Simulation of Diffusion Processes
Uploaded by
seanwu95Simulation of Diffusion Processes
Uploaded by
seanwu952) if we know the transition density
New question: Converges in what sense? And how fast?
p(t, y|x, s),
ie. y 7 p(t, y|x, s) is the density of X(t) given X(s) = x
Ex: Ornstein-Uhlenbeck/Vasicek (easy; Gaussian) Cox-IngersollRoss (hard; non-central-2)
For simulation purposes, wed especially like 1). That rarely happens.
Strong convergence of order if
E(|X(T ) Y n(T )|) = O((1/n) )
Weak convergence of order wrt. a function g
|E(g(X(T )) g(Y n(T )))| = O((1/n) )
Can we construct approximations, Y n, such that
Y n is close/converges to X
Result: The Euler-scheme converges strongly with order 1/2 and
weakly (wrt. all polynomials) with order 1.
the Y ns can be simulated (note: each Y n is a stochastic process)
Strong schemes have some path-wise convergence properties. Enough
said.
Simulation of Diffusion Processes (Seydel Ch.
3)
Diffusion: Soln of SDE
dX(t) = (X(t), t)dt + (X(t), t)dW, X(0) = x0
where and are functions (of appropriate dimensions; but lets say
everything is 1-d).
Soln is a stochastic process. When can we reasonably say we know
the solution?
1) if we know a deterministic function f such that
Assume and not (calendar) time-dependent, fix some T , an n,
put n = T /n and tn
i = in.
Natural idea: dW s are approximately iid N (0, dt). This leads to the
Euler scheme
n n
n n
n n
Y n(tn
i ) = Y (ti1 ) + (Y (ti1 ))n + (Y (ti1 ))W (ti ),
where (of course) Y n(0) = x0 and W (ti ) = W (ti ) W (ti1 ) (so the
W s are iid N (0, n)).
Notational nightmare when written in full, but easy to simulate. Good
news: The Euler scheme converges as n . (Closely connected to
how SDE solns are defined.)
t : X(t) = f (W (t), t).
Geo. BM: f (x, t) = x0 exp(( 2 /2)t + x)
1
Simulations for Geo. BM:
Now do the same for SDEs: Ito-Taylor-expansion.
Fairly clear that theres both weak and strong convergence.
Itos formula says that for any C 2-function f we have
f (X(t)) = f (x0 ) +
Evident that the strong order is 1/2.
Less evident that the weak order, but with a little suggestive
graphics ...
Whats with the different orders? Different forms of convergence are
different, so ...
Whats with the half-order? Non-integer order is something were
used to for BM; dW dt1/2 .
Z t
0
(L0f )(X(s))ds +
(L1f )(X(s))dW (s),
1
(L0 f )(x) = (x)f 0 (x) + 2 (x)f 00 (x) and (L1f )(x) = (x)f 0 (x).
2
Using this for f (x) = x gives the almost tautology
X(t) = x0 +
Z t
0
(X(s))ds +
Z t
0
(X(s))dW (s),
8
Can we improve the orders? Yes. But we have to study closely the
terms we have thrown away.
0.02
|E(X(T))E(Y(T)))|
0.03
0.04
0.03
|E(X(T))E(Y(T))|
0.03
Thrown away? Recall a Taylor-expansion:
0.01
0.01
0.01
0.02
0.02
0.04
0.04
0.05
0.05
where the operators L0 and L1 are defined trough
|E(X(T))E(Y(T))|
Z t
0.05
0.10
0.15
0.20
0.25
10
15
20
25
30
0.05
#steps, n
0.10
0.15
(t)2
(t)n
+...+f (n)(t0 )
+O((t)n )
2
n!
How to prove this: Fundamental theorem of calculus over & over.
f (t0 +t) = f (t0 )+f 0 (t0 )t+f 00(t0 )
0.25
stepsize, h
0.8
0.6
E(|X(T)Y(T)|)
1.0
0.8
E(|X(T)Y(T)|)
Difeqn-link: Sometimes we can easily express RHS in terms of the
coefficients. Ex: If f 0 = a(t) then f (t0 + t) f (t0 ) + a(t0 )t +
a0(t0 )(t)2 /2
0.4
0.6
0.8
0.6
0.4
E(|X(T)Y(T)|)
1.0
1.0
stepsize, h
0.05
0.10
0.15
stepsize, h
0.20
0.25
10
15
20
#steps, n
25
30
0.05
0.10
#steps, n
0.15
0.25
But this can be calculated explicitly (classical exercise)
Z t
0
W (s)dW (s) =
1
(W (t)2 t)
2
Clear strategy: Apply Ito to the 3 remaining double integrals in the
remainder. If we do that we have to look at terms like
Z tZ s
1
duds = t2
2
0 0
Z Z
Z
This suggests we use the (or: a) Milstein scheme:
n n
n n
n n
Y n(tn
i ) = Y (ti1 ) + (Y (ti1 ))n + (Y (ti1 ))W (ti )
1
0
n n
2
+ (Y n(tn
i1 )) (Y (ti1 ))(W (ti ) n)
2
0 0
Z tZ s
0
dW (u)ds =
dudW (u) =
0
Z t
0
W (s)ds := Z(t) N (0, t3 /3) (small exercise)
sdW (s) = W (t)t Z(t) (Ito right to left)
Result: The Milstein scheme has weak order 1 (wrt. polynomials)
and strong order 1.
10
12
0
1
0
where (for instance) (L ) = . Collecting all the double integrals
Z tZ s
0
dW (u)dW (s) = (x0) 0 (x0)
Z t
0
0.02
0.05
0.10
0.01
stepsize, h
0.05
0.10
|E(X(T)Y(T))|, Milstein
0.020
0.020
stepsize, h
0.005
|E(X(T)Y(T))|,Euler
0.02
0.001
(x0 ) 0(x0)
0.05
0.7
0.5
0.01
in a remainder R, and then throwing that away, we get the Euler
scheme.
We can apply the Ito-trick again to see that the R-term contains
among other things
0.02
(L )(X(u))dW (u) dW (s),
E(|X(T)Y(T)|), Milstein
0.4
(L1)(X(u))dW (u) ds
0.01
(L )(X(u))du +
Z s
Z s
0.010
(L0 )(X(u))du +
0.3
(x0 ) +
Z s
Z s
0.005
(x0) +
E(|X(T)Y(T)|),Euler
Z t
Z t
0.2
X(t) = x0 +
0.10
But now use the result for f = , inside the integrals:
W (s)dW (s)
0.01
0.02
0.05
stepsize, h
0.10
0.01
0.02
0.05
0.10
stepsize, h
11
Remarks
You can do this till the cows come home. Is there a pattern
(like for regular Taylor-expansion)? Not a simple one: The operators L0 and L1 dont commute and we get n-fold recursive
BM-integrals.
In theory you only have to hard-code a particular scheme once.
Then you just specify and functions and some of their derivatives. Of course in practice ...
Higher dimensions: Everything but the Euler-scheme gets very
hard or messy.
14
Can show that (W (t), Z(t)) is 2-d Gaussian and that cov(W (t), Z(t)) =
1 t2 . In simulations (t t) this can be achieved by using
2
1
1
(t)3/2 (1 + 2 ),
2
3
where s independent standard normal.
W =
t1 and Z =
In this way you get Seydels (3.13). This scheme has weak order 2
and (still only) strong order 1.
Variance reduction along the path in a scheme does not make
sense. But using a W -path and its negative is OK; the same
is simulation a known SDE-solution based on the W -path.
Computational efficiency trade-off in MC-pricing: Duffie & Glynn
shows that for a weak order scheme it is computationally optimal
to have
#paths # discretization steps2
Moral: For (say) an order-2 scheme, the discretization error is not
our main concern.
16
Prefer strong schemes if something must be calculated along the
path.
Despite all our calculations we have at best given an indication a
local truncation error order. Real convergence (order) proofs are
left to the experts (Kloeden & Platen, Talay).
1
00
0 2
2
(But if you add 1
2 b(bb + (b ) ))( 3 W )W it gets strong order
1.5.)
13
15
