Professional Documents
Culture Documents
Contents
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
5.2 The Backwards Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
5.3 Infinite Horizon Discounted Reward . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
5.4 Expected Reward to a Hitting Time . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
5.5 Bounding Expected Reward to a Hitting Time . . . . . . . . . . . . . . . . . . . . . 5
5.6 PDEs for Diffusions with Absorbing Boundaries . . . . . . . . . . . . . . . . . . . . . 7
5.7 PDEs for Diffusions with Reflecting Boundaries . . . . . . . . . . . . . . . . . . . . . 9
5.8 The Generator of a Diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
5.9 Integro-PDEs for Diffusions with Jump Boundaries . . . . . . . . . . . . . . . . . . . 14
5.1 Introduction
In Section 1, we saw how we can heuristically derive the partial differential equations (PDEs)
that need to be solved in order to compute various probabilities and expectations associated with
a given diffusion process. In this chapter, we show how the Itô-Doeblin calculus can be used to
rigorously validate this connection. This rigorous discussion will also serve to clarify the role of
boundary conditions in uniquely specifying the probabilistically interesting solution to the PDE.
Step 1: Assume that Ex [|f (X(t))|] < ∞ for all x ∈ Rd . If we set Y = (f (X(t)) : t ≥ 0), then
E [Y |X(u) : 0 ≤ u ≤ s]
E [Y |X(u) : 0 6 u 6 s] = u∗ (t − s, X(s)).
Hence, we know that (u∗ (t−s, X(s)) : 0 6 s 6 t) is a Px - martingale for each x ∈ Rd . Furthermore,
u∗ (0, x) = f (x).
1
§ SECTION 5: THE CONNECTION BETWEEN SDE’S AND PDE’S
Step 2: In view of Step 1, we should seek conditions on u so that u(0, x) = f (x) and (u(t −
s, X(s)) : 0 6 s < t) is a Px -martingale. Assume that u ∈ C 1, 2 . In view of Itô’s formula, we find
that
d
∂u(t − s, X(s)) X ∂u(t − s, X(s))
du(t − s, X(s)) = − + µi (X(s))
∂t ∂xi
i=1
d 2
1 X ∂ u(t − s, X(s))
+ bij (X(s)) ds (5.2.2)
2 ∂xi ∂xj
i, j=1
d n
X ∂u(t − s, X(s)) X
+ σij (X(s))dBj (s).
∂xi
i=1 j=1
Since we know the stochastic integral is a local martingale, we can force u(t − s, X(s)) to be a local
martingale by requiring that u satisfy the PDE
∂u
= L u, (5.2.3)
∂t
where
d d
X ∂ 1 X ∂2
L = µi (x) + bij (x) .
∂xi 2 ∂xi ∂xj
i=1 i, j=1
So, if u is bounded on [0, t] × Rd , then the left-hand side is bounded, so that the stochastic integral
∂u
is a martingale. Similarly if ∂x i
is bounded over [0, t] × Rd for 1 6 i 6 d, the growth condition on
σ and the fact that X is square-integrable ensure that the right hand side is a martingale. If the
stochastic integral is martingale, then
so that
u(t, x) = Ex [f (X(t))]
We conclude that if u ∈ C 1, 2 satisfies
∂u
= Lu
∂t (5.2.4)
s.t. u(0, x) = f (x)
∂u
and if either u is bounded over [0, t] ∈ Rd or ∂x i
is bounded over [0, t] × Rd for 1 6 i 6 d, then
u(t, x) = Ex [f (X(t))] (= u∗ (t, x)). In other words, if u satisfies the PDE (5.2.4) and is either itself
bounded or has bounded (spatial) derivatives, then u must be the expectation of interest.
and provide sufficient conditions under which the solution to the PDE is the required expectation.
2
§ SECTION 5: THE CONNECTION BETWEEN SDE’S AND PDE’S
Let X satisfy (5.2.1), where µ and σ satisfy the usual growth and Lipschitz conditions.
In order that (5.3.1) be representable as a stochastic integral (and hence a local martingale), we
require that u satisfy the PDE
αu − L u = f. (5.3.2)
If u satisfies (5.3.2), then
Z t
e−αs f (X(s))ds + e−αt u(X(t))
0
Z t d n (5.3.3)
−αs
X ∂u(X(s)) X
= e σij (X(s))dBj (s).
0 ∂xi
i=1 j=1
Note that if f is bounded by kf k, then |u∗ (x)| 6 kf k/α. Hence, if f is bounded, we should seek
bounded solutions u to (5.3.2). If u and f are both bounded, then the left-hand-side of (5.3.3) is a
martingale. Hence Z t
Ex e−αu f (X(u))du + e−αt u(X(t)) = u(x).
0
3
§ SECTION 5: THE CONNECTION BETWEEN SDE’S AND PDE’S
Z t Z ∞
−αs −αt
Ex e f (X(s))ds → Ex e f (X(t))dt
0 0
αu − L u = f,
Suppose X = (X(t) : t ≥ 0) satisfies (5.2.1), where µ and σ satisfy the usual growth and
Lipschitz conditions. For f : Rd → R+ and C c ⊆ Rd , our goal is to find the PDE for
Z τ
∗
u (x) = Ex f (X(s))ds ,
0
E [Y |X(u) : 0 6 u 6 t]
and hence
Z τ ∧t
f (X(s))ds + u(X(τ ∧ t)) − u(X(0))
0
Z τ ∧t Z d
τ ∧t X m (5.4.2)
∂u X
= [(L u + f )(X(s))]ds + (X(s)) σij (X(s))dBj (s).
0 0 ∂xi
i=1 j=1
Since f and u are bounded with Ex τ < ∞, the Bounded Convergence Theorem yields the identity
Z τ
Ex f (X(s))ds + Ex u(X(τ )) = u(x)
0
L u = −f (5.4.4)
is Px -martingale for each x ∈ C. This is needed, for example, to prove that (5.4.1) is a martingale
(rather than just a local martingale). We now offer a method, based on construction of a suitable
local supermartingale, for bounding Z τ
Ex f (X(s))ds
0
when f is non-negative. Of course, such bounds are of significant interest in their own right.
5
§ SECTION 5: THE CONNECTION BETWEEN SDE’S AND PDE’S
We start from (5.4.2), which holds for u ∈ C 2 . Suppose now that rather than requiring L u = −f
on C , we demand only that u be a non-negative C 2 function satisfying the differential inequality
L u 6 −f on C. In this case,
Z τ ∧t
f (X(s))ds + u(X(τ ∧ t))
0
Then,
Z τ ∧Tn ∧t
Ex f (X(s))ds + Ex u(X(τ ∧ Tn ∧ t)) 6 u(x)
0
Letting first n → ∞ and then t → ∞, we conclude that from the Monotone Convergence Theorem
that Z τ
Ex f (X(s))ds ≤ u(x) (5.5.1)
0
One often finds such Lyapunov functions through an educated guess of some kind. Lyapunov
functions that are frequently employed as guesses are: kxkp , exp(kxkp ) and log(1 + kxkp ) (p > 2).
But, in complex mathematical settings, one may need to be much more creative in finding a suitable
choice of Lyapunov function.
Remark 5.5.1 To prove that (5.4.3) is a martingale and to pass limits through the expectation
Ex u(X(τ ∧ t)), we need to also show that (u(X(τ ∧ t) : t > 0)) is suitably integrable. In particular,
if τ < ∞ Px a.s. and
sup Ex u(X(τ ∧ t))p < ∞ (5.5.2)
t>0
for p > 1, this suffices to ensure that u(X(τ ∧ t)) is integrable (so that (5.4.3) is a martingale) and
that
Ex u(X(τ ∧ t)) → 0 = Ex u(X(τ ))
as t → ∞. The method of supermartingales/Lyapunov functions also works in this setting.
6
§ SECTION 5: THE CONNECTION BETWEEN SDE’S AND PDE’S
Specifically, suppose that one has solved (5.4.4) but the solution u is unbounded. To verify
(5.5.2), we attempt to find a nonnegative Lyapunov function v ∈ C 1,2 such that v(0, x) > u(x)p
and
∂v
(L v)(t, x) 6 (t, x)
∂t
for t > 0 and x ∈ C. If (Tn : n > 1) is the appropriately chosen localizing sequence, we conclude
that for 0 6 s 6 t,
and hence
Ex u(X(t ∧ Tn ∧ τ ))p
= Ex u(X(t ∧ Tn ))p I(τ > t ∧ Tn )
6 Ex v(0, X(t ∧ Tn ))I(τ > t ∧ Tn )
6 Ex v(t − (t ∧ Tn ∧ τ ), X(t ∧ Tn ∧ τ ))
6 v(t, x)
for t > 0 and x ∈ C, from which (5.5.2) can now potentially be verified.
where µ and σ satisfy the Lipschitz and growth conditions of Section 4. For C c ⊆ Rd , let T = {t >
0 : Y (t) ∈ C c } be the exit time from C. Set
X(t) = Y (T ∧ t)
We start by deriving the backwards equation for X. Let f be a bounded function, and set
u∗ (t, x) = Ex f ((X(t)) for x ∈ C. As in Subsection 5.2, let u ∈ C 1, 2 be bounded and satisfy
∂u
(t, x) = (L u)(t, x) (5.6.1)
∂t
for t > 0 and x ∈ C (the closure of C), subject to
where µ(·) and σ(·) satisfy the regularity hypothesis described in Section 4 that guarantee exis-
tence/uniqueness, and L̃, Ũ are the minimal nondecreasing processes such that
We start by deriving the backwards equation for X. Let f be bounded and set u∗ (t, x) =
Ex f (X(t)). Recall that (u∗ (t − s, X(s)) : 0 6 s 6 t) is then a martingale adapted to (X(s) : 0 6
s 6 t). For u ∈ C 1, 2 , observe that
d(u(t − s, X(s)))
∂u
= − (t − s, X(s)) + (L u)(t − s, X(s)) ds
∂t
∂u ∂u ∂u
+ (t − s, X(s))dL̃(s) − (t − s, X(s))dŨ (s) + (t − s, X(s) σ(X(s))dB(s)
∂x ∂x ∂x
In order that u(t − s, X(s)) be a martingale, we require that
ut = L u (5.7.3)
for α > 0.
u0 (a) = −1 (5.7.7)
0
u (b) = 0 (5.7.8)
We conclude that if u ∈ C 2 is a bounded solution to (5.7.6) subject to (5.7.7) and (5.7.8), then
u = u∗ .
We have seen that the rate matrix Q plays a key role in the analysis of Markov jump processes,
as does the differential operator L in the setting of diffusion processes. In fact, the equations that
arise in computing, probabilities and expectations for the two classes of processes are essentially
identical, provided that one substitutes the linear differential operator L in the diffusion setting
for the (linear) operator/matrix Q that occurs in the jump process context.
The basic reason, of course, is that both Q and L describe the short-time (infinitesimal)
dynamics of the corresponding Markov processes. Consider first a jump process X = (X(t) : t ≥ 0)
taking values in a finite state space S having rate matrix Q. For any function f : S → R,
Note that the limit (5.8.2) is guaranteed to exist for all functions f . The rate matrix Q is often called
the generator of X, because Q “generates” the transition probabilities (P (t, x, y) : x, y ∈ S, t > 0)
via the Kolmogorov backwards and forwards differential equations, which in turn uniquely specify
the finite-dimensional distributions of X.
In the diffusion setting, we can proceed similarly. Roughly speaking, we define the generator to
be the operator A defined through the limit
Ex f (X(t)) − f (x)
lim = (A f )(x) (5.8.3)
t&0 t
Note that if the limit (5.8.3) exists (for all x ∈ Rd ) for the functions f and g, the limit also exists
for f + g and A (f + g) = A f + A g, so that A is a linear operator. The domain of A , denoted
D(A ), is clearly a vector space.
A key distinction between diffusions and jump processes is that not all functions f lie in D(A ),
when X is a diffusion.
Example 5.8.1 Let X = B, and put f (x) = I(x 6 0). Then, E0 f (B(t)) = P0 (B(t) 6 0) = 21 ,
where f (0) = 1. In this case, E0 f (B(t)) is not even continuous in t as t → 0, let alone differentiable
(as required by (5.8.3)).
Hence, the set of functions f for which A f can be defined through the limit (5.8.3) (i.e. the
“domain of A ”) is a proper subspace of the space of real-valued functions defined on Rd . The
Itô-Doeblin calculus allows us to now identify sufficient conditions under which a function f lies in
the domain.
In particular, suppose that X satisfies (5.2.1), which µ and σ satisfies the usual Lyapunov and
growth conditions. If f ∈ C 2 , Itô’s lemma asserts that
Z t
f (X(t)) − f (X(0)) − (L f )(X(s))ds (5.8.4)
0
is a Px -local martingale. If f and L f are bounded, then (5.8.4) is, in fact, a martingale. Because
L f is continuous, it is evident that Ex (L f )(X(t)) → (L f )(x) as t → 0, so
Z t
Ex (L f )(X(s))ds = t(L f )(x) + o(t) (5.8.5)
0
as t → 0. Hence, a sufficient condition for f to lie in the domain of A is that f ∈ C 2 with f and
L f bounded, in which case A f = L f .
We turn next to studying the role of the boundary behavior of X in restricting the domain of
A . Suppose, for example, that X is a process on R+ satisfying
where µ and σ satisfy the conditions of Lions and Snitzman (1984) that ensure existence of a unique
solution to (5.8.6), and L is the minimal nondecreasing process such that I(X(t) > 0)dL(t) = 0.
11
§ SECTION 5: THE CONNECTION BETWEEN SDE’S AND PDE’S
is a Px -local martingale. If f and L f are bounded and f 0 (0) = 0, then an argument identical to
that just given for diffusions without boundaries establishes that for each x > 0,
Ex f (X(t)) − f (x)
lim = (L f )(x).
t&0 t
We conclude that for such reflecting diffusions, the domain includes all functions f ∈ C 2 for which
f 0 (0) = 0 and f , L f are bounded, in which case A f = L f .
Remark 5.8.1 When modeling a jump process, the typical starting point is the specification of
the rate matrix Q. The first mathematical question to be answered is whether there exists a unique
jump process with the given rate matrix Q. For finite-state jump processes, this question is easily
answered.
The modern approach to diffusions is to model the process of interest by specifying a SDE.
One then studies the question of whether there exists a unique solution to the SDE. But another
approach (and one that was extensively followed in the early days of diffusions) is to model the
process by specifying the generator and its domain. One then needs to establish that there exists
a family of transition probabilities (P (t, x, ·) : t > 0, x ∈ Rd ) such that
Z
P (t + s, x, dy) = P (t, x, dz)P (s, z, dy) (5.8.7)
Rd
Equation (5.8.7) is known as the semigroup property, and is essential to proving that the stochastic
process that is constructed from the generator has finite-dimensional distributions that are consis-
tent with the Markov property. Given a generator and its domain, the key question is therefore
showing that there exists a unique semigroup that is generated by that generator and associated
domain. This is a question that can be answered by appealing to a functional analysis result due
to Hille and Yosida. Because this methodology takes advantage of Banach space ideas, one needs
to strengthen the definition (5.8.3) so that it exists as a Banach space limit. For example, if the
underlying Banach space is chosen to be the space of continuous functions equipped with supremum
norm, one would require that A f be a continuous function for which
Ex f (X(t)) − f (x)
sup − (A f )(x) → 0
x t
as t & 0. When this (strong) definition is followed, it leads to a definition of the generator under
which many functions f fail to be in the domain. On the other hand, it offers the ability to apply
function analytic tools to constructing diffusion processes.
Remark 5.8.2 One classical means of enlarging the class of functions lying in the domain of the
generator is to consider instead the characteristic operator A. The characteristic operator is defined
by letting Tε be the first time that kX(t) − X(0)k exceeds ε, and defining Af via
Ex f (X(Tε )) − f (x)
lim = (Af )(x) (5.8.8)
ε&0 Ex τε
12
§ SECTION 5: THE CONNECTION BETWEEN SDE’S AND PDE’S
the domain of a is hte class of functions f for which the limit (5.8.8) exists for all x ∈ Rd . Note
that if f is C 2 , it follows that (5.8.4) is a local martingale that is bounded for t 6 Tε , proving
the existence of the limit (5.8.8) for such functions, with Af = L f . Hence, we no longer need to
require that f and L f be (globally) bounded, in order that f lie in the domain of A.
The modern definition of the generator of a Markov process proceeds as follows (and leads to a
domain that is very large). Given a Markov process X = (X(t) : t ≥ 0), we say that f lies in the
domain of the generator A (and write f ∈ D(A )) if there exists a function g such that
Z t
f (X(t)) − g(X(s))ds (5.8.9)
0
This leads to a very large domain D(A ). In addition to containing all C 2 functions f , this
definition of the domain may include even non-smooth functions. Suppose, for example, that X
satisfies (5.2.1) and is a positive recurrent diffusion. Then, there typically exists a stationary
distribution π such that Z
Ex h(X(t)) → h(x)π(dx)
S
as t → ∞, for all functions h for which the right-hand side is integrable. Set
Z
g(x) = h(x) − h(y)π(dy).
S
If we assume that Z ∞
| Ex g(X(t))|dt < ∞
0
for all x ∈ Rd (as would generally occur when the rate of convergence to equilibrium is exponentially
fast), then Z ∞
f (x) = − Ex g(X(t))dt
0
is a well-defined (finite-valued) function. Furthermore, it is easily verified that (5.8.9) is then a
Px -martingale for each x, so f ∈ D(A ). Note that g need not be smooth, and there is no guarantee
that f is smooth. So D(A ) may include functions that lie (for example) outside C 2 . Because this
definition leads to a very large domain, it has become increasingly popular in the literature.
Again, the boundary behavior of X plays a role in defining the domain D(A ). Consider, for
example, a diffusion that is subject to absorption at the exit time from C. Such a diffusion can be
represented as X(t) = Y (t ∧ T ), where T is the exit time from C. Note that for t > T ,
Z t
f (X(t)) − g(X(s))ds
0
Z T
=f (Y (T )) − g(Y (s))ds − g(Y (T ))(t − T )
0
Hence, in order that this process be a local martingale, we must require that g = 0 on ∂C (or
equivalently, A f = 0 on ∂C). Hence if the boundary is accessible (i.e. Px (T < ∞) > 0), f ∈ D(A )
requires that A f = 0 on ∂C.
Exercise 5.8.1 Suppose that X satisfies (5.8.6). Show that f ∈ D(A ) if f ∈ C 2 and f 0 (0) = 0.
13
§ SECTION 5: THE CONNECTION BETWEEN SDE’S AND PDE’S
Z t Z t N (t)
X
X(t) = X(0) + µ(X(s))ds + σ(X(s))dB(s) + ξi ,
0 0 i=1
We assume that the ξi ’s are iid positive rvs. In other words, X has state space [0, ∞), and X
jumps into (0, ∞) according to the distribution P (ξ1 ∈ ·) every time X hits the origin.
To illustrate the computation of probabilities and expectations for this process, consider the
computation of Z ∞
Ex e−αt r(X(t))dt,
0
Note that X evolves continuously between the jumps, so separating out these intervals and applying
Itô’s formula (assuming that u ∈ C 2 ) yields
Z t
−αt
e u(X(t)) = u(X(0)) + e−αs ((L u)(X(s)) − αu(X(s)))ds
0
Z t
+ e−αs u0 (X(s))σ(X(s))dB(s) (5.9.2)
0
N (t)
X
+ (u(X(Ti )) − u(X(Ti −)))e−αTi .
i=1
Indeed,
Finally,
Z t Z t
u(X(t))e−αt −u(X(TN (t) ))e−αTN (t) = e−αs ((L u)(X(s))−αu(X(s)))ds+ e−αs u0 (X(s))σ(X(s))dB(s).
TN (t) TN (t)
Now,
N (t) N (t)
X X
(u(X(Ti )) − u(X(Ti −)))e−αTi = (u(X(Ti )) − u(X(Ti −)) − E(u(X(Ti )) − u(X(Ti −)) | X(s−) : 0 ≤ s ≤ Ti ))e−αTi
i=1 i=1
N (t)
X
+ E(u(X(Ti )) − u(X(Ti −)) | X(s−) : 0 ≤ s ≤ Ti )e−αTi .
i=1
Note that the stochastic integral is a local martingale, and it can be shown that
N (t)
X
(u(X(Ti )) − u(X(Ti −)) − E(u(X(Ti )) − u(X(Ti −)) | X(s−) : 0 ≤ s ≤ Ti ))e−αTi
i=1
is a martingale adapted to (X(t−) : t ≥ 0). Hence, in order that (5.9.1) be a martingale, we require
that u satisfy
(L u)(x) − αu(x) = 0, x ≥ 0
subject to Z ∞
u(y)P (ξ ∈ dy) = u(0).
0
If we can find a bounded solution to this system, then our standard argument establishes that
Z ∞
u(x) = Ex e−αt r(X(t))dt
0
for x ≥ 0.
15