April 14, 1999

THE DALANG–MORTON–WILLINGER THEOREM

FREDDY DELBAEN

Contents
1. Introduction 1
2. Statement of the theorem 1
3. Reduction to the case T = 1 2
4. Some measure theory 2
5. Topological version of the no-arbitrage condition 8
6. Schachermayer’s proof 11
7. Rogers’ proof 11
References 13

1. Introduction
Roughly speaking, the Dalang–Morton–Willinger theorem states that for a finite
sequence of price vectors (S0 , S1 , . . . , ST ) in Rd , the no-arbitrage condition im-
plies the existence of an equivalent martingale measure. Different proofs exist [1],
[3], [5], [6], [7], and even [2]. Rogers’ original proof is extremely elegant but it
uses the existence of regular conditional distributions, the original proof of [1] uses
the measurable selection theorem and Schachermayer’s proof needs a good deal of
functional analysis. The present presentation closely follows [6]. It gives additional
results. The present note is not intended for publication in a journal. It is only a
“didactical note”. I give proofs of intermediate results in integration and measure
theory. These results are, I think, interesting in themselves.

2. Statement of the theorem

Let (Ω, FT , P) denote a probability space equipped with the finite filtration F0 ⊂
F1 ⊂ · · · ⊂ FT . For each t ∈ {0, . . . , T } the random variable St : Ω → Rd denotes
the price vector prevailing at time t. We suppose that each St is Ft -measurable,
i. e., the process S = (S0 , . . . , ST ) is adapted to (Ft )0≤t≤T . The inner product in
Rd is denoted by hx, yi or x0 y since we will frequently represent elements of Rd as
column vectors. If an economic agent buys the quantities α0 , . . . , αT −1 at times
0, . . . , T − 1, then of course the decisions have to be taken using the information
available at these times, more precisely, αt : Ω → Rd has to be Ft -measurable for
t ∈ {0, . . . , T − 1}. The total gain (or loss) using such a strategy is then given by
PT −1
the sum t=0 hαt , St+1 − St i. The set of all these sums forms the vector space
½ TX−1 ¯ ¾
¯
K= ¯
hαt , St+1 − St i ¯ αt : Ω → R is Ft -measurable for t ∈ {0, . . . , T − 1} .
d

t=0
1
2 F. DELBAEN

K is a subspace of L0 (Ω, FT , P), the space of all real-valued random variables, where
two random variables equal P-almost surely are identified.
Definition 2.1. With the notation introduced above, we say that the process S =
(S0 , . . . , ST ) satisfies the no-arbitrage condition, (NA) for short, if K ∩ L0+ = {0}.
The symbol L0+ denotes the cone of non-negative random variables in L0 (Ω, FT , P).
Theorem 2.2 (Dalang, Morton, Willinger). With the above notations: S satisfies
the no-arbitrage condition if and only if there is a probability measure Q equivalent
to P, such that under Q the process S is a martingale, i. e.:
(a) For all t ∈ {0, . . . , T } we have St ∈ L1 (Q).
(b) For all t ∈ {0, . . . , T − 1} we have St = EQ [St+1 | Ft ] where EQ denotes
expectation with respect to Q.
Moreover Q can be chosen so that dQ/dP ∈ L∞ , i. e., the density of Q with respect
to P is uniformly bounded.
If such an equivalent martingale measure Q exists, then EQ [f ] = 0 for every
f ∈ K. Since EQ [f ] = 0 for f ∈ L0+ if and only if f = 0 almost surely, we get
K ∩ L0+ = {0}. Therefore, we only treat the non-trivial direction of the Dalang–
Morton–Willinger theorem in the following.

3. Reduction to the case T = 1

We will prove the Dalang–Morton–Willinger theorem using induction. So we sup-
pose that the theorem is proved for T = 1 and we now show the induction step.
From the induction hypothesis we deduce the existence of a measure Q1 ∼ P with
bounded density dQ1 /dP such that (S1 , F1 ), . . . , (ST , FT ) is a Q1 -martingale, i. e.,
EQ1 [St+1 | Ft ] = St for all t ∈ {1, . . . , T − 1}. We remark that when Q1 is replaced
by a measure Q ∼ Q1 such that
• f ≡ dQ/dQ1 is uniformly bounded and
• f is F1 -measurable,
then Q also has the property EQ [St+1 | Ft ] = St for all t ∈ {1, . . . , T − 1}. Since the
theorem is true for T = 1, we can take f so that EQ [S1 | F0 ] = S0 . This completes
the reduction of the Dalang–Morton–Willinger theorem to the case T = 1.
So we have to find a probability measure Q ∼ P such that
dQ
∈ L∞ , S0 , S1 ∈ L1 (Q) and EQ [S1 | F0 ] = S0 .
dP
We will reduce the problem further. Suppose we have Q1 ∼ P with dQ1 /dP ∈ L∞
so that for X ≡ S1 − S0 we have X ∈ L1 (Q1 ) and EQ1 [X | F0 ] = 0. In this case
we define a new measure Q ∼ Q1 by dQ/dQ1 = c exp(−kS0 k), where c is the
proper normalizing constant. Clearly dQ/dQ1 ∈ L∞ and hence dQ/dP ∈ L∞ .
Furthermore, X ∈ L1 (Q) because X ∈ L1 (Q1 ) and dQ/dQ1 ∈ L∞ . But also S0 ∈
L1 (Q) and hence kS1 k ≤ kXk + kS0 k ∈ L1 (Q). Since dQ/dQ1 is F0 -measurable, it
is easily seen that EQ [X | F0 ] = 0, which implies that EQ [S1 | F0 ] = S0 , because S0
and S1 are in L1 (Q).

4. Some measure theory

For the next lemma we suppose that A = (Ai,j )i,j∈{1,...,d} is a matrix-valued map-
ping A : Ω → (Rd )d such that A(ω) is symmetric and non-negative definite for
 THE DALANG–MORTON–WILLINGER THEOREM 3

almost all ω ∈ Ω. We also suppose that Ai,j ∈ L1 (P) for all i, j ∈ {1, . . . , d}.
1/2 1/2
From linear algebra we know that |Ai,j | ≤ Ai,i Aj,j and hence the Cauchy–
Schwarz inequality
Pd implies that Ai,j ∈ L1 (P) for all i, j ∈ {1, . . . , d} if and only
if Trace(A) = i=1 Ai,i ∈ L (P). We are only interested in the case T = 1 and
1

hence we suppose that all random variables are F1 -measurable. This in particular
applies to A. We now define
Bi,j = EP [Ai,j | F0 ], i, j ∈ {1, . . . , d}. (4.1)
Lemma 4.2. For almost all ω the (d×d)-matrix B(ω) given by (4.1) is symmetric
and non-negative definite.
Proof. Clearly, for i 6= j we have Bi,j = Bj,i almost surely. For each q ∈ Rd with
rational coordinates we find that q 0 Bq = EP [q 0 Aq | F0 ] ≥ 0 almost surely. It follows
that on a set Ωq of measure one we have q 0 Bq ≥ 0. The density of the rational
numbers in R then implies that x0 Bx ≥ 0 for every x ∈ Rd on the countable
intersection \
Ωq .
q∈Rd
q rational
These properties imply that almost surely B is symmetric and non-negative definite.

Lemma 4.3. If B is a symmetric, non-negative definite, (d × d)-matrix-valued

random variable, then
Z ∞
P ≡ B e−λB dλ (4.4)
0

is well defined and represents the orthogonal projection of Rd onto the range of B.
It follows from the F0 -measurability of B that P is also F0 -measurable.
Proof.
R∞ If C is an arbitrary non-negative definite, symmetric (d × d)-matrix, then
−λC
0
Ce dλ represents the orthogonal projection onto its range. This is easily
seen using an orthonormal basis where C is diagonal. We only have to apply this
elementary result from linear algebra to each matrix B(ω).
Exercise 4.5. If we start with a non-negative definite, symmetric, integrable, ma-
trix-valued mapping A, then the mapping B in (4.1) depends on the probability
measure P we use to calculate the conditional expectations (4.1). Show that the
projection P from (4.4), however, is independent of this probability measure in the
sense that we obtain the same projection P if we replace P by an equivalent proba-
bility measure P1 such that A remains integrable. Also show that the matrix-valued
mapping P is the same for A as for the matrix-valued mapping PA , where PA de-
notes the orthogonal projection onto the range of A. This allows to get rid of the
integrability conditions and reduces the proof to the case of projection operators.
This exercise, although not difficult, requires a lot of verification.
Hints 4.6. Check that (I − P )B(I − P ) = 0 almost surely, hence
EP [(I − P )A(I − P )] = 0.
This means Trace((I − P )A(I − P )) = 0 and therefore (I − P )A(I − P ) = 0 almost
surely. Consider now EP1 [ · | F0 ].
4 F. DELBAEN

Lemma 4.7. With the notation of the previous lemmas we have that for every
F0 -measurable function α : Ω → Rd
(a) β ≡ P α is F0 -measurable,
(b) Bβ = Bα almost surely,
(c) β 0 Bβ = α0 Bα almost surely,
(d) (α − β)0 B(α − β) = 0 almost surely.
Proof. Apply the results from linear algebra for each ω ∈ Ω separately.
Lemma 4.8. With the same notations: if (αn )n∈N is a sequence of F0 -measurable
functions so that αn0 Bαn −→
a.s.
0, then βn ≡ P αn −→
a.s.
0.
Proof. We may apply the results from linear algebra for each ω separately.
P P
Corollary 4.9. With the same notations: if αn0 Bαn −→ 0, then βn −→ 0.
Proof. We have to show that from each subsequence (nk )k∈N we can select a further
subsequence (nkl )l∈N , so that βnkl −→
a.s.
0. This follows from Lemma 4.8 by selecting
0
(nkl )l∈N so that αnk Bαnkl −→ 0 as l → ∞.
a.s.
l

Lemma 4.10. Let (fn )n∈N be a sequence of F0 -measurable, non-negative, real-

valued random variables. Then there is a probability measure P1 ∼ P such that
dP1 /dP is uniformly bounded and EP1 [exp(λfn )] < ∞ for all λ ≥ 0 and all n ∈ N.
Proof. Take εn > 0 so small that P[εn fn ≥ 1] ≤ 2−n . From the Borel–Cantelli
lemma it follows that h ≡ supn∈N εn fn < ∞ almost surely. Define the probability
measure dP1 = c exp(−h2 ) dP, where c is the proper normalizing constant. Then
P1 ∼ P and
EP1 [exp(λfn )] = EP1 [exp(λε−1
n εn fn )]

≤ EP1 [exp(λε−1 −1 2 −2
n h)] = c EP [exp(λεn h − h )] ≤ c exp(λ εn /4),
2

where we used ax − x2 ≤ a2 /4 for all a, x ∈ R in the last step.

Lemma 4.11. If X : Ω → Rd is F1 -measurable, then there is a F0 -measurable
mapping P : Ω → (Rd )d such that
(a) Almost surely P (ω) is an orthogonal projection.
(b) For every F0 -measurable α : Ω → Rd
hα, Xi = hP α, Xi almost surely.
(c) For every sequence (αn )n∈N of F0 -measurable Rd -valued random variables:
P P
hαn , Xi −→ 0 if and only if P αn −→ 0.
Proof. From Exercise 4.5 we learnt that the projection operator P does not depend
on the choice of an equivalent probability measure, so we will change this measure
freely. Take P1 ∼ P so that X ∈ L2 (P1 ), which is possible by Lemma 4.10. Take
now A ≡ XX 0 and apply Lemma 4.2 to find B ≡ EP1 [A | F0 ]. By Lemma 4.3, the
orthogonal projection P (ω) onto the range of B(ω) is still F0 -measurable. This
proves (a).
We now show (b). Let α : Ω → Rd be F0 -measurable and suppose first that α is
bounded. Since B(α − P α) = 0 almost surely, EP1 [(α − P α)0 B(α − P α)] = 0. Since
B = EP1 [XX 0 | F0 ], we have
EP1 [(α − P α)0 XX 0 (α − P α)] = 0,
 THE DALANG–MORTON–WILLINGER THEOREM 5

which means
EP1 [hα − P α, Xi2 ] = 0,
hence hα, Xi = hP α, Xi almost surely. The general case is reduced to the bounded
case by replacing α by α/(kαk + 1).
P
It remains to prove (c). Suppose that hαn , Xi −→ 0 and let (nk )k∈N be any
subsequence. It suffices to show the existence of a further subsequence (nkl )l∈N so
P
that P αnkl −→ 0.
By selecting a further subsequence if necessary, we may assume in the following
that hαnk , Xi −→a.s.
0. Note that h ≡ supk∈N |hαnk , Xi| < ∞ almost surely. We may
replace P by an equivalent probability measure P2 with dP2 /dP = c exp(−h). Then
h has moments of all orders, in particular h ∈ L2 (P2 ). Exercise 4.5 shows that P
is not affected by such a replacement.
The choice of P2 implies that EP2 [hαnk , Xi2 ] → 0, hence EP2 [hP αnk , Xi2 ] → 0 by
part (b). Therefore, EP2 [(P αnk )0 B2 (P αnk )] → 0, where B2 ≡ EP2 [XX 0 | F0 ]. Since
P
(P αnk )0 B2 (P αnk ) ≥ 0 by Lemma 4.2, this implies (P αnk )0 B2 (P αnk ) −→ 0. But
Exercise 4.5 precisely means that P is almost surely also the orthogonal projection
P
onto the range of B2 , hence P αnk −→ 0 by Corollary 4.9. The reverse direction
follows from part (b).
Corollary 4.12. If (αn )n∈N is a sequence of F0 -measurable, Rd -valued functions.
If the sequence (hαn , Xi)n∈N is bounded in probability, i. e.,
∀ ε > 0 ∃ c > 0 ∀ n ∈ N : P[|hαn , Xi| > c] ≤ ε,
then the sequence (P αn )n∈N , is also bounded in probability.
Proof. Suppose not, then there is ε > 0 and a subsequence (nk )k∈N such that
P[kP αnk k > k] ≥ ε for all k ∈ N. (4.13)
Look now at the sequence hαnk /k, Xi. This sequence tends to zero in probability
P
and hence P αnk /k −→ 0 by Lemma 4.11(c). This is a contradiction to (4.13).
Lemma 4.14 (Stricker’s lemma). If X : Ω → Rd is F1 -measurable, then
K = { hα, Xi | α : Ω → Rd F0 -measurable } ⊂ L0 (Ω, F1 , P)
is closed for the convergence in probability.
P
Proof. If hαn , Xi −→ k, where (αn )n∈N is a sequence of F0 -measurable functions
with values in R , then
d

lim hαm − αn , Xi = 0 in probability.

m,n→∞

Lemma 4.11(c) shows that

lim P (αm − αn ) = 0 in probability.
m,n→∞

In other words, βn ≡ P αn defines a Cauchy sequence, which converges in probability

to a random variable β : Ω → Rd . Clearly, as the almost sure limit of a subsequence,
β can be chosen F0 -measurable. It satisfies
hβ, Xi = lim hβn , Xi = lim hP αn , Xi = lim hαn , Xi = k,
n→∞ n→∞ n→∞

where all limits are taken in probability and Lemma 4.11(b) is used.
6 F. DELBAEN

We recall that if f : Ω → R+ is F1 -measurable, then we can always define g =

EP [f | F0 ] so that EP [g1C ] = EP [f 1C ] ∈ R+ for all C ∈ F0 .
Lemma 4.15. If the random variable f : Ω → R+ and the constant c > 0 are so
that P[f > c] ≥ δ, then for g = EP [f | F0 ] we have P[g > cδ/2] ≥ δ/4.
Proof. Let k ≡ min{f, 2c} and l ≡ EP [k | F0 ]. Clearly g ≥ l and P[k > c] ≥ δ. So it
suffices to prove the lemma for the bounded function k and for δ ≥ 0. Since l ≤ 2c,
cδ
EP [l ] = EP [l 1{l≤cδ/2} ] + EP [l 1{l>cδ/2} ] ≤
+ 2c P[l > cδ/2].
2
On the other hand, EP [l ] = EP [k] ≥ cδ. Putting together the inequalities yields the
desired estimate P[l > cδ/2] ≥ δ/4.
Lemma 4.16 (Schachermayer’s lemma). Let f1 , . . . , fN : Ω → R+ be measurable
PN
and λ1 , . . . , λN ∈ R+ with i=1 λi = 1. Suppose that P[fi > c] ≥ δ for all
i ∈ {1, . . . , N }. Then
X
N
g= λi fi satisfies P[g > cδ/2] ≥ δ/4.
i=1

Proof. Deduce this from the above Lemma 4.15 by considering the product space
Ω0 ≡ Ω × {1, . . . , N } with F10 ≡ F1 ⊗ P({1, . . . , N }), F00 ≡ F0 ⊗ {∅, {1, . . . , N }} and
P0 (A × {j}) ≡ P(A)λj for A ∈ F1 and j ∈ {1, . . . , N }. Check that g = EP0 [f | F00 ],
PN
where f (ω, j) ≡ i=1 fi (ω)1{i} (j) on Ω0 .
Lemma 4.17. Let f1 , . . . , fN : Ω → R+ be F1 -measurable and λ1 , . . . , λN ∈ R+
PN
with i=1 λi = 1. Suppose that P[fi > c] ≥ δ for all i ∈ {1, . . . , N } for some
constant c > 0. Then g = EP [λ1 f1 + · · · + λN fN | F0 ] satisfies
· ¸
cδ 2 δ
P g> ≥
16 16
Proof. By Schachermayer’s lemma, P[λ1 f1 + · · · + λN fN > cδ/2] ≥ δ/4. Hence
· ¸
cδ δ 1 δ
P g> ≥
2 42 16
by a simple application of Lemma 4.15.
Lemma 4.18. Let (fn )n∈N be a sequence of non-negative measurable functions
such that the set C = conv(fn , n ∈ N) of finite convex combinations is bounded in
L0 (Ω, F1 , P). This means
∀ ε > 0 ∃ c > 0 ∀ h ∈ C : P[h > c] ≤ ε. (4.19)
P
Then there exists g : Ω → R+ as well as gn ∈ conv(fm , m ≥ n) such that gn −→ g.
Proof. Take ϕ(x) ≡ 1 − e−x , a strictly concave function. Observe, for later use,
that for all x, y ∈ R
³ x + y ´ ϕ(x) + ϕ(y) ³ 1 1 ´
|x−y| |x−y|/2
ϕ − = + e −e e− max{x,y}
2 2 2 2 (4.20)
|x − y|2 − max{x,y}
≥ e
8
 THE DALANG–MORTON–WILLINGER THEOREM 7

by series expansion. For each n ∈ N define

an = sup{ E[ϕ(g)] | g ∈ conv(fm , m ≥ n) }.
Since 0 ≤ ϕ ≤ 1 on R+ , we have 0 ≤ an ≤ 1. Also the sequence (an )n∈N is
decreasing. For each n ∈ N take gn ∈ conv(fm , m ≥ n) such that
EP [ϕ(gn )] ≥ an − 2−n .
I claim that (gn )n∈N converges in probability to a function g : Ω → R+ . Indeed,
for m ≥ n we have
gm + gn
∈ conv(fn , fn+1 , . . . )
2
and hence h ³ g + g ´i
m n
EP ϕ ≤ an .
2
On the other hand by (4.20),
h ³ g + g ´i 1 ¡ ¢ 1 £ ¤
≥ EP [ϕ(gm )] + EP [ϕ(gn )] + EP |gm − gn |2 e− max{gm ,gn } .
m n
EP ϕ
2 2 8
It therefore follows that
EP [|gm − gn |2 e− max{gm ,gn } ] → 0 as m, n → ∞.
Given ε > 0 take c > 0 so that P[gn > c] ≤ ε for all n ∈ N, this is possible because
every gn is in C and C satisfies (4.19). Since
P[|gm − gn | > ε] ≤ P[gm > c] + P[gn > c] + P[max{gm , gn } ≤ c; |gm − gn | > ε],
Tchebychev’s inequality then yields
ec £ ¤
P[|gm − gn | > ε] ≤ 2ε + 2
EP |gm − gn |2 e− max{gm ,gn } .
ε
This shows that
lim P[|gm − gn | > ε] = 0
m,n→∞
in probability and the lemma is proved.
Remark 4.21. We cannot drop the positivity of the functions. Indeed, take an i. i. d.
sequence (fn )n∈N of 1-stable variables, meaning that they have the characteristic
function E[exp(itfn )] = exp(−|t|) for all t ∈ R. (Exercise: Check that every fn has
a Cauchy distribution.) An easy exercise shows that each element g in C has the
same distribution as the elements fn . Hence the set C is bounded in probability,
but no sequence gn ∈ conv(fm , m ≥ n) can converge in probability. Indeed, the
limit g ≡ limn→∞ gn would have the same distribution as every fn but it also has
to be degenerate by Kolmogorov’s 0–1-law.
Corollary 4.22. If (fn )n∈N is a sequence of measurable functions and
h ≡ sup |fn | < ∞ a. s.
n∈N

then there is a measurable function g : Ω → R and a sequence of convex combina-

P
tions gn ∈ conv(fm , m ≥ n) such that gn −→ g.
Proof. Define Fn = h + fn ≥ 0. Since Fn ≤ 2h we certainly have for each k ∈ C ≡
conv(Fn , n ∈ N) that k ≤ 2h. It follows that C is bounded in probability. The
corollary now follows from the previous lemma.
8 F. DELBAEN

Let us mention two additional results, which we don’t need in the following:
Corollary 4.23. If (fn )n∈N is a sequence of measurable functions such that
c ≡ sup EP [|fn |] < ∞,
n∈N
then there is a measurable function g : Ω → R as well as a sequence of convex
P
combinations gn ∈ conv(fm , m ≥ n) such that gn −→ g.
Proof. If fn ≥ 0 and EP [fn ] ≤ c, then for all h ∈ C ≡ conv(fn , n ∈ N) we have
EP [h] ≤ c. The set C is bounded in L1 (P) and hence bounded in probability. This
shows the corollary for non-negative functions as a consequence of Lemma 4.18.
The general case is a repeated
P application of the non-negative case. Indeed, first
+
find convex combinations m≥n λnm fm that converge almost surely. Now apply
P −
the first part to the sequence of functions qn ≡ m≥n λnm fm . This gives convex
P n
combinations m≥n µm qm that converge almost surely. Since convex P combinations
n +
of
P convex combinations are
Pmconvex combinations, we find that m≥n νm fm and
n −
ν
m≥n m m f with ν n
m ≡ µn k
λ
k=n k m both converge almost surely. (Why did we
use almost sure convergence and not just convergence in probability?)
The above corollary is valid in a much more refined way as shown by Komlós.
Theorem 4.24 (Komlós’ Theorem [4]). If (fn )n∈N is a bounded sequence in the
space L1 (P), then there is a function f ∈ L1 (P) and a subsequence (fnk )k∈N such
that for every subsequence (nkl )l∈N we have
1 X
N
fnkl −→
a.s.
f as N → ∞.
N
l=1

It is an excellent exercise in real analysis to study and especially to appreciate

the non-trivial proof of this theorem.

5. Topological version of the no-arbitrage condition

We use the same notation as in the previous section.
Theorem 5.1. If (αn )n∈N is a sequence of F0 -measurable Rd -valued functions
such that
hαn , Xi− −→
P
0
and if the no-arbitrage condition holds for X, then also
P
hαn , Xi+ −→ 0.
Proof. By Lemma 4.11(b), we may assume αn = P αn in the following.
Since we may take subsequences, we suppose that hαn , Xi− −→a.s.
0 and by re-
placing the measure as in the proof of Lemma 4.11(c) if necessary, we may even
suppose that
h ≡ sup hαn , Xi− ∈ L1 (P).
n∈N

Let A ≡ {supn∈N kαn k = ∞} ∈ F0 . For each n ∈ N and m ≥ n we define

An,n = {kαn k > n} ∩ A
and
An,m = {kαn k ≤ n, . . . , kαm−1 k ≤ n, kαm k > n} ∩ A.
 THE DALANG–MORTON–WILLINGER THEOREM 9

Clearly A is the disjoint union of (An,m )m≥n . For each n we now define
X αm
gn = 1A + αn 1Ac . (5.2)
kαm k n,m
m≥n

Since the sets form a countable partition of Ω, we have that

X hαm , Xi−
hgn , Xi− = 1An,m + hαn , Xi− 1Ac
kαm k
m≥n
and hence
h
hgn , Xi− ≤ 1A + hαn , Xi− 1Ac .
n
It follows that
hgn , Xi− −→
a.s.
0. (5.3)
But we also have kgn k ≤ supm∈N kαm k < ∞ on Ac an kgn k = 1 on A and therefore
sup kgn k < ∞ almost surely. (5.4)
n∈N

Suppose now that hgn , Xi+ does not tend to zero in probability and suppose by
selecting a subsequence if necessary, that for some c > 0 and for all n ∈ N
P[hgn , Xi+ > c] ≥ δ > 0. (5.5)
Due to (5.3) we can find n0 ∈ N so that for all n ≥ n0
· ¸
− cδ δ
P sup hgm , Xi > < . (5.6)
m≥n 4 8
If a1 , . . . , an ∈ R and µ1 , . . . , µn ∈ [0, 1] with µ1 + · · · + µn = 1, then we have
µX n ¶+ X n X n Xn X n
− − −
µi ai ≥ µi ai = µi ai −
+
µi ai ≥ i − max{a1 , . . . , an }.
µi a+
i=1 i=1 i=1 i=1 i=1
P
Now take any kn ∈ conv(gm , m ≥ n). We can write kn = m≥n λnm gm with λnm ≥ 0
P
satisfying m≥n λnm = 1. Then we get for every n ≥ n0
· ¸ ·X ¸
cδ − cδ
P hkn , Xi >
+
≥P λm hgm , Xi − sup hgm , Xi >
n +
4 m≥n 4
m≥n
·X ¸ · ¸
cδ − cδ
≥P λm hgm , Xi >
n +
− P sup hgm , Xi > (5.7)
2 m≥n 4
m≥n
δ δ δ
≥
− = ,
4 8 8
where we used (5.5), Schachermayer’s lemma (Lemma 4.16) and (5.6) for the last
estimate.
Due to (5.4), we may apply Corollary 4.22 repeatedly for every of the d dimen-
P
sions to find kn ∈ conv(gm , m ≥ n) so that kn −→ k for some F0 -measurable
function k : Ω → R . By convexity and (5.3) we have
d

hkn , Xi− ≤ sup hgm , Xi− −→

a.s.
0.
m≥n
Therefore
hk, Xi = lim hkn , Xi = lim hkn , Xi+ in probability.
n→∞ n→∞
10 F. DELBAEN

But then we also have from (5.7)

P[hk, Xi ≥ cδ/4] ≥ δ/8
P
and this contradicts the no-arbitrage condition. We therefore have hgn , Xi+ −→ 0.
P
We deduce from Lemma 4.11(c) and gn ∈ Range(P ) that kgn k −→ 0. This
P
shows two things. First P[A] = 0 since kgn k = 1 on A and second kαn k −→ 0 since
+ P
gn = αn on A . This clearly implies hαn , Xi −→ 0.
c

Theorem 5.8. Suppose that X satisfies the no-arbitrage condition. If (αi )i∈I is
a collection of F0 -measurable Rd -valued functions and if { hαi , Xi+ | i ∈ I } is
bounded in probability, then { hαi , Xi− | i ∈ I } is bounded in probability, too. It
follows that { hαi , Xi | i ∈ I } is bounded in probability and hence { kP αi k | i ∈ I }
(with the orthogonal projection P as in Lemma 4.11) is also bounded in probability.
Proof. Suppose that { hαi , Xi− | i ∈ I } is not bounded and select ε > 0 and a
P
sequence (αn )n∈N so that P[hαn , Xi− > n] ≥ ε. Since hαn /n, Xi+ −→ 0, we have
− P
by the previous theorem hαn /n, Xi −→ 0, a contradiction. The last part of the
theorem is proved in a similar way using Lemma 4.11(c).
The next two theorems are needed for Schachermayer’s proof in Section 6;
Rogers’ proof in Section 7 works without them.
Theorem 5.9. Suppose that X satisfies the no-arbitrage condition. Then K − L0+
is closed in probability.
Proof. Take (αn )n∈N F0 -measurable and hn ≥ 0 so that fn ≡ hαn , Xi − hn → f
in probability. By selecting a subsequence we may suppose that this sequence
even converges almost surely. Clearly this implies hαn , Xi− ≤ fn− and hence
supn∈N hαn , Xi− ≤ k for some measurable function k. It is clear that all the ele-
ments of the set
{ hβ, Xi− | β ∈ conv(αn , n ∈ N)},
remain bounded by the same function k. From Theorem 5.8 it then follows that
the set
{ hβ, Xi | β ∈ conv(αn , n ∈ N) } = conv(hαn , Xi, n ∈ N)
is also bounded in probability. Lemma 4.18 applied to conv(hαn , Xi + k, n ∈ N)
then implies the existence of a sequence βn ∈ conv(αm , m ≥ n), n ∈ N, say
X
βn = λnm αm ,
m≥n

so that hβn , Xi converges in probability to a random variable g. Stricker’s lemma

(Lemma 4.14) implies that g = hα, Xi for some F0 -measurable α : Ω → Rd . We
now take the same convex combinations for the functions hn , i. e.,
X
ln = λnm hm .
m≥n
P
m≥n λm fm , n ∈ N,
n
Since (fn )n∈N converges almost surely to f , the sequence
converges also almost surely to f . It follows that (ln )n∈N converges in probability
to a, necessary non-negative, random variable h. It is then clear that f = hα, Xi−h
is in K − L0+ .
 THE DALANG–MORTON–WILLINGER THEOREM 11

Theorem 5.10 (Yan’s separation theorem). If G is a closed convex cone in L1

such that G ∩ L1+ = {0} and if G contains the cone of non-positive functions, then
there is a strictly positive function h ∈ L∞ such that E[gh] ≤ 0 for all g ∈ G.
Proof. The proof is a combination of the Hahn–Banach Theorem and an exhaustion
argument. For each n ∈ N we define Zn as the set of all functions z such that
E[z] ≥ 1/n and 0 ≤ z ≤ 1. Clearly G ∩ Zn = ∅. Because Zn is a weakly compact
convex subset of L1 (prove this), we may separate G and Zn strictly. This means
that there is a function hn ∈ L∞ = (L1 )∗ so that for all g ∈ G and all z ∈ Zn :
E[hn g] < E[hn z]. Because G itself is a cone and contains the negative cone, we
obtain that hn ≥ 0 and E[hn g] ≤ 0 for all g ∈ G, hence E[hn z] > 0 for all z ∈ Zn .
The last expression shows that E[hn 1A ] > 0 for every set
P A such that P[A] ≥ 1/n.
We may suppose that khn k∞ ≤ 2−n . If we define h = n∈N hn , then we find that
h has all the desired properties.
We can now proceed in two ways. The first approach is Schachermayer’s proof,
the second is Rogers’ proof. We will give both.

6. Schachermayer’s proof
Proof of Theorem 2.2 for T = 1. Take Q ∼ P so that dQ/dP ∈ L∞ and kXk ∈
L1 (Q). From Theorem 5.9 it now follows that G ≡ L1 (Q) ∩ (K − L0+ ) is closed in
L1 (Q) and from the no-arbitrage condition it follows that G ∩ L1+ (Q) = {0}. We
now apply Yan’s separation theorem in order to obtain a function ϕ ∈ L∞ with
ϕ > 0 almost surely and
∀ f ∈ K ∩ L1 (Q) : EQ [ϕf ] ≤ 0.
Since K ∩ L1 (Q) is a vector space, this implies EQ [ϕf ] = 0 for all f ∈ K ∩ L1 (Q).
We normalize ϕ so that dQ1 = ϕ dQ defines a probability measure. Clearly Q1 ∼ P,
dQ1 /dP ∈ L∞ and kXk ∈ L1 (Q1 ). Obviously for all A ∈ F0 and all coordinates
i ∈ {1, . . . , d},
1A X i ∈ K ∩ L1 (Q).
We therefore have EQ1 [1A X i ] = 0 for all A ∈ F0 , which means EQ1 [X | F0 ] = 0.

7. Rogers’ proof
Proof of Theorem 2.2 for T = 1. We first change to the probability measure Q ∼ P
given by dQ/dP = exp(−kXk2 )/EP [exp(−kXk2 )]. Then EQ [exp(kXk2 )] < ∞. For
each F0 -measurable α : Ω → Rd we define
£ ¤
ϕ(α) = EQ exp hα, Xi .
The function ϕ takes values in R+ . Since ϕ(0) = 1, we have that
γ ≡ inf{ ϕ(α) | α F0 -measurable, Rd -valued } < ∞.
We want to show that the infimum γ is attained. Take a sequence (αn )n∈N so
that ∞ > ϕ(αn ) → γ. We may of course suppose that αn = P αn as Lemma 4.11(b)
shows. Using x+ ≤ ex for x ∈ R, the convexity of the exponential function and
Jensen’s inequality, it follows that for each n ∈ N and each β ∈ conv(αm , m ≥ n)
we have that
EQ [hβ, Xi+ ] ≤ ϕ(β) ≤ sup ϕ(αm ) = γ + εn ≤ γ + sup εm < ∞, (7.1)
m≥n m∈N
12 F. DELBAEN

where εn & 0. Hence the set { hβ, Xi+ | β ∈ conv(αn , n ∈ N) } is bounded in L1 (Q)
and therefore bounded in probability. Theorem 5.8 then shows that the convex set
{ hβ, Xi | β ∈ conv(αn , n ∈ N) } = conv(hαn , Xi, n ∈ N)
is bounded in probability. The estimate (7.1) also shows that
conv(hαn , Xi+ , n ∈ N)
is bounded in L1 (Q) and therefore bounded in probability. Since
hαn , Xi− = hαn , Xi+ − hαn , Xi,
we have
conv(hαn , Xi− , n ∈ N) ⊂ conv(hαn , Xi+ , n ∈ N) + conv(−hαn , Xi, n ∈ N),
hence conv(hαn , Xi− , n ∈ N) is bounded in probability, too. By Lemma 4.18 we
can therefore
P find convex combinations,
P defined by non-negative
P sequences (λnm )m≥n
with m≥n λnm = 1, such that m≥n λnm hαm , Xi+ and m≥n λnm hαm , Xi− both
P
converge in probability. If we put δn ≡ m≥n λnm αm ∈ conv(αm , m ≥ n), then
we obtain that (hδn , Xi)n∈N converges in probability. By Stricker’s lemma (Lemma
P
4.14) there exists an F0 -measurable δ : Ω → Rd such that hδn , Xi −→ hδ, Xi. By
Lemma 4.11(b) we may assume that δ = P δ. Since δn = P δn , Lemma 4.11(c)
P
implies that δn −→ δ. By passing to a subsequence if necessary, we may assume
δn −→ δ. Of course, by Fatou’s lemma and (7.1),
a.s.
£ ¤ £ ¤
ϕ(δ) = EQ exp hδ, Xi ≤ lim inf EQ exp hδn , Xi ≤ γ + lim εn = γ.
n→∞ n→∞

It follows that ϕ(δ) = γ.

Since δ minimizes ϕ, we can now use a variational argument. For each ε ∈ [−1, 1],
each A ∈ F0 , each n ∈ N and each coordinate i ∈ {1, . . . , d}, we put
δA,i,n (ε) = δ + εei 1A∩{kδk<n} ,
where ei ∈ Rd is the vector whose i-th coordinate is 1, and the other coordinates
are 0. Since hδA,i,n (ε), Xi ≤ (n + 1)kXk on {kδk < n}, we may differentiate under
the expectation and obtain that for all n ∈ N, i ∈ {1, . . . , d} and all A ∈ F0
d ¯ £ ¤
0= ϕ(δA,i,n (ε))¯ε=0 = EQ ehδ,Xi X i 1A∩{kδk<n} . (7.2)
dε
Put £ ¤ £ ¤
kn ≡ 1 + EQ ehδ,Xi kXk 1{n−1≤kδk<n} ≤ 1 + EQ e(n+1)kXk < ∞
and define
X
ehδ,Xi 1{n−1≤kδk<n} kn−1 e−n
2
/4
ψ= . (7.3)
n∈N

What are the properties of ψ? First EQ [ψkXk] < ∞, because

£ ¤ X £ ¤ X
EQ ehδ,Xi kXk 1{n−1≤kδk<n} kn−1 e−n /4 ≤ e−n /4 < ∞.
2 2
EQ ψkXk ≤
n∈N n∈N
£ ¤
Second, for all i ∈ {1, . . . , d} and all A ∈ F0 , we have EQ ψ1A X i = 0 due to (7.2)
and the above integrability of ψX. Third, ψ > 0 almost surely by (7.3). Finally,
dQ
ξ≡ψ
dP
 THE DALANG–MORTON–WILLINGER THEOREM 13

is in L∞ , because
ehδ,Xi 1{n−1≤kδk<n} kn−1 e−n /4 −kXk2
= e−(kXk−n/2) ≤ 1.
2 2
/4−kXk2 2
e ≤ enkXk−n
If we normalize ξ so that EP [ξ] = 1 and define Q1 so that dQ1 /dP = ξ, we have
dQ1 /dP ∈ L∞ and EQ1 [X | F0 ] = 0. This completes the proof.

References
[1] R. C. Dalang, A. Morton, and W. Willinger, Equivalent martingale measures and no-arbitrage
in stochastic securities market models, Stochastics Stochastics Rep. 29 (1990), no. 2, 185–201.
[2] F. Delbaen and W. Schachermayer, The fundamental theorem of asset pricing for unbounded
stochastic processes, Math. Ann. 312 (1998), 215–250.
[3] Yu. M. Kabanov and D. O. Kramkov, Large financial markets: asymptotic arbitrage and con-
tiguity, Theory Probab. Appl. 39 (1994), no. 1, 182–187.
[4] J. Komlós, A generalization of a problem of Steinhaus, Acta Math. Acad. Sci. Hungar. 18
(1967), 217–229.
[5] L. C. G. Rogers, Equivalent martingale measures and no-arbitrage, Stochastics Stochastics
Rep. 51 (1994), no. 1–2, 41–49.
[6] W. Schachermayer, A Hilbert space proof of the fundamental theorem of asset pricing in finite
discrete time, Insurance Math. Econom. 11 (1992), no. 4, 249–257.
[7] A. N. Shiryaev, Essentials of Stochastic Finance, World Scientific, London, 1999.

Departement Mathematik, ETH Zentrum, CH-8092 Zürich, Switzerland

E-mail address: [email protected]
URL: http://www.math.ethz.ch/~delbaen