892 CHAPTER 33. QUADRATIC VARIATION AND STOCHASTIC INTEGRATION

= c∫

∫ T

0d [M]dP≤

∫Ω

(M∗)2 dP < ∞ (33.16)

Hence for a.e. ω, limk,m→∞

∫ T0 | fk (s)− fm (s)|2 d [M] (s) = 0. Also from the fact these

elementary functions are all bounded, that inside integral on the right in 33.15 is no morethan K [M (T )] for a constant K which comes from the upper bound of all these elementaryfunctions. This is a function in L1 (Ω) by 33.16. Now if A is a measurable set,∫

A

∫ T

0| fn+k (s)− fn (s)|2 d [M] (s)dP≤ K

∫A[M] (T )

and [M] (T ) is a function in L1 so this collection of functions of ω ,∫ T

0| fk (s)− fm (s)|2 d [M] (s)

is uniformly integrable. By the Vitali convergence theorem,

limk,m→∞

∫Ω

∫ T

0| fk (s)− fm (s)|2 d [M] (s)dP = 0.

By 33.15,{∫ t

0 fkdM}∞

k=1 is a Cauchy sequence in M 2T (U) and so there is a unique martin-

gale I (t)≡∫ t

0 f dM such that∫ t

0 fkdM→ I (t) in M 2T (U) . Also by Proposition 31.7.2, there

is a subsequence which converges uniformly on [0,T ] for a.e. ω . In case f is of boundedvariation in addition to being continuous because in this case, the Stieltjes sums

∫ t0 fnk dM

converge to∫ t

0 f dM. This proves the following interesting relationship.

Proposition 33.1.1 In the above context where M∗ ∈ L2 (Ω) , M a martingale with val-ues in U a separable Hilbert space, suppose t → f (t,ω) is of bounded variation and iscontinuous and ω→ f (t,ω) is adapted to the filtration Ft . Also suppose (t,ω)→ f (t,ω)is bounded. Then the ordinary Stieltjes integral

∫ t0 f (s)dM (ω) is a martingale.

In the above argument, it was not necessary that t→ f (t,ω) have bounded variation sothe I (t) is in a sense more general than the Stieltjes integral but it extends the idea of theStieltjes integral.

What if M is only a local martingale? Then you could let σm be the first hitting time ofm by ∥M (t)∥ and you could repeat everything and get∫ t

0f (s)dMσm =

∫ t∧σm

0f (s)dM

is a martingale. The approximating sums in this case would be

mn−1

∑k=0

f (tnk )(Mσm

(t ∧ tn

k+1)−Mσm (t ∧ tn

k ))→∫ t

0f (s)dMσm

mn−1

∑k=0

f (tnk )(Mσm

(t ∧ tn

k+1)−Mσm (t ∧ tn

k ))

=

(mn−1

∑k=0

f (tnk )(M(t ∧ tn

k+1)−M (t ∧ tn

k )))σm

→∫ t∧σm

0f (s)dM

892 CHAPTER 33. QUADRATIC VARIATION AND STOCHASTIC INTEGRATION=e [atmjars [ (par <« (33.16)Hence for ae. @,limgmsoo Jo |fe(s) — fin (s)|?d[M](s) = 0. Also from the fact theseelementary functions are all bounded, that inside integral on the right in 33.15 is no morethan K [M (T)] for a constant K which comes from the upper bound of all these elementaryfunctions. This is a function in L' (Q) by 33.16. Now if A is a measurable set,r 2| ff town (s)— ols) Pala (s)ar <x f(r)and [M](T) is a function in L! so this collection of functions of @ ,[iy nyais uniformly integrable. By the Vitali convergence theorem,Ttim ff \fiels) —Jin(s) Pa (| (9) dP = 0.kim JQ JOBy 33.15, {fo fed}, is a Cauchy sequence in .4 (U) and so there is a unique martin-gale I(t) = {5 fdM such that {5 f,dM — I(t) in. (U). Also by Proposition 31.7.2, thereis a subsequence which converges uniformly on [0,7] for a.e. @. In case f is of boundedvariation in addition to being continuous because in this case, the Stieltjes sums fo tndMconverge to Jo fdM. This proves the following interesting relationship.Proposition 33.1.1 In the above context where M* € L? (Q) , M a martingale with val-ues in U a separable Hilbert space, suppose t — f (t,@) is of bounded variation and iscontinuous and @ — f (t,@) is adapted to the filtration ¥;. Also suppose (t,@) > f (t, @)is bounded. Then the ordinary Stieltjes integral Jj f (s)dM (@) is a martingale.In the above argument, it was not necessary that t > f (t,@) have bounded variation sothe /(t) is in a sense more general than the Stieltjes integral but it extends the idea of theStieltjes integral.What if M is only a local martingale? Then you could let 6,, be the first hitting time ofm by ||M (t)|| and you could repeat everything and get| " £ (s)dM% = [ ee” 65) dMis a martingale. The approximating sums in this case would bemn—1YF lf) (Mm (rata) Me" (etp)) + [f(ak=0mn—1|de F(a) (Mom (HA 1) — MO" (tA)k=0my—1 om t\Om= ( y f(g) (a en) ten) >| f (s)dMk=0 0