RETROSPECTIVE EXACT SIMULATION 195
The right-hand side is bounded below by (1/2)(
β
2
/8 −r/2forallu ∈ R. Hence
(u)=
1
2
r
2
β
2
−2
r
2
β
2
K
exp(−
β
u)+
r
2
β
2
K
2
exp(−2
β
u)
. (10.24)
The function (u) was designed so that (u) > 0. When u > 0 (u) ≤
(1/2)(r
2
/
β
2
), but that leaves the situation when u < 0 to worry about.
So use the ideas of the last section: find the minimum value of B
t
over [0,T ],then
that gives a maximum value of (B
t
) over [0,T ]. With that maximum value in place,
it once again becomes possible to generate a Bernoulli with mean exp(−I) where
I =
T
0
(B
t
) dt.
Therefore the techniques of the last section can be applied as long as A(u) does
not grow too quickly. So consider A(u)=
x
0
α
(s) ds.Here
A(u)=
β
2
−
r
β
u −
r
β
2
K
exp(−
β
u). (10.25)
Since A(u) grows more slowly than quadratically, it will be possible to normalize the
density of B
T
.
To be specific, for x
0
= −ln(v
0
)/
β
, it is necessary to be able to sample B
T
from
density
exp(A(u) −(u −x
0
)
2
/(2T )) ∝ exp(−(u −g
1
)
2
/(2T ) −g
2
exp(−
β
u)), (10.26)
where g
1
= x
0
+ T (
β
/2 −r/
β
) and g
2
= r/(
β
2
K) are constants. This can be accom-
plished quickly by using AR from a normal with mean g
1
+ T
β
g
2
exp(−
β
g
1
) and
variance T .
10.3.3 Unbounded
To deal with the case where is an unbounded function, it is necessary to find a value
M > 0 such that the Brownian motion stays between M and −M from [0,T ].In[10],
Beskos et al. found such a method by generating a Brownian bridge whose extreme
values were bounded.
An alternate method comes from [22]. Recall from Section 2 .6.1 that for a Brow-
nian motion {B
t
} with T
1
= inf{t : |B
t
| = 1}, it is possible to simulate from T
1
in
finite expected time using AR.
Suppose that the goal is to find the solution to the SDE at time T .IfT
1
> T ,then
|B
t
|≤1forallt ∈ [0 , T ], and the value of can be bounded appropriately using this
fact.
On the other hand, if T
1
< T ,letT
2
= inf{t > T
1
: |B
t
−B
T
1
| = 1}. Then for all
t ∈ [T
1
,T
2
], B
T
1
−1 ≤ B
t
≤B
T
1
+ 1. In g eneral, letting T
i+1
= inf{t > T
i
: |B
t
−B
T
i
|=
1} gives a sequence of passage times where the Brownian motion has moved at most
one away from its previous value. By the symmetry of Brownian motion, B
T
i+1
is
equally likely to be eith er B
T
i
+ 1orB
T
i
−1. See Figure 1 0.1 for an illustration.
Recall that to generate the Brownian motion {X
t
}that is a candidate for a solution