Math 6261 23 -

04 -
2/

ihm
$5.3 Doob's Martingale convergence

Def. Let (, 5,P) be a

prob. space, F. *Fut... and an

sequence of
increasing sub-r-algebras of 5.
A
sequence X, X2,..., of rvis with Xn being En-measurable
and ElXulso, is said to

a
submartingale if E)Xn+1/Fn)?Xn as

a
martingale if E(Xn+1/Fn)
=

Xn as

a
supermartingale if
E (XnxilEn) Xn
= as.

the
fundamental result about martingale is the following
theorem
of Doob:

Thi 5.6 (Submartingale convergence Thm)

* bounded submartingale (i.e. Sup ElXn)<0)
Every
converges a.s.

Below we
prove
this result by following the
approach
in the book "Ergodic theory" of K. Petersen.
we first give the following.
↳e5] (Krickeberg Decomposition (

Every2*-bounded submartingale [Xn) difference is the

of non-negative martingale [Mn)

a
a
non-negative and

supermartingale 9Sn). That is, Xn=

Ma-Sr.

Pf. Notice that Xu"is a

non-negative submartingale.
Indeed,

ESXnE,(Fn) = E(Xn+15n) Xr.

=

Since E(Xnt)Fn) =0, we obtain E(Xut/In) = Xut,

thus, (X) is a
non-negative submartingale.

Next fix m. Let ncm. Then

E(XnE(*m) E) E(Xni/En)(Fm)
=

E) X) Fm).
I

Hence E)X/5m) increases in n.

 I
Write Mm= E(X)Fm).
We claim
that Mm is a
martingale. Observe that

E(Mmx( En)= 5
(IE) X) Fnx))Fn)
imE(E(XI)(In)
-

IE) XR)Fn)
=

=>
Mn.

Hence Mn is a
Martingale.

Notice that Mutht, since E)Mn) =

limnE(E(X*nl)
Tim ESXR
=

Sup
<
E(XR) s0.

Finally, let Sn=Mn-Xn.

Recall that Mn=E) XutilFu) E(XnxlIn) =Xn,

=

So Sn20. Moreover, E(Sn) =

E(Mn) +E((Xnl)
E(XR)
syp
-> >
<0.
Hence SntL". Now

E)Sn+11Fn) E(Max)Fn) =
-

E)XatIFn)
-Mn-Xn= Sn.

So Sn is a

supermartingale. H

5.8
m
(optional sampling).
Let SXnd be a
non-negative supermartingale wirt on
Let
5, I be
stopping times, that is,

Sw: swssneFn and

[we(wisneEn
for m =1,2, ....
Define

G Xt(w)
X (w)
ifthermites
=
=

(w)

and
similarlyfor XU(w). Suppose that&
SI, then

E)Xe) = E(Xrl.
Pf. Fix mes. For him, we have

S50 =
my
X=anP=
/srem, zin
X-
dp +

((r n)Xndp
m,
=
x

Sgr=m,tan X-d+/r=m,t n} Xaxd

*

x
Since (W=m, t> nyeEn,
Xn?
E(XntilEn)(
=>

Ssrm, zany Xinn+dP +

S[r m, z
= ngXx(n+1)4P
-Scramb X=xcn+1 dP. (1)

It
follows that for ncm,

SaremyYordP=Ss=msXmP=S,w=myX+ mdP
=Sar=my XtandP (byck).

=Saram, or Xand
Thus

Ssr-my Xrap" Ssram,to an ap

 Sar
-

m,
=

sw
XedP

=Scrimy Xedp. (since Xt 0 =

g2 +03)
=

Sumining on in gives

E Xr) E(Xe).
)

=
1

of of male
convergence ihm:
show that nt-bounded
ByLem 5.7, it is enough to
every
non-negative supermartingle converges as

Let (And be a
non-negative (*-bounded
supermartingale.
suppose
on the
contrarythat [Xn) does
converge
not on a set
Then
with
positive prob. I OS2SBs such that

E [w:
=

I Xn(w) a
<- mXn(wIS
has
positive prob.

Define of stopping
times (Tibit by
sequence
a

To =
1,
 [ci+(w) inf)n> Tailw):
=
Xn(w) >BY if i co

and

[2i (w) if[n<Tzin(r): Xn(w)$G]

=

if is
1,

where we
inf4=+0. Then

To E, I I... .

Het p, PSW: j(w)sc) Then

= we have Paj Pajis.
Moreover

Paitl ) *FaitdP= S Xenit P

3 Tzi+, <0]
->
SX 2,d =

S
=

XTz;dP
St2rSc}
↳ 2 Pai

Hence
Paitis
Pai Pins
=
...
()"P to.

But Pait?PCE) 20, leading

to a contradiction.

H
 Remark: · Under the
assumption of thin 5.6, we
onlyhave.
SXn3 converges are.

·
It is
possible that Xu does not converge in L*.

Def. A collection of rv.'s Xi, ie, is said to be

uniformlyintegrable if
Hi i Sixisms IXildp) 0
=

is

that the
·

If we
pick M
large enough so
sups1,
then
EIXil=
SsIXiki3 Ixilap+/S/iKR IXi/dp
* I M
+

for all is1.

The 5.9 Let [Xn] be a

uniformlyintegral submartingale.
Then it and in it.
converges
as

Pf. It
follows from Thm5s and the Vital
convergence ihm. A