186 STOCHASTIC DIFFERENTIAL EQUATIONS
Definition 10.3. Given p
∗
∈(0,1] and a function f : [0, p
∗
] →[0,1],aBernoulli fac-
tory is a computable function A that takes as input a number u ∈ [0,1] together with
a sequence of values in {0,1}, and returns an output in {0,1} where the following
holds. For any p ∈ [0, p
∗
],X
1
,X
2
,... iid Bern(p), and U ∼ Unif([0, 1]),letTbethe
infimum of times t such that the value of A (U, X
1
,X
2
,...) only depends on the values
of X
1
,...,X
t
.Then
1. T is a stopping time with respect to the natural filtration and P(T < ∞)=1.
2. A (U,X
1
,X
2
,...) ∼ Bern( f (p)).
Call T the running time of the Bernoulli fa ctory.
Densities for SDE’s employ Girsanov’s Theorem [42], which gives the density
of the form exp(−p) for a specific p. Therefore, what is needed for these types of
problems is an exponential Bernoulli factory where Y ∼ Bern(exp(−p)).Beskoset
al. [11] gave the following elegant way of constructing such a Bernoulli factory.
Recall how thinning works (discussed in Sections 2.2 and 7.2.) Take P, a Poisson
point process of rate 1 on [0 , 1], and for each point in P flip a coin of probability p.
Then let P
be the points of P that received heads on their coin. P
will be a Poisson
point process of rate p. So the number of points in P
is distributed as Poisson with
mean p,whichmakesP(#P
= 0)=exp(−p).
More generally, suppose that A ⊂ B are sets in R
n
of finite Lebesgue measure.
Let P be a Poisson point process of rate 1 on B.ThenP ∩A is a Poisson point process
of rate 1 on A, and the chance that there are no points in A is exp(−m(A)),where
m(A) is the Lebesgue measure of A.
With this in mind, we are now ready to perfectly simulate some SDE’s.
10.3 Retrospective exact simulation
Retrospective exact simulation is a form of acceptance/rejection introduced by
Beskos et. al [11] to handle SDE’s. Start with Equation (10.1).
The SDE has a unique solution, that is, there exists a unique distribution of {X
t
}
for t ∈ [0, T ], given that the functions a and b satisfy some regularity conditions.
These conditions ensure that the SDE does not explode to infinity in fin ite time. See
Chapter 4 of Kloeden and Platen [83] for details.
Our original SDE had the form dV
t
= a(v) dt + b(v) dB
t
.Thefirst step to note
is that for many a and b it is possible to restrict our attention to SDE’s where b(v) is
identically 1.
To see th is, consider the differential form of It¯o’s Lemma (see for instance [110]).
Lemma 10.2 (It¯o Lemma (1st form)). Let f be a twice differentiable function and
dV
t
= a(v) dt + b(v) dB
t
. Then
df(V
t
)=[a(v) f
(v)+b(v)
2
f
(v)/2] dt + b(v)f
(v) dB
t
. (10.4)
Now let
η
(v)=
v
z
b(u)
−1
du for any z in the state space Ω.Then
η
(v)=b(v)
−1
and
η
(v)=−b
(v)/b(v)
2
. Hence for X
t
=
η
(V
t
),
dX
t
=[a(v)/b(v)−b
(v)/2] dt + dB
t
. (10.5)