a

r
X
i
v
:
1
1
0
1
.
1
1
4
9
v
1


[
m
a
t
h
.
P
R
]


6

J
a
n

2
0
1
1
The Annals of Applied Probability
2011, Vol. 21, No. 1, 332350
DOI: 10.1214/10-AAP698
c Institute of Mathematical Statistics, 2011
BOUNDARY CONDITIONS FOR THE SINGLE-FACTOR TERM
STRUCTURE EQUATION
By Erik Ekstr om
1,2
and Johan Tysk
1
Uppsala University
We study the term structure equation for single-factor models
that predict nonnegative short rates. In particular, we show that the
price of a bond or a bond option is the unique classical solution to a
parabolic dierential equation with a certain boundary behavior for
vanishing values of the short rate. If the boundary is attainable then
this boundary behavior serves as a boundary condition and guaran-
tees uniqueness of solutions. On the other hand, if the boundary is
nonattainable then the boundary behavior is not needed to guaran-
tee uniqueness but it is nevertheless very useful, for instance, from a
numerical perspective.
1. Introduction. When calculating prices of dierent interest rate deriva-
tives, such as bonds and bond options, stochastic methods seem to be more
commonly used than PDE methods; compare for instance [5]. This is in
contrast to the case of stock option pricing where PDE methods are used
extensively, in particular for low dimensional problems. We believe that one
possible explanation for this phenomenon is that the correspondence be-
tween the risk neutral valuation approach and the pricing equation (hence-
forth referred to as the term structure equation) with appropriate boundary
conditions is not fully developed.
For the BlackScholes equation, the boundary condition is of Dirichlet
type which corresponds to the underlying asset being absorbed if reaching
zero; compare [13]. In contrast, for most interest rate models this is not
the case since the short rate typically would not stay zero if the value zero
is reached. Consequently, it is not clear what boundary conditions should
Received November 2008; revised December 2009.
1
Supported by the Swedish Research Council (VR).
2
Supported by the Swedish National Graduate School of Mathematics and Computing
(FMB).
AMS 2000 subject classications. Primary 91B28; secondary 35A05, 35K65, 60J60.
Key words and phrases. The term structure equation, degenerate parabolic equations,
stochastic representation.
This is an electronic reprint of the original article published by the
Institute of Mathematical Statistics in The Annals of Applied Probability,
2011, Vol. 21, No. 1, 332350. This reprint diers from the original in pagination
and typographic detail.
1
2 E. EKSTR

OM AND J. TYSK
be specied for the term structure equation. In fact, the recent monograph
[6] draws attention to this issue in a section entitled The thorny issue of
boundary conditions. Moreover, in [12], two dierent solutions to the term
structure equation are presented in the case of the CIR-model. They are
both bounded and with the same terminal condition, but naturally they ex-
hibit dierent boundary behavior for vanishing interest rates. The authors of
that paper take the view that these solutions represent alternative possible
prices. We on the other hand regard only the solution given by the stochas-
tic representation as the price, and the purpose of the present paper is to
identify the boundary condition that this stochastic representation satises.
We consider the classical case of a single-factor model predicting non-
negative values of the short rate. More precisely, the rate X(t) is modeled
directly under the pricing measure as
dX(t) =(X(t), t) dt +(X(t), t) dW,
where W is a Brownian motion and (0, t) = 0 and (0, t) 0. As indicated
above, the option price u corresponding to a payo function g is given using
risk neutral valuation by
u(x, t) =E
x,t
[e

T
t
X(s) ds
g(X(T))].
Note that if the payo g 1, then bond prices are obtained. Also note that
the set-up covers the case of bond options. The corresponding term structure
equation is
u
t
(x, t) +
1
2

2
(x, t)u
xx
(x, t) +(x, t)u
x
(x, t) =xu(x, t)
with terminal condition u(x, T) =g(x). If the price u is twice continuously
dierentiable up to and including the boundary x = 0, then plugging in x = 0
in the equation would give the boundary behavior
u
t
(0, t) +(0, t)u
x
(0, t) = 0. (1)
Even though there is extensive literature on equations with degenerating
coecients, compare the classical reference [14], the C
1
-regularity of u at
the boundary is not available in the generality that is needed here. In fact,
one of the solutions in the example in [12] referred to above is bounded and
continuous, but fails to be C
1
up to the boundary. (This solution, however,
is not the one given by stochastic representation.)
In the present paper, sucient regularity of the option price u for (1) to
hold is established using the Girsanov theorem and scaling arguments; see
Sections 3 and 4. Let us emphasize that (1) is the correct boundary behavior
of the option price regardless if the boundary is hit with positive probability
or not. If the boundary can be reached with positive probability, then this
boundary behavior serves as a boundary condition for the term structure
THE SINGLE-FACTOR TERM STRUCTURE EQUATION 3
equation and guarantees uniqueness. On the other hand, if the boundary
is reached with probability zero, equation (1) is not needed to identify the
solution given by the stochastic representation, and the term boundary
condition is perhaps misleading. However, it is still valid and certainly
useful, for instance, from a numerical perspective. Indeed, the results of
the present paper have already been implemented numerically in [7]. For
simplicity of the exposition, we will refer to (1) as the boundary condition
regardless if the boundary can be reached or not.
In Section 2 the assumptions on the model and our main result, Theo-
rem 2.3, are presented. In Sections 3, 4 and 5, we establish regularity prop-
erties of the value function, and use them to prove Theorem 2.3. Finally, in
Section 6 we also provide the link between the stochastic problem and the
term structure equation for models dened on the whole real line.
2. Assumptions and the main result. When studying the term structure
equation on the positive real axis, we consider models which are specied
so that the short rate automatically stays nonnegative, that is, there is no
need to impose any boundary behavior of the underlying diusion process.
Throughout Sections 25 we work under the following hypothesis:
Hypothesis 2.1. The drift C([0, ) [0, T]) is continuously dier-
entiable in x with bounded derivative, and (0, t) 0 for all t. The volatility
C([0, ) [0, T]) is such that (x, t) :=
1
2

2
(x, t) is continuously dier-
entiable in x with a Holder continuous derivative, and (x, t) = 0 if and only
if x = 0. The functions , and
x
are all of, at most, linear growth:
|(x, t)| +|(x, t)| +|
x
(x, t)| C(1 +x) (2)
for all x and t. The payo function g : [0, ) [0, ) is continuously dif-
ferentiable with both g and g

bounded.
Let W be a standard Brownian motion on a ltered probability space
(, F, (F
t
)
t0
, P). Since is continuously dierentiable, is locally Holder
(1/2) in x. It follows that there exists a unique strong solution X(t) to
dX(t) =(X(t), t) dt +(X(t), t) dW (3)
for any initial point x 0; compare Section IX.3 in [15]. Moreover, it follows
from monotonicity results for stochastic dierential equations with respect
to the drift coecient (e.g., Theorem IX.3.7 in [15]), that X remains nonneg-
ative at all times. Indeed, if is replaced with 0, then the corresponding
solution to (3) is absorbed at zero, so X is nonnegative if (0, t) 0. The
option price u: [0, ) [0, T] [0, ) corresponding to a payo function
g : [0, ) [0, ) is given by
u(x, t) =E
x,t
[e

T
t
X(s) ds
g(X(T))], (4)
4 E. EKSTR

OM AND J. TYSK
where the indices indicate that X(t) =x. As described in the Introduction,
the corresponding term structure equation is given by
u
t
(x, t) +
1
2

2
(x, t)u
xx
(x, t) +(x, t)u
x
(x, t) =xu(x, t) (5)
for (x, t) (0, ) [0, T), with terminal condition
u(x, T) =g(x). (6)
Moreover, by formally inserting x = 0 in the equation we get the boundary
condition
u
t
(0, t) +(0, t)u
x
(0, t) = 0 (7)
for all t [0, T), since (0, t) = 0 by assumption. One of the main eorts in
this paper is to show that the option price u is continuously dierentiable up
to the boundary x = 0, and that it indeed satises the boundary condition
(7) in the classical sense.
Definition 2.2. A classical solution to the term structure equation is
a function v C([0, ) [0, T]) C
1
([0, ) [0, T)) C
2,1
((0, ) [0, T))
which satises (5), (6) and (7).
Our main result in this article is the following:
Theorem 2.3. In addition to Hypothesis 2.1, also assume that Assump-
tion 3.1 below holds. The option price u as given by (4) is then the unique
bounded classical solution to the term structure equation.
Example. Classical short rate models such as the CoxIngersollRoss
model
dX(t) = (a bX(t)) dt +
_
X(t) dW, (8)
and the Dothan model
dX(t) =aX(t) dt +X(t) dW, (9)
have boundary conditions at x = 0 that are immediate to write down. These
conditions are
u
t
+au
x
= 0 and u
t
= 0,
respectively. We note that the boundary condition u
t
= 0 for the Dothan
model means that u is constant along the boundary, that is, u(0, t) =g(0).
This is the same type of boundary condition that appears for options on
stocks in [13], which can be explained by the fact that the Dothan model
THE SINGLE-FACTOR TERM STRUCTURE EQUATION 5
is a geometric Brownian motion. Theorem 2.3 also covers the HullWhite
model
dX(t) = (a(t) b(t)X(t)) dt +(t)
_
X(t) dW (10)
(which is a time-dependent generalization of the CoxIngersollRoss model),
and models of, for example, the form
dX(t) = (b aX(t)) dt +X

(t) dW, (1/2, 1], (11)

which also would be natural to consider for bond pricing.
Remark. It seems that many of the classical models for the short rate
are proposed for their analytical tractability. In particular, if the drift
and the diusion coecient
2
are ane, then the model admits an ane
term structure. It is easy to check that known explicit formulas for bond
prices and bond options satisfy the boundary condition (7). In particular,
for models admitting an ane term structure, it is a consequence of the
associated Riccati equations (see [4], equation 22.25) that these boundary
conditions are fullled.
Remark. The assumption that g is continuously dierentiable is sat-
ised for bonds, but not in general for bond options. However, using the
Markov property, Theorem 2.3 readily extends to bounded Lipschitz payos
provided one can show that the corresponding option price x u(x, T )
is continuously dierentiable on (0, ) for any >0. The regularizing eect
of parabolic equations guarantees continuous dierentiability on (0, ), so
the main diculty is to show that x u
x
(x, T ) is continuous also at
0. If the model is convexity preserving, this is easily done in certain cases
including, for example, call options written on bond prices. (Note that the
corresponding payo function g is bounded since bond prices are bounded.)
For details on which short rate models are convexity preserving, see [8]. To
our knowledge, all models used in practice belong to this class.
One should note that the dierentiability of the option price up to the
boundary x = 0 is not valid without some Lipschitz bound of g at 0. To see
this, consider the contract function g(x) = e
2

x
. Then, with (x, t) =
1
2
x
and (x, t) =

2x, it is straightforward using the Ito formula to show that

the process
Y (s) =e

s
t
X(r) dr
g(X(s))
is a martingale. Consequently, the option price u is given by u(x, t) =g(x)
for all t, which fails to be a classical solution to the term structure equation
since it is not dierentiable at x = 0. One might argue, though, that the
boundary condition u
t
(0, t) +(0, t)u
x
(0, t) = 0 is satised in a weak sense.
6 E. EKSTR

OM AND J. TYSK
The proof of Theorem 2.3 is carried out in several steps.
Proof of uniqueness. Let v
1
and v
2
be two bounded classical solu-
tions to the term structure equation, and dene
v(x, t) =v
1
(x, T t) v
2
(x, T t).
Then v(x, t) is a bounded solution to
_
_
_
v
t
=
1
2

2
v
xx
+v
x
xv,
v(x, 0) = 0,
v
t
(0, t) =(0, t)v
x
(0, t).
(12)
Now consider the function
h(x, t) = (1 +x)e
Mt
,
where M is a positive constant. For M large enough, depending on (0, t)
and the growth rate of , h is a super-solution to (12) which tends to innity
at spatial innity. Thus, according to the maximum principle, the function
v is bounded above by h and below by h for any >0. It follows that
v 0, which demonstrates uniqueness of bounded classical solutions to the
term structure equation.
Proof of continuity. To show that u is continuous, denote by X
x,t
the solution to (3) with initial condition X
x,t
(t) =x. Let (x, t) and (y, r) be
two points in [0, ) [0, T]. Then, if r t, we have
|u(y, r) u(x, t)| E[e

T
r
X
y,r
(s) ds
|g(X
y,r
(T)) g(X
x,t
(T))|]
+E[g(X
x,t
(T))|e

T
r
X
y,r
(s) ds
e

T
t
X
x,t
(s) ds
|]
E[|g(X
y,r
(T)) g(X
x,t
(T))|] (13)
+C
_
T
t
E[|X
y,r
(s) X
x,t
(s)|] ds
+C
_
t
r
E[X
y,r
(s)] ds
for some constant C, where we have used that g is bounded. A similar
expression can be derived if r > t. It follows from Remark 1 in Section 8,
Chapter 2 in [11] that X
y,r
(t) x in L
2
as (y, r) (x, t). Therefore, from
Theorem 2.1 in [2] we have
E
_
sup
tsT
(X
y,r
(s) X
x,t
(s))
2
_
0
as (y, r) (x, t). (Theorem 2.1 in [2] also holds in the case of random starting
points.) Since g is assumed continuous and bounded, all three terms on the
THE SINGLE-FACTOR TERM STRUCTURE EQUATION 7
right-hand side of (13) tend to 0 as (y, r) (x, t). Thus u is continuous on
[0, ) [0, T].
Proof that u C
2,1
((0, ) [0, T)) and satisfies (5). For a given
point (x, t) (0, ) [0, T), let
R= (x
1
, x
2
) [t
1
, t
2
) (0, ) [0, T)
be a rectangle which contains (x, t), where x
1
>0. Since u is continuous, it
follows from standard parabolic theory, see [9], that there exists a unique
solution U C
2,1
(R) to the boundary value problem
_
U
t
+
1
2

2
U
xx
+U
x
xU = 0, in R,
U =u, on
p
R,
where
p
R= ([x
1
, x
2
] {t
2
}) ({x
1
, x
2
} [t
1
, t
2
]) is the parabolic boundary
of R. From Itos formula, the process
Z(s) =e

s
t
X
x,t
(r) dr
U(X
x,t
(s), s)
is a martingale on the time interval [t,
R
], where

R
= inf{s t : X
x,t
(s) / R}
is the rst exit time from the rectangle R. Therefore,
U(x, t) =E[e

R
t
X
x,t
(r) dr
u(X
x,t
(
R
),
R
)] =u(x, t),
where the second equality follows from the strong Markov property. Con-
sequently, u C
2,1
((0, ) [0, T)). Since u U on R, we also see that u
satises (5).
It remains to show that u is continuously dierentiable up to the spatial
boundary x = 0, and that it satises the boundary condition (7). This is
done in Sections 35.
3. Continuity of the rst spatial derivative. In this section we investi-
gate regularity of the spatial derivative u
x
at the boundary x = 0. To do
this we study the stochastic representation of the terminal value problem
obtained by formally dierentiating the term structure equation. We show
that this stochastic representation indeed is the derivative of u and that it
is continuous.
Recall that
x
is assumed to be continuous on [0, ) [0, T], where
(x, t) =
1
2

2
(x, t). Let the process Y be modeled by the stochastic dieren-
tial equation
dY (t) = (
x
+)(Y (t), t) dt +(Y (t), t) dW. (14)
Rather than specifying precise conditions under which (14) has a unique
solution, we simply assume what we need.
8 E. EKSTR

OM AND J. TYSK
Assumption 3.1. The coecients and are such that, path-wise,
uniqueness holds for equation (14).
Remark. Note that Assumption 3.1 holds for example if is twice
continuously dierentiable in space, since then the drift
x
+ is locally
Lipschitz continuous. Moreover, if and are time-independent, then it
follows from [1, 3] and Section IX.3 in [15] that Assumption 3.1 automatically
holds. Thus the CoxIngersollRoss model (8), the Dothan model (9), the
HullWhite model (10) and the model (11) all satisfy Assumption 3.1.
Also note that since (0, t) = 0, we have
x
(0, t) 0. Thus Y remains
nonnegative since it has the same volatility as X but a larger drift at 0.
Next, dene the function v by
v(x, t) =E
_
g

(Y (T)) exp
__
T
t

x
(Y (s), s) Y (s) ds
__
(15)
E
__
T
t
exp
__
s
t

x
(Y (r), r) Y (r) dr
_
u(Y (s), s) ds
_
,
where Y is the solution to (14) with initial condition Y (t) =x.
If the term structure equation (5) is formally dierentiated with respect
to x, then the derivative u
x
satises
(u
x
)
t
+(u
x
)
xx
+(
x
+)(u
x
)
x
+(
x
x)u
x
u = 0
with terminal condition u
x
(x, T) = g

(x). The function v dened in (15) is

the corresponding stochastic representation. In Theorem 3.4 below we show
that v indeed equals the spatial derivative of u.
Proposition 3.2. The function v(x, t) is continuous on [0, ) [0, T].
Proof. The result follows along the same lines as the continuity of u
above. Indeed, let (x
n
, t
n
) converge to (x, t), where t
n
t, and let Y and Y
n
be dened by
_
dY (s) = (
x
+)(Y (s), s) ds +(Y (s), s) dW,
Y (t) =x
and
_
dY
n
(s) = (
x
+)(Y
n
(s), s) ds +(Y
n
(s), s) dW,
Y
n
(t
n
) =x
n
,
respectively. Also dene
I(s) := exp
__
s
t

x
(Y (u), u) Y (u) du
_
THE SINGLE-FACTOR TERM STRUCTURE EQUATION 9
and
I
n
(s) := exp
__
s
tn

x
(Y
n
(u), u) Y
n
(u) du
_
.
Then
|v(x
n
, t
n
) v(x, t)| E[|I
n
(T)g

(Y
n
(T)) I(T)g

(Y (T))|]
+
_
T
t
E[|I
n
(s)u(Y
n
(s), s) I(s)u(Y (s), s)|] ds
+
_
t
tn
E[I
n
(s)u(Y
n
(s), s)] ds.
The rst term and the integrand in the second term are similar to the type
of terms treated when proving the continuity of u. Moreover, the integrand
of the third term is bounded. Thus it follows from bounded convergence that
v is continuous.
We also need a continuity result in the volatility parameter. To formulate
it, let {
n
(x, t)}

n=1
be a sequence of functions satisfying Hypothesis 2.1
uniformly in n, that is, with the same constant C in the bound (2). Moreover,
assume that
n
(x, t) converges to (x, t) and
n
x
converges to
x
uniformly
on compacts as n , where
n
=
1
2
(
n
)
2
. Let u
n
and v
n
be dened as u
and v but using the volatility function
n
instead of . More explicitly,
u
n
(x, t) =E[e

T
t
X
n
(s) ds
g(X
n
(T))]
and
v
n
(x, t) =E
_
g

(Y
n
(T)) exp
__
T
t

x
(Y
n
(s), s) Y
n
(s) ds
__
(16)
E
__
T
t
exp
__
s
t

x
(Y
n
(r), r) Y
n
(r) dr
_
u
n
(Y
n
(s), s) ds
_
,
where X
n
and Y
n
satisfy
_
dX
n
(s) =(X
n
(s), s) ds +
n
(X
n
(s), s) dW(s),
X
n
(t) =x
and
_
dY
n
(s) = (
n
x
+)(Y
n
(s), s) ds +
n
(Y
n
(s), s) dW(s),
Y
n
(t) =x,
respectively.
10 E. EKSTR

OM AND J. TYSK
Proposition 3.3. The functions u and v are continuous in the volatility
parameter. More precisely, u
n
(x, t) u(x, t) and v
n
(x, t) v(x, t) as n
for any xed point (x, t) [0, ) [0, T].
Proof. It follows from Theorem 2.5 in [2] that
lim
n
E
_
sup
s[t,T]
(X(s) X
n
(s))
2
_
= 0.
Therefore,
|u
n
(x, t) u(x, t)| E[|e

T
t
X
n
(s) ds
e

T
t
X(s) ds
|g(X
n
(T))]
+E[e

T
t
X(s) ds
|g(X
n
(T)) g(X(T))|]
C
_
T
t
E[|X(s) X
n
(s)|] ds +E[|g(X
n
(T)) g(X(T))|]
0
as n . Thus u is continuous in the volatility function.
The continuity of v in the volatility function is similar. Indeed, let
I(s) := exp
__
s
t

x
(Y (r), r) Y (r) dr
_
and
I
n
(s) := exp
__
s
t

x
(Y
n
(r), r) Y
n
(r) dr
_
.
Then
|v
n
(x, t) v(x, t)| E[|I
n
(T)g

(Y
n
(T)) I(T)g

(Y (T))|]
+
_
T
t
E[|I
n
(s)u
n
(Y
n
(s), s) I(s)u
n
(Y (s), s)|] ds
+
_
T
t
E[I(s)|u
n
(Y (s), s) u(Y (s), s)|] ds.
The rst term and the integrand of the third term are similar to the terms
appearing in the rst part of the proof. The integrand of the second term
can be dealt with using the fact that each u
n
is Lipschitz continuous in x,
uniformly in n since the Lipschitz property is inherited by the value function.
Thus all terms tend to zero as n , so v is continuous in the volatility.

Theorem 3.4. We have u

x
(x, t) = v(x, t) on [0, ) [0, T]. Conse-
quently, u
x
is continuous on [0, ) [0, T].
THE SINGLE-FACTOR TERM STRUCTURE EQUATION 11
Proof. It suces to prove u
x
(x, 0) =v(x, 0). We rst assume that is
continuously dierentiable in x with a bounded derivative. It then follows
from Section 5.5 in [10] or Section 8 in [11] that the derivative
(t) :=
X(t)
x
of X(t) =X
x,0
(t) with respect to the initial point x exists and is continuous,
and it satises
_
d(t) =(t)
x
(X(t), t) dt +(t)
x
(X(t), t) dW(t),
(0) = 1.
Moreover,
u
x
(x, 0) = E
_
g

(X(T))(T) exp
_

_
T
0
X(s) ds
__
E
_
g(X(T)) exp
_

_
T
0
X(s) ds
__
T
0
(s) ds
_
(17)
=: I
1
I
2
.
We claim that I
i
=J
i
, i = 1, 2, where
J
1
=E
_
g

(Y (T)) exp
__
T
0

x
(Y (s), s) Y (s) ds
__
and
J
2
=E
__
T
0
exp
__
s
0

x
(Y (r), r) Y (r) dr
_
u(Y (s), s) ds
_
,
compare (15) above. Here Y is dened as in (14) with initial condition
Y (0) =x.
To show that I
1
=J
1
, dene a new measure Q on F
T
by dQ=M(T) dP,
where the process M is dened by
M(t) =(t) exp
_

_
t
0

x
(Y (s)) ds
_
. (18)
By Itos formula,
dM(t) =M(t)
x
(X(t)) dW(t),
so M is a martingale since
x
is bounded. In particular, E[M(T)] = 1, so Q
is a probability measure. From Girsanovs theorem it follows that

W(t) =W(t)
_
t
0

x
(X(s)) ds
12 E. EKSTR

OM AND J. TYSK
is a Q-Brownian motion, and
dX = (
x
+)(X(t), t) dt +(X(t), t) d

W.
Here
x
=
x
, so by weak uniqueness, the Q-law of X is the same as the
law of Y under P. Consequently,
I
1
=E
_
g

(X(T))
x
(T) exp
_

_
T
0
X(s) ds
__
=E
Q
_
g

(X(T)) exp
__
T
0

x
(X(s), s) X(s) ds
__
=J
1
.
To prove I
2
=J
2
, note that
I
2
=E
_
g(X(T)) exp
_

_
T
0
X(s) ds
__
T
0
(s) ds
_
=
_
T
0
E
_
exp
_

_
s
0
X(r) dr
_
(s)
E
_
g(X(T)) exp
_

_
T
s
X(r) dr
_

F
s
__
ds
=
_
T
0
E
_
exp
_

_
s
0
X(r) dr
_
(s)u(X
s
, s)
_
ds
by the Markov property. Dene a new measure Q=Q
s
on F
s
by
dQ=M(s) dP,
where M is dened as in (18). Girsanovs theorem yields
E
_
exp
_

_
s
0
X(r) dr
_
(s)u(X
s
, s)
_
=E
Q
_
exp
__
s
0

x
(X(r), r) X(r) dr
_
u(X
s
, s)
_
=E
_
exp
__
s
0

x
(Y (r), r) Y (r) dr
_
u(Y
s
, s)
_
.
Consequently, I
2
=J
2
, which nishes the proof in the case of continuously
dierentiable .
The general case follows by approximation. Let
n
, u
n
and v
n
be as
described before Proposition 3.3, with each
n
being continuously dier-
entiable in x with bounded derivative. From above, we then know that
v
n
(x, t) = u
n
x
(x, t) at all points. Moreover, by Proposition 3.3, v
n
(x, t)
v(x, t) point-wise as n .
THE SINGLE-FACTOR TERM STRUCTURE EQUATION 13
On the other hand, since u
n
converges to u point-wise and is uniformly
bounded, it follows from standard parabolic theory that also u
n
x
converges
to u
x
point-wise for all points (x, t) with x > 0. Consequently, v = u
x
on
(0, ) [0, T]. Since v is continuous on [0, ) [0, T] by Proposition 3.2, it
is easy to check that u
x
(0, t) exists and that we have v =u
x
everywhere on
[0, ) [0, T]. The continuity of u
x
thus follows.
4. An estimate of the second spatial derivative. Since the function v
dened in (15) is continuous, it follows that [by a similar argument as in the
proof that u satises (5)] it indeed solves the dierentiated equation
v
t
=v
xx
+ (
x
+)v
x
+ (
x
x)v u
on (0, ) [0, T). In this section we use interior estimates to show that
v
x
0 as x 0. Since v = u
x
by Theorem 3.4, this shows that the term
u
xx
in (5) approaches zero close to the boundary.
Proposition 4.1. The function v =u
x
satises
lim
(x,t)(0,t
0
)
(x, t)v
x
(x, t) = 0
for any t
0
. Consequently, lim
(x,t)(0,t
0
)
(x, t)u
xx
(x, t) = 0.
Proof. Let {(x
n
, t
n
)}

n=1
(0, ) [0, T) be a sequence of points con-
verging to (0, t
0
), where t
0
[0, T). Dene new coordinates (y, s) by letting
y = kx and s = k(t t
0
), where k is specied more precisely below. Then
the function w dened by
w(y, s) =v(x, t)
satises
w
s
= w
yy
+

w
y
+w +h, (19)
where
(y, s) =
_
y
k
, t
0
+
s
k
_
k,

(y, s) = (
x
+)
_
y
k
, t
0
+
s
k
_
,
(y, s) =
1
k

x
_
y
k
, t
0
+
s
k
_

y
k
2
and
h(y, s) =
1
k
u
_
y
k
, t
0
+
s
k
_
.
14 E. EKSTR

OM AND J. TYSK
Now consider a region R=R
n
which contains the point (x
n
, t
n
), and such
that
1 (x, t)k 2 (20)
in R. Since
x
(x, t) is continuous up to the boundary, the region R in
(y, s)-coordinates does not collapse as n , but it can rather be chosen
to consist of a rectangle of xed size; the location of the rectangle is not
necessarily xed though. In this rectangle, the coecients of the equation
(19) satisfy
1 (y, s) 2,
|

(y, s)| C,
|(y, s)| C
and
|h(y, s)| C/k
for some constant C which is independent of n. Since w(y, s) = v(x, t) we
have that w converges to the constant v(0, t
0
) =u
x
(0, t
0
) uniformly on R as
n . By interior Schauder estimates, w
y
tends to 0 as n . Since
(x, t)v
x
(x, t) = (y, s)w
y
(y, s),
and since (y, s) is bounded on R, the conclusion follows.
5. The time derivative at the boundary. It follows from Proposition 4.1
and (5) that
lim
(x,t)(0,t
0
)
u
t
(x, t) +(0, t
0
)u
x
(0, t
0
) = 0 (21)
for any t
0
[0, T). In this section we show that the boundary condition (7)
also holds at the boundary, that is, not merely in the limit.
Proposition 5.1. The function u
t
(x, t) +(x, t)u
x
(x, t) denes a con-
tinuous function on [0, ) [0, T). Moreover, it vanishes for x = 0.
Remark. Note that Proposition 5.1 nishes the proof of Theorem 2.3.
Proof of Proposition 5.1. In view of (21) above, it suces to show
that u
t
exists at the boundary and that it equals u
x
. To do this, x a
point on the boundary with coordinates (0, t
0
). For notational simplicity we
assume that t
0
= 0. The time (left) derivative u
t
at the boundary is dened
by
u
t
(0, 0) = lim
k
k
_
u(0, 0) u
_
0,
1
k
__
, (22)
THE SINGLE-FACTOR TERM STRUCTURE EQUATION 15
provided the limit exists. To determine u
t
(0, 0), we let X
k
be dened by
_
dX
k
=(X
k
(t), t) dt +(X
k
(t), t) dW,
X
k
(1/k) = 0.
However, instead of considering the process X with dierent starting times,
we perform a change of variables so that the starting time is independent of
k. We thus introduce the process Y
k
(s) by
Y
k
(s) =kX
k
_
s
k
_
.
With respect to the time variable s, the dynamics of Y
k
has the form
_

_
dY
k
(s) =
_
1
k
Y
k
(s),
s
k
_
ds +

k
2
_
1
k
Y
k
(s),
s
k
_
dW
k
,
Y
k
(1) = 0,
(23)
where W
k
(s) denotes some Brownian motion. By the Markov property,
u(0, 1/k) =E[e

0
1/k
X
k
(s) ds
u(X
k
(0), 0)]
=E
_
e

0
1
(1/k
2
)Y
k
(s) ds
u
_
1
k
Y
k
(0), 0
__
.
Hence,
u
t
(0, 0) = lim
k
kE
0,1
_
u(0, 0) e

0
1
(1/k
2
)Y
k
(s) ds
u
_
1
k
Y
k
(0), 0
__
= lim
k
E
0,1
_
k
_
u(0, 0) u
_
1
k
Y
k
(0), 0
___
,
where the second equality follows using the inequality e
x
1 x since
E
0,1
_
ku
_
1
k
Y
k
(0), 0
_
|e

0
1
(1/k
2
)Y
k
(s) ds
1|
_
C
1
k
E
0,1
__
0
1
Y
k
(s) ds
_
0
as k . Now, dene the process Y by
_
dY =(0, 0) ds +
_
2
x
(0, 0)Y dW,
Y (1) = 0,
and redene Y
k
as in (23) above but using the same Brownian motion W
(this does not change the law of Y
k
). Since

k
(y, s) :=
_
y
k
,
s
k
_
(0, 0)
16 E. EKSTR

OM AND J. TYSK
and

k
(y, s) :=

k
2
_
y
k
,
s
k
_

_
2
x
(0, 0)y
uniformly on compacts as k (here we used the assumption that is
continuously dierentiable in space), it follows from [2] that Y
k
(0) Y (0) in
L
2
as k . From Theorem 3.4 above we know that u is dierentiable in x,
so k(u(0, 0) u(
y
k
, 0)) converges to u
x
(0, 0)y. By dominated convergence,
we have
kE
_
u(0, 0) u
_
1
k
Y (0), 0
__
u
x
(0, 0)E[Y (0)] =(0, 0)u
x
(0, 0)
as k . Moreover, the Lipschitz property of u yields that
kE
_
u
_
1
k
Y (0), 0
_
u
_
1
k
Y
k
(0), 0
__
CE[|Y (0) Y
k
(0)|] 0
as k . It follows that
u
t
(0, 0) +u
x
(0, 0)(0, 0) = 0.
As t
0
= 0 was chosen only for notational convenience, we have that
u
t
(0, t) +(0, t)u
x
(0, t) = 0
for any t. To be precise, we have shown the result above only for the left
t-derivative. However, this left t-derivative is continuous by the equation
above, so it follows from a simple calculus lemma that in fact u is dieren-
tiable in time, thus nishing our proof.
6. Models allowing negative interest rates. For models in which the
short rate can fall below zero with positive probability, the connection be-
tween the option price, given by a risk-neutral expected value, and the term
structure equation is more straightforward than for models with nonnega-
tive rates. Nevertheless, we have not been able to nd a precise reference for
this case, so for completeness we provide such a result in this section.
The assumptions needed on the drift and volatility are presented below,
where x
+
= max(x, 0). These assumptions now replace those of Hypothe-
sis 2.1. To our knowledge, all models used in practice that allow negative
interest rates satisfy these requirements. For example, the Vasicek model, in
which
dX(t) = (a bX(t)) dt + dW,
where a, b and are positive constants, is covered.
THE SINGLE-FACTOR TERM STRUCTURE EQUATION 17
Hypothesis 6.1. The drift C(R [0, T]) and the volatility
C(R[0, T]) are both Lipschitz continuous in x. Moreover,
0 <|(x, t)| C(1 +x
+
)
and
|(x, t)| C(1 +|x|) (24)
for some positive constant C.
Remark. The assumption that is strictly positive is not a strong
assumption. Indeed, let us for simplicity consider a time-homogeneous model
dX(t) =(X(t)) dt +(X(t)) dW(t) with (a) = 0 for some a R. If X(0)
a and (a) 0, then the process X cannot take values smaller than a, so we
are essentially in the situation handled in Sections 25 (but with the point
0 replaced with a). If b(a) <0, then X can take values below a, but if this
happens then the process will stay below a forever, and we are then again
back in the previous situation.
The bound on the volatility for negative rates guarantees that bond prices
are nite. For models in which grows faster than
_
|x| for negative rates,
bond prices can be innite; compare Theorem 4.1 in [16].
For a given continuous payo function g : R [0, ), dene the corre-
sponding option price u: R[0, T] by
u(x, t) =E
x,t
_
exp
_

_
T
t
X(s) ds
_
g(X(T))
_
,
where
_
dX(s) =(X(s), s) ds +(X(s), s) dW(s),
X(t) =x.
We require that the payo function is bounded for positive interest rates
and of, at most, exponential growth for negative rates, that is,
0 g(x) Kmax{1, e
Kx
} (25)
for some positive constant K. The corresponding term structure equation is
given by
u
t
(x, t) +
1
2

2
(x, t)u
xx
(x, t) +(x, t)u
x
(x, t) =xu(x, t)
on R[0, T), with terminal condition u(x, T) =g(x). By a classical solution
to this equation we mean a solution which is continuous up to the boundary
t = T and with all derivatives appearing in the equation being continuous
functions on the set t <T. The following is our main result about the term
structure equation on the whole real line.
18 E. EKSTR

OM AND J. TYSK
Theorem 6.2. Assume Hypothesis 6.1 and the bound (25). Then the
option price u(x, t) satises
u(x, t) K

max{1, e
K

x
} (26)
for some constant K

. Moreover, u(x, t) is the unique classical solution to the

term structure equation satisfying this growth assumption for some constant
K

.
Remark. Note that the bound (25) is natural for models on the whole
real line. In fact, even if g was bounded, the option price u would be of expo-
nential growth for negative rates. Also note that, for example, call options
on a bond are covered by Theorem 6.2. Indeed, the payo of a bond call
option with maturity T
1
is given by g(x) = (u(x, T
1
) K)
+
, where u is the
price of a bond maturing at T
2
>T
1
. Since u satises (26) by Theorem 6.2,
the payo g satises (25). Bond put options are trivially covered since they
have bounded payo functions.
Proof of Theorem 6.2. The bound (26) follows from Corollary 3.3
in [8]. To prove uniqueness of solutions, assume that v is a solution to
the term structure equation with boundary value g = 0 such that |v|
K

max{1, e
K

x
} for some constant K

. Let
h(x, t) =e
M(Tt)
(e
f(t)x
+x)
for some large constant M. Here
f(t) =
e
C(Tt)
1
C
+Ke
C(Tt)
,
where C is the constant appearing in (24) and K >K

. Then the set

{(x, t) R[0, T] : h(x, t) <v(x, t)}
is bounded, and
h
t
+
1
2

2
h
xx
+h
x
xh <0
at all points provided M is chosen large enough. Standard methods used
to prove the maximum principle yield that v = 0 at all points. Thus we
have uniqueness of solutions to the term structure equation in the class of
functions satisfying (26).
To show that u is a classical solution to the term structure equation,
we carry out an approximation argument. We consider the term structure
equation but with the discount factor x replaced by bounded functions that
agree with x inside some large compact sets. We also replace the payo
function g with functions of, at most, polynomial growth that agree with g
THE SINGLE-FACTOR TERM STRUCTURE EQUATION 19
on large compact sets. The corresponding equation is in the standard class
and its stochastic solution is known to be continuous and hence is a classical
solution. Now let the functions approximating the discount factor grow up
to x for large positive x and then decrease down to x for large negative x.
By the monotone convergence theorem, the corresponding stochastic solu-
tions converge to the stochastic solution above denoted u. Interior Schauder
estimates yield interior regularity of the limiting solution and continuity at
the boundary is established using the maximum principle.
REFERENCES
[1] Athreya, S. R., Barlow, M. T., Bass, R. F. and Perkins, E. A. (2002). Degen-
erate stochastic dierential equations and super-Markov chains. Probab. Theory
Related Fields 123 484520. MR1921011
[2] Bahlali, K., Mezerdi, B. and Ouknine, Y. (1998). Pathwise uniqueness and ap-
proximation of solutions of stochastic dierential equations. In Seminaire de
Probabilites, XXXII. Lecture Notes in Math. 1686 166187. Springer, Berlin.
MR1655150
[3] Bass, R. F. and Perkins, E. A. (2003). Degenerate stochastic dierential equa-
tions with Holder continuous coecients and super-Markov chains. Trans. Amer.
Math. Soc. 355 373405 (electronic). MR1928092
[4] Bj ork, T. (2004). Arbitrage Theory in Continuous Time, 2nd ed. Oxford Univ. Press.
[5] Brigo, D. and Mercurio, F. (2001). Interest Rate ModelsTheory and Practice.
Springer, Berlin. MR1846525
[6] Duffy, D. J. (2006). Finite Dierence Methods in Financial Engineering. Wiley,
Chichester. MR2286409
[7] Ekstr om, E., L otstedt, P. and Tysk, J. (2009). Boundary values and nite dier-
ence methods for the single factor term structure equation. Appl. Math. Finance
16 253259. MR2589641
[8] Ekstr om, E. and Tysk, J. (2008). Convexity theory for the term structure equation.
Finance Stoch. 12 117147. MR2362153
[9] Friedman, A. (1964). Partial Dierential Equations of Parabolic Type. Prentice-Hall,
Englewood Clis, NJ. MR0181836
[10] Friedman, A. (1975). Stochastic Dierential Equations and Applications 1. Proba-
bility and Mathematical Statistics 28. Academic Press, New York. MR0494490
[11] G

Ihman,

I.

I. and Skorohod, A. V. (1972). Stochastic Dierential Equations.
Springer, New York. MR0346904
[12] Heston, S., Loewenstein, M. and Willard, G. (2007). Options and bubbles. Rev.
Financ. Stud. 20 359390.
[13] Janson, S. and Tysk, J. (2006). FeynmanKac formulas for BlackScholes-type
operators. Bull. London Math. Soc. 38 269282. MR2214479
[14] Olenik, O. A. and Radkevi c, E. V. (1973). Second Order Equations with Non-
negative Characteristic Form. Plenum Press, New York. MR0457908
[15] Revuz, D. and Yor, M. (1999). Continuous Martingales and Brownian Motion, 3rd
ed. Grundlehren der Mathematischen Wissenschaften [Fundamental Principles
of Mathematical Sciences] 293. Springer, Berlin. MR1725357
[16] Yong, J. (2006). Remarks on some short rate term structure models. J. Ind. Manag.
Optim. 2 119134. MR2208398
20 E. EKSTR

OM AND J. TYSK
Department of Mathematics
Uppsala University
Box 480, SE-75106 Uppsala
Sweden
E-mail: ekstrom@math.uu.se
johan.tysk@math.uu.se