Exact simulation

Etienne Tanr
INRIA - Team Tosca
2011-2012
Etienne Tanr (INRIA - Team Tosca) Exact simulation 2011-2012 1 / 16
Outline
1
The Beskos and Roberts Exact Simulation Method
Etienne Tanr (INRIA - Team Tosca) Exact simulation 2011-2012 2 / 16
Calibration of Monte Carlo methods
Let X be a stochastic process. To test the quality of a Monte Carlo and/or
discretization method, one often compare Ef (X

) with its numerical approximation

in cases where the law of the process X is explicitly known. However,
S.D.E.s which have explicit solutions have often very particular structures
which may hide the diculties presented by realistic models.
Most of them are one dimensional.
Most often, one only knows the law of the random variable X
T
, but the joint
distribution of vectors of the type (X
T
1
, . . . , X
T
n
) cannot be known explicitly.
Is it possible to calibrate Monte Carlo methods by using another approach?
We present an interesting paper by Beskos and Roberts who propose an exact
simulation method of the law of the process X in one dimensional models.
Etienne Tanr (INRIA - Team Tosca) Exact simulation 2011-2012 3 / 16
Rejection methods for real random variables
To simulate random vectors with density f , one can use rejection methods:
choose a density g satisfying:
Random variables with density g are easy to simulate(e.g., g is Gaussian).
C > 0 such that h(x) :=
f (x)
Cg(x)
1 for all x.
The rejection method is as follows:
Sample Y with density g,
Sample U with uniform law on (0, 1) and independent of Y,
If U h(Y) =
f (Y)
Cg(Y)
then accept the sample of Y,
Otherwise resample Y, etc.
Notice that one rejects a random number of (independent) samples before
accepting.
Etienne Tanr (INRIA - Team Tosca) Exact simulation 2011-2012 4 / 16
Let M() denote the number of iterations of the procedure before accepting.
Then the random variable Y
M()
() has density f .
Indeed, notice that, since
_
f (z)dz = 1,
P
_
U <
f (Y)
Cg(Y)
_
=
_
f (z)
Cg(z)
g(z)dz =
1
C
.
We then have
P(Y
M
A) =

k1
P(Y
k
A, M = k)
=

k1
P(Y
k
A, U
k

f (Y
k
)
Cg(Y
k
)
)
k1
=1
P(U

>
f (Y

)
Cg(Y

)
)
=

k1
P(Y
k
A, U
k

f (Y
k
)
Cg(Y
k
)
)(1
1
C
)
k1
=
_
A
f (z)
Cg(z)
g(z)dz

k1
(1
1
C
)
k1
=
_
A
f (z)dz.
Etienne Tanr (INRIA - Team Tosca) Exact simulation 2011-2012 5 / 16
A generalization
Proposition.
Let (S, S) be a measurable space. Let and be probability measures on this
space such that is absolutely continuous w.r.t. . One desires to sample .
Assume
> 0, h(x) :=
d
d
(x) 1 a.s.
Let (Y
n
, I
n
) be a sequence of i.i.d. random variables taking values in S (0, 1).
The Y
i
s have law and
P(I
1
= 1 | Y
1
) = h(Y) a.s.
Let
M := min{k 1; I
k
= 1}.
Then the law of Y
M
is .
Etienne Tanr (INRIA - Team Tosca) Exact simulation 2011-2012 6 / 16
A generalization
Remark:
In the above elementary situation, we had:
(dx) = f (x)dx,
(dx) = g(x)dx,
P(U
f (Y)
Cg(Y)
) =
1
C
d
d
(Y).
Etienne Tanr (INRIA - Team Tosca) Exact simulation 2011-2012 7 / 16
The ingredients of the exact simulation method for a one
dimensional diusion
Consider
X
t
= x
0
+
_
t
0
(X
s
)ds + W
t
.
We aim to simulate (X
t
1
, . . . , X
t
n
) exactly.
Fix a density h
0
. Sample this density and get a value (). Set b := () and
a = x
0
. Now, choose a discretization step T/n and simulate the Brownian bridge
W
a,T,b
at times T/n, . . . , T T/n by using its time marginal densities (see the
above reminders).
Now, given the density h
0
, dene a new probability measure Z on the space of
continuous functions by
dZ
dP
W
() :=

2Th
0
(
T
) exp
_

2
T
2T
_
.
Etienne Tanr (INRIA - Team Tosca) Exact simulation 2011-2012 8 / 16
For all bounded functional on the space of continuous functions we have
_
()dZ =
_
()
dZ
dP
W
()dP
W
()
=

2T
_
()
h
0
(
T
)
exp
_

2
T
2T
_dP
W
()
=

2TE
_
(W

)h
0
(W
T
) exp
_
W
2
T
2T
__
=
_
E((W
0,T,b

) h
0
(b) db.
Therefore, the law Z can be desintegrated owing to the density h
0
and the laws of
Brownian bridges.
Etienne Tanr (INRIA - Team Tosca) Exact simulation 2011-2012 9 / 16
We also have (remember the reminders on Girsanovs theorem)
dP
X
dZ
() =
dP
X
dP
W
()
dP
W
dZ
()
=
C
h
0
(
T
)
exp
_
A(
T
)

2
T
2T

1
2
_
T
0
(
2
(
s
) +

(
s
))ds
_
,
where the normalizing constant C depends on A(0), etc.
Choose
h
0
(x) :=
exp(A(x)
x
2
2T
)
_
exp(A(z)
z
2
2T
)dz
.
Then, for some function ,
dP
X
dZ
() = exp(
_
T
0
(
s
)ds).
Etienne Tanr (INRIA - Team Tosca) Exact simulation 2011-2012 10 / 16
Gather all the pieces of the puzzle:
Given a functional F on the space of continuous functions, one has
EF(X

) =
_
F()dP
X
()
=
_
F()
dP
X
dZ
()dZ()
=
_
F() exp(
_
T
0
(
s
)ds)dZ()
=
_
E
_
F(W
0,T,b

) exp(
_
T
0
(W
0,T,b
s
)ds)
_
h
0
(b) db.
Etienne Tanr (INRIA - Team Tosca) Exact simulation 2011-2012 11 / 16
An approximate version of the exact simulation algorithm
A theoretical rejection method:
Consider the above general rejection method, and use the identication P
X
,
Z. Remember
dP
X
dZ
() = exp(
_
T
0
(
s
)ds).
To simplify the notation, suppose that exp(
_
T
0
(
s
)ds) < 1 a.s.
Sample the density h
0
(and thus get a random value b), and then the
approximate trajectory of the corresponding Brownian bridge. You have got a
sample of a process Y whose law is Z.
Compute an approximation of h(Y) := exp(
_
T
0
(W
0,T,b
s
)ds).
Sample the uniform law on [0, 1]. If U h(Y), then accept the trajectory.
Otherwise resample h
0
, etc.
Etienne Tanr (INRIA - Team Tosca) Exact simulation 2011-2012 12 / 16
Remark
To sample the density h
0
, one may need to use a rejection method.
To sample Y at a nite set of times iT/n, one can use the above description
of the joint law of (W
0,T,b
T/n
, . . . , W
0,T,b
TT/n
).
The law of this simulation is (approximately) equal to the law of
(X
0
, X
T/n
, . . . , X
T
).
Beskos and Roberts improve the accuracy of the algorithm owing to a procedure
which allows one to avoid the eects of the time discretization: this procedure
uses a random partition of (0, T) and provides exact skeletons of X at any given
nite collection of times 0 < t
1
< . . . < t
n
< T.
The method can be extended to produce exact simulations of the running
maximum of X, or of its hitting times, etc.
Etienne Tanr (INRIA - Team Tosca) Exact simulation 2011-2012 13 / 16
A rst summary and extension
Algorithm
Simulate a random variable with p.d.f h
0
Simulate a Brownian bridge (Y
t
, 0 t T) such that Y
T
=
You accept this trajectory with probability exp
_

1
2
_
T
0

(Y
t
) +
2
(Y
t
)dt
_
Etienne Tanr (INRIA - Team Tosca) Exact simulation 2011-2012 14 / 16
Point Poisson Process (P.P.P.)
Let us consider a set D (e.g. [0, T] [0, K]). A realisation of a P.P.P. on D with
intensity I(t, x) is a set of points of D. For all subset F of D, we denote by N
F
the number of points of the P.P.P. which are in F.
Caracterization of a Point Poisson Process
for all F D, N
F
is a random variable (with value in N) with Poisson law of
parameter
_
F
I(t, x)dtdx,
for all subset F and G with empty intersection, N
F
and N
G
are independant
random variables.
A particular case is: the intensity I is egal to 1. For all subset F of D, the number
N
F
of points of the P.P.P. in F is a Poisson random variable with parameter equal
to the volume of F.
Etienne Tanr (INRIA - Team Tosca) Exact simulation 2011-2012 15 / 16
Point Poisson Process
Recall that a Poisson random variable X of parameter has the following law:
P(X = k) = exp()

k
k!
.
In particular, P(X = 0) = exp().
Let us consider a process (X
t
; 0 t T) and a non negative function ,
bounded from above by K. In order to simulate an event with probability
exp(
_
T
0
(X
s
)ds), we can use Point Poisson Processes. Indeed, the probability
that the P.P.P. on [0, T] [0, K] have no point below the curve t (X
t
) is
precisely exp(
_
T
0
(X
s
)ds).
Etienne Tanr (INRIA - Team Tosca) Exact simulation 2011-2012 16 / 16