Article

Affine Transformations of Itô Diffusions and their Transition Densities

Sanae RUJIVAN

Division of Mathematics, School of Science, Walailak University, Nakhon Si Thammarat 80161, Thailand

(Corresponding author; e-mail: [email protected])

Received: 22nd December 2010, Revised: 7th April 2011, Accepted: 27th April 2011

Abstract
For a given Itô diffusion, we derive the forward Kolmogorov equation (FKE) associated with the
adjoint operator of the infinitesimal generator of an affine transformation of the given Itô diffusion. The
fundamental solution obtained by solving the FKE is, in fact, the transition density of the transformed
diffusion. Moreover, we prove that the transition density can be represented in terms of a product of two
functions, a Jacobian term and a composition of the transition density of the given Itô diffusion and the
inverse of the transformation. Finally, we present an application of our results in parameter estimation in
commodity markets in which the commodity prices are assumed to follow an extended Black-Scholes
model.
Keywords: Itô diffusions, forward Kolmogorov equation, transition densities

Introduction
Itô stochastic differential equations (Itô SDE) are known as the drift coefficient and the diffusion
are a natural choice to model the time evolution of coefficient, respectively. These functions are
dynamic systems which are subject to random known except the unknown parameter vector θ
influences [1-2]. For example, in physics the which is assumed to belong in a particular
dynamics of ions in super-ionic conductors are parameter space Θ ⊆ r for some positive integer
modeled via Langevin equations [3], and in r and E [ζ 2 ] < ∞. Under Lipschitz and the linear
engineering the dynamics of mechanical devices
growth conditions on the coefficients b and a,
are described by differential equations under the
influence of process noise as errors of measure- there exists a unique non-exploding strong solution
ment [4]. Other applications are in biology [5], of the above Itô SDE, called the Itô diffusion
medicine [6], econometrics [7], finance [8], geo- process or simply the Itô diffusion, which is a
physics [9], and oceanography [10]. Nevertheless, continuous strong Markov semi-martingale. The
the models are less used if they contain unknown drift and the diffusion coefficients are respectively
parameters that are without estimation. the instantaneous mean and instantaneous standard
It is natural that a model contains unknown deviation of the process. It should be noted that the
parameters. We consider the model as the Itô SDE diffusion coefficient is almost surely determined
by the process, i.e. it can be estimated without any
dX t = b(t, X t ; θ)dt + a (t, X t ; θ )dWt , t ≥ 0, X 0 = ζ (1) error if observed continuously throughout a time
where {Wt , t ≥ 0} is a one-dimensional Wiener interval (see [11,12]). However, continuous
observation of an Itô diffusion is a mathematical
process. The functions idealization because the path of the Itô diffusion is
b : [0, T ] × ×Θ → and very jagged and no measuring device can follow
an Itô diffusion trajectory continuously. Hence, the
a : [0,T ] × ×Θ → \ {0} observation is always discrete in practice.
Consequently, research on discretely observed Itô

Walailak J Sci & Tech 2011; 8(1): 71‐79.

Affine Transformations of Itô Diffusions and their Transition Densities S RUJIVAN

diffusions is growing recently with a powerful where for a probability space (Ω, F , Pθ ),
theory of statistical inference for Itô diffusions.
Statistical inference for Itô diffusions based DX = {x ∈ | ∃t ∈ [0,T ], Pθ (X t = x ) = 1}
on discrete-time observations is based on the denotes the domain of X . We assume for
likelihood function (see for example in [13-15]).
simplicity that DX is a bounded domain, i.e. an
The literature has mainly concentrated on the
maximum likelihood (ML) approach. Suppose that open connected bounded subset of . In view of
an Itô diffusion X = {Xt , t ≥ 0} is observed at the theory of partial differential equations of a
times ti ’s with 0 = t 0 < t1 < t2 < ... < tn = T . parabolic type, pX which satisfies the FKE (3) and

Since X is Markov, hence if the transition the condition (4) is, in fact, a fundamental solution
probability density function or simply the of the FKE (3) (see Chapter 1 of Friedman [20]).
transition density of X is known and denoted by In view of the theory of Markov processes, we
have
pX ≡ pX (t, x , s, x 0 ; θ ), one can use the log
likelihood function Pθ (X t ∈ B | X s = x 0 ) = ∫ B
pX (t, x , s, x 0 ; θ )dx
n

ln (θ) = ∑ ln p X
(ti , X t , ti −1, X t ; θ) (2) for every Borel set B ⊆ (see Theorem 5.4 of
i =1
i i −1
Chapter 6 of Friedman [21]). One can see from (2)
that, in order to follow the method of ML, the FKE
to estimate θ. Under the regularity conditions, the
has to be solved to obtain a closed form of pX or
corresponding maximum likelihood estimate θˆ is n

known to have usual good properties such as the at least the approximates of pX at the observed
consistency and asymptotic normality property
(see for example in [16-19]). points (ti , X t ), for i = 0, 1, ..., n .
i

According to the theory of Markov pro- In terms of analytical methods for solving the
cesses, under a set of conditions on the coefficients FKE (3) subject to the condition (4), for several
b and a and for a fixed parameter vector θ, pX special cases of the drift and diffusion coefficients,
the corresponding transition densities can be
satisfies the forward Kolmogorov equation (FKE),
derived in closed form. For instance,
for every fixed (s, x 0 ) ∈ [0, T ) × DX : a (t, x ) = σ, and b(t, x ) = μ
∂pX where σ > 0, μ are constants (the corresponding
(t, x , s, x 0 ; θ) = At* pX (t, x , s, x 0 ; θ)
∂t Itô diffusion is called a Brownian motion with drift
for all (t, x ) ∈ (s,T ] × DX , (3) μ ), its transition density is
subject to the condition pX (t, x , s, x 0 ; θ) =
⎛ 2⎞
lim ∫ pX (t, x , s, x 0 ; θ)f (x )dx = f (x 0 )
t ↓s 1
exp ⎜⎜−
(
⎜⎜ x − (x 0 + μ (t − s ) ⎟⎟
⎟⎟ , ) (6)
DX
⎜ 2(t − s )σ 2 ⎟⎟
⎜⎜
⎠⎟⎟
2π(t − s ) σ
for all f ∈ C 0 (DX ), (4) ⎝
where C 0 (D ) denotes the set of bounded
where θ = (μ, σ ). One can verify that pX as
continuous functions on D ⊆ . The operator At*
expressed in (6) satisfies the FKE (3) and the
on the RHS of (3), known as the adjoint operator
of the infinitesimal generator of X starting at condition (4), namely, pX is the fundamental
Xt = x ∈ DX , is given by solution of the FKE (3) for which

1 ∂
2
∂ 1 2 ∂2 ∂
*
At p(x ) =
2 ∂x 2
(a (t, x )p(x )) − ∂x (b(t, x )p(x ))
2
At* ≡ σ
2 ∂x 2
−μ
∂x
.

for all p ∈ C 2 ( ), (5)

72 Walailak J Sci & Tech 2011; 8(1)

Affine Transformations of Itô Diffusions and their Transition Densities S RUJIVAN

In many cases, however, a set of data which Assumption 2: The coefficients a (t, x ) and
relates to the transformation of X, i.e. b(t, x ) satisfy the linear growth condition in the
{F (ti , X t ; θ), i = 0, 1, 2, ..., n },
i space variable x , i.e.
2 2 2 2
is observed instead of the data of X . Although, for | a (t, x ) | + | b(t, x ) | ≤ K (1+ | x | )
each time ti , the transformation is one-to-one in for all (t, x ) ∈ [0, T ] × DX , where K is a positive
the space variable x , the data of X cannot be constant.
retrieved since the transformation depends on
unknown parameters. Fortunately, the method Assumption 3: There is λ>0 such that
of ML still works by replacing the transition a (t, x ) ≥ λ for all (t, x ) ∈ [0, T ] × DX . In other
density pX in (2) by the transition density of the words, a (t, x ) is uniformly elliptic on for all
transformed process Yt = F (t, X t ; θ). (t, x ) ∈ [0,T ] × DX .
The aim of this paper is to derive the
transition density of a transformed process Assumption 4: The partial derivatives
Y = {Yt , t ≥ 0} which is an affine transformation
2
∂a ∂a ∂b
, , and
of X, defined as follows: for a fixed parameter ∂x ∂x 2
∂x
vector θ, are Lipschitz continuous in both the time variable t
Yt = m(t ; θ )X t + c(t ; θ ) (7) and space variable x on [ 0,T ] × DX .
The following theorem is an important
for all t ∈ [0,T ], where m(t ; θ ) and c(t ; θ ) are implication of Friedman’s work in [20] and [21].
deterministic functions depending on the time t He established a connection between fundamental
and parameter θ. We provide a set of sufficient solutions of parabolic partial differential equations
conditions for proving that the transition density of and transition densities of Itô diffusions and
Y , denoted by pY , can be represented in terms of investigated some of their properties. We state the
theorem below without proof, because the proof is
a product of two functions, a Jacobian term and a quite intricate and needs several pages.
composition of pX and the inverse of the Theorem 1: Under Assumptions 1 - 4, there is a
transformation. unique fundamental solution of the FKE (3).
Moreover, if pX is the fundamental solution then it
Affine transformations of Itô diffusions and
their transition densities is nonnegative and its partial derivatives
In the present section, we provide sufficient ∂p X ∂p X
2
∂ pX
conditions to guarantee the existence and , , and
uniqueness of the transition density of Y . Note ∂t ∂x ∂x
2

that from this point forward, the parameter vector are continuous in the variables t, x , s, and x 0 on
θ is considered to be fixed and we omit writing θ
as an argument of the relevant functions. E X = {(t , x , s, x 0 ) | 0 ≤ s < t ≤ T , x , x 0 ∈ DX }.
We begin with the sufficient conditions to In addition,
guarantee the existence and uniqueness of the
transition density of X. ∫p X
(t, x , s, x 0 )dx = 1
D
X

Assumption 1: The coefficients a (t, x ) and

for all 0 ≤ s < t ≤ T and all x 0 ∈ DX .
b(t, x ) are Lipschitz continuous in both the time
Remark 1: We recommend Chapters 1 and 9 of
variable t and space variable x on [ 0,T ] × DX . Friedman [20] and Chapter 6 of Friedman [21] for
the proofs of the obtained results in the theorem.

Walailak J Sci & Tech 2011; 8(1) 73

Affine Transformations of Itô Diffusions and their Transition Densities S RUJIVAN

Remark 2: The condition in Assumption 3 can be DX as previously assumed, DY is also a bounded

relaxed somewhat by making it the local
domain in .
requirement (the condition A3b′ in [22]) as
From the previous section, if the transition
follows: density of Y exists then it must satisfy the FKE,
∞
Assumption 3′: There is a sequence (Dn )n =1 of for every fixed (s, y 0 ) ∈ [0,T ) × DY :

bounded domains with Dn ⊆ DX such that ∂pY

(t, y, s, y 0 ) = Aˆt pY (t, y, s, y 0 )
*

∞ ∂t
DX = ∪D n
and for each n there is λn > 0
for all (t, y ) ∈ (s, T ] × DY , (11)
i =1

such that a (t, x ) ≥ λn for all (t, x ) ∈ Dn . subject to the condition

Replacing Assumption 3 by Assumption 3′, lim

t ↓s
∫p Y
(t, y, s, y 0 )f (y )dy = f (y 0 )
the conclusions of Theorem 1 still hold. The D
Y

following assumption is required for preserving Y 0

for all f ∈ C ( ). (12)
as a random process and for applying Itô’s lemma.
The operator Aˆt
*
denotes the adjoint operator
Assumption 5: The real-valued functions m (t )
2
of the infinitesimal generator of Y starting at
and c(t ) belong to C ([0,T ]) and m(t ) is either
Yt = y ∈ DY ,
negative or positive on [0,T ].
2
Under Assumption 5, one can easily see that 1 ∂ ∂ ˆ
Y is a random process because m(t ) never
Aˆt p(y ) =
*

2 ∂y 2
2
(
aˆ (t, y )p(y ) −
∂y
)
(b (t, y )p(y ))
reaches zero for all t ∈ [0, T ]. Furthermore, m (t ) 2
for all p ∈ C ( ). (13)
and c(t ) are smooth on [0, T ], and hence, we can
apply Itô’s lemma to Y as defined in (7). Then, Corollary 1: Under Assumptions 1-5, there is a
we have Y that is also an Itô diffusion and unique fundamental solution of the FKE (11).
satisfies the Itô SDE Moreover, if pY is the fundamental solution then it
dYt = bˆ(t,Yt )dt + aˆ(t,Yt )dWt , t ≥ 0, is nonnegative and its partial derivatives
2
∂pY ∂pY ∂ pY
Y0 = m (0)X 0 + c(0), (8) , , and
∂t ∂y ∂y
2

where the coefficient functions in (8) are given by,

are continuous in the variables t, y, s, and y 0 on
for any (t, y ) ∈ [0, T ] × DY ,
⎛ y − c(t ) ⎞⎟ EY = {(t, y, s, y 0 ) | 0 ≤ s < t ≤ T , y, y 0 ∈ DY }.
aˆ(t, y ) = m(t )a ⎜⎜t, ⎟, (9) In addition,
⎜⎝ m(t ) ⎠⎟
⎛ y − c(t ) ⎞⎟ ∫p Y
(t, y, s, y 0 )dy = 1
bˆ(t, y ) = m ′(t ) ⎜⎜ ⎟ + c ′(t ) D

⎜⎝ m(t ) ⎠⎟
Y

(10) for all 0 ≤ s < t ≤ T and all y 0 ∈ DY .

⎛ y − c(t ) ⎞⎟
+ m(t )b ⎜⎜t, ⎟. Proof. Suppose that Assumptions 1-5 are fulfilled.
⎜⎝ m(t ) ⎠⎟ It is easy to show that the coefficient functions
Here, DY denotes the domain of Y and m ′(t ) and aˆ(t, y ) and bˆ(t, y ) of the Itô diffusion Y as
written in (9) and (10) satisfy the conditions in
c ′(t ) signify the derivatives of m(t ) and c(t ) Assumptions 1-4 with respect to the variables t
with respect to t , respectively. It should be noted and y . It should be pointed that bˆ(t, y ) is
that, under Assumption 5 and the boundedness of

74 Walailak J Sci & Tech 2011; 8(1)

Affine Transformations of Itô Diffusions and their Transition Densities S RUJIVAN

Lipschitz continuous since m (t ) and c(t ) belong Lemma 1: pY satisfies the FKE (11) on EY .
to C ([0,T ]) which implies that m ′(t ) and c ′(t )
2

are Lipschitz continuous. Hence, by Theorem 1, Lemma 2: pY satisfies the condition (12).
the conclusions of the corollary are obtained.
The proofs of Lemmas 1 and 2 are provided in
Appendices A and B, respectively.
Main results Now, we state the main theorem.
Theorem 2: Let X be a random variable for Theorem 3: Under Assumptions 1-5,
which the probability density function (p.d.f.) is fX pY = pY on EY .
and Pr(a < X < b ) = 1. Let Y = g (X ), and Proof. By Lemma 1 and 2, pY satisfies the FKE
suppose that g (x ) is continuous and either strictly (11) and the condition (12). From the existence
increasing or strictly decreasing on [a, b ]. and uniqueness of the fundamental solution in
Suppose also that a < X < b if and only if Corollary 1, pY must coincide with pY on EY .
α < Y < β, and let X = h (Y ) be the inverse
function for α < Y < β. Then the p.d.f. of Y is Application in parameter estimation in
specified by the relation commodity markets
⎧⎪ dh(y ) In this section, we consider parameter
⎪⎪ f (h (y )) for α < y < β
fY (y ) = ⎨ dy X estimation of an extended Black-Scholes model
⎪⎪ described by the Itô SDE as follows:
0 otherwise.
⎪⎪⎩
In particular, if Y = mX + c, for some constants dS t = (r − δ(t ))S t dt + σS tdWt , t ≥ 0, (15)
m ≠ 0 and c ∈ then the p.d.f. of Y is given K

by δ(t ) = α0 + ∑α k
sin(2πkt ), (16)
⎪⎧⎪ 1 ⎛ y − c ⎞⎟ k =1

⎪ f ⎜⎜ ⎟ for min(α, β ) < y < max(α, β )

fY (y ) = ⎨| m | X ⎜⎝ m ⎠⎟
⎪⎪ where S = {S t , t ≥ 0} is a price process, δ(t ) is
0 otherwise,
⎪⎩⎪
a deterministic function of time t, r is a risk-free
where α = ma + c and β = mb + c.
Proof. The proof of the theorem is provided in interest rate, and σ > 0, αi , for i = 0, 1, ..., K
Chapter 2 of [23]. for some integer K ≥ 1, are unknown parameters.
We now return to the transformation (7). One can verify that the drift and diffusion
Theorem 2 guides us to construct a nonnegative coefficients of S sufficiently satisfy the
function which is claimed to coincide with the
transition density of Y as follows. For a fixed conditions in Section 2, i.e. Assumptions 1, 2, 3′,
4, and 5. The models (15) and (16) describe,
(s, y 0 ) ∈ [0, T ) × DY , we define
respectively, dynamics of commodity prices and
⎛ y − c(t ) y − c(s ) ⎞⎟ seasonal variation of convenience yields (see
1
pY (t, y, s, y 0 ) = pX ⎜⎜t, ⎟ Chapter 2 in [24] for the definition of convenience
⎜⎝ m(t ) , s, m(s ) ⎠⎟⎟
0

| m(t ) | yields). From the Introduction, we need the

transition density of S for constructing the log-
for (t, y ) ∈ (s,T ] × DY . (14) likelihood function, and hence, for calculating the
It should be noted that the first term on the maximum likelihood estimates of the unknown
RHS of (14) is the absolute value of the Jacobian parameters. The associated FKE of S is
of the inverse of the transformation (7) and 2
∂p S 1 ∂ pS ∂p S
the second term is the composition of pX and =
2
σy
2
− (r − δ(t )) , (17)
∂t 2 ∂y
2
∂y
the inverse of the transformation (7).
Under Assumptions 1-5, we have the for every fixed (s, y 0 ) ∈ [0, T ) × DS ,
following two lemmas.

Walailak J Sci & Tech 2011; 8(1) 75

Affine Transformations of Itô Diffusions and their Transition Densities S RUJIVAN

where pS ≡ pS (t , y, s, y 0 ) and DS denote, the where pX is the transition density of X which is

transition density of S and the domain of S , known in closed form as
respectively. To avoid solving the FKE (17), we pX (t, x , s, x 0 ; θ) =
notice that S is, in fact, an affine transformation 1 ⎛ (ln(x / x ) − (r − 1 σ 2 )(t − s ))2 ⎞⎟
of a log-normal diffusion: exp ⎜⎜⎜− 0 2
⎟⎟ ,
⎝⎜ ⎠⎟
2
2π(t − s ) σx 2σ (t − s )
S t = m (t )X t , t ≥ 0 (18)
where θ = (σ, α0 , α1 , ..., αK ).
where

m(t ) = exp − ( ∫
0
t
δ(η )d η , ) Conclusions
This paper provides a set of sufficient
and X = {X t , t ≥ 0} is a log-normal diffusion conditions for the existence and uniqueness of the
(the Black-Scholes model), i.e. transition densities of affine transformations of a
given Itô diffusion. Moreover, the paper has
dX t = rX tdt + σ X tdWt . established a general formula for the transition
densities of the transformed Itô diffusions,
Applying Theorem 3 to the transformation (18), providing the transition density of the given Itô
for discretely observed data {(ti , S t ), i = 0, 1, ..., n }, diffusion. An application to parameter estimation
i
in commodity markets has been demonstrated.
we obtain Finally, an extension of these results to affine
(
pS ti , S t , ti −1 , S t ; θ =
i i −1
) transformations of multivariate Itô diffusions will
be investigated in future work.
1
×
exp − ( ∫ 0
t
i

δ(η )d η ) Acknowledgements
⎛ ⎞⎟ This research has been financially supported
⎜⎜ St St ⎟⎟ by the Institute of Research and Development,
pX ⎜⎜ti , , ti −1 , i
; θ ⎟⎟ , i −1

⎜⎜ t

⎜⎝ exp −∫0 δ(η)d η ( i

0
t

)
exp −∫ δ(η )d η
⎠⎟
⎟⎟
( i −1

) Walailak University under the contract WU52204.

76 Walailak J Sci & Tech 2011; 8(1)

Affine Transformations of Itô Diffusions and their Transition Densities S RUJIVAN

Appendix A (Proof of Lemma 1)

Proof. Assume that m (t ) is positive on [0, T ]. From Theorem 1, the continuities of the partial derivatives
of pX allow us to apply the chain rule to the formula of pY . The results obtained are the following
relations:
∂pY ⎛ 1 ⎞⎟
d 1 ⎛⎜ ∂pX ∂p dx ⎞⎟
= ⎜⎜⎜ ⎟ pX + ⎜ + X ⎟⎟
∂t dt ⎝ m(t ) ⎠⎟ m(t ) ⎜⎝ ∂t ∂x dt ⎠⎟

m ′(t ) 1 ∂p X
⎛c(t )m ′(t ) − c ′(t )m(t ) m ′(t ) ⎞⎟ ∂p
⎜ ⎟⎟ X ,
= − 2 pX + + ⎜⎜ − (m (t )x + c (t ))
⎜⎝ ⎟
m(t ) ∂t ⎠⎟ ∂x
3 3
m (t ) m (t ) m (t )
2 2
∂pY 1 ∂p X ∂ pY 1 ∂ pX
= , and = , where for any (t , y ) ∈ [s, T ] × DY ,
∂y m (t ) ∂x
2
∂y
2 3
m (t ) ∂x
2

y − c(t )
pY ≡ pY (t, y, s, y 0 ), pX ≡ pX (t, x (t ), s, x (s )), and x ≡ x (t ) = .
m(t )
We use the above relations to compute the partial derivatives contained in Aˆt pY . Thus we have
*

∂ ⎪⎧ m ′(t ) ∂b ⎛⎜ y − c(t ) ⎞⎫
⎟⎟⎪⎪⎬ p
{bˆ(t, y )p } = ⎪⎨⎪ + ⎜⎜t,
∂x ⎝ m(t ) ⎠⎟⎪
Y Y
∂y ⎩⎪ m(t ) ⎭⎪
⎧ ⎛ y − c(t ) ⎞⎟
⎪ ⎧⎪c ′(t )m(t ) − c(t )m ′(t )⎫⎪ m ′(t ) ⎫
⎪ ∂pY
+⎪⎨b ⎜⎜⎜t, ⎟⎟ m(t ) + ⎪⎨ ⎪
⎬+ y⎪
⎬
⎪
⎪ ⎝ m(t ) ⎠
⎩ ⎪
⎪
⎩ m(t ) ⎪
⎪ m(t ) ⎪
⎭ ⎪ ∂y
⎭
m ′(t ) ⎪⎧⎪c(t )m ′(t ) − c ′(t )m(t ) m ′(t ) ⎫ ∂p
⎪
⎪ X
= pX − ⎨ − (m (t )x + c (t ))⎬
2
m (t ) ⎪⎪ 3
m (t )
3
m (t ) ⎪
⎪ ∂x
⎩ ⎭
1 ∂
+ {b(t, x )p } X
m(t ) ∂x

∂
2 ⎪⎧ 2 ⎪ ⎧ ∂a ⎛⎜ y − c(t ) ⎞⎫ ⎪
2
⎛ y − c(t ) ⎞⎟ ∂ 2a ⎛ y − c(t ) ⎞⎟⎪⎪⎫
{aˆ (t ,
2
y )p ⎪ ⎪ t , ⎟ ⎪
} ⎨⎪ 2 ⎨⎪ ∂x ⎜⎜⎝ m(t ) ⎠⎟⎬⎪ ⎝⎜⎜⎜t, m(t ) ⎠⎟⎟ 2 ⎝⎜⎜⎜t, m(t ) ⎠⎟⎟⎬⎪ pY
= ⎟ + a
2 Y
∂y ⎪⎩⎪m (t ) ⎩⎪ ⎪
⎭ ∂x ⎪⎭⎪
⎧⎪ 4 ⎛ y − c(t ) ⎞⎟ ∂a ⎛ y − c(t ) ⎞⎫ ⎪ ∂pY ⎛ y − c(t ) ⎞⎟ ∂ 2 pY
+⎪ ⎜ ⎜ ⎟⎪ 2 ⎜
⎨ a ⎜⎜t , ⎟ ⎜ t , ⎟⎬ + a ⎜⎜t, ⎟
⎪
⎪
⎩ m (t ) ⎝ m (t ) ⎟
⎠ ∂ x ⎜
⎝ m (t ) ⎟
⎠⎪
⎪
⎭ ∂ y ⎝ m(t ) ⎠⎟ ∂y 2

1 ⎪
⎧ ⎪ ⎧ ∂a ⎫
⎪
2
∂a
2 ⎫
⎪ 4 ∂a ∂p
= ⎪
⎨ ⎨
2 ⎪ (t , x )⎪
⎬ + 2 a (t , x ) (t , x )⎪
⎬ pX + a (t, x ) (t, x ) X
m(t ) ⎪ ⎪
⎪ ∂x ⎪⎭
⎪ ∂x
2
⎪⎪ m(t ) ∂x ∂x
⎪ ⎩
⎪
⎩ ⎪
⎭
2
1 ∂ pX 1 ∂
2

m(t )
+
2
a (t, x )
m(t ) ∂x 2
= 2
{a (t, x )p } .
2
X
∂x
Replacing the partial derivatives contained in the following FKE with the above relations gives us
∂pY ∂pY ∂ 1 ∂
2

∂t
− Aˆt pY =
*

∂t
+
∂y
{bˆ(t, y )pY } −
2 ∂y 2
{aˆ (t, y )p }
2
Y

Walailak J Sci & Tech 2011; 8(1) 77

Affine Transformations of Itô Diffusions and their Transition Densities S RUJIVAN

m ′(t )
⎧c(t )m ′(t ) − c ′(t )m(t ) m ′(t )
⎪ ∂p X ⎫ ∂p
⎪
⎪
1
+⎨
⎪ X
=− pX + − (m (t )x + c (t ))⎬
2
m (t ) m(t ) ∂t ⎪
⎪
3
m (t )
3
m (t ) ⎪
⎪ ∂x
⎩ ⎭
⎧
⎪ ′ ′ ⎫ ∂p
m ′(t ) ⎪c(t )m (t ) − c (t )m(t ) − m ′(t ) (m(t )x + c(t ))⎪ ⎪ X + 1 ∂ b(t, x )p
+ 2 pX − ⎨
⎪
⎬
⎪
{ X
}
⎪ ∂x m(t ) ∂x
3 3
m (t ) ⎪ m (t ) m (t )
⎩ ⎭
⎧ 1 ∂2 2
1⎪ ⎪⎫
− ⎪
⎨ {a (t, x )pX }⎪⎬
2⎪
⎪ m(t ) ∂x
⎩
2
⎪⎭⎪
1 ∂p X 1 ∂ 1 ∂
2

=
m(t ) ∂t
+
m(t ) ∂x
{b(t, x )p } − X
2m(t ) ∂x 2
{a (t, x )p }
2
X

⎧⎪ ∂p ⎫
=
1 ⎪⎨ X − A * p ⎪⎪⎬ = 0.
m(t ) ⎪ ⎪⎭⎪
t X
⎩⎪ ∂t
In the case m(t ) being negative on [0, T ], one can show by following the above steps that pY satisfies

the FKE (11) on (s,T ] × DY . This completes the proof.

Appendix B (Proof of Lemma 2)

0
Proof. Let f ∈ C (DY ). From Theorem 1, pX (t, x , s, x 0 ) is continuous in both t and x for which
(t , x ) ∈ (s, T ] × DX and satisfies the condition (4). Moreover, m(t ) and c(t ) are continuous on [0, T ].
Thus we can compute the following limit:
1 ⎛ y − c(t ) y − c(s ) ⎞⎟
∫ pY (t, y, s, y 0 )f (y )dy = lim ∫ | m(t ) | pX ⎜⎜t, ⎟
lim
⎜⎝ m(t ) , s, m(s ) ⎠⎟⎟ f (y )dy
0

t ↓s t ↓s
D D
Y Y

⎪⎧⎪ 1 ⎛ y − c(τ ) y − c(s ) ⎞⎟ ⎪⎫⎪

⎜⎜t, ⎟
= lim ⎨lim
t ↓s ⎪⎪ τ ↓s ∫ | m (τ ) | pX ⎜
⎝ m(τ )
, s , 0
⎟
m(s ) ⎠⎟
f (y ) dy ⎬
⎪⎪
⎩ D
Y ⎭
⎧⎪ 1 ⎛ y − c(τ ) y − c(s ) ⎞⎫ ⎟⎟⎪⎪
∫ lim ⎪ ⎨ pX ⎜⎜t,
= lim
⎜⎝ m(τ ) , s, m(s ) ⎠⎟⎟⎪⎬ f (y )dy
0

τ ↓s ⎪| m (τ ) |
⎩⎪ ⎭⎪
t ↓s
D
Y

⎛ ⎞
1 ⎜t, lim ⎜⎛ y − c(τ ) ⎞⎟⎟ , s, y 0 − c(s ) ⎟⎟ f (y )dy
= lim
t ↓s
∫ | m(s ) | p ⎜
X ⎜ ⎜
⎝ τ ↓s ⎝⎜ m(τ ) ⎠⎟ m(s ) ⎠⎟
⎟
D
Y

1 ⎛ y − c(s ) y − c(s ) ⎞⎟
= lim ∫ pX ⎜⎜t, ⎟
⎜⎝ m(s ) , s, m(s ) ⎠⎟⎟ f (y )dy
0

t ↓s
D
| m(s ) |
Y

1
= lim ∫ pX (t, x , s, x 0 ) f (m(s )x + c(s )) J X dx
t ↓s
D
| m(s ) |
X

= lim ∫ pX (t, x , s, x 0 ) f (m(s )x + c(s )) dx

t ↓s
D
X

78 Walailak J Sci & Tech 2011; 8(1)

Affine Transformations of Itô Diffusions and their Transition Densities S RUJIVAN

= f (m(s )x 0 + c(s ))

= f (y 0 ),

y 0 − c(s )
where J X denotes the Jacobian of the transformation x y = m(s )x + c(s ) and x 0 = .
m(s )
The proof is now complete.

References
[1] L Arnold. Stochastic Differential Equations, [13] Y Aït-Sahalia. Maximum likelihood
Wiley, New York, 1974, p. 1-288. estimation of discretely-sampled diffusions: a
[2] NG Van Kampen. Stochastic Processes in closed form approximation approach.
Physics and Chemistry, North Holland, Econometrica 2002; 70, 223-62.
Amsterdam, 1981, p. 1-451. [14] A Egorov, H Li and Y Xu. Maximum
[3] W Dieterich, P Fulde and I Peschel. Likelihood of Time Inhomogeneous
Theoretical models for super-ionic Diffusion. J. Econometrics 2003; 114, 107-
conductors. Adv. Physics. 1980; 29, 527-605. 39.
[4] A Gelb. Applied Optimal Estimation, MIT [15] Y Aït-Sahalia. Closed-Form Likelihood
Press, Cambridge, MA, 1974, p. 51-84. Expansions for Multivariate Diffusions. Ann.
[5] RI Jennrich and PB Bright. Fitting systems of Stat. 2008; 36, 906-37.
linear differential equations using computer [16] P Billingsley. Statistical Inference for
generated exact derivatives. Technometrics Markov Processes, Chicago University Press,
1976; 18, 385-92. Chicago, 1961, p. 1-67.
[6] RH Jones. Fitting multivariate models to [17] MJ Crowder. Maximum Likelihood
unequally spaced data, In: E Parzen (ed.). Estimation for Dependent Observations.
Time Series Analysis of Irregularity Journal of Royal Statistical Society, Series B
Observed Data. Lecture Notes in Statistics 25 1976; 38, 45-53.
Springer-Verlag, New York, 1984, p. 155-88. [18] IV Basawa and BLS Prakasa Rao. Statistical
[7] AR Bergstrom. Statistical Inference in Inference for Stochastic Processes,
Continuous Time Economic Models, North- Academic Press, New York, 1980, p. 67-143.
Holland, Amsterdam, 1976, p. 267-327. [19] D Dacunha-Castelle and D Florens-Zmirou.
[8] F Black and M Scholes. The pricing of Estimation of the coefficients of a diffusion
options and corporate liabilities. J. Polit. from discrete observations. Stochastics 1986;
Econ. 1973; 81, 637-54. 19, 263-84.
[9] M Arato. Linear Stochastic Systems with [20] A Friedman. Partial Differential Equations
Constant Coefficients: A Statistical of Parabolic Type, Prentice-Hall, Englewood
Approach. In: Lecture Notes in Control and Cliffs, New Jersey, 1964, p. 1-268.
Information Sciences 45, Springer-Verlag, [21] A Friedman. Stochastic Differential
New York, 1982, p. 1-309. Equations, Vol. I, Academic Press, New
[10] R Adler, P Muller and BL Rozovskii. York, 1975, p. 128-50.
Stochastic Modeling in Physical [22] D Heath and M Schweizer. Martingales
Oceanography, Birkhauser, Boston - Berlin, versus PDEs in Finance: An Equivalence
1996, p. 1-349. Result with Examples. J. Appl. Probab.
[11] JL Doob. Stochastic Processes, John Wiley, 2000; 37, 947-57.
New York, 1953, p. 46-560. [23] G Casella and LB Roger. Statistical
[12] V Genon-Catalot and J Jacod. Estimation of Inference, Duxbury, California, 2002, p. 50-
Diffusion Coefficient for Diffusion 1.
Processes: Random Sampling. Scand. J. [24] H Geman. Commodities and Commodity
Statist. 1994; 21, 193-221. Derivatives, John Wiley & Sons Ltd., West
Sussex, 2005, p. 1-22.

Walailak J Sci & Tech 2011; 8(1) 79