introduction for a number of years nonstandard measure theory … · 2020. 4. 22. · equilibrium...
TRANSCRIPT
- i -
Introduction
For a number of years nonstandard measure theory was in a
somewhat problematic state. Although many interesting papers
appeared (e.g. Bernstein and Wattenberg [5] , Loeb [14] ,[15] and
Bernstein and Loeb [4]), the theory suffered from the fact that
nonstandard measures refuse to be countably additive. This made
application to and comparison with standard measure- and proba
bility-theory difficult.
Many of these difficulties were resolved when Peter A. Loeb
in [16] showed how to turn *-measures into real-valued,countably
additive measures. This opened the way for the use of nonstandard
constructions in probability theory, and Loeb himself gave the first
example by constructing a Poisson process. Shortly afterwards R.M.
Anderson [1] used Loeb's technique to make a nonstandard construction
of a Brownian motion in a very simple and attractive way. Anderson
also gave a nonstandard treatment of stochastic integration with
respect to this Brownian motion process, and this theory was used
by H.J. Keisler ln [10] to obtain new results on stochastic differ~
ential equations.
Keisler [9] (see also Hoover [7]) based his hyperfinite model
theory on the Loeb-measure, and A.E. Hurd [8] used it in analyzing
equilibrium states in statistical mechanichs. In his thesis [2],
Anderson further developed the theory of Loeb-spaces, found a repre
sentation of Radon-spaces as inverse images of hyperfinite Loeb-spaces,
and gave applications to economics. This list does by no means exhaust
the applications of the Loeb-measure, but it should be long and
varied enough to indicate the fruitfulness of the idea.
In this paper we shall carry on the work on nonstandard stochas
tic integration begun by Anderson ln [1 ]. Let us first give a short
- ii -
sketch of standard stochastic integration. The problem is simply
to give a precise meaning to the heuristic formula fXdM where
X,M: :Rt x Q + :R are suitable processes on a probability space Q •
This is nontrivial since "typical" examples of M (e.g. Brownian
motions) have paths of unbounded variation, and the integral fXdM
thus can not be defined by pathwise Stieltjes integration. However,
it was realized by N. Wiener and K. Ito that in the Brownian motion
case the integral fXdM can be defined as an L2 -limit, and this
has later been extended to more general classes of processes M
(so called local L 2 -martingales; see Meyer [20] or Metivier [18]
for expositions.)
We now turn to Anderson' s work: Let n E *N ':N , and let
T = { k/ n: k E *:N, k < n 2} • Anderson constructed a hyperfini te
*-probability space Q and an internal process x : T x Q + *.R -
which in fact was nothing but a hyperfinite random walk - such that
the process 13 : :R+ x Q + :R defined by 0 -S(t,w) = x<t,w) , was a
Brownian motion (here t is the least element of T larger than t .
For a definition of Anderson's pi'Ocess, see Example 1 below). If
Y:TxQ+
integral
*.R
fYdx
is an internal process, he defined the stochastic k-1 . . .
to be the process (k/n ,w) + E Y(l/n ,w)(x(l+1 /n ,w)-x(l/n,w)); i=O
i.e. as a pathwise "Stiel tj es integral'~ Anderson was now able to
show that given a process X such that fXdS was defined by the
standard theory, he could find an internal process Y (called a
2-lifting of X ) such that Thus he could do by
nonstandard methods what had proved impossible by standard ones; to
write the stochastic integral as a pathwise Stieltjes integral. This
intuitive definition of the stochastic integral enabled him to
give a simple proof of Ito's Lemma. The work by Keisler [10] on
stochastic differential equations has later confirmed the importance
- iii-
of this approach to stochastic integration.
The purpose of this paper is to extend Anderson's results
about x to a larger class of internal processes, which we shall
call local A2 -martingales. We shall deal with properties of these
processes such as finiteness and (right- and left- ) continuity of
the paths; we shall define stochastic integrals with respect to
them in a way similar to that of Anderson; and we shall derive the
basic properties of these integrals, and prove nonstandard versions
of the transformation-formula and Ito's formula.
In another paper [12], we shall prove that the nonstandard
theory for stochastic integration with respect to such hyperfinite
martingales, is equivalent to the standard theory for stochastic
integration with respect to '~the standard parts '1 of the martingales.
In a third paper [13], we shall show that all standard local
L2 -martingales can - in a suitable sense - be represented as "the
standard part" of a local A2 -martingale; and we shall argue that
this combined with the results in [ 1 2] shows that anything that can be
done by the standard theory for stochastic integration, can be done
by the nonstandard theory. The hope is -of course - that the non
standard theory will be simpler and more intuitive to work with.
This and the subsequent papers are based on a common previous
version, the [11] of the references. That version is at times a
little more comprehensive and contains some additional material.
On the other hand it is rather unreadable.
Our main references will be Metivier [18] on stochastic inte
gration, and Stroyan and Luxemburg [23] on nonstandard analysis.
This paper and the other two mentioned above constitutes my
cand.real.-thesis written at the University of Oslo under the super-
- iv -
vision of Professor Jens Erik Fenstad. I would like to thank
Professor Fenstad for all his encouragement and good advice, and
I would also like -to ext..:.nd my thanks to all the membert?· of the
seminar on nonstandard analysis in Oslo during the last years.
- 1 -
1. Setting
We shall work with polysaturated models for nonstandard ana
lysis. That is we begin with a set K large enough to include all
the standard entities we shall need (e.g. the real numbers F, the
relevant measure spaces), and we construct the superstructure V(K)
over K. vJe then take *V(K) to be a card(V(K))-saturated non-
standard model of V(K) . This will be our polysaturated model.
For a proof of the existence of such models and a survey of their
properties, see Stroyan and Luxemburg [ 2 3 ] , sections 7. 4 - 7. 6.
Let <Q,G,P> be a *-measure space. We may define a finitely
additive measure 0 P on the algebra G by 0 P(A) = st(P(A)) for
all A EG • Peter A. Loeb proved in [ 16] that 0 P satisfies the
conditions of Caratheodory's Extension Theorem (Royden [22], page257)~
and that 0 P thus may be extended to a measure on the a-algebra
a (G)
L(P)
generated by G . The completion of this measure we denote by
and call the Loeb-measure of P . The a-algebra a (G) is
called the Borel-algebra of G and its completion L(G) with respect
to L(P) is called the Loeb-algebra of G . The measure space
<Q, L(G), L( P) > is called the Loeb-space of <Q ,G , P > . When °P( Q)
is finite, it follows from the uniqueness part of Caratheodory's
Theorem that L(P) is uniquely determined by P; C.Ward Henson has
proved that this still holds when °P(Q) = "" • If Q is *-finite
and P(Q) = 1, <n,G,P> is called a hyperfinite probability space.
The measure- and integration-theory of Loeb-spaces were devel
oped in Loeb [ 16] and Anderson [ 1 ] . Expositions are given in Loeb
[17] and (without proofs) in Keisler [10].
- 2 -
2. Hyperfinite Processes
We now define the basic concepts in the theory of hyperfinite
processes:
A hyperfinite time-line is a hyperfinite subset T of *:R +
such that 0 E T and for each a E :R+ there is a t E T such that
tRJa. If S and T are hyperfini te time-lines with S c T , we
call S a subline of T .
By a hyperfinite stochastic process X we mean an internal
mapping X : T x Q + *m, where T is a hyperfini te time-line and
<n,G,P> is a hyperfinite probability space such that the mappings
w + X( t, w) are G -measurable for all t E T .
If T 1s a hyperfinite time-line, an internal basis
<n,{Gt}tET'P> indexed by T consists of a hyperfinite set n ,
an increasing *-sequence {G t} tET of *-algebras of subsets of n ,
and a *~probability measure p on where t co
is the largest
element in T . For convenience we shall often assume that Gt is co
the internal power-set of n such that each one-point set { w} is
G t -measurable. co
Let us illustrate these definitions by an example:
Example 1: In [ 1 ], Anderson constructed the following process:
Choose an n E *N,JJ, and let the time-line be T = {k:kE *N, k<n 2 L n
The set n is the set { -1 ~ 1} T of all internal functions from T to
{-1,1}. If wEn , we shall write w. 1
for i w (-) • The internal n
algebra Gk will be the /n
*-algebra generated by the equivalence
classes of the relation ~ k '
where W"' w' k if w. =w.' 1 1
for all
1 < k • The probability measure P 1s the uniform measure on
-n2 n ; P{ w} = 2 for all w E n . Then <Q, {G t}, P > is an internal
- 3 -
basis. We define a hyperfini te stochastic process x : T x n + ii'JR by
k-1 I: w./.;n.
. 0 l l=
Anderson proved that if we define 13 : JR x n + JR by + 0 -(3(w,t) = x<w,t),
where t is the least element of T larger than t , then 8 is a
Brownian motion.
We see that for each t E T , w + x ( t, w) is G t -measurable, and
this motivates our next definition:
A hyperfinite process X : T x n +*JR is said to be adapted to
the internal basis <n,{Gt},P> if w + X(t,w) is Gt-measurable
for all t E T •
Notation: We shall often write Xt(w) or - suppressing w - X t for X(t,w). Let 0 = t 0 <t 1 < ... <t~ =too be the elements ofT
in increasing order. If X : T x n + *JR is a process, we shall write
6X(ti,w) for X(ti+ 1 ,w)- X(ti,w) . We use the following notation
for summation: If s = t. ' t = t. l J
and
t :E X(r,w) s
j -1
for :E X(tk,w) . k=i
i<j, we write
t Notice that the term X(t,w) lS not included in the sum :E X(r,w) .
s We use similar notation when the limits s and t of the summation
depends on w .
By the quadratic variation of a process X : T x n + *JR , we mean
the process [X] :Txn +*JR defined by
t [X](t,w) = r (6X(s,w)) 2
0
Observe that if X is adapted to a basis <n, {G t}, P > , then so is
- 4 -
[X,]. If x 1s the process of Example 1 , then [ x ]( t ,w) = t .
An internal stopping time adapted to the internal basis
<Q,{Gt},P> 1s an internal mapping T: Q + T such that for all
t E T , the set { w E Q : T ( w) ~ t} is in G t
Definition 2: A hyperfini te process M : T x Q + *JR is called a
hyperfinite martingale (hyperfinite sub-martingale, hyperfinite
super-martingale) adapted to an internal basis <Q,{Gt} ,P>, if
M is adapted to <Q,{Gt},P> as a process and for all s,t ET,
s < t , and all AEG s
In this definition the expectation E(1A(Mt-Ms)) 1s just a
hyperfinite sum I: 1A(w)(Mt(w)-M (w))P{w}. wEQ s
This paper is a study of hyperfinite martingales, and we have
already encountered one; Anderson's process x is a martingale with
respect to the basis <Q,{Gt},P> constructed in Example 1.
We shall need the following technical lemma relating the expec-
tation of the quadratic variation of a martingale to the expectation
of the square of the martingale.
Lemma 3: Let M : T x Q + *JR be a hyperfini te martingale adapted
to the internal basis <Q,{Gt}P>, and let a,T be two internal
stopping times adapted to the same basis. Assume that a(w) > T(w)
for w E Q . Then:
o(w) E(M(a(w),w) 2 -H(T(w),w) 2 ) = E((M(o(w),w) -Mh(w),w)) 2 ) ::; E( I: ~M(s,w) 2 ).
-r(w)
Proof: We have:
E((M(a(w),w)-M(T(w),w))2) = E(M(o(w),w) 2-2M(a(w),w)M(T(w),w)+M(T(w),w) 2 ) = = E(M(o(w) ,w )2-M( T (w) ,w ) 2-2M( T (w) ,w )(M( a(w) ,w )-M( T(w) ,w) )) •
- 5 -
To prove the first equality in the lemma, it is thus enough
to prove that E(M(T(w),w)(M(cr(w),w)-M(T(w),w))) = 0
For each s E T let As = { wEQ : T ( w) = s} . Then As EG s , and
E(M(T(w),w)(M(cr(w),w)-M(T(w),w))) = E( I: 1A (w)M(s,w)(M(cr(w),w)-M(s,w))) sET s
cr(w) = E( I: 1A (w)M(s,w) I: t:.M(t,w)) = I: I: E(1A (w)M(s,w)1[t 1t:.M(t,w)),
sET ·D s SET t~s ·D <cr
where [t<cr] = { w: t<cr ( w)} . Here 1 A is G t -measurable and so s
Moreover, [t<cr] = Q '{w:cr(w)~t} EGt. By the martingale
property all the terms in the double sum above are zero, and the
first equality of the lemma follows.
To prove the second equality, observe that:
cr(w) E((M(cr(w) 7 w) -M(T(w),w))2) = E(( I: t:.M(t))2) =
T(w)
Rewriting the last term, we get:
E(I:{t:.M(s,w)t:.M(t,w): T(w).::; s < t < cr{w)})
cr(w) E( I: t:.M(t) 2) +E(2I:t:.M(s)t:.M(t)).
T(w) T$S<t<cr
= E(I:{1h$s]t:.M(s,w) 1[t<cr]t:.M(t,w) I sET,t>s}).
Here 1 [ T~S] , t:.M( s ,w) and 1 [t<cr] are all G t -measurable, and
consequently the expectation of each term lS zero. This proves
the second equality and hence the lemma.
We end this section by quoting two results from the standard
theory of martingales which we shall use in the sequel. The first
is a famous inequality due to J.L. Doob which is proved in almost
any text on stochastic processes; see e.g. Doob [ 6] or Metivier [18]:
Proposition 4: (Doob's Inequality). Let X: J x Z + JR be a positive
(standard) submartingale with respect to a stochastic basis
<Z, { F t} tEJ, ~ > , where J c JR + Let I be a countable subset of J
- 6 -
and assume that b = sup IE J . Then for all p E <1 ,m > such that
E(Xh) < m , we have:
II sup Xtl! < _E_1 I! xb !l . t E I p - p- p
In this paper we shall use the *-transfer of Proposition 4
to deal with hyperfinite martingales, while in [12] and [13] we
shall need the standard version above.
The next proposition is *-transfer of well-konwn inequalities
1n the theory for discrete martingales.
Proposition 5: Let H : T x Q + *JR be a hyperfini te martingale.
Then for all p E <1 ,oo> and all t E T
11/ M~+ [H] Ct) II _:: 1 Op!lsupCM.JII p s ~t p
and
II sup ( Mt )II _:: p/121!/M"~ + [ M] ( t) liP . s ~t p
For a proof, see Neveu [21] (Corollary vm-3-18, page 199) or
Meyer [19]. Lest the reader should feel discouraged by the length
and difficulty of these proofs, we hasten to assure him that we shall
only need the result for p = 4 , and for some constants c 4 ,d 4 E JR in
stead of 10o4 and 4/12 . A simple proof may then be obtained from
Meyer's comments in [19] , and this was done in [11] .
- 7 -
3. The structure of A2 -martingales
We define the classes of hyperfinite martingales that will
be our principal objects of study in this paper:
Definition 6: Let H : T x n + *:R be a hyperfini te martingale
adapted to the internal basis <D,{Gt},P>. M is called a
A2 -martingale if for all finite t E T, E(M~) is finite. M is
called a local A2-martingale if there exists an increasing sequence
{ • n }nE:N of internal stopping times adapted to <D, {G t}, P > such that
0 T + ~ a.e. in Loeb-measure, and such that for each n the martinn
gale M defined by M (t~w) = M(T (w)At,w) •n •n n
is a A2 -martingale.
The sequence {Tn} 1s then called a localizing sequence for M.
In this section we shall use Lemma 3 and Proposition 4 to
derive some properties of the paths of A 2 -martingales. An obvious
consequence of Lemma 1 is that a hyperfinite martingale is a A2 -
martingale if and only if E(M~+[M](t)) is finite for all finite
tET.
Proposition 7: Let M: T xQ + *:R be a local A2 -martingale. Then
the set of all w E Q such that M(t ,w) and [M] (t ,w) are finite
for all finite t E T has Loeb-measure one.
Proof: Let {T } be a localizing sequence for M. n For each finite
tET and each nE:N, E([M ](t)) •n
is finite. Consequently
[M ](t) is finite a.e .. Since [MT ] is an increasing process •n n
max[M ](s) is finite L(P)- a.e.. Now 0 Tn + ~ L(P)- a.e. and s::;t •n hence max[M](s) is finite L(P) -a.e .. Since this is true for an
s::;t arbitrary finite t E T, the [M]-part of the proposition follows.
- 8 -
By Proposition 3
ECmax M~n(s)) ~ 4E(M~n(t)) s~t
and consequently 2 max M, (s) is finite L(P)- a.e. for all n E:N s<t n
and all finite t E T . Repeating the argument of the first part of
the proof, we prove the M-part of the proposition.
Proposition 7 tells us that the local A2-martingales live
where we want them to live; in the finite part of *F. Our next
result deals with the behaviour of the paths as we approach a point
t in T . We need some definitions.
Definition 8: If f: T + *F is an internal function we define the
upper S-left-limit of f at t by:
S-llm f(s) = inf sup{ 0 f(r): s < r< t, r( mon(t)} s+t s<t
s(mon(t)
The lower S-left-limit of f at t is defined by:
S-lim f(s) = sup inf{ 0 f(r): s ~ r< t, r (mon(t)} s+t ~<t
sl[mon(t)
The upper S-right limi! s-11m f(s) and the lower s-1-t
S-right limit S-lim f(s) is defined in a similar way. s-1-t
The function f 1S said to have an S-left limit
S-lim f(s) = S-lim f(s) =I= ±co and an S-right limit if s+t s+t
S-lim fW = S-lim f(s) :j: ±co • s-1-t s+t
Notice that these limits only depend on the monad
not on which point in that monad we use.
at
of
t if
t and
- 9 -
A process X : T x Q + *JR lS said to have S-left limits
( S-right limits) a. e. if for L( P) -almost all w E Q the function
t + X(t,w) has S-left limits (S-right limits) for all finite t E T.
We have the following:
Theorem 9: Let M be a local A2 -martingale; then M has S-left-
and S-right limits a.e.
Proof: We prove the S-left limit case; the S-right limit case
is similar.
There are three ways in which M can fail to have S-left
limits on a path s + M( s ,w) , there must exist a finite t E T
such that either S-llm M(s,w) = ~ or stt
that
S-lim M(s,w) = -~ or such stt
S-lim M(s,w) - S-lim M(s,w) > e stt stt
for an e EJR , e > 0 •
In the first two cases max !M<s,w)l must be infinite and s~t
according to Proposition 7 this can only happen on a set of
measure zero.
In the third case we may for each n EJ-.J find a sequence
t 1 < t 2 < ••• < tn < t such that !MCti+l ,w)- M(ti,w) I ~ £ , for
1 ,:: i < n , by definition of S-llm and S-lim • If we can prove stt stt
that this only happens on a set of measure zero, we have proved the
theorem. Since it will be useful in the sequel, we formulate this
statement as a separate lemma:
Lemma 1 0 : Let M : T x Q + *JR be a local A 2 -martingale. Let A
be the set of all w such that there exists an e EJR , £ > 0 , and
a finite t E T such that for all n EJ-.J there are n elements
- 10 -
1 < i < n • Then L(P)(A) = 0 .
Proof: It is enough to prove the lemma for A2-martingales, s1nce
we can then use localizing sequences of stoppingtlines to prove it
for local A2-martingales. Moreover, it is enough to prove that
for all £ > 0 and all finite t E T , the set B of all w E Q
such that for all n EJN there are elements t 1 < t 2 < ••• < tn < t
in T with !MCti+ 1 ,w) -H(ti,w)l > £ has Loeb-measure zero.
So let us assume that M is a A 2-martingale with £ , t and B
as above. Define,an internal sequence of stopping times by: T0 (w) = 0
and
For k E:N let Bk be the set of all w E Q such that there
exist t 1 < t 2 < ••• < tk < t with !M<ti+ 1 ,w)- M(ti,w) I >
<r2,{Gt},P> is the internal basis of M, then BkEGt
£ • If
and is hence
measurable. For each w E Bk and each t. l
w as an element of there is an m E*N
and l.f ~m(w) _< ti+ 1 then ~ (w) < t. L L m+ 1 - 1+2 °
T k- 1 ( W ) ::: tk < t ,
By Lemma 3 we have:
E( I: (M(Tn+1(w),w) ~M(T (w),w))2) nE*N n
in the sequence defining
such that t. < T (w) < t. , 1- m - 1+1
Consequently
= E([M](t))
(These sums are really hyperfinite s1nce •n(w) = t for some nE*N ,
and so there is no problem of convergence.)
For w E Bk , we have by definition of
Tk 1 (w) < t : l: (M(T + (w) ,w)- M(T (w) ,w) ) 2 - nE*N n 1 n
Consequently: P(Bk)o(k-1)(~) 2 ::: E([M](t))
•n+ 1 and the fact that
£ 2 ~ (k-1)(2) •
or P(B ) < 4E([M](t)) k - £2(k-1)
Since M is a A2-martingale E([H](t)) is finite and hence
~ 11 -
L(P)(Bk) = 0 P(Bk) + 0 as k + oo. Since B = n Bk , L(P)(B) = 0 .
This proves the lemma, and completes the proof of Theorem 9.
4. S -continuous :>t 2 -martingales
Definition 11: An internal function f: T + *R is called
S-continuous if for all finite s, t E T such that s ~ t , the values
f(s) and f(t) are finite and infinitesimally close. A hyper-
finite process x : T x n + *JR lS S-continuous if for all w in a
set of Loeb-measure one the function t + X(t ,w) is S-continuous.
Using the internal definition principle it is easy to show
that f : T + *R is S-continuous if and only if f(t) is finite
for all finite t E T and for all such t and all positive e: ER
there is a o ER, o > 0, such that for all sET with !s-tl < o,
we have lfCs)-f(t) l < e: •
In this section we want to prove that a local :>t 2 -martingale
is S-continuous if and only if its quadratic variation is S-con
tinuous. This is an example of the main strategy of this paper; to
relate properties of a hyperfinite martingale to properties of its
quadratic variation. Since it is an increasing process, the quad
ratic variation is often much easier to work with than the wildly
oscillating martingale. A first feeble attempt in carrying out
this programme was made ln Le~ma 3 and some of its consequences
gathered in Lemma 10 and Theorem 9. Using Proposition 5 we shall
now launch a more serious attack leading up to the result announced
above (Theorem 14). We shall see more consequences later when we
use Theorem 14 to prove Theorem 21.
- 12 -
Let us return to the problem in question. We shall use a
sequence {Tn} of internal stoppingt~s defined by
T ( w) = min { t E T : I M ( t, w) ! > n} , n
and then apply Proposition 4 to the martingales MT with p = 4 • n
We shall thus consider expressions of the form E([MT ](t) 2 ). n
However, this may be infinite, and in that case we should not be
able to complete our calculations. Hence we look for conditions
that will ensure us that it is finite. One of them is obviously
the following:
A process X : T x n -+ *JR lS said to have infinitesimal
increments if for all finite t E T and all w E n , t.M( t, w) R:t 0 •
We shall show that all S-continuous A2-martingales (or A2-
martingales having S-continuous quadratic variation) can be modi
fied into A2-martingales with infinitesimal increments, and then
use Proposition 5 and the stopping times above to prove our results.
We shall need some technical preliminaries: Recall from
Anderson [ 1 ] , that a G -measurable internal function f on a hyper-
finite measure space <O,G,m> is called S-integrable if:
(i) 0 fl£1dm < oo
(ii) If A EG and m(A) R:t 0 '
then f If I dm R~ 0 • AI
(iii) If A EG and f(x) R:t0 for all xEA '
then J 1 t ! ctm R:l o . A
Clearly, (iii) is redundant if <O,G,m> is a probability space.
Anderson further defined SLP(n,G,m) to be the set of all internal
G-measurable functions such that !tiP lS S-integrable.
Lemma 12: Let <n,G,P> be a hyperfinite probability space. Let
f : n -+ *JR be a G-measurable internal function such that
0 f I f ' 2 dP < 00 • Then f is S-integrable.
- 13 -
Proof: It is enough to prove that f satisfies (i) and (ii) above.
By Holder 1 s inequality we have for A E G :
Taking A = n we have (i), and letting P(A) ~ 0 we get (ii).
This proves the lemma.
Proposition 13: Let M :Txn+ *JR be a A. 2 -martingale adapted to
the internal basis <n,{Gt},P>, and assume that the set
{wE rl : 3t E T ( t finite and 0 6M(t,w) * 0)} has Loeb-measure zero.
Then there exist an internal probability measure P on Gt and OJ
a >.. 2 -martingale adapted to
(i) M has infinitesimal increments.
(ii) For all B EGt , P(B) ~P(B) . OJ
(iii) There is a subset n' of Q of Loeb-measure one, such
that for all wEn' and all tET, M(t,w) = M(t,w).
Proof: Let a E JR+ and t a finite element of T . We shall
construct a process M such that 16M( s ,w)! ~a for all w E Q
and all s < t , and a probability measure on such that
M 1s a martingale with respect to <n, {G t}, P > . For all
we shall have f[M](s)dP:: f[:M](s)dP. Finally, replacing
s E T , ,.., M and
,.., P by M and P respectively, (ii) and (iii) hold.
Let us first show that if we can construct such M
for all a and t ' then we can prove the proposition:
the set of all n E*:N such that there exist Mn and
the following properties:
-n + *JR a hyperfinite and -n M : T X Q lS process, p
internal probability measure on Gt OJ
such that
pn
is
··n M
and p
Consider
with
an
is a
- 14 -
hyperfinite martingale adapted to
1 exists a set A EGt , P(A) <-, n oo n n
and all w (A , H(t,w) -n = M (t,w). n
have I p (B)- :pn ( B )j < 1. ' and for all n
> f [ Mn ]( t ) d :pn . Finally, for all -I -n I 1 t~M Ct,w) ~n.
There
such that for all t < n
For all B EGt , we 00
t<n '
f [M](t)dP > -wEn and all t < n ,
This is an internal set, and putting 1 a = -n
and t equal to the
least element in T larger than n ) we see that it contains all
finite n. Thus it contains an infinite y E *JN ' and taking
"" f1Y "" -y proposition. M = p = p and n' = n ...... A '
we get the y
Let a and t be given; we shall construct M and p . The
proof is long and tedious, and we divide it into several steps:
1 Definition of M : Let A be the subset of n consisting of
all w such that there exists an s < t with jt~M( s ,w) I >a . The
set A is internal and Gt-measurable. For wE A , let
the first element s 1n T such that lt~M( s, w) I > a .
If wEn '-A, let M(s,w) = M(s,w) for all sET.
If w EA, let M(s,w) = M(s,w) for s < s - w
let
let llM(s,w) = 0
and
for s > s w
s be w
By hypothesis, A has Loeb-measure zero and M thus satisfies (iii).
2 Definition of P : We must change P into P such that M
becomes a martingale with respect to cn,{Gt},P>. This is done by t.
defining a measure P 1 for all t. E T , t. < t , by induction. The l l -
measure Pt obtained by this procedure will be our measure P . t·
Let P0 = P . Assume that P 1 is defined, we shall construct
- 15 -
For each wEn, let [w]t· = n{B EGt· :wEB}. Given such an 1. l
equivalence class we define e: [ ] E *JR by: w t· 1.
.... t . f'V ,..,.
= l:{(I~M(ti,w)j-a)P 1 {w}:wE[w]ti" SO;
t· ti .... If Zi'E[wlt. and s,..,. * t. ' let p 1.+ 1 ({Zi'}) = p {w }/1 +e:
w 1. [w) t. 1. 1.
= t.} l
t· I ,..,. I t. f'V 1 ~M(ti,w) If Zi' E [ w lt. and let p 1.+ 1 ( {Zi'}) )•P 1 {w} sf'V = t. = a(1+e:[ ] w 1. ' 1. w t· 1.
This defines ti+1 t P , and hence P = P by induction.
3 P is a probability measure:
We write:
B = {Zi' E n
c = {wEn
: Zi' E [ w ] t " s,..,. * t . } • w l l
: w E [ t:l ] t . A sw = t i } 1.
16M( ti ,Zi') I a(1+e:[wlt.)
1.
t· Let us caluclate P 1 +1([wlt.)
1.
and
We have:
t· Thus P 1 +1([w]t·) =
1.
t· P 1 ([wlt.), and summing over all equi-
1. pti is a probability measure, so valence classes we see that if
ti+1 is P By induction is a probability measure.
- 16 -
i M is a martingale with respect to <rl, {Gt} ,P > : It is enough
to prove I: flM(t. ,w)Ptr;;:n = o for all wEn and all tiE T . wE [wlt. 1
l
By definition of M and P this is obvious if t. > t l-
or t. > s ' l IJ)
and hence we only have to prove the formula for t 1 <t, t.<s. l - w
In that case
P ti{w} = P{w} I a , where a = . rr • ( 1 + e: [ ] ) , for all w E [ w J t . • J<l w tj l
Let D={[W'J.r. :W'E[wlt.}. :L+l l
Then:
- t...... - t I: tlMCt. ,W')P {w} = I: tlMCt. ,W')P < [wlt > wE [w]t. 1 [W'Jt.E D 1 i+l
l l+l
- ,..., ti+l ,..., = :r l!M(t.,w)P ([w]t ) , [w] E.D 1 i+l
t. 1 l+
since it follows from .the calculation in ~ that Pt<[wl > = ti+l t·
= P l+l<[W'J ) for t > t. ti+I - l+ 1
Let now Dl = { ['Wlt. ED : s ..... * t. } l+l w l
and D2. = {[W'] t. ED : s .... = t.} l+l w l
We get
- ,..., ti+l "' I: l!M(t.,w)P ([w]t ) =
[wlt. ED 1 i+1 l+l
t· t· I: tlMCt. ,W'> P l+l < [W'lt > + I: flMCt. ,Ul') P l+l c [wl > =
["'] ED 1 1·+1 [-:vlt.€ D2 1 tl.+l w t. 1 w l+l l+l
1 I: tlM(t. ;un P c r'WJt >
[w\. ED 1 i+1 l+l
since M is a martingale. This proves that M is a martingale
with respect to <rl,{Gt},P>.
- 17 -
i An estimate: To prove that P(B) ~P(B), we shall need the
following estimate (where
n-1 < :t
i=o
t = t ) : n
n-1 :t
1 t· :r [II C1+e:[ 1 >l<-:t{(!f~M(t.,W'>!-a>·P l{w}:'WE[wlt 1\ g,...=t.}) =
i=o [] ,., .. wt a l . w l w t c.6 J<l . l . J
l
1 n-1
= - :t a i=o :t {(jf!M(t.,W')j-a)oP{w} :'WE [wlt A g,...=t.}
[wlt.cn l i w l l
_lf C!f!M(s ,w>!-a)dP<lfjflM(s ,w)!dP<3.Jrnax!M(s,w)ldP a · w -a w -a · A A A s~t
By Doobis inequality we see that E(max!M<s,w)j 2 ) is finite, s~t
and it follows by Lemma 12 that max!M(s,w>l is 8-integrable. s~t
Since P(A)~o, we get lf max!M<s,w)!dP~O, and consequently
n-1 a A s~t
:t ( .:t e:[ 1 )P{w} ~a. w€n'A l=O w t·
l
Q. P(B)~P(B) for all B€Gt : Let C be the set of all w€n'A CXI
n-1 such that .:t e:[ 1t ~~. It follows from.§_ that P(C) ~1.
l=O w i
Let D c C be an element of Gt . CXI
For each w € D , we have
t t P {w} = P{w}/ II (1+e:[ ] ) , and consequently P (D)=: P(D) • S<t W S
But on the other hand
This may be seen as follows:
< elog( 1 +e:) = 1+e: since the logarithm is concave.
- 18 -
M ( 1 ) n 1 n(n-1) 2+n(n-1)(n-2) 3+ ; ( )k 1 oreover, +~:: = +ne: + 1 •2 e: 1•2•3 e: ;•• ~ "- ne: = 1-ne:
k=o
(since and the sum converges). tAle get
n-1 Pt{D) ~ r P{w}(1-ne:) = P(D)- I: P{w} I: e:[ ] F.::$
wED wED i=o w t. 1
P(D) by 5. Hence
Pt(D) F-::$P(D) .
This proves the assertion when B c C • Let now B be an arbi-
trary element of G t ; then B = ( B n C) ll ( B'C) • By what we have (X)
just proved P(C) F-::$1 , and since P is probability measure
Since BnC c C , we have already seen that
1 f[M](r)dP?:: J[M] (r)dP: It is enough to prove
f t.M(ti,w) 2dP ~ f t.M(ti,w) 2dP for each [wlt. [w]t·
t. < t 1
and all equi-
1 1 valence classes [ w] t. . We define
1
E1 = { [ wl t c r w 1 . : s .... * t.} • -c • w 1 1+1 1
E2 = {[wlt. c[wlt.: s'il:l' = ti} 1+1 1
Then f tiM(t 1. ,w) 2dP = I: t.M(t. ,w) 2P[w]t + I: tiM(t. ,w) 2P[w]t 1 . 1 .
[wlt. [w] EE 1+1 [w] EE 1+1 t. 1 1 t. 2 1 1+ 1+1
If s < t. ' then t.M( t. 'w) = 0 w 1 1 for all wE[wlt. and the inequality
1
is obvious. If not, we have
2 2 t· r t~MCt. ,w) P[W'lt = . rr. <1+e:[ 1 > I: t~M<ti,W'> P 1+ 1[w~.
[w] EE 1 i+1 J <1 w t. [w]t EE. 1+1 ti+l 1 - J i+l 1
since t~A(t.,~) = tiM(t. ,~) by definition of M and 1 1
t· P 1 + 1 [w""] . b h · · 3 W 1 . y t e computat1on 1n . e a so have
ti+l
= 19 -
= . IT • ( 1 +e: [ ] ) ]<1 w tj
-( ,..., 2-["'] = :r t.M t. ,w) P w t [wJt. EE2 1 i+l
1+1
>
Combining our results we obtain the inequality. We have now proved
all assertions about M and P made in the beginning of the proof,
and we are through.
Notice the use of Proposition 13 in the proof of the main
result of this paper:
Theorem 14: Let M : T X n -+ *:R be a local A 2 -martingale. Then
M is S-continuous if and only if [M] 1s.
Proof: (i) Assume that [M] 1s S-continuous. Using a localizing
sequence of stopping times we may assume that M is a A2 -martingale.
Then M satisfies the hypothesis of Proposition 13, and M of that
proposition is S-continuous if and only if M is ; and [~] is
S-continuous since [M] 1s. Let Tn be the internal stopping time
Since
than
,..., M has
If r E *:JR
r. For
A (t) = m,n
Tn(w) = min{tET: CM](t,w) ~ n}
infinitesimal increments, [ f=f ] < n+1 . Tn
-let r be the smallest element of T larger
m,n E *:IN and t E T , let
where [t•n] is the integer part of ton . If A is the set of
all w such that s -+ M ( s ,w) is not S-continuous, then Tl
- 20 -
A = u u n A(k) Since -+ 00 a.e. as 1 -+ 00 ' it 1S kEIN mEJN nEIN m,n
enough for us to prove L(P)( n A(k))=O for all k,m,lE N; and to nEJN m,n
prove this it 1S enough to prove P(A(k))F:::~O for k,m,lEJN, m,y
y E*N '-N . We have:
[ky] . -< md 4 I: E < < [ M l < 1 + 1 1 y ) - [ M ] < i 1 y ) ) 2 )
i=O Tl Tl
[ky] -- --= E ( md 4 I: ( [ M ] ( i + 1 I y ) - [ M J ( i 1 Y ) ) 2 )
i:Q Tl Tl
(By Prop 5, for sui table d 4 E:R)
Since [M ] < 1+1 , we have: Tl
IJ<y] --- - 2 [ky] --rrrl4 I: ([M ]( i +1 1 y) - [11 ]( il y) ) ~ m:14 < 1 +1 ) I: < [!-1 ]( i +1 1 y) - [M ]( il y))
i=O Tl Tl i=O Tl Tl
2 ~ rrrl4 (1+1) 0
There exists a set of Loeb-measure one where
[M ] ( i+1 I y) - 01 ]( il y) < o for an infinitesimal o , when Tl Tl
i+1/y ~ [ky]. For w in this set
[ky] [ky] md 4 I: ([M l<1+11y,w)- ['M ](ily,w)f ~ m~o :E([M ](i+17y,w)-[M ](ilpw))
i:Q Tl Tl i:O Tl Tl
< mod 4 < 1+1 ) F:::l o .
Thus the expression inside the expectation
[ky] E(md 4 I: ([M ](i+1ly)- [M ](ily)) 2 )
i:Q Tl Tl
is finite everywhere and infinitesimal a.e., and hence the expecta
tion is infinitesimal. It follows that P(A(k)) F:::t0 and as we have m,y '
already noticed, this proves one part of the theorem.
- 21 -
(ii) Assume that M is S-continuous. As above we may assume
that M is a A2 -martingale, and since M satisfies the hypothesis
of Proposition 13, M is S-continuous. It is enough to prove that
[M] is S-continuous.
We define the sets B~~~ for m,n E *JJ , kEN by
B(k) = {wEn: 3i :5. [kon](([M](i+1/n)- [M](i/n)) 2 ) > 1/m)}. m,n
As above it is enough to prove for all m,k E lJ
y E *N ':N .
For g1ven y E *:IN 'N we define a new martingale M( Y) by:
For each w E Q such that there is an i < [ky J with
suP < I 11 < s , w ) - 'f-1( ify , w ) I ) :::. 1 , l?Y:s s:si+17y
let s be the least s w
M(y)(s,w) = M(s,w)
satisfying this formula. We put
for s < s , but let t.M( Y) ( s ,w) = 0 - w
Then M(y) is obviously a martingale, and since and
for s > s w
M(y) have
the same
to prove
paths outside a set of infinitesimal measure, it is enough
PCC(k)) ~o where C(k) is the set obtained from M(y) m,y ' m,y
in the same way that B ( k) was obtained from M . m, Y
Let S = {i/y: iE*JH be a subline of T. Let N : Sxn -+ *R
be the restriction of M( Y) to S . We define a sequence {on} of
internal stopping times by
on ( w) = min { s E S [ N] ( s , w) > n}
By the definitions of
Since -+ 00 as n -+ 00 '
for each n ' where c<k) m,y,n
M(y) and on, we have [N 0 ] :5. n+2 . n (k)
we only have to prove that P( C ) ~ 0 m,y,n
lS the set obtained from M(y) in the on
same way that was obtained from M. We have
- 22 -
0 < P(C(k) ) m,y,n
l)<y] ...... ( ) ""'( ) - 2 < I: P{w: ([My ](i+1/y)- [My ](i/y)) > 1/m}
. a a 1=0 n n
(by Prop. 5)
:: mc4 (4/3)'+ [!J E((M~y)(i+1/y) -M~y)(i/y)) 4 ) (by Doob's inequality) 1=0 n n
[ky] •. [ky] - 4 = mc,,(4/3)'+ I: E((N <i+1/y) -N (i/y)) .. = E(mc,,(4/3)'+I: (N (1+1/y)-N (i/y)) ). .. . a a .. . a a
1=0 n n 1=0 n n
Now (N Ci+1/y) -N (i/y)) 2 < [N ] :: n+2, and thus the an an - an
expression inside the last expectation is always less than:
4 [kyJ - 2 4 2 mc,,(4/3) (n+2) I: (N <i+1/y)-N (i/y)) ::mc,,(4/3) (n+2) which is .. . a a ..
1=o n n ( finite. On the set of measure one where M y) 1s S-continuous
an each (N (1+1/y) -N (i/y)) 2
an an is infinitesimal, and repeating the
argument we see that the expression inside the last expectation is
infinitesimal a.e.
It follows that the expectation is infinitesimal, and so is
P(C(k) ) . Hence [MJ is S-continuous and so is [M] . m,y,n
proves the theorem.
This
We have replaced the original martingale M by M in the
proof above to be able to prove that
[ky] 2 md4 :r C[M J<i+1/y)- [M J<ilr>>
i=O Tl Tl
and
are finite everywhere. This we need to know to deduce from the fact
that these expressions are infinitesimal a.e. that their expectations
- 23 -
are infinitesimal. The difficulty is peculiar to the nonstandard
theory since in standard integration theory integrals over sets of
measure zero are always zero. In other contexts - e.g. in the
theory for (Schwartz-) distributions - this peculiarity is an
advantage.
We end this section with an example:
Example 1 5: Let x be Anderson's process from Example 1 , and let
us see how much information we may obtain about x using the theory
we have developed so far: Obviously x is a martingale and
[x ](t) = t. By Lemma 3 x is a A2-martingale, and from Proposition 7
we see that x is finite a.e.. Since [xJ obviously is S-continuous
it follows from Theorem 14 that so is x .
5. The stochastic integral
Definition 16: Let X, Y : T x n -+ *JR be two hyperfini te processes.
By the stochastic integral of X with respect to Y , we shall mean
the process fXdY defined by
vJe shall write
process fXdY
t (t,w) -+ L X(s,w)6Y(s,w)
0
t t f 0 X(s,w)dY(s,w) or f 0 XdY for the value of the
t 2 2 at time t. Obviously, [fXdY](t)) =EX (s)6Y(s) .
0
We shall mainly be concerned with stochastic integrals of the
form fXdM where M is a hyperfinite martingale with respect to an
internal basis <S'2;{Gt} ,P>, and X is a process adapted to the
same basis. It follows from the martingale property that fXdM
then is a martingale adapted to <Q,{Gt},P>.
- 24 -
The following simple identity will be useful in [12] and [13] .
For X= x (the process of Example 1) it was proved by Anderson in
[1]; the proof in the general case is similar to his:
Proposition 17: Let X be a hyperfinite process. Then
Proof: He have
k-1 k-1 k-1 k-1 i-1 = X~+2X 0 .~ ~xt.+<.r ~~.) 2 -x~-2x 0 .r ~xt.-2.r <.r ~xt.)~Xt.
l= 0 l l= 0 l l= 0 l l= 0 J = 0 J l
k-1 k-1 i-1 k-1 i-1 = < r ~xt.) 2 -2 r r ~xt.~xt·
i=o 1 i=o j=o 1 J Now we have 2 L .'- ~xt.~Xt.
i=o J=O 1 J
=
Thus
which proves the theorem.
k-1 < r ~xt.) 2 =
i=O l
k-1 k-1 r .r ~xt.~xt.
i=O ]=0 l ]
If M is a A2 -martingale, we define A2 (M) to be the set
of all those internal processes X
of M such that 0 E((ftXdM) 2 ) <oo 0
fXdM is a hyperfinite martingale
this is by Lemma 2 equivalent to
adapted to the internal basis
for all finite t E T . Since
with quadratic t
0 E(EX2 ~M2 ) < CX)
0 for all finite t E T .
If M is a local A2 -martingale, we define A(M) to be the
~et of all X adapted to the basis of M such that there exists a
localizing sequence {Tn}nEJN
n El-T •
for M with XEA 2 (M ) for each 'n
The classes A2 (M) and A(M) are perhaps too large to be
really useful, and we now construct two smaller classes SL 2 (M) and
- 25 -
SL(M) which consist - in some sense - of the square S-integrable
elements of A2 (M) and A(M):
For each t E T , let Tt = T n * [ 0, t > and let T t be the
internal power-set of Tt If M is a A2 -martingale with respect
to <S1, {G t}, P > , we define an internal measure
*-measurable space <Tt x S1, T t x G t > by: 00
"M t on the
2 v M { < s , w >} = P { w } ( tJ. M ( s , w ) ) for w E S1 , s E T t . t
If t is finite:
t 2 "H (TtxQ) = L L tJ.H(s,w) P{w} = E([H](t))
t wEQ 0
which is finite.
Definition 18: Let M : T x Q -+ *JR. be a A2 -martingale adapted to
and let X be a process adapted to the same basis. Then <S1,{Gt},P>
XESL2 (M) if and only if for all
finite t E T .
If M is a local A 2-martingale, X E SL(M) if and only if there
is a localizing sequence {Tn}
all n EJN .
for M such that XESL 2 (HT ) n
for
Since the process X 1n Definition 18 is adapted, each XtTtx S1
is measurable with respect to a *-algebra Pt much smaller than
T t X Gtoo •
X~Tt x n E
xt Tt x S1 E
The question naturally arises whether . 2
SL (Ttxa,TtxGt '"M) if and only if 00 t
SL 2 (TtxS1,Pt'"M tP~). The answer is not quite obvious t
since there are more sets of infinitesimal measure 1n Tt xGt 00
than in Pt, but it is not difficult to prove:
Lemma 19: Let <S1,G,~> be a hyperfinite measure-space, and let B
be a *-sub-algebra of G . Let f : S1 -+ *JR. be B-measurable. Then
if and only if 1 f E SL (a, B, ~) •
= 26 -
Proof: Left to the reader.
The next proposition shows us that the class SL(M) is
large enough for most purposes.
Pro:eosition 20: Let H be a local A.2-martingale and X a process
adapted to the internal basis of M. Assume that on a set of Loeb-
measure one max I X( s) I is finite for all finite tET. Then s .st
XESL(M). In particular if X is a local A. 2-martingale,
then X E SL01)
Proof: Let {~n} be a localizing sequence for M, and define
another sequence {crn} of internal stopping times· by
cr ( w ) = min { s E T : I X ( s , w ) ! ~ n} n
By the hypothesis, 0 ~ ( w) n
-+ 00 a. e ..
lS a localizing sequence for M and
Let ~I : n
jX(s,w)J <n for
Consequently X E SL 2 (M~,) ? and X E SL(M) . n
then {~~}
s <~'(w) n
If X is a local A.2-martingale, then X satisfies the hypo-
thesis of Proposition 20 by Proposition 7.
Before we turn to the next theorem let us recall some results
about S-integration. Let <~ ,G , ~ > be a hyperfini te measure space
with ~ ( ~) finite. Anderson ( 1 ] showed that an internal G -mea sur-
able function f : ~ -+ *R is S-integrable if and only if there
exists a sequence {fn} of G-measurable finite functions such that
1 im 0 ( f I f- f n I d ~ ) = 0 •
If f lS *-non-negative it follows from the proof that we may
take f = f 1\ n n Anderson also proved that if f is G-measurable
and *-non-negative, then °ffd~ ~f 0 fdL(~) with equality if and only
if f is S-integrable.
'
- 27 -
We now use Theorem 14 to prove:
Theorem 21: Let M be an S-continuous local A2-martingale, and
let X E SL(M) Then fXdM is S-continuous.
Proof: It is clearly enough to prove the theorem when M is a
A2-martingale and XESI/(M).
For each n EJ.J define an adapted process X : T X Q -+ *:R by: n
X (t,w) = X(tqw) if jXCt,w)! <In; X (t,w) = Tn if X(t,w) > Tn; n ' - n
and X ( t , w ) = - fn if X ( t , w ) < -In . n
Since M is S-continuous, it follows from Theorem 13 that
[M] lS S-continuous. Now [JXndMJ(t)- [JXndM] (s) =
~ X 2 t:~M 2 < n f l:IM 2 = n ( [ M ]( t)- [ M] ( s) s n s
for s,tET, s<t and hence
we see that [JXndM] lS S-continuous for each nEJ.J
By the last of the results of Anderson mentioned above:
0 s s 0 < E(sup ([f XdM]-[J X dM])) <
0 s s E(maxC[J XdM]-[J X dM]))
S5t o o n - s ~t o o n
By definition of X E SL 2 ( M) and the first of the results of
Anderson above, 0 JCx 2-X2 )dvM -+ 0 as n-+ oo. n t
A well-known result
of measure-theory (see e.g. Bauer [3] , Satz 15.5), tells us we may
o Js Js find a subsequence {sup ([ XdM]-[ Xn dM])}kElli which converges s~t o o k
to zero a.e ..
Using the E-o-formulation of S-continuity (recall the comment
following Definition 11) one may prove exactly as in the real-valued
case that the S-uniform limit of a sequence of S~continuous func-
tions is S-continuous. It thus follows that [fXdM] is S-continuous.
- 28 -
Applying Theorem 14 again we see that fXdM lS also S-continuous,
and the theorem is proved.
If v.Je replace X E SL(M) by X E A.(M) ln Theorem 21, the
resulting statement will be false as is easily seen.
Anderson [ 1 ] proved the theorem above for M = x (the process
of Example 1) when X was in a certain subclass of SL 2 (x) (the
class of H2-liftings"). Keisler [1 O] used Bernstein's inequality
to prove the theorem when M = x and X is a finite adapted process.
6. The transformationforrnula and Ito's formula
In this section we prove a nonstandard version of what may
be called the fundamental theorem of stochastic analysis.
If r E T , we write r and + r for respectively the prede-
cessor and the successor of r ln T
Theorem 22: (Nonstandard version of ·the Transformationformula).
Let cp : *JR -.*JR be an internal function and assume that l.D,tp' ,(!)"
all exist and are S-continuous and finite for finite arguments.
Let M: T x Q -+ *JR be a local A. 2 -martingale. Then for all finite tET:
o lP ( Mt ) - o tp ( M o ) - o f; tp v ( M ) dM + ~ o£t<P" ( M ) d [ M]
t +:Z::: 0 (<.;J(M +)-<p(M )-tp'(M )(M +-M )+l<.p"(M )(M -M ) 2 ) a.e.,
0 s s s s s 2 s s+ s
where the last sum is absolutely convergent.
Proof: We begin by enumerating the points s E T such that
0 -6M(s ,w) o~:O. By Lemma 10 this is possible for almost all w E Q ,
since for each finite tET and each positive e: E JR there are only
a finite number of s E T , s < t , such that 6M( s-, w) > e: • To make
- 29 -
the definition precise, it is convenient to define a bijection
f: N+ + N+ x~+ and let (f(n)) 1 denote the first component of
f(n) . Define a sequence {Tn}nElli of internal stopping times by
T n ( w) = min { s € T: I L\M( s-, w) I > ( f (~)) 1 " s * T k ( w) for k<n } .
Let Q 1 be the set of Loeb-measure one where M(t) is finite for
each finite t E T and where the sequence { T n ( w)} enumerates all
finite sET such that 0 L\M(s-,w) * 0.
internal extension of {Tn}nEJN such that Tk * T when k * m . , m
For each finite t E T and each v E *:N '-JN , let
A = {sET: s+ ( {Tn(w):nsv}}. t,v
For all w E Q 1 we get from Taylor's formula:
t lP ( Mt ) - <P ( M 0 ) = I: ( <,o ( M + ) - lP ( M ) )
0 s s
= (q>(M +) -t,o(M )) s s
= :E (r.p(M t) -<.p(M- t)) + r {<P' (M )(M +-M ) T " Tn" SEA s s s nsv n t,v
+ l!1.., 11 (M )(M +-H ) 2 + l(,n''(r: ) •n"(H ))(M M ) 2 } 2 ., s s s 2 't' .., s - "¥ s ~- 8 +- -s '
where I; lies between H + and Ms s s
By the S-continuity of <.p 11 the sum over the very last term
is infinitesimal for w E Q 1 and thus
tp(Nt)- <p(Mo) ~ !: (<.p(M t) -q>(M- t)) n< Tn" Tn" _v
+ I: {<P' (M )(H +-M ) + ~W"(M )(M +-H ) 2 } for wE Q 1 • sEA s s s s 8 s
t,v
Adding and subtracting
- 30 -
we get: t t
<P ( Mt ) - <P ( M 0 ) F:$ f <P I ( M ) dM + ~ f <P II ( M ) d [ M] 0 0
+ L {tp(MT At) -!p(MT-At) -q>Y(MT-At)(MT At-MT-At) n~v n n n n n
on Q' •
Since this J.s true for all v E *JJ '-JJ , the series
00 0 2 :r {tp(M t)- q>(M _ t) - ~,p' (M _ t)(M t-M _ t)- ~"(M _ t)(M t-M - t) } 0 T A T A T A T A T A T A '[ A T A n n n n n n n n
The convergence is absolute since the order of the terms in the
sequence {•n} does not matter. But the sum above is equal on Q 1
to the sum in the theorem, and this completes the proof.
In the S-continuous case Theorem 22 becomes
Corollary 23: (Nonstandard version of Ito's formula). Let
<P : *R-. "'JR be an internal function and assume that lP,!.P' ,<P" all
exist, and are S-continuous and finite for finite arguments. Let
M : T x n -+ *JR. be an S-continuous local A. 2 -martingale. Then for
all finite t E T
t t <P ( Mt ) - lP ( M 0 ) ~ f lP 1 ( M) dM + ! f <P" ( M) d [ M J a . e .
0 0
Applying Corollary 23 to a suitable integral of Anderson's
process x , one may obtain a nonstandard version of Ito's lemma,
as Anderson did in [ 1 ] .
For standard versions of the transformationformula and Ito's
formula, see Metivier [18] , page 264 and page 49 respectively. We
shall deduce these results from the above versions in [13 ].
- 31 -
7. Doob-Meyer decomposition
A process X : T x r2 -+ *JR t
is said to be of finite variation
if I X 0 I + E I ~X( s) I is finite 0 .
a.e. for all finite t ET. Such
a process is easily decomposed into two monotone processes x+
and X , by defining and t E(~X(s,w) AO). 0
Integration with respect to monotone processes
is easy to deal with, and since we in this paper have developed
the theory for integration with respect to martingales, we should
now be able to treat integration with respect to processes of the
form X + M where X is of finite variation and M is a local
:\2-martingale.
Definition 24: Let Y : T x S1 -+ *JR be a hyperfini te process adapted
to the internal bas is <S1, {G t} , P > . Y is said to have a Doob-
Meyer decomposition if there exist a process X of finite variation
and a local :\ 2-martingale M both adapted to <r2,{Gt},P> such
that Y = X+ M .
Proposition 25: Let Y : T x Q -+ "'JR be a hyperfinite process adapted
to the internal basis <r2,{Gt},P> such that Y(O) is finite a.e ..
Assume that there exists a sequence
times such that 0 , -+ oo a.e., and
n
all n E:N and finite t E T . Assume
{•n} of internal stopping tAT
E( E n(~Y(s)) 2 ) is finite for 0 t
further that E!E(~Y(s)!G )! 0' s
is finite for finite t E T, L(P)- a.e.. Then Y has a Doob-Meyer
decomposition.
Before we prove the proposition, we make two remarks on the
conditions which Y must satisfy:
- 32 -
Remark 26: The conditional expectation E(llY(s) !G ) . s
hyperfinite case just an average: As in Section 4 let
n{BEG :wEB} s
Then we have:
E(llY(s)!G Xw) = ( :r llY(s,w)P{w})/P([w] ) s ;e[w] s
s
lS ln the
This definition makes some properties of the conditional expectation
very easy to prove, e.g. will the inequality
which we shall need in the proof of the proposition, reduce to the
algebraic inequality
[(a1p + o•o +a p )/(p + •oo +p )]7.(p + ooo +p ) < a 2p + ooo +a2p 1 n n ·1 n 1 n - 1 1 n n
t Remark 2 7: The condition ~I EC~Y( s) !G s)! <"' a. e. in the proposition
is not much stronger than necessary~ and we show that it is satisfied
when Y has a Doob-Meyer decomposition Y = X + M where X has
reasonable integration properties: Assume E( x;) <"' and E( -X~) < "'
for all finite t ET. Then E(llY(s)!Gs) = E(llX(s)!Gs) +E(llM(s)jGs)
= E ( ll X ( s ) I G s ) and E ( r l E ( ll Y ( s ) I G s ) I ) = E ( r I E ( ll X ( s ) I G s ) I ) ~ E(rEC I llX(s)! ICGs)) = E(r! t~X(s) I) = E(X;) + E(-X~) < "' • Thus t rjE(llY(s)!G )j < "'a.e .. This condition seems to correspond in the 0 s
standard case to the condition that Y is a quasimartingale, i.e.
that the Doleans-measure of Y is of finite variation (see Metivier
[18]). tAT
The condition that E( r nllY(s) 2 ) is finite is just to ensure ·o
that the martingale we construct is a local A2-martingale. Without
this condition we can only prove that M is a martingale.
- 33 -
Proof of Proposition 25: Define M by M(O) = 0 and
t1M(t) = t1Y(t) -E(t1Y(t)jGt) M is obviously a martingale. Define t
X(t) = Y(O) + L:E(t1Y(s) !G ); by hypothesis X is of finite variation. 0 s
Obviously Y = X + M •
It remalns to prove that M is a local A 2 -martingale, and to tAT
do this it is enough to prove that E( [MT ](t)) = E( L: nt1M(s) 2 ) is n 0
finite for all n EJJ and all finite tET.
The first term is finite by assumption and the third is finite since
we saw in Remark 26 that E(E(t1Y(s)!G ) 2 ) < E(t1Y(s) 2 ). s
inequality on the second term we get
Using Holder's
tAT tAT 1 tA~ 1
E( L: 0 lt1Y(s)E(L~Y(s)!G >!) < (E( L: 0 6Y(s) 2 )}2 (E( L:.llE(t1Y(s)!G )2 ))"=2 0 s - 0 0 s
tAT < E( L: n~Y(s) 2 ) , which proves that the second term is finite. - 0
This proves the proposition.
As a consequence we get that all 1'reasonable" hyperfini te
sub-martingales have Doob-Meyer decompositions:
Corollary 2 8: Let Y : T x Q + *JR. be a hyperfini te sub-martingale
such that there exists an increasing sequence {Tn} of internal
stopping times such that:
(i)
(ii)
0 , (w) + oo for almost all w • n tAT
E( L: n(t1Y(s)) 2 ) is finite for all n EN and all finite t€ T. 0
(iii) E(Y(tATn)-Y(O))<oo forall nElJ, andallfinite tET.
(iv) Y(O) is finite a.e ..
Then Y has a Doob-Meyer decomposition.
Proof: Since Y tAT
r nE ( ll Y ( s ) I G ) o
0 s
is a
Now
tAT E niECliY(s)IG )I = O tAT S
E ( Y ( t A T ) - Y ( 0 ) ) = E ( E nE ( ll Y ( s ) I G ) ) n 0 s
tAT
*-sub-martingale
and hence condition (iii) implies that E n!ECliY(s) IG ) I is finite 0 s
aoe. for all nE:N and all finite tET. t
But 0 T
n
it follows that for all finite tET, r!ECt~Y(s)IG )i 0 s
The corollary now follows by Proposition 25o
-+ oo a o e . , and
is finite a.e ..
- 35 -
References
1. R.M. Anderson: A Non-standard Representation of Brownian
Motion and Ito Integration. Israel J. Math. 25
(1976) pp. 15-46.
2. R.M. Anderson: Star-finite Probability Theory. Ph.D.-thesis,
Yale University, 1977.
3. H. Bauer: Wahrscheinlichkeitstheorie und Grundzuge der Mass
theorie. 2. erweiterte Auflage, Walter de Gruyter,
1 97 4.
4. A.R. Bernstein and P.A. Loeb: A Nonstandard Integration Theory
for Unbounded Functions. In A. Hurd and P. Loeb (eds);
Victoria Symposium on Nonstandard Analysis,
Springer-Verlag 1974, pp. 40-50.
5. A.R. Bernstein and F. Wattenberg: Nonstandard Measure Theory.
In W.A.J. Luxemburg (ed.): Applications of Model
Theory to Algebra, Analysis and Probability, Holt,
Rinehart and Winston, 1969, pp. 171-185.
6. J.L. Doob: Stochastic Processes, John Wiley and Sons, 1953.
7. D.N. Hoover: Probability Logic. Ann. of Math. Logic 14 (1978)
pp . 2 8 7- 31 3 .
8. A.E. Hurd: Nonstandard Analysis and Lattice Statistical
Mechanics: A Variational Principle. University of
Victoria, 1977.
9. H.J. Keisler: Hyperfinite Model Theory. In R.O. Gandy and
J.M.E. Hyland (eds): Logic Colloqium 76.
North-Holland 1977, pp. 5-110.
10. H.J. Keisler: An Infinitesimal Approach to Stochastic Analysis.
Preliminary Version, 1978.
11. T.L. Lindstr¢m: Nonstandard Theory for Stochastic Integration.
Unpublished.
- 36 -
12. T.L. Lindstr¢m: Hyperfinite Stochastic Integration II:
Comparison with the Standard Theory. Matematisk
institutt, Universitetet i Oslo, ?reprint Series, 1979.
13. T.L. Lindstr¢m: Hyperfinite Stochastic Integration IIT:
Hyperfinite Representations of Standard Martingales.
Matematisk institutt, Universitetet i Oslo, ?reprint
Series, 1979.
14. P.A.Loeb: A Non-standard Representation of Measurable Spaces,
L , and L * . In W. A. J. Luxemburg and A. Robinson ( eds): (X) (X)
Contributions to Non-standard Analysis,North-Holland
1 9 7 2 ' pp. 6 5-8 0 .
15. P.A.Loeb: A Non-standard Representation of Borel Measures and
a-finite Measures. In A. Hurd and P. Loeb (eds):
Victoria Symposium on Non-standard Analysis,
Springer-Verlag 1974, pp. 144-152.
16. P.A. Loeb: Conversion from Non-standard to Standard Measure
Spaces and Applications in Probability Theory.
Trans. Amer. Math. Soc., 211 (1975), pp. 113-122.
17. P.A. Loeb: An Introduction to Non-standard Analysis and
Hyperfinite Probability Theory. University of Illinois.
18. M. M~tivier: Reelle und Vektorwertige Quasimartingale und die
Theorie der Stochastischen Integration. LNM 607,
Springer-Verlag 1977.
19. P.-A. Meyer: Martingales and Stochastic Integrals I, LNM 284,
Springer-Verlag 1972.
20. P.-A. Meyer: Un Cours sur les Integrales Stochastiques. In
P.-A. Meyer (ed.): Seminaire de Probabilite X.
LNM 511, Springer-Verlag 1976, pp. 245-400.
21. J. Neveu: Discrete-Parameter Martingales. North-Holland 1975.
22. H.L. Royden: Real Analysis. Second Edition, Macmillan 1968.
23. K.D. Stroyan und W.A.J. Luxemburg: Introduction to the Theory
of Infinitesimals. Academic Press 1976.