in place of y where
i i i
sw x
2( s s 2( s
δτ δ≡ +τ − ) τ − )τ
ξ . (3.27)
After some algebraic manipulation, it is easy to verify that
( )
2 2 22x x
w2 s 2s 2
δ − δ δ+ = +
τ − τξ ξ
. (3.28)
Denoting the integral in (3.26) by I, we obtain the desired result
I ( ) ( )d
2d d d2 a b2 2 2
d 0
x 2s sexp exp w dw ( s) s ds
2 2
− − − −τ δ τ − δ = − − τ − π τ δτ ∫∫R
(3.29)
( ) ( )a b 1 a b a b
0 0Z x ( s) s ds Z x (1 ) dδ − − − − δ − −
τ τ
τ 1= τ − = τ − ρ ρ ρ∫ ∫
( ) ( )1 a bB 1 a,1 b Z x− − δτ= − − τ ,
where the substitution s
ρ =τ
was used in the second equality. <
Theorem 8 – Let 0 1< δ ≤ and 0 Tτ <τ≤ψ ≡ ψ be a continuous family of Z - boundedδ
functions. Then, [ ]U τ ψ is well-defined for every (0,T]τ∈ . In fact, [ ] 0 T
Uτ <τ≤ψ is a
continuous family of Z - boundedδ functions.
proof: From Lemma 7, it follows that there exists a positive constant C such that
[ ]Uτ ψ [ ] ( ) ( )s s s sd0 0
Z ds Z x d dsτ− τ−
τ τ≤ ψ ≤ − ψ∫ ∫ ∫R
ξ ξ ξ (3.30)
( ) ( ) ( )s sd0
C Z x Z d ds CTZ xδ δ δτ− τ
τ≤ − ≤∫ ∫R
ξ ξ ξ
for every (0,T]τ∈ . <
This result will now be used to establish (3.25).
51
Corollary 9 – Let 0 1< δ ≤ , f δ∈A , and 0 Tτ <τ≤ψ ≡ ψ be a continuous family of
Z - boundedδ functions. Then, for every (0,T]τ∈ we have
[ ] [ ][ ]U f U fτ τ ψ = ψ . (3.31)
proof: It follows from (3.24) that [ ] 0 T
fτ <τ≤ψ is a continuous, uniformly bounded
family of functions from 2 2δ ⊆A A . Consequently, we deduce from Theorem 5 that the
LHS of (3.31) is well-defined. In fact, [ ] ( )dU f C (0,T] ψ ∈ × R and [ ] 2U fτ δ ψ ∈ A
for each (0,T]τ∈ . Fix ( )x, (0,T]τ ∈ . From Theorem 8 we have that [ ]U τ ψ is well-
defined. Furthermore, we apply Fubini’s theorem to obtain
[ ][ ]( )U f xτ ψ [ ] ( ) [ ]( )s s0U f x Z ds f(x)
τ
τ τ−= ψ ∗ = ψ ∗∫ (3.32)
[ ]( ) ( )d s s0Z x f dsd
τ
τ−= ψ −∫ ∫Rξ ξ ξ
( ) ( ) ( )d d s s0Z x y y f dydsd
τ
τ−= − − ψ∫ ∫ ∫R Rξ ξ ξ
( ) ( ) ( )d d s s0Z x y y f dyd ds
τ
τ−= − − ψ∫ ∫ ∫R Rξ ξ ξ
( ) ( ) ( )d d s s0Z x w w f dwd ds
τ
τ−= − ψ −∫ ∫ ∫R Rξ ξ ξ
( ) ( ) ( )( )d ds s0Z x w w f d dwds
τ
τ−= − ψ −∫ ∫ ∫R R ξ ξ ξ
( ) [ ]d s s0Z x w f dwds
τ
τ−= − ψ∫ ∫R
[ ] ( ) [ ] ( )s s0Z f x ds U f x
τ
τ− τ = ψ = ψ ∫ . <
The final objective of this section is to prove that continuous, Z - boundedδ
functions are in the domain of the L-potential for every 1δ ≤ . In particular, the analogue
of (3.25) for the L-potential will be established. We will first prove a result for
continuous, Z - boundedδ functions that is similar to Theorem 3.
52
Lemma 10 – Let 0 1< δ ≤ and 0 Tτ <τ≤ψ ≡ ψ be a continuous family of Z - boundedδ
functions. Then, [ ]( ) ( )dL Z C (0,T]ψ ∈ ×R . In fact, for every ( ) dx, (0,T]τ ∈ ×R we have
[ ]( )( ) ( )dL Z x LZ x ( )dτ τ τψ = − ψ∫R
ξ ξ ξ . (3.33)
proof : We proceed as in the proof of Theorem 3. Clearly, the integrand in (3.33) is a
continuous function of ( )x,τ on d (0,T]×R . Hence, to prove (3.33) as well as show that
[ ]( ) ( )dL Z C (0,T]ψ ∈ ×R , it suffices to show that the integral in (3.33) is locally,
uniformly bounded with respect to ( ) dx, (0,T]τ ∈ ×R .
Let ( )x, M [q,T]τ ∈ × for some q 0> and compact dM ⊆ R . It follows from (3.29)
and (3.15) that there exists a positive constant C such that
( ) ( ) ( )d d
1 1 1
2 2 2LZ x ( ) d C Z x Z ( )d C Z x Cq C− − −δ δ δ
τ τ τ τ τ− ψ ≤ τ − = τ ≤∫ ∫R R%ξ ξ ξ ξ ξ ξ , (3.34)
where ( ) ( ) [ ] C sup Z x : x, M q,Tδτ≡ τ ∈ ×% . We deduce that the integral in (3.33) is
locally, uniformly bounded with respect to ( ) dx, (0,T]τ ∈ ×R from which we obtain the
desired conclusions. <
This lemma allows us to prove the following L-potential analogue of Theorem 8.
Theorem 11 – Let 0 1< δ ≤ and 0 Tτ <τ≤ψ ≡ ψ be a continuous family of Z - boundedδ
functions. Then, [ ]Vτ ψ is well-defined for every (0,T]τ∈ . In fact, [ ] 0 T
Vτ <τ≤ψ is a
continuous family of Z - boundedδ functions.
proof: From (3.15) and Lemmas 10 and 7, it follows that there exists a C 0> such that
[ ] [ ]( ) ( ) ( )ds s s s0 0V L Z ds LZ x d ds
τ τ
τ τ− τ−ψ ≤ ψ ≤ − ψ∫ ∫ ∫Rξ ξ ξ (3.35)
( ) ( ) ( )d
1
2s s0
C s Z x Z d dsτ − δ δ
τ−≤ τ − −∫ ∫R ξ ξ ξ
( ) ( )1CB 1, Z x CZ x
2δ δτ τ
= τ ≤
%
for every (0,T]τ∈ . <
53
The final result in this section is the L-potential analogue of Corollary 9.
Corollary 12 – Let 0 1< δ ≤ , f δ∈A , and 0 Tτ <τ≤ψ ≡ ψ be a continuous family of
Z - boundedδ functions. Then, for every (0,T]τ∈ we have
[ ] [ ][ ]V f V fτ τ ψ = ψ . (3.36)
proof: We proceed as in the proof of Corollary 9. It follows from (3.24) that
[ ] 0 T
fτ <τ≤ψ is a continuous, uniformly bounded family of functions from 2 2δ ⊆A A .
Consequently, we deduce from Theorem 5 that the LHS of (3.31) is well-defined. In fact,
[ ] ( )dV f C (0,T] ψ ∈ × R and [ ] 2V fτ δ ψ ∈ A for each (0,T]τ∈ . Fix ( )x, (0,T]τ ∈ .
From Theorem 11 we have that [ ]Vτ ψ is well-defined. Furthermore, we apply Fubini’s
theorem to obtain
[ ][ ]( )V f xτ ψ [ ] ( ) [ ]( )( )s s0V f x L Z ds f(x)
τ
τ τ−= ψ ∗ = ψ ∗∫ (3.37)
[ ]( )( ) ( )d s s0L Z x f dsd
τ
τ−= ψ −∫ ∫Rξ ξ ξ
( ) ( ) ( )d d s s0LZ x y y f dydsd
τ
τ−= − − ψ∫ ∫ ∫R Rξ ξ ξ
( ) ( ) ( )d d s s0LZ x y y f dyd ds
τ
τ−= − − ψ∫ ∫ ∫R Rξ ξ ξ
( ) ( ) ( )d d s s0LZ x w w f dwd ds
τ
τ−= − ψ −∫ ∫ ∫R Rξ ξ ξ
( ) ( ) ( )( )d ds s0LZ x w w f d dwds
τ
τ−= − ψ −∫ ∫ ∫R R ξ ξ ξ
( ) [ ]d s s0LZ x w f dwds
τ
τ−= − ψ∫ ∫R
[ ]( )( ) [ ] ( )s s0L Z f x ds V f x
τ
τ− τ = ψ = ψ ∫ . <
54
2.4 The Derivatives of the Gaussian Potential
In this section we turn our attention to the second term on the RHS of (3.19). In
particular, we will derive the following relationship between the Gaussian and L-
potentials expressed in (2.24):
[ ]( ) [ ] [ ]L U f V f fτ τ τ ψ = ψ − ψ (4.1)
for every f δ∈A , 1δ ≤ , and continuous family of Zδ -bounded functions 0 Tτ <τ≤ψ ≡ ψ
for which [ ]fψ is locally, uniformly Hölder continuous in space with exponent 1β < .
Theorem 1 – Let 1δ ≤ and (0,T]
f fτ τ∈= be a uniformly bounded, continuous family of
functions from δA that is locally, uniformly Hölder continuous in space with exponent
1β < . Then, for j 1,2∈ we have that [ ] ( )j
dji
U f C (0,T]x
∂∈ ×
∂R . In fact,
[ ] [ ]j j
s sj ji i0U f Z f ds
x xτ τ−
τ∂ ∂=∂ ∂∫ (4.2)
and [ ]j
2ji
U fx τ δ∂
∈∂
A for every (0,T]τ∈ .
proof: From Theorem 3.2 we have that ( ) [ ]( )j
s sji
x, ,s Z f xx τ−∂
τ∂
a is continuous on
( ) dE x, ,s : x ,0 s T≡ τ ∈ < < τ ≤R ; However, there is a singularity at s = τ . Hence, to
prove (4.2) as well as show that [ ] ( )j
dji
U f C (0,T]x
∂∈ ×
∂R , it suffices to show that the
integral in (4.2) is locally, uniformly bounded in ( ) dx, (0,T]τ ∈ ×R . We break up the
integration in (4.2) as follows:
[ ] [ ] [ ]j j j
s s s s s sj j ji i i
q
0 0 qZ f ds Z f ds Z f ds
x x xτ− τ− τ−
τ τ∂ ∂ ∂= +∂ ∂ ∂∫ ∫ ∫ , (4.3)
where q (0,T]∈ is fixed.
55
Let ( ) ( ) qx, ,s E x, ,s E : x M, s qτ ∈ ≡ τ ∈ ∈ ≤ < τ where dM ⊆ R is compact. For
the first integral on the RHS of (4.3), we apply Theorem 3.2, Lemma 3.1, and Lemma 2.2
to obtain the estimate
[ ]( )j
s sji
Z f xx τ−∂∂
( ) ( )d
j
s sji
Z x f dx τ−∂≤ −∂∫R
ξ ξ ξ (4.4)
( ) ( ) ( )d
j2
21 f sC s exp h Z x d
− ετ−≤ τ − −∫R
ξ ξ ξ
( )( )j
22
2 fC exp h x s−≤ τ −% ,
where f fh 2h<% . This implies that
[ ]( ) ( ) ( ) ( )j j
2 22
s s 2 f 3 fji
q q
0 0Z f x ds C exp h x s ds C exp h x
x−
τ−∂ ≤ τ − ≤∂∫ ∫% % . (4.5)
Now, let ( ) ( ) qx, ,s E x, ,s E : x M,0 q sτ ∈ ≡ τ ∈ ∈ ≤ < < τ% . For the second integral
on the RHS of (4.3), we define dB : 1≡ ∈ <Rξ ξ and rewrite (4.4) as
[ ]( )j
s sji
Z f xx τ−∂∂
( ) ( )C
j j
s s s sj ji iB BZ x f ( ) d f (x) Z x d
x xτ− τ−∂ ∂≤ − + −∂ ∂∫ ∫ξ ξ ξ ξ ξ (4.6)
( ) ( ) ( )j
s s sjiBZ x f f x d
x τ−∂+ − −∂∫ ξ ξ ξ .
For the first integral on the RHS of (4.6), it follows from (3.11) and (3.12) that
( ) ( )C
j2
s s 4 fjiBZ x f ( ) d C exp h x
xτ−
∂ − ≤∂∫ %ξ ξ ξ . (4.7)
We use the divergence theorem to obtain the following bound for the second integral:
( ) ( ) ( )j j 1
s s ij j 1i iB BZ x d Z x cos dS
x x
−
τ− τ−−∂
∂ ∂− = − γ
∂ ∂∫ ∫ ξξ ξ ξ (4.8)
( )j 1
sj 1iB
Z x dSx
−
τ−−∂
∂≤ −
∂∫ ξξ ,
56
where iγ is the angle between iξ and the outwardly directed normal to the boundary ∂B,
and dSξ is the ( )d 1− -dimensional volume element on ∂B. Since x − ξ = 1 on ∂B, we
use Lemma 3.1 with j 1
2
−µ = to deduce that
( ) ( )j j 1
s sj j 1i iB BZ x d Z x dS
x x
−
τ− τ−−∂
∂ ∂− ≤ −∂ ∂∫ ∫ ξξ ξ ξ (4.9)
( )2
d j 1
25
B
xC s exp dS
s
+ −−
∂
−≤ τ − −δ
τ − ∫ ξξ
d j 1
2
6 7B
C exp dS Cs s
+ −
∂
δ δ ≤ − ≤ τ − τ − ∫ ξ ,
where we used the fact that the expression in square brackets is uniformly bounded. It
follows that the second term on the RHS of (4.6) is bounded by
( ) ( ) ( )j
2s s 8 fj
iBf x Z x d C exp h x
x τ−∂ − ≤∂∫ ξ ξ . (4.10)
Since f is locally, uniformly Hölder continuous in space with exponent 1β < , we
have that there exists a positive constant 9C such that
( ) ( )s s 9f x f C xβ− ≤ −ξ ξ (4.11)
for every x and ξ in M and q s T≤ ≤ . Without loss of generality, we may assume that
B M⊆ . It follows from Lemma 3.1 with 1 12
β− < µ < that the final integral on the RHS
( ) ( ) ( )( )
2
j
s s s 9 d j 2jiB B
xexp
sZ x f f x d C d
x s xτ− + − µ−βµ
−−δ τ −∂ − − ≤
∂ τ −∫ ∫% ξ
ξ ξ ξ ξ− ξ
(4.12)
( )
( )910d j 2
B
C dC s
s x
−µµ + − µ−β≤ ≤ τ
τ −∫ ξ −− ξ
.
57
Combining (4.7), (4.10), and (4.12) with (4.6), we obtain
[ ] ( ) ( ) ( )j
2 2
s s 4 f 8 f 10ji
Z f C exp h x C exp h x C sx
−µτ−
∂ ≤ + + τ∂
% − (4.13)
( ) ( )2
11 fC s exp 2h x−µ≤ τ − .
Since 1µ < , it follows that
[ ] ( ) ( ) ( )j
2 2s s 11 f 12 fj
iq qZ f ds C exp 2h x s ds C exp 2h x
x−µ
τ−
τ τ∂ ≤ τ − ≤∂∫ ∫ . (4.14)
Therefore, we deduce from (4.3), (4.5), and (4.14) that for every ( )x, M (0,T]τ ∈ ×
we have
[ ]( ) ( ) ( ) ( )j
2 2 2
s s 3 f 12 f 13 fji0Z f x ds C exp h x C exp 2h x C exp 2h x
x τ−
τ ∂ ≤ + ≤∂∫ % (4.15)
We conclude that [ ] ( )j
dji
U f C (0,T]x
∂∈ ×
∂R , [ ]
j
2ji
U fx
τ δ∂
∈∂
A , and (4.2) holds for every
(0,T]τ∈ . < We have a similar result for the time derivative of the Gaussian potential.
Theorem 2 – Let 1δ ≤ and (0,T]
f fτ τ∈= be a uniformly bounded, continuous family of
functions from δA that is locally, uniformly Hölder continuous in space with exponent
1β < . Then, we have that [ ] ( )dU f C (0,T]∂
∈ ×∂τ
R . In fact, we have
[ ] [ ]s s0
U f Z f ds fτ τ− τ
τ∂ ∂= +∂τ ∂τ∫ (4.16)
from which we conclude that [ ] 2U fτ δ∂
∈∂τ
A for every (0,T]τ∈ .
proof: Let dM ⊆ R be compact, ( )x, M (0,T]τ ∈ × , and 0∆τ > . We apply the mean
value theorem to the following difference quotient to obtain
58
[ ]( )U Uf xτ+∆τ τ−
∆τ [ ]( ) [ ]( )s s s s
0 0
1Z f x ds Z f x dsτ+∆τ− τ−
τ+∆τ τ = −
∆τ ∫ ∫ (4.17)
[ ]( ) [ ]( )s ss s s
0
1 Z ZZ f x ds f x dsτ+∆τ− τ−
τ+∆τ−
τ+∆τ τ
τ
− = + ∆τ ∆τ ∫ ∫
[ ]( ) [ ]( )s s s s0
1Z f x ds Z f x dsτ+∆τ− τ−
τ+∆τ τ
τ
∂= +∆τ ∂τ∫ ∫ %
for some ( ),τ∈ τ τ + ∆τ% . In case the notation in the last line is unclear, we note that
[ ]( ) [ ]( )s s s sZ f x Z f xτ− τ−τ=τ
∂ ∂=
∂τ ∂τ%% (4.18)
We will show that the RHS of (4.17) approaches the RHS of (4.16) as ∆τ decreases
to zero. A symmetric argument can then be made to show that this result also holds as
∆τ increases to zero. This will establish (4.16) since the limit of the LHS of (4.17) as
0∆τ → is the LHS of (4.16).
Since [ ]( )s ss Z f xτ+∆τ−a is continuous on [ ],τ τ + ∆τ , it follows from the mean value
theorem for integrals that the first integral on the RHS of (4.17) may be rewritten as
[ ]( ) [ ]( )s s s s
1Z f x ds Z f xτ+∆τ− τ+∆τ−
τ+∆τ
τ=
∆τ∫ % % (4.19)
for some ( )s ,∈ τ τ + ∆τ% . Letting 0∆τ → yields
[ ]( ) [ ]( ) ( )s s s s0 0
1Z f x ds Z f x f xlim limτ+∆τ− τ+∆τ− τ
τ+∆τ
∆τ→ ∆τ→τ= =
∆τ∫ % % . (4.20)
For the second integral on the RHS of (4.17), it remains to be shown that
[ ]( ) [ ]( )s s s s0 0 0Z f x ds Z f x dslim τ− τ−
τ τ
∆τ→
∂ ∂=∂τ ∂τ∫ ∫% . (4.21)
59
Let 0ε > , 1 12
β− < µ < ,
( )1
1
1
11
2 3C
−µ− µ ε ε ≡
, and ( )1 ,ετ ∈ τ − ε τ , where the positive
constant C will be determined below. Consider
I [ ]( ) [ ]( )s s s s0 0
Z f x ds Z f x dsτ− τ−
τ τ∂ ∂≡ −∂τ ∂τ∫ ∫% (4.22)
[ ]( ) [ ]( ) [ ]( )e e
s s s s s s s0
Z Z f x ds Z f x ds Z f x dsε
τ− τ− τ− τ−
τ τ τ
τ τ
∂ ∂ ∂ ∂ ≤ − + + ∂τ ∂τ ∂τ ∂τ ∫ ∫ ∫% % .
For the first integral on the RHS of (4.22), fix s (0, ]ε∈ τ and note that
s 0ετ − > τ − τ >% . We deduce from Theorem 3.2 that [ ]( )q s sq Z f x−∂∂τ
a is continuous on
[ ],Tτ . It follows that there exists a 2 1ε ≤ ε such that for 2∆τ < ε we have
[ ]( )s s sZ Z f x3Tτ− τ−
∂ ∂ ε − < ∂τ ∂τ % (4.23)
from which we deduce
[ ]( )s s s0
Z Z f x ds3
ε
τ− τ−
τ ∂ ∂ ε − < ∂τ ∂τ ∫ % . (4.24)
To evaluate the second and third integrals on the RHS of (4.22), recall that Zτ is a
parametrix for L. This implies that
2Z 1Z
2τ
τ∂
= ∇∂τ
. (4.25)
Let 1 12
β− < µ < and N denote the diameter of M. Combining (4.25) with (4.13) yields
[ ]( ) ( )( ) ( )( ) ( )2 2ss f f
Zf x Cexp 2h x s Cexp 2h N s C s
−µ −µ −µτ−∂≤ τ − ≤ τ − ≤ τ −
∂τ% % (4.26)
for every [ ]s ,Tε∈ τ , where C% is some positive constant and ( )2fC Cexp 2h N≡ % .
60
We use this estimate to see that for ∆τ < δ and q ,∈ τ τ% we have
[ ]( )q s sZ f x dsε
−
τ
τ
∂∂τ∫ ( ) ( ) ( )1 1C
C q s ds q q1
ε
−µ −µ −µε
τ
τ ≤ − = − τ − − τ − µ∫ (4.27)
( ) ( ) ( ) 11C Cq q
1 1
−µ−µε ε≤ − τ ≤ − τ + τ − τ − µ − µ
( ) ( )1 1C C2
1 1 3
−µ −µε
ε≤ ∆τ + τ − τ < δ = − µ − µ
.
Combining the estimates (4.24) and (4.27) with (4.22) yields I < ε when ∆τ < δ% from
which we deduce (4.21). It follows from (4.20) and (4.21) that
[ ]( ) [ ]( )s s s s0 0
1Z f x ds Z f x dslim τ+∆τ− τ−∆τ↓
τ+∆τ τ
τ
∂+ ∆τ ∂τ ∫ ∫ % (4.28)
[ ]( ) ( )s s0
Z f x ds f xτ− τ
τ ∂= +∂τ∫ .
A symmetric argument can be made to establish the analogous result for the left-handed
limit. Combining this with (4.17) yields the desired result (4.16).
The remaining assertions follow from Theorem 3.2 and (4.16). In fact, it follows that
[ ]( )U f xτ∂∂τ
[ ]( ) ( )s s0
Z f x ds f xτ− τ
τ ∂≤ +∂τ∫ (4.29)
( ) ( ) ( )2 2 2
1 f 2 f 3 fC exp 2h x C exp h x C exp 2h x≤ τ + ≤
for every ( ) dx, (0,T]τ ∈ ×R from which we deduce that [ ] ( )j
dji
U f C (0,T]x
∂∈ ×
∂R and
[ ]j
2ji
U fx τ δ∂
∈∂
A . <
61
We combine Theorems 1 and 2 to obtain the desired relation (4.1) between the
Gaussian and L-potentials. Let 0 Tτ <τ≤ψ ≡ ψ be a continuous family of Z - boundedδ
functions and f δ∈A for some positive 1δ ≤ . It follows from (3.25) that [ ] 0 T
fτ <τ≤ψ is
a continuous, uniformly bounded family of functions from 2δA . Assuming that [ ]fψ is
locally, uniformly Hölder continuous in space with exponent 1β < , we deduce from
Theorems 1 and 2 that for every ( )x, (0,T]τ ∈ we have
[ ]( ) [ ]( ) [ ] [ ] [ ]s s0L U f L Z f ds f V f f
τ
τ τ− τ τ τ ψ = ψ − ψ = ψ − ψ ∫ . (4.30)
2.5 A Series Representation of the Fundamental Solution We will now apply relation (4.30) to the problem of representing the fundamental
solution for L by [ ]Z Uτ τ τΓ = + ϕ (5.1)
for some continuous family 0 Tτ <τ≤ϕ ≡ ϕ of functions from ( )dC R that satisfies
[ ]( ) [ ]( ) [ ][ ]( )0 L f L Z f L U fτ τ τ= Γ = + ϕ (5.2)
for every (0,T]τ∈ and f ∈A . Let f ∈A and assume that ϕ is Z - boundedδ for some
1δ ≤ . If [ ]fϕ is locally, uniformly Hölder continuous in space (with exponent 1β < ),
then we deduce from Corollary 3.9 and (4.30) that
[ ][ ]( ) [ ]( ) [ ] [ ]L U f L U f V f fτ τ τ τ ϕ = ϕ = ϕ − ϕ . (5.3)
Combining this with (5.2) implies that ϕ must satisfy the following Volterra integral
equation for every (0,T]τ∈ and f ∈A :
[ ] [ ]( ) [ ]f L Z f V fτ τ τ ϕ = + ϕ . (5.4)
62
2.5.1 Convergence and Continuity
The goal of this sub-section is to show that the solution to (5.4) is given by
[ ]m
m 0
V LZ∞
τ τ=
ϕ = ∑ , (5.5)
where mVτ is an operator defined by
0V Iτ ≡ (5.6) and m 1 mV V V+
τ τ τ≡ o (5.7)
for every (0,T]τ∈ and m 0≥ . We will see in the proof of Theorem 2 below that
[ ]mV LZτ is well-defined for every (0,T]τ∈ and m 0≥ . Furthermore, it is shown that
the series representation (5.5) is a continuous family of Zδ -bounded functions for every
1δ < . Finally, our objective is obtained in Corollary 3, where we deduce that (5.5) is a
solution to (5.4). We begin with a preliminary result:
Lemma 1 – LZ is Zδ -bounded for every positive 1δ < .
proof : Let0 1< δ < . From the definition of L we have
( ) ( ) ( ) ( ) ( ) ( )d
2
ii 1 i
LZ x x Z x x Z x x x Z xxτ τ τ τ
=
∂ θ = −θ + Ψ = + Ψ ∂ τ ∑ (5.8)
and
21
(x) x2
+ Ψ = γ − σ
% . (5.9)
From Lemma 3.1 with 0µ = , it follows that there is an A 0> such that for every
( ) dx, (0,T]τ ∈ ×R we have
( )0 LZ (x) AZ (x) Z x CZ (x)δ δτ τ τ τ≤ ≤ θ + γ ≤ , (5.10)
where C A≡ θ + γ . <
63
Throughout the remainder of this section, let ϕ be given by the series (5.5).
Theorem 2 – The series ϕ converges absolutely. Moreover, ϕ is a continuous, Zδ -
bounded function for every 1δ < .
proof : Let 0 1< δ < . The first step is to prove by induction on m that [ ] m
0 TV LZτ <τ≤
is a continuous family of Z - boundedδ functions such that for some C 0> we have
[ ] ( ) ( )m m 1
m 2C
V LZ Z xm 1
2
−δ
τ τ≤ πτ+ Γ
. (5.11)
From Lemma 1 we see that [ ] 0
0 TV LZτ <τ≤
is a continuous family of Z - boundedδ
functions that satisfies (5.11) for some C 0> . To use induction, assume that this holds
for some m 0≥ . Since [ ] m
0 TV LZτ <τ≤
is a continuous family of Z - boundedδ functions,
it follows from (3.36) and (5.11) that
[ ]m 1V LZ (x)+τ [ ] ( ) [ ]m m
s sd0
V V LZ x LZ (x )V LZ ( ) d dsτ τ−
τ
= ≤ − ∫ ∫Rξ ξ ξ (5.12)
( ) ( ) ( ) ( )m 1 1 m 1
2 2s s
d0
Cs s Z x Z d ds
m 1
2
+ −− δ δτ−
τ
≤ τ − π −+ Γ
∫ ∫Rξ ξ ξ
( ) ( )m 1 m
2C 1 m 1
B , Z xm 1 2 2
2
+δτ
+ = πτ + Γ π
( ) ( )m 1 m
2C
Z xm 2
2
+δτ= πτ
+ Γ
.
Therefore, [ ] m
0 TV LZτ <τ≤
is a continuous family of Z - boundedδ functions that
satisfies (5.11) for every m 1≥ . To obtain the desired bound on ϕ, we will break the
series apart as follows:
64
[ ]( ) [ ]( ) [ ]( )m m m
m 1 m 1 m 2mevenmodd
(x, ) V LZ x V LZ x V LZ xτ τ τ
∞ ∞ ∞
= = =
ϕ τ ≤ ≤ +∑ ∑ ∑ . (5.13)
For the odd series we have
[ ]( ) [ ]( ) ( ) ( )( )
n2n 1m 2n 1
m 1 n 0 n 0modd
CV LZ x V LZ x Z x
n 1
++ δ
τ τ τ
∞ ∞ ∞
= = =
πτ= ≤ Γ +
∑ ∑ ∑ (5.14)
( ) ( ) ( ) ( ) ( )n2n
21
n 0
CCZ x, Cexp C Z x, C Z x,
n!δ δ δ
∞
=
πτ≤ τ ≤ π τ τ ≤ τ
∑ ,
where ( )21C Cexp C T≡ π . For the even series we have
[ ]( ) [ ]( ) ( ) ( )1
n2n 2m 2n
m 1 n 1 n 1meven
CV LZ x V LZ x Z x
1n
2
−δ
τ τ τ
∞ ∞ ∞
= = =
πτ = ≤
Γ +
∑ ∑ ∑ (5.15)
( ) ( )( ) ( ) ( )
1n n2n 2n2
2
n 1 n 0
C C2Z x 2Z x C
n 1 ! n!
−δ δτ τ
∞ ∞
= =
πτ πτ ≤ ≤ πτ − ∑ ∑
( ) ( ) ( )2 222C exp C Z x C Z xδ δ
τ τ≤ πτ π τ ≤ ,
where ( )2 22C 2C Texp C T≡ π π . By inserting (5.14) and (5.15) into (5.13) we conclude
that ϕ is Zδ -bounded. Furthermore, since [ ] ( )m dV LZ C (0,T]∈ ×R for every m 1≥ and
the series representation for ϕ is uniformly bounded on compact subsets of d (0,T]×R ,
we also have that ( )dC (0,T]ϕ∈ ×R . <
Corollary 3 – ϕ is a solution to the following Volterra integral equation for every
(0,T]τ∈ and f ∈A :
[ ] [ ]( ) [ ]f L Z f V fτ τ τ ϕ = + ϕ . (5.16)
65
proof : It follows from Theorem 2 that for every (0,T]τ∈ we have
[ ]Vτ ϕ [ ]( ) [ ]( ) [ ]m ms s s s
m 0 m 00 0L Z ds L Z V LZ ds V V LZ
∞ ∞
τ− τ− τ= =
τ τ = ϕ = = ∑ ∑∫ ∫ (5.17)
[ ] [ ] [ ]m 1 m 0
m 0 m 1
V LZ V LZ V LZ LZ∞ ∞
+τ τ τ τ τ τ
= =
= = = ϕ − = ϕ −∑ ∑ .
Therefore, we deduce from Corollary 3.12 that for every (0,T]τ∈ and f ∈A we have
[ ] [ ][ ] ( )[ ]V f V f LZ fτ τ τ τ ϕ = ϕ = ϕ − . (5.18)
Hence, we conclude that (5.16) holds for the series representation (5.5) for ϕ. <
We would like to be able to claim that
[ ]( )L f 0τΓ = , (5.19)
where [ ]Z Uτ τ τΓ = + ϕ (5.20)
for every (0,T]τ∈ and f ∈A . However, it remains to be shown that [ ]fϕ is locally,
uniformly Hölder continuous in space with exponent 1β < . As stated in the beginning of
this section, this condition combined with Theorem 2 and Corollary 3 is sufficient to
conclude that (5.19) holds for every (0,T]τ∈ and f ∈A .
2.5.2 Hölder Continuity The first objective of this sub-section is to show that ϕ is uniformly Hölder
continuous in space with exponent 1β < . We deduce from (5.17) that
( ) ( ) ( ) ( ) [ ]( ) [ ]( )x y LZ x LZ y V x V yτ τ τ τ τ τϕ − ϕ ≤ − + ϕ − ϕ . (5.21)
for every x and y in dR and (0,T]τ∈ . Our goal will be obtained by proving the desired
result for each of the terms on the RHS of (5.21). We begin with a preliminary lemma:
66
Lemma 4 – Assume that 2
x y− < τ . Suppose that ζ lies along the line segment
connecting x and y. Then, for any positive constant δ there is a C 0> such that
( ) ( )Z CZ yδ δτ τζ ≤ . (5.22)
proof : Since ζ lies on the line segment connecting x and y, we have ζ = λx + (1 – λ)y
for some λ ∈ (0,1). Hence,
2 2 2
x (1 )y (x y) yζ = λ + − λ = λ − + (5.23)
2 2
y x y≥ − λ − 2y≥ −λτ .
It follows that
( )( )
( )d d 222 2 y
Z exp exp CZ y2 2 2 2
δ δτ τ
δ −λτ δ ζδ δ ζ = − ≤ − = πτ τ πτ τ
, (5.24)
where C exp2
δλ ≡
.
Theorem 5 – Let x and y be elements of the compact set dM ⊆ R , d∈Rξ (0,T]τ∈ ,
0 1< β < , and 0 1< δ < . Then, there exists a positive constant C such that
( ) ( )( )1
2x yL Z (x ) L Z (y ) C x y Z x Z y
+β−β δ δ
τ τ τ τ− − − ≤ − τ − + −ξ ξ ξ ξ , (5.25)
where the subscripts x and y denote the spatial variable of the action of the operator L.
proof: We divide this proof into two cases:
Case 1 – 2
x y− ≥ τ
From (3.15) it follows that for some C 0> we have
( ) ( )1 1
2 2 2xL Z x ) C Z x C x y Z x
β +β +β− −βδ δ
τ τ τ( − ≤ τ τ − ≤ − τ −ξ ξ ξ . (5.26)
A similar inequality holds for yL Z (y )τ − ξ , thus we obtain (5.25) in this case.
67
Case 2 – 2
x y− < τ
Throughout this case we will use the following result for every [ ]a 0,1∈ + β :
a 1 a 1 1
12 2 2 2x y x y x y C x y+β− +β +β− − −−β β β
τ − = τ − τ − ≤ τ −
. (5.27)
Consider the inequality
x y i ii i
d
i 1
L Z (x ) L Z (y ) x Z (x ) y Z (y )x y
τ τ τ τ
=
∂ ∂− − − ≤ θ − − −
∂ ∂∑ξ ξ ξ ξ (5.28)
( ) ( ) ( ) ( )x Z x y Z yτ τ+ Ψ − − Ψ −ξ ξ .
We first break up the summation and find that
i i i ii i i
d d
i 1 i 1
x Z (x ) y Z (y ) x y Z (x )x y xτ τ τ
= =
∂ ∂ ∂− − − ≤ − −
∂ ∂ ∂∑ ∑ξ ξ ξ (5.29)
ii i
d
i 1
Z (x ) Z (y ) yx yτ τ
=
∂ ∂+ − − −
∂ ∂∑ ξ ξ .
For the first term on the RHS of (5.29), we see from Lemma 3.1 with1
2µ = and
(5.27) that there exists a constant 1C 0> such that
( ) ( )1 1
2 2i i 1 1
i
d
i 1
x y Z (x ) C x y Z x C x y Z xx
+β− − βδ δτ τ τ
=
∂− − ≤ τ − − ≤ τ − −
∂∑ ξ ξ ξ . (5.30)
Applying the mean value theorem to the second term on the RHS of (5.29), it follows
that there exists a point ζ lying on the line segment between x − ξ and y − ξ such that for
some 2C > 0 we have
ii i
d
i 1
Z (x ) Z (y ) yx y
τ τ
=
∂ ∂− − −
∂ ∂∑ ξ ξ ( )2
i i i2i
d
i 1
Z x y yτ
=
∂≤ ζ −
∂ζ∑ (5.31)
( )2
2 2i
d
i 1
C x y Zτ
=
∂≤ − ζ
∂ζ∑ .
68
It follows from Lemma 3.1 with 1µ = that for some 3C 0> we have
( ) ( )2
132
i
d
i 1
Z C Z− δτ τ
=
∂ζ ≤ τ ζ
∂ζ∑ . (5.32)
Since 2
x y− < τ , we deduce from Lemma 4 that
( )2
2i
d
i 1
Zτ
=
∂ζ
∂ζ∑ ( ) ( )1 1
1 2 24 4C Z y C Z y
+β −β− −− δ δ
τ τ≤ τ − = τ τ −ξ ξ (5.33)
( )1
(1 )24C x y Z y
+β− − −β δ
τ≤ τ − − ξ .
Combining this with (5.31) yields
( )1
2i 5
i i
d
i 1
Z (x ) Z (y ) y C x y Z yx y
+β− β δτ τ τ
=
∂ ∂− − − ≤ τ − −
∂ ∂∑ ξ ξ ξ . (5.34)
Finally, by combining (5.30) and (5.34) with (5.29) we obtain
( ) ( )( )1
2i i 6
i i
d
i 1
x Z (x ) y Z (y ) C x y Z x Z y ,x y
+β− β δ δτ τ τ τ
=
∂ ∂− − − ≤ τ − − + − τ
∂ ∂∑ ξ ξ ξ ξ . (5.35)
We now consider the second term on the RHS of (5.28). Recalling the definition of
Ψ given in (5.9), we have
( ) ( ) ( ) ( ) ( ) ( )x Z x y Z y Z x Z xτ τ τ τΨ − − Ψ − ≤ γ − − −ξ ξ ξ ξ . (5.36)
Applying the mean value theorem, Lemma 3.1, Lemma 4, and (5.27) to the RHS of
(5.36), it follows that there exists a point ζ lying on the line segment between x − ξ and
y − ξ such that
( ) ( ) ( ) ( )12
i i 7i
Z x Z y x y Z C x y Z− δ
τ τ τ τ∂γ − − − = γ − ζ ≤ τ − ζ
∂ζξ ξ (5.37)
( )1
28C x y Z y
+β− β δ
τ≤ τ − − ξ .
Combining (5.35), (5.36), and (5.37) with (5.28), we see that (5.25) holds for case 2. <
69
Applying this result with 0=ξ , we deduce that for every q 0> and compact
dM ⊆ R , there exists positive constants C and C% such that
( ) ( )( )1 1
2 2LZ (x) LZ (y) C x y Z x Z y Cq x y+β +β
− −β βδ δτ τ τ τ− ≤ − τ + ≤ −% (5.38)
for every x and y in M and [ ]q,Tτ∈ . Hence, LZ is locally, uniformly Hölder continuous
in space with exponent 1β < . We proceed to establish this result for the second term on
the RHS of (5.21).
Corollary 6 – [ ]V ϕ is locally, uniformly Hölder continuous in space with exponent
1β < . More precisely, for every compact dM ⊆ R and positive 1δ < , we have
[ ]( ) [ ]( ) ( ) ( )( )V x V y C x y Z x Z yβ δ δ
τ τ τ τϕ − ϕ ≤ − + (5.39)
for every x and y in M and ( ]0,Tτ∈ .
proof : Let x and y be elements of the compact set M, 0 1< β < , and 0 1< δ < . Since ϕ
is Zδ -bounded, it follows from Theorem 5, Theorem 2, and Lemma 3.7 that for some
C 0> we have
[ ]( ) [ ]( )V x V yτ τϕ − ϕ [ ]( )( ) [ ]( )( )( )s s s s0
L Z x L Z y dsτ− τ−
τ≤ ϕ − ϕ∫ (5.40)
( ) ( )d
x s y s s
0L Z (x ) L Z (y ) d dsτ− τ−
τ
≤ − − − ϕ∫ ∫Rξ ξ ξ ξ
( ) ( )d
12
s s
0C x y ( s) Z Z x
+β−β δ δτ−
τ
≤ − τ − ∫ ∫Rξ −ξ
( )sZ y d dsδτ− + −ξ ξ
( ) ( )( )1
21
CB ,1 x y Z x Z y2
−ββ δ δ
τ τ
− β = − τ +
( ) ( )( )C x y Z x Z yβ δ δ
τ τ≤ − + . <
70
We conclude from (5.21), (5.38), and Corollary 6 that for every compact dM ⊆ R
and positive 1δ < , there is a positive constant C such that
( ) ( )( )1
2(x) (y) C x y Z x Z y+β
−β δ δτ τ τ τϕ − ϕ ≤ − τ + (5.41)
for every x and y in M and ( ]0,Tτ∈ . Hence, ϕ is locally, uniformly Hölder continuous
in space with exponent 1β < . This result allows us to prove the following theorem from
which we will deduce that [ ]L f 0τΓ = for every f ∈A and ( ]0,Tτ∈ .
Theorem 7 – [ ]fϕ is locally, uniformly Hölder continuous in space with exponent 1β <
for every f ∈A . More precisely, for every compact dM ⊆ R there is a positive constant
C such that we have
[ ]( ) [ ]( ) ( ) ( )( )1
2 22f ff x f y C x y exp h x exp h y
+β−β
τ τϕ − ϕ ≤ − τ +% % (5.42)
for every x and y in M and ( ]0,Tτ∈ , where f f0 h 2h< <% .
proof: Let x and y be elements of the compact set M, 0 1< β < , and f ∈A . It follows
from (5.41) that for some positive constants C and δ with 1
12
< δ < we have
[ ]( ) [ ]( )f x f yτ τϕ − ϕ ( ) ( )( ) ( )d
x y f dτ τ≤ ϕ − − ϕ −∫Rξ ξ ξ ξ . (5.43)
( ) ( )( ) ( )122
fd
C x y Z x Z y exp h d+β
−β δ δτ τ≤ − τ − + −∫R
ξ ξ ξ ξ
Since 2f δ∈ ⊂A A , it follows from Lemma 2.2 that there exists positive constants C%
and fh% with f fh 2h<% such that
[ ]( ) [ ]( ) ( ) ( )( )1
2 22f ff x f y C x y exp h x exp h y
+β−β
τ τϕ − ϕ ≤ − τ +% %% . < (5.44)
71
2.6 The Solution to the Cauchy Problem Before proceeding to prove the final results necessary to provide the solution to the
Cauchy problem, let us summarize the work we have completed in the past few sections.
We began by looking for a fundamental solution for L of the form
[ ]Z Uτ τ τΓ = + ϕ (6.1)
for some function ( )dC (0,T]ϕ∈ ×R . It was shown in Section 4 that
[ ][ ]( ) [ ] [ ]L U f V f fτ τ τ ϕ = ϕ − ϕ , (6.2)
provided that ϕ is Zδ -bounded for some 1δ ≤ , and that [ ]fϕ is locally, uniformly
Hölder continuous in space (with exponent 1β < ) for every f ∈A . Combining
(6.1) and (6.2) with the condition that [ ]( )L f 0τΓ = for every (0,T]τ∈ and f ∈A , it
was deduced in Section 5 that ϕ must satisfy the following Volterra integral equation:
[ ] [ ]f LZ V fτ τ τϕ = + ϕ . (6.3)
Finally, it was shown that the series [ ]m
m 0
V LZ∞
τ τ=
ϕ = ∑ satisfies (6.3) and the conditions
necessary for (6.2) to hold from which we deduced that [ ]( )L f 0τΓ = for every (0,T]τ∈
and f ∈A . Thus, to conclude that τΓ is a fundamental solution for L, it suffices to show
that [ ] [ ]( )00
f lim f fττ→Γ ≡ Γ = for every f ∈A . Furthermore, from (6.1) and Corollary 3.9
we have that
[ ] [ ]( )[ ] [ ]0 0 0
lim f lim Z U f f limU fτ τ τ ττ→ τ→ τ→ Γ = + ϕ = + ϕ (6.4)
for every f ∈A . Therefore, it suffices to prove the following theorem:
Theorem 1 – For every f ∈A , we have [ ]0
limU f 0ττ→ ϕ = .
proof: Let f ∈A . It follows from Theorem 5.2 and (3.24) that [ ] 0 T
fτ <τ≤ϕ is a
72
continuous, uniformly bounded family of functions from 2A . Hence, we deduce from
Lemma 2.2 that for some positive constants C and [ ] [ ]f fh 2hϕ ϕ<% we have
[ ] ( ) [ ] [ ]( )2
s s f0
U f x Z f ds C exp h x 0 as 0τ τ− ϕ
τϕ ≤ ϕ ≤ τ → τ → ∫ % . < (6.5)
Therefore, τΓ is a fundamental solution for L. It follows that the solution to our
Cauchy problem may be represented by
( ) [ ]( ) [ ]( ) [ ] ( )u x, g x Z g x U g xτ τ τ τ = Γ = + ϕ , (6.6)
where ( )2ˆg(x) exp x≡ α is the initial condition, provided that 1ˆ
32 dα <
ρ.
Finally, we must verify that u satisfies the hypotheses of the Feynman-Kac Theorem
from which we will deduce that
( ) ( ) ( )xP s0
u x, E g X exp dsτ
τ τ = Ψ ∫ (6.7)
for every ( ) [ ]dx, 0,Tτ ∈ ×R . From (6.6), Theorem 3.2, Theorem 4.1, and Theorem 4.2
we have that u is of class ( )2,1 dC [0,T]×R . Furthermore, by assuming that 1ˆ
32 dα <
ρ,
we see that g ∈A . Hence, we deduce from Theorem 5.2 and (3.24) that [ ] 0 T
gτ <τ≤ϕ is
a continuous, uniformly bounded family of functions from 2A . Consequently, it follows
from Lemma 2.2 and Theorem 3.5 that 4u ∈A , thus there exists positive constants A and
1h
8 d<
ρ such that
( ) ( )2
0 Tmax u x, Aexp h x
≤τ≤τ ≤ (6.8)
for every dx ∈R . We conclude from the Feynman-Kac Theorem that (6.7) holds for
every ( ) [ ]dx, 0,Tτ ∈ ×R . Therefore, we deduce from (1.5) that the arbitrage price of the
risk-spread option is given by
( ) ( )( ) ( ) [ ]( ) ( )t
t t s t0C exp r exp ds g X B t,Tτπ = − ατ + κ Ψ Γ −∫ . (6.9)
73
CHAPTER 3 NUMERICAL RESULTS AND APPLICATIONS
3.1 The Fourier and Laplace Transforms The series solution of convoluted potentials presented in the previous chapter is an
elegant exhibition of the ability of the modern mathematical prose to convey a relatively
simple method in a clear, concise manner. In fact, for those pure mathematicians who are
familiar with the antiquated notation used to present the parametrix method in Friedman
(1964), the refined potential theoretic approach is a significant contribution to the
literature. However, it remains to be shown that this solution is of practical value.
Although the iterated convolutions prevent us from directly integrating the series,
they are ideal for calculating the Fourier and Laplace transforms of our solution. Recall
that these transforms are defined by
[ ]( ) ( ) ( ) ( )d
d
2f 2 exp i x f x dx−= π − ⋅∫RF ξ ξ (1.1)
and
[ ]( ) ( ) ( )0
h s exp st h t dt∞
≡ −∫L (1.2)
for every df : →R R and h : + →R R for which these integrals are defined. A useful
result is that these transforms change convolutions into products. In fact, we have
[ ] ( ) [ ] [ ]d
21 2 1 2f f 2 f f∗ = πF F F (1.3)
and [ ] [ ] [ ]1 2 1 2h h h h∗ =L L L h . (1.4)
74
Recall from (2.6.6) that the solution to our Cauchy problem may be represented by
( ) [ ]( ) [ ]( ) [ ] ( )u x, g x Z g x U g xτ τ τ τ = Γ = + ϕ , (1.5)
where ( )2ˆg(x) exp x≡ α and [ ]m
m 0
V LZ∞
τ τ=
ϕ = ∑ for every ( ) [ ]dx, 0,Tτ ∈ ×R . It follows
from (2.2.5) that the first term on the RHS of (1.5) is given by
[ ]( ) ( )d
22
ˆˆZ g x 1 2 exp x
ˆ1 2
−τ
α = − ατ − ατ . (1.6)
The remainder of this section is devoted to the Gaussian potential in (1.5). Moreover, we
will consider the case where d 1= henceforth.
Since the Fourier transform of g does not exist, we will consider the truncated
approximation
( ) ( )n
g x if x ng x
0 if x n
≤≡ >
. (1.7)
Hence, we approximate the Gaussian potential in (1.5) by the sequence [ ] n n 1U g
∞
τ = ϕ .
We proceed to take the Fourier and Laplace transforms of the approximated Gaussian
potential in the spatial and temporal variables, respectively. We recall that
[ ] [ ]n s s n0U g Z g ds
τ
τ τ− ϕ = ϕ ∫ . (1.8)
It follows from (1.3) that for every ( ) [ ], 0,Tτ ∈ ×Rξ we have
[ ] ( ) ( )( ) [ ]( )n q q n0U g 2 Z dq g
τ
τ τ− ϕ = π ϕ ∫F F F Fξ ξ ξ . (1.9)
By defining Z 0τ ≡ for every 0τ ≤ , we see that
[ ]( ) [ ]( ) [ ] [ ]( )( )s s0Z ds Z ,
τ
τ− ϕ = • ϕ τ∫ F F F Fξ ξ ξ (1.10)
where • denotes convolution in time. Thus, we deduce from (1.4) that
75
[ ] ( )nU g ,s ϕ L F ξ [ ] ( ) [ ] ( ) [ ]( )n2 Z ,s ,s g = π ϕ L F L F Fξ ξ ξ (1.11)
[ ]( ) [ ] ( )n
2
2 2 g,s
2s
π = ϕ +
F L Fξ ξξ .
From the series representation for ϕ, we see that
[ ] [ ]m
m 0
V LZ∞
=
ϕ = ∑L F L F . (1.12)
It follows from the definition of the L-potential that
[ ]( ) [ ] ( ) [ ]( )m 1 m ms s0
V LZ x V V LZ x L Z V LZ dsτ+
τ τ τ− = = ∫ (1.13)
( ) [ ]( )mx s s0
L Z x y V LZ y dydsτ ∞
τ−−∞= −∫ ∫
( ) [ ]( )mx s s0
L Z x y V LZ y dsdy∞ τ
τ−−∞= −∫ ∫
( ) [ ]( )( )( )mxL Z x y V LZ y dy
∞
−∞= − • τ∫
for every ( ) [ ]x, 0,Tτ ∈ ×R and m 0≥ . Applying (1.4) to (1.13), we deduce that
[ ] ( ) [ ]( ) [ ] ( )m 1 mxV LZ x,s L Z x y,s V LZ y,s dy
∞+
−∞ = − ∫L L L . (1.14)
Let ( ) [ ]x, 0,Tτ ∈ ×R and y∈ R . From the definition of L, we have that
( )xL Z x yτ − ( ) ( ) ( )x x y x Z x yτθ = − + Ψ − τ
(1.15)
( ) ( ) ( )2 2x x y R x Z x y2
+
τθ σ = − + − − τ
%,
where 2
Rγ≡
σ% . Ideally, we would like to express ( )xL Z x yτ − as a function of x y−
and y so that we may rewrite (1.14) as a convolution in x. However, the plus operator in
the second term on the RHS of (1.15) makes this impossible. For the first term, we will
use the following identity:
( ) ( ) ( )2x x y x y y x y− ≡ − + − . (1.16)
76
Inserting this identity into (1.15) yields
( ) ( ) ( )( ) ( ) ( )2
xL Z x y x y y x y x Z x yτ τθ − = − + − + Ψ − τ
. (1.17)
We will write the Laplace transform of ( )xL Z x y− in terms of the following functions:
( ) ( ) ( )2 ˆF x,s x Z x,s x exp x 2s ≡ θ = θ − L (1.18)
and
( ) ( ) ( ) ( )ˆG x,s x Z x,s sgn x exp x 2s ≡ θ = θ − L , (1.19)
where
( ) ( )1Z x, Z xττ ≡
τ. (1.20)
It follows from (1.17) that
( ) ( ) ( ) ( )xL Z x y F x y,s yG x y,s xτ − = − + − + Ψ L . (1.21)
Let m 0≥ . Combining (1.21) and (1.14) yields
[ ] ( ) ( ) [ ] ( )m 1 mV LZ x,s F x y,s V LZ y,s dy+∞
−∞ = − ∫L L (1.22)
( ) [ ] ( )( )mG x y,s y V LZ y,s dy∞
−∞ + − ∫ yL
( ) [ ]( ) [ ] ( )mx Z x y,s V LZ y,s dy∞
−∞ +Ψ − ∫ L Ly .
The first two integrals may be expressed as convolutions. In fact, we have
( ) [ ] ( ) [ ]( )( )m mF x y,s V LZ y,s dy F V LZ x,s∞
−∞ − = ∗ ∫ L L (1.23)
and
( ) [ ] ( )( ) [ ]( )( )( )m mG x y,s y V LZ y,s dy G x V LZ x,s∞
−∞ − = ∗ ∫ yL L . (1.24)
For the third integral on the RHS of (1.22), we have
( ) [ ]( ) [ ] ( ) ( ) [ ] [ ]( )( )m mx Z x y,s V LZ y,s dy x Z V LZ x,s∞
−∞ Ψ − = Ψ ∗ ∫ L Ly L Ly . (1.25)
77
We now compute the Fourier transform. Applying (1.3) to (1.23) and (1.24) yields
[ ] ( ) [ ] [ ] ( )m mF V LZ ,s 2 F V LZ ,s ∗ = π F L F F Lξ ξ (1.26)
and
[ ]( ) ( ) [ ] [ ] ( )m mG x V LZ ,s 2 G x V LZ ,s ∗ = π F L F F Lξ ξ (1.27)
[ ] [ ]( )( )m2 G i V LZ ,s∂ = π ∂
F F L ξξ ,
respectively. The last equality follows from the fact that
( ) [ ]( )n
n n
nx f i f
∂ = ∂F Fξ ξξ (1.28)
for every function f for which these Fourier transforms are defined.
To compute the Fourier transform of (1.25), we will also need the following property:
[ ] ( ) [ ] [ ]( )d
21 2 1 2f f 2 f f
−= π ∗F F F (1.29)
Applying (1.28), (1.29), and (1.3) to (1.25) yields
[ ] [ ]( ) ( ) ( ) [ ] [ ]( ) ( )m 2 2 mZ V LZ ,s R x Z V LZ ,s2
+σ Ψ ∗ = − ∗ %F L Ly F L Lyξ ξ (1.30)
[ ] [ ] [ ]( ) ( )2
2 mR,R2
R 1 Z V LZ ,s2 −
σ ∂ = + ∗ ∂
% F L Ly ξξ
[ ] [ ] [ ]( )( )2
2 mR,R2
R 1 * Z V LZ ,s2 −
σ ∂ = + ∂
% F F L F Ly ξξ .
Combining (1.26), (1.27), and (1.30) with (1.22) yields the following iterative relation for
every m 0≥ :
[ ] ( )m 1V LZ ,s+ F L ξ [ ] [ ] [ ] ( )m2 F i G V LZ ,s ∂ = π + ∂ F F F L ξξ (1.31)
[ ] [ ] [ ]( )( )2
2 mR,R2
R 1 * Z V LZ ,s2 −
σ ∂ + + ∂
% F F L F Ly ξξ .
78
Computing the various Fourier transforms in (1.31), we have
[ ]( )( )
2
22
2 2sF ,s
2s
−= θ
π +F ξξ
ξ, (1.32)
[ ]( ) 2
2i G ,s
2s= θ
π +F ξξ ξ , (1.33)
[ ] ( ) ( )R,R
sin R21 −
= πF ξξ ξ , (1.34)
and
[ ] 2
2 1Z
2s = π +
F L ξ . (1.35)
Inserting these results into (1.31) yields
[ ] ( )m 1V LZ ,s+ F L ξ (1.36)
( )
( ) [ ] ( )2 2 m22
22s 2s V LZ ,s
2s
θ ∂ = − + + ∂ +F Lξ ξ ξ ξξξ
( )( )
( )( ) [ ] ( )2
2 m2 2
sin y RR V LZ y,s dy
2 y 2s y
∞
−∞
− σ ∂ + + ∂ − + ∫% F Lyξ
ξ ξ .
Combining (1.11), (1.12), and (1.6), we see that (1.5) becomes
( ) ( )d
22ˆ
ˆu x, 1 2 exp xˆ1 2
− α τ = − ατ − ατ (1.37)
[ ]( ) [ ] ( )n1 1 m
2nm 0
2 2 glim V LZ ,s
2s
∞− −
→∞ =
π + + ∑
FF L F Lξ ξξ ,
where the terms of the series are given by (1.36). Although this is not as aesthetically
pleasing as (1.5), we have removed the temporal convolution in the Gaussian potential
term. Unfortunately, the spatial convolution remains in the iterative relation (1.36).
Hence, we will set aside the numerical analysis of the risk-spread option for a future
project.
79
3.2 Delta Hedging with the Risk-Spread Option The introduction of a financial derivative invariably leaves the writer of the option
with the task of hedging against the risk he incurs by assuming a short position in the new
derivative. In the case of a call option on a stock, the hedging of the option writer against
changes in the underlying stock price is commonly known as delta hedging. The delta of
a portfolio is the first derivative of the portfolio with respect to the stock price.
Moreover, the portfolio is said to be delta neutral when it is insensitive to changes in the
stock price, that is, when it has a delta of zero. If we denote the delta of a call option by
∆, then a portfolio consisting of a short position in the option and ∆ shares of the
underlying stock is delta neutral. In fact, denoting the call option by C and the portfolio
value by Π , we find C SΠ = − + ∆ which implies 0.S
∂Π= −∆ + ∆ =
∂
For more exotic derivatives, delta hedging refers to the act of protecting the writer of
the option against changes in various risk factors. Unlike the call option on a tradable
security such as a stock, the risk-spread option must be hedged against the untradable risk
factors that drive the interest rates. However, there are numerous traded bonds that are
affected by the interest rate risk factors. Assuming that the number of risk factors is d,
then a portfolio of d distinct bonds together with a short position in the risk-spread option
that is delta neutral can be constructed.
Let C denote the risk-spread option and ii
C
x
∂∆ ≡
∂ for each of the d risk factors.
Consider d distinct bonds jB with jij
i
B
x
∂δ ≡
∂ and a portfolio with value Π . The hedging
problem is to find the amount jh of bond jB holdings such that the portfolio is delta
neutral. The portfolio value is given by
d
j ji 1
C h B=
Π = − + ∑ . (2.1)
80
This implies that the bond holdings may be determined by setting
d
i j ijj 1i
h 0x =
∂Π= −∆ + δ =
∂ ∑ (2.2)
for each i. We represent this by Dh = ∆ , where
( )d
ij i , j 1D
=≡ δ , (2.3)
( )T
1 dh h ,...,h≡ , (2.4)
and
( )T
1 d,...,∆ ≡ ∆ ∆ . (2.5)
Hence, the vector of bond holdings is given by 1h D−= ∆ , provided that D is invertible.
Returning to the example of the previous sections, consider the case of two risk
factors. We will construct a hedging strategy using the risky and riskless bonds. Fix
[ ]t 0,T∈ and recall the prices of the risky and riskless bond from Section 1.4:
( )B t,T ( ) ( )( )d
2t
ˆˆ1 2 V exp V r−τ τ= − α − +ατ (2.6)
and
( )B t,T% [ ] ( )( ) ( )( )d2
ttˆ1 B t,T 1 2 V exp V
−
τ τν>= − β − λ +βτ% . (2.7)
where V
Vˆ1 2 Vτ
ττ
≡− α
,V
Vˆ1 2 Vτ
τ
τ
≡− β
% , and ( )21V 1 e
2− θτ
τ ≡ −θ
. Also, we recall that
2
t t
1r X
2= σ (2.8)
and
2
t t
1X
2λ = σ% . (2.9)
Hence,
( ) ( )ii
B ˆt,T x V B t,Tx τ
∂= −σ
∂ (2.10)
and
( ) ( )ii
Bt,T x V B t,T
x τ∂
= −σ∂
% % % . (2.11)
where tX x= almost surely.
81
From (2.10) and (2.11), we see that the matrix D in this simple example is singular.
Hence, under the assumption that that there is more than one risk factor, we see that it is
not possible to simultaneously hedge the risk-spread option against all of the risk factors
using a portfolio consisting of only riskless and risky bonds. Consequently, we must add
a different type of interest rate derivative to our portfolio to make it delta neutral.
On the other hand, if we assume that there is a single risk factor, then it follows from
(2.2), (2.10), and (2.11) that we may hedge the risk-spread option by holding Bth riskless
bonds, or Bth% risky bonds, where
( )
Bth ˆxV B t,Tτ
∆= −
σ (2.12)
and
( )
Bth
xV B t,Tτ
∆= −σ
%%% . (2.13)
Alternatively, we may be able to hedge the risk-spread option against the riskless spot
rate directly, if it can be shown that the risk-spread option only depends on the risk
factors through the riskless spot rate. In this case, we compute
( )BV B t,T
r τ∂
= −∂
(2.14)
and
( ) ( )Bt,T V B t,T
rτ
∂ σ= −
∂ σ
% % % % (2.15)
from which we deduce that holdings of either
( )
Bth
V B t,Tτ
∆= − (2.16)
in the riskless bond, or
( )
Bth
V B t,Tτ
σ∆= −σ
%% %% (2.17)
in the risky bond will provide the desired hedge.
82
3.3 Numerical Properties of the Yield Curve Recall that the state-variable process tX from Section 1.4 satisfies
t t tdX dW X dt= − θ (3.1) for some positive parameter θ, where tW is a d-dimensional Brownian motion on
( ), ,PΩ F . Continuing with the example of Section 1.4, we also recall that the riskless
spot rate and risk spread are given by
2
t t
1r X
2= σ (3.2)
and
2
t t
1X
2λ = σ% , (3.3)
respectively, where ˆ ˆ4 ( )σ ≡ α θ − α , ˆ ˆ4 ( )σ ≡ β θ − β% , ˆd
αα ≡ , and ˆ
d
ββ ≡ are positive
constants. We will show that tr follows the Cox-Ingersoll-Ross (CIR) process given by
( )t t t tdr a b r dt 2 r dM= − + σ (3.4)
for some positive constants a and b, where tM is a one-dimensional Brownian
Motion on ( ), ,PΩ F with respect to the natural filtration tG of tW (Elliot & Kopp, 1999).
In fact, from ˆIto's lemma we deduce from (3.1) and (3.2) that
d d
i i it t t t
i 1 i 1
dr X dX d X2= =
σ= σ +∑ ∑ (3.5)
d
2i it t t
i 1
dX dW X dt dt
2=
σ= σ −θσ +∑
i idt t
t ti 1 t
d X dW2 r dt X
2 X=
σ = − θ + σ ∑
( )t t ta b r dt 2 r dM= − + σ ,
where a 2≡ θ , d
b4
σ≡
θ, and
i is s
ts
d
i 1
t
0
X dWM
X=≡ ∑∫ .
83
It remains to be shown that tM is a one-dimensional Brownian Motion on ( ), ,PΩ F
with respect to tG . We first assert that tM is a continuous martingale. In fact, we have
2i
isP s
s
T
0
XE dW T
X
≤ < ∞ ∫ (3.6)
for every i. Furthermore, it follows from ˆIto's lemma that
2i
2 st s s s s s ss
s
d
i 1
t t t tt
00 0 0 0
XM 2 M dM d M 2 M dM ds 2 M dM t
X=
= + = + = +
∑∫∫ ∫ ∫ ∫ . (3.7)
We deduce that 2tM t− is a martingale from which we conclude that tM is a standard
Brownian motion with respect to tG . Hence, we have shown that tr follows the CIR
process given by (3.4). Similarly, we have that
( )t t t td a b dt 2 dMλ = − λ + σλ% % , (3.8)
where d
b4
σ≡
θ%% .
The CIR process has two properties that correspond with empirical spot rates and risk
spreads. First, it is shown in (Lamberton & Lapeyre, 1996) that the CIR processes (3.4)
and (3.8) are almost surely positive, provided that d ≥ 2 , 0r 0> , and 0 0λ > almost
surely. Unfortunately, in the one dimensional case we have that the probability that these
CIR processes vanish for infinitely many times is one.
The second ideal property of the CIR process is that of mean reversion. Consider the
following ordinary differential equation:
( )dx a b x dt= − . (3.9)
The solution of (3.9) is given by
( ) ( )x t exp at b= − + . (3.10)
84
This function exponentially decays to the mean reversion level b at the mean reversion
rate a as t tends to infinity. Comparing this with (3.4) and (3.8), we deduce that the CIR
process has a mean-reverting drift term.
We conclude this chapter with a graphical comparison of the yield curves of the
riskless and risky bonds. Recall from Section 1.4 that the riskless and risky yield curves
are given by
( ) ( ) ( )t
1 1 dˆ ˆY t,T lnB t,T V r ln 1 2 V2τ τ
≡ − = + α τ + − α τ τ (3.11)
and
( ) ( ) [ ] ( ) ( )tt
1 1 d ˆY t,T lnB t,T 1 Y t,T V ln 1 2 V2
τ τν >
≡ − = + λ +βτ+ − β τ τ % % % , (3.12)
respectively. For fixed t, we see from the definitions following (2.7) that
( ) ( )( ) ( ) ( )( )t
1 exp 2 ˆ1 dY t,T r ln 1 1 exp 2
ˆ ˆ2 2 exp 2 2
− − θτ α = + α τ + − − − θτ τ θ − α + α − θτ θ (3.13)
and
( ) [ ] ( )tY t,T 1 Y t,Tν>=% (3.14)
[ ]( )
( ) ( )( )( )tt
ˆ1 exp 21 d1 r ln 1 1 exp 2
ˆ ˆ 22 2 exp 2ν>
− − θτ β + +βτ+ − − − θτ τ θθ − β + β − θτ
.
It follows that
( )limY t,Tτ→∞
= α (3.15)
and ( )limY t,T
τ→∞= α + β% . (3.16)
In the graphs below, we present the initial riskless and risky yield curves for various
parameter values. In particular, we see that the rate of convergence in (3.15) and (3.16)
depends on the values of θ, ˆθ − α , and ˆθ − β .
85
Fig. 1 The Risky and Riskless YTM with d = 5, θ = .05, α = .09, and β = .04
4.50%
6.50%
8.50%
10.50%
12.50%
14.50%
16.50%
18.50%
20.50%
22.50%
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Number of Years Until Maturity
RISKY YTM 20.8% RISKY YTM 16.8% RISKY YTM 12.81% RISKY YTM 8.81%
RISKLESS YTM 11% RISKLESS YTM 9% RISKLESS YTM 7% RISKLESS YTM 5%
86
Fig. 2 The Risky and Riskless YTM with d = 5, θ = .05, α = .09, and β = .04
4.50%
6.50%
8.50%
10.50%
12.50%
14.50%
16.50%
18.50%
20.50%
22.50%
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Number of Years Until Maturity
RISKY YTM 20.8% RISKY YTM 16.8% RISKY YTM 12.81% RISKY YTM 8.81%
RISKLESS YTM 11% RISKLESS YTM 9% RISKLESS YTM 7% RISKLESS YTM 5%
87
Fig. 3 The Risky and Riskless YTM with d = 5, θ = .05, α = .09, and β = .04
4.50%
6.50%
8.50%
10.50%
12.50%
14.50%
16.50%
18.50%
20.50%
22.50%
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Number of Years Until Maturity
RISKY YTM 19.91% RISKY YTM 15.91% RISKY YTM 11.92% RISKY YTM 7.92%
RISKLESS YTM 11% RISKLESS YTM 9% RISKLESS YTM 7% RISKLESS YTM 5%
88
Fig. 4 The Risky and Riskless YTM with d = 5, θ = .05, α = .09, and β = .04
4.50%
6.50%
8.50%
10.50%
12.50%
14.50%
16.50%
18.50%
20.50%
22.50%
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Number of Years Until Maturity
RISKY YTM 19.91% RISKY YTM 15.91% RISKY YTM 11.92% RISKY YTM 7.92%
RISKLESS YTM 11% RISKLESS YTM 9% RISKLESS YTM 7% RISKLESS YTM 5%
89
Fig. 5 The Risky and Riskless YTM with d = 5, θ = .05, α = .09, and β = .04
4.50%
6.50%
8.50%
10.50%
12.50%
14.50%
16.50%
18.50%
20.50%
22.50%
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Number of Years Until Maturity
RISKY YTM 19.91% RISKY YTM 15.91% RISKY YTM 11.92% RISKY YTM 7.92%
RISKLESS YTM 11% RISKLESS YTM 9% RISKLESS YTM 7% RISKLESS YTM 5%
90
Fig. 6 The Risky and Riskless YTM with d = 5, θ = .05, α = .09, and β = .04
4.50%
6.50%
8.50%
10.50%
12.50%
14.50%
16.50%
18.50%
20.50%
22.50%
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Number of Years Until Maturity
RISKY YTM 19.91% RISKY YTM 15.91% RISKY YTM 11.92% RISKY YTM 7.92%
RISKLESS YTM 11% RISKLESS YTM 9% RISKLESS YTM 7% RISKLESS YTM 5%
91
CHAPTER 4 SUMMARY AND CONCLUSIONS
4.1 Summary of Results
The potential theoretic framework developed in Chapter 1 extends the work of Rogers
(1997) to include the case of defaultable bonds. Among the numerous examples of both
the riskless and risky spot rates of interest that may be generated from this procedure, the
familiar, tractable example of the Ornstein-Uhlenbeck process was used as an illustration.
This example is realistic in the sense that the resulting Cox-Ingersoll-Ross model of the
spot rates is strictly positive, provided the dimension of the driving Markov process is at
least two. Furthermore, this model exhibits the mean-reverting behavior that has been
observed empirically.
This example was carried forward to the Cauchy problem for the risk-spread option
treated in Chapter 2. The potential theoretic parametrix method used to develop a series
solution to the Cauchy problem represents a significant contribution to mathematical
literature. To gain an appreciation for the modern language, the interested reader should
compare this method with the parametrix method outlined in the first chapter of
(Friedman, 1964).
In Chapter 3, the Fourier and Laplace transforms were used to derive an expression
for the risk-spread option. This removed the temporal convolution in the Gaussian
92
potential term; however, the remaining spatial convolution forced us to postpone the
numerical analysis of our solution. In the second section of this chapter, it was shown
how to delta hedge the risk-spread option using a portfolio of riskless and risky bonds.
Finally, we examined the properties of the spot rates through the graphs of the yield
curves in Section 3.3.
4.2 Future Projects and Model Extensions
The work presented in this dissertation has laid down the foundation for future
development in pricing derivatives on the risky spot rate such as the risk-spread option.
It also presents a new framework within which various models of the risky spot rate may
be produced. One way to extend the model developed in this dissertation is to relax the
independence assumption used to obtain the representation (1.3.41) of the risky bond
price. Another extension that might prove interesting is to allow the mean of the
exponential random variable used in modeling the time of default to be a free parameter.
The procedure outlined in Section 1.2 for constructing a positive supermartingale to
model the state-price density may be modified by rewriting
( )( )
ttt
0
f Xe
f X−αζ = (2.1)
as
( )( )
tt t
0
f XM
f Xζ = , (2.2)
where tM is a multiplicative functional such as ( )( )t
s0exp X ds− α∫ . Naturally, this could
also be done to the risky state-price density.
93
The simple Ornstein-Uhlenbeck process that we have considered is one of many
Markov processes that may be used to obtain tractable results. A slight modification of
this example is to replace the parameter θ with a d d× matrix to allow for interactions
between the components of the Markov process. In addition, the function f in (2.1) could
be redefined using a symmetric positive-definite matrix A and a vector dc∈R :
( )Tf(x) exp (x c) A(x c)= − − . (2.3)
The interested reader is referred to (Rogers, 1997) for more examples.
Regardless of the state-variable used in this framework, the calculation of the
effective risk-spread insurance level is a nice problem that naturally follows from the
risk-spread option. Recall from Section 1.5 that for a given risk-spread insurance level γ,
the effective risk-spread insurance level γ% is the minimum rate of return that the holder
of a risky bond and a risk-spread option expects to receive. It was suggested in Section
1.5 that γ% be defined by
( )( ) ( )( ) ( ) ( )( )P T P T T 0E E r ln B 0,T CT
+ + ∂γ − λ = γ − λ − − + π
∂%% . (2.4)
An investigation of this relation will make an interesting future project.
The final extension that will be discussed follows the work of Lando (1998) in the
area of risk classes. By constructing a different risky state-price density for each risk
class, the Markov chain model used by Lando could be adapted to our potential theoretic
framework. In fact, the rate of transition between risk classes should follow a process
that is similar to the risk-spread. After establishing a model of these transition processes,
it would be interesting to price options on the transition between risk classes using the
potential theoretic parametrix method of Chapter 2.
94
REFERENCES
Björk, T. (1997) Interest rate theory. In: Financial Mathematics, Bressanone, 1996, W. Runggaldier, ed. Lecture Notes in Mathematics 1656. Springer-Verlag, New York, 53-122.
Blumenthal, R. M. & Getoor, R. K. (1968) Markov Processes and Potential Theory.
Academic Press, New York. Chung, K. L. (1974) A Course in Probability Theory. Academic Press, New York. Chung, K. L. & Williams, R. J. (1990) Introduction to Stochastic Integration.
Birkhäuser, Boston. Delbaen, F. & Scachermayer, W. (1994) A general version of the fundamental theorem of
asset pricing. Mathematische Annalen 300, 463-520. Delbaen, F. & Scachermayer, W. (1998) The fundamental theorem of asset pricing for
unbounded stochastic processes. Mathematische Annalen 312, 215-250. Elliot, R. J. & Kopp, P. E. (1999) Mathematics of Financial Markets. Springer-Verlag,
New York. Friedman, A. (1964) Partial Differential Equations of Parabolic Type. Prentice-Hall,
Englewood Cliffs, NJ. Hull, J. C. (1997) Options, Futures, and Other Derivatives. Prentice-Hall, Upper Saddle
River, NJ. Karatzas, I. & Shreve, S. E. (1991) Brownian Motion and Stochastic Calculus. Springer-
Verlag, New York. Lamberton, D. & Lapeyre, B. (1996) Introduction to Stochastic Calculus Applied to
Finance. Chapman & Hall, New York. Lando, D. (1998) On Cox processes and credit risky securities. Review of Derivatives
Research 2, 99-120.
95
Musiela, M. & Rutkowski, M. (1998) Martingale Methods in Financial Modelling. Springer-Verlag, New York.
Renardy, M. & Rogers, R. C. (1992) An Introduction to Partial Differential Equations.
Springer-Verlag, New York. Revuz, D. & Yor, M. (1991) Continuous Martingales and Brownian Motion. Springer-
Verlag, New York. Rogers, L. C. G. (1997) The potential approach to the term structure of interest rates and
foreign exchange rates. Mathematical Finance 7 (2), 157-176. Stoer, J. & Bulirsch, R. (1993) Introduction to Numerical Analysis. Springer-Verlag,
New York.
96
BIOGRAPHICAL SKETCH
At the age of seventeen, I finished my junior year of high school as well as sixty
hours of college credits at the local community college. These credits included the two-
semester freshman physics and chemistry sequences, calculus I, II, and III, and a first
course in differential equations. Since I did not require a senior year of high school, I
entered the University of Florida in the fall of 1990 without a high school diploma. I
chose to major in physics because of my passion for scientific truth and wanted a deeper
understanding of this universe into which I have been born. Since I entered the university
as a junior, I received a bachelor’s degree in May of 1992 as I turned nineteen years of
age.
After obtaining a bachelor’s degree three years ahead of schedule, I realized that I
might enjoy a couple of semesters away from academia while I considered the
appropriate path to follow in graduate school. In the fall of 1992, I took a job at 102.5
FM in St. Petersburg, where I worked until the following summer. I still had not decided
what to do about graduate school, but I was ready to leave the radio station and return to
academia in June of 1993. During that summer I worked on a project in a solid state
physics lab under the guidance of Dr. Tanner at the University of Florida. Although I
enjoyed experimental physics, I realized that a deep understanding would only come
from a more theoretical perspective.
97
In the fall of 1993, I was back at UF as a post-baccalaureate student taking the few
remaining classes needed to get a second bachelor’s degree in mathematics. During the
Spring semester of 1994, I had a revelation in Dr. Groisser’s Advanced Calculus II class.
It became apparent to me that mathematics is nothing less than the language of physics.
Furthermore, I knew that I had to obtain at least a master’s in mathematics before I could
understand physics. So, after obtaining a bachelor’s degree in mathematics in 1994, I
proceeded to enter graduate school in mathematics the following fall.
The first two years of graduate school seemed to go by rather quickly, and I received
a master’s degree in the spring of 1996. During the second year, I was influenced the
most by Dr. Dinculeanu in his two-semester course on measure and integration theory.
Dr. Dinculeanu taught me how to write a mathematical proof. His style and teaching
mannerisms have left lasting impressions in me and guided me toward my ultimate
decision to work in the area of probability theory and stochastic processes.
In the spring of my third year of graduate school, I became somewhat disillusioned
with the poor job outlook for a mathematical physicist, and took a class called The
Mathematics of Financial Derivatives with my current advisor Dr. Glover. As I was
maturing mathematically, I realized that mathematics is more than the language of
physics. Mathematics is the language of order and structure itself. All structural
concepts begin as vague, creative thoughts which can only be tested when measured, and
can only be measured when expressed in the proper mathematical framework. Therefore,
I decided that I could write a dissertation in mathematical finance as well as gain an
understanding of the probabilistic nature of the universe. Of course, the prospect of a
starting six-figure salary on Wall Street had nothing to do with this decision at all.