Professional Documents
Culture Documents
2
p(w, t )
w
2
Formal solution to LE is called an Ornstein-Uhlenbeck process
v(t ) = v
0
e
t
+e
t
_
t
0
e
s
dw(s)
Intro to SDEs with with Examples
Stochastic Differential Equations
Brownian Motion
A Simple Stochastic Differential Equation
What does dw(t ) mean?
w(t ) = w
1
+w
2
+ +w
n
each increment is independent, and
E{w
i
w
j
} =
ij
t
or innitesimal version
Edw(t ) = 0
E{dw(t ) dw(s)} = (t s) dt ds
Intro to SDEs with with Examples
Stochastic Differential Equations
Brownian Motion
The Langevin Equation
Solution to LE has properties
Ev(t ) = v
0
e
t
+e
t
_
t
0
e
s
Edw(s)
= v
0
e
t
and
E(v(t ))
2
= (v
0
)
2
e
2t
+
2
e
2t
e
2t
1
2
2
2
as t
Something familiar about this?
m
2
E(v)
2
=
m
2
2
2
=
1
2
kT
Intro to SDEs with with Examples
Stochastic Differential Equations
It Calculus
It Calculus
It calculus for multi-dimensional version
dw(t )
2
dt or dw
i
(t )dw
j
(t )
ij
dt
In non-isotropic case, system
dz = b(z) dt +(z) dw(t ) (SDE)
is shorthand for
z(t ) = z
0
+
_
t
0
b(z
s
) ds +
_
t
0
(z
s
) dw
s
.
It rule for Stochastic integral:
E{
_
t
0
(z
s
)dw
s
} = 0,
and
E{
_
t
0
(z
s
)dw
s
}
2
=
_
t
0
T
(z
s
)ds.
These integrals are martingales.
Intro to SDEs with with Examples
Stochastic Differential Equations
It Calculus
A Standing Martingale
Figure: A standing martingale
Intro to SDEs with with Examples
Stochastic Differential Equations
Numerical Solution of SDEs
Numerical Solution of SDEs
Simulation? First,
Ef (z(t ))
1
N
N
i =1
f (z
[i ]
(t ))
for sample of N paths z(t ). Paths {z
[1]
, z
[2]
, ..., z
[N]
} integrated by
some rule, e.g. Euler Two criteria two versions of solution
z(t ), z(t )
are equivalent (
z(T)) Ef (z(T))| = 0
Intro to SDEs with with Examples
Stochastic Differential Equations
Types of Solutions to SDEs
Weak Solutions
Example: weak simulation (m 0):
dx = x|x|
m1
dt +dw(t )
has solution whose distribution law satises Kolmogorov equation
p(x, t )
t
=
x
_
1
2
x
+x|x|
m1
_
p(x, t ) 0
when t . That is, x(t ) becomes stationary. p(x, t ), properly
normalized, is
p(x, ) = N
m
e
2
m+1
|x|
m+1
.
Two examples
p(x, ) = e
2|x|
for m = 0
p(x, ) =
1
e
|x|
2
for m = 1
Intro to SDEs with with Examples
Stochastic Differential Equations
Types of Solutions to SDEs
Strong Solutions
Example: a strong test,
dx = xdt +xdw(t )
having formal solution
x(t ) = x
0
exp(( +
2
2
)t +w(t )). (1)
Notice x(t ) 0 as t . Many authors (Mitsui et al, Higham, ...)
have studied stability regions, , , for asymptotic stability x(t
n
) 0,
when
t
n
= h
1
+h
2
+. . . +h
n
may have varying stepsizes. Cases
t = T
1
= n h,
and h h/2
m
= h
,
t = T
m
= n2
m
h
) =
1
w(T
m
+h
+ h
) =
1
+
2
. . .
w(T
m
+h) =
m
k=1
k
w(T
1
+h) = w(T
m
+h)
Here, one follows the pathwise convergence as m is changed. See
Kloeden and Platen, chapt. 9, p. 309. One compares "exact" solution,
equation (1), with simulation values at points T
1
= T
m
.
Intro to SDEs with with Examples
Stochastic Differential Equations
Types of Solutions to SDEs
Strong Solutions
Numerical criteria similar: discrete times t
k
= kh, h = step size,
T = Mh, and
z
k
= numerical approx.,
strong order :
(E max
0kM
|z
k
z(t
k
)|
2
)
1/2
K
1
h
(z
k
)(W
2
h)
is strong order = 1, weak order 1
Intro to SDEs with with Examples
Stochastic Differential Equations
Higher-Order Methods
Higher-Order Methods
Higher order weak methods require modeling
I
ij
=
_
h
0
w
i
dw
j
I
i 0
=
_
h
0
w
i
(s)ds
I
ijk
=
_
h
0
w
i
w
j
dw
k
I
ii 0
=
_
h
0
w
2
i
ds
For example, for Runge-Kutta type methods
I
ij
1
2
j
+
h
2
ij
,
I
i 0
h
2
i
,
I
ijk
h
2
ij
k
I
ii 0
h
2
2
i
ij
is a model for
_
w
i
dw
j
w
j
dw
i
.
Intro to SDEs with with Examples
Stochastic Differential Equations
Higher-Order Methods
Examples
w = is approximately gaussian
E = 0, E
2
= h, E
3
= 0, E
4
= 3h
2
.
Do N sample paths per time-step - one for each z
[i ]
. A simple w is
=
3h with probability
1
6
,
=
3h with probability
1
6
,
= 0 with probability
2
3
.
Important facts about these bounded increments:
3h, where
k Z
d
.
3h with probability
1
6
,
=
3h with probability
1
6
,
= 0 with probability
2
3
.
Important facts about these bounded increments:
3h, where
k Z
d
.
Langevin dynamics
Langevin dynamics
Langevin dynamics
2
)t +w(t )
.
Intro to SDEs with with Examples
Stochastic Differential Equations
Some Applications
Langevin Dynamics
Langevin dynamics: we want some physical quantity
Ef =
_
p(x)f (x)d
n
x =
_
e
S(x)
f (x)d
n
x
_
e
S(x)
d
n
x
.
To nd a covering distribution q(x), q(x) p(x), but 1 is not
large - difcult if n large.
Alternative is Langevin dynamics:
dx(t ) =
1
2
S
x
dt +dw(t ),
and use
Ef = lim
T
1
T
_
T
0
f (x(t ))dt .
The following is sufcient for convergence: if |x| big,
x
S
x
> 1
Intro to SDEs with with Examples
Stochastic Differential Equations
Some Applications
A Simple Example
A simple example: dx = sign(x)dt +dw, whose p.d.f as t is
p(x, ) = e
2|x|
.
Intro to SDEs with with Examples
Stochastic Differential Equations
Some Applications
Stochastic Dyer-Roeder
Stochastic Dyer-Roeder: Sachs equations for shear (), ray
separation , in free space with scattered point-like particles:
d
ds
+2 = F
d
ds
+
2
+||
2
= 0
is complex, F is the Weyl term, and s is an afne parameter -
related to redshift z.
=
1
2
d
dz
ln(A)
where A D
2
is the beam area, get two eqs.,
d
ds
+2
1
D
dD
ds
= F
1
D
d
2
D
ds
2
+||
2
= 0.
Intro to SDEs with with Examples
Stochastic Differential Equations
Some Applications
Stochastic Dyer-Roeder
In Lagrangian coordinates (contract with redshift z), the Weyl term to
1
st
order has derivatives of the gravitational potential (x, y), with
z = x +i y:
F =
1
c
2
(1 +z)
2
d
2
dz
2
.
Light "sees" shearing forces orthogonal to congruence. Problem is
essentially 2-D:
x
y
light ray
scattering plane
Figure: 2-D character of light scattering
Intro to SDEs with with Examples
Stochastic Differential Equations
Some Applications
Stochastic Dyer-Roeder
Correlation length is about 7 cells, i.e. 7 Mpc at z = 0. Softened
(2-3 cells) shears are normal in < 128 Mpc.
Figure: Shearing forces, from H. Couchmans code
Intro to SDEs with with Examples
Stochastic Differential Equations
Some Applications
Stochastic Dyer-Roeder
More useful form for 1st:
D
2
=
_
s
0
D
2
(s
)F(s
)ds
.
Expressing the afne parameter in terms of the redshift
s =
_
z
0
d
(1 +)
3
1 +)
Yields a generalized Dyer-Roeder eq.
(1 +z)(1 +z)
d
2
D
dz
2
+(
7
2
z +
2
+3)
dD
dz
+
|(z)|
2
(1 +z)
5
D = 0.
Intro to SDEs with with Examples
Stochastic Differential Equations
Some Applications
Stochastic Dyer-Roeder
Shear can be well approximated by
(z) =
3
8(D(z))
2
_
z
0
(D())
2
(1 +)(1 +)
1
2
dw()
where w(z) is a complex (2-D) B-motion. Constant 0.62 was
determined by N-body simulations.
Intro to SDEs with with Examples
Stochastic Differential Equations
Some Applications
Stochastic Dyer-Roeder
Figure: Shear free Dyer-Roeder D(z)
Intro to SDEs with with Examples
Stochastic Differential Equations
Some Applications
Stochastic Dyer-Roeder
Figure: D(z) histograms at 0 z 5. Non-linear integration. Scales for the
abscissas are: 10
6
for z = 1/2, 10
5
for z = 1, 2, 3, 4, 5.
Intro to SDEs with with Examples
Stability
Weak Solutions
Weak Simulations
Recall some basic rules of the It calculus
Edw(t ) = 0
E{dw(t ) dw(s)} = (t s) dt ds
Multi-dimensional version
dw
i
(t )
2
dt or dw
i
(t )dw
j
(t )
ij
dt
Usual z(t ) C
0
process:
dz = b(z) dt +(z) dw(t ) (SDE)
is shorthand for
z(t ) = z
0
+
_
t
0
b(z
s
) ds +
_
t
0
(z
s
) dw
s
.
Stochastic integral is non-anticipating. Important thing about It rule:
E{
_
t
0
(z
s
)dw
s
} = 0.
Intro to SDEs with with Examples
Stability
Weak Solutions
Weak Simulations
Taking the expression for z(t ) for one step t t +h,
z(t +h) = z
t
+
_
t +h
t
b
s
ds +
_
t +h
t
s
dw
s
,
and substituting z(s) from the right-hand side into the left side
integrals, e. g.
_
t +h
t
b(z
s
) ds =
_
t +h
t
b(z
t
+
_
s
t
b
u
du +
_
s
t
u
dw
u
) ds.
Since t u s t +h and
_
s
t
u
dw(u) = O((s t )
1/2
)
an expansion gives, including the
_
dw term, Picard-fashion, a
stochastic Taylor series (due to Wolfgang Wagner)
Intro to SDEs with with Examples
Stability
Weak Solutions
Weak Simulations
Truncating Taylor series to O(h) accuracy, we get Milsteins method
(scalar case):
z(t +h) = z(t ) +hb(z(t )) +(z(t ))
+
1
2
(
2
h)
Again
=
h
where =zero-centered, univariate normal:
E = 0, E
2
= 1.
Notice that because E
2
= h, Milsteins term preserves the
Martingale property
E
1
2
t
(
2
h) = 0.
Intro to SDEs with with Examples
Stability
Weak Solutions
Weak Simulations
It is not hard to modify this for vector case:
z
t +h
= z
t
+hb
t
+
t
w+
1
2
t
Where matrix is a model
_
t +h
t
=
h
2
(
) >
=
h
2
(
1
+
) <
=
h
2
((
1
)
2
1) =
Additional variables
.
Intro to SDEs with with Examples
Stability
Higher-Order Schemes
Higher-Order Schemes
Here is a second order accurate method. Writing b = A +B,
z
t +h
= z
t
+
h
2
(A
(z
t +h
) +B
(z
t
+
t
1
+(A
t
+B
t
)h)
+A
(z
t
) +B
(z
t
))
+
1
2
{
(z
t
+
_
1
2
0
+
h
2
(A
t
+B
t
))
+
(z
t
_
1
2
0
+
h
2
(A
t
+B
t
))}
1
+(
t
)
.
The rst A(z
t +h
) is implicit.
Intro to SDEs with with Examples
Stability
Examples
An Example
Lets take a simple case, M > 0 (stable matrix),
dz = Mzdt +dw
and write M = A +B, where I +hA is easy to invert. The semi-implicit
algorithm is
(I +hA) z
t +h
= (I hB) z
t
+ w
or
z
t +h
= (I +hA)
1
((I hB) z
t
+w)
In particular case A = B =
1
2
M,
z
t +h
= (I +
h
2
M)
1
((I
h
2
M)z
t
+w).
Stability of procedure will depend on L
2
norm
||(I +
h
2
M)
1
(I
h
2
M)|| < 1.
Intro to SDEs with with Examples
Stability
Examples
An Example
Even in scalar case, when h is large enough (h > 2/M), |1 hM| > 1,
but
|(1 hM/2)/(1 +hM/2)| 1
for all h > 0.
Two dimensional case when scales of e.v.s are very different:
_
dx
dy
_
=
1
2
_
1
+
2
,
1
2
,
1
+
2
_ _
x
y
_
dt
+
_
dw
1
(t )
dw
2
(t )
_
.
which converges for
i
> 0, but if
1
2
, the stepsize
h < 2/
1
- too small to be useful. This is stiffness, just
like in the ODE case.
Intro to SDEs with with Examples
Stability
Examples
Another Example
For real, SPD matrix M:
dz = Mzdt +dw.
The solution is formally
z(t ) = e
Mt
z(0) +
_
t
0
e
M(st )
dw(s).
Large t corr. matrix approximates
1
2
M
1
:
Ez
i
()z
j
() =
1
2
[M
1
]
ij
.
For big M, actual computational method is
Ez
i
()z
j
()
1
T
_
T
0
z
i
(t )z
j
(t )dt
as T gets big, from the ergodic theorem.
Intro to SDEs with with Examples
Stability
Examples
Non-Symmetric Case
Non-symmetric case:
dX = MXdt +dw,
dY = M
T
Ydt +dw,
initial conditions X(0) = Y(0) = 0. Same nD w for both X(t ), Y(t ).
From formal solutions, extract X, Y covariance
EX(t )Y
T
(t )
1
2
M
1
as t .
Again, splitting M = A +B, a stabilized and cheap procedure for each
X(t ), Y(t ) is
z
t +h
= (I +hA)
1
(I hB) z
t
+w
where in the diffusion term, we ignore the O(h
3/2
) contribution.
Examples: A = diag(M), or A = tridiag(M)
Intro to SDEs with with Examples
Stability
Examples
A Test Problem
Test problem: M = U
T
TrU, where Tr = upper triangular,
diag(Tr ) = (1, . . . , N), [Tr ]
i ,j
N(0, 1), j > i . Random orthogonal U
by Pete Stewarts procedure: S = diag(sign(u
1
))
U = SU
0
U
1
. . . U
N2
where
U
k
=
_
I
k
H
Nk
_
H
j
= Householder transforms,
H
j
= I
j
2
uu
T
||u||
2
with j length vectors u
u = x ||x||e
1
,
each x
i
N(0, 1), i = 1, . . . , j . Also, cond(M) N.
Intro to SDEs with with Examples
Stability
Numerical Examples
Convergence of the Euler Method
1
Intro to SDEs with with Examples
Stability
Numerical Examples
Convergence of a Second-Order Method
1
Intro to SDEs with with Examples
Stability
Numerical Examples
More Examples
More general problems? Some has been done. Talay, Tubaro, and
Ballys Euler estimates
|Ef (z(T)) Ef (z
n
(T))| h
K(T)||f ||
T
q
h = T/n = time step, q > 0 constant, and K(T) is non-decreasing.
Optimal choice of T is unclear. Example of Langevin dynamics,
dz(t ) = b(z)dt +dw(t ), (2)
want z to converge to stationary. For large |z(t )|,
E|z +z|
2
E|z|
2
.
From eq. (2),
2z b(z) 1
Discretization errors O(h) for Euler, O(h
2
) for 2nd order RK.
Intro to SDEs with with Examples
Stability
Numerical Examples
A Final Example
A nal example model problem, where m Z
+
dx = x|x|
m1
dt +dw(t )
Two procedures: =
h ,
x
h
= XTR(x
0
, )
= x
0
h
2
(x
euler
|x
euler
|
m1
+x
0
|x
0
|
m1
)
+
x
h
= ITR(x
0
, )
= x
0
h
2
(x
h
|x
h
|
m1
+x
0
|x
0
|
m1
)
+
Intro to SDEs with with Examples
Stability
Numerical Examples
A Final Example
Intro to SDEs with with Examples
Bibliography
Selected Bibliography
[1] P. E. Kloeden and E. Platen, Numerical Solution of Stochastic
Differential Equations, Springer, 1999.
[2] G. N. Milstein Numerical Integration of Stochastic Differential
Equations, Kluwer, 1995.
[3] W. Petersen, J. Comp. Physics, 113(1), July 1994.
[4] W. Petersen, J. Stoch. Analysis and Applications, 22(4):
9891008, 2004.
[5] W. Petersen, SIAM Journal on Numerical Analysis, 35(4): 1439,
1998.
[6] F. Buchmann and W. Petersen, Preconditioner..., SAM Report,
2003.
[7] G. W. Stewart, SIAM Journal on Numerical Analysis, 17(3), 1980.