This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
0, the set {x £ E : \f{x)\ > e} is a compact subset of E. Equipped with the supremum norm, Co is a Banach space. Since E is second countable, there exists a countable collection C of open relatively compact subsets of E such that every open subset O of E is a countable union of open sets from C. Let C = {Uj : j > 1} be an enumeration of C and choose a,j G Uj, j > 1. Let ifj G Co, j > 1 be a sequence of continuous functions on E such that 0 < ifij < 1 and ) G O, and the time shift operators on the space CI are given by ds( [x T 1
t > S} belongs to the cr-algebra a(S,Xs)Va(Xp: u for all u G [0, T]. Let A1 = Ai(r) = { G <&}. We will first prove that any random variable ip(S) G $ is a ^-stopping time. Then we will establish that for all Si G M and S2 E M with r < Si < S2 < T, the stopping time S2 is measurable with respect to the ex-algebra a (Si) C Q^1 ' v . Let us first notice that if tp G $, then the inverse function ip_1 exists on the interval [ip(r),T] and maps [cp(r),T] onto [r, «(?)]. Moreover, tp-1 (tp(t)) = t (ifi U if 2 ) + II ' llm, associated with the function l. (2) For any a with 0 < a < 1, the time-dependent measure A 7 (a, /?) belongs to the class Vm,ip associated with the function (p(r, t) = (t — r ) ° if and only if either a + a + ^ < 1, and 0 < / 3 < o o , o r 0 < d + 2 a - < 5 < 2 , a + a+ ^ = 1, and /3 > 0.
for all j > 1. It is not hard to see that the function p : E x E —> [0,1] denned by oo
p (x, y) = J2 2~j 1 ^ 0 ) - Vi(v)\.
(x,y)£Ex
E,
j=i
is a metric on the space E. This metric generates the topology of E. Hence, the space E is metrizable. It is also clear that E is a cr-compact space. 1.1.4
Stochastic
Processes
Let (Q, f, P) be a probability space, and let (E, £) be a state space. Suppose that Xs : Cl H-> E, S G /, is a family of random variables where I is an index set. Then the family Xs, s G J, is called a stochastic process. Throughout the book, we will use a bounded interval [a, b] or the halfline [0, oo) as the index set J, unless specified otherwise. The stochastic
6
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
process Xs, s £ I, can be identified with the mapping X : (s, w) H-> XS(LJ), (s, u>) £ I x £1. In addition, the process Xs can be considered as a random variable X : 0 -* E1 denned by X(w) = (X a (w)) s 6 / . For every u £ fi, the function s — i > X s (w), s £ /, is called a sample path. The path space E1 is equipped with the product cr-algebra ®/£. This u-algebra is generated by finite-dimensional cylinders of the form Ylsei E*, where Es £ £, s £ I, and Es ^= E for only a finite number of elements s £ I. Let Xs, s £ I, be a stochastic process, and let J = ( s i , . . . , sn) be a finite subset of / . The finite-dimensional distribution of the process Xs corresponding to the set J is the distribution of the random variable Xj : 0, —• EJ given by Xj(u) = (X3l(u),.. .,XSn(u)). It is defined as follows: »J(B)
= F[XJ€B],
B£BEJ.
Here BEJ denotes the Borel er-algebra of the space EJ. The distribution of the process Xs, s £ I, that is, the distribution of the random variable X is determined by the family of all finite-dimensional distributions. Each member of the family {(EJ,BEJ,(IJ)
:JCI,
J finite}
is a probability space, and the family {/jy : J finite} is a projective (or consistent) system of probability measures. This means that for finite subsets J and K oi I with J C K, the equality HK[(PK)-X{B)\=HJ{B)
holds for all B £ BEJ. given by
Here the projection mapping p^j : EK —> EJ is
PJ ( K W ) =* (<"«)sej -
(U').eK
G
E
*
(see Section 5.2). Let X] and X^, s £ I, be stochastic processes on the probability spaces (fii, Ti, Pi) and {Q.2, -^2, P2), respectively, and suppose that both processes have the same state space (E,£). The processes X^ and Xf are called stochastically equivalent if their finite-dimensional distributions coincide, i.e., if the equality Pi [X*j £B]=P2
[Xj £ B]
Transition Functions and Markov
Processes
7
holds for all finite subsets J of 7 and all B £ BEJ- Note that in the definition of the stochastic equivalence, it is not necessary to assume that the processes X^ and X% are defined on the same probability space. If the processes X] and X% are denned on the same probability space (fi, J7, P), then it is said that the process X% is a modification of the process X\ if for every s £ 7, P [X\ / Xf\ = 0. The processes X} and X% are called indistinguishable if P [X} ^ X% ] = 0 for all s £ I. It is not hard to see that if XJ and X^1 are continuous stochastic processes and X% is a modification of XI, then X} and X^ are indistinguishable (see [Revuz and Yor (1991)], p. 18).
1.1.5
Filtrations
Let (fljJ7) be a measurable space. A filtration Qt, t £ [0,T], is a family of sub-cr-algebras of T such that if 0 < t\ < £2 < T, then Qtl C Gt2- A two-parameter filtration QJ, 0 < r < £ < T, is a family of sub-cr-algebras of T that is increasing in t and decreasing in T. A stochastic process X s , s G [0,T], with state space {E,£) is called adapted to the nitration Qt, t G [0,T], provided that for every s G [0,T], X , is Ts/£-measurable. The following filtration is associated with the process Xs: Ft=a{Xs:Q<s
0
Recall that a (Xs : 0 < s < t) is the smallest cr-algebra containing all sets of the form f]s€J X~* (Bs), where J is any finite subset of [0, t], and Bs G £, s G J. Instead of Borel subsets of E, one can take, e.g., open subsets, or any other 7r-system which generates the cr-algebra £ (see Subsection 5.1 for the definition of a 7r-system). The cr-algebra a (Xs : 0 < s < t) is the smallest cr-algebra such that all the state variables X3 with 0 < s < t are measurable. The process Xs, s £ [0, T], generates the following two-parameter filtration: Tl =a(Xs:r
<s
0
The filtration T[ is sometimes called the internal history of the process Xs, s£[0,T].
Let (ft, T, P) be a probability space, and let Gt be a nitration. It will be often implicitly assumed that for every t £ [0,T], the cr-algebra Qt is complete with respect to the measure P. This means that if A £ J7, B C A, and F(A) — 0, then B £ Qt. In other words, all subsets of P-negligible sets belong to every cr-algebra Qt with t £ [0, T). If the filtration Qt is not
8
Non-Autonomous Kato Classes and Feynman-Kac Propagators
complete, then we can always augment it by the family M = {B C n : B C A, AeJ7,
F(A) = 0} .
More precisely, the augmentaion means that one passes from the cr-algebra Gt to the cr-algebra Qt = a {Gt,N). We refer the reader to Section 1.7 for more information on completions of u-algebras. 1.2
Markov Property
In this section we introduce Markov processes and formulate several equivalent conditions for the validity of the Markov property for a general stochastic process. Let (fi, T, P) be a probability space, and let Xt with t G [0, T] be a stochastic process on Q with state space E. Recall that the state space E is a locally compact Hausdorff topological space satisfying the second axiom of countability. Definition 1.1
A stochastic process Xt is called a Markov process if E[f(Xt)\F3]=E[f(Xt)\a(Xs)}
(1.1)
P-almost surely for all 0 < s < t < T and all bounded Borel functions / on E. The next lemma is well-known. It provides several equivalent descriptions of the Markov property. For the definition of two-parameter filtrations see Subsection 1.1.5. Lemma 1.2 Let Xa be a stochastic process on (fi,.F, P) with state space E. Then the following are equivalent: (1) Condition (1.1) holds. (2) For all t € [0, T], all finite sets { n , . . . , rn} with 0 < r\ < r 2 < • • • < r„ < t, and all bounded Borel functions f on E, the equality E [f(Xt)
\a(Xri,...,Xrn)]=E
[f(Xt)
| a(Xrn)]
holds P-a.s. (3) For all s G [0,T] and all bounded real-valued ^-measurable variables F, the equality E [F | Ts] - E [F | a(Xs)] holds F-a.s.
random
Transition Functions and Markov Processes
9
(4) For all s £ [0, T], and all bounded real-valued random variables G and F such that G is Fs-measurable and F is ^-measurable, the equality E[GF]=E[GE[F\a(Xs)]] holds. (5) For all s e [0,T], A € Ts, and B € T^, the equality P [A n B | a(Xs)} =F[A\
a(Xs)} P [B \ a(Xs)}
holds F-a.s. We refer the reader to [Blumenthal and Getoor (1968)] for the proof of Lemma 1.2. A similar lemma concerning the reciprocal Markov property will be established in Section 1.10 (see Lemma 1.20). Condition (5) in Lemma 1.2 states that for a Markov process the future and the past are conditionally independent, given the present. The aalgebra !FS is often interpreted as information from the past before time s, while the cr-algebra Fj, contains the future information. The time s is considered as the present time. In a sense, a Markov process forgets its past history. R e m a r k 1.1 If Xt, 0 < t < T, is a Markov process with respect to the probability measure P, then the time reversed process X1 = Xr-t, t E [0,T], is also a Markov process with respect to the same measure P. Indeed, it is not hard to see that condition (5) in Lemma 1.2 is invariant with respect to time-reversal. Therefore, Lemma 1.2 implies that the timereversed process Xt possesses the Markov property. The next assertion follows from Lemma 1.2: Lemma 1.3 Let Xs be a stochastic process on (O,^ 7 , P) with state space E. Then the following are equivalent: (1) For all 0 < s < t < T, and all bounded Borel functions f on E, the equality E[f(Xs)\^]=E[f(X3)\cr(Xt)] holds P-o.s. (2) For all t with 0 < t < T, all finite sets {u\,..., un} with t < u± < ti 2 < • • • < un < T, and all bounded Borel functions f on E, the equality E [/(X t ) | a(XUl,...,XUn)]
= E [f(Xt)
| a(XUl)}
10
Non-Autonomous Koto Classes and Feynman-Kac Propagators
holds P-a.s. (3) For all s with s 6 [0,T] and all bounded real-valued Ts-measurable random variables F, the equality E[F\J%\
=
E[F\a(Xs)]
holds P-a.s. (4) For all s with s € [0, T], the equality E[FG\ =
E[FE[G\
holds for all bounded real-valued random variables G and F such that G is Fs-measurable and F is J-^-measurable. (5) For all s £ [0, T], A e Ts, and B <E T?, the equality P [A n B | a(Xs)] =P[A\
a(Xs)] P [B \ a(Xs)}
holds F-a.s. 1.3
Transition Functions and Backward Transition Functions
This section is devoted to non-homogeneous transition functions. We will first introduce a forward transition probability function, or simply a transition probability function. Definition 1.2 A transition probability function P{r, x; s, A), where 0 < r < s < T, x 6 E, and A G £, is a nonnegative function for which the following conditions hold: (1) For fixed r, E. (2) For fixed r, (3) P(r, x; s, E) (4) P(r, x; s, A)
s, and A, P is a nonnegative Borel measurable function on s, and x, P is a Borel measure on £. — 1 for all r, s, and x. = JE P(r, x; u, dy)P(u, y; s, A) for all r < u < s, and A.
A function P satisfying Condition (3) in Definition 1.2 is called normal, or conservative. Condition (4) is the Chapman-Kolmogorov equation for transition functions. In applications, a transition function P describes the time evolution of a random system. The number P(r, x; s, A) can be interpreted as the probability of the following event: The random system located at x € E at time r hits the target A C E at time s.
Transition Functions and Markov
Processes
11
The next definition concerns backward transition probability functions. Definition 1.3 A backward transition probability function P(T, A; t, y), where 0
For fixed r , A, and t, P is a Borel function on E. For fixed r , t, and y, P is a Borel measure on £. The normality condition P(T, E; t, y) — 1 holds for all T, t, and y. The Chapman-Kolmogorov equation, that is, P(T, A; t,y)= f P{T, A; A,x)P(X,dx; t,y), JE
holds for all r < A < t, A G £, and y £ E. There is a simple relation between forward and backward transition probability functions in the case of a finite time-interval [0,T]. Here we need the time reversal operation 11-> T — t, t € [0, T]. It is easy to see that P is a backward transition probability function if and only if P{T,x;t,A) = P(T-t,A;t-T,x)
(1.2)
is a transition probability function. If a function P satisfies conditions (1), (2), and (4) in Definition 1.2, but does not satisfy the normality condition, then it is called a transition function. Similarly, a function P satisfying conditions (1), (2), and (4) in Definition 1.3 is called a backward transition function. If the condition P(T,x;t,E)
(1.3)
holds instead of condition (3) in Definition 1.2, then P is called a transition subprobability function. Similarly, if P(T,x;t,E)
(1.4)
then P is called a backward transition subprobability function. If P is such that (1.3) holds, then one can define a new state space EA = E U {A} where A is an extra point. If E is a compact space, then A is attached to E as an isolated point. If E is not compact, then the topology of E is that of a one-point compactification of E. The Borel cr-algebra of EA will
12
N'on-Autonomous
be denoted by £A.
PA(r,x;t,A)
Koto Classes and Feynman-Kac
Propagators
Put
= <
1,
ifx = A a n d A e A .
0,
if x = A and A £ .4
P(r,x;t,A),
if x € £ and A £ A
l-P(T,x;t,£),
if x € £ and A = {A}.
(1.5)
L e m m a 1.4 Lei P be a transition function satisfying (1.3), and let PA be defined by (1.5). Then PA is a transition probability function; moreover, the functions P A and P coincide on E.
Proof. It is clear that only conditions (3) and (4) in Definition 1.2 need to be checked for the function P A . For x £ E, (1.5) implies
P A (T, X; t, EA) = P(T, X; t, E) + P A (r, x; t, A) = P ( T , x ; t , E ) + l - P(T,x;t,E)
= \.
If x = A, then (1.5) gives P A (T, A ; i , £ A ) = 1. Therefore, the function P A is normal. Our next goal is to prove that the function P A satisfies the ChapmanKolmogorov equation. For x ^ A and A £ £, this fact follows from the Chapman-Kolmogorov equation for P . Let x € EA, T < r < t, A e £A, AG A, and put A = A\A. Then
PA(T,x;r,dz)PA(r,z;t,A)
J
= I PA(r,x;r,dz)PA(r,z;t,A)
+ f
JE
+ [ J{A}
+ /
PA(T,x;r,dz)PA(r,z;t,{A})
JE A
A
P (T,x;r,dz)P (r,z;t,A) V
PA(T,x;r,dz)PA(r,z;t,{A}).
/
(1.6)
Transition Functions and Markov
Processes
13
If x £ E, then (1.6), (1.5), and the Chapman-Kolmogorov equation for P give PA(T,x;r,dz)PA(r,z;t,A)
J =
P{T, X;
t,A)+
[ P{r,
X;
r, dz){\ - P(r, z; t, E)) + l -
P{T, X;
r, E)
JE
= P{T, x;t,A) + l - P(T, X; t, E) =
PA(T,
x-1, A) + P A ( r , x; t, {A}) =
PA{T,
X;
t, A).
(1.7)
It is easy to see that if x £ E and A £ A, then Lemma 1.4 follows from (1.7). A similar reasoning can be used in the case where x = A and A £ A, and in the case where x = A and A ^ A. • A nonnegative Borel measure m on (E, £) will be fixed throughout the book. The measure m is called the reference measure. We will write dx instead of m(dx) and assume that 0 < m(A) < oo for any compact subset A of E with nonempty interior. It is said that a transition function P possesses a density p, provided that there exists a nonnegative function p(r, x; s, y) such that for all 0 < r < s < T, the function (x, y) —> p (r, x; s, y) is £ <8> ^-measurable, and the condition P(r,x;s,A)=
/
p(r,x;s,y)dy
holds for all A € £. The Chapman-Kolmogorov equation for transition densities is P(T, X; t,y)=
/ p(r, X; r, z)p(r, z; t, y)dz JE
for all 0 < r < r < t and m x m almost all (x, y) £ E x E.
1.4
Markov Processes Associated with Transition Functions
Let E be a locally compact space such as in Section 1.1, and let P be a transition probability function. Our first goal is to construct a filtered measurable space (Q, J-, TJ), a family of probability measures PT,x, x £ E, T € [0,T], on the space (Q,^), and a Markov process Xt, t £ [0, T], such
14
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
that T[ = cr(Xs:T<s
0 < T
(1.8)
and Pr,x(X t G A) = P{T, x;t,A),
0
Ae £.
(1.9)
We begin with the construction of what is called the standard realization of such a process. Let f2 = J5' 0,T ' be the path space consisting of all functions mapping [0, T] into E. The space Q, is equipped with the cylinder cr-algebra T, which is the smallest a-algebra containing all sets of the form {UJ G Q :u(ti) G ^ i , . . . , ^ ^ ) G Ak}, 0 < ti <•••< tk < T and Ai e £ for all 1 < i < k. Such sets are called finite-dimensional cylinders. The process Xt is defined on the space Q. by Xt(u>) = u>(t). The cr-algebra T[ is defined by formula (1.8). The symbol PT,x, where 0 < T < T and x S E, stands for the probability measure on T^ such that P r , x (XT e A0, Xtl e Alt • • • , Xtk € Ak) = PT,X (w e fl: w(r) £ A0,iv(ti)
€AU---,
= XA0{x) /
P(ti,xr,t2,dx2)
P(T,X;tudxx)
J Ax
/
/
u(tk) G Ak) •••
JA2
P(tfc_i,Xfc_i;tfe,dxfc)
(1.10)
JAk
for all r < t\ < t2 < • • • < tk < T and Ai G £, 1 < i < k. Such a measure exists by Kolmogorov's extension theorem (see Appendix A). The process Xt is a Markov process with respect to the family of measures P r , x , that is, Er,x [f(Xt)\J^]
= Ea,x.f{Xt)
PT,x-a.s.
(1.11)
for all T < s < t, and all bounded Borel functions / on E. Recall that the symbol E T:X in (1.11) stands for the expectation with respect to the measure P T)X , and the left-hand side of (1.11) is the conditional expectation of f(Xt) with respect to the tr-algebra TTS. Taking the conditional expectation with respect to the cr-algebra o (Xs) on both sides of equality (1.11) and using the a (X s )-measurability of the expression on the right-hand side of (1.11), we see that condition (1.1) holds for all measures P r , x and all t, s with T < s < t < T. In a sense, the Markov property in (1.11) means that the past history of the process Xt does not affect predictions concerning the future of Xt. It is not hard to see that condition (1.9) holds for the process Xt. The expression on the left-hand side of (1.9) is called the marginal
Transition Functions and Markov
Processes
15
distribution of Xt, while the expression on the right-hand side of (1.10) is called the finite-dimensional distribution of Xt. It is not hard to see that for a Markov process, the finite-dimensional distributions can be recovered from the marginal ones. Next, we discuss general non-homogeneous stochastic processes (Xt,Gt^-r,x) defined on a general sample space (£1, Q) equipped with a two-parameter filtration Gl,0
PTliE-a.S.
(1.12)
for all r < s < t, and all bounded Borel functions / on E. In the definition of the standard realization of the process Xt, the distribution of the state variable XT with respect to the measure P TiX is Sx, the Dirac measure at x. Let / i b e a probability measure on (E,£). Then, replacing (1.10) with P r , M (u; efl : w(r) e A ) , ^ ) eAu--= j
dfi(x) /
J AQ
P(T,X;ti,dxi)
JAI
I P(tk-i,xk-i;tk,dxk),
/
,u(tk-i) P(ti,xi;t2,dx2)
£ Afc_i,a>(*fc) € Ak) •••
J A2
(1-13)
JAk
we get a stochastic process with the initial distribution at t — T equal to fi. Any two Markov processes associated with the same transition function P are called stochastically equivalent. Such processes may be defined on different sample spaces. Next, we will consider backward transition probability functions. Let P be such a function, and let t € (0, T] and y G E. Define a family of finitedimensional distributions on the path space Q = E^°'T^ equipped with the
16
Non-Autonomous Koto Classes and Feynman-Kac Propagators
cylinder a-algebra T by the following formula: ^
y
(u G ft : u){ti) G Alt • • • ,w(**-i) G i4fc_i,w(tfc) G Ak,w(t) P ( t i , da;i; £2, x2) • • • P(tk-i,dxk-i;
/
G A fc+1 )
tk, xk) (1.14)
P(tk,dxk;t,y)xAk+1{y),
where 0 < * i < t 2 < - - < t f e < t < T ' and At £ £, 1 < i < k+ 1. Here we use Kolmogorov's extension theorem to establish the existence of the measure P t , ! / . For all t G [0,T] and y £ E, the measure P*'y is defined on the <7-algebra J-^. Consider the standard realization Xs (u>) = u (s) on the path space (ft,.?-"). Then the process Xs has P as its backward transition function, that is, P{T,A;t,y)
=
Ft>*[XreA]
for all 0 < T < t < T and A G £. Moreover, the backward Markov property holds for Xs. This means that £*•» [/ (XT) | Ff] = E'-" [/ (X T ) | a (X s )] = E s - X »/ (XT) P ^ - a . s . for all bounded Borel functions / and all 0 < r < s < t < T. The process Xs has the terminal distribution at s = t equal to the Dirac measure Sy. In the case of a prescribed terminal distribution v at s = t, the finite-dimensional distributions satisfy P'-" (LOG CI: u(ti) GAi,---, /
P(ti,dxi\t2,
oj(tk-i) x2)---
GAk-i, w(tk) G Ak,w{t) G P(tk-i,
Ak+l)
dxk-i; tk, xk)
J' A i X -- xx A A jt ++ i
P(tk,dxk;t,y)du(y).
(1.15)
Since the relation described in (1.2) is a one-to-one correspondence between forward and backward transition functions, a backward transition probability function P generates a backward Markov process X1, 0 < t < T, with respect to the family of cr-algebras J-\ = a (Xr : t > r > T) = J-TZl and the family of measures P*'1. This means that Et,x
/(x T ) T\
= E*'X
[/(^)
Xs
E's.X'
/ (*T)
T)t,X
a.s.
for all 0 < T < s < t < T and all bounded Borel functions / on E. It is easy to see that the backward Markov property for the time reversed
Transition Functions and Markov Processes
17
process Xs — XT-S follows from the Markov property for the process Xt. Therefore, if Xt is a Markov process with transition function P, then the process Xr-t is a backward Markov process with backward transition function P. In general, any property of Markov processes can be reformulated for backward Markov processes. We simply let time run backward from T to 0. For instance, if P is a backward transition function satisfying the subnormality condition, then we can use the construction in Section 1.3 to extend P to a backward transition probability function on the space EA.
1.5
Space-Time Processes
A transition function P is called time-homogeneous if P(T,X;t,
A) = P((T + h)AT,x;
{t + h)AT, A)
(1.16)
for all 0 < T < t < T, h > 0, x e E, and A E £. The values of a time-homogeneous transition function depend only on the time span t — r between the initial and final moments. We will write P(t—r, x, A) instead of P(T, X; t, A) in the case of a time-homogeneous transition function P. The minimum (T + h) A T appears in formula (1.16) because the parameters T and t vary in a bounded interval. Definition 1.5 Let (CtjJ7) be a measurable space. A family of measurable mappings fls : Cl —> Cl, s > 0, such that i W . = 0t+., #o = I
(1-17)
for all t > 0 and s > 0, is called a family of time shift operators. The symbol / i n (1.17) stands for the identity mapping on CI. If a stochastic process Xt with state space (E, £) is given on a probability space (CI, P), then it is natural to expect the process Xt to be related to the time shift operators 6S as follows: Xt o # s = X(t+s)/\T
(1-18)
for all t £ [0,T] and s € [0, T]. Equality (1.18) is often used in the theory of time-homogeneous stochastic processes. For some sample spaces, it is clear how to define the family of time shift operators {"ds}. For instance, if Cl = E$
18
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
Let P be a transition probability function, not necessarily timehomogeneous, and let (Xt,J-[,FT>x) be a Markov process on (Q,F) associated with P. Our next goal is to define the space-time process Xt = (t, Xt), t G [0, T]. It is clear that the state space of this process must be the space [0, T] x E. However, it is not clear what is an optimal choice of the sample space for the space-time process and what is the transition function of the process Xt. It is natural to expect that the sample space for the process Xt has to be the space [0, T] x CI. We will show below that this is the case under certain restrictions. Let B be a Borel set in [0, T] x E. For every s G [0,T], the symbol (B)s will stand for the slice of the set B at level s, defined by (B)s = {x G E : {s,x) G B}. Consider the following function P (t, (r, x), B) = P (r, x- (r +1) A T, (B)( T + t ) A T ) , where (T,X) G [0,T] x E, t G [0,T], and B G be rewritten in the following form: P (t, (r, x), dsdy) = P{T, X; S,
B[ 0 ,T]X.E-
(1.19)
Equality (1.19) can
dy)d5{T+t)AT(s),
where 6 is the Dirac measure. If the transition probability function P has a density p, then formula (1.19) becomes P (t, (r, x), B)= f
P{T,
X; (T
+ t) A T, y ) ^
-' ( B ) ( T + I ) A T
= 11
P(i~,x;s,y)dyd6(T+t)AT(s).
(1.20)
The time variable in the definition of P is £ and the space variables are {T,X) G [0,T] x £ and B G £[0,r]x,ETheorem 1.1 The function P given by (1.19) is a time-homogeneous transition probability function. Proof. It follows from (1.19) that the function P is normal. Next, we will show that the function P satisfies the Chapman-Kolmogorov equation, which in the case of time-homogeneous transition functions has the following form: P ( ( t i + t 2 ) A T , ( T , a ; ) , B ) = [[
P(t1,(T,x),dsdy)P(t2,{s,y),B)
Jj[0,T]xE
(1.21)
Transition Functions and Markov Processes
19
for all (T,X) € [0,T] x E, B E B[0>T]xE, 0 < h < T, and 0 < t2 < T. By using the Chapman-Kolmogorov equation for P and equality (1.19) twice, we see that P ((ti + t2) A T, (r, x),B) = P (r, i ; (r + h +12) A T, ( B ) ( T + t l + t a ) A T ) = / P(T,x;(r
+
ti)AT,dy)
JE
P =
((T
/ /
+ ti) A T, y; „.
(T
+ ti +1 2 ) A T, (B) ( _p(T>x''s'dy)p(s'y''(s+t2)hT,(B)is+t2)AT)ds{T+tl)AT(s)
[0,T]xE
J he[o,r)xB
P (ti, (r,x),dsdy)P(t2,
(s,y),B).
This establishes (1.21). Therefore, the function P defined by (1.19) is a transition probability function. • Note that even if the transition function P has a density p, the measure B — i > P(t,(T,x),B) on B[otx]xE is singular with respect to the measure dtdm. The function P will play the role of the transition function of the space-time process Xt. There are several possible choices of the sample space fl for the spacetime process Xt- For instance, one can choose the full path space, that is, the space n = ([0,T\xE)[o'Ti, to be the sample space of the space-time process Xt. space-time process is defined by
(1.22) In this case, the
Xt(
(1.23)
Since P is a transition probability function, the Kolmogorov extension theorem implies the existence of a family of measures P( r , x ) indexed by
20
Non-Autonomous Kato Classes and Feynman-Kac Propagators
(T, X) e [0, T]x E such that P(T,X)
Xti
G Ai,
• • • , Xtk
G Ak
= P(T,x) [fa (*i), w («i)) G A 1 ; • • • , (^ (tk), u (t fe )) G Ak] • r,x
^(T+tOAT G ( ^ ^ ( ( r + t ^ A T ) ' • • • '
^(r+tfc)AT G (^fc)v((r+tfc)AT)J
(L24)
for all 0 < ti < i 2 < • • • < tk < T and A{ G S[ 0 ,T]X£I 1 < i < k. This construction results in a time-homogeneous Markov process ( Xt,^ r t T ,P( T x ) j on the space Q,. The process Xt is our first version of the space-time process. However, it may happen so that the first component t H-» tp(t) of an element (
n,x (K1 (n*)) = i-
(i-25)
In formula (1.25), the symbol P * x stands for the outer measure on the space £,!°'T1 generated by the measure PT,X> {^T} denotes the family of time shift operators on 2?[°'T1 given by i?T(w)(t) = u>((r + t) A T), and I?~ 1 (J4) denotes the inverse image of the set A under the mapping i9T. Let fi* be the subspace of the space Q. consisting of all pairs (^)c,w) £ Q, such that
T/ie following equality holds for all (T,X) G [0,T] X £ : P^ TIX)
(n*) = i.
(i.26)
Transition
Functions and Markov
Processes
21
Proof. Given 0
{xtleAu---,xtkeAk}
= {(
.
(1.27)
For every T G [0,T], (1.27) gives d-1(C)={(ip,u):w((T
+
h)AT)e(A1)v{{T+ti)AT),
• • • , w ((r + tfc) A T) G ( A 0 „ ( ( T + t j k ) A r ) } .
(1.28)
Let 7To : £,[°,T1 H-> fi be the following mapping: 7To(w) = (<£o,w) where w G -Bl°'Tl and
= [w : w ((r + *0 A T) G M 0 ( T + t l ) A T , •••,W((T + tfe)AT)G(^0(.+tfc)Ar}-
(1-29)
Suppose that the set O* is covered by a countable family {Cj} of cylinders such as in (1.27). Then the family of cylinders WQ1 \$~x(Ci)\
>
defined in (1.29) covers the set 7r0-
1
(^- 1 (n*))=^.
Using (1.24) and (1.25), we get OO
OO
£P ( r ,z) (d) = £ P r , * ( V ($7\Ci))) > 1.
(1.30)
1=1
Now it is clear that equality (1.26) follows from (1.30). This completes the proof of Lemma 1.5.
•
It follows from equality (1.26) that one can restrict the probability space structure from (fl, J ^ P ^ x ) ) t o the set fi*. The resulting probability space will be denoted by f fi*,^j.,P(T]X) J. Summarizing what has already been accomplished, we see that the space-time process Xt can be defined on the probability space f fi*,^-j.,P(T>:t:) J as follows: Xt(
22
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
Here w e ft*, c > 0, and <^c is the function on [0,T] given by <£c(£) = (c + t) A T. Using the identification j of the spaces ft* and [0, T] x ft*, we see that the space-time process Xt can be defined on the space [0, T) x ft* by
X,(c)w) = ((c+t)AT ) w(t)). The state space of this process is the space ([0,T] x E, Kfcrjxs)The space-time process Xt can also be defined for a general Markov process {Xt,J-t,WT,x) o n t n e sample space ft provided that there exists a family {fis} of time shift operators on ft satisfying condition (1.18). In this case, the space-time process Xt is defined on the sample space ft = {fc}o
(vc(t),Xt(uj)).
The family of time shift operators on the space ft is given by ds {
It is not difficult to check that Xt o dT = X( T + t ) A T . As before, we can identify the space ft with the space [0, T] x ft, by using the mapping j : ft —> [0,T] x ft defined by j(
Xt(c,w) = ((c + i ) A T , X t H ) , and the time shift operators # s can be written as follows: 0,(c,w) = ((c + s ) A T , 0 , ( w ) ) . The state space of the space-time process Xt is [0, T] x E. Our next goal is to define the measure P(T,x) o n the er-algebra Tl. For the cylinders C defined in (1.27), we can use formula (1.24). For a general set A € F^, we extend this formula to P(r,x)(^)=Pr,x(7ro1(^1(A))),
where n0 : ft —> ft is defined by TTQ{UJ) = {(j>o,ui). Note that the time shift operator $T is ^ / L A / ^ /.Fj?-measurable. Moreover, since the equality Xt o n0 = (t, Xt) holds, the mapping TT0 is TIJT%measurable. It also follows that for any bounded .^-measurable random
Transition Functions
and Markov
Processes
23
variable F, E{TtX)[F] = Er,x
F O fiT O
7TQ
The space-time process Xt is a Markov process. The Markov property of the process Xt can be formulated as follows: E.(T,X) / [X{t+s)AT)
I ?% = E( S) x.) /
[Xtj
for all t G [0, T], s G [0, T], x G E, and all bounded Borel functions / on the space [0, T] x E. An equivalent formulation of the Markov property of the process Xt is E(T,*) [/ ( X t ) O #8 \F.
s,X.) f
\Xt)
Remark 1.2 The space-time process Xt associated with the given Markov process Xt is not simply the process t \-> (t,Xt). For instance, if the space-time process Xt is defined on the sample space [0, T] x 0, then Xt(c,ij)
= ((c +
t)AT,Xt(Lj))
where t G [0, T] and (c, u) G [0, T] x O. For c = 0, we get Xt (0, w) = (t,Jf t (w)). Our next goal is to discuss space-time processes associated with backward transition probability functions. Let P(T, A; t, x) be such a function, and put P(T, (i, x), B) = P ((t - T) V 0, (B) (t _ T) vo; t, x).
(1.31)
Here r G [0, T] plays the role of the time-variable, whereas (t, x) G [0, T]xE and 2? G B[O,T]XB a r e the space variables. The equality in (1.31) can be rewritten as follows: P(T, (t, x), dsdy) = P (s, dy; t, x) eW(t_T)Vo(s).
(1.32)
Theorem 1.2 Let P(T, A; t, x) be a backward transition probability function. Then the function P defined by (1-31) is a time-homogeneous transition probability function. Proof. It follows from equality (1.31) that P is normal. Next let T\ G [0, T], r 2 G [0, T], {t, x) G [0, T] xE, and B G B [0 ,T]XJ5- Then, using formulas
24
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
(1.31), (1-32), and the Chapman-Kolmogorov equation for the function P, we obtain P((r1+T2)AT,(t,x),B) = P((t-
(TI + r 2 ) A T) V 0, ( B ) ( t _ ( T l + T 2 ) A T ) v 0 ; t, x)
= [ P((t-
( n + r 2 ) A T) V 0, ( £ ) ( t _ ( r i + T . 2 ) A T ) v o ; (t - n ) V 0,2/)
P((t-Ti)V0,d»;t,x) = / / J
P((s-T2)V0,(B)(s-T2)vo;s,y)P(s,dy;t,x)dd{t_Tl)s,Q(s)
J[0,T]xE
-a
P{TU(t,x),dsdy)P(T2,{s,y),B).
(1.33)
l[0,T]xE
In (1.33), we used the equality
((* - n) v o - n) v o = (t - (n + r2) A T) V O. It is clear that (1.33) implies the Chapman-Kolmogorov equation for P. This completes the proof of Theorem 1.2. • Theorem 1.2 allows us to define the space-time process XT associated with the backward transition probability function P. Arguing as in the case of transition probability functions, we first choose the space Q defined by (1.22) as the sample space of the space-time process XT(ip,uj) = (
GA1,---,XTkGAk\
= p".*) [(
(
= P*'x (^ ( t _ r i )vo G (^i) v ((t-Ti)vo) - ' ' ' >^(t-rfc)v0 G (-^fc)v,((t-Tit)vo)J • (1.34) In (1.34), P t , x is the family of measures denned by (1.14). The Markov process (XT, FT, p(*'x) J on the sample space fi is our first version of the spacetime process associated with the backward transition probability function P. The following family of time shift operators on Cl can be used in this
Transition Functions and Markov
Processes
25
case: ?.(^W) = M(T-*)VO)IW((T-J)VO)),
S>0.
(1.35)
The operators -d3 are backward shifts by s with respect to the time variable T. The time shift operators ds are connected with the space-time process XT as follows: Xrotfs = X(T_s)v0
(1.36)
for all s > 0 and r G [0, T], It is also possible to define space-time processes on smaller sample spaces as it has already been done in the case of transition probability functions. For instance, let P be a backward transition probability function, and let (XT, F[, P*'x J be a backward Markov process defined on the sample space Q and with P as its transition function. Assume that there exists a family of time shift operators i?5 on the sample space Cl such that XT o d„ = X ( r _ a ) v 0 for all s > 0 and r G [0,T]. Then the space-time process
(1.37) (xT,TT,P{t'xA
can be defined on the sample space [0, T] x fl by Xr{c,w)=
((T-C)V0,XT(C))
where (c, w) G [0, T]xCl and r G [0, T]. The time shift operators ^ s on the sample space [0, T] x f2 of the space-time process are defined by ?a(c,w)= ( ( T - C ) V 0,0,(2)) where s > 0 and (c, w) G [0, T] x Cl. It is clear that condition (1.36) holds for XT and i? s . 1.6
Classes of Stochastic Processes
In this section we introduce and discuss various classes of stochastic processes. Let (Xt, FJ, PT,X) be a stochastic process on (fi, T) with state space (E, £). A sample path of the process (Xt, FJ, PT,X) corresponding t o w e f i is the function s\-^> Xs (LJ) defined on the interval [0, T]. For a given u G O, the sample path s >—> XS(UJ) is often called the realization of LJ, or a realization of the process Xt.
26
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
Definition 1.6 (1) A process (Xt,?J,FTtX) is called right-continuous if its sample paths are right-continuous functions on the interval [0,T). (2) A process (Xt, J-J, PT,x) is called left-continuous if its sample paths are left-continuous functions on the interval (0,T]. (3) It is said that a process (Xt, !F[,FT,x) is right-continuous and has left limits if its sample paths are right-continuous functions on the interval [0, T) and have left limits on the interval (0, T]. (4) It is said that a process (Xt,!Fl,'PTtX) is continuous if its sample paths are continuous functions on the interval [0,T]. If the process (Xt, FJ,fr,x) is right-continuous and has left limits, then the following notation will be used: \imX3(u>) = Xa-(u).
(1.38)
Since the process Xt is right-continuous, we also have \imXs(uj) = Xs(w). sit
(1.39)
Definition 1.7 A process (Xt, J^,FTiX) is called stochastically continuous if for all x G E, T G [0, T], and e > 0, , lim
FT,x(p(X!l,Xt)>e)=0.
Recall that the symbol p in Definition 1.7 stands for the distance on E x E. For e > 0 and y G E, put Gc(y) = {xeE:
p(x, y) > e} .
(1.40)
Then an equivalent condition for the stochastic continuity of the process Xt is as follows: for all x G E, r G [0, T], and e > 0, •
, Km
t — sl0;r<s
/ p ( s , y ; i , G e ( y ) ) P ( r , : r ; M 2 / ) = 0.
(1.41)
Jg
Definition 1.8 It is said that a process (Xt, T\', P T ,i) is strongly stochastically continuous provided that for all e > 0, *
. J ^ T
sup
Pr,x(p(Xa,Xt)>e)=0.
Transition Functions
and Markov
Processes
27
An equivalent condition for the strong stochastic continuity of the process Xt is as follows: for all e > 0, lim
sup
f P(s,y;t,Ge(y))P(r,x;s,dy)=0.
(1.42)
Definition 1.9 Let (fi,.F, P) be a probability space equipped with a filtration Tt,ti
The following assertions hold:
(1) Let Zt, 0 < t < T, be a right-continuous ft-martingale
with respect to
28
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
the measure P, and suppose that sup E\Zt\ < oo. t€[0,T)
Then the limit ZT = lim t |T Zt exists and is finite P-a.s. (2) Let Zt, 0 < t < T, be a right-continuous Tt-martingale with respect to the measure P, and suppose that the family of random variables Zt, t e [0,T), is uniformly integrable. Then ZT E L 1 , and Zt tends to XT in L1. Moreover, for all t £ [0, T],
Zt=E
[ZT I Ft]
F-a.s. The next assertion concerns martingales associated with transition probability functions. Lemma 1.6 Let P be a transition probability function, and let (Xt, Tl,PT,Z) be a Markov process with P as its transition function. Then for allO
Er,Xr [P (s, Xs; t, A)}
= / P(r, Xr; s, dz)P(s, z; t, A) = P{r, Xr; t, A) JB
P riX -a.s. This proves the first part of Lemma 1.6. The proof of the second part is similar. This completes the proof of Lemma 1.6. D
1.7
Completions of cr-Algebras
This section is devoted to completions of cr-algebras with respect to families of measures.
Transition Functions
Definition 1.11
and Markov
Processes
29
A measure space (X, J7, fj.) is called complete if B&T,
n(B) = 0, A c 5 = > A e JF.
Let (X, T, n) be a measure space, and define a family of sets by M = {A : there exists B £ J7 such that A C B and fx(B) - 0} . Definition 1.12 The completion J7^ of the cr-algebra J7 with respect to the measure /i is the cr-algebra given by
A
30
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
Hence x £ A\, and the inclusion A\ C A holds. Next, we will prove that A C A2. Let x £ A and x £ A2. Then x £ A' and x £ B. Therefore, x £ A\A' C AAA' C B, which is a contradiction. It follows that A C A2. We also have A2\Ai C B, and hence y, {A2\A\) = 0 . • Remark 1.3 It is not difficult to prove that the family A consisting of all sets A satisfying any of the equivalent conditions in Lemma 1.7, is a cr-algebra. Using condition (a) in Lemma 1.7, we can show that A = T^. Indeed, since the inclusions T C A and H C A hold, we have T^ C A. Conversely, if A £ A, then condition (a) in Lemma 1.7 holds for the set A, and hence A £ F*1. Definition 1.13 Let (X, J7) be a measurable space, and let V be a family of measures on T. The completion Tv of the cr-algebra T with respect to the family V is defined as follows: nev A cr-algebra T satisfying T = J-v is called V-complete. Let (X, T, fi) be a measure space. Then the measure /x can be extended to a measure [IQ on T^ by setting fio(A) = fi (Ai), where A £ J711 and A\ is a set such as in part (b) of Lemma 1.7. It is not difficult to see that the number Ho(A) does not depend on the choice of the set A\ in (b). We will often use the same symbol \i for the extension \i§ of \i. The measure space (XJJ^J/J.) is called the completion of the measure space {X,!F,^) with respect to the measure /u. If V is a family of measures on T, then every measure fj. £ V can be extended to the cr-algebra !FV as above, and we will use the same symbol V for the family {/io : H £ V} consisting of the extensions of measures ^L £ V. It is not hard to see that the cr-algebra Tv is ^-complete. Next, we will discuss completions of nitrations generated by Markov processes. Let P be a transition probability function, and let Xt be a Markov process on (fi, J7) associated with P. The process Xt generates the following families of cr-algebras: J=rt =a(Xs:T<s
and
^
=
f]
FJ
(1.43)
s:t<s
where 0 < r < t < T. Both families FJ and J-J+ are increasing in t and decreasing in r. For every T £ [0,T], consider the family of measures on
Transition Functions and Markov Processes
31
the cr-algebra Fj, given by VT = {Pa,x : 0 < s < T, x G E}.
(1.44)
Suppose that {Gl)-> 0 < r < i < T , i s a family of cr-algebras such that it is increasing in t, decreasing in r, and satisfies the condition G[ C \Ttf*
(1.45)
for all 0 < T < t < T. In (1.45), VT is the family of measures defined by (1.44). For the sake of simplicity, we will denote the completion [GI]VT of the a-algebra G\ with respect to the family of measures VT by the symbol G\. Since the families of a-algebras {FJ} and {J~[+} satisfy condition (1.45), the families {£[} and {^7+} a r e well-defined. L e m m a 1.8
The following assertions hold:
(a) A set A belongs to the a-algebra Tl if and only if for all (s, x) E [0, r] x E there exist sets ASiX G F[, A's<x G [f%f'-x, and A's\x G [Jr^f"-X such that A U A'ttX = Aa>x U A'lx
and
F.,x (A's,x) = Fs<x «
x
) = 0.
(b) A set A belongs to the a-algebra J-^+ if and only if for all (s, x) G [0, r] x E there exist sets Aa%x G F[+, A!ax G [T^f"-X, and A!'sx G [JFf]P",x such that AUA'S<X = AStX U A's\x
and
P s , x ( i ^ x ) = Fs<x ( <
x
) = 0.
Proof. Let (X, T, y) be a measure space, and let G be a cr-algebra satisfying G C T^. Let A be a set for which there exists A' G G such that AAA' G T» and y(AAA') = 0. Then there exist A" G G, Ax G .P 4 , rx and A2 € ^ ' such that /J, (Ax) = 0, y, (A2) = 0, and AU Ax = A" U A2. Indeed, it is sufficient to take A" = A', Ai = A'\A, and A2 = A\A'. Then Ai C AAA', A2 c AAA', and hence Ai G P" and ^42 G ^ with /j,(A1) + n(A2) = 0. Now let A be a set such that there exist sets A", A\, and A2 for which the conditions formulated above hold. Set A' = A". Since AAA' = AAA" C Ail) A2, we have AAA' € T*1 and /i (AAA') = 0. Next, arguing as above, we see that Lemma 1.8 holds. •
32
Non-Autonomous Koto Classes and Feynman-Kac Propagators
Lemma 1.9 Let P be a transition probability function, and let Xt be a corresponding Markov process on (Q, IF). Then for allO < r
n c n+
(i-46)
and u:t
Proof. The inclusion in (1.46) is straightforward. Next, we will prove the equality in (1.47). Let A G JJ+, and let As,x G C\U:tf',x, and A'lx G [F£]r'-X be the sets from part (b) of Lemma 1.8. Then it follows from part (a) of Lemma 1.8 that A G C\u:t
fl
fl
(1.48)
u:t
holds. Let A G f]U:t oo, and put 4 S , X = limsup^ U n , a i X , A'StX = limsupyi; n]S]X , and A"tX = l i m s u p A ^ ^ . n—*oo
n—*oo
n—»oo
Since
for all n > 1, we see that -^ U A s , x
=
A J , X U -^s,x-
Therefore, At,t G J^ + , ^ ) X G [ ^ ] p - , < x G [ ^ ] p - « , and P s , x (A'tiX) = P s>x (A"iX) = 0. It follows from part (b) of Lemma 1.8 that A G FJ+. Moreover,
fl
^c-^+-
(!-49)
u:t
Now we see that (1.47) follows from (1.48) and (1.49). This completes the proof of Lemma 1.9.
•
Transition Functions
1.8
and Markov
Processes
33
Path Properties of Stochastic Processes: Separability and Progressive Measurability
It will be established in this section and in Section 1.9 that under certain restrictions on a transition probability function P, there exists a Markov process associated with P and with prescribed measurability or continuity properties of its sample paths. Let P be a transition probability function, and let (Xt, J-[,PT,x) be a Markov process on (f2,.F) with P as its transition function. If Xt is a modification of Xt, then the process Xt is not necessarily adapted to the filtration T\. However, if we replace the family J-J by its completion 3FJ, then it is not hard to prove that the process (X t ,.F t T ,P r]X ) is a Markov process, and any modification Y% of Xt is an ^"-adapted process. We will always assume that filtrations are complete when dealing with modifications of stochastic processes (see Subsection 1.1.5). Let J be a countable dense subset of the interval [a, b], and let g be an 2J-valued function defined on [a, b]. The function g is called minimally continuous with respect to the set J if for every t € [a, b] there exists a sequence tk & J such that tk —» t and g{tk) —> g(t) as k —» oo. Similarly, the function g is called minimally right-continuous with respect to J if for every t G [a, b) there exists a sequence tk € J such that tk I t and 9(tk) -> g{t) as k -> oo. The next definitions concern the separability and measurability properties of Markov processes. Definition 1.14 A Markov process Xt is called separable if there exists a countable dense set J C [0, T] such that the sample paths 11—> Xt(to) are minimally continuous with respect to J PTiX-almost surely for all (r, x) £ [0, T] x E. Similarly, a Markov process Xt is called separable from the right if there exists a countable dense subset J C [0, T) such that the sample paths t H+ X ( (LJ) are minimally right-continuous with respect to J PTtXalmost surely for all (r, x) G [0, T] x E. Definition 1.15 (a) A stochastic process Xt on a measurable space (Q, T) with state space (E,£) is called a measurable process if the function (s,u) H-> Xs(w) is BmtT] ® .F/£-measurable, where B[O,T] stands for the Borel a-algebra of the interval [0,T]. (b) Let Xt be a stochastic process on a measurable space (CI, J-) with state space (E, £), and let Tt be a filtration such that Xt is ^-adapted. The
34
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
process Xt is called .^-progressively measurable if for every t £ [0,T], the restriction of the function (S,OJ) >—• XS(OJ) to the set [Q,t] x CI is B[ott] ® .F t /£-measurable. Definition 1.16 Let P be a transition probability function, and let [Xt, Tl, P r ,x) be a corresponding Markov process with state space (E,£). The process Xt is called ^"/"-progressively measurable if for every r and t with 0 < r < t < T, the restriction of the function (s, w) — i > Xs (w) to the set [r,t] x f i i s B[r,t] ® J7"/5-measurable. The next result states that measurable processes always have progressively measurable modifications. Theorem 1.4 Let Xt be a measurable stochastic process on a probability space (fi, J-, P) and with (E, £) as its state space. Let Tt be a filtration such that the process Xt is Ft-adapted. Then there exists an J-t-progressively measurable modification Yt of the process Xt. Proof. Any measurable process can be identified with a i?[o,r] ® F /£measurable function (t,uj) i-> Xt(u>). Denote by ME the space of classes of equivalence of Tj£-measurable functions from the space Q into the space E, equipped with the metric d(f, g) = inf (e + P [w : p(f(u), g{u>)) > e]). Then the convergence in the metric topology of the space ME is equivalent to the convergence in probability. Moreover, if / € ME and / „ G ME are such that oo
£d(/,/„)
(1-50)
n=l
then / n (a;) —> f(u>) P-almost surely on Q. Any #[o,r] ® -F/f-measurable function / generates a function / : [0, T] —> . M E defined as follows: for t 6 [0, T], /(£) is the class of equivalence in ME containing the function U) H->
f(t,Lj).
A simple function is a function from the space [0, T] x Q into the space E assuming only finitely many values, each on a S[o,r] ® ^-measurable set. A simple product-space function is a simple function such that each value is assumed on a set that can be represented as a finite disjoint union of direct products of sets from S[O,T] a n d T. An elementary measurable process Yt is a stochastic process on the space (fi, J7, P) such that there exist a partition
Transition Functions and Markov Processes
35
{Afc : k > 1} of the interval [0, T] into Borel measurable sets and a sequence {fk • k > 1} of T/E -measurable mappings of the space ft into the space E such that Yt = fk for all t e AkDenote by DE the class consisting of all $[O,T] ® ^"/5-measurable functions / , for which the function J: [0,T] -> ME is B[0,T]/BJ^ -measurable and has a separable range. Here Bj^ denotes the Borel cr-algebra of the space ME- An equivalent definition of the class DE is as follows. A S[O,T] ® .F/f-measurable function / belongs to the class DE if the function / can be approximated by a sequence of elementary measurable processes in the sense of pointwise convergence on ft uniformly in t G [0,T]. Lemma 1.10 The class DE coincides with the class of all B[O,T] ®T IEmeasurable functions. Proof. Let us first prove the lemma in the case where E = R. It is not difficult to see that the class £>R is closed under pointwise convergence of sequences of functions and contains all simple product-space functions. By the monotone class theorem for functions, Lemma 1.10 holds for E = R. Our next goal is to prove Lemma 1.10 for any finite subset of R. Let Ro = {ci,c 2 , • • • ,Cn} be a finite subset of R equipped with the metric inherited from the space R. Next, using Lemma 1.10 for E = M, we see that if / : [0,T] x ft — i > Ro is a JB[O,T] ® F/Buo-measurable function and i s N , then the range < f(t) : 0 < t < T > of the function / can be covered by a countable disjoint family A\, k G N, of Borel subsets of the space MR so that f~l (A\) G B[otT], and the diameter of any set A\ is less than \. Therefore, there exists a sequence / , of elementary measurable processes such that fi{t) G MR0 for all t G [0,T], and moreover sup d(f(t,-)-fi(t,-))
(1.51) l
for all i € N. It follows from the second definition of the class DR 0 that / G -DR 0 . This establishes Lemma 1.10 for E = Ro. Next, we will pass from a finite subset Ro — {ci, C2, • • • , c„} of the space R to a finite subset EQ = {x\, X2, • • • , xn} of the given state space E. Let g : [0,T] x ft — i > EQ be a B[O,T] <8> J 7 //BB 0 -measurable function. Then for every 1 < j < n, we have g(t, w) = Xj on a set Bj e S[O,T] ® J7- The sets -Bj may be empty. It is also true that the nonempty sets Bj are disjoint and cover [0,T] x ft. Let us consider a #[O,T] ®T/B^o-measurable function defined by f(t, u>) = Cj on the set Bj with 1 < j < n. By the previous part
36
Non-Autonomous Kato Classes and Feynman-Kac Propagators
of the proof, there exists a sequence fa of elementary measurable processes such that fi(t) £ MRQ and inequality (1.51) holds. Replacing Cj by Xj in the function /,, we get a function g^. Taking into account (1.51) and the fact that p(xm,Xfc) < c cm — Ck\ for 1 < m < k < n, where c > 0 is a finite constant, we see that lim sup d(ff(i,-)>ffi(t>0) = 0. °te[o,T]
,_>0
(1.52)
Since gi is an elementary process and (1.52) holds, we have g £ DE0- It follows that any simple function s : [0, T] x Q —-> E belongs to the class DENow let / be a S[O,T] ® T/E -measurable function. Then the function / can be approximated pointwise by simple functions, and since the class DE is closed under pointwise convergence, we have / £ DEThis completes the proof of Lemma 1.10. • Let us continue the proof of Theorem 1.4. By Lemma 1.10, the class DE coincides with the class of all B\O,T\ ® .F/f-measurable functions. Approximating 6[o,T]/Sjq-measurable functions with separable range by simple functions, we see that for any measurable process Xt, there exists a sequence Y" of elementary measurable processes such that
<*PW)<2^T
(L53)
for all t £ [0,T] and n > 1. Recall that to every elementary process Ytn there corresponds a partition {A% : k > 1} of the interval [0, T] into Borel measurable sets and a sequence {/£ : k > 1} of T/E-measurable mappings of the space Q, into the space E such that Y™ = f% for all t £ A%. Fix n > 1 and 5 £ (0, T). Our next goal is to modify the process Y™ as follows. Put si = inf {t : t £ A%}. If sn £ An, then the new process Z? is denned for t £ Al by Ztn = X s n. If s£ £ A£, then we fix f£ e A% such that *£ - sfe < *> a n d P u t %t = -XtJJ f ° r a u * G -4fe- It i s c l e a r that the new processes J?" are elementary measurable processes. Since Xs is an adapted process, it is easy to see that for every t £ [0, T — 5], the restriction of the function (s, w) i-» Z?(w) to the space [0, t] x £1 is B[o,t] ®Tt+s/£-measurable. Moreover, inequality (1.53) implies d{XuZ?)<±;
(1.54)
for all t £ [0,T] and n > 1. By (1.50), Zt"(w) -> X;(w) as n -> oo for all £ G [0, T] almost surely on (1 Fix XQ £ E, and put I t (a;) = lim Z"(w)
Transition Functions and Markov Processes
37
if the limit exists, and Yt(u)) = xo otherwise. Then the process Yt is a modification of the process Xt. We will next show that the process Yt is progressively measurable. It is clear from the definition of the process Yt that it suffices to prove that every process Z " is progressively measurable. Since the process Xt is adapted, we see that for every u £ [0,T — S], the restriction of the function (s,u>) t-> Z™(u>) to the space [0,u] x 0, is #[o,u] <8> Fu+s/S-measurable. Now let t € [0,T]. Then using the previous assertion with um = t — ^ , m > mo, we see that the restriction of the function (s, w) — i > Z™{w) to the space [0,t) x Cl is S[o,t) <8> Tt/^-measurable. Since the process Xt is adapted, we see that the process Z " is progressively measurable for all n > 1. This completes the proof of Theorem 1.4. • Every sample path of a measurable process is a Bp^}/^ measurable function. The next assertion provides examples of progressively measurable stochastic processes. Theorem 1.5 measurable.
Every left- or right-continuous process is progressively
Proof. Let X be a right-continuous process. Fix r and t with 0 < T < t
7rfc = { r = s (
fc
)<4 fc )<...< s f
= f}
such that the mesh of the partition 7i> tends to zero as A; —> oo. For every k > 1, define a simple process Xk on [r,t] as follows: X% = X oo, if s € [SJ ', Sj + '), and X$ = Xt. It is clear that the function (s, a;) — i > X*(w) defined on [T, t] x Q, is B\Ttt] <8> F[ /£-measurable. It follows from the rightcontinuity of the process Xt that lim
X*{LJ)=XS(W)
fc—+00
for all s € [r,t] and w e fi. Hence, the function (s,w) i—»-X"s(w) defined on [T, i] x fi is #[T,t] ® .F t T /£-measurable. This means that the process Xt is progressively measurable. The proof of Theorem 1.5 is thus completed for right-continuous processes. The proof for left-continuous processes is similar. • The next result concerns separable processes and stochastic equivalence. Theorem 1.6 Let P be a transition probability function satisfying condition (1.41)- Then there exists a separable process (Xt,^,PT,x) on (Q,J-)
38
Non-Autonomous Koto Classes and Feynman-Kac Propagators
with state space (E, £) such that the transition function of Xt with P. Proof.
coincides
The following lemma will be used in the proof of Theorem 1.6:
Lemma 1.11 Let P be a transition probability function, and let Xt be a Markov process associated with P. Then for every (r, x) G [0, T] x E there exists a separable P r , x -modification XSI r < s
The next assertion is an important part of the proof of Lemma
Lemma 1.12 Let B G £, r G [0,T], and x e E. Then, there T exists a finite or countable sequence tk G [ ,T] depending on B for which ¥TiX[N{t,B)] = 0, t G [T,T], X G E, where N(t,B) = {Xtk G B : k>\,XtiB}. Proof. We will use the method of mathematical induction in the proof of Lemma 1.12. Let t\ be any number in [r, T]. If the numbers t\, £2, • • •, tk have already been chosen, then we put 7fe
= sup P r , x [Xtj € B : 1 < j < k, Xt i B] . t€[r,T]
If 7fc = 0, then we are done. If jk > 0, then there exists tk+i £ [r, T] such that sup P T;X [XtjeB:l<j<
k, Xtk+1
tB]>^.
te[r,T)
z
Put Nk(B) - {Xtj £B:l<j
Xtk+1 $ B) .
It is clear that the sets Nk(B) are disjoint. Therefore,
k
k
and hence, -jk —» 0 as k —> 00. Since P r ? x [N(t, B)] < 7^ for any k > 1, we have P r , x [N(t,B)] = 0 for all t G [T,T\. This completes the proof of Lemma 1.12. •
Transition Functions and Markov Processes
39
Now let Bi be a sequence of Borel sets in E. It follows from Lemma 1.12 that for every i > 1 there exists a sequence t\, k > 1, such that PT,x[N(t,Bi)]=0
(1.55)
for all t G [r, T\. By enumerating the set {t\ : i > 1, k > l } , we see that the sequence tk can be chosen independently of i. It is not hard to prove that the sets N(t,Bi) constructed for this sequence are subsets of the similar sets in (1.55). This means that (1.55) holds for the new sequence. Next we will show that more is true. Lemma 1.13 Let r G [0,T], x G E, and let Bt be a sequence of Borel sets in E. Then there exists a sequence tk G [T,T], k > 1, and for every t G [r, T] there exists a set N(t) G Tj- such that Pr,x [N(t)] = 0,
(1.56)
N(t,B)cN(t)
(1.57)
and
for all t G [r, T] and for all sets B which can be represented as a countable intersection of elements of the family {Bi}. Proof.
Put N(t) = ( J N(t, Bt) (here we use the sequence tk constructed
after equality (1.55)). Let B = f|. Biy N(t, B)c\J
{Xtk zBfk>\,Xti
Then Btj}
j
c\J{Xtk 3
G B^, k>\,Xti
Bij}=\jN(t,Bij)
c N(t).
(1.58)
3
Now it is clear that (1.56) follows from (1.55), while (1.57) follows from (1.58). • Let us return to the proof of Lemma 1.11. Denote by Cj the family of all open balls of rational radii centered at the points of a countable dense subset of E, and put Bi = E\d. It is clear that the family of sets which are representable as countable intersections of the sets from the family {Bi} contains the family of all closed subsets of E. Applying Lemma 1.13, we see that there exists a sequence tk G [r, T], and for every t G [r, T] there exists a set N(t) such that Pr,x[7V(i)] = 0 and N(t, C) C N(t) for all closed subsets C of E and all t G [r, T\.
40
Non-Autonomous Kato Classes and Feynman-Kac Propagators
Let Yt be a separable process on ft, and define a new process Xt as follows: Xt(w) = Xt{w) if t G {tk} or u £ N(t), and Xt(u>) = Yt(w) otherwise. Our next goal is to show that
Xt=Xt
=1
(1-59)
for all t G [T, T] and also to prove that Xt is a separable process. Indeed, if t G {U}, then lxt = Xt\ = ft. If t € {**}, then lxt ^ Xt\ C JV(i). This gives (1.59). Now we are ready to prove the separability of the process Xt- If t = ti for some i > 1, then Xti = Xti, and hence for all UJ G ft, the corresponding sample path is minimally continuous with respect to {U}. lit $. {tk} and ui G -/V(i), then X t coincides with a {ij}-separable process. If t £ {t,} and w $. N(t), then Xt(u>) = X t (w), and we proceed as follows. Suppose that X t (w) cannot be approximated by a subsequence of the sequence Xti{nv). Then there exists a ball C centered at Xt{u) such that Xti(w) £ C for all i > 1. Moreover, for every i > 1, we have Xtt{u>) G B and Xt{uj) £ B where B = E\C. Hence, w G N(t,B) C iV(i), which is a contradiction. Therefore, Xt (u>) can be approximated by a subsequence of the sequence Xti(uj), and this implies the separability of the process Xt. This completes the proof of Lemma 1.11. • Note that we have not yet employed the stochastic continuity condition in the proof of Theorem 1.6. This condition is needed to guarantee that any countable dense subset of [r, T] can be used as a separability set. Lemma 1.14 Suppose that P is a transition probability function satisfying the stochastic continuity condition (1-7), and let Xt be a corresponding Markov process. Fix (T,X) G [0,T] X E. Then any countable dense subset of [T,T] can be used as a separability set in Lemma 1.11. Proof. By Lemma 1.11, there exists a separable process Xt on [T,T] associated with a separability set {tk}- Let {SJ} be any countable dense subset of [r,T}. Next, we will show that for PTiX-almost all u> G ft and all k > 1, the element Xtk{w) of E belongs to the set A{UJ) consisting of all limit points of the set < XSi (w) >. By Fatou's Lemma and the stochastic
Transition Functions and Markov Processes
41
continuity of the process Xt, we see that for every k > 1, liminfp(xtt,X,4) >0] < lim P x
\iminfp(xtk,XSi)
< lim liminfP TiX n—»-oo i—>oo
> >
= 0.
(1.60)
Condition (1.60) means that for P r x -almost every w € Q, and every k > 1, the element Xtk(uj) of £ belongs to the set A(LJ). Therefore, the process Xt is separable with respect to the set {s;}. This completes the proof of Lemma 1.14. D The next result (Lemma 1.15) will allow us to get rid of the dependence of the process Xs in Lemma 1.11 on the variables r and x. The conditions in Lemma 1.15 are as follows. A family of stochastic processes X\T,X parameterized by (r, x) 6 [0, T] x E is given, and it is known that the sample paths of these processes possess a certain property. Our goal is to construct a single non-homogeneous process Xt from the processes X\T so that the sample paths of Xt possess the same property. Lemma 1.15 Let P be a transition probability function and suppose that for every pair (r, x) G [0, T]x E, a stochastic process is given on (fi,!F). Suppose also that X,
(r,x)
eB
P{j,x;t,B)
(1.61)
for all t with r < t < T and all B £ £. Let F be a class of E-valued functions defined on [0,T], and assume that the sample paths of all the processes X t ' T ' x) belong to F. Then there exists a Markov process Xt such that its sample paths belong to the class F and P is its transition function. Proof. Consider a new sample space (l = [ 0 , T ] x £ x ( l . This space will be equipped with the cr-algebra T=
{ACQ
: AT>X G T for all (T,X) G [0,T] X £?} .
Here AT^X = iu> : (r, x, ui) € A >. Define a stochastic process on il by Xt(r,x,u)
X\T
Ft = * (X.
T
{UJ) and consider the family of ^--algebras given by
<
s
•~fl
fined as follows: P.
Ar,
for every A £ T. It is not hard to
42
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
see from the definition of the process Xt that its sample paths belong to F. It remains to prove that Xt is associated with P. Let B £ £ and r < t. Then
'u = (u,y,w) eft : Xt(T'x) (w) e B
Xt£B
w : X t (r,x) (w) e f l ] = PT)X [w : X t (w) G B].
(1.62)
It follows from (1.61) and (1.62) that P is the transition function of the process Xt. This completes the proof of Lemma 1.15. • Finally, we are ready to finish the proof of Theorem 1.6. Let P be a given transition probability function satisfying the stochastic continuity condition, and let Yt be a Markov process on (fl, J7) with transition function P. By Lemma 1.11, for every (r, x) there exists a separable process i y , T < t < T, that is stochastically equivalent to the process Yt, r < t < T. Put
*
\x,
ifO
Then it is clear that the process Xj:T'x' is separable for every r and x. Since the stochastic continuity condition holds, Lemma 1.14 and the definition of the process X\T'X' imply that any countable dense subset of [0, T] serves as a separability set for the processes X^T'x'. Moreover, condition (1.61) holds. Let usfixa separability set {£&}, and define a class of functions F as follows. The class F contains all functions on [0, T] with values in E which are minimally continuous with respect to {tfc}. Now it is clear that we can apply Lemma 1.15, and get a separable process Xt with transition function P. This completes the proof of Theorem 1.6. • The next theorem concerns the existence of a progressively measurable Markov process associated with a given transition probability function P. T h e o r e m 1.7 Let P be a transition probability function satisfying the strong stochastic continuity condition: ,J.ini ^ t~slO;s
SU
P
( T ) a ; ) € [o i 4 ] x JS J
P{s,y;t,Gt(y))P{T,x;s,dy) = 0 E
Transition Functions and Markov Processes
43
for all £ > 0. Then there exists a progressively measurable process (Xt,Tl,Pr,x) on the space {£l,F) with P as its transition function. Proof. The strong continuity condition was discussed in Section 1.6 (see (1.42) and Definition 1.8). Let Xt be a strongly stochastically continuous Markov process with P as its transition function and assume that FJ is the filtration a (Xs : T < s
T]xE}.
Consider a sequence of partitions 0 = tft < t" < • • • < tm —T such that max (i? - *"_,) - > 0 a s n oo, and define a sequence of stochastic processes by if *!?_! < s < t™ with 1 < j < m„
X"
if s = T.
x7
It is clear that every process Xn is ^"/-progressively measurable. It follows from the definition of Xn and Definition 1.8 that lim
sup
sup
"->oo(T,x)e[0,T]xBs:0<s
Prx
lp(X?,Xs)>e}=0
(1.63)
for all e > 0. Next, using (1.63) and reasoning as in the standard proof of the fact that the convergence in measure of a sequence of functions implies the existence of an almost everywhere convergent subsequence, we see that there exists a sequence in f oo such that for every s € [0,T], X%sn converges to Xs P T)X -a.s. for all (T,X) £ [0,T] x E. Indeed, from (1.63) we see that there exists a sequence in | oo such that P(xi",xs)>
1 - 2"
for all s G [0, T], T G [0, T], and x e E. Therefore,
nul'W.*-)^} j>ln>j
for all s G [0, T], the sequence X\n Without loss Let A be the set
^
'
r G [0, T], and x G E. It follows that for every s G [0, T], converges P r , x -a.s. to Xs for every r G [0, T] and x G E. of generality, we can assume that in = n for all n > 1. of all (s, w) G [r, T] x f2 such that linin-joo X™(ui) exists.
44
Non-Autonomous Koto Classes and Feynman-Kac Propagators
It is clear that the set A is (fi[o,rj ® ^-measurable. Moreover, for every se[0,T], PT,X{W:(S,W)G.4} = 1
(1.64)
for all (T, X) € [0, T) x £\ Indeed, if s G [0, T] and (r, x) G [0, T] x E, then the set {ui : (s, u>) G .4} contains an ^"-measurable set AST such that ¥r<x(As(T,x)) = 1. This follows from the fact that if s G [0,T], then X£ converges to Xs P T)X -a.s. for every (T, X) G [0, T] x E. Define a new process Xs on Q by X,(w) = lim X™(UJ), if (s, u>) G ^4, and
Xs(bj) = xo for all (S,UJ) $ A, where xo is a fixed point in E. Since for all (r, x) G [0, T] x E, the process X* is an P^-modification of the process Xt, it is clear that PT)X Xt e B = P (T, X; t, B) for all 0 < T < t < T and all B G £. Moreover, since the processes Xn are .^-progressively measurable and (1-64) holds, the process Xt is ^-progressively measurable. This completes the proof of Theorem 1.7. • 1.9
P a t h Properties of Stochastic Processes: Continuity and Continuity
One-Sided
In this section we continue our exploration of the properties of paths of Markov processes. Specifically, we will study the processes with continuous sample paths and the processes for which the sample paths have only jump discontinuities. Theorem 1.8 Let P be a transition probability function, and let Xt be a Markov process with transition function P. Suppose that for all (T, X) G [0, T) x E the following condition holds: lim
PT,x-ess sup P (s, Xs; t, Ge (xs) ) = 0
t-si0)T<s
\
(1.65)
\ J J
for all e > 0, where Ge(y) is defined by (1-40)- The essential supremum in (1.65) is taken with respect to the measure PT,X- Then there exists a Markov process (X t ,^7,P T , x ) on (Q,,^) such that Xt is right-continuous, has left limits, and the transition function of Xt coincides with P. Corollary 1.1 Let P be a transition probability function satisfying the following condition: lim
supP(*,3/;*,G £ (j/))=0
(1.66)
Transition Functions
and Markov
Processes
45
for all e > 0, where Ge(y) is defined by (HO). Then there exists a Markov process (Xt,J-[,FT,x) on (Q.jJ7) such that Xt is right-continuous, has left limits, and the transition function of Xt coincides with P. Remark 1.5 It can be shown that condition (1.66) implies condition (1.41). Indeed, if a transition probability function P satisfies condition (1.66), then we have / P (s, y; t, Ge(y)) P(T, X; S, dy) < sup P (s, y; t, Ge(y)), JE
y&E
and hence condition (1.41) holds. Proof. The structure of the proof of Theorem 1.8 is similar to that of Theorem 1.6. We start with the following lemma. L e m m a 1.16 Let P be a transition probability function, and let Xt be a Markov process with transition function P. Suppose that condition (1.65) holds for a fixed pair (T,X) 6 [0,T) x E. Then there exists a ¥T>Xmodification Xt of the process Xt, r < t < T, which is right-continuous and has left limits. The process Xt depends on T and x. Proof. Without loss of generality, we may assume that Xt is a separable process on (fi,.F) (see Lemma 1.11). Fix r and x. The following random variables will be used in the proof: tp(e,r,8,t) = P TlX [p(Xs,Xt)
> e | J7]
(1.67)
where r < r < s
sup
P r>x -ess sup
(s,t):0
(1.68)
u>
where 0 < 5 < T — r and e > 0. Using the properties of conditional expectations, we get
P T , x a.s.
(1.69)
Indeed, tp(e, r, s, t) = E T j I [PTjX [p(Xs,Xt)
>e\j^]\FJ]<
a(e, t - s).
Moreover, using the Markov property, we obtain a(e,6)<
sup (s,t):0
P r , x -esssup P {s, X3;t,Gc(Xs)). '
ui
(1.70)
46
Non-Autonomous Kato Classes and Feynman-Kac Propagators
Indeed, since f o r O < T < r < s < £ < T w e have T^ C J-J, it follows that a(e, S) <
sup
ess sup PStxs [p(Xs, Xt) > e]
(s,t):0
=
sup
u
P TiX -esssup P{s, Xs;t,
(s,t):0
Ge(Xs)).
u
By condition (1.65) for (r, x), we see that lima(e,<5) = 0
(1.71)
610
for every e > 0. In order to continue the proof of Lemma 1.16, we will need the following definition. Definition 1.17 Let / b e a subset of [r, T] and / be a function defined on J and taking values in E. Let e > 0 and k > 1. It is said that the function / has at least k e-oscillations on the set / provided that there exists a finite subset J = {si < S2 < • • • < Sk < Sfc+i} of the set / such that p(f(si),f(si+i)) > e for all 1 < i < k. Denote by Zk,e the class of all functions having at least k e-oscillations on /. We will say that the function / has a finite number of e-oscillations on / provided that there exists k £ N such that / ^ Zk>cLemma 1.17 For every subinterval I = [a,b] of the interval [T,T], the class of functions on I having a finite number of e-oscillations for every e > 0 coincides with the class of functions on I with no discontinuities of the second kind. Proof. Then
Suppose that the left limit of / does not exist at a point t € (a, b]. C=
inf
sup p(f(s)J(r))>0.
(1.72)
s:a<s
Now let e be such that 0 < e < (. Then it follows from (1.72) that / has an infinite number of e-oscillations on I. The proof of this fact in the case of the right-hand limit is similar. Conversely, let us assume that there exists e > 0 such that / has an infinite number of e-oscillations on I. Then there exists a sequence h = {ti < t\ < • • • < £fc+i}i k > 2, of finite subsets of [a, b] such that P (/ (*j) ' / (*i+i)) — e f° r a ^ 1 — J — ^- This implies the existence of a sequence (i£-i» i £ , 4 + i ) s u c h t h a t 4 + i - i i L - i -^ 0 as A; ^ oo. By passing to a subsequence, we may assume that t\ has a one-sided limit t. With
Transition Functions and Markov
Processes
47
no loss of generality, we may assume that tkk | t. This is a contradiction, since tkik T t, tkk_, -> t, and p ( / (%_-,) , f (tkk)) > e for all k > 2. This completes the proof of Lemma 1.17. D Let us continue the proof of Lemma 1.16. For every e > 0, k > 1, and any Borel subset H of the interval [r, T], define the following events: Tfc(e, H) = {w : the function t •-» Xt(uj) has at least k e-oscillations on H} (1.73) and roo(e,H)=f)Tk(e,H).
(1.74)
fc>i
Our goal is to prove that Pr,x[roo(€,[r,T])] = 0
(1.75)
for every e > 0. Then • T,X
|Jr0O(e,[T,ri) =o,
and hence, by Lemma 1.17, FTiX {u> : the function t >—> Xt (w) has no discontinuities of the second kind on [T, 2*1} = 1.
(1.76)
Let I = {ti < ti < • • • < tn} be a finite subset of the interval [T, T], Put $fc(e,/)=PT,x[rfc(e,J)|J71]
(1.77)
(3k{e,I) = ess sup $fc(e, J)(w).
(1.78)
and
L e m m a 1.18
For every k > 1, the following estimate holds: /3fc(e,/)<(2a(|,tn-i1))fc.
(1.79)
Proof. For any i with 1 < i < n, denote by A\ (e, /) the event consisting of all u G Ti(e,/) such that p(Xtj,Xtm) > e for some j and m with i < j < m < n. In addition, let us denote by Blk_l(e,I) the event consisting of all u G rfe_i(e,/) such that w has at least k — 1 e-oscillations on the set (
48
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
than k — 1. Then the events B\_l (e,I) are disjoint and ^-measurable, n
[J Bl-i
(e> -0 = r fe-i (e> -0> a n d moreover,
i=l n
rfe(e,/)cU[£i-iM)n4M)]. i=l
For every 1 < i < n, put /* = {tm : i <m
It follows that
n
$k(e,/)<52ETlI
-"r,x
XA^e./jXa^^e,/) | JTi | Ft!
n
K
XA\(e,I) i=l n
< X > - i ( e , J«)PTlX [i?UM) | JTJ i=i
= max
^-i^JOE^frfc-iCc,/)!^]
i:l
=
'
max Afc_i(e,/i)$i(e,/).
(1.80)
i:l
Now we see that in order to prove (1.79), it suffices to show that /31(e,J)<2a(|,sm-s1),
(1.81)
for any set J = { s i , . . . , sm} with T < s\ < • • • < sm < T. Define the following events: Ki(e,J)
= [p(XSl,X3j)
< | for all 1 < j < i - 1, and p(XSl,XSi)
> |}
for all 2 < i < m, and
Li(e,J)={p(X.t,X.m)>l} for all 1 < i < m. It is clear that the events Ki(e, J) are disjoint and Tl -measurable. Moreover, m
r x (6, J) C L^e, J) U ( J (tf<(e, J ) n L4(e, J ) ) . t=2
(1.82)
Transition Functions
and Markov
Processes
49
Indeed, using the triangle inequality, we obtain m i=2
Let UJ e Ti(e,J). Then CJ S Ki(e,J) for some i with 2 < i < m. Here i depends on CJ. If w ^ £•;(£, J), then it follows from
P(xai(w),xam(w))>p(x.M,xM)-p(x.M,xM)
>\
that w e L ^ e , J ) . This gives (1.82). It is not hard to see from (1.82) that *i(e, J) < Pr,x [Li(c, J) | ^ J J + X;P T ,x [Ki(e, J ) n ^ ( e , J ) | J ^ J i=2 m
< P r , x [Li(C, J) | J7J + £ E r , x [Er,x [x^(e,J)XLi(£,J) I ^ J I K] < FT>X [Liit, J) | J7J + £ > , , * [Xif((e,J)Er,« [XLi(e,J) I JT4] I - ^ J • i=2
(1.83) Therefore, taking the essential supremum in (1.83), we get A(e. •/) <
a
Q> s m - s i ) + a ( | , s m - s i ) ^ P r ,
x
[^(e, J) | J ^ J
j=2
< 2a K , s m - s i J . Now it is clear that (1.81) holds. Moreover, (1.81) and (1.80) imply (1.79). This completes the proof of Lemma 1.18. • Next, we will prove conditions (1.75) and (1.76). Fix e > 0. Since (1.71) holds, there exists 5 > 0 such that oc{\,5)<\.
(1.84)
Given e > 0, fix a number S for which (1.84) holds. Let [a, b] be any subinterval of [r, T] such that b — a < S. Then (1.79) gives lim sup {0k(e, I) • I finite and / C [a, b]} = 0. fe—»oo
50
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
Let us subdivide the interval [T,T] into a finite family {Aj}, 1 < j < m, of intervals of length at most S. Recall that we assumed the separability of the process Xt. Moreover, any countable dense subset J = {st : £ > 1} of [T, T] may serve as a separability set for Xt on [r, T]. Fix such a set J and put Jn = {st • 1 < (• < n}, Jnj = J„ n A j , and anj = min{s : s G Jn,j}Then for every k > 1, we have
Pr,x [roc (e, Jn,j)\
< Pr,x [Tfc (e, J n > i )] = E r , x
Pr
roo(e,J)= |J roofoJnAj), l<j'<m
implies PT)X [Too (e, J)] = 0. Since J is a separability set for Xt on the interval [T,T], we see that (1.75) and (1.76) hold. Denote by Q the event consisting of all w G Q for which t — i > Xt(u) has no discontinuities of the second kind on [r, T], and define a stochastic process by Xt(u>) = Xt(uj) if LJ £ fi, and by Xt(u) = x if u) £ Q\Q. It is clear that the process Xt is separable, and that its sample paths have no discontinuities of the second kind. Moreover, Xt is stochastically equivalent to Xt on [r, T]. It follows from the properties of the process Xt that the limit Xt(w) = lirrin-Kx, Xt+x (w) exists for all t G [r, T) and UJ e Q. The process X t is right-continuous on [0,T) and has left limits on a fixed t G [r, T), we have
(T,T}.
{x t ^x t }= |J i P (x t ,x t )>-i. m=1 ^
'
Moreover, for
Transition Functions
and Markov
Processes
51
It follows that • T,X
Xtj=XA<
lim
FT,x\p(Xt,Xt)> m
lim P T .
(xt,
lim
Xt+x
\ < lim
lim P TiX P(xt,xt+1)>-
m—>oo n—*oo
Now using the stochastic continuity of the process Xt, we get = 0. Therefore, the process Xt, [T,T], is a modification T,X Xt ^ Xt 1
of the process Xt, [T,T], and it follows that the process Xt has P as its transition function. This completes the proof of Lemma 1.16. • Finally, using Lemma 1.16 and Lemma 1.15 for the class F consisting of all right-continuous functions on [0, T] with left limits, we see that Theorem 1.8 holds. • The next result concerns the continuity of the sample paths. Theorem 1.9 Let P be a transition probability function, and let Xt be a Markov process with transition function P. Suppose that for all (r, x) S [0, T) x E the following condition holds: lim
PT,x-esssupP(s,Xs;t,Ge(Xs))
t-siO-,T<s
=0
\
(1.85)
)
for all e > 0. Then there exists a continuous process (Xt, J-^,FT}X) with P as its transition function. The next corollary follows from Theorem 1.9. Corollary 1.2 Let P be a transition probability function satisfying the following condition: ,. lim sup P(s,y;t,Gt(y)) t-s|0;0<s
= 0n
for all e > 0. Then there exists a continuous process (Xt,!Fl,FTtx) as its transition function. Proof.
(1.86) with P
The following lemma will be used in the proof of Theorem 1.9:
52
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
Lemma 1.19 Let P be a transition probability function, and let Xt be a Markov process with transition function P. Suppose that for a given pair (T,x)e[0,T)xE, lim
Vrx-esssupP(s,Xs;t,GJXs))
t-slO;r<s
t —S
'
V
=0
(1.87)
/
for all e > 0. Then there exists a continuous PT)X -modification Xt of the process Xt. The process Xt depends on r and x. Proof. By Lemma 1.16, there exists a PTiX -modification Yt of the process Xt which is right-continuous and has left limits. Let TTi = {r = t« < t\ < • • • < 4 _ x < ^ = T } ,
i>1
be a sequence of partitions of the interval [r, T] such that 7TJ-|_I is a refinement of 7Tj for a l i i > 1. Let us also assume that 5i = m a x { 4 + i - t \ :0
tends to 0 as i tends to cc. Note that condition (1.87) still holds if we replace Xt by Yt. Put z(e,<S) =
sup
PT,X- ess sup P
(s,Ys;t,Gc(Ys))
(s,t):0
and ?(e, (5) =
sup
Pr,ar ess sup
P( 5 ,n;i,G e (F s )) f-s
(s,t):0
for all e > 0 and 5 > 0. Then, using (1.67)-(1.70), we obtain Tli — 1
EMK^'^+J^ fc=0 n;-l
= £ ET,X [PT>I [/> (Yti,Yti+i) > e | ^ ] ] < £ a (e,4 +1 - 4) fc=o
< E - (e> «i+i - 4) = E Ci+i - 4) fe=0
k=o
i,fc+1 t%
fcj
— fl
rii-1
< E (4+1 - 4) % ffc+1 - 4) < (T - T)Z(€, 5i) < Tz(e, 5t fc=0
Transition Functions
and Markov
Processes
53
fcE^K^'^J^^0-
(L88)
By (1.85), z (e, Si) —> 0 as i —• oo. Therefore Tli-l
fe=0
For every z > 1 and e > 0, put Bi(e)
= {, e n: f e max n / (y t ,M,r t . + i H) >
and oo
Cj(e) = \jBi(e). i=j
It is easy to see, using (1.88) and passing to a subsequence of the sequence 7Tj, that without loss of generality we may assume that lira P T , x [C>(e)]=0. j—>oo
The sequence of events Cj(e) is decreasing. Hence, for C(e) = O
Cj(e),
we have PT,X [C(e)j = 0. The complement D(e) of the event C(e) can be described as follows: u £ D(e) if and only if there exists j > 1 depending on u) and such that
for all i > j . Recall that the process Yt is right-continuous and has left limits. For all t € (r,T], put Yt-(cj) — limSft Ys(ui). Our next goal is to show that if u £ D(e), then p(Yt-{u),Yt{Lj))<2e
(1.90)
for all t E (r, T}. Indeed, let w G £>(e), and let j be the integer corresponding to UJ in the definition of D(e). Assume that p(Ys_(uJ),Ys(u))>2e for some s € (r, T]. It is not hard to see that this inequality contradicts the inequality in (1.89). Hence, (1.90) is satisfied. Moreover, if w € H D(e), then (1.90) holds for any e > 0. This establishes the continuity condition for all a; £ Q D(e) and all t G {T,T\. Since the process Xt is right-continuous £>0
54
Nov.-Autonomous
Kato Classes and Feynman-Kac
Propagators
at t = r and P T]X [ne>oD(e)j = 1, there exists a continuous modification Xt of the process Yt. This completes the proof of Lemma 1.19. • It is not hard to see that Lemma 1.15 with the class F consisting of all continuous functions on [0,T] implies Theorem 1.9. • Remark 1.6 Note that if r and x in Theorem 1.9 are fixed, and if Xt is a Markov process with transition function P, then there exists a continuous PTjX-modification Xt of the process Xt. Remark 1.7 Let P be a transition probability function, and let Xt be a Markov process associated with P. Fix r with 0 < T < T and x e E, and let A be a closed subinterval of the interval [r, T\. Suppose that lim
FTX-esss\ipP(s,Xs;t,Ge(Xs))
t—s|0;s,teA
'
=0
\
/
for all e > 0. Then, arguing as in the proof of Lemma 1.16, we see that there exists a PT>x-modification Xt of the process Xt, which is right-continuous and has left limits on the interval A. Similarly, if P is a transition probability function such that lim
P r x- ess sup P (s, Xs; t, GJXS))
t-«i.0;s,t6A t — S
\
=0 I
for all e > 0, then there exists a PTiX-modification Xt of the process Xt, which is continuous on the interval A. Next, we will formulate two results concerning the path properties of supermartingales (Theorems 1.10 and 1.11 below). The proofs of these results will be omitted, and we refer the reader to Section 2.9 in [Yeh (1995)] for more information. Theorem 1.10 Let Z% be a supermartingale on a filtered probability space (fi, J-, Tu P) where 0 < t < T. Denote by Q the set of all rational numbers in [0, T], Then there exists a set A £ T such that P(A) = 0, and the limits lim Zr(w)
and
lim
ZJw)
exist for all UJ £ fi\A. The set A in the formulation of Theorem 1.10 can be chosen independently o f t e [0,T].
Transition Functions and Markov Processes
55
Theorem 1.11 Let Zt be a super-martingale on a filtered probability space (f2,.F, .F t ,P), and assume that Tt is an augmented right-continuous filtration. Then the function t i—• E [Zt] is right-continuous if and only if there exists a modification Zt of the process Zt, which is right-continuous and has left limits. It is clear that a martingale is also a supermartingale, and for a martingale Zt, the function t >—> E [Zt] is equal to a constant. Hence, Theorem 1.11 implies that every martingale on a filtered probability space satisfying the conditions in Theorem 1.11 has a modification that is right-continuous and has left limits. 1.10
Reciprocal Transition Functions and Reciprocal Processes
Let Xt be a Markov process with transition function P. Then the time reversed process Xt = Xr-t is a backward Markov process with backward transition function P given by P(T,A;t,y)
= P(T -t,y;T
-T,A)
(see Section 1.4). If the motion of a random system described by the process {Xs : T < s
F? = o-(Xs
:T<S
or v
<s
The family of c-algebras {J^ V J7?} will play an important role in the present section. Definition 1.18 The process Xs, s £ [r, t], is called a reciprocal Markov process on the interval [r, t] provided that for all u, s, and v with r < u <
56
Non-Autonomous Kato Classes and Feynman-Kac Propagators
s < v < t and all bounded Borel functions / on E, the equality E [/ (X3) \FZvr?]=E[f(X.)\
(Xu, Xv)]
(1.91)
holds P-a.s. Condition (1.91) is called the two-sided Markov property of the process Xs. The o--algebra a(Xu,Xv) in (1.91) is interpreted as the present information about the random system described by the process Xs, the cr-algebra J-£ V J^ contains the information about the system before the moment u in the past and the moment v in the future, and finally the cralgebra !F% contains the information about the system after the moment u in the past and the moment v in the future. The following lemma gives several equivalent conditions for the given stochastic process Xs to be a reciprocal Markov process on the interval
M]Lemma 1.20 Let Xs be a stochastic process on (Q, !F, P) with state space (E,£). Then the following are equivalent: (1) Condition (1.91) holds. (2) For all u and v with r < u < v
| a(Xri,...,Xrn,Xu,Xv)]
= E [f(Xs)
\ a(Xu, Xv)]
holds for any bounded Borel function f on E. (3) For all u and v with r
=V[A\ a(Xu, Xv)] P [B \ and B € F%.
a(Xu,Xv)]
Transition Functions and Markov
Processes
57
Remark 1.8 Condition 5 in Lemma 1.20 states that the future and the past are conditionally independent if the present is known. Here the present, the past, and the future are interpreted in the sense of reciprocal processes. Condition 5 was used in [Jamison (1970); Jamison (1974)] as the definition of a reciprocal process. Proof.
(1) ==• (2).
This implication follows from the inclusions cr(Xu, Xv) C a (Xri,...,
XTn, Xu, Xv) C T^ V T%.
(2) = • (3). Assume that condition (2) in Lemma 1.20 holds. In order to prove the implication (2) = » (3), we will first show that the equality E[F\a(Xri,...,Xrn,Xu,Xv)]
=E[F\
(1.92)
holds if the random variable F has a special form. Let m
F=
Y[fj(Xaj), i=i
where fj, 1 < j < m, is a bounded Borel function on E and u < s\ < S2 < • • • < sm < v. We will prove equality (1.92) by induction with respect to m. Since we assumed that condition (2) holds, (1.92) is valid for m = 1. Next, suppose that (1.92) is true for m G N and for all finite sets {SJ : 1 < j' < rn} with u < s\ < S2 < • • • < sm < v. Let {s\,..., s m + i } be a subset of [u, v] with m + 1 elements. Without loss of generality, we can assume that u < sj < S2 < • • • < s m +i < v. Put x — a (Xri,..., Xrn, Xu, Xv) and 7i — a {Xri,...,
XTn,XSl,...,
XSm,Xu,
Xv).
Then the tower property of conditional expectations gives m+l
E
n h (*.,) i ^ =
= E
E
n/j(^)E[/„,+i(xara+1)|w]|^ J'=l
(1.93)
58
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
It follows from condition (2) in Lemma 1.20 that E [fm+i (XSm+1)
\H]=E
[fm+i (XSm+1)
\a(X3m,Xv)]
.
(1.94)
Therefore, (1.93) gives m+l
E
n fj(x») i ^
= E l[fi(Xaj)E
[fm+1 {XSm+1)
\a(XSm,Xv)]
| f
(1.95)
Our next goal is to prove that there exists a bounded Borel function g on E x E such that E [fm+i (XSm+1)
\a(X3rn,Xv)]=g(XSrn,Xv)
(1.96)
P-almost surely. Actually, we will prove a more general known assertion: If F is a bounded a (XSm, Xv)-measurable random variable on O, then there exists a bounded Borel function g on E x E such that F =
g(XSm,Xv)
(1.97)
P-almost surely. Denote by H the class of all real functions F on Q for which there exists a bounded Borel function g on E x E such that equality (1.97) holds. The class H is a vector space containing the constant functions. Moreover, the product of any two functions from 7i belongs to H, and any function of the form XA (XSm)xB (Xv) with A G £ and B G £ belongs to TC. Next, let hn € 71, n 6 N, be an increasing uniformly bounded sequence of positive functions. Then
K = gn{xSm,xv), neN, where gn is a bounded Borel function on E x E. It is easy to see that without loss of generality we can assume that the sequence gn is increasing as n —> co (otherwise, we can replace the sequence gn by the sequence max gi). Then, Ki
sup K = ( sup£„ ) (XSm, n
\ n
J
Xv),
Transition Functions and Markov
Processes
59
and thus sup n hn £ H. It follows from the monotone class theorem for functions (Theorem 5.2) that any bounded a (XSm, X„)-measurable random variable on fi belongs to H. Therefore, equality (1.96) holds. It is clear that (1.95) and (1.96) imply that m+l
E
= E 3= 1
n / J W S (*.».*») i-F
(1.98)
3=1
P-almost surely. Applying the induction hypothesis to the right-hand side of (1.98) and using (1.94) and the tower property of conditional expectations, we get m+l
E
ruwi^
E X\f3(X3.)g{XSm,Xv)\a(Xu,Xv)
3= 1
E
JJfjiX^E
3= 1
[fm+1 (XSm+1)
| a{XSm,Xv)]
|
a(Xu,Xv)
3= 1
:E
Y[fj(XSj)E
[fm+1 (XSm+1)
| 7i] |
a(Xu,Xv)
3= 1
m+l
E E
n f*x'i) i n
a \XU, Xv)
3= 1
m+l
= E
(1.99)
J\f3{XSj)\a{Xu,Xv) 3= 1
By the monotone class theorem and (1.99), E{F\
T]
=E[F\a(Xu,Xv)}
(1.100)
for any bounded ^-measurable random variable F. Next, applying the monotone class theorem again and using the definition of conditional expectations, we obtain E[F\j^\/J^}=E[F\a(Xu,Xv)}. Therefore, condition (3) in Lemma 1.20 holds. (3) =* (4). This implication follows from the definition of conditional expectations, the
60
Non-Autonomous Kato Classes and Feynman-Kac Propagators
formula poo
G=
roo
X{G->x}d\, (1.101) Jo Jo and Tonelli's theorem. (4) = • (5). Suppose that condition (4) in Lemma 1.20 holds, and let A G T^ V T%, B G F%, and D G o(Xu,Xv). Then, using condition (4) with G = XA and F = XBXD, we obtain E
X{G+>x}d\-
[XCXAXB]
= E
[XAE [X£>XB
I o'v-Xui-Xu)]]
= E [XDXAE [XB J ff(X„, X„)j] .
(1.102)
Therefore, E
[XAXB
I
= E [XAE [XB I < T ( X U > X „ ) ] | a(X u ,X„)] = P [A | a(Xu,Xv)} P [B | ff^.X,,)] , (1.103)
and hence condition (5) in Lemma 1.20 holds. (5) = > (1). Suppose that condition (5) in Lemma 1.20 holds. Then, comparing (1.102) and (1.103), we see that (1.102) holds. Let D = Q and G = f(Xs). It follows from (2.110) that E\XAf{X3)}
= E [XAE [f(Xs)
| a{Xu,Xv)}}
.
Next, using the definition of conditional expectations, we get E [f(Xs)
| Tl V F t \ = E [/(*,) |
a(Xu,Xv)]
for all s with u < s < v. Therefore, condition (1) in Lemma 1.20 holds. This completes the proof of Lemma 1.20.
•
It has already been established in Section 1.2 that the Markov property is invariant under time-reversal (see Lemma 1.2 and Lemma 1.3). In fact, more is true. Lemma 1.21 Let (£],.F, P) be a probability space, and let Xt with 0 < t < T be a Markov process with respect to the measure P. Then Xt is a reciprocal process with respect to P. Proof. We will prove that equality (1.91) holds. Suppose that 0 < r < u<s
Transition Functions
and Markov
Processes
61
J^-measurable random variable. Let / be a bounded Borel function on E. It follows from the properties of conditional expectations that E[Yf(Xs)Z\a(Xu,Xv)] = E [E [Yf(Xs)Z = E[f{X,)E
| a (XU,XS,XV)}
\ a (Xu, Xv)]
[YZ \a(Xu,Xs,Xv)]
\a(Xu,Xv)]
.
(1.104)
Next, using the properties of conditional expectations and the Markov property of the processes Xr and XT-T (see Lemma 1.2 and Lemma 1.3), we see that E[YZ\a(Xu,Xs,Xv)] = E [YE [Z | a (Y, Xu,X„
Xv)] \ a
= E [YE [Z | a {Xv)] | a {XU,XS, = E [Z | a (Xv)} E[Y\a
Xvj\
(XU,X3,XV)]
=
E[Z\a(Xv)]E[Y\a(Xu)]
=
E[Z\cr(Xv)]E[Y\a(Xu,Xv)]
= E[YE[Z\a(Xv)}
(XU,XS,Xv)]
\a(Xu,Xv)]
= E [YE [Z | a (Y, Xu, Xv)} | a (Xu, Xv)}
= E [E [YZ J a(Y,XU,XV)}
|
a{Xu,Xv)}
= E[YZ\a{Xu,Xv)}.
(1.105)
Therefore, (1.105) and (1.104) give E [Yf(Xs)Z
| a(XU,XV)]
= E [YE [Z \ a(Xu,Xv)} = E[Y\a
\
{Xu, Xv)} E[Z\a
a(Xu,Xv)} (Xu, Xv)] . (1.106)
Now it is clear that (1.106) implies (1.91), and hence Xr is a reciprocal Markov process. This completes the proof of Lemma 1.21. • The remaining part of the present section is devoted to reciprocal transition probability functions. Such a function describes the evolution of a random system as follows. The value of a reciprocal transition probability function is the probability of finding the system inside a set B € £ at an intermediate moment of time knowing the location of the system at the initial and the final moments of time.
62
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
Definition 1.19 A function Q(T, X; S, B; t, y) where 0 < r < s < t
Q(T,x;v,dz;t,y)Q(T,x;u,C;v,z)
JD
= / Q (T,x;u,dw;t,y)Q Jc
(u,w;v,D;t,y).
(1.107)
The equation in (iii) is an analogue of the Chapman-Kolmogorov equation for transition probability functions. The following equalities follow from (1.107): Q(T,x;v,D;t,y)=
/ Q(r,x;u,dw;t,y)Q(u,w;v,D;t,y)
(1.108)
JE
for all 0 < r < u < v < T and D £ £; and Q(T,x;u,C;t,y)=
/ Q (r,x;u,C; v,z)Q (r,x; v,dz\ t,y)
(1.109)
JE
for all 0 < T < u < v < T and C £ £. The equality in (1.107) can be interpreted as follows. Suppose we know that a random system with the transition law Q is at x £ E at the initial moment r and at y £ E at the final moment t. Let u and v be such that r < u < v < t, and let C £ £ and D £ £. Then the probability of the system hitting the target C at the moment u and then the target D at the moment v in forward motion is equal to the probability of hitting the target D at the moment v and then the target C at the moment u in backward motion. L e m m a 1.22 Let Q(T,x;s,A;t,y) be a reciprocal transition probability function. Then for all (t, y) £ [0, T] x E, the function Pt,v{T, x; s, A) = Q(T, X; S, A; t, y)
Transition Functions and Markov Processes
63
is a transition probability function, while for all (r, x) e [0, T] x E, the function PT,x{s,A;t,y)=Q(T,x;s,A;t,y) is a backward transition probability function. Proof. we get
Using condition (iii) in Definition 1.19 with C = E and D = A,
Pt,v(T, x; s, A) = Q(T, X; s, A; t, y) " /
Q (r, x; s, dz; t, y) Q (r, x; u, E; s, z) Q
•
(T, X; U,
dw; t, y) Q (u, w; s, A; t, y)
/
- /
Pt,y (r, x; u, dw) PttV (u, w; s, A).
(1.110)
On the other hand, using condition (iii) in Definition 1.19 with C = A and D — E, we obtain PT,x(s, A-1, y) = Q(T, X; s, A; t, y) = / JA •
/
Q(T,x;s,dw;t,y)Q(s,w;v,E;t,y) Q
(T, X;
v, dz; t, y) Q (r, x; s, A, v, z)
= I PTiX{v,dz;t,y))PTyX{s,A;v,z).
(1-1H)
JE
Now it is clear that Lemma 1.22 follows from (1.110) and (1.111).
•
Let Q be a reciprocal transition probability function. Next, we will construct a stochastic process Xt associated with Q. Fix (T, X) € [0, T] x E and (t, y) 6 [0, T] x E with 0 < r < t < T, put Q, = E^°'T\ and consider the process Xs{ui) = w(s) for all ui £ Q, and 0 < s < T. Let T be the cr-algebra generated by finite-dimensional cylinders. Then there exists a probability
64
Non-Autonomous Kato Classes and Feynman-Kac Propagators
measure P(T,t),(z,j/) °n i^,^7)
sucn
P(r,t),(x,v) [XT G A0, XSl
= XA0(x)XAn+1(y)
/
that
£ Ai,...,
XSn
G An, Xt
G
An+i]
Q{T,x;S!,dzi;t,y)
JA1
(1.112)
where Ai G 5 for all 0 < i < n + 1 and T < s\ < • • • < sn < t. The existence of the measure P(r,t),(x,y) follows from the Kolmogorov extension theorem, since the expression on the right-hand side of (1.112) defines a projective system of measures. This fact is a consequence of the equalities in (1.108) and (1.109). More precisely, suppose that for some k with 1 < k < n we have Ak = E. Then (1.109) implies the equality
Q (sk-i, Zk-i; " /
sk+1,C;t,y)
Q(sk-i,Zk-i;sk,dzk;t,y)Q(sk,Zk;sk+i,C;t,y),
C G £.
In other words, the number of integrations on the right-hand side of (1.112) can be reduced by one. If k = n, then applying (1.108) we arrive at the same conclusion. Now it is not hard to see that (1.112) defines a projective system of measures. A special case of (1.112) is
p
(T,t),(*,») [ * « € A\ = Q(T> x< s> A\ *. v)
(L113)
for all 0 < T < s < t < T, x € E, y G E, and A G £. Formula (1.113) corresponds to the case n — 1. The initial distribution of the process X3, T < s < t, is the Dirac measure 5X, while the final distribution of X8 is 5y. Since the present state of the reciprocal process X3, T < s < t, can be interpolated from the past and future information, the joint distribution of XT and Xt plays an important role. Next, we will construct a reciprocal process Xt with a prescribed joint distribution /i of XT and Xt. Let ^ be a probability measure on £
Transition Functions and Markov Processes
65
(Q, F) for which P(r,t),/i [XT 6 AQ, XS1 S A i , . . . , XSn € An, Xt € A n + 1 ] = // dfi(x,y) J JA0xAn+i JAi
Q{T,x;sudz1;t,y) d^jt.y)
•Ma
(1.114)
-M„
where Aj £ £ for all 0 < i < n + 1 and T < Si < • • • < sn < t. The existence of the measure P(T,t),/n follows from the Kolmogorov extension theorem. Note that our previous notation P(T,t),(x,y) is nothing else but P(T,t),*rx<5B- We will call the measure p, the initial-final distribution of the process Xs on the interval [r, t]. T h e o r e m 1.12 Let Q be a reciprocal transition probability function, and let the measures P(r,i),(x,y) be defined by (1.112). Then the following conditions hold for the process Xs: (1) For allO
xeE,yeE,
Ae £, and B € £,
(r,t),(x,y) [XT € A, Xt € B) = 5X X Sy(A X B).
(2)ForallO
E(u,vUxM
[f(Xs)},
(1.116)
and E :r,t),(»,v) [/(*.) I K v J7] = j f(y)Q («, *«;», <*y;«, *„) • (i-H7) (4) For allO
66
N'on-Autonomous Koto Classes and Feynman-Kac Propagators /loWP( Ti t),(z,y)-a.S..-
E(T,0,<*,„) [F I K V f%\ = E(r,t),(x,„) [^ | a (Xu, XV)]
(1.118)
and E(r,t),(x,„) [^ | ^
VT*t] = E{u>vUXM
[F].
(1.119)
Proof. Condition (1) in Theorem 1.12 is a corollary of (1.112). Our next goal is to prove that condition (2) holds. Fix T < u < s < t, and let AQ,AI,..., An+i be an arbitrary finite family of Borel subsets of E. By the monotone class theorem and the definition of the conditional expectation, condition (2) follows from the following assertion: The equality P(r,t),(x,j/) [XT € A), Xri G Ai,...,
Xrk_1 £ Ak-i, Xrk G Ak,
Xrk+i € Ak+1, • • • , -X'rn £ ^ n , Xt G A n + i J = E(r,t),(a;,i/) [Q (W. Xu] S, Ak\V,Xv),XT Xrk~i
£ Ak-i, Xrk+1
holds for all finite subsets {r\,..., that
£ A0,Xri
€ A\, . . . ,
G Afc+i,. . •, XTn G v4 n , Xt G A n + i J (1.120)
r n } of the set (r, t) \ {(u, s) U (s, v)} such
r < r i < • • • < rfc_i = u < r f e = s < w = r f c + i < • • • < rn < t. It is not hard to see that the equality in (1.120) is a consequence of (1.112). We will next prove condition (3) in Theorem 1.12. Note that (1.101) with / instead of G and Tonelli's theorem imply equality (1.117). The fact that the expressions on the right-hand side of (1.116) and (1.117) are equal follows from (1.112). This shows that the function (x,y) >-> E(T)t))(X)j,) [/ (Xs)} is Borel measurable. Therefore, the right-hand side of (1.116), that is, the expression TE>(u,v),(xu,xv) [f {Xs)], is measurable with respect to the aalgebra a (XU,XV). Using the definition of the conditional expectation and formula (1.112), we see that the expressions on the right-hand side of (1.115) and (1.116) are equal. For any s with u < s < v, put F — f (Xs). Then the equalities in part (4) of Theorem 1.12 can be obtained from part (3) of this theorem. Equality (1.118) follows from part (3) of Lemma 1.20. The fact that the expressions on the right-hand side of (1.118) and (1.119)
Transition Functions
and Markov
Processes
67
coincide is a consequence of (1.112) and the monotone class theorem. More precisely, (1.112) implies that E
(r,t),(*,v) [F I o-(XU,XV)]
= KMl(Xu,Xv)
[F]
(1.121)
for F — XA-, where A = {XT G AQ, Xn G Ai,...,
Xrk^
G A fc -i, Xrk G Afc,
-^i-fc+i G j4fc+i, • • •, -Xrn S A n , Xt G A n + i } . Here each Aj, 0 < j < n + 1, is a Borel subset of E and r < r\ < • • • < T-fc_i = u
AeS,
and B
e£,
P(T,t),M [*r €A,Xt£B]=
n(A x B).
(2) For all 0 < r < u < s < v
P (r>t)>/1 -a.s.
(3) The two-sided Markov property with respect to the measure P(r,t),M holds for the process Xs. This means that for allO
I K V J ? ] = E(T,t),M [f{X.)
| o- (Xu, Xv)]
P(T,t),M-o.a.
The next part of the present section is devoted to the relations between Markov processes and reciprocal Markov processes. Let Q be a reciprocal transition probability function. It is said that Q possesses a density q if there exists a function q (T, X; s, Z; t, y) such that for all r, s, and t with 0
/ q(r,x;s, JA
z;t,y)dm(z)
(1.122)
68
Non-Autonomous Kato Classes and Feynman-Kac Propagators
holds. In (1.122), m is the reference measure (see Section 1.3). The following construction goes back to Schrodinger (see [Schrodinger (1931)]). Let P be a transition function possessing everywhere positive density p. Here we do not assume that P is normal. Consider the so-called derived density 9(T l;S 2;i y)
'
'
'
=
p(T,»;t,y)
'
(1 123)
'
where 0 < T < s
*-('. *;*.») =[2n(t-r)]^eXP{Jir7j}-
(L124)
The reciprocal transition probability function Q with derived density q given by (1.123) is associated with a reciprocal Markov process Xs satisfying the conditions in Theorem 1.12. The density q is nothing else but the conditional density of the process Xs subject to the conditions XT = x and Xt = y. Indeed, for any open neighborhood U of the point y € E, we have P
\X ^A\X
GlJ]-
Vr,AXs€A,Xt€U]
pr,x [x.eA\xteir\_
T
IAXUP( >
X
S
Z
Prx [Xt e u}— S
> ' I)P( '
Z 1
^'
z
2)dzidz2
Jup{r,x;t,z2)dz2 By shrinking U to y, we get the following formula: P
rx- <= A I y - „1 -
IAP(T'x's'z^P(s,zi;t,y)dz1
Let jtx be a Borel probability measure on E x E, and let Xs be the reciprocal Markov process on [r, t] with derived density q and such that /i is its initial-final distribution; that is, l*(A xB)=
P (Tit)iM [XT G A, X t E B]
(1.125)
for all Borel subsets A and B of E. We will denote by fiT and /xt the marginal distributions of /x. They are defined as follows: fiT(A) = /i(A x E)
Transition Functions and Markov Processes
69
and Ht{B) = ^{E x B). For all u and v with r < u < v < t, the symbol HUtV will stand for the joint distribution of the random variables Xu and Xv with respect to the measure P(T)t),M. Jamison posed and solved the following problem in [Jamison (1974)]: Determine under what restrictions on the initial-final distribution fi a reciprocal Markov process Xs on the interval [T, t] associated with the derived density q is a Markov process with respect to the measure P(T,t),/xT h e o r e m 1.14 Let p(r,x;s,y) be a transition density, and let q(T,x;s,z;t,y) be the derived density given by (1.123). Fix r and t with 0 < T < t < T, and let Xs be a reciprocal process on [T, t] with transition density q and the initial-final distribution \x given by (1.125). Then the following are equivalent: (1) Xs is a Markov process with respect to the measure P(T)t),M(2) There exist a-finite Borel measures uT and vt on E such that KG) -
If
P(T, x-1, y)dvT{x)dvt(y)
(1.126)
for all Borel measurable subsets G of E x E. Proof. (2) =$• (1). Suppose that there exist measures vr and ut for which condition (1.126) holds. Let T < si < S2 < • • • < sn < t and Ai G £ for all 1 < i < n. Given n > 1, put E x. E x Ai x ••• x An-i
= J4^_!.
Then, using equalities (1.123) and (1.126) and making cancellations, we get P(T,t),„ [XSi e Au 1 < * < n] = /
q{T,x;si,zi;t,y)...q(sn-i,zn-i;sn,zn;t,y) q (sn-i, Zn-i; sn, zn; t, y) d[i(x, y)dzi...
L
dzn-idzn
p (r, x; si, z i ) . . .p (s„_i, Zn-i; sn, zn)
A„-i
p(s
xA„
\t,y)dvT{x)dvt{y)dzi.
..dzn-idzn.
(1.127)
Denote by / the following function: f(zn-i)
=
J A n X E p ( s n - i , 2n-i! sn, zn)p(sn, zn\ t, JEP(sn-l,Zn-i;
t, y)dvt{y)
y)dzndvt(y)
70
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
The function / depends on s n _i, sn, and An. It follows from (1.127) that P(r,t),,i [X,i € At, 1 < i < n) = /
dvT(x)dvt(y)p(T,x;si,Zi)
. . .p (sn-2,
L
Zn-2', Sn-i,
Zn-i) p (sn-i,
Zn-i;t,y)
f(zn-i)dZi
.
..dZn-i
dp,(x,y)q(T,x;si,z1;t,y)
.. .q (sn-2, zn-2; sn-i, zn-i\t,y) = E (Tit)iM [/ {Xtn_x)
f(zi)dzi
...dzn-i
,XSieAi,l
(1.128)
Therefore, (1.128) gives P(r.o,/x [Xsn G An | Tln_^
= /(XSn_J
= P ( T i t ) i / 1 [XSn G A , | a ( X S n _ J ] .
(1.129)
The cases where s\ = r and s n < i; or si > r and s n = t; or si = r and sn = t are similar. Now it is not hard to see that the Markov property holds for Xs. (1) = » (2). Let us suppose that the reciprocal process Xs is simultaneously a Markov process on [r, t] with respect to the measure P(T,t),M. For all s with r < s
l(s, A, B, C) = p(Ti0iM [XT eA,xse
B,
xteC\.
Then the following equalities hold: I(s,A,B,C)=
/
dp.(x,y) /
JAxC
= 1 JAXC
q(T,x;s,z;t,y)dz
JB dli(Xiy)
f JB
P(r,*;s,z)p(s,z,t,y)dz_
P\T,x;t,y)
On the other hand, since Xs is a Markov process, there exists the Markov transition P*(u, Xu;s, A). It follows that / dfj,T(x) / Pf{T,x;s,dz)Pf(s,z;t,C). (1.131) JA JB Let u and s be such that T < u < s < t. Then, since the process Xs is simultaneously reciprocal and Markov, and the Borel cr-algebra of E is I(s,A,B,C)=
Transition Functions and Markov
Processes
71
countably generated, we have pf (u,x;s,B)
=
pf (u,x;t,dy)
q(u,x;s,z;t,y)dz
(1.132)
for /i u -almost all x £ E and all B £ £. Let us denote by AQ the set of all x £ E such that (1.132) holds for all B &£. Then /x„ (S\A 0 ) = 0. Our next goal is to prove that for /i u -almost all x £ E, the measure B \r-> P? (u,x;s,B) and the reference measure m are mutually absolutely continuous. Indeed, let m(B) = 0. Then (1.132) gives Pf(u, x; s, B) = 0 for ^ - a l m o s t all x G E. On the other hand, if x £ AQ and P? (u, x; s, B) = 0, then (1.132) implies Pf (u, x; t, dy) q(u, x; s, z\ t, y) = 0.
/ dz
Since p is a strictly positive function and P* (u,x;t,E) = 1, we get m(B) = 0. This completes the proof of the mutual absolute continuity of the measures B H-> pf (u, x; s, B) and m. It follows from the reasoning above that Pf(u,x;s,B)=
pf(u, x; s, z)dz
(1.133)
JB
for all x G A0 and B G £. In (1.133), pf(u,x;s,z)=
/ Pf(u,x;t,dy)q{u,x;s,z;t,y). JE We will next prove that the measure D >-> P? (u, z; t, D) is absolutely continuous with respect to the measure fit for /i u -almost all z G E. Indeed, we have Ht{D) = f dfj,u(x)Pf (u, x; t, D)
(1.134)
for all D G £. Suppose that C G £ is such that /ut(C) = 0. Then it follows from (1.134) with D = C that Pf (u,x;t,C)=0
for /x„-almost all x £ E.
(1.135)
Using the Markov property, we see that for every D £ £, Pf (u,x;t,D)=
Pf (u,x;s,dy)Pf
(s,y;t,D)
(1.136)
/uu-almost everywhere on E. Let us denote by A\ the set of those x £ E for which (1.136) holds for all D £ £. It follows from the separability of £
72
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
that /u„ (E\Ai) = 0. Fix xi £ A0r\Ai such that (1.135) holds for xi. Note that xi depends on the set C. Then (1.136) and (1.135) with x = x\ show that Pf(s,y;t,C)
=0
(1.137)
for Pf (u,xi; s, -)-almost all y G E. Since xx £ AQ, equality (1.137) holds for m-almost all y G E. Now we see that for all XQ G AQ n A\, equality (1.137) holds for P ' (u, x0; s, -)-almost all y & E. Then (1.136) with D = C gives Pf(u, xo; t, C) = 0. We have thus completed the proof of the fact that the measure D H-» P ^ (U, Z; £, Z?) is absolutely continuous with respect to the measure fit for /i u -almost all z G E. Indeed, the equality Pf(u, xo; t, C) = 0 holds for all C with Ht(C) — 0 for all XQ belonging to the set AQC\A\. This set is independent of C and such that fiu (E\ {AQ n A{)) = 0. It follows that there exists a function P such that Pf («, z; t,D)=
[ Pf (u, z; t, y) d(it(y)
(1.138)
JD
for all £> G £ and for /i u -almost all z £ E. Now it is not hard to show that for every u G [r, £) there exists a function (z,y) >—> P* (u,z;t,y) such that it is £ (g> immeasurable and coincides with the function (z, y) i-> P (u, z; i, y) /xu x jut-almost everywhere. Indeed, using (1.134) and (1.138), we get / dnu(z) JE
/ P (u,z;t,y)dnt(y)=
Ht(E) = l,
JE
and hence, there exists a probability measure v on £ ® £ such that iv(Ax.D)= / d/xu(z) / P"'
(u,z;t,y)dfj.t(y).
JD
JA
This measure is defined as follows. For any set G G £ <8> £, v{G)=
P~f
/ d// u (z) / JE
(u,z;t,y)dfit(y)
JGZ
where Gz = {y G E : (z,y) G G}. It is easy to see that the measure v is absolutely continuous with respect to the measure fiu x fit. Let us denote An
~
the Radon-Nikodym derivative —. r by PA Then for all u with d (/xu x lit) T < u < t, the function (z, y) H-> P-f (U, Z; i, y) is £ <8> ^-measurable, and moreover, / dnu{z) \ P JA
JD
(u,z;t,y)dni (y)
Transition Functions and Markov
73
Processes
= f dfiu(z) f Pf(u,z;t,y)dtH(y) JA
(1.139)
JD
for all Borel sets A and D in E. It follows from (1.139) and from the separability of the cr-algebra £ that for every u with r
= Pf
(u,z;t,y)
for fj,u x //t-almost all (z, y) G E x E. Let us fix s with r < s < t. Then, using (1.130), (1.131), and (1.133), we get [
f
MXiy)
JAXC
P(r,x;s,z)p(s,z;t,y)dz
JB
P(T,x;t,y)
I pf(T,x;s,z)dzPf(s,z;t,C).
= I d(ir(x) JA
(1.140)
JB
Now it would be natural to use equality (1.138) with u = s in (1.140). However, we only know that equality (1.138) holds for /^-almost all z £ E. The following equality justifies the use of (1.138) with u = s in (1.140): fi3(B) =
dnT(x)pf
(T,X;S,B)
— / dz / d^T{x)pf
(T,X;S,Z)
.
Therefore, f
Mx,y)
JAXC
[
P(r^s/z)p(s,z;t,y)dz
JB
P{T,x]t,y) f
= / dfj,T(x) \ p (r,x;s,z)dz JA
JB
/ P (s,z;t,y)dfj,t(y).
(1-141)
JC
It is not hard to see that (1.141) implies , , . p(r, X; S, z)p(s, z; t, y) , , , . , . . f, .—f. ., dn(x, y) —7 dz = diiT{x)dpLt{y)pI(r, x; s, z)P (s, z; t, y)dz. P(T,x;t,y) Hence, there exists ZQ E E such that the following representation holds for
J i \ i 4. \Pf(T>x'>s>zo) . / \ „ p (s,zo;t,y) dfi(x,y) = p{T,x;t,y)—± r-dfiT(x) x — -—Ldfit{y). p{T,x;s,z0) p{s,z0;t,y) This shows that the implication (1) =>• (2) holds. The proof of Theorem 1.14 is thus completed.
•
74
Non-Autonomous Koto Classes and Feynman-Kac Propagators
R e m a r k 1.9 Assume that the conditions in Theorem 1.14 hold. Then (1.129) implies 1Rr P I T-rl - SBP(s>x>'>t>y)dvt(v)El>.tUx.,y)[F] Hr,t),n [? \
»
, . (1.142)
where the random variable F is measurable with respect to the cr-algebra Tl- Indeed, the function / appearing in the proof of Theorem 1.14 can be represented as follows: fAnXEp(sn-i,zn-i;t,y)q(sn-1,zn_1;sn,zn;t,y)dzndi>t(y) J {z-n-i) = '—~ fEP(Sn-l,
Zn-1] t,
y)dvt(y)
s
JEP {sn-i, Zn-i; t, y) dvt{y)Q ( s n - i , Zn-ii n, An; t,y) dz, IEP(
;t,y)dvt{y)
(1.143)
It is not hard to see that (1.129) and (1.143) imply • (r,t),A»
•X-Sr, G Ann \ J•> s„ a.
fEp(sn-i,XSn_1;
t, y) dvt{y)Q (sn-i,XSn_1; fEp(sn-i,
sn, An; t, y) dzn
X8n_1; t, y)dut{y)
IEP{s^-^XSn_1]t,y)dvt{yyS'(sn_1,t),(xSri_1,y)
[XSn <S An] .
(1.144)
JE P(«n-i, ^ „ _ ! ; t, y)dut(y) Therefore, (1.142) follows from (1.144) and from the Markov property of the process Xs. A formula, similar to formula (1.142), holds in the case of a general probability measure \i on £ x £. Lemma 1.23 Suppose that the conditions in Theorem 1.14 hold. Then the following are true: (1) For all T < r < s < t and all J-* -measurable random variables F,
%,*),, [F I JT] /
JEXE
P{T,X;r,Xr)p(s,Xa;t,y)E{s,tUxs,y)
[
JEXE
[F]
p(r,x;r,Xr)P(s,Xs;t,y)^^
d x
^ 'f>
P\J-,x\t,y)
P{r,x;t, y)
(1.145)
Transition Functions and Markov
Processes
(2) For all T < r < s
75
random variables F,
E(r,t),M [F | F.] [
p(T,x;r,Xr)p(s,Xt;t,y)E(T
JExE
f{x'f
[F] d
p(T,x;r,Xr)p(s,Xs;t,y)
^ f
p[T,x;t,y)
Note that the measurability properties of the random variable F in parts (1) and (2) of Lemma 1.23 are different. Proof. Part (1). Let r < s\ < S2 < • • • < Sfc < Sfc+i • • • < sn < t, and let Ai G £ for all 1 < i < n. Then, using (1.123) and (1.126) and making cancellations, we get P(T.t).M [X"i
£Ai,l
= /
q{T,x;si,Zi;t,y)...q(sn-i,Zn-i;sn,zn;t,y) q (s„_i, z n _i; sn, zn; t, y) dfj,(x, y)dzx...
dzn^xdzn
d^(x,y) p (T, X; SI, zi).. .p (s„_i, Zn-i; sn, zn) p(r,x;t,y)
I
ExExAiX-xAn
V ( 5 n , zn \t,y)dzi...dzn-\dzn.
(1-147)
For the sake of shortness, we will use the following notation: B = E x E x Ak+i
diiix
v]
x • • • x An and du(x, y) = , .. p (T, X; t, y)
Let us define the function / by f{zi,Zk) I p {T,X; s\, zi)p (sk, zk; sk+i, zk+i) • • •p(sn,zn;t,y)dvdzk+i _
• --dzn
JJB
/
p(T,x;si,zi)p(sk,zk;t,y)di/
JExE
This function depends on si, Si with k < i < n and Ai with k + 1 < i < n. It is easy to see from (1.147) that W(T,t),v, [X3i G Ai, 1 < i < n] = /
p(r,x;si,zi)...
JExExAi_x---xAk
p (sfc, zk; t, y) f(zi, zk)dv(x, y)dzx
...dzk
76
Non-Autonomous
J
Koto Classes and Feynman-Kac
Propagators
ExExAiX--xAk
...q(sfc-i,2fc-i;sk,zk;t,y)
f(zi,zk)dfi(x,y)dzi
...dzk
= E(T,t),M [/ (XS1 ,XSk),X3ieAul
(1.148)
It follows from (1.148) that P ( r , t ) i / I [X3k+1 G Ak+1,...,
X3n G An | J £ ] = / (X S1 , XSk).
(1.149)
Moreover, we have / p (r, x; si, ^i)p (sfe, Zjt; s fe+ i, z fe+ i) • • -p (sn, zn; t,y) du(x, y)dzk+i
---dzn
JB f
„, „. _ . x P(T,X;S!,ZI)
IEXE JE
dfj,(x,y)
p(r,x;t,y)
/ P (sfc, zk;t, y) q {sk,zk; sk+i, zk+i; t, y) J' A, Ak+iX---xAn • q(sn-i,zn-i;sn,zn;t,y)dzk+i • • -dzn
L
ExE
p(r,x; SI, zi)p{sk,
zk;t,y)
dfi(x, y) P{T,x;t,y)
[XSk+lGAk+l,...,X3neAn].
(1.150)
Now the definition of the function / , (1.149), and (1.150) in the case where si = r and sn = s give (1.145) for all random variables of the form n
F= n xAAxSi). i=k+l
By the monotone class theorem, equality (1.145) holds for all bounded ^"/-measurable random variables F. It is not hard to see that this implies equality (1.145) for all ^/-measurable random variables F. Equality (1.146) can be obtained from equality (1.145) by using time-reversal and noting that the reciprocality is preserved under time-reversal. This completes the proof of Lemma 1.23. • Let VQ and v? be c-finite Borel measures on E, and let p be a strictly positive transition density (not necessarily normal). A pair (fo, v-r) is called an entrance-exit law provided that f J ExE
p (0, z; T, y) du0(x)duT{y)
= 1.
(1.151)
Transition Functions and Markov Processes
77
It follows from Jamison's theorem (Theorem 1.14) that there exists a one-toone correspondence between entrance-exit laws and initial-final conditions fi on £ x £, for which the process X3 associated with the derived reciprocal density q is Markovian with respect to the measure P(O,T),M- This one-toone correspondence is given by K
[
VT) <=>H(AXB)=
p(0, x; T, y)du0{x)duT{y)
(1.152)
JAxB
for all A € £ and B e£. Given an entrance-exit law (^o, VT), we define the functions h and h by h{s,x)=
I
p(s,x;T,y)dvT(y)
JE
for all s with 0 < s < T and x € E; and f>-(s,y) = /
p(0,x;s,y)dvo(x)
JE
for all s with 0 < s < T and y £ E. These functions are strictly positive, but it is not excluded that h(s, x) or h(s, x) may be infinite. However, for all s with 0 < s < T, the function x H-> h (s, x) is finite m-almost everywhere. This follows from the equalities / dv0(z) / dxp(0, z; s,x)h(s,x) JE
= /
JE
p(0, z;T,y)dv0{z)dvT{y)
= 1-
JExE
Here we used the strict positivity of p, the Chapman-Kolmogorov equation, and the definition of the entrance-exit law. If s = 0, then we can only claim that h(0,x) is finite i^o-alrnost everywhere. Similarly, for all 0 < s < T, the function x — i > h(s,x) is finite m-almost everywhere, and the function I H / I (T, x) is finite j/r-almost everywhere. Next, we will introduce the transition functions P 1 and Pi. Unlike the transition functions in Section 1.3, the functions P 1 and Pi are defined almost everywhere. More exactly, we put P 1 (r, x; t, A) = —^— H[T,X)
/ p (r, x; t, y) h {t, y) dy
(1.153)
jA
for 0 < r < t < T, x £ E, and A € £. Moreover, we set P 1 (r, x; T, A) = — ^ -
/ p (r, i ; T, y) ^ T ( y )
fi (T> a:) 7v4
(1.154)
78
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
for 0 < r < T and x E E. For every T and t with 0 < T < t < T and every A E £, the function x H-> P 1 (r, x; t, A) is defined m-almost everywhere with the exceptional set independent of t and A. On the other hand, for all 0 < T < T, the function x H-+ P 1 (T, X; T, A) is defined i/0-almost everywhere with the exceptional set independent of A. Put
P1(T,A;t,y)=
/ /i(r,a;)i)(r,a;;t,2/)da;^JA n(t,y)
(1.155)
for 0 < r < t < T and i / e £ , and
P1(0,A;t,y)
= [ p(0,x;t,y)dvo(x)T-!—7A h(t,y)
(1.156)
for 0 < t < T and i / e £ . Then for every r and £ with 0 < T < t < T and every A E £, the function y >—• P\ (T, A; i, y) is defined m-almost everywhere with the exceptional set independent of r and A. On the other hand, for all 0 < t < T, the function y — i » Pi (0, A; t, y) is defined i/^-almost everywhere with the exceptional set independent of A. It is easy to check that the function P 1 is a transition probability function and the function P\ is a backward transition probability function, if we exclude the exceptional sets described above. Theorem 1.15 Let p be a strictly positive transition density, and let (VQ, VT) be an entrance-exit law with respect to p. Let \L be the probability measure defined in (1.152), and denote by q the derived reciprocal transition probability density associated with p. Let P 1 and P\ be given by (1.153)(1.156). Then the process Xt(ui) = u{t), u> € Cl, 0 < t < T, where Q. = E\°'T>, has the following three realizations: a reciprocal process with the entrance-exit law (i and the reciprocal transition density q; a Markov process with the initial condition h (0, x) duo (x) and the transition function P 1 ; and a backward Markov process with the final condition h (T, x) dvT (x) and the backward transition function P\. Proof. Let 0 < s\ < s2 < • • • < sn < T and Ai E £ with 0 < i < n + 1. Then it is not difficult to see that the finite-dimensional distributions of all
Transition Functions and Markov Processes
79
the processes described in the formulation of Theorem 1.15 are given by
/
dfj,(x,y) /
JA0xAn+i
J
q{0,x;s1,z1;T,y)q(s1:zi;s2,z2;T,y) AiX--xAn
•••q (s n _i, zn-i; sn, zn; T, y) dz\... = /
dzn
du0(x)p(0,x;si,z1)p(si,z1;s2,z2)
JA0x---xAn+i
• • -p(sn, z„; T, y) dzx...
dzndvT(y).
This completes the proof of Theorem 1.15.
•
Theorem 1.15 is taken from [Nagasawa (2000)] (see Theorem 3.3.1 in [Nagasawa (2000)]). A triplicate nature of the stochastic process Xt is clearly seen from Theorem 1.15. The three representations of the process Xt, that is, the reciprocal, the Markov, and the backward Markov are called the Schrodinger, the forward Kolmogorov, and the backward Kolmogorov representation of the process Xt, respectively. Many of the ideas discussed in this section go back to Schrodinger, Bernstein, and Kolmogorov. These ideas found applications in quantum mechanics (see the references in Section 1.13).
1.11
P a t h Properties of Reciprocal Processes
Let p be a strictly positive transition density, and denote by q the corresponding derived transition probability density given by (1.123). Fix (r, x) € [0, T] x E and (t, y) e [0, T] x E, and suppose that the measure H = 5X x 5y is the initial-final distribution in formula (1.114). The resulting measure will be denoted by f(T,t),(x,y)- This measure satisfies
P(r,t),(T,y) [XT e AQ, XSl e A\,..., = XA0(x)XAn+1(y)
/
XSn e An, Xt s
Q(T, X\ si,dzx ;t,y)
J Ax Sn — 1) Zn — 1 i &n i
dzn;t,y) JA2
J An
An+X\
80
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
(see (1.112)). Using the definition of the derived density q and making cancellations, we obtain J
(r,t),(x,i/) [XT £ AQ, XSl £ Ai,..., P(r,x;t,y)
XSn £ An, Xt £ An+{\
JAlx-xAn
• • -p{sn,zn;t,y)dzi
•••dzn.
(1.157)
If the transition density p satisfies the normality condition, then an equivalent form of the previous equality is W(T,t),(x,y) [XT £ AQ, XSl £ Ai,..., =
XSn £ An, Xt £ An+i]
XA
° i g ) X A ; + 1 ! V ) Er,x [p (aw, X.n; t, y), XSl £ Ax,..., P\T, x, i, y)
XSn £ An].
Moreover, using the monotone class theorem, we get P r„T P ™'<*»> [ A ] ~
ET,x[XAP(s,Xs;t,y)} P(r,x;t,y)
(1 158)
'
for all A £ ?J with T < s < t. If, in addition, a strictly positive function p(r, x;t,y) is simultaneously a forward and a backward transition probability density, then
*™<*-v)
[B]
-
(1 159)
p(r,x;t,y)
'
for all B £ Tl with r < s
°
n
and UT =
**>•
p{0,x0;l,yo) Then it is easy to see that (vo, VT) is an entrance-exit law. Hence Theorem 1.15 can be applied. Since h (s, x)=p (s, x; T, yo) and Ji (s, y) =
p (0, x0; T, y0)'
we see that the process Xs(w) = w(s), w £ Q, 0 < s < T, in Theorem 1.15 has the following three properties. It is a reciprocal process with M = $x0 x $vo a s * n e entrance-exit law and q as the reciprocal density; a
Transition Functions and Markov Processes
81
Markov process with the initial condition 6Xo and the transition function P
T,VO
s i v e n °y P
T,yo (T,x;t,A)=
„.T,,\
p(T,x;t,y)p(t,y;T,y0)dy
= f q(T,x;t,y;T,y0)dy
(1.161)
JA
for 0 < T < t < T, x G £ , and 4 G £, and by P^yo(r,a;;r,J4)=x^(2/o)
(1.162)
for 0 < r < T, x G £ , and A G £\ and, finally, a backward Markov process with the final condition Syo and the backward transition function Px 'x° given by P°,X0 (T,i4;t,y) = — —- / p{0,x0;t,y) JA
p(0,xo;T,x)p(T,x;t,y)dx
L
q(0,xo;T,x;t,y)dx
(1.163)
for 0 < T < t < T and y G E, and by ??>Xo(0,A;t,y)
= XA(xo),
(1.164)
for 0 < t < T, y G E, and A G £. As before, we denote by P(o,r),(s0,i/o) t n e measure on Q = E^°'T^ associated with the entrance-exit law (1.160) and the reciprocal transition density q; by PQ'%° the measure on fi corresponding to the transition function PTIVO given by (1.161) and (1.162) and the initial condition 5Xo; and by P 0 '"° the measure on fi associated with the backward transition function P®>x° defined by (1.163) and (1.164) and with the final condition 5Vo. If p is a transition density, then we denote by Po,x0 the measure on Q. associated with the density p and the initial distribution SXo. If p is also a backward transition probability density, then we denote by P r ' y o the measure on Q associated with the backward transition density p and the final distribution 5yo. By Theorem 1.15, the measures P(o,T),(xo,2/o)> ^o,'xa' anc ^ ^o,'xo° c ° i n c i d e on the (7-algebra Tj.. It is not hard to prove that if p is a transition probability density, then the measure P 0 'x° is absolutely continuous with respect to the measure Po,x0 o n every cr-algebra 3-^ with 0 < r < T. Similarly, if, in addition, p is a backward transition probability density, then the measure PQ'"" is
82
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
absolutely continuous with respect to the measure FT'yo on every c-algebra T^ with 0 < r < T. Suppose that under certain restrictions on the transition density p the following conditions hold: (a) There exists a continuous P(O,T),(X0,I/O)" modification of the forward Kolmogorov representation of the process Xt on the half-open interval [0,T); (b) l i m t j r X ( = Vo ]P(o,T),(:ro,!/o)~amiost ev~ erywhere. Then the same restrictions guarantee the existence of a continuous modification of the reciprocal process Xt on the closed interval [0, T]. Similarly, suppose that under certain restrictions on p, the following conditions hold: (a) There exists a P( 0) T),(xo,!/o)" mo ^ mcat i on °f t n e forward Kolmogorov representation of the process Xt, which is right-continuous and has left limits on the half-open interval [0, T); (b) l i m t j r X t = j/o F>(o,T),(x0,2/o)"ahTlost everywhere. Then the same restrictions imply the existence of a P(o,T),(x 0 ,!/o)" mo< ^ mcat i on °f the reciprocal process Xs that is right-continuous and has left limits on the closed interval [0, T]. We will use these ideas in the proof of the following assertion. T h e o r e m 1.16 Let p(r,x;t,y) be a strictly positive function that is simultaneously a forward and a backward transition probability density, and let XQ € E and yo £ E be given. Denote by q the derived reciprocal transition probability density associated with p, and consider the Schrddinger representation of the process X t (w) = w(i), u> £ fi, 0 < t < T, on the space £1 — J5[°>T] with respect to entrance-exit law (1.160) and the reciprocal transition density q. Suppose that for all e > 0, lim / *i°
lim
p(0,x0;t,y)dy
= 0,
(1.165)
JGt(x0)
sup /
p(r,x;t,z)dz
— 0,
t-TlO;0
lim
sup /
(1.166) V
p(T,z;t,y)dz
= 0,
;
(1.167)
t-rl0;0
and lim/
p(t,y;T,yo)dy
= 0.
(1.168)
Then there exists a W,(Q>T),(x0,ya)-rn°dification Xt of the process Xt that is right-continuous and has left limits on [0,T].
Transition Functions
and Markov
Processes
83
Proof. Conditions (1.165) and (1.166) in the formulation of Theorem 1.16 guarantee the existence of a Po^xo'^odification Yt of the process Xt that is right-continuous and has left limits on the half-open interval [0, T) (see Remark 1.7). Since the measure P 0 'x° is absolutely continuous with respect to the measure P 0]Xo on every cr-algebra J® where 0 < r < T, there exists a P0'^"-modification Yt of the process Xt which is right-continuous and has left limits on the half-open interval [0,T). Since P(0,T),(*o,yo)=<xo>
the process Yt is a P(o,T),(xo,2/o)"m°dification of the process Xt on [0,T). On the other hand, since p is a backward transition probability density, conditions (1.167) and (1.168) imply the existence of a P T ' yo -modification Zt of the process Xt that is left-continuous and has right limits on the half-open interval (0,T] (see Remark 1.7). Using the absolute continuity of the measure Pj£'"° with respect to the measure FT'yo on every cr-algebra Fj, with 0 < r < T and the equality P(0,T),(x o ,j/o) = P 0 , x o ° '
we see that there exists a P(o,T),(To,yo)"mo(lification Zt of the process Xt on the interval (0, T] that is left-continuous and has right limits on the interval (0, T}. Using the fact that the processes Yt and Zt are P(o,T),(x0,3/o)~ modifications of the process Xt, we see that there exists a P(o,T),(x0,?/o)" modification Xt of the process Xt having right and left limits on the closed interval [0,T]. Finally, by redefining the process Xt as we did at the end of the proof of Lemma 1.16, we get a process Xt satisfying the conditions in Theorem 1.16. • The next assertion concerns continuous modifications of reciprocal processes. Its proof is similar to that of Theorem 1.16, and we leave it as an exercise for the reader. Theorem 1.17 Let p (r, x; t, y) be a strictly positive function which is simultaneously a forward and a backward transition probability density, and let xo e E and yo £ E. Denote by q the derived reciprocal transition probability density associated with p, and consider the Schrodinger representation of the process -Xt(w) = w{t), OJ € £1, 0 < t < T, on the space Q, = £ , I 0,r l with respect to the entrance-exit law (1.160) and the reciprocal transition
84
Non-Autonomous Kato Classes and Feynman-Kac Propagators
density q. Suppose that the following conditions hold for every e > 0:
lim- /
p(0,x0;t,z)dz
= 0,
(1.169)
lim
sup/
p(T,y;t,z)dz
= 0,
(1.170)
lim
sup /
p(T,z;t,y)dz
= 0,
(1-171)
t-TlO;0
and !lm^7
/
p{t,z;T,yo)dz = 0.
(1.172)
TTien t/iere exists a continuous P(o,r),(x0iVo)-modification Xt of the process In [Jamison (1975)], a different condition is used to ensure the validity of the equality limXt = yo- In the next theorem, Jamison's condition is employed instead of the conditions in (1.167), (1.168), (1.171), and (1.172): T h e o r e m 1.18 Let p (T, X; t, y) be a strictly positive transition probability density, and let XQ E E and yo 6 E. Denote by q the derived reciprocal transition probability density associated with p, and consider the Schrodinger representation of the process Xt(ui) = u>(t), u> G Q., 0
sup
^TyeGc(yo)
p(t,y;T,yo)
=0
for all e > 0. Then there exists a F^o,T),(x0,y0)'mo^fica^on Xt that is right-continuous and has left limits on [0,T].
(1.173) Xt of the process
T h e o r e m 1.19 Let p (r, x; t, y) be a strictly positive transition probability density, and let XQ G E and yo £ E. Denote by q the derived reciprocal transition probability density associated with p, and consider the Schrodinger representation of the process Xt(w) — w(t), w £ Q,, 0 < t < T, on the space £1 = £[°. r ] with respect to the entrance-exit law (1.160) and the reciprocal transition density q. Suppose that conditions (1.169), (1.170), and (1.173)
Transition Functions and Markov Processes
hold. Then there exists a continuous ¥^otT),(x0,vo)'mo^fica^on process Xt-
85
%t of the
Remark 1.10 It is not assumed in Theorems 1.18 and 1.19 that p is a backward transition probability density. Proof. Arguing as in the beginning of the proof of Theorem 1.16, we see that there exists a P ^ T ) , ^ , ^ - m o d i f i c a t i o n Y of the process Xt that is right-continuous and has left limits on the interval [0,T). We will next show that Jamison's condition (1.173) implies the equality lim.t-\T Xt = Vo P(o)T)i(Xo,!/0)-almost everywhere. The following lemma has an independent interest. A special case of this lemma will be used in the proof of Theorem 1.18. Lemma 1.24 —, '
T y
For every T and t with 0 < r < t < T, the process
' ' y w T < s < t, is a P(o i r),(a;o,yo)" mar ^ n 3 a ' e with respect to
the filtration TTS V a (Xt), r < s
We will first show that E (0,T),(xo,!/o)
p(T,XT;t,Xt)
[p(s,Xs;t,Xt)\
1.
"
(L174)
Indeed, using the normality condition for p and the Chapman-Kolmogorov equation twice, we get E,(0,T),(xo,2/o)
'p(T,XT;t,Xty p(s,X3;t,Xt)_
q (0, x 0 ; r, zx\ T, y0) q (r, zx;s, z2; T, y0) q (s, z2; t, z3\ T, y0) JE*
p(r,zi-t,z3)
p(s,zr,t,z3)
dz\dz2dz3
p (T, Z\ ; t, z3) p (0, x 0 ; T,z{)p (r, zx;s, z2) p (t, z3; T, y0) ,n r N dzldz2dz3 JB p (0, x0; T, 2/0) 1 ,n rr s dzidz3p{T,zi;t,z3)p{Q,Xo;T,zi)p(t,z3;T,y0) p(0,xo;T,yo) JBxE 3
=
=
—
7F, r—V / P(*> Z 3;T,2/0)^3 / p(0,xo;l,yo) JB JE
~H\ ^—V / p(t,z3;T,yo)p(0,x0;t,z3)dz3 p{0,xo;T,yo) JB This gives (1.174).
p(T,zi;t,z3)p(0,x0;T,zi)dzi = 1.
86
N'on-Autonomous
Kato Classes and Feynman-Kac
We will next prove that for all r < Si < s2 E,(0,T),(x0,yo)
p(r,XT;t,Xt) p(s2,Xa2;t,Xt)
Propagators
Kvv(Xt)
p(T,XT;t,Xt) p(si,XSl;t,Xt)
(1.175)
P(o,T),(xo,2/o)"a-s- Indeed, using the normality condition for p, the measurability of the random variable p (r, XT; t, Xt) with respect to the
(0,r),(x 0 ,2tt)
p(T,XT;t,Xt) p(s2,XS2;t,Xt)
^VoiXt)
= P (T, X T ; t, Xt) E(0,T),(xo,y0) -j———— P(s2,XS2;t,Xt) 1 p(s2,X32;t,Xt) 1 = P (T, XT\t, Xt) E{sut),(xn,xt) p(s2,XS2;t,Xt)_ q{s1,XSl;s2,z;t,Xt) = p(T,XT;t,Xt) / — dz JE P{S2,z;t, Xt) p(T,XT;t,Xt) ' ' v\ / P(si,XSl;s2,z)dz p(si,Xs i\l-> A t ; JE _ p(r,XT;t,Xt) p(sx,XSl;t,Xt)' = P (T> XT; t, Xt)
E(O,T),(X0,!/O)
\r3iVo{Xt)
.£?,
VJ4
This establishes (1.175). It follows from (1.174) and (1.175) that Lemma 1.24 holds.
(1.176)
D
We are now ready to finish the proof of Theorem 1.18. Using Lemma 1.24 with T = 0 and t — T, we see that the process — ——- is a p(s,Xs;T,y0) nonnegative P ^ r ^ ^ ^ - m a r t i n g a l e . By Theorem 1.3 and equality (1.174), this martingale converges to a finite limit as s f T P(o,r),(x0,s/o)"a'most: surely. It follows that p(s, XS;T, yo) converges to a strictly positive limit IP(o,r),(:ro,i/o)~amlost s u r ely as s T T. By condition (1.173), we see that this can only happen if Xs converges to yo P(o,r),(xo,vo)"a^mos* s u r e i v a s s | T. Combining the previous assertion with the fact that there exists a P(o,r),(xo,i/o)~mocufication Ya of the process Xs that is right-continuous and has left limits on the interval [0, T), we see that there exists a P(o,T),(x0,3/o)~ modification X3 of the process Xs that is right-continuous and has left limits on the closed interval [0,T].
87
Transition Functions and Markov Processes
This completes the proof of Theorem 1.18. The proof of Theorem 1.19 is similar. • If the transition density p is not normal, then the measures P r > x o and W'Vo employed in the proof of Theorem 1.16 do not exist. However, by changing the assumptions in Theorem 1.16 and taking into account the remarks before the formulation of Theorem 1.16, we see that the following assertion holds. Theorem 1.20 Let p(r,x;t,y) be a strictly positive transition density, and let xo € E and yo € E. Denote by q the derived reciprocal transition probability density associated with p, and consider the Schrodinger representation of the process Xt(u) = w(£), LJ £ CI, 0 < t < T, on the space ft = E^°'T' with respect to the entrance-exit law (1.160) and the reciprocal transition density q. Suppose that for all e > 0, lim/
lim t-Tio-fl
P(0,xo;t,y)p(t,y;T,yo)dy
suP-7 xeE
^
p ( r , x ; T , y0)
7/
= 0,
p(T,x;t,z)p(t,z;T,y0)dz
(1.177)
= Q,
JGe(x)
(1.178) lim sup-— —-r / p(0,xo;T,z)p(T,z;t,y)dz t-T|o ; o
= 0, (1.179)
and lim/ ^TJGe(yo)
p(0,xo;t,z)p{t,z;T,yo)dy
= 0.
(1.180)
Then there exists a P(0,T),(x0,y0)~modification Xt of the process Xt that is right-continuous and has left limits on [0,T]. A similar theorem holds for continuous modifications. It is based on Theorem 1.17. We leave this case as an exercise for the reader. In the next definition, we introduce pinned measures. Definition 1.20 Let (T,X) e [0,T] x E, (t,y) £ [0,T] x E, and let p be a strictly positive transition density on E. Denote by Xs the process T v XS{OJ) = w(s) on the space O = E^°' \ Then the measure p^ x on (Q, !FJ)
88
Non-Autonomous Koto Classes and Feynman-Kac Propagators
satisfying /#» (XT €A0,X3leAi,...,XSn
G An,Xt
= p(r,a;;i,y)P( T ,t),(x,y) [XT eA0,XSl = XA0{x)XAn+1(y)
/ J
G An+1)
eAi,...,X3n
eAn,Xt
€ An+i]
P(T,X;SI,ZI)P(SI,Z1-,S2,Z2) AI*....XA„
•••p(sn,zn;t,y)dzi...dzn
(1.181)
where r < si < S2 < • • • < sn < t and Ai G £ for 1 < i < n, is called the pinned measure associated with p, (T,X), and (£, y). It follows from (1.157) and (1.181) that M r " = P ( r . a1! *> 2/) P (r,t),(x,y)-
Remark 1.11 If the density JD is normal, then in addition to (1.181) the following formula holds:
/ 4 * (XT eAo,xaieAi,...,xSn G A„,xt e > W i ) s = XA0 0*0x,*„+1 (y)Er,x b ( n, -X"Sn;«, y), XSl e Ax,...,
x3n e A n ] . (1.182)
The pinned measure ^yx is defined on the measurable space ( E [ ° ' T 1 , !FJ). However, if Xs is a Markov process on a smaller path space, e.g., on the space of continuous paths or on the space of left- or rightcontinuous paths, then certain difficulties may arise at the endpoints r and t. To avoid these difficulties, the measure fi^x is usually restricted to the CT-algebra F[_ = a (Xs : T < s < t) in the case of right-continuous processes Xs having left limits, while in the case of left-continuous processes having right limits, the measure /4'x is restricted to the cr-algebra F{+ =a{Xs:r
<s
Lemma 1.25 Let 0
=
ETtX[Fp(u,X»]t,y)].
Transition Functions
and Markov
Processes
89
The proof of Lemma 1.25 is not difficult, and we leave it as an exercise for the reader.
1.12
Examples of Transition Functions and Markov Processes
In this section we discuss several well-known examples of stochastic processes (see Section 1.13 where more references can be found). We also introduce pinned processes or "bridges".
1.12.1
Brownian
motion
and Brownian
bridge
The state space E of Brownian motion is d-dimensional Euclidean space E.d equipped with the Borel cr-algebra B^d. The d-dimensional Lebesgue measure plays the role of the reference measure m on B.d. Recall that the d-dimensional Gaussian transition probability density Pd is defined by Pd(r, x; t, y) = ^
_* ^
exp { ~ ^ y
}
(see Section 1.11). Since pd depends only on the differences s = t — T and z = x — y, we will use the notation gd (s, z) = pd (T, X; t, y). The function Pd is a transition probability density associated with a time- and spacehomogeneous transition probability function. The density pd is normal, since ixi 2
r
(27r)d/2
/
e
2
dx = 1.
The fact that gd satisfies the Chapman-Kolmogorov equation can be obtained from the following well-known formula for the Fourier transform of the Gaussian density:
*<«•«- (dps
e K dx
L ~^ "
IKJi 2
For any (T, X) € [0, oo) x R d , there exists a measure P TjX on the path space
90
Non-Autonomous Koto Classes and Feynman-Kac Propagators
'°° with the finite-dimensional distributions given by
PTiX[XT eA0,XtleAi,...,xtn = XA0(X)
e An\
/
9d(ti,xi-x)gd(t2-ti,x2-xi)---
JA!X---xA„
9d (tn -tn-i,xn-
xn-i) dxi--- dxn,
(1.183)
where r < ii < t2 < • • • < tn, x e R d , and At e SRd for 0 < i < n. The Gaussian transition function P is defined by P{T,x;t,A)=
/ JA
gd(t-T,x-y)dy.
The function P satisfies condition (1.86) in Corollary 1.2. Indeed,
P(s,y;t,Ge(y))
1
/
__J__
f Js-yf,
^-^x:|x-y|>e(2^-S))^eXP\
*-*
=
2(t-S)/to
* ^ /:w>« (27r(t - *))-/» exp \~2(rr7)}dz
*ih ( ^ 1 M >^>- exp {"^} "u(1.184) and it follows from (1.184) that r hm
sup
4-s-^O+^jjd
P(s,y;t,Ge{y)) —^ = 0 t - S
for all e > 0. By Corollary 1.2, there exists a continuous stochastic process (Bt,T{ ,WTtX) with state space Rd such that its finite-dimensional distributions are given by (1.183). The space Q of all continuous paths from [0, oo) into Rd can be taken as the sample space in this case. The process Bt is called Brownian motion in M.d. It follows from (1.183) that PTtX [BT &Ao,BtleAlt...,Btn = P0,o [-Bo eAo-x,Btl-r
e An) eAi-x,...,BU-T
£
An-x\.
The measure P 0iX is denoted by P x . Brownian motion starting at x — 0 at moment T = 0 is called a standard Brownian motion, and the measure
Transition Functions and Markov
Processes
91
space (C (R+; Rd) , C, P 0 ,o), where C is the cylinder cx-algebra of C (R+; R d ), is called the Wiener space. We denote the components of Brownian motion Bt by B\, 1 < i < d. For a given pair (r, x) £ [0, oo) x Rd, Brownian motion has the following properties: (1) Every component B\ of Bt is a one-dimensional Brownian motion with transition density gi. The components of Bt are PTjX-independent. (2) P T > X [BT = x} = l.
(3) For T < ti < t2 < • • • < tn, the increments Btl, Bt2 — Btl, •••, Btn — Bt„_i of Brownian motion are P TiX -independent. (4) The increments of Bt are stationary; that is, for all r < s < t and h>Q,Bt—B3 has the same P^-distribution as Bt+h — Bs+h(5) Brownian motion is a Gaussian process with mean ET<xBt = x, r < t, x G R d , and the correlation matrix given by E r , a (Bj - x') (fl£ - x^') = (s -
T)
A (t -
T)
,
where r < s, r < i, 1 < i, j < d, and x = (a; 1 ,.. .,x d ) G Md (see [Revuz and Yor (1991)] for the properties of Gaussian processes). If Bt is a d-dimensional Brownian motion starting at x at time r, then —Bt is a d-dimensional Brownian motion starting at —x at time r. For a standard Brownian motion Bt, the finite-dimensional distributions of the processes Bt and tBi/t, t > 0, coincide. They also coincide with the finitedimensional distributions of the process Bt+h — Bh for every h > 0. Moreover, for any a > 0, the finite-dimensional distributions of the process Bt are the same as those of the process a~1^Bat (scale-invariance). If Q is an orthogonal dxd matrix, then the process QBt is a d-dimensional standard Brownian motion (orthogonal invariance). If Brownian motion Bt starts at x at time r, then Bt + y is a Brownian motion starting at x + y at time r (translation-invariance). A d-dimensional Brownian motion Bt is a P TlX -martingale with respect to the filtration F[. There are several other interesting martingales related to Brownian motion, for instance, if Bt is Brownian motion starting at x at time r, then the process |-B t | 2 —dt,t> T, is a P TjX -martingale with respect to the filtration Tl', r < t. If d = 1, then for every a > 0, the process exp {aBt - \aH) is a P T);c -martingale with respect to the filtration J~1 •, T
92
Non-Autonomous Koto Classes and Feynman-Kac Propagators
Next, we will discuss the reciprocal process and the pinned measure associated with Brownian motion. Fix T > 0. Since the Gaussian density is strictly positive, we can define the derived transition probability density qd using formula (1.123). An important link between the reciprocal process with transition density qd and the original Brownian motion is provided by the following formula: qd(T,x;s,z;t,y) (t — s)x + (s — r)y 2 z—
1 d/2
(S-T)(t-S) 2TT
exp <
t-T
\
>. (1.185)
,(s-r)(t-s) t-T
t-T
This formula can be established as follows: qd (r, x; s, z; t, y) 9d(s-T,zx) gd(t- s,zgd(t-T,xy) 2TT(S - r)(t -
y) \x-y\2 \x - z\2 2(t - T) ~ 2(s -T)~
exp
s)
\z-y\2' 2(t - s) _
(t — s)x + (s — r)y 2 ^ z—
1 2TT
(S - T)(t - S)
d/2
exp <
t-T Cs-T)(t-s) t-T
t-T
where 0
x) gd(t-
= 9d(t-T,x-y)
/
s,z-
y) dz
qd(T,x;s,z;t,y)dz
JRd
=
9d(t-T,x-y).
Let (r, x) e [0, T] x Rd and (t, y) G [0, T] x Rd. Since the density gd satisfies the conditions in Theorem 1.17, there exists a continuous reciprocal stochastic process B,^vl, T < s < t, such that its reciprocal transition probability density is equal to the derived density qd obtained from the
Transition Functions and Markov
Processes
93
density gd as in formula (1.123). The process B]fcy\ 3 is called the Brownian bridge between (r, x) and (t,y), or Brownian motion pinned at x at time T and at y at time t. Lemma 1.26
The finite-dimensional distributions of the process
*.=*{$,..
r<s
with respect to the measure W(T,t),(x,y) coincide with the finite-dimensional distributions of the following processes: (a) The process Xl = V t ^ B ^ ^
+ ( l - ^ - \ x +
S
^ y ,
r<s
(1.186)
r<s
(1.187)
with respect to the measure F(o,i),(o,o)(b) The process x
l
= ^-B(.-r)(t-r) t—T t-s
+ ( 1 - ^ ^ J x + ^^y, y t—TJ t—T
with respect to the measure Po,o(c) The process X ^ V t ^ ^ B ^ - ^ B . y ^ l - ^ y + ^ y ,
r<s
with respect to the measure Po,oProof. The finite-dimensional distributions of the process Xs, r < s < t, are given by F
(r,t),(x,v) [*** € i4j : 1 < i < n] .
n—1
= /
qd{T,x;si,zi;t,y)Y[qd(si,Zi\Si+i,zi+i;t,y)dzi...dzn
JA1x--xAn
i = 1
(1.189) where r < si < s 2 < • • • < sn < t and Ai £ Bua for 1 < i < n. Put t — Si \ Wi =
t —T ( Zi t — Si V
t —T t — Si t —T
t —T x -
J
Si — T \ t —T y = Zi - x t —T I t — Si
Si — T t — Si
y,
94
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
and (Si-T)(t-T) t - Si
for 1 < i < n. Then the finite-dimensional distributions of the process X% are given by n Bj.j-THt-r) t-si
'0,0
P 0 ,o
G
n—1
9d (Si, wi) TT gd($i+i -5i,wi+1 r*
i ^ /•
TT -.— n
r
/ -i d
FT :
i=l
-wi)dw1...dwn
n—i
gdih^^Y]
9d(si+i-s,
u>i+i -Wi)dzi..
.dzn
»
/
dzi...dzn
' (si — r)(t — r ) £ — zr ffd I : .T i t — s\ t — si n-l
: l < i < n
G Ci : 1 < i < n
P
/ n
T
I A t — Si Si — T . Ai x y t — Si \ t —T t — T
Bj.i-rHt-T)
t ~
x
- -.
51—7" V
t — s\
(t-r)2(si+i-Si)
(t-r)zi+i
(t - s i + i ) ( i - Si)
t - s.»+i
(t-r)zi t - Si
(t - r)(s i + i - s»)y ( i - s i + 1 ) ( i - Sj)
(1.190) where r < s\ < S2 < • • • < sn < t and Ai G B^d for 1 < i < n. Put pi =
t — T t — Si
where 1 < i < n. Then, it is not hard to see that
formula (1.185) implies the equalities qd(r,x;si,zi;t,y)
= pdgd (pi(si
-r),pizi
- x - ~—y)
(1.191)
and qd(si,Zi;si+i,Zi+i;t, Pi+l9d
y)
I Pi+lPi (Si+l - Si) , pi+iZi+i
- PiZi -
pi+lp,
(Si+l -
Si)
t-T
(1.192)
Transition Functions and Markov Processes
95
for 1 < i < n—1. It follows from formulae (1.189)—(1.192) that the processes Xs and X% are stochastically equivalent. By the previous results, the process -Bj^o) x is stochastically equivalent to the process (1 - X)B * . Hence, the process B^
^^
is stochasti-
t —s
cally equivalent to the process
Bs=j_. Now the scaling properties of t — T *-' Brownian motion imply that the process yjt — T B 1 ' 0 S^J_ is stochastically equivalent to the process
(0.0), t_T
t —s
BU-TW-T) t — T t-s
. It follows that the process X]
is stochastically equivalent to the process X%. Finally, we will prove the stochastic equivalence of the processes X2S and X^. In the proof, we will use the fact that the process (1 — A) B \ is stochastically equivalent to the process B\ — XBi where 0 < A < 1. Indeed, it follows from the properties of Brownian motion that the processes B\ — XBx, \Bi — XBx, \Bi _!, and (1 — X)B x are all stochastically equivalent. S — 7"
Moreover, the process B ^ B\ is stochastically equivalent to the t—r t —T t—s process Ba=j_. This follows from the property of Brownian motion t — T *-« formulated above. Here we take A = s — r . Multiplying by y/t — T and t —T using the scaling property of Brownian motion, we see that the processes Xg and X% are stochastically equivalent. This completes the proof of Lemma 1.26. D The stochastic equivalence of the processes Xs and X% in Lemma 1.26 provides an alternative proof of the existence of continuous versions of Brownian bridges.
1.12.2
Cauchy process
and Cauchy
bridge
The transition density p^ of a d-dimensional Cauchy process is defined by pcd(T,x;t,y)
= cd{t-T,x-y),
0
xGMd,
y&Rd,
where
<*,(«, *) = r ( g ± l )
'
(s,,)e[o,oo)x^.
96
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
It is known that the following formula holds for the Fourier transform of cd(t,£) = e-m,
t>0,££Md.
(1.193)
It is not hard to see using (1.193) that the density pd is normal and satisfies the Chapman-Kolmogorov equation. A Markov process associated with the transition density pd is called a Cauchy process. Since the Cauchy density satisfies the conditions in Corollary 1.1, there exists a realization Xt of a Cauchy process that is right-continuous and has left limits. Cauchy processes have the following scaling property. For every a > 0, the process a~1Xat is P^o-stochastically equivalent to the process XtThe next lemma provides an example of a martingale related to a Cauchy process. Lemma 1.27 Let Xt be a Cauchy process. Then for every £ £ Rd, x G B.d, and T > 0, the process t >—> exp {i£ • Xt + £|£|} is a complex P T)X martingale on the interval [r, oo). Proof. By the Markov property and formula (1.193), we see that for 0 < s < t, ET,X [exp {i£ • Xt + i|£|} | Fs] = e'leiE.,*. [e***] i e*l«l1 / e*vCd (tfe-s,X y)dy = e *l€l e «-*. e -(t-.)iei ^cd(t-s,X ss-y)dy
= exp{i£-Xa + 8\t\}. This completes the proof of Lemma 1.27.
•
Our next goal is to define a pinned Cauchy process. Since the density pd is strictly positive, formula (1.123) can be used for a Cauchy process. According to this formula, the derived transition probability density qd is given by
,S(r,*;S,.;«,9) = r ( i ± I ) < i ^ i > ( i - T ) 2 + |x-y|2
7T ((* - T)2 + \Z - X\*) {(t - S)2 + \y - *P) _ (1.194)
Transition Functions and Markov Processes
97
where 0 < r < s < t < T and x, y, z e Rd. Fix (r,x) e [0,T] x Rd and (t, y) € [0, T] x Rd with T < £. Since the Cauchy density a satisfies the conditions in Theorem 1.16, there exists a reciprocal process -X/*'"\ , T < s
1.12.3
Forward bridges
Kolmogorov
representation
of
Brownian
Brownian motion and Cauchy process are homogeneous Markov processes. Next we give an example of a non-homogeneous transition density and nonhomogeneous Markov process. Such examples can be obtained from the forward Kolmogorow representation of Brownian bridges. Similar examples can be constructed using the forward Kolmogorow representation of Cauchy bridges and other pinned processes. Let XQ 6 Rd and yo 6 K d . In Subsection 1.12.1, we defined the Brownian bridge B^'^3, s £ [0,T], between (0,x 0 ) and {T,y0). In this case, the reciprocal transition density is given by the formula qd (r, x; s, z; t, y) (t - s)x + (s-
1 2 T ('-;W-«))
d/2
exp <
r)y 2 1
t-T
,(s-r)(t-5) t-T
>, (1.195)
where 0
T,y0{T^x'^^A)
= I JA
Qd{T,x;t,y;T,y0)dy
where 0 < T < t < T and qd is defined in (1.195).
98
1.13
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
Notes and Comments
(a) The following list is a sample of books devoted to Markov processes [Dynkin (1960); Dynkin (1965); Dynkin (1973); Dynkin (1982); Dynkin (2000); Dynkin and Yushkevich (1969); Bhattacharya and Waymire (1990); Blumenthal and Getoor (1968); Chung (1982); Chung and Zhao (1995); Doob (2001); Ethier and Kurtz (1986); Gihman and Skorohod (1974); Gihman and Skorohod (1975); Gihman and Skorohod (1979); Jacob (2001); Jacob (2002); Jacob (2005); Meyer (1966); Revuz and Yor (1991); Sharpe (1988); Stroock (2005)]. Our presentation of the path properties of Markov processes in Sections 1.8 and 1.9 is similar to that in [Gihman and Skorohod (1975)]. (b) Kolmogorov's papers [Kolmogorov (1931); Kolmogorov (1933)] are important early contributions to the theory of non-homogeneous Markov processes. (c) Progressively measurable stochastic processes were first introduced and studied in [Chung and Doob (1965)]. Theorem 1.4 concerning the existence of progressively measurable modifications of measurable processes was established in [Chung and Doob (1965)] (see also [Meyer (1966); Doob (2001)]). (d) Reciprocal processes were studied in [Jamison (1970); Jamison (1974); Jamison (1975)]. The concept of a reciprocal process and a reciprocal transition function goes back to Schrodinger (see [Schrodinger (1931)]) and Bernstein (see [Bernstein (1932)]). Reciprocal processes are used in stochastic quantum mechanics to model some aspects of the behavior of quantum mechanical systems. The readers who would like to learn more about reciprocal processes and their use in quantum mechanics may consult the following books: [Nelson (1967); Nelson (1988); Aebi (1996); Nagasawa (1993); Nagasawa (2000); Chung and Zambrini (2003)], and the following articles: [Cruzeiro and Zambrini (1994); Cruzeiro, Wu, and Zambrini (2000); Privault and Zambrini (2004); Privault and Zambrini (2005); Roelly and Thieullen (2002); Roelly and Thieullen (2005); Thieullen (1993); Thieullen (1998); Thieullen and Zambrini (1997); Thieullen (2002); Truman and Davies (1988); van Casteren (2000)]. An important paper on time reversal of Markov processes is [Chung and Walsh (1969)]. For more information on time symmetries of Markov processes see [Chung and Walsh (2005)]. (e) Brownian motion is probably the most popular example of a Markov process. For more information on Brownian motion see [Chung (1982);
Transition Functions and Markov Processes
99
Chung and Zhao (1995); Chung and Walsh (2005); Durrett (1984); Johnson and Lapidus (2000); Kahane (1997); Kahane J.-P. (1998); Karatzas and Shreve (1991); Revuz and Yor (1991)]. (f) Cauchy processes belong to the class of Levy processes. This means that they are right-continuous, have left limits, and their increments are independent and stationary. More precisely, a Cauchy process is a symmetric stable process of index 1 (see [Bertoin (1996); Sato (2000); Barndorffet al (2001); Schoutens (2003); Applebaum (2004)] for more information on Levy processes).
This page is intentionally left blank
Chapter 2
Propagators: General Theory
2.1
Propagators and Backward Propagators on Banach Spaces
Propagators are two-parameter families of bounded linear operators on a Banach space satisfying the flow condition (forward propagators) or the backward flow condition (backward propagators). Let B be a Banach space, and denote by L (B, B) the space of all bounded linear operators on B. The symbol / will stand for the identity operator on B. Definition 2.1
A two-parameter family {W (t, T) e L (B, B) : 0 < r < t < T}
is called a propagator on B provided that the following conditions hold: (1) W(t, T) = W(t, X)W(X, T) for 0 < r < A < t < T. (2) W{T,
T)
= I for 0 <
T
< T.
Conditions (1) and (2) in Definition 2.1 are called the flow conditions. Definition 2.2
A two-parameter family of operators {Q(T, t) € L(B, B):0
is called a backward propagator on B provided that the following conditions hold: (1) Q(r, t) = Q(T, X)Q(X, t) for 0 < T < X < t < T. (2) Q(t, t) = / for 0 < t < T. Conditions (1) and (2) in Definition 2.2 are called the backward flow conditions. 101
102
N'on-Autonomous
Kato Classes and Feynman-Kac
Propagators
There are simple relations between propagators and backward propagators. If T > 0 is given, and Q is a backward propagator on a Banach space B, then the family of operators defined by W(t,T) = Q(T-t,T-T),
0
(2.1)
is a propagator. Moreover, if Q is a backward propagator on a Banach space B, and a family of operators is defined by W (t, r ) = Q* (r, t) where Q* (T, t) is the adjoint of Q (r, t), then W is a propagator on the space B*. Here B* stands for the dual space of B. Similarly, if W is a propagator on B, and Q (r, t) = W* (t, T), then Q is a backward propagator on B*. A propagator W is called strongly continuous if for every x £ B, the B-valued function (£, r) —• W (t, r) x, 0 < r < t < T, is continuous. A propagator W is called uniformly bounded if \\W(t,T)\\B^B<M
for allO
If r and t are such that 0 < r < t < oo, and for every compact subset K of the set {(r, t) : 0 < r < t < oo}, the estimate \\W(t,r)\\B^B<MK holds for all (£,r) e If, then W is called a locally uniformly bounded propagator. A propagator W is called separately strongly continuous if for every fixed t and x € B, the function r —• W(t,r)x is continuous on [0,t], and for every fixed r and x £ B, the function £ —> M/(£, r)x is continuous on [r, T] (if T = oo, then we consider the interval [t, oo) instead of the interval [t, T}). Similar definitions apply in the case of backward propagators. The next theorem states that the joint continuity and the separate continuity are equivalent if forward or backward propagators are locally uniformly bounded. T h e o r e m 2.1 Let W be a propagator on a Banach space B. following are equivalent for W:
Then the
(i) The strong continuity, (ii) The strong separate continuity and the uniform local boundedness. The same assertion holds for a backward propagator Q on B. Proof. We will prove Theorem 2.1 for a backward propagator Q. The case of propagators is similar. By the uniform boundedness principle, condition (i) implies condition (ii). Next, let Q be a strongly separately continuous and locally uniformly
Propagators:
General
Theory
103
bounded backward propagator. Let 0 < r < t < T, and suppose t' and r' are close to t and r, respectively. We will first assume that t > r. Then for T' close to r-, we have t > T'. Using the local uniform boundedness condition and assuming that t' > t, we see that for every x £ B,
I=\\Q(r',t')x-Q(T,t)x\\B < \\Q(T\ t')x - Q(T', t)x\\B + 113(7-', t)x - Q(T, t)x\\B < \\Q(T', t)(Q(t, 0 * -x)\\B + \\Q(r', t)x - Q(T, t)x\\B < M \\Q{t, t')x - x\\B + \\Q(T', t)x - Q(T, t)x\\B . It follows from the separate continuity condition that lim
7 = 0.
(2.2)
t'—>t,T'—>T
If t' < t, then I < \\Q(T', t')x - Q(T', t)x\\B + \\Q(T', t)x - Q(T, t)x\\B < \\Q(r', t')(x - Q(t', t)x)\\B + \\Q(T', t)x - Q(T, t)x\\B < M \\Q(t', t)x - x\\B + \\Q(T', t)x - Q(T, t)x\\B , and we see that formula (2.2) also holds for t' < t. Finally, let r = t < T' < t'. Then the separate continuity condition implies that for every e > 0 there exists A > 0 such that A > r and ||Q(r,A)ar-ar|| B < e .
(2.3)
If follows from the local uniform boundedness condition and from (2.3) that I=\\Q(T,,t')z-z\\B < \\Q(T', t')x - Q(T', \)X\\B
+ \\Q(T', X)x - Q(r, \)x\\B
+ \\Q(T, X)X -
< | | Q ( T ' , t')(x - Q(t', \)x\\B
+ \\Q(T', X)X - Q(T, X)X\\B
+e
< M \\Q(t', X)x - x\\B + < M \\Q{t', X)x -
\\Q(T', X)X
Q(T, X)X\\B
+ \\Q(T',X)x-Q(T,X)x\\B < M \\Q(t', X)x + (M + l)c.
Q(T, X)X\\B
-
Q(T, X)X\\B
+ M \\Q(T,
X)X
x\\B
+e
- x\\B
+e + \\Q(T',
X)X
-
Q(T,
X)X\\B
(2.4)
In (2.4), M depends on t. Next we get from (2.4) and from the separate continuity condition that there exists S > 0 such that for r < T' < t' < T+S, we have I < (2M + 2)e. Therefore, (2.2) holds for r = t < T' < t'.
104
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
This completes the proof of Theorem 2.1.
D
A simple example of a propagator is as follows. Let St be a semigroup of linear operators on a Banach space B, and consider the following twoparameter family of operators: St-T, 0
=
St-T.
Then U is a propagator on B, and Y is a backward propagator on B.
2.2
Free Propagators and Free Backward Propagators
Recall that we denoted by E a locally compact Hausdorff topological space satisfying the second axiom of countability. Let p : E x E —> [0, oo) be a metric generating the topology of E, and let £ be the cr-algebra of Borel subsets of E. The symbol m will stand for the reference measure. The measure m is a nonnegative Borel measure on (E, £), and we always assume that 0 < m(A) < oo for any compact subset A of E with nonempty interior. We will next define various function spaces on the space E. The symbol BC will stand for the space of all bounded continuous functions on E equipped with the norm ||/|| 0 0 = 8 u p | / ( x ) | ,
feBC
(2.5)
x€E
We denote by Co the subspace of the space BC consisting of all bounded continuous functions on E vanishing at infinity. More precisely, a function / belongs to the space Co if for every e > 0 there exists a compact set Kcj in E such that |/(ai)| < e for all x G E\Kej. It is not hard to prove that Co is a closed subspace of the space BC. The symbol BUC will stand for the space of all bounded uniformly continuous functions on E. More precisely, a function / belongs to the space BUC provided that for every e > 0 there exists 5 > 0 such that for all x € E and y £ E with p (x, y) < S, the inequality \f{x) — f(y)\ < e holds. The space BUC is a closed subspace of BC. For 1 < r < oo, we denote by U£ the space of all Borel functions on E such that 11/11, = {J
\f(x)\rdx}
^oo.
(2.6)
The symbol Iff will stand for the space of all bounded Borel functions on
Propagators: General Theory
105
E equipped with the norm sup | / ( x ) | . xeE We denote by Lr, 1 < r < oo, the Lebesgue space on E with respect to the measure TO. The norm on the space U is defined by (2.6). We denote by £m the cr-algebra obtained by completing the Borel c-algebra £ with respect to the reference measure m. As usual, it is assumed that the elements of Lebesgue spaces are classes of equivalence modO of £mmeasurable functions on E with respect to the measure m. For r = oo, the norm of a function / G L°° is given by ll/lloo = e s s s u p x e s |/(a:)| where esssup^gg |/(x)| = inf {a : \f(x)\ < a
for m-almost all i £ f i } .
In the remaining part of the present section, we discuss backward propagators generated by transition functions. Let P (r, x; t, A) be a transition subprobability function, and define a family of operators on Lf by Y{TMX)
f/s/(»)^.*;Mv),
ifo
\/(a0,
ifr = t a n d O < r < T
for all x e E and / 6 L^. It follows from the subnormality condition that Y is a family of contraction operators on Lf. Moreover, the ChapmanKolmogorov equation shows that Y is a backward propagator on Lf. The family Y defined by (2.7) will be called the free backward propagator associated with the transition function P. Suppose that the transition subprobability function P possesses a density p; that is, there exists a nonnegative function p(r, x; s, y) such that JA
for all A £ £. In this case, the free backward propagator Y is defined on the space L°° by
Y(T1t)m
= {s*ny)p{T'x't'v)dv> ^/(x),
* ° ^ < ' *
r
(2.8)
if r = t and 0 < r < T
for all x e E and / e L°°. The operator Y(T, t) maps the space L°° into the space Lf.
106
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
If P is a backward transition subprobability function, then we define the free propagator associated with P by
[f{y),
\IT — t and 0 < r < T
for all x £ E and / £ Lf. If p is a backward subprobability transition density, then the free propagator has the following form:
[f(y),
IIT = t and 0 < r < T
for all x £ E and / £ L°°. Let us recall that in the case of a finite timeinterval [0,T], there exists a one-to-one correspondence P(T,x;t,A)
= P(T-t,A;T-T,x),
0 < T
between forward and backward transition functions. Therefore, if P is a given backward transition subprobability function, then the free propagator associated with P can be defined by U{t,T) = Y{T-t,T-r),
(2.11)
where Y is the backward free propagator associated with P.
2.3
Generators of Propagators and Kolmogorov's Forward and Backward Equations
In this section we continue the study of propagators on Banach spaces. Our goal is to show that propagators are related to non-homogeneous evolution equations. We will first discuss general propagators and then consider free propagators generated by transition functions. The non-homogeneous equations which are studied in the present section are called Kolmogorov's forward and backward equations. Let B be a Banach space and let ip be a B-valued function on the interval (0,T). If the function
symbol ——(r) will stand for its derivative from the right, that is, or dr
MO
h
Propagators:
General
Theory
107
Similarly, if the function ip is differentiable from the left, then the symbol — — ( T ) will stand for its derivative from the left, or v ;
dr
hio
-h
Suppose that Q = Q{r, t), 0 < r < t < T is a backward propagator on a Banach space B. For every r with 0 < r < T, consider a linear operator on B defined as follows: A
i \
i.
Q(T
— h,r)x
—x
.„ „ .
A-(T)x
= hm— ~ . (2.12) Mo h The domain of this operator is the set D(A-(T)) of all x £ B for which the limit in (2.12) exists. The operators A_(r), 0 < r < T, are called the left generators of the backward propagator Q. For every t € (0,T], denote by D-(t) the set of all x e B such that the function T H-» Q(r,t) is differentiable from the left on (0,t), and by .F(i) the set of all x £ B for which l i m Q ( t - M ) a : = x.
(2-13)
The next result explains in what sense the left generators are related to the backward propagator Q. Theorem 2.2 Let Q be a backward propagator on B, and let t G (0,T]. Then for every x £ D-(t)C\F(t), the function u (r) = Q(r,t)x is a solution to the following final value problem on (0, i):
f
£M--MrWr).
pi4)
limu(T) = a;. Proof. Let a; € £>-(£) ("1 F(t). Then, using the properties of backward propagators, we obtain d~u, , ,. Q(T - h,t)x -Q(T,t)x —— (T) = lim '-j5r fcio -h -- lim ^ Mo
T
~ H' T^T'
^X ~ g ^ T ' -/i
^
_lim(Q(r-/i,r)-/)Q(r,^ MO = -A-{j)Q{r,t)x
-h =
-A-(T)U{T).
108
Non-Autonomous Koto Classes and Feynman-Kac Propagators
In addition, the equality limu(r) = x follows from the definition of the set Tit
F(t). This completes the proof of Theorem 2.2.
•
Our next goal is to introduce the family of right generators of the backward propagator Q. For every T with 0 < r < T, consider a linear operator on the space B given by A+(T)X v
=
Q (r, r + h) x — x h
lim-
hj.o
(2.15)
The domain D (A+(T)) of this operator is the set of points x e B for which the limit in (2.15) exists. The operators A+(T), 0 < T < T, are called the right generators of the backward propagator Q. For every t € (0,T], denote by D+(t) the set of all x G B such that the function T H-> Q(r,t) is differentiable from the right on (0, t). T h e o r e m 2.3 Let Q{r,t), 0 < T < t < T, be a strongly continuous backward propagator on B, and fix t with 0 < t < T. Then for every x e D+(t), the function U(T) = Q (T, t) x is a solution to the following final value problem on (0, t): ( d+U
-(r) = dr limzi(r) = x.
-A+(T)U{T),
(2.16)
T\t
Proof. Let x € D+ (t). Then, using the strong continuity of the backward propagator Q, the Banach-Steinhaus theorem, and the definition of the set D+(t), we obtain d+u ,. _,. —— =hmQ(T,T OT
MO
,^d+u + h)—— OT
d+u Q(r + h,t)x-Q(T,t)x dr h Q(T + h,t)x - Q{r,t)x + limQ(T,T + /i) Mo h Q(T + h,t)xQ(r,t)x — lim Q(T,T + h) Mo h Q(T, t)x - Q(T, T + h)Q(r, t)x lim Mo h Q(r,T + h)-I = —lim Q{r,t)x. Mo h
= lim Q(T,T + h) Mo
(2.17)
Propagators: General Theory
109
It follows from (2.15) and (2.17) that Q(T, t)x G D (A+(T)) and the equation in (2.16) is satisfied. In addition, the equality limu(r) = x follows from the T"T*
strong continuity of Q. This completes the proof of Theorem 2.3.
•
Now we turn our attention to propagators on B. The generators in this case are defined exactly as in the case of backward propagators. Suppose that W is a propagator on a Banach space B. For every t with 0 < t < T, consider a linear operator on the space B given by 7 /^
i.
W(t
+ h,t)x-x
.„.,„,
A+(t)x = lim — i -^ . (2.18) v Mo h ' The domain D (A+ (t) J of this operator is the set of points x £ B for which the limit in (2.18) exists. The operators A+(t), 0 < t < T, are called the right generators of the propagator W. For every T e [0,T), denote by D+(T) the set of all x £ B such that the function t H-> W(t,r)x is differentiable from the right on (r, T), and by F(T) the set of all x € B for which limt^(T + /i,T)a; = ar. (2.19) Mo Theorem 2.4 Let W be a propagator on B, and fix T with 0 < r < T. Then for every x £ D+(T) n F(T), the function u{i) = W(t,r)x is a solution to the following initial value problem on (r, T):
^
w
= i + ( «<),
(22o)
limu(t) = x. Ur
Proof. Let x £ D+(T) n F(T). Then, using the properties of propagators and the definition of the set D+(T), we obtain d+~,.s ,. W(t + h,t)x-W(t,T)x -K7u(t) = hm —i '— dt MO h ,. W(t + h,t)W(t,r)x-W(t,T)x = hm ; MO h _ {W{t + h,t)-T)W(t,T)x Mo h = A+(t)W(t,T)x = A+(t)u(t).
110
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
In addition, the equality limu(i) = x follows from the definition of the set t\.T
F(r). This completes the proof of Theorem 2.4.
•
Let W b e a propagator on B. For every t with 0 < t < T, consider a linear operator on the space B given by r ,N ,. W(t,t — h)x — x A_(t)a; = lim————• . w MO h The domain D (A-(t))
,„.,, (2.21)
of this operator is the set of points x G B for which
the limit in (2.21) exists. The operators A-(T), 0 < t < T, are called the left generators of the propagator W. For every T G [0, T), denote by D-(T) the set of all x G B such that the function t H-> W(t, r) is differentiable from the left on (r, T). Theorem 2.5 Let W be a strongly continuous propagator on B, and fix r with 0 < r < T. Then for every x G D- (r), the function u (t) = W (t, T) X is a solution to the following initial value problem on (r, T): (
^ ( t )
=
A_(t)u(t),
dt
(2.22)
limw(£) = x. tJ.T
Proof. Let x G D-(t). Then, using the strong continuity of the propagator W, the Banach-Steinhaus theorem, and the definition of the set -D-(i), we obtain
HO
K
' [ dt
-h
hiO
-h
llmW(t,t-h)W^-h^X-W^X
= hiO
V
'
'
-h
W(t, T)X ~ W(t, t - h)W(t, T)X h.10
-h
J
Propagators:
General
Theory
111
It is not hard to see that (2.21) and (2.23) imply W(t,r)x G D (A-(t)Y Moreover, the equation in (2.22) is satisfied. In addition, the equality limu(t) = x follows from the strong continuity of W. This completes the proof of Theorem 2.5.
•
Our next goal is to discuss what happens if we differentiate Q and W with respect to "wrong" time variables. Let Q be a backward propagator on a Banach space B. For every r with 0 < r < T, put D*+{T)=
p|
D(A+(t)),
t:r
where D(A+(t))
is the domain of the operator A+(t) defined by (2.15).
Theorem 2.6 Let Q be a backward propagator on B, and fix r with 0 < r < T. Then for every x G D\(T) and t with r
=
Q(T,t)A+(t)x.
Proof. Let x G D*+{T). Then, using the properties of backward propagators and the definition of D+(T), we obtain d+Q(T,t)x dt
_
Q(T,t +
h)x-Q(T,t)x
hio
h Q(T,t)Q(t,t
hio
h)x-Q(T,t)x
Q(r,t)]imQ{t't+uh)x-X hio h
= =
+ h
Q{T,t)A+{t)x.
This completes the proof of Theorem 2.6.
•
For a propagator W an & Banach space B and t G (0, T], put
Dl(t)=
f|
D(A-(T)),
T:0
where D
(A-(T)J
is the domain of the operator A_(r) defined by (2.21).
Theorem 2.7 Let W be a propagator on B, and fix t with 0 < t < T. Then for every x G D*_{i) and r with 0 < T < t, d~W(t,r)x ^
=
,-r , , -W(t,T)A-(T)x.
112
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
Proof. Let x e D*_{t). Then, using the properties of propagators and the definition of D*_(t), we obtain d~W(t,T)x dr
_
W(t,T-h)x-W(t,r)x h
hio _
W(t, T)W(T,
r-h)x-
hio
W(t, T)X
h -W(t,r)hmW^T-k)X-X h
=
K
=
-W(t,T)A-(r)x.
' hio
This completes the proof of Theorem 2.7.
•
Next we will discuss Kolmogorov's forward and backward equations for transition probability functions. Given such a function P, let us consider the Banach space B = Lf and the free backward propagator Y defined by formula (2.7). Instead of the strong generators A_(r) (see formula (2.12)) and A+(T) (see formula (2.15)), we will consider generators in a weaker sense. The new generators have larger domains than those of the generators A_(r) and A+(T). We will use the topology a (Lf ,M) on the space Lf, where the symbol M stands for the space of all finite signed Borel measures on E. For every r € [0,T), denote by Dw ( A ^ ( r ) ) the set of all / e Lf for which the a (Lf, M)-limit
AM(T)f =
*(Lf,M)-limY{T-h'hT)f-f
exists. This means that there is a function A^(r)f
€ Lf
//"w^°a//(r"T)/"^
such that
<2-24>
for all v e M. The operators A^(T), 0 < r < T, are called the left a (Lf, M)-generators of the backward propagator Y. The next simple lemma provides an equivalent condition for the convergence of a sequence of functions in the topology a (Lf ,M). Lemma 2.1 A sequence of functions hk € Lf converges to a function h G Lf in the topology a(Lf,M) if and only if lim hk(x) exists for all k—HX>
x e E, and sup \hk(x)\ < oo. k,x
Propagators: General Theory
Proof. that
113
Suppose that hk —> ft in the topology a(Lf,M).
This means
lim / hkdu = / hdu ^<*>JE
(2.25)
JE
for all ;/ G M. Put 1/ = 5X. Then the sequence ftfc converges to ft pointwise on E. Moreover, the uniform boundedness of the sequence {hk} follows from the Banach-Steinhaus theorem. Now assume that the sequence hk converges pointwise on E to ft, and moreover, it is uniformly bounded. Then h £ Lf, and by the dominated convergence theorem, we see that hk —* h in the topology a (Lf,M). • For every r with 0 < r < T and every / £ Dw (A™ (r)), we have lim Y(r-h,r)f(x)-f(x) hio h
= AM{T)f{x)
(2i2g)
and sup |y^-^^)/(')-/WI (/i,x)6(0,r)x£ ft
< oo.
(2.27)
This follows from Lemma 2.1. By (2.26), (2.27), and the dominated convergence theorem, for all / £ Dw (A^(T)), the pointwise derivative from the left i (x) of the function r H-> K(T, t)f(x) coincides with its a (^|°, M)-derivative from the left. This means that for all u € M, lim / y(r-h,t)f(X)-Y(r,t)f(x) hioJE ft Y(T-h,t)f(x)-Y(r,t)f(x) I im Y{r-h,mx)-Y(r,mx)Mx) = / llim h IE JE L° ft Let t G (0,T], and denote by F™(t) the set of all functions / G Lf such that KmY(t-h,t)f(x)
= f(x)
for all x €. E. The uniform boundedness of the family of functions Y (t — h,t) f in the space Lf follows from the definition of the free backward propagator Y. By Lemma 2.1, Y (t — h,t)f —> / in the topology a (Lf,M) as ft | 0. For any t with 0 < t < T, denote by D™(t) the set of all functions / G Lf such that the function r i-» Y(T, t)f is cr (Lf,M)differentiable from the left on (0, £).
114
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
Theorem 2.8 LetO
?-1(T,X)
=
-AM(T)U(T)(X)
9T limu(r, x) = f(x)
(2.29)
Tit
for all x G E. Remark 2.1
In final value problem (2.29), the derivative ——(r, x) and or the equality limw(T, x) = f(x) are understood in the sense of pointwise convergence, or in the sense of convergence in the space a (I/g°, M). Proof. Let / e D™ (t) n F™{t). Then, using the properties of backward propagators and the definition of the set D^(t), we see that for every v e M, t u{r-h)-u{r)dv hioJEE -hn —
Hm
Y{r-h,t)f-Y(r,t)fdiy h (Y(T-h,T)-I)Y(T,t)f = - lim / dv nh IE
=
_]]mr HoJ E h-lUJE
-L
A?(T)Y(T,t)fdl>
= - f A"(T)u(r)dv.
(2.30)
JE
Now it is easy to see that Theorem 2.8 follows from (2.26), (2.28), (2.30), and the definition of the set F^{t). This completes the proof of Theorem 2.8. • The equation in (2.29) is called Kolmogorov's backward equation. Kolmogorov's equation is a linearization of the Chapman-Kolmogorov equation. One can find examples of transition probability functions by solving Kolmogorov's backward equation. This can be done, for instance, if the set D^(t) is large enough, and if final value problem (2.29) is uniquely solvable. More precisely, suppose that for every open set O C E, there exists a sequence /„ G D^(t) such that sup |/„(a;)| < oo and lim fn(x) — Xo(x) for n,x
n—>oo
all x £ E. By Lemma 2.1, the sequence / „ converges to xo in the topology a (X/£°,M). Therefore, the sequence Y (r,t) fn(x) converges as n —> oo for
Propagators:
General
Theory
115
all x G E and 0 < r < t. Here we use the dominated convergence theorem. Put
Y{T,t)Xo(x)=
lim
Y(r,t)fn(x).
Then it is not hard to see that Y (r, t) xo does not depend on the approximating sequence / „ . It is also clear that
P{T,x-t,0)
=
Y(T,t)Xo{x).
Fix r , t, and x such that 0
*{Lf>,M)-1im nliJ
Y(r,T + h
h)f-f
(2.31)
exists. The operators A+ (r) are called the right a (Lf, M)-generators of the backward propagator Y. For any t with 0 < t < T, denote by D+(t) the set of all functions / S Lf such that the function T —> Y(r,t)f is c (L~, M)-differentiable from the right on (0, t). Theorem 2.9 Let Y be an o (Lf ,M)-continuous backward propagator on the space Lf, and fix t with 0 < t < T. Then for every f G D^(t), the function U(T) =Y (T, t) f is a solution to the following final value problem
on(0,t): ( d+u (T,X) dr
=
l i m u ( r , x) =
-A¥(T)U(T)(X),
(2.32) f(x)
for all x G E. R e m a r k 2.2
d+i
The derivative ——(r, x) and the equality or limu(T)(a;) = f(x)
116
Non-Autonomous Kato Classes and Feynman-Kac Propagators
in final value problem (2.32) are understood in the sense of convergence in the space (Lf, a (Lf, M)). Proof. Let / £ D^(t), and suppose that hn is a sequence of positive numbers such that hn j 0. Let v e M be a nonnegative Borel measure on E. Then, using the fact that Y is a a (L|?,M)-continuous backward propagator on the space Lf and reasoning as in the proof of (2.17), we get r g+u r g+u / ——dv= lim / Y (r, r + h„) —— dv JE dr n^ooJB dr = hm / Y
(T, T
+ hn)
+ lim / F ( T , T . / E.
—
+ /I„)
d^
y(T + /ln,*)/-K(T,t)/
dv. (2.33)
/in
Our next goal is to prove that lim / Y(T,r + hn)
d+u dr
Y(T +
hn,t)f-Y(T,t)f hn
di/ = 0. (2.34)
It follows from the a (Lf, M)-continuity of Y that for every n G N, the set function vn(B)=
i Y{r,T + hn)XB{x)du{x),
Be£,
JE
is a Borel measure. It is not hard to see that the set function \i defined by
L
n=\
is a finite Borel measure on E. Define a sequence of functions on E by 9n
d+u dr
Y(r +
hn,t)f-Y(r,t)f hn
Then,
d+u
[ Y(r,T + hn) dr JE
Y(r +
hn,t)f-Y(T,t)f h„
dv=
\ JE
gn{x)dvn{x). (2.35)
Since / e D+(t) and Lemma 2.1 holds, lim gn(x) = 0
(2.36)
Propagators:
General
Theory
117
for all x € E, and su
( 2 - 37 )
Plffnloo < °°-
n
It follows from (2.36) and Egorov's theorem that for every e > 0 there exists B £ £ such that H(E\B) < e and \gn{x)\ < e, x£B,
n e N.
(2.38)
Moreover, since every measure vn is absolutely continuous with respect to the measure fi, and lim vn(B) = v(B) for all B € £, the Vitali-Hahn-Saks n—>oo
theorem (see Section 5.5) implies that lim sxxpvJB) = 0 .
(2.39)
M(B)iO n
Now it is not hard to see that equality (2.34) follows from (2.35), (2.37), and (2.39). Next, using (2.34) and the properties of backward propagators, we see that
[p*,JE
dT
lim /r(x, T + M r ( T + " - ' ' / - n r ' t ) / ^ h
n->°°JE
= lim
n
/ Y(T,t)f-Y(T,T
+
hn)Y(T,t)f
dv
n->ooJE
~ ~."St/, YiT-rlK)-'Y(r,
W,
(2.40)
It follows from (2.40) that Y(r,t)f € Dw (A*£(r)) and that the equation in (2.32) holds in the space (Lf, a (Lf, M)). In addition, the equality lim / u{r)du = I fdv follows from the continuity of Y on the space T T* JE JE
{Lf,
= [ d(i{x) [ P(T,x;t,dy) JE
JB
= [ JE
P(T,x;t,B)dfi{x),
118
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
and the last expression is a Borel measure on E. Let 0 < t < T, and denote by Dw ( ^ ( i ) ) the set of all /x G M such that the limit m
= lhnW(t
m + v ;rA
'
h,t)KB)-m h
+
hio
exists for all B G £. Then, by the corollary to the Vitali-Hahn-Saks theorem formulated in Subsection 5.5 of the Appendix (see Corollary 5.2), A+(t)/x is a Borel measure on E for all ix G Dw (A+(t)\. We will denote by F£ (T) where 0 < r < T the set consisting of all /x G M such that \imW(T
+
h,T)fi(B)=fi{B)
for all B G £. For every r G [0,T), denote by D + ( T ) the set of measures /tx G M such that the function £ — i > iy(£, T)/X is setwise differentiable from the right on {T,T). The next result concerns Kolmogorov's forward equation. Theorem 2.10 Let Y be the free backward Lf -propagator corresponding to a transition probability function P, and let W be the propagator on the space M given by W(t,r) = Y(r,t)*, 0 < r < t < T. Suppose that 0 < r < T and /x G D+(T) f]F£(r). Then the M-valued function v defined for all t G (T,T) by the formula v{t){B) = W (t,r)/x(B), B G £, is a solution to the following initial value problem: ( d
~(t)(B) = A£+(t)u(t)(B), M Unu/(t)(B) = M(B)
(2.42)
tlr
for all B G £. Proof. We have already established that the family W is a propagator on the space M. If xx G x9^.(r), then using the properties of propagators and the definition of the set x9^(r), we get 9+V
- lim W{t
+ M)/i(B)
- lim W ^
+ hl
T)/i(jB)
(t)(B)
- l i m hio
(
^
+
^W^ /l f)
'
~
W
" ^(*'TM3) ^
"/)ff(f'T);j(g) h
r)M(B)
Propagators:
General
Theory
119
A£tW(t,T)(i(B)=A£tv(t)(B)
=
for all B € £. In addition, the equality limi/(t)(B)
= n(B), B e £, follows
from the definition of the set F£{T). This completes the proof of Theorem 2.10.
D
The equation in Theorem 2.10 is called Kolmogorov's forward equation, or the Fokker-Planck equation. If the Dirac measure Sx belongs to the set D+{r) for every x £ E, and if problem (2.42) is uniquely solvable for all /i € D+(T), then we can recover the transition function P from the formula P(r,x;t,A)
=
W(t,T)St(A).
Let Y be the free backward propagator on the space Lf associated with a transition probability function P. For every r with 0 < r < T, put
2?f*(r)=
f| ^ « W ) . t:r
where the operator A?f(t) is defined by (2.31). Then the following theorem holds: T h e o r e m 2.11 Let Y be the free backward propagator on the space Lf associated with a transition probability function P. Then for every r with 0
£
for all t e
Proof. Um
hlOjE
=Y(T,t)A™(t)f
(2.43)
{T,T).
Let / e f Y(r,t
+
D+'*(T).
h)f-Y{r,t)f h
Then for every /z <E M , _ ^
/ Y{r,t)(Y(t,t
hlOJE
+ h)f - f) h
(2.44) By Lemma 2.1, the cr(Lf, M)-convergence of a sequence hk & Lf to a function h £ Lf is equivalent to the pointwise convergence and the uniform boundedness of the sequence hk in the space Lf'. It is not hard to prove that
„(L?,M)_m"(MXr(M + W - / ) „ r ( r , t ) A , m
(2 . 45)
120
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
Next, passing to the limit under the integral sign on the left-hand side of equality (2.44) and using (2.45), we get
J
Y{T,t + h)f-Y{T,t)f H
=
h
JE W
I
m
JE
for all (i G M. It is clear that (2.46) implies (2.43). This completes the proof of Theorem 2.11.
•
Finally, let Y be the free backward propagator on the space L|° associated with a transition probability function P, and let W(t, r ) = Y(T, t)*. Then W is a propagator on the space M (see the discussion before the formulation of Theorem 2.10). For every r with 0 < T < T, consider the operator
K(TMm . J j f c M M , n|0
B e £,
(2.4T)
fl
and denote by D (As_(r)\ the subspace of the space M consisting of all measures fJ. for which the limit in (2.47) exists for all B € £. By the corollary to the Vitali-Hahn-Saks theorem formulated in Subsection 5.5 of the Appendix (see Corollary 5.2), for // £ D (A^_(T)), yl£(r)/i is a Borel measure on E. Put D£J*(t)=
f)
D(A£_{T)).
(2.48)
T:0
Then the following theorem holds: Theorem 2.12 Let Y be the free backward Lf -propagator associated with a transition probability function P, and let W be the propagator on the space M given by W(t,r) = Y(r,t)* where 0 < r < t < T. Then for any t with 0
-^llli
Proof.
= -W{t,r)Ai(r),.
(2.49)
Let ( i £ M . Then we have W{t,r)n(B)=
I P(T,x;t,B)dn(x)
(2.50)
JE
for all B S £. It follows from (2.50) that if a sequence /u^ £ M converges to \x G M setwise, then the sequence W(t, r)fik converges to W(t, T)H setwise.
Propagators:
General
Theory
121
For/iG D£J*{t), we have W(t,
T - fe)/x - W(t, T)H
_
W(t,
T)(W(T,
T - ft)/i -
/x)
Now using the previous remark and passing to the limit as h [ 0 in (2.51), we get (2.49). This completes the proof of Theorem 2.12. •
2.4
Howland Semigroups
Propagators are two-parameter generalizations of semigroups. Conversely, if a propagator is given, then, under certain restrictions, one can define a semigroup by introducing an extra time variable. For instance, let Q be a backward propagator on a Banach space B, and let T > 0 be a positive number. Denote by Lg D ([0,T], B) the space of all bounded Bvalued strongly measurable functions on [0, T] equipped with the norm
11/11^= sup ||/(t)||B. te[o,T]
Definition 2.3 Let Q be a backward propagator on a Banach space B, and let T > 0. The Howland semigroup SQ(£) on the space Ug(\Q,T],B) is defined by SQW(T)
=
Q(T,
(r +1) A T)f((r
+ t)AT),
t £ [0,T],
(2.52)
where f e Lf
([0,T],B).
Lemma 2.2 Lf([0,T\,B).
The family of operators 5g(t) is a semigroup on the space
Remark 2.3 Note that the semigroup SQ(t) is defined on a finite interval [0, T]. If T = oo, then the Howland semigroup is defined by SQ(t)f{T)=Q(T,T Proof.
+ t)f(T + t),
0
0
We will first prove the following two assertions:
(1) If / € Lg'([0,ri,B),then SQ(t)(f) e Lf([0,T\,B). (2) For all s > 0 and t > 0, SQ({S +1) A T) = SQ(S) O SQ{t), and moreover SQ (0) = /, where / stands for the identity operator on the space Lf([0,T\,B).
122
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
Let / € L g ( [ 0 , T ] , 5 ) . Then there exists a sequence of simple functions sn : [0,T] -> B such that for all u £ [0,T], \\f(u)-sn(u)\\B -> 0 as n —> oo. Fix £ € [0, T], and define a sequence of simple functions by Sn(r) = Q(T, (r +1) A r ) s „ ( ( r + t ) A T ) where T € [0,T]. Then, for all r E [0,T], l|5Q(t)/(r)-S-n(r)||B = \\Q(r, (r + t)A T)f((r <
\\Q(T, (T
+t)AT)-
Q(T, (r + t)A T)sn((r
+1) A T)\\B
+ t) A T ) | | B ^ B ||/((r + t) A T) - sn((r + t) A T ) | | B - 0
as n —» oo. Therefore, the B-valued function Sq{t)f is strongly measurable. Moreover, l|5gW/(r)||B<||Q(r,(r + i ) A T ) | | B ^ B | | / ( ( r + i ) A r ) | | B , and hence, sup \\SQ(t)f(T)\\B
< co.
T€[0,T]
It follows that 5 Q ( < ) / e Lf([0, T\,B). This establishes condition (1) above. We will next prove condition (2). It is clear that £Q(0)/(T)=Q(T,T)/(T)=/(T).
Moreover, using the properties of backward propagators, we see that for all s€ [0,T] a n d i e [0,T], SQ(s) O SQ(t)f(T)
= Q(r, (T + s) A T) (5 Q (t)/) ((T + s) A T)
= Q(r, (r + S) A T ) Q ( ( T + s) A T, ((r + s) A T + t) A T)
f(((r + s)AT
+
= Q(r, (r + s) A T)Q((T
= Q(T, (T +
8
t)AT) + s) AT,(T
+ t) A T)f((r
+ s + t) A T)f{{r
+ s +1) A T)
+ s + t ) A T ) = 5 Q ((s +1) A T)f(r).
This establishes condition (2). The proof of Lemma 2.2 is thus completed.
•
An element x of the space B can be identified with the constant function / x ( r ) = x, T € [0, T\. By taking this identification into account, we get the
Propagators:
General
Theory
123
following formula connecting the backward propagator Q with the Howland semigroup SQ: SQ{t)fx{T)
= Q(T,(T + t)AT)x
(2.53)
for all t G [0, T], r e [0, T], and x e B. Next, we will discuss Howland semigroups associated with propagators. Definition 2.4 Let W be a propagator on a Banach space B, and let T > 0. The Howland semigroup Sw on the L%>([0, T], 5 ) is defined for all r > 0 by Sw(r)f(t)
= W{t, (t-r)V
0)/((t - r) V 0)
(2.54)
where / G Lg>([0,T],£). Arguing as in the proof of Lemma 2.2, we can show that SW(T) is a semigroup on the space L^([0,T],B). The propagator W is related to the semigroup Sw as follows: Sw(T)fx{t) = W(t,(t-T)V0)x
(2.55)
for all t G [0, T], r G [0, T], and a; G B. Howland semigroups associated with free propagators and free backward propagators admit a probabilistic description. Let P be a transition probability function, and let Xt be a corresponding Markov process with state space (E,£). Recall that in Section 1.5 we discussed space-time processes associated with the process Xt (see Section 1.5 for the definition of the transition function P, the family of measures P(T)a;), and the sample space Cl of the space-time process Xt). Let Y be the free backward propagator on the space Lf associated with the transition function P. As we already know, the backward propagator Y can be expressed in probabilistic terms as follows: K(T, *)/(*) = E r , * / ( * t )
(2-56)
for all 0 < r < t < T, x G E, and / G Lf. In addition, for the Howland semigroup Sy(t) associated with the backward propagator Y, the following probabilistic characterization is valid: SY(t)F(T,x)
= ETiXF ((r + t ) A T , X ( T + t ) A T ) = E ( T , X ) F ( x t )
(2.57)
for all r G [0,T], t G [0,T], x G E, and F G L% ([0,T], L£°). This can be derived from (2.52) and (2.56).
124
Non-Autonomous Kato Classes and Feynman-Kac Propagators
Now let P be a backward transition probability function, and let XT be a corresponding backward Markov process. In Section 1.5, we constructed a time-homogeneous transition probability function P and the family of measures pC- 1 ). The function P was used as the transition function of the space-time process XT. Let U be the free propagator on the space Lf corresponding to the backward transition probability function P. Then the following formula is valid: U{t,T)f{x)=&xf(xT)
(2.58)
for all 0 < r < t < T, x 6 E, and / G Lf. For the Howland semigroup Sir (T) associated with the propagator U, we have Sv(T)F(t,
x) = E*'XF ((t - r) V 0, X ( t _ T ) v 0 ) = E{t'x)F
(xT)
(2.59)
for all T G [0,T], t G [0,T], i G E, and F G L ^ ([0,T],Lf). This follows from (2.54) and (2.58).
2.5
Feller-Dynkin Propagators and the Continuity Properties of Markov Processes
Let us recall that in Section 2.2 we defined the following spaces of continuous functions on the space E: the space BC of all bounded continuous functions on E; the space BUC of all bounded uniformly continuous functions on E, and the space Co of all functions from BC which vanish at infinity. Definition 2.5 A backward BC-propagator is called a backward Feller propagator. A backward Co-propagator is called a backward Feller-Dynkin propagator. If a backward L|°-propagator Q is such that Q(T,t)eL(L?,BC) for all 0 < r < t < T, then it is said that Q possesses the strong Feller property. If a backward Lf -propagator Q is such that Q(T,t)eL{Lf,BUC) for all 0 < r < t < T, then it is said that Q satisfies the strong BUCcondition.
Propagators:
General
Theory
125
Remark 2.4 If Q is a backward L°°-propagator, then one can replace the space Lf by the space L°° in the definition of the strong Feller property and the strong B[/C-condition. It is not hard to see how to define forward Feller and Feller-Dynkin propagators, and also propagators possessing the strong Feller property, or satisfying the strong i?£/C-condition. It this section, we begin the study of Markov processes associated with free backward Feller-Dynkin propagators. Let P be a transition probability function, and let Y be the corresponding free backward propagator. Our first goal is to prove that if Y is a strongly continuous backward FellerDynkin propagator, then the class of all Markov processes associated with P contains a process Xt such that its sample functions are right-continuous and have left limits. In the following sections, we will show that the process Xt possesses additional properties. Given a transition probability function P, we start with the standard realization Xt of a corresponding Markov process on the probability space (n,F[,FTtX} where ft = £ l 0 ' r l . It is denned by Xt(u) = u(t) where u> £ Ct. By Q, will be denoted the space of all E-valued functions defined on the interval [OjT], which are right-continuous and have left limits in E. Then we have Q C ft. Put Xt{w) = u(t), w £ ft, t £ [0,T], and let J J , 0 < T < t < T, be the er-algebra generated by Xs with T < s < t. For every r € [0, T) and x £ E, denote by P T|X the probability measure on TT determined by FT,X [Xtl £Bu...,Xtn£Bn)=
PT,X [ x t l 6 B i
Xtn £ £ „ ] .
(2.60)
Here r < h < • • • < tn < T and Bj £ £, 1 < j < n. Let us denote by !F^+, t £ [T,T), the cr-algebra defined by Tl+=
Q
T[.
(2.61)
t<s
Theorem 2.13 Let P be a transition probability function such that the corresponding free backward propagator Y is a strongly continuous backward Feller-Dynkin propagator. Then the processes {Xt^T'ljr,Pr,x) and ( Xt, Ft •> ^T,X ) are stochastically equivalent. Remark 2.5 It follows from Theorem 2.13 that the process (Xt, .FtT+, PT,X) is a Markov process with P as its transition function. Moreover, all the sample paths of Xt are right-continuous and have left limits.
126
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
It also follows from the proof of Theorem 2.13 that for every w G O , the set {Xt(u>) :t£ [0, T]} is relatively compact in E. Proof.
For every function f £ Co and every e with 0 < e < T, put Ne(f)(t,z)
= f
Y(t,s)f(z)ds
Jt+e
where 0 < t < T - e and z € E. Lemma 2.3 Fix r with 0 < r < T and x G E. Then for every nonnegative function f £ Co and every e with 0 < e < T — T, the process t — i > Nc(f) (t,Xt), t e [T,T — e], is a supermartingale with respect to the filtration Tl and the measure PT>X. Proof.
It is clear that the process t i-> Ne(f) (t,Xt\
filtration T[. implies that E.r,x
is adapted to the
Moreover, for r < t\ < ti < T — e, the Markov property
Ne(f) (t2,Xt2) | Ft] =E t i ] j f t i [Ne(f) (t2,Xt2)] .
(2.62)
Since Y is a backward propagator, E ti,Z
Ne(f)(t2,Xt2)}=
= /
[
Y{t2,s)f(xt2)
E-
Y (h, t2) Y (t2, s) f(z)ds=
f
Jt2+e
<[
Y (h,s)
ds f(z)ds
Jt2+e
Y(t1,s)f(z)ds
= Ne(f)(t1,z)
(2.63)
for all z € E. Finally, (2.63) with z = Xtl and (2.62) show that Lemma 2.3 holds. • Let us continue the proof of Theorem 2.13. It follows from Lemma 2.3 and Theorem 1.10 that for every f £ Co and T £ [0,T), there exists a set A r G T such that PT)X (AT) = 0, and for every integer n > 1, the process Z?(f)(uj) = nJ
n
Y(t,s)f(xt(u))ds,
T
has left and right limits, tonZr(/)(w) sjt.sGQ
and
Urn Z ? ( / ) ( w ) , sit,seQ
(2.64)
Propagators:
General
Theory
127
for all u G fi\Ar and £ G [T,T - i ] . The set A r does not depend on n and i. Moreover, AT can be chosen independently of / € Co- Indeed, let /& be an everywhere dense countable subset of Co- Then it is not hard to see that there exists an exceptional set A r G f with PT)X(A) = 0 such that the limits in (2.64) with / replaced by fk exist for all w G fi\AT, t e [T,T- ^ ] , and k > 1. It is also easy to prove that if a sequence gk G Co converges uniformly to a function g, then we have Y(t,s)gk(z)ds->n Y (t, s) g (z) ds •/ uniformly in z as k —> oo. It follows that the exceptional set A r can be chosen independently of / G Co, n > 1, and t with T
J
lim
sup
n->OO
zG£,0
Y(t,s)f(z)ds-f(z)
0
for every / G Co- Hence, there exists a set AT G T for which P T X(AT) = 0, and for every f £ Co the limits lim f(xs(u))
and
lim f (xJu))
(2.65)
exist for all ui G fi\A T and all t with T < t < T. The exceptional set AT does not depend on / G Co and t G [r, 71]. The following well-known fact concerning nonnegative supermartingales will be used in the proof of Theorem 2.13. L e m m a 2.4 Let (fi, F, Tt, P) be a filtered probability space, and let Zt be a nonnegative Tt-supermartingale. Denote by Q a countable dense subset of the interval [0, T]. Then Zt > 0,
inf Zs = 0 = 0 s
for all t G [0, T]. Proof. Fix t with 0 < t < T, and let Q5 = {sj < • • • < s ^ } increasing sequence of finite subsets of the set Q D [0, t] such that
Qn[0,t] = ( j Q i .
be an
128
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
For all j > 1 and n > 1, put Tj>n = m m
|4:
i
In addition, if LJ G
min
fi\
Z, < —
then we put Tj,„(w) = T. Next, using the properties of supermartingales and the Tsi -measurability of the event < r^n = s°k \, we get E Zu
min Zi < - = E Zt, TjyTl < T, ZTjn < — l
"k
n
= fc=i E E •^t, ''"j.n — S , Z k
= EE E
Zt I ^
Tjn
<
j T j . n — Sk, Z j < —
rrij
<Efc=iE
Z.j i Tin — Sk, Z j < — "k J k n
< - . (2.66) n
Passing to the limit in (2.66) as j —» oo and then as n —> oo, we get E Zu
inf Zs = 0
0.
(2.67)
s
Now it is clear that Lemma 2.4 follows from (2.67).
•
Let us return to the proof of Theorem 2.13. Fix a countable dense subset Q of the interval [0, T], and for every r G [0,T), consider the event TT consisting of all w G CI for which there exists r G Q n [r, T) such that the set Xs(o»), s G Q n [r, r], is not contained in any compact subset of E. Then the following lemma holds: Lemma 2.5 Let f be a strictly positive function from CQ. Then for every T with 0 < r < T, TT=
\J
i w :I
reQn[r,T) I
inf
/
s€Qn[r «n[r,r],/s
Y(r,u)f(xr{u))du>0,
^r
Y(s,u)f(xs(w))du = o\. v
'
(2.68)
Propagators: General Theory
Proof.
129
Our first goal is to prove that for every r G [r, T) and z £ E, /
Indeed, suppose that / /
Y(r,u)f(z)du>Q.
(2.69)
Y (r, u) f(z)du = 0. Then the equality
Y (r, u) f(z)du =
Jr
du Jr
f(y)P (r, z; u, dy), JE
and the strict positivity of the function / imply that for almost all u € [r, T] with respect to the Lebesgue measure, we have P (r, z; u, E) = 0. This contradicts the definition of transition probability functions. Therefore, inequality (2.69) holds. Next, fix a sequence sn € [T,T] such that lim sn = s where r < s < T. We will show below that for a sequence zn G E, the following two conditions are equivalent: (1) There exists a compact subset C C E such that zn £ C for all n > 1. (2) The inequality inf /
Y(sn, u)f(zn)du>0
(2.70)
holds. Indeed, if condition (1) does not hold, then there exists a subsequence znk of zn such that lim g(znk) = 0 for all g G Co- It follows from the strong fe—>oo
continuity of Y on Co and from the condition lim sn = s that f
Y(s,u)f(znk)du-
Js
f
Y(snk,u)f(znk)du^0
(2.71)
JSnk
as k —» oo. Next, the strong continuity of Y on Co gives lim f
Y(s,u)f(znk)du
= 0.
(2.72)
By (2.71) and (2.72), we see that inequality (2.70) does not hold. Hence, the implication (2) =*• (1) is valid. On the other hand, if condition (1) above holds, and condition (2) does not hold, then there exists a sequence nk such that lim znk = z where k—*oo
130
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
z € C; moreover, lim /
Y(snk,u)f(znk)du
= 0.
(2.73)
k
^°°Jsnk
Since Y is a strongly continuous backward propagator on Co, and (2.69) holds, we have lim /
Y(snk,u)f(znk)du=
Y (s,u) f{z)du > 0.
k—>oo /„
/.
This contradicts (2.73). Therefore, the implication (1) =>• (2) holds. The proof of Lemma 2.5 is thus completed.
•
It follows from Lemma 2.3 with e = 0, Lemma 2.4, and Lemma 2.5 that for all r G [0, T) and all cc G £ , the equality PT,X [TT] = 0 holds. Therefore, PT,X fi\rr = 1. Moreover, for all w G f2\r T and r G Q ("1 [r,T), there exists a compact set C C E such that Xj(w) G C for all s G Q n [T, r]. The set C depends on r and u>. Given r G [0,T), put
S r = n \ (AT u r T ), where A r is the complement of the event consisting of all u> G tl for which the limits in (2.65) exist for all f € Co and all t with T
lim XJw)
(2.74)
sj.t,«6Q
exist for all t with T
of the first limit in (2.74) is similar. Let us consider the subset Q of the set f2, consisting of all w G E^°'T^ such that the following conditions hold: the limits lim ui(s) and lim LJ(S) *lt,s€Q
sRsGQ
exist for all t G [0, T), and for every r G Q n [0, T) there exists a compact subset CT{u)) of E for which w(s) G C r (w) where s G Q n [0, r]. It is not
Propagators:
General
Theory
131
difficult to show that for all r and x, (2.75)
n
where P * x denotes the outer measure generated by the measure P r , x - Indeed, for every r G [0,T) and x G E, the set fi contains the set £lT,x, consisting of those w G fir for which w(s) = x for 0 < s < T. Since fiT 1, we have P* fiT = 1. Now it is clear that (2.75) holds. Therefore, one can restrict the probability space structure from Q, to O. The resulting probability space is denoted by (Q, J7, PT)X J. Let us define a stochastic process on Q, by lim Xs, Xt = { »lt,st
H0
XT,
Then, it is not difficult to see that the process Xt is right-continuous and has left limits on the interval [0,T). It is also clear that the process Xt, t G [T,T), is .F t r + -adapted. Here
*T+
n %• t<.s
Lemma 2.6 The process lXt,^+,PT>x) sition function P.
is a Markov process with tran-
Proof. Let r < s < t < T, and choose sn G n (s, t) and tn G Q n (t, T) with sn I s and tn [ t as n —> oo. Let .4 G ^ J + . Then A G ^ J n for all n > 1, and since X u is a Markov process,
Er,x [/ (*t„) XA] = Er,« [ESn,^sn [/ ( X )
XA
^T,x[Y(sn,tn)f(^XSn)xA]
(2.77)
for every f £ Co- It follows from the strong continuity of Y on Co and from (2.77) that E. ;x
[f (Xt) XA] = Kx [Y (S, t) f ( £ ) X^
132
Non-Autonomous Kato Classes and Feynman-Kac Propagators
for all A G J~l+- Here we used the fact that ET]X is the restriction of ET>I to ft. Therefore, E.T,X
'f(xt)\^+]=Y(s,t)f(xs)
(2.78)
for all / G Co- In the case where t = T, (2.78) is also true. Indeed, we may assume tn =T for all n > t, and use the equality XT = XTFor all x G E and 0 < r < t < T, we have E T, , [/ (Xt)}
= ET)X [ s U m Q / ( * . ) ] =
lim
ETtX\f(xs)]
s
H m Q I r , x [/ ( x s )
lim
r(r,S)/(x)
sit,s€
Y(T,t)f(x)=ET,x[f(xt)'
(2.79)
Here we used the strong continuity of Y. In the case where t = T, (2.79) also holds since XT = XT- Next, using Urysohn's Lemma, the monotone class theorem, and approximating bounded Borel functions by simple functions, we see that (2.78) and (2.79) hold for all bounded Borel functions / on E. Consequently, the process Xt is an P riX -Markov process with respect to the filtration !F[+, r
and •
Lemma 2.7
The process ( Xt, Ft+i ^T,X ) is a modification of the process
(Xt,T[,FT,x);
that is, for every 0 < r < t < T and x G E, the equality
P r x Xt = Xt
= 1 holds.
Proof. Let / and g be functions from the space Co, and fix x G E, t G [0, £], and T with 0 < T < t. Let sn be a sequence in Q n (t,T) such that sn 11. Then the Markov property of the process Xu, 0
Er,x [/ (*.„) 9 (Xt) = Er>x Er,x [/ (Xsn) | fl] 9 (Xt) = ET|X K~xt [/ {*.„)] 9 (Xt)} = E r x \Y(ttsn)f(Xt))g(Xt)].
(2.80)
Propagators:
General
Theory
133
Passing to the limit in (2.80) as n —> oo and taking into account that Y is strongly continuous on Co, we obtain Kx
[/ ( * t ) 9 ( * t ) ] = ^m%,x
= lim^^x
[f (*«„) g
(Xtj
[(Y (t, sn) f ( X t ) ) g {xt)\
= ET,X [/ (Xt) g ( X t ) ] = ET,X [/ (Xt) g ( * t ) ] •
(2.81)
Next, formula (2.81) with g = f and the stochastic equivalence of the processes Xt and Xt (see Lemma 2.6) give
f{xt)-f(xt) f (X t ) 2 - 2ET;X [/ (X t ) / ( x t ) = Er
x
f [Xtf
- 2ET,X [/ (xt) f (Xt)
/ ( X t ) 2 ] = 0 (2.82)
for all / G Co, 0 < r < f < T, and a; 6 £ . Let //c G Co with HAU^ < 1, k > 1, be a sequence such that the closure of its linear span coincides with Co- For all x, y G E, put oo
p(a;,y) = ^ 2 - f c | A ( a : ) - / f e ( j / ) | . Then p is a metric on E x E generating the same topology as the metric p. It follows from (2.82) that ET,x[p(xt,Xt)] = 0 for all 0 < r < t < T and x G E. This completes the proof of Lemma 2.7.
(2.83)
•
Now we are ready to finish the proof of Theorem 2.13. By equality (2.60) and Lemma 2.7, PT,X [Xtl eBlt...,XtnGBn]=
Pr,x [Xtl G £ i , . . . , X t „ G £ „ ]
(2.84)
for all T < ti < • • • < tn < T and Bj G £ with 1 < j < n. It is not hard to see that Theorem 2.13 follows from (2.84). •
134
2.6
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
Stopping Times and the Strong Markov Property
In this section we discuss the Markov property with respect to random times. In the next definition, an important family of random times is introduced. Definition 2.6 Let {Q.,Q,Qt,W), r < t < T, be a filtered probability space. A function S : Q —» [T,T] is called a ^-stopping time if for every t e [r, T], the event {S < i) belongs to the cr-algebra Qt- If T = oo, then it is assumed that S takes values in the extended real half-line R+. We will often write "a stopping time" instead of "a C/f-stopping time" if the filtration is fixed. It is not hard to see that 5 is a stopping time if and only if the process t H—> X{S
u> € ft.
Definition 2.7 Let (Q,£,£ ( ,P), T < t < T, be a filtered probability space, and let 5 be a St-stopping time. The cr-algebra Qs is defined as follows: An event A £ Q belongs to Qs if and only if A n {5 < t) £ Qt for every t G [r, T]. It is not hard to see that Qs is a cr-algebra. The cr-algebra Qs is usually interpreted as the information contained in the process Xt before or at time 5. If S\ and 52 are stopping times such that S\ < S2, then QSl^Qs2-
(2.85)
Indeed, if A G QSl, then A n {5 2 < t} = A n {Si < t} n {5 2 < t}, and since the events An {Si < t} and {S2 < t} belong to the cr-algebra Qu the event A n {S 2
Propagators: General Theory
Gg* and G$l
are
135
defined as follows:
gll=a(xs.:S'ESil) and GSsl=o-(s',Xs,:S'£Sssl). Put Gs = Gs and Gs = G%- It is not hard to see that for all stopping times Si and S2 with Si < S2 the following inclusion holds: £f* C Sf 1 . On the other hand, it is not clear whether Gs C Gs- However, if the process Xt is progressively measurable, then Gs C Gs C Gs-
(2.86)
The inclusions in (2.86) follow from the fact that S is C/s-measurable and from the following lemma: Lemma 2.8 Let (Xt,Gl), 0 < r
(2.87)
Since the composition of measurable functions is measurable, the random variable XsA(tAu) 1S ^tA«/^- m e a s u r a t>le. It follows that the event {^SA(tAu) € -B} belongs to the cr-algebra GJAW a n d therefore it also belongs to the cr-algebra GZ,- In addition, the event {S At
136
Non-Autonomous Kato Classes and Feynman-Kac Propagators
Our next goal is to define and study cr-algebras related to the future behavior of a stochastic process Xt after a stopping time S, and also aalgebras containing the present information about Xt. It seems promising to choose the cr-algebras Qj. and GT to represent the future of the process after the stopping time S. However, this is not a good choice, if we want to use these cr-algebras in the formulation of the strong Markov property (see Definition 2.14 below), since the events belonging to the cr-algebras GT and GT m a v depend on the past history of the process before the stopping time S. It is easy to see that GT C GT- O*1 * n e other hand, it is not clear whether these cr-algebras coincide, or if the cr-algebras in (2.86) coincide. Note that the stopping time S is not necessarily measurable with respect to the cr-algebra a (Xs). As a result, we can choose the cr-algebra a (Xs) or the cr-algebra a (S, Xs) as a storage of information contained in the process Xt at the stopping time S. Next, we turn our attention to cr-algebras representing the future of a stochastic process Xt after the stopping time S. One possibility to define such a cr-algebra is to imitate the definition of the cr-algebra Gs representing the past before S. Let us first note that if a stochastic process Xt is progressively measurable, then Gs=
f^
{AeGT--
An{S
(2-88)
t:0
Indeed, equality (2.88) follows from the equality
{AeGT-
An{S
= {AeGT-
An{S
XSM) VGt}, 0 < t < T,
and from the fact that for progressively measurable processes, tr{SAt,XSM)cg°SMC$ (see Lemma 2.8). For a stopping time S, define the family of events GS by the following: AeGS <^>An{S
(2.89)
for all t with 0 < t < T. It is not hard to see that the family Gs in (2.89) is a cr-algebra. Note that an equality similar to the equality in (2.88) does not hold for this cr-algebra. Indeed, the event {S < t} does not necessarily belong to the a-algebra GT- The next equality shows that the cr-algebra GS
Propagators:
General
Theory
137
is not large enough to store the information about the process Xt between S and T: gs =
(2.90)
The equality in (2.90) can be obtained as follows. By (2.89) with t = T, we see that if A G Qs, then Aea(S,Xs,XT). Moreover, if A € a (S, Xs, XT), then A£a{S,Xs)VQT for all t e [0, T]. On the other hand, we have {S
An{S
(2.91)
Condition (2.91) is equivalent to the following condition: for all ti and t2 with 0 < tx < t2 < T, {S' >t2>t1>S}ea(S,Xs)V
9%.
(2.92)
Indeed, this equivalence can be easily established by taking into account the equality
{S' >t2>h>s}
= {h > s} n {s'
>t2>s}.
Example 2.1 Let S be a stopping time, and let a random variable S' be defined by S' = <j> (S, Xs) where <j> is a Borel function on [0,T]xE. Assume that S' is a stopping time such that S < S'. Then we have 5" £ N(S). This follows from the fact that the event {S < t < S'} belongs to the cr-algebra a (S, Xs)- As an example, we can take S' = (a + S) A T where a > 0. Another example is given by S' = S V p where 0 < p < T.
138
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
Definition 2.10 A stopping time S is called a terminal stopping time if for every pair of fixed times (£1,^2) with r < ti < t?, < T, the event {ti < S < £2} belongs to the cr-algebra GilRemark 2.6 Not all stopping times are terminal. For instance, the stopping times (a + S) A T, where a > 0 and S is a given stopping time, are not necessarily terminal stopping times. E x a m p l e 2.2 Let S and 5" be stopping times with T < S < S' < T. Assume that S' is a terminal stopping time in the sense of Definition 2.10. Then S' G M(S). Indeed, {S
= {S
{S' >t}ea
(S, Xs) V QlT.
Lemma 2.9 Let Si, S2, and S3 be stopping times such that Si < S2 and Si < S3. If S2 G M (Si) and S3 G TV (Si), then S2 V S3 G M (Si) and S2AS3eM(Si). Proof.
Lemma 2.9 for the stopping time S2 V S3 follows from
{S 2 V S3 > t > Si} = {S 2 > t > Si} U {S3 > t > Si} G a (SuXSl)
V $,.
The proof of Lemma 2.9 for the stopping time S2 A S3 is equally simple. • The next definition will be important in our presentation of the strong Markov property (see Theorems 2.17 and 2.20 below). The cr-algebra Q^y in Definition 2.11 is a good candidate to represent the future information about the process Xt after the stopping time S. Definition 2.11 Let S be a stopping time. Then the a-algebra Q^'v is defined as follows: Q^y = a (S V p, XSvP • 0 < p
The c-algebra QT' is a special case of the
and QT (S)),
(
: S'
>S}.
' are defined as follows:
G?{S)
= a (S\ Xs> : S' G M
(S)). (2.93)
Propagators:
General
Theory
139
It is clear that if S is a stopping time and M = {SV p : 0 < p
Let S be a stopping time. Then the cr-algebra Q^^
@pS)=a(S,,Xs>
is
: S' &N (5).).
For the
Let S be a stopping time. Then
gs c QST* c g*(s) c g$.
(2.94)
Proof. The first inclusion in (2.94) follows from equality (2.90). Since SV pe Af(S) for all p e [0, T], we get the second inclusion in (2.94). The third inclusion is obvious. This completes the proof of Lemma 2.10. • It is interesting to notice that for a right-continuous stochastic process Xt on (fi, F) and the two-parameter filtration F[ generated by the process Xt, the second and third cr-algebras in (2.94) coincide. Lemma 2.11 Let Xt be a right-continuous stochastic process on (Cl,F) with state space E, and let S be a F\-stopping time. Then Ql»=Q$W.
(2.95)
Proof. The inclusion Q^,'v C QT follows from Lemma 2.10. In order to prove the opposite inclusion, let us consider the family Tit, t € [0,T], of vector spaces of random variables on O defined as follows. For every t e [OjT1], the vector space Ht consists of all bounded random variables F on Q such that F is measurable with respect to the cr-algebra a (S, Xs) V a (Xp : t < p
140
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
Ht contains all random variables of the form n
fo{S,Xs)Y[fj{pj,XPj), 3= 1
where fj : [ 0 , T ] x £ - * R , 0 < j < n , are bounded Borel functions, and t < pi < • • • < pn < T. By the monotone class theorem, Ht contains all bounded random variables which are measurable with respect to the cr-algebra a (S, Xs) V a (XSVp • t
t
It is not hard to see that the event {S' > t > S} belongs to the cr-algebra a (S, Xs) V<7(SVp, XsvP
:t
Hence, the same is true for the event {S' > t} = {S' > t > S} U {S > t}. It follows that the stopping time S' is measurable with respect to the cralgebra cr (S V p, XSvp •• 0 < p < T). By Theorem 2.20 (this theorem will be obtained below), we see that the state variable Xs> is also measurable with respect to the cr-algebra a(SVp, XsvP -0 < p
gfs) c gp holds. This completes the proof of Lemma 2.11.
•
The next definition concerns the strong Markov property in the case of a family of stopping times. Definition 2.14 Let M be a family of stopping times that is closed with respect to the operations A and V. A stochastic process (Xt, 0, Qt, P) with state space E is called a strong Markov process with respect to the family M provided that the following conditions hold: (1) The process Xt is progressively measurable. (2) For every pair of stopping times Si, S2 € M with 52 > S\ and every bounded Borel function / on [0, T] x E, the equality E [f(S2,XS2)
I gSl]=E[f(S2,Xs2)
\a(SuXSl)]
(2.96)
Propagators:
General
Theory
141
holds P-almost surely. The progressive measurability of the process Xt is assumed in Definition 2.14, because this implies the measurability of X$ with respect to the aalgebra Gs (see Lemma 2.8). There are more versions of the strong Markov property with respect to the family M, namely E [f(XS2)\G3l]=E[f
(XS2) | o (XSl)]
P-a.s.
(2.97)
for all bounded Borel functions / on E, or E f(S2,XS2)
| QSl] =E[f{S2,XS2)
\o-(SuXSl)]
P-a.s.
(2.98)
for all bounded Borel functions / on [0, T] x E, or E [f(XSa)
\gSl]=E
[f(XS2)
\a(S1,XSl)}
P-a.s.
(2.99)
for all bounded Borel functions / on E. Recall that it is assumed in equalities (2.97), (2.98), and (2.99) that S 2 > Si. The next theorem provides several equivalent conditions for the validity of the strong Markov property with respect to the family M (see Definition 2.14). This theorem is related to Lemmas 1.2 and 1.20. Recall that for S G .M, we put M (S) = {S' G M : S' > S}. Theorem 2.14 Let M be a family of stopping times that is closed with respect to A and V, and let Xs be a progressively measurable stochastic process on (fi, T, P) with state space (E, £). Then the following are equivalent: (1) For all pairs Si e M. and S2 € M. with S2 > Si and all bounded Borel functions f on [0, T] x E, the following equality holds: E [f (S2, XS2) I QSl] = E [/ (5 2 , XS2) I a (Si,XSl)} (2) For any S € M and any bounded gT F, the equality
(
F-a.s.
(2.100)
-measurable random variable
E[F\gs]=E[F\a(S,Xs)] holds P-a.s. (3) For any S G M, any bounded gs-measurable random variable G, and any bounded real-valued 0T -measurable random variable F, the equality E[GF]=E[GE[F\a(S,Xs)}]
142
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
holds. -M(S) (4) For any S £ M, A £ Gs, and B £ QT y ', the equality P [A n B I a(S, XS)] =F[A\
a(S, Xsj] P [B \ a(S, Xs)]
holds P-a.s. Proof. (1) = • (2). Suppose that condition (1) in Theorem 2.14 holds. The most important step in the proof of the implication (1) = > (2) is to show that the equality E[F\gs]=E[F\
(2.101)
a(Xs)]
holds P-almost surely provided that m
F=
Ylfj{Si,XSi)
(2.102)
3= 1
where for any 1 < j < m, fj is a bounded Borel function on [0, T] x E, and Sj is a stopping time from M. {S), satisfying S < S\ < ••• < Sm < T. We will obtain equality (2.101) using the method of mathematical induction. Since we assumed that condition (1) holds, equality (2.101) is satisfied for m = 1. Next, assume that (2.101) is true for a given integer m > 1 and for all finite sets {fy :l<j <m\ from M (5) with 5 < Si < S2 < • • • < Sm < T. Let Si < • • • < Sm+i from M. (S), and put
be an increasing family of stopping times
H = cr [QsiS\,X§i,...,Sm,^sm)
•
Then, using the tower property of conditional expectations, we get m+l
E m
JJfj (Sj,X§.) I fm+l [Sm+l,Xgm+iJ E
[[fj ( s , j . ^ 5 j E fm+i [sm+i,Xgm+ij
I Gs
| n
Gs
(2.103)
Propagators: General Theory
143
It follows from condition (1) in Theorem 2.14 and from the inclusions Xo C H C Qa that E :E
(2.104)
Therefore, (2.103) gives m+l
E
ii/»-(^^)i^ 3=1
= E 3= 1
(2.105) L e m m a 2.12 that
There exists a bounded Borel function g on [0, T]x E such
E fm+i (sm+i,Xsm+i)
| a (5m,X§mjJ = g
[Sm,Xgmj
Proof. Lemma 2.12 follows from the Radon-Nikodym theorem. Indeed, consider the following Borel measures: (J-i {B) = E fm+i
ISm+i, Xgm+1 J , {Sm, Xgm J e B
and ^(B)=¥[(sm,X-Sm)eB where B belongs to the Borel cr-algebra of [0,T] x E. It is easy to see that the measure [i\ is absolutely continuous with respect to the measure /i2- Denote the Radon-Nikodym derivative of the measure ^i with respect to the measure ^2 by g. Then the equality in the formulation of Lemma 2.12 holds. Note that H2 is the image measure of the measure P under the mapping ^Sm,XgmJ. This completes the proof of Lemma 2.12. •
144
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
Let us return to the proof of Theorem 2.14. It follows from Lemma 2.12 and equality (2.105) that m+l
E I m
E I I U (Shx~s)) 9 (Sm,XSm)
| Gs
(2.106)
j=i
Applying the induction hypothesis to the right-hand side of (2.106) and using (2.104) and the tower property of conditional expectations, we get m + lL n-t-
E
n
I I fi (§> X~Si) 9 (Sm, X~Sm) I a (S, XS) 3=1
E I J fj [Sj,Xg. j E / m + i [Sm+i,Xgm+i =E
j | a {Sm,X§mJ
\a(S,Xs)
m
T{ f3 (sjtXSj)E[fm+1
( 5 m + l l X § m + i ) | n] |
a(S,Xs)
m+l
:E E
UfifaXsJlH
v(S,Xs)
3=1
m+l
Ylfi(sjtXh)\a{S,Xs)
(2.107)
3=1
This establishes condition (2) in Theorem 2.14 in the case where the random variable F is given by (2.102). Our next goal is to prove the equality in (2) in the case where F = X {(~SuXSi)eBl}x---xX{Csm,XsJeBmy
(2.108)
In (2.108), S i , . . . , Bm are Borel subsets of [0, T] x E and Si,..., Sm are stopping times from the class M (S). The sequence of stopping times Si, 1 < i < m, is not necessarily increasing. For 1 < k < m, define an increasing sequence of stopping times S'k G M (S) by S'k = m i n { 4 V • • • V Sjk : 1 < ji < • • • < jk < m} .
Propagators: General Theory
145
Let n be a permutation of the set { 1 , . . . , m}, and define the event A„ by
^ = {(%^a.) eBl
(^^O6*"}-
Then it is not hard to see that {(SUX~S1)
e B1,...,{srn,X~Sm)
6 Bm)
= (jAr-
Next, using the inclusion-exclusion principle, we get F =
X
{{~SuX§i)eBl}X---XX{(~sm,x~sJeBm}
= 2 ^ XA„ if
2_^
XA„nA„,
7r, 7r':7T^7r'
+
Y.
XA.rv^ru,,,, - • •' •
(2.109)
Since S^ is an increasing sequence of stopping times, (2.107) and (2.109) give E[F\gs]=E[F\a(S,Xs)], where F is defined by F = X
x
{{~SuXSi)eBl}
---xX{(sm,x§jEBm}-
It follows from /»00
G
=
/ Jo
/-DO
X{G+>A}dA- / Vo
X{G->x}dX
(2.110)
that condition (2) holds for all functions F such that m
F = l[fj(sj,X~Sj). j=i
Finally, using the monotone class theorem, we obtain condition (2) as it is formulated in Theorem 2.14. (2) = » (3). This implication follows from the properties of conditional expectations. (3) = > (4). Suppose that condition (3) in Theorem 2.14 holds, and let A £ Qs, B €
146
Non-Autonomous
§™{S\ and D e a(S,Xs). F = XBXD, w e obtain E
[XDXAXB]
Kato Classes and Feynman-Kac
Propagators
Then, applying condition (3) with G = XA and
=E
[XAE [XDXB
\ <* (S,Xs)]]
= E[XDXAE[XB\
(2.111)
Therefore, E [XAXB
I a (5, Xs)]
= E [ X A E [XB | * (S, Xs)]
= F[A\a(S,
Xs)] F[B\a
\ o (S,
Xs)]
(S, Xs)] .
(2.112)
This implies condition (4) in Theorem 2.14. (4) => (1). Suppose that condition (4) in Theorem 2.14 holds. Then it is not hard to show that (2.111) is valid. Let Si G M and S2 £ M. be stopping times such as in condition (1). Taking D = Q in (2.111) and using (2.110) with G = f (S2, Xs2), we obtain E
[XA!
(5 2 , XS2)] = E
[XAE
[/ (5 a , X 5 2 ) | a (Si, X S l )] ] .
It follows from the definition of conditional expectations that E [f(S2,XS2)
I GSl] =E[f(Si,XSt)
\a(SuXSl)]
.
This implies condition (1) in Theorem 2.14. The proof of Theorem 2.14 is thus completed.
•
The next theorem concerns the strong Markov property with respect to the cr-algebra Qs- Theorem 2.15 below contains an extra condition (condition (2)) in comparison with Theorem 2.14. The reason why we did not include this condition in the formulation of Theorem 2.14 is that we do not know whether Qs = QsTheorem 2.15 Let M be a family of stopping times that is closed with respect to A and V, and let Xs be a progressively measurable stochastic process on (fi, T, IP) with state space (E, S). Then the following are equivalent: (1) For all stopping times Si £ M and S2 £ M with Si < S2, E \f(XS2)
\gSl]=E
[f(XS2)
I a (XSl)]
F-a.s.
Propagators:
General
Theory
147
(2) For all stopping times S G M and S £ M with S < S, any finite set of stopping times Si, 1 < i < n, such that Si < S for all i, and any bounded Borel function f on E, the equality E [f (XS) | a (XSl ,...,XSn,
XS)} = E [/ (Xs) |
holds P-a.s. (3) For any stopping time S € M. and any bounded QT random variable F, the equality E F\Gs
=
(
-measurable
E[F\a(Xs)]
holds P-a.s. (4) For any stopping time S G M., any bounded Qs -measurable random variable G, and any bounded QT ^ -measurable random variable F, the equality E[GF]=E[GE[F\a(Xs)}] holds. (5) For any stopping time S £ M., the equality P [A n B I a(Xs)]
=F[A\
holds P-a.s. for all A e Qs and B £
<J(XS)} P [B \ a(Xs)] g^(s).
Proof. We will only prove the implications (1) =>• (2) and (2) = » (3). The remaining implications can be obtained as in Theorem 2.14. (1) = • (2). Suppose that condition (1) in Theorem 2.15 holds, and let S, S, Si,..., S„ be such as in condition (2). It is not hard to see that condition (2) follows from condition (1) applied to the pair S and S and from the following inclusions:
a(Xs)Ca(XSl,...,Xsn,Xs)cgs. (2) = » (3). Suppose that condition (2) in Theorem 2.15 holds. Then, reasoning as in the proof of the implication (1) =>• (2) in Theorem 2.14, we see that the equality E [F \o-(XSl,.
..,XSn,Xs)]
=E[F\
(2.113)
148
Non-Autonomous Koto Classes and Feynman-Kac Propagators
holds for any G? -measurable random variable F. Next, replacing the cr-algebra cr (Xst,..., Xs„, Xs) by the cr-algebra Gs in formula (2.113) and using the monotone class theorem, we get condition (3) in Theorem 2.15. This completes the proof of Theorem 2.15. • Remark 2.7 By using the inclusion-exclusion principle as we have already done in the proof of the implication (1) =»• (2) in Theorem 2.14, we can show that the following condition is equivalent to conditions (l)-(5) in Theorem 2.15: (2') For any pair of stopping times S € M. and S £ A4 with S < S, any increasing finite sequence of stopping times {Si : 1 < i < n} such that Si < S for 1 < i < n, and any bounded Borel function / on E, the equality E [/ (XS) \a(XSl,...,
XSn,Xs)]
= E [/ (X§) | a (Xs)]
holds P-a.s. An assertion, similar to Theorem 2.15, holds for the cr-algebras Gs and GT • The proof of this fact is similar to the proof of Theorems 2.14 and 2.15, and we leave it as an exercise for the reader. Theorem 2.16 Let Ad be a family of stopping times that is closed with respect to A and V, and let Xs be a progressively measurable stochastic process on (fi, !F, P) with state space (E, S). Then the following are equivalent: (1) For all pairs of stopping times S\ 6 M and S% € M with S2 > Si, the following equality holds: E f(S2,XS2)
J a S l ] = E [f(S2,XS2)
I a(SuXSl)}
F-a.s.
for all bounded Borel functions f on [0, T] x E. (2) For any pair of stopping times S G M and S € M with S < S, any finite set of stopping times {Si : 1 < i < n] such that Si < S for 1 < i < n, and all bounded Borel functions f on [0, T) x E, the equality E f [S,Xgj f(s,Xs)
I a(S\,Xs1,.
..,Sn,XSn,S,Xs)
\a(S,Xs)
holds F-a.s. (3) Let S 6 M.. Then for any bounded GT F, the equality E F\Gs]=E[F\a(S,Xs)]
-measurable random variable
Propagators: General Theory
149
holds P-a.s. (4) For any S £ M, any bounded Gs-fneasurable random variable G, and any bounded real-valued QT -measurable random variable F, the equality E[GF}=E[GE[F\a{S,Xs)]] holds. (5) For any S e M, A^Gs, P [A n B |
and B e Q^S\ =P[A\
the equality
(7(5, XS)] P [B \ a(S, Xs)]
holds P-a.s. R e m a r k 2.8 Arguing as in the proof of the implication (1) ==> (2) in Theorem 2.14, we can show that the following condition is equivalent to conditions (l)-(5) in Theorem 2.16: (2') For all pairs of stopping times S e M and S £ M with S < S, all increasing finite sequences of stopping times {Si : 1 < i < n} such that Si < S for 1 < i < n, and all bounded Borel functions / on [0, T] x E, the equality E [/ (S,Xs)
\
= E [/ (s,X§)
|
a(S,Xsj
holds P-a.s. We conclude the present section by the definition of the strong Markov property of a stochastic process. It is based on condition (2) in Theorem 2.14 and on our choice of the future cr-algebra. Definition 2.15 A stochastic process {Xt,G,Gt,1?) with state space E is called a strong Markov process if for every stopping time S and every C w
bounded GT -measurable random variable F, the equality E[F\Gs)=E[F\a(S,Xs)}
holds P-almost surely.
150
2.7
Non-Autonomous Kato Classes and Feynman-Kac Propagators
Strong Markov Property with Respect to Families of Measures
This section is a continuation of Section 2.6. It is devoted to strong Markov processes with respect to families of stopping times and families of measures. Let (Xt,QJ,PT,x) be a non-homogeneous Markov process on {Q.,J-) with state space (E,£). Fix T € [0,T], and consider a pair (Si,S2) of Q\stopping times with T < S\ < S2 < T. One of the possible ways to define the strong Markov property with respect to the pair (Si, 52) and the family of measures {Ps,^ : r < s < T, y € E} is the following: ET,X [f (S2,XS2)
\GrSl]= E S l , x S l [/ (S2,XS2)}
P TlX -a.s.
(2.114)
for every bounded Borel function / on [r, T] x E. However, in order for the expressions in equality (2.114) to be well-defined, one has to impose certain restrictions on the process Xt and the stopping time S2. Let us first assume that Xt is an ^-progressively measurable process (see Definition 1.15). Then, since Si is a stopping time with T < S\ < T, the random variable X$ is QTS -measurable. Indeed, by the progressive measurability of the process Xt, the function (t, ui) 1—> Xt (OJ) is #[T'r]
Propagators:
General
Theory
151
stochastic process implies its progressive measurability (see Theorem 1.5), and the measurability of 52 with respect to the cr-algebra ^ T J ' V implies the measurability of Xs2 with respect to the same cr-algebra (see Theorem 2.17 below). Now we are finally ready to define the strong Markov property with respect to a pair of stopping times. Definition 2.16 Let P be a transition probability function, and let (Xt,Gl,fT<x) be a corresponding adapted Markov process. Fix r e [0,T], and suppose that Si and 52 are S^-stopping times with r < Si < S2 < T. Then the process Xt is called a strong Markov process with respect to the pair of stopping times (5i,52) and the family of measures {PS]!/ : r < s < T, y £ E} provided that the following conditions hold: (1) The process Xt is right-continuous. (2) For every B £ £, the function (T,x,t) >—> P(r,x;t,B) S[o, T] - measurable. (3) The stopping time S2 is £ T 1,v -measurable.
is
#[O,T]
®£ ®
(4) The equality ETtX[f(S2,XS2)\gTSi]=Es1,xSl[f(S2,Xs,)}
P T , x -a.s.
(2.115)
holds for all bounded Borel functions / on [T, T]x E. Recall that our choice for the cr-algebra representing the future after the stopping time Si is the cr-algebra QTl' . Hence, condition (3) in Definition 2.16 can be interpreted as follows: the stopping time S2 resides in the future after the stopping time Si. It remains to show that under the restrictions described in conditions (l)-(3) in Definition 2.16 the function on the right-hand side of (2.115) makes sense. Moreover, we will establish that this function is measurable with respect to the u-algebra a (Si, Xgj). We will prove this fact for more general functions of the form E s ^ s [F], where F is a £/ T 1,v -measurable random variable. Note that by condition (3) in Definition 2.16, 52 is QT1,S/measurable. Moreover, using condition (1) in Definition 2.16 and Theorem 2.17 below, we see that Xs2 is 5 T 1,v -measurable. Therefore, f(S2,Xs2) is GT1,V-measurable for any bounded Borel function / on [T,T] X E. Lemma 2.13 Let P be a transition probability function, and let (Xt, GJ, FV,x) be a corresponding Markov process. Fix r € [0, T] and x € E, and suppose that S is a Q\ -stopping time with r < S < T and F is a bounded GT'W-measurable random variable. Suppose also that condition (2)
152
N'on-Autonomous Koto Classes and Feynman-Kac Propagators
in Definition 2.16 holds. Then the expression Es,xs \F] *s well-defined, and moreover, there exists a bounded Borel function g on [0, T] x E such that Es,xs[F}
= g(S,Xs).
(2.116)
Remark 2.9 It follows from (2.116) that the random variable Ms,xs [F] is a (S, Xs)-measurable. Proof. Let "H be the vector space of all bounded QT' v -measurable random variables F satisfying the following conditions: (i) ^s,xs [F] exists PTiX-almost surely. (ii) There exists a bounded Borel function j o n [ 0 , T ] x £ such that the equality in (2.116) holds. It is not hard to see that H contains the constant functions and is closed under the convergence of uniformly bounded nonnegative nondecreasing sequences of its elements. Moreover, for any random variable F of the form F = / (5 V p,Xsvp), where / is a bounded Borel function on [T,T] X E and T < p < T, we have F € H. Indeed, the random variable F can be rewritten in the following form: ESM,X S ( I J ) (UO
[F] = Es(oi),x S M H [/ ( 5 M
= Y ( S M , S( W ) \fp)f
(S(w) Vp)
v
P>xs(u,)\/p(-))]
(XS{u){w))
= f f (5(w) V p, y) P (S{u), XS(UJ) ( w ); S(u) V p, dy) .
(2.117)
JE
Next, put 9(s,z)=
/ / (s V p, y) P (s, z; s V p, dy). JE
Then it follows from (2.117) and condition (2) in Definition 2.16 that F Let F be a random variable defined by
eH.
n+l
F=Y[fj(SVpj,XsvPj),
n>0,
3= 1
where T < p\ < p2 < • • • < pn < pn+i < T, and for every j with 1 < j < n + l, fj is a, bounded Borel function on [T,T] X E. Our next goal is to prove that any such random variable F belongs to H. We will use the method of mathematical induction in the proof. For n = 0, the assertion
Propagators:
General
Theory
153
above follows from (2.117). By the Markov property for constant times, we have n+1
E S(u>),Xs(u)(u)
I]/;(%)V ft ,Is M v ft (.)) 3= 1
= E S M , x S M ( w ) Hfj(S(u)VPj,Xs{u)s,Pi(-)) j=i
E
=
S ( w ) V p „ , X s ( „ ) V p n ( t j ) [/n+1 {S(U}) V
Pn+l,XS{oj)Vpn+1(-))]
(2.118)
ES(CJ),X,S(u>) M
j=l
where gj — fj for 1 < j < n — 1 and 9n{s,y)
= fn(s,y)Es,y
[/„+i (sV/)B+l,XsVp„+1(-))] •
Next, using condition (2) in Definition 2.16, we see that the function gn is a bounded Borel function. Taking into account (2.118) and the fact that F S H for n = 0, we see that induction gives F € H for all n > 0. By the monotone class theorem for bounded functions, we conclude that the space H contains all bounded C/ r ' v -measurable random variables. This completes the proof of Lemma 2.13. • The next theorem has already been used in the beginning of the present section. In our opinion, Theorem 2.17 has an independent interest. Theorem 2.17 Let Xt, r < t < T, be a right-continuous stochastic process with state space E, and let (Si, £2) be a pair of QJ -stopping times with r < Si < S2 < T. Suppose that S2 is measurable with respect to the a-algebra QT1: . Then the state variable Xs2 is measurable with respect to the same a-algebra. Proof.
Define a family of random variables by S2,n = Si +
T-Si 2"
'2"(52-5i)"
T-Si
, n> 1.
We will need the following lemma in the proof of Theorem 2.17.
154
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
L e m m a 2.14 For every n > 1, the random variable S2,n is a stopping time. Moreover, S2 < S ^ . n + l < S2,n
< 52 +
(2.119)
—.
Proof. It is not difficult to see that (2.119) follows from the definition of the function x 1—> \x\ (see Subsection 1.1.1). The fact that S2,n is a stopping time can be established as follows. For every r < t < T, we have 2"
{S2,n
f
u{
2"(52-5i)' T-S!
k \ n {52,„ < t)
fc=0 v-
S = {5i = S2 < t} U j j S.S1 + ^ - ^ (fc - 1) < S2 < Si + T -k
2™
fc=i *•
(2.120) It is also true that {S2 < t}\{Si
=
=S2
|J
= {Si
<S2
{Si
(2.121)
r€Qn[r,t]
Now it is not hard to see that {S\ = S2 < t} 6 Q\. For every k > 1, we have
{si + 7^(k-i)<S2<s1 Si + ^iJ1 = {ft + ^ ^
+
^^k
{k - 1) < S2 < t J n J S2 < S, + 1 ^k
n\sl + ^-¥^k
+ ^-^-k<S2
(2.122)
We will next prove that for every k > 0, the random variable Si H k is a stopping time. We have already established this fact for k — 0. The case where k = 2" is also simple. In the case where 1 < k < 2™, we have
{si + ^k
- fcT = {Sl<( *2" 2"-fc
This implies the assertion above.
A^
G &.
Propagators: General Theory
155
Let us go back to the proof of Lemma 2.14. By taking into account the fact that the random variable S\ H
——k is a stopping time and
reasoning as in the proof of (2.121), we see that the event on the right-hand side of (2.122) belongs to the cr-algebra Qt. Now it is clear that Lemma 2.14 can be obtained from (2.120), (2.121), and (2.122). • It follows from the right-continuity of the process Xt and from (2.119) that in order to establish the £ T 1,v -measurability of the state variable Xs2, it suffices to prove that the state variables Xs2iTV, n > 1, are QTl,vmeasurable. Consider the following events:
{*.» = Sl + ^ A f c } = [S2,n = (l " £ ) * + YnT) (2"123) where 0 < k < 2". It is not hard to see that the events in (2.123) belong to the cr-algebra QT1' . On these events, the state variables Xs2,n coincide with X
S1 + ^p~k
= X
(l-2*k)S1
+ &T-
Consider the following approximating sequence for Si: T •
2m{Sl~T)
S\,m —T-\ Since the process Xt is right-continuous, it suffices to prove that the state variable Xi, k \a , * ^ is 0^ x ' v -measurable. ^1— j r r j J l , m + 27r J
J
Let 0 < £ < 2 m . Then it follows from £ \ £ 2m J 2
m
~~
that on the event
{^=(1-^)-+^},
(2.124)
the equality
holds. Here we take into account that p (k, £) > Si. Therefore, on the event in (2.124) we have X
(}-Jk)Si.m+&T
= xs1vP(k,e)-
(2.125)
156
Non-Autonomous Kato Classes and Feynman-Kac Propagators
Now it is not hard to see that the QT1' -measurability of the state variable Xs2 follows from the ^| 1 , v -measurability of the events in (2.123) and (2.124) together with equality (2.125). This completes the proof of Theorem 2.17. • The next definition is motivated by condition (3) in Definition (2.16). Definition 2.17 Let Xt be a stochastic process, and let r £ [0,T]. The relation •< is defined on the set Sj. x <Sj. as follows: S x < S2 «=> Si < S2 and S2 is ££ 1,v -measurable.
(2.126)
Lemma 2.15 Suppose that Xt is a right-continuous process, and let r £ [0, T]. Then the relation •< defined in (2.126) is a partial ordering on the set o<ji x o^p. Proof. The reflexivity and antisymmetry of the relation X are clear. Next let Si, S2, S 3 £ Sj. x Sf with Si X S 2 d S 3 . Then, it is clear that Si < S3. Since S2 is GT1,y-measurable, S2 V a is also G? ' v -measurable for all a € [T, T]. It follows from the right-continuity of Xt that the function (S2 V a,Xs2va) is QT1,y-measurable (apply Theorem 2.17 to S2 V a and Si). Therefore, the cr-algebra QT2'V is contained in the cr-algebra QT1,V. Now the measurability of S3 with respect to the <j-algebra QT2'V implies that S3 is also G^1' -measurable. This completes the proof of Lemma 2.15. D Let P be a transition probability function, and let {Xt,Gt,^T,x) be a corresponding Markov process on Q. with state space E. Suppose that for every r £ [0, T], a family FT<X, x £ E, of probability measures is defined on the measurable space (fi, C/J.). Put
M°T=
[j
Sf.
re[0,T]
Here the symbol Sf stands for the family of all ^-stopping times where T
DSf.
Definition 2.18 A family M C Mj> is called an admissible family of stopping times provided that the following two conditions hold: (i) For every T £ [0,T], the family M(T) is closed with respect to the operations V and A.
Propagators: General Theory
157
(ii) For every r G [0,T], the conditions Si G M(T), S\ < S2 imply the condition Si •< S2.
S2 G M(T),
and
The next definition is a generalization of Definition 2.16. Definition 2.19 Let P be a transition probability function, and let (Xt,Gt^T,x) be a corresponding Markov process. Suppose that M is an admissible family of stopping times. Then the process Xt is called a strong Markov process with respect to the family of stopping times M and the family of measures {PT)X : 0 < r < T, x G E} provided that the following conditions hold: (1) The process Xt is right-continuous. (2) For every B £ 8, the function (T,x,t) i-» P(T,x;t,B) S[o,T]-measurable. (3) The equality ET,X [f (S2,XS2)
I FSl] = E S l , X s i [f(S2,XS2)}
is £[O,T] ® £ ®
P T , x -a.s.
(2.127)
holds for all r G [0,T], a; G £;, 5i G M{r), S2 G A ^ ( T ) with Si < S2, and all bounded Borel functions / on [r, T] x £\ Lemma 2.16 Let Xt be a strong Markov process with respect to an admissible family of stopping times M. and the family of measures PTiX. Then formulae (2.96) and (2.98) hold for every measure P r , x and all QJ -stopping times Si€M (T) and S2 G M (r) with r < Si < S2 < T. Proof. Let Xt be a strong Markov process with respect to the family M and the family PT<X, and let Si and #2 be stopping times such as in the formulation of Lemma 2.16. Then the random variable F = f (S2,Xs2) is Qp,v^measurable (see Definition 2.18 and Theorem 2.17). Therefore, equality (2.127), Remark 2.9, and the properties of conditional expectations give E S l ,x S l [f(S2,Xs2)}=ETtX
[f(S2,XS2)
I GTSl]
= ET,X [ET,X [f(S2,XS2) =
I GTSl]
\a(Si,XSl))
ET>x[f(S2,XS2)\a(Si,XSl)].
This implies condition (2.96). Condition (2.98) follows from (2.96), since
o-iSuXs^cg^cg^. This completes the proof of Lemma 2.16.
•
158
N'on-Autonomous
Koto Classes and Feynman-Kac
Propagators
The next theorems provide equivalent conditions for the validity of the strong Markov property with respect to a family of stopping times M and the family of measures PT)X- Recall that if M. is a family of stopping times, then the^family M(S) is denned by M(S) = {5" £ M(T) : S' > S} and the symbol QT ( stands for the a-algebra given by gP{s)
=a{S',Xs,:S'
£M(S)).
Theorem 2.18 Let P be a transition probability function, and let (Xt, Ql, P r ,x) be an adapted Markov process with P as its transition function. Suppose that Ad is an admissible family of stopping times, and assume that conditions (1) and (2) in Definition 2.16 hold. Then for every r £ [0,T] and x £ E, the following are equivalent: (1) For all QJ -stopping times S\ £ M (T) and S2 £ M (T) with r < Si < S2 < T and all bounded Borel functions f on [r, T] x E, the equality
Er,x [f{S2,XS2) I SJJ =Es l l X s i [f{S2,XS2)\ holds FTiX-a.s. (2) For all stopping times S £ M{T) and all bounded real-valued QT ^ measurable random variables F, the equality ETtX [F I Ql] = Es,xs [F] holds FT:X-a.s. (3) For any stopping time S £ M (r), any bounded Gg-measurable random variable G, and any bounded QT -measurable random variable F, the equality E r>x [GF] = ETtX [GEStXs [F]} holds. (4) For any A£gTs,
B £ g™{S), and S £ M (r), the equality
P r , x [Af)B\
a{S, Xs)} = Fr>x [A | a(S, Xs)} Fs,Xs [B]
holds FTiX-a.s. Proof. We will only prove the implications (1) =£• (2) and (4) =4> (1). The remaining implications can be obtained as in Theorem 2.14. (1) = * (2).
Propagators: General Theory
159
Suppose t h a t condition (1) in Theorem 2.18 holds. It will be first shown t h a t the equality E r , x [F | Ql] = Es,xs
[F]
(2.128)
holds for all random variables F given by
where for every 1 < j < m, fj is a bounded Borel function on [0, T] x E, and Sj £ M (S) is a stopping time such t h a t S < Si < • • • < Sm < T. We will use the m e t h o d of mathematical induction in the proof of equality (2.128). It follows from condition (1) in Theorem 2.18 t h a t (2.128) holds for m = 1. Next, let m > 1 and assume t h a t (2.128) is true for all 1 < k < m and for all finite subsets <Sj : 1 < j < k > of M (S) with S < Si < S2 < • • • < Sk < T. For any family of stopping times Si < • • • < Sm+i contained in the class M. (S), denote TC = a [Gs,Si,X^,...,Sm,Xg^j
with m + 1 elements
.
Then, using the tower property of conditional expectations, we get m+1 E
r,x
n
i=i
m = ET 3=1
/
..A
= ET,:
Y[fj {Sj,Xs.)ET,x
[/m+i (sm+i,Xgm+i)
| W| | Qs
Now it follows from condition (1) and from the inclusions
*(sm,X~Sm)cHcQlm
. (2.129)
160
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
that
/m+l [Sm+l,Xgm+i) I ftj
ET,
[fm+1 (5 m + i,X § m + i jJ.
= E
(2.130)
Next, we obtain from (2.129) and (2.130) that m+l
n/i(^'x5,)i^
ET,
" 1
m
ET
J J / j ^5j,X g jEg m)Xgm |/ m + i ^S m+ i,Xg m+ jJ | ^5 i=i
(2.131) Since M. is an admissible family, the stopping time 5 m +i is QsTmm',v - measurmm,v able. Moreover, Theorem 2.17 implies that Xg is Q->S T ' - measurable. It follows from Lemma 2.13 that there exists a bounded Borel measurable function g on [T, T] X E such that
g\Sm,XSm)
- E § m X . m |/m+i
[Sm+1,X§m+ij^
(2.132)
Therefore, (2.131) and (2.132) give m 771 + - J -l1
E.T , X
n m
— Er,:r
Ylfi{Si,xh)g(sm,xSm)\rs
(2.133)
By equality (2.130), the tower property of conditional expectations, and the induction hypothesis applied to the right-hand side of (2.133), we have m+l
ET,
Ilfi{Si,x§j)\gi
_j = l
m ••Es,. XS j= l
m
Es,x s
J J / j ( S ^ X g J E ^ ^ [/m+1 (S' m+1 ,X §m+i J J'=I
Propagators: General Theory
161
771
=
E
S,XS
1 1 /?' ( S i>- X 'sJ E S.*s fm+l \Sm+i,Xgm+i)
|W
j=i
n+1 = Es,x s
E
II £($.*§,) I w
S,XS
3=1
m+1
= Es, x s
iH(^*s,)
(2.134)
J' = l
It follows from (2.134) that equality (2.128) holds for any random variable F given by 771
F=
Hfj(sj,X~Sj)
where fj, 1 < j < m, is a finite sequence of bounded Borel functions on [T,T] x E, and Sj G M (S), 1 < j < m, is a finite sequence of stopping times satisfying r < S < Si < ••• < Sm < T. Our next goal is to prove the equality in condition (2) in Theorem 2.18 for a function F given by F = X
{ (§i.*Sl)eBl}
X
" 'X
X
«S-**.>M'
where B\,..., Bm are Borel subsets of [r,T] x £ , and Si,...,Sm are stopping times from the class M (S) satisfying Sj > S for all 1 < j < m. Define an increasing sequence of stopping times Sk by S'k = min lsjl
V • • • V Sjk : 1 < j i < • • • < j k < m j ,
1 < k < m.
It is clear that S'k £ M (S). For any permutation 7r of the set { 1 , . . . , TO}, put
^ = {(s;(1)I^(i))eBi,...>(s;(m)>^{m))€Bm}. Then we have
77
Next, using the inclusion-exclusion principle, we get F
= *{ ( § . . * * ) € * }
X
• • •X
X
{(Sm,XsJ^Bm}
162
Non-Autonomous
Kato Classes and Feynman-Kac
22
= Z2XA^TV
7r,
+
Propagators
XA„nA„,
7Tr:ir^nf
X)
XA^A„,nA„„ - • • • •
7T, 7 r ' , TT"-.TTy^TV1 ,-n^lt"
(2.135)
,7r' T ^ 7 T "
Since S£ is an increasing sequence of stopping times from the class M (S), equalities (2.134) and (2.135) imply that ET,I [F I Gs] = Es,xs [F]. By formula (2.110), the previous equality holds for all functions F which can be represented as follows: m
F=
l[fj(sj,X-S]).
Now using the monotone class theorem, we see that condition (2) in Theorem 2.18 holds. (4) = » (1). Suppose that condition (4) in Theorem 2.18 holds, and let A £ (?£ , B G §¥iSl), Si € M(T), and S 2 G M(T) with r < Si < S 2 < T. Then, using condition (4) in Theorem 2.18 with S = Si, we get FT,X [AnB\a(SuXSl))
= ET,X
[B]
[XAFSUXSI
\
P T):r -a.s. Taking the expectation ETyX in the previous equality, we see that P r , x \Br\A}= for all A e GTSl and B G G^(Sl)•
ET,X [Fsi,xSl [B] XA] It follows that
E r , x [/ (S 2 , X S a ) x^] = ET,X [ESuXsi
[/ (S 2 , XS2)]
Xi4 ]
,
and hence ET,X [f(S2,XS2)
| 0 J J =ESuXsi
lf(S2,XS2)}
This establishes condition (1) in Theorem 2.18. The proof of Theorem 2.18 is thus completed.
P r ,*-a.s.
D
The next theorem is similar to Theorems 2.14 and 2.18. Its proof is omitted.
Propagators:
General
Theory
163
Theorem 2.19 Let P be a transition probability function, and let (Xt,Gt^T,x) be an adapted Markov process with P as its transition function. Suppose that M. is an admissible family of stopping times, and assume that conditions (1) and (2) in Definition 2.16 hold. Then, for every T G [0, T] and x G E, the following are equivalent: (1) For all Qf -stopping times S\ G M. (T) and S2 G M. (T) with r < S\ < S2 < T, and all bounded Borel functions f on [r, T] x E, the equality I GTSl]
ET,x [f(S2,XS2)
=ESuxsJ(S2,Xs2)
holds P T)X -a.s. (2) For all stopping times S G M(T) and S G M (S), all finite families of stopping times {Si : 1 < i < n} such that r < Si < S for 1 < i < n, and all bounded Borel functions f on [r, T] x E, the equality o~ (Si, Xs1 ,...,Sn,
ET =
Xsn, S, Xs)
ESlxs[f(s,X§)~
holds P r>x -a.s. (3) For any stopping time S € M(T) random variable F, the equality T Er,x F\G S
and any bounded QT
-measurable
=Es,xs[F}
holds ¥r<x-a.s. (4) For any stopping time S £ A4 (T), any bounded Gs-measurable random variable G, and any bounded real-valued GT -measurable random variable F, the equality ET,X \GF] = ET,X [GEs,xs [F]} holds. (5) For A G Gs, B & GT PT,X [AHB\
> and any stopping time S G M (T), the equality a(S,Xs)]
= PT,X [A I a(S,Xs)]
Fs,xs [B]
holds P r>x -a.s. Finally, we are ready to define the strong Markov property of a stochastic process with respect to a family of measures.
164
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
Definition 2.20 Let P be a transition probability function, and let (Xt,Ql,PTtX) be an adapted Markov process associated with P. It is said that the process Xt is a strong Markov process with respect to the family of measures PT,X if for every r e [0, T], x e E, every ^-stopping time S with C w
T < S
(2.136)
for all T
[F]
(2.137)
where F is any bounded Gp -measurable random variable. Let Y be the free backward propagator associated with P. It is defined for all 0 < r < t < T and / e Co by the following formula: Y{r,t)f(x)=
[
f(y)P(r,x;t,dy).
JE
Recall that for a given stopping time S, the symbol £^' v stands for the cr-algebra a (S V p, Xsvp '• 0 < p < T) (see Definition 2.11). It is clear that for a stopping time S with r < S
0 < p
:
T
Theorem 2.20 Let P be a transition probability function, ( X t , ^ , P T i X ) be an adapted Markov process associated with P. that the following conditions hold:
and let Suppose
(1) The process Xt is right-continuous. (2) For every B € £, the function (T, X, t) i-> P (r, z; t, B) is B[o,r] ® £ ® #[o,r] -measurable.
Propagators:
General
Theory
165
(3) For any function f e Co and t € (0,T], the function (T,X) H-» Y (r, t) f(x) is continuous from the right in r on the interval [0, i) and continuous in x on E. Then Xt is a strong Markov process with respect to the family {P T)X }. Remark 2.10 ing condition:
Condition (3) in Theorem 2.20 is equivalent to the follow-
lim
EtiX[f(Xt)]=Et0,X0[f(Xt)]
siso,x—*xo
for all XQ S E, T < SQ < t, and f £ Co- It follows from conditions (1) and (3) in the formulation of Theorem 2.20 that lim ESiXs [/ (Xt)} = E S0 , X
[/ (Xt)],
P TiX -a.s.
(2.138)
for all x € E, f e C0, and 0 < r < s 0 < t < T. Proof. Let T € [0,T], x € E, and let S be a ^-stopping time with T < S
r,x
[F | gTs] = Es<Xs
[F]
(2.139)
holds PT)X-almost surely for any bounded QT' -measurable random variable F. We will derive equality (2.139) from the following assertion. For all r e [0,T], x e E, all bounded Borel functions / on [T,T] X E, and all T
ET,x[f(svp2,xSvP2)\gTSVPi] = E s v p 1 , x S V p i / ( 5 v / 9 2 , X S V p 2 ) Pr,x-a.s.
(2.140)
Indeed, if (2.140) holds, then the validity of (1) = > (2) in Theorem 2.18, where the family M is defined by .M = { S V p : r < p < T } , implies equality (2.139). It remains to prove equality (2.140). Let r < p\ < P2 < T, and put S = S V p\. Then S V p2 = S V p2- Consider the following approximations from above of the stopping time S: ~ Sm=T
T-T
+ ——
•(S-0 T-T
m> 1.
166
Non-Autonomous Kato Classes and Feynman-Kac Propagators
Next we will prove that for all m > n and all functions f £ Co ([r, T] x E),
f(snvp2,x~Snyp2)\gzm
Er, :
= E~sm,x-sj{SnVP2,X~Snyp2)
P T , x -a.s.
(2.141)
Let us first show that (2.141) implies (2.140). Assume that (2.141) holds, and let w G fi. Then the right-hand side of (2.141) evaluated at w is equal to E
§m(„),XSmluy{u)f
{Sn{")V
P2,X~Sn{ul)yp2{-))
= Y (Sm (w), Sn (to)) f (Sn(u>) V p2) (Xg n ( w ) V p 2 ) ,
(2.142)
where the function / (Sn (u) V p2 J is defined by / (Sn(w) V P2) (y) = / (Sn(w) V p2, y) , » G £ . By the right-continuity of the process Xt, condition (2) in Theorem 2.20, and equality (2.142), we see that lim Es , , „ B ) ( B ) / ( g , ( w ) v f t l X j , ( u ) V B ( . ) ) m-»oo ,5'"^a'''Asm(, • EW l M ( ^ ( ^ ) V f t - ^ ) v W 0 ) c
(2-143)
for P r , x -almost all w £ f l . Let A G 0 | . Then A £ QZ for all m > n, and hence (2.141) gives /(5„VP2,^Vp2),A = E.
Ee
Y_
/(5„Vp2,^snVp2
,A
(2.144)
Passing to the limit as m —> oo in (2.144) and using (2.143), we get E. -,x[f(snVp2,X~s^p2),A = ET
E
§[f(snVP2,X-SnVp2),A
(2.145)
for all / G CQ([T,T] X E). NOW taking into account the continuity of the function / and the right-continuity of the process Xt, we see that the dominated convergence theorem can be applied in (2.145). Therefore, ET,:
'f(svP2,X~SVp2),A
Propagators: General Theory
= ET
E
3.*s
f{sVp2,X~SVp2),A]\.
167
(2.146)
The random variable S is (/--measurable (use the definition of the cr-algebra Q-,). Moreover, the state variable Xs is also C/i-measurable. This follows from Lemma 2.8 and from the fact that the right-continuity of a stochastic process implies its progressive measurability. Hence, the random variable ^s,x- f \S v P2^svP2) *s ^ s - m e a s u r a b l e . Now, using the definition of conditional expectations and equality (2.146), we see that equality (2.140) holds for all f e Co ([r, T] x E). By Urysohn's Lemma and the monotone class theorem, equality (2.140) holds for all bounded Borel functions. This establishes the implication (2.141) ==>• (2.140). Next we will prove equality (2.141). The following lemma will be needed in the proof. Lemma 2.17 Let n be a given nonnegative integer. Then for every integer m with m > n and every integer k with 0 < k < 2m, there exists a unique integer (.^m such that
k < 2m-ne<£]m < 2m-n + jfe - 1. Proof. It is easy to see that Lemma 2.17 holds for m = n. Now suppose that m ^ n. Then the inequalities in the formulation of Lemma 2.17 are equivalent to the inequalities k2n-m
< e(n)^
< ^n-m
+
1
_ yi-m_
(2.147)
It is clear that the interval [fc2"~m, k2n~m + 1 - 2™~m] contains at most one integer. We will show that
4*2, = [k2n-m + 1 - 2"" m J = \2n~mk] .
(2.148)
Let I = \2n-mk\. Then 2n~mk <£< 2n~mk + 1, and hence k < 2m~n£ < k + 2m~n. It follows that k < 2m~nt
< £ < k2n-m
+ 1 - 2n~m
and i < [k2n~m + 1 — 2n~m\. Combining the previous equality with the facts that I = [2 n-m fc] and there is at most one integer between k2n~m and k2n~m + 1 - 2 n " m , we see that (2.147) holds. This completes the proof of Lemma 2.17. •
168
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
We are finally ready to finish the proof of Theorem 2.20. Note that it only remains to establish equality (2.141). For n > 1, m > n, and 0 < k < 2 m , choose v£m as in Lemma 2.17. Then for Sm defined by
(S-r)
T-T
Sm = r +
T-T
we have
{S™ = r + ^ f c } C {§ n = r + ^ C }
•
(2.149)
Indeed, the event on the left-hand side of (2.149) is the same as the event f 2™ ( S - r ) ) < k — 1 < —•£ — < k >, while the event on the right-hand side of T T
"
I
J
, .
2" f 5 - T)
'- < C ,)
(2.149) coincides with the event { 4?m - 1 < — £
fc,m
Now
we see that (2.149) follows from Lemma 2.17. Next put f(Jfc,m) = r + ^ - = ^ f c =
(I-1-)T+^-T
and «(«)
t (n, k,m) = T+
Z^ffrl km
2"
^'
-•
~ \
fc,m
«(«) I
2"~ /
.
k,m.rp
2"
'
Then, it follows from the properties of the number P£'m that t (k, m) < t(n,k,m). Therefore, (2.149) gives = t ( f c , m ) | C JS„ = t(n,
fc,m)|.
Let A G gz . Then, using (2.150), we get
2m
- £ E T , S [/ (§n V p 2 , X § n V p 2 ) , A n { s m = tffc.m)}]
(2.150)
Propagators: General Theory
] T E r , x f(t(n,k,m)Vp2,Xt fc=0 2m = ^2^r,x [Er,i / (t (n, fc, m) V p2, ' =0 k=0
169
),An{sm=t{k,m)}} Xt(nXm)yp2),
An{sm = t(k,m)}\gj{k:,m)
(2.151)
It follows from the definition of the cr-algebra Q\
that
An{sm = t(k,m)}egiik<m). Therefore, (2.151) gives
ET,x[f(snVP2,X~Snyp2),A =
^^T,X ' =o fe=0
| E T , I / (* (", fc, m)\J p2, X t ( n i f c i m )v P 2 ) | £tTt(fc,m) (
An{§ m = t(fc,m)}] .
(2.152)
Next, applying the Markov property for constant times and using (2.150) in (2.152), we obtain
, [f(snVp2,X~Snyp2),A
E.T x
= ET , £
YlEt(k>m)>XHk,m)
[f (l (n> fc> m )
V
^2, *t(n,fc,m)Vp2)]
.fc=0
X{Sm=t(fc,m)}> ^ •
(2.153)
Since the r a n d o m variable 2m t(fe,m),Xt(fcm) [/ (i ( l , k, m) V P2> -^t(n,fc,m)VP2)] X{s m= t(fc,m)} fc=0 is C/~ -measurable, inclusion (2.150) and equality (2.153) give E.
-4f{snvpi,x~SnVp2)\gzm_ ^Z^t(k,m),xt(k,m) k=0
[f (t (n, k, m) V p2, X t ( n > f c ) m ) V p 2 )] X { § m = t ( f e , m ) }
170
Non-Autonomous
= HE§m,X-Sm
Kato Classes and Feynman-Kac
[f (Sn V p 2 , X g n V p 2 ) ]
Propagators
X{sm=t(k,m)}
k=0
= E
f(snvP2,x~SnVp2)'
This implies equality (2.141). The proof of Theorem 2.20 is thus completed.
•
Remark 2.11 Fix r € [0, T] and x £ E, and denote by Q^'y the completion of the cr-algebra Q^y with respect to the measure P r , x . Arguing as in the proof of Theorem 2.20, we see that the ^y' v -measurability assumption for the random variable F in the formulation of the strong Markov property can be replaced by the QT1,V -measurability of F. This means that under the conditions in Theorem 2.20, the GT' -measurability of the random variable F implies the strong Markov property in the following form: ^r,x[F\Gs1]=^S1,XSl[F] P TiX -almost surely. The next assertion follows from Theorem 2.20. Corollary 2.1 Let P be a transition probability function, (Xt,Gl,¥TtX) be an adapted Markov process associated with P. that the following conditions hold:
and let Suppose
(1) The process Xt is right-continuous. (2) For every B e £, the function (T, X, t) t-> P (r, x; t, B) is S[ 0 ,T] ® £ ® S[o,r] -measurable. (3) For every f E Co and t 6 (0,T], the function (T,X) t-> Y (r,t) f{x) is continuous from the right in T on the interval [0, t) and continuous in xeE. Let M. be an admissible family of stopping times. Then the process Xt is a strong Markov process with respect to the families M and {PT,x}Next we will give examples of families of stopping times which can be used as the families M(T) in Corollary 2.1. More examples will be given in Section 2.10. Example 2.3 Let P be a transition probability function, and let {Xt,Gl,PT,x) be a corresponding Markov process. Assume that the process Xt is right-continuous, and fix r £ [0,T]. For a given (/^-stopping time S with T < S < T, consider the following family of stopping times:
Propagators:
General
Theory
171
M = M(T) = {(S + p) A T : p > 0}. It is not difficult to prove that the family M satisfies the following condition: for all S\ G M and 52 G M with r < Si < S2 < T, the stopping time S2 is measurable with respect to the ex-algebra a (Si) C GTl'y. In the next example, we will consider a more general case. E x a m p l e 2.4 In this example, the family M = M(T) consists of certain functions of a given stopping time. Let r G [0, T], and let S be a ^-stopping time with r < S < T. Suppose that the process Xt is right-continuous, and consider the family $ of all functions cp : [r, T] —• [r, T] satisfying the following conditions: (1) ip is continuous on [T,T]. (2) There exists a point u(tp) G [r, T] such that the function
for all t G [r, u(tp)].
(2.154)
In order to show that any random variable
MS) < t} = n G etr. If r < t <
C/?(T),
then
MS)
then
M S ) < <} = ( v t o < v(5) < t } = { 5 £ [ r , ^ - 1 ^ ) ] } G 0J_ 1 ( t ) C 0tT. Therefore,
for all r G S ( f i ) .
(2.155)
172
Non-Autonomous Kato Classes and Feynman-Kac Propagators
Consider the following function on the interval [<£I(T),T]:
• (-)-{*? (WW)
SUM™
P-156)
Then 4* is a Borel function on [
for all u e f l .
(2.157)
Indeed, if S(u) G [T,U(V?I)), then (2.157) follows from (2.154) and (2.156). If S(u)) G [u((pi),T], then cpi(S(u)) = T, and (2.156) implies the equality $(»Ji(5(w))) = T. On the other hand, (2.155) gives
the stopping time
2.8
Feller-Dynkin Algebras
-measurable.
Propagators and Completions of cr-
In this section we continue our study of Markov processes associated with Feller-Dynkin propagators. Let us first recall the construction of the process (Xt,Jrt+,PT,x) in Theorem 2.13. For a given transition probability function P , we consider the standard realization Xt of a corresponding Markov process on the space (_El°'Tl, .Ff, Fs,x j . Then, by restricting the process Xt to the space Q of all right-continuous E-valued functions on [0, T] having left limits, we get the process Xt that is stochastically equivalent to the process Xt- It follows from Theorem 2.13, Lemma 2.6, and Lemma 2.7 that if the free backward propagator Y associated with P is a strongly continuous backward Feller-Dynkin propagator, then Xt is a Markov process with respect to the filtration TJ+. This means that Es,x [f (Xu) \^+]
= EtiXt [f (Xu)}
(2.158)
P S)X -a.s. for all 0 < s < t < u < T and all bounded Borel functions / on E. Since Xt is a Markov process with respect to the filtration .FtT, we have Es,x [f (Xu) \^+]
= Es,x [/ (Xu) \F?}
(2.159)
Propagators:
General
Theory
173
P S|X -a.s. for all 0 < s < t < u < T and all bounded Borel functions / on E. It follows from (2.158) and (2.159) that Es,x [f (Xu) \FJ+] = Es,x [f (Xu) \FZ\ = E t ,x, [/ (Xu)}
(2.160)
P s , x -a.s. for all 0 < s < r < t < u < T, x e E, and all bounded Borel functions / on E. In (2.160), the measure P SiX is restricted to the cr-algebra [FT] T • Recall that the symbol [FT] r stands for the completion of the cralgebra FT with respect to the family of measures defined by VT = {fs,x
(2.161)
:0<S
We will also need the cr-algebras F[ and F[+ which are the completions of the cr-algebras FJ and F[+ with respect to the family of measures in (2.161) (see Section 1.7). Next, suppose that r
Ui+i < • • • < Un < T,
and let fk be bounded Borel functions on E. Then, using (2.160) and the equivalence (1) •*=> (3) in Theorem 2.16, we get E,
[F\F?+] =
Hfk(XUk)EStX I I fk(XUk)\FT+ Lfc=i+1
fc=i
ll fk fc=i
(XUk)Es,x
n
h{xUk)\Fi
= E,,x
[F\Fl\,
(2.162)
.k=i+l
By the monotone class theorem, (2.162) implies that the equality E s , x [F\FJ+] = Es,x [F\Fl]
(2.163)
holds P S)X -a.s. for a ! 1 0 < s < r < f < T ' and all ^y-measurable random variable F. It is assumed in (2.163) that the measure P 5)X is restricted to the cr-algebra [FT] T. L e m m a 2.18 Let (Xt,F[+,FT<x) be the process constructed in the proof of Theorem 2.13. Then for allO < r
174
N'on-Autonomous Koto Classes and Feynman-Kac Propagators
Proof.
The inclusion F[ C Ft+
(2.164)
has already been established (see (1.46)). Next we will prove the opposite inclusion. Let A £ J^+. Then equality (2.163) with (s,x) £ [0,r] x E and F = XA shows that the function XA is equal to an ^-measurable function almost surely with respect to the measure P s , x restricted to the cr-algebra [!FT\ T • It follows that for every pair (s,x) £ [0,r] x E, there exists a set As,x € Tl such that AAAS,X € [F£fs'x and Fs<x (AAASyX) = 0. This means that for all (s, x) € [0, r] x E, the set J4 belongs to the completion [??}V''X of the cr-algebra Tl in the cr-algebra [ J T r ] P s ' x . It is not hard to prove that for all (s, x) £ [0, r] x £ , we have
J7+c[J7]p-. Therefore, [ ^ + ] P ^ C [^T] P "* .
(2.165)
By intersecting the sets in (2.165) with respect to all (s, x) £ [0, r] x E , we obtain the inclusion Ft+ C fl. Now it is clear that Lemma 2.18 follows from (2.164) and (2.166). 2.9
(2.166) •
Feller-Dynkin Propagators and Standard Processes
It is natural to expect that under certain restrictions on a transition probability function P , the class of all Markov processes associated with P contains a process with special properties. For instance, the process Xt constructed in the proof of Theorem 2.13 is such a process. In this section we continue our discussion of the behavior of this process. Let us recall that the process Xt is defined on the space Q of all E'-valued functions on the interval [0, T] which are right-continuous on the interval [0, T) and have left limits on the interval (0, T]. Define a family {&[} of cr-algebras by GI =?J,
0
(2.167)
where JFtr is the completion of the cr-algebra TJ with respect to the family of measures VT = {PSjX : 0 < s < r, x £ E}. The family {QJ} is a two-
Propagators: General Theory
175
parameter filtration. It is not hard to see that flcgiC
[^
•
(2.168)
For all r € [0,T], the process Xt, r < t < T, is adapted to the filtration {QJ}. It is also clear that for every pair (T,X) 6 [0,T] x £J and t > T, the measure PT,X can be extended to the <7-algebra Q\. By (1.47) and Lemma 2.18, we see that Q\ = Gl+, 0
0
xGE.
(2.169)
It follows from (2.160) that the process (.Xt, 5^,P r ,x) is a Markov process, that is, ! . , * [/ (Xu) \QTt) = Et,xt [f (Xu)} for all 0 < s < T < t < u < T, x G E, and all bounded Borel functions / on E. Moreover, using Corollary 2.1, we see that Xt is a strong Markov process. Recall that this means the following. Let M be an admissible family of stopping times. Then for all r £ [0, T], x G E, Si £ M(r), and S2 e M(T) with r < Si < S2 < T, the equality Er,x [f(S2,XSa)
I GTSl] =KsltxSl
lf(S2,XS2)}
(2.170)
holds PTi3;-almost surely for any bounded Borel function / on [r, T] x E. The process in (2.169) has left limits but is not necessarily leftcontinuous. However, it satisfies a weaker condition which is called the quasi left-continuity. The next definition explains what the quasi left-continuity means for a general stochastic process. Definition 2.21 A stochastic process ( ^ t , ^ , P r , x ) adapted to the filtration Q1 is called quasi left-continuous provided that for all (r, x) G [0, T] x E, all ^-stopping times S, and all non-decreasing sequences Sn of <7(T-stopping times such that r < Sn < S < T and S = lim Sn, the n—>oo
equality lim Xsn = Xs holds P r x -a.s. n—+00
'
T h e o r e m 2.21 Let P be a transition probability function such that the corresponding free backward propagator Y is a strongly continuous backward Feller-Dynkin propagator. Then the process (Xt,Gt,PT,x) in (2.169) is quasi left-continuous.
176
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
Proof. Let Sn and S be as in Definition 2.21. It follows from Theorem 2.13 that the process Xt has left limits on (T,T\. Set Xs- = lim Xsnn—*oo
Then Corollary 2.1 and Example 2.3 show that for every n > 1, the process Xt is a strong Markov process in the sense of equality (2.170). Here the family of stopping times M is defined by M = {(Sn + p) AT : 0 < p
\GTSn] = E5„,x Sn [/ (X{Sn+p)AT)]
(2.171)
P TiX -a.s. for all p > 0, n > 1, and all bounded Borel functions / on E. Now let / and g be functions from the space CQ. Then, using the rightcontinuity of the process Xt and equality (2.171), and taking into account the restrictions on the backward free propagator Y in Theorem 2.13, we get E r , x [f(XS-)g(Xs)\
= lim lim E r , x [f(XSn)g
= lim lim ET,X [f(XsJET,x
(X(Sn+p)AT)}
[g (X{Sn+p)AT) | S S J]
= lim lim ET,X [/ ( X S J F ( S n , (5„ + p) A T) (XSn)\ pj.0 n—KX>
= limE T , X [/ ( X s - ) F ( S , ( 5 + p ) A T ) 5 ( X s _ ) l = ET,x[f(XS-)g(XS-)].
(2.172)
Taking /(x) = 1, x € £J, in (2.172) and replacing g by g2, we get Er,x[52(Xs)]=ET,x[92(Xs-)] for all g £ Co. Moreover, taking / = g in (2.172), we obtain E r , x [ 5 (Xs-)g (Xs)} = ET,X [52
(XS-)]
for all g € C0. It follows that ET,X [( 5 (X 5 _) - 9 (Xs))2]
= ET)X [g2 (XS-)]
- 2ET,X [ ( X s _ ) ff (X s )]
+ E T , X [, 25 2 (X S )]=0
Propagators: General Theory
177
for all 5 € Co- Hence, g(XS-)=g(Xs)
(2.173)
PT)1:-a.s. for all g e Co. Since the space Co is separable, the exceptional set in (2.173) can be chosen independently of g. Since the functions from Co separate points, we have Xs~ = X$ PT)a;-a.s. This completes the proof of Theorem 2.21. • The next definition concerns general adapted stochastic processes (Xt,0l,PT,x)It is based on the properties of the process in (2.169). Definition 2.22 An adapted stochastic process (Xt, Q\, PT,X) is called a standard process provided that (1) The process Xt is right-continuous and has left limits.
(2) gi = gi+ = g[. (3) For every admissible family M of stopping times, the process Xt is a strong Markov process with respect to M. and {Pi-,*}. (4) The process Xt is quasi left-continuous. R e m a r k 2.12 Let (Xt, Ql,PT,x) be a standard process. Recall that for fixed r € [0, T] and x G E, the symbol (? r ' v stands for the completion of the cr-algebra £7T'V = a (S V p, Xsvp '• 0 < p < T) with respect to the measure P T)X . Then, for all pairs of stopping times (Si, Si) such that 52 is ^ T 1 , v measurable, all T < Si < S2 < T, and all bounded Borel functions / on [T, T] x E, the equality ETtX [f(S2,XS2)
I QTSl] =ESuXai
[f(S2,Xs2)}
holds PT)X-almost surely (see Remark 2.11). Now we are ready to formulate one of the main results of this chapter. Theorem 2.22 Let P be a transition probability function such that the corresponding free backward propagator Y is a strongly continuous backward Feller-Dynkin propagator. Then there exists a standard process (Xt,Qt,PT,x) with P as its transition function. Proof. Since the stochastic process in (2.169) satisfies conditions (l)-(4) in Definition 2.22, it is a standard process. It is also clear that this process has P as its transition function. This completes the proof of Theorem 2.22. •
178
Non-Autonomous
2.10
Kato Classes and Feynman-Kac
Propagators
Hitting Times and Standard Processes
In this section we study how fast a Markov process reaches a Borel subset of the state space. Definition 2.23 Let (Xt,Gt^T,x) be a Markov process on Q with state space E, and let fibea Borel subset of E. Let r £ [0, T), and suppose that S : O —> [T, T] is a ^-stopping time. For the process Xt, the entry time of the set B after time S is defined by ' inf {t : t > S, Xt € B}
on
D%
[j
{S < t, Xt € B} , (2.174)
T
T elsewhere. The pseudo-hitting time of the set B after time S is defined by ' inf {t : t > S, Xt &B}
on
(J
{5 < t, Xt £ B} , (2.175)
T
D% = < T elsewhere.
Finally, the hitting time of the set B after time S is defined by ' inf {t: t > S, Xt £ B}
on
T§
\J
{S < t, Xt E B} ,
T
(2.176)
T elsewhere. It is not hard to prove that (J
{S
XtGB}=
t-.T
(J
{SVt
Xsvt&B}
t:r
and U
{S
t:r
(J
{SVt
t:r
We also have DSB < DSB < T§. equalities hold:
Next we will show that the following
Ti=i„f{ D r w }=^{^ + s , A r }-
<"">
Indeed, the first equality in (2.177) can be obtained using the inclusion {t > (e + S) A T, XtGB}c{t>
S, Xt e B}
Propagators:
General
Theory
179
and the fact that for every t G [T,T) and u> G {S < t, Xt G B} there exists e > 0 dependent on w and such that w G {(e + S) A T < t, Xt G B). The second equality in (2.177) follows from the monotonicity of the entry time DQ with respect to 5. Our next goal is to prove that under certain restrictions on the process (Xt,Gl,PT,x), the entry time Df, the pseudo-hitting time D%, and the hitting time Tf are stopping times. Throughout the present section, the symbols K and O will stand for the family of all compact subsets and the family of all open subsets of the space E, respectively. The Choquet capacitability theorem will be used in the proof of the fact that Dg, Dg, and T§ are stopping times. We will restrict ourselves to positive capacities and the pavement of the space E by compact subsets. For more general cases of the Choquet theorem, we refer the reader to [Doob (2001); Meyer (1966)] (more references can be found in Section 2.11). Definition 2.24 A function / from the class P(E) of all subsets of E into the extended real half-line R + is called a Choquet capacity provided that (i) If Ai G V(E) and A2 G V(E) are such that Ai C A 2 , then
I(A1) 1, and A e V(E) are such that An | A, then I (An) —»1(A)
as n —> oo.
(iii) If Kn G /C, n > 1, and K G K, are such that Kn j K, then I (Kn) —> I(K)
as n —> oo.
Definition 2.25 A function ip : K. —+ [0, oo) is called strongly subadditive provided that the following conditions hold: (i) If Kx G K and K2 G K, are such that Ki C K2, then tp (K{) <
(Ki) +
(2.178)
The following construction allows one to define a Choquet capacity starting with a strongly subadditive function. Let ip be a strongly subadditive function satisfying the following additional condition:
180
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
(iii) For all K G K. and e > 0, there exists G G O such that K c G and
sup
(2.179)
and define a set function 7 : V(E) —> M.+ by 1(A) =
Ggmf cG/*(G),
4£P(£).
(2.180)
It is known that the function J is a Choquet capacity. It is clear that for any G G O, we have 7(G) = I*(G). Moreover, it is not hard to see that for any K G K,
(2.181)
Now we are ready to formulate the Choquet capacitability theorem (see, e.g., [Doob (2001); Dellacherie and Meyer (1978); Meyer (1966)]). Theorem 2.23 Let (p : K —> [0, oo) be a strongly subadditive function satisfying condition (iii), and let I be the Choquet capacity generated by (p. Then every analytic subset of E, and in particular, every Borel subset of E is I-capacitable. The definition of analytic sets can be found in [Doob (2001); Dellacherie and Meyer (1978)]. We will only need the Choquet capacitability theorem for Borel sets. The symbol P(E) will stand for the collection of all Borel probability measures on the space E. For B G £ and (J, £ P(E), we put Pr, M (£)= f
FT,x(B)dn(x).
For instance, if /J, = 6X is the Dirac measure concentrated at x € E, then Pr,5x = Pr,x-
Lemma 2.19 Let r G [0,T], and let ( X t , ^ , P T , x ) be an adapted, rightcontinuous, and quasi left-continuous stochastic process. Suppose that S is a Ql -stopping time such that T < S < T. Then, for any t G [T,T] and
Propagators:
General
Theory
181
fj, G P{E), the following functions are strongly subadditive on K. and satisfy condition (Hi): K ^ PT,M [1%
and K^
P Ti/t \D% < t\ ,
K G K.
(2.182)
Proof. We will check conditions (i) and (ii) in Definition 2.25 and also condition (iii) for the set functions in (2.182). Let K\ G K. and K% G K. be such that Ki C K2. Then D^ > D^2, and hence Pr,„ [DSKl
T,MP&<*]<
inf
< inf
sup
sup
PT,„ [ £ & < * ]
P T M [Dfc, < t]
n&K'^K-.K'CGn
< inf PT>/1 [DsGn
(2.183)
It follows from (2.183) that Pr,„[DSK
inf GeO-.GDK
sup
P Ti/1 [£>&,< t] .
(2.184)
K,eic:K'cG
Now it is clear that the equality in (2.184) implies property (iii) for the function K H-> P T I M [D^ < t]. The proof of (iii) for the function K t-> PT>/i DK < t is similar. Here we use part (d) of Lemma 2.22 (note that part (d) of Lemma 2.22 holds under the restrictions in Lemma 2.19). Next we will prove that the function K i-> PTiM [D^ < t] satisfies condition (ii). In the proof, the following simple facts will be used: for all Borel subsets B\ and B? of E, DSBlUB2=DSBlADsB2,
(2.185)
DSBlnB2>DSBlVDsB2.
(2.186)
and
182
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
By (2.185) and (2.186) with Kx G K. and K2 G K, instead of Bx and B2, respectively, we get {DSKlUK2 s
< t} \ {DSK2
= {D Kl < t) \ {D
{DSK2 < t}) \ {DSK2 < t)
s
K2
= {DsKx < t) \ {DsKl V DSK2
{DsKl < t) \ {DsKinK2
< t} . (2.187)
It follows from (2.187) that PV,M [DkuK2 < t] +PT,M [DSKlnK2 < t] < Pr,M [DsKl
(2.188)
Now it is clear that (2.188) implies condition (ii) for the function K i-> PT]M [Dfc < t]. The proof of condition (ii) for the second function in Lemma 2.19 is similar. This completes the proof of Lemma 2.19. • The next theorem states that under certain restrictions, the entry time £>§, the pseudo-hitting time DQ, and the hitting time Tg are stopping times. Recall that we denoted by (% the completion [QJ] r of the a-algebra Ql with respect to the family of measures VT = {P s>x : 0 < s < r, x € E} (see Section 1.7). Theorem 2.24 Let (Xt,Qf,PTiX) fying the following conditions:
be an adapted stochastic process satis-
(i) The process Xt is right-continuous and quasi left-continuous. (H) Ql = GI+ = GTt forO
[T,T],
the
Proof. We will first prove Theorem 2.24 assuming that it holds for all open and all compact subsets of E. The validity of Theorem 2.24 for such sets will be established in Lemmas 2.20 and 2.21 below. Let B b e a Borel subset of E, and suppose that we have already shown that for any e > 0 the random time D^ 'A is an (/^-stopping time. Since T§ =
inf £>o, £ eQ+
D%+S)AT
Propagators:
General
Theory
183
(see (2.177)), we see that T§ is a ^ - s t o p p i n g time. Therefore, in order to prove that T§ is a (^.-stopping time, it suffices to show that for every Borel subset B of E, the random time Dg is a FTifl [D^ < t]. Therefore, there exists an increasing sequence Kn £ /C, n £ N, and a decreasing sequence Gn £ O, n £ N, such that Kn c Kn+i
C B c Gn+1 c G „ , n e N,
and 8upPT,^[£>^
A^>s=\J{DsKn
and A^' s = f) {^„ < t} •
Then Lemma 2.20 implies Al'^s £/^+. Moreover, we have
t2-^)
£ QJ+, and Lemma 2.21 gives A[' M,S G
A[ , " , s c {DB
AJ"'5,
(2.190)
and •^2
infPT,M[Z?Sn
n£N
sup PT,M [£>^n < t] = PT,M fAl'"' s ] .
(2.191)
It follows from (2.190) and (2.191) that Ar,M,S •T,H
\
A r,/x,5
= 0.
Then, using (2.190) again, we see that the event {Dg < t} belongs to the a-algebra Ql+. Therefore, the random time Dg is a 5J"+-stopping time. As we have already observed, it also follows that the random time Tg is a ^.-stopping time. A similar argument with Dg replaced by Dg shows that the random time Dg, B £ 8, is a (^.-stopping times.
184
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
This completes the proof of the fact that Theorem 2.24 for all open and all closed sets implies the general case. • Next we will prove that Theorem 2.24 holds for all open sets. Lemma 2.20 Let S : fi —> [r, T] be a QrtJr-stopping time, and let G £ O. Then the random times DG, DG, and TG are Ql+-stopping times. Proof.
It is not hard to see that
{DG
*•
•*
Q
(J
{S
m€NT
(2.192) The last event in (2.192) belongs to the cr-algebra and hence the random time DG is a (/^-stopping time. The fact that DG is a ^.-stopping time follows from
{DG
i +
i \ = p|
m€N *•
'
U
{S
m6Npg(Tit+i)nQ+
The equality (2.177) with G instead of B implies that TG is a (^-stopping time. • Finally, we will establish that Theorem 2.24 holds for all closed sets. Lemma 2.21 Let S : Cl —> [r, T] be a Ql+-stopping time, and let K £ /C. Then the random times D^, DSK and T% are Ql+-stopping times. Proof. Let K be a compact subset of E, and let Gn, n £ N, be a sequence of open subsets of E satisfying the following conditions: K C Gn+i C Gn and f]neNGn = K. Then, every random time DG is a C/^-stopping time (see Lemma 2.20), and for every fi £ P(E) the sequence of random times DGn, n £ N, increases PTj^-almost surely to Df,. This implies that the random time T ^ is a £/L.-stopping time. It is not hard to see that the equality (2.177) with K instead of B implies that T j | is a ^.-stopping time. Our next goal is to show that the sequence DG , n £ N, is PT)M-almost surely convergent. Put DK = sup„ DGn. Since DGn < DGn+1 < D%, we have DK < D^. By Lemma 2.20, the random times DG , n £ N, are ^ . - s t o p p i n g times. Therefore, the quasi left-continuity of the process Xt, t £[0,T], implies that lim XDs
= XDK
PT)A1-a.s.
Propagators:
General
X D K ef]Gn
=K
Theory
185
Moreover, PTiM-a.s.
n
It follows from DK > S that DK < DK PT,M-almost surely, and hence DK = DK P r ,^-almost surely. This proves that DQU —> DK PTiM-almost surely as n —» oo. In order to finish the proof of Lemma 2.21, we will establish that for every p, e P(E), the sequence of random times DQ^ increases P Ti ^-almost surely to DK. Put DK = sup„ Z ) ^ . It follows from D^n < Dan+1 < DK that DK < DK- Since the process Xt, t € [0, T], is quasi left-continuous, lim X^s
= X„
P r ,^-a.s.
Therefore XbKef]Gn=K
PT,M-a.s.
It follows from DK > S that DK < DK P r , M -almost surely, and hence DK = DK PT)M-almost surely. This equality shows that the random time DK is a QJ+ -stopping time. This completes the proof of Lemma 2.21. • Let us return to the study of standard processes (see Definition 2.22). It was established in Section 2.9 that if P is a transition probability function such that the backward free propagator Y associated with P is a strongly continuous backward Feller-Dynkin propagator, then there exists a standard Markov process (Xt,Ql,FT:X) with P as its transition function (see Theorem 2.22). By Theorem 2.24, for a standard process (Xt, GJ, P T , X ), the random variables D § , Z)§, and T§ are (^-stopping times. Since a standard process is always a strong Markov process in the sense of condition (3) in Definition 2.22, the stopping times in Examples 2.3 and 2.4 can be used in the formulation of the strong Markov property of the process Xt. Our next goal is to construct more examples of such families of stopping times. Let (Xt, Ql, PT,x) be a Markov process and let r S [0, T). Suppose that 5 is a (/^"-stopping time with r < S
186
Non-Autonomous Kato Classes and Feynman-Kac Propagators
Theorem 2.25 Let (Xt,GJ,FTiX) be a standard process, and let B £ £. Then the stopping times D%, D § , and T§ are measurable with respect to the a-algebra
The proof of Theorem 2.25 is based on the following two lemmas.
Lemma 2.22 Let K £ K, r £ [0,T), and suppose that Gn £ O, n £ N, is a sequence such that K C Gn+\ C Gn and f]n^Gn — K. Then the following assertions hold: (a) For every /i £ P(E), the sequence of stopping times DG increases and tends to DK FTtlJ,-almost surely. (b) For every t £ [r, T], the events {DGn
[T,T],
the events | 5 § n
n £ N, are
G^y-
measurable, and the event < D^
= XDK P r ,u-a.s.
G„
^
Therefore, XDK
ef)Gn
= K PTiM-a.s.
n
Since DK > S, we have DK < DK PT,M-a.s., and hence D^. = DK P T ,^-a.s. (b) Fix t £ [r, T) and n £ N. By the right-continuity of paths, we have
{DsGn
men ^ = 0
U
"iGNpg[ r ) t + -L)nQ+
meNpg[T,(+j-)
{SVp
XSyPeGn}.
(2.193)
Propagators:
General
Theory
187
It follows that
{DsGn
0
Next, using assertion (a), we see that the events {DK < i} and flnGN {^Gn — *} coincide P r , M -almost surely. It follows that {DK < i] e QT'V. This proves assertion (b). (c) Since the sets Gn are open and the process Xt is right-continuous, the hitting times T j and the entry times DG coincide. Hence, the first part of assertion (c) follows from assertion (b). In order to prove the second part of (c), we reason as follows. By assertion (b), for every r € Q + , the stopping time D% ' A is Q^ 'A ' v -measurable. Our next goal is to prove that for every e > 0,
4e+S)AT'v c $ v .
(2.194)
Fix e > 0, p € [T, T], and put Si = S and S2 = ((e+S)/\T)Vp. Observe that for t £ [0, T], the events {((e + S) A T) V p < t} and {S V (p - e) < t - e} coincide. Therefore, S2 is QTf -measurable. Since the process Xt is rightcontinuous, it follows from Theorem 2.17 that Xs2 is QT' -measurable. This implies inclusion (2.194). It follows that a
(f+5)AT,VcaS,V;
( 2 1 9 5 )
and we see that the stopping times D^ ' , e > 0, are C?T'v-measurable. Since the family D^ ' , e > 0, decreases to TK, the hitting time TK is C/T'v-measurable as well. (d) Fix p € P(E), and let K € K. and Gn € O, n e N, be as in assertion (a). Put DK = sup n £>c . Since DsGn
= XuK
Pr //-almost surely.
Gn
Therefore, Xjj
G I ) Gn = K PTiM-almost surely. n
Now DK > S implies that D^ < DK PT)M-almost surely, and hence DK = DK Pi-./j-almost surely.
188
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
(e) Fix t £ [T, T) and n £ N. By the right-continuity of paths,
{5i
U
{SVP
XsvP€Gn}.
(2.196)
eNp€(T,t+i)nQ+
It follows that < DQU < t> £ QT' . Next, using assertion (d), we see that the events I D^
and flneN j ^Gn — * f
comc
i d e PTiM-almost surely.
Therefore, <-Df- < t > £ Q^ • This proves assertion (e).
•
Let us return to the proof of Theorem 2.25. We will first prove that for any Borel set B, the entry time Dg is measurable with respect to the cr-algebra GT'V • The same assertion holds for the hitting time Tg. Indeed, if Dg is QT' -measurable for all stopping times S, then for every e > 0, the stopping time Dg ' is measurable with respect to the u-algebra g(e+5)AT,Vt B y (2.195), D<£+S)*T is ^ - m e a s u r a b l e . Now (2.177) implies the QT' -measurability of T j . Fix t £ [T,T), fi £ P(E), and B £ £. By Lemma 2.19, the set B is capacitable with respect to the capacity K \—> P Tj/i [D^ < t]. Therefore, there exists an increasing sequence Kn £ fC, n £ N, and a decreasing sequence Gn £ O, n £ N, such that Kn c Kn+1 C B C Gn+1 cGn,
n£ N,
and sup P r „ \Dl
< t] = inf P r
u
\D% < t].
Consider the following events: AT^S=\J{DsKn
and
A^'s
= f] {DsGn < t} .
(2.197)
These events are QT' -measurable. Moreover, we have Ar,n,S
c
i^S
< ^
and
Al*s
; inf P r „ \£>£ < t]
c
AT,„,S
^
m
j
Propagators: General Theory
189
s u P P T , M [ D | n < t ] = P r , M A[' M ' S < t
(2.199)
Now (2.198) and (2.199) give P r , M A£ M,S \ A[,M,S1 = 0 . It follows from (2.198) that the event {DB < t} is measurable with respect to the cr-algebra 0 T ' V . This establishes the T ' v -measurability of the entry time DB and the hitting time TB. The proof of Theorem 2.25 for the pseudo-hitting time DB is similar to that for the entry time DB. The proof of Theorem 2.25 is thus completed. • Corollary 2.2 Let (Xt,Ql,WTtX) be a standard process, and let A and B be Borel subsets of E with B C A. Then the entry time DB is measurable with respect to the a-algebra QTA' . Moreover, the hitting time TB is measurable with respect to the a-algebra QTA' . Proof.
By Theorem 2.25, it suffices to show that the equalities DDA
DB
and DBA = TTB
hold P r>/J -almost surely for all fi e P{E). follows from
(J {D\<s, XS&B}= T<S
(2.200)
The first equality in (2.200)
U
{XSGB},
T<S
while the second equality in (2.200) can be obtained from
U {TX<s,xseB}= T<S
U
{^eB}
T<S
This proves Corollary 2.2.
• T
Corollary 2.2 implies that the families {D A : A € £} and {TjJ : A € £} can be used in the definition of the strong Markov property in the case of standard processes. The next theorem states that the strong Markov property holds for entry times and hitting times of comparable Borel subsets. T h e o r e m 2.26 Let (Xt,GZ,FT,x) be a standard process, and fix r € [0,T]. Let A and B be Borel subsets of E such that B C A, and let f : [T,T] x E —> R be a bounded Borel function. Then the equalities E T ,x [/ (Dg,XDrB) | 0Lr ] = E ^ , * ^ [/ (D*B, X Dr)]
190
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
and ET
f(TTB,XTrB)
I g^] = E r ; , x T J
[f(TTB,XTh)}
hold FTtX-almost surely. Proof.
2.11
Theorem 2.26 follows from Corollary 2.2 and Remark 2.12.
•
N o t e s and Comments
(a) Propagators (evolution families) are two-parameter generalizations of semigroups. However, propagator theory is not yet as complete as semigroup theory. We refer the reader to [Pazy (1983); Goldstein (1985); Engel and Nagel (2000); Demuth and van Casteren (2000); van Casteren (2002)] for more information on semigroup theory. Various results concerning propagators can be found in [Pazy (1983); Nagel (1995); Nickel (1997); Nagel and Nikel (2002); Schnaubelt (2000/2001); Schnaubelt (2000); Liskevich, Vogt, and Voigt (2005)]. Under certain restrictions, propagators generate solutions to Cauchy problems for non-autonomous evolution equations. Such results go back to Sobolevskii (see [Sobolevskii (1961)]) and Tanabe (see [Tanabe (1960a); Tanabe (1960b); Tanabe (1997)]). Important discoveries in the theory of non-autonomous evolution equations were made by Acquistapace and Terreni (see the survey [Acquistapace (1993)] and the references therein). (b) Theorem 2.1 was first formulated without proof in [Gulisashvili (2004a)]. The proof can be found in [Gulisashvili (2004b); Gulisashvili (2004c)]. This theorem was also obtained independently but later in [Liskevich, Vogt, and Voigt (2005)]. (c) Howland semigroups were introduced in [Howland (1974)] (see [Chicone and Latushkin (1999)] for more information on Howland semigroups). (d) The Feller property and the Feller-Dynkin property are discussed in [Rogers and Williams (2000a); Rogers and Williams (2000b); Revuz and Yor (1991); Demuth and van Casteren (2000); van Casteren (2002)]. (e) Our presentation of the strong Markov property of non-homogeneous processes has certain similarities with that in [Dynkin (1994); Dynkin (2002)]. For more information on the strong Markov property of non-homogeneous stochastic processes see [Dynkin (1973); Kuznetsov (1982)].
Propagators:
General Theory
191
(f) Time-homogeneous standard processes are discussed in [Blumenthal and Getoor (1968)]. (g) For the proof of Choquet's capacitability theorem see [Bourbaki (1956); Dynkin (1965); Meyer (1966)].
This page is intentionally left blank
Chapter 3
Non-Autonomous Kato Classes of Measures 3.1
Additive and Multiplicative Functionals
In this section we study two-parameter additive and multiplicative functionals. We will first give examples of additive and multiplicative functionals of Markov processes. In these examples, the functionals are generated by a Borel function V on [0, T] x E. More complicated examples, where the functionals are associated with time-dependent measures, will be discussed in Sections 3.9 and 3.10. The additive and multiplicative functionals considered in the present section and in Sections 3.9 and 3.10 will be used in the definition of one of the main objects of our study in the present book, namely, the Feynman-Kac propagator. In a sense, the Feynman-Kac propagator is a perturbation of a free propagator by a multiplicative functional. Definition 3.1 A two-parameter family A = A (T, t), 0 < T < t < T, of random variables on a filtered probability space (fi, J7, ,F t r , P r ,x) with values in the extended real half-line M+ is called an additive functional provided that the following conditions hold: (1) For all r and t with 0 < r < t < T and all x £ E, the random variable A(T, t) is finite P r)X -almost surely and ^"-measurable. (2) For all r , A, and t with 0 < r < A < t < T, and all x e E, A(T, t) = A(T, A) + A(X, t)
PTia.-a.s.
(3) For all 0 < r < T and all x € E, A(T, T) = 0
P r , x -a.s.
Let P be a transition probability function, and let (JsTt,^7,P r , x ) be a Markov process associated with P. Then a typical example of an additive 193
194
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
functional of the process Xt is given by the following. Let V > 0 be a Borel function on [0, T) x E, and define a family of random variables by
Av(r,t) = l£VM)d8'
*/><'.*•>*«*>
10,
(3.1}
otherwise.
It will be shown in the sequel that under certain restrictions on the process Xt and the function V, the family Ay in (3.1) is an additive functional. This functional satisfies several additional conditions. For instance, Ay is non-decreasing and continuous. It will be assumed below that the process Xt is progressively measurable. This condition is needed to guarantee the measurability of the integrand in (3.1). Definition 3.2 A two-parameter family M = M (r, t), 0 < r < t < T, of random variables with values in the extended half-line M+ is called a multiplicative functional provided that (1) For all r and t with 0 < r < t < T and all x G E, the random variable M(T, t) is finite PTiX-almost surely and ^"-measurable. (2) For all r, A, and t with 0 < r < A < t < T and all x € E, M(T, t) = M(T, X)M(\, t)
P T , x -a.s.
(3) For all 0 < r < T and all x <E E, M(T, T) = 1 P Tl3r a.s. If A is an additive functional, then the functional M defined by M(r,t) — eA^T,t' is a multiplicative functional. An important example of a multiplicative functional is the Kac functional My where V is a Borel function on [0, T) x E. This functional is obtained by exponentiating the functional Ay defined in (3.1). More precisely, , Mv{T,t)
x
[exp{-[tV{8,X a)ds\, = l I JT V ' I 1,
;
J
if JTf' I V(s,X„) U s < oo ' v ' ' (3.2) otherwise.
The functional My defined in (3.2) will be used in the definition of the backward Feynman-Kac propagator Yy associated with the function V (see Section 4.2). In Section 3.9, we will study the functionals A^ and MM corresponding to a time-dependent measure (i = {/x(i) : 0 < t < T}.
Non-Autonomous
Kato Classes of Measures
195
Definition 3.3 A family \x = {n{t) : 0 < £ < T } of signed Borel measures of locally bounded variation on (E, £) is called a time-dependent measure provided t h a t for every set B G £ the function s H-> ^ ( S , B) is Borel measurable on [0, T ] . Recall t h a t if v is a signed measure, then at least one of the measures v+ and v~ is finite. In Section 3.9, we will study the f u n c t i o n a l in (3.1) and (3.2) in the case of a time-dependent measure /x. T h e following construction is used in Section 3.9 t o define t h e functional A^: t h e time-dependent measure fi is approximated in a special sense by a sequence of functions Vk on [0, T] x E so t h a t the sequence Ayk converges in a certain topology. T h e additive functional A^ is defined by
A„(T,t)
lim / = < k-^°cJT
Vk (s, Xs) ds,
if t h e limit exists, (3-3)
0,
otherwise.
In addition, the Kac functional M M is given by MM(r,t)=exp{-AM(r,£)}.
(3.4)
We conclude the present section by the following examples. Let 5 be a terminal stopping time (see Definition 2.10). Then A{T,t)
=
X{r<s
is an additive functional, while M(T,t)
=
l-X{r<S
is a multiplicative functional.
3.2
P o t e n t i a l s of T i m e - D e p e n d e n t A u t o n o m o u s K a t o Classes
Measures
and
Non-
Let P be a transition probability function, and let Y be the corresponding free backward propagator. If V is a Borel measurable function on [0, T] x E such t h a t T
/
Y(T,s)\V(s)\(x)ds
196
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
for all (r, x) e [0, T] x E, then the function N(V)(T,t,x)=
f Y(T,s)V(s)(x)ds
(3.5)
is defined for all 0 < r < t < T and x 6 E. The function N(V) is, in a sense, a potential of V. It is obtained by integrating the free propagator with respect to the time variable. Now let fj.be a time-dependent measure. In order to define the potential of the measure /x, we suppose that the transition function P has a density p with respect to the reference measure m. Let us denote by |/x(t)| the variation of the measure /x(t), and assume that the following condition holds: T
/ for all (T,X)
Y(T,s)\n(s)\(x)ds
G [0,T] X E.
gral / p(r,X;S,y)d\n(s)\.
(3.6)
In (3.6), Y(T,S)\H(S)\(X)
stands for the inte-
If inequality (3.6) holds for a time-dependent
JE
measure //, then the potential N(fi) of fi is defined by N(fi)(r,t,x)=
/ Y(T,s)n(s)(x)ds= JT
ds JT
p(r,x;s,y)dfj.(s)
(3.7)
JE
for all 0 < T < t < T and x € E. Next, we will introduce various classes of functions and measures. These classes are generalizations of the Kato class of potential functions (see Section 4.1). Note that in the case of classes of functions in Definition 3.4 we do not require the existence of a transition density p, while for the classes of time-dependent measures, the existence of the density p is assumed. In Definition 3.4, the subscripts / and m distinguish the classes of functions from the classes of time-dependent measures. Definition 3.4 Let P be a transition probability function. class V*j is defined as follows: V£V*f<=^
sup 0
swpN(\V\)(T,t,x)
x€E
Let V e V). Then the class V) is defined by VeK«=>
lim sup7V(|V|)(T,t,x) = 0. t—rJ.0x€B
Then the
Non-Autonomous Kato Classes of Measures
197
Suppose that the transition probability function P has a density p. Then the class V^ is defined as follows: H G V^
<=>
sup 0
sup N(\(I\)(T,
t, x) < oo.
x€E
If fi G Vm, then the class V^ is given by M G P ^ < = > lim sapN(\ii\)(r,t,x)
= 0.
We call the class V% (V^) the extended non-autonomous backward Kato class of functions (measures), while the class V% (P^) is called the nonautonomous backward Kato class of functions (measures). Remark 3.1 Suppose that the transition probability function P has a density p. Then the classes V% and V} can be identified with subclasses of the classes V^ and V^ as follows: V(T, X) <=*> dfi(T,x) — V(r,x)dx. For a function V e P J and a time-dependent measure /x G P ^ , denote \\V\\}=
sup
mpN(\V\)(T,t,x),
0
and llMllm=
sup 0
swpN(\(i\)(r,t,x). x€E
It is clear that ||V||}=
sup
SUPN(\V\)(T,T,X)
r:0
and IHIm=
sup T:0
supiV(|/z|)(T,T,:r). XGE
Next, we will show that the classes in Definition 3.4 are normed spaces. We will first make several remarks concerning the equivalence relations mod 0 for the elements of these classes. Let I denote the Lebesgue measure on the cr-algebra B[o,r] of all Borel subsets of [0,T]. For every r G [0, T] and i £ £ , define a measure £T,X on the
= 11
P(T,x;u,dy)du,
U G B[T>T] ®£.
198
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
It follows that for V G P), the condition \\V\\*f = 0 means that for all r G [0, T) and x £ E, the equality V(u, y) — 0 holds £T>x-a.e. on [r, T] x E. If P has a density p such that p(r, x; u, y) > 0
(3.8)
for all r, x, u, and y, then we get the following equivalent condition: V(u, y) = 0 I x m-a.e. on [0, T] x E. If there exists a density p, and if \x G 7 ^ , t n e n t n e condition | |A*| [{^ = 0 means that /
/ p(T,x;u,y)d\n(u)\(y)du
=0
for all T and x. If p satisfies (3.8), then we get the following equivalent condition: fi(u) = 0 for i-a.a. u G [0, T]. Taking into account the identifications described above, we see that the spaces [Pf, \\ • \\*A and f?^,, || • ||^J are normed spaces. Next, we will prove that they are Banach spaces. Lemma 3.1 Let P be a transition probability function. Then {P*., || • ||*.) is a Banach space, and (Pf, \\ • \\*j) is a closed subspace ofPf. Moreover, if P has a strictly positive density p, then (P^, \\ • ||^) is a Banach space, and Vm is a closed subspace ofP^. Proof. We will prove that if p is strictly positive, then the space P^ is complete, and P^ is a closed subspace of P^. The proof of Lemma 3.1 for the spaces PJ and P^ is similar, and we leave it as an exercise for the reader. Let fik G P^, k > 1, be a sequence of time-dependent measures such that y"llMfc|lm = y " k = l
k
SU
P
T:0
SU
P /
x£E
du
p{T,x;u,y)d\nk{u,y)\<
JT
oo. (3.9)
JE
Then for every x G E, / du I p(0,x;u,y)dy2\fik(u,y)\ Jo JE k=l
< oo.
Hence, there exists a Borel set Ux G [0,T] such that 1{UX) = T and r.
OO
p(0,x;u,y)d'^2\iJ,k(u,y)\
(3.10)
Non-Autonomous
Kato Classes of Measures
199
for all u G Ux. Fix x G E. Then (3.10) implies that for every j > 1 and
ueUx, oo
^2W{u)\(AjtU)
where Aj )U — {y E E :p(Q,x;u, y) > j ~ 1 } . Hence, ^ zzfc (u) is a finite signed Borel measure on every set AjtU for all u € Ux. Since the strict positivity oo
of p implies M AjtU = £7 for all u 6 (7X, the measure p,(u) — J^MfcC") i s j=i
a signed Borel measure on 1? for all u £ Ux, and hence Z-a.e. on [0,T]. It follows from (3.9) that /z G P ^ . Moreover, it is not difficult to prove using (3.9) that the series Y^kLi Mfc converges to /x in the space P^. This establishes the completeness of the space Vm. Now let /ifc G P „ , fc > 1, be such that //& —> /x in V^. Then / y(r,u)|/i(u)|(x)du< /
Y(T,U)\(I(U)
-
fik(u)\(x)du
+ f Y(T,u)\fik(u)\(x)du.
(3.11)
It follows from (3.11) that /z G V^. Hence, the class V^ is a closed subspace of the space P^. This completes the proof of Lemma 3.1. • The next result provides a description of the classes Vf and V^ in terms of the potential operator N. Note that a function V G Pj and a timedependent measure /J, G P^ are characterized by the uniform boundedness of the functions iV(|V|) and JV(|//|), respectively. For the classes Vj and 7-^, the following lemma holds. Lemma 3.2 (a) Let V G Vj.
Then V G V) if and only if
lira sup sup[N(\V\)(T,t',x)-N(\V\)(T,t,x)} t'-t-»o+ T-.o
= 0.
(3.12)
Then \i G V^ if and
200
Non-Autonomous
Proof.
Koto Classes and Feynman-Kac
Propagators
Let V GV). Then we have N(\V\)(T,t',x)-N(\V\)(r,t,x)
=J =
Y(T,u)\V(u)\(x)du Y(r,t)N(\V\)(t,t')(x),
and it follows that sup[N(\V\)(r, t', x) -
N(\V\)(T,
t,x)} < sup N(\V\)(t, t', x).
x€E
x€E
It is clear that the previous estimate implies (3.12). Now assume that (3.12) holds. Then we have lim supN(\V\)(T,t,x)=
lim sup[N(\V\)(T,t,x)-N(\V\)(T,T,X)}
*—r^°x€E
= 0.
t T
- -^°x€E
This implies V GV*f. The proof of part (b) of Lemma 3.2 is similar. Renicirk 3.2 Let V G Vf. 3.2, we see that lim
sup
•
Then, reasoning as in the proof of Lemma
sup\N(V)(T,t',x)-N(V){T,t,x)\
= 0.
t ' - t - > 0 + T : 0 < T < t x&E L e t V £ VJ.
T h e following f u n c t i o n w i l l b e u s e d i n S e c t i o n 3.9:
M(V)(r,t)=
sup
sup \N(V)(r,t,x)\,
r:r
0
x£E
Similarly for (i £ V^, we put M(n){T,t)=
sup
sup \N(n)(r,t, x)\,
r:r
3.3
0
x€E
Backward Transition Probability Functions and NonAutonomous Kato Classes of Functions and Measures
Let P(T, A; t, y) be a backward transition probability function (see Definition 1.3 in Chapter 1). Recall that the free propagator U associated with P is defined on the space L£° by U(t,T)g(y)=
f JE
U(t,t)g(y)
= g(y)
g(x)P(T,dx;t,y)
Non-Autonomous Kato Classes of Measures for all 0 < r < t < T, y £ E, and g £ Lf.
201
If P has a density p, then
U(t, r)g{y) = / g{x)p(r, x; t, y)dx JE
U(t,t)g(y)
= g(y)
for all 0 < r < t < T, y € E, and g £ L°°. Suppose that P is a backward transition probability function. If V is a Borel function on [0, T] x E such that /
U(t,s)\V(s)\(x)ds<
CO
for all 0 < r < t < T and x £ E, then we define the potential of the function V by N{V)(t,T,x)=
f
U(t,s)V(s)(x)ds.
Similarly, if P has a density p, then for a time-dependent measure /x satisfying J
U(t,s)\v(s)\(x)d.S
< CO
for all 0 < r < t < T and x £ E, we define the potential of the measure /x by N((i)(t,T,x)=
/
U(t,s)fi(s)(x)ds
for all 0 < r < t < T and x S £ . Definition 3.5 Let P be a backward transition probability function. Then the class Vf is defined as follows: V £pf
«=>
sup 0
sup N(\V\)(t, T, x) < co. xeE
Let V £ Pf. Then the class Vf is defined by V £Pf
=> lim sup N(\V\)(t, T, x) = 0. t-rj.0a.gjc;
202
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
Suppose that the backward transition probability function P has a density p. Then the class Vm is defined as follows: H € Vm <=$• sup 0
sup N(\fi\)(t, T, x) < oo. xeB
If li € Vm, then the class Vm is defined by / j e P m ^
lim sup JV(|/i|)(£,r,a;) = 0. t-r|0l€£
The class Vf {Vm) is called the extended non-autonomous Kato class of functions (measures). Similarly, the class Vf {Vm) is called the nonautonomous Kato class of functions (measures). For the classes in Definition 3.5, the norms are defined by the following: \\V\\f=
sup
supN(\V\)(t,r,x)
0
x€E
SUp 0
xeE
and |M|m=
SUpN(\fi\)(t,T,x).
It is clear that ||V||/=
sup T:0
supN(\V\)(T,T,x) x€E
and ||/x|| m =
sup T:0
supiV(|/i|)(T,T,x). X€E
As in the case of backward Kato classes of functions and measures in the previous section, we should take into account the equivalence modO for functions and measures. For instance, if P is a backward transition probability function, we define a measure £t,x by Zt,v(B)
=
P(t,dx;u,y)du
where B € B[ott] x £. Then for V €Vj the condition ||V||/ = 0 means that for all t € [0, T) and y £ E, we have V(u, x) = 0 £i,2/-a.e. on [0, t] x E. Now it is clear how to define the above-mentioned equivalence relation in the case of functions. The remaining cases are similar. Taking into account
Non-Autonomous Kato Classes of Measures
203
the identifications described above, we see that the spaces [Vf, || • ||/J and [Vm, || • ||m) are normed spaces. Moreover, they are Banach spaces. Lemma 3.3
Let P be a backward transition probability function. 5a
\Pf-> II' 11/) *
Then
Banach space, and (Vf, || • ||/) is a closed subspace ofVf.
Moreover, if P has a strictly positive density p, then [Vm,\\ • \\m) is a Banach space, and Vm is a closed subspace of Vm • The proof of Lemma 3.3 is similar to that of Lemma 3.1. The next lemma is analogous to Lemma 3.2. Lemma 3.4 (a) Let V GPf. Then V <EVf if and only if lim T-T'10
sup[JV(|V|)(*,T,x)-JV(|V|)(t,T / ,a;)] = 0.
sup
(3.13)
f.Q
(b) Suppose that P has a density p, and let \x £ Vm. Then fi £ Vm if and only if (3.13) holds with [i instead ofV. Remark 3.3
T
lim
Let V £ Vf. Then it is easy to see that sup
sup N(V)(t,T,x)-N(V)(t,T',x)
= 0.
—r' 1° t:0
3.4
Weighted Non-Autonomous Kato Classes
Let
/
Y{T,s)\V(s)\{x)cp(T,s)'lds
< oo
204
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
for all (T, X) € [0, T] x E, then the function NV{V){T,t,x)
=J
Y(r,s)V(s)(x)V(r,s^ds
(3.14)
is defined for 0 < T < t < T and x e E. Next assume that the transition function P has a density p. Let /it be a time-dependent measure, and suppose that the following condition holds: /
Y(T, s)\n(s)\(x)(p(T, S) 1ds < oo
for all (T, X) e [0, T] x E. Then we put NV{H)(T,
t,x)= J Y(T, S)VL{8)(X)
(3.15)
for all 0 < r < t < T and x € £ . Now we are ready to define weighted non-autonomous Kato classes. As in Sections 3.2 and 3.3, we use the subscript / in the case of classes of functions and the subscript m for classes of time-dependent measures. Definition 3.6 Let P be a transition probability function. class V% v is defined as follows: V £ V)
«=>
sup
sup NV(\V\){T,
Then the
t, x) < oo.
0
Let V € V}iip. Then the class V*fip is defined by V € V}iV
*=*
lim sup NV(\V\)(T, t~Tl0 xeE probability function
Suppose that the transition the class P^^ is defined as follows: M e Pm,*
<s
=^
SU
SU
P
t, x) = 0.
P has a density p. Then
P Nv>(\^\)(T^ t,x) < ° ° .
0
If n e Pmtip, then the class ^,v> *s g i v e n by M€^m, ¥ , < ; =^
lim
supNv>(\ij.\)(T,t,x)
t-rl0xeE
= 0.
Non-Autonomous Kato Classes of Measures
205
Now let P be a backward transition probability function, and V be a Borel function on [0, T) x E such that I U(t,s)\V{s)\{x)(p{s,t)-lds
for all 0 < r < t < T and x € £ . Then we put JNf„00(t,T,x)= /" ^(t,s)V(s)(a:Ma > t)- 1 ds. Similarly, if P has a density p, then for a time-dependent measure fi satisfying U(t,s)\n(s)\(x)(p(s,t)-1ds<-oo
I
for all 0 < r < i < T and a; € E, we put ^(^(t.r.a;) = / for all 0 <
T
f/(t,s)/i(s)(a;)
< i < T and x e £ .
Definition 3.7 Let P be a backward transition probability function. Then the class P/l¥> is defined as follows: ,v ^=>-
sup
supJV v (|V|)(£,T, x) < oo.
0
Let V £ P/l¥>- Then the class Pf,v is defined by VeVf,v<=>
lim s u p J \ L ( | V | ) ( t , T , i ) = 0 .
Suppose that the backward transition probability function P has a density p. Then the class Vm,ip is defined as follows: Vm,ip ^=>
sup 0
sup Nv(\n\)(t,T,x)
< oo.
x&B
If /i £ P m , then the class P miV , is defined by A* G Vm
206
Non-Autonomous Koto Classes and Feynman-Kac Propagators
For a function V G V}
and a time-dependent measure fj, G V^
put |MI},„=
sup 0
supNv(\V\)(T,t,x) x€E
and SU
IMIm, v =
P
supAT^d/iD^t.a;).
0
Similarly, for a function V G VflV> and a time-dependent measure \x £ Pmiip, we put ||V||/lV=
sup
supA^(|V|)(*,T,:r)
0
and ||Ml|m, V =
SUP 0
S\ipNv(\lJ,\){t,T,x). x€E
As in Sections 3.2 and 3.3, the symbol / stands for the Lebesgue measure on the cr-algebra B[O,T]- For every r G [0, T] and a; G E, we define a measure £T,X,VJ on the cr-algebra S[ T ,T] <8> £ as follows. For [/ G <8[T,T] ® £, £T,X I¥ >(^) = /
/
Pfaxiv^dyfyfau^du.
Then for V G 7>£v the condition \\V\\*f
(3.16)
for all r, a;, u, and y, then we can formulate an equivalent condition: V(u, y) =0 I x m-a.e. on [0, T] x E. If the density p exists and fi G P ^ , then the condition ||/u||^ i( . = 0 means that / / p(T,x;u,y)d\fi(u)\(y)(p(T,u)~1du =0 JT JE for all r and a:. If p satisfies condition (3.16), then we get the following equivalent condition: fi(u) = 0 for Z-a.a. u G [0, T]. If we take into account the identifications described above, then the spaces ( ^ / ^ l l ' l l / m ) a n d ym,
Non-Autonomous Kato Classes of Measures
Lemma
3.5
(PfiV, || • ll/,u,J
Let P be a transition ls a
207
probability function.
Banach space, and [P}^, || • 11/m)
ls a
Then
closed subspace
of the space V% . Moreover, if P has a strictly positive density p, then ('Pm,ip> II" llm.v) *s a Banach space, and V^^ space
is a closed subspace of the
P^.
Lemma 3.6
Let P be a backward transition probability function.
[P/iVi II ' ll/.v)
?5 a
of the space Pf,v.
Then
Banach space, and (Pf,v, || • \\f,tp) is a closed subspace Moreover, if P has a strictly positive density p, then
\Pm,yi || • ||m,¥>) is a Banach space, and Pm,
Examples of Functions and Measures in Non-Autonomous Kato Classes
In this section we give examples of functions and time-dependent measures from the classes V*j, V^, Vf, and Vm. Recall that E denotes a locally compact Hausdorff topological space satisfying the second axiom of countability. It is equipped with a metric p : E x E —> [0, oo). The symbol Br{x) stands for the open ball of radius r > 0 centered at x € E. It is denned by Br(x) = {y€E:p(x,y)
(3.17)
208
Non-Autonomous Kato Classes and Feynman-Kac Propagators
Let fi be a time-dependent measure, and assume that there exist a number S > 0 and a positive function C on [0, T] such that \fi(s)\(Br(x))
(3.18)
for all 0 < s < T, r > 0, and x £ E. Moreover, suppose that /•oo
/ Jo
\6d(-$(\))
(3.19)
and lim /
,7I(T,S)
C(s)'-
ds = 0.
(3.20)
Then the time-dependent measure /i belongs to the class V^. L e m m a 3.8 Let p be a backward transition probability density, and suppose that there exist positive Borel functions 71 and 72 onO
(2) The following estimate holds: p{s,x;t,y)
<7i(s,t)$(j2(s,t)P(x,y))
(3.21)
for allO <s
(3.22)
for all 0 < s < T, r > 0, and y € E. Moreover, suppose that
I
00
5, X°d ( - $ ( A ) ) < 00
(3.23)
and lim / t-TlOjT
v 7i(s.i) C{s)ds = 0.
'li{s,t)
Then the time-dependent measure JI belongs to the class V,
(3.24)
Non-Autonomous
Proof.
Koto Classes of Measures
209
We will only prove Lemma 3.7. The proof of Lemma 3.8 is similar.
It follows from (3.17) and (3.18) that / p(r, X; S, y)d \fj,(s)\(y) = / JB
\fi(s)| (y : p(r, x; s, y) > A) dX
JO /•*(0)7i(T,a)
< / Jo
\fi(s)\ {y : 7i(T, S)9
/•*(0)7I(T,S)
(72(T,
/
1
/r*(0)-yi(r,a) • * ( 0 ) 7 i (T,S)
1
s)p(x, y)) > A) d\
//
A \
/
A
\ \0«
=c(s) (% i(7?)5
(325)
^r ~ ^-
-
By (3.25), sup /
/ p(-r,a;;s,3/)d|/i(s)|(y)
x€EjT
JB
<
/•*(0)
1
A
$-i(r,)5dV
/"*
7i (V S)
C(3)-§-!-{ds.
Now we see that (3.19), (3.20), and (3.26) imply Lemma 3.7.
(3.26) •
The next two lemmas allow us to construct examples of time-dependent measures not belonging to the backward Kato class Vm associated with p, or to the forward Kato class Vm associated with p*. Lemma 3.9 Let p be a transition probability density, and suppose that there exist positive Borel functions 71 and 72 on 0 < T < s < T and a positive function $ on [0,00) satisfying the following conditions: (1) The function $ is bounded, strictly decreasing, and such that $(A) —> 0 as A —> 0 0 .
(2) The estimate p(r,x;s,y) holds for allO
> 71 (r, s)$ (72(T, s)p(x,y))
(3.27)
and y € E.
Let ix be a time-dependent measure, and assume that there exist a number 5 > 0, a positive Borel function u 1—• r(u) on [0, T] with values in the extended half-line K + , a positive Borel function u 1—> C(u) on [0,T], and
210
Non-Autonomous Kato Classes and Feynman-Kac Propagators
a family Ds, 0 < s < T, of Borel subsets of E such that the function (s, x) i-» XD3 (X) is Borel measurable and |/i(S)|(Sr(x))>C(S)r5 for allO<s
(3.28)
r(s), and x 6 Ds. Moreover, suppose that
i sup lim s u p<^ ssup u p / / XDAX)C(S) -Ti.0 | l t-HO I x iJTT
l
72 (r,s)r(s)
S . ' .ds 72 (r> s)
"|
\sd(-$(\))\ >0.
(3.29)
T/ien t/ie time-dependent measure \i does not belong to the class V^. Lemma 3.10 Letp be a backward transition probability density, and suppose that there exist positive Borel functions 71 and 72 onO
(3.30)
and y G E.
Let Jl be a time-dependent measure, and assume that there exist a number 5 > 0, a positive Borel function u \—> r{u) on [0,T] with values in R+, a positive Borel function u i-> C{u) on [0,T], and a family Ds, 0 < s < T, of Borel subsets of E such that the function (s, x) 1—> Xf> (x) is Borel measurable and \Jl(s)\(Br(y))>C(s)rs for allO<s
(3.31)
r(s), and y £ Ds. Moreover, suppose that
{
/"'
~
7i (s t)
SU
f
P / XD,(X)C(S)~S) ~l2(s,t)r(s)
Xsd{ (-5(A))
'
lds
I > 0.
(3.32)
Jo
Then the time-dependent measure /I does not belong to the class Vm.
Non-Autonomous
Koto Classes of Measures
211
Proof. We will only prove Lemma 3.9. The proof of Lemma 3.10 is similar. Let s e [0,T]. Then, assuming that x € Ds and using (3.27), we obtain / p(r, X; S, y)d \fx{s)\ (y) = / JE
|/x(s)| (y : p(r, x; s, y) > X) dX
JO /•*(0)7i(r,«)
> / Jo
\fi(s)\ {y : 7i(T, S ) $ (TJJ(T, a)/»(x, »)) > A) dX
= /
|M(*)| (J/ : P(x,y) < —p
Jo
V
r-*- 1 ( — ^ - ^ ) ) dA
72(r, s)
/•*(0)7i(r,«)
/
\1I(T,S)JJ 1
\
= 7i(T, S) / \n(s)\ y : p(a:,y) < — r®'1 (77) dr). (3.33) Jo V 72 (T, S) J Our next goal is to use estimate (3.28) in (3.33). However, the lower estimate in (3.28) holds only if 72(T,S)
-X1fa)< r(s),
or equivalently, T] > $ ( 7 2 ( 7 - , s ) r ( s ) ) .
Taking this into account, we see that (3.33) implies the estimate ,71^*)
/ p(r, x; s, y)d | M (s)| (y) > C(s)^f^-
r
S
/-*(0)
/
(r,)S dr, (3.34)
^
/$(72 (r,s)r(s)) 72V > J J*(72(T,s)r(a))
JE
for all x e Ds. Now it is clear that Lemma 3.9 follows from (3.29) and (3.34). • Assume that 72(r, s)r(s) > e\ for all T and s such that s — r < £2 where ei and £2 are strictly positive constants. Then (3.29) is equivalent to the following condition:
SU / X£».(^)C(g) V,T ' S ( r f s r limsup^sup t-rjO
'{ // l x JT
,7I(T,5)
72 ( > )
> 0
-
( 3 - 35 )
Similarly, (3.32) is equivalent to the condition limsup (sup f XD(x)C(a)^f^-ds\ > 0. (3.36) t-rlO I x JT 72 ( M ) J R e m a r k 3.4 Conditions (3.17), (3.21), (3.27), and (3.30) are modelled after the two-sided Gaussian estimates for the heat kernels. Conditions
212
Non-Autonomous Kato Classes and Feynman-Kac Propagators
(3.18), (3.22), (3.28), and (3.31) are often used in Geometric Measure Theory. In a sense, they mean that the measure is 5-dimensional. In the case of a function V defined on [0,T] x E, condition (3.18) becomes /
\V(s,y)\dy
JBr(x)
Example 3.1 Let E be d-dimensional Euclidean space R d equipped with the metric p(x, y) = \x — y\, where | • | stands for the Euclidean norm on R d . The reference measure m in this case is the Lebesgue measure md- Let Gt be the Gaussian density on Rd given by f Ixl2 Gt (x) = T exp < — U ; (27rf)i I 2* where t > 0 and x G Rd. The function Gt generates a transition probability density p and a backward transition probability density p as follows: 1
p(r, X; t, y) = p{r, X; t, y) = Gt-T{x - y). Next we choose 7i (r, s) = 7i (r, s) = — - j , (27r(s — r))2 72 ( T , S) = 72 ( T , S ) =
, y/S — T
and $(A) = $(A) = e x p | ~ | . Here we use the notation in Lemmas 3.7-3.10. It follows that conditions (3.19) and (3.23) hold for the functions $ and $. Indeed, r°°
/
\5d{-$(\))d\=
r°°
= f
/
~
\
\sd(-$(\))d\
A5+1exp|-y|rfA
Let us consider the following family of functions on [0, T] x Rd: 1 a talnPTe]^ **>,(> A***) = t ln'
^ 3 - 37 )
Non-Autonomous Kato Classes of Measures
213
where a > 0, 3 > 0, and 0 < v < d. Note that for all a > 0 and 3 > 0, there exists a number a(a, 3) > 0 such that i? In" — < a{a, 3)q In" —
(3.38)
if 0 < ti < *2 < T. Indeed, we have
5HT)-"-'°'-1?(<"'"T-<' /TTI
Therefore, if a > 3, then the function 1i-> t a In" — is increasing on [0,T].
If a < 3, then the maximum of this function is attained at
t = Texp <
>, and (3.38) easily follows. Property (3.38) will be used
in the estimates below. We have
JBr(x)
\y\V
Jy:\y\
for all_z G Rd and r > 0. Hence, if £ = d - v, r{u) = oo for 0 < u < T, DS=DS = {0} for all 0 < s < T, and C(s) = C(s) = cvA-
l
s^ln"
then conditions (3.18), (3.22), (3.28), and (3.31) hold. Next we will determine for what values of the parameters a, 3, and v the remaining conditions in Lemmas 3.7-3.10 also hold. Let us start with condition (3.20). By property (3.38), i-
f ni
\7I(T",S)
limsup / C(s)—r-
72(r> )
t - r l O JT
<
Cl
t-ria
(27r)2
f -
Cv,d ,. /"' lim sup
(2TT)*
CVJ
Jo
J0
T
V^iv 2 Tj I
ds
f
-ds =—^hmsup (2ir)i t-rio S
t-riO
/ (s-T)*S° Jr Jr
in0 &•
dr) + r ^ l n
0
(a, /?, i/, d) lim sup [' ^ . ej.o Jo •na+? ln p ^
^
(3.39)
It follows from (3.39) that if a + | < 1 and 0 < 3 < oo, then condition (3.20) holds. This condition also holds for a + | = 1 and 3 > 1.
214
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
Next, we turn our attention to condition (3.24). Reasoning as in the proof of (3.39) and using (3.38), we see that limsup / C m —r-.—-as = — ^ l i m s u p / t-Ti0
752(s,t)
JT
(2TT)*
(27T)*
t - r i O Jr
—
^^r-
(t - s)% S<* III0
^
»?*(«-»?)a In" g j
TniFJo
Cu,d ,. f€ dr) —r lim sup sup / — _ „ a 13 (2TT)* do t-e
=
(3.40)
Let us assume that 0 < v < 2. Then we have lim sup / eio io
— < 2 lim sup 5-=- / a „ ^(e-T))aln^ ealn/3^./o ejo
—V*
,1-a-*
c3(a,i/) limsup . . «|o ln p ^
(3.41)
Moreover, d V „ _ < „*,. *7 „—=lim sup /T — 2 2 lim sup e- * 2 T/ e a Ui ° 40 P A ryf (e - 77)" In" -e-», £- " « A ( " V) In" e-r, £ ^ " 5 V- '// "*
£ 1 —2a — *-
< Ci(a, t<) lim s u p — „ • . ej.0
lnP —
(3.42)
It follows from (3.40), (3.41), and (3.42) that condition (3.24) holds provided that a + I < 1 and 0 < / ? < o o , o r 0 < i / < 2 , a + | = l, and /3>0. Next, we will analyze the range of applicability of condition (3.35) in the case of the functions AQ)ig „. Reasoning as in the proof of (3.39), we get limsup (sup / XDS (s)C(fl) 7 ^ T ' S\ ds\ = lim sup f t-rlO I x JT 72( r - s ) J t-rjO JT c«,,d .. /"* ds - Jk .
(27r)i t-rid JT (S —T)*< cu,d ,. /"* ds QQ+ (27r)f uo Jo 7o ss +2 .f l n " ^ '
Cisfl^'^ds 72 (T) S )
Non-Autonomous
Koto Classes of Measures
215
Therefore, condition (3.35) holds provided that a + | > 1 and 0 < /3 < oo, ora + f = landO?
J
=limsup t t-rlO JT
C(s)^^-ds 72 (*, 0
-=-^p± r)2(t-r))alnt-v -
5- lim sup ' j- lim sup / (2TT)*
eio
> ^ 1 _limsup
7O
—z
a
^ ( e - ^ h r
*
f1 **.
_ 3
^
(3.43)
It follows from (3.43) that condition (3.36) holds provided that either v > 2, or 0 < v < 2, a + § > 1, and 0 < /3 < oo, or 0 < i> < 2, a + | = 1, and (5 = 0. Therefore, Lemmas 3.7-3.10 imply that (1) The function Aa^tI/ belongs to the class V% if and only if either a + | < 1 and 0 < / 3 < o o , o r a + f = l and /? > 1. (2) The function Aa,p,v belongs to the class Vf if and only if either a + f < 1 and 0 < / 3 < o o ; o r O < i / < 2 , a + | = l, and /3 > 0. The previous characterizations show that in the case of the Gaussian transition density the classes V*. and Vj do not coincide. Indeed, the function A i J x belongs to the class Vf, but does not belong to the class V*f. This result was first obtained in [Gulisashvili (2002a)]. E x a m p l e 3.2 This example is a continuation of Example 3.1. Here we consider a family A 7 (a,/3) of time-dependent measures defined by
where a > 0, (3 > 0, and 7 is a signed Radon measure of locally bounded variation on R d . Although the family of time-dependent measures in (3.44) is more general than the family of functions in (3.37), the methods used
216
Non-Autonomous Kato Classes and Feynman-Kac Propagators
in example (3.37) can also be employed in the case of the time-dependent measures defined by (3.44). Let us suppose that there exists a number 5 > 0 such that N (Br(x)) < Clrs
(3.45)
for all x 6 Rd and r > 0, where C\ is a positive constant. It is also assumed that there exist numbers ro > 0, c% > 0, and a point xo £ Kd such that N (Br(xo)) > c2rs
(3.46)
for all 0 < r < ro- It is not hard to see, reasoning as in Example 3.1, that the following assertions hold (1) The time-dependent measure A 7 (a,/3) belongs to the class V^ if and only if either a + ^ < 1 and 0 < /3 < co, or a + ^ = 1 and (3 > 1. (2) The time-dependent measure A 7 (a,/3) belongs to the class Vm if and only if either a + ^- < l a n d 0 < / 3 < oo, or 0 < d-8 < 2 , a + ^ = 1, and /? > 0. Let us suppose that j — 5Xo, where 5Xo is the Dirac measure concentrated at xo € K d . Then we can take S = 0 in conditions (3.45) and (3.46). It is not difficult to show using assertions (1) and (2), that for such a measure 7, A 7 (a,/3) € V^ or A 7 (a,/3) £ Vm only if d = 1. Moreover, A 7 (a, (3) € V^ if and only if either 0 < a < | and 0 < / 3 < o o , ox a — \ and /? > 1. Similarly, A 7 (a,/?) € P m if and only if either 0 < a < | and 0 < / 3 < o o , o r a = 5 and f3 > 0. In the case where d > 1, examples of time-dependent singular measures in non-homogeneous Kato classes can be constructed utilizing the measures satisfying conditions (3.45) and (3.46) with 0 < S < d. For instance, if 5 is an integer, we can take any <5-dimensional subspace of R d and restrict the <5-dimensional Hausdorff measure Hs to it. If <5 is fractional, then the Hausdorff measure H& restricted to the standard Cantor set of Hausdorff dimension 5 in the unit cube of Rd is an example of a measure 7 satisfying conditions (3.45) and (3.46). It follows that for such a measure 7 assertions (1) and (2) provide necessary and sufficient conditions for A 7 (a,/3) e V^ orA7(a,/3)G7?m.
Non-Autonomous
3.6
Koto Classes of Measures
217
Transition Probability Densities and Fundamental Solutions to Parabolic Equations in Non-Divergence Form
Parabolic initial and final value problems are rich sources of transition densities. For instance, under rather general conditions, a fundamental solution p to a second order parabolic conservative initial value problem is also a backward transition probability density. Similarly, if p is a fundamental solution to a second order parabolic conservative final value problem, then p is a transition probability density. Let us consider the following partial differential equations on ddimensional Euclidean space Rd: ^
+ Lu = 0
(3.47)
— -lm = 0.
(3.48)
and
In equations (3.47) and (3.48), the symbol L stands for the differential operator given by L
=E^.»)^J+i:Mr,x)^.
(3-49)
It is said that the operator L in formula (3.49) is in non-divergence form. Let us assume that the coefficients a y and bi are Borel measurable on [0, T] x R d , and that the uniform parabolicity condition holds for the second order part of the operator L. This means that there exist constants 71 > 0 and 72 > 0 such that for all (T, X) € [0, T] x Rd and all collections of real numbers Ai, • • • , A^, the following inequalities hold: d
^Z
A
d
d
* ^ E OijfoaOAiAj < 72 £ \l
(3.50)
Inequalities (3.50) imply the boundedness of the coefficients a^. In the present section, we will study final value problems of the following form: du — + Lu = 0, 0
(3.51)
218
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
where / is a function on Rd. The function / in (3.51) is called the final condition. We will also consider initial value problems given by §-Lu U{T)
= 0,0
(352)
= g.
The function g in (3.52) is called the initial condition. Definition 3.8 Suppose that L is a differential operator given by (3.49). A function p(r,x;t,y) where 0 < r < t < T, x £ Rd, and y £ Rd, is called a fundamental solution to equation (3.47) provided that the following conditions hold: (1) For all (t,y) £ (0,T] x Rd, the function (r,x) H-> p(r,x;t,y) is continuously differentiable in T on the interval (0, t) and twice continuously differentiable in x on Rd. Moreover, p satisfies equation (3.47). (2) For any bounded and continuous function / on R d , f(x) = lim /
f(y)p(r, x; t, y)dy
uniformly on compact subsets of Rd. Definition 3.9 Suppose that L is a differential operator given by (3.49). A function p(r,x;t,y) where 0 < r < t < T, x £ M.d, and y € Rd, is called a fundamental solution to equation (3.48) provided that the following conditions hold: (1) For all (T, X) £ [0,T) x R d , the function (t,y) H-» p(r,x;t,y) is continuously differentiable in t on the interval (r, T) and twice continuously differentiable in y on Rd. Moreover, p satisfies equation (3.48). (2) For any bounded and continuous function g on Rd, g(y) = lim /
g(x)p{r, x; t, y)dx
uniformly on compact subsets of Rd. The next result is known (see the references in Section 3.12). It concerns the existence of fundamental solutions and the Gaussian estimates. Theorem 3.1
Let L be as in formula (3.49), and assume that
(a) The functions aij and bi are bounded and measurable on [0,T] x Rd, and the matrix (aij) is symmetric, i.e. aij(t,y) — aji(t,y).
Non-Autonomous
Kato Classes of Measures
219
(b) The uniform parabolicity condition (3.50) holds. (c) There exists a constant 5 with 0 < 6 < 1 such that d
d
^T \Oij (n,Xi)
- dij (r 2 , X2)\ + Y^ \bi (rl> ^l)
i,j = l
_ b
i ( r 2, a;2)|
i=l
< C[\xi
for all ( n , x\),
-x2\s + \n
-T2\SJ
(r2, x2) € [0, T] x R d (the Holder continuity
condition).
Then there exists a fundamental solution p(T,x;t,y) to equation (3.48). The function p is jointly continuous, strictly positive, and the following estimates hold:
^;^)<^^exp{-^},
(3.53)
(the upper Gaussian estimate for p); p(r,X; t,y) > —
]
-7exp{
-—^-}
M(t-T)?
(3.54)
t-T
(the lower Gaussian estimate for p); dp(r,x;t,y) dt
-
( i
M _T)f+i
e X P
f \x - y\2 } \ a(t-r)j
(the upper Gaussian estimate for the time-derivative dp(T,x;t,y) dyi
M M eXP
(t - T) -7^rW
f (
(3 55)
-
ofp);
2 \x -\x-y\ y\2 }
V^T^)f
(3.56)
for all 1 < i < d (the upper Gaussian estimate for the first order spacederivatives ofp); and d2p(r,X; t,y) dVidyj
M -(i_T)f+i
<
-*{-w%} (357) _ (
\x - y\
for all 1 < i < d and 1 < j < d (the upper Gaussian estimate for the second order space-derivatives ofp). In estimates (3.53)-(3.57), M and a are positive constants. Remark 3.5
The fundamental solution p in Theorem 3.1 is unique.
220
Non-Autonomous Kato Classes and Feynman-Kac Propagators
Let p be the fundamental solution to equation (3.48) in Theorem 3.1, and define a function p by P(T,
X; t, y) = p(T -t,y;T-r,x).
(3.58)
Then ^I^lM
=
-D3p(T-t,y;T-T,x)
=
-mT-t,y;T-T,x)
= ~Lp(r, x; t, y) where D3 denotes the partial derivative with respect to the third variable, and the operator L acts on the variable x. It follows that the function p given by (3.58) is the unique fundamental solution to equation (3.47). Moreover, formulas (3.58), (3.53), and (3.54) imply the validity of the twosided Gaussian estimates p(r, x;t,y)<
M m t
(t~)
^ exp <j - ^
^
\
(3.59)
and P(r, x; t, y) > — -7 exp{ M(t-r)?
'- ^ } t-r
(3.60)
for the function p. The Gaussian estimates (3.55)-(3.57) for the derivatives of p are also satisfied. The following assertion holds: T h e o r e m 3.2 Let L be the operator given by (3.49), and suppose that L satisfies conditions (a)-(c) in Theorem 3.1. Then the unique fundamental solution p to equation (3.47) is a transition probability density, while the unique fundamental solution p to equation (3.48) is a backward transition probability density. Proof. We will prove the first part of Theorem 3.2. The proof of the second part is similar. It is known that for every final condition / € BC, where BC is the space of bounded continuous functions on M.d, the final value problem in (3.51) has a unique solution given by U(T,X)=
f(y)p(T,x;t,y)dy. Jw.d
(3.61)
Non-Autonomous
Kato Classes of Measures
221
We will first prove that the function p satisfies the condition p(T,x;t,y)dy=l
/
(3.62)
for all 0 < T < t < T and x € M.d. Fix 0 < t < T, and consider the following functions: (T, X) I-> 1 and (T, X) H-> J"Rd p (r, x; t, j/) dy. It is not hard to see that these functions are solutions to final value problem (3.51). In order to prove that the second function is a solution to final value problem (3.51), we differentiate under the integral sign. This is possible, since the upper Gaussian estimates (3.55), (3.56), and (3.57) hold. Now the uniqueness result implies that (3.62) is true. Next, we will establish the Chapman-Kolmogorov equation for p\ that is, the equation P (r, x; v, y) = /
p (T, X;
S, Z)
p (s, z; v, y) dz, xeRd,
0
and y e Rd.
(3.63) d
Consider the following two functions on the open set (0, s) x R : (T,x)y-*p(T,x\v,y)
and
(T,x)t->/
p(r,x;
s,z)p(s,z;v,y)dz.
These functions are solutions to the final value problem in (3.51) with t = s and f(x) — p(s,x;v,y). Here we use the upper Gaussian bounds for the function p in order to justify the differentiation under the integral sign. By the uniqueness of solutions to final value problem (3.51), the ChapmanKolmogorov equation holds. • Let Vm be the non-autonomous Kato class of time-dependent measures associated with p, and Vm be the non-autonomous Kato class associated with p. Next, we will continue our discussion of the properties of the family of time-dependent measures defined by
^a>®
=
(3-64)
^F£i
where a > 0, /3 > 0, and 7 is a signed Borel measure of locally bounded variation on Md. We assume that there exists a number 6 > 0 such that | 7 | (Br(x)) <
5
Clr
for all x £ Rd and r > 0, where c\ is a positive constant. Moreover, we also assume that there exist numbers ro > 0, C2 > 0, and a point XQ e Rd such
222
Non-Autonomous Kato Classes and Feynman-Kac Propagators
that |7l (Br(xo)) > c2rs for all 0 < r < TQ. It follows from the two-sided Gaussian estimates for p and p and from the results obtained in Examples (3.1) and (3.2) that A
R
A 7 (a,/?)eP£ l «<=>.a + - = - < l , or a H
0?
— = 1, p > 1,
and A 7 (a, /3) G P m «=K* + —^— < 1, 0 < p < oo, or 0
3.7
< 2, aH
— = 1, /3 > 0.
Transition Probability Densities and Fundamental Solutions to Parabolic Equations in Divergence Form
In this section we discuss equations (3.47) and (3.48) with L given in divergence form. This means that the operator L can be represented by the following formula: d
d
L
d
d
= E a^ ^ f o ^ f l d + 5 > ( T ' X W 7.1 =
1
L
J
J
7=
(3 65)
-
1
The fundamental solutions to equations (3.47) and (3.48) with L in divergence form are defined in the weak sense. Let 0 < a < b < T, and let CQ° ((a, b) x R d ) denote the space of infinitely differentiable functions with compact support contained in the set (a, b) x Rd. We denote by C£° (R d ) the space of all functions / e C£° (R d ) such that / is infinitely differentiable on Rd and all partial derivatives of / are bounded on M.d. The symbol C ([0, t]\ L2 (R d )) will stand for the space of all L?-valued continuous functions on the interval [0, t], and H1 (Rd) will denote the Sobolev space of all functions g £ L2 (Rd) such that / e l
2
(R d ) for all 1 < i < d. Here the
partial derivatives are understood in the sense of distributions. The space of all locally square integrable functions on R d is denoted by L2oc (Rd), and Hioc (Rd) denotes the local Sobolev space, consisting of all functions from the space L2oc (Rd) with all first order generalized partial derivatives in the
Non-Autonomous
Kato Classes of Measures
223
space Lfoc (Rd). The symbol L2 ((0,t) ;H1 (Rd)) stands for the space of all strongly measurable functions u on the interval (0, t) with values in the space H1 (Rd) for which / ll u ( s )lljji^ s < °°Jo We denote by L2oc ((0,t); Hloc (Rd)) the space of all strongly measurable functions u on the interval (0, t) with values in the space H\oc (Rd) such that r-t-e
/
ds /I
I , ,
2, - ,n , . , 92\}dx . (\u(s,x)\ +\Vu(s,x)\
< CO
for every e with 0 < e < t and every compact subset C of Rd. We will also need the space Lq ((0, T); Lp (B(0, r))) where 9 > 1, p > 1, and B(0, r) stands for the ball of radius r in Rd centered at the origin. This space consists of all strongly measurable functions u defined on the interval (0, T) with values in the space IP (5(0, r)), and such that /
\\u(s)\\lp(B(0,r))ds
<°°-
Now we are ready to introduce the notion of a weak solution. Definition 3.10 Suppose that L is a differential operator given by (3.65). Let 0 < t < T and / £ Lfoc (Rd). A function U(T, X) where 0 < r < t < T and x G Rd, is called a weak solution to final value problem (3.51) provided that the following conditions hold: (1) The function u belongs to the space L2oc ((0, t); Hloc (2) The function d
/-.
(T,X) ^ ^ 6 i ( r , x ) —OU- ( r , a ; ) i=l
belongs to the space L\oc ((0,i) x Rd). (3) For all test functions tp € C%° ((0, t) x
224
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
(4) The following equality holds: lim /
u(r,x)rp(x)dx
= /
f(x)ip(x)dx
for all L2-functions ip on Rd with compact support. Definition 3.11 Suppose that L is a differential operator given by (3.65). Let 0 < r < T and g € L2oc (Rd). A function u(t, y) where T
((T,T)
; H\oc
belongs to the space L^ c ((T,T) X Rd). (3) For all test functions
u(f, y)ip(y)dy = /
g(y)ip(y)dy
for all L 2 functions ip on R d with compact support. Definition 3.12 Suppose that L is a differential operator given by (3.65). A Borel function p(r,x;t,y) where 0 < T < t < T, x e Rd, and y 6 Md, is called a weak fundamental solution to equation (3.47) if for every t with 0 < t < T and every bounded measurable function / with compact support, the function u defined by U(T,X)=
I f(y)p{T,x;t,y)dy
(3.66)
is a bounded weak solution to final value problem (3.51). Definition 3.13 Suppose that L is a differential operator given by (3.65). A Borel function p(r, x; t, y) where 0
Non-Autonomous Kato Classes of Measures
225
0 < r < T and every bounded measurable function g of compact support, the function u defined by «(*. V) = /
9(X)P(T, X;
t, y)dx
(3.67)
is a bounded weak solution to initial value problem (3.52). Definitions 3.10, 3.11, 3.12, and 3.13 can be found in [Porper and Eidel'man (1984)]. In [Liskevich and Semenov (2000)], the authors studied weak solutions and weak fundamental solutions in the global case. The definitions used in [Liskevich and Semenov (2000)] are as follows: Definition 3.14 Suppose that L is a differential operator given by (3.65). Let 0 < t < T and / e L1 (Rd) n L°° (R d ). A function u(r,x) where 0 < T < t < T and x £ E d , is called a weak solution to final value problem (3.51) provided that the following conditions hold: (1) The function u belongs to the space C ([0, t];L2 (Md)) n L2 ((0, t); H1
(Rd)).
(2) The function p.
d
(T,X) I-*
^2bi(T,x)~(T,x), i=\
*
belongs to the space L1 ((0,t) x Rd). (3) F o r a l l ^ G C ^ ° ( ( 0 , i ) x R r f ) ,
LL
[u^+^d^dvr^1biipdVl)dxdT=0-
(4) u(t, x) = f{x) m-almost everywhere. By condition (1), the function x \—> u(t, x) belongs to the space L2 (Rd). Therefore, the equality in condition (4) makes sense. Definition 3.15 Suppose that L is a differential operator given by (3.65). A Borel function p(r, x; t, y) where 0 < r < t < T, x € Rd, and y € Rd, is called a weak fundamental solution to equation (3.48) provided that the following conditions hold:
226
N'on-Autonomous
Kato Classes and Feynman-Kac
Propagators
(1) For every 0 < r < T and g e L1 (Rd) C\L°° (Rd), the function u defined by (3.67) belongs to the space C
([T,
T];L2 (Rd)) n L2
((T,
T); JT1 (R d )) .
(2) The function
a2A
»=i
belongs to the space L1
((T,T)
X
Rd).
(3) For all
Jr h*y dt . ^
J
a^% ^
dyij
(4) u(r, y) = g(y) m-almost everywhere. Definition 3.16 Suppose that L is a differential operator given by (3.65). A Borel function p(r,x;t,y) where 0 < r < t < T, x G Rd, and y € Md, is called a weak fundamental solution to equation (3.47) if for every t with 0 < t < T and / e L1 (Rd) n L°° (R d ), the function u defined by U(T,X)=
f(y)p(T,x;t,y)dy
(3.68)
is a weak solution to final value problem (3.51). Definition 3.17 Suppose that L is a differential operator given by (3.65). A Borel function p(r, x; t, y) where 0 < r < t < T, x £ Rd, and y £ Rd, is called a weak fundamental solution to equation (3.48) if for every r with 0 < r < T and every g € L1 (Rd) n L°° (R d ), the function u defined by w(*. V) =
9{x)p{r, x; t, y)dx
(3.69)
is a weak solution to initial value problem (3.52). The following result concerns the existence and uniqueness of weak fundamental solutions (see [Porper and Eidel'man (1984)], Theorem 6.1). Theorem 3.3 Let L be given by (3.65), and suppose that the following assumptions hold: (1) The coefficients aij satisfy the uniform parabolicity condition (3.50).
Non-Autonomous
Kato Classes of Measures
227
(2) For some ti0 € (0,1), r 0 > 0, s > — V , and 0t = ]- - £- - - > % , 1—171
2
2s
q
2
t/ie coefficients {bi} belong to the space Lq ((0, T); L s (B (0, ro))). f3j For almost all x eRd with \x\ > r0 and all t G [0,T], the coefficients {bi} satisfy |6j(t,a;)| < M, 1 < i < d, where M is a positive constant. Then there exists the unique weak fundamental solution p to the equation du
„
— — Lu = 0 in the sense of Definition 3.13 such that the two-sided Gaussian estimates (3.53) and (3.54) hold for p with the constants M and a depending only on T, d, the uniform parabolicity constants ji and 72, and the constants q, s, ro, M, and i?oFundamental solutions in Theorem 3.3 are unique almost everywhere with respect to the measure m<j x m j . A theorem, similar to Theorem 3.3, holds for weak fundamental solutions in the case of equation (3.48). Here we take into account formula (3.58). It is also true that bounded solutions to initial value problem (3.52) are unique. Next we formulate a recent result from [Liskevich and Semenov (2000)] concerning the existence of the unique weak fundamental solution p to the du _ _ equation — Lu = 0 and the two-sided Gaussian estimates for p under at more general conditions on the coefficients {6j} than those in Theorem 3.3. The existence theorem in [Liskevich and Semenov (2000)] was obtained in the case where the coefficients Borel measurable on [0, T] x Rd and satisfy the uniform parabolicity condition (3.50), while the restrictions on the coefficients bi are expressed in terms of the function W(t,x)=f^
i,j=l
(afrx)+«'{t'x)) V
Z
' (t,x)bi(t,x)bj(t,x), '
(3.70)
ij
where a*(t,x) denotes the transpose of the matrix a(t,x). More precisely, it is assumed that the function w belongs to the intersection of the extended weighted non-autonomous Kato classes Vf>ip and V*, . These classes are associated with a scaled Gaussian transition density 1
P\(T,x;t,y)
= —
— exp \~ 2\(t - T) j '
(27rA(i-r))?
^ 3 ' 71 ^
and the function tp is such that
228
Non-Autonomous Kato Classes and Feynman-Kac Propagators
and satisfies the conditions ip(Q) = 0, / Jo
Z-^ds
[(p'(s)],ds < oo, and / —-— < oo. Jo
< oo, / Jo
Theorem 3.4 Let L be given by (3.65), and suppose that the following assumptions hold: (1) The coefficients aij satisfy condition (3.50). (2) There exist a constant A > 0 and a function y> such that the conditions in (3.72) are satisfied and w£VftipnV*ftV.
(3.73)
In (3.73), the function w is defined by formula (3.70), and the classes Vj,v and Vjv are associated with the function
229
N'on-Autonomous Kato Classes of Measures
conditions (2) and (3) in Theorem 3.3 hold for the coefficients {6j}. Then there exist a number A > 0 and a function ip such that the conditions in (3.72) are satisfied and condition (2) in Theorem 3-4 holds. „ , _,. . . a(t, x) + a*(t,x) , „ Proof, bince the matrix is symmetric and satisfies the uniform parabolicity condition with the same constants 71 and 72 as the matrix a(t, x), its inverse matrix I -^———
!
— J
is bounded by 7J"1
uniformly with respect to t and x. It follows that a(t,x) + a*(t,x)
-v
2
1 < — 7i
for all 1 < i < d, 1 < j < d, t e [0,T], and x £ Rd. By formula (3.70), d
\w{t,x)\<
MdY^^S.x). i=l
It suffices to prove that there exist a number A > 0 and a function
l
(3.74)
Take A = 1 and ip(r, t) = (t — T) C , where e > 0 is a small number that will be chosen later. By Definition 3.6, we see that in order to establish (3.74), it suffices to show that SU
P
SU
fT dv f ( P / ~, ^TT / b2i(v,y)exp\-]
\x — v\2 1 y Wdy
(3.75)
and /** sup sup /
dv f ( \x — y\21 2 —-j / b (v,y)exp\\ dy < 00
(3.76)
for all 1 < i < d. We will prove only inequality (3.75). Inequality (3.76) can be obtained similarly. It follows from condition (2) in Theorem 3.3 that q > 2 and s > 2. It is also clear from (2) and (3) in Theorem 3.3 that every function b2 with 1 < i < d can be represented as the sum b2 = rji + Q of two functions such that ^ e L* ((0,T); L* (M.d)) and & € L°° ([0,T] x Rd). Now, using
230
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
Holder's inequality twice, we get
^^lViiV'y)eXP{-^)}dy
/
m y)Uy
^ fT
2
v^AL ^ Y
{i^i^is^if-r)}^} ' 2
c
^>
7 JT
^ T \
i
Vi(v,y) dy
(V — T) e i "» {JR*
^IMlL§((0,T);Li (*>){/ ( ^ T F J where A = [ e-\— I
V
(3 ?7)
-
. By condition (2) in Theorem 3.3, we have
sJq-2
> 0. Therefore, —; r < 1, and for small values of e we have 2 2s q ' s{q - 2) A < 1. It follows that the last integral in (3.77) is finite. Hence, inequality (3.75) holds with r}i instead of b\. We also have fT dv f / \x-y\2\ , JT
(V -TY+2
JW
I
2{V~T))
( 3 - 78 )
•
The next theorem is similar to Theorem 3.2. Theorem 3.5 Let L be the operator given by (3.65), and suppose that the conditions in Theorem 3.4 hold. Then the unique weak fundamental solution p to equation (3-47) is a transition probability density, while the unique weak fundamental solution p to equation (3.48) is a backward transition probability density. Proof. The normality condition can be established using the results obtained in [Liskevich and Semenov (2000)]. We have already mentioned that in the proof of Theorem 1 in [Liskevich and Semenov (2000)], the unique fundamental solution p is approximated in the space L}oc (Rd x Rd, dxdy)
Non-Autonomous Kato Classes of Measures
231
by a double sequence rn>m(T, x; t, y) of classical fundamental solutions. The limit in this construction is uniform with respect to the variables r and t with 0 < T < t < T. This was established in the remark after the proof of Lemma 4.2 in [Liskevich and Semenov (2000)]. Moreover, the upper Gaussian estimate holds for rn>m uniformly with respect to n and m. By Theorem 3.2, J rntTn(T,x;t,y)dy = 1. Therefore, the dominated convergence theorem implies that / p(r, x; t, y)dy — 1. Our next goal is to prove that the Chapman-Kolmogorov equation holds for p. Let 0 < T < s < v < T and ip e Ll (Rd) n L°° (Rd). Then the function ip(y)p(T,x;v,y)dy
UI(T,X)=
is a weak solution to the final value problem (3.51) with t = v and with the final condition u\{v) = tp. It follows from the properties of weak solutions that the function u\ is a weak solution to final value problem (3.51) with t = s and the final condition given by ui(s,x)=
ip(y)p(s,x;v,y)dy. JUL*
Moreover, the function U2(T,X)=
Jm.d
=
i>(y)dy
p(T,x;s,z)p(s,z;v,y)dz
JRd
p{r,x;s,z)dz J'SLd
ip(y)p{s,z;v,y)dy JRd
is a weak solution to final value problem (3.51) with t = s and the final condition given by u2(s,x)=
ij)(y)p(s,x;v,y)dy. JRd
By the upper Gaussian estimate, u2(s) e L1 (R d ) n L°° (Rd). The uniqueness of weak solutions established by Liskevich and Semenov yields that for all r with 0 < r < s, the equality ui(r, x) = U2(T,X) holds m-almost everywhere on Rd. This implies that for all i e i 1 (Rd) n L°° (Rd) and 0 < r <s
= / ip(y)dy / JR
p(T,x;s,z)p(s,z;v,y)dz.
JR
Therefore, p satisfies the Chapman-Kolmogorov equation.
232
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
This completes the proof of Theorem 3.5.
•
Let A 7 (a,/3) be the family of functions denned in Example 3.2. Since the two-sided Gaussian estimates hold for all transition probability densities discussed in this section, assertions (1) and (2) in Example 3.2 hold for these densities. Similar results are true for the classes Vmtp and Vm)ip where the function tp is defined by ip(r, t) = (t — r)a with 0 < a < 1. Let us assume that the class Vmip is associated with one of the transition probability functions P discussed in this section, while the class Vm,
3.8
Diffusion Processes and Stochastic Differential Equations
This section is devoted to stochastic processes generated by second order partial differential operators. Such processes are called diffusions (see Definition 3.18 below). Let P be a transition probability function on R d , and consider the following second order partial differential operator with timedependent coefficients:
i,j = l
J
i=l
where r £ [0, T) and x G Rd. We use the subscript T in (3.79) to emphasize the time-dependence. It is assumed that the following conditions hold: (1) The coefficients a j j and bi are Borel measurable and locally bounded functions on [0,T] xRd. (2) The matrix field a,ij{T, x) is symmetric; that is, aitj(r,x) = ajti{r,x).
Non-Autonomous
(3) The matrix field aitj(r,x)
Koto Classes of Measures
233
is nonnegative; that is, d
for all (r, x) G [0, T] x R d and all £ = (£i,...,£d) G R d . In the next definition, we introduce non-homogeneous diffusion processes. Definition 3.18 A Markov process (Xt, T[', PT,x) with state space Rd is called a diffusion process with generator LT provided that (1) The process Xt has continuous paths. (2) For any function / G CQ°, the equality Er,*[/(^t)]=/(l)+Er,x
J Lsf(Xs)ds
(3.80)
holds for all 0 < r < t < T and x G : The matrix field a^- in Definition 3.18 is called the covariance or the diffusion coefficient of the process Xt, while the vector field bi is called the drift oiXt. The next assertion provides an equivalent martingale characterization of a diffusion process. Lemma 3.12
The following are equivalent:
(1) Condition (2) in Definition 3.18 holds for the process Xt. (2) For any r G [0, T], x G Rd, and f G C£°, the process
MTtJ = f {Xt) - f (XT) - J Lsf (Xs) ds, t G [T,T],
(3.81)
is a Tl -martingale with respect to the measure ¥TtX. Proof. Suppose that condition (2) in Lemma 3.12 holds. Then the expression Er<x [M£>f] does not depend on u G [r, T]. By substituting u = t and u — T into this expression, we see that condition (1) in Lemma 3.12 holds. Conversely, suppose that condition (1) holds and let r < t\ < ti < T. By the Markov property, we get ET
[M£' I ^] - M £ ' = ET,X [ M £ ' - Ml? \ Tl]
234
Non-Autonomous
= E.TtX
Koto Classes and Feynman-Kac
\f (Xt2) -f{Xtl)-
'ti,Xti
Propagators
T LJ (Xs) ds | n,
f(Xt2)-f(Xt
Lsf(Xs)ds
(3.82)
By substituting T = t\, x = Xtl, and t = ti into (3.80), we see that (3.82) yields condition (2) in Lemma 3.12. This completes the proof of Lemma 3.12. D Let Y be the free backward propagator on the space Lf defined by Y(T,t)f(x)= [ f(y)P(T,x;t,dy). It is clear that formula (3.80) in Definition 3.18 is equivalent to the following formula: Y(T, t)f(x) = f(x) + J Y(T, s)LJ(x)ds
(3.83)
for all 0 < T < t < T, x G Rd, and / € C£°. In Section 2.3, we defined the right a (Lf, M)-generators A+(T), 0 < r < T, of the backward propagator Y. They are given by the formula A^(T)f = a(Lf,M)-Um
Y(T,T +
h)f-f
(3.84)
where / belongs to the subspace D (A+(T)) of the space Lf', consisting of all functions for which the limit on the right-hand side of formula (3.84) exists. The next lemma states that under certain restrictions, the generator of the diffusion process Xt coincides with the right generator of the backward propagator Y. L e m m a 3.13 Suppose that the free backward propagator Y in (3.83) is a strongly continuous backward Feller-Dynkin propagator, and let Xt be a diffusion process with generator LT. Suppose also that the coefficients aij and bi in (3.79) are continuous functions. Then LTf(x) = A+{r)f{x) for all T £ [0, T), x G Rd, and f G Cg. Proof.
It follows from equality (3.83) with t = r + h that Y(r,r + h)f(x)-f(x)
i rr+h rT+n = -J Y(r,s)Lsf(x)ds
(3.85)
Non-Autonomous
Kato Classes of Measures
235
Passing to the limit as h [ 0 in equality (3.85) and using the assumptions in Lemma 3.13, we see that llm
y(r,r
+
HO
ft)/(x)-/(»)
= K
h
V
'
'
for all x G Rd. Moreover, equality (3.85) and the fact that the backward propagator Y is a family of contractions on the space Lf, imply sup h:0
Y(T,T + h)f(x)-f(x) sup —i-! '-^-i i±-i- < d
x€M
"
lr
sup
(s,x)S[r,T]xK<1
tl
\Lsf(x)\
s.
< oo (3.87)
for all / G Cg. It follows from (3.86) and (3.87) that hiO
h
for all / G Cg. Hence Cg C £» ( J 4 ^ ( T ) ) for all r e [0,T), and AM(T)/ L T / for all r G [0, T) and / G Cg.
= •
What second order partial differential operators on R d can be generators of diffusion processes? This question will be partially answered in the sequel. We will need the basic facts from the theory of stochastic integration and stochastic differential equations. Some of the results gathered below are formulated without proofs. We refer the reader to Section 3.12 for the bibliography concerning stochastic integrals and stochastic differential equations. Let us consider the following stochastic differential equation: dXt = b(t,Xt)dt
+ a(t,Xt)dBu
(3.88)
where Bt is a standard Brownian motion. Since Brownian paths are not of finite variation, the equation in (3.88) cannot be integrated pathwise using Riemann-Stieltjes integral. If we integrate the stochastic differential equation in (3.88) informally, the resulting integral equation is as follows: Xt=XT+
f b (s, X„) ds+ j
a (s, X.)
dBs.
(3.89)
Hence, in order to understand the meaning of the stochastic differential equation in (3.88), one should know how to integrate stochastic processes with respect to Brownian motion. The theory of stochastic integration and stochastic differential equations was developed by Ito in 1950's. Next
236
Non-Autonomous Kato Classes and Feynman-Kac Propagators
we will discuss the elements of this theory. Most of the proofs of the results formulated below are omitted, and we refer the reader to Section 3.12 for the bibliographical information on stochastic integration and stochastic differential equations. Let (BtjJ^.P) be a standard one-dimensional Brownian motion (see Subsection 1.12.1). A stochastic process Hs, s € [T}T], is called a simple process if there exists a partition r = so < si < • • • < sm = T of the interval [T,T] and a finite sequence of random variables hi, 0 < i < m — 1, such that hi is !FSi -measurable and Hs = hi for all Sj < s < Si+\. The stochastic integral or the Ito integral of the process H3 is the stochastic process defined as follows:
/ H.dB, = ^2 hi (BSi+1 -BSi)+ JT
hj (Bt - BSj)
(3.90)
t=o
for Sj < t < Sj+x, 0 < j < m — 1. The next lemma concerns Ito integrals of simple processes. Lemma 3.14 Let Hs, r < s
(2) For all t with
is a continuous
[T,T].
r
:E
y>>!
ds
(3.91)
(3) The following inequality holds: E
sup
(i(H)ty < 4 E
t:r
r
(Hs)2 ds
(3.92)
Next we will explain how to integrate more general stochastic processes. Let r € [0,T), and denote by HT the class of all stochastic processes Ht, t £ [r, T], such that Ht is an ^J-adapted process, and E
i:
(Hsy ds
< 00.
The next assertion concerns the existence of Ito integrals of stochastic processes from the class HT.
Non-Autonomous
Koto Classes of Measures
237
Theorem 3.6 There exists a unique linear mapping J from the space HT into the space of all continuous T\ -martingales on the interval [T,T] such that (1) For any simple process Ht, t £
[T,T],
J(H)t = J* H dB . s
(2) For all t with r
s
ds
/ > > '
The mapping J in Theorem 3.6 is unique. More precisely, if J\ and J
f HsdBs
=E
f
(Hs) ' ds
(3.93)
The next inequality for Ito integrals is a generalization of the inequality in (3.92): 21
E
sup ( I HsdBs) sup
<4E
(Hs)2ds
/ JT
(3.94)
It is also known that for all r € [0,T], q > 0, and H £ HT, there exist positive constants aq and Aq depending only on q and such that aqE
tH
'I -
ds
<E
sup t:r
i:
TIH-
H.dB,
' ds
(3.95)
This result is a special case of the Burkholder-Davis-Gundy inequalities (see [Ikeda and Watanabe (1989); Revuz and Yor (1991)]). Our next goal is to explain what is the precise definition of a solution to the stochastic differential equation in (3.89). Let d and n be positive integers, and let aij{r,x), 1 < i < d, 1 < j < n, be a matrix field on
238
Non-Autonomous Kato Classes and Feynman-Kac Propagators
[0, T] x R d with values in the space oi dxn matrices. In this section we use column vectors in the spaces R n and R d . For a d x n matrix a, the symbol laI stands for the norm of a defined by I |2
V^
V
l
Let b(r, x) — (&I(T, a;),..., bd(r, x))* be a vector field on [0, T] x R d , and let Bt = (i?t,.. -,-B") D e a standard Brownian motion in R n , where v* denotes the transpose of the row vector v. Next we explain when a stochastic process Xt = (Xj,...,Xd) on Q, with values in Rd is a solution to the stochastic differential equation in (3.88). The process Xt should be adapted to the filtration J-t generated by Bt, satisfy the conditions E
J \bi(s, X,)\ds
< oo, 1 < i < d,
and E
J 4 (*>*<•) ds
< oo, 1 < i < d, 1 < j
and be such that
x\ = x ; + f bi (s, x.) ds + Yl [ ua (s- x°)
dB
i
(3.96)
P-almost surely for all 1 < i < d and 0 < r < t < T. Next we formulate and prove an existence and uniqueness theorem for stochastic differential equations. Note that by Tl is denoted the two-parameter filtration generated by Brownian motion. Theorem 3.7 Suppose that a is a continuous matrix field and b is a continuous vector field on the space [0, T] x R d . Suppose also that the following conditions hold: (1) There exists a constant c\ > 0 such that \b(t, y) - b{t, z)\ + \a{t, y) - a(t, z)\ 0 such that |6(«,l/)| + k ( t , l / ) | < C 2 ( l + |y|)
z\
Non-Autonomous
Kato Classes of Measures
239
for all te [0,T] and y 6 R d . £e£ r G [0,r], q > 1, and /ei Z 6e an Rd-valued !F^-measurable random variable on CI, where c is such that 0 < c < r . Assume that E[\Z\q]
+ f b (s, X.) ds+ J a (s, Xs) dBs.
(3.97)
The process Xt, r < t < T, is continuous, and for every t e [T,T], the random variable Xt is measurable with respect to the a-algebra a (Z, B32 — BSl : T < s\ < s% < t) augmented by the events A such that F(A) = 0. Moreover, the following inequality holds: E
\Xt\9
sup
(3.98)
< OO.
,f.T
/ / Xt and Yt are two solutions to equation (3.97) satisfying the conditions in Theorem 3.7, then Xt and Yt are indistinguishable. Proof. Given r € [0,T] and c € [0, T], consider the vector space S!* c consisting of all ^ - a d a p t e d continuous processes Xt on [r, T] with state space R d such that
1X11 =
sup
|X t |«
< OO.
(3.99)
t:r
Then |-| is a norm on the space «S* e . It is not hard to prove that the space 5* c equipped with the norm defined in (3.99) is a Banach space. Next we will show that the function
f
b{s,Xs)ds+
f
a(s,Xs)dBs
(3.100)
where Z is the initial condition in the formulation of Theorem 3.7, maps the space <S
240
Non-Autonomous Koto Classes and Feynman-Kac Propagators
3.6). Using the linear growth condition for b, we get
f b(s,Xs) ds
sup f.T
< (T - r)qE
\b(s,Xs)\Q
sup S:T<8
\JT
<4(T-r)m\
sup
(1 + |X.|)«
<
(3.101)
00.
\_S:T<S
In addition, using the linear growth condition for a and inequality (3.95), we obtain
J (s,Xs) dBs
sup
(fkh X )rds
s
t.T
< Agc92(T - T)%E
(l + \Xs\)q\
sup
(3.102)
,S:T<S
Now it is not hard to see that the condition $ ( X ) G <S^C follows from (3.101), (3.102), and the estimate E [\Z\q] < oo. Our next goal is to establish that for small values of the difference T — T, the mapping $ : Sqc —> Sqc is a contraction. Let X e Sqc and Y e 5 ^ . Then \\$(X)-$(Y)\\l <2g~1
<E
sup
( / \b(s,X3)-b{s,Ys)\ds
• r
+E
\JT
(s,Xs)-a(s,Ys))dBi
sup
(3.103)
t:r
Next, using the same ideas as in the proof of (3.101) and (3.102), we see that condition (1) in Theorem 3.7 implies E
sup
(f
t:T
\b(s,Xs)-b(s,Ys)\ds
\JT
< ( T - T)«:
\b(s,Xs)-b(s,Ys)\q
sup ,S:T<S
\XS-Ys\q
sup S:T<S
=
<%(T-T)"\\X-Y
19
I?
(3.104)
Non-Autonomous
Kato Classes of Measures
241
and E
I f* sup
/
(a(s,Xs)-a(s,Ys))dBs
t:T
sup
\a(s,Xs)-a(s,Ys)\q
IS:T<S
sup
\XS-Ya\q
S:T<S
=
Aqc\{T-T)i\\X-Y\\l.
(3.105)
Next, combining (3.103), (3.104), and (3.105), we see that ||$(X) - $(Y)\\q
< 2 ^ (c\(T - T )« + Aqc\{T - T ) * )
?
||X - y | | , . (3.106)
Estimate (3.106) shows that if 2*-1(c?(r-T)« + v ! C r - T ) * ) < i , then the mapping $ : Sj? c —> <S^C is a contraction. Let us denote by Z the stochastic process defined by Zt = Z, r < t < T. Then it is clear that Z € <S*C. It follows from the fixed point theorem that the sequence $"(.£) converges in the space S%c to a stochastic process Xt, T < t < T, which is a solution of equation (3.97). Moreover, the random variables JT b (s,
242
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
Lemma 3.15 Let f be a continuous function on the interval [r,t\, and suppose that there exist a and (3 such that —oo < a < oo, P > 0, and + pJS
f(s)
ae^-T\
for all s e [r,t]. Then f(t) < Proof.
f{u)dv
The case @ = 0 is trivial. Next suppose that 0 > 0. Set /
f{u)du.
Then g'(s) = e-M'-r'*
-a
\
f(u)du +
M
<
s T
fte-«
")
for all r < s < t. Integrating the previous inequality from T to t, we get 9(t)
e-^dv
= - ( l - e-^-^)
.
Therefore,
£ f{u)du < | (J*-') - l) , and it follows that /(<) <<* + / ? / f{u)du
This completes the proof of the Gronwall lemma.
Let Xt and Yt be two solutions to equation (3.96). Then, using the Lipschitz estimates for b and a, Holder's inequality, and equality (3.93), we see that for every t 6 [r, T], 2
E \Xt-Yt\
<2E + 2E <2{t-r)
(b(s,Xs)-b(s,Ys))ds
/ J
(o-(s,Xs)-a(s,Ys))dB< J
E[|6(s,Xs)-6(s,ys)|2]ds
+ 2 / E \a(s,Xs) / '
-a(s,Ys]
ds
Non-Autonomous
243
Koto Classes of Measures
<2cl(T-T
+l
Jf E
\XS — Ys\
ds.
(3.107)
It is not hard to see that the function
fit) = E \Xt-Y,\ is continuous on the interval
[T,T\.
Therefore, (3.107) implies that
/(*) < P I /'
f(s)ds
for all t G [r, T] where /3 = 2c2 (T - r +1). By the Gronwall lemma (Lemma 3.15), /(£) = 0 for all t £ [T,T], It follows from the previous assertion that for all t £ [T,T], Xt = Yt P-almost surely. Since Xt and Yt are continuous processes, they are indistinguishable (see Subsection 1.1.4). This proves the uniqueness of solutions in Theorem 3.7. • Let x G Rd, and assume that XT{w) — Z(w) = x for P-almost all w £ ( l . In this case, we will use the symbol X\'x for the unique solution of equation (3.96) in Theorem 3.7. Note that for fixed r and x, the process Xj'x is defined P-almost surely with the exceptional set depending on T and x. Theorem 3.8 There exists a modification of the process (r, t, x) — i > Xj'x such that its sample paths are locally Holder continuous of order a < \. Proof. Fix a pair (r, a;) G [0, T] x R d and put Xl'x = x for all 0 < t < r. Then the function t — i > XJ'X is continuous on the interval [0,T] P-almost surely. Here the exceptional set depends on T and x. For r £ [0, T], x G Md, and y G M.d, consider the processes X\,x and Xj'v where r < t < T. Since these processes are solutions to the stochastic differential equation in Theorem 3.7, we have xr,x
_ xr,y
=
x
_
y +
/'
y*
[ft ( S ) Xr,x)
_
fc (fl> X J
,B)]
d s
[a^XJ^-tr^XJ^ldS.
(3.108)
R d -almost surely. The exceptional set in (3.108) depends on r , x, and y. For q >2, set /(t) = / ( t ; r , x , y , g ) = E
sup ,u:r
|X^
X. T,y\Q
r
(3.109)
244
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
Then / is a continuous function on [T, T]. Indeed, for 0 < t < s < T, we have
sup
f(s) - f(t) < E
{x^-x^i"
u:t
Now the continuity of the process Xt, estimate (3.98), and the monotone convergence theorem imply the continuity of / . It follows from (3.108), (3.109), the Lipschitz estimate in (1) of Theorem (3.7), Holder's inequality, and (3.95) that t
f(t)<3"-1\x-y\9
+ 3('-1c\E \XJ'X
+ 3«-1C?J4,E
<39_1|a:-2y
+ V-'cl
\XI'X
-XI>y\ds
\ i
-XJ'y\Us
((t - T ) " - 1 + Aq (t -
T)2*2)
pt
E
'ds
< 39"-i \x _ y\i + 3 9 - i c ? ( ( t _
T ) « - i +Ag(t~
r)2^) J
f(s)ds. (3.110)
By (3.110) and the Gronwall lemma, E
sup
\X2X-XZV\9
u:r
< 3 " - 1 \x - y\q exp ^ " M ((t - T)" +Aq(t-
T)*) }
for all i € [T, T], x G R d , and y S R d . Hence, E
sup
W-XW
\x-y\'
(3.111)
TJ,-.T
for all 0 < r < t < T, x € R d , and j/ € Rd, where the constant « i depends only on q, T, and c\. Fix i € l l i and t £ [0, T]. Then for all T\ and T2 with 0
Y'T2'X
/ [b (s, X? •*) - b (s, X?•*)] ds + r J To.
JTI
b (s, X? •*) ds
Non-Autonomous Kato Classes of Measures
+ f [a (s, X?'x)
- a (3, X?>x)} dBs + P a (s, X^x)
245
dBs
(3.112)
P-almost surely. The exceptional set in (3.112) depends on n , r 2 , and x. For every q > 2, put a(t) =g(t-,T1,T2,x)
=E
sup \x?-* - x?>x\q
r2
(3.113)
,U-T2
Then, using (3.112), (3.113), the Lipschitz and the linear growth estimates in conditions (1) and (2) of Theorem (3.7), Holder's inequality, and (3.95), we obtain 9(t)<4'-1
X™
- X?*\9
da
V T2
+ v-1cq2(T2-T1y-lE\n(i 4"-1Aqcql{t-T2)^1E
+
+ \x,n,x\\q
ds
( \X?'X - X?'x\q
ds
VT2
A"-1AQcq2(T2-T1)^E
+
r\i+\x?'x\)qds
< tf-'cl ((r 2 - n ) " - 1 + Ag (r 2 +
n
) ^ ) E ^
^-^cq1^t-T2)g-1+Aq(t-T2)^1)J
(1 + \X?>X\)9 ds
g(s)ds.
(3.114)
It follows from (3.114) and the Gronwall lemma that E
sup
)X?"
- X?**]'1
U:T2
< 4"- 1 c« ( ( T 2 - n ) " - 1 +Aq (T2 - n ) * ^ ) E
P
(1 +
exP{4"-1c?((i-r2)'7 + A q ( i - r 2 ) ^ }
\X?'x\)qds (3.115)
for alHG {T2,T}. Next let TI < u < T2. Then X?*
- X?'* = X?*
= [Ub(s,X?>x)ds+
- x f1 (T(s,X?>x)dBs
(3.116)
Non-Autonomous Kato Classes and Feynman-Kac Propagators
246
P-almost surely. Using (3.116) and reasoning as in the proof of (3.114), we see that \X?>X-X?>x\q
sup
E
U:TI
< 2«- 1 c§ ((r 2 - n)"-1 + Aq (r 2 - n ) ^ ) E \fj
(1 + \X?>x\f'ds . (3.117)
It follows from (3.115) and (3.117) that E
\X?*-X?'X\9
sup u:T\
< 2<-14 ((r 2 - n ) 9 " 1 + Aq (r 2 - n ) 2 ^ ) E [ C (1 + \X?<X\)9 ds x ( 2 ' - 1 exp {4"-lc\
((« - r 2 ) 9 + Aq(t-
75)*) } + l )
(3.118)
for all t G [ T I , T ] . Moreover, for all 0 < T < h < t2 < T and x G Md, the
following estimate holds: E[\Xlix-X^\9] < 2"-lcl
((i 2 - Uf-1
+ Aq (t2 - i j ) ^ 1 ) E
f ' (1 + IXJ^D" ds (3.119)
We also have I YT'x\q
our. SU
E
P
T
\^u I
39~
< 3 9 "11 UI9 |x
T,X\\Q
< 3 9 " 1 |z| 9 + e ^ c 9 ((* - r ) 9 + A, (t + r
1
4 ( ( t - T r
1
T)*)
+ A , ( t - r ) ^ ) E [ f
^
ds
'
LJT
sup
K'
ds
U:T
Therefore, by the Gronwall lemma, E
sup
9 1 l | x ; > T < 3 " \x\" + 6"- 4
((* - r ) 9 + A,(t - r ) § ) ]
x exp { e ^ c 9 ((t - r ) 9 + j4*(t - T ) * ) } .
(3.120)
Non-Autonomous
Kato Classes of Measures
247
It follows from (3.118) and (3.120) that E
sup
X?' x l
\X?<*
< a2 ((r 2 - n)" +
(T 2
- T i ) ? ) (a 3 + |x|*)
U-.Tl
(3.121) for all 0 < Ti < r 2 < T, n < * < T, and x £ Md. The constant a 2 > 0 in (3.121) depends only on q, T, c\ and C2, while the constant a$ > 0 depends only on q, T, and c 2 . It is not hard to see that (3.119) and (3.120) imply E [\Xl;x-Xl;x\q]
< aA ((t 2 -txy
+ ( * 2 - t i ) * ) (as + \x\q)
(3.122)
for all 0 < r < ti < t2 < T and x £ Rd. The constants a 4 > 0 and a5 > 0 in (3.122) depend only on q, T, and c 2 . Let 0 < TI < r 2 < T, n < ti, n < t2, x £ Kd, j / £ Rd, and suppose that |T2 — Ti| < 1 and |£2 - *i| < 1. Since V T l ' X l | * " \VT2,X2
\yT2,X2
|^t2
- A
t l
| < |At2 +
YT2,Xl\
- A
A
| t2
t 2
_ A
t l
I I V T 2' X 1
| + |At2
V r l >X1 I
- A
t 2
|
|'
the estimates in (3.111), (3.121), and (3.122) imply E [|A-£'Xa - X ^ 1 1 9 ] < 3«" 1 ai |x 2 -
q
Xl\
+ 2 x S ' " 1 ^ |T 2 - n | * (a 3 + \xi\9) + 2 x 3*- J a 4 |*2 - * i | f («5 + | x i | 9 ) . Hence, there exists a constant a > 0, depending only on q, T, c\, and c 2 , for which
E[|x-t7'Xa-A-t™n < a {\x2 -
q
Xl\
+ (|T2 - TI|* + |*2 - i l l 1 ) (1 + l n H } .
(3.123)
In order to complete the proof of Theorem 3.8, we need a multidimensional version of Kolmogorov's criterion for the existence of continuous modifications of a stochastic process. This criterion concerns stochastic processes indexed by subsets of Euclidean space E". Let C = [oi, b\] x • • • x [o„, bn) be a rectangular box in Rn, and consider a stochastic process Xs, s £ C, on a probability space (fi,^ 7 , P) with state space Rd. The following assertion is a multi-dimensional version of Kolmogorov's criterion (see, e.g., [Revuz and Yor (1991)]).
248
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
Theorem 3.9 Suppose that there exist three strictly positive constants q, c, and e such that
E{\Xs-Xr\l}
(3.124)
for all s G C and r £ C. Then there is a modification Xs of the process Xs ~ £ such that the paths of X„ are Holder continuous of order a with 0 < a < -. The symbols |-|d and |-| n in the formulation of Theorem 3.9 stand for the Euclidean norms in Rd and M", respectively. Next we will rewrite inequality (3.123) using the notation in the formulation of Theorem 3.9. Let k > 1, and denote by Ck a rectangular box in M.d+2 given by Ck = [0, T]2 x Dk, where Dk is the cube in Rd defined by Dk = {x € Rd : \xi\
(3.125)
for all q > 2, s G Ck and r G Ck- The constant @k in (3.125) depends only on k, q, T, c± and C2. It follows that if n = d + 2, q > 2(d + 2), and £ = | — d — 2, then Kolmogorov's criterion can be applied. Therefore, there exists a modification X* of the process Xs such that the paths of X* are Holder continuous of order a with a < \. Here we assume that s £ Ck- Since any two continuous modifications of a stochastic process are indistinguishable, it is not hard to construct a required modification of the process (r,t,x) — i > X\'x in Theorem 3.8 using the processes X^. This completes the proof of Theorem 3.8. • The next assertion concerns the Markov property of the process X\,x with respect to the Brownian filtration. The uniqueness of solutions of equation (3.96) will be used in the proof. Theorem 3.10 Suppose that b and a satisfy the conditions in Theorem 3.7, and let X\,x be the unique solution to equation (3.96). Then for all T < s < t
(3.126)
We will first establish the flow property of the process Xl'x\ that 3 x x xr,x = x t' °'
P-almost surely.
(3.127)
Non-Autonomous Kato Classes of Measures Indeed, using the fact that Xj'x we get X?* =x + Jb(p,
is the unique solution to equation (3.96),
X;<x) dp+ J
= x + Jb(p,X;>x)
249
dp+j
a (p, X^')
dBp
a (p,XTp<x) dBp
+ f b (p, x;>*) dP+ fa (p, x;>*) dBp Js
Js
= XZ>X + f b(p, *;•*) dp + f a (p, XTp>x) dBp. Js
(3.128)
Js
Therefore, (3.96) with the initial value equal to XZ'X implies that Xst'Xl'* = XI<* + J
b (p,XSP'X:'X)
dp+ f a (p,Jtfx'")
dBp.
(3.129)
Now (3.128) and (3.129) imply that the processes Xj'x and Xt'X''° the same stochastic differential equation
satisfy
Xt = XZ>x + f b(p,Xp)dp+ I Js
a(p,Xp)dBp.
Js
By the uniqueness of solutions in Theorem 3.6, equality (3.127) holds. It follows from the flow property in (3.127) and Theorem 3.7 that for T < s < t < T, the random variable Xpx is measurable with respect to the cr-algebra a (XZ'X, B32 — B3l : s < s\ < s2 < t) augmented by the family of P-negligible sets. Therefore, for every Borel function / on E d , the random variable / (XT'X) is measurable with respect to the same cr-algebra. Since the Brownian increments are independent, and the process XZ'X is adapted to the Brownian filtration !F7, the random variable XJ'X is independent of the Brownian increments after time s. Next let Ai £ a (XZ'X) and A2 € a (BS2 — BSl : s < si < s2 < t). By the properties of conditional expectations, we get E [XMXA2 | T[\ = XA^
[XA2 I ?Ts] =
XAJ?[A2]
(3.130)
and E[XMXA2
| v(XZ<x)]
=XAM*M
I " (*;•*)] =XA^[A2}.
(3.131)
250
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
By (3.130) and (3.131), the Markov property holds for the functions XA-LXAI- It follows from the monotone class theorem that the first equality in (3.126) holds. The proof of the second equality in (3.126) is similar. This completes the proof of Theorem 3.10. • The next result is the celebrated Ito's formula. It is known that Ito's formula holds for more general stochastic processes Xt and functions / than those in the formulation of Theorem 3.11 below (for more information, see the references in Section 3.12). Theorem 3.11 Let (r, x) € [0, T] x Rd, and suppose that b and a satisfy the conditions in Theorem 3.7. Let Xt = Xl'x be the unique solution to the stochastic differential equation in Theorem 3.7. Then for every function f € C 1 ' 2 ((T, T) x R d ) such that the partial derivatives —— are bounded for all 1 < i < d, the following formula holds: rt
f(t,Xt)=f(T,XT)+J +
^(s,Xs)ds
E /
bi(s,Xs)^-(s,Xs)ds
=i
i
»=i j=i
1
/"'
+ 2E i,j = l
a
s X
d2 f
/T H ( > *) -foT^T. (*' **) ds> x
( 3 - 132 )
3
where aij(s,y) = {o-a*)^ (s,y) =^2aik(s,y)ujk{s,y).
(3.133)
k=l
The last term on the right-hand side of formula (3.132) is called the Ito correction. The next lemma follows from the Ito formula and from the fact that under the conditions in Theorem 3.6, stochastic integrals are continuous martingales. Lemma 3.16 Let (T,X) G [0,T] X Rd, and suppose that b and a satisfy the conditions in Theorem 3.7. Let Xt = Xj'x be the unique solution to the stochastic differential equation in Theorem 3.7. Then for every function f G C 1 ' 2 ((r, T) x Rd) such that the partial derivatives —— are bounded for OXi
Non-Autonomous
Kato Classes of Measures
251
all 1 < i < d, the process f (t, Xt) - f (r, XT) - f Laf (s, X.) ds-
J
~
(s, XB) ds
(3.134)
is a continuous Tl -martingale on [T,T]. In (3.134), the operator Ls is defined by (3.79) withaij given by formula (3.133). Our next goal is to show that the operator LT in (3.79) generates a diffusion process (see Definition 3.18). Theorem 3.12 Let b and a satisfy the conditions in Theorem 3.7, and let a = era*. Then the family of operators LT defined by (3.79) is the generator of a diffusion process Xt such that its sample paths are Holder continuous of order a < | . Proof. We will first show that it suffices to construct a continuous Markov process (Xt,J-[,PT,x), 0 < r < t < T, with state space Rd such that ETlX\f(Xt)]=E[f(Xl'x)]
(3.135)
for all 0 < T < t < T, x G M.d, and all bounded Borel functions / on Rd. Indeed, if / is a Cg-function on M.d, then Lemma 3.16 implies that the process
/ (xz>x) - f (x;>x) - J* LJ (xj-*) ds is a martingale with respect to the measure P. It follows from (3.135) that formula (3.80) holds. Our next goal is to prove equality (3.135). Put P(T, X; t,B) = P (Xpx e B),
(3.136)
where 0 < T < t < T, x £ Rd, and B £ Bu
Y(T, t)f{x) = [
(3.137)
d
JR
where 0 < r < t < T, x € Rd, and / is a Borel function on Rd. We will need the following lemma: Lemma 3.17 Suppose that b and a satisfy the conditions in Theorem 3.7, and let XJ'X be the unique solution to the stochastic differential equation
252
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
in Theorem 3.7. Then the function P defined by (3.136) is a transition probability function, and the family of operators Y(r,t) given by (3.137) is a backward Feller propagator. Proof. It follows from Theorem 3.8 that without loss of generality we may assume that the sample paths of the process (T, t, x) — i > Xj'x are continuous. The family of operators defined by (3.137) is a backward propagator on the space Z/£°. Here the symbol £ stands for the Borel cr-algebra of the space R d . This assertion can be obtained from Theorem 3.10. Moreover, it is not hard to see that the function P defined by (3.136) is a transition probability function. It only remains to prove that Y is a Feller propagator. Let / be a bounded continuous function on R d . Then for all 0 < r < t < T, the function Y(T, t)f(x) is continuous on R d . This fact can be obtained using the continuity of the process (r, t, x) H-> X \ , X and equality (3.137). This completes the proof of Lemma 3.17. • Let us go back to the proof of Theorem 3.12. Since the function P defined by (3.136) is a transition probability function (see Lemma 3.17), there exists a Markov process (Yj, T[, P r ,x) with state space R d such that P is its transition function. Here Tl[ = a (Ys : T < s < t). It follows from (3.136) and the Markov property in Theorem 3.10 that ET,x[F(Ytl,...,YtJ] = E[F(X^,...:X^)]
(3.138)
for all r < t\ < • • • < tm < T, x € R d , and all positive Borel functions F on [Rd]m. Now it is clear that (3.122) and (3.138) imply Er,x{\Xt1-Xt2\«}=E[\x;;*-Xj;*\q} < a4 ((£ 2 - h)9 + (t2 - £ i ) § ) (a5 + \x\q)
(3.139)
for all r < t\ < t 2. Reasoning as in the proof of Theorem 3.8, we see that there exists a modification Xt of the process Yt such that its sample paths are Holder continuous of order a < | . It follows from (3.138) that equality (3.135) holds. Therefore, the process Xt is a diffusion process satisfying the conditions in Theorem 3.12. The proof of Theorem 3.12 is thus completed. • The next lemma concerns the Feller-Dynkin property of the backward propagator Y. In our opinion, this lemma has an independent interest. Lemma 3.18 Suppose that b and a are bounded on R d and satisfy the conditions in Theorem 3.7. Let XJ'X be the unique solution to the stochastic
Non-Autonomous
Kato Classes of Measures
253
differential equation in Theorem 3.7. Then the family of operators Y(T, t) defined by (3.137) is a backward Feller-Dynkin propagator. Proof. By Lemma 3.17, it suffices to prove that for all 0 < r < t < T, f £ C0 => Y(r, t)f € C 0 . Let / € C 0 . Then for all M > 0, we have \Y(T,t)f(x)\
< |y(T,t)/XB(x,M)(x)| +
<
sup
\f(y)\ +
\Y(T,t)fx*\B(x,M)(X)\
\\f\\00F[\Xr~x\>M]
y:\x-y\<M
<
T.X |/(y)| + U ^ f •E IX t
sup
M
y.\x-y\<M
j
~
X
\
(3.140)
Since XTt'x -x=
f b(a,XTS'X)ds+
f a(s,XTs'x)dBs,
(3.141)
<2(T27l2+T722),
(3.142)
we have sup E x€R
d
\xl
•x
where 71 =
sup
|6(s,x)|
and 7 2 =
(s,i)e[o,r]xiy
sup
(a,i)e[o,T]xi d
|a(s,x)|.
In the proof of estimate (3.142), we used (3.141) and (3.93). It follows from (3.140) and (3.142) that \Y(T,t)f(x)\
<
sup y:\x-y\<M
J | / ( y ) | + 7,11/lloc M2 '
(3.143)
where the constant 7 > 0 does not depend on x. Now it is clear that (3.143) implies Jim
M
sup
\Y(T,t)f(x)\=0.
-*°°x:\x\>M
This completes the proof of Lemma 3.18.
•
In the proof of Theorem 3.12, we implicitly solved the so-called martingale problem. This problem was formulated and studied by Stroock and Varadhan (see [Stroock and Varadhan (1979)]). Let Q, = C ([0,T],Rd), where C([0,T],R d ) is the space of all R d -valued continuous functions on the interval [0,T], and denote by Xt the coordinate process on f2. Put Tl =a(Xs :T <S
254
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
Definition 3.19 For (r, x) G [0,T] x Rd, a probability measure WT>X on Q is a solution to the martingale problem TT(T,X), provided that (1) P r , x [XT =X] = 1.
(2) For any function / G Cg°, the process
t .-> / (Xt) - f (XT) - J L,f (Xs) ds is an ^"-martingale on [T,T] with respect to the measure P T , X . By Theorem 3.12, if (Xt,fl,!FJ,FT>x) is a diffusion process generated by LT, then for every pair (T,X) G [0, T] X Rd, the measure ¥T>X is a solution to the martingale problem -K(T,X). On the other hand, suppose that the operator LT is given, and the problem is to determine whether there exists a diffusion process Xt generated by the operator LT. In order to find such a process, one can start with the martingale problem IT(T,X). If the solution PT,a; exists for all pairs (r, x) G [0, T] x Rd, then one can try to prove that the coordinate process Xt is a Markov process with respect to the family {P T>X }. It is known that if for every pair (T, X) G [0, T] x Rd, the martingale problem TT(T, X) is uniquely solvable, then the Markov property holds (see, e.g., [Stroock and Varadhan (1979)]). We refer the reader to [Stroock and Varadhan (1979)] for more information on the martingale problem. Next, we will show that the diffusion process in Theorem 3.12 possesses the strong Markov property. Let us first recall what was established in Corollary 2.1. Let P be a transition probability function, and let (Xt,QJ,Pr,x) be an adapted right-continuous Markov process on fi associated with P. Suppose that the following conditions hold: (1) For every B G £, the function (r, x,i) H-> P(r,x;t,B) is S[0,r] ® £ ® S[o,r]-measurable. (2) For any / G Co and t G (0,T], the function (T,X) H-> Y (r,t) f(x) is right-continuous in r on the interval [0, t) and continuous in x G E. Then Xt is a strong Markov process with respect to the family of measures {Pr,x} (see Definition 2.20). This means that for any admissible family M of stopping times, r G [0,T], x G E, Si G M(T), S2 G M{T) with T < Si < S2
Er,x [f(S2,Xs2) I GTSl] =ESuXsi
[f(S2,XS2)}
Non-Autonomous
Koto Classes of Measures
255
holds P TiX -almost surely. The next assertion concerns the diffusion process Xt in Theorem 3.12. Theorem 3.13 Suppose that the conditions in Theorem 3.12 hold, and let Xt be the diffusion process whose existence was established in Theorem 3.12. Then Xt is a strong Markov process with respect to the family of measures {P r ,x}Proof. By Theorem 3.12, the process Xt is continuous. Moreover, condition (3) in Corollary 2.1 holds. Indeed, let x0 € M.d, r < so < T, and / € Co- By equality (3.134), condition (3) in Corollary 2.1 is equivalent to the equality (
lim
E [/ (Xi<*)] = E [/ (XSt°'X0)}.
(3.144)
Since the process XJ'X is continuous with respect to the variables r , t, and x, the equality in (3.144) can be established using the dominated convergence theorem. Next, applying Corollary 2.1, we see that the process Xt is a strong Markov process with respect to the family {P T)X }. This completes the proof of Theorem 3.13. •
3.9
Additive Functionals Associated with Time-Dependent Measures
This section is devoted to additive functionals of non-homogeneous progressively measurable Markov processes. The existence of additive functionals associated with time-dependent measures has already been mentioned (see Section 3.1). The next definition introduces an important class of transition probability functions. Definition 3.20 The class VM of transition probability functions is defined as follows: P £ VM if and only if there exists a progressively measurable Markov process Xt with P as its transition function. Given a transition probability function P e VM, we will always choose a progressively measurable process Xt to represent P. Suppose that V is a nonnegative function from the non-autonomous Kato class V*. Then for all T and t with 0 < r < t < T, the random variable (s, u>) \—> V (s, X$(UJ)), T < s < t, is B[Ttt\ <8> J7"-measurable. Moreover, the random variable w— i>/ V
(s,Xs(oj))ds
256
Non-Autonomous Kato Classes and Feynman-Kac Propagators
is ^-measurable and finite PT>x-almost surely. It follows that the functional Ay defined by
Av(r,t)JS'V^d'-
«/X*.W<°°
10,
(3 . 145)
otherwise,
satisfies the following conditions: (1) For all r and t with 0 < r < t < T, the random variable AV{T, t) is finite everywhere on ft and ^"-measurable. (2) For all 0 < r < T, AV(T, T) = 0 everywhere on ft. (3) For all r and t with 0 < r < t < T, the process t H-> AV(T, t),T
= AV(T,X)
+
Av(X,t)
(5) For all 0 < T < t < T and x G E, ETtXAv{T,t)
=
N{V){T,t,x).
For any V €Pj, we define the functional Ay associated with V by AV(T,t)
= AV+{T,t)
-
AV-(T,t).
This definition makes sense since both terms on the right-hand side are finite. Lemma 3.19 Let V e V%. Then the functional Ay defined by (3.145) satisfies the following condition: For allO < r
T
< s < t).
Proof. It is easy to see that without loss of generality we can assume that V > 0. Let Uk be a sequence of numbers such that r < Uk < t and Uk T t. Since the random variable Ay {r,Uk) is T^k-measurable, it is also ^L-measurable. Moreover, for all w € Q, with / V (s, Xa(cj)) ds < oo, we have lim /
V(s,Xs(w))ds=
f V(s,Xs(u;))ds.
(3.146)
Non-Autonomous Kato Classes of Measures
257
It is not hard to see that in order to prove Lemma 3.19, it suffices to show that the set Cj = L
: f V{s,Xs{ij))ds
<j \
(3.147)
belongs to the cr-algebra Tl_ for all j > 1. This fact easily follows from the equality oo
pit,
(•
= n{-/
Cj = f ) I u : I
V(s,Xs(uj))ds<j
This completes the proof of Lemma 3.19.
•
Our next goal is to construct the additive functional A^ associated with a time-dependent measure \x from the non-autonomous Kato class V^. The first step in the construction is to approximate /z by a sequence of functions Vk in a special sense. We will need the potentials N(V) and N([i) and the functions M(V) and M(fi) defined after Remark 3.2. Recall that for V e Vj the function M(V) is given by M(V)(T,*) =
sup sup\N(V)(r,t,x)\, r:r
0
x£E
Similarly, for /x e V^ the function M(/x) is defined by M(/u)(r,t)=
sup r\r
sup\N(fj,)(r,t,x)\,
0
x£E
Lemma 3.20 Let P be a transition probability function, and let V S V*. For k > 1, 0 < r < T, and x € E, put Vk(r, x) = kN(V)
(T,
min
(T+^,T),X\.
(3.148)
Then the following conditions hold: Vk G V) for all
(3.149)
k>\; lim
sup
sup|JV(V-Vfc)(T,t,a:)|=0;
(3.150)
fc-»oo0
and lim supsup7V(|y fc |)(T,i,:r) = 0.
t T
- l°k>lxeE
(3.151)
258
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
Lemma 3.21 Let P be a transition probability function possessing a density p, and let fi G V^. For k > 1, 0 < r < T, and x G E, put Vk{r,x) = kN(fi) fr.min
(3.152)
(T+^TYX).
Then conditions (3.149)-(3.151) in Lemma 3.20 hold with \x instead
ofV.
Proof. We will only prove Lemma 3.20. The proof of Lemma 3.21 is similar. It is not hard to see that N(Vk)(r,t,x)
=k
Y(T,s)ds rt
= k
/
Y(s,u)V{u)(x)du
,min( S + £,T)
ds
Y(T,u)V(u)(x)du
,min(t+*T,T)
= kj
/•*
Y(T,u)V(u){x)duJ
xck(u)(s)ds,
(3.153)
where Xck{u) is the characteristic function of the set Ck(u) = \s:s
< m i n ( s + - , T j 1.
It follows from (3.153) that ,min(r+£,T)
N(Vk)(T,t,x)
= kj
,t
Y(T,u)V(u){x)dul
•[
Xck(u)(s)ds
Y(T,u)V(u)(x)du /.min(t+£,T)
+ k
rt
Y(T,u)V(u)(x)du
is. Xck(u)(s)ds
(3.154) Next (3.154) gives \N(V-Vk)(r,t,x)\
r min ( T +£.«) < I
Y(T,u)\V(u)\(x)du
/•m>n(r+£,t)
+ kj
rt
Y(r,u)\V(u)\(x)duJ ,min(t+£,T)
+ kj
xck(u)(s)ds ft
Y(T,u)\V(u)\(x)duJ
xck(u)(s)ds
Non-Autonomous Kato Classes of Measures
259
Since the Lebesgue measure of the set Ck(u) does not exceed \, we get \N(V~Vk)(T,t,x)\<2N(\V\)(r,min(r
+
^,t\,x]
,min(t+i,T)
+ Y(r,t)
Y(t,u)\V(u)\(x)du
= 2N{\V\) (T,nan(T + Y(r,t)N(\V\)
+
(t,mm(t
^,tYx\ + ^ T ) ) (X).
Therefore, sup |JV {V - Vk) (r, t,x)\ < 2 sup N(\V\) (T,min f r + \ , t) ,x + sup 7V(|F|) U , min U + i , T J , x J . (3.155) Now it is clear that condition (3.150) in Lemma 3.20 follows from condition (3.155) and the definition of the class Vf. Since ft N(\Vk\)(T,t,x)
/-min(s + £,T) Y{T,s)ds
,t
Y{s,u)\V{u)\(x)du
/ .min( s + i,T)
= fc / ds I
Y(T,u)\V(u)\(x)du
,min(t+|,T)
= /c/
,
y(r,u)|V(u)|(a;)du /
Xc*(u)(s)ds,
we get N(\Vk\)(T,t,x)
+
^,T),x
(3.156) It is easy to see that (3.149) follows from (3.156). It remains to show that condition (3.151) holds. Let e > 0. Then (3.156) and Lemma 3.2 imply that there exist 5i > 0 and fco > 1 such that sapN(\Vk\)(T,t,x)<e x€E
(3.157)
260
N'on-Autonomous
Kato Classes and Feynman-Kac
Propagators
for alii — r < <5i and k > fc0- Moreover, since Vfc e V*, there exists 82 > 0 such that #2 < ^i and (3.157) holds for alii — r < 82 and k < ho- Therefore, (3.157) holds for all fc > 1 and t - r < 62. This establishes equality (3.151) and completes the proof of Lemma 3.20. • Remark 3.6 Suppose that the conditions in Lemma 3.21 hold. Then it is not hard to see that (3.150) implies lim M(Vk)(T,t)
=
M(n)(T,t).
K—»00
Moreover, (3.156) gives limsupJV(|Vfc|)(T,i,z) <
N(\n\)(T,t,x)
k—*oo
and limsupM(|V f c |)(T,t)<M(|/i|)(T,t). fc—>oo
In the next definition, we introduce a special type of approximation based on the formulas in Lemmas 3.20 and 3.21. Definition 3.21 Let P be a transition probability function and let V £ VZ. By definition, a sequence of functions Vjt € V**, k > 1, approaches a function V in the potential sense if lim
sup
sup\N(V-Vk){T,t,x)\->0,
fc->oo0
and lim sup sup iV(|Vfc|)(r, t, x) = 0. t-rlO k>l
x€E
Suppose that P possesses a density p and let /J £ R . By definition, a sequence of time-dependent measures Hk £ Vm, k > 1, approaches a timedependent measure fi in the potential sense if lim
sup
sup \N(fi — /ife)(r, t,x)\ —> 0,
*:-*°° 0 < T < t < T x£E
and lim sup sup AT(|/ijt|)(r, t, x) = 0. '-TlOfe^lxGfi
Non-Autonomous
Kato Classes of Measures
261
It follows from (3.148), (3.152), and from the definition of the class V*, that the functions Vk in Lemmas 3.20 and 3.21 are bounded. Hence, for any function V 6 Vj (any time-dependent measure fi G V^J there exists a sequence of bounded functions approximating the function V (the timedependent measure n) in the sense of Definition 3.21. L e m m a 3.22 that
Let Vk be a sequence of Borel functions on [0, T] x E such
lim sup sup iV(|Vfej) (T, t, x) = 0.
t-rl°fc>ixe£
Then supfe ||Vjt|L < oo. The same result is true for time-dependent measures. Proof. for t -T
By the assumption in Lemma 3.22, there exists 5 > 0 such that < 5 and k > 1, supsupJV(|V f c |)(T,t,a:)
(3.158)
x£Ek>l
Moreover, for all 0 < r < t < T, there exists a partition r = to < • • • < tn = t of the interval [r, t] such that max {\tj+i — tj\ : 0 < j < n — 1} < S and n<5~lT. It follows that N{\Vk\) (r,t,x)
= J2 f ^
Y(T,S)
IVfc| (x)ds fJ+1Y(tj,s)\Vk\(x)ds
= Y^Y(T,tj) Jt
j=0
i
n-1
=
Y,Y(T,tj)N(\Vk\)(tj,tj+1)(x).
Therefore, (3.158) implies n-l
rp
sup sup N (\Vk\) (T, t, x) < Y\ N (| JVfc|) (tj, tj+1, x) < n< k>ixes k^ This completes the proof of Lemma 3.22 in the case of functions. The case of time-dependent measures is similar. D The next theorem concerns the existence of the additive functional A^ associated with a time-dependent measure /x € V*m.
262
N'on-Autonomous
Kato Classes and Feynman-Kac
Propagators
Theorem 3.14 Let P G VM. be a transition probability function possessing a density p, and let Xt be a corresponding progressively measurable Markov process. Then for every time-dependent nonnegative measure /J, from the class V^, there exists afunctional AM(r,t) of the process Xt such that (1) For all T and t with 0 < r < t < T, the random variable A^ (r, t) is nonnegative and finite everywhere on 0 . (2) For all T and t with 0 < r < t < T, the random variable A^ (T, t) is Tl_ -measurable. (3) For allO < T
T
PT,x-a.s.
=
N(fx)(T,t,x).
Proof. Let Vk be the sequence of functions in formula (3.152) corresponding to the time-dependent measure fi. Our first goal is to show that lim
sup
supE T]:c
fc,j->ooT:0
'
sup
\Avk{r,t)
— AvAT,t)\
=0
(3.159)
t:r
for every n > 1. This will allow us to define the functional A^ as the limit of the sequence Ayk. Lemma 3.23 Let P e VM and V G V*f. Then for all 0 < r < t < T, x G E, and all integers n > 2, \ET,XAV(T,
Proof. that
t)n\ < n\N(\V\)(r,
t, X)M(\V\)(T,
t)n-2M(V)(r,
t).
(3.160)
It follows from the Markov property and the condition V G V}
ET,xAv(r,t)2 = 2E r , x f V(s,Xs)ds
f
V(u,Xu)du
2EiTTlX f V (s, Xs) ds f ETiX (V (u, Xu) | Tl) du
Non-Autonomous
= 2E r , x J
V (s, Xs) ds J
Koto Classes of Measures
263
ESi2V (u, Xu) du \z=Xa
< 2 / dsY{T,s)\V{s)\{x)ds
sup
JT
sup f
s:r<s
Y(s,u)V(u)(y)du
(3.161)
Js
Now it is clear that (3.161) implies (3.160) with n = 2. Next, let n > 2 be any positive integer. By induction, we get ET,xAv(r,t)n ft
pt
= nlEr^ / V (h,Xtl)
ft
dh / V (t 2 , Xt2) dt2 • • • /
I Y(T,s)\V(s)\{x)ds
sup
JT
sup
r-.T
/
V (tn, XtJ
dtn
y(r,«)|V(u)|(j/)du
lJr
sup sup / Y(r,u)V(u)(y)du . r-.T
\AV(T,t)\n
< V ( n - l ) ! ( n + l ) . W ( | V | ) ( r , t, X)M(\V\)(T, Proof.
t)n-2M(V)(r,
t).
(3.162)
If n > 3 is odd, then ETlX \Av(T,t)\n
< {ET,xAv{T,t)n-l}h
{ETtXAv(r,t)n+1y
Now it is clear that (3.162) follows from Lemma 3.23.
. •
Lemma 3.23 and Corollary 3.1 provide pointwise estimates for the expression KTtX\Av(T,t)\n. The next lemma shows that stronger uniform estimates hold. Lemma 3.24 Let P € VM and V € V). Then for any r with 0 < r < T, any 6 > 0 such that r + S
'
sup tlT
AV(T, t)n
264
Non-Autonomous
Kato Classes and Feynman-Kac
< cnM(\V\)(T, T + S^MiV)^,
Propagators
r + 5),
(3.163)
where n
(3.164)
' „! + l
Moreover, for any odd integer n > 3, supE TiX x£E
\Av(T,t)\n
sup t:T
< cjlffl V|)(T, r + Sr^MiV)^, where Cn = Remark 3.7
T + 6)
(3.165)
^/Cn-icn+\. For n = 1, the following estimate holds: supE TjX xeE
sup
|Av(r, i)|
' t-.T
< {C2M{\V\)(T,T
+ S)M{V){T,T
+ 5)}K
(3.166)
Estimate (3.166) follows from (3.163) with n = 2. Proof. We will prove estimate (3.163). Estimate (3.165) follows from (3.163) and Holder's inequality. Let n > 2 be an even integer, and let V 6 V}. Given x € E and r, 5, and t with 0 < r < t < r + S < T, put M t = E T i X ( i 4 v ( T , T + 5) | ^ 7 ) . It follows from (3.160) that Mt is an .^-martingale from the space Ln. By the Markov property, for every t with 0 < r
Mt = AV(T, t)+
ET,X (V (s, X,)\7t)
ds
fT+S
= Av(T,t)+l
Y(t,s)V(s)(Xt)ds
P r x - a . s . Hence, Mt is a modification of the functional Mt = Av(r,t)
+ N(V) (t,r +
6,Xt).
Let us fix a partition r = to < t\ < • • • < tk = r + 6 of the interval [T, T + 5]. By Doob's inequality (see Subsection 5.6), we get ET:X
sup j:0<j
Av(T,tj)n
Non-Autonomous Kato Classes of Measures
<2"E T , X
sup M " + 2"ET)X
<2n
n n-1
< 2" (^—)
5,Xtj)\n
sup \N(V) ( i j , r +
j:0<j
265
3-0<j
+ 5)n+2n sup
ETtXAv(T,T
sup
z€E
\N(V)(s, T + 5, z)
S:T<S
E T > X A V ( T , T + 8)n + 2nM(V)(r,
T + 5)n.
(3.167)
It follows from (3.160) and (3.167) that ET,X
sup Av(T,tj)n
+ 6)n-1M(V)(T,T
+ 6)
(3.168)
j:0<j
for all x E E where Cn is defined by (3.164). Now choose a sequence of successive refinements of the partition T = to < h < • • • < tfc = T + (5 on the left-hand side of (3.168), for which the maximum length of the partition intervals tends to 0. Passing to the limit in (3.168) and using the monotone convergence theorem and the continuity of the functional Ay (T, t) with respect to the variable t, we get estimate (3.163). This completes the proof of Lemma 3.24. • Let us continue the proof of Theorem 3.14. For a nonnegative timedependent measure // G V^, define the sequence Vk by (3.152). Then (3.163) and (3.165) imply supEr,x xGE
\AVk-Vj(T,t)\n
sup t-.T
< CnM (\Vk - Vj\) (T, r + Sr-'M
(Vk - Vj) (r, r + 5).
(3.169)
It follows from Lemma 3.21 that formula (3.159) holds. Therefore, lim
sup supE T]X
k,j->o°T:0
x£E
sup \Avk{r,t)
—
Avj{r,t)\=Q,
t:r
and hence, there exists a functional AM(r, t) of the process Xt such that for all 0 < r < t < T, Ap(T,u) -
lim sup ET]X sup fc-»oo X£E
'
AVk(T,U) = 0.
(3.170)
u:r
The random variable A^(T, t) in (3.170) is measurable with respect to the aalgebra T[. It follows from (3.170) that there exists an increasing sequence nk of positive integers such that lim
fc +0o
-
sup
u:T<«
AM(T,u)-AVnk(T,u)
=0
(3.171)
266
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
P r , x -almost surely for all x £ E. The sequence n/t in (3.171) depends on T and t. Moreover, it can be chosen independently of x. This can be shown using (3.170) and reasoning as in the standard proof of the fact that the convergence in measure implies the existence of an almost everywhere convergent subsequence (see, e.g., the proof of Proposition 18 on pages 95-96 in [Royden (1988)]). We are finally ready to define the functional A^. For all 0 < r < t < T, put f Jim Av
(T, t),
^0,
if . Iim
sup
Av
(r,t) - Av
(r, t) = 0
otherwise.
(3.172) Then for all k > 1, Ayn is nondecreasing and continuous everywhere on fi. Moreover, the random variable Ayn (T, t) is .T^-measurable. Indeed, for any V G VJ we have Av(r,t)
r
= lim / k—tao
V{s,Xs)ds,
and hence Av(r, t) is ./^-measurable. It follows from the properties of the functionals Ay mentioned above that the functional J4M defined by (3.172) satisfies conditions (l)-(6) in Theorem 3.14. This completes the proof of Theorem 3.14. • The next assertion concerns the uniqueness of the functional A^. Lemma 3.25 Let fi be a non-negative time-dependent measure from the class V^, and let A\ and A% be two functionals satisfying conditions 1-5 in Theorem 3.14- Then for every 0 < T
It is not hard to see, using the conditions in Theorem 3.14, that ET,x[A1{T,t)-A2(r,t)]2 2
ft
= 2 J ] (-ly+'E^ / 2
[Ai(T,t) - Mr, s)] dAj(r, s) ,t
= 2 J ] ( - l ) t + J E T i I / Ms,t)d^j(T,s).
(3.173)
Non-Autonomous
267
Kato Classes of Measures
By condition (1) in Theorem 3.14, the random variable Ai(t,s) is T^measurable for i = 1,2. Next, using the Markov property in (3.173), we obtain ETiX[A1(T,t)-A2(T,t)}2 2
,i
= 2^(-l)i%r,
1
2
/ EZtSAi(s,t)
\z=x
dAj(T,s)
,t
= 2 J2 (-1)4+JET,X / 7VM (s^XjdAfas)
= 0.
(3.174)
It follows from (3.174) that for given r and x, the process A2 is a modification of A\. Since both processes are continuous, they are indistinguishable. This completes the proof of Lemma 3.25. • Definition 3.22 Let \i e P^, and denote by fx+ and JJ,~ the positive and the negative variation of the time-dependent measure n, respectively. Then the functional A^ is defined as follows: A^T,
t) = A^+ (r, t) - Ap- (r, t).
Lemma 3.26 Let P e PM. be a transition probability function possessing a density p, and let fi € P^. Then estimates (3.163) and (3.165) hold with fj, instead of V. Proof. We will prove estimate (3.163) for a time-dependent measure //. It is clear that estimate (3.165) for /i follows from estimate (3.163). Let (i G Pm, and let Vk be the sequence constructed for /u in Lemma 3.21. Then Vk € P%. Applying estimate (3.163) to the sequence Vk, we see that supE T , x X&E
\AVk(r,t)\n
sup
+ 5)n~lM(Vk)(T,T
+ 5).
t:T
(3.175) It follows from Remark 3.6 that limsupM(|V fe |)(T,T + J)
(3.176)
k—+oo
and M(Vk)(T,T
+ 5)^M(n)(T,T
+ 5)
(3.177)
as k -> oo. Now using (3.169), (3.172), (3.175), (3.176), and (3.177), we see that (3.163) holds for fi.
268
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
This completes the proof of Lemma 3.26.
•
Next, we turn our attention to additive functionals corresponding to time-dependent measures from non-autonomous Kato classes. The timereversal and the fact that if P is a backward transition probability function, then P(T,X;
t,B) = P{T - t, B;T-T,X)
(3.178)
is a transition probability function will be used in the definition of the additive functional. Recall that the following notation was used for the time reversed objects: Ff'y = FT-t,v, X* = XT-t, and f\ = T%Z\, where the forward objects correspond to the transition probability function P, while the backward ones are associated with the backward transition probability function P (see Section 1.4). We will say that a transition probability function P belongs to the class VM. if there exists a backward Markov process Xf that is progressively measurable with respect to the filtration T\. and has P as its backward transition function. It is not hard to see that the process Xt is .T-J-progressively measurable if and only if the time reversed process X1 is ^-progressively measurable. Therefore, p g VM <=> P G VM. For a function V eVf,
the functional Ay is defined by
AV(T, t)=
f V (s, Xs) ds.
(3.179)
P t,v -a.s. Now let /J, G Vm and denote by Ji the time-reversal of fj, given by Jl(t) = n(T — t). Then the functional A^ is defined as follows: A^(T,t)=i4p(T-t,r-T).
(3.180)
The functional A^ satisfies the following conditions: (1) For all T < t, the random variable A^T, t) is ^"'-measurable. (2) For all t and y G E, A^t, t) = 0 P'^-a.s. (3) For all r < t and y G E, the process r H-> A M (T, t), 0 < r < t, is non-increasing and continuous P t,y -a.s. (4) For all r < A < t, A^T, t) = A^(r, A) + A^X, t) P'^-a.s. (5) For all r < t and y G E, E^A^T,
t) = iV(/i)(r, t, y).
Non-Autonomous
Kato Classes of Measures
269
Note that the free propagator U(t, r ) associated with P is related to the backward free propagator Y(T, t) corresponding to P by the formula Y(T-t,T-r)
= U(t,T)
(see formula (2.11) in Section 2.2).
3.10
Exponential Estimates for Additive Functionals
In this section we study the multiplicative functionals exp{— J V(s, Xs)}ds and exp{—AM(t, r)}. These functionals are called the Kac functionals. The next result is a generalization of Khas'minski's Lemma to non-autonomous Kato classes. The exponential estimate in Lemma 3.27 is a useful result. For instance, Lemma 3.27 implies that the backward Feynman-Kac propagators Yy and VM with V € VJ and /i € Vi^ are uniformly bounded on the space U^ (see Theorem 4.2 below). For the proof of Khas'minski's lemma in the autonomous case see [Simon (1982)]. Lemma 3.27
The following assertions hold:
(a) Let P e VM., V € VJ, and let the numbers r and t with 0 < r < t < T be such that M ( | V | ) ( T , £ ) < 1. Then %P*r*<«p{f'\V(s,X.)\ds} <
x
_
M
'
m M
-
(3-181)
(b) Suppose that P € VM has a density p. Let \i € V"^, and let the numbers T and t with 0
(3-182)
It follows from estimate (3.160) that ETtXAlVli^t)n
< M(\V\)(T,t).
(3.183)
It is not hard to see that (3.183) implies (3.181). Moreover, reasoning as in the proof of Lemma 3.26, we see that (3.181) implies (3.182). D The next lemmas contain more exponential estimates: Lemma 3.28
The following assertions hold:
270
Non-Autonomous Koto Classes and Feynman-Kac Propagators
(a) Let P G VM, V G V*j, and let the numbers r and t with 0 < r
{2N(\V\)(T,t,x)M(V)(r,t)}?
2^3 ~3
N(\V\)(r,t,x)M(V)(T,t) l-M(\V\)(r,t) •
(3J84)
(b) Suppose that P £ VM has a density p. Let fx £ V^, and let the numbers r and t with 0
\/(m-l)!(m + l)! < -~(m!) for all m > 3. The proof of part (b) is similar. Here we reason as in the proof of Lemma 3.26. • Lemma 3.29
The following assertions hold:
(a) Let P G VM, q > 1, 1 < r < oo, - + - = 1, V € V), and W € V). Let T and t with 0
- exp
{Aw(T,t)}\q
(l-M{rq\W\)(r,t))r x i [2N (r'q\V -W\)(T,t,x)M
+
(r'q(V
-W))(T,t)]?
2 ^ 3 N (r'q\V - W\) (r, t, x)M {r'q{V - W)) (r, t) \ * 3 l-M(r'q\V-W\)(T,t) j '
(b) Suppose that P G VM
[6
'
has a density p, and let q > 1, 1 < r < oo,
- + — = 1, \i G Vm, and v G V^. If the numbers r and t with 0 < r < t
are such that M(rq\v\)(r,t)
ET,X |exp {A^T,
t)} - exp {AV(T,
< 1 and M(r'q\fi - v\) (r,t) < 1,
t)}\q
N'on-Autonomous Koto Classes of Measures
(1 - M(rq\v\)(T,t))T
271
[
2V3 N (r^l/i - v\) (T, t, x)M (r'qfr - u)) (r, f) ) * +
Proof. &> 1,
3
l-M(r'^-«/|)(r,0
J '
^ ^
It is not hard to see that for all a and 6 with —oo < a < oo and |ea_l|h<eM<>_l.
(3.187)
By (3.187) and Holder's inequality, ET,x |exp {AV(T,
t)} - exp {AW{T,
< {ET,xexp{ArgW(T,t)}}r
t)}\q
{KriX\exp{AV-W(T,t)}
< {ET)Xexp{Arqm{T,t)}}7
- lf'9}^ - l}77 . (3.188)
{ET,xexp {\Ar,q{v_w)(T,t)\}
Now it is clear that (3.185) follows from (3.181), (3.184) and (3.188). This completes the proof of part (a) of Lemma 3.29. Next we will prove part (b) of Lemma 3.29. By Lemma 3.21, there exist the approximating sequences gk (for fi) and hk (for i/). It follows from the assumptions in part (b) of Lemma 3.29 that M(rq\u\)(T, t) < 1 and M {r'q\n — v\) (r, t) < 1. Therefore, there exists a sequence {&'} of positive integers such that M(rq \hk.\) (r,t) < 1, M (r'q \gk, - hk>\) (r,t) < 1, lim A9k,
(T, t)
k —*oo
= An (r, t), fc'
lim AK, (r, t) = Av —»-oo
limsupM(|^|)(r,f)<M(|/i|)(r,t), fc'—too
limsupM(|0 f c '-/i f e <|)(r,t) < M(|/x - i/|)(r,f), fc'—>oo
limsupiV(|g fc , - hk>\) (r, t,x) < iV(|/u - i/|)(r, t,a;), fc'—*oo
and lim M (5fc, - /ifcO (r, t) = Af (/x - v){r, t). k' —*OQ
(T,
t),
272
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
It follows from (3.185) that ET,X |exp {A9k,(r, t)} - exp {Ahk,(r, 1
<
t)}\"
r
(l-M(rq\hk'\)(T,t))' <[2N(r,q\gk,-hk.\){T,t,x)M{r'q(ak>-hk.))(T,t)]i
+
2V3N(r'g\gk. 3
•
- hk.\) (r,t,x)M (r'g(gk, - hk.)) (r,t) \ * l-M(r>q\9k,-hk,\)(T,t) j •
{
•°
j
Now using Fatou's Lemma in (3.189), we see that estimate (3.186) holds. This completes the proof of Lemma 3.29. D It is not hard to see from formula (3.164) that c„ < c2"n!. However, this inequality does not allow us to get an exponential estimate for the functional Ay from inequalities (3.163) and (3.165). Next, we will obtain such an estimate by modifying the proof of Lemma 3.24. Lemma 3.30
The following assertions hold:
(a) Let P £ VM. and V e V%. Then for every r with 0 < r < T and every 6 > 0 such that r + 5 < T and M(\V\)(T,T + 6) < 1, the following estimate holds: supE T j X exp< X&E
'
sup
< exp {M(V)(T, M(\V\)(T,T +C
|Ay(T,t)\ >
[t-.T
J
T + 6)} (l
+ c{M(\V\)(T,
+ 5)M(V)(T,T
l-M(\V\)(r,r
+
T + S)M(V)(T,
T +
6)}?
6)\
+ S)
^i9Uj
) '
(b) Suppose that P £ VM. has a density p, and let fi £ V^. Then for every r with 0 < r < T and every 8 > 0 such that r + 8 < T and M(\n\){r, T + 5) < 1, the following estimate holds: sup E r , x exp \ x€E
sup
|A M (r,f)|i'
{t:T
)
< exp {M(/i)(r, r + 5)} I 1 + c{M(|/i|)(r, M(\H\)(T,T
+c-
+ 5)M(H)(T,T
1 - M ( | M | ) ( T , T + J)
+ 6Y
r + 5)M{H){T,
T + <5)}'
Non-Autonomous Koto Classes of Measures
273
Proof. In the proof of Lemma 3.30, we will use the same notation as in the proof of Lemma 3.24. By Doob's inequality (see Subsection 5.6), for every n > 2, ET,X
sup \Mt,\n <(-?—) j:0<j
n
\
_
i
ET,x\Av(r,r
+ 6)\n
/
n\M{\V\){T,T + 8)n-lM{V){T,T
< (^~[)
+ 5).
Next, dividing the previous inequality by n\, adding up the resulting inequalities, and using (3.163) and the equality Mt = AV(T, t) + N(V) (t, t + 6, Xt) we get E r > x exp{
sup \Mt\}
[j:0<j
J
sup lMtA+cM(\V\)(r,r + S)M(V)(r,r + 5) 3-.o
sup sup | A y (\Ar v,(T,tj)\+E t , ) | + E r TtX ,x j:0<j
+c
M{\V\){T,T
sup \N(V)(tj,T
+ 5,Xtj)
j-Xj'-0<j
1-M(\V\)(T,T
+ 5)
+ 5)
I2
< l + { E r , x sup [ j--0<j
. + ^ +
\Av(T,tj)\ 2 I J M(\V\)(T,T
+ M(V)(r, r + S) + c
< 1 + {c 2 M(| V | ) ( T ,
T
+ 6)M(V)(T,T
1_Mm){TtT
+ 5)M(V){T,
T
+ 6)
+ S)
+ 5)}i
We also have E Tj:c exp I sup \MtA > [j:0<j
sup \Av(T,tj)\[j:0<j
sup \N(V) (tj,T+ 3-0<j
S,Xtj)\
274
Non-Autonomous Kato Classes and Feynman-Kac Propagators
> e x p { - M ( l / ) ( r , r + <5)}ET,xexp^
sup \AV (T,tj)\)
.
(3.192)
j:0<j
It follows from (3.191) and (3.192) that
E T)X exp<
sup
\Av(T,tj)\
[j:0<j
<exp{M(V)(T,T +M(V)(T
+ 6)} 1 + {C2M(\V\)(T,
T + 5) + cM(W\)(r,r
+
T + S)M(V)(T,
S)M(V)(r,r
+
T + 6)}-
S)
Therefore,
E TjX exp< sup |Av (v, tj)| > [j:0<j
T + 6)} (l + c{M(\V\)(T, +
1-M(\V\)(T,T
T + 5)M(V)(T,
T + 5)}?
6)M(V)(T,T+5)\ + 8)
J'
[d 196)
-
Finally, consider a sequence of successive refinements of the partition r = t0 < t\ < • • • < tk = T + 5 of the interval [r, r + 6] such that the maximum length of the partition intervals tends to 0. It follows from the monotone convergence theorem and the continuity of the functional AV(T, t) that one can pass to the limit in (3.193). This establishes estimate (3.190). The proof of part (b) of Lemma 3.30 is similar. Here we reason as in the proof of part (b) of Lemma 3.29. • Let P be a backward transition probability function. Recall that for a Borel function V on [ 0 , T ] x £ w e denoted by Ay the additive functional associated with V (see formula (3.179)). Similarly, for a time-dependent measure /x we denoted by A^ the additive functional corresponding to /x (see formula (3.179)). It is not difficult to see that all the results for the functionals Ay and A^ obtained in this section can be reformulated for the functionals Ay and A^.
Non-Autonomous
3.11
Kato Classes of Measures
275
Probabilistic Description of N o n - A u t o n o m o u s K a t o Classes
A probabilistic characterization of non-autonomous Kato classes can be obtained from property (5) of the functionals Ay, -AM, Ay, and A^ in Definitions 3.4 and 3.5. For all T < t and x G E, we have ET,xAv(r,t)
&xAv(T,t)
= N(V)(T,t,x),
E r > x A M (r,i) = N(n)(r,t,x),
and E^A^t)
=
N(V)(r,t,x), =
N(ji)(r,t,x).
It follows that the following lemmas hold. Lemma 3.31
Let P be a transition probability function. V G V*f 4=>
sup
Then
sup E r xA\v\(r, t) < oo.
0
LetV eV}.
Then V G V) «=>
lim t—r->
sup ET XA\V\ (r, t) = 0. 0+xeE
Suppose that the transition probability function P has a density p. Then fj, G Vm <=>
sup
sup ETtXA\,j.\ (r, t) < oo.
0
Let fie Vm.
Then
l * e P ; < ^ , lim sup E TiX ^| M | (r, t) = 0. Lemma 3.32 Let P be a backward transition probability function. V G Vf <==>• sup supE*' x ^|v|(r, t) < oo. 0
LetV eVf.
Then
xeE
Then V G Vf 4=>
lim t~T^0+
sup E*' x li V | (r, f) = 0. xeE
Suppose that the backward transition probability function P has a density p. Then lieVm<=^>
sup
s u p E ^ A i ^ r , * ) < oo.
0
276
Non-Autonomous
Let fj, £ Vm.
Kato Classes and Feynman-Kac
Propagators
Then H^Vm
<*=>
lim
sup E*'XA|M| (T, t) = 0.
t-r^0+x6E
3.12
Notes and Comments
(a) The reader, interested in additive and multiplicative functionals of time-homogeneous Markov processes, may consult [Blumenthal and Getoor (1968); Revuz and Yor (1991)]. The additive functional AM considered in Sections 3.1 and 3.9 was studied in [Gulisashvili (2004b); Gulisashvili (2004c)]. It is known that under certain restrictions on measures and processes there is a correspondence between additive functionals and measures (the Revuz correspondence, see [Revuz and Yor (1991); Beznea and Boboc (2000); Beznea and Boboc (2004)]). For more information on additive and multiplicative functionals associated with measures and the corresponding Schrodinger semigroups see [Blanchard and Ma (1990a); Blanchard and Ma (1990b); Albeverio, Blanchard, and Ma; Albeverio and Ma (1991); Glover, Rao, and Song (1993); Glover, Rao, Sikic, and Song (1994)]. (b) The exponential estimates for the additive functionals Ay and A^ (see Section 3.10) were obtained in [Gulisashvili (2004b); Gulisashvili (2004c)]. The approximation in the potential sense discussed in Section 3.9 was denned in [Gulisashvili (2004b); Gulisashvili (2004c)]. (c) Non-autonomous Kato classes were studied in [Sturm (1994); Qi Zhang (1996); Qi Zhang (1997); Schnaubelt and Voigt (1999); Nagasawa (2000); Rabiger, Rhandi, Schnaubelt, and Voigt (2000); Gulisashvili (2002a); Gulisashvili (2002b); Gulisashvili (2004a); Gulisashvili (2004b); Gulisashvili (2004c); Gulisashvili (2005)]. Our presentation of the non-autonomous Kato classes follows [Gulisashvili (2004b); Gulisashvili (2004c)]. (d) For more information on fundamental solutions of parabolic initial and final value problems in the case of differential operators in nondivergence form see [Dressel (1940); Dressel (1946); Friedman (1964); Il'in, Kalashnikov, and Oleinik (1962); Ladyzenskaja, Solonnikov, and Ural'ceva (1968); Porper and Eidel'man (1984); Eidel'man (1969); Eidel'man and Zhitarashu (1998)]. (e) For the results concerning weak fundamental solutions and the Gaussian estimates in the case of differential operators in divergence form
Non-Autonomous Koto Classes of Measures
(f) (g)
(h)
(i)
(j)
277
see [Nash (1958); Aronson (1967); Aronson (1968); Fabes (1993); Fabes and Stroock (1986); Liskevich and Semenov (2000); Porper and Eidel'man (1984); Semenov (1999); Qi Zhang (1996); Qi Zhang (1997)]. The relations between fundamental solutions and transition densities are discussed in [Porper and Eidel'man (1984)]. There exists a rich literature on stochastic integrals and stochastic differential equations, for example, [Friedman (1975); Friedman (1976); Chung and Williams (1990); Ikeda and Watanabe (1989); Metivier and Pellaumail (1980); Metivier, M. and Viot, M. (1987); Lamberton and Lapeyre (1996); 0ksendal (1998); Protter (2005)]. More information on the Ito formula can be found in [Bhattacharya and Waymire (1990); Durrett (1996); Chung and Williams (1990); Ikeda and Watanabe (1989); Karatzas and Shreve (1991); Lamberton and Lapeyre (1996); Revuz and Yor (1991); Stroock (2003)]. The martingale problem was introduced and studied by Stroock and Varadhan (see [Stroock and Varadhan (1979)]). Liggett used the martingale problem to study interacting particle systems (see [Liggett (2005)]). Diffusion processes are discussed in [ito and McKean (1965); Dynkin (2002); Eberle (1999); Revuz and Yor (1991); Ikeda and Watanabe (1989); Nagasawa (1993); Rogers and Williams (2000a); Rogers and Williams (2000b); Stroock (1987); Stroock and Varadhan (1979); Carlen (1984)].
This page is intentionally left blank
Chapter 4
Feynman-Kac Propagators
4.1
Schrodinger Semigroups with Kato Class Potentials
Feynman-Kac propagators are two-parameter analogues of Schrodinger semigroups. In this section we gather several known results concerning Schrodinger semigroups with Kato class potentials. Some of these results are formulated without proof. We refer the reader to [Simon (1982)] for more information on Schrodinger semigroups. Let V be a Lebesgue measurable function on R d , and consider the following initial value problem for the perturbed heat equation:
* f M _ | A « ( t > * ) + V(*)«(t>*) = 0 U(T,X)
=
(41)
g(x),
where 0 < r < t < T and x € Rd. In (4.1), g is a Lebesgue measurable function on R d , and the symbol A stands for the Laplace operator defined d d2 by A = 2_. ~^~2 • I t 1 S known that under certain restrictions on the potential function V, there exists the Schrodinger semigroup S(t) = e'tH,
t > 0,
where H is the Schrodinger operator given by H = — ^A + V. The Schrodinger semigroup is bounded on the space Lp with 1 < p < oo. Moreover, the function u defined by u(t, x) = S(t — r)g(x) is a solution to initial value problem (4.1). It is also known that the Feynman-Kac formula holds 279
280
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
for the Schrodinger semigroup, that is, u(t,x) = e-(t-T)Hg(x)=Exg(Bt-T)expl.-
J
V(B.)
(4.2)
where Bt is Brownian motion. One can also consider the following final value problem for the perturbed backward heat equation: ^
^
+ l-Au(r,x) - V(x)u(r,x)
=0
u(t,x) = f(x) where 0
J
V(B,)d*j.
(4.4)
If V = 0 in (4.1) and (4.3), then we get the classical heat equation and the backward heat equation. In this case we have u(t,x) = Exg(Bt-T)
(4.5)
and u(T,i)=EI/(Bt_T).
(4.6)
Formulae (4.5) and (4.6) can be derived from the following two facts: (1) The equality —— = -Agd holds for the Gaussian density gs.. (2) The formula
P x [Bt £ A] = J gd{t, y)dy, t>0,
AeBud,
holds for one-dimensional distributions of Brownian motion. Recall that we discussed Brownian motion in Subsection 1.12.1. Definition 4.1 For 1 < p < oo, the local uniform L p -space L\oc of all Lebesgue measurable functions V on E d for which sup / xeRJy.\x-y\
\V{y)\pdy < oo.
consists
Feynman-Kac
Propagators
281
Definition 4.2 A locally integrable function V on the space Rd with d > 3 belongs to the Kato class K4, provided that . \V(y}}
lira sup / a
9dy
= 0.
i°xm"Jy.\x-v\
If d = 2, then the Kato condition is lim sup / a
WxeMdJy.\x-y\
\V(y)\m
1
\&y — 0-
F ~ V\
For d = 1, the Kato class coincides with the space L\oc u. It is clear from Definition 4.2 that the Kato condition utilizes the convolution of the function \V\ with the truncated Riesz potential kernel. For d > 3, the Kato class Kj, equipped with the norm
Kd = sup d zeR
\v(y)\ _ Jn T^^hdy,
f
x
Jy.\x-y\
is a Banach space (see [Voigt (1986)]). Similar result holds for the classes K\ and KiThe next assertion provides several equivalent characterizations of the Kato class. They are expressed in terms of Brownian motion and the heat semigroup eitAf(x)
= f f(y)gd(t, x - y)dy = Ex [/ (Bt)}.
(4.7)
JRd
Theorem 4.1 Let V be a locally integrable function on E d . following are equivalent:
Then the
(1) The function V belongs to the Kato class Kd(2) The equality lim sup Ex / \V{Bs)\ds
=0
(4.8)
lim sup / eisA \V\ (x)ds = 0
(4.9)
Uo
xSKd
Jo
holds. (3)
The
equality
*J-° x€Rd Jo holds.
282
Non-Autonomous Kato Classes and Feynman-Kac Propagators
Proof. The equivalence of conditions (2) and (3) in Theorem 4.1 follows from (4.7). Next, we will prove the equivalence of conditions (1) and (3) for d > 3. The cases where d = 1 and d = 2 are similar. Let us compare the truncated Riesz potential kernel —
' _ „ — and
the kernel eXP Jo (2ns)i (2TTS)* {
I
dS a d 2s] ^\x\ 27rf -2\x\ ~'-*J\x J\x\ym
2s J
u2
2
e
u
du,
appearing in condition (3). We will show that the following estimates hold: 1
fc
1
!TT* \x\ad~* ~2 7|x|V(2t) L 27rf
^~2e-udu
< j^-2,
(4.10)
^ r ^ - 2 f «*-ae-d« > ^ f ^ , d 2 27ri | z | " 7|xp/(2t) '~'^2
(4.11)
and 1
/>00
1
2iri \Ad
2
/ w^- 2 e-"du < c^e-Al x|2 2 J\x\*/(2t) JM /(2t)
for |x| > yft.
(4.12)
Inequality (4.10) is obvious since d > 2. Inequality (4.11) is a consequence of the inequality /•OO
ui'2e~udu
/
> 0.
Finally, inequality (4.12) can be obtained from the estimate /'OO
/
/»00
J-2e~udu<e-ia
Ja
ui-2e-^udu,
a > 0.
Ja
It is not hard to see that the equivalence of conditions (1) and (3) in Theorem 4.1 can be derived from estimates (4.10)-(4.12). This completes the proof of Theorem 4.1. • Next, we will formulate several assertions concerning the Kato class and Schrodinger semigroups. The following inclusions hold for the Kato class Kd: •kfocu
CK
d
Cilocu,
P>
r-
Feynman-Kac
Propagators
283
The Schrodinger semigroup e~iH with V G Kj is bounded on the space IP with 1 < p < oo and strongly continuous on the space i p with 1 < p < oo. It is (L p -L 9 )-smoothing, that is, e-tH
-LV->Lq
for all 1 < p < q < oo and t > 0.
Moreover, e~tH is a strongly continuous semigroup on the spaces Co and BUC. The Kato class can be defined for more general homogeneous Markov processes (see Sections 3.1 and 3.2 in [Chung and Zhao (1995)]). Definition 4.3 below is based on the probabilistic characterization (4.8) of the Kato class KdDefinition 4.3 Let P G VM. be a time-homogeneous transition probability function, and let (Xt, Ft^x) be a corresponding progressively measurable time-homogeneous Markov process on (fi,^-") with state space (E,£). Then it is said that a Borel function V on E belongs to the Kato class J if limsupE x / \V(Xs)\ds *loxeB J0
= 0.
Let Xt be a time-homogeneous Markov process. Then the semigroup associated with Xt is denned by
S(t)f(x)=Exf(Xt). Chung and Zhao studied the semigroup S(t) and its perturbations by functions from the Kato class J (see [Chung and Zhao (1995)]).
4.2
Feynman-Kac Propagators
Let P € VM. be a transition probability function, and let Xt be a corresponding progressively measurable Markov process. Recall that for a function V G V}, the Kac functional My is defined as follows: Mv(r,i)=exp|- f
V(s,Xs)ds\
where 0 < r < t < T. The progressive measurability of the process Xt implies the J^-measurability of the random variable MV(T, t). UP has a density p, and \i G V^ is a time-dependent measure, then the Kac functional
284
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
MM is given by M^T,
t) = exp {-A^T,
*)} ,
(4.13)
where A^ is the additive functional associated with \i (see Theorem (3.14)). Now we are finally ready to define backward Feynman-Kac propagators. Recall that the class VM of transition functions was introduced in Definition 3.20. Definition 4.4 (a) Let P G VM be a transition probability function, and let V G V*tThen the family of operators given by YV{T, t)g(x) = ETiXg(Xt) e x p { - J
V {s, X3) ds},
0
is called the backward Feynman-Kac propagator associated with the transition probability function P and the function V. (b) Let P G VM be a transition probability function possessing a density p, and let /x G Vi^. Then the family of operators defined by Y„(T,t)g(x) = £; T , x 5 (Xt)exp{-i4 M (T,t)}, 0 < r < t < T, is called the backward Feynman-Kac propagator associated with the transition probability function P and the time-dependent measure \x. Our next goal is to introduce forward Feynman-Kac propagators. Here we need the definitions from Sections 1.4 and 3.3. For a backward transition probability function P G VM, a progressively measurable backward Markov process X1 associated with P, and a function V G Vf, the functional Ay is defined by Av(r,t)
= J
v(s,xAds
P'^-a.s. In the case of a time-dependent measure /i G Vm, the functional Ap is given by AIM(T,t) =
An(T-t,T-T).
Here Jl is the time-reversal of the time-dependent measure /x, defined by Jl(t) = fi(T - t) (see Section 3.9).
Feynman-Kac
Propagators
285
Definition 4.5 (a) Let P G VM be a backward transition probability function, and let V G Vf. Then the family of operators given by Uv(t, r)g{x) = E**g ( x t ) e x p { - f V (s, Xs"j ds},
0
is called the Feynman-Kac propagator associated with the backward transition probability function P and the function V. (b) Let P G VM be a backward transition probability function possessing a density p, and let /x G Vm. Then the family of operators defined by £/ M (T, t)g(x) = E^g
( x r ) exp {-!„(*, r ) } , 0 < r < t < T,
is called the Feynman-Kac propagator associated with the backward transition probability function P and the time-dependent measure \x. 4.3
The Behavior of Feynman-Kac Propagators in Spaces
Lp-
The backward Feynman-Kac propagators Yy and Y^ with V G V*t and H G V^ inherit various properties of the free backward propagator Y. In this section and in Section 4.4, we study the relations between Y, Yy, and Y^. We follow [Gulisashvili (2004b); Gulisashvili (2004c)] in our discussion of the inheritance problem. The first result in the present section concerns the L°°-boundedness of backward propagators. Theorem 4.2
Let P G VM.
Then the following assertions hold:
(a) For any V G Vf, Yy is a backward propagator on Lf. (b) If P has a density p, then for every V £ Vf, Yy is a backward propagator on L°°. (c) If P has a density p, then for every /J, G V^, Y"M is a backward propagator on L°°. Proof. Let g G Lf. Then it is not hard to prove that the function Yy(r, t)g is Borel measurable. By Lemma 3.27, sup \Yv(T,t)g(x)\
< llsll^supE^expf.VitT,*)} < H g ^ ^ _
Mny\)(T,t)
286
Non-Autonomous Kato Classes and Feynman-Kac Propagators
for all r, t with Af (| V | ) ( T -t)
+ l\\
,
where 5 > 0 is any number such that p(S) = sup {M(\V\)(n, A) : A -
V
< 5} < 1
and a — In
— . Similar estimates hold under the conditions in parts 1 - p(6) (a) and (c) of Theorem 4.2.
The next result concerns approximations in the potential sense and the uniform convergence of Feynman-Kac propagators. Theorem 4.3 Let P £ VM. and V £ V*.. Suppose that a sequence Vk EV*, k > 1, approaches V in the potential sense. Then lim
sup
^°°
(T,t):0
k
\\Yv(T,t)-YVk(T,t)\\Lr^Lr>=0. £
Suppose that P G VM. has a density p and let \i £ V^- Suppose also that a sequence \Xk £ V^, k>\, approaches \i in the potential sense. Then lim k
~*°°
sup
\\Y^T,t)-Yllk(T,t)\\oo^oo=0.
(r,t):0
Proof. We will prove only the second part of Theorem 4.3. The proof of the first part is similar. Let /i and Hk be such as in the formulation of Theorem 4.3, and let / € L°°. Then, by part (b) of Lemma 3.29 with q = 1 and r = 2, there exists 5Q > 0 such that \Y^T,t)f{x)-Y»k{T,t)f{x)\ < all/Hoc ({M (/x - /ifc) (r, t)}l+M(ji-
/*) (r, t))
(4.14)
Feynman-Kac Propagators
287
for all r and t with t — r < So and all x € E. In (4.14), the constant a does not depend on x, t, r, and k. It follows from (4.14) that lim fc
-*°°
\\Y/1(T,t)-Ylik(T,t)\\L-^L<~=0-
sup (r,t):0
( 4 - 15 ) £
Next we will get rid of the restriction t — r < So in (4.15). Consider a partition 0 — to < h < ^2 < • • • < tn = T of the interval [0, T] such that tj+i — tj < So for all j with 0 < j < n — 1. Then the estimate in (4.14) holds, provided that t and r belong to the same interval [t,,i,+i], 0 < j < n — 1. Next, we will show how to complete the proof of the second part of Theorem 4.3, using the previous assertion and the properties of backward propagators. We will consider only a special case, where tj < T < tj+\ < t < tj+2 with 0 < j < n — 2, and leave the rest as an exercise for the reader. By the uniform boundedness of the propagators Y^, k > 1, on the space L°° (this follows from Remark 4.1), we have
< \\Y» (T,t j + i)y„(tj+ut) -yw
(r,ij+1)yMt
fe+i,*)!^^
< \\Y» (r, tj+i) Y^ (tj+1, t) - yM (r, ti+1) Y^ (tj+1, + ||KM (r, tj+i) Yn (tj+x,t)
t ) ^ ^
- Y^ (r, tj+1) Y^ (tj+l,
t ) ^ ^ ,
- Y^ (T,tj+1)\\00^00
.
Since t — tj+i < So and tj+i—r < So, we can apply the special case of (4.15) that has already been established. It follows that equality (4.15) holds for all tj
fc_>0
°
Let P 6 VM, V G V}, and suppose that Vk is defined by sup (T,t):0
\\Yv(T,t)-YVk{T,t)\\L?,_Lr
£
£
= 0.
Let P e VM, H € Vm> and suppose that Vk is defined by (3.152). Suppose also that P has a density p. Then lim fc
sup
\\YM(r, t) - YVk (T, t ) ^ ^ = 0.
-*°°(r,t):0
Since the sequence Vk defined by (3.148) approaches V in the potential sence and, similarly, the sequence Vk defined by (3.152) approaches /x in
288
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
the potential sense (see Lemmas 3.20 and 3.21), Corollary 4.1 follows from Theorem 4.3. The next lemma will be important in the sequel. Lemma 4.1
Let P £ VM-
Then the following assertions hold:
(a) For any V £ V), UrnJYv(T,t)-Y(T,t)\\Lr^L
=0.
(4.16)
(b) If P has a density p, then for every V £ V**, lim ||>V(r,i) -Y(r,t)\\oo^oo=0.
(4-17)
t — TJ.U
fcj //"P has a density p, then for every \i G V^} tlimo||yM(T,t)-y(r,t)||00_too
= 0.
(4.18)
Proof. It follows from part (a) of Lemma 3.27 and from the definition of the class VJ that limsup||y v (T,t) - y(T,OH^^oo < limsupsup (E T , x exp {A\V\{T,t)\ t-ri.0
t—rj.0
- l)
x&B
M(\V\)(T,t) u < hm sup sup ,y\' ' , = 0. " t-na xeEl-M(\V\)(T,t) This implies equality (4.17). The proof of (4.16) and (4.18) is similar. In the proof of (4.18), we use part (b) of Lemma 3.27. • The next result provides sufficient conditions for the boundedness of backward Feynman-Kac propagators on the space Ls. Theorem 4.4 Suppose that 1 < s < oo, 1 < r < s, and P £ VM. the following assertions hold:
Then
(a) Let V £ Vj, and suppose that the free backward propagator Y satisfies the condition Y(r,t) £ L(Lr£,Lr£) for all 0 < r < t < T. Then Yv is a backward propagator on L£. If, in addition, Y is uniformly bounded on L£ and strongly continuous on L£, then Yy is a strongly continuous backward propagator on L£. (b) If P has a density p, and if Y(T, t) £ L(Lr, Lr) for all 0 < r
Feynman-Kac
289
Propagators
(c) Suppose that P has a density p, and let \i € V^. IfY(r, t) € L (Lr, Lr) for all 0 < r < t < T, then Y^ is a backward propagator on L3. If, in addition, Y is uniformly bounded on U and strongly continuous on Ls, then Y^ is a strongly continuous backward propagator on Ls. Remark 4.2 We do not know whether Theorem 4.4 holds for r = s. In the case of the heat semigroup, Theorem 4.4 may fail for s = 1. This was established in [Gulisashvili (2005)]. Proof. We will prove only part (b) of Theorem 4.4. Parts (a) and (c) are similar. Assume that the conditions in part (b) hold, and let g € Ls. By Holder's inequality, 3—r
\Yv(r,t)g(x)\ < {ET,x\g(Xt)\ry U^exp i-L-A{vl(T,t)\\
'
i
=
{Y(T,t)\g\i(x)}i{Y1±m(T,t)l(x)y
< Y^lvl(T,t)\
s
{Y(r,t)\g\Hx)y
(4.19)
for all t and r with 0 < r < t < T. It follows from Remark 4.1 that y_*_|v(T,*) *
r
(4.20)
oo—>oo
where c > 1 depends on s, r, and V. Next (4.19) and (4.20) give \Yv(T,t)g(x)\
(4.21)
Now we see that (4.21) implies \\Yv(T,t)\\s^s
(4-22)
Therefore, Yy is a propagator on Ls. Remark 4.3 Inequality (4.22) provides an estimate for the norm of the backward Feynman-Kac propagator on the space Ls. Let us return to the proof of part (b) of Theorem 4.4. The next result will be needed in the proof. Theorem 4.5 Let l<s
l
and P € VM. Then the
290
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
(a) Let V G V*f, and suppose that the free backward propagator Y is uniformly bounded on Lr£. Then limJYv(T,t)-Y(T,t)\\L,
L,=0.
t — T[0
E
(4.23)
£
(b) Suppose that P has a density p, and Y is uniformly bounded on U'. Then for any V GP}, ]im\\Yv(T,t)-Y{T,t)\\s~s
= 0-
(4.24)
t —TJ.0
(c) Suppose that P has a density p, and Y is uniformly bounded on U Then for any [i £ V^, lim | | Y M ( T , i ) - Y ( T , 0 | | ^ s = 0.
(4.25)
t~ TlU
Proof. We begin with the proof of part (b) of Theorem 4.5. It follows from part (b) of Theorem 4.4 that Yy is a backward propagator on the space Ls. Let g £ Ls. By Holder's inequality and inequality (3.187), we obtain \Yv(T,t)g(x)-Y(T,t)g(x)\
< JEr>x \g (Xt)\"Y <{y(r,t)\g\
Hx)y
\ET>X |exp{Av(r,i)} - 1| — } ~ | E T I X exp
Am{r,t))-l s—r
(4.26)
It follows from (4.26), part (a) of Lemma 3.27, the definition of the class Vj-, and the uniform boundedness of Y on 17 that lim sup 1 1 ^ ( 7 " , * ) - ^ ^ t—r|0
< a(s, r, V) lim sup sup <ETX exp t-rj.0 x€E I
< c(s,r,V) lim sup t-riO
'
A\V\(T,t)\-l [S — r
l^M(\V\)(r,t) *-M(\V\)(T,t)
= 0.
This implies equality (4.24). Equalities (4.23) and (4.25) can be obtained similarly. Part (b) of Lemma 3.27 is used in the proof of equality (4.25) instead of part (a) of Lemma 3.27. •
Feynman-Kac Propagators
291
Let us continue the proof of part (b) of Theorem 4.4. Suppose that the free backward propagator Y is locally uniformly bounded on U and strongly continuous on U. We have already shown above that Yv is a backward propagator on Ls. Moreover, Ys is uniformly bounded (see estimate (4.22)). Therefore, in order to prove the strong continuity of Yy, it suffices to establish the separate strong continuity (see Theorem 2.1). Let 0 < r < t < T, and suppose that t' >t and g € Ls. Then \\Yv{T,t')g-Yv(T,t)g\\a = \\Yv(r,t)
(Yv (t, t>) g - g)\\a < M \\YV (t, f) g - g\\s
<M\\g\\s\\Yv(t,t')~Y(t,t')\\s+M\Y(t,t')g-g\s. It follows from Theorem 4.5 and from the strong continuity of Y that lim \\YV (T, t') g - YV(T, t)g\\a = 0.
(4.27)
hm\\Yv(r,t')g-Yv(T,t)g\\s=0.
(4.28)
Similarly, we get
Suppose that r ' < r . Then \\Yv(T',t)g-Yv(T,t)g\\s
=
\\(Yv(r,T')-I)Yv(r,t)g\\3
<\\Yv(T',T)-Y(r',T)\\s^s\\Yv(r,t)g\\s +
\\Y(T',r)Yv(r,t)g-Yv(r,t)g\\s.
It follows from (4.22), Theorem 4.5, and from the strong continuity of Y that lim \\YV (r', t)g-
YV{T, t)g\\s = 0.
(4.29)
T'fT
Finally, let r < r ' < t, and let A be such that r ' < A < t. Then \\Yv(T,,t)g-Yv(T,t)g\\a
\\(Yv(T,,\)-Yv(T,\))Yv(\,t)g\\a
=
<\\(Y(r',X)-Y(T,X))Yv(X,t)g\\s +
\\Yv(r',X)-Y(T',X)\\s^s\\Yv(X,t)g\\s
+ ||Y v (r,X)
-Y(T,\)\\„JYv(\,t)g\\a
<\\(Y(T',X)-Y(T,X))Yv(X,t)g\\s + +
M\\Yv(r',X)-Y(T',X)\\a^a\\g\\s M\\Yv(T,X)~Y(T,X)\\s^\\g\\a
292
Non-Autonomous Kato Classes and Feynman-Kac Propagators
= h+h
+ h-
(4.30)
For every e > 0, fix A such that r < A < t and h + h < | e for all r ' with r < T' < A. This can be done using Theorem 4.5. Then the strong continuity of Y implies the existence of 6 > 0 such that I\ < 4e for all r ' with T
Yv(T,t)g\\s
= 0,
(4.31)
T'IT
and it follows from (4.27), (4.28), (4.29), and (4.31) that Yv is separately strongly continuous. This completes the proof of Theorem 4.4. D The next result is an (Ls-Lq )-smoothing theorem for backward Feynman-Kac propagators. T h e o r e m 4.6 Let I < s < q < 00, 1 < r < s, and P G VM. following conditions hold: (a) Let V G Vj, and suppose that Y(r,t)
£ L (LrE, Lf}
Then the
for all 0 < T <
t < T. Then YV{T, t) £ L (L£, L%) for all0
for all 0 < for all 0 <
Proof. We will prove part (b) of Theorem 4.6 in the case q ^ 00. The proof in the case q = 00 is similar. Let g G L3. By estimate (4.21),
l|iv(r,t)sll,
(4.32) 3
Now it is clear that part (b) of Theorem 4.6 follows from (4.32). The proof of parts (a) and (c) is similar. D
Feynman-Kac
4.4
Propagators
293
Feller, Feller-Dynkin, and BJJC-Property of FeynmanKac Propagators
In this section we study the behavior of backward Feynman-Kac propagators on the spaces of continuous functions. Recall that a backward .BC-propagator is called a backward Feller propagator. A backward Copropagator is called a backward Feller-Dynkin propagator. If a backward L^-propagator Q is such that
Q(j,t)eL(Lf,BC),
0
then it is said that Q satisfies the strong Feller condition. If a backward Z,|°-propagator Q is such that Q{T, t)eL
(Lf, BUC),
0
then it is said that Q satisfies the strong Sf/C-condition. Theorem 4.7 hold:
Let P € VM and V £ V* • Then the following assertions
(a) IfY satisfies the strong Feller condition, then Yy also satisfies the same condition. (b) IfY satisfies the strong BUC-condition, then Yy also satisfies the same condition. Proof. We will prove part (a) of Theorem 4.7. The proof of Part (b) is similar. The next lemma will be used in the proof of part (a). Lemma 4.2
The following assertion holds for P £ VA4:
(a) Let V eVJ. Then for all x,x' e E, 0 < r < t
and A > 0
\Yv(T,t)g(x')-Yv(T,t)g(x)\ <2\\Yv(T,T
+ X)-Y(T,T
+ \Y(T, T + X)YV(T
+ X)\\00^00\\Yv(r
+ X,t)g\\00
+ X, t)g (a/) - Y(T, T + X)Yv{r + X,
t)g(x)\.
(4.33) (b) Suppose that P has a density p. Then inequality (4-33) holds for all geL°°.
294
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
(c) Suppose that P has a density p. If /i £ V^, then for all x £ E, x' £ E, 0
< 2 | | y M ( r , r + A ) - y ( r , r + A)|| o o ^ o o ||y / t (r + A,%|| 0 0 + \Y(T,T + X)Ylx(T + \,t)g(x')-Y(r,r + X)Yll(r + X,t)g(x)\. Proof. We will prove only part (a) of Lemma 4.2. The proof of parts (b) and (c) is similar. We have \Yv(T,t)g(x')-Yv(r,t)g(x)\ =
\YV(T,T
+ X)Yv(r + X,t)g(x')
-
YV(T,T
<
\YV(T,T
+
-
Y(T,T
+
\Y(T, T
+
\YV(T,T
X)YV(T
+
< 2 \\YV(T,T
X)YV(T
+
X)YV
+ X,t)g(x')
+ X, t)g (x') (T
+ X,t)g(x) -
+ X)YV(T
+
+ X)YV(T
X,t)g(x')\
Y(T, T
+ X)Yv(r + A, t)g(x)\
Y(T,T
+ X)YV(T
+ A) - Y(r,r + A ) ^ ^ \\Yv(r +
+ \Y(T, T + X)YV(T
+ X,t)g(x)\
+ X, t)g (x') - Y(T, T + X)YV(T
+ X,t)g(x)\
X,i)^^ + A,
t)g(x)\.
This completes the proof of Lemma 4.2.
•
Let us go back to the proof of part (a) of Theorem 4.7. Suppose that Y satisfies the strong Feller condition, and let g £ Lf. Since Yy is a uniformly bounded backward Lg'-propagator (see Remark 4.1), we have ll>V(r,t)||00^00<M
(4.34)
for all 0 < r < t < T. It follows from (4.34) and Theorem 4.5 that for every e > 0 there exists A > 0 such that T + X < t and 2 | | y v ( r , r + A ) - F ( r , r + A)|| 0O _ +0O ||y v (r + A , ^ | | o o < | .
(4.35)
Moreover, for a number A such as above and any fixed x £ E, there exists 6 > 0 such that \Y(T, T
+ X)YV(T
+ A, t)g (x') -
Y(T, T
+
X)YV(T
+ A, t)g{x)\ < |
(4.36)
for all x' with p (x', x) < S. This follows from (4.34) and from the assumption that y is a backward strong Feller propagator. Now it is easy to see that part (a) of Theorem 4.7 follows from (4.35), (4.36), and Lemma 4.2. This completes the proof of Theorem 4.7. •
Feynman-Kac
Propagators
295
The next three theorems are the inheritance results in the case of the Feller property, the Feller-Dynkin property, and the BUC-property. Since it is not known whether these properties are inherited by Yy and Y^ from Y, we impose an additional restriction on the free backward propagator Y. Theorem 4.8 Let P G VM, V £ Pj, and suppose that Y satisfies the strong Feller condition. Then the following assertions hold: (a) If Y is a backward Feller-Dynkin propagator, then the same is true for Yv. (b) IfY is a strongly continuous backward Feller-Dynkin propagator, then the same is true for Yy. Proof,
(a) Let g e Co- Then, by part (a) of Theorem 4.7, YV(T,
t)g eBC
for all 0 < r < t < T.
For every e > 0, there exists a compact set Ke such that \g(x)\ < e for all x G E\KC. Moreover, Urysohn's Lemma implies that there exists a continuous function ge on E with compact support such that gc{x) = 1 for all x € Ke. It follows that \Yv(r,t)g(x)\
< \Yv(T,t)gg£(x)\
+ \Yv(r,t)g
(1 - g£) (x)\
296
N'on-Autonomous
Kato Classes and Feynman-Kac
Propagators
strong continuity of Yy on V in Theorem 4.4. Here we use the space BC instead of the space L3. O The next theorem provides sufficient conditions for the continuity of the function (T, X) —> YV(T, t)g(x) on the set [0,£) x E. Let £ denote the topology on the space BC generated by the uniform convergence of functions on compact subsets of the space E. Theorem 4.10 conditions:
Let P G VM., and suppose that Y satisfies the following
(i) Y is a backward strong Feller propagator. (ii) For every function h G BC such that h = Yy (r, s)g with 0 < r < s < T and g G BC, the mapping (u,v) i-> Y(u,v)h of the space {(u, v) : 0 < u < v < T} into the space (BC, £) is continuous. Then for any V G V*f, t G (0, T], and g G Lf, YV(T, t)g(x) is continuous on the space [0,t) x E.
the function (T,X) H->
Proof. Suppose that the conditions in Theorem 4.10 hold. Using part (a) of Theorem 4.7, we see that Yy is a backward strong Feller propagator. Given t G (0,T] and g G Lf, fix x G E and r with 0 < r < t. Suppose that T' is close to r and x' G U(x), where U(x) is a relatively compact neighborhood of x. Then we have \Yv(T',t)g(x')-Yy(T,t)g(x)\ < \YV (T', t)g (x') - YV(T, t)g (x')\ + \Yv(r, t)g (x') - YV(T, t)g(x)\ = Ji + J2.
(4.37)
Since Yy is a backward strong Feller propagator, lim J 2 = 0.
(4.38)
x' —*x
Next, we will estimate the quantity
sup J\. Let us first suppose that x'eu{x)
T' < T. Then we have SUp J i < sup x'eu(x) x'eu(x) < sup x'et/(x) +
sup x'eu(x)
\(Yy(T',T)-I)Yy(T,t)g(x')\ \(Yy(r',r)-Y(T',T))Yy(T,t)g(x')\ \(Y(T',T)-I)Yv(T,t)g(x')\
<M\Yv{iJ,T)-Y(T',T)\oo^O0\g\O0
Feynman-Kac
+
297
Propagators
sup \(Y(T',T)-I)Yv(r,t)g(x')\. x'eu(x)
(4.39)
Put h = YV(T, t)g. Then for any small e > 0 we have h = YV(T, t - e)Yv(t - e, t)g = YV(T, t - e)hi. Since condition (ii) in Theorem 4.10 holds, the function hi belongs to the space BC. It follows from (4.39), condition (ii) in Theorem 4.10, and Lemma 4.1 that lim sup Ji = 0. 'TTx'e(7(x)
(4.40)
T
Next, suppose that T < r'. Then for every A with r ' < A < t, sup
Ji<
x'£U(x)
sup
\(Yv(T',X)-Yv(T,X))Yv(X,t)g(x')\
x'€U(x)
<
sup x'eu(x) +
\(Yv(T':X)-Y(T',X))Yv(X,t)g(x')\
sup
\(Yv(T,X)~Y(T,X))Yv(X,t)g(x')\
x'€U(x)
+
sup
\(Y(r',X)-Y(r,X))Yv(X,t)g(x')\
x'GU(x) <M\9UYV(T',X)-Y(T',X)\00^00
+ +
M\g\00\Yv(r,X)-Y(r,X)\00_i00 sup
\(Y(T',X)-Y(T,X))Yv(X,t)g(x')\
x'£U(x)
= Ci + C2 + C3.
(4.41)
Using Lemma 4.1, we see that the following statement holds: for every e > 0 there exists A e (r, t) such that if r < r ' < A, then C1+C2 < \E. Moreover, Yv(X,t)g
= Y(X,t~
6)Yv(t - S,t)g = Y{X,t-
5)h
(4.42)
where h 6 BC. Now condition (ii) in Theorem 4.10 and (4.42) imply that there exists 77 > 0 such that r < r ' < T + TJ < X and C3 < \e. Hence, (4.41) gives lim sup Ji = 0. 'J-T x'eu(x)
T
(4.43)
Now it is clear that Theorem 4.10 follows from (4.37), (4.38), (4.40), and (4.43). •
298
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
Corollary 4.2 Let P £ VM, and suppose that Y is a backward strong Feller propagator. Suppose also that for every g £ BC, the mapping (u,v) i-> Y(u, v)g of the space {(u,v) : 0 < u < v < T} into the space (BC,£,) is continuous. Then for any V € Vf, t £ (0, T], and g £ Lf, the function (T,X) —> YV(T, t)g(x) is continuous on the set [0,t) x E. Corollary 4.3 Let P £ VM, and suppose that Y is a strongly continuous backward BUC-propagator. Suppose also that Y possesses the strong BUCproperty. Then for any V G V%, t G (0, T], and g E Lf, the function (T,X) —> YV{T, t)g{x) is continuous on the set [0,t) x E. It is not hard to see that Corollaries 4.2 and 4.3 follow from Theorem 4.10. The proof is left as an exercise for the reader. Remark 4.4 If a transition probability function P G VM. has a density p, then Theorems 4.7-4.10 and Corollaries 4.2 and 4.3 hold for any timedependent measure \x from the class V^. The proofs of these results for fi G Vfn are similar to the proofs for V £V*f.
4.5
Integral Kernels of Feynman-Kac Propagators
Backward Feynman-Kac propagators Yy with V £ V*. possess integral kernels. This will be shown in this section. The kernels of the backward Feynman-Kac propagators Yy and YM are measures. These measures are absolutely continuous with respect to the reference measure m provided that the transition probability function P has a density p. Let P be a transition probability function from the class VM, and let Xt be a corresponding progressively measurable process. Recall that the free backward propagator Y associated with P is the family of integral operators on the space Lf defined by Y(T,t)f(x)=
f
f(y)P(r,x;t,dy)
JE
for all 0 < r < t < T, x £ E, and f £ Lf. In general, the kernel of such a propagator is a Borel measure. If the transition probability function P possesses a density p, then Y(T,t)f(x)=
f JE
f(y)p(T,x;t,y)dy
Feynman-Kac
299
Propagators
for all 0 < r < t < T, x £ E, and / £ Lf. Therefore, in this case, the kernel of Y is a Borel function. Let V be a function from the Kato class V%. Then the corresponding backward Feynman-Kac propagator Yy is a uniformly bounded family of linear operators on the space Lf. This is also true for the backward Feynman-Kac propagator Y^ associated with a time-dependent measure fi from the class V^, provided that the transition probability function P has a density p (see Theorem 4.2 and Remark 4.1). We have already mentioned above that the free backward propagator Y is a family of integral operators. It will be shown below that if the transition density p is strictly positive, then the backward Feynman-Kac propagators Yy and Y^ associated with V £ V*t and /J, £ P^ are also families of integral operators. In the formulation of the next theorem, the symbol ^['y stands for the pinned measure on the cr-algebra T[_ = a (Xs : r < s
then the function PV{T, x-1, A) = YV(T, t)XA(x)
where 0 < T < t < T, x £ E, and A 6 £, is a transition Moreover,
(4.44) function.
Pv(T,x;t,A) = limE T)X exp j - /
V (s, Xs)ds\p
(u, Xu; t, A)
(4.45)
for allO
= ( f(y)Py
(r, x-1, dy)
(4.46)
JE
for allO
(4.47)
300
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
where 0
= / f(y)Plx (r, x; t, y) dy.
(4.48)
JE
Moreover, if the transition probability function P has a strictly positive density p, then for all T and t with 0 < r < t < T and x G E, P»(T, X> t' V) ~
limE
r , x exp {-A„(T,U)}p(U,
XU; t, y)
u\t
= ^exp{-A(r,t)}d^
(4.49)
for m-almost all y € E. Proof. Let V £V}, and let Pv be the for fixed T, t, and A, the function x i-> on E. This follows from the fact that Yy space Lf1 (see Theorem 4.2). Next fix r, Yv(T,t)xA(x)=ETtXXA{Xt)expl-
function denned by (4.44). Then Pv{r,x;t,A) is a Borel function is a backward propagator on the t, and x. Since J
V(s,Xs)ds\,
(4.50)
the function A *-> Py(r, x; t, A) is a Borel measure on £. Moreover, the fact that the function Py satisfies the Chapman-Kolmogorov equation follows from the properties of Yy. Our next goal is to prove the equality in (4.45). By (4.50), the Markov property, and the properties of conditional expectations, Py (T, x; t, A) = YV(T, t)xA{x) = ET,xXA(Xt)expl-[
V(s,Xs)ds\
= limETtxXA(Xt)exp!.-f
V{s,Xs)ds\
= limE r , x exp j - J
V (s, X.) ds\ E T , T [XA (Xt) \ TTU\
= limE T , x exp 1 - j
V (s, Xs) ds\ E u , X u [XA (Xt)]
Feynman-Kac Propagators
= limE r , x exp j -
301
f V (s, X,) ds\ P (u, Xu; t, A).
Therefore, formula (4.45) holds. Moreover, if / € Lf and / > 0, then /•OO
Yv(r,t)f(x)
=
Yv(T,t)X{f>x}(x)dX.
(4.51)
Jo It follows from (4.51) that the function Py is the integral kernel of the backward Feynman-Kac propagator Yy. This completes the proof of part (1) of Theorem 4.11. Next, we will prove part (2) of Theorem 4.11. Suppose that P has a density p, and let \i & V^. We have Y„(T, t)XA(x)
= ET<xXA (Xt) exp {-A M (r, t)} .
(4.52)
It follows from formula (4.52) that for all 0 < r < t < T and x e E, the function A^P^(T,x;t,A)
(4.53)
is a Borel measure on £. Since Y^ is a backward propagator on the space Lf (see Theorem 4.2), the set function x i-> P^(r,x;t, A) is a Borel function on E for all 0 < T < t < T and A £ £. Moreover, PM satisfies the Chapman-Kolmogorov equation. By the continuity of the functional A M (T, t) with respect the variable t, the Markov property, and the properties of conditional expectations, we get PM (r, x; t, A) = Y^T,
t)XA{x)
= UmET,xexp{-A)M{T,u)}ETtX
[XA{Xt)
I Tl\
= liraE T , X exp {-A M (r,u)}E T i a ; [XA (Xt) | K] ujt
= limE r , I exp{-A M (r,u)}E„,x„ \XA (Xt)] = limE T , x exp {-A^T,
U)} Y(U,
t)XA (Xu)
u\t
= limE r x exp {-AJT,
U)}
P
(U, XU; t,
A)
= lim / dyET,xexp {-A M (r, u)}p (u, Xu\ t, y) "Tt JA
(4.54)
for all 0 < r < £ < T, a; G £ , and AG £. It follows from (4.54) that the
302
N'on-Autonomous
Kato Classes and Feynman-Kac
Propagators
measure in (4.53) is the setwise limit of the measures AH
/ dyET<xe-x.r,{-A^{T,u)}p{u,Xu;t,y).
(4.55)
J A
Since the measures in (4.55) are absolutely continuous with respect to m, the measure in (4.53) is also absolutely continuous with respect to m. By the Radon-Nikodym theorem, the transition function P^ has a density p M . It is not difficult to prove that pM satisfies the Chapman-Kolmogorov equation for transition densities. By (4.47) and the formula /•OO
y*(7,t)f(x)=
/ Jo
rM(T,*)X{/>A}(x)rfA,
where / is a nonnegative function from L£°, we see that formula (4.48) holds. Our next goal is to establish the equalities in (4.49). We have exp {-AM (r, t)} < exp {A^- (r, t)}
(4.56)
for all 0 < r < t < T. Since yT € Vm, the family of functions defined by {y t-> ET)X exp {-A^T, u)} p(u,Xu;t,y) : r
(4.58)
Next, let Uk be a sequence of numbers such that T < Uk
= lim / dy / exp {A^- (T, uk)} k->oojE j n
dyETiXexp{A^-(T,uk)}p{uk,XUk;t,y)
= supE T , x exp{A M - (r,u fe )} < E T , x exp {A„- ( T , * ) } < oo. k>l
dfiH
(4.59)
Feynman-Kac
Propagators
303
Therefore, J exp {AM- (T, 0 } d/x[,'* < oo
(4.60)
for m-almost all y £ E. Next (4.56), (4.58), (4.60), and the dominated convergence theorem give lim ET,X exp {-A^ (r, uk)}p(uk,XUk;t,
y)
k—>oo
= /'exp{-A M (T,t)}d/i t T ;J
(4.61)
for m-almost all y € E. Note that the exceptional set in (4.61) does not depend on the sequence uk. Taking this into account, we see that (4.61) implies limE r , x exp {-A^ (T, u)}p(u, Xu; t, y) = /exp{-AM(r,i)}dM^
(4.62)
for m-almost all y G E. Moreover, it follows from (4.54), (4.56), (4.58), (4.59), (4.61), (4.62), and the dominated convergence theorem that / Pn (T, X; t, y) dy = lim / JA k->ooJA = /
lim
Er<xexp{~A^(T,uk)}p(uk,XUk;t,y)dy ET,xexp{-Ali(T,uk)}p(uk,XUk;t,y)dy
JAk-HX>
= / ]\mETtXexp J A "T*
{-A,j.{T,u)}p(u,Xu;t,y)dy
= f /exp{-^(r,i)}dM[; y x J A Jii
for any Borel subset A of the space E. Therefore, Pfj,(T,x;t,y)
= lim E 7 - j X exp{-A / i (r,u)}p(u,X u ; t,y) k—*oo
/ exp{-i4 M (T,t)}d/iJ;JQ
for m-almost all y S E. This establishes formula (4.49). The proof of Theorem 4.11 is thus completed.
•
304
4.6
Non-Autonomous Kato Classes and Feynman-Kac Propagators
Feynman-Kac Propagators and Howland Semigroups
In Section 2.4, we discussed Howland semigroups associated with propagators and backward propagators and considered special examples of Howland semigroups on the space Lf ([0,T], Lf) corresponding to free propagators and free backward propagators on the space Lf. Let us recall that for a free backward propagator, Y{r, t)f(x) = ETtXf(Xt), associated with a transition probability function P, the Howland semigroup Sy(t) is given by the following formula: SY(t)F(r,
x) = Er,xF
((r + t)AT,
X( T+t)AT)
for all T G [0,t], t G [0,T], x G E, and F G Lf (\0,T),Lf). Similarly, let P be a backward transition probability function, and let U be the corresponding free propagator on the space Lf. Then the Howland semigroup SC/(T) is defined by
Su(T)F(t, x) = Ef'xF ((* - r ) V 0, X ( t _ T ) v 0 ) for all r G [0, T], t G [0, T], x G E, and F G Lf ([0, T\,Lf). Next, assume that the transition probability function P belongs to the class VM, and let V be a Borel function on [0,T] x E from the nonautonomous Kato class Vf. Recall that the backward Feynman-Kac propagator Yy is defined on the space Lf by the formula YV(T, t)f(x)
= Er,xf pf t ) exp j - J
V (s, Xs) ds\ ,
where 0
- Er,xf (Xt) exp {-A^T,
t)}
where 0 < T < ( < r , a ; e £ , and / G Lf (seeJDefinition 4.4 and Theorem 4.2). For a backward transition function P G M and V G Vf, the FeynmanKac propagator Uy is given by Uv(t,T)f(x)
=El>xf ( X r ) exp j - f V ( s , X s ) ds\
where 0 < r < t < T, x G E, and / G Lf. If P has a density p, then for every \x G Vf, the Feynman-Kac propagator U^ is given by the following
Feynman-Kac Propagators
305
formula: t/ M (t,r)/(x) = £<•*/ ( X ) exp { - 1 M ( i , r ) } where 0 < r < t < T, x G E, and / G Lf (see Definition 4.5). Since Yy and YM with V £ Vf and /z G 7>^ are backward propagators on the space Lf, the corresponding Howland semigroups are defined by the following formulas: SYv(t)F(r,x) ( /.(r+t)AT = ET,XF ((T + t ) A T , X ( r + t ) A T ) exp I - I V(s,Xs)ds
\ \
and SY.(t)F(T, x) = E T ) X F
((T
+ t) A T, X ( T + t ) A T ) exp { - A M ( T ,
(T
+1) A T)}
where 0 < T < t < T, x £ E, and F £ Lf {[0, T),Lf). Moreover, since Uv and £/M with V £ Vf and /i G P m are propagators on the space L|°, the corresponding Howland semigroups are defined by the following formulas: SUv{r)F{t,x) = El'xF ((t - r ) V 0, X ( t _ T ) v 0 ) exp J - j
V (s, X . ) ds 1
and SUlt(r)F(t,x)
= E'-'F ((t-r)VO,X(t-r)vo) e x p { - A ^ t , ( t - r ) V O ) }
where 0 < T < t < T, x € E, and F € Lg* ([0, T], LgP). The next assertion shows that the non-autonomous Kato class V*t coincides with the Kato class J associated with the Howland semigroup Sy (t) (see Definitions 3.4 and 4.3). Lemma 4.3 Let P be a transition probability function from the class VM., and let Xt be a corresponding progressively measurable Markov process. Let Xt be a space-time process associated with Xt- Then VJ — J, where V% is the non-autonomous Kato class corresponding to the process Xt, and J is the Kato class associated with the process Xt • Proof.
Let V be a Borel function on [0, T] x E. By Definition 3.4, VeV}<=>
lim s u p E T X / \V(s,Xa)\ds t-T{oxeE ' J
= 0.
306
Non-Autonomous Kato Classes and Feynman-Kac Propagators
It follows that /•(U+T)AT
VeV}<=>\im
sup ETiX «i°(T,x)e[0,T]x£ ' JT
\V(s,Xs)\ds
= Q.
(4.63)
On the other hand, Definition 4.3 shows that VeJ^^lim
sup E(TX) '1° (r,x)e[0,T]xE (r,i)£[0,T]xE 'Jo
V V
(XS) ds = 0. '
It is not difficult to see from the previous equivalence that AU+T)AT
VeJ<=Mim
sup
"J-°(T,X)€[0,T]X£;
ETX /
| V ( s , X s ) | d s = 0.
(4.64)
' JT
Now it is clear that Lemma 4.3 follows from (4.63) and (4.64).
•
In the book by Chung and Zhao [Chung and Zhao (1995)] various results concerning the inheritance of properties of free semigroups by Feynman-Kac semigroups are discussed. One may be tempted to derive the inheritance results obtained in Sections 4.3 and 4.4 of the present book from the similar results for Feynman-Kac semigroups established in Section 3.2 of [Chung and Zhao (1995)]. An encouraging motivation for this approach is the coincidence of the non-autonomous backward Kato class V% and the Kato class J associated with the space-time process Xt (see Lemma 4.3). However, this approach often fails. Next, we will give two examples illustrating the restricted applicability of Howland semigroups to the study of the inheritance problem for propagators. Let us first compare Theorem 4.7 in Section 4.4 of the present book and Proposition 3.12 in [Chung and Zhao (1995)]. Both results concern the inheritance of the strong Feller property. It is assumed in Proposition 3.12 in [Chung and Zhao (1995)] that the transition probability function P has a density p that is a symmetric bounded and time-homogeneous. It is also assumed that the function V belongs to the Kato class J (see condition (15) on page 70 in [Chung and Zhao (1995)]). Chung and Zhao established that under such restrictions, the FeynmanKac semigroup inherits the strong Feller property from the free semigroup associated with the transition probability density p. On the other hand, Theorem 4.7 in the present book states that if P is a non-homogeneous transition probability function, then the backward Feynman-Kac propagator Yv associated with a function V from the non-autonomous Kato class V*t inherits the strong Feller property from the free backward propagator Y. Let us try to obtain a special case of Theorem 4.7 using Proposition
Feynman-Kac
Propagators
307
3.12 in [Chung and Zhao (1995)]. Suppose that p is a transition probability density such that the corresponding free backward propagator, Y(T, t)f(x) = / f(y)p(T, x; t, y)dy,
(4.65)
JE
possesses the strong Feller property. This means that for every bounded Borel measurable function / and all 0 < r < t < T, the function x— I > Y(T, t)f(x) belongs to the space BC of bounded continuous functions on E. The Howland semigroup Sy(t) associated with the free backward propagator Y has the following form: SY{t)F(r,
x) =
Y(T,
(T
+1) A T)F((T
+ t)A T){x)
(4.66)
for all t > 0, (T, x) G [0,T] X E, and all bounded Borel functions F on [0, T] x E. Even if we forget that the transition probability function for the Howland semigroup has a singular component, and it is not necessarily symmetric and bounded, we still need to establish that the semigroup Sy (t) possesses the strong Feller property, that is, the function on the righthand side of (4.66) is continuous on the space [0,T] x E for every t > 0. However, the validity of the previous assertion is not clear, since we only know the continuity of the function x H-> Y(T, t)f(x) in the variable x, and no continuity assumption in the variable r is imposed. Our next example concerns the strong continuity of semigroups and propagators on the space LP. Even if the free propagator Y is strongly continuous on the space LP(E), it is not clear how to prove that the corresponding Howland semigroup Sy(t) is strongly continuous on the space LP ([0,T] x E). More precisely, it is difficult to expect that the strong continuity of Sy(t), that is, the condition lim / / *i° J
\F(T,
x) -
Y(T,
(r + t)A
T)F((T
+1) A T)(x)\v drdx = 0
J[0,T}xE
for all F £ Lp([0,T] x E), can be obtained from the strong continuity of the free backward propagator Y on the space LP(E).
4.7
Duhamel's Formula for Feynman-Kac Propagators
Duhamel's formula is an important link between the free backward propagator Y and the backward Feynman-Kac propagators Yv and YM. This formula shows that backward Feynman-Kac propagators can be obtained by
308
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
solving a Volterra type integral equation. It will be established below that for a function V £ Vj and a time dependent measure fi £ Vm, Duhamel's formula holds pointwise. Theorem 4.12 (a) Let P e VM, f £ Lf, and V £ V*f. Fix t £ (0,T], and define a function on [Q,t] x E by U(T,X) = Yv(r,t)f(x). Then for every T £ [0, £], the function x *—> U(T,X) belongs to the space Lf. Moreover, the following Volterra type integral equation holds: U(T, X) = Y{T, t)f(x) -
f Y{T, S) [V{S)U(S)}
(x)ds
(4.67)
for allxGE andO
(4.68)
for all x £ E and 0 < r
The equations in (4.67) and (4.68) can be rewritten as
t)f(x)
= Y(T, t)f(x) - J
Y(T, S) [V(S)YV(S,
t)f] (x)ds
(4.69)
Y^(T, t)f(x)
= Y(T, t)f(x) - J
Y(T, S) [V(s)Yli(s, t)f] (x)ds.
(4.70)
and
Proof, (a) Let V £ V). By Theorem 4.2, for every T £ [0, t] the function x H-> U(T, X) belongs to the space Lf. Moreover, under the assumptions in part (a) of Theorem 4.12, we have
/
Y(T,s)[V(s)Yv(s,t)f](x)ds J
ETiXV(s,Xs)Es,xJ(Xt)exr,{-
J
V(X,Xx)d\\
ds,
Feynman-Kac
Propagators
309
where Xt is a progressively measurable process with P as its transition function. By the Markov property, t
It
Y(T,s)[V(s)Yv(s,t)f}(x)ds
= J
E T , I V ( a , X , ) K r , x ( / ( X t ) e x p j - /"
= J
ET,xf(Xt)V(s,Xs)expi-[
= J
ET<xf(Xt)-^exp(-J
= ET,xf(Xt) =
V{X,Xx)dx\\Ts)ds V(X,Xx)dx\ds
V(X,Xx)dx\ds
- ETlXf(Xt)
exp | - J
V(X:
Xx)dx\
Y(r,t)f(x)-Yv(T,t)f(x).
This gives part (a) of Theorem 4.12. (b) Suppose that the assumptions in part (b) of Theorem 4.12 hold, and let /Lt £ V*. In the proof of part (b) we employ the approximation result from Section 3.9 (Lemma 3.21). Let 14 be the sequence of functions defined by (3.152). It follows from Lemma 3.21 and Theorem 4.3 that lim
sup
sup\Yll(T,t)f(x)-YVk(T,t)f(x)\=0
fc
(4.71)
^°°0
for all / G Lf. The functions 14 belong to the class V*f (see Lemma 3.21). Therefore, (4.69) gives YVk (r, t)f(x) = Y(T, t)f(x) - j
Y(T, S) [Vk(s)YVk (s, t)f] (x)ds.
It follows from the properties of backward propagators that
i:
Y(T,s)[Vk(s)YVk(s,t)f}(x)ds
= f Y(T, s)k fS ft
JT
*
Y(s, X) [fi(X)YVk (s, t)f]
(x)dXds
/-(S+£)AT
J kj --
(*+i)*r
Y(T,X)HX)YVk(s,t)f](x)dXds dXk
Y(T, X) HX)YVk (s, t)f) (x)ds J(X-i)\/T
(4.72)
310
Non-Autonomous Kato Classes and Feynman-Kac Propagators
-{t+i)^T /
f d\Y(T,X)
/-A n(X)k /
YVk(s,t)fds (x).
(4.73)
Next, passing to the limit as k —> oo in (4.73) and using (4.71) and the definition of the class P^,, we get lim / Y(T,s)[Vk(s)YVk(s,t)f}(x)ds=
f
Y(r,X)[^(X)Y^X,t)f}(x)dX.
(4.74) Now it is clear that part (b) of Theorem 4.12 follows from (4.71), (4.72), and (4.74). This completes the proof of Theorem 4.12. D Duhamel's formula shows that backward Feynman-Kac propagators generate solutions to Volterra type integral equations. Next, we will see that under certain restrictions, backward Feynman-Kac propagators generate solutions to final value problems. We will first reason informally. It will be assumed in the remaining part of this section that all functions are differentiable as many times as needed. Let t and r be such that 0 < r < t < T. Fix 7 with 0 < 7 < r and apply the operator ^ ( 7 , r) to Duhamel's formula. Then, using the properties of backward propagators, we get Y(y,T)Ylt(T,t)f(x)=Y(1,t)f(x)-J
Y(1,s)[fi(s)Yti(s,t)f](x)ds
(4.75)
where / G Lf. Differentiating the equation in (4.75) from the right with respect to the variable r on the interval (7, t), we obtain ^
[y( 7 , T)Y,(T, t)f(x)} = F ( 7 , r ) [/i(r)^(r, t)f] (x).
It follows that |^Y(7,T))
Y,(r,t)f(x)
+y(7,r)
(|^(T,*)/(*))
= Y{7,T)\/i(T)Yll{T,t)f](x).
(4.76)
Our next goal is to apply Theorem 2.11 to the first term on the left-hand side of equality (4.76). By Theorem 2.11, if Y^^feD^in)
(4.77)
Feynman-Kac
Propagators
311
for all 0 < T\ < Ti < t, then F ( 7 ) r ) ^ ( r ) K M ( r , t)f(x) + Y(j, r ) ( f ^ ( r , *)/(*)) = Y(1,T)[fx(r)Yfl(T,t)f}(x).
(4.78)
The symbol D+'* (n) in (4.77) stands for the set defined by
<•>!)=
pi a™ «(*)), t:n
where Dw (A+(t)) denotes the subspace of the space Lf consisting of all functions for which the limit in formula (2.31) exists. Passing to the limit as 7 | T in (4.78), we see that if condition (4.77) holds, then the function U(T, X) = l^(r, t)f(x) is a solution to the following final value problem: — U ( T , X) + A% (T)U(T)Or)
- M(T)U(T,
X)=0,
u(t,x) = f(x). In the next section we will explain what can be done if the differentiability conditions in the reasoning above are not satisfied. It will be shown that under certain restrictions there exist viscosity solutions to final value problem (4.79). 4.8
Feynman-Kac Propagators and Viscosity Solutions
Viscosity solutions to partial differential equations were introduced in [Crandall and Lions (1983)]. See also [Crandall, Ishii, and Lions (1992)] for more information on viscosity solutions. The main results in this section are Theorems 4.15 and 4.16. These theorems provide sufficient conditions for the solvability of the final value problem in (4.79) in the viscosity sense. We will first establish several preliminary results. Theorem 4.13 Let /J, G V*m, f G L°°, and fix t such that 0 < t Then the following assertions hold:
(a) Suppose that ip is a bounded continuous function on [0, T] x E, and let (TO, X0) G [0, t) x E and 5 > 0 with TQ + S < t be such that Yli{T0,t)f{xo)-i}){To,xo)=
min (r,x)e[T 0 ,T 0 +<5]x£
(Y^(T,t)f(x)
-
ip(r,x)).
312
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
Then for every 0 < e < 6, Y
+ e) V (TO + e) (JC0) - ip (TO,
(TO, TO I
X0)
rro+e
e
fT0
e
Y(TO,s)[n(s)Yli(s,t)f}(x0)ds<0.
(4.80)
•>Tn
(b) Suppose that ip is a bounded continuous function on [0, T] x E, and let (TO, XO) G [0, t) x E and 5 > 0 with To + S < t be such that Y^{T0,t)f(x0)-ip(To,xo)=
max
(Y^T,
t)/(x) -
V(T,:T))
(T,^)e[ro,r 0 +(J]x£'
/or some
(TO,XO)
e [0, t) x E. Then for every 0 < e < 5,
y
+ e) ip (r 0 + e) (x 0 ) - ^ (r 0 , x 0 ) e
(TO, TO 1
- - /
/-To+e
y(7i,,a)[^(s)y M (s,*)/](a:o)da>0.
(4.81)
Proof. We will prove only part (a) of Theorem 4.13. The proof of part (b) is similar. Let ip be any bounded Borel function on [0, T] x E, and let M be any real number. Put G(T, X) = yM (T, t)f(x) - V(T, Z) - M.
(4.82)
Lemma 4.4 Le£ /x G 7-£j, and let e > 0 be such that T + e < t. Then the following equality holds for the function G defined by (4-82): G(T,x)-Y(r,T
+ e)G(T + e)(x)
= Y(T, T + C)V(T + e)(x) - IP(T, X)- J'
' Y(T, S) [/i(s)y^(s, t)f] (x)ds. (4.83)
Proof.
We have y M (r, t)f(x) - Y(T, T + e)Yli(T + e, t)f(x) = (Y^T,T + e)-Y(r,T
+ e))Yfi(r + e,t)f(x).
Using formula (4.68) in (4.84), we get Y^T, t)f{x) - Y(T, T + e)Yli(T + e, t)f(x)
I
Y{T, S) \fi(8)Yli(s, T + e)Yfi(T + e,t)f] (x)ds
(4.84)
Feynman-Kac
Propagators
313
= - JT ' Y(T, S) \H(S)Y^ (S, t)f] (x)ds.
(4.85)
Now it is clear that (4.83) follows from (4.85). This completes the proof of Lemma 4.4.
•
Let us return to the proof of Theorem 4.13. Suppose that ip is a function such as in the formulation of Theorem 4.13, and put M=
min
(F M (T,i)/(x)-V(T,z)).
(r,x)e[ro,ro+<5Jx£
Define the function G by (4.82). Then, Lemma 4.4 with r = To, x = xo, and e > 0 such that TQ + e < t implies that G (r 0 , X 0 ) - Y (T0, TQ + e) G (r 0 + e) (x0) = Y (T0, T0 + e)ip (T0 + e) (x0) - tp (T0, X0) C0
£
Y(T0,S)[»(s)YM(s,t)f}(x0)ds.
(4.86)
JTC 'TO
Next, dividing (4.86) by e and using the facts that G(TO,XO) = 0 and G (TQ + e, y) >0 for all e < <5 and y G E, we get estimate (4.80). This completes the proof of Theorem 4.13. • The next theorem is a local version of Theorem 4.13. Theorem 4.14 Let fj, g V*m, f £ L°°, and fix t such that 0 < t Then the following assertions hold:
(a) Suppose that ij) is a bounded continuous function on [0,T] x E and (TO, So) is a point in \0,t) x E. Suppose also that there exists 5 > 0 with To + 5 < t and a relatively compact neighborhood Q of XQ in E such that Y/J.(To,t)f(xo)-tp(To,x0)=
min
_(Y/1(T,t)f(x)
-
tp(T,x)),
(T,X)€[TO,T 0 +<5]XQ
where Q denotes the closure of Q in E. Then for every 0 < e < 5, Y (r 0 , TQ + e) tl> (T0 + e) (x) - if> (r 0 , x0) e T0
- -
f e
< - / e
V (TQ, S) [fi(s)Y^(s, t)f] (x0)ds
JTO
p(To,x0;T0 + 6,y)dy, JE\Q
(4.87)
314
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
where a > 0 and 6 > 0 do not depend on e. (b) Suppose that tp is a bounded continuous function on [0,T] x E, and let (TO,XO) be a point in [0, t) x E. Suppose also that there exists a relatively compact neighborhood Q of (TO,XQ) in [0,i) x E such that y^ (TO, t) f(xQ)
- ip (T 0 , XO) =
max (Y^(T, t)f(x)
- tp{r, x)).
(T,X)£Q
Then for every small e > 0, y (TO, TQ + e) tp (T 0 + e) (x) -ipJTp, x0) e Jr. >
/ e
Y(T0,s)[n(s)Yti{s,t)f}{x0)ds p(To,x0;T0+e,y)dy,
(4.88)
JE\Q
where a > 0 and 5 > 0 do not depend on e. Proof. We will only prove part (a) of Theorem 4.14. The proof of part (6) is similar. Suppose that the conditions in part (a) of Theorem 4.14 are satisfied. Define G by formula (4.82) with ip as in the formulation of Theorem 4.14 and with the number M given by M =
min
(rM(T, t)f(x) - V(T, X)).
(r,x)e[r 0 ,To+5]xQ
Then, using equality (4.83) with T = To, x = Xo, and e > 0 such that e < 5 and taking into account that G (TO, XO) = 0, G(T, X) > 0 for (T, X) G [TO, TO + S] x Q, and | G ( T , X)| < a, we get y
(TO, T 0
+ e) V
(TO
+ e) (x0) - ip (T 0 ,
X0)
To+e
£ = -Y
Y(T0,s)[ii(s)Y^(s,t)f](x0)ds JT, (TO, T 0 + e) XQ (T 0 + e) (x0)
-Y(T0,T0
+ e) XE\QG
(T 0 + c) (a;0)
< - / p(T0,x0-T0 + e,y)dy. e VB\Q Therefore, estimate (4.87) holds. This completes the proof of Theorem 4.14.
•
Our next goal is to make several simplifications in inequalities (4.80), (4.81), (4.87), and (4.88). The following lemma concerns the first term
Feynman-Kac
Propagators
315
on the left-hand side of estimates (4.80), (4.81), (4.87), and (4.88). For he BC and (T,X) € [0,T) x E, we p u t A+{r)h{x)
= lim
Y{T, T + e)h{x) -
h{x)
(4.89)
e—»0+
provided t h a t t h e limit in (4.89) exists and is finite. We will say t h a t a bounded continuous function ip on [0, T] x E is differentiable from t h e right at To € [0, T) uniformly with respect t o y £ E, if there exists a function D^tp(T0r)eBC such t h a t lim sup
i>(ro + e,y)
-ip{T0,y)
-D+iP(T0,y)
= 0.
(4.90)
L e m m a 4.5 Suppose that the free backward propagator Y satisfies the conditions in Theorem ^.10, and let ip be a bounded continuous function on [0, T] x Rn. Let TQ e [0,t) and xo £ E be such that ip is differentiable from the right at TQ S [0, T) uniformly with respect to y £ E, and A+ (TO) ip (TQ) (XQ) exists and is finite. Then y
lim
(TO, TO
+ e) V (TO + e) (x 0 ) - ip ( T 0 ,
I0)
(4.91)
= D+iP (T 0 , X0) + [A+ ( T 0 ) ip ( T 0 ) ] ( X 0 ) . Proof. y
We have (TQ, TQ
+ e) ip
= Y(TO,ro
+
(TQ
+ e) ( i 0 ) - ip ( T 0 , e
e){^T0
X0)
+ e
]-^T0)
-DliP(r0)}(x0)
+ [Y (T 0 , TO + e) D+iP ( T 0 ) ] ( X 0 ) +
Y{T0,T0
+
e)-I •
= h + h + h-
(x 0 ) (4.92)
Since Y is a family of contraction operators on L | / i | < sup
ip{To +
£,y)-ip(T0,y)
yeE
Dti>(T0,y)
It follows from (4.90) t h a t lim h = 0. £ —0
(4.93)
316
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
Since D^ip (TQ, •) € BC, and the conditions in Theorem 4.10 are satisfied, we get lirah = Dt^ {T0, x0).
(4.94)
Finally, (4.89) implies that lim h = [A+ (r0) i> (TO)] (X 0 ).
(4.95)
Now it is clear that Lemma 4.5 follows from (4.92)-(4.95).
D
Next, we turn our attention to the second term on the right-hand side of estimates (4.80), (4.81), (4.87), and (4.88). For fj, G V^, consider its Radon-Nikodym-Lebesgue decomposition d^,(s) = V(s)dm + d\(s), where A(s) is the singular part of fi(s) with respect to m. It is clear that V 6 V% and A € V^. Let x0 € E and r 0 G [0,t) be given, and suppose that Ck with — oo < k < oo is a strictly increasing sequence of Borel sets of positive measure m such that XQ G Cfc for all k, diam(Cfc) + m(Ck) —» 0 asfc—» —oo, and M Ck = E. For every integer j , put fc=0
7j(s)=
sup p ( T 0 , x 0 ; s , y ) ,
and define the majorant p* of p with respect to the family {Ck} as follows: P*(T0,X0;S,Z)
=jj(s)
where j is the unique integer such that z G Cj+\\Cj. Let us also recall that the function VM(s, t)f(x) is bounded on [0, t] x £ . Moreover, it is continuous on [0,t) x E, by Theorem 4.10 and Remark 4.4. The following conditions will be used in the sequel: sup S:T0<S
— ^
f
fn(yk)
Jck
\V(s, y)-V
(TO,X 0 )
\dy - • 0
(4.96)
as k —> —oo, where S > 0 is a number such that TO + 6 < t; sup
sup
—^— f
k:k>j S:T0<S
™\yk)
\V(s,y)\dy<MUj
(4.97)
JCk
for all j € Z;
sup
™u„
S:T 0 <S
"HOfcJ
(4.98)
Feynman-Kac
Propagators
317
as k —> —oo; sup
sup
k:k>j S:T0<8
|AQQ|(C ) ^ ^ fc < M2tj
(4.99)
WHOfcJ
for all j € sup
/
:TO<S
p* (r 0 ,x 0 ; s, z) dz < M3ik
(4.100)
JCk
for all k S Z, and lim / P* (To,xo;s,z)dz s^r 0 + JEB\C fc
=0
(4.101)
for all /c € Remark 4.6 Condition (4.96) means that XQ is a Lebesgue point of the function V(T, •) uniformly with respect to T near TO- Condition (4.97) resembles a uniform local integrability condition for V. Similarly, condition (4.98) is a differentiability condition for the singular part A of //, while (4.99) is a uniform local integrability condition for A. Conditions (4.100) and (4.101) are expressed in terms of the majorant p* of the transition density p. They are based on the integrability condition for the majorant of an approximation of the identity (see [Stein and Weiss (1971)], Theorem 1.25). Condition (4.100) concerns the local integrability of p*, while (4.101) is a stochastic continuity condition for p*. Lemma 4.6 Let ji € V^, and assume that the free backward propagator Y satisfies the conditions in Theorem 4-10. Let TO € [0,t), XQ € E, and let {Ck} be such that conditions (4-96)-(4-101) hold. Then
£
limi/
Y (r 0 ,
S) [ M ( S ) ^ ( S >
*)/)] (*o)ds = V
(TO,
ar0) Y„ (T 0 , t) f(x0).
^°+ e J-n
Proof.
(4.102)
Put D(s,y) = V(s,y)Yli(s,t)f(y)
and dv(s) =
Yp{s,t)f(y)d\{s).
Since the function Yft(s,t)f(y) is continuous on [0,i) x E and bounded on [0, t] x E, it follows from (4.96) that sup S:T0<S
—±— [ "HCfcj
JCk
\D{s,y)-D{T0,xo)\dy^0
(4.103)
318
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
as k —•> —oo. In (4.103), 8 > 0 is a number such that To + 5
sup
—-—
j
^GfcJ
JCk
z:k>j S:T0<S
\D{s,y)\dy < Mij
(4.104)
for all j G Z;
MflM^^o
sup S:TO<S
(4,05)
m\yk)
as k —> —oo; and sup
|t/(5
sup
fc:fc>js:T0<s
, ^ f c ) < MSij
(4.106)
W(Ofe)
for all j € Z. Our next goal is to show that lim - f ° ~*°+
e
y (T 0 , a) [|D(«) - D (r 0 , ar0)|] (z 0 )ds = 0
(4.107)
JI To To
and lim - /
y(r 0 ,s)|i/(s)|(a;o)ds = 0.
(4.108)
•'To
Put T
(«.y) = l-D(s, y) - -D (r 0 ,io) |.
Then we have Y(T0,S)T(S){X0)=
/
T{s,y)p(T0,x0;s,y)dy
JE
< / T{s, y)p*
(T 0 , Z 0 ;
s, y) dy
JE
= / JO
dX [
T(s,y)dy.
(4.109)
•'{yp"(To,xo;s,y)>X}
Since 7it(s) is a non-increasing sequence, it follows from (4.109) that there exists 5 > 0 such that Y(-n},a)T(a)(xo)
f
T(s,y)dy.
(4.110)
Feynman-Kac Propagators
319
It is not hard to see that for any j e Z, (4.110) gives Y(T0,S)T(S)(X0)
<
sup
S:T0<S
Yl (7fe(*)-7fc+i(a))m(Cfc)-77TT / j.r^ rn{Ck) JCk
T(s,y)dy
oo
+ E
(7*(s)-7fc+i(s))m(C* fc )^77T /
r(s,»)dy = Ji(j) +
J2(j,s).
k=j+i
(4.111) We have Ji(j) < \
sup sup 7TT C I S:TO<S
/
/
|-D(S,J/)--D(T0,XO)|^1
Jck
J
p*(T0,x0;s,y)dy.
(4.112)
r.T0<S
Moreover, for every j e Z, oo
J2U^)<(M4,J+I
+ \D(TO,X0)\)
Y,
(lk(s) -
Jk+i(s))m(Ck)
k=j+i
<(M 4 l J - + l + |I>(7i),Xo)|) lj+i(s)m(Cj+i)
+ ]T
lk(s)m(Ck\Ck-i)
k=j+2
< ( M 4 J + i + |£>(r 0) xo)|) <(Mitj+1
+
\D{T0,x0)\)
M C j + l) m(CJ+1\Cj)
+1 /
P*
m(Cj+i) miCj+^Cj)
+1 /
P*
(To,x0;s,y)dy
JBXCJ
(r0,x0;s,y)dy.
JE\d
(4.113) It follows from (4.111), (4.112), and (4.113) that Y(T0,S)[\D(S)-D(T0,X0)\}(X0)
<\
sup
sup
• /
ys:Ta<s
sup
/
\D(s,y)
-D(r0,x0)\dy
JCk
J
p*(T0,x0;s,y)dy
s:To<s <S
+ (M4ij + 1 +
\D(T0,Xo)\)
m(Cj+1) MCJ+ACJ)
+1
/ JE\Cj
P*(T0,X0; s,
y) dy
320
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
for all j £ Z. Now it is not difficult to show that conditions (4.103) and (4.104) imply lim
Y (TO, S) [\D(S)
- D (r 0 , x0) |] (io) = 0.
S—>T0 +
This gives equality (4.107). The proof of equality (4.108) is similar. Here we use (4.105) and (4.106) instead of (4.103) and (4.104). It is clear that (4.107) and (4.108) imply (4.102). This completes the proof of Lemma 4.6. • Now we are ready to formulate the main results of the present section. The first of them concerns viscosity solutions in the case of global maxima or minima. Theorem 4.15 Let \i € Vm, f e L°°, 0 < t < T, and suppose that the transition density p is such that the corresponding free backward propagator Y satisfies the conditions in Theorem J^.10. Then the following two assertions hold: Let (TO, XQ) € [0, t) x E, 6 > 0 with TQ + 5 < t, and ip be such that: (1) (2) (3) (4)
ip is a bounded continuous function on [0, T] x E; ip is differentiable from the right at TO uniformly with respect to y e E; A+ (TO) ip (TO) (XO) exists and is finite; There exists a sequence of sets Ck such that conditions (4-96)-(4-99) hold; (5) The equality Y» (TO, t) f(x0) - V (TO, zo) =
, min
(Y^T, t)f(x) - ip(r, x))
holds. Then Dti> (T0, XO) + [A+ (TO) V (TO)] (io) - V (T 0 , XQ) Y„ (T 0 , t) f(x0)
< 0.
Suppose that conditions 1~4 in part (a) of Theorem 4-15 are satisfied. Suppose also that Ytl(T0,t)f(xo)-i)(To,xo)
=
i
max
(Yll(T,t)f(x)
- V>(r,i)).
Then Dtl> (TO, XQ) + [A+ (TO) i) (T 0 )] (X 0 ) - V (T 0 , X0) Y^ (T 0 , t) f(x0) > 0.
Feynman-Kac
Propagators
321
It is clear that Theorem 4.15 follows from Theorem 4.13, Lemma 4.5, and Lemma 4.6. Our next result concerns viscosity solutions in the case of local maxima and minima. Theorem 4.16 Let n £ V*m, f G L°°, 0 < t < T, and suppose that the transition density p is such that the corresponding free backward propagator Y satisfies the conditions in Theorem 4-10. Then the following two assertions hold: Let (ro, xo) € [0, t) x E, 5 > 0 with TQ + 6 < t, and ip be such that: (1) (2) (3) (4)
ip is a bounded continuous function on [0, T] x E; ip is differentiate from the right at TQ uniformly with respect to y € E; A+ (TO) ip (TO) (XQ) exists and is finite; There exists a sequence of sets Ck such that conditions (4-96)-(4-99) hold; (5) The equality lim - / p(To,x0;T0+e,y)dy e-^o+ e JE\Q
=0
(4.114)
holds for every relatively compact neighborhood Q of XQ; (6) There exists a relatively compact neighborhood Q of (TO, XQ) in [0, t)xE such that Y,j,(To,t)f(xo)-ip(To,xo)=
min
_(Yll(T,t)f(x)
-
tp(r,x)).
(T,X)£[TO,TO+6]Q
Then D+i> (TO,X0) + [A+ (TQ) V (TO)] (X0) - V (T0,X0) Y„ (T0,t) f(x0)
< 0.
Suppose that conditions 1-5 in part (a) are satisfied. Suppose also that there exists a relatively compact neighborhood Q of (ro, xo) in [0, t)xE such that y^(To,t)f(xo)-tp(T0,x0)=
max
_(YIJ.(T,t)f(x) -
ip(r,x)).
(r,x)e[T0,T0+S]Q
Then Dfi/J (T0, x0) + [A+ (T0) $ (TO)] (XO) - V (T0, X0) Y^ (T0, t) f(x0) > 0. Theorem 4.16 follows from Theorem 4.14, Lemma 4.5, and Lemma 4.6.
322
Non-Autonomous
Kato Classes and Feynman-Kac
Propagators
E x a m p l e 4.1 Let E be d-dimensional Euclidean space Rd. We will assume that the reference measure m coincides with the Lebesgue measure md on WLd. Suppose that p is a fundamental solution of a second order parabolic partial differential equation with time-dependent coefficients such as in Sections 3.6 and 3.7. Then p is a transition probability density (see Theorems 3.2 and 3.5). Recall that the density p in these theorems satisfies the upper Gaussian estimate; that is, p(r, X; t, y) < a.x9d (oc2(t -r),x-y),
(4.115)
where ot\ and ca-i are positive constants. In estimate (4.115), gd stands for the rf-dimensional Gaussian density given by
Define the radial majorant of the transition density p by P*{r,x;t,y)=
sup
p{r,x;t,z).
z:\z — x\>\y—x\
It is clear that if estimate (4.115) holds for p, then P*(T, X; t, y) < axgd (a2(t - T), x - y), and hence, conditions (4.100) and (4.101) hold for p*. It is not difficult to prove that condition (4.114) also holds. We will assume that the sets Ck in the formulation of Theorem 4.15 and Theorem 4.16 are given by Ck = B(x0, rk) where r^ I 0 as k —> - c o and rk T oo as k —> oo. The next assertion follows from Theorem 4.15. Corollary 4.4 Let p be a transition probability density on Rd such that estimate (4.115) holds for p. Let p, £ V*, f S L°°, 0 < t < T, and suppose that Y satisfies the conditions in Theorem 4-10. Then the following two assertions hold: Let (ro, XO) e [0, t) x E, 5 > 0 with TO + d < t, and tp be such that: (1) (2) (3) (4)
ip is a bounded continuous function on [0, T] x E; ip is differentiable from the right at TQ uniformly with respect to y € E; A+ (ro) ip (TO) (ZO) exists and is finite; Conditions (4-96)-(4.99) hold with Ck = B(x0,rk) where rk are such that rk | 0 as k —> —oo and rk 1 oo as k ^ oo;
Feynman-Kac
Propagators
323
(5) There exists a relatively compact neighborhood Q of XQ in E such that Y»(To,t)f{xo)-iP(T0,x0)=
min
{Y^T,t)f{x)
-
ip(r,x)).
(T,x)e[T0,T0+S]xQ
Then D+V(r, x) + [A + (r)V(r)] (X) - V (TO, I 0 ) ^ fa, *) /(*<>) < 0. Suppose that conditions 1-4 in part (a) are satisfied. Suppose also that there exists a relatively compact neighborhood Q of XQ in E such that y^(To,t)f{x0)-ip(T0,x0)^
max
{Y^T,t)f(x)
-
I/}(T,X)).
(T,X)6[TO,T0+«]XQ
Then D+1>(T, X) + [A+(T)1>{T)]
4.9
(X) - V (TO, X0) YM (T 0 , t) f(x0)
> 0.
N o t e s and Comments
(a) The Kato class of potential functions was introduced and studied in [Aizenman and Simon (1982); Simon (1982)]. The definition of the Kato class in these papers is based on a condition used in [Kato (1973)]. Similar classes were studied in [Stummel (1956)] and [Schechter (1971)]. More information on the Kato classes of functions and measures can be found in [Johnson and Lapidus (2000); Demuth and van Casteren (2000); Gulisashvili (2002c)]. (b) Schrodinger semigroups are discussed in [Aizenman and Simon (1982); Simon (1979); Simon (1982); Carmona (1974); Chung and Zhao (1995); Blanchard and Ma (1990a); Blanchard and Ma (1990b); Davies (1997); Johnson and Lapidus (2000); Demuth and van Casteren (2000); Zhang (2001); Gulisashvili and Kon (1996); Gulisashvili (2000)]. See also [Carmona, Masters, and Simon (1990)]. (c) The Feynman-Kac formula in (4.2) goes back to Kac (see [Kac (1949); Kac (1951); Kac (1959)], see also [Kac (1979)]), who was inspired by Feynman's ideas. We refer the reader to [Johnson and Lapidus (2000); Kleinert (2004)] for more information on the Feynman integral and related topics. (d) Kato classes and Feynman-Kac semigroups associated with general Markov processes are discussed in [Chung and Zhao (1995)].
324
Non-Autonomous
Koto Classes and Feynman-Kac
Propagators
(e) The results in Sections 4.3 and 4.4 concerning the inheritance of properties of free semigroups or propagators by their Feynman-Kac perturbations are taken from [Gulisashvili (2004b); Gulisashvili (2004c)]. For the case of Feynman-Kac propagators associated with the heat semigroup see [Gulisashvili (2005)]. See also [Ouhabaz, Stollmann, Sturm, and Voigt (1996)] for earlier results concerning time-independent perturbations of semigroups on the space L 1 . Measure perturbations of semigroups of operators were studied in [Getoor (1999)]. (f) The reader may consult [Demuth and van Casteren (2000)] for more information on the integral kernels of Feynman-Kac semigroups. (g) Viscosity solutions of partial differential equations were introduced in [Crandall and Lions (1983)] (see also [Crandall, Ishii, and Lions (1992)]). The results in Sections 4.7 and 4.8 concerning the generation of viscosity solutions by Feynman-Kac propagators are taken from [Gulisashvili and Van Casteren (2005)].
Chapter 5
Some Theorems of Analysis and Probability Theory 5.1
Monotone Class Theorems
In this section we formulate monotone class theorems for sets and functions. These theorems are due to Dynkin. Definition 5.1 Let O be a set and let S be a collection of subsets of $7. Then S is called a d-system if it has the following properties: (a) fle5. (b) If A and B belong to <S and if A 2 B, then A \ B belongs to S. (c) If An, n € N, is an increasing sequence of elements of S, then the union I X L I -An belongs to S.
Definition 5.2 Let Q, be a set and let S be a collection of subsets of 0 . Then S is called a 7r-system if it is closed under finite intersections. It is clear that the intersection of any family of d-systems is a d-system. Let S be a collection of subsets of CI. Then the smallest d-system containing S is called the d-system generated by S. The next assertion is the monotone class theorem for sets. Theorem 5.1 Let M be a K-system of subsets of Si. Then the d-system generated by Ai coincides with the a-algebra generated by Ai. Next we formulate the monotone class theorem for functions. Theorem 5.2 Let Q be a set and let Ai be a ir-system of subsets of CI. Let H be a vector space of real valued functions on Cl satisfying the following condition: (i) The constant function 1 belongs to 7i. (ii) For any A e M, XA £ W . 325
326
Non-Autonomous Koto Classes and Feynman-Kac Propagators
(Hi) If fn, n G N, is an increasing sequence of non-negative functions in H such that f = sup n e N / „ is finite (bounded), then f £ H , Then Ti contains all real valued functions (all real valued bounded functions) on fi, which are a(A4)-measurable. Theorems 5.1 and 5.2 are often used in the following setting. Let fl be a set, and let (Ei, Ei)i£l be a family of measurable spaces, indexed by a set I. Suppose that for every i G 7, a 7r-system Si of subsets of Ei generating Ei is given. Suppose also that for every i G I, fi is a mapping from fi into Ei. Then the following two assertions hold. Theorem 5.3 Let M be the collection of all sets of the form C\i€J fr1(J^i)> where A{ € Si, i G J, and J is a finite subset of I. Then M is a n-system, and moreover, a(M) =
a(fi-.ieI).
Theorem 5.4 Let Jibe a vector space of real valued functions on CI such that the following conditions hold: (i) The constant function 1 belongs to Ti. (ii) Ifhn,n€N, is an increasing sequence of non-negative functions in Ti such that h = sup„ hn is finite (bounded), then h belongs to Ti. (Hi) Ti contains all products of the form Y\i€J XAi°fi, where Ai G Si, i G J, and J is a finite subset of I. Then Ti contains all real valued functions (all bounded real valued functions) which are measurable with respect to the a-algebra CT(/, : i G I). In our presentation of the Markov property, I = [0,T], Et = E for all t G [0,T] where E is the state space of the process, and the maps ft, t G [0,T], are the state variables Xt. We refer the reader to [Blumenthal and Getoor (1968); Sharpe (1988)] for more information on the monotone class theorems. 5.2
Kolmogorov's Extension Theorem
Let £ be a locally compact second countable Hausdorff topological space equipped with the Borel cr-algebra £. Let I = [0,T]. For every finite subset J of the set I, let P j be a probability measure on the measurable space (EJ, BEJ). A family {Pj : J c I, J finite} of such probability measures is
Some Theorems of Analysis and Probability
Theory
327
called a projective or consistent family on E1 = ]JteI E, provided that for any pair J C K of finite subsets of /, and for any set A G BE J , the equality Pj(A) = PK {(wj)jeK
G EK : ( W j ) i € J G A}
(5.1)
holds. For J c i f as above, define the projection K
pK :E
^ EJ
(5.2)
by the following:
Then equality (5.1) can be rewritten as follows:
VJ(A) =
VK{(rfy1(A)}.
Note that it is not necessary to assume that the sets in (5.2) are finite. The set E1 will be equipped with the cr-algebra T generated by the mappings {pj : J
Uniform Integrability
Let (S, A, v) be a measure space. Definition 5.3 A family of functions {fj : j G J } in L1 (S, A, v) is called uniformly integrable if for every e > 0 there exists 5 > 0 such that sup / \fj\du jeJ J A whenever A G A and v(A) < 5.
<s
328
Non-Autonomous Koto Classes and Feynman-Kac Propagators
It is clear that if the family {fj : j £ J} is uniformly integrable, and if {gj : j £ J } is such that for every j £ J, \gj\ < \fj\ v-almost everywhere, then the family {gj : j £ J } is uniformly integrable. It is also clear that Cauchy sequences in L1 (S, A, v) are uniformly integrable. Next we give an example of a family of functions which is not uniformly integrable. Let / > 0 be a function from the space L 1 (R , B^d,m) where m is the Lebesgue measure on Rd. Suppose / f(x)dm(x) > 0 and limn^O0ndf(nx) = 0 for all x ^ 0, and put /„(x) = ndf(nx), n £ N. Then the sequence / „ is not uniformly integrable (see [Meyer (1966)] for more information on the uniform integrability of families of functions).
5.4
Radon-Nikodym Theorem
Let (S, A) be a measurable space, and let v\ and v^ be two measures on A. By definition, the measure v
Some Theorems of Analysis and Probability
Theory
329
The random variable G is called the conditional expectation of the random variable F given the c-algebra FQ. It is defined P-almost surely, and is denoted by G = E [F | To]. We refer the reader to [Folland (1999)] for more information on the Radon-Nikodym theorem.
5.5
Vitali-Hahn-Saks Theorem
Let (S, A) be a measurable space, and let v be a signed real-valued measure on A. By \u\ will be denoted the variation of v. By Corollary 5 on p.127 in [Dunford and Schwartz (1988)], the measure v is of bounded variation. The Vitali-Hahn-Saks theorem concerns setwise convergent sequences of signed measures of bounded variation which are absolutely continuous with respect to a non-negative measure A. T h e o r e m 5.7 Let (£,.4) be a measurable space, and let X be a nonnegative finite measure on A. Suppose that vn is a sequence of signed realvalued measures on A such that for every n > 1, the measure vn is absolutely continuous with respect to the measure A and the limit lim„_ 0 o vn(A) exists for every A € A. Then lim vn(A) = 0 uniformly for n > 1. |i/(A)|-»0
The following corollary to Theorem 5.7 is useful. It is due to Nikodym. Corollary 5.2 Let £ n be a sequence of signed real-valued measures on (S,A) such that the limit ^{A) = lim £n(A) exists for all A £ A. Then n—>oo
the sequence {£n} is uniformly countably additive and the set function £ is countably additive on A. We refer the reader to [Dunford and Schwartz (1988)] for more information on the Vitali-Hahn-Saks theorem. 5.6
Doob's Inequalities
The next assertion contains Doob's inequalities for right-continuous martingales (see, e.g., [Revuz and Yor (1991)]). T h e o r e m 5.8 Let Xt, 0 < t < T, be a right-continuous martingale on (fi, ?, Gu P). Then for allp>\ and A > 0, ApP
sup \Xt\ > A < E [ | X T | P ] . te[o,T]
330
Non-Autonomous Kato Classes and Feynman-Kac Propagators
Moreover, for allp > 1, sup \Xt\ te[o,T]
^£-=TII*HI„.
Bibliography
Acquistapace, P. (1993). Abstract Linear Nonautonomous Parabolic Equations: A Survey, in: Differential Equations in Banach Spaces: Proc. of the Bologna Conference (eds. G. Dore, A. Favini, E. Obrecht, and A. Venni), Led. Notes in Pure and Appl. Math. 148, Marcel Dekker, Inc., New York, pp. 1-19. Aebi R. (1996). Schrodinger Diffusion Processes, Probability and its Applications, Birkhauser, Basel. Aizenman, N. and Simon, B. (1982). Brownian motion and Harnack's inequality for Schrodinger operators, Comm. Pure Appl. Math. 35, pp. 209-273. Albeverio, S., Blanchard, Ph., and Ma, Z. M. (1991). Feynman-Kac semigroups in terms of signed smooth measures, in: Proc. Random Partial Differential Equations, Internat. Series of Numerical Math., 102, Birkhauser, Boston, pp. 1-31. Albeverio, S. and Ma, Z. M. (1991). Additive functionals, nowhere Radon and Kato class smooth measures associated with Dirichlet forms, Osaka J. Math. 29, pp. 247-265. Applebaum, D. (2004). Levy Processes and Stochastic Calculus, Cambridge Studies in Advanced Mathematics 93, Cambridge University Press. Aronson, D. G. (1967). Bounds on the fundamental solution of a parabolic equation, Bull. Amer. Math. Soc. 73, pp. 890-896. Aronson, D. G. (1968). Non-negative solutions of linear parabolic equations, Ann. Scuola Norm. Sup. Pisa (3) 22, pp. 607-694. Barndorff-Nielsen, O. E., Mikosch, T., Resnick, S. I. (Editors) (2001), Levy Processes, Birkhauser. Bernstein, S. (1932). Sur les liaisons entre les grandeurs aleatoires, Vehr. des intern. Mathematikerkongr., Band 1, Zurich. Bertoin, J. (1996). Levy Processes, Cambridge Univ. Press, Melbourne, NY. Beznea, L. and Boboc, N. (2000). Excessive kernels and Revuz measures, Probab. Theory Relat. Fields 117, pp. 267-288. Beznea, L. and Boboc, N. (2004). Fine densities for excessive measures and the Revuz correspondence, Potential Analysis 20, pp. 61-83. Bhattacharya, R. N. and Waymire E. C. (1990). Stochastic Processes with Applications, Wiley Series in Probability and Mathematical Statistics, A Wiley331
332
Non-Autonomous Koto Classes and Feynman-Kac Propagators
Intescience Publication, John Wiley & Sons, New York. Blanchard, Ph. and Ma, Z. M. (1990a). Semigroup of Schrodinger operators with potentials given by Radon measures, in: Stochastic Processes-Physics and Geometry (eds. S. Albeverio, et al.), World Scientific Publishing Co., Inc., Teaneck, NJ, pp. 160-195. Blanchard, Ph. and Ma, Z. M. (1990b). New results on the Schrodinger semigroups with potentials given by signed smooth measures, in: Stochastic Analysis and Related Topics, II (Silivri, 1988) Led. Notes in Math. 1444, Springer-Verlag, Berlin, pp. 213-243. Blumenthal, R. M. and Getoor, R. K. (1968). Markov processes and potential theory, Pure and Applied Mathematics 29, Academic Press, New York. Bourbaki, N. (1965). Elements de Mathematiques, Livre III, topologie generale, 2nd edition, Chapter 9, Hermann, Paris. Carlen, E. A. (1984). Conservative diffusions, Comm. Math. PhysicsQA, pp. 293315. Carmona, R. (1974). Regularity properties of Schrodinger and Dirichlet semigroups, J. Fund. Anal. 17, pp. 227-237. Carmona, R., Masters, W. C , and Simon, B. (1990). Relativistic Schrodinger operators: asymptotic behavior of the eigenfunctions, J. Fund. Anal. 91, pp. 117-142. Cerrai, S. (2001). Second Order PDE's in Finite and Infinite Dimension. A Probabilistic Approach, Springer-Verlag, Berlin. Chicone, C. and Latushkin, Y. (1999). Evolution Semigroups in Dynamical Systems and Differential Equations, American Mathematical Society, Providence, RI. Chung, K. L. (1982). Lectures from Markov Processes to Brownian Motion, Grundlehren der Mathematischen Wis sens chaften 249, Springer-Verlag, New York. Chung, K. L. and Doob, J. L. (1965). Fields, optionality and measurability, Amer. J. Math. 87, pp. 397-424. Chung, K. L. and Walsh, J. B. (1969). To reverse a Markov process, Ada Math. 123, pp. 225-251. Chung, K. L. and Walsh, J. B. (2005). Markov Processes, Brownian Motion, and Time Symmetry, Grundlehren der Mathematischen Wissenschaften 249, Second edition, Springer-Verlag, New York. Chung, K. L. and Williams R. J. (1990). Introduction to Stochastic Integration, Probability and its Applications, Birkhauser, Boston, MA. Chung, K. L. and Zambrini, J.-C. (2003). Introduction to Random Time and Quantum Randomness, Monographs of the Portuguese Mathematical Society 1, New edition, World Sci. Publishing, River Edge, NJ. Chung, K. L. and Zhao, Z. X. (1995). From Brownian Motion to Schrodinger's Equation, Grundlehren der Mathematischen Wissenschaften 312, SpringerVerlag, Berlin. Crandall, M. G. and Lions, P. L. (1983). Viscosity solutions of Hamilton-Jacobi equations. Trans. Amer. Math. Soc. 277, pp. 1-42. Crandall, M. G., Ishii, H., and Lions, P.-L. (1992). User's guide to viscosity
Bibliography
333
solutions of second order partial differential equations, Bull. Amer. Math. Soc. (N.S.) 27, pp. 1-67. Cruzeiro, A. B. and Zambrini, J.-C. (1994). Euclidean Quantum Mechanics. An outline, A. L. Cardoso et al. (eds.), Stochastic Analysis and Applications in Physics, Kluwer Academic Publishers, pp. 59-97. Cruzeiro, A. B., Wu, Liming, and Zambrini, J. C. (2000). Bernstein processes associated with a Markov process, in: Stochastic Analysis and Mathematical Physics, ANESTOC'98, Proceedings of the Third International Workshop, Ed. R. Rebolledo, Trends in Mathematics Series, Birkhauser, Boston, 2000. Davies, E. B. (1997). L p -spectral theory of higher-order elliptic differential operators, Bull. London Math. Soc. 29, pp. 513-546. Dellacherie, C. and Meyer, P.-A. (1978). Probabilities and Potential, NorthHolland, Amsterdam. Demuth, M. and van Casteren, J. A. (2000). Stochastic Spectral Theory for Selfadjoint Feller Operators: A functional integration approach, Probability and its Applications, Birkhauser, Basel. Doob, J. L. (2001). Classical Potential Theory and Its Probabilistic Counterpart, Classics in Mathematics, Reprint of the 1984 edition, Springer-Verlag, Berlin. Dressel, F. G. (1940). The fundamental solution of the parabolic equation, Duke Math. J. 7, pp. 186-203. Dressel, F. G. (1946). The fundamental solution of the parabolic equation, II Duke Math. J. 13, pp. 61-77. Dunford, N. and Schwartz, J. T. (1988). Linear Operators, Part I: General Theory, A Wiley-Interscience Publication, John Wiley & Sons, New York. Durrett, R. (1984). Brownian Motion and Martingales in Analysis, Wadsworth Mathematics Series, Wadsworth International Group, Belmont, CA. Durrett, R. (1991). Probability: Theory and Examples, The Wadsworth & Brooks/Cole Statistics/Probability Series Wadsworth & Brooks/Cole Advanced Books & Software, Pacific Grove, CA. Durrett, R. (1996). Stochastic Calculus, A practical introduction, Probability and Stochastic Series, CRC Press, Boca Raton. Dynkin, E. B. (1960). Theory of Markov Processes, Pergamon Press, OxfordLondon-New York-Paris. Dynkin, E. B. (1965). Markov processes. Vols. I, II, Grundlehren der Mathematischen Wissenschaften 121, 122, Academic Press Inc., Publishers, New York. Dynkin, E. B. (1973). Regular Markov processes (Russian), Uspekhi Mat. Nauk 28, pp. 35-64. English translation in: Markov Processes and Related Problems of Analysis, London Math. Soc. Lecture Notes Ser. 54, Cambridege Univ. Press, Cambridge, 1982, pp. 187-218. Dynkin, E. B. (1982). Markov Processes and Related Problems of Analysis, London Mathematical Society Lecture Note Series 54, Cambridge Univ. Press, Cambridge. Dynkin, E. B. (1994). An Introduction to Branching Measure-Valued Processes, CRM Monograph Series 6, American Math. Society, Providence, Rhode
334
Non-Autonomous Koto Classes and Feynman-Kac Propagators
Island. Dynkin, E. B. (2000). Selected papers of E. B. Dynkin with commentary, Edited by A. A. Yushkevich, G. M. Seitz and A. L. Onishchik, Amer. Math. Soc, Providence, R.I. Dynkin, E. B. (2002). Diffusions, Superdiffusions and Partial Differential Equations, American Mathematical Society Colloquium Publications 50, Amer. Math. Soc, Providence, RI. Dynkin, E. B. and Yushkevich, A. A. (1969). Markov Processes: Theorems and Problems, Plenum Press, New York. Eberle, A. (1999). Uniqueness and Non-uniqueness of Semigroups Generated by Singular Diffusion Operators, Springer, Lecture Notes in Math. 1718, Springer-Verlag, Berlin. Eidel'man, S. D. (1969). Parabolic Systems, North-Holland Publishing Company, Amsterdam-London, Wolters-Noordhoff Publishing, Groningen. Eidel'man, S. D. and Zhitarashu, N. V. (1998). Parabolic Boundary Value Problems, Birkhauser, Basel. Engel, K.-J. and Nagel, R. (2000). One-Parameter Semigroups for Linear Evolution Equations, Graduate Texts in Mathematics, 194, With contributions by S. Brendle, M. Campiti, T. Hahn, G. Metafune, G. Nickel, D. Pallara, C. Perazzoli, A. Rhandi, S. Romanelli and R. Schnaubelt, Springer-Verlag, New York. Ethier, S. N. and Kurtz, T. G. (1986). Markov Processes: Characterization and convergence, Wiley Series in Probability and Mathematical Statistics: Probability and Mathematical Statistics, John Wiley & Sons Inc., New York. Fabes, E. B. (1993). Gaussian upper bounds on fundamental solutions of parabolic equations; the method of Nash, in: Dirichlet Forms (Varenna, 1992), Lecture Notes in Math. 1563, Springer, Berlin, pp. 1-20. Fabes, E. B. and Stroock, D. W. (1986). A new proof of Moser's parabolic Harnack inequality using the old ideas of Nash, Arch. Rat. Mach. Anal. 96, pp. 327338. Folland, G. B. (1999). Real Analysis: Modern techniques and their applications, A Wiley-Interscience Publication, John Wiley & Sons Inc., New York. Freidlin, M. (1985). Functional Integration and Partial Differential Equations, Ann. of Math. Studies 109, Princeton University Press, Princeton, N.J. Friedman, A. (1964). Partial Differential Equations of Parabolic Type, PrenticeHall Inc., Englewood Cliffs, N.J. Friedman, A. (1975). Stochastic Differential Equations and Applications, Vol. 1, Probability and Mathematical Statistics 28, Academic Press [Harcourt Brace Jovanovich Publishers], New York. Friedman, A. (1976). Stochastic Differential Equations and Applications, Vol. 2, Probability and Mathematical Statistics 28, Academic Press [Harcourt Brace Jovanovich Publishers], New York. Fukushima, M., Oshima, Y., and Takeda, M. (1994). Dirichlet Forms and Symmetric Markov processes, de Gruyter Studies in Mathematics 19, Walter de Gruyter & Co., Berlin. Getoor, R. K. (1999). Measure perturbations of Markovian semigroups, Potential
Bibliography
335
Anal. 11, pp. 101-133. Gihman, I. I. and Skorohod, A. V. (1974). The Theory of Stochastic Processes I, Springer-Verlag, Berlin. Gihman, I. I. and Skorohod, A. V. (1975). The Theory of Stochastic Processes II, Springer-Verlag, New York. Gihman, I. I. and Skorohod, A. V. (1979). The Theory of Stochastic Processes III, Springer-Verlag, Berlin. Glover, J., Rao, M., and Song R. (1993). Generalized Schrodinger semigroups, Seminar on Stochastic Processes (Seattle, WA, 1992), Progr. Probab. 33, Birkhauser, Boston, pp. 143-172. Glover, J., Rao, M., Sikic, H., and Song, R. (1994). Quadratic forms corresponding to the generalized Schrodinger semigroups, J. Fund. Anal. 125, pp. 358-378. Goldstein, J. A. (1985). Semigroups of Linear Operators and Applications, Oxford Mathematical Monographs, Oxford University Press, Oxford. Gulisashvili, A. (2000). Sharp estimates in smoothing theorems for Schrodinger semigroups, J. Fund. Anal. 170, pp. 161-187. Gulisashvili, A. (2002a). On the heat equation with a time-dependent singular potential, J. Fund. Anal. 194, pp. 17-52. Gulisashvili, A. (2002b). Classes of time-dependent measures and the behavior of Feynman-Kac propagators, C. R. Math. Acad. Sci. Paris 334, pp. 445-449. Gulisashvili, A. (2002c). On the Kato classes of distributions and the BMOclasses, in: Differential Equations and Control Theory, Led. Notes in Pure and Applied Mathematics 225 (S. Aizicovici and N. Pavel, eds.), Marcel Dekker, New York, pp. 159-176. Gulisashvili, A. (2004a). Free propagators and Feynman-Kac propagators, Seminar of Mathematical Analysis, Proceedings, Universities of Malaga and Seville (Spain), September 2003-June 2004, Daniel Girela Alvarez, Genaro Lopez Acedo, Rafael Villa Caro (editors), Secretariado de Publicaciones, Universidad de Sevilla, pp. 47-64. Gulisashvili, A. (2004b). Markov processes and Feynman-Kac propagators, Centre de Recerca Matemdtica Preprint Series bf 573, April 2004, 48 pages. Gulisashvili, A. (2004c). Markov processes, classes of time-dependent measures, and Feynman-Kac propagators, submitted for publication. Gulisashvili, A. (2005). Nonautonomous Kato classes of measures and FeynmanKac propagators, Trans. Arner. Math. Soc. 357, pp. 4607-4632. Gulisashvili, A. and Kon, M. (1996). Exact smoothing properties of Schrodinger semigroups, Amer. J. Math. 118, pp. 1215-1248. Gulisashvili, A. and Van Casteren, J.A. (2005). Feynman-Kac propagators and viscosity solutions, J. Evol. Equ. 5, pp. 105-121. Howland, J. S. (1974). Stationary scattering theory for time-dependent Hamiltonians Math. Ann. 207, pp. 315-335. Ikeda, N. and Watanabe, S. (1989). Stochastic Differential Equations and Diffusion Processes, Second edition, North-Holland Mathematical Library 24, North-Holland, Amsterdam. Il'in, A. M., Kalashnikov, A. S., and Oleinik, O. A. (1962). Linear Equations of
336
Non-Autonomous Kato Classes and Feynman-Kac Propagators
the Second Order of Parabolic Type, Russian Math. Surveyes 17, pp. 1-143. Ito, K. (1951). On Stochastic Differential Equations, Mem. Amer. Math. Soc. 4, Amer. Math. Soc, Providence, R.I. Ito, K. (1987). Selected papers, Edited and with an introduction by S. R. S. Varadhan and Daniel W. Stroock, Springer, New York. Ito, K. and McKean, Jr., H. P. (1965). Diffusion Processes and Their Sample Paths, Grundlehren der Mathematischen Wissenschaften Band 125, Academic Press Inc., Publishers, New York. Jacob, N. (2001). Pseudo Differential Operators and Markov Processes. Vol. I, Fourier Analysis and Semigroups, Imperial College Press, London. Jacob, N. (2002). Pseudo Differential Operators and Markov processes. Vol. II, Generators and Their Potential Theory, Imperial College Press, London. Jacob, N. (2005). Pseudo Differential Operators and Markov Processes. Vol. Ill, Markov Processes and Applications, Imperial College Press, London. Jacod, J., and Shiryaev, A. N. (1987). Limit Theorems for Stochastic Processes, Springer-Verlag, 1987. Jamison, B. (1970). Reciprocal processes: The stationary Gaussian case, Ann. Math. Statist. 4 1 , pp. 1624-1630. Jamison, B. (1974). Reciprocal processes, Z. Wahrscheinlichkeitstheorie und Verm. Gebiete 30, pp. 65-86. Jamison, B. (1975). The Markov processes of Schrodinger, Z. Wahrscheinlichkeitstheorie und Verw. Gebiete 32, pp. 323-331. Johnson, G. W. and Lapidus, M. L. (2000). The Feynman Integral and Feynman's Operational Calculus, Oxford Mathematical Monographs, Oxford Science Publications, Clarendon Press, Oxford. Kac, M. (1949). On distributions of certain Wiener functionals, Trans. Amer. Math. Soc. 65, pp. 1-13. Kac, M. (1951). On some connections between probability theory and differential and integral equations. In: Proc. Second Berkeley Symposium on Mathematical Statistics and Probability (ed. J. Neyman), University of California Press, Berkeley, pp. 189-215. Kac, M. (1959). Probability and Related Topics in Physical Sciences, With special lectures by G. E. Uhlenbeck, A. R. Hibbs, and B. van der Pol. Lectures in Applied Mathematics, Proceedings of the Summer Seminar, Boulder, Colorado 1, Interscience Publishers, London-New York. Kac, M. (1979). Mark Kac: Probability, Number Theory, and Statistical Physics, Mathematicians of Our Time 14, Selected papers, Edited by K. Baclawski and M. D. Donsker, MIT Press, Cambridge, Mass.-London. Kahane, J.-P. (1997). A century of interplay between Taylor series, Fourier series and Brownian motion, Bull. London Math. Soc. 29, pp. 257-279. Kahane, J.-P. (1998). Le mouvement brownien, Seminaires et Congres 3, pp. 123— 155. Karatzas, I. and Shreve, S. (1991). Brownian Motion and Stochastic Calculus, Graduate Texts in Mathematics 113, Second edition, Springer-Verlag, New York. Kato, T. (1973). Schrodinger operators with singular potentials, Israel J. Math.
Bibliography
337
13, pp. 135-148. Kleinert, H. (2004). Path Integrals in Quantum Mechanics, Statistics, Polymer Physics, and Financial Markets, Third edition, World Scientific Publishing Co. Inc., River Edge, N.J. Kolmogorov, A. N. (1931). Uber die analytischen Methoden in der Wahrscheinlichkeitsrechnung, Math. Ann. 104, pp. 415-488. English translation in: Selected Works of A. N. Kolmogorov, Vol. II, Kluwer Acadamic Publishers, pp. 62-108. Kolmogorov, A. N. (1933). Zur Theorie der stetigen zufalligen Prozesse, Math. Ann. 108, pp. 149-160. English translation in: Selected Works of A. N. Kolmogorov, Vol. II, Kiuwer Acadamic Publishers, pp. 156-168. Kuznetsov, S. E. (1982). Nonhomogeneous Markov processes (Russian), Sovremennye Problemy Matematiki 20, VINITI, Moscow, pp. 37-178. English translation in: J. Soviet Math. 25 (1984), pp. 1380-1498. Ladyzenskaja, O. A., Solonnikov, V. A., and Ural'ceva, N. N. (1968). Linear and Quasilinear Equations of Parabolic Type, Translations of Mathematical Monographs 23, Amer. Math. Soc, Providence, R.I. Lamberton, D. and Lapeyre, B. (1996). Introduction to Stochastic Calculus Applied to Finance, Chapman & Hall/CRC, London. Liggett, T. M. (2005). Interacting Particle Systems, Springer, New York. Liskevich, V. and Semenov Y. (2000). Estimates for fundamental solutions of second-order parabolic equations, J. London Math. Soc. (2) 62, pp. 521543. Liskevich, V., Vogt, H. and Voigt, J. (2005). Gaussian bounds for propagators perturbed by potentials, submitted for publication. Metivier, M. and Pellaumail, J. (1980). Stochastic Integration. Probability and Mathematical Statistics, Academic Press [Harcourt Brace Jovanovich Publishers] , New York. Metivier, M. and Viot, M. (1987). On weak solutions of stochastic partial differential equations, Lecture Notes in Math. 1322, Springer-Verlag, pp. 139-150. Meyer, P.-A. (1966). Probability and Potentials, Blaisdell, Waltham. Nagasawa, M. (1993). Schrodinger Equations and Diffusion Theory, Monographs in Mathematics 86, Birkhauser, Basel. Nagasawa, M. (2000). Stochastic Processes in Quantum Physics, Monographs in Mathematics 94, Birkhauser, Basel. Nagel, R. (1995). Semigroup methods for non-autonomous Cauchy problems, in: Evolution Equations, eds. G. Ferreyra et al. Led. Notes Pure Appl. Math. 168, Marcel Dekker, New York, pp. 301-316. Nagel, R. and Nikel, G. (2002). Wellposedness for Nonautonomous Abstract Cauchy Problems, in: Progress in Nonlinear Differential Equations and Their Applications, 50, Birkhauser Verlag, Basel, pp. 279-293. Nash J. (1958). Continuity of solutions of parabolic and elliptic equations, Amer. J. Math. 80, pp. 931-954. Nelson, E. (1967). Dynamical Theories of Brownian Motion, Math. Notes, Princeton University Press, Princeton, N.J. Nelson, E. (1985). Quantum Fluctuations, Princeton Series in Physics, Princeton
338
Non-Autonomous Kato Classes and Feynman-Kac Propagators
University Press, Princeton, N.J. Nickel, G. (1997). Evolution semigroups for nonautonomous Cauchy problems, Abstract and Applied Analysis 2, pp. 73-95. 0ksendal, B. (1998). Stochastic Differential Equations, Universitext, Fifth edition, Springer-Verlag, Berlin. Ouhabaz, E. M., Stollmann, P., Sturm, K.-Th., and Voigt, J. (1996). The Feller property for absorption semigroups, J. Fund. Anal. 138, pp. 351-378. Pazy, A. (1983). Semigroups of Linear Operators and Applications to Partial Differential Equations, Springer-Verlag, New York. Porper, F. O. and Eidel'man, S. D. (1984). Two-sided estimates of the fundamental solutions of second-order parabolic equations, and some applications, Uspekhi Mat. Nauk 39, pp. 107-156 (in Russian). English translation in Russian Math. Surveyes 39, (1984), pp. 119-178. Privault, N. and Zambrini, J.-C. (2004). Markovian bridges and reversible diffusion processes with jumps, Ann. I. H. Poincare - PR 40, pp. 599-633. Privault, N. and Zambrini, J.-C. (2005). Euclidean Quantum Mechanics in the momentum representation, J. Math. Phys. 46, 25 p. Protter, P. (2005). Stochastic Integration and Differential Equations, Second Edition, Stochastic Modelling and Applied Probability 21 Springer-Verlag, Berlin. Rabiger, F., Rhandi, A., Schnaubelt, R., and Voigt, J. (2000). Non-autonomous Miyadera perturbations, Differential Integral Equations 13, pp. 341-368. Revuz, D. and Yor, M. (1991). Continuous Martingales and Brownian Motion, Grundlehren der Mathematischen Wissenschaften 293, Springer-Verlag, Berlin. Roelly, S. and Thieullen, M. (2002). A characterization of reciprocal processes via an integration by parts formula on the path space, Probab. Theory Related Fields 123, pp. 97-120. Roelly, S. and Thieullen, M. (2005). Duality formula for the bridges of a Brownian diffusion: applications to gradient drifts, Stoch. Proc. Appl. 115, pp. 16771700. Rogers, L. C. G. and Williams, D. (2000a). Diffusions, Markov Processes, and Martingales. Vol. 1. Foundations. Cambridge Mathematical Library, Cambridge University Press, Cambridge. Rogers, L. C. G. and Williams, D. (2000). Diffusions, Markov Processes, and Martingales. Vol. 2. Ito Calculus. Cambridge Mathematical Library, Cambridge University Press, Cambridge. Royden, H. L. (1988). Real Analysis, Third edition, Macmillan Publishing Company, New York. Sato, K. (2000). Levy Processes and Infinitiely Divisible Distributions, Cambridge University Press, Cambridge. Schechter, M. (1971). Spectra of Partial Differential Operators, North-Holland, Amsterdam. Schnaubelt, R. (2000/2001). Well-posedness and asymptotic behavior of nonautonomous linear evolution equations, Tiibinger Berichte zur Funktionalanalysis 10, pp. 195-218.
Bibliography
339
Schnaubelt, R. (2000). Semigroups for nonautonomous Cauchy problems, in: K. Engel and R. Nagel, One-Parametes Semigroups for Linear Evolution Equations, Springer-Verlag, pp. 477-496. Schnaubelt, R. and Voigt, J. (1999), The non-autonomous Kato class, Arch. Math. 72, pp. 454-460. Schoutens, W. (2003). Levy Processes in Finance. Pricing Financial Derivatives, John Wiley & Sons. Schrodinger (1931). Uber die Umkehrung der Naturgesetze. Sitzungsber. Preuss. Akad. Wiss. Berlin Phys. Math. Kl. 8 / 9 , pp. 144-153. Semenov, Yu. A. (1999). On perturbation theory for linear elliptic and parabolic operators; the method of Nash, Applied Analysis, J. R. Doroh, G. Riuz Goldstein, J. A. Goldstein, M. M. Tom, Editors, AMS, Providence, RI, Contemporary Mathematics 221, pp. 217-284. Sharpe, M. (1988). General Theory of Markov processes, Pure and Applied Mathematics 133, Academic Press Inc., Boston, MA. Simon, B. (1979). Functional Integration and Quantum Physics, Pure and Applied Mathematics 86, Academic Press Inc. [Harcourt Brace Jovanovich Publishers], New York. Simon, B. (1982), Schrodinger semigroups, Bull. Amer. Math. Soc. (N.S.) 7, pp. 445-526. Sobolevskii, P. E. (1961). Equations of parabolic type in a Banach space, Trudy Moscov. Math. Obsc. 10, pp. 297-350 (Russian); English transl.: Amer. Math. Soc. Transl. 49 (1965), pp. 1-62. Stannat, W. (2004). Time-dependent diffusion operators on L1, J. Evol. Equ., 4, pp. 463-495. Stein, E. and Weiss, G. (1971). Introduction to Fourier Analysis on Euclidean Spaces, Princeton Mathematical Series 32, Princeton University Press, Princeton, N.J. Stroock, D. W. (1987). Lectures on Stochastic Analysis: Diffusion Theory, London Mathematical Society Student Texts 6, Cambridge University Press, Cambridge. Stroock, D. W. (1993). Probability Theory, an Analytic View, Cambridge University Press, Cambridge. Stroock D. W. (2003). Markov processes from K. Ito's perspective, Ann. of Math. Stud. 155, Princeton Univ. Press, Princeton, NJ. Stroock D. W. (2005). An Introduction to Markov Processes, Graduate Texts in Mathematics 230, Springer-Verlag, Berlin, 2005. Stroock, D. W. and Varadhan, S. R. S. (1979). Multidimensional Diffusion Processes, Grundlehren der Mathematischen Wissenschaften 233, SpringerVerlag, Berlin. Stummel, F. (1956). Singulare elliptische differentialoperatoren in Hilbertschen Raumen, Math. Ann. 132, pp. 150-176. Sturm, K.-Th. (1994). Harnack's inequality for parabolic operators with singular low order terms, Math. Z. 216, pp. 593-611. Tanabe, H. (1960a). Remarks on the equations of evolution in a Banach space, Osaka Math. J. 12, pp. 145-166.
340
Non-Autonomous Kato Classes and Feynman-Kac Propagators
Tanabe, H. (1960b). On the equations of evolution in a Banach space, Osaka Math. J. 12, pp. 363-376. Tanabe, H. (1997). Functional Analytic Methods for Partial Differential Equations, Marcel Dekker, Inc., New York. Thieullen, M. (1993). Second order stochastic differential equations and non gaussian reciprocal diffusions, Probab. Theory Related Fields 97, pp. 231-257. Thieullen, M. (1998), Reciprocal diffusions and symmetries, Stochastics Stochastics Rep. 65, pp. 41-77. Thieullen, M. (2002). Reciprocal diffusions and symmetries of parabolic PDE: the nonfiat case, Potential Anal. 16, pp. 1-28. Thieullen, M. and Zambrini J. C. (1997). Probability and quantum symmetries 1. The theorem of Noether in Schrodinger's Euclidean quantum mechanics, Ann. Inst. H. Poincare 67, 3, pp. 297-338. Truman, A. and Davies, I. M. (1988). Stochastic Mechanics and Stochastic Processes, Lecture Notes in Mathematics 1325, Truman, A. and Davies, I. M. (editors), Springer-Verlag, Berlin. van Casteren, J. A. (2000). Some problems in Stochastic Analysis and Semigroup Theory, Progress in Nonlinear Differential Equations and Their Applications, Vol. 42, Birkhauser-Verlag, Basel, pp. 43-60. van Casteren, J. A. (2002). Markov pocesses and Feller semigroups, Conferenze del Seminario di Matematica dell'Universita di Bari 286, Aracne, Roma. Voigt, J. (1986). Absorption semigroups, their generators, and Schrodinger semigroups, J. Fund. Anal. 67, pp. 167-205. Voigt, J. (1995), Absorption semigroups, Feller property, and Kato class, in: Partial Differential Operators and Mathematical Physics (Holzhau, 1994). Oper. Theory Adv. Appl. 78, pp. 389-396, Birkhauser, Basel. Wilansky, A. (1983). Topology for Analysis, Robert E. Krieger Publishing Co. Inc., Melbourne, FL. Yeh, J. (1995). Martingales and Stochastic Analysis, Series on Multivariate Analysis 1, World Scientific, Singapore. Zhang, Qi (1996). On a parabolic equation with a singular lower order term, Trans. Amer. Math. Soc. 348, pp. 2811-2844. Zhang, Qi (1997). On a parabolic equation with a singular lower order term II. The Gaussian bounds, Indiana Univ. Math. J. 46, pp. 989-1020. Zhang, Qi (2003). A sharp comparison result concerning Schrodinger heat kernels, Bull. London Math. Soc. 35, pp. 461-472. Zhang, T. S. (2001). Generalized Feynman-Kac semigroups, associated quadratic forms and asymptotic properties, Potential Analysis 14, pp. 387-408.
Index
7r-system, 7, 325 cr-algebra, 2 0 | ' v , 138 generated by a stopping time, 134 d-system, 325
backward flow conditions, 101 backward Kolmogorov representation, 79 backward propagator, 101 Feller, 124, 252 Feller-Dynkin, 253 left generators, 107 right generators, 108 backward transition function, 11 Brownian bridge, 93 Brownian motion, 90 orthogonal invariance, 91 scale invariance, 91 standard, 90 translation invariance, 91 Burkholder-Davis-Gundy inequalities, 237
additive functional Av, 256 exponential estimate, 269 additive functional AM, 257 existence, 261 exponential estimate, 269 uniqueness, 266 admissible family of stopping times, 156 examples, 170 approximation in the potential sense, 260, 286 augmentation, 8
Cauchy bridge, 97 Cauchy process, 96 scale invariance, 96 Chapman-Kolmogorov equation, 10, 62 for densities, 13 Choquet capacitability theorem, 180 Choquet capacity, 179 completion of a cr-algebra with respect to a family of measures, 30 with respect to a measure, 28 conditional expectation, 3, 328
backward Feynman-Kac propagator, 284 L s -boundedness, 288 (L s -L 9 )-smoothing property, 292 L°°-boundedness, 285 strong f?£/C-property, 293 strong Feller property, 293 backward Feynman-Kac propagator £[/C-property, 295 continuity properties, 296 DuhamePs formula, 308 Feller-Dynkin property, 295 integral kernels, 299 341
342
Non-Autonomous Kato Classes and Feynman-Kac Propagators
derived density, 68 differential operator divergence form, 222 non-divergence form, 217 diffusion process, 233, 251 covariance, 233 drift, 233 generator, 233 martingale characterization, 233 strong Markov property, 255 distribution, 2 Doob's inequality, 264, 273 entrance-exit law, 76 entry time, 178 Feynman-Kac formula, 279, 280 filtration, 7 final condition, 218 final value problem, 217 finite-dimensional distributions, 6, 15 floor and ceiling functions, 1 flow conditions, 101 Fokker-Planck equation, 119 forward Kolmogorov representation, 79 free backward propagator, 105 weak right generators, 115 weak left generators, 112 free propagator, 106 function space, 104 BC, 104 BUC, 104 Co, 104 LT, 105 LrE, 104
functional additive, 193, 255 Kac, 194 multiplicative, 194 fundamental solution, 218 weak, 224, 225 Gaussian density, 68 Gaussian estimates, 219
Gronwall lemma, 241 hitting time, 178 Howland semigroup associated with a backward propagator, 121 associated with a free backward propagator, 123 associated with a free propagator, 123 associated with a propagator, 123 associated with backward Feynman-Kac propagators, 305 independent
Index marginal distributions, 15 Markov process, 8, 15 progressively measurable, 255 reciprocal, 55 separable, 33 Markov property, 8, 9 reciprocal, 9 martingale, 27 Martingale Convergence Theorem, 27 martingale problem, 253 measure space complete, 29 monotone class theorem for functions, 325 for sets, 325 Nikodym's theorem, 329 partial ordering ^ , 156 path properties of Markov processes continuity, 51 one-sided continuity, 44 progressive measurability, 42 separability, 37 path properties of reciprocal processes continuity, 83 one-sided continuity, 82 path space, 6 pinned measure, 88, 299 associated with Brownian motion, 92 potential of a function, 196, 201 of a time-dependent measure, 196, 201 probability space, 2 projective system of measures, 6 propagator Feller-Dynkin, 172 Feynman-Kac, 284 left generators, 110 locally uniformly bounded, 102 right generators, 109 separately strongly continuous, 102 strongly continuous, 102 uniformly bounded, 102
343
pseudo-hitting time, 178 Radon-Nikodym theorem, 143, 302 Radon-Nikodym theorem, 328 random variable, 2 random variables independent, 3 reference measure, 13, 104 sample path, 6, 25 sample space, 2 Schrodinger operator, 279 Schrodinger representation, 79 Schrodinger semigroup, 279 space-time process, 18, 23 sample space, 19 standard process, 185 standard realization of a Markov process, 14 stochastic differential equation, 235, 237 unique solvability, 238 stochastic process, 8 adapted, 7, 27 continuous, 26 left-continuous, 26 measurable, 33 modifications, 7 progressively measurable, 34 quasi left-continuous, 175 right-continuous, 26 right-continuous with left-hand limits, 26 standard, 177 stochastically continuous, 26 strongly stochastically continuous, 26 time reversed, 9 stochastic processes, 5, 15 indistinguishable, 7 stochastically equivalent, 6, 15 stopping time, 134 terminal, 138 strong .St/C-condition, 124 strong Feller property, 124 strong Markov process, 149
344
Non-Autonomous Kato Classes and Feynman-Kac Propagators
strong Markov property with respect to families of stopping times and measures, 140, 150, 157 with respect to hitting times, 189 strongly subadditive function, 179 submartingale, 27 supermartingale, 27, 55 time reversal, 11 time shift operators, 17 time-dependent measure, 195 transition function, 10, 11 normal, 10 reciprocal, 62 time-homogeneous, 17 transition probability density, 13 transition probability function, 10 backward, 11 two-parameter filtration, 7 uniform integrability, 327 uniform parabolicity condition, 217 Urysohn's Lemma, 4 viscosity solutions, 311, 320, 321 Vitali-Hahn-Saks theorem, 118 Vitali-Hahn-Saks theorem, 329 weak solutions, 223, 225 Wiener space, 91
NON-AUTONOMOUS KATO CLASSES OP' This book provides an introduction to propagator theory. Propagators, or evolution families, are two-parameter analogues of semigroups of operators. Propagators are encountered in analysis, mathematical physics, partial differential equations, and probability theory. They are often used as mathematical models of systems evolving in a changing environment. A unifying theme of the book is the theory of Feynman-Kac propagators associated with time-dependent measures from non-autonomous Kato classes. In applications, a Feynman-Kac propagator describes the evolution of a physical system in the presence of time-dependent absorption and excitation. The book is suitable as an advanced textbook for graduate courses.
1
-
2
0
0
www.woridscientiJic.com