martingales - moodle 2020-2021
TRANSCRIPT
APA lecture 12
Martingales
Def : Let ( r,
F,
P ) be a probability space
A filtration is a sequence of sub - r - fields off( Fh
,n EIN )
such that I Ctu er
,the N
.
Ex : D= [ o,
a ],
F = B ( Co,i ] )
,D= lebesgue measurea for ]
Xn I w ) = nth decimal of co E ( 0,1 ] n 21
Fo = { ¢,
r },
FICK ),
ferch,
ka ) - - . I -- orch -
. x !,
W = 0,17392 - - -
e m
I
Def : Let ( I,
new ) be a filtration on Cr ,F
,P )
A square - int. martingale wit ( In
,new ) is a discrete - time
stochastic process ( Mn,
n EIN ) such that
( i ) E( Mi ) c too th EIN
( ii ) Mn is I - measurable KEN ( adopted process )→ Ciii ) E ( Mine .
Ifn ) = Mn as.
then
Rink : Because EC Mine ,IF ) is Fu - measurable
by definition,
condition Cii ) is actually redundant.
Rink : martingale # Markov process
Example The simple symmetric random walk.
Let ( x,
use ) be iidr.u.rs st.
PCEX -- eel) - PHX = -131=7
So = o,
{ = Xt - - .tl/nhZ1
to - { do,
r },
Fu = Ha,
. . .
,x ) n 21
Then § ,new ) is a sq - int
. martingale iv. rt ( I,
he N )
Proof :
( i ) E (d) = n.EC/h2)=nctosVnz1
( ii ) Sn is Fu - measurable : yes ,as Sext .
-
t Xn ✓
( iii ) E ( Sue ,IF ) = E ( Et Xen IF ) = ECG II ) e EHealth )
= Su t E ( Xue .) = Sn as . #
t
Fu - measurable↳
the,
I I
Remark : If (
Mn,
new ) is a sq - out . martingale ,then
F- C then ) = ECE.ch#IIzD--ECrh)tuerNSo E ( oh ) = E ( Mn - a) =
- . - . = HE (Mo ) th E IN
So a martingale is a process with constant
expectation ; but it is also much more than that !
Ex : The process ( X,
u za ) with X,
i id r.
V.
has constant
expectation ,but it is not a martingale : EHealth ) - Elka ) = o
+ Xn:Illustration for Guin EN )
su;!Ekaanti
sie.
n nee
Property: let ( Mn
,nze ) be a square . integrable martingale
. E ( Muerte ) - F- ( E ( Mutz IIe.) IF ) = Ether IE )
= Mn
- - - - - - - -
IFI Ee .
. E ( Mum II ) = Mn th,
men ,-
i
Remark : martingale t Marka Su .
+martingale property : E c the ,th ) = Mn I nos
. Marka property : B ( { Man EB } Itn ) =P ({ thereBllrh )equivalently : HBEBGR )
Ecg
cnn.it/Tn)--ECgCrh-u)lrh)fgegcuz
)
Sub- and supermanhhgales ( square - integrable )
Defy : A process ( Mn,
new ) is a
✓
SEPT.
} martingaleWar
.
t.
a filtration ( Tn,
new ) if :
( i ) ECT) cease knew
→ ( ii ) Mr is In - measurable Vn E N
( iii ) E ( Mn ,I Tn ) Z Mn as
.
the N
Rem : a ) if M is asub - gnarlingale
,then E ( the . ) Z Eth )
.super - E
2) Not every process is either a sub martingale ,
or a martingale,an asupermartihgale .
Example ( simple asymmetric randan walk )
Soo,Sn=Xnt . . .th with X, use ) i id r
.v. 's with
PC { teen 3) =p = e - P ( { Xi - R) apse
& ,new ) is a safe. } martingale if { PPE
If: E ( See Itn ) = ECS II ) e Eke . II ) = St ECx
Property : If M is a square - integrable martingale⇐& ✓
& Cf : Nsr is a convex function st Ely (oh )%tos
Then @(Mn ),
ne IN ) is a sub martingale .
KEN
PI : E Cyane . ) Itn ) 2 4 CE Cohn.
II )) = 4cm )✓
Jensen
Stopping times & optional stopping theorem
Deff: A random time is a random variable T : R - s N
Let ( I,
ne IN ) be a filtration on Cr,
f,
P ).
Def: A stopping time w.r.t.CI,
n c- IN ) is a random true
such that { WE r : Tla ) = u } E Tu th EN
Exi let ( xn,
ne N ) be a discrete - time stochastic process
I = 5C Xo,
X,
- - -
,X ) new
Then Ta = if { n e N : Anza } is a stopping time
a c- OR W.r. t ( I,
u EN )
In
qAER -
° !
an° .o
°
i
• !⑦ ⑧
8
•.
I n•
Ta
Proof.
To be checked :{Ta -
- n } E I knew
{ Ta -
- n } = { Xue a toe Ken - a & X za }
=L?. .
Kusa } ) n { X za } e E-
-
C- Fec Fn C- I ✓
Def : If ( xn,
new ) is a discrete - time
stochastic process and T is a random time,
then we define
Xtlw.tl/twCw)=eIXlw)1Et=n3lw)Def:We say that a randomtime is
boundedif
there exists N E N such thatTcu ) EN touch
T
( a PIET en } ) =L )
Optional stoppingtheorem
. Let M be a ( square - integrable ) martingaleWent
. a filtration ( Fh,
NEIN )
Let T be a
ftp.p.h.g.h.k.e.w.r.t. ( Fu
,new )
such that F N E IN withE.E III.EItakeThenE (Mo ) = E ( Mt ) = E ( Mn )
T T
. If M is a ( square - integrable ) sub martingale ,
then ( Echo ) E)E ( Mt ) E E ( Mn )- - - - - -
-
pnot proven here
Proof :
EE.
As of Tla ) EN KWEI,
we have
Mtlw ) = Mta ,Cw)=⇐n Malu ) . lEt=n3lw )
-EEE' - ⇐ retainer.⇒÷EE¥¥:¥f÷i!÷
,
= !⇐E(E( Mn - let.us/FnH=IEECMn-1q-=nD--ECMn. let.mg/=ECMn)--ECMo )
Is- - - - -
p
( M = martingale )#
When does the optional stopping theorem fail ?
. If T is not a stopping time :
Ex : T = value of n E { o . .
N } such that Mn ? Mm
V-mcfa.nlThen both E ( Mt ) s Echo ) & ECM
,- I > ECMN)
. If T is not bounded :
Ex : Consider ( X,
n za ) be a sequence of iidr.u.rs
s.t.PK/h--eeH=PGxr--eD=I;Mo--o,Mn=4t.-exn,nnM-martingale , Ta = ihf { ne N : MnZa } a c- IN#
Then E ( Mta) = a so = Echo ) .
= a
Why is it an interesting theorem ?
Consider again the previous example with N -
- simple
Symmetric random walk : Men = hit - -t Xn
.
X 's= can tosses I?FI ⇒ Mn = cumulated gain at time h
AN 21,
E c Mn ) = E ( Mo ) = o
Let I mirinf { n e IN : Mn Z to },
N }
E ( Mt ) = E ( no )- - - -
+ to with high probability↳{ I- x with tiny probability ( when T= N )