Projektbereich B

Discussion Paper No. B–247

Variance-Optimal Hedging
in Discrete Time

by

∗)
Martin Schweizer

February 1993

∗)
Financial support by Deutsche Forschungsgemeinschaft,
Sonderforschungsbereich 303 at the University of Bonn, is
gratefully acknowledged.

(Mathematics of Operations Research 20 (1995), 1–32)

This version: 04.05.1995

 Variance-Optimal Hedging in Discrete Time

Martin Schweizer
Universität Göttingen
Institut für Mathematische Stochastik
Lotzestraße 13
D-37083 Göttingen
Germany

Abstract: We solve the problem of approximating in L2 a given random variable H by

stochastic integrals GT (ϑ) of a given discrete-time process X. We interpret H
as a contingent claim to be paid out at time T , X as the price evolution of some
risky asset in a financial market, and G(ϑ) as the cumulative gains from trade
using the hedging strategy ϑ. As an application, we determine the variance-
optimal strategy which minimizes the variance of the net loss H − GT (ϑ) over
all strategies ϑ.

Key words: hedging, financial mathematics, contingent claims, mean-variance criterion,

signed martingale measures

AMS 1980 subject classifications: 90A09, 49A60, 60K30

0. Introduction
Consider a stochastic process X = (Xt )t∈T defined on a probability space (Ω, F , P ) and
adapted to a filtration IF = (Ft )t∈T , with a time index set T ⊆ [0, T ] for some T > 0, and
an FT -measurable random variable H. Think of Xt as the price at time t of some risky asset
(e.g., a stock) in a financial market and of H as a contingent claim, i.e., a random loss suffered
at time T by a small agent in this market. The standard example is given by a European
call option written on X with expiration date T and strike price K, where H = (XT − K)+ .
One of the central problems in financial mathematics is the pricing and hedging of contingent
claims by means of dynamic trading strategies based on X. Such a strategy is described by
an IF -predictable process ϑ = (ϑt )t∈T , with ϑt representing the number of shares of X held
by the agent at time t. Predictability of ϑ is a mathematical formulation of the obvious
informational constraint that ϑ is not allowed to anticipate ¡ ¢the movement of X. To any
trading strategy ϑ then corresponds a process G(ϑ) = Gt (ϑ) t∈T of cumulative gains from
trade. If c denotes the agent’s initial capital, a good hedging strategy should thus have the
property that the total gain c + GT (ϑ) is in some sense as close to H as possible. Moreover,
the initial capital c∗ of an optimal pair (c∗ , ϑ∗ ) provides a natural candidate for a price of H
based on the chosen criterion of closeness.
In this paper, we solve this problem under three assumptions: frictionless trading, a
quadratic criterion for measuring closeness, and discrete time. The first assumption implies
that the gains process G(ϑ) of any self-financing trading strategy is given by the stochastic
integral
Zt
Gt (ϑ) = ϑs dXs for t ∈ T ;
0

see Harrison/Kreps (1979) and Harrison/Pliska (1981) for motivation and precise definitions.
Since we are working in discrete time with T = {0, 1, . . . , T }, the above integral is actually a
finite sum of the form
t
X
Gt (ϑ) = ϑj ∆Xj for t = 0, 1, . . . , T ,
j=1

where ∆Xj := Xj − Xj−1 denotes the increment of X at time j. We assume that each Xt
is square-integrable, and we define the set Θ of admissible strategies to consist of all those
IF -predictable processes ϑ such that G(ϑ) is square-integrable. Finally, using a quadratic
criterion means that we want to solve, for fixed c ∈ IR and H ∈ L2 (P ), the optimization
problem
h¡ ¢2 i
(0.1) Minimize E H − c − GT (ϑ) over all ϑ ∈ Θ.

Put differently, we are looking for an approximation of H in L2 (P ) by stochastic integrals of

X.
Unfortunately, it turns out that (0.1) has no solution in general; this is shown by an
explicit counterexample in section 5. We therefore introduce in section 1 a nondegeneracy
condition on X. We assume that X has a bounded mean-variance tradeoff in the sense that
2
(E[∆Xt |Ft−1 ])
(0.2) is P -a.s. bounded, uniformly in t and ω.
Var[∆Xt |Ft−1 ]

1
Under the much stronger assumption that this ratio is deterministic, (0.1) was previously
solved by Schäl (1994). For an informal discussion of (0.1), see also Hipp (1993). In the present
paper, we prove the existence of an optimal strategy ξ (c) for (0.1) under the assumption (0.2).
In fact, we prove in section 2 that (0.2) implies that GT (Θ) is closed in L2 (P ). Moreover,
this result is sharp; a counterexample in section 5 shows that condition (0.2) is in general
indispensable. Having established existence, we then proceed to analyze the structure of
ξ (c) in more detail. The main tool for this purpose is provided by the adjustment process
β = (βt )t=1,...,T of X; this process is defined in section 2 and studied in detail in Schweizer
(1993a). A first structural result on ξ (c) is given in Theorem 2.4 where we obtain an expression
for ξ (c) in feedback form. As a corollary, we deduce explicit formulae for the mean and the
variance of the net loss H − c − GT (ξ (c) ) under the optimal strategy. The second structural
result is Theorem 2.8 which links ξ (c) to the locally risk-minimizing strategy for H introduced
in Schweizer (1988). This yields a decomposition of ξ (c) into three components, namely a pure
hedging demand, a demand for mean-variance purposes and a demand for hedging against
stochastic fluctuations in the mean-variance ratio (0.2). In particular, this extends previous
results by Schäl (1994) and Schweizer (1992).
Section 3 contains applications to several optimization problems with quadratic criteria.
We give explicit expressions for the optimal choice of strategy and initial capital, and we
determine the strategy which minimizes the variance of the net loss H − GT (ϑ), both with
and without the constraint of a fixed mean. In section 4, we discuss the simplification of our
results in special cases. If X is a martingale, problem (0.1) reduces to a simple application
of the Kunita-Watanabe projection theorem. If H is attainable in the sense that it can be
written as the sum of a constant and a (discrete-time) stochastic integral of X, we provide
a closed-form expression for ξ (c) in terms of the adjustment process β and the integrand
in the representation of H. Finally we examine the case where the mean-variance tradeoff
in (0.2) is deterministic. Under this assumption, we rederive the results of Schäl (1994) as
simple consequences of our approach, and we also provide an alternative proof which can
be generalized to a continuous-time framework. Section 5 concludes the paper with some
explicit examples and counterexamples.

1. The basic problem

Let (Ω, F , P ) be a probability space, T > 0 a fixed natural number and IF = (Fk )k=0,1,...,T
a filtration, i.e., an increasing family of sub-σ-algebras of F . We shall assume that F = FT .
Let X = (Xk )k=0,1,...,T be a real-valued, IF -adapted, square-integrable process, i.e., each Xk
is Fk -measurable and in L2 (P ). We denote by

∆Xk := Xk − Xk−1 for k = 1, . . . , T

the increments of X. A process ϑ = (ϑk )k=1,...,T is called predictable if ϑk is Fk−1 -measurable

for each k.

Definition. We denote by Θ the set of all predictable processes ϑ such that ϑk ∆Xk ∈ L2 (P )
for k = 1, . . . , T . For ϑ ∈ Θ, G(ϑ) is the process defined by

k
X
Gk (ϑ) := ϑj ∆Xj for k = 0, 1, . . . , T.
j=1

2
We shall use throughout the conventions that a sum over an empty set is defined to be 0, a
product over an empty set is defined to be 1, and 00 = 0.

The basic problem addressed in this paper is the following:

(1.1) Given H ∈ L2 (P ) and c ∈ IR,

h¡ ¢2 i
minimize E H − c − GT (ϑ) over all ϑ ∈ Θ.

Interpretation. Think of Xk as the (discounted) price at time k of some risky asset (e.g.,
a stock) in a financial market. The process ϑ describes the trading strategy of a small agent
in this market, where “small” means that his actions do not influence prices. The random
variable ϑk is thus interpreted as the number of shares held during the time interval (k − 1, k],
and predictability is imposed so that the decision about the choice of ϑk at time k − 1 must
be made without exact knowledge of the evolution of X in the next time interval. If we also
assume the existence of some riskless asset whose discounted value is 1 at all times, then
any ϑ determines a unique self-financing trading strategy, and G(ϑ) describes the process
of (discounted) gains from trade. Roughly speaking, “self-financing” means that a strategy
neither requires nor generates funds between dates 0 and T . For a detailed exposition, see
Harrison/Kreps (1979), Harrison/Pliska (1981) or Duffie (1988). If we now interpret the
contingent claim H as a random loss suffered by the agent at time T , then H − c − GT (ϑ) is
the agent’s net loss if he starts with initial capital c and follows the strategy ϑ. As an example,
H could be a financial obligation resulting from the sale of some financial instrument with
expiration date T . A European call option on X with strike price K, for instance, would
lead to H = (XT − K)+ . The goal in the basic problem is then to minimize the expected net
square loss by choosing a trading strategy.

Remark. As pointed out by the referees, the assumption of frictionless trading is crucial
here. Any inclusion of transaction costs would destroy the linearity of the gains G(ϑ) in ϑ
and thus make (1.1) intractable in the present generality. It would be an interesting and
challenging problem to find processes X for which (1.1) can also be solved in the presence of
transaction costs, but this is beyond the scope of the present paper.

Without additional assumptions on X, we obtain in general neither uniqueness nor ex-

istence of a solution to (1.1). Nonuniqueness obviously prevails for instance if X is constant,
and an explicit example in section 5 shows that nonexistence can also occur. In the sequel,
we shall be concerned with the existence and structure of an optimal strategy for (1.1). To
this end, we introduce the following condition on X:

Definition. We say that X satisfies the nondegeneracy condition (ND) if there exists a
constant δ ∈ (0, 1) such that
¡ ¢2 £ ¯ ¤
E [∆Xk |Fk−1 ] ≤ δ E ∆Xk2 ¯Fk−1 P -a.s. for k = 1, . . . , T .

Note that by Jensen’s inequality, we always have

¡ ¢2 £ ¯ ¤
E [∆Xk |Fk−1 ] ≤ E ∆Xk2 ¯Fk−1 P -a.s.

The point of condition (ND) is to ensure a strict inequality uniformly in ω.

3
 To obtain other formulations of (ND), we now write X in its Doob decomposition as

X = X0 + M + A,

where M = (Mk )k=0,1,...,T is a square-integrable (P, IF )-martingale with M0 = 0 and A =

(Ak )k=0,1,...,T is a square-integrable predictable process with A0 = 0. It is well known that
this decomposition is unique and given by

∆Ak := E [∆Xk |Fk−1 ] ,

∆Mk := ∆Xk − ∆Ak .

e = (λ
The predictable process λ ek )k=1,...,T is then defined by

ek := ∆Ak
(1.2) λ ¯
£ ¤ for k = 1, . . . , T ;
E ∆Xk2 ¯Fk−1

e is well-defined due to Jensen’s inequality. Finally we recall the notation

by our conventions, λ
£ ¯ ¤ ¡ ¢2
Var [∆Xk |Fk−1 ] := E ∆Xk2 ¯Fk−1 − E [∆Xk |Fk−1 ]

for the conditional variance of ∆Xk given Fk−1 and note that
£ ¯ ¤
Var [∆Xk |Fk−1 ] = E ∆Mk2 ¯Fk−1 P -a.s.

b = (K
Definition. The process K b j )j=0,1,...,T defined by

j
X 2
b j := (E[∆X` |F`−1 ])
K for j = 0, 1, . . . , T
Var[∆X` |F`−1 ]
`=1

is called the mean-variance tradeoff process of X.

Using the above definitions, it is straightforward to check that (ND) is equivalent to each
of the following conditions:

(1.3) ek ∆Ak ≤ δ
λ P -a.s. for k = 1, . . . , T , with a constant δ ∈ (0, 1).

(1.4) b is P -a.s. bounded, uniformly in ω and j.

The process K

£ ¯ ¤ £ ¯ ¤
(1.5) E ∆Mk2 ¯Fk−1 ≥ (1 − δ)E ∆Xk2 ¯Fk−1 P -a.s. for k = 1, . . . , T ,
with a constant δ ∈ (0, 1).

2
(E[∆Xk |Fk−1 ])
(1.6) is P -a.s. bounded, uniformly in ω and k.
Var[∆Xk |Fk−1 ]

Condition (ND) was introduced by Schweizer (1988) and also used by Schäl (1994) in the
equivalent form (1.6). In a continuous-time version of (1.1), a condition of the form (1.4)

4
plays an important role; see Schweizer (1993c). The term “nondegeneracy” is explained by
the equivalent formulation of (ND) that
£ ¯ ¤
E ∆Mk2 ¯Fk−1 ≥ L ∆A2k P -a.s. for k = 1, . . . , T , with a constant L < ∞,

which intuitively states that on each time interval (k − 1, k], any drift A of X must be
counterbalanced by a sufficiently strong diffusive behaviour of M . For a similar condition in
a continuous-time framework, see Schweizer (1991).

Example 1. Suppose that X is an event tree in the sense that X0 is a constant, each Xk takes
only finitely many values and IF is the filtration generated by X (i.e., Fk = σ(X0 , X1 , . . . , Xk )
for k = 0, 1, . . . , T ). This corresponds to a situation where at each time for each price, there
are only finitely many possible values for the next price. Intuitively,¡ an event
¢ tree may be
pictured as a graph whose nodes are given by the date-price pairs k, Xk (ω) ; see section 5 for
a graphical illustration. Call X nondegenerate if for each k, the conditional distribution of Xk
given Fk−1 is P -a.s. not concentrated in one point. This means that from each node, there
are at least two branches going to the right. Then it is easy to see that every nondegenerate
event tree X satisfies (ND). In fact, we may take Ω finite without loss of generality, and since
nondegeneracy of X implies
Var [∆Xk |Fk−1 ] (ω) > 0,
hence ¡ ¢2 £ ¯ ¤
E [∆Xk |Fk−1 ] (ω) < E ∆Xk2 ¯Fk−1 (ω)
for each ω and each k, we can choose δ ∈ (δ 0 , 1) with
¡ ¢2
E [∆X k |F k−1 ] (ω)
δ 0 := sup £ 2¯
¯ ¤ < 1,
k,ω E ∆Xk Fk−1 (ω)

since both k and ω run through a finite set only.

2. Existence and structure of an optimal strategy

The basic result for proving the existence of a solution for (1.1) is

Theorem 2.1. If X satisfies (ND), then GT (Θ) is closed in L2 (P ).

¡ ¢
Proof. Let (ϑn )n∈IN be any sequence in Θ such that GT (ϑn ) n∈IN is a Cauchy sequence in
¡ ¢
L2 (P ), and denote by Y the limit in L2 (P ) of GT (ϑn ) . For any ϑ ∈ Θ,

(2.1) GT (ϑ) = (GT −1 (ϑ) + ϑT ∆AT ) + ϑT ∆MT ;

since ϑ, A are predictable and M is a martingale, the two terms on the right-hand side of
(2.1) are orthogonal in L2 (P ), and therefore
h¡ ¢2 i h¡ ¢2 i h £ ¯ ¤i
E GT (ϑ) = E GT −1 (ϑ) + ϑT ∆AT + E (ϑT )2 E ∆MT2 ¯FT −1 .

But this implies that the sequence of FT −1 -measurable random variables

q £ ¯ ¤
ψTn := ϑnT E ∆MT2 ¯FT −1

5
is also a Cauchy sequence in L2 (P ), hence convergent in L2 (P ) to some ψT∞ which is again
FT −1 -measurable. Set

ψT∞
ϑ∞
T := I {E [∆MT2 |FT −1 ]>0} q £ ¯ ¤.
2 ¯
E ∆MT FT −1

Since X satisfies (ND), we have by (1.5)

h¡ ¢2 i h £ ¯ ¤i
2
E ϑ∞
T ∆XT − ϑnT ∆XT =E − (ϑ∞
T E ϑnT ) ∆XT2 ¯FT −1
1 h £ ¯ ¤i
∞ n 2 2¯
≤ E (ϑT − ϑT ) E ∆MT FT −1 ,
1−δ

and the last term tends to 0 as n → ∞ by the definition of ϑ∞ ∞

T . This shows that ϑT ∆XT is
in L2 (P ), since ϑn ∈ Θ, and that

ϑnT ∆XT −→ ϑ∞
T ∆XT in L2 (P ) as n → ∞.

Hence GT −1 (ϑn ) converges to Y −ϑ∞ 2

T ∆XT in L (P ) as n → ∞, and by iterating the preceding
arguments we conclude that

T
X
Y = ϑ∞
k ∆Xk P -a.s.
k=1

for some ϑ∞ ∈ Θ. Thus GT (Θ) is indeed closed in L2 (P ).

q.e.d.

Remark. It is natural to ask if GT (Θ) is still closed in L2 (P ) if one abandons the assumption
that X satisfies (ND). A counterexample due to W. Schachermayer shows that the answer is
negative in general; the same counterexample also shows that (1.1) will not have a solution
in general. For a detailed account, see section 5.

Theorem 2.2. Suppose that X satisfies (ND). For any H ∈ L2 (P ) and any c ∈ IR, there
exists a strategy ξ (c) ∈ Θ which solves (1.1).

Proof. By Theorem 2.1, GT (Θ) is a closed linear subspace of the Hilbert space L2 (P ); hence
we can project H − c on GT (Θ).
q.e.d.

Let us pause here a moment to emphasize the generality of Theorem 2.2. Under the
sole assumption that X satisfies (ND), we can solve (1.1) for any contingent claim H in
L2 (P ). Previous work by Schäl (1994) relied crucially on the additional assumption that
the mean-variance tradeoff process K b is deterministic. Moreover, Theorem 2.2 is essentially
the best possible result: the counterexample in section 5 shows that condition (ND) is in
general indispensable. Notice also that our existence argument does not require any structural
assumptions on X or H, in contrast to the corresponding continuous-time problem studied
in Duffie/Richardson (1991), Schweizer (1992) and Schweizer (1993b).

6
 In order to describe the structure of the optimal strategy ξ (c) in more detail, we first
recall from Schweizer (1993a) the definition of the adjustment process associated to X. This
is the predictable process β = (βk )k=1,...,T defined by
" ¯ #
Q
T ¯
E ∆Xk (1 − βj ∆Xj )¯¯Fk−1
j=k+1
(2.2) βk := " ¯ # for k = 1, . . . , T.
QT ¯
E ∆Xk2 (1 − βj ∆Xj )2 ¯¯Fk−1
j=k+1

The next result summarizes those properties of β we shall use in the sequel; see Schweizer
(1993a) for a proof.

Proposition 2.3. β is well-defined by (2.2) and has the property that for k = 1, . . . , T

T
Y
(2.3) (1 − βj ∆Xj ) ∈ L2 (P ),
j=k

T
Y
(2.4) ∆Xk (1 − βj ∆Xj ) ∈ L2 (P ),
j=k+1

T
Y
(2.5) βk ∆Xk (1 − βj ∆Xj ) ∈ L2 (P )
j=k+1

and
 ¯   ¯ 
T
Y ¯ T
Y ¯
¯ ¯
(2.6) E (1 − βj ∆Xj )2 ¯Fk−1  = E  (1 − βj ∆Xj )¯Fk−1  ≤ 1 P -a.s.
¯ ¯
j=k j=k

In particular, the random variable

T
Y
(2.7) Ze0 := (1 − βj ∆Xj )
j=1

is in L2 (P ) and satisfies
£ ¤
(2.8) 0 ≤ E[Ze0 ] = E (Ze0 )2 ≤ 1.

Furthermore, Ze0 has the property that

h i
(2.9) e 0
E Z GT (ϑ) = 0 for all ϑ ∈ Θ

or equivalently
h ¯ i
¯
(2.10) E Ze0 ∆Xk ¯Fk−1 = 0 P -a.s. for k = 1, . . . , T .

7
Remark. Our subsequent arguments rely heavily on computations involving conditional ex-
pectations. We should like to point out here that the integrability properties of β summarized
in (2.3) – (2.5) allow us to verify the existence of all these conditional expectations, and to
justify rigorously all manipulations below. For a detailed example of the type of reasoning to
be used, we refer to the proofs in Schweizer (1993a).

Throughout the rest of this section, we shall assume that X satisfies (ND), and we fix a
contingent claim H ∈ L2 (P ) and a constant c ∈ IR. In order to elucidate the structure of the
corresponding optimal strategy ξ (c) , we introduce the predictable process % = (%k )k=1,...,T
defined by
" ¯ #
Q
T ¯
E H∆Xk (1 − βj ∆Xj ) ¯¯Fk−1
j=k+1
(2.11) %k := " ¯ # for k = 1, . . . , T .
QT
2 ¯
E ∆Xk2 (1 − βj ∆Xj ) ¯¯Fk−1
j=k+1

Note that % is well-defined due to (2.4) and the Cauchy-Schwarz inequality, and that
 ¯   ¯ 
YT ¯ YT ¯
¯ ¯
(2.12) E Hβk ∆Xk (1 − βj ∆Xj ) ¯Fk−1  = E %k ∆Xk (1 − βj ∆Xj ) ¯Fk−1 
¯ ¯
j=k+1 j=k+1

P -a.s. for each k by (2.2) and (2.11).

Theorem 2.4. For k = 1, . . . , T , we have P -a.s.

(2.13) H − c − GT (ξ (c) )
T
X T
Y ³ ´Y
T
=H− %j ∆Xj (1 − β` ∆X` ) − c + Gk−1 (ξ (c) ) (1 − β` ∆X` )
j=k `=j+1 `=k

and
³ ´
(c)
(2.14) ξk = %k − βk c + Gk−1 (ξ (c) ) .

Proof. We show (2.13) and (2.14) simultaneously by backward induction. By the projection
theorem (see for instance Luenberger (1969), Theorem 3.3.1), a strategy ξ ∈ Θ solves (1.1) if
and only if £¡ ¢ ¤
E H − c − GT (ξ) GT (ϑ) = 0 for all ϑ ∈ Θ
or equivalently
£¡ ¢ ¯ ¤
(2.15) E H − c − GT (ξ) ∆Xk ¯Fk−1 = 0 P -a.s. for k = 1, . . . , T .

Since ξ (c) is in Θ and optimal, we obtain for k = T

h³ ´ ¯ i
¯
0 = E H − c − GT (ξ (c) ) ∆XT ¯FT −1
h ³ ´ ¯ i
(c) ¯
= E H∆XT − ξT ∆XT2 − c + GT −1 (ξ (c) ) ∆XT ¯FT −1

8
and therefore (2.14) with k = T by (2.2) and (2.11). This yields in turn
³ ´
(c)
H − c − GT (ξ (c) ) = H − ξT ∆XT − c + GT −1 (ξ (c) )
³ ´
(c)
= H − %T ∆XT − c + GT −1 (ξ ) (1 − βT ∆XT )

which is (2.13) for k = T . Suppose now that (2.13) and (2.14) hold for j = k + 1, . . . , T .
Then (2.15) implies
h³ ´ ¯ i
(c) ¯
0=E H − c − GT (ξ ) ∆Xk ¯Fk−1
  ¯ 
T T ¯
X Y ¯
= E ∆Xk H − %j ∆Xj (1 − β` ∆X` ) ¯¯Fk−1 
j=k+1 `=j+1 ¯
" ¯ #
³ ´ Y T ¯
¯
− E ∆Xk c + Gk (ξ (c) ) (1 − β` ∆X` )¯Fk−1
¯
`=k+1
 ¯ 
XT Y T ¯
¯
= E[H∆Xk |Fk−1 ] − E ∆Xk %j ∆Xj (1 − β` ∆X` )¯Fk−1 
¯
j=k+1 `=j+1
" ¯ #
YT ¯
(c) ¯
− ξk E ∆Xk2 (1 − β` ∆X` )¯Fk−1
¯
`=k+1
" ¯ #
³ ´ Y T ¯
¯
− c + Gk−1 (ξ (c) ) E ∆Xk (1 − β` ∆X` )¯Fk−1
¯
`=k+1
  ¯ 
X T YT ¯
¯
= E H∆Xk 1 − βj ∆Xj (1 − β` ∆X` ) ¯¯Fk−1 
j=k+1 `=j+1 ¯
" ¯ #
YT ¯
(c) 2 ¯
− ξk E ∆Xk (1 − β` ∆X` )¯Fk−1
¯
`=k+1
" ¯ #
³ ´ Y T ¯
¯
− c + Gk−1 (ξ (c) ) E ∆Xk (1 − β` ∆X` )¯Fk−1
¯
`=k+1
" ¯ # " ¯ #
Y T ¯ YT ¯
¯ (c) ¯
= E H∆Xk (1 − β` ∆X` )¯Fk−1 − ξk E ∆Xk2 (1 − β` ∆X` )¯Fk−1
¯ ¯
`=k+1 `=k+1
" ¯ #
³ ´ Y T ¯
¯
− c + Gk−1 (ξ (c) ) E ∆Xk (1 − β` ∆X` )¯Fk−1
¯
`=k+1

by (2.13) with k + 1 instead of k, (2.12) with j instead of k (after conditioning on Fj−1 ⊇ Fk

for each summand) and the identity

T
Y T
X T
Y
(2.16) (1 − c` ) = 1 − cj (1 − c` ).
`=k+1 j=k+1 `=j+1

9
Using (2.6), (2.2) and (2.11), we obtain (2.14) for k. This implies
(c)
c + Gk (ξ (c) ) = c + Gk−1 (ξ (c) ) + ξk ∆Xk
³ ´
= %k ∆Xk + c + Gk−1 (ξ (c) ) (1 − βk ∆Xk )
and therefore by (2.13) for k + 1 instead of k
H − c − GT (ξ (c) )
T
X T
Y ³ ´ Y
T
(c)
=H− %j ∆Xj (1 − β` ∆X` ) − c + Gk (ξ ) (1 − β` ∆X` )
j=k+1 `=j+1 `=k+1
T
X T
Y ³ ´Y
T
(c)
=H− %j ∆Xj (1 − β` ∆X` ) − c + Gk−1 (ξ ) (1 − β` ∆X` )
j=k `=j+1 `=k

which is (2.13) for k.

q.e.d.

Corollary 2.5. We have

 
h i T
Y
(2.17) E H − c − GT (ξ (c) ) = E (H − c) (1 − βj ∆Xj ) = E[H Ze0 ] − cE[Ze0 ]
j=1

and
·³ ´2 ¸
(c)
(2.18) E H − c − GT (ξ )
 2 
T
X T
Y
 
= c2 E[Ze0 ] − 2cE[H Ze0 ] + E H − %j ∆Xj (1 − β` ∆X` )  .
j=1 `=j+1

Proof. By (2.13) with k = 1, we have

 
h i XT T
Y T
Y
E H − c − GT (ξ (c) ) = E H − %j ∆Xj (1 − β` ∆X` ) − c (1 − β` ∆X` )
j=1 `=j+1 `=1
   
T
X T
Y T
Y
= E H 1 − βj ∆Xj (1 − β` ∆X` ) − c (1 − β` ∆X` )
j=1 `=j+1 `=1

from (2.12) after conditioning the j-th summand on Fj−1 . Using (2.16) and (2.7) then gives
(2.17). Again by (2.13) with k = 1,
·³ ´2 ¸
(c)
E H − c − GT (ξ )
   2 
YT XT YT
2  
= c2 E  (1 − βj ∆Xj )  + E H − %j ∆Xj (1 − β` ∆X` ) 
j=1 j=1 `=j+1
  
T
X T
Y T
Y
− 2cE H − %j ∆Xj (1 − β` ∆X` ) (1 − βk ∆Xk ) .
j=1 `=j+1 k=1

10
But the first term equals c2 E[Ze0 ] by (2.8), and for each j,
 ¯ 
T
Y T
Y ¯
¯
E %j ∆Xj (1 − β` ∆X` ) (1 − βk ∆Xk ) ¯Fj−1 
¯
`=j+1 k=1
 ¯ 
YT ¯ j−1
Y
2 ¯
= E ∆Xj (1 − βj ∆Xj ) (1 − β` ∆X` ) ¯Fj−1  %j (1 − βk ∆Xk )
¯
`=j+1 k=1
" ¯ #
YT ¯
¯
= %j E ∆Xj (1 − βk ∆Xk ) ¯Fj−1 = 0 P -a.s.
¯
k=1

by (2.6) (after conditioning on Fj ) and (2.10). Thus the third term equals −2cE[H Ze0 ] by
(2.7), and this proves (2.18).
q.e.d.

Remark. Theorem 2.4 suggests an alternative way to solve (1.1): we could define a pre-
dictable process ξ = (ξk )k=1,...,T recursively by
¡ ¢
(2.19) ξk = %k − βk c + Gk−1 (ξ)

and then try to show that ξ is optimal. If (and this is the crucial point) ξ is in Θ, then it
is not too hard to show that ξ solves (1.1). One first proves by induction as in Theorem 2.4
that
XT YT YT
H − c − GT (ξ) = H − %j ∆Xj (1 − β` ∆X` ) − c (1 − β` ∆X` ) .
j=1 `=j+1 `=1

A similar argument as in the proof of Corollary 2.5 then shows by using (2.10), (2.12) and
(2.16) that
£¡ ¢ ¯ ¤
E H − c − GT (ξ) ∆Xk ¯Fk−1 = 0 P -a.s. for k = 1, . . . , T

which implies optimality as above. However, we have so far not been able to prove that (2.19)
automatically implies ξ ∈ Θ. In the special case where the mean-variance tradeoff process
b is deterministic, this is indeed true, as was shown by Schäl (1994); see also section 4. In
K
general, however, we do not know if (1.1) can be solved by this approach.

We can obtain additional information on the structure of ξ (c) if we examine H in more

detail. The following result essentially goes back to Schweizer (1988); see Schäl (1994) or
Schweizer (1993b) for a proof. Recall the assumption that X satisfies (ND).

Proposition 2.6. Every contingent claim H ∈ L2 (P ) admits a decomposition

T
X
(2.20) H = H0 + ξjH ∆Xj + LH
T P -a.s.,
j=1

where H0 is a constant, ξ H ∈ Θ and LH = (LH k )k=0,1,...,T is a square-integrable (P, IF )-

H
martingale with E[L0 ] = 0 which is strongly orthogonal to M (i.e., the product LH M is a

11
(P, IF )-martingale null at 0). The processes ξ H and LH can be obtained as
à ¯ !
PT ¯
Cov H − ξjH ∆Xj , ∆Xk ¯¯Fk−1
j=k+1
ξkH := for k = 1, . . . , T
Var[∆Xk |Fk−1 ]

(with the obvious notation for the conditional covariance) and

 ¯   
XT ¯ XT
 ¯
LHk := E H − ξjH ∆Xj ¯Fk  − E H − ξjH ∆Xj  for k = 0, . . . , T .
¯
j=1 j=1

The constant H0 in (2.20) can be obtained explicitly. To that end, we define the process
Z = (Zbk )k=0,1,...,T by
b

k k
à !
Y ej ∆Xj
1−λ Y ej
λ
(2.21) Zbk := = 1− ∆Mj .
ej ∆Aj
1−λ ej ∆Aj
1−λ
j=1 j=1

Since X satisfies (ND), (1.3) implies that Zb is a square-integrable (P, IF )-martingale. If we

define the signed measure Pb on (Ω, F ) by setting

YT ej ∆Xj
dPb b 1−λ
:= ZT = ,
dP e
j=1 1 − λj ∆Aj

we thus obtain
" ¯ #
dPb ¯ h ¯ i
¯
(2.22) E ∆Xk ¯¯Fk−1 = E Zbk ∆Xk ¯Fk−1
dP
Zbk−1 h³ ´ i
= ek ∆Xk ∆Xk |Fk−1
E 1−λ
1−λek ∆Ak
=0 P -a.s. for k = 1, . . . , T

e If we assume that E[Ze0 ] > 0, we can define another signed measure Pe

by the definition of λ.
on (Ω, F) by setting

dPe Ze0
(2.23) := Ze := .
dP E[Ze0 ]

By (2.22) and (2.10), both Pb and Pe are then signed martingale measures for X in the sense
of the following

Definition. A signed measure Q on (Ω, F ) is called a signed martingale measure for X if

Q[Ω] = 1, Q ¿ P on F with dQ 2
dP ∈ L (P ), and X is a (Q, IF )-martingale in the sense that
· ¯ ¸
dQ ¯
E ¯
∆Xk ¯Fk−1 = 0 P -a.s. for k = 1, . . . , T .
dP

12
Q is called an equivalent martingale measure (with square-integrable density) if in addition,
Q is a probability measure and Q ≈ P on F .

We point out that in contrast to most situations encountered in financial mathematics,

we are working here with signed instead of equivalent martingale measures. This notion was
introduced in Müller (1985) and appears here quite naturally in the study of the solution of
(1.1). An explicit example in section 5 shows that Pb and Pe will typically be signed measures,
i.e., Zb and Ze will also take negative values. Pe is called the variance-optimal signed martingale
measure for X due to Theorem 3 in Schweizer (1993a), and Pb is the discrete-time version
of the minimal martingale measure introduced by Föllmer/Schweizer (1991). If X has a
deterministic mean-variance tradeoff process K, b we shall see in section 4 that β coincides
with λ e and Pe coincides with Pb. In general, however, the two measures are different; this will
be shown by an explicit example in section 5.

Now introduce the process Vb by setting

(2.24) Vbk := H0 + Gk (ξ H ) + LH
k for k = 0, 1, . . . , T.

Then
Vbk = E[H|F
b k] P -a.s. for k = 0, 1, . . . , T
in the sense that
" ¯ #
dPb b ¯¯
E ∆Vk ¯Fk−1 = 0 P -a.s. for k = 1, . . . , T
dP

(i.e., Vb is a “(Pb, IF )-martingale”) and VbT = H P -a.s. In fact, (2.24), the martingale property
of Zb and (2.21) imply that
" ¯ #
dPb b ¯¯
E ∆Vk ¯Fk−1
dP
h ¯ i
¯
= E Zbk (ξkH ∆Xk + ∆LH ) F
k ¯ k−1
" ¯ # Ã !
d b
P ¯ £ ¯ ¤ ek
λ £ ¯ ¤
= ξkH E ∆Xk ¯¯Fk−1 + Zbk−1 E ∆LH ¯
k Fk−1 − E ∆LH ¯
k ∆Mk Fk−1
dP ek ∆Ak
1−λ
=0 P -a.s.,

since Pb is a signed martingale measure for X and LH and LH M are (P, IF )-martingales. For
an economic interpretation of Vb as the intrinsic value process associated to H, we refer to
Hofmann/Platen/Schweizer (1992).

Lemma 2.7. In the decomposition (2.20), the constant H0 is given by

 
T
Y e
1 − λj ∆Xj  b
(2.25) H0 = E  H = E[H].
1−λ ej ∆Aj
j=1

13
If E[Ze0 ] > 0, then
£ ¤
E LH e0
e T Z
(2.26) V0 := E[H] = H0 + .
E[Ze0 ]

Proof. Since LH is a martingale and strongly orthogonal to M , we have for each k by (2.24)
£ ¯ ¤ £ ¯ ¤ h ¯ i
¯ ¯ b ¯
0 = E ∆LH k ∆M k F k−1 = E ∆L H
k ∆X k F k−1 = E (∆ V k − ξ H
k ∆X k )∆X F
k ¯ k−1

and therefore £ ¯ ¤
E ∆Vbk ∆Xk ¯Fk−1
ξkH = £ ¯ ¤ .
E ∆Xk2 ¯Fk−1
Since LH is a martingale, this implies
£ ¯ ¤
0 = E ∆LH ¯
k Fk−1
£ ¯ ¤
= E ∆Vbk ¯Fk−1 − ξkH E[∆Xk |Fk−1 ]
£ ¯ ¤
= E ∆Vbk (1 − λ ek ∆Xk )¯Fk−1
£ ¯ ¤
= E Vbk (1 − λ ek ∆Xk )¯Fk−1 − Vbk−1 (1 − λ
ek ∆Ak )

and therefore " ¯ #

1 − ek ∆Xk ¯
λ
Vbk−1 = E Vbk ¯Fk−1 .
¯
e
1 − λk ∆Ak

Since VbT = H and

H0 = E[Vb0 − LH b
0 ] = E[V0 ],

we obtain (2.25) by iteration. Since

H = H0 + GT (ξ H ) + LH
T

by (2.20), (2.9) implies

E[H Ze0 ] = H0 E[Ze0 ] + E[LH e0
T Z ],

hence (2.26) by (2.23).

q.e.d.

Remark. We shall show in section 4 that H0 and V0 coincide if the mean-variance tradeoff
b is deterministic; an example in section 5 will show that H0 and V0 differ in general.
process K

To relate the optimal strategy ξ (c) for (1.1) to the strategy ξ H , we now introduce the
predictable process γ = (γk )k=1,...,T defined by
" ¯ #
Q
T ¯
E (LH H
T − Lk−1 )∆Xk (1 − βj ∆Xj )¯¯Fk−1
j=k+1
γk := " ¯ # for k = 1, . . . , T .
Q
T ¯
E ∆Xk2 (1 − βj ∆Xj )2 ¯¯Fk−1
j=k+1

14
Due to (2.4) and the Cauchy-Schwarz inequality, γ is indeed well-defined.

Theorem 2.8. For every fixed c ∈ IR, the solution ξ (c) of (1.1) satisfies
³ ´
(c) H b (c)
(2.27) ξk = ξk + βk Vk−1 − c − Gk−1 (ξ ) + γk P -a.s. for k = 1, . . . , T .

Proof. By Theorem 2.4, ξ (c) satisfies for each k

³ ´
(c) (c)
ξk = %k − βk c + Gk−1 (ξ ) P -a.s.

By (2.11), (2.20) and (2.24), the numerator of %k is given by

 ¯   ¯ 
YT ¯ YT ¯
¯ ¯
E H∆Xk (1 − βj ∆Xj )¯Fk−1  = Vbk−1 E ∆Xk (1 − βj ∆Xj )¯Fk−1 
¯ ¯
j=k+1 j=k+1
 ¯ 
YT ¯
¯
+ ξkH E ∆Xk2 (1 − βj ∆Xj )¯Fk−1 
¯
j=k+1
 ¯ 
XT YT ¯
¯
+ E ξ`H ∆X` ∆Xk (1 − βj ∆Xj )¯Fk−1 
¯
`=k+1 j=k+1
 ¯ 
YT ¯
¯
+ E (LH H
T − Lk−1 )∆Xk (1 − βj ∆Xj )¯Fk−1  .
¯
j=k+1

But every summand in the third term on the right-hand side equals 0 P -a.s. by (2.10) (after
conditioning on F`−1 ⊇ Fk ), and dividing by the denominator of %k implies by (2.6) that

%k = βk Vbk−1 + ξkH + γk P -a.s.,

hence (2.27).
q.e.d.

Theorem 2.8 has a very interesting and intuitive interpretation. To explain this, we
rewrite (2.27) as
³ ´ µ ³ ´ ¶
(c) H e b (c) e b (c)
ξk = ξk + λk Vk−1 − c − Gk−1 (ξ ) + (βk − λk ) Vk−1 − c − Gk−1 (ξ ) + γk .

It is known from the results of Schweizer (1988, 1991) that ξ H determines a unique locally risk-
minimizing strategy for the contingent claim H. The first term in the above decomposition
of ξ (c) can therefore be interpreted as a pure hedging demand. In analogy to Schweizer
(1992), the second term can be viewed as a demand for mean-variance purposes. Finally,
the third term corresponds to a demand for hedging against the stochastic fluctuations in the
mean-variance ratio
2
(E[∆Xk |Fk−1 ])
.
Var[∆Xk |Fk−1 ]

15
 b is a
We shall see in section 4 that the third term vanishes if this ratio (or equivalently K)
deterministic process. In general, however, the interplay between the non-hedgeable part LH
of H and the stochastic mean-variance tradeoff induces an additional term to the solutions
given by Schweizer (1992) and Schäl (1992).

3. Applications

Throughout this section, we assume that X satisfies (ND), and we consider a fixed contingent
claim H ∈ L2 (P ). With the help of the solution of (1.1), we shall solve several optimization
problems with quadratic criteria. It should be mentioned that some of the techniques used
are very similar to those in Duffie/Richardson (1991); one major difference is that our com-
putations do not depend on the claim H under consideration and do not require a particular
structure for X.

3.1. The optimal choice of strategy and initial capital

The first problem we consider is

h¡ ¢2 i
(3.1) Minimize E H − V0 − GT (ϑ) over all ϑ ∈ Θ and all V0 ∈ IR.

This can be interpreted as choosing an initial capital V0 and a self-financing strategy ϑ in

such a way that the expected net quadratic loss is minimized. We first prove a general result
which holds for any set Θ 6= ∅ and any mapping GT : Θ → L2 (P ). In particular, it uses
neither the discrete-time structure of X nor the precise definition of Θ and can thus also be
applied in a continuous-time framework.

Lemma h¡ 3.1. If (1.1) has a¢2solution

i ξ (c) ∈ Θ for every c ∈ IR, and if c∗ minimizes the function
¡ ∗ ¢
c 7→ E H − c − GT (ξ (c) ) , then the pair c∗ , ξ (c ) solves (3.1).

Proof. For any pair (c, ϑ), we have

h¡ ·³ ´2 ¸ ·³ ´2 ¸
¢2 i (c) ∗ (c∗ )
E H − c − GT (ϑ) ≥ E H − c − GT (ξ ) ≥ E H − c − GT (ξ )

by the definitions of ξ (c) and c∗ , respectively.

q.e.d.
¡ ¢
Corollary 3.2. The solution of (3.1) is given by the pair V0 , ξ (V0 ) with

E[H Ze0 ] e
(3.2) V0 = = E[H]
E[Ze0 ]

(the second equality only holds if E[Ze0 ] > 0).

h¡ ¢2 i
Proof. By Corollary 2.5, the function c 7→ E H − c − GT (ξ (c) ) is minimized by c∗ with
c∗ E[Ze0 ] = E[H Ze0 ], i.e., by V0 ; note that V0 is well-defined due to (2.8). Hence the assertion
follows from Theorem 2.2 and Lemma 3.1.
q.e.d.

16
Remarks. 1) Corollary 3.2 shows a feature common to many optimization problems in
financial mathematics: the optimal initial capital
¡ is the expectation of H under a suitable
¢
martingale measure for X. It is tempting and was in fact suggested by Schäl (1994) to
interpret V0 as a fair hedging price for H. However, this is not always appropriate; an example
in section 5 illustrates the problem which may arise fromh such a definition. i
¡ ¢2
2) If E[Ze0 ] = 0, then Corollary 2.5 shows that E H − c − GT (ξ (c) ) does not de-
¡ ¢
pend on c. Hence every pair V0 , ξ (V0 ) with V0 ∈ IR solves (3.1), and (3.2) implies by our
conventions that we choose V0 = 0. An analogous comment also applies to Corollary 3.4
below. ¡ ¢
3) By Corollary 2.5, the optimal pair V0 , ξ (V0 ) also satisfies
h i
E H − V0 − GT (ξ (V0 ) ) = 0.
¡ ¢
As in Lemma 3.8 of Schäl (1994), this implies that V0 , ξ (V0 ) can be extended to a mean-
self-financing strategy. We refer to Schäl (1994) and Schweizer (1988) for precise definitions
and more details on this question.

3.2. The variance-optimal strategy

As a second problem, consider

(3.3) Minimize Var[H − GT (ϑ)] over all ϑ ∈ Θ.

The next result is again valid for any set Θ 6= ∅ and any mapping GT : Θ → L2 (P ).

Lemma h¡ 3.3. If (1.1) has a¢2solution

i ξ (c) ∈ Θ for every c ∈ IR, and if c∗ minimizes the function
∗
c 7→ E H − c − GT (ξ (c) ) , then ξ (c ) solves (3.3).

Proof. For every ϑ ∈ Θ, we have

h¡ ¢2 i
Var[H − GT (ϑ)] = E H − E[H − GT (ϑ)] − GT (ϑ)
·³ ´2 ¸
(E[H−GT (ϑ)])
≥ E H − E[H − GT (ϑ)] − GT (ξ )
·³ ´2 ¸
∗ (c∗ )
≥ E H − c − GT (ξ )
h i
∗ (c∗ )
≥ Var H − c − GT (ξ )
h ∗
i
= Var H − GT (ξ (c ) ) ,

where the first inequality uses the definition of ξ (c) with c := E[H − GT (ϑ)] and the second
the definition of c∗ .
q.e.d.

Corollary 3.4. The solution of (3.3) is given by ξ (V0 ) with V0 as in (3.2).

Proof. This follows immediately from Lemma 3.3 and the proof of Corollary 3.2.
q.e.d.

17
3.3. The mean-variance frontier

The third problem we consider is

(3.4) Given m ∈ IR, minimize Var[H − GT (ϑ)] over all ϑ ∈ Θ

satisfying the constraint E[H − GT (ϑ)] = m.

Lemma 3.5. For every c ∈ IR, ξ (c) is H-mean-variance efficient in the sense that
h i
Var H − GT (ξ (c) ) ≤ Var[H − GT (ϑ)]

for every ϑ ∈ Θ such that

h i
E[H − GT (ϑ)] = E H − GT (ξ (c) ) .

£ ¤
Proof. Let m = E H − GT (ξ (c) ) and take any ϑ ∈ Θ with E[H − GT (ϑ)] = m. Then the
definition of ξ (c) implies

Var[H − GT (ϑ)] = Var[H − c − GT (ϑ)]

h¡ ¢2 i
= E H − c − GT (ϑ) − (m − c)2
·³ ´2 ¸ ³ h i´2
(c)
≥ E H − c − GT (ξ ) − E H − c − GT (ξ (c) )
h i
= Var H − c − GT (ξ (c) )
h i
= Var H − GT (ξ (c) ) .

q.e.d.

Corollary 3.6. Suppose that X is not a martingale. For every m ∈ IR, the solution of (3.4)
is then given by ξ (cm ) with

m − E[H Ze0 ]
(3.5) cm = .
1 − E[Ze0 ]

Proof.
£ Fix m ∈ that there exists c ∈ IR with
¤ IR. By Lemma 3.5, it is enough to show(c)
(c)
E H − GT (ξ ) = m, since the corresponding strategy ξ will then solve (3.4). But
Corollary 2.5 implies that for every c ∈ IR
h i ³ ´
E H − GT (ξ (c) e 0 e 0
) = E[H Z ] + c 1 − E[Z ] ,

and this equals m if c is given by cm in (3.5). Because X is not a martingale, E[Ze0 ] 6= 1 by

Theorem 5 of Schweizer (1993a), and so cm is well-defined.
q.e.d.

18
Remark. If X is a martingale, then E[GT (ϑ)] = 0 for every ϑ ∈ Θ; hence (3.4) only makes
sense for m = E[H]. In that case, the solution is given by ξ (E[H]) , since
h¡ ¢2 i
Var[H − GT (ϑ)] = E H − E[H] − GT (ϑ)
·³ ´2 ¸
(E[H])
≥ E H − E[H] − GT (ξ )
h i
= Var H − GT (ξ (E[H]) )

for all ϑ ∈ Θ. As a matter of fact, the strategy ξ (c) does not depend on c in the martingale
case, but only on H; see subsection 4.1.

4. Special cases
In this section, we return to the basic problem (1.1) and indicate the simplifications arising
in several special cases.

4.1. The martingale case

If X is a (P, IF )-martingale (and as usual square-integrable), (1.1) becomes very simple. First
of all, the process A is identically 0 so that (ND) is trivially satisfied. Furthermore, it is clear
that GT (Θ) is closed in L2 (P ) since X as a martingale has pairwise orthogonal increments.
The adjustment process β is identically 0; it coincides with λ, e and the measures Pe, Pb and P
all coincide. By Theorem 2.4, the optimal strategy for fixed c and H is

(c) E [H∆Xk |Fk−1 ]

ξk = %k = £ ¯ ¤;
E ∆Xk2 ¯Fk−1

note that this does not depend on c, which justifies the remark after Corollary 3.6. The
decomposition (2.20) is the well-known Kunita-Watanabe decomposition of H with respect to
the martingale X; see for instance Dellacherie/Meyer (1982), Theorem VIII.51. In particular,
we obtain

(4.1) ξ (c) = % = ξ H for every c ∈ IR.

Lemma 2.7 yields

H0 = V0 = E[H],
and ·³ ´2 ¸ h¡ i
(c) 2 ¢2
E H − c − GT (ξ ) = (E[H] − c) + E LH
T

since
H − c − GT (ξ (c) ) = E[H] − c + LH
T
¡
by (2.20) and (4.1). In¢ particular, the minimal expected net quadratic loss or in the termi-
nology of Schäl (1992) minimal total risk
h¡ ¢2 i
J0 := min E H − c − GT (ϑ)
ϑ∈Θ,
c∈IR

19
is given by
h¡ ¢2 i
(4.2) J0 = E LH
T .

4.2. The case where H is attainable

In this subsection, we assume that the contingent claim H is attainable in the sense that
LHT = 0 P -a.s. in the decomposition (2.20). This means that H can be represented as

T
X
(4.3) H = H0 + ξjH ∆Xj P -a.s.,
j=1

i.e., as the sum of a constant and a (discrete-time) stochastic integral with respect to X. We
shall impose no special conditions on X, except as usual that (ND) is satisfied. Assumption
(4.3) implies that
Vb = H0 + G(ξ H )
and γ ≡ 0 by (2.24). Hence Theorem 2.8 yields for each k
 
k−1
X ³ ´
(c) (c)
ξk − ξkH = βk H0 − c − ξj − ξjH ∆Xj  P -a.s.
j=1

and therefore by an induction argument

k−1
Y
(c)
(4.4) ξk = ξkH + (H0 − c)βk (1 − β` ∆X` ) P -a.s. for k = 1, . . . , T .
`=1

This implies for each k that

k
Y
Vbk − c − Gk (ξ (c) ) = (H0 − c) (1 − βj ∆Xj ) P -a.s.
j=1

In particular,
T
Y
H − c − GT (ξ (c)
) = (H0 − c) (1 − βj ∆Xj ) = (H0 − c)Ze0
j=1

which implies by (2.8) that

h¡ ·³ ´2 ¸
¢2 i 2
min E H − c − GT (ϑ) =E H − c − GT (ξ (c)
) = (H0 − c) E[Ze0 ].
ϑ∈Θ

The solution of (3.1) is therefore given by the pair (H0 , ξ H ), since ξ (H0 ) = ξ H by (4.4).
Alternatively, this can be deduced from Corollary 3.2, since V0 = H0 by Lemma 2.7 and the
assumption (4.3). In particular, we see that the minimal total risk J0 becomes 0; this is of
course obvious from (4.3).

20
4.3. The case where X has a deterministic mean-variance tradeoff

In this subsection, we consider the special case where X has a deterministic mean-variance
tradeoff in the sense that

(4.5) b is deterministic.
the process K

Under this assumption, (1.1) was solved by Schäl (1994). Note that (4.5) is equivalent to
ek ∆Ak )k=1,...,T is deterministic, since
saying that the process (λ

à !−1
2 2
ek ∆Ak = 1 − (E[∆X |F ]) Var[∆Xk |Fk−1 ] (E[∆Xk |Fk−1 ])
(4.6) 1 − λ £ k2 ¯ k−1 ¤ = £ ¯ ¤ = 1+ .
¯
E ∆Xk Fk−1 E ∆Xk2 ¯Fk−1 Var[∆Xk |Fk−1 ]

Remark. Assumption (4.5) implies that for each k, there is a constant δk ∈ [0, 1] such that
ek ∆Ak = δk P -a.s.; furthermore, δk = 1 if and only if Var[∆Xk |Fk−1 ] = 0 P -a.s. Thus (4.5)
λ
implies (ND) under the additional nondegeneracy condition that for each k,

Var[∆Xk |Fk−1 ] > 0 with positive probability;

see also Example 1.

In the remainder of this subsection, we shall assume that X satisfies (ND), and we
consider a fixed contingent claim H ∈ L2 (P ). The basic result underlying the subsequent
simplifications is then

Lemma 4.1. If X has a deterministic mean-variance tradeoff, then

 ¯ 
T
Y ¯ T
Y
e ¯ ej ∆Aj )
E 
(1 − λj ∆Xj )¯Fk−1 = (1 − λ P -a.s. for k = 1, . . . , T .
¯
j=k j=k

Proof. For fixed k,

 ¯ 
T
Y ¯ −1
TY
E ej ∆Xj )¯¯FT −1  = (1 − λ
(1 − λ eT ∆AT ) ej ∆Xj )
(1 − λ
¯
j=k j=k

by the definition of A. By (4.5), the first term is deterministic, and so the assertion follows
by successive conditioning on FT −2 , . . . , Fk−1 .
q.e.d.

Corollary 4.2. If X has a deterministic mean-variance tradeoff, the adjustment process β

e and the measures Pe and Pb coincide.
coincides with λ,

e By (1.2) and (2.2), βT = λ

Proof. We first show by backward induction that β ≡ λ. eT . For

21
k < T, " #
¯
Q
T ¯
E ∆Xk (1 − βj ∆Xj )¯¯Fk−1
j=k+1
βk = " ¯ #
QT ¯
E ∆Xk2 (1 − βj ∆Xj )2 ¯¯Fk−1
j=k+1
" ¯ #
Q
T ¯
E ∆Xk ej ∆Xj )¯Fk−1
(1 − λ ¯
j=k+1
= " ¯ #
Q
T ¯
E ∆Xk2 ej ∆Xj )¯Fk−1
(1 − λ ¯
j=k+1

Q
T
ej ∆Aj )
E[∆Xk |Fk−1 ] (1 − λ
j=k+1
=
£ ¯ ¤ Q
T
E ∆Xk2 ¯Fk−1 ej ∆Aj )
(1 − λ
j=k+1
ek
=λ
by the induction hypothesis, (2.6) and Lemma 4.1 (after conditioning on Fk ) and (4.5). This
implies by Lemma 4.1
 
T
Y YT
E[Ze0 ] = E  (1 − λ ej ∆Xj ) = ej ∆Aj ) > 0
(1 − λ
j=1 j=1

by (1.3). Thus Ze is well-defined and equals ZbT by (2.23) and (2.21), so that Pe = Pb.
q.e.d.

Proposition 4.3. If X has a deterministic mean-variance tradeoff, the solution ξ (c) of (1.1)
satisfies
³ ´
(c) H e b (c)
(4.7) ξk = ξk + λk Vk−1 − c − Gk−1 (ξ ) P -a.s. for k = 1, . . . , T .

Proof. By Theorem 2.8 and Corollary 4.2, it is enough to show that (4.5) implies
(4.8) γk = 0 P -a.s. for k = 1, . . . , T.
But for every fixed k, we have for j > k
" ¯ #
YT ¯
¯
E ∆LH j ∆Xk (1 − β` ∆X` )¯Fj−1
¯
`=k+1
  ¯ 
¯ ¯
YT ¯ ¯ j−1
Y
 H e  e ¯ ¯  e` ∆X` )
= E ∆Lj (1 − λj ∆Xj )E (1 − λ` ∆X` )¯Fj ¯Fj−1  ∆Xk (1 − λ
¯ ¯
`=j+1 ¯ `=k+1

h ¯ i Y
T j−1
Y
=E ∆LH e ¯
− λj ∆Xj ) Fj−1 e
(1 − λ` ∆A` )∆Xk e` ∆X` )
(1 − λ
j (1
`=j+1 `=k+1

=0 P -a.s.,

22
where the first step uses Corollary 4.2, the second Lemma 4.1 and (4.5), and the third the
fact that LH is a martingale and strongly orthogonal to M . In the same way, we obtain
" ¯ #
YT ¯ £ ¯ T
¤ Y
¯ ¯ e` ∆A` ) = 0
E ∆LH k ∆X k (1 − β ` ∆X ` ) ¯ F k−1 = E ∆L H
k ∆X k F k−1 (1 − λ
¯
`=k+1 `=k+1

P -a.s. and therefore by summing over j ≥ k and conditioning on Fk−1

" ¯ #
¡ H ¢ YT ¯
H ¯
E LT − Lk−1 ∆Xk (1 − β` ∆X` )¯Fk−1 = 0 P -a.s.,
¯
`=k+1

hence (4.8) by the definition of γ.

q.e.d.

The result in Proposition 4.3 was previously obtained by Schäl (1994). However, it
should be emphasized that his method of proof is completely³ different
´ from the approach
(c)
taken here. He starts by defining a predictable process ψ (c) = ψk recursively by
k=1,...,T
³ ´
(c) ek Vbk−1 − c − Gk−1 (ψ (c) )
(4.9) ψk = ξkH + λ

and then shows that ψ (c) is in Θ and solves (1.1). Both these arguments rely on the condition
that X has a deterministic mean-variance tradeoff; see also the remark following Corollary
2.5. Note that (4.9) has exactly the same structure as (4.7); we could therefore recover the
results of Schäl (1994) by showing that ψ (c) is in Θ, since this implies by (4.9) and (4.7) that
ψ (c) and ξ (c) coincide, and in particular by Proposition 4.3 that ψ (c) solves (1.1). We prefer to
give an alternative proof which also works in a continuous-time framework and which seems
a bit more elegant than the proof in Schäl (1994). The inspiration for this argument comes
from Duffie/Richardson (1991); see also Schweizer (1992, 1993c).

Theorem 4.4. Suppose that X has a deterministic mean-variance tradeoff. For every fixed
c ∈ IR, the process ψ (c) defined by (4.9) is in Θ and solves (1.1). Furthermore,
h¡ ·³ ´2 ¸
¢2 i (c)
(4.10) min E H − c − GT (ϑ) = E H − c − GT (ψ )
ϑ∈Θ

³ £ ¤´ Y
T T
X £ T
¤ Y
2
= (H0 − c) + E (LH 2 ej ∆Aj ) +
(1 − λ (∆LH 2 ej ∆Aj ).
(1 − λ
0 ) E k )
j=1 k=1 j=k+1

The solution of (3.1) is given by the pair (H0 , ψ (H0 ) ), and the minimal total risk is
T T T
£ ¤Y X £ ¤ Y
(4.11) J0 = E (LH 2 ej ∆Aj ) +
(1 − λ (∆LH 2 ej ∆Aj ).
(1 − λ
0 ) E k )
j=1 k=1 j=k+1

Finally, ψ (c) and ξ (c) coincide.

Proof. 1) We first show by induction that ψ (c) ∈ Θ. For every k,

£ ¯ ¤
e2 E ∆X 2 ¯Fk−1 = λ
λ ek ∆Ak
k k

23
is deterministic. This implies by (4.9) that
·³ ´2 ¸ ·³ ´2 ¸
(c) e b
E ψ1 ∆X1 =E ξ1H ∆X1
+ λ1 ∆X1 (V0 − c)
£ H ¤ h £ ¯ ¤i
2 b 2 e2 2¯
≤ 2E (ξ1 ∆X1 ) + 2E (V0 − c) λ1 E ∆X1 F0 < ∞,

(c)
since ξ H ∈ Θ and Vb is square-integrable. Now suppose that ψj ∆Xj ∈ L2 (P ) for j =
1, . . . , k − 1. Then Vbk−1 − c − Gk−1 (ψ (c) ) ∈ L2 (P ) and an analogous argument as above shows
(c)
that ψk ∆Xk ∈ L2 (P ), thus completing the induction.
2) To show that ψ (c) solves (1.1), we fix ϑ ∈ Θ and define the function
 
³ ´ T
Y
f (k) := E  Vbk − c − Gk (ψ (c) ) Gk (ϑ) ej ∆Xj )
(1 − λ for k = 0, 1, . . . , T.
j=k+1

Since (4.9) yields by (2.24)

³ ´
Vbk − c − Gk (ψ (c) ) Gk (ϑ)
³ ´
b (c) H (c) H
= Vk−1 − c − Gk−1 (ψ ) + (ξk − ψk )∆Xk + ∆Lk (Gk−1 (ϑ) + ϑk ∆Xk )
µ³ ´ ¶
= b (c) e H
Vk−1 − c − Gk−1 (ψ ) (1 − λk ∆Xk ) + ∆Lk (Gk−1 (ϑ) + ϑk ∆Xk ) ,

Lemma 4.1, (2.10) and Corollary 4.2 imply that

f (k) = f (k − 1),

because LH is a martingale and strongly orthogonal to M . Hence

h³ ´ i
(c)
E H − c − GT (ψ ) GT (ϑ) = f (T ) = f (0) = 0

for every ϑ ∈ Θ, so that ψ (c) solves (1.1).

3) It is clear from (4.9) and (4.7) that ψ (c) and ξ (c)h must coincide. Furthermore, the
¡ ¢ i
(c) 2
solution of (3.1) is obviously obtained by minimizing E H − c − GT (ψ ) over c, and
thus it only remains to prove (4.10). To that end, we define the function
 
³ ´2 Y
T
g(k) := E  Vbk − c − Gk (ψ (c) ) ej ∆Xj )
(1 − λ for k = 0, 1, . . . , T.
j=k+1

Then (4.9) implies as before that

³ ´
ek ∆Xk ) + ∆LH
Vbk − c − Gk (ψ (c) ) = Vbk−1 − c − Gk−1 (ψ (c) ) (1 − λ k

24
and therefore by Lemma 4.1, since LH is a martingale and strongly orthogonal to M ,
 
³ ´2 T
Y
ek ∆Xk )2
g(k) = E  Vbk−1 − c − Gk−1 (ψ (c) ) (1 − λ ej ∆Xj )
(1 − λ
j=k+1
 
T
Y
+ E (∆LH 2 ej ∆Xj )
(1 − λ
k )
j=k+1
T
£ ¤ Y
= g(k − 1) + E (∆LH 2 ej ∆Aj ).
(1 − λ
k )
j=k+1

Here, we have used Corollary 4.2 and (2.6) to simplify the first term and Lemma 4.1 for the
second one. Hence we obtain
·³ ´2 ¸ T
X T
£ ¤ Y
E H − c − GT (ψ (c)
) = g(T ) = g(0) + E (∆LH 2 ej ∆Aj ),
(1 − λ
k )
k=1 j=k+1

and since
 
T
Y T
£ ¤Y
g(0) = E (Vb0 − c)2 e  H 2
(1 − λj ∆Xj ) = E (H0 − c + L0 ) ej ∆Aj )
(1 − λ
j=1 j=1

by Lemma 4.1, we obtain (4.10) after noting that E[LH

0 ] = 0.
q.e.d.

Remarks. 1) Theorem 4.4 contains the main results of Schäl (1994); note £ Hthat ¤ due to (4.6),
2
his formula for J0 agrees with ours. The additional term involving E (L0 ) in (4.10) and
(4.11) is due to the fact that we have not assumed F0 to be trivial.
2) To obtain the solution of (3.1), we could also have used Corollary 3.2 and the fact
that H0 = V0 by Corollary 4.2 and Lemma 2.7.
e ≡ 0 and
3) If X is a martingale, then the expression (4.11) for J0 reduces to (4.2), since λ
H
L as a martingale has pairwise orthogonal increments. If X does not have a deterministic
mean-variance tradeoff, an explicit formula like (4.10) does not seem to be available.

5. Explicit examples

The purpose of this section is to illustrate the previously developed concepts by means of
several examples where explicit computations are possible.

Example 2. Suppose that X0 = 0 and that ∆X1 takes the values +1, 0, −1 with probability
1 1
3 each. Given that X1 6= +1, ∆X2 takes the values ±1 with probability 2 each. The
conditional distribution of ∆X2 given X1 = +1 is denoted by ν, and we shall assume that
Z∞
(5.1) x2 ν(dx) < ∞
−∞

25
and
¡ ¢
(5.2) ν {0} < 1.

The filtration IF will be that generated by X. See Figure 1 for a graphical illustration of X.

(Insert Figure 1 here)

To simplify the notation, we shall denote the value of any F1 -measurable random variable
Y on the sets {X1 = +1}, {X1 = 0}, {X1 = −1} by Y (+) , Y (0) and Y (−) , respectively. Thus
we have for instance
(−)
∆A2 = E[∆X2 |X1 = −1] = 0.
It is then easy to check that
e(−) = λ
e1 = λ
λ e(0) = 0
2 2

and
R∞
x ν(dx)
e(+) E[∆X2 |X1 = +1] −∞
(5.3) λ = £ ¯ ¤= ∞ ;
2 2 ¯
E ∆X2 X1 = +1 R
x2 ν(dx)
−∞

this is well-defined by (5.1) and Jensen’s inequality, and

à !2
R∞
x ν(dx)
(+) (+) −∞
(5.4) e ∆A
λ = <1
2 2 R∞
2
x ν(dx)
−∞

by (5.2) and Jensen’s inequality. In particular, we see that (ND) is satisfied.

Next we compute the adjustment process β. By its definition, β2 = λ e2 and

E [∆X1 (1 − β2 ∆X2 )]
(5.5) β1 =
E [∆X12 (1 − β2 ∆X2 )2 ]
h i
E ∆X1 (1 − λ e2 ∆A2 )
= h i
E ∆X12 (1 − λ e2 ∆A2 )
à !2
R∞
− x ν(dx)
−∞
= Ã !2 ,
R∞ R∞
2 x2 ν(dx) − x ν(dx)
−∞ −∞

e2 , conditioning on F1 and (2.6), and the second uses (5.4)

where the first step uses β2 = λ
and the structure of X. This shows that the processes β and λe are different as soon as
Z∞
x ν(dx) 6= 0,
−∞

26
which in the present setting is equivalent to saying that X is not a martingale. Furthermore,
it is clear from (5.3) and (5.4) that both

Ze0 = (1 − β1 ∆X1 )(1 − β2 ∆X2 )

and
e1 ∆X1 )(1 − λ
(1 − λ e2 ∆X2 )
Zb2 =
e1 ∆A1 )(1 − λ
(1 − λ e2 ∆A2 )
will become negative with positive probability if supp ν is unbounded. This shows that both
Pe and Pb will in general not be measures, but only signed measures.

Example 3. Now consider the special case of Example 2 where

1 1 1
ν= δ{+2} + δ{+1} + δ{−1} ,
3 2 6
where δ{x} denotes a unit mass at the point x; see Figure 2.

(Insert Figure 2 here)

Then (5.3) – (5.5) simplify to

e(+) = 1 ,
λ 2
2
e(+) ∆A(+) = 1 ,
λ 2 2
2
1
β1 = − .
3
If we number the trajectories ω as ω1 to ω7 , starting with ω1 = {∆X1 = +1, ∆X2 = +2},
ω2 = {∆X1 = +1, ∆X2 = +1} and so on, we can write the random variable Ze0 as a vector,
µ ¶
2 2 2
Ze0 = 0, , 2, 1, 1, , .
3 3 3

Hence
7
E[Ze0 ] =
9
and
µ ¶
Ze0 6 18 9 9 6 6
(5.6) Ze = = 0, , , , , , .
E[Ze0 ] 7 7 7 7 7 7

Similarly, we obtain

(5.7) Zb2 = (0, 1, 3, 1, 1, 1, 1)

which clearly shows that Ze and Zb2 , hence also the measures Pe and Pb, do not agree. The
explicit expressions for Pe and Pb in terms of their transition probabilities are given in Figures
3 and 4, respectively. Note that both Pe and Pb are martingale measures for X, but not
equivalent to P .

27
 (Insert Figure 3 here)

(Insert Figure 4 here)

This example also illustrates another point. Since

h i
E e
Z 0
I
e Pe [{ω}] {ω}
Z(ω) = =
P [{ω}] E[Ze0 ]P [{ω}]

for every ω, Corollary 3.2 shows that P [{ω}] Z(ω)e gives the optimal initial capital V0 corre-
sponding to the contingent claim H := I{ω} . Similarly, Lemma 2.7 shows that P [{ω}] Zb2 (ω)
ω

equals the constant H0 in the decomposition (2.20) of H ω . A comparison of (5.6) and (5.7)
thus reveals that H0 and V0 will not agree in general. Moreover, there will be no general or-
dering between H0 and V0 ; both H0 > V0 and H0 < V0 can occur, as well as H0 = V0 . Ze and
Zb2 are often called state prices or state price densities with respect to Pe and Pb, respectively;
see for instance Back (1991).
Another interesting feature of Example 3 is the fact that it provides us with an example
of a contingent claim, namely

H ω1 = I{∆X1 =+1,∆X2 =+2} ,

which is bounded, nonnegative and positive with positive probability, and yet has both H0 = 0
and V0 = 0. This shows that an interpretation of either H0 or V0 as a fair price of H does
not always make sense from an economic point of view. The terminology “fair hedging price”
suggested by Schäl (1994) should therefore not be used carelessly. Note that the problem here
is not due to any inherent
¡ pathology of X; it is obvious
¢ that X admits an equivalent martingale
measure and therefore see Harrison/Kreps (1979) provides no arbitrage opportunities.
To round off the example, we now compute the optimal strategy ξ (c) and the mean and
second moment of the net loss H − c − GT (ξ (c) ) for the contingent claim H ω1 = I{ω1 } . First
of all, % in (2.11) is computed to be given by

(−) (0) (+) 1

%1 = %2 = %2 = 0 , %2 = .
3
Using Theorem 2.4, we then obtain ξ (c) as
(c) c
ξ1 = ,
3
³ ´(−) ³ ´(0)
(c) (c)
ξ2 = ξ2 = 0,
³ ´(+) 1 − 2c 1 c
(c)
ξ2 = = +c− .
3 3 3
This tells us that at time 0, we should buy 3c shares. If the stock goes up, we should buy
another 31 + c shares at time 1; otherwise, we simply sell our 3c shares at time 1. At time
2, we liquidate whatever
³ we have. We´ remark that ξ (c) could of course also be obtained by
(−) (0) (+)
parametrizing ξ as ξ1 , ξ2 , ξ2 , ξ2 and then minimizing the quadratic function
h¡ ¢2 i ³ ´
(−) (0) (+)
E H − c − G2 (ξ) = h ξ1 , ξ2 , ξ2 , ξ2 .

28
For the total net loss, Corollary 2.5 yields
h i 7c
E H − c − G2 (ξ (c) ) = −
9

and ·³ ´2 ¸
(c) 1 7c2
E H − c − G2 (ξ ) = +
27 9
with an obvious minimum at c = 0 = V0 .

Example 4. Our final example shows that if X does not satisfy (ND), then GT (Θ) need
not be closed in L2 (P ) and (1.1) may fail to have a solution. This counterexample is due to
Walter Schachermayer (private communication).
Let Ω = [0, 1] × {−1, +1} with its Borel σ-algebra F; elements of Ω will be denoted
by ω = (u, v) with u ∈ [0, 1], v ∈ {−1, +1}, and we denote by U (ω) := u the first and by
V (ω) := v the second coordinate. Let F0 = F1 = σ(U ), F2 = F , and let P be the measure
on (Ω, F ) such that U is distributed uniformly on [0, 1] and the conditional distribution of V
given U is U 2 δ{+1} + (1 − U 2 ) δ{−1} . Finally, let X0 = 0, ∆X1 = 1 and

∆X2 = V + (1 + U ) − 1 = V + U − V − ,

so that
∆X2 (u, v) = uI{v=+1} − I{v=−1} .
This model can be interpreted as follows. At time 0, we observe the value of a random
variable U distributed uniformly on [0, 1]. Whatever the value of U , X0 = 0 and X1 = 1. At
time 2, we toss a coin with (random) probability U 2 of getting heads. If the coin turns up
heads, ∆X2 = U ; otherwise, ∆X2 = −1.
Consider now the contingent claim

1 1
H=( + 1)V + = V + (1 + U ).
U U

Then H ∈ L2 (P ), since
· ¸
£ ¤ 1 £ + 2¯ ¤ £ ¤
E H 2 ¯
= E ( + 1) E (V ) U = E (1 + U )2 < ∞.
2
U

If ξ is a predictable process with G2 (ξ) = H P -a.s., then

1 + ¡ ¢
(5.8) V (1 + U ) = H = ξ1 ∆X1 + ξ2 ∆X2 = ξ1 + ξ2 V + (1 + U ) − 1
U

implies that
1
ξ1 = ξ2 = P -a.s.
U
by considering (5.8) separately on {V = +1} and {V = −1}. However,

1
ξ1 ∆X1 = / L2 (P )
∈
U
29
shows that ξ ∈
/ Θ, and since there is no other predictable process ϑ with G2 (ϑ) = H P -a.s.,
we conclude that
H∈ / G2 (Θ).
But if we set
1
ξ n := ξI{U ≥ 1 } = I 1 ,
n U {U ≥ n }
then
1
ξ1n ∆X1 = I ∈ L2 (P )
U {U ≥ n }
1

and
1 ¡ + ¢
ξ2n ∆X2 = V (1 + U ) − 1 I{U ≥ 1 } ∈ L2 (P ),
U n

n
hence ξ ∈ Θ for every n ∈ IN , and
1 +
G2 (ξ n ) = V (1 + U )I{U ≥ 1 } = HI{U ≥ 1 }
U n n

converges to H in L2 (P ). This shows that G2 (Θ) is not closed in L2 (P ) and that (1.1) does
not have a solution for this contingent claim H and for c = 0.
To conclude the example, we show that X violates condition (ND). In fact,
£ ¯ ¤
E[∆X2 |F1 ] = E V + (1 + U ) − 1¯U = U 3 + U 2 − 1

and £ ¯ ¤ £ ¯ ¤
E ∆X22 ¯F1 = E (V + U − V − )2 ¯U = U 4 − U 2 + 1,
and therefore the ratio
2
(E[∆X2 |F1 ]) (U 3 + U 2 − 1)2
£ ¯ ¤ =
E ∆X22 ¯F1 U4 − U2 + 1
is not uniformly bounded away from 1, since the right-hand side tends to 1 as U approaches
0.

Acknowledgement. This work is based on a part of my Habilitationsschrift at the Univer-

sity of Göttingen. Financial support by Deutsche Forschungsgemeinschaft, Sonderforschungs-
bereich 303 at the University of Bonn, is gratefully acknowledged.

References

K. Back (1991), “Asset Pricing for General Processes”, Journal of Mathematical Eco-
nomics 20, 371–395
C. Dellacherie and P.-A. Meyer (1982), “Probabilities and Potential B”, North-Holland
D. Duffie (1988), “Security Markets. Stochastic Models”, Academic Press
D. Duffie and H. R. Richardson (1991), “Mean-Variance Hedging in Continuous Time”,
Annals of Applied Probability 1, 1–15
H. Föllmer and M. Schweizer (1991), “Hedging of Contingent Claims under Incomplete
Information”, in: M. H. A. Davis and R. J. Elliott (eds.), “Applied Stochastic Analysis”,
Stochastics Monographs, Vol. 5, Gordon and Breach, London/New York, 389–414

30
 J. M. Harrison and D. M. Kreps (1979), “Martingales and Arbitrage in Multiperiod
Securities Markets”, Journal of Economic Theory 20, 381–408
J. M. Harrison and S. R. Pliska (1981), “Martingales and Stochastic Integrals in the
Theory of Continuous Trading”, Stochastic Processes and their Applications 11, 215–260
C. Hipp (1993), “Hedging General Claims”, Proceedings of the 3rd AFIR Colloquium,
Rome, Vol. 2, 603–613
N. Hofmann, E. Platen and M. Schweizer (1992), “Option Pricing under Incompleteness
and Stochastic Volatility”, Mathematical Finance 2, 153–187
D. Luenberger (1969), “Optimization by Vector Space Methods”, Wiley
S. M. Müller (1985), “Arbitrage Pricing of Contingent Claims”, Lecture Notes in Eco-
nomics and Mathematical Systems 254, Springer
M. Schäl (1994), “On Quadratic Cost Criteria for Option Hedging”, Mathematics of
Operations Research 19, 121–131
M. Schweizer (1988), “Hedging of Options in a General Semimartingale Model”, Diss.
ETHZ No. 8615
M. Schweizer (1991), “Option Hedging for Semimartingales”, Stochastic Processes and
their Applications 37, 339–363
M. Schweizer (1992), “Mean-Variance Hedging for General Claims”, Annals of Applied
Probability 2, 171–179
M. Schweizer (1993a), “A New Characterization of the Martingale Property”, SFB 303
discussion paper No. B–246, University of Bonn
M. Schweizer (1993b), “Approximating Random Variables by Stochastic Integrals, and
Applications in Financial Mathematics”, Habilitationsschrift, University of Göttingen
M. Schweizer (1993c), “Approximating Random Variables by Stochastic Integrals”, pre-
print, University of Göttingen, to appear in Annals of Probability

31