a tutorial introduction to stochastic differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf ·...
TRANSCRIPT
![Page 1: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/1.jpg)
A Tutorial Introduction toStochastic Differential Equations:
Continuous-time Gaussian Markov Processes
Chris Williams
Institute for Adaptive and Neural ComputationSchool of Informatics, University of Edinburgh, UK
Presented: 9 December, minor revisions 13 December 2006
1
![Page 2: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/2.jpg)
AR Processes: Discrete-time Gaussian Markov Processes
A discrete-time autoregressive (AR) process of order p:
Xt =
p∑k=1
akXt−k + b0Zt ,
where Zt ∼ N (0, 1) and all Zt ’s are iid.
AR(2) example:
. . . . . .
Linear combinations of Gaussians are Gaussian
2
![Page 3: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/3.jpg)
From discrete to continuous time
In continuous time, have not only the function value but alsop of its derivatives at time t
apX(p)(t) + ap−1X
(p−1)(t) + . . . + a0X (t) = b0Z (t),
where Z (t) is a white Gaussian noise process with covarianceδ(t − t ′), and ap = 1.
This is a stochastic differential equation (SDE)
Applications in many fields, e.g. chemistry, epidemiology,finance, neural modelling
We will consider only SDEs driven by Gaussian white noise;this can be relaxed
3
![Page 4: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/4.jpg)
Vector processes
An AR(p) process can be written as a vector AR(1) process ifone stores Xt and the previous p − 1 values in Xt
Similarly for the pth order SDE
X (p)(t) + ap−1X(p−1)(t) + . . . + a0X (t) = b0Z (t),
X1(t) = X (t)
X2(t) = X1(t) = X (t)
...
Xp(t) = Xp−1(t) = X (p−1)(t)
Xp(t) + ap−1Xp(t) + . . . + a1X2(t) + a0X1(t) = b0Z (t)
4
![Page 5: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/5.jpg)
or, in matrix form
X(t) = FX(t) + BZ(t)
for Z(t) being a p-dimensional white noise process, with
F =
0 1 0 . . . 0 00 0 1 . . . 0 0...
......
......
0 0 0 . . . 0 1−a0 −a1 −a2 . . . −ap−2 −ap−1
and
B = diag(0, 0, . . . , 0, 1)
5
![Page 6: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/6.jpg)
Overview
Wiener process
SDEs and simulation
Stationary processes and covariance functions
Inference (Gaussian process prediction)
Fokker-Planck equations
3 views: SDE vs covariance function vs Fokker-Planck
6
![Page 7: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/7.jpg)
The Wiener Process
W (t) is continuous and W (0) = 0
W (t) ∼ N(0, t)
Independent increments: W (t + s)−W (s) ∼ N(0, t) and isindependent of the history of the process up to time t
cov(W (s),W (t)) = min(s, t)
Interpret dW (t) = W (t + dt)−W (t)
7
![Page 8: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/8.jpg)
Discretized Wiener Process
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−3
−2.5
−2
−1.5
−1
−0.5
0
0.5
1
1.5
t
W(t)
8
![Page 9: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/9.jpg)
Gaussian Processes
For a stochastic process X (t), mean function is
µ(t) = E[X (t)]
Covariance function
k(t, t ′) = E[(X (t)− µ(t))(X (t ′)− µ(t ′))]
Gaussian processes are stochastic processes defined by theirmean and covariance functions
9
![Page 10: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/10.jpg)
SDEs
Consider the SDE
X(t) = FX(t) + BZ(t)
This is a Langevin equation
A problem is that we want to think of Z(t) as being thederivative of a Wiener process, but the Wiener process is(with probability one) nowhere differentiable ...
The “kosher” way of writing this SDE is
dX(t) = FX(t)dt + B dW(t)
where W(t) is a vector of Wiener processes
10
![Page 11: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/11.jpg)
Simulation of an SDE
Times t0 < t1 < t2 < . . . < tn, ∆ti = ti+1 − ti
Xi+1 = Xi + FXi∆ti + BZi
√∆ti
where Zi ∼ N(0, I )
This is the Euler-Maruyama method; higher-order methods are alsopossible (Milstein)
11
![Page 12: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/12.jpg)
Stochastic Integration
Riemann sum∫ T
0h(t)dt = lim
N→∞
N−1∑j=0
h(tj)(tj+1 − tj)
for tj = jT/NIto stochastic integral∫ T
0h(t)dW (t) = m.s. limN→∞
N−1∑j=0
h(tj)(W (tj+1)−W (tj))
Example ∫ T
0W (t)dW (t) =
1
2W (T )2 − 1
2T
Mnemonically we have dW (t)2 = dt
12
![Page 13: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/13.jpg)
General form of a Diffusion process
dX(t) = a(X, t)dt + B(X, t) dW(t)
where the functions a(X, t) and B(X, t) must be non-anticipating,corresponding to the integral form
X(t)− X(0) =
∫ t
0a(X(t ′), t ′)dt ′ +
∫ t
0B(X(t ′), t ′) dW(t ′)
a(X, t) is the drift vector, B(X, t) is the diffusion matrix
Sample paths of a diffusion process are continuous
dX(t) = F (t)X(t)dt + B(t) dW(t)
is the most general form that is a Gaussian process
13
![Page 14: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/14.jpg)
Simple Examples
Wiener process
dX = dW X (t) = X (0) + W (t)
Wiener process with scaling and drift
dX = adt + σdW X (t) = X (0) + at + σW (t)
Ornstein-Uhlenbeck process
dX = −aXdt+σdW X (t) = X (0)e−at+σ
∫ t
0e−a(t−t′)dW (t ′)
14
![Page 15: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/15.jpg)
Infinitesimal moments
∆X(t) = X(t + ∆t)− X(t) = a(X, t)∆t + B(X, t)Zt
√∆t
First moment: drift
lim∆t→0
E[∆X(t)]
∆t= a(X , t)
Second moment: diffusion
lim∆t→0
E[(∆X(t))(∆X(t))T ]
∆t= B(X , t)BT (X , t)
Notice that E[(√
∆tZt)(√
∆tZt)T ] = (∆t)I
15
![Page 16: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/16.jpg)
Stationary Processes
Assume time-invariant coefficients of univariate SDE of orderp
If the coefficients are such that eigenvalues of F are in the lefthalf plane (negative real parts) then the SDE will have astationary distribution, such that E[X (t)X (t ′)] = k(t − t ′)
Can generalize this to vector-valued processes, when k is amatrix-valued function
16
![Page 17: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/17.jpg)
Fourier Analysis
Sinusoids are eigenfunctions of LTI systems
X (s) =
∫ ∞
−∞X (t)e−2πist dt, X (t) =
∫ ∞
−∞X (s)e2πist ds,
X (k)(t) =
∫ ∞
−∞(2πis)k X (s)e2πist ds.
p∑k=0
akX (k)(t) = b0Z (t),
p∑k=0
ak(2πis)k X (s) = b0Z (s)
17
![Page 18: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/18.jpg)
Power spectrum of SDE
Wiener-Khintchine Theorem
k(τ ) =
∫S(s)e2πis·τ ds, S(s) =
∫k(τ )e−2πis·τ dτ .
so〈X (s1)X
∗(s2)〉 = S(s1)δ(s1 − s2)
and thus
S(s) =b20
|A(2πis)|2
where A(z) =∑p
k=0 akzk . Require that roots of A(z) lie in lefthalf plane for stationarity
18
![Page 19: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/19.jpg)
Examples
First order SDE
X (t) + a0X (t) = b0Z (t), S(s) =b20
(2πs)2 + a20
, k(t) =b20
2a0e−a0|t|
Damped simple harmonic oscillator (second order SDE)
X (t)+a1X (t)+a0X (t) = b0Z (t), S(s) =b20
(a0 − (2πs)2)2 + a21(2πs)2
if a21 < 4a0 (weak damping) then
k(t) =b20
2a0a1e−α|t|(cos(βt) +
α
βsin(β|t|))
where α = a1/2, and α2 + β2 = a0.
19
![Page 20: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/20.jpg)
0 1
−2
−1
0
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−1.5
−1
−0.5
0
0.5
1
1.5
OU process Damped harmonic oscillatora0 = 5 a0 = 500, a1 = 5
20
![Page 21: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/21.jpg)
Vector OU process
dX(t) = −AX(t) + BdW(t)
solution is
X(t) = exp(−At)X(0) +
∫ t
0exp(−A(t − t ′))B dW(t ′)
For stationary solution remove X(0) dependence
〈X(t)XT (s)〉def= Σ(t − s)
=
∫ min(t,s)
−∞exp(−A(t − t ′))BBT exp(−AT (s − t ′)) dt ′
21
![Page 22: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/22.jpg)
Can show that
AΣ(0) + Σ(0)AT = BBT
and
Σ(t − s) = exp(−A(t − s))Σ(0) for t > s
and Σ(t − s) = ΣT (s − t)
Can also do spectral analysis of vector OU process
See Gardiner (1985, §4.4.6) for more details
22
![Page 23: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/23.jpg)
Mean square differentiability
apX(p)(t) + ap−1X
(p−1)(t) + . . . + a0X (t) = b0Z (t),
SDEs of order p are p − 1 times mean square differentiable
This is easy to see intuitively from the above equation, asX (p)(t) is like white noise
Note that a process gets rougher the more times it isdifferentiated
23
![Page 24: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/24.jpg)
Relating Discrete-time and Sampled Continuous-timeGMPs
Discrete time ARMA(p, q) process
Xt =
p∑i=1
Xt−i +
q∑j=0
bjZt−j
A continuous-time ARMA process has spectral density
S(s) =|B(2πis)|2
|A(2πis)|2
24
![Page 25: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/25.jpg)
Theorem (e.g. Ihara, 1993): Let X be a continuous-timestationary Gaussian process and Xh be the discretization ofthis process. If X is an ARMA process then Xh is also anARMA process. However, if X is an AR process then Xh isnot necessarily an AR process
A discretized continuous-time AR(1) process is a discrete-timeAR(1) process
However, a discretized continuous-time AR(2) process is not,in general, a discrete-time AR(2) process.
25
![Page 26: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/26.jpg)
Inference
Given observations of X at times t1, t2, . . . , tn, computeposterior distribution at t∗
Note that for OU process, the Markov property means that weneed only condition on tP and tF , the nearest times to thepast and future of t∗
Caveat: observations must be noise free, otherwise allobservations will count
This is just Gaussian process prediction:
26
![Page 27: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/27.jpg)
X (t∗)|X (t1), . . . X (tn) ∼ N (µ(t∗), σ2(t∗))
with
µ(t∗) = (k∗P , k∗F )
(kPP kPF
kPF kFF
)−1 (XP
XF
)σ2(t∗) = k∗∗ − (k∗P , k∗F )
(kPP kPF
kPF kFF
)−1 (k∗Pk∗F
)where k∗P = k(t∗, tP) etc
Vector process works similarly
27
![Page 28: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/28.jpg)
Fokker-Planck Equations
Consider the transition pdf pdef= p(x, t|x0, t0). This evolves
according to the (forward) Fokker-Planck equation
∂tp = −∑
i
∂i (ai (x, t)p) +1
2∂i∂j [B(x, t)BT (x, t)]ijp]
corresponding to the SDE
dX(t) = a(X, t)dt + B(X, t) dW(t)
This is just the differential form of the Chapman-Kolmogorovequation
There is also a “backward” equation
28
![Page 29: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/29.jpg)
Simple example: Wiener process with drift
Wiener process with scaling and drift
dX = adt + σdW X (t) = X (0) + at + σW (t)
p(x , t|x0, 0) =1√
2πσ2texp
(−(x − x0 − at)2
2σ2t
)
29
![Page 30: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/30.jpg)
Fokker-Planck Boundary Conditions
Feller, 1952
Regular
AbsorbingReflecting...
Exit
Entrance
Natural0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
−2
−1.5
−1
−0.5
0
0.5
1
1.5
t
W(t)
30
![Page 31: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/31.jpg)
Parameter Estimation
If we have observations X = (X (t1), . . . ,X (tn))T of a
Gaussian process at some set of finite times t1, . . . , tn, then
log p(X|θ) = −1
2log |Kθ|−
1
2(X−µθ)
TK−1θ (X−µθ)−
n
2log(2π)
Can use e.g. numerical methods to optimize parameters θ
For continuous observations, see e.g. Feigin (1976)
31
![Page 32: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/32.jpg)
Summary
Relationship of SDEs driven by Gaussian white noise toGaussian Markov processes
Formal mathematical framework of stochastic integration
As Gaussian processes we can compute their mean andcovariance functions, and do inference
Markov properties are to the fore for Fokker-Planck equations
Extend to allow observation noise: continuous-time Kalmanfilter (Kalman and Bucy, 1961)
Challenges of the workshop: nonlinear dynamics, nonlinearobservation
32
![Page 33: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/33.jpg)
References
C. W. Gardiner, Handbook of Stochastic Methods (secondedition) Springer-Verlag, 1985 [3rd ed, 2004]B. Øksendal, Stochastic Differential Equations,Springer-Verlag, 1985D. J. Higham, An Algorithmic Introduction to NumericalSimulation of Stochastic Differential Equations, SIAM Review43(3) 525-546, 2001P. E. Kloeden and E. Platen, The Numerical Solution ofStochastic Differential Equations Springer-Verlag, 1999P. D. Feigin, Maximum likelihood estimation forcontinuous-time stochastic processes, Adv. Appl. Prob. 55,468-519, 1976C. E. Rasmussen and C. K. I. Williams, Gaussian Processes forMachine Learning [see Appendix B], MIT Press, 2006
33
![Page 34: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/34.jpg)
Stratonovich stochastic integral
Ito stochastic integral∫ T
0h(t)dW (t) = m.s. limN→∞
N−1∑j=0
h(tj)(W (tj+1)−W (tj))
Stratonovich integral∫ T
0h(t)dW (t) = m.s. limN→∞
N−1∑j=0
h(tj + tj+1
2)(W (tj+1)−W (tj))
Some authors use 12(h(tj) + h(tj+1))(W (tj+1)−W (tj))
instead
34
![Page 35: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/35.jpg)
Ito’s formula
Let the stochastic process X satisfy
dX = a(X , t)dt + b(X , t)dW
Then Y = f (X , t) satisfies
dY =
(a(X , t)fx(X , t) +
1
2b2(X , t)fxx(X , t) + ft(X , t)
)dt
+ (b(X , t)fx(X , t))dW
Example: Y (t) = X (t)2, dX = dW (Wiener process)
dY = dt + 2√
Y dW
35
![Page 36: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/36.jpg)
∫ T
0 W (t)dW (t)
N−1∑j=0
W (tj)(W (tj+1)−W (tj))
=1
2
N−1∑j=0
(W (tj+1)
2 −W (tj)2 − (W (tj+1)−W (tj))
2)
=1
2
W (T )2 −W (0)2 −N−1∑j=0
(W (tj+1)−W (tj))2
Last term has expected value T and variance O(δt), Thus∫ T
0W (t)dW (t) =
1
2W (T )2 − 1
2T
for the Ito integral
36
![Page 37: A Tutorial Introduction to Stochastic Differential ...homepages.inf.ed.ac.uk/ckiw/talks/sde1.pdf · This is the Euler-Maruyama method; higher-order methods are also possible (Milstein)](https://reader031.vdocuments.us/reader031/viewer/2022030402/5a78c6a07f8b9a83238c4442/html5/thumbnails/37.jpg)
Geometric Wiener Process
dX = X (µdt + σdW )
X (t) = exp(σW (t) + (µ− 1
2σ2)t)
An essential part of the Black-Scholes model for option pricing
37