Professional Documents
Culture Documents
By
M. H. A. Davis
By
M. H. A. Davis
Notes by
K. M. Ramachandran
Springer-Verlag
Berlin Heidelberg New York Tokyo
1984
Author
M. H. A. Davis
Department of Electrical Engineering
Imperial College of Science and Technology
London SW 7
United Kingdom
c
Tata Institute of Fundamental Research, 1984
Printed in India
Preface
These notes comprise the contents of lectures I gave at the T.I.F.R. Cen-
tre in Bangalore in April/May 1983. There are actually two separate
series of lectures, on controlled stochastic jump processes and nonlin-
ear filtering respectively, and the corresponding two parts of these notes
are almost disjoint. They are united however, by the common philoso-
phy (if that is not too grand a work for it) of treating Markov processes
by methods of stochastic calculus, and I hope the reader will, at least,
be convinced of the usefulness of this and of the extended generator
concept in doing calculations with Markov precesses.
The first part is aimed at developing optimal control theory for a
class of Markov processes called piecewise-deterministic (PD)proce-
sses. These were only isolated rather recently but seen general enough
to include as special cases practically all the non-diffusion continuous
time processes of applied probability. Optimal control for PD processes
occupies a curious position just half way between deterministic and Sto-
chastic optimal control theory in such a way that no standard theory
from either side is adequate to deal with it. The only applicable theory
that exists at all is very recent work of D. Vermes based on the gener-
alized dynamic programming ideas of R.B. Vinter and R.M. Lewis, and
this is what I have attempted to describe here. Undoubtedly, further de-
velopment of control theory for PD processes will be a fruitful field of
enquiry.
Part II concentrates on the pathwise theory of filtering for diffu-
sion processes and on more sophisticated extensions of it due primarily
to H. Kunita. The intriguing point here is to see how stochastic partial
v
vi
II Filtering Theory 63
0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 65
1 Linear and Nonlinear Filtering Equations.... . . . . . . . 65
2 Pathwise Solutions of Differential Equations . . . . . . . 81
3 Pathwise Solution of the Filter Equation . . . . . . . . . 86
Bibliography 101
vii
Part I
1
Chapter 1
0 Introduction
3
4 1. Stochastic Jump Processes
Service
Demand
Figure 1.2:
x1
S (K) =
([s, ] X)U([o, s] {z0 }) if z0 K and Z(E A) K =
((s, ] X)U([o, s] K) if z0 K and Z(E A) K , .
[o, s] K if z0 < K.
where E = R+ X A.
Clearly [0, t] X cannot be in the - algebra generated by sets of
the above form. So T is not an Ft0 stopping time. Let B = X {z0 }. By
assumption, P(A) = 0; so A Ft .
0
x1
t (B) = [0, t] X A F t .
5 So
[0, t] X Ft .
But {T t} = [0, t] X. Hence T is an Ft stopping time.
FT = {G F : G (T t) Ft , t}.
Clearly
FT = F .
Proof. Let S n be stopping times such that S n a.s. Then MtS n is u.i.
martingale. Let Mtn = MtS n , n. Then by optional sampling theorem
E[MSn FT S n ] = MTn ;
but FT = F .
So 6
E[MSn | FT ] = MSn .
Also
lim MTn = MT
n
and lim MSn = MS .
n
So
MS = MT a.s.
(i) F = Fto
F A (t) = ([t, ] A)
and F(t) = F X (t) = P[T > t].
Note that F(0) = 1 and F(.) is monotone decreasing and right con-
tinuous. Define
inf {t : F(t) = 0}
c=
+ if {t : F(t) = 0} = .
Proposition 1.5. Let q(t, A) = p(t, A) p(t, A). Then (q(t, A))t0 is an Ft
- martingale, i.e., p(t, A) the compensator of the point process p(t, A).
Proof. (Direct calculation). Take t > s, then 9
F A (s) F A (t)
E[p(t, A) p(s, A) | F s ] = I(s<T ) .
F(s)
E[ p(t, A) p(s, A) | F s ] = I s<T
Z Z Z
F(t) dF A (u) 1 dF A (u)
dF(r)
F(s)
F(u) F(s) F(u)
[s.t] [s.t] [s.r]
and
Z Z Z Z
dF A (u) 1
dF(r) = dF(r)dF A (u)
F(u) F(u)
[s,t] [s,r] [s,t] [u,t]
10 1. Stochastic Jump Processes
Z
1
= (F(t) F(u))dF A (u)
F(u)
[s.t]
Z
dF A (u)
= F(t) + F A (t) F A (s).
F(u)
[s.t]
So
E[q(t, A) q(s, A) | F s ] = 0
Another expression for p(t, A): We have F A (.) << F(.) (i.e. F A (.) is
absolutely continuous w.r.t. F(.)). So there exists a function (s, A) such
that Z
A A
F (0) F (t) = (s, A)dF(s).
]o.t]
In fact
(s, A) = P(Z A | T = s).
Suppose X is such that a regular version of this conditional probabil-
dF A (s)
ity exists (which is the case, since X is Borel space). Then =
F(s)
dF(s)
(s, A)d(s) where d(s) .
F(s)
Then Z
p(t, A) = (s, A)d(s).
]o,T T ]
10 Stochastic Integrals
Let I denote the set of measurable functions g : R such that
g(, z ) = 0.
jump points and they are only countable; the function Nt defines
P
a random measure on (R, B(R)) say, = T i where X is the
i
Dirac measure at x. Similarly, the one jump process can be iden-
tified with the random measure (T,x R T ) on R+ X. So we can
define Stieltjes integrals of the form g(t, x)p(dt, dx) for suitable
integrands g I as
Z
g(t, x) p(dt, dx) = g(T, xT ).
We say g L1 (p) if
Z
E |g(t, x) | p(dt, dx) <
and denote
Z
|| g ||L1 (p) = E | g(t, x) | p(dt, dx)
Proposition 1.6.
|| g ||L1 (p) =|| g ||L1 ( p)
and so
L1 (p) = L1 ( p).
Proof.
Z Z
1
|| g ||L1 ( p) = | g(s, x) | d(s, x)dF(T )
F(s )
R+ [o,T ]
Z Z
1
= | g(s, x) | ( dF(t))d(s, x)
F(s )
[s,]
Z
= | g(s, x) | d(s, x)
=|| g ||L1 (p) .
Define
Lloc
1 (p) = {g I | g(s, x)I st L1 (p), t < c}
Lloc
1 ( p) = {g I | g(s, x)I st L1 (p), t < c}. Clearly
Lloc loc
1 (p) = L1 ( p).
and Z
l
Mt = I(tT ) h(T, Z) I(t<T ) h(s, z)(ds, dz) (2)
F(t)
]0,t]X
= h(T, Z) a.s.
Mt = E[h(T, Z)|Ft ]
Z
1
= I(tT ) h(T, Z) + I(t<T ) h(s, x)(ds, dx). (4)
F(t)
]t,]X
For g Lloc
I (p), the stochastic integral
Z
g
Mt = g(s, x)q(ds, dx)
]0,t]X
is defined by
Z Z
g
Mt = I(st) g(s, x)p(ds, dx) I(st)g(s,x) p(ds, dx).
R+ X R+ X
g
Then the question is whether Mt given by (2) is equal to Mt for
some g. As a motivation to the answer consider the following example.
14 1. Stochastic Jump Processes
Then
ZT Z
g
1
Mt = I(tT )
g(T, Z) g(s, x)(s, x)dxds
F(s)
0 R
t
Z Z
1
I(tT )
g(s, x)(s, x)dxds
(5)
F(s)
0 R
g
14 If My given by (5) is equal to Mt given by (2), then the coefficients
of I(tT ) and I(t<T ) must agree. Comparing the coefficients of I(t>T ) , we
require
Zt Z
1
h(t, z) = g(t, z) g(s, x)(s, x)dxds.
F(s)
0 R
Let
(t) = h(t, z) g(t, z).
Define Z
(t) = hdx
R
and Z
f (t) = dx.
R
Then
ZT Z
1
(t) = h(s, z) + (s) (s, x)dx)ds
F(s)
0 R
Zt Zt
1 1
= (s)ds + (s) f (s)ds;
F(s) F(s)
0 0
1. Martingale Theory for Jump Processes 15
that is
d f (t) 1
(t) = n(t) + (t)
dt F(t) F(t)
(o) = 0
which has a unique solution
Zt
1
(t) = (t, s) (s)ds,
F(s)
0
where
Zt
f (u)
(t, s) = exp du
F(u)
s
F(s) dF(t)
= , since f (t) = .
F(t) dt
So 15
Zt
1
(t) = (s)ds.
F(t)
0
Hence
Zt Z
1
g(t, z) = h(t, z) + h(s, x)(s, x)dxds (6)
F(t)
0 R
Now it can be checked that with this choice of g the coefficients of I(t<T )
in (2) coincides with that of (5). So Mt = Mtg .
Now we can prove the general case given in Proposition 1.7.
M s = E[h(T, Z)|F ]
Hence g Llog
1 (p).
1. Martingale Theory for Jump Processes 17
log g
Conversely, suppose g L1 (p). Then it can be checked that Mt is
a local martingale.
[(T 1 , Z1 ) ] = 1 ()
[(S i , Zi ) |wi1 = ] = i ( : )i = 2, 3
Notice that as in the single jump case, here (iii) ensure that two
jump times T i1 , T do not occur at one and that the process xt does
effectively jump at its jump times (iv) ensures that [Zk = z |T k = ] =
1.
As before, Fto = {xs , s t} and
Ft = completion of Fto with all -null sets of F o .
The idea here is to reduce everything to one jump case. That is, the
process restarts at each T k . We need the following result.
Proposition 1.9.
F(T kl +t) T k = FT kl V x(T kl +s) T k , o s t) .
if U = (T k1 + to )T k , then
k1
Y Y
FU = yi ykto Yi ,
i1 i=k+1
where ykto = S B[o, to ] (x [to , ]) .
X
2
xt = I(tT i ) .
i=1
Then
(a) Show that
Ft = Borel sets in {S 1 + S 2 t}
\
+ (A R+ ) B + [t, ] R+ , A B(R)
where \
B = {S 1 + S 2 t} {S 2 t}.
where 20
F 1 = 1 ([t, ] X)
and Z
1
kA (wk1 ; s) = dF kA (u)
F k (u )
[o,s]
where
F kA (u) = k (wk1 ; [u, ] A).
Now define
p(t, A) = 1A (T 1 ) + 2A (w1 ; S 2 ) + Aj (w j1 ; t T j1 ())
for t ]T j1 , T j ].
Exercise 1.3. Consider a renewal process
X
xt = I(tT i )
i
where
dF kA
(kl , s, A) = (s)
dF k
g loc
1 (p) if there exists a sequence of stopping times k T a.s. and
gI(tn L1 (p), n. For g L1oc
1 (p) we define
Z
g
Mt = g(s, x) q (ds, dx)
[o,t]X
Z Z
= g(s, x) p (ds, dx) g(s, x) p (ds, dx).
[o,t]X [o,t]X
x1t = MtT 1
xkt M(t+T kl )T k MT k1 , k = 2, 3, . . .
Not every Markov process has a transition function, but usually one
wants to start with transition function and construct the corresponding
process. This is possible if (E, E) is a Borel space (required to apply
Kolmogorov extension theorem; refer Wentzel [27]). One constructs a
Markov family,
{P x,s , (x, s) E R+ }px,s
being the measure for the process starting at xs = x. All measures P s,x
have the same transition function p. Denote by E x,s integration w.r.t
24 1. Stochastic Jump Processes
25 Define
T s,t f (x) = E x,s | f (xt )|], s t.
T s,t is an operator on B(E) such that
for
26 So
2. Some Discontinuous Markov Processes 25
d
Af = T t f (x)|t=0
dt
Take f D(A). Then
1 1
lim (T t T s f T s f ) = lim (T t+s f T s f )
t0 t t0 t
1
= lim T s (T t f f )
t0 t
= T s A f.
So f D(A) implies T s f D(A) and AT s f = T s A f . So we get
backward Kolmogorov equation
d
T s f = A(T s f ). (1)
ds
The main results of the analytic theory of Markov semigroups are the
following;
NB: The domain DA provides essential information.
Integrating (1), we get Dynkins formula
Zt
T t f (x) f (x) = T s A f (x)ds
0
Zt
i.e., E x f (xt ) f (x0 ) = E x A f (xs )ds, f D(A).
0
26 1. Stochastic Jump Processes
Proposition 2.1. If f D(A) then the process
Zt
Ctf = f (xt ) f (x0 ) A f (xs )ds
0
is a martingale.
27 Proof. For t s
Zt
E|Ctf C sf |F s ] = E f (xt ) f (xs ) A f (xu )du|F s
s
Zt
= E xs f (xt ) f (xs ) E xs A f (xs )ds. 0.
s
So C 2 D(A) and
X Zt
f 1X 2 f
A f (x) = bi (x) + ( )i j Cf = f dw.
i
xi 2 i j xi x j t
0
where i 0 and
X
i i = r < .
i=1
Find the extended generator of xt
(This is also an example where jump times are not isolated).
28 1. Stochastic Jump Processes
Here again an = 0, bn = +.
Example 2.5. VWT with renewal process arrivals. (The GI/G/I queue) 31
Suppose the inter arrivals times in Example 2.4 are not exponential,
but form a renewal process with inter arrival density b(.). Now the ap-
propriate structure is is 0 or 1 as before, d(1) = 2, d(0) = 1. (When
= 1 we have to remember both the value of VMT and the time since
the last arrival.)
Example 2.6. Shot noise. This has sample functions similar to the VWT
process except that decay between arrivals is exponential rather than
linear (fig. 2.1).
30 1. Stochastic Jump Processes
Figure 2.1:
32 Example 2.7. A model for capacity expansion. Suppose that the demand
for some utility is monotone increasing and follows a Poisson process
with rate . Each unit of supply provides q units of capacity. These
are built one at a time at a cost of Rs.p. Investment takes place at a rate
of Rs.u(t)/week and u(t) constant. When
Zt
u(s)ds = p,
0
E = U M = {(, ); K, M }.
K
Let
E = A ; A B(M)
K
d
f ( (t, z)) = X f ( (t, z))
dt
(o, z) = z
for every smooth function f , and (t, z) exists for all t o. Let M
be the boundary of M . M is those points in M at which integral
curves exit from M , i.e., M = {z M : (t, ) = z for some
(t, ) R M }.
Let
= {, z : : K, z M }.
So is the set of points on the boundary at which jumps may take
place. For x = (, z) E, denote
P[T k+l > s|T k , T k+l > t] = P[T k+l T k > s T k |T k , T k+l T k > t T k ]
sT
Z k
= exp (, (u, ))du
tT k
st
Z
= exp (t , t (u, t ))du
o
Example 2.8. Here has only a single value; so delete it. takes values
in M = [0.1] R+ , = 0 and X = . Then = {(l, y); y R+ }. Let
1
1 1
Q(., (l, y)) = (1 y, y).
2 2
Then starting at x0 = (0, 1) we have 36
X
n
1
Tn = so that lim T n < .
k=1
k1 n
Figure 2.2:
Proof. This follows form the fact that there is a one-to-one mapping
from x[o,t] to z[o,t] .x z is given by (2). Conversely, if z[o,t] is given
then x[o, t] can be constructed since the motion in the interval [T k , T k+1 [
is deterministic.
NB:
is an Ft -martingale.
But t 38
R
exp (xs )ds
t < t1 (x)
Ft =
0
0 t t1 (x).
dFt
Thus - = (xt )dt for t < t1 (x) and
Ft
Ftt
= 1.
Ft1
given by (3). Since p(t, A) and p(t, A) are monotonic increasing func-
tions and T n a.s. E(p(t, E)) < , taking the limits, we have
is a martingale.
Zt Z Z Z
B f (y, s, w) p(ds, dy) = ( f (y) f (xs ))Q(dy; xs )(xs )ds
o E [0,t] E
Z Z
+ ( f (y) f (xs ))Q(dy; xs )dss . (5)
[0,t] E
Then the second integral in (5) is zero. The following result charac-
terizes the extended generator A of (xt ).
(i) For each (n, z) E the function t f (n, n (n, z)) is absolutely
continuous for t [0, t (n, z)[.
(iii) Bf Lloc
1 (p).
Now, 40
X X
f (xT i ) f (xT i ) = ( f (xT i ) f (XT i1 )) + f (xt ) f (xT n )
T i t T i t
X
( f (xT i ) f (xT i1 )) + f (xt ) f (xT n )
T i t
where T n is the last jump time before t. The first bracket is ( f (xt )
f (xo )). Note that
ZT i
f (xT i ) f (xT i1 ) = XTi1 f (T i1 Ti1 (T i1 , s)ds a.s).
T i1
Rt
So the second bracket is equal to x f (xs )ds and
0
38 1. Stochastic Jump Processes
Z
B f dq = f (xt ) f (x0 )
Zt Z Zt
( f (y) f (xs ))Q(dy, xs )(xs )ds X f (xs )ds.
o E o
Rt
So A f is given by (7) and Ctf = Bf dq. Conversely, suppose
0
f D(A). Then there exists a function h such that s h(xs ) is
Rt
Lebesgue integrable and Mt = f (xt ) f (x0 ) h(xs )ds is a local mar-
0
g
tingale. By the martingale representation theorem, Mt = Mt for some
g
g L1oc
1 (p). Now the jumps of Mt and Mt must agree, these only occur
when t = T i for some i and are then given by
Mt = Mt Mt = f (xT i ) f (xT i ).
g g g
Mt = Mt Mt
Z
= g(xt , t, w) g(y, t, w)Q(dy, xt )I(xt )
E
at t = T i . It follows that
g(x, t, w)I(xt </) = ( f (x) f (xt ))I( xt<)
41 except possibly on a set G E p such that
Z
Ey IG p(dt, dx) = 0 for all y E.
R+ E
for all x except a set A E such that Q(A, z) = 0. Since only the first
terms on the left and right involve x it must be the case that
f (x) = g(x, t, ) + f(t, )
2. Some Discontinuous Markov Processes 39
Z
and f (z) = g(y, t, )Q(dy; z) + f(t, w)
E
Hence we get
Zt
Mt = f (o , o (t, o )) f (o , o ) h(xs )ds
o
Zt Z
g
Mt = ( f (y) f (xs ))Q(dy; xs )(xs )ds.
o E
f (s, x)
+ A f (s, x) (s, x) f (s, x) + c(s, x) = 0 (9)
s
(s, x) [o, t[ E.
43 Then
t s
Z Z
f (o, x) = Eo,x exp (u, xu )du c(s, xs )ds
0 o
t
Z
+ exp (u, xu )du (xt ) (10)
o
2. Some Discontinuous Markov Processes 41
Then
So
Z
m(t) = E[Nt |T 1 = s] f (s)ds
o
Zt
m(t) = (1 + m(t s)) f (s) ds. (11)
o
42 1. Stochastic Jump Processes
Z
f(p) = ept f (t)dt
o
etc., we get
1
m(p) = m(p) f(p) + f(p).
p
So
1
p f(p)
m(p) = .
1 f(p)
In particular, for the Poisson process f (t) = et ,
f =
+ p
will give
m(p) =
p2
to get
m(t) = t.
A f (, ) = f (, ) + ()[ f ( + 1, 0) f (, 0)].
2. Some Discontinuous Markov Processes 43
f (0, , ) = E(,) t .
Clearly
f (s, + 1, ) = f (s, , ) + 1.
Define
f (s, 0, ) = h(s, ).
Then the equation for f (or h) becomes
h(s, ) + h(s, ) + ()[1 + h(s, 0) h(s, )] = 0 (12)
s
h(t, ) = 0.
Define
z(u) = h(u, u).
Then
d
z(u) = (u)[1 + h(u, 0) z(u)]
du
z(t) = 0.
where
Z
F f
(u) = = , F(u) = f (s) ds.
F F
u
Zt
z(o) = h(o, o) = f (u)[1 + h(u, o)]du (14)
o
Define
m(s) = h(t s, o).
Then (14) coincides with the renewal equation (11). Having deter-
mined h(u, o) o u t, h(s, ) for s , , o can be calculated from
(12). The result will be equivalent to that of Exercise 2.4.
Chapter 2
Optimal Control of pd
Processes
45
46 2. Optimal Control of pd Processes
49 So
t
Z
V(o, x) E x es c(xs , us )ds + et (xt )
o
= Jx (u).
Objections:
x = b(xt , u(t)).
d
V() = g(V(.)).
d
Ztl
minimize (xt , t, ut )dt
to
then
< t + x f, x,u >= (xl , tl ) (xo , to ).
Define
= {C (A ) : (i) and (ii) are satisfied}.
(W) (S ).
Proposition 4.
Sketch of proof:
Then
if < W
q () =
i i
lim( (x1 , t1 ) (xo , to )) if W
i
where = lim i and i = it + ix f.
i
53
(D) = sup((xo , to ))
Proof of the main results now follow easily. The strong problem has
a solution since (S ) = (D).
(S ) = lim(i (xo , to ))
E = (Eo Eo ) ET .
f : Eo Y Rd ;
: Eo Y R+ ,
and Q : (Eo UE ) Y m1 (Eo )
xt = f (xt , yt ).
xt = Xn(t),z(t)
Zz
where Xn,z = n + f (Xn,s , y)ds.
o
0
(Z )
(x, n, z) min (, , o)Q(d, y, x) + o (x, y) (x, n, z) E (4)
yY
where A E .
QoT (A) = P xoo [ xT A]
for A ET .
Comparing with deterministic case, the necessary and sufficient con-
dition there was that (xt , t ) is optimal if and only if
Z n o
i
t (xt , t) + ix (xt , t) f (xt , t, u) (xt , t, u) t (du) o in Ll (t0 , t1 ).
62 then
Z X Z
E x g( xs )ds = ci E x Ai ( xs )ds
o i 0
X
= ci Q (Ai )
Zi
= g() Q(d).
Eo
t+s = t + s ot
t, pt , I(t) are some example of additive functionals.
Zt1
(x1 , t1 ) (xo , to ) = (t + x f )ds
to
E x (x , n , z ) ( x)
Z Z
= E x Ay (xt , nt , zt )(dy; nt , zt )dt
o Y
Z Z
+ By(xt , nt , zt )(dy, zt , nt )dpt
o Y
Z Z
= Ay (x, n, z)(dy; n, z)Qo (dx, dn, dz)
Eo Y
Z Z
+ By(x, n, z)(dy; x, z)Qo (dx, dn, dz)
Ed Y
Now Z
E x (x , n , z ) = (x, n, z)QT (dx, dn, dz).
ET
60 2. Optimal Control of pd Processes
\ Z Z
M (S 1 S 2 ) = QT S 1 ET + (dy; n, z)Qo (dx, dn, dz)
T
S1 Eo S 2
Z Z
+ (dy; n, z)Q (dx, dn, dz).
T
S1 E S 2
1. M = Mo + M + MT
where MT m1 (ET ).
M m(E Y)
MO m(Eo Y).
61
R
2. ( xo ) = LdM, C 1 (E).
65 From this point on, the development follows the Vinter-Lewis argu-
ments closely. We reformulate the weak problem as a convex optimiza-
tion problem by incorporating the constrains in the cost function and
obtain the characterization of optimality by studying the dual problem.
The reader is referred to Vermes [25] for the details.
Then
Qo () = h( xo ).
The results outlined above are the first general results on optimal
control of PD processes. Obviously much work remains to be done;
natural next steps would be to determine necessary conditions for op-
timality of Pontrjagin type; to develop computational methods; and to
study optimal stopping and impulse control for PD processes. For
some related work, see van der Duyn Schouton [29], Yushkevich [30]
and Rishel [31].
Part II
Filtering Theory
63
0. Introduction 65
0 Introduction
Suppose {xt } is a signal process which represents the state of a system, 66
but cannot be observed directly. We observe a related process {yt }. Our
aim is to get an expression for the best estimate of xt , given the history
of {yt } upto time t.
In Section 1, we give quick derivations of the Kalman filter for
the linear systems, and nonlinear filtering equations, that of Fujisaki,
Kallianpur and Kunita and Zakais equation for unnormalized condi-
tional density (Kallianpur [19], Davis and Marcus [8]). In section 2,
we will study pathwise solutions of differential equations. In section
3, we will study the Robust theory of filtering as developed by Clark
[5], Davis [10] and Pardoux [20]. Here the above filtering equations are
reduced to quasi-deterministic form and solved separately for each ob-
servation sample path. Also, we will look here into some more general
cases of filtering developed by Kunita [17], where the existence of con-
ditional density functions is proved using methods related the theory of
Stochastic flows.
yt = {Y s , s t}.
d xt = A xt dt + P(t)H dt
xo = 0 (4)
E xt s = E xt s
t
Z
= E g(t, u)dvu s
o
Zt
= g(t, u)du.
o
Hence
d
g(t, s) = E xt s .
ds
68
dt = H xt dt + dWt where xt = xt xt .
Zs
E xt s = E(xt x u ))H du.
o
Now
Zt
xt = (t, u)xu + (t, r)CdVr
u
where is the transition matrix of A. So
Zs
E xt s = (t, u)( xu x u )H du
o
Zs
= (t, u)P(u)H du
o
g(t, s) = (t, s)P(s)H .
So
Zt
xt = (t, s)P(s)H d s .
o
But this is the unique solutions of (4).
Important Points:
(2) In the strict sense version xt is a sufficient statistic for the condi-
tional distribution of xt given (ys , s t), since this distribution is
N( xt , P(t)) and P(t) is nonrandom.
Now show that the Kalman filter gives the same result.
Nonlinear Filtering
Suppose signal xt is a Markov process and observation yt is
given by
dyt = h(xt )dt + dWt ,
generally h is a bounded measurable function (extra smoothness condi-
tion will be added later). Assume that for each t, xt and (Wu Wv ), u, v
t are independent, which allows for the feedback case. Our objective
is to calculate in recursive from the estimates of xt . to do this, it is
necessary to compute the condition of xt given
yy = {ys , s t}.
ZT
E z2s ds < (6)
o
zt = E[zt |yt ].
70
ZT
g2s ds < a.s.
o
and
Zt
Mt = gs dS .
o
y y
This is true even if Ft , Ft , but note that (gs ) is adapted to Ft , not
necessarily to Ft .
73 A General Filtering Formula: Take an Ft -martingale nt , process (t )
satisfying
ZT
E | s |2 ds <
o
Zt
t = o + s ds + nt . (8)
o
1. Linear and Nonlinear Filtering Equations.... 71
t = E[t |yt ].
Proof. Define
Zt
t = t o = s ds
o
Now for t s, h i
E t yt t yt |ys = 0.
So t
Z
E (dz
u u ) z
u u + u u du|y s = 0.
s
Let
V(u) = d
u zu u zu + u u .
This is, Z
V(u)dudP = 0s, t s, A F s .
A[s,t]
Zt
f
Ct = f (xt ) f (xs ) A f (xu )du
s
is a martingale. Suppose
Zt
f
< C , W >t = Z f (xs )ds
o
f
Now apply (9) with t = f (xt ); A f (xs ) = s , Ct = nt and zt = h(xt )
to get Fujisaki-Kallianpur-Kunita filtering formula
Zt Zt
t ( f ) = o ( f ) + s (A f )ds + s (D f ) s (h) s ( f ) dvs (11)
o o
where
D f (x) = Z f (x) + h(x) f (x).
If we interpret t as the conditional distribution of xt given yt , so
that Z
t ( f ) = f (x)t (dx) = E[ f (xt )|yt ],
f
Note that, in general, Ct is a semi-martingale under Po but < ., . > is
invariant under absolutely continuous change of measure. Also if Z = o,
74
Now
Eo [t yt t yt |ys ] = o for t s,
so we get
d t ) := Eo [t h(xt )|yt ].
t = (t)h(x
So (13) becomes ag 78
Z
t = 1 + s s (h)dys (14)
Then
d < Xk f, Bk >t = Xk2 f (xt )dt.
Then Ito version of (16) is
!
1X 2 j
d f (xt ) = Xo + X j f (xt )dt + X j f (xt )dBt .
2
So
1X 2
A = XO + Xj
2
and2
Zt
f j
Ct = X j f (xs )dB s .
o
Proof.
Z
f
d < C ,W > = d < X j f dB j , W >
= X j f d < B j, W >
= j X j f (xt )dt.
So 81
D=Z+h
= i Xi + h.
2
We sometimes use the convention of implied summation over repeated indices.
78
Proof.
D2 f = (i Xi + h)( j X j f + h f )
= i j Xi X j f + i Xi (h f ) + j hX j f + h2 f
= i j Xi X j f + hZ f + Dh f.
1 1X 2 1
A D2 = Y hZ + Xo Dh
2 2 i i 2
1X 2 1
= Y + Yo Dh
2 i i 2
(I ) = 1 ( )2
1||2
0.
P
So 2 t = EUWt , where U = i
i Bi .
||4 t2 EU 2 t = ||2 t2 .
dt ( f ) = t (A f )dt + t (D f )dyt
dt ( f ) = t (L f ) + t (D f )odyt
80
So under Po ,
j j
dVt = dBt + j h(xt )dt
are independent Brownian motions, but V j is not independent of yt , in
fact < V j , y >t = j t.
j j
Now define bt = Vt j yt , then < b j , y >= 0, and this implies b j , y
are independent. But the b j are now not independent. In fact,
j k t for k , j
j k
< b , b >t =
1 ( ) t for k = j
j 2
So
h i
< b j , bk >t = < b j , bk >t
= (I )t.
Let (I ) = as before and define bt = 1 bt .
Then bt = by and < b >t = It. So
d f (xt ) = Xo f (xt )dt + X j f (xt )odBi
= Xo f (xt )dt + X j f (xt )o( j h(xt )dt + j dyt + dbtj )
j
= (Xo f (xt ) hZ f (xt ))dt + Z f (xt )odyt + Y j f (xt )odbt
j
= Yo f (xt )dt + Z f (xt )odyt + Y j f (xt )o d bt
84 where Yo f = Xo f hZ f and Y = X.
The so called Kallianpur- Striebel Formula gives the solution t of
the Zakai equation as a function space integral in the following form,
where xt is functional of (y, b1 , . . . br ).
t
Z Zt
1
t ( f ) = Eo f (xt ) exp h(xs )dys h (xs )ds yt
2
2
o o
xt = (w(t), x).
Define
1
(t Xo ) f (x) = Xo ( f ot1 )(t (x)).
Then the equation for (t) can be expressed as
d 1
f (t ) = (t Xo ) f (t )
dt
o = x,
86 for
1 d
(t Xo ) f (t ) = b(t (x)) f ( 1 (x))|t (x)
dx t
d d
= b(t (x)) f (x) (t1 (x))|t (x)
dx dx
d
= b(t (x)) f (x)( x (x))1 . ()
dx
d 1 d
( ((x))) (x) = 1.
dx dx
When x Rd , then
X
d
f (x)
Xo f (x) = bi (x)
i=1
xi
X
d
f (x)
X1 f (x) = gi (x)
i=1
xi
X
d
1
(t Xo ) f (t ) = bi (x) { f t1 }|t (x)
i=1
xi
2. Pathwise Solutions of Differential Equations 83
X
d X
d
f (x) ( 1 ) jt
= (x).
j=1 i=1
xi xi
So the same results apply for xt Rd , but generally not for more
than one input, i.e., for vector w(t).
Interpretation: xt defined by (2) makes sense for any w(.) C(R+ ).
In particular, if w(t) is a sample path of Brownian motion, then what 87
equation does xt satisfy?
Answer: the Stratonovich equation
Exercise 2.1. Expand xt given by (2) using Itos calculus and show that
it satisfies (3).
The following examples show that the pathwise solution idea cannot
generally be extended to multi-input equations.
x = g1 (x)w1 + g2 (x)w2
x(o) = x.
and
So we must have
g1 g2x = g2 g1x .
88 Define the Lie bracket [X1 , X2 ] = X1 X2 X2 X1 . Now
!
1 d 2df
X1 X2 f = g g
dx dx
g1 g2x fx + g1 g2 fxx .
Therefore
[X1 , X2 ] f = (g1 g2x g2 g1x ) fx .
So a necessary condition for (4) to hold is that
[X1 , X2 ] = 0
i.e., X1 X2 = X2 X1 .
Let i (t, x) be the flow of gi and ti (x) = i (wit , x). Then show that
xt = t1 ot2 o . . . otn
if [X i , X j ] = 0 i, j.
[A, B] = AB BA , 0.
" #
n j1 j
Partition [0, 1] into n equal intervals I j = , , j = 1, 2, . . . , n.
n n
Partition each I nj into four equal intervals I nj,i , i = 1, 2, 3, 4. Define w1,n
to be equal to 4n1/2 for t I nj,1 to 4n1/2 for I nj,2 , and to zero for all other
t. Similarly, let w2,n be equal to 4n1/2 for t I nj,2 to 4n1/2 for t I nj,4 ,
and to zero for all other t.
Then Z t
i,n
w (t) = wi,n (s)ds, i. = 1, 2.
o
Clearly wi,n
converges to zero uniformly as n , i = 1, 2. Let
s=n1/2 , then
1 1
C = A + B + [A, B] + {[[B, A], A] + [[B, A], B]} +
2 12
we get
1
= [B, A] + o(1/n).
n
So 90
xn (1) = en xo
= e([B,A]+o(1/n)) xo
Hence
lim xn = et[B,A] xo .
n t
86
Then
xt = t ot
= (y(t), (t)).
The generator of is
1 X 1 2
At = t
1
Yo + Yj .
2 j t
where D = Z + h.
Notation: For any diffeomorphism : M M, : C (M) C (M)
is given by
f (x) = f o(x) = f ((x)).
So
f (xt ) = t f (t ),
3. Pathwise Solution of the Filter Equation 87
" Z t Z !#
(b) 1
and t ( f ) = E t f (t ) exp s h( s )odys Dh( s )ds
o 2 o
Notation:
gs (x) = H(y(s), x)
yj = t
1
Yj
Z s !
B s f (x) = s f (x) exp u h(x)du .
o
Finally, we get 92
rt = rs ts .
In particular,
T s,t 1 = E s,x [ts ].
93 It is a Markov (or Sub-Markov) semigroup when
ts = t+r
s+r
o r
where r xt = xt+r .
Then
ts = ots o s .
So denoting t = ot , the m. f . property is
t+s = t . s o t .
T t f (x) = E x ( f (xt ) t ).
Zt
f (xt ) f (xo ) A f (xs )ds
o
So equivalently, f D(A ) if
Zt
t ( f (xt ) f (x) s A f (xs )ds
o
So
A f (x) = A f (x) V(x) f (x).
90
a(xt )
Example 3.2. Let t = where a D(A) and a(x) > o x. Then 95
a(xo )
1
T t g(x) = T t (a f )(x).
a(x)
If we define
dPx
= t ,
dP x
then
j j
dwt = dwt + X j g(xt )dt
is a Px - Brownian motion. Thus
X
j
d f (xt ) = (Xo f (xt ) X j g(xt )X j f (xt ))dt + X j f (xt )o dwt .
j
96 So
X 1X 2
A f (x) = Xo X j g(xt )X j f + X f.
j
2 j j
3. Pathwise Solution of the Filter Equation 91
t = t t
with
1 X
V(x) = Ag(x) + (X j g(x))2
2 j
a(x) = eg(x) .
So
1 X
A f (x) = eg A(eg f ) Ag (X j g)2 f (x).
2 j
So
X 1 X
eg A(eg f ) = A f (X j g)X j f Ag + (X j g)2 f.
2 j
97 This factors ts
into product of a Girsanov m. f . and a Feynman-
Kac m. f . Hence the corresponding generator is
X 1 X 1
Y j gt Y j f + (Y j gt )2 A gt t Dh f.
y
At f = At f
j
2 j 2
Proposition 3.1. Ayt f = Byt A 21 D2 B1
yt .
It is easy to see that this must be the right formula. The calculations
y
have been carried out for arbitrary y C[o, T ] but At depends only on
y y
y(t). So At = At where y(s) y(t) (t fixed). Now
t
Z Zt
1
t ( f )(y) = E f (xt )eh(xt )yt exp y(s)dh(xs ) h2 (xs )ds
2
o o
Zt
1
= E f (xt )eh(xt )yt exp y(t)h(xs ) + y(t)h(xo ) h2 (xs )ds
2
o
Zt
exp (y(t)h(x )) 1
= E f (xt )eh(xt )yt . exp h2 (xs )ds .
t
exp(y(t)h(xo )) 2
o
Theorem 3.1.
h f, t i = hByt f, t i (3)
where
Byt f (x) = h f, t i
and t , t satisfy the equations
100 Proof. If L = o, then (4) is the same as (2); so the solution of (4) is
t
Z
< f, t >= f (xt ) exp h(xs ) dys
o
where xt satisfies
d f (xt ) = Z f (xt )odyt .
But this has pathwise solution xt = (yt , x). The previous definition
of B was yt
Z
Byt f (x) = f ((yt , x)) exp h((u, x))du .
o
Now, yt
Z
d h((u, x))du = h((yt , x))odyt .
o
= Byt L Bl
yt Byt f, t > dt+ < Byt D f, t > odyt
by (5) and (4).
=< Byt L f, t > dt+ < Byt D f, t > odyt .
d 1
where | | is the Jacobian of the map V 1 (yt , V).
dV
Proof. If t has a density qt , then
Z
< f, t > = Byt f (z)qt (z, x)dz
Z Zyt
= f ((yt , z)) exp( h((u, z))du)qt (z, x)dz.
o
Now
t1 1 1
[Yi , Y j ] = [t Yi , t Y j ] etc .
with solution
X
< f, t > = exp yi (t)h(x) f (x)
i
3. Pathwise Solution of the Filter Equation 97
Y
= exp(yi (t)h(x)) f (x)
i
xo = x.
So 1
s,t (x) is well defined by (8). Now verify that
Zt
i
hi x1
r,t (p) dyr p=x s,t (x)
s
3. Pathwise Solution of the Filter Equation 99
Zt
= hi (xs,r )odyir .
s
[1] R.M. BLUMENTHAL and R.K. GETOOR, Markov processes and 106
Potential theory, Academic press, New York 1968.
[2] J.M. BISMUT, Martingales, The Malliavin calculus and Hypoel-
licity under general Hormander conditions: Z. Wahrscheinlichkeit-
stheorie 56 (1981) 469-505.
[3] J.M. BISMUT and D. MICHEL, Diffusions conditionelles I, II, J.
Functional Analysis 44 (1981), 174-211, 45 (1982) 274-292.
[4] R. BOEL, P. VARAIYA and E. WONG, Martingales on jump pro-
cesses. I Representation results, SIAM J. Control 13 (1975) 999-
1021.
[5] J.M.C. CLARK, The design of robust approximations to the
Stochastic differentials equations of nonlinear filtering, Communi-
cation Systems and Random Process Theory, ed. J.K. Skwirzynski,
NATO Advanced Study Institute Series, Sijthoff and Noordhoff,
Alphen aan den Rijn, 1978.
[6] M.H.A. DAVIS, The representation of Martingales of jump pro-
cesses, SIAM J. Control and Optimization 14 (1976) 623-638.
[7] M.H.A. DAVIS, Piecewise-deterministic Markov Processes, a gen-
eral class of non-diffusion Stochastic models, to appear in J. Royal
Statist. Soc. (B).
[8] M.H.A. DAVIS and S.I. MARCUS, An Introduction to nonlin-
ear filtering, Stochastic Systems: The Mathematics of Filtering
101
102 BIBLIOGRAPHY
[26] R.B. VINTER and R.M. LEWIS, A necessary and sufficient con-
dition for optimality of Dynamic Programming type, Making no
apriori assumption on the controls, SIAM J. Control and Optimiza-
tion 16 (1978) 571-583.
[29] F.A.VAN DER DUYA SCHOUTON, Markov Decision Drift Pro- 109
cesses, Mathematics Centre Tract, Amsterdam 1983.
104 BIBLIOGRAPHY