2Martingales in discrete time

III Advanced Probability



2.3 Martingale convergence theorems
One particularly nice property of martingales is that they have nice conver-
gence properties. We shall begin by proving a pointwise version of martingale
convergence.
Theorem
(Almost sure martingale convergence theorem)
.
Suppose (
X
n
)
n0
is a super-martingale that is bounded in
L
1
, i.e.
sup
n
E|X
n
| <
. Then there
exists an F
-measurable X
L
1
such that
X
n
X
a.s. as n .
To begin, we need a convenient characterization of when a series converges.
Definition
(Upcrossing)
.
Let (
x
n
) be a sequence and (
a, b
) an interval. An
upcrossing of (
a, b
) by (
x
n
) is a sequence
j, j
+ 1
, . . . , k
such that
x
j
a
and
x
k
b. We define
U
n
[a, b, (x
n
)] = number of disjoint upcrossings contained in {1, . . . , n}
U[a, b, (x
n
)] = lim
n→∞
U
n
[a, b, x].
We can then make the following elementary observation:
Lemma.
Let (
x
n
)
n0
be a sequence of numbers. Then
x
n
converges in
R
if and
only if
(i) lim inf |x
n
| < .
(ii) For all a, b Q with a < b, we have U[a, b, (x
n
)] < .
For our martingales, since they are bounded in
L
1
, Fatou’s lemma tells us
E lim inf |X
n
| <
. So
lim inf |X
n
|
= 0 almost surely. Thus, it remains to show
that for any fixed
a < b Q
, we have
P
(
U
[
a, b,
(
X
n
)] =
) = 0. This is a
consequence of Doob’s upcrossing lemma.
Lemma (Doob’s upcrossing lemma). If X
n
is a super-martingale, then
(b a)E(U
n
[a, b(X
n
)]) E(X
n
a)
Proof.
Assume that
X
is a positive super-martingale. We define stopping times
S
k
, T
k
as follows:
T
0
= 0
S
k+1
= inf{n : X
n
a, n T
n
}
T
k+1
= inf{n : X
n
b, n S
k+1
}.
Given an
n
, we want to count the number of upcrossings before
n
. There are
two cases we might want to distinguish:
b
a
n
b
a
n
Now consider the sum
n
X
k=1
X
T
k
n
X
S
k
n
.
In the first case, this is equal to
U
n
X
k=1
X
T
k
X
S
k
+
n
X
k=U
n
+1
X
n
X
n
(b a)U
n
.
In the second case, it is equal to
U
n
X
k=1
X
T
k
X
S
k
+(X
n
X
S
U
n
+1
)+
n
X
k=U
n
+2
X
n
X
n
(ba)U
n
+(X
n
X
S
U
n
+1
).
Thus, in general, we have
n
X
k=1
X
T
k
n
X
S
k
n
(b a)U
n
+ (X
n
X
S
U
n
+1
n
).
By definition,
S
k
< T
k
n
. So the expectation of the LHS is always non-negative
by super-martingale convergence, and thus
0 (b a)EU
n
+ E(X
n
X
S
U
n
+1
n
).
Then observe that
X
n
X
S
U
n
+1
(X
n
a)
.
The almost-sure martingale convergence theorem is very nice, but often it is
not good enough. For example, we might want convergence in
L
p
instead. The
following example shows this isn’t always possible:
Example. Suppose (ρ
n
)
n0
is a sequence of iid random variables and
P(ρ
n
= 0) =
1
2
= P(ρ
n
= 2).
Let
X
n
=
n
Y
k=0
ρ
k
.
Then this is a martingale, and
EX
n
= 1. On the other hand,
X
n
0 almost
surely. So kX
n
X
k
1
does not converge to 0.
For
p >
1, if we want convergence in
L
p
, it is not surprising that we at least
need the sequence to be
L
p
bounded. We will see that this is in fact sufficient.
For
p
= 1, however, we need a bit more than being bounded in
L
1
. We will need
uniform integrability.
To prove this, we need to establish some inequalities.
Lemma
(Maximal inequality)
.
Let (
X
n
) be a sub-martingale that is non-
negative, or a martingale. Define
X
n
= sup
kn
|X
k
|, X
= lim
n→∞
X
n
.
If λ 0, then
λP(X
n
λ) E[|X
n
|1
X
n
λ
].
In particular, we have
λP(X
n
λ) E[|X
n
|].
Markov’s inequality says almost the same thing, but has
E
[
|X
n
|
] instead of
E[|X
n
|]. So this is a stronger inequality.
Proof.
If
X
n
is a martingale, then
|X
n
|
is a sub-martingale. So it suffices to
consider the case of a non-negative sub-martingale. We define the stopping time
T = inf{n : X
n
λ}.
By optional stopping,
EX
n
EX
T n
= EX
T
1
T n
+ EX
n
1
T >n
λP(T n) + EX
n
1
T >n
= λP(X
n
λ) + EX
n
1
T >n
.
Lemma (Doob’s L
p
inequality). For p > 1, we have
kX
n
k
p
p
p 1
kX
n
k
p
for all n.
Proof. Let k > 0, and consider
kX
n
kk
p
p
= E|X
n
k|
p
.
We use the fact that
x
p
=
Z
x
0
ps
p1
ds.
So we have
kX
n
kk
p
p
= E|X
n
k|
p
= E
Z
X
n
k
0
px
p1
dx
= E
Z
k
0
px
p1
1
X
n
x
dx
=
Z
k
0
px
p1
P(X
n
x) dx (Fubini)
Z
k
0
px
p2
EX
n
1
X
n
x
dx (maximal inequality)
= EX
n
Z
k
0
px
p2
1
X
n
x
dx (Fubini)
=
p
p 1
EX
n
(X
n
k)
p1
p
p 1
kX
n
k
p
(E(X
n
k)
p
)
p1
p
(H¨older)
=
p
p 1
kX
n
k
p
kX
n
kk
p1
p
Now take the limit k and divide by kX
n
k
p1
p
.
Theorem
(
L
p
martingale convergence theorem)
.
Let (
X
n
)
n0
be a martingale,
and p > 1. Then the following are equivalent:
(i) (X
n
)
n0
is bounded in L
p
, i.e. sup
n
E|X
i
|
p
< .
(ii)
(
X
n
)
n0
converges as
n
to a random variable
X
L
p
almost surely
and in L
p
.
(iii) There exists a random variable Z L
p
such that
X
n
= E(Z | F
n
)
Moreover, in (iii), we always have X
= E(Z | F
).
This gives a bijection between martingales bounded in
L
p
and
L
p
(
F
),
sending (X
n
)
n0
7→ X
.
Proof.
(i)
(ii): If (
X
n
)
n0
is bounded in
L
p
, then it is bounded in
L
1
. So by
the martingale convergence theorem, we know (
X
n
)
n0
converges almost
surely to X
. By Fatou’s lemma, we have X
L
p
.
Now by monotone convergence, we have
kX
k
p
= lim
n
kX
n
k
p
p
p 1
sup
n
kX
n
k
p
< .
By the triangle inequality, we have
|X
n
X
| 2X
a.s.
So by dominated convergence, we know that X
n
X
in L
p
.
(ii) (iii): Take Z = X
. We want to prove that
X
m
= E(X
| F
m
).
To do so, we show that
kX
m
E
(
X
| F
m
)
k
p
= 0. For
n m
, we know
this is equal to
kE(X
n
| F
m
)E(X
| F
m
)k
p
= kE(X
n
X
| F
m
)k
p
kX
n
X
k
p
0
as
n
, where the last step uses Jensen’s. But it is also a constant. So
we are done.
(iii)
(i): Since expectation decreases
L
p
norms, we already know that
(X
n
)
n0
is L
p
-bounded.
To show the “moreover” part, note that
S
n0
F
n
is a
π
-system that
generates F
. So it is enough to prove that
EX
1
A
= E(E(Z | F
)1
A
).
But if A F
N
, then
EX
1
A
= lim
n→∞
EX
n
1
A
= lim
n→∞
E(E(Z | F
n
)1
A
)
= lim
n→∞
E(E(Z | F
)1
A
),
where the last step relies on the fact that 1
A
is F
n
-measurable.
We finally finish off the
p
= 1 case with the additional uniform integrability
condition.
Theorem
(Convergence in
L
1
)
.
Let (
X
n
)
n0
be a martingale. Then the follow-
ing are equivalent:
(i) (X
n
)
n0
is uniformly integrable.
(ii) (X
n
)
n0
converges almost surely and in L
1
.
(iii) There exists Z L
1
such that X
n
= E(Z | F
n
) almost surely.
Moreover, X
= E(Z | F
).
The proof is very similar to the L
p
case.
Proof.
(i)
(ii): Let (
X
n
)
n0
be uniformly integrable. Then (
X
n
)
n0
is bounded
in
L
1
. So the (
X
n
)
n0
converges to
X
almost surely. Then by measure
theory, uniform integrability implies that in fact X
n
L
1
.
(ii) (iii): Same as the L
p
case.
(iii)
(i): For any
Z L
1
, the collection
E
(
Z | G
) ranging over all
σ-subalgebras G is uniformly integrable.
Thus, there is a bijection between uniformly integrable martingales and
L
1
(F
).
We now revisit optional stopping for uniformly integrable martingales. Recall
that in the statement of optional stopping, we needed our stopping times to be
bounded. It turns out if we require our martingales to be uniformly integrable,
then we can drop this requirement.
Theorem.
If (
X
n
)
n0
is a uniformly integrable martingale, and
S, T
are arbi-
trary stopping times, then E(X
T
| F
S
) = X
ST
. In particular EX
T
= X
0
.
Note that we are now allowing arbitrary stopping times, so
T
may be infinite
with non-zero probability. Hence we define
X
T
=
X
n=0
X
n
1
T =n
+ X
1
T =
.
Proof. By optional stopping, for every n, we know that
E(X
T n
| F
S
) = X
ST n
.
We want to be able to take the limit as
n
. To do so, we need to show
that things are uniformly integrable. First, we apply optional stopping to write
X
T n
as
X
T n
= E(X
n
| F
T n
)
= E(E(X
| F
n
) | F
T n
)
= E(X
| F
T n
).
So we know (
X
T
n
)
n0
is uniformly integrable, and hence
X
nT
X
T
almost
surely and in L
1
.
To understand E(X
T n
| F
S
), we note that
kE(X
nT
X
T
| F
S
)k
1
kX
nT
X
T
k
1
0 as n .
So it follows that E(X
nT
| F
S
) E(X
T
| F
S
) as n .