File C 8

dimH
1 Hitting time for SRW

2 Non-negative superharmonic functions of martingale
Markov Processes and Martingales 3 Martingal convergence
4 Sums of zero-mean independent r.v.
Károly Simon 5 Doob decomposition
Department of Stochastics
6 Closing
Institute of Mathematics
Technical University of Budapest
7 Preparation for the Uniform families
www.math.bme.hu/~simonk
8 Uniform families
C File 9 LIL
10 Doob’s Lp inequality
11 Kakutani’s Theorem on Product Martingales
Károly Simon (TU Budapest) Markov Processes & Martingales C File 1 / 188 Károly Simon (TU Budapest) Markov Processes & Martingales C File 2 / 188
Hitting time for SRW Hitting time for SRW cont.

Let X be a r.v. with
è é 11 θ 2
1 E eθX = e + e−θ .
P (X = 1) = P (X = −1) = . 2
2 So,
d è é
Let X1 , X2 , . . . be iid, Xi = X . Let S0 := 0 and (2) E (sech θ · eθXn ) = 1,
Sn := X1 + · · · + Xn for n ≥ 1. We define
where sech θ = 1/ cosh θ = 2/(eθ + e−θ ). (Read:
(1) T := min {n : Sn = 1}
hyperbolic secant). So, M θ = (Mnθ ),
Our goal on the next slides is to find the probability
è é (3) Mnθ = (sech θ)n eθSn .
generating function α Ô→ E αT of T .
We write Fn := σ(X1 , . . . , Xn ) = σ(S0 , S1 , . . . , Sn ). Then is a martingale by A File Example 1.8. Using that T is a
S = (Sn ) is adapted to {Fn }. Clearly, stopping time:
Hitting time for SSRW cont. Hitting time for SSRW cont.
Note that eθ·ST = eθ by the definition of T . So we
è é è é obtained:
(4) E MTθ ∧n = E (sech θ)T ∧n eθST ∧n = 1 ∀n. è é
(6) E (sech θ)T = e−θ , θ > 0.
Assume that θ > 0. Then
exp (θ · ST ∧n ) ≤ eθ . So, MTθ ∧n ≤ 1 (this follows Observe that
from(3) since sech θ ≤ 1). 
 1, if T < ∞;
T
lim M θ = MTθ , where MTθ = 0 if T = ∞. (7) lim (sech θ) = 
n→∞ T ∧n θ→0+ 0, if T = ∞.
Using (4) and the Dominated Conv. Thm.
è é è é
Using Dominated Conv. Thm., by (6), (7)
(5) E MTθ = 1 = E (sech θ)T eθ , è é è é
ü ûú ý 1 = lim+ E (sech θ)T = E ✶{T <∞} = P (T < ∞) .
0 if T =∞ θ→0
Hitting time for SSRW cont. Explanation of (9)

Recall that for an |x | < 1 we have
In (6) let α := sech θ. Then using (6) we get  
∞
Ø β ∞
Ø β(β − 1) · · · (β − n + 1) n
β  x n
∞
Ø è é (1 + x ) = = x .
(8) n
α · P (T = n) = E α T n=0 n n=0 n!
n=1
5 √ 6
Using this with β = 1/2 and x = −α2 we get
= e−θ = α−1 1 − 1 − α2 .
 
−1
3 √ 4 ∞
Ø 1
Hence, (10) α 1− 1− α2 =  2 (−1)n+1 α2n−1 .
n=1 n
112
m+1
(9) P (T = 2m − 1) = (−1) · m
2 . Putting together (8) and (10) yields (9). The heart of
the matter was (8). Now we give an alternative proof for
(8).
An easier proof for (8) An easier proof for (8) cont.
Fact 1.1
Consider a (simple symmetric) random walker who starts
from −1. Let This is immediate from the Strong Markov Property of
T1 be the first time she reaches 0, SSRW. (We learned about the Strong Markov Property
T2 be the time she needs to get to 1 AFTER she on slide # ?? of File A of the course Stochastic
reached 0. Processes.
(So, she needs time T1 + T2 to get to 1 from −1.) Then
d d
(a) T1 = T2 = T ( T was defined in (1)).
(b) T and T1 are independent .
An easier proof for (8) cont. dimH

è é
Let α ∈ (0, 1]. we define f (α) := E αT . ( T was 3 Martingal convergence
defined in (1)). Then 4 Sums of zero-mean independent r.v.
1 è é 1 è é 5 Doob decomposition
f (α) = E αT |X1 = 1 + E αT |X1 = −1 6 Closing
2 2
1 1 è 1+T1 +T2 é 7 Preparation for the Uniform families
= α+ E α = 12 α + αf (α)2 .
2 2 8 Uniform families
è √ é
This implies that f (α) = α−1 1 − 1 − α2 , that is (8) 9 LIL
holds. 10 Doob’s Lp inequality
Superharmonic functions Superharmonic functions cont.

First recall from File A: Definition 2.2, Theorem 2.6 and
Theorem 6.6 (Optional stopping thm). Let S be a Recall that for a function h : S → R+ we defined the
countable set and let P = (p(i, j))i,j∈S be stochastic function Ph : S → R+ by
matrix. Let µ be any measure on S. Let Z = (Zn )∞ n=0 be q
the Markov chain corresponding to the transition (Ph)(i) = p(i, j)h(j), i ∈S.
j∈S
probability matrix P and initial distribution µ. That is
Assume that h is P-superharmonic . That is
(11)
n−1
Ù (13) (Ph)(i) ≤ h(i), ∀i ∈ S.
Pµ (Z0 = i0 , Z1 = i1 . . . , Zn = in ) = µi0 · p(ik , ik+1 ).
k=1
It follows from File A Theorem 2.6 that h(Zn ) is a
Let Fn := σ(Z0 , . . . , Zn ). Then supermartingal.
(12) Pµ (Zn+1 = j|Fn ) = p(Zn , j).
Superharmonic functions cont. Superharmonic functions cont.

Proof
Namely, Assume that the chain Z = (Zn ) is irreducible and
Ø recurrent. Then fij := Pi (Tj < ∞) = 1, ∀i, j ∈ S ,
Eµ [h(Zn+1 )|Fn ] = h(j) = (Ph)(Zn ) ≤ h(Zn ).
p(Zn ,j) where Tj := min {n ≥ 1 : Zn = j} . Let h be a
superharmonic function. Consider the process: h(Zn ) .
Lemma 2.1 By Theorem 2.6, h(Zn ) is a supermartingale. Then by
File A, Theorem 6.8, (a corollary of the Optional
The Markov chain Z = (Zn ) is irreducible and recurrent Stopping Theorem) we have
iff every non-negative superharmonic function is constant.
è é
We prove only the =⇒ implication. h(j) = Ei h(ZTj ) ≤ Ei [h(Z0 )] = h(i).
That is h is constant.
dimH Uncrossings
2 Non-negative superharmonic functions of martingale Given a process X = (Xn ) and numbers a < b. We
3 Martingal convergence define the upcrossings of [a, b] by Xn by time N ∈ N as
follows: Let S0 := 0, a < b and let us define the
4 Sums of zero-mean independent r.v. following stopping times:
5 Doob decomposition
Rk : = min {n ≥ Sk−1 : Xn > b}
6 Closing
Sk : = min {n ≥ Rk : Xn < a} .
8 Uniform families The upcrossings of [a, b] by time N is
9 LIL
UN [a, b](ω) := max {k : Rk (ω) ≤ N} .

Now we construct a game which corresponds to this
process as we did on slide # 24 in File A.
Upcrossing cont. Uncrossing cont., the definition of Cn

Imagine that somebody plays games at times Starting from time n = 0, we define Cn the stake at time
k = 1, 2, . . . . Let Xk − Xk−1 be the net winnings per n as follows:
unit stake in game n. C1 := {X0 <a}
Cn is the player’s stake at time n (specified on the next Cn = + .
{Cn−1 =1}∩{Xn−1 ≤b} {Cn−1 =0}∩{Xn−1 <a}
slide) which is decided based upon the history of the
game up to time n − 1. The winning on game n is On the next Figure we color (n, Xn ) ∈ R2
Cn (Xn − Xn−1 ). The total winning after n game is white if Cn = 0,
q
black if Cn = 1.
(14) Yn := Ck (Xk − Xk−1 ) =: (C • X )n . In summary:
1≤k≤n
The first black ball appears where we go below a for the
By definition: Y0 (ω) ≡ 0. (C • X )0 = 0 and first time. The only way to get a black ball is:
Fn := σ(X1 , . . . , Xn ). So, we assumed above that either after a white ball which is below a or
Cn ∈ Fn−1 . Clearly, Yn − Yn−1 = Cn (Xn − Xn−1 ). after a black ball which is below b .
Uncrossing cont.
Observe that
Remark 3.1
(a) All the increase of Y is due to a run of black
balls which does not end at N.
(b) All the decrease of Y is due to a run of black
balls which ends at N.
Uncrossing cont.
Remark cont.
(c) every run of black balls which ends earlier
than N follows a white ball which is below a
(let say Xk ) and ends with a black ball above
b (let say Xℓ ). Then there is exactly one
uprcossing between k and ℓ which corresponds
to this run of black balls.
(d) The maximum decrease in Y caused by the
last run of black balls is (XN − a)− , where
(XN − a)− := max {0, −(XN − a)} .

Figure: This and the previous Figure is from Williams’ book.
Uncrossing cont. Uncrossing cont.
It follows from Remark 3.1 that

Proof.
− Using that C is previsible, bounded and non-negative we
(15) YN (ω) ≥ (b − a)UN [a, b](ω) − [XN (ω − a)] .
may apply Theorem 4.2 of File A to obtain that
Y = C • X is also a supermartingale . Then
Lemma 3.2 (Doob’s Upcrossing Lema)
Let X = (Xn ) be a supermartingale . Then E [YN ] ≤ E [Y0 ] = 0 .
è é
(16) (b − a)E [UN [a, b]] ≤ E (XN − a)− . This and (15) together imply that (16) holds.
Uncrossing cont. Uncrossing cont.

Corollary 3.3 Proof.
By (16), for every N we have
Let X = (Xn ) be an L1 -bounded supermartingale:
(17) sup E [(|Xn |)] < ∞ . (20) (b − a)E [UN [a, b]]
n ≤ |a| + E [|XN |] ≤ |a| + sup E [|Xn |] .
n
For an a < b let U∞ [a, b] := lim UN [a, b] Then
N→∞ Let N ↑ ∞ and use Monotone Convergency
Theorem.
(18) (b − a)E [U∞ [a, b]] ≤ |a| + sup E [|Xn |] .
n
Definition 3.4
Hence,
For definiteness for a general X = (Xn ) we define
(19) P (U∞ [a, b] = ∞) = 0. X∞ (ω) := lim sup Xn (ω) .
n→∞
Doob’s Forward Convergence Theorem Forward Convergence Theorem cont.

Theorem 3.5 (Doob’s Forward Convergence Proof.
Theorem) Û
I J
Λ= ω : lim inf Xn (ω) < a < b < lim sup Xn (ω)
n→∞
Let X = (Xn ) be an L1 -bounded ( (17) holds) a,b∈Q,
a<b ü ûú
n→∞
ý
supermartingale. Then Û
Λa,b
⊂ {ω : U∞ [a, b](ω) = ∞},
ü ûú ý
lim Xn , and X∞ < ∞ a.s. .
X∞ = n→∞ a,b∈Q,
a<b å a,b
Λ
Let Λ be the set of ω ∈ Ω for which Xn (ω) does NOT 1 2

By (19) we have P Λå a,b = 0, hence P (Λ) = 0. So, for
converge to a limit in [−∞, ∞]. So there is no limit
almost all ω the following limit exists:
even is we alow that the limit can be infinite. So, for an
ω ∈ Λc the limit n→∞
lim XN (ω) ∈ [−∞, ∞] exist. lim Xn (ω) ∈ [−∞, ∞].
X∞ = n→∞
Forward Convergence Theorem cont. Forward Convergence Theorem cont.

Corollary 3.6
proof cont.
Now we use Fatau Lemma: Assume that X = (Xn ) is a non-negative
5 6
supermartingale. Then the limit X∞ := n→∞
lim Xn exists
E [|X∞ |] = E lim inf |Xn | almost surely. (This was Theorem ?? in File E on the
n→∞
course "Stochastic Processes".)
≤ lim inf E [|Xn |] ≤ sup E [|Xn |] < ∞ ,
n→∞ n
Proof.
by assumption. This shows that P (X∞ < ∞) = 1 . So X is L1 -bounded. In deed
the limit exists and less than ∞ almost everywhere.
E [|Xn |] = E [Xn ] ≤ E [X0 ] .
Be careful. It can happen that the limit does not exists
in L1 . Then we apply Theorem 3.5.
Pythagorean formula for L2 martingales Pythagorean formula for L2 martingales
cont.
Below we always assume that M = (Mn ) is a martingale
in L2 : Namely,
n
Ø
(21) Mn ∈ L2 ∀n. (23) Mn = M0 + (Mk − Mk−1 )
k=1
Then the Pythagorean formula holds: and the increments are orthogonal: ∀s ≤ r ≤ t ≤ z we
have
è é è é n
q è é
(22) E Mn2 = E M02 + E (Mk − Mk−1 )2 . (24) E [(Mr − Ms )(Mz − Mt )] = 0.
k=1
and
This follows from the orthogonality (in L2 ) of the
increments (see Theorem ?? in File E). (25) E [Ms (Mz − Mt )] = 0.
So, we take squares of both sides in (23) to get (22).
Martingales bounded in L2 Martingales bounded in L2 cont.
Definition 3.7 Theorem 3.8

Assume that M is an L2 -bounded martingale. (That is
M = (Mn ), Mn ∈ L2 is bounded in L2 if
(26) holds). Then
è é
(26) sup ëMn ë2 < ∞ that is sup E Mn2 < ∞.
n n (28) Mn → M∞ both a.s. and in L2 .
By the Pythagorean formula
Proof.
∞ è é
(27) M is L2 bounded ⇐⇒
Ø
E (Mk − Mk−1 )2 < ∞. Assume that M is L2 -bounded. Then M is also
k=1 L1 -bounded. So, we can apply Doob’s Convergence
Theorem (Theorem 3.5).
Martingales bounded in L2 cont. Martingales bounded in L2 cont.
Proof cont. Proof cont.

This implies that the following limit exists and finite
è é ∞
Ø è é
lim M = M∞ , a.s. (30) E (M∞ − Mn )2 ≤ E (Mk − Mk−1 )2
n→∞ n k=n+1
By the Pythagorean thm.: Hence,

è é n+r è é è é
q
(29) E (Mn+r − Mn )2 = E (Mk − Mk−1 )2 (31) lim E (M∞ − Mn )2 = 0.
n→∞
k=n+1
That is Mn → M∞ also in L2 .
Let r → ∞ on both sides to obtain:
Martingales bounded in L2 cont. dimH

3 Martingal convergence
We remark that it follows from putting together (3) and 4 Sums of zero-mean independent r.v.
(31) that there is equality in (30). That is 5 Doob decomposition
è é ∞
Ø è é
6 Closing
(32) E (M∞ − Mn )2 = E (Mk − Mk−1 )2 . 7 Preparation for the Uniform families
k=n+1
8 Uniform families
9 LIL
Sums of zero-mean independent r.v. Sums of zero-mean independent r.v. cont.
Theorem 4.1
In the proof we use the following
Assume that X1 , X2 , . . . are independent and
σk2 := Var(Xk ) < ∞. Definition 4.2
∞
q ∞
q
(a) If σk2 < ∞ then Xk converges a.s.. Fn := σ {X1 , . . . , Xn } , F0 := {∅, Ω}
k=1 k=1
(b) If X = (Xn ) is bounded (that is ∃K s.t. Mn := X1 + · · · + Xn , M0 := 0
∞
q
∀n, ∀ω, we have |Xn (ω)| < K ) and Xk Further, A0 := 0, N0 := 0 and let
k=1
∞
q n
Ø
converges a.s. then σk2 < ∞ . (33) An := σk2 , Nn := Mn2 − An .
k=1
k=1
Note that it follows from Kolomogov’s 0 − 1 law that

P (Xk converges ) = 0 or 1.
Sums of zero-mean independent r.v. cont. Sums of zero-mean independent r.v. cont.
Proof of part (a)
We know that M is a martingale with
Proof of part (b)
è é è é Similarly to (34) we have
2
(34) E (Mk − Mk−1 ) = E Xk2 = σk2 . (36)è é è é è é
E (Mk − Mk−1 )2 |Fk−1 = E Xk2 |Fk−1 = E Xk2 = σk2 .
So, by (22) we get
n
since Xk is independent of Fk−1 . Using this and the fact
è é Ø
(35) E Mn2 = σk2 = An . that Mk−1 ∈ Fk−1 , we obtain that
k=1 è é
∞
2
σk2 = E Mk2 |Fk−1 − 2Mk−1 E [Mk |Fk−1 ] + Mk−1
q è é
If σk2 < ∞ then M = (Mn ) is bounded in L2 , so 2
= E Mk2 |Fk−1 − Mk−1 .
k=1
lim Mn exists almost surely.
n→∞
Proof of part (b) cont. Proof of part (b) cont.

è é
That is E Mk2 |Fk−1 − σk2 = 2
Mk−1 . This implies that For a fixed c > 0, we define the stopping time
 
T := inf {r : |Mr | > c} .
 
 Øn  n−1
Ø
 
(37) E Mn2 − 2
σk |Fn−1  2
= Mn−1 − σk2 . We defined the stopped process on slide # 35 of File A
 
ü k=1
ûú ý  ü ûú
k=1
ý as
Nn Nn−1 NnT (ω) := NT ∧n (ω).
So, we have just proved that Nn (defined in (33)) is a It follows from Theorem 6.1 of File A that N T is also a
martingale since N is a martingale as we pointed out
martingale .
above.
Proof of part (b) cont.
So, by Theorem 6.1, we have Proof of part (b) cont.
è é è é On the other hand, if T (ω) ≤ n then
(38) 0 = E [N0 ] = E NnT = E (MnT )2 − AT ∧n .
(40) MT (ω)∧n (ω) = MT (ω) (ω).
Now we prove that
It follows from our assumption that whenever T is finite
(39) |MnT | = |MT ∧n | ≤ K + c, ∀n.
we have
Namely, if n < T (ω) ≤ ∞ then
|MT (ω) − MT (ω)−1 (ω)| = |XT (ω)| < K .
|MnT (ω)| = |Mn (ω)| ≤ c by the definition of T . So, in
this case (39) holds.
Proof of part (b) cont.
However, by the definition of T , whenever T < ∞: Proof of part (b) cont.
|MT −1 | ≤ c. Using (39) and (38) we get
Using the last two inequalities and the triangular E [AT ∧n ] ≤ (K + c)2 , ∀n.
inequality we have that whenever T < ∞ we have
n
q
Remember that we assumed that Xk = Mn a.s.
|MT | ≤ K + c. k=1
convergent. This means that the partial sums
Putting together this and (54) we obtain that (39) holds {Mn (ω)}∞
n=1 is bounded for a.a. ω.
also for those ω for which T (ω) ≤ n. So we have verified
(39)
Sums of zero-mean independent r.v. cont. dimH

Proof of part (b) cont. 2 Non-negative superharmonic functions of martingale
So, there is an L such that on a set H of positive 3 Martingal convergence
measure, P (H) > 0, the partial sums are smaller than L 4 Sums of zero-mean independent r.v.
in modulus. So for any c > L we have 5 Doob decomposition
T (ω) = ∞, ∀ω ∈ H. 6 Closing
Hence for all ω ∈ H and for all n,
8 Uniform families
n
Ø by (38) è é
σk2 = An = AT ∧n (ω) = E Mn∧T
2
< (K + c)2 . 9 LIL
k=1 10 Doob’s Lp inequality
In the last proof we used only that X is bounded and the 11 Kakutani’s Theorem on Product Martingales
partial sums are bounded

Károly Simon (TU Budapest)
on a set of positive measure.
Markov Processes & Martingales C File 51 / 188 Károly Simon (TU Budapest) Markov Processes & Martingales C File 52 / 188
Doob decomposition Doob decomposition cont.

Theorem 5.1 (Doob decomposition) The last comment means that: If
æ å
Given a filtered probability space (Ω, F, {Fn } , P). Let X = X0 + M +A
X = (Xn ) be an adapted process with Xn ∈ L1 for all n. is another decomposition then
Then X has a Doob decomposition: 1 2
æ å
P Mn = M n , An = An , ∀n = 1.
(41) X = X0 + M + A ,
Corollary 5.2
where
M = (Mn ) is a martingale with M0 = 0 X is a submartingale iff A in its the Doob
A = (An ) is previsible (that is An ∈ Fn−1 ) with decomposition (41), is an increasing process . That is
A0 = 0. (An is called compensator of Xn ).
(42) P (An ≤ An+1 ) = 1.
The decomposition is unique mod zero.
Doob decomposition cont. Doob decomposition cont.

Proof Thm 5.1
Strategy: Find our what the compensator A should be. Proof Thm 5.1 cont.
If X has decomposition given by (41)
Namely, let Mn := Xn − An . Then M = (Mn ) is a
martingale:
(43) E [Xn − Xn−1 |Fn−1 ]
= E [Mn − Mn−1 |Fn−1 ] + E [An − An−1 |Fn−1 ] E [Mn |Fn−1 ] = E [Xn − An |Fn−1 ]
= 0 + An − An−1 . E [Xn |Fn−1 − An ]
E [Xn |Fn−1 ] − An−1 − E [Xn |Fn−1 ] + Xn−1
So, the decomposition in (41) comes from:
Mn−1 .
n
q
(44) An := E [Xk − Xk−1 |Fk−1 ] .
k=1
Remark 5.3
Assume taht X = (Xn ) is an L2 martingale, with X0 = 0.
Proof Corollary 5.2 Then (by Theorem ?? of File E of the Course Stochastic
X = (Xn ) is a submartingale if Processes), ϕ(Xn ) is a submartingale for any convex
function ϕ. In particular,
E [Xk − Xk−1 |Fk−1 ] ≥ 0.
(45) Xn martingale =⇒ Xn2 submartingale.
On the other hand, by (43) this happens when
Let A = (An ) be the compensator of X . That is
An ≥ An−1 .
Xn2 − An is a martingale.
We say that (An ) is the quadratic variation of (Xn ).


Remark 5.4 Remark 5.5
Let X = (Xn ) be an L2 martingale. Then the quadratic Let X = (Xn ) be an L2 random walk that is a martingale.
variation is Then the quadratic variation is An = nσ 2 . The proof is
n
a home work # 22.
Ø è é
An = E (Xk − Xk−1 )2 |Fk−1 .
k=1 Remark 5.6
Recall: Let Xn be an L2 martingale with quadratic variation
Xn = Z 0 + Z 1 + · · · + Z n A = (An ). Let Cn be previsible and also L2 . Then the
quadratic variation B = (Bn ) of their martingale
is a random walk if Zn are iid and L1 . Let µ := E [Xi ].
Then Xn − nµ is a martingale . This was Example ?? in transform Y = C • A is B = C 2 • X . The proof is
File E in the course "stochastic Processes". assigned as home work # 23.
The angle bracket process The angle bracket process cont.

Let M be an L2 martingale with M0 = 0. We write Since E [Nn ] = E [N0 ] = 0, we have
2 è é
(46) M = N + A, E Mn2 = E [An ] .
where N is a martingale and A is previsible and
So, by the monotone convergence theorem:
increasing with
(47) N0 = 0 and A0 = 0. (49) M is bounded in L2 ⇐⇒ E [A∞ ] < ∞.
We write Note that
(48) < M >:= A .
è é
So it makes sense to define (50) An − An−1 = E Mn2 − Mn−1
2
|Fn−1
è é
A∞ := n→∞
lim An , a.s.. = E (Mn − Mn−1 )2 |Fn−1 .
Convegence of Mn vs. A∞ < ∞ Convegence of Mn vs. A∞ < ∞ cont.
Theorem 5.7 Proof of Theorem 5.7 part (a)

Let M be an L2 martingale with M0 = 0 and let A be a Recall: A is previsible, that is
version of < M >. Then
(51) Az+1 ∈ Fz , ∀z ≥ 0.
(a) For almost all ω for which A∞ (ω) < ∞ we
have n→∞
lim Mn (ω) exists. Hence
(b) Assume that M has uniformly bounded S(k) := inf {n : An+1 > k}
increments. Then for almost all ω for which
is a stopping time. S(k) = ∞ means that Ar ≤ k for all
lim M (ω) exists, we have A∞ (ω) < ∞.
n→∞ n r ∈ N.
Convergence of Mn vs. A∞ < ∞ cont. Convergence of Mn vs. A∞ < ∞ cont.
Proof of Theorem 5.7 part (a) cont. Proof of Theorem 5.7 part (a) cont.
The relevance of S(k): To prove the Fact we need to verify that
Û
(53) An∧S(k) (B) ∈ Fn−1 , ∀B ∈ R, ∀n ≥ 1.
(52) {A∞ < ∞} = {S(k) = ∞} .
k To see this, we fix an nî and a B ∈ R
ï (Borel set on the
So, the right hand side above is a partition of line) and we represent An∧S(k) ∈ B as:
{A∞ < ∞}. î ï
An∧S(k) ∈ B = F1 ∪ F2 ,
Fact 5.8
where F1 , F2 correspond to S(k) = r < n, S(k) ≥ n
The stopped process AS(k) is previsible . respectively. That is:

Proof of Theorem 5.7 part (a) cont.
n−1
Û
F1 : = {S(k) = r ; Ar ∈ B} Ü
r =0 F2 = {An ∈ B} {S(k) ≥ n}
 
n−1
Û r
Ü Ü Ü n
Ü Ü
= {A
r ∈ B} {Aℓ ≤ k} {Ar +1 > k} . = {An ∈ B} {Ar ≤ k} .
r =0 ℓ=1 r =1
Using that Ar +1 ∈ Fr for every r , (A is previsible), and Clearly, all of the events in the previous line are in Fn−1 .
r + 1 ≤ n above, we get that This completes the proof of (53). So, we have verified
Fact 5.8 above.
(54) F1 ∈ Fn−1 .

So, we know that Proof of Theorem 5.7 part (a) cont.
AS(k) is previsible (Fact 5.8), By the uniqueness of Doob decomposition of
N S(k) is a martingale . This is so because (M S(k) )2 = martingale + previsible

(a) N = M 2 − A (defined in (46)) is a
martingale (by definition), and since
(b) S(k) is a stopping time (M S(k) )2 = N S(k) + AS(k) ,
so, N S(k) is a martingale by Theorem 6.1 of File A we obtain that
(this theorem says that in general, a martingale
raised to the power of a stopping time is a (55) < M S(k) >= AS(k) .
martingale).

By the definition of S(k) we have that Proof of Theorem 5.7 part (a) cont.
AS(k)
n ≤ k, ∀n. Finally, we can apply Theorem 3.8 (L2 bounded
martingale convergence thm.) to get
So, è é
S(k)
E A∞ < ∞. (56) lim Mn∧S(k) exists almost surely.
n→∞
Using this, we can apply (49) for M S(k) instead of M to

obtain that M S(k) is an L2 bounded martingale.
Proof of Theorem 5.7 part (a) cont. Proof of Theorem 5.7 part (b)
However, We argue by contradiction. Assume that part (b) of
Û
Theorem 5.7 is not true. Then we have
(57) {A∞ < ∞} = {S(k) = ∞} . A B
k
(58) P A∞ = ∞, sup |Mn | < ∞ > 0 .
n
So part (a) follows from the combination of (56) and
A B
(57). Namely, if A∞ (ω) < ∞ then by (57) there exists a
k = k(ω) such that S(k)(ω) = ∞. Then Then ∃c s.t. P A∞ = ∞, sup |Mn | < c > 0. That is
n
Mn∧S(k)(ω) (ω) = Mn (ω) ∀n. (59) P (A∞ = ∞, T (c) = ∞) > 0 ,

Hence by (56) we obtain that part (a) holds. where T (c) = inf {r : |Mr | > c}

Proof of Theorem 5.7 part (b) cont.
By Theorem 6.1 of File A we have
Proof of Theorem 5.7 part (b) cont.
That is, ∃K s.t.
èé
(60) 0 = E [N0 ] = E NTn (c)
è é è é è é (62) |Mn (ω) − Mn−1 (ω)| < K , ∀ω and ∀n.
= E NT (c)∧n = E MT2 (c)∧n − E AT (c)∧n .
Observe that if n < T (c) then
Now we prove that there exists a K s.t.
|MT (c)∧n | = |Mn | < c.
(61) |MT (c)∧n | ≤ K + c.
On the other hand, if n ≥ T (c) then by the definition of
Namely, by assumption the increments of M are T (c) and by (62)
uniformly bounded.
Convergence of Mn vs. A∞ < ∞ cont. Cesáro’s Lemma

Given two sequences of real numbers {bn }∞ n=1 and
Proof of Theorem 5.7 part (b) cont. {vn }∞
n=1 satisfying the following conditions:
|MT (c)∧n | = |MT (c) | ≤ K + c 0 < b1 , for all n : bn < bn+1 , lim b = ∞.
n→∞ n
This and (60) imply that
and
è é lim v = v∞ ∈ R.
(63) E AT (c)∧n ≤ (c + K )2 , ∀n. n→∞ n
Then
However, by monotone convergence Theorem we get
1 q n
that (59) and (63) contradict to each other. (64) bn (bk − bk−1 ) · vk → v∞ .
k=1
Kronecker’s Lemma Strong law under variance cond.
Let
Theorem 5.9
{bn }∞
n=1 be defined as in Cesáro’s Lemma,
∞
{xn }n=1 be an arbitrary sequence of real numbers Let X1 , X2 , . . . be independent r.v. s.t.
sn := x1 + · · · + xn . ∞
Ø Var(Xn )
(66) E [Xn ] = 0, and < ∞.
Then n=1 n2
A B
Then
∞ x 3 4
q sn
(65) n
converges =⇒ →0 . X1 +···+Xn
n=1 bn bn (67) n→∞
lim n =0 a.s..
Strong law under variance cond. cont. Strong law for L2 martingales
Theorem 5.10
Proof of Theorem 5.9 Let M be an L2 martingale with M0 = 0 and let
By Kronecker’s Lemma we only need to prove that A =< M > (defined in (48)). Then
∞
Ø Xn
, converges a.s.. Mn (ω)
(68) A∞ (ω) = ∞ =⇒ n→∞
lim = 0.
n=1 n An (ω)
We get this if we apply part (a) of Theorem 4.1 for the
r.v. Xn /n since their mean is zero and
Proof
First we claim that
Var(Xn /n) = Var(Xn )/n2 .
n
Ø Mk − Mk−1
(69) Wn := is a martingale.
k=1 1 + Ak
Strong law for L2 martingales cont. Strong law for L2 martingales cont.
Proof cont. î ï Proof cont.
This is so because for Cn := (1 + An )−1 the process Since < W > is a non-negative increasing process, in
C = (Cn ) is a bounded and previsible. Hence order to prove Fact 5.11, it is enough to show that
(70) Wn = C • M n
Ø
(72) < W >n = (< W >k − < W >k−1 )
k=1
is a martingale too.
by (50) Øn è é
Fact 5.11 = E (Wn − Wn−1 )2 |Fn−1
k=1
≤ 1.
(71) < W >∞ ≤ 1. In the one but last step we used (50).
Proof cont.
Proof cont. è é

(Mn − Mn−1 )2

To verify (72) we prove another claim about the yellow E (Wn − Wn−1 )2 |Fn−1 = E  |Fn−1 
(1 + An )2
part of the previous formula:  
Now we claim that  
  

 
è é    
  
(73) E (Wn − Wn−1 )2 |Fn−1 ≤ 1
1+An−1 − 1
1+An . = (1 + An )−2 · 
2  2
E  Mn |Fn−1  − Mn−1 

 üûúý ü ûú ý 
 Nn−1 +An−1 
ü Nn +Aûún
ý 
To verify this, note that: Nn−1 +An
= (1 + An )−2 · (Nn−1 + An − (Nn−1 + An−1 ))

= (1 + An )−2 · (An − An−1 )
Proof cont.
Proof cont. Observe that in (74) we have a telescopic sum on the
However, using that An is increasing we get right hand side. This immediately implies that (72) holds
which, in turn, implies that Fact 5.11 holds.From (71)
An − An−1 1 1 and from Theorem 5.7 part (a) we get that
(74) 2
≤ − .
(1 + An ) 1 + An−1 1 + An
lim Wn exists.
n→∞
This and the argument on the previous slide verifies that
(73) holds. We can use Kronecker’s Lemma if bk = 1 + Ak → ∞
that is A∞ = ∞. In this case by Kronecker’s Lemma we
get n→∞
lim Mn /An = 0.
Borel Cantelli Lemma Borel Cantelli Lemma cont.
Let En be events on the probability space (Ω, F, P). We Corollary 5.13 (Infinite Monkey theorem)
write
u t∞ Monkey typing random on a typewriter for infinitely time
E∞ := lim sup En := ∞ n=1 k=n En .
n→∞ will type the complete works of Shakespeare eventually.
Clearly, E∞ is the set of ω for which infinitely many of En
occur.
Lemma 5.12 (Borel-Cantelli Lemma)
∞
q
(a) If P (En ) < ∞ then P (E∞ ) = 0.
k=1
∞
q
(b) If {Ei }i are independent and P (En ) = ∞
k=1
then P(E∞ ) = 1. Figure: Picture is from Wikipedia. Proof is a homework
Levy’s extension of BC Lemma Proof of BC Lemma as a corollary of Thm

5.14
Theorem 5.14
Given a filtered space (Ω, F, {Fn } , P) and a sequence of
Proof of BC Lemma as a corollary of Thm 5.14.
events En ∈ Fn . Let
Using that E [ξk ] = P (Ek ) we can see that part (a) of
n n
Ø Ø Theorem 5.14 immediately implies the first part of BC
Zn := Ek , ξk := P (Ek |Fk−1 ) , Yn := ξk .
k=1 k=1 Lemma.
Then for almost all ω On the other hand if {En } are independent and
(a) If Y∞ (ω) < ∞ then Z∞ (ω) < ∞. Fn := σ(E1 , . . . En ) then ξk = P (Ek ). Hence the second
part of BC follows from part (b) of Theorem 5.14.
lim YZnn(ω)
(b) If Y∞ (ω) = ∞ then n→∞ (ω) = 1.
Levy’s extension of BC Lemma cont. Levy’s extension of BC Lemma cont.

Proof of Thm 5.14 Proof of Thm 5.14 cont.
Observe that Z = (Zn ) is submartingale with Z0 := 0. That is Z = M + Y be the Doob decomposition of Z .
Namely, E [Zn+1 |Fn ] = E [ En+1 |Fn ] + Zn ≥ Zn . Now Now we observe that
observe that n
Ø
Y is previsible since ξk ∈ Fk−1 . (75) An :=< M >n = ξk (1 − ξk ) ≤ Yn .
k=1
M := Z − Y is a martingale since by Ek ∈ Fk :
To see this we substitute Xk for Mk2 into (44) and then
E [Zn+1 − Yn+1 |Fn ] an immediate calculation yields (75).
Now we prove part (a): If Y∞ (ω) < ∞ then
= Zn + E [ En+1 |Fn ] − (Yn + E [ En+1 |Fn ])
A∞ (ω) < ∞. So, by Theorem 5.7 the limit n→∞ lim Mn (ω)
= Zn − Yn . exists. In this way Z∞ < ∞ since Y∞ < ∞ by
assumption and n→∞lim Mn (ω) exists.
Levy’s extension of BC Lemma cont. dimH

Proof of Thm 5.14 cont. 3 Martingal convergence
Now we prove part (b): 4 Sums of zero-mean independent r.v.
If Y∞ = ∞ and A∞ < ∞ then n→∞ lim Mn exists and 5 Doob decomposition
it is clear that Zn /Yn → 1. 6 Closing
If Y∞ = ∞ and A∞ = ∞ . It follows from 7 Preparation for the Uniform families
Theorem 5.10 that Mn /An → 0. So, by the last part 8 Uniform families
of (75) we have Mn /Yn → 0. From here and from
the definition of M we get Zn /Yn → 1.
9 LIL
Closure Closure cont.
This Section is built on Chapter 10.8.5 of Resnik’s book Example 6.1
[20]. Here we discuss the following problem:
Let Zn be a branching process and assume that the mean
Let X = (Xn ) be a positive martingale w.r.t. the of the offspring distribution µ < 1. Then the population
filtration Fn . Then we know by Corollary 3.6 that dies out so, Z∞ = 0. On the other hand Wn := Zn /µn is
Xn → X∞ a.s.. However, in general this does NOT a non-negative martingale so it converges to W∞ = 0.
mean that Xn = E [X∞ |Fn ]. However, Zn /µn Ó= 0 for all n.
In this short Section first we give a counter example
Definition 6.2
which verifies the last assertion. Then we state a useful
theorem (Theorem 6.3) related to this problem and an A martingale {Xn , Fn } is closed (or right-closed) if
important corollary of this theorem. The proof of there exists an L1 r.v. X∞ ∈ F∞ = σ(Fn , n ∈ N) s.t.
Theorem 6.3 is available in [20, Section 10.8.5]. We omit
Xn = E [X∞ |Fn ] for all n.
this proof here.
Closure cont. Closure cont.

Theorem 6.3 Corollary 6.4
Given a p ≥ 1 and a rv. X ∈ Lp . Also given a filtration
For p ≥ 1 the class of Lp convergent non-negative
Fn and we write F∞ := σ(Fn , n ∈ N). Then for
martingales is equal to the class:
(76) Xn := E [X |Fn ] and X∞ := E [X |F∞ ]
C := {(E [X |Fn ])∞ p
n=0 , X ∈ L , X ≥ 0} .
we have Xn is a closed martingale which converges a.s.
and in Lp : Proof of the Corollary
If (Xn ) ∈ C then by Theorem 6.3 Xn is Lp convergent.
(77) Xn = E [X∞ |F∞ ]
Conversely, let Xn ≥ 0 be an Lp convergent martingale.
If r ≥ n then E [Xr |Fn ] = Xn .
(78) Xn → X∞ a.s. and in Lp .
Closure cont. dimH

Proof of the Corollary cont. 3 Martingal convergence
Lp
By assumption that Xr → X∞ and E [•|Fn ] is continuous 4 Sums of zero-mean independent r.v.
in Lp metric (this follows from the Lp -non expansive
5 Doob decomposition
property (property (e)) of the conditional expectation,
(see #133 slide if File "Some basic Facts from 6 Closing
probability theory" ). So, as r → ∞ we have 7 Preparation for the Uniform families
Lp
8 Uniform families
Xn = E [Xr |Fn ] → E [X∞ |Fn ] .
9 LIL
So Xn = E [X∞ |Fn ] .
Recall Fatou Lemma Reverse Fatou Lemma
From now we follow again Williams book [23, Chapter Lemma 7.2 (reverse Fatou Lemma)
14] about the uniformly integrable martingales. First we
Let {Xn } be sequence of r.v. on (Ω, F, P) s.t. ∃ a
recall some very well known lemmas:
non-negative r.v. Y ∈ F such that Xn ≤ Y and
Lemma 7.1 (Fatou Lemma) E [Y ] < ∞. Then
Let X1 , X2 , . . . be non-negative r.v. on (Ω, F, P). Then C D
5 6
E lim sup Xn ≥ lim sup E [Xn ] .
n→∞ n→∞
E lim inf Xn ≤ lim
n→∞
inf E [Xn ] .
n→∞
Proof.
May be both sides are infinite. Apply Fatou Lemma for Y − Xn .
Theorem BDD Theorem BDD cont
We will need the following version of Bounded
Convergence Theorem (abbreviated as BDD): Proof of Thm BDD
We claim that
Theorem 7.3 (Theorem BDD)
(79) |X | ≤ K a.s..
Let Xn , X be rv. on (Ω, F, P). We assume that
P
(a) Xn → X (Xn tends to X is Probability) AND Namely, for any k ∈ N
1 2 1 2
(b) ∃K s.t. ∀ω ∈ Ω, |Xn (ω)| < K . (That is the P |X | > K + k −1 ≤ P |X − Xn | > k −1 , ∀n,
process X = (Xn ) is bounded.)
Then Using assumption (a) the right hand side tends to zero
L1 for every k. This yields that (79) holds.
Xn → X , that is E [|Xn − X |] → 0.
Theorem BDD cont Theorem BDD cont
Proof of Thm BDD cont.

Proof of Thm BDD cont. C D
For a given ε choose n0 s.t. for all n > n0 we have ε
E [|Xn − X |] =E |Xn − X |; |Xn − X | >
A B C
3 D
1 ε ε
P |Xn − X | > ε < , for n ≥ n0 . + E |Xn − X |; |Xn − X | ≤
3 3K A B
3
ε ε
So, whenever n ≥ n0 we have =2K P |Xn − X | > +
3 3
≤ ε .
a.s. convergence vs conv. in prob. Absolute continuity

The proof of the following theorem is available in [14, Theorem 7.5
p.137].
Let X ∈ L1 (Ω, F, P). Then ∀ε > 0 ∃δ > 0 s.t.
Theorem 7.4
Given X , Xn be rv on (Ω, F, P). Let (81) If F ∈ F and P (F ) < δ =⇒ E [|X |; F ] < ε.
Yn := sup |Xk − X |. Proof

k≥n
We argue by contradiction. Assume that ∃ε0 > 0 and
Then {Fn } s.t. Fn ∈ F s.t.
a.s. P
P (Fn ) < 2−n and E [|X |; Fn ] > ε0 .
(80) Xn −→ X ⇐⇒ Yn → 0.
Absolute continuity cont. Absolute continuity cont.

Proof cont. Corollary 7.6
Let
H := lim sup Fn . Let X ∈ L1 (Ω, F, P) and ε > 0. Then ∃K ∈ [0, ∞) such
n→∞
that
That is ✶H = lim sup ✶Fn . Then by Borel-Cantelli Lemma,
n→∞
(82) E [|X |; |X | > K ] < ε.
P (H) = 0.
Proof.
On the other hand by the reversed Fatou Lemma Choose δ for ε as in Theorem 7.5. Choose a K s.t.
applied for the indicator functions of Fn we have: |]
P (|X | > K ) ≤ E[|XK < δ. Then by the previous theorem
E [|X |; |X | > K ] < ε.
E [|X |; H] > ε0 .
Homework # 31 is an extension of this Corollary.
The last two displayedMarkov
Károly Simon (TU Budapest)
formulas contradict.
Processes & Martingales C File 111 / 188 Károly Simon (TU Budapest) Markov Processes & Martingales C File 112 / 188
dimH UI martingales, definition
2 Non-negative superharmonic functions of martingale Definition 8.1
Let C be a class of random variables on the probability
4 Sums of zero-mean independent r.v. space (Ω, F, P). We say that C is uniformly integrable
5 Doob decomposition ( UI ) if
6 Closing
(83) ∀ε > 0, ∃K s.t. ∀X ∈ C, E [|X |; |X | > K ] < ε.
8 Uniform families Example 8.2 (Example of non-UI martingale)
9 LIL Let the probability space be ([0, 1], R[0, 1], L1 ) and
10 Doob’s Lp inequality Xn = n · ✶[0,n−1 ] . ∀K if n > K then E [Xn ; |Xn | > K ] = 1
11 Kakutani’s Theorem on Product Martingales still Xn → 0 a.s. (but Xn Ó→ 0 in L1 ).
UI families cont. UI families cont.
Lemma 8.4
Remark 8.3 Let C is a class of random variables of (Ω, F, P). Then
We have seen above that an L1 -bounded family is NOT either of the following two conditions imply that C is UI.
necessarily UI but it is obvious that every UI family is (a) If ∃p > 1 and A ∈ R such that E [|X |p ] < A
L1 -bounded. for all X ∈ C. (Lp bounded for some p > 1.)
How to check if a family is UI? Two simple ways are as (b) ∃Y ∈ L1 (Ω, F, P), s.t. ∀X ∈ C we have
follows: |X (ω)| ≤ Y (ω). (C is dominated by an
integrable (non-negative) r.v..)
The proofs are easy and left as homeworks.

Theorem 8.5 Proof of the if part (a)
1
Let X , Xn ∈ L (Ω, F, P) for every n. The the following Assume that (a) and (b) hold. For a K ∈ [0, ∞) we
two assertions are equivalent: define
L1 
1 Xn → X (that is Xn → X in L1 that is 

 K , if x > K ;
E [|Xn − X |] → 0). ϕK (x ) :=  x, if −K ≤ x ≤ K ;


2 Both of the following conditions hold: −K , if x < −K .
P
(a) Xn → X (that is ∀ε > 0:
Fix an ε > 0. Now we apply the fact that Xn is UI and
lim P (|Xn − X | > ε) = 0) AND
n→∞ Corollary 7.6 in this order to conclude that ∃K s.t.
(b) The sequence {Xn }∞ n=1 is UI.
ε ε
The complete proof is available: [23, p. 131]. Here we (84) E [|ϕK (Xn ) − Xn |] < ; E [|ϕK (X ) − X |] < .
prove only the "if" part. 3 3
Proof of the if part (a) cont. Proof of the if part (a) cont.
We claim that Hence for a fixed ε > 0 we can find n0 s.t. for n ≥ n0 we
L1 have
(85) ϕK (Xn ) −→ ϕK (X ). ε
E [|ϕK (Xn ) − ϕK (X )|] < .
3
Using that |ϕK (x ) − ϕK (y )| ≤ |x − y | and using that Putting together this and the two inequalities in (84) we
P P
Xn → X , we obtain that ϕK (Xn ) → ϕK (X ). Observe obtain that for n ≥ n0 :
that the process {ϕK (Xn )} is bounded (by K ), so we can
apply Theorem BDD which yields that (85) holds. E [|Xn − X |] < ε.
UI martingales definition Conditional expectation vs. UI
Theorem 8.7
Definition 8.6 (UI martingale)
Let X ∈ L1 . Then the following family is UI:
M = (Mn ) is a UI martingale on a filtered probability
space (Ω, F, {Fn , P}) if (86) C := {E [X |G] : G is a sub-algebra of F} .
M is a martingale,
{Mn }∞n=0 is a UI family.
More precisely, C consists of the versions of E [X |G], for
some G ⊂ F sub-σ-algebra.
Conditional expectation vs. UI cont. Conditional expectation vs. UI cont.

Proof Proof cont.
Fix an ε > 0. Using Theorem 7.5 there exist δ > 0 s.t. So, E [|Y |] ≤ E [|X |] and
(87) ∀F ∈ F, P (F ) < δ =⇒ E [|X |; F ] < ε. K P (|Y | > K ) ≤ E [|Y |] ≤ E [|X |] .
Choose a K s.t. From here and (88) we get
(88) K −1 E [|X |] < δ. P (|Y | > K ) < δ.
Let G be a sub-σ-algebra of F and let Y be a version of Clearly, {|Y | > K } ∈ G, so by (89) we get
E [X |G]. Then by Jensen inequality:
E [|Y |; |Y | ≥ K ] = E [|E [X |G] |; |Y | ≥ K ]
(89) |Y | ≤ E [|X ||G] , a.s. ≤ E [|X |; |Y | ≥ K ] < ε.
UI martingales cont. UI martingales cont.

Theorem 8.8
Let M be a UI martingale. Then Proof cont.
We know that almost sure convergence implies
(90) lim Mn exists a.s. and in L1 .
M∞ := n→∞
convergence in probability. So, both of the conditions of
the second point of Theorem 8.5 are satisfied. Hence
Further, L1
Mn −→ M∞ . This completes the proof of (90).
(91) Mn = E [M∞ |Fn ] . To prove (91), we only need to verify:
Proof (92) E [M∞ ; F ] = E [Mn ; F ] , ∀F ∈ Fn .

It follows from Theorem 3.5 that M∞ := n→∞
lim Mn a.s.
exists since M is an L1 -bounded martingale.
UI martingales cont. Lévy’s Upward Theorem
Theorem 8.9 (Lévy’s Upward Theorem)

Proof cont.
Let r > n. Then E [Mr |Fn ] = Mn . Hence, for ∀F ∈ Fn Let X ∈ L1 and on the filtered probability space
we have (Ω, F, {Fn } , P) and let
(93) E [Mn ; F ] = E [Mr ; F ] → E [M∞ ; F ] Mn := E [X |Fn ] and Y := E [X |F∞ ] ,
since Mr → M∞ is L1 . This implies that (92) holds, where F∞ := σ(Fn , n ∈ N). Then
which completes the proof of (91). (a) M = (Mn ) is a UI martingale and
(b) Mn → Y a.s. and in L1 .
Lévy’s Upward Theorem cont. Lévy’s Upward Theorem cont.
So, M is UI. Then by Theorem 8.8 ∃M∞ s.t.
Note that Theorem 8.9 is weaker than Theorem 6.3
(which we did not prove), but we prove Theorem 8.9. (94) Mn −→ M∞ a.s. and in L1 .
Proof of Thm 8.9 part (a) So, we only need to prove that
M is a martingale since by the tower property: (95) Y = M∞ .

 
 
Proof of Thm 8.9 part (b)
E [Mn+1 |Fn ] = E E [X |Fn+1 ] |Fn  = Mn . WLOG (acronym for without loss of generality) we can
ü ûú ý
Mn+1 (and do) assume that X ≥ 0 . Let ν1 , ν2 be measures on
the measurable space (Ω, F∞ ) defined by:
M is UI by Theorem 8.7.
ν1 (F ) := E [Y ; F ] and ν2 (F ) := E [M∞ ; F ] , F ∈ F∞ .
Lévy’s Upward Theorem cont. Lévy’s Upward Theorem cont.

Proof of Thm 8.9 part (b) cont.
Proof of Thm 8.9 part (b) cont. Both Y and M∞ are F∞ measurable. So
First observe that by tower property:
F := {ω : Y (ω) > M∞ (ω)} ∈ F∞ .
E [Y |Fn ] = E [X |Fn ] = E [M∞ ; F ] , ∀F ∈ Fn .
Hence E [Y ; F ] = ν1 (F ) = ν2 (F ) = E [M∞ ; F ] . That is
Namely, the first equality follows from the tower
property, the second one was checked in (92). E [Y − M∞ ; Y − M∞ ] = 0 .
So, ν1 , ν2 coincide on the π-system ∪Fn (which is
actually an algebra). So, ν1 is equal to ν2 also on F∞ . That is P (Y > M∞ ) = 0. Similarly we can see that
P (M∞ > Y ) = 0 and this completes the proof of (95).

Kolomorov’s 0 − 1 law Martingale proof for Kolomorov 0 − 1 law
Proof
Theorem 8.10 (Kolmogorov’s 0 − 1 law) Let
Fn := σ(X1 , . . . , Xn ) ,
Let X1 , X2 , . . . be a sequence of independent rv.
Fix an F ∈ T , and put Y := ✶F . Using the fact that
∞
Ü Y ∈ F∞ in the first equality and Levy’s upward Thm. in
Tn := σ(Xn+1 , Xn+2 , . . . ), T := Tn . the second one we get:
n=1
We say that T is the tail σ-algebra. Then (97) Y = E [Y |F∞ ] = n→∞

lim E [Y |Fn ] , a.s.
(96) P (F ) = 0 or 1, ∀F ∈ T . On the other hand, ∀n:
(98) Y ∈ Tn =⇒ Y is independent of Fn .
Martingale proof for Kolomorov 0 − 1 law cont. Levy’s Downward Theorem
Proof cont. Theorem 8.11 (Levy’s Downward Theorem)

Let {G−n : n ∈ N} be a sequence of sub-σ-algebras of F
(99) E [Y |Fn ] = E [Y ] = P (F ) , a.s. on the probability space (Ω, F, P) satisfying:
∞
Ü
Putting together (97) and (99) we obtain that G−∞ := G−k · · · ⊂ G−(m+1) ⊂ G−m · · · ⊂ G−1 .
k=1
Y = P (F ) .
Let X ∈ L1 (Ω, F, P) and let M−n := E [X |G−n ] . Then
Since by definition Y is either zero or one we obtain that
lim M−n exists a.s. and in L1 ,
(a) M−∞ := n→∞
P (F ) is also either zero or one holds for all F ∈ T∞ .
(b) M−∞ = E [X |G−∞ ] a.s.
Levy’s Downward Theorem cont. Levy’s Downward Theorem cont.
Proof Proof cont.
Observe that Recall from the Figure on slide # 103 of the File "Some
basic Facts in Probability Theory" that
(100) (Mk , Gk : −N ≤ k ≤ −1) . (102)
a.s. convergence =⇒ convergence in probability.
is a FINITE martingale sequence. So, we can apply
Doob’s Upcrossing Lemma (slide # 25) for it. Clearly, We apply this for M−n and Theorem 8.7 (in this order)
E [|M−k |] ≤ E [|X |] so M−n is L1 bounded. Therefore we to get
can apply the steps of both Corollary 3.3 and Doob’s P
(a) M−n −→ M−∞
Forward Convergence Theorem (slide # 29) to obtain
that ∃M−∞ s.t. (b) (M−n ) is UI.
L1
Using Theorem 8.5 we conclude that M−n −→ M−∞ .
(101) M−∞ = n→∞
lim M−n , a.s. This completes the proof of part (a).
Levy’s Downward Theorem cont. Levy’s Downward Theorem cont.

Proof part (b) Proof of part (b) cont.
In particular this implies that However, this follows from (103) if we can check that
(103) lim E [M−r ; G] = E [M−∞ ; G] , ∀G ∈ G∞ . (106) E [X ; G] = E [M−r ; G] , ∀G ∈ G−∞ ⊂ G−r .

r →∞
In order to verify that This is immediate from the definition of conditional

expectation since by definition M−r = E [X |Gr ]. This
(104) M−∞ = E [X |G−∞ ] completes the proof of (104).
it is enough to show that ∀G ∈ G−∞ : Now we give a martingale proof for SLLN (Strong Law
of Large Numbers). At the end of the course Probability
(105) E [M−∞ ; G] = E [X ; G] = E [E [X |G−∞ ] ; G] . 1 you have seen a proof for the special case when the
4-th moments existed.
Martingale proof for SSLN Martingale proof for SSLN cont.

Below we mention an important Corollary. It follows
Theorem 8.12 (SLLN) easily from SLLN. This proof is assigned as homework #
Let X1 , X2 , . . . be iid. rv. with E [|X1 |] < ∞. Let 35.
Corollary 8.13
µ := E [Xk ] , and Sn := X1 + · · · + Xn .
Let X1 , X2 , . . . be iid. rv. with E [X + ] = ∞ and
Then E [X − ] < ∞. (Recall X = X + − X − and X + , X − ≥ 0.)
We can use SLLN to prove that
(107) lim Sn
n→∞ n
= µ , a.s. and in L1 .
Sn /n → ∞ , a.s.
Also in the course Probability I, only the almost sure
convergence was stated. where Sn := X1 + · · · + Xn .
Martingale proof for SLLN cont. Martingale proof for SLLN cont.
Proof of Thm 8.12 Proof of Thm 8.12 cont.
Let
∞
Sn
Ü (109) M−n := E [X1 |G−n ] = , a.s.
G−n := σ (Sn , Sn+1 , . . . ) , G−∞ := G−n . n
n=1
We apply Levy’s Downward Theorem for (M−n ) to obtain
Clearly, that the following limits exists:
E [X1 |G−n ] = E [X1 |Sn ] .
However, it is easy to see (assigned as homework # 36) (110) lim Snn .
L := n→∞
that
Observe that for each k
(108) E [X1 |Sn ] = Sn /n. Xk+1 + · · · + Xn
(111) L = n→∞
lim .
So we obtain that n
Martingale proof for SLLN cont. Doob’s submartingale inequality
Theorem 8.14 (Doob’s submartingale inequality)

Proof of Thm 8.12 cont. Let Z be a non-negative submartingale and L > 0.
That is L ∈ TK := σ(XK +1 , Xk+2 , . . . ). So, L ∈ T∞ . Then
Using Kolmogov’s 0 − 1 law this means that the rv L is (112)
1 2 è é
almost sure constant. Then it cannot be anything but L · P supk≤n Zk ≥ L ≤ E Zn ; supk≤n Zk ≤ E [Zn ] .
C D
Sn
L = E [L] = n→∞
lim E
n
= µ. Proof
F0 := {Z0 ≥ L},
Fk := {Z0 < L}∩{Z1 < L}∩· · ·∩{Zk−1 < L}∩{Zk ≥ L} .
Doob’s submartingale inequality Recall a fact learned earlier

Proof cont.
That is for T := min {k : Zk ≥ k} we have Remark 8.15
Fk = {T = k}. Recall that for a
Clearly, Fk ∈ Fk and Zk ≥ L on Fk . That is for k ≤ n: convex function ϕ and a
(113) martingale M = (Mn )
E [Zn ; Fk ] = E [E [Zn |Fk ] ; Fk ] ≥ E [Zk ; Fk ] ≥ LP (Fk ) .
it follows from conditional Jensen’s inequality ( see slide
To complete the proof it is enough to sum up for k ≤ n # 133 of File "Some basic facts from probability theory")
since Zn ≥ 0 and that
 
  h h h E [|ϕ(Mn )|] < ∞ =⇒ ϕ(Mn ) is a submartingale.

sup Zk ≥ L = F0 F1 ··· Fn .
k≤n
Kolmogorov’s inequality Kolmogorov’s inequality cont.

Lemma 8.16 (Kolmogorov’s inequality)
Given a sequence of rv. (Xn , n ≥ 1). We assume that
(Xn , n ≥ 1) are independent, Proof.
E [Xi ] = 0, Fn := σ {X1 , . . . , Xn }. Then S = (Sn ) is a martingale.
Xi ∈ L2 . So, by Remark 8.15, S 2 is a non-negative submartingale.
Hence, we can apply the Submartingale inequality for
We define
qn S 2.
Sn := X1 + · · · + Xn , Vn := Var(Sn ) = σk2 . Then for
k=1
any L > 0 we have
1 2
(114) L2 · P supk≤n |Sk | ≥ L ≤ Vn .
dimH An estimate on normal distribution

2 Non-negative superharmonic functions of martingale Fact 9.1
3 Martingal convergence Let X ∼ N (0, 1) and Φ and ϕ are the CDF and the
4 Sums of zero-mean independent r.v. density of X respectively. That is
5 Doob decomposition 1 − x2
6 Closing ϕ(x ) = e 2,
2π
7 Preparation for the Uniform families Ú∞
8 Uniform families P (X > x ) = 1 − Φ(x ) = ϕ(y )dy .
x
9 LIL
Then for x > 0 we have
(a) P (X > x ) ≤ x −1 ϕ(x )
(b) P (X > x ) ≥ (x + x −1 )−1 ϕ(x )
An estimate on normal distribution cont. An estimate on normal distribution cont.
Proof
It is easy to check that Proof cont
(115)  ′
To verify (b), we use the second part of (115):
ϕ(x ) 
ϕ′ (x ) = −x ϕ(x ) and  = −(1 + x −2 )ϕ(x ). Ú∞
x
x −1 ϕ(x ) = (1 + y −2 )ϕ(y )dy
x
Using the first equality we get that for x > 0: ∞
s
≤ (1 + x −2 ) ϕ(y )dy .
Ú∞ Ú∞ x
ϕ(x ) = y ϕ(y )dy ≥ x ϕ(y )dy .
x x which yields (b).
Which yields (a).
Kolmogov’s Law of Iterated Logarithm LIL

Theorem 9.2 (Kolmogov’s Law of Iterated
logarithm (LIL))
Given a sequence of rv X = (Xn , n ≥ 1) satisfying:
X1 , X2 , . . . are iid ,
E [Xi ] = 0 ,
Var(Xi ) = 1 .
As usual, we write Sn := X1 + · · · + Xn . Then almost
surely,
(116) Figure: A plot of the average of n Bernoulli trials (each taking a
Sn Sn value of +/- 1). Plot of (red), its variance given by CLT (blue) and
lim sup √ = 1, lim inf √ = −1.
n→∞ 2n log log n n→∞ 2n log log n its bound given by LIL (green). Figure is from Wikipedia.
A crude heuristics Proof of LIL in a special case

By CLT: We give the proof in the special case when
S
√n −→
D
N (0, 1).
n (117)
D
Xi = N (0, 1) .
Sn
If √
n
was a N ∼ N (0, 1) then
Ø 3 ñ 4
P N > (2 + ε) log log n < ∞ Proof of LIL assuming (117)
n
Let ñ
q 1 ñ 2
and P N > (2 − ε) log log n = ∞ so the first one h(n) := 2n log log n, for n ≥ 3.
n
happens finitely many times and the second one happens First we verify that for every c > 0 and n ≥ 3 we have:
infinitely many times by BC Lemmas. (This is not a 1 2 1 2
proof just a heuristics. The second part of BC Lemma (118) P supk≤n Sk ≥ c ≤ e− 2 c /n
.
(Lemma 5.12) holds only if the vents are independent.)
Proof of LIL in a special case cont. Proof of LIL in a special case cont.
Proof of LIL assuming (117) cont.
Namely, we have learned that the moment generating Proof of LIL assuming (117) (upper bound)
function of the standard normal distribution Xi is We choose a K > 1 (actually K will be close to 1). Let
è é 1 2 è é 1 2
E eθXi = e 2 θ . Hence, E eθSn = e 2 θ n < ∞ . For every
cn := Kh(K n−1 ).
fixed θ ∈ R the function x → eθx is convex. So,
eθSn is a submartingale . So, we can use Submartingale Then
inequality for θ > 0:  
 
1 2 è é
1 2
P supk≤n Sk ≥ c = P sup e ≥ e  ≤ e−θc E eθSn
θSk θc
P  sup Sk ≥ cn  ≤ exp −cn2 /2K n
k≤n
k≤K n
1 2
−θc 2θ n
= e e . = (n − 1)−K (log K )−K .
For c > 0 choosing θ = c/n we obtain that (118) holds.

b
Proof of LIL assuming (117) (upper bound) cont.

Proof of LIL assuming (117) (lower bound)
From this and from the first BC Lemma: almost surely
for all n ≥ n0 (ω) and for k ∈ [K n−1 , K n ] we have Fix N > 1, N ∈ N and ε ∈ (0, 1). Define the events:
î 1 2ï
(119) Sk ≤ sup Sk ≤ cn = Kh(K n−1 ) ≤ Kh(k), a.s. Fn := SN n+1 − SN n > (1 − ε)h N n+1 − N n .
k≤K n
So, for K > 1: lim sup h(k)−1 Sk ≤ K a.s. By letting Then by Fact 9.1, for
k→∞
K ↓ 1 we get that ñ
y := (1 − ε) 2 log log (N n+1 − N n )
lim sup h(k)−1 Sk ≤ 1 a.s.
Proof of LIL assuming (117) (lower bound) cont. Proof of LIL assuming (117) (lower bound) cont.
we have The events Fn are independent so, we can apply BC
1 Lemma (Lemma 5.12). This yields that almost surely
P (Fn ) = 1 − Φ(y ) ≥ √ exp(−y 2 /2). infinitely many of Fn occurs. That is for infinitely many
2π · (y + y −1 ) n:
By ignoring the logarithmic terms we see that 1 2 1 2
(121) S N n+1 > (1 − ε)h N n+1 − N n + S(N n ) .
2
P (Fn ) ∼ (n log N)−(1−ε)
Observe that (119) is also valid if we substitute S with
This yields that −S. For K = 2 this yields
(120)
q
P (Fn ) = ∞ . S(N n ) > −2h(N n ).
Proof of LIL in a special case cont. dimH

Proof of LIL assuming (117) (lower bound) cont. 2 Non-negative superharmonic functions of martingale
So, by the last twi displayed formulas: 3 Martingal convergence
1 2 1 2 4 Sums of zero-mean independent r.v.
S N n+1 > (1 − ε)h N n+1 − N n − 2h(N n ) 5 Doob decomposition
From here, simple algebraic manipulations yield: 6 Closing
lim sup h(k)−1 Sk ≥ lim sup h(N n+1 )−1 S(N n+1 ) 8 Uniform families
k→∞ n→∞
√ 9 LIL
≥ (1 − ε) 1 − N −1 − √2 .
N 10 Doob’s Lp inequality
This completes the proof since this holds for all N. 11 Kakutani’s Theorem on Product Martingales
An Auxiliary Lemma Proof of Lemma 10.1

Lemma 10.1
Given X , Y ≥ 0 rv. staisfying: Proof
(122) cP (X ≥ c) ≤ E [Y ; X ≥ c] , ∀c > 0. Ú∞
L := pc p−1 P (X ≥ c) dc
For every p > 1 we define the conjugate q of p by c=0
(122) Ú∞
1 1 ≤ pc p−2 E [Y ; X ≥ c] dc =: R
(123) + = 1.
p q c=0
Then we have By Fubini Thm. we get
(124) ëX ëp ≤ q · ëY ëp .
Proof of Lemma 10.1 Proof of Lemma 10.1
Proof cont. Proof cont.
  (127) è é
Ú∞ Ú
(125) L = 
✶  p−1 E [X p ] = L ≤ R = E qX p−1 Y ≤ qëY ëp ëX p−1 ëq .
 {X ≥c} (ω)dP(ω) pc dc
c=0 Ω
  WLOG we may assume that ëY ëp < ∞ .
Ú XÚ(ω)

=  pc dc 
p−1
 dP(ω) = E [X p ] . If additionally, ëX ëp < ∞ then by (p − 1)q = p we get
Ω c=0
Similarly we get ëX p−1 ëq = E [X p ]1/q .

è é
(126) R = E q · X p−1 · Y . Using this and (127) we get
By Hölder inequality: (E [X p ])1−1/q = ëX ëp ≤ qëY ëp .

Proof of Lemma 10.1 Doob’s Lp inequality

Theorem (Doob’s Lp inequality)
Fix a p > 1 and let q its conjugate, defined by
Proof cont. 1 1
p + q = 1. Given the process Z = (Zn ) satisfying:
For general X we do the same for X ∧ n to conclude that
non-negative,
ëX ∧ nëp ≤ qëY ëp , ∀n. submartingale,
sup ëZn ëp < ∞. (Z is Lp -bounded.)
n
Then we apply Monoton Conv. Thm. which concludes
We use the standard notation:
the proof of the Lemma.
Z ∗ := sup Zk .
k
Then Z ∗ ∈ Lp .
Doob’s Lp inequality cont. Doob’s Lp inequality cont.

Theorem (Doob’s Lp inequality cont.)
More precisely, Theorem (Doob’s Lp inequality cont.)
(128) ëZ ∗ ëp ≤ q · sup ëZr ëp . In the special case when Zn = |Mn | for a martingale
r M = (Mn ) which is bounded in Lp then
So, the non-negative submartingale Z is Lp -dominated (131) lim Mn exists a.s. and in Lp .
M∞ := n→∞
(by Z ∗ ∈ Lp ). Further,
and
(129) lim Zn exists a.s. and in Lp and
Z∞ := n→∞
(132) Z∞ = |M∞ | a.s.
(130) ëZ∞ ëp = sup ëZr ëp =↑ rlim

→∞
ëZr ëp .
r

Proof of Doob’s Lp inequality Proof of Doob’s Lp inequality cont.
First we prove that (128) holds. Let to obtain that
Zn∗ := sup Zk . ëZn∗ ëp ≤ qëZn ëp ≤ q sup ëZr ëp .
k≤n r
Doob’s submartingale inequality yields: That is Z ∗ is Lp -bounded. Observe that −Z is an

Lp -bounded supermartingale. Hence, −Z is also an
(133) cP (Zn∗ ≥ c) ≤ E [Zn ; {Zn∗ ≥ c}] L1 -bounded supermartingale. Doob’s Forward
Convergence Thm. (Theorem 3.5) yields that
Now we apply Lemma 10.1 with
(134) Z∞ := n→∞
lim Zn exists a.s.
X = Zn∗ and Y = Zn .
Proof of Doob’s Lp inequality cont. Proof of Doob’s Lp inequality cont.
On the other hand, Namely, consider the convex function
|Zn − Z∞ |p ≤ (2Z ∗ )p ∈ Lp . ϕ(x ) := |x |p .
Using Dominated Conv. Thm. we get By the Conditional Jensen’s inequality (File Some Basic
Facts from Probability Theory, ...133 ) we have:
Lp
Zn −→ Z∞ .
ëZr +1 ëpp = E [ϕ(Zr +1 )] = E [E [ϕ(Zr +1 )|Fr ]]
Now we verify that cond.Jensen
≥ E [ϕ (E [Zr +1 |Fr ])]
(135) {ëZr ëp }r is monotone increasing. ≥ E [ϕ (Zr )] = ëZr ëpp .
Doob’s Lp inequality cont. dimH

4 Sums of zero-mean independent r.v.
Proof of Doob’s Lp inequality cont. 5 Doob decomposition
To verify (131) and (132) is a homework.
6 Closing
8 Uniform families
9 LIL
Product Martingales Product Martingales cont.

Kakutani’s Theorem Kakutani’s Theorem cont.
Given X = (Xn ) satisfying: Then the following conditions are equivalent
X1 , X2 , . . . are independent (a) E [M∞ ] = 1,
L1
Xk are non-negative for all k. (b) Mn −→ M∞ ,
(c) M is UI.
E [Xk ] = 1 for all k. ∞
r
n
r (d) an > 0,
Then for Mn := Xi is a non-negative martingale (see n=1
i=1 ∞
q
File E of the course Stochastic processes). Hence (e) (1 − an ) < ∞.
n=1
M∞ := n→∞lim Mn exists a.s. We introduce: If any of these conditions does not hold then
5ñ 6
(136) an := E Xn . (137) P (M∞ = 0) = 1.
Product Martingales cont. Product Martingales cont.

Proof of Kakutani’s Theorem cont. Proof of Kakutani’s Theorem
First observe that by Jensen’s inequality: By independence and (138):
5ñ 6 ñ è é 1 1
(138) 0 ≤ = an = E Xn ≤ E [Xn ] = 1 . (140) E Nn2 ≤ n
r
≤ ∞
r
< ∞.
ak2 ak2
k=1 k=1
Assume that (d) holds. Let
√ That is Nn is L2 bounded. So, we can apply Doob’s L2
n Xk
(139) Nn :=
Ù
. inequality (Theorem 172 with p = 2) for Nn2 in the
k=1 ak second step below:
(141)
Clearly, Nn is a martingale (product of independent C D C D
è é
nonnegative rv having mean 1). E sup |Mn | ≤ E sup Nn2 ≤ 4 sup E Nn2 < ∞ .
n n n
Product Martingales cont. [1] BalÃązs MÃąrton, TÃşth BÃąlint
ValÃşszÃŋnÅśsÃľgszÃąmÃŋtÃąs 1. jegyzet matematikusoknak Ãľs ﬁzikusoknak
BÃąlÃązs MÃąrton Honlapja, 2012. Az internettes vÃąltozatÃľrt kattintson ide.
[2] P. Billingsley
Convergence of probability measures
Proof of Kakutani’s Theorem [3]
Wiley, 1968
B. Driver
That is Analysis tools with examples
Lecturenotes, 2012. Click here.
M ∗ := sup |Mn | ∈ L1 . [4] R. Durrett

Essentials of Stochastic Processes, Second edition
n Springer, 2012. Click here
[5] R. Durrett
That is M is a UI martingale which implies that Probability: Theory with examples, 4th edition
Cambridge University Press, 2010.
properties (a), (b), (c) hold. [6] R. Durrett
Probability: Theory and Examples
Click here
On the other hand, [7] D.H. Fremlin

Measure Theory Volume I
∞
r Click here
Assume that an = 0. Then N defined in (139) is a [8] D.H. Fremlin
n=1 Measure Theory Volume II
nonnegative martingale, so its limit N∞ exists a.s. This Click here
r [9] O. van Gaans

implies that M∞ = 0 a.s. since an = 0. Probability measures on metric spaces
Click here
[10] I.I. Gihman, A.V. Szkorohod

BevezetÃľs a sztochasztikus folyamatok elmÃľletÃľbe
MÅśszaki KÃűnyvkiadÃş1975, Budapest, 1985
[11] S. Karlin, H.M. Taylor
A ﬁrst course in stochastic processes
Academic Press, New York, 1975
Sztochasztikus Folyamatok [20] S.I. Resnik
Gondolat, Budapest, 1985 A Probability Path
Birkhäuser, 2005
A second course in stochastic processes [21] S. Ross
, Academic Press, 1981 A First Course in Probability, 6th ed.
Prentice Hall, 2002
[14] A.F. Karr
Probability [22] TÃşth BÃąlint Sztochasztikus folyamatok jegyzet
Springer-Verlag, 1993 TÃşth BÃąlint JegyzetÃľrt kattintson ide
[15] G. Lawler [23] D. Williams

Intoduction to Stochastic Processes Probability with Martingales
Chapmann & Hall 1995. Cambridge 2005
[16] D.A. Levin, Y. Peres,E.L. Wilmer

Markov chains and mixing times
American Mathematical Society, 2009.
[17] Major PÃľter
Folytonos idejÅś Markov lÃąncok http://www.renyi.hu/~major/debrecen/debrecen2008a/markov3.html
[18] P. Mattila Geometry of sets and measure in Euclidean spaces. Cambridge, 1995.
[19] RÃľnyi AlfrÃľd

ValÃşszÃŋnÅśsÃľgszÃąmÃŋtÃąs, (negyedik kiadÃąs)
TankÃűnyvkiadÃş Budapest, 1981.

File C 8

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

File C 8

Uploaded by

Copyright:

Available Formats

dimH

1 Hitting time for SRW

11 Kakutani’s Theorem on Product Martingales

Hitting time for SRW Hitting time for SRW cont.

Hitting time for SSRW cont. Explanation of (9)

An easier proof for (8) cont. dimH

11 Kakutani’s Theorem on Product Martingales

Superharmonic functions Superharmonic functions cont.

Superharmonic functions cont. Superharmonic functions cont.

11 Kakutani’s Theorem on Product Martingales

Upcrossing cont. Uncrossing cont., the definition of Cn

(XN − a)− := max {0, −(XN − a)} .

It follows from Remark 3.1 that

Uncrossing cont. Uncrossing cont.

Doob’s Forward Convergence Theorem Forward Convergence Theorem cont.

Let Λ be the set of ω ∈ Ω for which Xn (ω) does NOT 1 2

Forward Convergence Theorem cont. Forward Convergence Theorem cont.

Martingales bounded in L2 Martingales bounded in L2 cont.

Definition 3.7 Theorem 3.8

Martingales bounded in L2 cont. Martingales bounded in L2 cont.

Proof cont. Proof cont.

By the Pythagorean thm.: Hence,

Martingales bounded in L2 cont. dimH

11 Kakutani’s Theorem on Product Martingales

Note that it follows from Kolomogov’s 0 − 1 law that

Proof of part (b) cont. Proof of part (b) cont.

Sums of zero-mean independent r.v. cont. dimH

partial sums are bounded

Doob decomposition Doob decomposition cont.

Doob decomposition cont. Doob decomposition cont.

We say that (An ) is the quadratic variation of (Xn ).

Doob decomposition cont. Doob decomposition cont.

The angle bracket process The angle bracket process cont.

Convegence of Mn vs. A∞ < ∞ Convegence of Mn vs. A∞ < ∞ cont.

Theorem 5.7 Proof of Theorem 5.7 part (a)

Convergence of Mn vs. A∞ < ∞ cont. Convergence of Mn vs. A∞ < ∞ cont.

Convergence of Mn vs. A∞ < ∞ cont. Convergence of Mn vs. A∞ < ∞ cont.

N S(k) is a martingale . This is so because (M S(k) )2 = martingale + previsible

Convergence of Mn vs. A∞ < ∞ cont. Convergence of Mn vs. A∞ < ∞ cont.

Proof of Theorem 5.7 part (a) cont.

Using this, we can apply (49) for M S(k) instead of M to

Mn∧S(k)(ω) (ω) = Mn (ω) ∀n. (59) P (A∞ = ∞, T (c) = ∞) > 0 ,

Convergence of Mn vs. A∞ < ∞ cont. Convergence of Mn vs. A∞ < ∞ cont.

Convergence of Mn vs. A∞ < ∞ cont. Cesáro’s Lemma

Kronecker’s Lemma Strong law under variance cond.

= (1 + An )−2 · (Nn−1 + An − (Nn−1 + An−1 ))

Levy’s extension of BC Lemma Proof of BC Lemma as a corollary of Thm

Levy’s extension of BC Lemma cont. Levy’s extension of BC Lemma cont.

Levy’s extension of BC Lemma cont. dimH

11 Kakutani’s Theorem on Product Martingales

Closure cont. Closure cont.

Closure cont. dimH

Recall Fatou Lemma Reverse Fatou Lemma

Theorem BDD cont Theorem BDD cont

Proof of Thm BDD cont.

a.s. convergence vs conv. in prob. Absolute continuity

Yn := sup |Xk − X |. Proof

Absolute continuity cont. Absolute continuity cont.

UI families cont. UI families cont.

UI families cont. UI families cont.

UI families cont. UI families cont.

Conditional expectation vs. UI cont. Conditional expectation vs. UI cont.