. (2.1)
Then in the riskneutral formulation of Heath, Jarrow and Morton (1992) term structure
models on (, F, (F
t
)
t[0, ]
, P), the forward rates are assumed to satisfy the stochastic
integral equation
1
f(t, ) = f(0, ) +
_
t
0
(s, )
(s, ) ds +
_
t
0
(s, )
dw(s), (2.2)
1
It is a consequence of the HJM forward rate drift restriction that the forward rate drift under the risk neutral
measure is (t, )
_
t
(t, u) du.
STATE VARIABLES AND AFFINE NATURE OF MARKOVIAN HJM MODELS 3
where f(0, ) is the initial forward rate curve, (t, ) is an adapted R
n
+
valued forward
rate volatility process, (t, ) =
_
t
(t, u) du and the superscript
denotes matrix trans
position. Note that a Heath, Jarrow and Morton, henceforth HJM, term structure model is
completely determined by its dimension n, the initial forward rate curve f(0, ) and the
forward rate volatility process (t, ).
Let d N
+
. Then an HJM model Mis said to be Markovian with d factors if there ex
ists a ddimensional (F
t
)Markov process z(t) such that the forward rates can be expressed
in the form f(t, , z(t)), where the explicit dependence on t and are deterministic and
the dependence on the Wiener path enters only through z(t). Many of the short rate mod
els such as Vasicek (1977), Cox, Ingersoll and Ross (1985) and Ho and Lee (1986) are
examples of Markovian HJM models.
A Markovian HJM model Ais said to be afne if the forward rate process can be written
in the form
f(t, , z(t)) = h
0
(t, ) + h(t, )
z(t), (2.3)
where h
0
(t, ) and h(t, ) are deterministic R and R
d
valued functions respectively. Note
that in view of (2.1) A is afne if and only if the bond price in A is exponential afne
2
in
the sense of Dufe and Kan (1996).
Throughout this paper, the Musiela (1993) parameterisation, = t + x, of the maturity
variable will often be used. Under this parameterisation the fundamental quantities are the
benchmark forward rates r(t, x) dened by the equation
r(t, x) = f(t, t + x). (2.4)
The benchmark forward rates satisfy the stochastic integral equation
r(t, x) = f(0, t + x) +
_
t
0
(s, t + x)
(s, t + x) ds +
_
t
0
(s, t + x)
dw(s), (2.5)
and the stochastic differential equation
dr(t, x) =
x
_
r(t, x) +
1
2
 (t, t + x)
2
_
dt + (t, t + x)
dw(t), (2.6)
where   denotes the standard Euclidean norm. If the corresponding benchmark bond
prices b(t, x) are dened by b(t, x) = p(t, t + x), then it follows from (2.1) that they
satisfy the equation
b(t, x) = exp
_
_
x
0
r(t, u) du
_
. (2.7)
3. REVIEW OF MARKOVIAN HJM MODELS
In this section, the class of Markovian HJM models introduced in Chiarella and Kwon
(2001b) is briey reviewed. As indicated above, this class includes the models considered
in Carverhill (1994), Ritchken and Sankarasubramanian (1995), Bhar and Chiarella (1997),
Inui and Kijima (1998) and de Jong and SantaClara (1999) as special cases.
Let m N, and assume given d
1
, . . . , d
n
N and x
1
< x
2
< < x
m
R
+
. Then
in addition to the standard HJM assumptions, suppose further that:
[A1] (t, ) is a function of t, and m benchmark forward rates r(t, x
1
), . . . , r(t, x
m
),
so that
(t, ) = (t, , r(t, x
1
), . . . , r(t, x
m
)), (3.1)
2
Recall that Dufe and Kan (1996) refer to p(t, ) as being exponential afne if there exists a ddimensional
Markov process z(t) such that p(t, ) = exp[k
0
(t, ) k(t, )
z(t)], where k
0
(t, ) and k(t, ) are deter
ministic R and R
d
valued functions respectively.
4 CARL CHIARELLA AND OH KANG KWON
[A2] for each 1 i n,
i
(t, ) is d
i
times differentiable with respect to and satises
a d
i
th order homogeneous linear differential equation of the form
3
L
i
i
(t, ) = 0, where L
i
=
d
i
d
i
d
i
1
j=0
i,j
()
j
j
(3.2)
and
i,j
() are deterministic functions.
Note that the volatility processes considered in Carverhill (1994), Ritchken and Sankara
subramanian (1995), Inui and Kijima (1998) and de Jong and SantaClara (1999) satisfy the
above assumptions with d
i
= 1 for all i, and the volatility process considered in Bhar and
Chiarella (1997) satisfy the above assumptions with d
i
= k 0 and L
i
= (/t
i
)
k
.
In order to establish that the HJM models satisfying assumptions [A1] and [A2] admit
nite dimensional Markovian realisations, a nite set of state variables that capture the
history of the Wiener path must be identied. Thus, for each 1 i n, 0 p q < d
i
,
0 l < d
i
and x [0, ), dene the stochastic processes
p,q
i,x
(t) and
l
i,x
(t) by
p,q
i,x
(t) =
_
t
0
i
(s, t + x)
x
p
i
(s, t + x)
x
q
ds, (3.3)
l
i,x
(t) =
_
t
0
i
(s, t + x)
x
l
_
t+x
s
i
(s, u) duds +
_
t
0
i
(s, t + x)
x
l
dw
i
(s). (3.4)
Theorem 3.1 (Chiarella and Kwon (2001b, Theorem 2.3)). Suppose that the forward rate
volatility process (t, ) satises the assumptions [A1] and [A2]. Then the corresponding
HJM model is Markovian with respect to the set
{r(t, x
k
),
p
i
,q
i
i,x
k
(t),
l
i
i,x
k
(t)} (3.5)
where 1 i n, 1 k m, 0 p
i
q
i
< d
i
and 1 l
i
< d
i
.
The proof of the above theorem relies on noting the interdependence of the dynamics
of r(t, x
k
),
p
i
,q
i
i,x
k
(t), and
l
i
i,x
k
(t) on one another, and recognising that when assumption
[A2] is satised then
p
i
,q
i
i,x
k
(t) and
l
i
i,x
k
(t) are expressible in terms of the processes in (3.5)
if p
i
d
i
, q
i
d
i
, or l
i
d
i
.
At this stage, the economic signicance of the state variables
p,q
i,x
k
(t) and
l
i,x
k
(t) intro
duced in (3.3) and (3.4) is unclear, and it is not immediately obvious if they are expressible
in terms of the quantities directly observed in the market. This problem is resolved in
Section 5, in which a clear link is established between the state variables and the market
forward rate curve.
4. ALTERNATIVE SET OF STATE VARIABLES
In this section, the HJM models that satisfy the assumptions [A1] and [A2] of Section 3
are considered with respect to an alternative set of state variables. These variables provide
an easier passage to linking the variables
p,q
i,x
k
(t) and
l
i,x
k
(t) in (3.5) to quantities that are
economically meaningful.
If it is assumed that
i,j
() in (3.2) are continuous, then it is known from the theory
of ordinary differential equations
4
that there exist d
i
linearly independent deterministic
functions
i,j
() such that L
i
i,j
() = 0 for all 1 j d
i
, and stochastic coefcient
processes c
i,j
(t) = c
i,j
(t, r(t, x
1
), . . . , r(t, x
m
)) such that
i
(t, ) =
d
i
j=1
c
i,j
(t)
i,j
(). (4.1)
3
A similar condition was obtained in the deterministic volatility case by Bj ork and Gombani (1999), and
generalised to the separable volatility case by Bj ork and Svensson (1999).
4
See for example Coddington and Levinson (1955, Theorem 5.1).
STATE VARIABLES AND AFFINE NATURE OF MARKOVIAN HJM MODELS 5
Equation (4.1) simply expresses the general solution
i
(t, ) of the ordinary differential
equation (3.2) as a linear combination of d
i
linearly independent solutions
i,j
() with
coefcients c
i,j
(t). The signicance of this expression is that the dependence of
i
(t, )
on the current time t and the maturity time are essentially separated.
As a motivation for the variables to be introduced, note that the benchmark forward
rates corresponding to the volatility process (4.1) satisfy the stochastic integral equation
r(t, x) = f(0, t + x) +
n
i=1
_
t
0
i
(s, t + x) dw
i
(s)
+
n
i=1
_
t
0
i
(s, t + x)
_
t+x
s
i
(s, u) duds
= f(0, t + x) +
n
i=1
d
i
j=1
i,j
(t + x)
_
t
0
c
i,j
(s) dw
i
(s)
+
n
i=1
d
i
j,j
=1
i,j
(t + x)
_
t
0
c
i,j
(s)c
i,j
(s)
_
t+x
s
i,j
(u) duds.
= f(0, t + x) +
n
i=1
d
i
j=1
i,j
(t + x)
_
t
0
c
i,j
(s) dw
i
(s)
+
n
i=1
d
i
j,j
=1
i,j
(t + x)
_
t
0
c
i,j
(s)c
i,j
(s)
__
t+x
0
i,j
(u) du
_
s
0
i,j
(u) du
_
ds.
= f(0, t + x) +
n
i=1
d
i
j=1
i,j
(t + x)
_ _
t
0
c
i,j
(s) dw
i
(s)
d
i
=1
_
t
0
c
i,j
(s)c
i,j
(s)
_
s
0
i,j
(u) duds
_
+
n
i=1
d
i
j,j
=1
i,j
(t + x)
_
t+x
0
i,j
(u) du
_
t
0
c
i,j
(s)c
i,j
(s) ds.
(4.2)
From this it can be seen that the quantities which play an important role in describing the
dynamic evolution of the benchmark forward rates are the deterministic functions
i,j
(t)
dened by
i,j
(t) =
_
t
0
i,j
(u) du, (4.3)
and the stochastic processes
j,j
i
(t),
j,j
i
(t) and
j
i
(t) dened by
j,j
i
(t) =
_
t
0
c
i,j
(s) c
i,j
(s) ds, (4.4)
j
i
(t) =
_
t
0
c
i,j
(s) dw
i
(s)
d
i
=1
_
t
0
c
i,j
(s) c
i,j
(s)
i,j
(s) ds, (4.5)
for 1 i n and 1 j, j
d
i
. One of the consequences of (4.2) is that the path
dependence of the forward rate curve is completely captured by the stochastic processes
j,j
i
(t) and
j
i
(t).
6 CARL CHIARELLA AND OH KANG KWON
The next lemma establishes the connection between the state variables
p,q
i,x
(t) and
l
i,x
(t) as dened in (3.3) and (3.4) and the stochastic processes
j,j
i
(t) and
j
i
(t).
Lemma 4.1. Let
p,q
i,x
(t) and
l
i,x
(t) be dened as in (3.3) and (3.4). Then
p,q
i,x
(t) =
1j,j
d
i
i,j
(t + x)
x
p
q
i,j
(t + x)
x
q
j,j
i
(t), (4.6)
l
i,x
(t) =
1jd
i
i,j
(t + x)
x
l
j
i
(t)
+
1j,j
d
i
i,j
(t + x)
l
i,j
(t + x)
x
l
j,j
i
(t).
(4.7)
In particular,
p,q
i,x
(t) and
l
i,x
(t) are linear in
j,j
i
(t) and
j
i
(t).
Proof. Consider rstly (4.6). Using the denition of
p,q
i,x
(t) from (3.3)
p,q
i,x
(t) =
_
t
0
i
(s, t + x)
x
p
q
i
(s, t + x)
x
q
ds
=
1j,j
d
i
i,j
(t + x)
x
p
q
i,j
(t + x)
x
q
_
t
0
c
i,j
(s)c
i,j
(s) ds
=
1j,j
d
i
i,j
(t + x)
x
p
q
i,j
(t + x)
x
q
j,j
i
(t).
where (4.4) was used in the last equality. Similarly, using the denition of
l
i,x
(t) from
(3.4)
l
i,x
(t) =
_
t
0
i
(s, t + x)
x
l
_
t+x
s
i
(s, u) duds +
_
t
0
i
(s, t + x)
x
l
dw
i
(s)
=
1j,j
d
i
i,j
(t + x)
x
l
_
t
0
c
i,j
(s)c
i,j
(s)
_
t+x
s
i,j
(u) duds
+
1jd
i
i,j
(t + x)
x
l
_
t
0
c
i,j
(s) dw
i
(s)
=
1j,j
d
i
i,j
(t + x)
x
l
_
t
0
c
i,j
(s)c
i,j
(s) [
i,j
(t + x)
i,j
(s)] ds
+
1jd
i
i,j
(t + x)
x
l
_
t
0
c
i,j
(s) dw
i
(s)
=
1jd
i
i,j
(t + x)
x
l
_
_
_
t
0
c
i,j
(s) dw
i
(s)
1j
d
i
_
t
0
c
i,j
(s) c
i,j
(s)
i,j
(s) ds
_
_
+
1j,j
d
i
i,j
(t + x)
x
l
i,j
(t + x)
_
t
0
c
i,j
(s) c
i,j
(s) ds
=
1jd
i
i,j
(t + x)
x
l
j
i
(t) +
1jj
d
i
i,j
(t + x)
l
i,j
(t + x)
x
l
l,j
i
(t)
where the last equality follows from (4.4) and (4.5).
Inversion of (4.6) and (4.7) to express
j,j
i
(t) and
j
i
(t) in terms of
p,q
i,x
(t) and
l
i,x
(t)
relies on the invertibility of a certain coefcient matrix whose precise denition requires
STATE VARIABLES AND AFFINE NATURE OF MARKOVIAN HJM MODELS 7
the introduction of some notation. Firstly, let
i
=
1
2
d
i
(d
i
+ 1),
P
i
= {(j, j
) N N  1 j j
d
i
},
and dene
i
: P
i
{1, 2, . . . ,
i
} to be the lexicographical ordering given by
i
(j, j
) =
1
2
(j 1)(2d
i
j) + j
, (4.8)
so that for example,
i
(1, j
) = j
and
i
(2, j
) = d
i
1 + j
and
i
(d
i
, d
i
) =
i
. Note
that
1
i
is well dened since
i
is a bijection. Next dene the coordinate projections
i
: N N N, for i = 1, 2, so that
1
(j, j
) = j and
2
(j, j
) = j
) = j
for 1 j
d
i
,
it follows that
1,i
(j
) = 1 and
2,i
(j
) = j
for 1 j
d
i
, and similarly since
i
(2, j
) = d
i
1 + j
for 2 j
d
i
, it follows that
1,i
(d
i
1 + j
) = 2 and
2,i
(d
i
1 + j
) = j
for 2 j
d
i
. Finally, dene
j,j
by
j,j
=
_
1
2
, j = j
1, j = j
. (4.9)
Now, for any x [0, ), dene a d
i
d
i
matrix A
(i,x)
, a d
i
i
matrix B
(i,x)
and an
i
matrix C
(i,x)
by
A
(i,x)
j,j
=
j1
i,j
(t, x)
x
j1
, (4.10)
B
(i,x)
j,j
=
1,i
(j
),
2,i
(j
)
i,
2,i
(j
j1
i,
1,i
(j
)
x
j1
+
1,i
(j
),
2,i
(j
)
i,
1,i
(j
j1
i,
2,i
(j
)
x
j1
(4.11)
C
(i,x)
j,j
=
1,i
(j
),
2,i
(j
1,i
(j)1
i,
1,i
(j
)
x
1,i
(j)1
2,i
(j)1
i,
2,i
(j
)
x
2,i
(j)1
+
1,i
(j
),
2,i
(j
1,i
(j)1
i,
2,i
(j
)
x
1,i
(j)1
2,i
(j)1
i,
1,i
(j
)
x
2,i
(j)1
.
(4.12)
and let M
(i,x)
be the (d
i
+
i
) (d
i
+
i
) matrix given by
M
(i,x)
=
_
A
(i,x)
B
(i,x)
0 C
(i,x)
_
(4.13)
Then (4.6) and (4.7) can be written in the matrix form
_
0
i,x
(t),
1
i,x
(t), . . . ,
d
i
1
i,x
(t),
0,0
i,x
(t),
0,1
i,x
(t), . . . ,
d
i
1,d
i
1
i,x
(t)
_
= M
(i,x)
_
1
i
(t),
2
i
(t), . . . ,
d
i
i
(t),
1,1
i
(t),
1,2
i
(t), . . . ,
d
i
,d
i
i
(t)
_
(4.14)
where the superscript
denotes matrix transposition and
p,q
i,x
(t) and
j,j
i
(t) have been
written in lexicographical order. The next lemma gives the conditions under which (4.6)
and (4.7) can be inverted.
Lemma 4.2. Suppose for each 1 i n there exists 1 k
i
, k
i
m such that A
(i,x
k
i
)
and C
(i,x
k
i
)
are invertible. Then
j,j
i
(t) and
j
i
(t) can be expressed as linear functions
of
p,q
i,x
k
i
(t) and
l
i,x
k
i
(t) with deterministic coefcients. In this case, the processes
j,j
i
(t)
and
j
i
(t), where 1 j j
d
i
, 1 j
d
i
and 1 i n, are contained in the
linear span of the Markovian system given in Theorem 3.1.
Proof. This is clear from (4.13) and (4.14).
Proposition 4.3. Let M be an HJM model satisfying the assumptions [A1] and [A2]. If
the conditions of Lemma 4.2 are satised then Mis an afne term structure model.
8 CARL CHIARELLA AND OH KANG KWON
Proof. Firstly, the set {r(t, x
k
),
p
i
,q
i
i,x
k
(t),
l
i
i,x
k
} forms a Markovian systemby Theorem3.1.
Next, from (2.5) and (3.4), the benchmark forward rate r(t, x) can be written in the form
r(t, x) = f(0, t + x) +
n
i=1
0
i,x
(t).
Using (4.7) to substitute for
0
i,x
(t) gives
r(t, x) = f(0, t + x) +
n
i=1
d
i
j
i
=1
i,j
i
(t + x)
j
i
i
(t)
+
n
i=1
d
i
j
i
,j
i
=1
j
i
j
j
i
,j
i
i
(t + x)
j
i
,j
i
i
(t)
(4.15)
where
j
i
,j
i
i
(t + x) =
_
i,j
i
(t + x)
i,j
i
(t + x) if j
i
= j
i
,
i,j
i
(t + x)
i,j
i
(t + x) +
i,j
i
(t + x)
i,j
i
(t + x) if j
i
< j
i
.
(4.16)
Since
i,j
i
(t + x) and
i,j
i
(t + x) are deterministic functions, r(t, x) is an afne function
of
j
i
,j
i
i
(t) and
j
i
i
(t) for all x R
+
, where 1 i n and 1 j
i
j
i
d
i
. But
since the conditions of Lemma 4.2 are satised by assumption,
j,j
i
(t) and
j
i
(t) can be
expressed as linear functions of
p,q
i,x
k
i
(t) and
i,x
k
i
(t) with deterministic coefcients. This
establishes that r(t, x) is an afne function of
p,q
i,x
k
i
(t) and
i,x
k
i
(t), and hence that Mis
afne.
The next corollary extends the InuiKijima bond price formula to the more general
Markovian models introduced in Chiarella and Kwon (2001b).
Corollary 4.4. Let M be an HJM model satisfying the assumptions of Proposition 4.3.
Then the benchmark bond prices in Mare exponential afne and are given by the formula
b(t, x) =
b(0, t + x)
b(0, t)
exp
_
n
i=1
d
i
j
i
=1
i,j
i
(t, x)
j
i
i
(t) +
n
i=1
d
i
j
i
,j
i
=1
j
i
j
j
i
,j
i
i
(t, x)
j
i
,j
i
i
(t)
_
,
(4.17)
where
i,j
i
(t, x) =
_
x
0
i,j
i
(t + u) du,
j
i
,j
i
i
(t, x) =
_
x
0
j
i
,j
i
i
(t + u) du.
Proof. From (2.7), the determination of b(t, x) involves the computation of the integral
_
x
0
r(t, u) du. But from (4.15) this, in turn, requires the integrals of deterministic functions
f(0, u),
i,j
i
(u) and
j
i
,j
i
i
(u), which results in (4.17).
Note that the bond price given in (4.17) is exponential afne in the sense of Dufe and
Kan (1996). However, note also that the functional form of the forward rate volatility has
not been restricted beyond (3.1) and (3.2). In particular, the components of (t, ) are
not restricted to be square root afne in the state variables, which is a restriction obtained
in Dufe and Kan (1996) for the class of afne models they consider. Thus the class of
afne models considered in this paper include those which do not necessarily fall under
STATE VARIABLES AND AFFINE NATURE OF MARKOVIAN HJM MODELS 9
the DufeKan characterisation. For a more complete discussion on the assumptions un
derlying the DufeKan characterisation of afne models see Kwon (2001).
5. STATE VARIABLES AND THE FORWARD RATE CURVE
The purpose of this section is to determine the explicit links between the state variables
and the forward rate curve for an HJM model Mwhich satises the assumptions [A1] and
[A2] of Section 3 and the conditions of Lemma 4.2. Firstly, recall from Theorem 3.1 that
the models which satisfy these assumptions are Markov with respect to
p,q
i,x
(t) and
l
i,x
(t).
5
But since these variables are expressible in terms of
j
i
,j
i
i
(t) and
j
i
i
(t) by Lemma 4.1, it
sufces to determine the links between the processes
j
i
,j
i
i
(t) and
j
i
i
(t) and the forward
rate curve. Now, (4.15) gives the forward rates as afne functions of
j
i
,j
i
i
(t) and
j
i
i
(t),
and so the processes
j
j
,l
i
i
(t) and
j
i
i
(t) can be expressed in terms of the forward rates by
inverting (4.15). The details of this inversion procedure and the sufcient conditions under
which this inversion is possible are given in the remainder of this section.
For x R
+
dene r
(t, x)
satises the equation
r
(t, x) =
n
i=1
d
i
j
i
=1
i,j
i
(t + x)
j
i
i
(t)
+
n
i=1
d
i
j
i
,j
i
=1
j
i
j
j
i
,j
i
i
(t + x)
j
i
,j
i
i
(t)
(5.1)
and consequently each x R
+
gives rise to a linear equation for r
(t, x) in terms of
j
i
,j
i
i
(t) and
j
i
i
(t) with deterministic coefcients. It is possible that some of the processes
j
i
,j
i
i
(t) and
j
i
i
(t) are linearly dependent on others or even deterministic. In order to
obtain a minimal set of stochastic processes with which to capture the state of the forward
rate curve, dene the set
S
,
= {
j
i
,j
i
i
(t),
j
i
i
(t)  1 i n, 1 j
i
j
i
d
i
, 1 j
i
d
i
}, (5.2)
let S
,
be the subset of deterministic elements in S
,
and let S
,
be a maximal linearly
independent subset of S
,
S
,
. Choose an ordering for S
,
and write
j
(t) for the
elements of this ordered set, where 1 j S
,
. Then by moving the deterministic
terms on the righthand side of (5.1) to the lefthand side, (5.1) can be rewritten in the form
r
(t, x) =
S
,

j=1
j
(t + x)
j
(t), (5.3)
where
j
(t + x) are the appropriate deterministic coefcient functions and
r
(t, x) = r
(t, x)
n
i=1
j
i
i
(t)S
i,j
i
(t + x)
j
i
i
(t)
i=1
j
i
,j
i
i
(t)S
j
i
,j
i
i
(t + x)
j
i
,j
i
i
(t).
(5.4)
For notational convenience, set n
= S
,
.
5
Along with a nite number benchmark forward rates.
10 CARL CHIARELLA AND OH KANG KWON
Proposition 5.1. For any
1
, . . . ,
n
R
+
, dene the matrix (t,
1
, . . . ,
n
) by
(t,
1
, . . . ,
n
) =
_
1
(t +
1
)
2
(t +
1
)
n
(t +
1
)
1
(t +
2
)
2
(t +
2
)
n
(t +
2
)
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
(t +
n
)
2
(t +
n
)
n
(t +
n
)
_
_
, (5.5)
and suppose there exist
1
, . . . ,
n
R
+
such that det (t,
1
, . . . ,
n
) = 0 for all t.
Then the state variables
j
(t) can be written as afne functions of the benchmark forward
rates r(t,
1
), . . . , r(t,
n
).
Proof. Let
1
, . . . ,
n
R
+
be as given. Then (5.3) holds for each x =
i
, and writing
the resulting equations in matrix form gives
[ r
(t,
1
), . . . , r
(t,
n
)]
= (t,
1
, . . . ,
S
) [
1
(t), . . . ,
n
(t)]
. (5.6)
Now, since det (t,
1
, . . . ,
n
) = 0 the last equation can be inverted and results in ex
pressions of the form
j
(t) =
n
=1
j,j
(t +
1
, . . . , t +
n
) r
(t,
j
) (5.7)
for 1 j n
, where
j,j
are the entries of (t,
1
, . . . ,
n
)
1
. Since
j,j
are deter
ministic, and r
(t,
j
) and r
(t,
j
) are related by (5.4), this completes the proof.
The above result establishes that the variables
j
i
i
(t) and
j
i
,j
i
i
(t), and hence also
p,q
i,x
(t) and
l
i,x
(t), are afne functions of a nite number of benchmark forward rates
r(t,
1
), . . . , r(t,
n
). Furthermore, since the forward rate curve is afne in
j
i
i
(t) and
j
i
,j
i
i
(t) by (4.15), the entire forward rate curve is afne in r(t,
1
), . . . , r(t,
n
). This is
the content of the next corollary.
Corollary 5.2. If the conditions of Proposition 5.1 are satised, then the forward rate
curve is afne with respect to a nite number of benchmark forward rates.
Note that substituting (5.7) into (4.17) allows the bond price to be expressed as an
exponential afne function of a nite number of benchmark forward rates. Although the
notation is somewhat cumbersome for the general case, the expression for the bond price
is easily obtained in specic cases as will be shown in Section 7.
6. STATE VARIABLES AS YIELDS
It was established in the previous section that the state variables in a large class of Mar
kovian HJM models can be expressed as afne functions of a nite number of forward
rates. However, the instantaneous forward rates are not directly observable in the mar
ket, and for the numerical implementation of these models, it is convenient to express the
state variables in terms of directly observed market quantities. Fortunately, for the models
considered in Section 4 and Section 5, it is possible to express the forward rates as afne
functions of the market observed yields and this, in turn, allows these models to be afne
with respect to a nite number of yields.
Let x R
++
. Then the xtenor yield, y(t, x), is dened by the equation
y(t, x) =
1
x
_
x
0
r(t, u) du. (6.1)
From (5.3), the xtenor yield satises the equation
y
(t, x) =
n
j=1
j
(t, x)
j
(t), (6.2)
STATE VARIABLES AND AFFINE NATURE OF MARKOVIAN HJM MODELS 11
where
y
(t, x) =
1
x
_
x
0
r
j
(t + u) du. (6.4)
The next result establishes a sufcient condition under which the variables
j
(t) are ex
pressible as afne functions of yields and is the direct analogue of Proposition 5.1.
Proposition 6.1. For
1
, . . . ,
n
R
+
, dene the matrix
(t,
1
, . . . ,
n
) by
(t,
1
, . . . ,
n
) =
_
_
1
(t,
1
)
2
(t,
1
)
n
(t,
1
)
1
(t,
2
)
2
(t,
2
)
n
(t,
2
)
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
(t,
n
)
2
(t,
n
)
n
(t,
n
)
_
_
, (6.5)
and suppose there exist
1
, . . . ,
n
R
+
such that det
(t,
1
, . . . ,
n
) = 0 for all t.
Then the state variables
j
(t) are afne in the yields y(t,
1
), . . . , y(t,
n
).
Proof. Apply the arguments in the proof of Proposition 5.1 to (6.2).
Corollary 6.2. If the conditions of Proposition 6.1 are satised, then the forward rate
curve is afne with respect to a nite number of benchmark yields.
As was the case with benchmark forward rates, the above results allow the bond price
to be expressed as an exponential afne function of a nite number of benchmark yields.
7. EXAMPLES
This section illustrates the general framework developed in the previous sections with
some simple examples. In particular, the results from Sections 5 and 6 are applied to
express the state variables for these examples in terms of a nite number of benchmark
forward rates or yields.
7.1. Extended Vasicek Model. The Hull and White (1990) extended Vasicek model M
Vas
is a 1dimensional HJM model corresponding to the forward rate volatility
(t, ) =
0
e
(t)
, (7.1)
where
0
and are constants. Since (t, )/ + (t, ) = 0, M
Vas
is a special case
of the Ritchken and Sankarasubramanian (1995) model and is consequently Markov. If the
forward rate volatility is rewritten in the form
(t, ) =
_
0
e
t
_
e
, (7.2)
then the coefcient and the deterministic function of in (4.1) can be identied as
c(t) =
0
e
t
and () = e
. (7.3)
The relevant state variables for M
Vas
are hence
11
(t) and
1
(t), where the subscripts
have been omitted for this example. But
11
(t) =
_
t
0
c(s)
2
ds =
1
2
2
0
_
e
2t
1
_
(7.4)
and so
11
(t) is deterministic in this case while
1
(t) =
_
t
0
c(s) dw(s)
_
t
0
c(s)
2
(s) ds =
0
_
t
0
e
s
dw(s)
2
0
2
(e
t
1)
2
(7.5)
12 CARL CHIARELLA AND OH KANG KWON
is stochastic. So for M
Vas
, S
,
= {
11
(t)} and S
,
= {
1
(t)}. It follows that M
Vas
is
a 1factor Markovian model with
1
(t) being the sole state variable, and the forward rate
in (4.15) can be written
r(t, x) = f(0, t + x) + e
(t+x)
_
1
2
2
0
_
e
2t
1
_
_
1 e
(t+x)
_
+
1
(t)
_
. (7.6)
The conditions of Proposition 5.1 are trivially satised, and
1
(t) = e
(t+x)
[r(t, x) f(0, t + x)]
1
2
2
0
_
e
2t
1
_
_
1 e
(t+x)
_
(7.7)
is the expression of
1
(t) in terms of any benchmark forward rate r(t, x). In particular,
setting x = 0 allows
1
(t) to be expressed in terms of the spot rate r(t)
1
(t) = e
t
_
r(t) f(0, t)
1
2
2
0
_
e
2t
1
_ _
1 e
t
_
_
. (7.8)
The expressions for the yield and bond prices of any maturity in this case are easily com
puted and are given by
y(t, x) =
0
(t, x) +
1
2
2
x
2
0
_
e
t
e
t
_ _
1 e
x
_
+
1
x
e
t
_
1 e
x
_
1
(t)
(7.9)
b(t, x) =
b(0, t + x)
b(0, t)
exp
_
1
2
2
2
0
_
e
t
e
t
_ _
1 e
x
_
_
exp
_
e
t
_
1 e
x
_
1
(t)
_ (7.10)
Once again, the expression (7.8) for
1
(t) can be substituted to obtain the yields and
bond prices in terms of the spot rate. Note that similar arguments using (7.9) instead
of (7.6) allows the bond price to be expressed in terms of a benchmark yield rather than a
benchmark forward rate.
7.2. RitchkenSankarasubramanian Model. The next example is a generic Ritchken
and Sankarasubramanian (1995) model M
RS
in which the forward rate volatility process
is given by
(t, ) = g(t, r(t))e
(t)
, (7.11)
where is a constant and g is a deterministic function. Note that M
RS
is a 1dimensional
HJM model and that (t, ) satises the equation (t, )/ +(t, ) = 0. Once again,
rewriting the forward rate volatility in the form
(t, ) = g(t, r(t))e
t
e
(7.12)
allows the coefcient and the function of in (4.1) to be identied as
c(t) = g(t, r(t))e
t
and () = e
. (7.13)
The state variables for M
RS
are
11
(t) and
1
(t) given by
11
(t) =
_
t
0
g(s, r(s))
2
e
2s
ds, (7.14)
1
(t) =
_
t
0
g(s, r(s))e
s
dw(s)
1
_
t
0
g(s, r(s))
2
e
2s
(1 e
s
) ds, (7.15)
where the subscripts have been omitted from
11
(t) and
1
(t) . In contrast to the situation
in the extended Vasicek model,
11
(t) and
1
(t) are both stochastic in this case and so
STATE VARIABLES AND AFFINE NATURE OF MARKOVIAN HJM MODELS 13
S
,
= and S
,
= {
11
(t),
1
(t)}. It follows that M
RS
is a 2factor Markovian
model and the forward rate in (4.15) can be written
r(t, x) = f(0, t + x) + (t, x)
1
(t) + (t, x)
11
(t), (7.16)
where (t, x) = e
(t+x)
and (t, x) =
1
e
(t+x)
_
1 e
(t+x)
.
Now, if
1
=
2
then
det
_
(t,
1
) (t,
1
)
(t,
2
) (t,
2
)
_
=
1
e
(3t+
1
+
2
)
_
e
2
e
= 0, (7.17)
and Proposition 5.1 gives
1
(t) =
(t,
2
)r
(t,
1
) (t,
1
)r
(t,
2
)
(t,
1
)(t,
2
) (t,
1
)(t,
2
)
, (7.18)
11
(t) =
(t,
1
)r
(t,
2
) (t,
2
)r
(t,
1
)
(t,
1
)(t,
2
) (t,
1
)(t,
2
)
(7.19)
as the expressions for the state variables in terms of the forward rates in this case. The
standard Markovian state variables for this model,
0,0
(t) and
0
(t), can now be obtained
from (4.6) and (4.7).
To express the state variables in terms of yields, note from (6.1) and (7.6) that
y(t, x) =
0
(t, x) + (t, x)
1
(t) +
(t, x)
11
(t), (7.20)
where (t, x) and
(t, x) are given by
(t, x) =
1
x
e
t
(1 e
x
),
(t, x) =
1
2
2
x
e
t
(1 e
x
)
_
2 e
t
e
(t+x)
_
.
Once again, if
1
=
2
then it is easily shown that
det
_
(t,
1
)
(t,
1
)
(t,
2
)
(t,
2
)
_
=
e
3t
2
3
2
(1 e
1
)(1 e
2
)(e
2
e
1
) = 0,
and so Proposition 6.1 gives
1
(t) =
(
2
)y
(t,
1
)
(
1
)y
(t,
2
)
(
1
)
(
2
)
(
1
) (
2
)
, (7.21)
11
(t) =
(
1
)y
(t,
2
) (
2
)y
(t,
1
)
(
1
)
(
2
)
(
1
) (
2
)
. (7.22)
where y
(t, x) = y(t, x)
0
(t, x). It should be noted that an expression of this form was
rst obtained in Bliss and Ritchken (1996). Finally, the bond price is given by
b(t, x) =
b(0, t + x)
b(0, t)
exp
_
x (t, x)
1
(t) x
(t, x)
11
(t)
. (7.23)
Equations (7.18) and (7.19) can be used to express the bond price in terms of benchmark
forward rates, and similarly (7.21) and (7.22) can be used to express the bond price in terms
of benchmark yields.
7.3. HJMModel with Humped Volatility. The next example considered is a 1dimensional
HJM model M
HV
corresponding to the forward rate volatility
(t, ) = r
(t) [
0
+
1
( t)] e
(t)
, (7.24)
where , ,
0
and
1
are constants. This represents a term structure model in which the
volatility is level dependent and contains a hump.
6
Since (/ + )
2
(t, ) = 0, M
HV
6
For a detailed discussion of this model with = 0, see Ritchken and Chuang (1999).
14 CARL CHIARELLA AND OH KANG KWON
is an example of the Markovian models considered in Chiarella and Kwon (2001b) and the
framework developed in the previous sections apply. Rewriting (t, ) in the form
(t, ) = (
0
t
1
)r
(t)e
t
e
+
1
r
(t)e
t
e
, (7.25)
the quantities in the decomposition (4.1) can be identied as
c
1
(t) = (
0
t
1
)r
(t)e
t
,
c
2
(t) =
1
r
(t)e
t
,
1
() = e
2
() = te
.
The state variables for M
HV
are
1
(t),
2
(t),
11
(t),
12
(t) and
22
(t) given by
11
(t) =
_
t
0
(
0
s
1
)
2
r
2
(s)e
2s
ds, (7.26)
12
(t) =
_
t
0
1
(
0
s
1
)r
2
(s)e
2s
ds, (7.27)
22
(t) =
_
t
0
2
1
r
2
(s)e
2s
ds, (7.28)
1
(t) =
_
t
0
(
0
s
1
)r
(s)e
s
dw(s)
_
t
0
(
0
s
1
)
2
r
2
(s)e
2s
(1 e
s
) ds
2
_
t
0
1
(
0
s
1
)r
2
(s)e
2s
[1 (1 + s)e
s
] ds,
(7.29)
2
(t) =
_
t
0
1
r
(s)e
s
dw(s)
_
t
0
1
(
0
s
1
)r
2
(s)e
2s
(1 e
s
) ds
2
_
t
0
2
1
r
2
(s)e
2s
[1 (1 + s)e
s
] ds,
(7.30)
where the subscripts have once again been omitted from the state variables. In this case
S
,
= and S
,
= {
1
(t),
2
(t),
11
(t),
12
(t),
22
(t)}. Now, equation (4.3) for
j
(t) gives
1
(t) =
1
(1 e
t
), (7.31)
2
(t) =
1
2
[1 (1 + t)e
t
], (7.32)
and the expression (4.15) for the forward rate in M
HV
is
r(t, x) = f(0, t + x) +
2
i=1
i
(t + x)
i
(t) +
2
i=1
i
(t + x)
i
(t + x)
i,i
(t)
+ [
1
(t + x)
2
(t + x) +
2
(t + x)
1
(t + x)]
1,2
(t).
(7.33)
It follows from the linear independence of the functions e
x
, xe
x
, x
2
e
x
, e
2x
and
xe
2x
that if
1
< <
5
, then the matrix
=
_
1
(
1
)
2
(
1
)
1
(
1
)
1
(
1
)
12
(
1
)
2
(
1
)
2
(
1
)
1
(
2
)
2
(
2
)
1
(
2
)
1
(
2
)
12
(
2
)
2
(
2
)
2
(
2
)
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
(
5
)
2
(
5
)
1
(
5
)
1
(
5
)
12
(
5
)
2
(
5
)
2
(
5
)
_
_
(7.34)
is nonsingular, where
12
() =
1
()
2
() +
2
()
1
() and
i
= t +
i
. Hence,
(7.33) can be inverted to express the state variables in terms of the forward rates r(t,
i
),
STATE VARIABLES AND AFFINE NATURE OF MARKOVIAN HJM MODELS 15
1 i 5. Formulae for the corresponding yields and bond prices can be obtained by
integrating (7.33) as in the previous examples.
7.4. TwoDimensional Extended Vasicek Model. The nal example considered is a 2
dimensional HJM model M
2D
in which the volatility is given by
i
(t, ) =
i
e
i
(t)
(7.35)
for i = 1, 2, where
i
and
i
are constants and
1
=
2
. This is the 2dimensional
extension of M
Vas
considered earlier. Since
i
(t, )/ +
i
(t, ) = 0, M
2D
falls
under the InuiKijima framework and is consequently Markov. Rewriting
i
(t, ) for each
i gives
i
(t, ) =
i
e
i
t
e
, (7.36)
and so the quantities in (3.2) are easily identied as
c
i,1
(t) =
i
e
i
t
and
i,1
() = e
. (7.37)
The state variables for M
2D
are
11
1
(t),
1
1
(t),
11
2
(t) and
1
2
(t). However, as was the case
for M
Vas
the variables
11
1
(t) and
11
2
(t) are deterministic since
11
i
(t) =
1
2
2
i
_
e
2
i
t
1
_
, (7.38)
while the variables
1
(t) and
2
(t) are stochastic since from (4.5)
11
i
(t) =
i
_
t
0
e
i
s
dw(s)
2
i
2
i
(e
i
t
1)
2
. (7.39)
So S
,
= {
11
1
(t),
11
2
(t)} and S
,
= {
1
1
(t),
1
2
(t)} for M
2D
. Hence M
2D
is a 2
factor Markovian HJM model with state variables
1
1
(t) and
1
2
(t), and (4.15) implies that
the forward rate in M
2D
is given by
r(t, x) = f(0, t + x) +
2
i=1
e
i
(t+x)
_
1
2
i
2
i
_
e
2
i
t
1
_
_
1 e
i
(t+x)
_
+
1
i
(t)
_
.
Since
1
=
2
, there exist
1
and
2
such that (
1
1
+
2
2
) (
1
2
+
2
1
) = 0. For
such a choice of
1
and
2
= det
_
e
1
(t+
1
)
e
2
(t+
2
)
e
1
(t+
2
)
e
2
(t+
1
)
_
= e
(
1
+
2
)t
_
e
(
1
1
+
2
2
)
e
(
1
2
+
2
1
)
_
= 0,
and so Proposition 5.1 applies to give
_
1
1
(t)
1
2
(t)
_
=
1
_
e
2
(t+
2
)
e
2
(t+
1
)
e
1
(t+
2
)
e
1
(t+
1
)
_ _
r
(t,
1
)
r
(t,
2
)
_
, (7.40)
where
r
(t,
j
) = r(t,
j
) f(0, t +
j
)
i=1
1
2
i
2
i
e
i
(t+
j
)
_
e
2
i
t
1
_
_
1 e
i
(t+
j
)
_
.
16 CARL CHIARELLA AND OH KANG KWON
The yields and bond prices are given by
y(t, x) =
0
(t, x) +
2
i=1
1
2
2
i
x
2
i
_
e
i
t
e
i
t
_ _
1 e
i
x
_
+
2
i=1
1
i
x
e
i
t
_
1 e
i
x
_
1
i
(t)
(7.41)
b(t, x) =
b(0, t + x)
b(0, t)
exp
_
i=1
1
2
2
i
2
i
_
e
i
t
e
i
t
_ _
1 e
i
x
_
_
exp
_
i=1
1
i
e
i
t
_
1 e
i
x
_
1
i
(t)
_
.
(7.42)
As noted previously, expressions for
1
1
(t) and
1
2
(t) from (7.40) can be substituted into
(7.42) to obtain a formula for the bond price in terms of two benchmark forward rates,
and similar arguments using yields in place of forward rates allows the bond price to be
expressed in terms of two benchmark yields.
8. CONCLUSION
The Markovian HJM models introduced in Chiarella and Kwon (2001b) generalised
many of the Markovian HJM models previously considered in the literature, but did not
address the problem of relating the state variables to market observed quantities, nor in
vestigate in depth the nature of these models. The results of this paper resolve this gap by
linking the state variables directly, and explicitly, to the forward rates and market observed
yields, and establishing that the models are in fact afne with respect to a nite number of
benchmark forward rates or yields.
Consequently, the setup in Chiarella and Kwon (2001b) provides a consistent frame
work for the systematic construction of a wide range of afne term structure models within
the HJM framework, and Chiarella and Kwon (2001a) provides an example of such a con
struction.
This paper also established an explicit formula for the bond price in terms of the state
variables which generalises the results of Ritchken and Sankarasubramanian (1995) and
Inui and Kijima (1998). In particular, it was shown that the bond price takes an exponential
afne form in a much broader class of models than the square root afne volatility models
considered in Dufe and Kan (1996).
The results obtained in this paper are of signicant value in implementing the models
in practice, and the research into the practical implementation, calibration and evaluation
of these models remains an ongoing project.
REFERENCES
Bhar, R. and Chiarella, C. (1997), Transformation of HeathJarrowMorton Models to Markovian Systems,
European Journal of Finance 3, 126.
Bj ork, T. and Gombani, A. (1999), Minimal Realizations of Interest Rate Models, Finance and Stochastics
3, 413432.
Bj ork, T. and Svensson, L. (1999), On the Existence of Finite Dimensional Realizations for Nonlinear Forward
Rate Models, Working paper, Stockholm School of Economics.
Bliss, R. and Ritchken, P. (1996), Empirical Tests of Two StateVariable HeathJarrowMorton Models, Journal
of Money, Credit, and Banking 28(3), 452481.
Carverhill, A. (1994), When is the Short Rate Markovian?, Mathematical Finance 4(4), 305312.
Chiarella, C. and Kwon, O. (2001a), Formulation of Popular Interest Models under the HJM Framework, Asia
Pacic Financial Markets (Forthcoming).
Chiarella, C. and Kwon, O. (2001b), Forward Rate Dependent Markovian Transformations of the HeathJarrow
Morton Term Structure Model, Finance and Stochastics 5(2), 237257.
STATE VARIABLES AND AFFINE NATURE OF MARKOVIAN HJM MODELS 17
Coddington, E. and Levinson, N. (1955), Intoduction to Ordinary Differential Equations, McGrawHill Publish
ing Inc., New York.
Cox, J., Ingersoll, J. and Ross, S. (1985), A Theory of the Term Structure of Interest Rates, Econometrica
53(2), 385407.
de Jong, F. and SantaClara, P. (1999), The Dynamics of the Forward Interest Rate Curve: A Formulation with
State Variables, Journal of Financial and Quantative Analysis 34(1), 131157.
Dufe, D. and Kan, R. (1996), A Yield Factor Model of Interest Rates, Mathematical Finance 6(4), 379406.
Heath, D., Jarrow, R. and Morton, A. (1992), Bond Princing and the Term Structure of Interest Rates: A New
Methodology for Contingent Claim Valuation, Econometrica 60(1), 77105.
Ho, T. and Lee, S. (1986), Term Structure Movements and Pricing Interest Rate Contingent Claims, Journal of
Finance 41, 10111028.
Hull, J. and White, A. (1990), Pricing Interest Rate Derivative Securities, Review of Financial Studies 3, 573
592.
Inui, K. and Kijima, M. (1998), A Markovian Framework in MultiFactor HeathJarrowMorton Models, Jour
nal of Financial and Quantitative Analysis 33(3), 423440.
Kwon, O. (2001), Characterisation of Afne Term Structure Models, Working Paper, School of Banking and
Finance, University of New South Wales.
Musiela, M. (1993), Stochastic PDEs and Term Structure Models, Journ ees Internationales des Finance, IGR
AFFI, La Baule .
Ritchken, P. and Chuang, Y. (1999), Interest Rate Option Pricing with Volatility Humps, Review of Derivatives
Research 3, 237262.
Ritchken, P. and Sankarasubramanian, L. (1995), Volatility Structures of Forward Rates and the Dynamics of the
Term Structure, Mathematical Finance 5(1), 5572.
Vasicek, O. (1977), An Equilibrium Characterisation of the Term Structure, Journal of Financial Economics
5, 177188.