RETROSPECTIVE EXACT SIMULATION 189
Now Radon-Nikodym derivatives behave m uch like regular derivatives, including
obeying the chain and inverse rules. So
dQ
dL {R
t
}
({R
t
})=
dQ
dW
x
({R
t
})
/
dL {R
t
}
dW
x
({R
t
})
∝
exp(A(R
T
) −(1/2)
T
0
[
α
(R
t
)+
α
2
(R
t
)] dt )
exp(A(R
T
))
= exp
−
T
0
(
α
(R
t
)+
α
2
(R
t
))/2 dt
.
At this point, to make acceptance/rejection work, assume that the function (
α
+
α
2
)/2 is bounded below by k
α
. Then setting
(u)=
α
(u)+
α
2
(u)
2
−k
α
(10.14)
makes (u) ≥ 0, and since exp(−k
α
T ) is a constant,
dQ
dL {R
t
}
({R
t
}) ∝ exp
−
T
0
(R
t
) dt
≤ 1, with probability 1 under L {R
t
}.
(10.15)
It is necessary here that A(u) be a function such that exp(A(u) −(u −x)
2
/(2T ))
is a normalizable density. This is equivalent to saying that there is a constant a
1
such
that A(u)=a
1
u
2
for sufficiently large or small u.Whena
1
is large, T might need to
be very small for the density to be normalizable.
When the density is nor malizable, if
α
2
+
α
is bounded below, it is possible in
theory to use an AR approach to generate draws from Q using draws of {R
t
}.
Beskos et al. [11] introduced a Poisson point process procedure for accom-
plishing this. Let A = {t,y : t ∈ [0,T ], y ∈ [0,(B
t
)]}. Recall that if P is a Poisson
point process over B where A ⊆ B,thenP(#(P ∩A)=0)=exp(−
ν
(A)).Using
ν
(A)=
T
0
(R
t
) dt then gives the desired result.
So how to determine if #(P ∩A)=0? First consider the case where the function
is bounded above by M.ThenA ⊆ [0,T ] ×[0,M], and it is possible to generate a
PPP P of constant intensity 1 over [0,T ] ×[0,M]. Consider a point (t
i
,y
i
) ∈ P.When
does this point fall into A?
Since B
t
and are continuous maps, this point falls into A if and only if y
i
≤(R
t
i
).
So to determine if any point of P falls into A, it is only necessary to evaluate R
t
at the
finite set of time values {t
1
,t
2
,...,t
#P
}.
Alternately, given the time value t
i
of the points of P, the chance that y
i
> R
i
is
max{0,(M −R
t
i
)/M}. So it is not necessary to draw y
i
explicitly.
When the PPP on [0,T ] ×[0,M] is projected onto the interval [0,T ] by removing
the second coordinate, the result is a one-dimensional PPP of rate T . It is possible
to generate such a PPP by adding exponential random variables of rate M to 0 until
the result is greater than T . An exponential random variable of rate M has the same
distribution as (1/M)ln(1/U),whereU is uniform over [0,1].
The authors referred to this method as retrospective sampling, since once the