Contemporary Mathematics

Optimal Derivatives Design under Dynamic Risk Measures

Pauline Barrieu and Nicole El Karoui

Abstract. We develop a methodology to optimally design a financial issue to

hedge non-tradable risk on financial markets. Economic agents assess their risk
using monetary risk measure. The inf-convolution of convex risk measures is
the key transformation in solving this optimization problem. When agents’ risk
measures only differ from a risk aversion coefficient, the optimal risk transfer
is amazingly equal to a proportion of the initial risk.
For dynamic risk measures defined through their local specifications using
BSDE’s, their inf-convolution is equivalent to that of their associated drivers.
In this case, it is also possible to characterize the optimal risk transfer.

Introduction
In recent years, a new type of financial instruments (among them, the so-called
“insurance derivatives”) has appeared on financial markets. Even though they
have all the features of financial contracts, they are very different from the classical
structures. Their underlying risk is indeed related to a non-financial risk (natural
catastrophe, weather event...), which may somehow be connected to more tradi-
tional financial risks. Their high level of illiquidity, deriving partly from the fact
that the underlying asset is not traded on financial markets, makes them difficult
to evaluate and to use. Several authors (see, for instance, D. Becherer [Be1], M.
Davis [Da2] or M. Musiela and T. Zariphopoulou [MuZ]) have been interested in
these new products, especially in their pricing. However, neither their impact on
“classical” investments nor their optimal design are mentioned in the literature.
On the other hand, this accrued complexity of financial products has naturally lead
to an increasing interest in quantitative methods of assessing the risk related to a
given financial position.
This paper focuses on these problems in a framework where economic agents may
take positions on two types of risk: a purely financial risk (or market risk) and a
(non-financial) non-tradable risk. The optimal structure of a contract depending
on the non-tradable risk and its price are determined.
Since the structure represents a new diversification instrument for any investor,

1991 Mathematics Subject Classification. Primary 60Gxx, 91B28, 91B90; Secondary 46N10,
90C39.
Key words and phrases. Static and dynamic risk measure, non-tradable risk, optimal design,
illiquidity, inf-convolution, BSDE.

c 0000 (copyright holder)

1
2 PAULINE BARRIEU AND NICOLE EL KAROUI

optimal wealth allocation becomes a more complex question and the question of
an efficient quantitative risk assessment becomes crucial. Different authors have
recently been interested in defining and constructing a coherent, in some sense, risk
measure (see, for instance, Artzner et al. [ADEH] or Föllmer and Schied [FS1]),
using a systematic axiomatic approach. The framework developed by these authors
will be that of this study.

This paper is structured as follows: in the first section, after having recalled
some basic properties of convex risk measures, we generate new risk measures as
the inf-convolution of convex risk measures. Then, in the second section, we solve
the problem of an optimal non-tradable risk transfer. In the third section, we
introduce dynamic risk measures defined through their local specifications with the
help of Backward Stochastic Differential Equations in order to propose a method
to characterize completely the optimal structure.

1. Risk transfer and inf-convolution of risk measures

In this section, we first present a general class of risk measures introduced by
Föllmer and Schied ([FS1] and [FS2]). Then, we generate new risk measures as
the inf-convolution of different risk measures. We finally apply these results to
the optimal design of a transaction based on a non-tradable risk. In particular,
we obtain a necessary and sufficient condition to the existence of an optimal risk
transfer.

1.1. Convex risk measures.

1.1.1. Definition and basic properties . We first recall the definition and some
key properties of the convex risk measures introduced by Föllmer and Schied ([FS1]
and [FS2]). In the following, X denotes a linear space of bounded functions in-
cluding constant functions, defined on the measurable space (Ω, F).
Definition 1.1. The functional ρ : X → R is a convex risk measure in the
sense of Föllmer and Schied if, for any X and Y in X , it satisfies the following
properties:
a) Convexity: ∀λ ∈ [0, 1] ρ (λX + (1 − λ) Y ) ≤ λρ (X) + (1 − λ) ρ (Y );
b) Monotonicity: X ≤ Y ⇒ ρ (X) ≥ ρ (Y );
c) Translation invariance: ∀m ∈ R ρ (X + m) = ρ (X) − m.
Intuitively, ρ (Ψ) may be interpreted as the amount the agent has to hold to
completely cancel the risk associated with her risky position Ψ
ρ (Ψ + ρ (Ψ)) = 0
The risk measure ρ induces a particular set of positions: the acceptance set, Aρ ,
defined as the set of all acceptable positions as they carry no positive risk:
(1.1) Aρ = {Ψ ∈ X , ρ (Ψ) ≤ 0}
We now present a key result obtained by Föllmer and Schied [FS2] (Theorem 4.12)
in the following Theorem.
Theorem 1.2. Let M1,f be the set of all additive measures on (Ω, F). Another
formulation of the convex risk measure is given in terms of a penalty function, α (Q)
 OPTIMAL DERIVATIVES DESIGN UNDER DYNAMIC RISK MEASURES 3

taking values in R ∪ {+∞}:

(1.2) ∀Ψ ∈ X ρ (Ψ) = sup {EQ (−Ψ) − α (Q)}
Q∈M1,f

By duality between M1,f and X ,

(1.3) ∀Q ∈M1,f α (Q) = sup {EQ (−Ψ) − ρ (Ψ)} (≥ ρ (0))
Ψ∈X

or equivalently
(1.4) ∀Q ∈M1,f α (Q) = sup EQ (−Ψ)
Ψ∈Aρ

In the following, we are especially interested in risk measures related to prob-

ability measures. In general, the assumption of decreasing continuity from below
is made and suffices to imply that the dual formulation of risk measure (Equation
(1.2)) is satisfied for Q ∈ M1 , where M1 is the set of all probability measures on
the considered space. In this case, the equations previously obtained concerning
the penalty function (Equations 1.4 and 1.3) still hold replacing M1,f by M1 .
Moreover, Föllmer and Schied have proven in [FS2] (Theorem 4.12) that there al-
ways exists a measure of M1,f such that the supremum in Equation (1.2) is reached.
When working with M1 , the supremum is reached under some conditions presented
in Theorem 4.22 of [FS2]. These results will be quite important in the following
as they ensure the existence of an “optimal” measure (or “optimal” probability
measure under some assumptions).
Example 1.3. A classical example of convex risk measure is the entropic risk
measure
(1.5)   
1
∀Ψ ∈ X eγ (Ψ) = sup (EQ (−Ψ) − γh (Q/P)) = γ ln EP exp − Ψ
Q∈M1 γ
where γ is the risk tolerance coefficient and h (Q/P) is the relative entropy1 of Q
with respect to the prior probability P.
1.1.2. Risk measure generated by a convex set and coherent risk measure . We
now introduce some particular convex risk measures generated by a convex set as
follows
Definition 1.4. Let H be a non-empty convex subset of X such that
inf {m ∈ R, such that ∃ξ ∈ H, m ≥ ξ} > −∞
Then the functional v H defined as
v H (Ψ) = inf {m ∈ R; such that ∃ξ ∈ H, m + Ψ ≥ ξ}
is a convex risk measure. The associated penalty function α is given by:
∀Q ∈M1,f αH (Q) = sup EQ (−H)
H∈H

1When finite (i.e. if Q  P), the relative entropy is defined by

 
dQ dQ
h (Q/P) = EP ln
dP dP
4 PAULINE BARRIEU AND NICOLE EL KAROUI

and the acceptance set is defined by

AvH = {Ψ ∈ X , ∃ξ ∈ H, m + Ψ ≥ ξ}
When H is a cone, the penalty function associated with v H can only take two
possible values
α (Q) = 0 if Q ∈QH and +∞ otherwise
where QH is the set of all additive measures such that ∀ξ ∈ H, EQ (ξ) ≥ 0. The
risk measure v H is then coherent 2 in the sense of Artzner et al. ([ADEH]) and its
dual formulation is simply given by
∀Ψ ∈ X v H (Ψ) = sup EQ (−Ψ)
Q∈QH

1.1.3. Inf-convolution of risk measures. Rockafellar ([Ro]) has given some sta-
bility properties of the inf-convolution of convex functions. The following Theorem
extends these results to the inf-convolution of convex functionals:
Theorem 1.5. Let ρ1 and ρ2 be two convex risk measures with respective
penalty functions α1 and α2 . Let ρ1,2 be the inf-convolution of ρ1 and ρ2 defined as
Ψ → ρ1,2 (Ψ) ≡ ρ1 ρ2 (Ψ) = inf {ρ1 (Ψ − H) + ρ2 (H)}
H∈X

and assume that ρ1,2 (0) > −∞. Then ρ1,2 is a convex risk measure, which is finite
for all Ψ ∈ X . The associated penalty function is given by
∀Q ∈M1,f α1,2 (Q) = α1 (Q) + α2 (Q)
Note that the convex risk measure ρ1,2 may also be defined as the value func-
tional of the program
ρ1,2 (Ψ) = inf {ρ1 (Ψ − H) , H ∈ Aρ2 }

Proof. Please refer to Barrieu-El Karoui [BEK2]. 

Moreover, using Subsection 1.1.2, the following result is a direct consequence

of Theorem 1.5:
Corollary 1.6. Let H be a cone of X and ρ be a convex risk measure with
penalty function α such that
inf {ρ (−H) , H ∈ H} > −∞
The inf-convolution of ρ and ν H , ρH ≡ ρν H , also defined as
ρH (Ψ) ≡ inf {ρ (Ψ − H) , H ∈ H} = sup {EQ (−Ψ) − α (Q)}
Q∈QH

is a convex risk measure with penalty function α on QH and +∞ otherwise.

2It satisfies indeed the positive homogeneity property ∀Ψ ∈ X , ∀λ ≥ 0, v H (λΨ) =

λv H (Ψ). (For more details, please refer to Föllmer and Schied [FS2], Remark 4.13). This
property simply translates the fact that the size of the transaction or exposure does not have any
particular impact.
 OPTIMAL DERIVATIVES DESIGN UNDER DYNAMIC RISK MEASURES 5

Remark 1.7. This property may be interpreted in terms of hedging strategies.

The inf-convolution ρH is simply the residual risk measure after having optimally
chosen the hedging strategy for Ψ with elements of H. For instance, one may see H
as the cone of all gain processes related to a given financial market. In the following,
we will refer to ρH (sometimes denoted ρm ) as the modified market risk measure,
when appropriate.
1.1.4. Market modified risk measure in the entropic framework.
Incomplete market. In this entropic framework (corresponding to an ex-
ponential utility function), this problem has been widely studied as an hedging
problem when H is the space of bounded gain processes written on locally bounded
semi-martingale price processes S, at a future time T ,
 
 ZT 
VT = ξT = hϕt , dSt i ; ϕt ∈ Kt
 
0

associated with financial strategies, ϕ, satisfying some geometric constraints de-

scribed by Kt . Several authors (for instance, Frittelli [Fr1], Delbaen et al. [DGRSS],
Becherer [Be1]) have solved the dual problem (1.6 to obtain the existence of an op-
timal hedging strategy. El Karoui-Rouge [EKR] have established the same type of
results and characterized the solution using a different approach based on BSDE’s
techniques presented in the last section of this paper. Musiela-Zariphopoulou
[MuZ] have also been interested in this problem and completely solved a particular
example using PDE’s.
Partial information. Some economic agents, unable to observe the non-
tradable risk, base their financial strategies only on the information contained in the
financial asset prices S. In particular, they can observe the filtration FTS generated
by σ(Su ; 0 ≤ u ≤ T ). A measurability constraint is then added to the geometric
constraint on the financial strategies. The space of gain processes is denoted by
VTS :
 
 ZT 
VTS = ξT = hϕt , dSt i ; ϕt ∈ FtS and ϕt ∈ Kt
 
0
In the entropic case, it is possible to solve the problem as previously, given that
 
em,S (Ψ) = inf{γ ln EP exp(− γ1 (Ψ − ξ)) ; ξ ∈ VTS }
 
= inf{γ ln EP exp(− γ1 (ΨS − ξ)) ; ξ ∈ VTS }

where    
S 1 S
Ψ ≡ −γ ln EP exp − Ψ /=T
γ
is the opposite of the conditional entropic risk measure of Ψ given FTS , assessing
the cost of partial information.
In the so-called “filtering framework”, the financial assets’ prices S are associated
with a risk premium depending on the non-tradable risk. Different authors (see
for instance Lakner [La1] [La2], Lefèvre [Le], Pham-Quenez [PhQ]) have shown
however that there exists a “risk-neutral” probability measure Q b T such that the
S H
= -market is complete. The set Q is then the set of all probability measures on
6 PAULINE BARRIEU AND NICOLE EL KAROUI

the considered measurable space (Ω, =) such that their restriction to =S is the risk-
neutral probability measure Qb T . In particular, the market modified risk measure
may be written as
    
m,S 1 S
e (γ, Ψ) = γEQb T ln EP exp − Ψ /=T
γ
For more details, please refer to Barrieu-El Karoui [BEK4].

2. Optimal design problem

In the following, we focus on the question of an optimal transaction between two
economic agents. These agents, respectively denoted A and B, are evolving in an
uncertain universe modeled by a probability space (Ω, =, P). At a fixed future date
T , agent A is exposed towards a non-tradable risk Θ for an amount X ≡ X (Θ, ω)
in the scenario ω. A wants to issue a financial product F ≡ F (Θ, ω) and sell it to
agent B for a forward price at time T denoted by π as to reduce her exposure. We
assume that X and F belong to X .

2.1. General framework. Both agents assess the risk associated with their
respective positions by a convex risk measure, denoted respectively ρA and ρB , with
associated penalty functions αA and αB .
The issuer, agent A, wants to determine the structure (F, π) as to minimize her
global risk measure
min ρA (X − F + π)
F ∈X ,π
while the issuer’s constraint related to the buyer’s interest in doing the transaction
may be written as
ρB (F − π) ≤ ρB (0)
This constraint now imposes a maximum threshold to the risk the buyer accepts to
bear.
We now consider a more general framework where both agents may also invest
in the financial market in order to reduce their respective exposure. They choose
(A) (B)
optimally their financial investments via two cones VT and VT , characterizing
the terminal gains associated with their respective admissible financial strategies.
This opportunity to invest optimally in a financial market has a direct impact on the
risk measure of both agents as previously mentioned. As a consequence, provided
the condition inf ξB ∈V (B) ρB (ξB ) > −∞ and inf ξA ∈V (A) ρA (ξA ) > −∞
T T
we are exactly in the framework of Corollary 1.6 and both agents simply assess their
non-tradable exposure using a market modified risk measure, denoted respectively
by ρm m
A and ρB . The optimization program related to the F -transaction simply
becomes
inf ρm
A (X − F + π) subject to ρm m
B (F − π) ≤ ρB (0)
F ∈X ,π

Using the cash translation invariance property and binding the constraint at the
optimum, the pricing rule of the F -structure is fully determined by the buyer as
(2.1) π ∗ (F ) = ρm m
B (0) − ρB (F )

It corresponds to an “indifference” pricing rule from the point of view of agent B’s
market modified risk measure.
 OPTIMAL DERIVATIVES DESIGN UNDER DYNAMIC RISK MEASURES 7

Using again the cash translation invariance property, the optimization program
simply becomes
inf (ρm m m
A (X − F ) + ρB (F ) − ρB (0))
F ∈X
We are almost in the framework of Theorem 1.5, apart from the constant ρm
B (0).
To deal with it, we consider the reduced program3
m
(2.2) RAB (X) = inf (ρm m m m
A (X − F ) + ρB (F )) = (ρA ρB ) (X)
F ∈X
m
The value functional RAB of this program may be interpreted as a measure of the
residual risk after all transactions.
A direct consequence of Theorem 1.5 is now given:
Proposition 2.1. The inf-convolution of the risk measures ρm m m
A and ρB , RAB (X),
4
is a convex risk measure with the penalty function given for any Q in Q ∩ Q(B)
(A)

by
m
αAB (Q) = αA (Q) + αB (Q) , +∞ otherwise
2.2. Characterization of the optimal structure.
2.2.1. Generalized entropic framework. In the case of the entropic risk measure
eγ defined by Equation (1.5), we easily obtain the following semi-group property
eγ eγ 0 = eγ+γ 0
More generally, let ρ be a convex risk measure with penalty function α. The risk
measure ργ with penalty function γα, is equal to the “right scalar multiplication”
of ρ defined by Rockafellar ([Ro]), more precisely:
 
1
(2.3) ∀Ψ ∈ X ργ (Ψ) = γρ Ψ
γ
In this family of convex risk measures, by duality, the inf-convolution defines a new
convex risk measure of the same family: for any (γ, γ 0 ), strictly positive, indeed,
the following stability property holds
ργ ργ 0 = ργ+γ 0
In this case, the optimal structure F ∗ realizing the inf-convolution (2.2) may be
explicitely obtained:
Proposition 2.2. When both agents have the same access to the financial
market and have market modified risk measures of the type described above by (2.3),
the optimal structure F ∗ is given by:
γB
F∗ = X P a.s.
γA + γB
Proof. The result is immediately obtained by checking that:
∗ ∗
ρm m
A (X − F ) + ρB (F ) = γA ρm ( γAX m X
+γB ) + γB ρ ( γA +γB )
m X
= (γA + γB )ρ ( γA +γB )
= (ρm m
A ρB ) (X)
Hence, the optimality of F ∗ is deduced. 
3The value functional obtained in this case should be translated by the constant −ρm (0) in
B
order to obtain the value function of the previous program.
4Note that Q(A) ∩ Q(B) is the set of all additive measures Q such that ∀ξ ∈ V (A) ∩V (B) ,
T T
EQ (ξ) ≥ 0.
8 PAULINE BARRIEU AND NICOLE EL KAROUI

Interpretation: when both agents have the same access to the financial market,
the underlying logic of this new asset class is that of insurance and is far away from
that of speculation. The issuer has an interest to sell a structure if and only if she
is initially exposed (or, more precisely, if her initial exposure differs from that of
the buyer). The underlying logic is that of insurance and hedging. It is by no way a
speculative logic and the sale of this type of contract aims to hedge a real exposure
towards a non-financial risk.
2.2.2. Characterization of the optimal structure in a general framework. We
now consider a more general case and find some conditions to have an optimal
structure F ∗ realizing inf-convolution RAB
m
(X) for a given X. First let us give two
definitions of optimality and precise the dual relationship between exposure and
additive measure:
Definition 2.3. Given a convex risk measure ρ and its associated penalty
function α, we say

i) that the measure QΨρ is optimal for (Ψ, ρ) if ρ (Ψ) = EQΨ
ρ
(−Ψ) − α QΨ ρ .
ii) that the exposure Ψ is optimal for (Q, α) if α (Q) = EQ (−Ψ) − ρ (Ψ).
Let QX m
AB be the optimal measure for for (X, RAB ), the existence of which has
been mentioned in Subsection 1.1.1.
The following Theorem presents a necessary and sufficient condition to have an
optimal structure F ∗ in terms of this optimal measure QX
AB .

Theorem 2.4. Let

n o
∗
QX m
QX m



CA AB = F ; αA AB = EQX
AB
(− (X − F )) − ρ A (X − F )
n o
∗ X m X m



CB QAB = F ; αB QAB = EQX
AB
(−F ) − ρ B (F )
The necessary and sufficient condition to have an optimal solution to the inf-
convolution problem described in the Program (2.2) is
∗ ∗
QX X



CA AB ∩ CB QAB 6= ∅
Moreover, denoting by F ∗ an optimal solution for the inf-convolution problem, the
following relationships prevail:
∗ ∗
X−F
QX
AB ” = ” QA ” = ” QF
B

Proof. Please refer to Barrieu-El Karoui [BEK3]. 

This Theorem gives a procedure to obtain the optimal structure. For a given
X, there exists an optimal measure QX m
AB for RAB . This measure is necessary
∗
and sufficient to have an optimal structure F . Hence, the solutions of the inf-
convolution problem are determined by the dual formulation of the residual risk
m
measure RAB . Note also that the last equalities translate the fact that both agents
valuate their respective residual risk using the same measure QX AB . This enables
the transaction.

3. Solving the inf-convolution problem in a dynamic framework

We come back in this section to the inf-convolution problem when considering
various classes of convex functionals. We adopt dynamic programing techniques,
in particular Backward Stochastic Differential Equations (BSDE’s), to study risk
measures defined by their local specifications and propose a method to characterize
 OPTIMAL DERIVATIVES DESIGN UNDER DYNAMIC RISK MEASURES 9

the optimal solution of the inf-convolution problem. This approach leads to a

particular definition of dynamic risk measures.
3.1. Localization of convex risk measures.
3.1.1. Introduction. On the probability space (Ω, =, P), let consider a Brownian
filtration (=t = σ (Ws ; 0 ≤ s ≤ t) ; t ≥ 0). This enables us to extend the notion of
static entropic risk measure to a more local and dynamic one
   
1
eγ,t (X) = γ ln EP exp − X /=t , X∈X
γ
with the terminal condition
eγ,T (X) = −X
The dynamics of the adapted process (eγ,t (X); t ∈ [0, T ]) is given by the follow-
1 2
ing BSDE with the quadratic driver f (t, z) = 2γ kzk , using stochastic calculus
arguments:
1 2
−deγ,t (X) = kzt k dt−hzt , dWt i with the terminal condition eγ,T (X) = −X
2γ
The idea is then to introduce families of solutions driven by convex generators
f (t, z) of the same kind. As we will see, they generate local convex risk measures.
Under some regularity assumptions5 for the function f , implying the existence and
uniqueness of a solution to this BSDE, the key tool is the so-called Comparison
Theorem. It corresponds to the maximum principle when working with PDEs.
Theorem 3.1. Consider the general BSDE with the solution (Yt , zt )
(3.1) −dYt = f (t, zt ) dt − hzt , dWt i with terminal condition YT = X
Let X1 and X2 be two elements of X and f 1 and f 2 two “regular” drivers. We
denote by (Y 1 , z 1 ) and (Y 2 , z 2 ) the associated
solutions.

We assume that X1 ≥ X2 P a.s. and f 1 t, zt2 ≥ f 2 t, zt2 dt × dP a.s.. Then we
have
∀t ≥ 0 Yt1 ≥ Yt2
3.1.2. Local specification. We are now able to generalize the notion of static
convex risk measure to a more dynamic notion, by considering the BSDE’s solutions
as functional of their terminal condition. More precisely, thanks to the Comparison
Theorem 3.1, properties as monotonicity, convexity of the drivers are also satisfied
by the solution.
Theorem 3.2. Suppose that the regular driver f (t, z) is convex w.r. to z.
The solution (ρt (X))t ≤ T of the BSDE (3.1) with terminal conditon −X.
−dρt (X) = f (t, zt ) dt − hzt , dWt i >, ρT (X) = −X
is for any time t a convex risk measure.
The convexity result has been proved in the study of pricing functionals with
constraints (see, for instance, El Karoui-Quenez [EKQ], Peng [P] or Gianin [Gi]).
The cash invariance property is a consequence of the independence between f and
ρ.
5For example, f uniformly Lipschitz or with quadratic growth, as shown by El Karoui-Quenez
[EKQ] , Kobylanski [Kob] or Lepeltier-San Martin [LSM a].
10 PAULINE BARRIEU AND NICOLE EL KAROUI

3.2. Inf-convolution. In this subsection, given f 1 and f 2 two regular convex

drivers, we compare the solution of different BSDE’s related to


(1) −dρ1t (Ψ) = f 1 t, zt1
dt − zt1 , dWt , ρ1T (Ψ) = −Ψ
2 2 2 2
(2) −dρt (Ψ) = f t, zt dt − zt , dWt , ρ2T (Ψ) = −Ψ
In particular, we will study for any t the inf-convolution of the convex functionals
ρ1t and ρ2t defined as
ρ1 ρ2 t (X) = inf ρ1t (X − F ) + ρ2t (F )


(3.2)
F
The first step is to introduce the following BSDE
−dρ1,2 1 2
ρ1,2


(3.3) t (X) = f f (t, zt ) dt − hzt , dWt i , T (X) = −X
where the driver f 1 f 2 (t, z) is the inf-convolution between both convex functions
f 1 (t, z) and f 2 (t, z). In that follows, we assume its regularity.
The next step is to verify that, under some additional assumptions, the solution
of BSDE generated by f 1 f 2 is the inf-convolution ρ1 ρ2 of the dynamic risk
measures ρ1 and ρ2 .
Theorem 3.3. For a given X ∈ X , let ρ1,2


t (X) , zt be the solution of (3.3)
with a regular driver f 1 f 2 and terminal condition −X.
Then, the following results hold:
i) For any (F, t) ∈ X × [0, T ], ρ1,2 1
t (X) ≤ ρt (X − F ) + ρt (F )
2
P a.s.
ii) If there exists an admissible zbt2 such that
f 1 f 2 (t, zt ) = f 1 t, zt − zbt2 + f 2 t, zbt2 .



∀t ≥ 0
then
ρ1,2 1 2


∀t ≥ 0 t (X) = ρ ρ t (X) P a.s
iii) Under this assumption, let F ∗ the structure defined by the forward equation
ZT ZT
F∗ = f 2 t, zbt2 dt − zbt2 , dWt

0 0
∗
Then F is an optimal solution for the inf-convolution problem (3.2) of the dynamic
risk measures.
Proof. i) By definition of inf-convolution,
f 1 f 2 (t, z) ≤ f 1 (t, z − y) + f 2 (t, y)


∀ (t, z, y)
On the other hand, for any F ∈ X , ρ1t (X − F ) + ρ2t (F ) satisfies



−d ρ1t (X − F ) + ρ2t (F ) = f 1 (t, zt1 ) + f 2 (t, zt2 ) dt −
hzt1 + zt2 , dWt i
= f 1 (t, zt − zt2 ) + f 2 (t, zt2 ) dt − hzt , dWt i
with terminal condition −X.
The second formulation expresses that ρ1t (X − F ) + ρ2t (F ) is solution of a BSDE
with terminal condition −X and a driver written as f 1 (t, zt − zt2 ) + f 2 (t, zt2 ) where
zt2 is fixed as the solution of the BSDE (2) with terminal condition −F .
This driver is always greater than that of the BSDE (3.3) and their respective
terminal conditions are identical. Thanks to the Comparison Theorem 3.1, the
result is obtained.
ii) and iii) Let us assume that there exists an admissible zbt2 such that
f 1 f 2 (t, zt ) = f 1 t, zt − zbt2 + f 2 t, zbt2




∀t ≥ 0
 OPTIMAL DERIVATIVES DESIGN UNDER DYNAMIC RISK MEASURES 11

where zt is the solution of (3.3) with the terminal condition −X.

We now introduce the structure F ∗ defined by the forward equation
ZT ZT
∗ 2 2
zbt2 , dWt


F = f t, zbt dt −
0 0
Rt
Rt 2
Let us observe that −Rt2 ≡ f2 u, zbu2 du − zbu , dWu satisfies
0 0
ZT ZT
∗ ∗
Rt2 2
u, zbu2 zbu2 , dWu


(F ) = −F + f dt −
t t

F is solution of the BSDE with driver f and terminal condition −F ∗ .

∗


Then Rt2 2

By uniqueness, this process is the dynamic risk measure ρ2t (F ∗ ).

We have seen in i) that ρ1t (X − F ∗ ) + ρ2t (F ∗ ) is solution of the BSDE with dri-
ver written

as f 1 (t, zt − zbt2 ) +
f 2 (t, zbt2 ) and

terminal condition −X. Given that
f f (t, zt ) = f 1 t, zt − zbt2 + f 2 t, zbt2 , by uniqueness, the following equality
1 2

holds
ρ1,2 1 2


∀t ≥ 0 t (X) = ρ ρ t (X) P a.s
The proof also gives the optimality for the Problem (3.2) of the structure
ZT ZT
∗ 2
t, zbt2 zbt2 , dWt


F = f dt −
0 0

Given the results of Proposition 2.2, it is natural to study which assumptions
on the driver of such dynamic risk measures lead to a non-speculative logic6.
To simplify the arguments, we now consider normalized risk measures, i.e.
∀t ρ1t (0) = ρ2t (0) = 0
Corollary 3.4. Assume that f 1 (t, 0) = f 2 (t, 0) = 0 and ∂z f 1 (t, 0) = ∂z f 2 (t, 0) =
0, then:

i) The inf-convolution f 1 f 2 (t, 0) and that of the associated risk measure (ρ1 ρ2 ) (t, 0)
are identically null.
Moreover, F ∗ ≡ 0 is an optimal solution for the inf-convolution problem (3.2).
ii) If both drivers f 1 and f 2 are strictly convex, then F ∗ ≡ 0 is the unique optimal
solution for the inf-convolution problem (3.2).
Proof. i) Since f 1 (t, 0) = f 2 (t, 0) = 0, we have
f 1 f 2 (t, 0) = inf {f 1 (t, −y) + f 2 (t, y)} ≤ 0


y

On the other hand,

∀y f 1 (t, −y) + f 2 (t, y) ≥ (−∂z f 1 (t, 0) + ∂z f 2 (t, 0))y = 0
using the assumption

∂z f 1 (t, 0) = ∂z f 2 (t, 0) = 0.
1 2
Hence f f (t, 0) = 0 and as a consequence (ρ1 ρ2 ) (t, 0) = 0.
Moreover, as y = 0 is a solution of f 1 (t, −y) + f 2 (t, y) = 0, by construction, F ∗ ≡ 0
6By non-speculative logic, we simply mean that the issuer has an interest to sell a structure
if and only if she is initially exposed. The underlying logic is that of insurance and hedging.
12 PAULINE BARRIEU AND NICOLE EL KAROUI

is an optimal solution for the inf-convolution problem (3.2).

ii) When both drivers f 1 and f 2 are strictly convex, then y = 0 is the unique
solution of f 1 (t, −y) + f 2 (t, y) = 0 and consequently F ∗ ≡ 0 is the unique optimal
solution for the inf-convolution problem (3.2). 

4. Conclusion
Standard diversification will occur in exchange economies as soon as agents
have non-speculative risk measures. The regulator has to impose very different
rules on agents as to generate risk measures, which are not non-speculative, if she
wants to increase the diversification in the market. In other words, diversification
occurs when agents are very different one from the other. This result supports for
instance the intervention of reinsurance companies on financial markets in order to
increase the diversification on the reinsurance market.

References
[ADEH] P. Artzner, F. Delbaen, J.M. Eber, D. Heath, Coherent Measures of Risk Mathematical
Finance 9 (1999), 203–228.
[B] P. Barrieu, Structuration Optimale de Produits Financiers en Marché Illiquide et Trois
Excursions dans d’autres Domaines des Probabilités Thèse de doctorat, Université de
Paris VI (2002).
[BEK1] P. Barrieu, N. El Karoui, Reinsuring Climatic Risk using Optimally Designed Weather
Bonds Geneva Papers, Risk and Insurance Theory 27 (2002), 87–113.
[BEK2] P. Barrieu, N. El Karoui, Optimal Design of Derivatives in Illiquid Markets Quantitative
Finance 2 (2002), 1–8.
[BEK3] P. Barrieu, N. El Karoui, Optimal Derivatives Design and Diversification in Financial
Market with non-tradable Risk Working Paper (2003).
[BEK4] P. Barrieu, N. El Karoui, Structuration optimale de produits financiers et diversification
en présence de sources de risque non-négociables Comptes Rendus de l’Académie des
Sciences, Série I 336 (2003), 493-498.
[Be1] D. Becherer, Rational Hedging and Valuation with Utility-Based Preferences, PhD The-
sis, Berlin University (2001).
[Be2] D. Becherer, Rational Hedging and Valuation of Integrated Risks under Constant Ab-
solute Risk Aversion, To appear in Insurance: Mathematics and Economics.
[BelF] F. Bellini, M. Frittelli, On the Existence of Minimax Martingale Measures, Mathemat-
ical Finance 12 (2002), 1–21.
[Bo] K. Borch, Equilibrium in a Reinsurance Market, Econometrica 30 (1962), 424–444.
[Bu] H. Bühlmann, Mathematical Methods in Risk Theory Springer Verlag (1970).
[BuDES] H. Bühlmann, F. Delbaen, P. Embrechts, A. Shiryaev, On Es-
scher Transforms in Discrete Finance Models Working Paper
(www.math.ethz.ch/˜delbaen/ftp/preprints/Esscher-BDES-ASTIN.pdf).
[Da1] M. Davis, Option Pricing in Incomplete Markets in Mathematics of derivative securities
(eds: M.A.H. Dempster and S.R. Pliska), Cambridge University Press (1997), 227–254.
[Da2] M. Davis, Pricing Weather Derivatives by Marginal Value Quantitative Finance 1
(2001), 1–4.
[DGRSS] F. Delbaen, P. Grandits, T. Rheinländer, D. Samperi, M. Schweizer, C. Stricker, Expo-
nential Hedging and Entropic Penalities, Mathematical Finance 12 (2002), 99–123.
[EG] L. Eeckhoudt, C. Gollier, Risk: Evaluation, Management and Sharing Harvester Wheat-
sheaf (1995).
[EKQ] N. El Karoui, M.C. Quenez, Non-linear Pricing Theory and Backward Stochastic Dif-
ferential Equations, Financial Mathematics (ed: W.J. Runggaldier), Lecture Notes in
Mathematics 1656, Springer Verlag (1996), 191–246.
[EKPQ] N. El Karoui, N. Peng, M.C. Quenez, Backward Stochastic Differential Equations in
Finance Mathematical Finance 7 (1997), 1–71.
[EKR] N. El Karoui, R. Rouge, Pricing via Utility Maximization and Entropy Mathematical
Finance 10 (2000), 259–276.
 OPTIMAL DERIVATIVES DESIGN UNDER DYNAMIC RISK MEASURES 13

[FS1] H. Föllmer, A. Schied, Convex Measures of Risk and Trading Constraints Finance and
Stochastics 6 (2002), 429–447.
[FS2] H. Föllmer, A. Schied, Stochastic Finance: An Introduction in discrete Time De
Gruyter Studies in Mathematics (2002).
[Fr1] M. Frittelli, The Minimal Entropy Martingale Measure and the Valuation in Incomplete
Markets Mathematical Finance 10 (2000), 39–52.
[Fr2] M. Frittelli, Introduction to a Theory of Value Coherent with the No-Arbitrage Principle
Finance and Stochastics, 4 (2000), 275–297.
[Gi] E.R. Gianin, Some Examples of Risk Measures via g-Expectations Working Paper
(2003).
[HN] S.D. Hodges, A. Neuberger, Optimal Replication of Contingent Claims under Transac-
tion Costs Review of Futures Markets 8 (1989), 222–239.
[K] G. Kallianpur, Stochastic Filtering Theory Springer Verlag (1980).
[KaKo] I. Karatzas, S.G. Kou, On the Pricing of Contingent Claims under Constraints Annals
of Applied Probability, 6 (1996), 321–369.
[Kob] M. Kobylanski, Backward Stochastic Differential Equations and Partial Differential
Equations with Quadratic Growth Annals of Probability, 28 (2000), 558-602.
[La1] P. Lakner, Utility Maximization with Partial Information Stochastic processes and their
applications, 56 (1995), 247-273.
[La2] P. Lakner, Optimal Trading for an Investor: The Case of Partial Information Stochastic
processes and their applications, 76 (1998), 77-96.
[Le] D. Lefèvre, An Introduction to Utility Maximization with Partial Observation Rapport
de recherche n◦ 4183, INRIA (2001).
[LSMa] J.P. Lepeltier, J. San Martin, Existence of BSDE with Superlinear Quadratic Coefficient
Stochastics and Stochastic Report, 63 (1998), 227-240.
[LiS] R.S. Lipster, A.N. Shiryayev, Statistics of Random Processes Springer Verlag (2001).
[Me] R. Merton, Optimum Consumption and Portfolio Rules in a Continuous Time Model
Journal of Economic Theory 3 (1971), 373–413.
[MuZ] M. Musiela, T. Zariphopoulou, Pricing and Risk Management of Derivatives Written
on Non-Traded Assets Working Paper (2001).
[P] S. Peng, Backward SDE and Related g-Expectations in Backward Stochastic Diffential
Equations (eds: N. El Karoui and L. Mazliak), Pitman Res. Notes Math. Ser. Longman
Harlow, 364 (1997), 141-159.
[PhQ] H. Pham, M.C. Quenez, Dynamic portfolio optimization in the case of partially observed
drift process Annals of Probability, 11 (2001), 210-238.
[QV] J.P. Quadrat, M. Viot, Introduction à la commande stochastique Cours polycopié de l’
Ecole Polytechnique (1996).
[Ra] A. Raviv, The Design of an Optimal Insurance Policy American Economic Review 69
(1979), 84–96.
[Ro] R.T. Rockafellar, Convex Analysis Princeton Landmarks in Mathematics (1970).

Pauline Barrieu, L.S.E., Statistics Department, Houghton Street, London, WC2A

2AE, United Kingdom
E-mail address: [email protected]

Nicole El Karoui, C.M.A.P., Ecole Polytechnique, 91128 Palaiseau Cédex, France

E-mail address: [email protected]