Technical note: Dependence and two-asset

options pricing

Grégory Rapuch
CREST & EHESS, Paris, France

Thierry Roncalli
Groupe de Recherche Opérationnelle, Crédit Agricole SA, France

In this short note, we consider some problems dealing with two-asset options
pricing. In particular, we investigate the relationship between options prices
and the “correlation” parameter in the Black–Scholes model. Then, we
consider the general case in the framework of the copula construction of risk-
neutral distributions. This extension involves results on the supermodular
order applied to the Feynman–Kac representation. We show that it could be
viewed as a generalization of the maximum principle for parabolic PDEs.

1 Introduction
In this paper, we address the problem of the relationship between the dependence
function and the price of two-asset options. For example, one important question
is whether the price of a spread option is a monotonic (decreasing or increasing)
function of the correlation parameter in the Black–Scholes model. Another ques-
tion is related to the (lower and upper) bounds of the option price with respect to
this correlation parameter.
The paper is organized as follows. In section 2, we consider the Black–Scholes
parabolic PDE to study the effect of the correlation parameter on the price of a
European option with two underlying assets when the volatilities are fixed. Using
the maximum principle, we show that if the cross-derivative of the payoff func-
tion is positive (resp. negative), then the option price is non-decreasing (resp.
non-increasing) with respect to the correlation parameter. Such a property is veri-
fied for a large class of traded two-assets options, for example spread and basket
options (see Table 1). These results provide simple bounds for such option prices
which correspond to the cases where the correlation parameter is equal to –1 and
1. In section 3, we generalize the foregoing results by using the martingale prop-
erty of the option price. Using a copula representation, we now study the effect of
the dependence function on the two-assets option price when the marginals are
fixed – ie, when the volatility smiles of each asset are fixed. Using a standard order
on the dependence functions, which is called the concordance order, we obtain

We acknowledge helpful discussions with Jérôme Busca, Nicole El Karoui, Jean-Frédéric

Jouanin and Gaël Riboulet and thank the anonymous referee for their suggestions and point-
ing us to the alternative proof of Proposition 2.

Volume 7/ Number 4, Summer 2004 23 URL: www.thejournalofcomputationalfinance.com

24 Grégory Rapuch and Thierry Roncalli

similar results to those in the Black–Scholes case: if the cross-derivative of the

payoff function is positive (resp. negative), then the option price is non-decreasing
(resp. non-increasing) with respect to the concordance order. Numerical illustra-
tions are considered in Section 4. Section 5 concludes and suggests that the results
of this paper may not be generalized in the multivariate case.

2 The case of the Black–Scholes model

In the Black–Scholes (BS) model with constant coefficients, the asset prices are
modeled as correlated geometric Brownian motions

d S1( t ) = µ1S1( t ) dt + σ1S1( t ) dW1 ( t ) (1)


d S2 ( t ) = µ 2 S2 ( t ) dt + σ 2 S2 ( t ) dW2 ( t )
where E[W1(t)W2 (t)] = ρt. Using the Feynman–Kac representation formula (see
Friedman, 1975), the price P(S1, S2, t) of the European two-asset option with
the payoff function G(S1, S2) is the (unique) solution of the following parabolic
PDE1

 1 σ 2 S 2 ∂2 P + ρσ σ S S ∂2 P + 1 σ 2 S 2 ∂2 P + b S ∂ P +
 2 1 1 1,1 1 2 1 2 1, 2 2 2 2 2,2 1 1 1

 b2 S2 ∂2 P − rP + ∂t P = 0
 (2)
P ( S1, S2 , T ) = G ( S1, S2 )


where b1 and b2 are the cost-of-carry parameters and r is the instantaneous con-
stant interest rate. First, we consider the case of the spread option and show that
the price is a non-increasing function of the correlation parameter ρ. Second, we
extend this result to other two-asset options. Moreover, we give explicit lower and
upper bounds for these option prices.

2.1 An example with the spread option

Before stating the main proposition, we recall the weak maximum principle for
parabolic PDEs (see for example Protter and Weinberger, 1967).

THEOREM 1 (Phragmen–Lindeloff principle) We consider the operator Lu(x, t)

= ∑ i, j a i, j (x, t)∂2i, j u (x, t) + ∑i b i (x, t)∂i u (x, t) + ∂ t u (x, t) + c (x, t) u (x, t) where
the functions ai, j (x, t), bi (x, t) and c(x, t) are continuous and bounded. Moreover,
we assume that (ai, j (x, t)) is a symmetric, positive definite matrix for all (x, t)
∈Rm × [0, T). Let w ∈C 2 (Rm × [0, T)) ∩ C 0 (Rm × [0, T] ) with | w(x, t) | ≤ β eα || x ||
for all (x, t) ∈Rm×[0, T] and some constants α and β. If Lw(x, t) ≤ 0 for t < T and
w(x, T) ≥ 0, then we have w(x, t) ≥ 0 for all (x, t) ∈Rm × [0, T]

1∂ is the partial derivation operator with respect to space variables Si.

i

URL: thejournalofcomputationalfinance.com Journal of Computational Finance

 Dependence and two-asset options pricing 25

In order to obtain qualitative properties of the spread option, it is more convenient

to deal with an elliptic operator. The change of variables S̃1 = ln S1 and S̃ 2 = ln S2
leads us to

 1 σ 2 ∂2 P + ρσ σ ∂2 P + 1 σ 2 ∂2 P + b + 1 σ 2 ∂ P +
 2 1 1,1 1 2 1, 2 2 2 2,2 1 2 1 1 ( )



1
( )
b2 + σ 22 ∂2 P − rP + ∂t P = 0
2
(3)

( ) (
 P S , S , T = exp S − exp S − K
 1 2 2 1
+
)
The operator L ρ u = 12– σ 12 ∂12, 1 u + ρσ1 σ2 ∂ 1,
2 u + 1– σ 2 ∂ 2 u + (b + 1– σ 2 ) ∂ u +
2 2 2 2, 2 1 2 1 1
(b 2 + 2– σ 2 )∂2 u – ru + ∂ t u is also elliptic for ρ ∈]–1, 1[. We can now establish the
1 2

following proposition.

PROPOSITION 1 The price of the spread option in the Black–Scholes model is a

non-increasing function of ρ for ρ ∈[–1, 1] .

Proof: The complete proof is given in Appendix A. We give here just an outline
of the proof. We first verify the exponential growth condition. Then, we consider
the case ρ1 < ρ2 and compute the difference function ∆(S̃1, S̃ 2, t) = Pρ1(S̃1, S̃ 2, t)
– Pρ2 (S̃1, S̃ 2, t). It turns out that ∆ is the solution of the following PDE:

( ) (
Lρ ∆ S1, S2 , t = ( ρ2 − ρ1) σ1σ 2 ∂12, 2 Pρ S1, S2 , t
 1 2
)

(   )
 ∆ S1, S2 , T = 0
In order to apply the maximum principle to ∆(S̃1, S̃ 2, t), we would like to show
that ∂ 12 , 2Pρ2 (S̃1, S̃ 2, t) ≤ 0. We show this by using again the maximum principle to
∂ 12 , 2Pρ2 (S̃1, S̃ 2, t), which is the solution of another PDE. Finally, ∆(S̃1, S̃ 2, t) ≥ 0.
So, we conclude that

( ) (
ρ1 < ρ2 ⇒ Pρ S1, S2 , t ≥ Pρ S1, S2 , t
1 2
) (4)

REMARK 1 The spread option price is a one-to-one mapping with respect to the
parameter ρ. To any price corresponds one only parameter ρ. This is the implied
BS correlation.2

2.2 Other two-asset options

We may extend the previous proposition to other two-asset options. We remark
that the key point in the proof is the sign of the cross derivative ∂ 12 , 2G(S1, S2) of the

2In fact, we can show that increasingness is strict by using a strong maximum principle
which is available as soon as the operator is strictly elliptic (Nirenberg, 1953).

Volume 7/ Number 4, Summer 2004 URL: www.thejournalofcomputationalfinance.com

26 Grégory Rapuch and Thierry Roncalli

payoff function. In general, this differential is a measure and G does not depend
on the parameter ρ.

PROPOSITION 2 Let G be the payoff function. If ∂ 12 , 2G is a non-positive (resp. non-

negative) measure then the option price is non-increasing (resp. non-decreasing)
with respect to ρ.

Let us investigate some examples. For the call option on the maximum of two
assets, the payoff function is defined as G(S1, S2) = (max(S1, S2) – K) +. We have
∂ 12 , 2G(S1, S2) = –δ{S = S , S > K} which is a non-positive measure. So, the option
1 2 1
price non-increases with respect to ρ. In the case of a BestOf call ⁄call option,
the payoff function is G(S1, S2) = max((S1 – K1) +, (S2 – K2) +) and we have
∂ 12 , 2G(S1, S2) = – δ{S – K – S + K = 0, S > K , S > K }. We have the same behavior
2 2 1 1 1 1 2 2
as in the Max option. For the Min option, we remark that min(S1, S2) = S1 + S2 –
max(S1, S2) . So, the price is a non-decreasing function of ρ. Other results can be
found in Table 1.

2.3 Bounds of two-asset options prices

The previous analysis leads us to define the lower and upper bounds of two-asset
options price when the parameter ρ is unknown. Let P – (S1, S2, t) and P + (S1, S2, t)
be the lower and upper bounds respectively:

P − ( S1, S2 , t ) ≤ Pρ ( S1, S2 , t ) ≤ P + ( S1, S2 , t ) (5)

We have the following result.

PROPOSITION 3 If ∂ 12 , 2G is a non-positive (resp. non-negative) measure, then

P – (S1, S2, t) and P + (S1, S2, t) correspond to the cases ρ = 1 (resp. ρ = –1) and ρ =
–1 (resp. ρ = 1). The two-dimensional PDE reduces then to a one-dimensional
PDE.

TABLE 1 Relationship between option prices and the parameter ρ.

Option type Payoff Increasing Decreasing

Spread (S2 – S1 – K) + ✓
Basket (α1S1+α2S2 – K) + α1α2 > 0 α1α2 < 0
Max (max(S1,S2 ) – K) + ✓
Min (min(S1,S2 ) – K) + ✓
BestOf call/call max((S1 – K1) +, (S2 – K 2 ) +) ✓
BestOf put/put max((K1 – S1) +, (K 2 – S2 ) +) ✓
WorstOf call/call min((S1 – K1) +, (S2 – K 2 ) +) ✓
WorstOf put/put min((K1 – S1) +, (K 2 – S2 ) +) ✓

URL: thejournalofcomputationalfinance.com Journal of Computational Finance

 Dependence and two-asset options pricing 27

2.4 Extension to Dupire’s local volatility model

We point out that all the previous results remain true in the case of the local vola-
tility model :

d S1( t ) = µ1S1( t ) d t + σ1( t , S1( t ) ) dW1 ( t )

 (6)
d S2 ( t ) = µ 2 S2 ( t ) d t + σ 2 ( t , S2 ( t ) ) dW2 ( t )

where E[W1(t)W2 (t))] = ρt. The proof is exactly the same as the one we followed
in Proposition 2, except one minor difference (see footnote 5).

REMARK 2 As pointed by the referee, there is a simpler and more intui-

tive method than the maximum principle to show Proposition 2. Using the
method described in Section 4 of Carr (2000), we can formally express the
price as P (S1, S2, t) = exp(L · (T – t))G(S1, S2), where L is the Black–Scholes
operator. Under some mild regularity assumption, we obtain ∂P ⁄ ∂ρ =
(T – t)σ1σ2 e –r(T – t) EQ [∂1, 2G]. This proves the desired results. However, this
method works only for constant coefficients.

3 The general case

In this section, we generalize the foregoing results. For that, we know that the
European prices of two-asset options are given by

P ( S1, S2 , t ) = e − r ( T − t ) E Q G ( S1 ( T ), S2 ( T ) ) Ft  (7)

with Q the martingale probability measure. Let F be the bivariate risk-neutral

distribution at time t. Using the copula construction,3 we have

F ( S1, S2 ) = C ( F1 ( S1 ), F2 ( S2 ) ) (8)

where F1 and F2 are the two univariate risk-neutral distributions. C is called the
risk-neutral copula.
In the Black–Scholes model and for the spread option, we have proved that
if ρ1 < ρ2, then Pρ1(S1, S2, t) ≥ Pρ2 (S1, S2, t) and that the lower and upper bounds
are reached respectively for ρ = 1 and ρ = –1. Now, we are going to give similar
results for the general case. Let us define the concordance order C1 ≺ C2 such that
C1(u1, u2) ≤ C2 (u1, u2) for all u1, u2 in [0, 1] 2. For the spread option, we will prove
that if C1 ≺ C2, then PC1(S1, S2, t) ≥ PC2 (S1, S2, t) and that the lower and upper
bounds are reached for the upper and lower Fréchet bounds. These results are all
based on properties of the supermodular order.

3 A copula function joins univariate distribution functions to form bivariate distribution

functions. It is itself a bivariate distribution on the unit square with uniform marginals (see
Nelsen, 1999).

Volume 7/ Number 4, Summer 2004 URL: www.thejournalofcomputationalfinance.com

28 Grégory Rapuch and Thierry Roncalli

3.1 Supermodular order

Following Müller and Scarsini (2000), we say that the function f is supermodular
if and only if

∆ ( 2 ) f := f ( x1 + ε1 , x2 + ε 2 ) − f ( x1 + ε1 , x2 ) − f ( x1 , x2 + ε 2 ) + f ( x1, x2 ) ≥ 0 (9)

for all (x1, x2) ∈ R2 and (ε1, ε2) ∈ R+2. If f is twice differentiable, then the condition
(9) is equivalent to ∂ 12 , 2 f (x1, x2) ≥ 0 for all (x1, x2) ∈ R2. We can then show the fol-
lowing relationship between the concordance order and supermodular functions.
THEOREM 2 Let F1 and F2 be the probability distribution functions of X1 and
X2. Let EC[ f (X1, X2)] denote the expectation of the function f (X1, X2) when the
copula of the random vector (X1, X2) is C. If C1 ≺ C2 , then E C1 [ f (X1, X2)] ≤
E C2 [ f (X1, X2)] for all supermodular functions f such that the expectations exist.
Proof: See Tchen (1980) and Müller and Scarsini (2000).

3.2 Concordance order and two-asset options prices

Using Theorem 2, we have directly the following proposition.
PROPOSITION 4 If the payoff function G is supermodular, then the option price
non-decreases with respect to the concordance order. More explicitly, we have

C1 ≺ C2 ⇒ PC ( S1, S2 , t ) ≤ PC ( S1, S2 , t ) (10)

1 2

Let us consider the simple case of the basket option where G(S1, S2) =
(S1 + S2 –K) +. We have that ∂ 12 , 2 G is a non-negative measure and that G is
continuous. This is enough to assert that G is supermodular (see Appendix B
for a proof). Using Proposition 4, if C1〈S1(T ), S2 (T )〉 ≺ C2 〈S1(T ), S2 (T )〉, we
find that E C1[G (S1(T ), S2 (T ))] ≤ E C2 [G (S1(T ), S2 (T ))]. So, we deduce that the
price of the basket option does not decrease with respect to the concordance
order and that the price of the spread option does not increase with respect to
the concordance order. We can prove this last statement in two different ways.
Indeed, for the spread option, we have C1〈–S1(T ), S2 (T )〉  C2〈–S1(T ), S2 (T )〉
because C〈 –X1, X2 〉 (u1, u2) = u2 – C〈X1, X2 〉 (1 – u1, u2) (Nelsen 1999). Hence, using
the supermodularity of the basket payoff function, E C1[G (–S1(T ), S2 (T ))] ≥
E C2[G (–S1(T ), S2 (T ))]. Another way to derive the result is to remark that ∂ 12 , 2 H is
a non-positive measure, where H(S1, S2) = G(–S1, S2). Thus, –H is supermodular
and the result is a consequence of the next proposition.
We can then generalize the results of Table 1 in this framework. To this aim,
we state the main proposition of this section.
PROPOSITION 5 Let G be a continuous payoff function. If the distribution ∂ 12 , 2 G
is a non-negative (resp. non-positive) measure then the option price is non-
decreasing (resp. non-increasing) with respect to the concordance order.

URL: thejournalofcomputationalfinance.com Journal of Computational Finance

 Dependence and two-asset options pricing 29

Proof: See Appendix B

REMARK 3 This last proposition is interesting because families of copulas are

generally totally (positively or negatively) ordered. For example, we know that the
parametric family Cρ of bivariate Normal copulas is positively ordered
ρ1 < ρ2 ⇒ Cρ ≺ Cρ (11)
1 2

As for the Black–Scholes model, the option price is non-increasing or non-

decreasing with respect to the “correlation” parameter ρ, depending on the
submodular or supermodular property of the payoff function. Therefore, the
results of the Black–Scholes model can be viewed as a special case of this frame-
work. Moreover, they remain true if the bivariate risk-neutral distribution is not
Gaussian, but only has a Normal copula. Other cases are considered in Coutant et
al (2001). For example, we obtain similar relationships with Ornstein–Uhlenbeck
diffusions.

3.3 Bounds of two-asset options prices

Let us introduce the lower and upper Fréchet copulas C– (u1, u2) = max(u1 + u2 –
1, 0) and C + (u1, u2) = min(u1, u2) . We can prove that for any copula C, we have
C– ≺ C ≺ C +. For any distribution F with given marginals F1 and F2, we find
that C– (F1(x1), F2 (x2)) ≤ F(x1, x2) ≤ C + (F1(x1), F2 (x2)) for all (x1, x2) ∈ R2+ . The
probabilistic interpretation of these Fréchet bounds are the following:
❑ two random variables X1 and X2 are countermonotonic – or C〈 X1, X2〉 = C – – if
there exists a random variable X such that X1 = f1(X) and X2 = f 2 (X) with f1 non-
increasing and f 2 non-decreasing;
❑ two random variables X1 and X2 are comonotonic – or C〈X1, X2〉 = C + – if there
exists a random variable X such that X1 = f1(X) and X2 = f 2 (X) where the func-
tions f1 and f 2 are non-decreasing.
We can now state the following proposition.

PROPOSITION 6 If ∂ 12 , 2 G is a non-positive (resp. non-negative) measure then

P – (S1, S2, t) and P + (S1, S2, t) correspond to the cases C = C + (resp. C = C –) and
C = C – (resp. C = C +).

4 Discussion
We conclude this paper with some remarks.

❑ We recall that the main results depend on the sign of ∂ 12 , 2 G. Using two different
points of view, we obtain the same condition. It appears that results obtained
with a maximum principle for the Black–Scholes model are a special case of the
supermodular order. It could be explained by the Feynman–Kac representation
of risk-neutral valuation.

Volume 7/ Number 4, Summer 2004 URL: www.thejournalofcomputationalfinance.com

30 Grégory Rapuch and Thierry Roncalli

FIGURE 1 Relationship between the price of the WorstOf option and the correlation
parameter.

S2 = 95, τ = 3M, σ1 = 30% S2 = 105, τ = 3M, σ1 = 30%

6.5 4

6.0 3

5.5 2

5.0 1
–1.0 –0.5 0.0 0.5 1.0 –1.0 –0.5 0.0 0.5 1.0
ρ ρ

S2 = 95, τ = 1M, σ1 = 30% S2 = 95, τ = 3M, σ1 = 10%

3.5 4.01

3.96

3.91
3.0
3.86

3.81

2.5 3.76
–1.0 –0.5 0.0 0.5 1.0 –1.0 –0.5 0.0 0.5 1.0
ρ ρ

❑ If ∂ 12 , 2 G is not a positive or negative measure, it means that the price is not

necessarily monotonous with respect to the concordance order. It will depend
on the parameters of the option contract (for example, the maturity) and the
market conditions (the univariate risk-neutral distributions). To illustrate
this, let us consider a WorstOf call/put option with a barrier G(S1, S2) =
1{S2 < 105} · max((S1 – 105) +, (95 – S2) +). In Figure 1, we have reported the
option prices for different values of the initial price S2 of the second asset, the
volatility σ1 of the first asset, and the maturity τ of the option in the case of the
Black–Scholes model.4 From a risk management point of view, it is then diffi-
cult to define a general rule to apply a conservative price for this type of option,
because the rule must depend on the market conditions and the contract.
❑ Similar problems have been already studied in actuarial sciences. For example,
Dhaene and Goovaerts (1996) show that the bounds of the stop-loss problem
are reached for the Fréchet bounds (see Genest et al, 2002, for a survey).
❑ Maximum principles are commonly used in the analysis of partial differential
equations. They can be used to study the dependence of a solution with respect

4 We assume that the volatility of the second asset is 20%, the initial price of the first asset is

100 and the interest rate is zero.

URL: thejournalofcomputationalfinance.com Journal of Computational Finance

 Dependence and two-asset options pricing 31

to the coefficients of the PDE. Such technics were used in Berestycki (2000) to
study the influence of the volatility parameter on a multi-asset European option.
This technique enabled the author to derive many qualitative properties, such
as the convexity of the option with respect to the asset price. This point of view
was generalized to American options in Rapuch (2003) and Touzi (1999) in the
context of viscosity solution.
❑ The natural following extension is to consider more than two assets. In the case
of the Black–Scholes model, the PDE becomes

1
2 ∑ ∑
σ i2 Si2 ∂i2, i P + ρi , j σ i σ j Si S j ∂i2, j P + ∑ bi Si ∂i P − rP + ∂t P = 0
 i i< j i (12)
 P ( S ,…, S , T ) = G ( S ,…, S )
 1 N 1 N

where ρi,j is the correlation between the Brownian motions of Si and Sj . If we fix

all the correlations ρi, j except one, we retrieve the same condition as in the two-
assets options case. Sometimes, the trader uses the same values for all ρi, j. Let
us denote ρ this parameter, which could be interpreted as the mean correlation.
We can give the following result.

Proposition 7 Assume that G is continuous. If ∑i < j σi σj ∂i,2 j G is a non-negative

(resp. non-positive) measure, then the price is non-decreasing (resp. non-
increasing) with respect to ρ.

In the case of the three-asset option with G(S1, S2, S3) = (S1+S2 – S3 – K) +, we
have ∑i < j σi σj ∂i,2 j G = (σ1σ2 – σ1σ3 – σ2σ3) δ(S1 + S2 – S3 – K = 0) . Hence, if σ1σ2 –
σ1σ3 – σ2σ3 > 0, the price non-decreases with ρ, and if σ1σ2 – σ1σ3 – σ2σ3 < 0,
the price non-increases. In addition, if σ1σ2 – σ1σ3 – σ2σ3 = 0, the price does
not depend on ρ. We could of course give similar results for the Max and for
more general basket options.
❑ Moreover, one might wonder if the method using the concordance order can
be generalized with more than two assets. Müller and Scarsini (2000) show
that the supermodular order is strictly stronger than the concordance order for
dimensions greater than three. So the method used for two-asset options cannot
be generalized here. This is not surprising if we consider the example above:
the condition about the sign of ∑i < j σi σj ∂i,2 j G involves the values of the volatili-
ties which are independent of the payoff function.

As a result, it is more difficult to define conservative prices for multi-asset options.

Understanding the relationship between the stochastic dependence and the price
of equity structured products is then a challenge for both the front office and the
risk management.

Volume 7/ Number 4, Summer 2004 URL: www.thejournalofcomputationalfinance.com

32 Grégory Rapuch and Thierry Roncalli

Appendix A – Proof of Proposition 2

We first prove that there exist two constants α and β such that | P(S̃1, S̃ 2, t) | ≤
β exp (α || S̃1, S̃ 2|| ). Using standard stochastic calculus (see, for example, Jaillet et
al, 1990, or Villeneuve, 1999), for all M > 0, there exists a constant C such that
E [sup{t ≤ s} exp (M || S̃ts, (s1̃, s̃2) || )] ≥ C exp (M || s̃1, s̃2 || ) where S̃ts, (s1̃, s̃2) = (S̃1(t))|S̃1(s) =
s̃1,S̃ 2 (t) | S̃ 2 (s) = s̃2). By noting that | P(S̃1, S̃ 2, T) | ≤ β exp (α || S̃1, S̃ 2||), the upper
bound follows from the Feynman–Kac representation.
We can then use the maximum principle. This part of the proof is adapted
from Berestycki (2000). Let –1 ≤ ρ1 < ρ2 ≤ 1 and Pρ be the solution of the PDE.
We consider ∆(S̃1, S̃ 2, t) = Pρ1(S̃1, S̃ 2, t) – Pρ2 (S̃1, S̃ 2, t). ∆ is then the solution of the
following PDE

( ) (
Lρ ∆ S1, S2 , t = ( ρ2 − ρ1) σ1σ 2 ∂12, 2 Pρ S1, S2 , t
 1 2
)

( 
 ∆ S1, S2 , T = 0 )
The weak maximum principle asserts that if L ρ1 ∆ ≤ 0 for (S̃1, S̃2, t) ∈ R2 × [0, T)
and if ∆(S̃1, S̃ 2, T) ≥ 0 for (S̃1, S̃ 2) ∈ R2, then ∆(S̃1, S̃ 2, t) ≥ 0 for (S̃1, S̃ 2, t) ∈ R2 ×
[0, T] . We assume that the solution is smooth (say C ∞ in the domain where t < T).
We can differentiate with respect to S̃1 and S̃ 2 the equation Lρ 2Pρ 2 (S̃1, S̃ 2, t) = 0
and we get5 Lρ 2 ∂ 12 , 2 Pρ 2 (S̃1, S̃ 2, t) = 0. For the terminal condition, we use a convo-
lution product with an identity approximation because the payoff is not smooth.
Let θ(x1, x2) be a positive function C ∞ (R2) with its support in B(0, 1) satisfying
∫∫B(0, 1)θ(x1, x2)dx1 dx2 = 1. We define θm (x1, x2) = m –2 θ(m –1x1, m –1x2). We now
consider ψm (S̃1, S̃ 2, t) = (∂ 12 , 2 Pρ 2 * θm)(S̃1, S̃ 2, t) = (∂ 12 , 2 θm * Pρ 2)(S̃1, S̃ 2, t). We
know that ψm is C ∞. By using the properties of the convolution product, we get
Lρ 2ψm (S̃1, S̃ 2, t) = 0. We just have to prove that ψm (S̃1, S̃ 2, T) ≤ 0 and | ψm (S̃1, S̃ 2, t) |
≤ β exp (α || S̃1, S̃ 2|| ) .
The first step can be done by calculating ∂ 12 , 2 Pρ 2 (S̃1, S̃ 2, T) using the jump
formula and the relationship ∂ 12 , 2 Pρ 2 (S̃1, S̃ 2, T) = S1S2 ∂ 12 , 2 Pρ 2 (S1, S2, T). We also
get ∂ 12 , 2 Pρ 2 (S1, S2, T) = – δ{S2 – S1 – K = 0} where δ is the Dirac measure. Because
∂ 12 , 2 Pρ 2 is a non-positive measure, ψm (S̃1, S̃ 2, T) ≤ 0.
To establish the bound, we observe that the support of the function ∂ 12 , 2 θm is
included in B(0, R) for some constant R and that there exists a constant M such
that | ∂ 12 , 2 θm (S̃1, S̃ 2) | ≤ M. We find that | ψm (S̃1, S̃ 2, t) | ≤ β′ exp(α′|| S̃1, S̃ 2 || ) where
α′ and β′ do not depend on time. So, ψm (S̃1, S̃ 2, t) ≤ 0 because of the maximum
principle. Moreover, ψm (S̃1, S̃ 2, t) converges pointwise to ∂ 12 , 2 Pρ 2 (S̃1, S̃ 2, t) because
∂ 12 , 2 Pρ 2 (S̃1, S̃ 2, t) is continuous for t < T. We finally obtain that Lρ 1 ∆(S̃1, S̃ 2, t) ≤ 0
because ∂ 12 , 2 Pρ 2 (S̃1, S̃ 2, t) ≤0. This completes the proof.

5 In the case of the local volatility model, we get L*∂ 1,

2 P (S̃ , S̃ , t) = 0 where L* is another
2 ρ2 1 2
elliptic operator.

URL: thejournalofcomputationalfinance.com Journal of Computational Finance

 Dependence and two-asset options pricing 33

Appendix B – Proof of Proposition 5

We just have to see that when ∂ 12 , 2 G exists and is continuous, we have ∆(2) G =
S +ε S +ε
∫S22 2∫S11 1∂ 12 ,2 G ≥ 0. When ∂ 12 ,2 G is a distribution, we use the same method as
in Appendix A. We consider the same kernel θm so that G * θm is smooth enough.
We get ∆(2)(G * θm) = ∫SS2 + ε2∫SS1 + ε1∂ 12 , 2 (G * θm) = ∫SS2 + ε2∫SS1 + ε1 (∂ 12 , 2 G) * θm ≥ 0. When
2 1 2 1
m tends to ∞, we obtain ∆(2) (G * θm) → ∆(2) G ≥ 0 because G is continuous. So the
function G is supermodular. Using Proposition 4, we get the first result.
If the distribution ∂ 12 , 2 G is a non-positive measure, –G is supermodular and we
have directly the result.

REFERENCES
Berestycki, H. (2000). Utilisation des EDP en finance. Université de Paris 6, DEA d’Analyse
Numérique.
Black, F., and Scholes, M. (1973). The pricing of corporate liabilities. The Journal of Political
Economy 81(1), 637–54.
Carr, P. (2000). Deriving derivatives of derivative securities. Journal of Computational
Finance 4(2), Winter 2000/2001, 5–29.
Dhaene, J., and Goovaerts, M. J. (1996). Dependency of risks and stop-loss order. Astin
Bulletin 26(2), 201–12.
Friedman, A. (1975). Stochastic Differential Equations and Applications. Academic Press.
Genest, C., Marceau, E., and Mesfioui, M. (2002). Upper stop-loss bounds for sums of pos-
sibly dependent risks with given means and variances. Statistics and Probability Letters
57, 33–41.
Jaillet, P., Lamberton, D. and Lapeyre, N. (1990). Variational inequalities and the pricing of
american options. Acta Applicandae Mathematicae 21, 263–89.
Müller, A., and Scarsini, M. (2000). Some remarks on the supermodular order. Journal of
Multivariate Analysis 73(1), 107–19.
Nelsen, R. B. (1999). An Introduction to Copulas. Lectures Notes in Statistics, 139. Springer-
Verlag.
Nirenberg, L. (1953). A strong maximum principle for parabolic equations. Communications
on Pure and Applied Mathematics 6, 167–77.
Protter, M. H., and Weinberger, H. F. (1967). Maximum Principles in Differential Equations.
Prentice-Hall.
Rapuch, G. (2003). American options and the free boundary exercise region: a PDE approach,
CREST & EHESS, Working paper.
Tchen, A.H. (1980). Inequalities for distributions with given marginals. Annals of Statistics
8(4), 814–27.
Touzi, N. (1999). American options exercise boundary when the volatility changes randomly.
Applied Mathematics and Optimization 39, 411–22.
Villeneuve, S. (1999). Options américaines dans un modèle de Black–Scholes multi-
dimensionnel. Université de Marne la Vallée. PhD Dissertation.

Volume 7/ Number 4, Summer 2004 URL: www.thejournalofcomputationalfinance.com