A short introduction to arbitrage theory and pricing in

mathematical finance for discrete-time markets with or

without friction.
Emmanuel Lépinette

To cite this version:

Emmanuel Lépinette. A short introduction to arbitrage theory and pricing in mathematical finance
for discrete-time markets with or without friction.. Master. France. 2019. �cel-02125685v2�

HAL Id: cel-02125685

https://hal.archives-ouvertes.fr/cel-02125685v2
Submitted on 3 Jul 2019

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 A short introduction to arbitrage
theory and pricing in mathematical
finance for discrete-time markets with
or without frictions
Emmanuel LEPINETTE1
1
Paris Dauphine university, PSL research university, Ceremade,
CNRS, UMR, Place du Maréchal De Lattre De Tassigny,
75775 Paris cedex 16, France and
Gosaef, Faculty of Sciences of Tunis, Tunisia.
Email: [email protected]
Abstract: In these notes, we first introduce the theory of arbitrage and
pricing for frictionless models, i.e. the classical theory of mathematical
finance. The main classical results are presented, i.e. the characterization
of absence of arbitrage opportunities, based on convex duality, and dual
characterizations of super-hedging prices are deduced. We then present
financial market models with proportional transaction costs. We discuss
no arbitrage conditions and characterize super-hedging prices as in the
frictionless case. Another approach based on the liquidation value con-
cept is finally introduced.
Keywords and phrases: Financial market models, No-arbitrage con-
dition, Pricing, European options, Liquidation value, Transaction costs.
2000 MSC: 60G44, G11-G13.

The following lectures have been written for the workshop organized from
Monday the 22th to the 26th of April 2019 by the laboratory Latao of the
Faculty of Sciences of Tunis and by the reasearch group Gosaef which gathers
researchers working on order structures, mathematical finance and mathe-
matical economics. These notes are devoted to graduate students and any-
one who wants to be initiated to arbitrage theory. The author thanks the
organizers, in particular Amine Ben Amor for his hearty welcome.

1
 / 2

1. Markets without frictions

1.1. Introduction

We consider a discrete-time stochastic basis (Ω, (Ft )t=0,1,··· ,T , P ). The set Ω

is the space of all possible states of the financial market we consider on the
period [0, T ]. A state ω ∈ Ω may be complicated; it may include risky asset
prices but also preferences of agents acting on the market. For every t, we
suppose that Ft is a σ-algebra, which is supposed to be complete, i.e. contains
the negligible sets for the probability measure P . Recall that, by definition,
the elements of Ft are subsets of Ω and we have the following properties:

(i) Ω, ∅ ∈ Ft ,
(ii) Ft ∈ Ft implies that Ftc := Ω \ Ft ∈ Ft , S
(iii) For all countable family (Ftn )n≥1 of Ft , n Ftn , n Ftn ∈ Ft .
T

Notice that (Ft )t=0,1,··· ,T is called a filtration in the sense that, for all t < u,
Ft ⊆ Fu . The σ-algebra Ft models the information available at time t.
Example Let us consider a financial market composed of d exchangeable
assets whose prices are given at time t by the vector St = (St1 , · · · , Std ). We
define
Ft = σ (Su : u ≤ t) , t ≥ 0,
as the smallest σ-algebra making the mappings Su : ω 7→ Su (ω), u ≤ t,
measurable with respect to Ft . Such a σ-algebra exists as an intersection
of any family of σ-algebras is a σ-algebra. We may verify that (Ft )t≥0 is a
filtration.
In finance, we generally suppose that the filtration (Ft )t≥0 is complete,
i.e. F0 contains the negligible sets for P . Actually, the classical case is to
consider F0 as the smallest σ-algebra containing the negligible sets. We may
show that X0 is F0 -measurable if and only if there exists a constant c such
that P (X = c) = 1, i.e. X = c a.s. (almost surely).
The family of random variables (Xt )t≥0 is said to be a stochastic pro-
cess adapted to the filtration (Ft )t≥0 if, for all t ≥ 0, Xt : Ω → Rd is
Ft -measurable. This means that, for all B in the Borel σ-algebra B(Rd ),
Xt−1 (B) ∈ Ft . Notice that, if Ft is the information available at time t on the
market, the Ft -measurability means that Xt is observed at time t.
 / 3

In the following, we describe a portfolio process by the quantities held by

an agent acting on the market. Precisely, a strategy θ̂ = (θ0 , θ) is such that
θt0 is the quantity at time t invested in some non risky asset whose price is
St0 at time t while θt = (θt1 , · · · , θtd ) is the vector of quantities θti invested in
risky asset number i = 1, · · · , d whose price is Sti at time t.

The asset S 0 is said non risky at time t if var(St0 ) = 0, i.e. there is no

uncertainty about the future prices St0 : out of a negligible set N , we have
St0 (ω) = St0 , for all ω ∈ N c . In particular, we know by advance the prices
(St0 )t∈[0,T ] . A classical modeling of S 0 is given by the deterministic dynamics
0
St+1 − St0
= r,
St0

where the interest rate r is a constant and S00 is given.

On the contrary, we say that the asset (St )t∈[0,T ] is risky at time t if
var(St ) > 0. This means that the mapping ω 7→ St (ω) is not constant. There-
fore, we do not know by advance the future values of St (ω) as it depends on
the market state ω ∈ Ω. A classical example is to suppose that

∆St+1 := St+1 − St = µSt + σSt Gt+1

where (Gt )t=1,··· ,T is a family of i.i.d. random variables with common distri-
bution N (0, 1) and σ, µ are two constants. This means that the returns are
normally distributed. Notice that when σ = 0, S is deterministic, i.e. is not
risky.
Remark 1.1. There exists a continuous version of the model. The non risky
asset satisfies the continous time dynamics

dSt0 = rSt0 dt, S00 = 1.

The solution is given by St0 = ert as it is the solution of the o.d.e. (St0 )0 =
dSt0
dt
= rSt0 . Notice that
 0
St+dt − St0

r = lim /dt.
dt→0 St0
This means that r is interpreted as an instantaneous interest rate.
 / 4

The risky asset is given by the Black and Scholes model, i.e. the price S
follows the dynamics:
dSt = µSt dt + σSt dWt , S0 is given.
The stochastic process W is supposed to be a (standard) Brownian motion,
i.e. W satisfies the following conditions:
1 For all t, Wt is Ft -measurable and W0 = 0.
2 With probability 1, the trajectories t 7→ Wt (ω), ω ∈ Ω, are continuous.
3 For all u < t, Wt − Wu is independent of Fu .
4 If t4 − t3 = t2 − t1 , then Wt4 − Wt3 and Wt2 − Wt1 are equally distributed
as N (0, t4 − t3 ) 1 .
Let us interpret the dynamics of (St )t∈[0,T ] . We introduce the discrete dates
tni = Tn i, i = 0, 1, · · · , n. We have ∆tni := tni − tni−1 = T /n. As n → ∞,
∆Stni+1 := Stni+1 − Stni ' µStni ∆tni + σStni ∆Wtni+1 , i ≥ 1,
p
where ∆Wti = T /n Gi with (Gi )i=1,··· ,n a family of i.i.d. random variables
with common distribution N (0, 1). This property is directly deduced from the
definition of W . Notice that, when σ = 0, St = S0 eµt is deterministic, i.e.
it is non risky. If σ > 0, the Black and Scholes model supposes that the
log-returns log(Stni+1 /Stni ) are normally distributed. Indeed, we may show that
σ∆W
t n +(µ−σ 2 /2)∆tn
i+1
Stni+1 = Stni e i+1 . The coefficient σ is called the volatility. The
larger σ is, the further could be St from the deterministic trajectory S0 eµt .
Actually, we may show that E(St ) = S0 eµt , t ≥ 0.
For readers interested in stochastic calculus, very good notes by Jeanblanc
M. are available in french [13] but also by Lamberton D. and Lapeyre B. in
english [19].

1.2. Financial market without frictions

A financial market is said without frictions if there is no transaction costs

when selling or buying risky assets. For a strategy θ̂ = (θ0 , θ), we define the
liquidation value at time t:
d
X
Vt = Vtθ̂ = θt0 St0 + θt · St = θt0 St0 + θti Sti .
i=1
1
It is a Gaussian distribution with mean 0 and variance t4 − t3
 / 5

The stochastic process V = V θ̂ is called the portfolio process associated to θ̂.

Here, θii is allowed to be non positive (short position), which corresponds to a
debt in the asset number i. The formulation above supposes that there is no
transaction costs. Indeed, otherwise, when selling or buying risky assets to
liquidate the positions given by θ̂, there should be a cost ct > 0 to withdraw
from the liquidation value.
In the following, we denote by L0 (Rn , Ft ), n ≥ 1, the set of all Ft -
measurable random variables with values in Rn .
Definition 1.2. An European option is a contract between two agents (seller
and buyer) allowing the option holder (buyer) to get a terminal wealth ξT
(called the payoff ) from the seller at some fixed maturity T > 0. Such a
contract is sold at time t = 0 at some price. The classical example is the so-
called Call option, i.e. such that ξT = (ST − K)+ where K ≥ 0 is a constant,
which is called the strike. This means that, if ST ≥ K, the option holder get
ST − K and 0 otherwise. Such a contract corresponds to the possibility for
the holder to buy the underlying asset S at price K at time T instead of the
real price ST . This is clearly interesting only if ST ≥ K, in which case the
gain of the transaction is ST − K ≥ 0.
A fundamental problem in mathematical finance is to determine a price
for an European option. To do so, we introduce the following definitions.
Definition 1.3. A portfolio process Vt = Vtθ̂ , t = 0, · · · , T , is said self-
financing if
0
θt−1 St0 + θt−1 · St = θt0 St0 + θt · St , t = 1, · · · , T.

For any stochastic process X, we introduce the following notations: ∆Xt =

Xt − Xt−1 for t ≥ 1 and the discounted value X̃t = Xt /St0 . We may easily
show the following:
Lemma 1.4. A portfolio process Vt = Vtθ̂ , t = 0, · · · , T , is said self-financing
0
if and only if ∆Vt = θt−1 ∆St0 + θt−1 · ∆St , t = 1, · · · , T .
Lemma 1.5. A portfolio process Vt = Vtθ̂ , t = 0, · · · , T , is said self-financing
if and only if ∆Ṽt = θt−1 · ∆S̃t , t = 1, · · · , T .
In the following, we denote by RT0 the set of all discounted terminal values
ṼT of self-financing portfolio processes starting from the initial value
 / 6

Ṽ0 = V0 = 0. Writing ṼT = Ṽ0 + Tt=1 ∆Vt , we have

P

( T )
X
RT0 = θt−1 · ∆S̃t : θt ∈ L0 (R, Ft ), t = 0, · · · , T − 1 .
t=1

We also introduce the set of super-hedgeable claims AT0 = RT0 − L0 (R+ , FT )

we may obtain from a zero initial endowment, i.e. ξT ∈ AT0 if and only if
there exists VT ∈ RT0 such that VT ≥ ξT a.s. In that case, we say that V
super-replicates (or super-hedges) ξT at time T . Moreover, if VT = ξT a.s. we
say that V replicates ξT .
Definition 1.6. A price for the payoff ξT is any initial value V0 of a self-
financing portfolio process V such that VT ≥ ξT a.s. We denote by P(ξT ) the
set of all prices for ξT .

1.3. One step financial market: T = 1 and d = 1

In this section, we consider the simplest case where T = 1 and d = 1. It

suffices to understand the main ideas in that case to extend them to the
generalncase, up to someotechnical difficulties. Here, observe that we have
RT0 = θ0 · ∆S̃1 : θ0 ∈ R . A price for the payoff ξ1 ∈ L0 (R, F1 ) is a value
p0 such that Ṽ1 = p0 + θ0 · ∆S̃1 ≥ ξ˜1 a.s. for some θ0 ∈ R. This is equivalent
to say that ξ˜1 − p0 ∈ A10 . Therefore, the question is whether ξ˜1 − p0 ∈ A10 or
not: ξ˜1 − p0 ∈/ A10 . The last condition may be related to a convex separation
problem as A10 is actually a convex cone. This is why, we prefer for A10 to
be closed. The natural problem is to find a condition under which this is the
case.
Notice that, if d = 1, S0 is a price for ξ1 = (S1 − K)+ , K ≥ 0. Indeed,
it suffices to follow the buy and hold strategy θ0 = (0, 1), i.e. buying one
unit of the risky asset at price S0 . At T = 1, we obtain Ṽ1 = S0 + θ0 ∆S̃1 =
S0 + (S̃1 − S0 ) = S̃1 . So, V1 = S1 ≥ (S1 − K)+ = ξ1 .
It is traditional to suppose the closedness of AT0 to characterize the super-
hedging prices. Of course, we need to precise the topology we use. In partic-
ular, L0 (Rd , FT ) is endowed with the topology of the convergence in proba-
bility, so that it is a metric space: d0 (X, Y ) = E (|X − Y | ∧ 1). The spaces
Lp (Rd , FT ), p ∈ [1, ∞] are endowed with the usual norms kXkp := (E|X|p )1/p
if p < ∞ and kXk∞ is the usual norm for bounded random variables X of
L∞ . With p1 + 1q , the topological dual of Lp (Rd , FT ) is Lq (Rd , FT ) for p ≤ 1
 / 7

but the dual of L∞ (Rd , FT ) is larger than L1 (Rd , FT ) except if we endow

L∞ (Rd , FT ) with the σ(L∞ , L1 ) topology.
Closedness of A10 in L0 (R, F1 ) for d = 1.
We denote by S the single risky asset. Let X n = θ0n ∆S1 − + 1
n ∈ A0 where
(θ0n )n≥1 is a sequence of R and (+ 0
n )n≥1 is a sequence in L (R+ , F1 ). Suppose
that X n → X a.s.
1st case: supn |θ0n | < ∞. In that case, the sequence (θ0n )n≥1 belongs to a
compact set and there exists a subsequence such that that θ0n → θ0 ∈ R.
Therefore, (+ + 0
n )n≥1 is almost surely convergent to some  ∈ L (R+ , F1 ). We
+ 1
conclude that X = θ0 ∆S1 −  ∈ A0 .
2nd case: supn |θ0n | = ∞. In that case, we may suppose that |θ0n | → ∞.
n
Let us define θ¯0 = θ0n /(1 + |θ0n |). We define similarly X̄ n and ¯+n . We have
n ¯ n + 1 ¯ n
X̄ = θ0 ∆S1 − ¯n ∈ A0 where |θ0 | ≤ 1. Therefore, we may apply the first
case and, as n → ∞, we deduce an equality of the type 0 = θ̄0 ∆S̃1 − ¯+ where
|θ̄0 | = 1. We deduce that θ̄0 ∆S1 ≥ 0 a.s.
As d = 1, θ̄0 = ±1. Consider the case where θ̄0 = 1, we have ∆S1 ≥ 0 a.s.
Otherwise, we have ∆S1 ≤ 0 a.s. At this stage, we can not conclude anything
about the closedness. Let us consider the case ∆S1 ≥ 0 a.s. Consider the
strategy θ̂0n = n(−S0 , 1). Then, starting from the zero initial endowement,
we get the terminal portfolio process Ṽ1n = 0 + θ0n ∆S1 = n∆S1 ≥ 0 a.s.
This means that from nothing (zero initial capital), we get a non negative
terminal wealth, i.e. we do not take any risk to face a loss. Moreover, if n
is large enough and if P (∆S1 > 0) there is a non null probability to get a
strictly positive gain n∆S1 > 0 as large as we want, i.e. we get what we call
an arbitrage opportunity. If the agents acting on this financial market are
well informed and rational, we may think that they all utilize this possibility
to get positive money without taking any risk. Therefore, they all buy the
risky asset to hold a position of type θ0n . Then, the risky asset price S0 should
go up and the condition ∆S1 ≥ 0 a.s. should fail. This leads to the absence
of arbitrage opportunity condition we now define.
Definition 1.7. An arbitrage opportunity is a terminal portfolio process
ṼT ∈ RT0 starting from the zero initial endowment such that ṼT ≥ 0 a.s.
and P (ṼT > 0) > 0.
Definition 1.8. We say that the NA condition (No Arbitrage opportunity)
holds if there is no arbitrage opportunity, i.e. RT0 ∩ L0 (R+ , FT ) = {0} or,
 / 8

equivalently, AT0 ∩ L0 (R+ , FT ) = {0}.

Proposition 1.9. If NA holds, then A10 is closed in L0 (R+ , FT ).
Proof. From above, it suffices to study the case where ∆S̃1 ≥ 0 a.s. or
∆S̃1 ≤ 0. In that case, ±∆S̃1 ∈ A10 ∩ L0 (R+ , F1 ) = {0} hence ∆S̃1 = 0 a.s.
hence X n = −+ 0
n → X ≤ 0 a.s. This implies that X ∈ −L (R+ , F1 ) ⊆ A0 .
1

The conclusion follows. 2

The following step is to characterize the NA condition. To do so, we first
recall a lemma, see [15, Lemma 2.1.3, Section 2.1.2]. This result may be seen
as a generalization of the Halmos-Savage theorem, see e.g. [17].
Lemma 1.10. Let G = (Γi )i∈I be a family of elements of a σ-algebra F such
that, for all Γ ∈ F, if P (Γ) > 0, there exists i ∈ I such that S
P (Γ ∩ Γi ) > 0.
Then, there exists a countable family (Γin )n≥1 such that Ω = ∞ n=1 Γin a.s.

Proof. We may suppose that G is stable under countable union. Indeed, in

the contrary case, it suffices to replace G by G̃, which is the set of all countable
unions of elements of G. Let us consider m = supi P (ΓS i ). For a countable
sequence, we have m = lim ↑ P (Γin ) = P (Γ̂) where Γ̂ = ∞ n=1 Γin . We claim
that P (Γ̂) = 1, which is enough to conclude. Suppose by contradiction that
P (Γ̂) < 1 hence P (Γ̂c ) > 0. By assumption, there exits i0 ∈ I such that
P (Γ̂c ∩ Γi0 ) > 0. Therefore, as G is stable under countable union,

m ≥ P (Γ̂ ∪ Γi0 ) = P (Γ̂) + P (Γ̂c ∩ Γi0 ) > P (Γ̂) = m.

We get a contradiction so we may conclude. 2

Theorem 1.11. Suppose that T = 1 = d. Then, NA holds if and only if
there exists Q ∼ P such that dQ/dP ∈ L∞ (R, F1 ) and EQ S̃1 = S0 .
Proof. Before presenting the proof, let us recall that the probability mea-
sures Q and P are equivalent (Q ∼ P ) means that they admit the same
negligible sets. By the Radon-Nikodym theorem, if Q ∼ P , there exists
ρ ∈ L1 ((0, ∞), P, F1 ) such that dQ/dP = ρ, i.e. Q(A) = EP (ρ1A ) for all
A ∈ F1 . In particular, we have EQ (X) = EP (ρX) for all X ∈ L1 (R, F1 , P ).
Suppose that NA holds. The property still holds under P 0 ∼ P . In partic-
ular, with dP 0 /dP = αe−|S̃1 | , we may suppose w.l.o.g. that S̃1 is integrable
under P . We know that A10 is a convex cone closed in L0 (R, F1 , P ) by Propo-
sition 1.9. We deduce that A10 ∩ L1 (R, F1 , P ) is also closed in L1 (R, F1 , P )
since the convergence in L1 implies the convergence in probability. By NA,
 / 9

/ A10 ∩ L1 (R, F1 ). By the Hahn-Banach sepa-

for all x ∈ L1 (R+ , F1 ) \ {0}, x ∈
ration theorem, we deduce the existence of ρx ∈ L∞ (R, F1 ) and c ∈ R such
that
E(ρx X) < c < E(xρx ), ∀X ∈ A10 .
As A10 is a cone, replace X by kX and make k → ∞. We get that E(ρx X) ≤ 0
for all X ∈ A10 . Since, −L0 (R+ , F1 ) ⊆ A10 , we then deduce that ρx ≥ 0 a.s.
With X = 0, we get that c > 0 and, as R10 is a vector space, E(ρx X) = 0
for all X ∈ R10 . As P (ρx > 0) > 0 (see the strict inequality above), we may
renormalize and suppose that kρx k∞ = 1.
Let us consider the family G = (Γx )x∈I where I = L1 (R+ , F1 ) \ {0} and
Γx = {ρx > 0}. For any Γ ∈ F1 such that P (Γ) > 0, x = 1Γ ∈ I. Therefore,
S∞ x 1Γ ) > 0 hence P (Γx ∩ Γ)
E(ρ P∞> 0. −i
By Lemma 1.10, we may write Ω =
Γ
i=1 xi . Let us define ρ = i=1 2 ρ xi . We have ρ > 0 a.s. and we may
∞ +
renormalize ρ such that ρ ∈ L (R , F1 ) and EP (ρ) = 1. We then define
Q ∼ P such that dQ/dP = ρ. We still have E(ρX) = 0 for all X ∈ R10 , in
particular with X = ∆S̃1 ∈ R10 , we may conclude that EQ (S̃1 ) = S0 .
Reciprocally, suppose the existence of Q ∼ P such that EQ (S̃1 ) = S0 .
Take ṼT = θ0 ∆S̃1 ∈ RT0 ∩ L1 (R+ , F1 ). We have EQ (ṼT ) = EQ (θ0 ∆S̃1 ) =
θ0 EQ (∆S̃1 ) = 0. As ṼT ≥ 0 a.s., we get that ṼT = 0, i.e. NA holds. 2

A probability Q as in Theorem 1.11 is called a risk-neutral probability

measure or (equivalent) martingale measure. Recall that, if ξT ∈ L1 (R, F1 )
is a payoff, a (super-replicating) price for ξT is an initial endowment p0 of a
portfolio process V satisfying VT ≥ ξT a.s. We say that V replicates ξT when
VT = ξT a.s. We denote by Γ(ξT ) the set of all prices for ξT .
Theorem 1.12. Suppose that T = 1 and NA holds. Let us consider the set
EMM6= ∅ of all equivalent martingale measure. Then, if ξT ∈ L1 (R, F1 ),
 
Γ(ξT ) = sup EQ (ξ˜T ), ∞ .
Q∈EM M

Proof. Consider p0 ∈ Γ(ξT ), i.e. there exists θ0 ∈ R such that p0 + θ0 ∆S̃1 ≥

ξ˜T a.s. Taking the Q-expectation, we get p0 ≥ EQ (ξ˜T ) since EQ (θ0 ∆S̃1 ) =
θ0 EQ (∆S̃1 ) = 0. It remains to show that p∗0 = supQ∈EM M EQ (ξ˜T ) ∈ Γ(ξT ).
Let us suppose by contradiction that p∗0 ∈ / Γ(ξT ), i.e. ξ˜T − p∗0 ∈
/ A10 . As the
1
latter set is a closed convex set in L , the Hahn-Banach separation theorem
 / 10

applies and we get Z ∈ L∞ (Rd , F1 ), c > 0, such that

E(ZX) < c < E(Z(ξ˜T − p∗0 )), ∀X ∈ A10 .

As in Theorem 1.11, we get that Z ≥ 0 and E(Z∆S̃1 ) = 0. Consider Z1 =

dQ1 /dP where Q1 ∈ EM M 6= ∅. We define ρ = α (βZ + Z1 ) where α, β > 0.
We have
 
E(ρ(ξ˜T − p∗0 )) = α βE(Z(ξ˜T − p∗0 )) + E(Z1 (ξ˜T − p∗0 )) .

As E(Z(ξ˜T − p∗0 )) > 0, we may choose β > 0 large enough in such a way
that E(ρ(ξ˜T − p∗0 )) > 0. We then fix α such that ρ > 0 defines an equivalent
probability measure Q ∼ P with dQ/dP = ρ. Moreover, by construction,
EQ (S̃1 ) = 0, i.e. Q ∈ EM M . It follows that p∗0 ≥ EQ (ξ˜T ). On the other
hand, EQ (ξ˜T ) > p∗0 by construction hence a contradiction. 2
A natural question is whether EMM is a singleton. This is related to the
concept of completeness for the market.
Definition 1.13. We say that the financial market is complete if for any
ξT ∈ L1 (R, FT ), there exists a self-financing portfolio process V such that
VT = ξT .
Proposition 1.14. Let T = 1. Suppose that NA holds. Then, the market is
complete if and only if EMM is a singleton.
Proof. Suppose that the market is complete. Let Q1 , Q2 ∈ EM M . Consider
A ∈ F1 . The payoff ξT = 1A is replicable by assumption, i.e. there exists a self-
financing portfolio process V such that VT = 1A . We have ṼT = V0 +θ0 ∆S̃1 for
some θ0 ∈ R, hence EQ1 ṼT = EQ2 ṼT = V0 . This implies that Q1 (A) = Q2 (A),
for all A, i.e. Q1 = Q2 .
Reciprocally, if EM M = {Q}, we know by Theorem 1.12, that p∗0 = EQ (ξ˜1 )
is a super-replication price, i.e. there exists a portfolio process V such that
Ṽ1 ≥ ξ˜1 . This means that ξ˜1 = p∗0 + θ0 S̃1 − + + 0
1 where 1 ∈ L (R+ , F1 ). We
deduce that EQ (ξ˜1 ) = p0 − EQ (1 ) hence EQ (1 ) = 0 and +
∗ + +
1 = 0. This
˜
implies that ξ1 is replicable. 2

1.4. General case: the Dalang-Morton-Willinger theorem

In this section, we generalize the results of the last section.

 / 11

Definition 1.15. Let Q ∼ P . We say that the stochastic process (Mt )t=0,··· ,T
is a Q-martingale if, for all t = 0, · · · , T , Mt is Q-integrable (EQ |Mt | < ∞)
and EQ (Mt+1 |Ft ) = Mt .
We shall need a generalized version of the conditional expectation which
allows to consider conditional expectation of non integrable random variables.
Recall that the conditional expectation of any non negative random variable
X exists and is defined as E(|X||G) = limn ↑ E(|X| ∧ n|G) where |X| ∧ n ∈
[0, n], n ≥ 1, is integrable.
Definition 1.16. Let G ⊆ F be two σ-algebras and X ∈ L0 (Rd , F), d ≥ 1.
We say that X admits a conditional expectation E(X|G) if E(|X||G) < ∞
a.s. In that case, we define
E(X|G) = E(X + |G) − E(X − |G) ∈ L0 (Rd , G).
We may show the following:
Lemma 1.17. Let XG ∈ L0 (Rd , G) and suppose that Y ∈ L0 (Rd , F) admits a
conditional expectation E(Y |G). Then, XG Y admits a conditional expectation
such that E(XG Y |G) = XG E(Y |G).
Proposition 1.18. Suppose that NA holds. Then, AT0 is closed in L0 .
Proof. We show the statement by induction. For two dates, let us consider
X = θTn −1 S̃T −n+
n T
T ∈ AT −1 converging a.s. to X as n → ∞. We suppose that
θT −1 ∈ L (R , FT −1 ) and n+
n 0 d 0
T ∈ L (R+ , FT ). We split Ω into two subsets:

a) On the set ΩT −1 = {lim inf n |θTn −1 | < ∞} ∈ FT −1 . By [15, Lemma 2.1.2,

Section 2.1.2 ], there exists a random sequence nk ∈ L0 (N, FT −1 ) such that
θTnk−1 converges almost surely to some θT −1 . Notice that
∞
X
θTnk−1 = θTj −1 1nk =j ∈ L0 (Rd , FT −1 ).
j=k

We deduce that θT −1 ∈ L0 (Rd , FT −1 ). At last, we get that n+ T → +

T ∈
0 + T
L (R+ , FT ). Finally, X1ΩT −1 = θT −1 1ΩT −1 S̃T − T 1ΩT −1 ∈ AT −1 .
b) On the set ΩcT −1 = {lim inf n |θTn −1 | = ∞}. We use the normalization
procedure, as in the last section, of the type X̄ = X/(1 + |θTn −1 |). Then, we
apply the first step a) to the sequence X̄ n . In limit, we get that θ̄T −1 S̃T −¯+ T =
0 for some ¯+ T ≥ 0 a.s. and θ̄T −1 ∈ L 0
(R d
, F T −1 ) such that |θ̄T −1 | = 1. As
+ T 0
θ̄T −1 S̃T = ¯T ∈ AT −1 ∩ L (R+ , FT −1 ), we get that θ̄T −1 S̃T = 0 by NA.
 / 12

Let us restrict ourselves to the case d = 1. We shall see the general case
below. As θ̄T −1 ∈ {−1, 1}, we get that ∆S̃T = 0 hence X n ≤ 0 and X ≤ 0.
Therefore, X1ΩcT −1 ∈ ATT −1 . We conclude that X = X1ΩT −1 +X1ΩcT −1 ∈ ATT −1 .
Suppose by induction that ATt is closed and let us show that ATt−1 is also
closed. To do so, consider a converging sequence X n = θt−1 n
∆S̃t + · · · +
n n+
θT −1 ∆S̃T − T → X.
n
c) On the set Ωt−1 = {lim inf n |θt−1 | < ∞} ∈ Ft−1 , we may suppose
w.l.o.g. (see the first step a)) that θt−1 → θt−1 ∈ L0 (Rd , Ft−1 ). Therefore,
n

θtn ∆S̃t+1 + · · · + θTn −1 ∆S̃T − n+

T is convergent by the induction hypothesis to
Xt,T ∈ At . It follows that X1Ωt−1 = θt−1 1Ωt−1 ∆S̃t + Xt,T 1Ωt−1 ∈ ATt−1 .
T

d) On the set Ωct−1 = {lim inf n |θt−1 n

| = ∞} ∈ Ft−1 , we use the normal-
ization procedure as in b) and we deduce an equality of the type γt−1 =
θ̄t−1 ∆S̃t + · · · + θ̄T −1 ∆S̃T − ¯+
T = 0. In the case where d = 1, θ̄t−1 ∈ {−1, 1}.
We split Ωct−1 = {θ̄t−1 = −1} ∪ {θ̄t−1 = 1}. On the set {θ̄t−1 = 1}, we observe
that

X n = X n − θt−1
n
γt−1 = θtn ∆S̃t+1 + · · · + θTn −1 ∆S̃T − n+ T
T ∈ At .

Using the induction hypothesis, we may conclude. Similar arguments apply

on the set {θ̄t−1 = −1}.
The general case where d > 1 needs to be thought component-wise. As
|θ̄t−1 | = 1, we split Ωct−1 into a partition (Bi )i=1,··· ,d of Ft−1 such that Bi ⊆
i
{θ̄t−1 6= 0}. On each Bi , we assume w.l.o.g. that θtni 6= 0 and we write
X = X n − αt−1
n n
γt−1 where αt−1 n
is chosen such that it is possible to rewrite
n ni
X in such a way that θt−1 = 0, i.e. we have strictly reduced the number
n
of non null components of θt−1 . We then go to step c) and, if necessary, we
n
still reduce the number of non null components of θt−1 . As d is finite, we
n
may conclude as, in the worst case, θt−1 is finally reduced to 0 so that the
induction hypothesis applies. Another technique is to define almost surely a
matrix P n such that P n θ̄t−1 = θtn and observe that X n = X n − P n γT −1 . In
that case, we need do show that P n is Ft−1 -measurable. This may be proven
by a measurable selection argument. 2
The following result is fundamental in the theory of arbitrage theory. A
complete version is given in [15]. We provide a proof which is not exactly
the original one. Note that there are other available proofs, see [26], [24] and
[23].
 / 13

Theorem 1.19 ( Dalang-Morton-Willinger theorem). The condition NA

holds if and only if there exists Q ∼ P such that (S̃t )t=0,··· ,T is a Q-martingale.
Proof. Suppose that NA holds. We know by Proposition 1.18 that AT0 is
closed in L0 . So, we apply the reasonings we did for T = 1 and we deduce
Q ∼ P such that EQ (X) = 0 for all X ∈ RT0 . In particular, for all t ≥ 1, for
all Ft−1 ∈ Ft−1 , 1Ft−1 ∆S̃t ∈ AT0 hence EQ (1Ft−1 ∆S̃t ) = 0. This implies that
EQ (∆S̃t |Ft−1 ) = 0, i.e. S̃ is a Q-martingale.
Reciprocally, suppose that S̃ is a Q-martingale. Consider ṼT ∈ AT0 ∩
L0 (R+ , FT ). We write ṼT = ṼT −1 + θT −1 ∆S̃T where θT −1 ∈ L0 (Rd , FT −1 ).
As ∆S̃T is Q-integrable, we deduce that ṼT admits a generalized conditional
expectation such that EQ (ṼT |FT −1 ) = ṼT −1 . We repeat the argument and we
get that EQ (ṼT |FT −2 ) = ṼT −2 . Finally, we have EQ (ṼT ) = Ṽ0 = 0. As ṼT ≥ 0
a.s., ṼT = 0, i.e. NA holds. 2
As in the case T = 1, we also deduce the following dual characterization
of the prices from the set EMM of all equivalent martingale measures Q ∼ P
under which S̃ is a Q-martingale.
Theorem 1.20. Suppose that NA holds. Consider ξT ∈ L1 (R, FT ). The set
of all super-hedging prices of ξT is

Γ(ξT ) = [ sup EQ (ξ˜T ), ∞).

Q∈EM M

We may also introduce the largest sub-hedging price p for ξT , i.e. the
largest price p such that p + VT ≤ ξT a.s. for some self-financing portfo-
lio process V . By symmetry, we have p = inf Q∈EM M EQ (ξ˜T ). Now, consider
an extended market model where the payoff ξT is quoted at price p(ξT ) at
time 0 and is available only at timePT at the price ξT . Therefore, a termi-
nal discounted claim is of the form Tt=1 θt−1 ∆S̃t + θ0 (ξ˜T − p(ξT )). Suppose
that there exists an arbitrage opportunity for this extended market. In par-
ticular, for some strategy (θ, θ0 ), Tt=1 θt−1 ∆S̃t + θ00 (ξ˜T − p(ξT )) ≥ 0 a.s. If
P
θ00 = 0, we get an arbitrage opportunity for the initial market contrarily to
the assumption. If θ00 > 0, divide by θ0 and take the expectation for any
Q ∈ EM M . We get that p(ξT ) ≤ EQ (ξ˜T ) hence p(ξT ) ≤ inf Q∈EM M EQ (ξ˜T ).
0 ˜
 if θ0 < 0, we get that p(ξT ) ≥ sup
Otherwise,  Q∈EM M EQ (ξT ). Therefore,
p(ξT ) ∈ inf Q∈EM M EQ (ξ˜T ), sup Q∈EM M EQ (ξ˜T ) implies that there is no ar-
bitrage opportunity.
 / 14

Conclusion: We have presented the main ideas of arbitrage theory without

frictions and in discrete-time, i.e. a no arbitrage condition NA is considered
to ensure the closedness of the set of hedgeable claims. The NA condition
is equivalent to the existence of a probability risk measure under which the
discounted prices are martingales. At last, it is possible to dually characterize
the super-hedging prices under NA via the dual elements, i.e. probability risk
measures. For a deeper study of arbitrage theory for frictionless models, we
send the readers to [15, Section 2].

2. Markets with frictions

2.1. Introduction

The theory we present in this section is rather recent. Most of the main results
of the literature have been developed in the last fifteen years. A pioneering
work is the paper by Jouini and Kallal [14] where bid and ask prices are
considered. We propose in this section an introduction to financial market
models with proportional transaction costs. In the following, we consider a
discrete-time stochastic basis (Ω, (Ft )t=0,1,··· ,T , P ). We denote by e1 the vector
of Rd , d ≥ 1, such that the only non null component is the first one which is
fixed to 1.
Example. Suppose that the market is composed of two assets. The first
one is non risky and its (discounted) value is St0 = 1 for all t ∈ [0, T ]. The
second asset is risky and the price is St at time t. As usual, we suppose
that S is a stochastic process adapted to the filtration. We suppose that we
need to pay proportional transaction costs when buying or selling the risky
asset. Precisely, when buying one unit of the risky asset, we pay the price
St (1 + ) = St + St . When selling one unit of the risky asset, we get the price
St (1 − ) = St − St . This means that the proportional transaction cost rate
is  > 0.
In this setting, we denote by V a portfolio process. Contrarily to the fric-
tionless model, V is expressed in physical units, i.e. V is the strategy θ̂ of
the last section. This choice is motivated by technical reasons: the dynamics
of a portfolio process is not trivial with transaction costs. In the sequel, a
financial position (x, y) describes the quantity x ∈ R and y ∈ R invested in
assets S 0 and S respectively.
Definition 2.1. The liquidation value at time t of the financial position (x, y)
 / 15

is
Lt ((x, y)) := x + y + St (1 − ) − y − St (1 + ).
This definition is clear. If y > 0, we liquidate the long position by selling
the y units of risky asset at price St (1 − ). If y < 0, we liquidate the short
position by buying the y − units of risky asset at price St (1 + ). Notice that
L is linear only in the first component. This linearity is used afterwards.
Definition 2.2. At time t, the solvency set is defined as
Gt (ω) := {z = (x, y) ∈ R2 : Lt (z) ≥ 0}.
Gt is the set of all positions we may liquidate without any debt. Indeed, if
z ∈ Gt , write z = z − Lt (z)e1 + Lt (z)e1 and observe that Lt (z − Lt (z)e1 ) = 0.
We may easily show that Gt is a closed convex cone. For y ≥ 0, z = (x, y) ∈
Gt if and only if x + ySt (1 − ) ≥ 0, i.e. zgt2∗ ≥ 0 where gt2∗ = (1, ySt (1 − )).
For y < 0, z = (x, y) ∈ Gt if and only if x + ySt (1 + ) ≥ 0, i.e. zgt1∗ ≥ 0
where gt1∗ = (1, ySt (1 + )). The vectors gt1∗ and gt2∗ are the generators of the
positive dual cone
G∗t = {z ∈ R2 : zgt ≥ 0, ∀gt ∈ Gt } = cone (gt1∗ , gt2∗ ).
Lemma 2.3. The solvency set is Ft -graph-measurable at time t:
graph Gt := {(ω, z) ∈ Ω × Rd : z ∈ Gt (ω)} ∈ Ft ⊗ B(Rd ).
Proof. It suffices to observe that (ω, z) ∈ graph (Gt ) if and only if zgt1∗ ≥ 0
and zgt2∗ ≥ 0. 2
Similarly, we have Gt = cone (gt1 , gt2 ), where gt1 = (St (1 + ), −1) and
gt2 = (−St (1 − ), 1). Therefore, G∗t is Ft -graph-measurable at time t since
(G∗t )∗ = Gt .
Proposition 2.4. For all z ∈ R2 ,
Lt (z) = max{α ∈ R : z − αe1 ∈ Gt }.
Proof. Consider α ∈ R such that z − αe1 ∈ Gt . Then, Lt (z − αe1 ) ≥ 0, i.e.
Lt (z) − α ≥ 0 hence α ≤ Lt (z). Moreover, Lt (z − Lt (z)e1 ) = 0 implies that
z − Lt (z)e1 ∈ Gt . The conclusion follows. 2
Definition 2.5. A self-financing portfolio process is a stochastic process
(Vt )t=0,··· ,T starting from an initial endowment V−1 = V0− such that, for all
t ≥ 0, ∆Vt ∈ −Gt a.s.
 / 16

The interpretation of the dynamics above is the following: we may write

Vt−1 = Vt +(−∆Vt ) so that it is possible to change Vt−1 into Vt as it is allowed
to let aside (−∆Vt ) whose liquidation value Pis non negative. Observe that the
terminal value of V is an element of V0− + Tt=0 L0 (−Gt , Ft ).
Several no arbitrage conditions have been considered in order to solve the
super-hedging problem, as in the frictionless case. To do so, we consider the
set of all terminal claims AT0 we may obtain from a zero initial capital. We
have
T
X
T
A0 = L0 (−Gt , Ft ).
t=0

We suppose that the prices are non negative. Therefore, R2+ ⊆ Gt a.s. hence
−L0 (R2+ , FT ) ⊆ AT0 .
Definition 2.6. NAw : AT0 ∩ L0 (R2+ , FT ) = {0}.
Proposition 2.7. Suppose that ST > 0 a.s. and  < 1. The condition NAw
holds if and only if, for all VT ∈ AT0 , LT (VT ) ≥ 0 implies that LT (VT ) = 0
a.s.
Proof. Suppose that NAw holds and consider VT ∈ AT0 such that LT (VT ) ≥
0. Since VT −LT (VT )e1 ∈ GT and GT is stable under addition, we deduce that
LT (VT )e1 = VT − (VT − LT (VT )e1 ) ∈ AT0 ∩ L0 (R2+ , FT ) = {0}, i.e. LT (VT ) = 0
a.s. Reciprocally, consider VT ∈ AT0 ∩ L0 (R2+ , FT ). Necessarily, LT (VT ) ≥ 0
hence LT (VT ) = 0. As VT ∈ R2+ , we have 0 = LT (VT ) = VT1 + VT2 ST (1 − ). It
follows that VT1 = VT2 = 0. 2
Clearly, the meaning of NAw is the same than the NA condition of the
frictionless case. In general, we shall see that stronger conditions are con-
sidered in presence of transaction costs to ensure the closedness of AT0 . In
the following, we introduce the stochastic preorder x ≥Gt y if and only if
x − y ∈ Gt , t = 0, · · · , T .
Definition 2.8. An endowment for the payoff ξT ∈ L0 (R2 , FT ) is a vector
p0 ∈ R2 which is the initial capital of a self-financing portfolio V such that
VT ≥GT ξT a.s.
Notice that p0 ∈ R2 is an endowment if p0 − Tt=1 gt = ξT + gT0 for some
P
gt ∈ L0 (Gt , Ft ), t ≤ T and gT0 ∈ L0 (GT , FT ). As GT is a convex cone, we
get that p0 + VT = ξT where VT ∈ AT0 . As in the frictionless case, let us see
whether AT0 may be closed. Let us start with T = 1.
 / 17

Lemma 2.9. Suppose that T = 1 and S1 is not deterministic. Then, A10 is

closed in probability under NAw .
Proof. Consider a convergent sequence X n = −g0n − g1n where gtn ∈ −Gt
a.s., t = 0, 1. We denote by X the limit of (X n )n≥1 .
1) First case: supn |g0n | < ∞. We may suppose that g0n → g0 ∈ G0 as
(Gt )t=0,··· ,T take closed values. Therefore, g1n → g1 ∈ L0 (G1 , F1 ) hence X =
−g0 − g1 ∈ A10 .
2) Second case: supn |g0n | = ∞. We may suppose that |g0n | → ∞. We normalize
the sequence by setting ḡ0n = g0n /|g0n |, X̄ n = X n /|g0n |, etc. By the first case, we
get an equality of the type g0 + g1 = 0 where gt ∈ Gt a.s., t = 0, 1. Therefore
−g0 ∈ A10 is such that L1 (−g0 ) = L1 (g1 ) ≥ 0 hence L1 (g1 ) = L1 (−g0 ) = 0
by NAw . Moreover, g1 = −g0 is deterministic. In the case where the second
component g12 of g1 is non negative, L1 (g1 ) = 0 means that g11 +g12 S1 (1−) = 0.
If g12 = 0, then g11 = 0 hence g0 = −g1 = 0 in contradiction with |g0 | = 1. So,
g12 6= 0 and S1 = −g11 /(g11 (1 − )). This leads to a contradiction. Similarly,
the case where g12 ≤ 0 is excluded. 2
Theorem 2.10. Suppose that T = 1 and S1 is not deterministic. Then, NAw
holds if and only if there exists a process (Zt )t=0,1 such that Z0 = E(Z1 ) and
Zt ∈ G∗t a.s., t = 0, 1.
Proof. Suppose that NAw holds. Then, by Lemma 2.9, we deduce that
A10 ∩ L1 (R2 , F1 ) is closed in L1 (R2 , F1 ). Moreover, for any x ∈ L1 (R2+ , F1 ) \
{0}, x ∈ / A10 by NAw . Therefore, by the Hahn-Banach separation theorem,
there exists Zx ∈ L∞ (R2 , F1 ) and c > 0 such that E(XZx ) < c < E(xZx )
for all X ∈ A10 . Note that Zx 6= 0 and we may assume that kZx k∞ = 1. As
A10 is a cone, we deduce that E(XZx ) ≤ 0 for all X ∈ A10 . With X = −g0
where g0 is chosen arbitrarily in G0 , we deduce that g0 E(Zx ) ≥ 0 for any
g0 ∈ G0 , i.e. E(Zx ) ∈ G∗0 . Similarly, we have E(Zx g1 ) ≥ 0 for all g1 ∈
L1 (G1 , F1 ). We deduce that Zx ∈ G∗1 ⊆ R2+ a.s. Indeed, otherwise, we may
construct pointwise g1 ∈ L0 (G1 , F1 ) with |g1 | = 1 such that Zx g1 ≤ 0 a.s. and
P (Zx g1 < 0) > 0, i.e. a contradiction. To do so, we apply [15, Theorem 5.4.1,
Section 5.4] which asserts that a F1 -measurable selection g1 exits as soon as
the existence holds pointwise. We now consider the family (Gx )x∈L1 (R2+ ,F1 )\{0}
with Gx := {Zx e1 > 0}. For any Γ such that P (Γ) > 0, consider x = 1Γ e1 ∈
L1 (R2+ , F1 ) \ {0}. As E(xZx ) > 0, we deduce that S P (Γ ∩ {Zx e1 > 0}) > 0.
Therefore, Lemma 1.10 applies: we have Ω = ∞ i=1 {Z x e1 > 0} for some
countable family (xi )i≥1 . We finally conclude with Z1 = i≥1 2−i Zxi > 0 and
P
 / 18

Z0 = E(Z1 ).
Reciprocally, suppose the existence of Z and consider V1 = −g0 − g1 ∈
A10 ∩ L1 (R2+ , F1 ). Then Z1 V1 ≥ 0 and Z1 V1 = 0 if and only if V1 = 0 as
Z1 ∈ G∗1 ⊆ int R2+ . On the other hand, E(Z1 V1 ) = −g0 Z0 − E(g1 Z1 ) ≤ 0 by
assumption. Therefore, Z1 V1 = 0 and V1 = 0. 2
Definition 2.11. A consistent price system (CPS) is a P -martingale (Zt )t=0,··· ,T
adapted to the filtration (Ft )t=0,··· ,T such that Zt ∈ G∗t \ {0} a.s. for all
t = 0, · · · , T .
The following theorem (see [9]) is a generalization of Theorem 2.10 for
d = 2, see [15, Theorem 3.2.15, Section 3].
Theorem 2.12 (Grigoriev’s theorem). Suppose d = 2 and T is arbitrarily
chosen. The following statements are equivalent:
1) NAw .
2) AT0 ∩ L1 (R2+ , F1 ) = {0}.
3) There exists a CPS.
In [15, Section 3], some counterexamples show that AT0 is not necessarily
closed under NAw . In that case, it is not possible a priori to characterize the
set of all super-hedging prices of a payoff.
Proposition 2.13. Suppose that NAw holds and AT0 is closed. Consider a
payoff ξT ∈ L1 (Rd , FT ). Then, the set of all super-replicating prices Γ(ξT ) of
ξT is given by
Γ(ξT ) = x0 ∈ Rd : x0 Z0 ≥ E(ZT ξT ), ∀Z CPS in L∞ .


Proof. Let us consider x0 ∈P Γ(ξT ), i.e. there exists VT ∈ AT0 such that
x0 + VT = ξT . We have Vt = − tu=0 gu where gu ∈ L0 (Gu , Fu ), u = 0, · · · , t
and t ≤ T . We have ZT ξT = ZT (x0 + VT ) = ZT (x0 + VT −1 − gT ). As ZT ∈ G∗T ,
we deduce that ZT ξT ≤ ZT (x0 + VT −1 ) hence, considering the generalized
conditional expectation, we get that
E(ZT ξT |FT −1 ) ≤ ZT −1 (x0 +VT −1 ) = ZT −1 (x0 +VT −2 −gT −1 ) ≤ ZT −1 (x0 +VT −2 ).
Repeating the reasonning, i.e. take the successive generalized conditional
expectations, we finally get that E(ZT ξT ) ≤ Z0 x0 .
Reciprocally, consider x0 ∈ Rd such that E(ZT ξT ) ≤ Z0 x0 for all CPS Z.
Suppose by contradiction that ξT − x0 ∈ / AT0 ∩ L1 (Rd , FT ). By the Hahn-
Banach separation theorem, there exists Ẑ ∈ L∞ (Rd , FT ) and c ∈ R such
 / 19

that
E(ẐX) < c < E(Ẑ(ξT − x0 )), ∀CPS Z.
As AT0 is a cone, we deduce that E(ẐX) ≤ 0 for all X ∈ AT0 and c > 0. With
X = −gt ∈ L0 (−Gt , Ft ) ⊆ AT0 , we have E(Ẑt gt ) ≥ 0 for any gt ∈ L0 (Gt , Ft ),
where Ẑt = E(Ẑ|Ft ). Arguing by contradiction with a measurable selection
argument, see [15, Theorem 5.4.1, Section 5.4], we deduce that Ẑt ∈ G∗t a.s.
Let us define Zt = αZ̄t + Ẑt where Z is a CPS. By construction Z is a CPS
if α > 0. Moreover, with α small enough, we get that E(Ẑ(ξT − x0 )) > 0 in
contradiction with the property satisfied by x0 . 2
In the literature, a stronger no-arbitrage condition NAr has been intro-
duced. This condition means that there is no arbitrage opportunity even if
the transaction costs are slightly smaller. Equivalently, this means that there
is no arbitrage opportunity when the solvency set is larger, i.e. the positive
dual is smaller. A CPS for this enlarged market is therefore in the interior of
the initial positive dual. This ensures the closedness of AT0 , see [15, Section
3.2.2], so that Proposition 2.13 applies. This condition NAr appears to be
crucial to derive a FTAP as in the papers [2], [20] and [8] among others.

2.2. A new approach based on the liquidation value

In the last section, we have seen that the set of all terminal claims AT0
is not necessarily closed under NAw . A natural question is to understand
whether this is the case for the liquidation values of these terminal claims.
We consider here the case d = 2. The first asset is riskless and defined by
the price St0 = 1, t = 0, · · · , T . The risky asset is defined by the bid and ask
prices Stb and Sta such that 0 < Stb ≤ Sta , t = 0, · · · , T . At time t, Stb and Sta
are respectively the prices we get when selling/buying one unit of the risky
asset. That corresponds to the best bid/ask prices in an order book. The
liquidation value process is then:
Lt ((x, y)) = x + y + Stb − y − Sta , t = 0, · · · , T.
We then define Gt := {z ∈ R2 : Lt (z) ≥ 0} as in the last section. Similarly,
we define
t
X
Atu : = L0 (−Gr , Fr ),
r=u
Ltu : = {Lt (Vt ) : Vt ∈ Atu } 0 ≤ u ≤ t ≤ T.
 / 20

In the following, we consider a technical condition:

E: For T ≥ 2, for all t ≤ T − 1 and u ≥ t + 1, Fu ∈ Fu ,

(i) If Sta = Stb on Fu , then there exists r ≥ u such that Sta ≥ Sra on Fu .
(ii) If Stb = Sta on Fu , then there exists r ≥ u such that Srb ≥ Stb on Fu .
In [27], we provide classical examples where Condition E is satisfied. Clearly,
it is satisfied under the efficient market hypothesis, i.e. when S b < S a . The
following is easy to prove:
Lemma 2.14. The condition NAw is equivalent to one of the equivalent
conditions:

LT0 ∩ L0 (R+ , FT ) = {0} ⇔ AT0 ∩ L0 (R2+ , FT ) = {0}.

The following theorem is new and proved in [27]. It may be seen as an

analog of the DMW Theorem 1.19.
Theorem 2.15. Suppose that condition E holds for 3 steps or more. Then,
the following conditions are equivalent:
(i) NAw .
(ii) LT0 is closed in probability and LT0 ∩ L0 (R+ , FT ) = {0}.
(iii) There exists Q ∼ P satisfying dQ/dP ∈ L∞ ((0, ∞), FT ) such that
EQ (LT (VT )) ≤ 0 for all LT (VT ) ∈ LT0 ∩ L1 (R, FT ).
Proof. We provide here the proof in the case where there are only two
steps. The general case is deduced by induction, see [27].
The implication (ii) ⇒ (iii) follows from the Hahn-Banach separation
theorem, see the last sections. The implications (iii) ⇒ (i) and (ii) ⇒ (i)
are trivial.
Closedness. It remains to show that (i) ⇒ (ii), i.e. LT0 is closed in probability.
With one time step, this is immediate as LTT = −L0 (R+ , FT ). We may show
that, for any γ ∈ LT0 , γe1 = −g0T ∈ AT0 where gut , u ≤ t,P is a general
notation we introduce for the sequel to designate a sum gu = tr=u gr with
t

gr ∈ L0 (Gr , Fr ), r P
≤ T . When considering a sequence of such elements, we
write them gut,n = tr=u grn with grn ∈ L0 (Gr , Fr ). In the following, we may
suppose w.l.o.g. that gr ∈ ∂Gt := Gt \ intGt for all t ≤ T − 1. To do so,
we withdraw Lr (gr ) ≥ 0 from gr that we add to GT . Recall that, by the
Grigoriev theorem, there exists a CPS Z.
 / 21

Two steps. Consider γT∞ = limn γTn where γTn e1 = −gTT,n n

−1 = −gT −1 − gT ∈
n

LTT −1 . Define the set ΓT −1 := {lim inf |gTn −1 | = ∞} ∈ FT −1 . Up to a ran-

dom subsequence, we may suppose that |gTn −1 | > 0. We normalize the se-
quences by setting γ̃Tn := γTn /|gTn −1 |, g̃Tn −1 := gTn −1 /|gTT,n n n T,n
−1 |, g̃T := gT /|gT −1 |.
As |g̃Tn −1 | = 1, we may assume that g̃Tn −1 → g̃T∞−1 ∈ GT −1 , see [15, Lem.
2.1.2]. As limn γ̃Tn e1 = 0, we deduce that g̃Tn → g̃T∞ ∈ GT and g̃T∞−1 + g̃T∞ = 0
where g̃T∞−1 ∈ ∂GT −1 and g̃T∞ ∈ GT . We set g̃T∞−1 = g̃T∞ = 0 on ΛT −1 =
Ω \ ΓT −1 ∈ FT −1 . Let Z be a CPS. As ZT (g̃T∞−1 + g̃T∞ ) = 0, we deduce that
ZT −1 g̃T∞−1 + E(ZT g̃T∞ |FT −1 ) = 0. By duality, i.e. using the property that a
CPS evolves in the positive dual of G, we get that ZT −1 g̃T∞−1 = ZT g̃T∞ = 0.
As g̃T∞ = −g̃T∞−1 is FT −1 -measurable, we get that 0 = E(ZT g̃T∞ |FT −1 ) =
ZT −1 g̃T∞ . So, ZT −1 g̃T∞ = ZT g̃T∞ hence ZT −1 ∈ (R+ ZT ) ∩ G∗T . Therefore,
ZT −1 γTn e1 = −ZT −1 gTn −1 − ZT −1 gTn ≤ 0 by duality. Since ZT −1 e1 > 0, we
deduce that γTn ≤ 0. So, γTn e1 = −ĝTT,n n n
−1 a.s., where ĝT −1 = gT −1 1ΛT −1 ∈
n n n 0
∂GT −1 and ĝT = gT 1ΛT −1 + (−γT e1 )1ΓT −1 belongs to L (GT , FT ). By con-
struction, lim inf n |ĝTn −1 | < ∞ hence we may suppose that ĝTn −1 → ĝT∞−1 ∈
L0 (GT −1 , FT −1 ) by [15, Lem. 2.1.2]. We deduce that ĝTn → ĝT∞ ∈ L0 (GT , FT )
hence γT∞ = −ĝTT,∞ −1 ∈ LT −1 . 2
T

In the following, we denote by M∞ (P ) the set of all Q ∼ P such that

dQ/dP ∈ L∞ and EQ LT (V ) ≤ 0 for all LT (V ) ∈ LT0 . For any contingent
claim ξ ∈ L1 (R, FT ), we define the set Γ(ξ) of all initial endowments of
portfolio processes whose terminal liquidation values coincide with ξ, i.e.

Γ(ξ) := {x ∈ R : ∃V ∈ AT0 : LT (xe1 + VT ) = ξ}.

Corollary 2.16. Suppose that condition E holds. Let ξ ∈ L0 (R, FT ) be such
that EP |ξ| < ∞. Then, under condition NAw , Γξ = [supQ∈M∞ (P ) EQ ξ, ∞).

The proof is very similar to the one for frictionless markets.

2.3. When the solvency set is not a convex cone

All the arguments we have used in the previous sections are possible because
the solvency sets are closed convex sets. This allows to deduce a dual charac-
terization of no-arbitrage conditions and super-hedging prices. In particular,
AT0 is a closed convex cone. Clearly, this classical principle in mathematical
finance is no more valid if G is not convex. In the following, we present a
 / 22

modest new contribution allowing to compute the super-hedging prices in a

non convex setting.
Let us consider the very simple example with two assets and two time
steps. The first one is St0 = 1, t = 0, 1, the second one is risky and defined
by the price St , t = 0, 1. We suppose that there are proportional transaction
costs to pay when buying/selling the risky asset. Moreover, a fixed cost c ≥ 0
is charged. We suppose that the agent accepts to pay c only if the liquidation
value of the risky position y is either negative or larger than the fixed cost.
Indeed, if 0 < yS1 (1 − ) ≤ c, it is not interesting for the agent to liquidate
the risky position y. Precisely, we suppose that

Lt ((x, y)) = x + (ySt (1 − ) − c)+ − y − St (1 + ) − c1y<0 , t = 0, 1.

As usual, we define the solvency set Gt := {z ∈ R2 : Lt (z) ≥ 0}, t = 0, 1.

We may easily observe that G is not convex. By [28], z 7→ Lt (z) is upper
semi-continuous.
A new approach is necessary to obtain the super-hedging prices of some
payoff ξ1 ∈ L0 (Re1 , F1 ). We suppose that ξ1 is of the form ξ1 = h(S1 )e1 where
h is a continuous function. The problem we propose to solve is to characterize
the set of all prices p0 ∈ Γ(h) such that p0 e1 − g0 − g1 = h(S1 )e1 for some
gt ∈ L0 (Gt , Ft ), t = 0, 1, i.e.

Γ(h) = {p0 ∈ R : p0 − h(S1 ) + L1 (−g0 ) ≥ 0}.

Notice that p0 ∈ Γ(h) if and only if p0 ≥ p0 (g0 ) = ess supF0 (h(S1 ) − L1 (−g0 )),
where the notion of essential supremum is given in [15, Section 5.3.1]. More-
over, Γ(h) is an interval. Our goal is to determine inf Γ(h). By [1, Proposition
2.13], we have
p0 (g0 ) = sup g(g0 , s), g0 ∈ G0 ,
s∈supp (S1 )

where supp (S1 ) is the support of S1 and

g(g0 , s) = h(s) + γ(g0 , s), g0 = (x0 , y0 ),

γ((x0 , y0 ), s) = x0 − (y0 s(1 − ) + c)− + y0+ s(1 + ) + c1y0 >0 .

In the following, we suppose that h(s) = (s − K)+ , K ≥ 0, and we use the

 / 23

notation g0 = (x0 , y0 ). We suppose that supp (S1 ) = [S1min , S1max ]. We have:

g(g0 , s) = g 1 (g0 , s) = x0 + y0 s(1 + ) + c + (s − K)+ , y0 > 0,

−c
= g 2 (g0 , s) = x0 + (s − K)+ , 0 < s ≤ , y0 < 0
y0 (1 − )
−c
= g 3 (g0 , s) = x0 + y0 s(1 − ) + c + (s − K)+ , s > , y0 < 0.
y0 (1 − )

Therefore,

p0 (g0 ) = g 1 (g0 , S1max ), y0 > 0,

 
2 −c min −1
p0 (g0 ) = g g0 , ∨ S1 , y0 ≤
y (1 − ) 1−
 0  
2 −c min 3 max −1
p0 (g0 ) = max g g0 , ∨ S1 , g (g0 , S1 ) , 0 > y0 > .
y0 (1 − ) 1−

Notice that g0 = (x0 , y0 ) ∈ G0 if and only if x0 + L0 ((0, y0 )) ≥ 0, i.e.

x0 ≥ δ(y0 ) := −L0 ((0, y0 )). Therefore,

p∗0 = inf Γ(h) = inf p0 (δ(y0 ), y0 ).

y0 ∈R

When computing p∗0 , we obtain the argmin y0 and x0 = δ(y0 ) such that
g0 = (x0 , y0 ). For instance, with c = 1.5,  = 5% and K = 50, we get
that g0 = (64.05, −0.61) and p∗0 = 74. With c =  = 0 and K = 50, we
get g0 = (56.99, −0.5699) and p∗0 = 65.27. In Table 2.3, minimal prices are
computed.

3. Conclusion

We have discovered the main arguments and tools allowing to characterize no-
arbitrage conditions and then deduce dual characterizations of super-hedging
prices. It was possible to do it because the set of terminal claims is a closed
convex cone under N A or other stronger condition. In practice, the transac-
tion costs are not necessarily linear so that the solvency set G is not a cone.
Then, new approaches need to be invented. One of them could be to use
the natural stochastic preorder generated by G, i.e. x ≥Gt y if and only if
x − y ∈ Gt , see [28] and [29]. We also presented a new approach that should
be generalized.
 / 24

Fig 1. The price function y0 7→ p0 (δ(y0 ), y0 ) for y0 ∈ [−1, 1]. The parameters are c = 1.5,
K = 50,  = 5%.

K=30 K=50 K=70 K=100

c =  = 0% p∗
0 = 85.27 p∗
0 = 65.27 p∗
0 = 49.96 p∗
0 = 28.21
∗ ∗ ∗ ∗
c = 1.5,  = 0% p0 = 85.72 p0 = 66.8 p0 = 50.28 p0 = 28.8
c = 1.5,  = 1% p∗0 = 88 p∗0 = 68 p∗
0 = 49.22 p∗0 = 34.5
c = 1.5,  = 5% p∗0 = 91.8 p∗0 = 74 p∗0 = 56.8 p∗
0 = 31.15
c = 1.5,  = 10% p∗0 = 98.8 p∗0 = 79.2 p∗0 = 60 p∗0 = 34.1

Fig 2. Numerical computation of the minimal prices for several parameters.

For readers who wish to deepen their knowledge on arbitrage theory, a

list of references is given in the bibliography. Among the very well known
authors among others 2 (currently) working on arbitrage theory, we may cite
Bouchard B, Campi L., Cherny A., Delbaen F. , Guasoni P., Kabanov Y. ,
Rásonyi M., Schachermayer W., Touzi N.

References

[1] Baptiste J., Carassus L. and Lépinette E. Pricing without martingale

measures. Preprint arXiv:1807.04612.
[2] Burzoni M. Arbitrage and Hedging in model-independent markets with
frictions. Siam Journal of Financial Mathematics, 2016, 7(1), 812-844.
[3] Campi L., Schachermayer W. A super-replication theorem in Kabanov’s
2
My apologies if you are not cited, I am sure that you are very famous.
 / 25

model of transaction Costs. Finance and Stochastics, 2006, 10, 4, 579-

596.
[4] Cherny A. General arbitrage pricing model. II. Transaction costs, in
’Séminaire de Probabilités XL’, 2007, Vol. 1899 of Lecture Notes in
Math., Springer, Berlin, 447–461.
[5] Dalang E.C., Morton A. and Willinger W. Equivalent martingale mea-
sures and no-arbitrage in stochastic securities market models. Stochas-
tics and Stochastic Reports, 1990, 29, 185-201.
[6] Delbaen F. and W. Schachermayer. The Mathematics of Arbitrage.
Springer Finance, 2006.
[7] De Vallière D., Kabanov Y. and Denis (Lépinette) E. Hedging of Amer-
ican options under transaction costs. Finance and Stochastics, 2009, 13,
1, 105-119.
[8] Grépat J, Kabanov Y. Small transaction costs, absence of arbitrage and
consistent price systems. Finance and Stochastics, 2012, 16, 3, 357-368.
[9] Grigoriev P. On low dimensional case in the fundamental asset pricing
theorem under transaction costs. Statist. Decisions, 23 (2005), 1, 33-48.
[10] Guasoni P., Lépinette E. and Rásonyi M. The fundamental theorem of
asset pricing under transaction costs. Finance and Stochastics, 2012,
16, 4, 741-777.
[11] Guasoni P., Rásonyi M. and Schachermayer, W. The fundamental the-
orem of asset pricing for continuous processes under small transaction
costs. Annals of Finance, 2010, 6, 2, 157-191.
[12] Harrison J.M. and S. Pliska. Martingales and Stochastic Integrals in
the Theory of Continuous Trading, Stochastic Processes and their Ap-
plications, 11, 215-260, 1981.
[13] Jeanblanc M. Cours de calcul stochastique. https://www.maths.
univ-evry.fr/pages_perso/jeanblanc/cours/M2_cours.pdf
[14] Jouini E., Kallal H. Martingales and arbitrage in securities markets
with transaction costs. J. Econ. Theory, 1995, 66, 178-97.
[15] Kabanov Y., Safarian M. Markets with transaction costs. Mathematical
Theory. Springer-Verlag, 2009.
[16] Kabanov Y., Lépinette E. Consistent price systems and arbitrage op-
portunities of the second kind in models with transaction costs. Finance
and Stochastics, 16, 2011, 1, 135-154.
[17] Klein I., Schachermayer W. A quantitative and a dual version of the
Halmos-Savage theorem with applications to mathematical finance.
The Annals of Probability, 1996, 24, 2, 867-881.
 / 26

[18] Kreps D. Arbitrage and equilibrium in economies with infinitely many

commodities. Journal of Mathematical Economics 8, 15-35, 1981.
[19] Lamberton D., Lapeyre B. Introduction to stochastic calculus applied
to finance. Chapman and Hall/CRC Financial Mathematics Series,
ed.2, CRC Press, 2011.
[20] Lépinette E. Robust no arbitrage of the second kind with a continuum
of assets and proportional transaction costs. SIAM Journal on Financial
Mathematics, 2016, 7, 1, 104-123.
[21] Pennanen T. Convex duality in stochastic optimization and mathe-
matical Finance. Mathematics of Operations Research, 36(2), 340-362,
2011.
[22] Przemyslaw R. Arbitrage in markets with Bid-Ask spreads. The fun-
damental theorem of asset pricing in finite discrete time markets with
Bid-Ask spreads and a money account. Annals of Finance, 2015, 11,453-
475.
[23] Rokhlin D.B. Martingale selection problem and asset pricing in finite
discrete time. Electronic Communications in Probability, 2007,12, 1-8.
[24] Rogers L.C.G. Equivalent martingale measures and no-arbitrage.
Stochastics and Stochastic Reports, 1994, 51,1-2, 41-49.
[25] Schachermayer W. The fundamental theorem of asset pricing under
proportional transaction costs in finite discrete time. Mathematical Fi-
nance, 2004, 14, 1, 19-48.
[26] Schachermayer W. A Hilbert space proof of the fundamental theorem
of asset pricing in finite discrete time. Insurance: Mathematics and
Economics, 1992, (11) 4, 249-257.
[27] Lépinette E., Zhao J. A complement to the Grigoriev theorem for the
Kabanov model. Preprint 2018, hal-01666860v6.
[28] Lépinette E., Tran Q.T. Arbitrage theory for non convex financial
market models. Stochastic Processes and Applications, 127 (2017), 10,
3331-3353.
[29] Lépinette E., Tran Q.T. General financial market model defined by a
liquidation value process. Stochastics, 88 (2016), 3, 437-459.
[30] Rásonyi M. Arbitrage with transaction costs revisited. Optimality
and Risk: Modern Trends in Mathematical Finance. Eds. Delbaen F.,
Ràsonyi M., Stricker Ch. Springer, Berlin–Heidelberg–New York, 2009.