4

Martingales in Discrete-Time

Suppose that (, F, P) is a probability space.

Definition 4.1. A sequence F = {Fn , n = 0, 1, . . .} is called a filtration if each Fn is a
sub--algebra of F, and Fn Fn+1 for n = 0, 1, 2, . . ..
In other words, a filtration is an increasing sequence of sub--algebras of F. For notational convenience, we define
!

[
F =
Fn
n=1

and note that F F.

Definition 4.2. If F = {Fn , n = 0, 1, . . .} is a filtration, and X = {Xn , n = 0, 1, . . .} is a
discrete-time stochastic process, then X is said to be adapted to F if Xn is Fn -measurable
for each n.
Note that by definition every process {Xn , n = 0, 1, . . .} is adapted to its natural filtration
{Fn , n = 0, 1, . . .} where Fn = (X0 , . . . , Xn ).
The intuition behind the idea of a filtration is as follows. At time n, all of the information
about that is known to us at time n is contained in Fn . As n increases, our knowledge
of also increases (or, if you prefer, does not decrease) and this is captured in the fact that
the filtration consists of an increasing sequence of sub--algebras.
Definition 4.3. A discrete-time stochastic process X = {Xn , n = 0, 1, . . .} is called a
martingale (with respect to the filtration F = {Fn , n = 0, 1, . . .}) if for each n = 0, 1, 2, . . .,
(a) X is adapted to F; that is, Xn is Fn -measurable,
(b) Xn is in L1 ; that is, E|Xn | < , and
(c) E(Xn+1 |Fn ) = Xn a.s.
From this definition, we can immediately conclude the following useful facts.
Theorem 4.4. Suppose that X = {Xn , n = 0, 1, . . .} is a martingale with respect to F.
(i) E(Xn ) = E(X0 ) for each n = 0, 1, 2, . . ..
(ii) E(Xn |Fm ) = Xm for every non-negative integer m n.
Proof. We begin first with the proof of (ii). Suppose that for a given n and m, we write
n = m + k. Since F = {Fn , n = 0, 1, . . .} is a filtration, repeated use of the tower property
of conditional expectation gives
E(Xn |Fm ) = E(E(Xn |Fn1 )|Fm ) = E(Xn1 |Fm )
= E(E(Xn1 |Fn2 )|Fm ) = E(Xn2 |Fm )
..
.
= E(E(Xn(k2) |Fn(k1) )|Fm ) = E(Xn(k1) |Fm )
= E(Xm+1 |Fm ) = Xm .

As for (i), we see that taking m = 0 in (ii) gives E(Xn |F0 ) = X0 . Thus taking expectations
we find
E(Xn ) = E(E(Xn |F0 )) = E(X0 )
completing the proof.
Remark. Some authors use part (ii) of the previous theorem in the definition of martingale
in place of our (c). For those authors, part (c) then follows as an easy corollary of (ii) in the
definition, just as for us (ii) followed as an easy corollary to our part (c).

4.1

Examples of martingales

Example 4.5. Suppose that X1 , X2 , . . . is a sequence of independent random variables with

E|Xi | < and E(Xi ) = 0 for each i. Let F0 = {, }, and for n 1 set Fn = (X1 , . . . , Xn )
so that F = {Fn , n = 0, 1, . . .} is a filtration. Define the stochastic process S by setting
S0 = 0, and
n
X
Sn =
Xi , n 1.
i=1

To show that S = {Sn , n = 0, 1, . . .} is a martingale with respect to F we need to verify the

three parts of the definition. Clearly Sn is Fn -measurable, and by assumption
E|Sn |

n
X

E|Xi | < .

i=1

Furthermore, since Sn is Fn -measurable, and since Xn+1 is independent of Fn , we conclude

E(Sn+1 |Fn ) = E(Xn+1 + Sn |Fn ) = E(Xn+1 |Fn ) + E(Sn |Fn ) = E(Xn+1 ) + Sn = Sn .
Taken together these show that S is a martingale with respect to F. Note that in this case
E(Sn ) = E(S0 ) = 0 for each n.
Example 4.6. As a slight generalization of the previous example, suppose that Z is a random
variable in L1 which is independent of X1 , X2 , . . .. Set X0 = Z, and for n = 0, 1, 2, . . ., take
Fn = (X0 , . . . , Xn ) and set
n
X
Sn =
Xi .
i=0

It follows as in the previous example that S is a martingale with respect to F = {Fn , n =

0, 1, . . .}, except this time E(Sn ) = E(S0 ) = E(Z) which need not equal 0.
Example 4.7. Suppose that S is as in Example 4.5, except assume further that E(Xi2 ) <
for each i. We show that Y = {Sn2 , n = 0, 1, 2, . . .} is not a martingale with respect to F.
Although Yn = Sn2 is Fn -measurable, and for each n 1,
E|Yn | =

E(Sn2 )

n
X
i=1

E(Xi2 ) < ,

we claim that E(Yn+1 |Fn ) 6= Yn . Since Xn+1 is independent of Fn , and since Sn is Fn measurable, it follows that
2
E(Yn+1 |Fn ) = E(Sn+1
|Fn ) = E((Sn+1 Sn + Sn )2 |Fn )
= E((Sn+1 Sn )2 |Fn ) 2E(Sn (Sn+1 Sn )|Fn ) + E(Sn2 |Fn )
2
|Fn ) 2E(Xn+1 Sn |Fn ) + E(Sn2 |Fn )
= E(Xn+1
2
= E(Xn+1
) 2Sn E(Xn+1 ) + Sn2
= 1 0 + Sn2 = 1 + Yn

Although Yn is not a martingale, we see that Zn = Sn2 n is a martingale since

2
|Fn ) (n + 1) = 1 + Sn2 (n + 1) = Sn2 n = Zn .
E(Zn+1 |Fn ) = E(Sn+1

Example 4.8. Suppose that X1 , X2 , . . . is a sequence of independent, non-negative random

variables with E(Xi ) = 1 for each i. Let F0 = {, }, and for n 1 set Fn = (X1 , . . . , Xn )
so that F = {Fn , n = 0, 1, . . .} is a filtration. Define the stochastic process M by setting
M0 = 1, and
n
Y
Mn =
Xi , n 1.
i=1

To show that M = {Mn , n = 0, 1, . . .} is a martingale with respect to F we need to verify

the three parts of the definition. Clearly Mn is Fn -measurable, and by assumption
E|Mn | =

n
Y

E|Xi | =

i=1

n
Y

E(Xi ) = 1 < .

i=1

Furthermore, since Mn is Fn -measurable, and since Mn+1 is independent of Fn , we conclude

E(Mn+1 |Fn ) = E(Xn+1 Mn |Fn ) = Mn E(Xn+1 |Fn ) = Mn E(Xn+1 ) = Mn 1 = Mn .
Taken together these show that M is a martingale with respect to F. Note that in this case
E(Mn ) = E(M0 ) = 1 for each n.
Example 4.9. Suppose that (, F, P) is a probability space, F = {Fn , n = 0, 1, . . .} is
a filtration, and Y : R is a random variable in L1 . Define the stochastic process
X = {Xn , n = 0, 1, . . .} by setting Xn = E(Y |Fn ) for each n = 0, 1, 2, . . .. The definition of
conditional expectation allows us to immediately conclude that Xn is Fn -measurable, and
from the conditional version of Jensens inequality we have
E|Xn | = E|E(Y |Fn )| E(E(|Y ||Fn )) = E|Y | <
so that Xn L1 for each n. Furthermore, the tower property of conditional expectation
gives
E(Xn+1 |Fn ) = E((E(Y |Fn+1 )|Fn ) = E(Y |Fn ) = Xn
which shows that X is a martingale with respect to F.
3

4.2

Stopping Times

Definition 4.10. A random variable T : N {+} is called a stopping time if

{T n} Fn for every n = 0, 1, 2, . . ..
Notice that the filtration F = {Fn , n = 0, 1, . . .} is an integrable part of the definition.
It is useful to think of a stopping time as the first time that a given random event happens
with the convention that T = + if it never happens. If T is a stopping time, and X =
{Xn , n = 0, 1, . . .} is a stochastic process, then define the random variable XT as XT () ().
Example 4.11. Suppose that S is a simple random walk on Z, and let T be the first time
that Sn = 4. That is,
(
min{n 0 : Sn = 4}, if Sn = 4 for some n N,
T =
+,
otherwise,
for in other words, T () = inf{n 0 : Sn () = 4}. (By writing inf, we stress that we follow
the convention that the infimum of the empty set is +.) In particular,
{T n} =

n
[

{Sk = 4} Fn

k=0

where {Sk = 4} Fk Fn for k n because {Fn , n = 0, 1, . . .} is a filtration. Hence, it

follows that T is a stopping time.
Definition 4.12. A stopping time T is said to be bounded if there exists a constant C <
such that P(T C) = 1, and is said to be finite a.s. if P(T < ) = 1.
Note that a bounded stopping time is necessarily finite a.s., although the converse need
not be true. For example, consider repeatedly flipping a fair coin, and let T denote the first
time a head is observed. Then T is finite a.s. but there is no C such that P(T C) = 1.
Indeed, P(T C) P(T dCe) = 1 2dCe < 1 for any constant C < .
Definition 4.13. If T is a stopping time, then the stopping time -algebra FT is defined as
FT = {A F : A {T n} Fn n}.
Exercise 4.14. Show that FT is, in fact, a -algebra.
Theorem 4.15. A random variable T : N {+} is a stopping time if and only if
{T = n} Fn for each n = 0, 1, 2, . . ..
Proof. Suppose that T is a stopping time so that {T n} Fn for each n. In particular,
{T n 1} Fn1 Fn since F = {Fn , n = 0, 1, . . .} is a filtration so that
{T = n} = {T n} {T n 1}c Fn .
On the other hand, if {T = n} Fn for each n, then {T = j} Fj Fn for each j n
since F is a filtration. Therefore,
{T n} =

n
[

{T = j} Fn

j=0

so that T is a stopping time.

4

Theorem 4.16. If X = {Xn , n = 0, 1, . . .} is a discrete-time stochastic process, and T is a

stopping time, then the random variable XT is FT -measurable.
Proof. In order to prove that XT is FT -measurable we must show that {XT B} FT for
every Borel set B B(R). In other words, we must show that {XT B} {T n} Fn
for every n. Since
{XT B} {T n} =

n
[

({XT B} {T = k}) =

k=0

n
[

({Xk B} {T = k})

k=0

and since {Xk B} {T = k} Fk Fn for every k n, we conclude that XT is

FT -measurable as required.
Exercise 4.17. If T is a stopping time, show that T is FT -measurable.
Recall from Theorem 4.4 that if X = {Xn , n = 0, 1, . . .} is a martingale, then for fixed
times n = 0, 1, 2, . . . it follows that E(Xn ) = E(X0 ). However, our goal is to determine when
E(XT ) = E(X0 ) for random times T . Even when the random time T is a stopping time, this
conclusion is not immediate as the following simple example shows.
Example 4.18. If S is a simple random walk on Z, and T = inf{n 0 : Sn = 4} as in
Example 4.11, then clearly ST = 4 so that E(ST ) = 4. Since E(S0 ) = 0, we conclude that
E(ST ) 6= E(S0 ).
Theorem 4.19. If T is a bounded stopping time, and X = {Xn , n = 0, 1, . . .} is a martingale, then E(XT ) = E(X0 ).
Proof. Since T is a bounded stopping time, we may assume without loss of generality that
T is bounded by N for some positive integer N . To begin, note that XT can be written as
XT () () =

Xn ()1{T ()=n} .

n=0

Since X is a martingale, we conclude that

!
!

N
N
X
X
X
E(XT ) = E
Xn 1{T =n} = E
Xn 1{T =n} =
E(Xn 1{T =n} )
n=0

N
X

n=0
N
X

E(E(XN |Fn )1{T =n} ) =

n=0

=E

n=0

E(E(XN 1{T =n} |Fn )) =

n=0
N
X
n=0

XN 1{T =n}

!
= E XN

N
X

E(XN 1{T =n} )

n=0

N
X

1{T =n} = E(XN ) = E(X0 )

n=0

where the last equality followed from Theorem 4.4.