a

rn l
u o
o f
J
arXiv:math/9806022v2 [math.PR] 3 Dec 1999

P
c r
o ni ob
abil
Elec t r ity

http://www.math.washington.edu/~ejpecp/

Concrete Representation of Martingales

Stephen Montgomery-Smith
Department of Mathematics
University of Missouri, Columbia, MO 65211
stephen@math.missouri.edu
http://math.missouri.edu/~stephen

Volume 3, paper number 15, 1998

http://www.math.washington.edu/~ejpecp/EjpVol3/paper15.abs.html

15 pages, submitted June 4, 1998, published December 2, 1998

Abstract: Let (fn ) be a mean zero vector valued martingale sequence. Then there exist
R1
vector valued functions (dn ) from [0, 1]n such that 0 dn (x1 , . . . , xn ) dx
Pnn = 0 for almost all
x1 , . . . , xn−1 , and such that the law of (fn ) is the same as the law of ( k=1 dk (x1 , . . . , xk )).
Similar results for tangent sequences and sequences satisfying condition (C.I.) are pre-
sented. We also present a weaker version of a result of McConnell that provides a Skorohod
like representation for vector valued martingales.

Keywords: martingale, concrete representation, tangent sequence, condition (C.I.), UMD,

Skorohod representation

A.M.S. Classification (1991): 60G42, 60H05.

Research supported in part by the N.S.F. and the Research Board of the University of
Missouri.
1. Introduction
In this paper, we seek to give a concrete representation of martingales. We will present
theorems of the following form. Given a martingale sequence (fn ) (possibly vector valued),
there is a martingale (gn ) on the probability space [0, 1]N , with respect to the filtration Ln ,
the minimal sigma field for which the first n coordinates of [0, 1]N are measurable, such
that the sequence (fn ) has the same law as (gn ). Thus any martingale may be represented
by a martingale
n
X
gn ((xn )) = dk (x1 , . . . , xk ),
k=0

R1
where 0 dn (x1 , . . . , xn ) dxn = 0 for n ≥ 1. (Here, as in the rest of this paper, the notation
(xn ) will refer to the sequence (xn )n∈N , where N refers to the positive integers.)
The value of such a result is perhaps purely psychological. However, we will also
present similar such results for tangent sequences, and also for sequences satisfying condi-
tion (C.I.). Tangent sequences and condition (C.I.), as defined in this paper, were intro-
duced by Kwapień and Woyczyński [KW]. The abstract definition of tangent sequences can
be, perhaps, a little hard to grasp. However, in this concrete setting, it is clear what their
meaning is. To demonstrate the psychological advantage that this view gives, we will give
a new proof of a result of McConnell [M2] that states that tangent sequence inequalities
hold for the UMD spaces.
We will also give a weaker version of a result of McConnell [M1] that provides a
Skorohod representation theorem for vector valued martingales, in that any martingale is a
stopped continuous time stochastic process. This means that many martingale inequalities
that are true for continuous time stochastic processes are automatically also true for general
martingales.

2
2. Representations of sequences of random variables
In this section, we present the basic result upon which all our other results will depend.
First let us motivate this result by considering just one random variable, that is a
measurable function f from the underlying probability space to a separable measurable
space (S, S). (A measurable space is said to be separable if its sigma field is generated by
a countable collection of sets.) We seek to find a measurable function g : [0, 1] → S that
has the same law as f , that is, given any measurable set A, we have that Pr(f ∈ A) =
λ(g ∈ A), or equivalently,
R given any measurable bounded function F : S → R, we have
that E(F (f )) = F (g) dλ. Here λ refers to the Lebesgue measure on [0, 1].
First, it may be shown without loss of generality that (S, S) is R with the Borel sets.
(The argument for this may be found in Chapter 1 of [DM], and is presented below.) Then
the idea is to let g be the so called increasing rearrangement of f , that is,

g(x) = sup{t ∈ R : Pr(f < t) < x}.

That g has the required properties is easy to show.

Next, suppose that f is nowhere constant, that is, we have that Pr(f = s) = 0 for all
s ∈ S. Then it may be seen that g is a strictly increasing function. In that case, it may
be seen that the minimal complete sigma field for which g is measurable is the collection
of Lebesgue measurable sets.
Now let us move on to state the main result. We will start by setting our notation.
We will work on two probability spaces: a generic one (Ω, F , Pr), and ([0, 1]N , LN , λ). Here
LN refers to the Lebesgue measurable sets on [0, 1]N, and λ refers to Lebesgue measure on
[0, 1]N .
Let us make some notational abuses. The reason for this is to make some expressions
less cumbersome, while hopefully not being too obscure. We will always identify [0, 1]n
with the natural projection of [0, 1]N onto the first n coordinates. Any function g on
[0, 1]n will be identified with its canonical lifting on [0, 1]N. The notation Ln refers to the
Lebesgue measurable sets on [0, 1]n , and also to their canonical lifting onto [0, 1]N. We will
let λ also refer to Lebesgue measure on [0, 1]n.
Given a random variable f , and a sigma field G, we will say that f is nowhere constant
with respect to G if Pr(f = g) = 0 for every G measurable function g. Let us illustrate
this notion with respect to a measurable function f on [0, 1]2 . It is nowhere constant with
respect to the trivial sigma field if and only if it is nowhere constant as defined above. Let
G be the sigma field generated by the first coordinate. Then f is nowhere constant with
respect to G if and only if for almost every x ∈ [0, 1] the function y 7→ f (x, y) is nowhere
constant. Now let G be the set of all Lebesgue measurable sets on [0, 1]2 . In this case, f
can never be nowhere constant with respect to G.
If f1 , . . . , fn are random variables taking values in a measurable space (S, S), we will
let σ(f1 , . . . , fn ) denote the minimal sigma field which contains all sets of measure zero,
and for which f1 , . . . , fn are measurable. Note that a simple argument shows that the
σ(f1 , . . . , fn ) measurable functions coincide precisely with the functions that are almost
everywhere equal to F (f1 , . . . , fn ), where F is some measurable function on S n .
Throughout these proofs we will use the following idea many times. Let (Ω, F , Pr) be
a probability space, and let (S, S) be a measurable space. Suppose that X : Ω → S is a

3
measurable function such that σ(X) is the whole of F . If Y is any random variable on Ω
taking values in R, then there exists a measurable map φ : S → R such that Y = φ ◦ X
almost surely. This is easily seen by approximating Y by simple functions.
Theorem 2.1. Let (fn ) be a sequence of random variables taking values in a separable
measurable space (S, S). Then there exists a sequence of measurable functions (gn :
[0, 1]n → S) that has the same law as (fn ).
If further we have that fn+1 is nowhere constant with respect to σ(f1 , . . . , fn ) for all
n ≥ 0, then we may suppose that σ(g1 , . . . , gn ) = Ln , for all n ≥ 1.
If even further we have that (S, S) is R with the Borel sets, then we may suppose
gn is Borel measurable (that is, the preimages of Borel sets are Borel sets), and that
gn (x1 , . . . , xn ) is a strictly increasing function of xn ∈ (0, 1) for almost all x1 , . . . , xn−1 .

Proof: We may suppose without loss of generality that S = R, and S is the Borel subsets
of R. To see that we may do this, see first that we may suppose without loss of generality
that S separates points in S, that is, if s 6= t ∈ S, then there exists A ∈ S such that s ∈ A
and t ∈
/ A. Let {Cn } be a countable collection of sets in S that generate S. Notice then
that the sequence {Cn } also separates points in S, that is, if s 6= t ∈ S, then there exists a
number n such that one and only one of s or t is in Cn . Define a map ϕ : S → R:
∞
X I(s∈Cm )
ϕ(s) = m
.
m=1
3

Clearly ϕ is injective. Further ϕ maps Cn to Dn ∩ ϕ(S), where

∞
( )
X em
Dn = m
: em = 0 or 1, en = 0 ,
m=1
3

and thus ϕ maps any element of S to a Borel subset of R intersected with ϕ(S). Conversely,
the preimage of Dn under ϕ is Cn , and hence the preimage of any Borel set under ϕ is in
S. (This argument may be found in Chapter 1 of [DM]).
Now apply the theorem to (ϕ ◦ fn ), to obtain (gn ). Since the law of gn is the same as
ϕ ◦ fn , the range of gn lies in the range of ϕ(S) with probability one. Then the sequence
(ϕ−1 ◦ gn ) will have the same law as (fn ).
So let us suppose as the induction hypothesis that we have obtained (gn : [0, 1]n →
R)1≤n≤N that has the same law as (fn )1≤n≤N , and that gn is Borel measurable, for all
n ≤ N , where N is a non-negative integer. (The induction is started with N = 0 in which
case the hypothesis is vacuously true. The arguments that follow simplify greatly in this
case.)
For each t ∈ Q, let
pt = E(I(fN +1 <t) |σ(f1 , . . . , fN )).
Since pt is σ(f1 , . . . , fN ) measurable, we may write

pt = qt (f1 , . . . , fN ) almost surely

4
for some Borel measurable function qt : RN → [0, 1]. Define the Borel measurable functions
rt by
rt = qt (g1 , . . . , gN ).

Since the sequence (rt )t∈Q has the same law as (pt )t∈Q , we see that there is a Borel
set B ⊂ [0, 1]N of full measure such that if (x1 , . . . , xN ) ∈ B, then rt (x1 , . . . , xN ) is an
increasing function of t ∈ Q, that tends to 0 as t → −∞, and tends to 1 as t → ∞.
If (x1 , . . . , xN ) ∈ B and xN+1 ∈ (0, 1), let

gN+1 (x1 , . . . , xN+1 ) = sup{t ∈ Q : rt (x1 , . . . , xN ) < xN+1 },

and let it be zero otherwise. We see that gN+1 is Borel measurable, since gN+1 on B ×(0, 1)
is the supremum of countably many Borel measurable functions (st )t∈Q where

t if rt (x1 , . . . , xN ) < xN+1
st (x1 , . . . , xN+1 ) =
−∞ otherwise.

It is easy to see that gN+1 is always finite.

Let us now show that (f1 , . . . , fN+1 ) has the same law as (g1 , . . . , gN+1 ). It is sufficient
to show that for any Borel set A ⊂ RN , and for t ∈ Q, that

Pr((f1 , . . . , fN ) ∈ A and fN+1 < t) = λ((g1 , . . . , gN ) ∈ A and gN+1 < t).

So let us begin.

λ((g1 , . . . , gN ) ∈ A and gN+1 < t)

= λ({(g1 , . . . , gN ) ∈ A and gN+1 < t} ∩ B × (0, 1))
≤ λ({(x1 , . . . , xN+1 ) : (g1 (x1 ), . . . , gN (x1 , . . . , xN )) ∈ A and rt (x1 , . . . , xN ) ≥ xN+1 }
∩ B × (0, 1))
= λ({(x1 , . . . , xN+1 ) : (g1 (x1 ), . . . , gN (x1 , . . . , xN )) ∈ A and rt (x1 , . . . , xN ) ≥ xN+1 })
Z Z 1
= I(rt ≥x) I((g1 ,...,gN )∈A) dx dλ
[0,1]N 0
Z
= rt I((g1 ,...,gN )∈A) dλ
[0,1]N
= E(pt I((f1 ,...,fN )∈A) )
= E(I(fN +1 <t) I((f1 ,...,fN )∈A) )
= Pr((f1 , . . . , fN ) ∈ A and fN+1 < t).

Similarly

λ((g1 , . . . , gN ) ∈ A and gN+1 ≤ t) ≥ Pr((f1 , . . . , fN ) ∈ A and fN+1 < t),

5
and so by Lebesgue’s monotone convergence theorem

λ((g1 , . . . , gN ) ∈ A and gN+1 < t) = lim λ((g1 , . . . , gN ) ∈ A and gN+1 ≤ s)

sրt
s∈Q

≥ lim Pr((f1 , . . . , fN ) ∈ A and fN+1 < s)

sրt
s∈Q

= Pr((f1 , . . . , fN ) ∈ A and fN+1 < t).

Thus we have shown that (f1 , . . . , fN+1 ) has the same law as (g1 , . . . , gN+1 ).
Now let us add the assumption that fn+1 is nowhere constant with respect to σ(f1 , . . . , fn )
for all n ≥ 0. Let us assume the inductive hypothesis that σ(g1 , . . . , gN ) = LN ,
Then there exist Borel measurable functions (αn : Rn → [0, 1])1≤n≤N such that for
1 ≤ n ≤ N we have

αn (g1 (x1 ), . . . , gN (x1 , . . . , xN )) = xn almost everywhere.

Let βn = αn (g1 , . . . , gN ), and let γn = αn (f1 , . . . , fN ).

Notice that gN+1 (x1 , . . . , xN+1 ) is an increasing function in xN+1 ∈ (0, 1) for fixed
(x1 , . . . , xN ) ∈ B. Let us show that gN+1 (x1 , . . . , xN+1 ) is a strictly increasing function
in xN+1 ∈ (0, 1) for (x1 , . . . , xN ) ∈ C where C is a subset of B of full measure. For
suppose otherwise. Then the following set has positive measure: the set of (x1 , . . . , xN+1 )
such that for some rational number t ∈ [0, 1] we have that gN+1 (x1 , . . . , xN , xN+1 ) =
gN+1 (x1 , . . . , xN , t). But this is equal to the set
[
{gN+1 = gN+1 (β1 , . . . , βN , t)}.
t∈[0,1]∪Q

But the measure of this set is the same as the probability of the set
[
{fN+1 = gN+1 (γ1 , . . . , γN , t)}.
t∈[0,1]∪Q

Since fN+1 is nowhere constant with respect to σ(f1 , . . . , fN ), it follows that this last set
has probability zero.
Let us now show that σ(g1 , . . . , gN+1 ) = LN+1 . For any a ∈ [0, 1] we have that

{gN+1 ≤ gN+1 (β1 , . . . , βN , a)} = [0, 1]N × [0, a] up to a set of measure zero.

Thus σ(g1 , . . . , gN+1 ) contains all set of the form [0, 1]N × [0, a]. Since it also contains
σ(g1 , . . . , gN ) = LN , the result follows.
Q.E.D.

6
3. Representation of martingales — 1

Theorem 3.1. Let (dn ) be a Bochner integrable martingale difference sequence taking
values in a Banach space X. Then there exists a sequence of Bochner measurable functions
(en : [0, 1]n → X) such that (dn ) has the same law as (en ), and such that for almost every
x1 , . . . , xn we have
Z 1
en (x1 , . . . , xn ) dxn = 0.
0

Proof: First we show that without loss of generality that we may assume that dn+1 is
nowhere constant with respect to σ(d1 , . . . , dn ) for all n ≥ 0. To do this, replace Ω by
Ω × [0, 1]N, replace X by X × R (where R is equipped with the sigma field of Borel sets),
and replace dn by (ω, (xn)) 7→ (dn (ω), xn − 12 ). Apply the theorem to this sequence. After
obtaining the resulting (en ), compose these functions with the natural projection of X × R
to X.
Further, since (dn ) is Bochner integrable, we may suppose without loss of generality
that X is separable.
Apply Theorem 2.1 to obtain the sequence (en ). Suppose that φ : [0, 1]n−1 → R is
any bounded measurable function. Then there exists a bounded Borel measurable function
θ : X n−1 → R such that φ(x1 , . . . , xn−1 ) = θ(e1 (x1 ), . . . , en−1 (x1 , . . . , xn−1 )) for almost
all x1 , . . . , xn−1 . Then
Z
φen dλ = E(θ(d1 , . . . , dn−1 )dn ),
[0,1]n

and this is zero because E(dn |σ(d1 , . . . , dn−1 )) = 0. The result follows.
Q.E.D.

7
4. Representation of tangent sequences
The following definition may be found in [KW]. Let (Fn )n≥0 be an increasing sequence of
sigma fields on (Ω, F , Pr), where F0 is the trivial sigma field. Two adapted sequences (fn )
and (gn ) taking values in a measurable space (S, S) are said to be tangent if for each n ≥ 1
we have that the law of fn conditionally on Fn−1 is the same as the law of gn conditionally
on Fn−1 , that is, E(I(fn ∈A) |Fn−1 ) = E(I(gn ∈A) |Fn−1 ) for any A ∈ S.
Notice then that if F : S 2n−1 → R is any measurable bounded function, that

E(F (f1 , g1 , . . . , fn−1 , gn−1 , fn )) = E(F (f1 , g1 , . . . , fn−1 , gn−1 , gn )).

This is easily seen by reducing it to the case where F is a finite linear combination of
characteristic functions of sets of the form A1 × . . . × A2n−1 where A1 , . . . , A2n−1 ∈ S.
Theorem 4.1. Let (S, S) be a separable measurable space. Let (fn ) and (gn ) be S-valued
measurable sequences adapted to (Fn )n≥0 that are tangent. Then there exists a sequence
(hn : [0, 1]n → S), and a sequence of Borel measurable functions (φn : [0, 1]n → [0, 1]) such
that the map φn (x1 , . . . , xn−1 , ·) is a measure preserving map on [0, 1] for almost every
x1 , . . . , xn−1 , such that the law of (fn , gn ) is the same as the law of (hn , kn ), where

kn (x1 , . . . , xn ) = hn (x1 , . . . , xn−1 , φn (x1 , . . . , xn )).

Further, if S is a separable Banach space, and (fn ) is a martingale difference sequence

with respect to Fn , then
Z 1
hn (x1 , . . . , xn ) dxn = 0
0

for almost all x1 , . . . , xn−1 .

Lemma 4.2. Let (S, S, µ), (T, T , ν) and (U, U, τ ) be measure spaces such that U is sepa-
rable. Let φ : S × T → U be a measurable map such that for A ∈ S and C ∈ U we have
that
µ ⊗ ν((s, t) : s ∈ A and φ(s, t) ∈ C) = µ(A)τ (C).
Then φ(s, ·) is a measure preserving map T → U for µ almost every s.

Proof: The hypothesis of the lemma can be recast as:

Z Z
ν(t : φ(s, t) ∈ C) dµ(s) = τ (C) dµ,
s∈A A

from which we conclude that for every C ∈ U, we have that ν(t : φ(s, t) ∈ C) = τ (C) for
µ almost every s.
Let C be a countable subcollection of U that generates U. Then there exists a set
A ∈ S of full measure such that ν(t : φ(s, t) ∈ C) = τ (C) for C ∈ C and s ∈ A. Let

M = {C ∈ U : ν(t : φ(s, t) ∈ C) = τ (C) for s ∈ A}.

8
Notice that M is a sigma field that contains C, and hence M = U.
Q.E.D.
Proof of Theorem 4.1: Without loss of generality (at least until we prove the last
part), let us suppose that S = R with the Borel sets. To do this, we use the same argument
that we gave at the beginning of the proof of Theorem 2.1.
Consider the sequence (f1 , g1 , f2 , g2 , . . .). By the kind of argument we presented at the
beginning of Theorem 3.1, we may suppose that each element of this sequence is nowhere
constant with respect to the sigma field generated by those elements in the sequence that
precede it.
Applying Theorem 2.1, we create a sequence of Borel measurable functions (h̃1 , k̃1 , h̃2 , k̃2 , . . .)
with the same law, where h̃n : [0, 1]2n−1 → R, k̃n : [0, 1]2n → R, and h̃n (x1 , y1 , . . . , xn−1 , yn−1 , xn )
is a strictly increasing function of xn ∈ (0, 1) for almost every x1 , y1 , . . . , xn−1 , yn−1 .
Define the function

rn (x1 , y1 , . . . , xn−1 , yn−1 , s) = sup{t ∈ Q ∩ (0, 1) : h̃n (x1 , y1 , . . . , xn−1 , yn−1 , t) < s},

where we will set the first supremum equal to zero if the set is empty. Notice that rn is the
supremum of countably many Borel measurable functions, and hence is Borel measurable.
Also, since h̃n (x1 , y1 , . . . , yn−1 , xn−1 , xn ) is a strictly increasing function of xn ∈ (0, 1) for
almost every x1 , y1 , . . . , xn−1 , yn−1 , we see that

rn (x1 , y1 , . . . , xn−1 , yn−1 , h̃n (x1 , y1 , . . . , xn−1 , yn−1 , xn )) = xn

for almost every x1 , y1 , . . . , xn . Define

φ̃n (x1 , y1 , . . . , xn , yn ) = rn (x1 , y1 , . . . , xn−1 , yn−1 , k̃n (x1 , y1 , . . . , xn , yn )).

Now let αn , βn be functions such that

xn = αn (h̃1 (x1 ), k̃1 (x1 , y1 ), . . . , h̃n (x1 , y1 , . . . , xn )) almost everywhere

yn = βn (h̃1 (x1 ), k̃1 (x1 , y1 ), . . . , k̃n (x1 , y1 , . . . , xn , yn )) almost everywhere,

and let γn , δn be random variables defined by

γn = αn (f1 , g1 , . . . , fn )
δn = βn (f1 , g1 , . . . , fn , gn ).

If A is any Borel subset of [0, 1]2n−2 , and C is a Borel subset of [0, 1], then since fn
and gn have the same law with respect to Fn−1 ,

λ((x1 , y1 , . . . , xn−1 , yn−1 ) ∈ A and φ̃n (x1 , y1 , . . . , xn , yn ) ∈ C)

= Pr((γ1 , δ1 , . . . , γn−1 , δn−1 ) ∈ A and rn (γ1 , δ1 , . . . , γn−1 , δn−1 , gn ) ∈ C)
= Pr((γ1 , δ1 , . . . , γn−1 , δn−1 ) ∈ A and rn (γ1 , δ1 , . . . , γn−1 , δn−1 , fn ) ∈ C)
= λ((x1 , y1 , . . . , xn−1 , yn−1 ) ∈ A and rn (x1 , y1 , . . . , xn−1 , yn−1 , h̃n (x1 , y1 , . . . , xn )) ∈ C)
= λ((x1 , y1 , . . . , xn−1 , yn−1 ) ∈ A and xn ∈ C)
= λ(A)λ(C).

9
Thus, by Lemma 4.2, we have that φ̃(x1 , y1 , . . . , xn−1 , yn−1 , ·, ·) is a measure preserving
map from [0, 1]2 to [0, 1] for almost every x1 , y1 , . . . , xn−1 , yn−1 .
Now let x 7→ (θ1 (x), θ2 (x)) be any Borel measurable measure preserving map from
[0, 1] to [0, 1]2 . Let

hn (x1 , . . . , xn ) = h̃n (θ1 (x1 ), θ2 (x1 ), . . . , θ1 (xn ))

kn (x1 , . . . , xn ) = k̃n (θ1 (x1 ), θ2 (x1 ), . . . , θ1 (xn ), θ2 (xn )).

Notice that
kn (x1 , . . . , xn ) = hn (x1 , . . . , xn−1 , φn (x1 , . . . , xn )),
where
φn (x1 , . . . , xn ) = φ̃n (θ1 (x1 ), θ2 (x1 ), . . . , θ1 (xn ), θ2 (xn )).
The last part of the theorem follows by exactly the same argument as for Theorem 3.1.
Q.E.D.

5. Representation of sequences satisfying condition (C.I.)

The following definition may be found in [KW]. Let (Fn )n≥0 be an increasing sequence of
sigma fields on (Ω, F , Pr), where F0 is the trivial sigma field. An adapted sequence (fn )
taking values in a measurable space (S, S) is said to satisfy condition (C.I.) if there exists
a sigma field G ⊂ F such that the law of fn conditionally on Fn−1 is the same as the law
of fn conditionally on G, that is, E(I(fn ∈A) |Fn−1 ) = E(I(fn ∈A) |G) for any A ∈ S, and if
the sequence (fn ) is conditionally independent with respect to G, that is, for any sequence
of sets An ∈ S we have

E(I(f1 ∈A1 ) · · · I(fn ∈An ) |G) = E(I(f1 ∈A1 ) |G) · · · E(I(fn ∈An ) |G).

It is shown in [KW] that given any sequence (fn ) adapted to some filtration, that
after possibly enlarging the underlying probability space, that there exists a sequence (f˜n )
that is tangent to (fn ) and that satisfies condition (C.I.).
Indeed, using Theorem 2.1, we can show a technically weaker but essentially identical
result as follows. Let (gn ) be the sequence constructed by Theorem 2.1. Enlarge the
probability space [0, 1]N to [0, 1]N × [0, 1]N, and define the sequences

un ((xn ), (yn )) = gn (x1 , . . . , xn−1 , xn )

vn ((xn ), (yn )) = gn (x1 , . . . , xn−1 , yn ).

Then (un ) has the same law as (fn ), and (vn ) is tangent to (un ) with respect to the filtration
(Ln ⊗ L0 ), and (vn ) satisfies condition (C.I.). (Here Ln is the minimal complete sigma
field on [0, 1]N for which the first n coordinate functions are measurable. In particular, L0
denotes the trivial complete measure sigma field on [0, 1]N .)
This is the motivation behind the next result, which shows that this previous con-
struction is the canonical representation.

10
Theorem 5.1. Let (S, S) be a separable measurable space. Let (fn ) be an S-valued
measurable sequence adapted to (Fn )n≥0 that satisfies condition (C.I.). Then there exists
a sequence (hn : [0, 1]n → S) such that if we define the functions un on [0, 1]N × [0, 1]N by

un ((xn ), (yn )) = hn (x1 , . . . , xn−1 , yn ),

then the laws of (fn ), (hn ) and (un ) are the same.
Further, if S is a separable Banach space, and (fn ) is a martingale difference sequence
with respect to Fn , then Z 1
hn (x1 , . . . , xn ) dxn = 0
0

for almost all x1 , . . . , xn .

Proof: As usual, we suppose without loss of generality that fn is nowhere constant with
respect to σ(f1 , . . . , fn−1 ). Let (hn ) be the sequence generated by Theorem 2.1. Then for
any A1 , . . . , An ∈ S we have that

Pr((f1 , . . . , fn ) ∈ A1 × · · · × An )
= E(E(I(f1 ∈A1 ) · · · I(fn ∈An ) |G))
= E(E(I(f1 ∈A1 ) |G) · · · E(I(fn ∈An ) |G))
= E(E(I(f1 ∈A1 ) |F0 ) · · · E(I(fn ∈An ) |Fn−1 )).

Now for each k = 1, . . . , n, let Fk : S k−1 → [0, 1] be the measurable function such that

E(I(fk ∈Ak ) |Fk−1 ) = Fk (f1 , . . . , fk−1 ) almost surely.

Notice that if G : S n−1 → R is bounded measurable, then

E(I(fk ∈Ak ) G(f1 , . . . , fk−1 )) = E(Fk (f1 , . . . , fk−1 )G(f1 , . . . , fk−1 )),

and hence

E(I(hk ∈Ak ) G(h1 , . . . , hk−1 )) = E(Fk (h1 , . . . , hk−1 )G(h1 , . . . , hk−1 )).

Since σ(h1 , . . . , hk−1 ) = Lk−1 , we see that

E(I(hk ∈Ak ) |Lk−1 ) = Fk (h1 , . . . , hk−1 ) almost everywhere.

Thus the sequence

(E(I(fk ∈Ak ) |Fk−1 ))1≤k≤n
has the same law as the sequence

(E(I(hk ∈Ak ) |Lk−1 ))1≤k≤n ,

11
and this last sequence is very easily seen to have the same law as the sequence

(E(I(uk ∈Ak ) |Lk−1 ⊗ L0 ))1≤k≤n .

Hence

E(E(I(f1 ∈A1 ) |F0 ) · · · E(I(fn ∈An ) |Fn−1 ))

Z
= (E(I(u1 ∈A1 ) |L0 ⊗ L0 ) · · · EI(un ∈An ) |Ln−1 ⊗ L0 ) dλ ⊗ λ,

and this last quantity is easily computed to be

λ ⊗ λ((u1 , . . . , un ) ∈ A1 × · · · × An ).

Finally, the last part of the theorem follows exactly as in Theorem 3.1.
Q.E.D.

6. Tangent sequences in UMD spaces

The concept of UMD spaces was introduced by Aldous [A], and extensively explored by
Burkholder [Bu]. It is from this second reference that the following definition of UMD
(along with many other equivalent definitions) may be found.
A Banach space X is said to be UMD if for some 1 < p < ∞ (equivalently all
1 < p < ∞), there exists a positive constant c depending only upon X and p such that if
(dn ) is a Bochner integrable X-valued martingale difference sequence, then for any signs
ǫn = ±1, we have that
XN XN
ǫn dn ≤ c dn .



n=1 Lp (X) n=1 Lp (X)

This section is devoted to providing a new proof of the following result. This is essentially
the same as the first part of Theorem 2.2 from the paper by McConnell [M2]. While the
following proof can be rewritten so as to avoid needing the representation theorems, to do
so would be to remove the motivation behind this proof.
Theorem 6.1. Let X be a UMD space. Then given 1 < p < ∞, there is a positive
constant C, depending only upon X and p, such that given a filtration (Fn )n≥0 , and two
martingales (fn ) and (gn ) adapted to this filtration that are X-valued Bochner integrable
and tangent to each other, we have that

XN XN
gn ≤ C fn .



n=1 Lp (X) n=1 Lp (X)

Proof: It is sufficient to prove the inequality in the case when one of the sequences
satisfies condition (C.I.), since the inequality may be then obtained by comparing both
sides with another sequence tangent to them both that satisfies condition (C.I.). We will

12
prove it in the case when the sequence (gn ) satisfies condition (C.I.), but it will be evident
that the proof also works when the sequence (fn ) satisfies condition (C.I.).
So applying Theorems 4.1 and 5.1, we see that there exists a sequence hn : [0, 1]n → X
of X-valued Bochner integrable functions such that

Z 1
hn (x1 , . . . , xn ) dxn = 0
0

for almost all x1 , . . . , xn−1 , and such that if we define the the functions dn , en : [0, 1]N ×
[0, 1]N → X by

dn ((xn ), (yn)) = hn (x1 , . . . , xn−1 , xn )

en ((xn ), (yn)) = hn (x1 , . . . , xn−1 , yn ),

then the sequence (f1 , g1 , f2 , g2 , . . .) has the same law as the sequence (d1 , e1 , d2 , e2 , . . .).
Now we define a martingale difference sequence rn on [0, 1]N ×[0, 1]N by r2n−1 = dn +en
and r2n = dn − en , where the filtration is G2n = Ln ⊗ Ln , and G2n−1 is the collection of
those sets in G2n generated by sets like

{((xn ), (yn )) : (x1 , . . . , xn ) ∈ A × C, (y1 , . . . , yn ) ∈ B × C},

where A, B ∈ Ln−1 , and C ∈ L1 .

PN P2N PN P2N
Then we see that n=1 dn = 21 n=1 rn and n=1 en = 12 n=1 (−1)n+1 rn , and thus
the result follows immediately from the fact that X is a UMD space.
Q.E.D.

13
7. Representations of martingales — 2
In this section, we present a version of a result of McConnell [M1]. This result provides a
Skorohod like representation for vector valued martingales. Our result is weaker than that
of McConnell in that his process is adapted to a filtration generated by a one-dimensional
Brownian motion. The process presented here is adapted to a two dimensional Brownian
motion.
Theorem 7.1. Let (dn ) be a Bochner integrable martingale difference sequence taking
values in a Banach space X. Then there exists a continuous time X-valued stochastic
process (Ft )t≥0 with continuous sample paths such that (dn ) has the same law as (Fn −
Fn−1 ). Furthermore, the process (Ft ) is adapted to a two dimensional brownian motion.

Proof: Let D = {z ∈ C : |z| < 1}, D̄ = {z ∈ C : |z| ≤ 1} and ∂D = {z ∈ C : |z| = 1},

and give ∂D normalized Lebesgue measure d|z|/2π. Since ∂D is measure equivalent to
[0, 1], we may apply Theorem 3.1 to obtain a sequence (en : ∂Dn → X) that has the same
law as (fn ), and such that
d|zn |
Z
en (z1 , . . . , zn ) =0
∂D 2π
for almost every z1 , . . . , zn−1 .
For each z1 , . . . , zn−1 , extend the function zn 7→ en (z1 , . . . , zn ) to its harmonic exten-
sion on D̄. We will identify en with this extension.
(n)
Now let (bt )n∈N be a sequence of independent Brownian motions into C with origin
(n)
at zero. Let τn be the stopping time τn = inf{t : |bt | ≥ 1}. Notice that τn is finite almost
surely.
By the Itô calculus, we have that
Z min{t,τn }
(1) (n)
en (bτ1 , . . . , bmin{t,τn } ) = ∇n en (b(1) (n−1) (n) (n)
τ1 , . . . , bτn−1 , bs ) · dbs ,
0

where ∇n refers to taking the gradient with respect to the nth coordinate. Define a process
n−1
!
(n)
X
Gn,t = em (bτ1 , . . . , bτm ) + en (b(1)
(1) (m)
τ1 , . . . , bmin{t,τn } ).
m=1

Let φ : [0, 1) → [0, ∞) be defined by φ(t) = t/(1 − t), and let

Ft = G[t+1],φ(t−[t]) ,
where [t] represents the largest integer less than t.
To see that (Ft ) may be adapted to a two-dimensional brownian motion, let b̃t be a
brownian motion taking values in C, and set
Z n−1+φ−1 (t) p
(n)
bt = φ′ (s) db̃s .
n−1

Q.E.D.

14
 Acknowledgments
I would like to mention the help of Gerald Edgar and P. Fitzsimmons in obtaining the
argument at the beginning of Theorem 2.1, where it is shown that without loss of generality
that (S, S) is R with the Borel sets, and for bringing the reference [DM] to my attention. I
also extend thanks to Donald Burkholder, who brought the reference [M1] to may attention.
Finally, I would like to thank Terry McConnell and an anonymous referee for many useful
remarks.

References
[A] D.J. Aldous, Unconditional bases and martingales in Lp (F ), Math. Proc. Cambridge
Philos. Soc. 85, (1979), 117–123.
[Bu] D.L. Burkholder, A geometrical characterization of Banach spaces in which martingale
difference sequences are unconditional, Ann. Probab. 9, (1981), 997–1011.
[DM] C. Dellacherie and P.-A. Meyer, Probabilities and Potential, North-Holland Mathe-
matics Studies 29, North-Holland, Amsterdam-New York-Oxford, 1978.
[KW] S. Kwapień and W.A. Woyczyński, Tangent sequences of random variables, in Almost
Everywhere Convergence, G.A. Edgar and L. Sucheston, Eds., Academic Press, 1989,
pp. 237–265.
[M1] T.R. McConnell, A Skorohod-like representation in infinite dimensions, Probability
in Banach spaces, V (Medford, Mass., 1984), 359–368, Lecture Notes in Math., 1153,
Springer, Berlin-New York, 1985.
[M2] T.R. McConnell, Decoupling and stochastic integration in UMD Banach spaces, Probab.
Math. Statist. 10, (1989), 283–295.

15