LOCAL VOLATILITY MODEL AND MULTI-DIMENSIONALITY

DANIEL MONTENEGRO DA FONSECA LIMA

Dissertation submitted for the MSc in Mathematical Finance

Department of Mathematics
University of York

24 April 2022

Supervisor: Maciej J. Capiński

1
2 DANIEL MONTENEGRO DA FONSECA LIMA

Contents
Acknowledgements 3
1. Introduction 4
2. Preliminaries 6
2.1. Stochastic processes and stochastic calculus 6
2.2. Black-Scholes model with dividends 12
2.3. Multi-dimensional Black-Scholes model 15
3. Dupire’s equation for a single asset 21
4. Dupire’s equation for several assets 25
4.1. Proof 26
4.2. Numerical example 32
5. Conclusion 35
References 36
 LOCAL VOLATILITY MODEL AND MULTI-DIMENSIONALITY 3

Acknowledgements
I would like to thank my supervisor Dr. Maciej Capiński for his constant guid-
ance and support which were key elements in my learning journey and in this
project. I would also like to thank Dr. Marek Capiński and Dr. Tomasz Za-
stawniak for sharing their knowledge and giving me the opportunity to join this
programme. This project would not have been possible without Benedikt Petko
and Dr. Fabio Rodrigues who helped me bridge the gap between my Economics
background and this programme. Thank you Dr. Rodrigo Lanna and Guilherme
Abbud for opening the doors that lead me to this fulfilling path.
I would like to thank my wife Julia, my parents Marcia and Edson, and my
brother Lazaro for their never-ending support and for always believing in me.
4 DANIEL MONTENEGRO DA FONSECA LIMA

1. Introduction
Since the Black-Scholes model was introduced in 1973, there have been multiple
extensions to the model. One of such extensions, proposed by Bruno Dupire, is
popularly known as the Local Volatility model and gives a method for recovering
the volatility of the underlying asset’s price based on the option smile. Dupire’s
model is widely used in the industry, usually in the context of barrier option pricing.
One of the limitations of the Local Volatility model is that it is only applicable for
derivatives with a single underlying asset. Given the importance of basket and
index options, there have been multiple proposals to generalise Dupire’s model for
the setting where there are several underlying assets. This text covers one of these
proposals in detail.
In the multi-dimensional case, we assume that the interest rate and the divi-
dend rates are time-dependent. Additionally, the model assumes the risky assets
follow a multi-dimensional Ito Process with a time-dependent drift coefficient and,
differently than the original Dupire model, where the volatility was also state de-
pendent, the volatility is assumed to be only time-dependent. When dealing with
several assets, there is also added complexity coming from the fact that the model
needs to take into account the correlations between the risky assets.
The objective is to provide a method to recover the covariance matrix from
the prices of basket options. This introduces an important variable that is not
present in the one dimensional case: the weight of each risky asset in the basket.
These variables are of great importance given that, for a given price path of the
risky assets, for every different set of weights you have a different basket and,
consequently, a different basket price path. This adds to the complexity of the
problem and to address it we introduce different spaces. These spaces work with
either different combinations of weights and price paths of each risky asset that
composes the basket, or they work directly with the basket price paths as variables.
With that, one of the main tools used in the derivation of Theorem 4.6 is the change
of measure, where we constantly shift between these spaces to simplify partial
derivatives and integrals. We find that the weights and the partial derivatives with
respect to the weights of the risky assets are a central part of the derivation. Just
like in the one dimensional case, integration by parts is employed to simplify higher
order partial derivatives. Lastly, the multi-dimensional Fokker-Planck equation also
plays a key role by providing a suitable expression for the option’s theta.
We have devised a numerical example to illustrate how the resulting Theorem
4.6 can be applied. In the example, we provide prices of basket options that would
in practice be given by the market and proceed to apply finite differences to ap-
proximate the value of the relevant derivatives. The first order partial derivatives
were calculated by forward difference approximation and the second order partial
derivatives by central difference approximation. The application of the Theorem
4.6 results in a system of equations that can be solved, allowing us to recover the
covariance matrix from the prices of the basket options.
The main body of the text is organised in three parts. The first one covers the
preliminaries. It uses concepts from analysis, measure theory and probability theory
to introduce stochastic processes and the tools of stochastic calculus. It then covers
the Black-Scholes model with dividends and the multi-dimensional Black-Scholes
model.
 LOCAL VOLATILITY MODEL AND MULTI-DIMENSIONALITY 5

The second part of the text covers the Dupire model. It introduces the Dupire
equation and provides a step by step proof of the result which is more detailed than
the referenced papers [4, 5].
The third and final part covers the model from the paper [6], where the au-
thors attempted to generalise the Local Volatility model to multiple dimensions
and proposed a theorem for the case where the volatility is time-dependent but not
asset-price dependent. While the original paper provides the proof, we find that
the free text structure and the omission of important elements such as intermediary
steps can make it challenging to the reader. With that, we provide a detailed step
by step proof, with well-structured definitions, propositions, lemmas and theorems.
In other words, the flowing narrative from [6] has been re-cast into a theorem-proof
style, where we identify the key results, take care to formulate explicitly all the
needed assumptions and provide step proofs. We have also subdivided the proof
of the main result into steps by formulation appropriate lemmas, to make the ar-
gument easier to follow. The example presented in the dissertation of the how
the method can be applied was not included in the source paper [6] and has been
devised independently.
6 DANIEL MONTENEGRO DA FONSECA LIMA

2. Preliminaries
The first half of this chapter consists in a review of the fundamental building
blocks of stochastic processes and stochastic calculus. This will serve as a base for
the second half, where we apply stochastic calculus in the context of derivatives
pricing by introducting the Black-Scholes model with dividends. We finish the
chapter by exploring the Multi-dimensional Black-Scholes model.
We provide some of the proofs. For the proofs not provided, please refer to the
relevant cited sources.
2.1. Stochastic processes and stochastic calculus. A continuous time stochas-
tic process is a family of random variables indexed by time.
Definition 2.1. Let (Ω, F, P ) be a general probability space, T = [0, T ], 0 < T ≤
∞ and B(T) be the σ-field of Borel subsets of T. A mapping X : T × Ω → R
measurable with respect to the σ-field B(T) × F = σ({A × B : A ∈ B(T), B ∈ F}]
is a stochastic process. For each fixed ω0 ∈ Ω, the mappings X(·, ω0 ) : T → R
are the paths of the stochastic process and, for each fixed t0 ∈ T, the mappings
X(t0 , ·) : Ω → R are random variables.
Definition 2.2. A partition of a set T = [0, T ] is a collection T1 , T2 , ..., Tn of
subsets of T such that:
Sn
(1) i=1 Ti = T
(2) Ti ∩ Tj = ∅ if i 6= j
Definition 2.3. A filtration is a sequence of σ-fields Fi , i = 0, 1, ..., n, such that
Fi ⊂ Fi+1 for each i.
Definition 2.4. The filtration FtX generated by a stochastic process X(t, ω) is the
smallest σ-field such that the random variables X(s) for all s ≤ t are measurable
with respect to it. That is,
\
FtX = {G : G is a σ-field, FX(s) ⊂ G for all s ≤ t}.

In essence, the filtration FtX captures all the information available from X(t, ω)
at time t.
Given that the focus of this text is the pricing of financial derivatives, it is
relevant to introduce the notion of adapted processes. This is a key concept for
mathematical finance since adaptedness ensures that we are unable to predict the
future states of a process, meaning that at time t < T we are limited by the
information available at t.
Definition 2.5. A stochastic process (X(t))t∈T is adapted to a filtration Ft if for
every t ∈ T, X(t) is Ft -measurable.
Another key notion of mathematical finance is the concept of martingales, su-
permartingales and submartingales.
Definition 2.6. For s < t, a process (X(t))t∈T adapted to a filtration Ft can be a

supermartingale if E [X(t)|Fs ] ≤ X(s)

martingale if E [X(t)|Fs ] = X(s) ,

submartingale if E [X(t)|Fs ] ≥ X(s)


with respect to a filtration Fs .

 LOCAL VOLATILITY MODEL AND MULTI-DIMENSIONALITY 7

Definition 2.7. Let (Ω, F, P ) be a probability space and T = [0, ∞). A mapping
W : T × Ω → R measurable with respect to the σ-field B(T) × F = σ({A × B : A ∈
B(T), B ∈ F}] is a Wiener process if it has the below properties:
(1) W (0) = 0 almost surely
(2) for 0 ≤ s < t < ∞, the increments (W (t) − W (s)) ∼ N (0, t − s)
(3) for all m and all 0 ≤ t1 ≤ t2 ≤ ... ≤ tm the increments (W (tn+1 ) − W (tn )),
n = 0, 1, ..., m − 1, are independent
(4) for almost all ω ∈ Ω, the paths are continuous
Lemma 2.8. [1] The Wiener process is a martingale with respect to the filtration
FtW .
We now introduce two concepts that are fundamental for the understanding and
application of the tools of stochastic calculus: variation and covariation (also called
cross-variation).
Definition 2.9. Let T = [0, T ] and consider a sequence of partitions 0 = t0 < t1 <
... < tn = T of T, where max |ti+1 − ti | → 0 as n → ∞. For any process X(t),
t ∈ T, let
n−1
X
VTp (n) = |X(ti+1 ) − X(ti )|p .
j=0
p
If there exists a process V p (t) such that V[0,T p
] (n) → V (t) in probability as n → ∞
(and, consequently, max |ti+1 − ti | → 0), the limit is called the variation of X(t) in
T of order p. When p = 2 we call it the quadratic variation and use the notation
[X, X](t).
Definition 2.10. Let T = [0, T ] and consider a sequence of partitions 0 = t0 <
t1 < ... < tn = T of T, where max |ti+1 − ti | → 0 as n → ∞. For any two process
X(t) and Y (t), t ∈ T, let
n−1
X
[X, Y ](t) = lim [X(ti+1 ) − X(ti )] [Y (ti+1 ) − Y (ti )] ,
max |ti+1 −ti |→0
j=0

where the limit is in probability. If such limit exists, it is called the covariation.
Lemma 2.11. [1] W (t) has infinite variation and [W, W ](t) = t in [0, t].
At this stage we can begin to define the stochastic integral. This will be done in
steps. The object is first defined for a class of processes we call simple processes,
then it is extended to more general classes.
Definition 2.12. A process X : Ω × [0, T ] is a simple process (X ∈ S 2 ) if there is a
partition 0 = t0 < ... < tn = T and FtW
k
-measurable random variables fk ∈ L2 (Ω),
k = 0, 1, ..., n − 1 such that
n−1
X
X(t, ω) = f0 (ω)1{0} (t) + fk (ω)1(tk ,tk+1 ] (t).
k=0

Definition 2.13. The stochastic integral of X ∈ S 2 is a random variable defined

by
n−1
X
I(X) = fk (W (tk+1 ) − W (tk )) .
k=0
8 DANIEL MONTENEGRO DA FONSECA LIMA

For [a, b] ⊂ [0, T ] , where 1[a,b] X ∈ S 2 , the stochastic integral is defined as:
Z b
X(t)dW (t) = I(1[a,b] X).
a
It is worth noting that, although the integral of a function
R is a number,
 the
b
stochastic integral of a process is a random variable, so a X(t)dW (t) : Ω → R.
We will now extend the stochastic integral to a wider class of processes.
Definition 2.14. Let
M2 = X : [0, T ] × Ω → R : X is FtW -adapted, X ∈ L2 ([0, T ] × Ω) .


2
Theorem 2.15. [1] For every continuous
hR process X ∈ M
i there exists a sequence
T 2
of Xn ∈ S 2 such that limn→∞ E 0 (X(s) − Xn (s)) ds = 0.

Definition 2.16. The stochastic integral of X ∈ M2 is defined by the following

sequence of steps:
(1) Take a sequence of Xn ∈ S 2 such that Xn → X in L2 ([0, T ] × Ω)
(2) Create a sequence of random variables by finding the stochastic integral
I(Xn ) ∈ L2 (Ω) of each Xn
(3) Since L2 (Ω) is complete and the sequence I(Xn ) is Cauchy in L2 (Ω) (see
proof on [1]), then the sequence converges to a limit as n → ∞
For [a, b] ⊂ [0, T ], the stochastic integral of X ∈ M2 is a random variable defined
by
Z b
X(s)dW (s) = I(1[a,b] X) = lim I(1[a,b] Xn ).
a n→∞
hR i
T
Theorem 2.17. [1] For X ∈ S 2 and X ∈ M2 we have E 0
X(s)dW (s) = 0.
Rt
Theorem 2.18. [1] If X ∈ M2 , then 0
X(s)dW (s) is a martingale with respect
to FtW .
Theorem 2.19 (Conditional Ito Isometry). [1] If X ∈ M2 and [a, b] ⊂ [0, T ], then
 !2  "Z #
Z b b
W 2 W
E X(s)dW (s) |Fa  = E X (s)ds|Fa .
a a

Once again, we will extend the stochastic integral to a wider class of processes.
Definition 2.20. Let
( )
Z T
2
P = X : [0, T ] × Ω → R : X is FtW -adapted, 2
X (s)ds < ∞ a.s. .
0

Definition 2.21. Let t ∈ [0, T ] and τ : Ω → [0, ∞). If, for a given filtration Ft , τ
satisfies
{ω : τ (ω) ≤ t} ∈ Ft ,
then τ is a stopping time for Ft .
Definition 2.22. A sequence τn of stopping times is localising for X ∈ M2 if it
satisfies the following conditions:
(1) For each n ≥ 1, τn ≤ τn+1
 LOCAL VOLATILITY MODEL AND MULTI-DIMENSIONALITY 9

(2) For each n ≥ 1, 1{t≤τn } X ∈ M2

(3) P (∪∞
n=1 {ω : τn (ω) = T }) = 1

Definition 2.23. For [a, b] ⊂ [0, T ], the stochastic integral of X ∈ P 2 is defined

by
Z b Z T
X(s)dW (s) = lim 1[a,b] (s)1[0,τn ] (s)X(s)dW (s),
a n→∞ 0
where τn is a sequence of stopping times localising for X ∈ M2 .
Definition 2.24. Let τn be a sequence of stopping times. We call Xτn (t) =
X(min{t, τn }) the stopped process.
Definition 2.25. If Xτn (t) is a martingale, for a sequence of stopping times τn
satisfying P (∪∞
n=1 {ω : τn (ω) = T }) = 1, then X(t) is a local martingale.

Definition 2.26. A semimartingale is a stochastic process that can be decomposed

as the sum of a local martingale with a finite variation process.
There is one type of semimartingale that is widely used to model asset prices in
the context of mathematical finance. These are known as Ito processes.
Definition 2.27. Let t ∈ [0, T ]. An Ito process is a process of the form
Z t Z t
(2.1) X(t) = X(0) + a(s)ds + b(s)dW (s),
0 0

where b ∈ P 2 is the diffusion

n coefficient, × [0, T ] → R is called the drift
and a : Ωo
RT
coefficient and satisfies P ω : 0 |a(s, ω)|ds < ∞ = 1.

An Ito process can also be represented by the differential notation, under which
(2.1) can be represented by
dX(t) = a(t)dt + b(t)dW (t).
This notation has no meaning and is just an abbreviation of the integral form given
in Definition 2.27.
Theorem 2.28. [1] If X1 = X2 and:
dX1 (t) = a1 (t)dt + b1 (t)dW (t)
,
dX2 (t) = a2 (t)dt + b2 (t)dW (t)
then, for almost all t, a1 (t) = a2 (t) and b1 (t) = b2 (t) almost surely. This means
that the representation of an Ito process is unique.
Theorem 2.29. [1] Let X(t) be an Ito process of the form dX(t) = a(t)dt +
b(t)dW (t). The quadratic variation of X(t) is
Z t
[X, X](t) = b2 (s)ds.
0

For the remainder of this text, for a function F : [0, T ] × R → R, the following
notation will be used:
∂F
Ft = ,
∂t
∂F
Fx = ,
∂tx
10 DANIEL MONTENEGRO DA FONSECA LIMA

∂2F
Fxx = ,
∂x2
and C 1,2 will denote the collection of functions which which have continuous partial
derivatives Ft , Fx and Fxx .
Theorem 2.30 (Ito formula). [1] Let F (t, x) ∈ C 1,2 be a mapping F : [0, T ] × R →
R. Also, let X(t) be an Ito process dX(t) = a(t)dt+b(t)dW (t). Then, for t ∈ [0, T ],
the process F (t, X(t)) is an Ito process:
Z t Z t
F (t, X(t)) = F (0, X(0)) + Ft (u, X(u)) du + Fx (u, X(u)) a(u)du
0 0
Z t
1 t
Z
+ Fx (u, X(u)) b(u)dW (u) + Fxx (u, X(u)) b2 (u)du.
0 2 0
Theorem 2.31. [1] Let X(t) and Y (t) be two Ito processes with characteristics
aX (t), bX (t) and aY (t), bY (t) respectively. The product (XY ) (t) is also an Ito pro-
cess with stochastic differential
d (XY ) (t) = X(t)dY (t) + Y (t)dX(t) + bX (t)bY (t)dt.
Definition 2.32. A function a : [0, T ] × R → R is said to have linear growth if, for
x ∈ R and t ∈ [0, T ], there is a C > 0 such that
|a(t, x) ≤ C(1 + |x|).
Theorem 2.33. [1] Consider the equation
dX(t) = a(t, X(t))dt + b(t, X(t))dW (t),
along with the initial condition that X(0) is a known constant.
If a(t, X(t)) and b(t, X(t)) are uniformly continuous with respect to the first vari-
able, Lipschitz continuous with respect to the second variable and have linear growth,
then the equation has a unique solution X(t) ∈ L2 ([0, T ]×Ω) with continuous paths.
Since we will discuss the Local Volatility model, it is also necessary to introduce
the Fokker-Planck equation, which is a partial differential equation that describes
the temporal evolution of the probability density of a stochastic process.
Theorem 2.34 (Fokker-Planck Equation). [3] Let X(t) be an Ito process of the
form
dX(t) = a(X(t), t)dt + b(X(t), t)dW (t),
and let Z
P (X (t) ∈ A) = p (x, t) dx.
A
The Fokker-Planck equation applied to the process X(t) is
∂p(x, t) ∂ 1 ∂2
b2 (x, t)p(x, t)


(2.2) =− (a(x, t)p(x, t)) + 2
∂t ∂x 2 ∂x
Definition 2.35. Take k Wiener processes defined on k probability spaces (Ωj , Fj , Pj ),
j = 1, ..., k, and consider the probability space (Ω, F, P ) where
Ω = Ω1 × Ω2 × ... × Ωk ,
F = F1 × F2 × ... × Fk ,
and, for Aj ∈ Fj ,
P (A1 × A2 × .. × Ak ) = P1 (A1 )P2 (A2 )...Pk (Ak ).
 LOCAL VOLATILITY MODEL AND MULTI-DIMENSIONALITY 11

The k−dimensional Wiener process in the probability space (Ω, F, P ) is defined by

W (t) = (W1 (t), ..., Wk (t)).
Definition 2.36. Let t ∈ [0, T ], i = 1, ..., n, j = 1, ..., k, x ∈ Rn and take a
k−dimensional Wiener process W (t). Define n Ito processes of the form
k
X
dXi (t) = ai (t)dt + bij (t)dWj (t).
j=1

An n−dimensional Ito process is a process X : Ω × [0, T ] → Rn of the form

dX(t) = a(t)dt + b(t)dW(t),
where X = [Xi ] and a = [ai (t)] are n−dimensional vectors, b = [bij (t)] is an n × k
matrix.
Theorem 2.37 (Multi-dimensional Ito formula). [2] Take n independent Wiener
processes W1 (t), ..., Wn (t), take a k × n matrix B(t) = [bij (t)] with bij (t) ∈ P 2 ,
i=n1, 2, R..., k, j = 1, 2, ..., n,o
and take k drift coefficients a1 (t), ..., ak (t) satisfying
T
P ω : 0 |ai (s, ω)|ds < ∞ = 1. For t ∈ [0, T ], define k Ito processes:
n
X
dXi (t) = ai (t)dt + bij (t)dWj (t).
j=1

Take a mapping F : [0, T ] × Rk → R of class C 1 in the first variable and C 2

in the others. Then, Y (t) = F (t, X1 (t), ..., Xk (t)) is an Ito process with stochastic
differential
k
X
dY (t) = Ft (t, X1 (t), ..., Xk (t)) dt + Fxi (t, X1 (t), ..., Xk (t)) ai (t)dt
i=1
k
X n
X
+ Fxi (t, X1 (t), ..., Xk (t)) bij (t)dWj (t)
i=1 j=i
n k
1XX
+ Fxi xl (t, X1 (t), ..., Xk (t)) bij (t)blj (t)dt.
2 j=1
i,l=1

For the purposes of derivatives pricing, it is easier to work with objects that are
martingales. To this end, we introduce the Girsanov’s Theorem which will be used
to find a probability measure under which our objects of interest are martingales.
Theorem 2.38. [Girsanov’s Theorem][2] Let W (t) = (W1 (t), ..., Wn (t)) be an n-
dimensional Wiener process on (Ω, F, P ) and ai (t) be an FtW −adapted process.
Let
Z t
Xi (t) = ai (s)ds + Wi (t)
0
and assume that
( n Z n
)
t Z t
X 1X
M (t) = exp − ai (s)dWi (s) − a2i (s)ds
i=1 0 2 i=1 0
12 DANIEL MONTENEGRO DA FONSECA LIMA

is a martingale. Let Z
Q(A) = M (T )dP.
A
Then X is an n−dimensional Wiener process on the probability space (Ω, F, Q).
Theorem 2.39 (Multi-dimensional Fokker-Planck Equation). [3] Take m Wiener
processes. Let x ∈ Rn , X(t) be an n-dimensional Ito process of the form
m
X
dX(t) = a(X(t), t)dt + b(X(t), t)dW (t),
j=1

and Z
P (X(t) ∈ A) = p (x, t) dx.
A
The Fokker-Planck equation applied to the process X(t) is
n
∂p(x, t) X ∂
(2.3) =− (ai (x, t)p(x, t))
∂t i=1
∂x i
n
n X m
!
X ∂2 1X
+ bik (x, t)bjk (x, t)p(x, t) .
i=1 j=1
∂xi ∂xj 2
k=1

2.2. Black-Scholes model with dividends. This section explores the impact of
adding a continuously compounded dividend to the underlying stock. We provide
some of the proofs. For the proofs not provided, please refer to the relevant cited
source.
Let the market be comprised of 1 risk-free asset A(t) and 1 risky asset S(t). Let
r be a the constant risk-free interest rate, then risk-free asset is described by:
(2.4) dA(t) = rA(t)dt.
We can assume A(0) = 1 with no loss of generality. This gives
(2.5) A(t) = ert .
The underlying probability space is (Ω, F, P ), under which W (t) is a Wiener
process.
Let δ be the continuously compounded dividend rate and assume all dividends
are reinvested in the stock.
The risky asset is described by
(2.6) dS(t) = µS(t)dt + σS(t)dW (t).
Definition 2.40. A set (y(t), x(t)) of FtW −adapted processes where y(t) is the
number of risk-free asset A(t) and x(t) is the number of risky asset S(t) held at
time t is called a strategy and has value
V (t) = y(t)A(t) + x(t)eδt S(t).
Definition 2.41. A strategy is self-financing if it satisfies
dV (t) = y(t)dA(t) + x(t)eδt dS(t) + δx(t)eδt S(t)dt.
Let the discounted value process be
Ṽ (t) = e−rt V (t).
 LOCAL VOLATILITY MODEL AND MULTI-DIMENSIONALITY 13

Definition 2.42. Let

(µ + δ − r)
WQδ (t) = t + W (t).
σ
Definition 2.43. Let
(µ + δ − r)
aδ = ,
σ
and write   
1
2
Qδ (A) = E 1A exp − aδ T − aδ W (T ) .
2
Proposition 2.44. [2]WQδ (t) is a Wiener process in the probability space (Ω, F, Qδ ).
Proof. This follows from a direct application of Girsanov’s theorem 2.38. 

Proposition 2.45. [2]We claim the discounted value process Ṽ (t) of a self-financing
strategy is a Qδ −local martingale described by
dṼ (t) = σS(t)x(t)e(δ−r)t dWQδ (t).
Proof. We have
d e−rt V (t)


dṼ (t) =
= −re−rt V (t)dt + e−rt dV (t)
= x(t)eδt S(t)e−rt [(µ + δ − r) dt + σdW (t)]
= S(t)x(t)e(δ−r)t [(µ + δ − r) dt + σdW (t)]
= σS(t)x(t)e(δ−r)t dWQδ (t).


Proposition 2.46. [2]The dynamics of S(t) under (Ω, F, Qδ ) are given by

dS(t) = (r − δ)S(t)dt + σS(t)dWQδ (t).
Proof. We have

dS(t) = µS(t)dt + σS(t)dW (t)

 
µ+δ−r
= (r − δ)S(t)dt + σS(t) dt + dW (t)
σ
= (r − δ)S(t)dt + σS(t)dWQδ (t).


Definition 2.47. Let

S δ (t) = eδt S(t).
Proposition 2.48. [2]The dynamics of S δ (t) under (Ω, F, Qδ ) are given by
dS δ (t) = rS δ (t)dt + σS δ (t)dWQδ (t).
Proof. The proof follows directly from Ito’s formula
14 DANIEL MONTENEGRO DA FONSECA LIMA

dS δ (t) d eδt S(t)

=
= δeδt S(t)dt + eδt dS(t)
= δeδt S(t)dt + eδt µS(t)dt + eδt σS(t)dW (t)
= (δ + µ) S δ (t)dt + σS δ (t)dW (t)
 
(δ + µ − r)
= rS δ (t)dt + σS δ (t) dt + dW (t)
σ
= rS δ (t)dt + σS δ (t)dWQδ (t).

δ −rt δ δ
Proposition 2.49. [2]We claim S̃ (t) = e S (t) is a Q −local martingale and
is described by
dS̃ δ (t) = σS δ (t)dWQδ (t).
Proof. The proof follows directly from Ito’s formula

dS̃ δ (t) = d e−rt S δ (t)

= −re−rt S δ (t)dt + e−rt dS δ (t)

= −re−rt S δ (t)dt + re−rt S δ (t)dt + σe−rt S δ (t)dWQδ (t)
= σ S̃ δ (t)dWQδ (t).

Definition 2.50. A strategy is admissible if there is a non-negative constant L ∈ R
such that V (t) > −L for all t, almost surely with respect to P , and Ṽ (t) is a
Qδ −martingale.
Definition 2.51. A path-independent European derivative is a random variable
H(T ) measurable with respect to FTW and with price process H(t) = h(S(t)) for a
given mapping h : R → R.
If we assume that a derivative with price process H(t) is traded in the market,
then we have an extended market with this additional asset.
Definition 2.52. A set (z(t), y(t), x(t)) of FtW −adapted processes where z(t) is
the number of derivatives H(t), y(t) is the number of risk-free asset A(t) and x(t)
is the number of risky asset S(t) held at time t is a strategy in the extended market
and has value
V (t) = z(t)H(t) + y(t)A(t) + x(t)eδt S(t).
Definition 2.53. An admissible self-financing strategy is an arbitrage if
(1) V (0) = 0,
(2) V (T ) ≥ 0 and P (V (T ) > 0) > 0.
Definition 2.54. The initial price H(0) is a constant number that does not lead
to arbitrage with constant z. The derivative price process H(t) is an Ito process
dH(t) = aH (t)dt + bH (t)dW (t),
n R o
T
with aH (t) satisfying P ω : 0 |aH (s)|ds < ∞ = 1 and bH (t) ∈ P 2 , such that
it does not generate arbitrage opportunities.
 LOCAL VOLATILITY MODEL AND MULTI-DIMENSIONALITY 15

A strategy in the extended market is self-financing if it satisfies

dV (t) = z(t)dH(t) + y(t)dA(t) + x(t)eδt dS(t) + δx(t)eδt S(t)dt.
Definition 2.55. A self-financing admissible strategy replicates a derivative H if
H(T ) = V (T ).
Definition 2.56. A self-financing strategy is a martingale strategy if Ṽ (t) is a
martingale in (Ω, F, Qδ ).
Assumption 2.57 (No Arbitrage Principle). Arbitrage opportunities do not exist.
Theorem 2.58. [2] If a derivative H is replicated by an admissible martingale
strategy and its price process H(t) is an Ito process, then the No Arbitrage Principle
implies that for all t ∈ [0, T ], H(t) = V (t).
Theorem 2.59. [Derivatives Pricing Formula] If we have an admissible martingale
strategy with value process V (t) replicating a derivative H(t), then we can write
H(t) = EQδ e−rT H(T )|FtW ,
 
(2.7)
where the expectation is taken with respect to the Qδ -measure.
Proof. Since it is a martingale strategy we have
V (t) = EQ e−rT V (T )|FtW ,
 

and theorem 2.58 gives the result. 

Theorem 2.60 (Black-Scholes with dividends PDE). [2] Assume there is a function
u : R → R such that H(t) = u(S(t)). Then u satisfies the PDE below, along with
the terminal condition
(
ut + (r − δ)zuz + 12 σ 2 z 2 uzz = ru t < T
(2.8)
u(T ) = H(T ) t=T
2.3. Multi-dimensional Black-Scholes model. In this section we introduce the
multi-dimensional Black-Scholes model which can be used as a framework for pricing
basket options. Differently than the one-dimensional case, in the multi-dimensional
model it is usually impossible to find a closed-form solution to price derivatives.
This problem is typically solved by Monte Carlo or Finite Difference methods. The
goal of this section is to derive the partial differential equation satisfied by the value
process of a derivative.
We provide the proofs only for the results that are not proven in the references.
For the proofs we do not provide, please refer to the relevant cited source.
The market is assumed to be comprised of 1 risk-free asset A(t) and n risky
assets Si (t), i = 1, 2, ..., n. Let r(t) be a time-dependent deterministic risk-free
interest rate, then risk-free asset is described by:
(2.9) dA(t) = r(t)A(t)dt.
We can assume A(0) = 1 with no loss of generality. This gives
Rt
r(s)ds
(2.10) A(t) = e 0 .
Let T = [0, T ] and take an n−dimensional Wiener process W(t) defined on the
probability space (Ω, F, P ). This is the underlying probability space of the model.
16 DANIEL MONTENEGRO DA FONSECA LIMA

Proposition 2.61. [2] For i = 1, 2, ..., n, Wi (t) is a martingale with respect to the
filtration FtW1 ,W2 ,...,Wn .
n R o
T
Let µi (t)Si (t) satisfy P ω : 0 |µi (t)Si (t)|ds < ∞ = 1 and C(t) = [cij (t)]
be an n × n invertible matrix with cij (t)Si (t) ∈ P 2 . The risky assets are described
by an Ito process of the form
n
X
(2.11) dSi (t) = µi (t)Si (t)dt + cij (t)Si (t)dWj (t).
j=1

Lemma 2.62. [2] Equation (2.11) can be solved for


Z t n Z
1X t 2
(2.12) Si (t) = Si (0) exp µi (s)ds − cij (s)ds
 0 2 0 j=1

n Z
X t 
+ cij (s)dWj (s) .
j=1 0 

Proof. This can be verified by fixing i and using the Ito formula with
F (t, x) = ex ,
Si (t) = F (t, Xi (t)),
Si (0) = F (0),
and Z t n Z t n Z t
1X X
Xi (t) = µi (s)ds − c2ij (s)ds + cij (s)dWj (s),
0 2 j=1 0 j=1 0

which gives
n n  
X X 1 1
dSi (t) = Si (t)µi (t)dt + Si (t)cij (t)dWj (t) + Si (t) c2ij (t)dt −
j=1 j=1
2 2
Xn
= µi (t)Si (t)dt + cij (t)Si (t)dWj (t).
j=1


Under the imposed conditions, more specifically, the invertibility of C(t), there
is a one-to-one relation between Si (t) and (W1 (t), ..., Wn (t)). This implies that
FtS1 ,S2 ,...,Sn = FtW1 ,W2 ,...,Wn .
Definition 2.63. A set (y(t), x1 (t), x2 (t), ..., xn (t)) of FtW1 ,W2 ,...,Wn −adapted pro-
cesses where y(t) is the number of risk-free asset A(t) and xi (t) is the number of
risky asset Si (t) held at time t is called a strategy and has value
X n
V (t) = y(t)A(t) + xi (t)Si (t).
i=1

Definition 2.64. A strategy is self-financing if it satisfies

Xn
dV (t) = y(t)dA(t) + xi (t)dSi (t).
i=1
 LOCAL VOLATILITY MODEL AND MULTI-DIMENSIONALITY 17

Definition 2.65. A strategy is admissible if there is a non-negative constant L ∈ R

such that V (t) > −L for all t, almost surely with respect to P .
Theorem 2.66. [2] Let γi (t) be an n-dimensional process satisfying
Pd
(1) i (t) = µi (t) − r
j=1 cij (t)γn
Pn R t Pn R t o
(2) M (t) = exp − i=1 0 γi (s)dWi (s) − 21 i=1 0 γi2 (s)ds is a martingale
Rt
R − r(s)ds
Let Q(A) = A M (T )dP . Then the discounted price process S̃i (t) = e 0 Si (t)
is a Q−martingale.
The probability space (Ω, F, Q) is called the risk-neutral probability space and,
by the Girsanov theorem, the processes
Z t
Q
Wi (t) = γi (t)dt + Wi (t)
0
are the coordinates of an n−dimensional Wiener process in the risk-neutral proba-
bility space and, in this space, the risky assets dynamics are described by
n
X
(2.13) dSi (t) = r(t)Si (t)dt + cij (t)Si (t)dWjQ (t).
j=1

We can follow the same process described above to solve equation (2.11) to get
(2.14)  
Z t n Z n Z t
1X t 2 X 
Si (t) = Si (0) exp r(s)ds − cij (s)ds + cij (s)dWjQ (s) .
 0 2 0
j=1 0 j=1


Proposition 2.67. Let

Rt
− r(s)ds
S̃i (t) = e 0 Si (t).
This gives
n
X
dS̃i (t) = cij (t)S̃i (t)dWjQ (t),
j=1

which means that the discounted risky assets prices are local martingales in the
risk-neutral probability space.
Proof. From equation (2.14) we have
 
 1X n Z t Xn Z t 
Q
S̃i (t) = Si (0) exp − c2ij (s)ds + cij (s)dWj (s) .
 2
j=1 0 j=1 0


The Ito formula with

F (t, x) = ex ,
S̃i (t) = F (t, X(t)),
and
n n
1X 2 X
dX(t) = − c (t)dt + cij (t)dWjQ (t)
2 j=1 ij j=1

gives the result:

18 DANIEL MONTENEGRO DA FONSECA LIMA

n n n
1X 2 X Q 1X 2
dS̃i (t) = − c (t)S̃i (t)dt + cij (t)S̃i (t)dWj (t) + c (t)S̃i (t)dt
2 j=1 ij j=1
2 j=1 ij
n
X
= cij (t)S̃i (t)dWjQ (t).
j=1


Corollary 2.68. If cij (t)Si (t) ∈ M2 , then, by Theorem 2.18, the discounted risky
assets prices are martingales in the risk-neutral probability space.
Rt
− r(s)ds
Proposition 2.69. Let Ṽ (t) = e 0 V (t). A strategy is self-financing if and
only if
X n
(2.15) dṼ (t) = xi (t)dS̃i (t).
i=1

Proof. Assume that the strategy is self-financing. The Ito product rule gives
 Rt  Rt Rt
− r(s)ds − r(s)ds − r(s)ds
dṼ (t) = d e 0 V (t) = −r(t)e 0 V (t)dt + e 0 dV (t).

Using Definitions 2.63 and 2.64 we get

Rt n Rt n
− r(s)ds − r(s)ds
X X
dṼ (t) = −r(t)e 0 xi (t)Si (t)dt + e 0 xi (t)dSi (t)
i=1 i=1
n  Rt Rt 
− r(s)ds − r(s)ds
X
= xi (t) −r(t)e 0 Si (t)dt + e 0 dSi (t)
i=1
n  Rt 
− r(s)ds
X
= xi (t)d e 0 Si (t)
i=1
n
X
= xi (t)dS̃i (t).
i=1

Now we assume (2.15) and show that it implies self-financing.

 Rt  Rt Rt
r(s)ds r(s)ds r(s)ds
dV (t) = d e 0 Ṽ (t) = r(t)e 0 Ṽ (t)dt + e 0 dṼ (t)
Rt n
r(s)ds
X
= r(t)V (t)dt + e 0 xi (t)dS̃i (t)
i=1
n
! Rt n
r(s)ds
X X
= r(t) y(t)A(t) + xi (t)Si (t) dt + e 0 xi (t)dS̃i (t)
i=1 i=1
n
! n
X X
= r(t) y(t)A(t) + xi (t)Si (t) dt + xi (t) (−r(t)Si (t)dt + dSi (t))
i=1 i=1
n
X
= y(t)dA(t) + xi (t)dSi (t).
i=1
This concludes the proof. 
 LOCAL VOLATILITY MODEL AND MULTI-DIMENSIONALITY 19

Proposition 2.70. The discounted value process of self-financing strategies are

local martingales in the risk-neutral probability space.
Proof. Propositions 2.67 and 2.69 give the result. 
Corollary 2.71. If cij (t)Si (t) ∈ M2 , then, by Theorem 2.18, the discounted value
process of self-financing strategies are martingales in the risk-neutral probability
space.
Definition 2.72. A path-independent European derivative is a random variable
H(T ) measurable with respect to FTW1 ,W2 ,...,Wn and with price process H(t) =
h(S1 (t), ..., Sn (t)) for a given mapping h : Rn → R.
If we assume that a derivative with price process H(t) is traded in the market,
then we have an extended market with this additional asset.
Definition 2.73. A set (z(t), y(t), x1 (t), x2 (t), ..., xn (t)) of FtW1 ,W2 ,...,Wn −adapted
processes where z(t) is the number of derivatives H(t), y(t) is the number of risk-
free asset A(t) and xi (t) is the number of risky asset Si (t) held at time t is a strategy
in the extended market and has value
X n
V (t) = z(t)H(t) + y(t)A(t) + xi (t)Si (t).
i=1

Definition 2.74. An admissible self-financing strategy is an arbitrage if

(1) V (0) = 0,
(2) V (T ) ≥ 0 and P (V (T ) > 0) > 0.
Definition 2.75. The initial price H(0) is a constant number that does not lead
to arbitrage with constant z. The derivative price process H(t) is an Ito process
dH(t) = aH (t)dt + bH (t)dW (t),
n R o
T
with aH (t) satisfying P ω : 0 |aH (s)|ds < ∞ = 1 and bH (t) ∈ P 2 , such that
it does not generate arbitrage opportunities.
A strategy in the extended market is self-financing if it satisfies
Xn
dV (t) = z(t)dH(t) + y(t)dA(t) + xi (t)dSi (t).
i=1

Definition 2.76. A self-financing admissible strategy replicates a derivative H if

H(T ) = V (T ).
Definition 2.77. A self-financing strategy is a martingale strategy if Ṽ (t) is a
martingale in the risk-neutral probability space.
Assumption 2.78 (No Arbitrage Principle). Arbitrage opportunities do not exist.
Theorem 2.79. [2] If a derivative H is replicated by an admissible martingale
strategy and its price process H(t) is an Ito process, then the No Arbitrage Principle
implies that for all t ∈ [0, T ], H(t) = V (t).
Theorem 2.80. [Derivatives Pricing Formula] If we have an admissible martingale
strategy with value process V (t) replicating a derivative H(t), then we can write
 RT 
− r(s)ds W
(2.16) H(t) = EQ e t H(T )|Ft ,
20 DANIEL MONTENEGRO DA FONSECA LIMA

where the expectation is taken with respect to the Q-measure (risk-neutral probability
space).
Proof. Since it is a martingale strategy we have
 RT 
− r(s)ds W
V (t) = EQ e t V (T )|Ft ,

and theorem 2.79 gives the result. 

Theorem 2.81 (Multi-dimensional Black-Scholes PDE). [2] Assume there is a
function u : Rn → R such that H(t) = u(S1 (t), ..., Sn (t)). Then u satisfies the
multi-dimensional Black-Scholes PDE along with the terminal condition
( Pn Pn
ut + i=1 uzi r(t)zi + 21 j,i,l=1 zi zl uzi zl cij (t)clj (t) = r(t)u t < T
(2.17)
u(T ) = H(T ) t=T
 LOCAL VOLATILITY MODEL AND MULTI-DIMENSIONALITY 21

3. Dupire’s equation for a single asset

In [4], Dupire extended the Black-Scholes model by proposing a model where the
dynamics of the underlying asset under the risk-neutral measure is described by

(3.1) dS(t) = γ(t)S(t)dt + σ(S(t), t)dW (t),

which means the volatility σ is not constant, but a deterministic function of both
time t and the stock price S(t). This is the essence of the local volatility model and
Dupire has shown that, if the prices of European options on the underlying asset
for all strikes and maturities are known, the local volatility can be recovered from
market data.
The risk-free asset follows the same dynamics described in the multi-dimensional
Black-Scholes section, so R r(t) is a time-dependent deterministic risk-free interest
t
r(t)dt
rate and A(t) = A(0)e 0 , where we can assume A(0) = 1 with no loss of
generality.

Theorem 3.1. [Dupire’s Equation] [6] Let C(t, S(t), T, K) be the undiscounted
European call option price, r(t) the risk-free interest rate, D(t) the dividend rate,
K the strike price, T the maturity and σ(K, T ) the local volatility. The Dupire
equation for a single asset is
∂2C
 
∂C 1 ∂C
(3.2) = σ 2 (K, T )K 2 + (r(T ) − D(T )) C − K .
∂T 2 ∂K 2 ∂K
The remainder of this section constitutes a proof to Theorem 3.1. Some propo-
sitions and lemmas introduced in the proof do not have a reference as they are
introduced by us as building blocks and intermediary results for the proof.

Definition 3.2. Let the discount factor be

RT
− r(s)ds
B(t, T ) = e t .

Definition 3.3. Let the transition probability density from S(t) to S(T ) be
Z
P (S(T ) ∈ A|S(t) = s) p (s, t, x, T ) dx,
A

where, in order to simplify notation, we may use p instead of p (s, t, x, T ).

Assumption 3.4. We assume p (s, t, x, T ) → 0 as x → ∞.

∂


Assumption 3.5. We assume (x − K) ∂x σ 2 (x, T )x2 p (s, t, x, T ) → 0 as x → ∞.

Assumption 3.6. There is an admissible martingale strategy replicating the Eu-

ropean Call.

Proposition 3.7. [4, 5, 6]The undiscounted price C(t, S(t), T, K) of an European

Call is given by
Z ∞ Z ∞
(3.3) C(t, s, T, K) = p (s, t, x, T ) xdx − K p (s, t, x, T ) dx.
K K

Proof. With Assumption 3.6, equation (2.59) applies and we have

22 DANIEL MONTENEGRO DA FONSECA LIMA

H(t)
C(t, s, T, K) =
B(t, T )
EQ 1{S(T )>K} (S(T ) − K) |FtW
 
=
Z ∞
= 1{x>K} p (s, t, x, T ) (x − K) dx
−∞
Z ∞
= p (s, t, x, T ) (x − K) dx
ZK∞ Z ∞
= p (s, t, x, T ) xdx − K p (s, t, x, T ) dx.
K K

Remark 3.8. In order to simplify notation, we will use C instead of C(t, S(t), T, K)
for the remainder of the section.
Lemma 3.9. [4, 5, 6]We have
Z ∞
∂C
(3.4) = − pdx
∂K K
∂2C
(3.5) = p (s, t, K, T ) .
∂K 2
Proof. Direct differentiation of (3.3) with respect to K gives the result. 
Proposition 3.10. We claim
(3.6) Z 
∞
1 ∂2

∂C ∂ 2 2


= − (γ(T )xp (s, t, x, T )) + σ (x, T )x p (s, t, x, T ) (x − K) dx.
∂T K ∂x 2 ∂x2
Proof. Differentiating (3.3) with respect to T gives
Z ∞
∂C ∂p (s, t, x, T )
= (x − K) dx.
∂T K ∂T
∂p
We can find ∂T by applying the Fokker-Planck equation to p with t = T . This
gives
∂p (s, t, x, T ) ∂ 1 ∂2
σ 2 (x, T )x2 p (s, t, x, T ) .


=− (γ(T )xp (s, t, x, T )) + 2
∂T ∂x 2 ∂x
Substituting this result in the above gives the claim. 
Proposition 3.11. We claim
Z ∞  
∂ ∂C
(3.7) − (γ(T )xp (s, t, x, T )) (x − K) dx = γ(T ) C − K .
K ∂x ∂K
Proof. Let Z ∞
∂
A=− (γ(T )xp (s, t, x, T )) (x − K) dx.
K ∂x
We integrate by parts using
u(x) = (x − K) ,
∂
du = (x − K) dx = dx,
∂x
 LOCAL VOLATILITY MODEL AND MULTI-DIMENSIONALITY 23

v(x) = γ(T )xp (s, t, x, T ) ,

∂
dv = (γ(T )xp (s, t, x, T )) dx
∂x
and, under assumption 3.4, we get
Z ∞
∂
A = − (γ(T )xp (s, t, x, T )) (x − K) dx
K ∂x
Z ∞
∞
= − [(x − K) γ(T )xp (s, t, x, T )]K + γ(T )xpdx
K
Z ∞
= γ(T ) xp (s, t, x, T ) dx.
K

By using (3.3) and (3.4) we have

Z ∞ Z ∞
∂C
C −K = p (s, t, x, T ) xdx + (K − K) p (s, t, x, T ) dx
∂K
ZK∞ K

= p (s, t, x, T ) xdx
K
A
= .
γ(T )

Multiplying by γ(T ) gives the claim. 

Proposition 3.12. We claim

∞
∂2 ∂2C
Z
1 1
σ 2 (x, T )x2 p (s, t, x, T ) (x − K) dx = σ 2 (K, T )K 2


(3.8) 2
.
2 K ∂x 2 ∂K 2

Proof. Let
∞
∂2
Z
1
σ 2 (x, T )x2 p (s, t, x, T ) (x − K) dx.


B= 2
2 K ∂x
We integrate by parts using

u(x) = (x − K) ,

∂
du = (x − K) dx = dx,
∂x

∂
σ 2 (x, T )x2 p (s, t, x, T ) ,


v(x) =
∂x

∂2
σ 2 (x, T )x2 p (s, t, x, T ) dx.


dv = 2
∂x
24 DANIEL MONTENEGRO DA FONSECA LIMA

Under assumption 3.5 we get

1 ∞ ∂2
Z
σ 2 (x, T )x2 p (s, t, x, T ) (x − K) dx


B = 2
2 K ∂x
1 ∞ ∂
Z
σ 2 (x, T )x2 p (s, t, x, T ) dx


= −
2 K ∂x
1 2 ∞
= − σ (x, T )x2 p (s, t, x, T ) K ,
2
1 2
= σ (K, T )K 2 p (s, t, K, T ) .
2
1 2 ∂2C
= σ (K, T )K 2 .
2 ∂K 2

By Propositions 3.11 and 3.12, equation (3.6) becomes

∂2C
 
∂C ∂C 1
= γ(T ) C − K + σ 2 (K, T )K 2 .
∂T ∂K 2 ∂K 2
Under the risk-neutral measure, the drift is γ(t) = r(t) − D(t). This substitution
which gives the Dupire equation (3.2) for a single asset:
∂2C
 
∂C ∂C 1
= (r(T ) − D(T )) C − K + σ 2 (K, T )K 2 .
∂T ∂K 2 ∂K 2
This concludes the proof of Theorem 3.1.
 LOCAL VOLATILITY MODEL AND MULTI-DIMENSIONALITY 25

4. Dupire’s equation for several assets

In [6], the authors attempted to generalise Dupire’s equation for several assets.
The main result of the paper is a theorem, denoted by Theorem 1 in the original
text, which is equivalent to the multi-asset Dupire’s equation for the case where the
volatility is time dependent but not asset-price dependent. The goal of this section
is to introduce the theorem, provide a step-by-step proof and a numerical example.
We begin by providing the setting under which the theorem is applicable.
Let i = 1, 2, ..., N , where N is the number of assets. The dynamics of the assets
under the risk-neutral measure is given by

N
X
(4.1) dSi (t) = γi (t)Si (t)dt + Si (t) σij (t)dWj (t),
j=1

where W (t) = (W1 (t), .., WN (t)) is an N −dimensional Wiener process under the
risk-neutral measure, σij (t) is the volatility matrix, γi (t) = r(t) − Di (t), Di (t) is
the continuously compounded dividend rate of Si (t).

Assumption 4.1. The volatility matrix σij (t) and the drifts γi (t) are smooth and
bounded. For any C ≥ 0, C ∈ R,

|γi (t)| ≤ C and |σij (t) ≤ C.

Assumption 4.2. The matrix A = (νij (T )) = 21 σσ t satisfies the uniform el-

lipticity condition, which means there exist constants λ, Λ > 0 such that, for all
y ∈ RN ,
N
X
λ|y|2 ≤ νij (t)yi yj ≤ Λ|y|2 .
i,j=1

Definition 4.3. A stock basket price is the weighted average of the prices of the
stocks that compose the basket. Let B(t) be the value of the basket at time t and
w = (w1 , ..., wN ) a vector of weights, where wi ≥ 0 is the weight of the stock i in
PN
the basket and i=1 wi = 1, then

N
X
(4.2) B(t) = wi Si (t).
i=1

Definition 4.4. A stock basket option is an option which has a stock basket as
the underlying asset.

Lemma 4.5. Consider a stock basket with N stocks and assume that there is an
admissible martingale strategy replicating an European call on the basket. Then,
undiscounted fair price Cw (t) of the European basket call is
" N
! #
X
(4.3) Cw (t) = EQ wi Si (T ) − K 1{PN wi Si (T )>K}
|S(t) = s .
i=1
i=1
26 DANIEL MONTENEGRO DA FONSECA LIMA

Proof. Since there is there is an admissible martingale strategy replicating the op-
tion, equation (2.16) is applicable and we have
RT
r(s)ds
Cw (t) = e t H(t)
RT " R N
! #
T
r(s)ds − r(s)ds
X
= e t EQ e t wi Si (T ) − K 1{B(T )>K} |S(t) = s
i=1
" N
! #
X
= EQ wi Si (T ) − K 1{B(T )>K} |S(t) = s .
i=1

We are now ready to introduce the main result of [6], refered to in the paper as
Theorem 1.
Theorem 4.6. [6] Assume that the volatility matrix σij is a deterministic locally
integrable function of time. Then the undiscounted fair price Cw of the European
basket call option satisfies
N N N
∂Cw X ∂Cw X X ∂Cw
(4.4) = γi (T )wi + νij (T )wi wj ,
∂T i=1
∂wi i=1 j=1
∂wi ∂wj

where A = (νij (T )) denotes the matrix given by A = 21 σσ t .

4.1. Proof. This subsection consists in a proof of Theorem 4.6. Some propositions
and lemmas introduced do not have a reference as they are introduced by us as
building blocks and intermediary results for the proof.
Let the vector of underlying asset prices be
S(t) = (S1 (t), ..., SN (t)).
Definition 4.7. The transition probability density of S(t) to S(T ) is
Z
P (S(T ) ∈ A|S(t) = s) p (s, t, x, T ) dx,
A
where, in order to simplify notation, we may use p instead of p (s, t, x, T ).
From equation (4.3) we have
" N ! #
X
Cw (t) = EQ wi Si (T ) − K 1{PN wi Si (T )>K}
|S(t) = s
i=1
i=1
Z N
!
X
(4.5) = p (s, t, x, T ) 1{PN wi xi >K}
w i xi − K dx.
RN i=1
i=1

Definition 4.8. Define the region under which the call will be exercised at expi-
ration as ( )
XN
N
Hw = x ∈ R | wi xi ≥ K .
i=1

We can rewrite (4.5) as

Z N
!
X
(4.6) Cw (t) = p wi xi − K dx.
Hw i=1
 LOCAL VOLATILITY MODEL AND MULTI-DIMENSIONALITY 27

We now define a new variable that will be used throughout the entire proof.
Definition 4.9. For i = 1, ..., N − 1, define
wi xi
Qi = .
B(T )
Definition 4.10. Let Q = (Q1 , ..., QN −1 ) and define
−1
( N
)
X
∆N = Q ∈ RN −1 : Qi ≥ 0, Qi ≤ 1 .
i=1

We apply the change of variables x 7→ (Q, B(T )) on (4.6). With the new variables
we can rewrite xi , i = 1, ..., N − 1 and xN :
Qi
(4.7) xi = B(T ),
wi
PN −1
(1 − i=1 Qi )
(4.8) xN = B(T ).
wN
Proposition 4.11. The determinant of the Jacobian matrix J of the change of
variables x 7→ (Q, B(T )) is
B(T )N −1
det J = .
w1 w2 ...wN
Proof. We have
∂x1 ∂x1 ∂x1
 
∂Q1 ... ∂QN −1 ∂B(T )
∂x2 ∂x2 ∂x2

∂Q1 ... ∂QN −1 ∂B(T )

J= .
 
 ... ... ... ... 
∂xN ∂xN ∂xN
∂Q1 ... ∂QN −1 ∂B(T )
The partial derivatives are
(
B(T )
∂xi wi i=j
= , for i, j = 1, ..., N − 1,
∂Qj 0 i 6= j

∂xN B(T )
=− ,
∂Qi wN
∂xi Qi
= , for i = 1, ..., N − 1,
∂B(T ) wi
and
PN −1 PN −1 wi xi
∂xN 1− i=1 Qi + B(T ) i=1 B 2 (T )
=
∂B(T ) wN
PN −1 PN −1 w i xi
1− i=1 Qi + i=1 B(T )
=
wN
PN −1 PN −1
1− i=1 Qi + i=1 Qi
=
wN
1
= .
wN
28 DANIEL MONTENEGRO DA FONSECA LIMA

With that, we have

N −1
∂xN Y ∂xi
det J =
∂B(T ) i=1 ∂Qi
1 B N −1
=
wN w1 w2 ...wN −1
B(T )N −1
= .
w1 w2 ...wN


We now introduce an important object that will be key throughout the proof:
the transition probability density in Q.

Definition 4.12. Write

Z Z
P (Q ∈ A, B(T ) ∈ B|Q = u, B(t) = c) pB (u, c, t, a, b, T ) dadb,
B A

where, in order to simplify notation, we may use pB instead of pB (u, c, t, a, b, T ).

With the change of variables x 7→ (Q, B(T )), (4.6) becomes

Z ∞Z
bN −1
(4.9) Cw (t) = pB (b − K) dadb.
K ∆N w1 w2 ...wN

From (4.9) we compute ∂C ∂wi . Since Q and B(T ) are the variables of integration
w

in (4.9), we recall Definition 4.9 using the same notation in (4.9). This gives
wi xi
ai =
b
and
∂xi ai b
=− 2.
∂wi wi
The remainder of the proof consist in a series of propositions that will start from
(4.9) and proceed to simplify the equation.

Proposition 4.13. We claim

N Z N N
!
X ∂Cw X ∂ X
(4.10) γi (T )wi =− (γi (T )xi p) w i xi − K dx.
i=1
∂wi Hw i=1 ∂xi i=1

Proof. From (4.9) we have

∞
bN −1
Z Z
∂Cw ∂pB ∂xi
= (b − K) dadb
∂wi K ∆N ∂xi ∂wi w1 w2 ...wN
Z ∞Z
1 bN −1
− pB (b − K) dadb,
w i K ∆N w1 w2 ...wN
which simplifies to
 LOCAL VOLATILITY MODEL AND MULTI-DIMENSIONALITY 29

∞
ai b bN −1
Z Z
∂Cw ∂pB
=− (b − K) 2 dadb
∂wi K ∆N ∂xi wi w1 w2 ...wN
Z ∞Z
1 bN −1
− pB (b − K) dadb.
w i K ∆N w1 w2 ...wN

Multiplying by wi gives us
∞
ai b bN −1
Z Z
∂Cw ∂pB
(4.11) wi =− (b − K) dadb
∂wi K ∆N ∂xi wi w1 w2 ...wN
Z ∞Z
bN −1
− pB (b − K) dadb.
K ∆N w1 w2 ...wN

Equation (4.11) holds for i = 1, ..., N − 1 but it can be extended to i = N by

PN −1
using aN = 1 − i=1 ai .
Reverting to the original variables with the change of variables (Q, B(T )) 7→ x
and get
Z   X N
!
∂Cw ∂p
(4.12) wi =− xi +p wi xi − K dx.
∂wi Hw ∂xi i=1

Multiplying by γi (T ) gives us
Z   X N
!
∂Cw ∂p
γi (T )wi = − γi (T )xi + γi (T )p wi xi − K dx
∂wi Hw ∂xi i=1
Z N
!
∂ X
= − (γi (T )xi p) wi xi − K dx.
Hw ∂xi i=1

Summing across i = 1, ..., N gives the claim. 

The next step is to use the multi-dimensional Fokker-Planck equation to simplify

(4.10).

Proposition 4.14. We claim

N
X ∂Cw ∂Cw
(4.13) γi (T )wi =
i=1
∂wi ∂T
N X
N N
!
∂2
Z X X
− (νij (T )xi xj p) wi xi − K dx.
Hw i=1 j=1 ∂xi ∂xj i=1

Proof. From the multi-dimensional Fokker-Planck equation applied to p, with t =

T , we get
N N N
X ∂ ∂p X X ∂ 2
− (γi (T )xi p) = − (νij (T )xi xj p) .
i=1
∂xi ∂T i=1 j=1
∂xi ∂xj
30 DANIEL MONTENEGRO DA FONSECA LIMA

Using this result in (4.10) gives

N Z N
!
X ∂Cw ∂p X
(4.14) γi (T )wi = wi xi − K dx
i=1
∂wi Hw ∂T i=1
N X
N N
!
∂2
Z X X
− (νij (T )xi xj p) wi xi − K dx.
Hw i=1 j=1 ∂xi ∂xj i=1

From (4.6) we compute

Z N
!
∂Cw ∂p X
= wi xi − K dx.
∂T Hw ∂T i=1
Using this result in (4.14) gives the claim. 
We can simplify the resulting expression from the previous Proposition, equation
(4.13), by focusing on the integral on the RHS.
Proposition 4.15. We claim that
N X
N N
!
∂2
Z X X
(νij (T )xi xj p) w i xi − K dx =
Hw i=1 j=1 ∂xi ∂xj i=1
Z N X N
X ∂
− (νij (T )xi xj p) wi dx.
Hw i=1 j=1 ∂xj

Proof. Let !
N X
N N
X ∂2 X
A= (νij (T )xi xj p) wi xi − K ,
i=1 j=1
∂xi ∂xj i=1
and consider the integral
N X
N N
!
∂2
Z Z X X
Adx = (νij (T )xi xj p) wi xi − K dx.
Hw Hw i=1 j=1 ∂xi ∂xj i=1

We integrate by parts using

∂
v(xi ) = (νij (T )xi xj p) ,
∂xj
∂2
dv = (νij (T )xi xj p) dxi ,
∂xi ∂xj
N
!
X
u(xi ) = wi xi − K ,
i=1

du = wi dxi .
and get
Z Z N X N
X ∂
Adx = − (νij (T )xi xj p) wi dx,
Hw Hw i=1 j=1 ∂x j

where assumptions 4.1 and 4.2 ensure the boundary terms vanish since p → 0 as
||x|| → ∞. 
 LOCAL VOLATILITY MODEL AND MULTI-DIMENSIONALITY 31

With Proposition 4.15, equation (4.13) becomes

N Z N X N
∂Cw X ∂Cw X ∂
(4.15) = γi (T )wi − (νij (T )xi xj p) wi dx.
∂T i=1
∂wi Hw i=1 j=1 ∂x j

The final step of the proof is to simplify the integral in (4.15) by finding a
suitable expression for the integral in the RHS. This is achieved by the next two
propositions. The first proposition is an intermediary result that will be required
in the second proposition.
Proposition 4.16. We claim
Z ∞Z
∂Cw ai b bN −1
(4.16) = pB dadb.
∂wi K ∆N wi w1 w2 ...wN
Proof. From (4.12) we have
Z N
!
∂Cw 1 ∂ X
(4.17) =− (xi p) wi xi − K dx.
∂wi wi Hw ∂xi i=1

Let !
Z Z N
∂ X
Adx = (xi p) w i xi − K dx.
Hw Hw ∂xi i=1
We integrate by parts using
v(xi ) = xi p,
∂
dv = (xi p) dxi ,
∂xi
N
!
X
u(xi ) = wi xi − K ,
i=1

du = wi dxi
and get
Z Z
Adx = − xi pwi dx,
Hw Hw

where assumptions 4.1 and 4.2 ensure the boundary terms vanish since p → 0 as
||x|| → ∞.
Using this result in (4.17) gives
Z
∂Cw
(4.18) = xi pdx.
∂wi Hw

Applying the change of variables x 7→ (Q, B(T )) gives the claim. 

Proposition 4.17. We claim

∂ 2 Cw
Z
∂
(4.19) νij (T )wi wj =− (νij (T )xi xj p) wi dx.
∂wi ∂wj Hw ∂xj
32 DANIEL MONTENEGRO DA FONSECA LIMA

Proof. From (4.16) we compute

Z ∞Z
∂ 2 Cw ∂pB ∂xj ai b bN −1
= dadb
∂wi ∂wj K ∆N ∂xj ∂wj wi w1 w2 ...wN
Z ∞Z
ai b 1 bN −1
− pB dadb,
K ∆N wi wj w1 w2 ...wN
which simplifies to
Z ∞Z
∂ 2 Cw ∂pB ai b aj b bN −1
=− 2 dadb
∂wi ∂wj K ∆N ∂xj wi wj w1 w2 ...wN
Z ∞Z
ai b 1 bN −1
− pB dadb.
K ∆N wi wj w1 w2 ...wN
Multiplying by wj gives
Z ∞Z
∂ 2 Cw ∂pB ai b aj b bN −1
wj =− dadb
∂wi ∂wj K ∆N ∂xj wi wj w1 w2 ...wN
Z ∞Z
ai b bN −1
− pB dadb.
K ∆N wi w1 w2 ...wN
Applying the change of variables (Q, B(T )) 7→ x and multiplying by νij (T ) gives
∂ 2 Cw
Z Z
∂p
νij (T )wj = − νij (T )xi xj dx − pνij (T )xi dx
∂wi ∂wj Hw ∂xj Hw
Z
∂
= − (νij (T )xi xj p) dx.
Hw ∂xj

Finally, multiplying by wi gives the claim. 

Using (4.19) on (4.15) we get
N N N
∂Cw X ∂Cw X X ∂ 2 Cw
= γi (T )wi + νij (T )wi wj .
∂T i=1
∂wi i=1 j=1
∂wi ∂wj

This concludes the proof of Theorem 4.6.

4.2. Numerical example. In this subsection we give a numerical example of how

Theorem 4.6 can be used to recover A = 12 σσ t = (νij (T )) from market prices of
options. We restrict the example to the simplest setting, where N = 2 and constant
γ(t) = γ.
We have
w1 ∈ [0, 1],

w2 = 1 − w1
and " #
σ1 (T ) q0
σ= ,
σ2 (T )ρ1,2 σ2 (T ) 1 − ρ21,2
where ρ1,2 is the correlation between the continuously compounded returns of assets
1 and 2.
 LOCAL VOLATILITY MODEL AND MULTI-DIMENSIONALITY 33

Let C = σσ t be the covariance matrix, then

σ12 (T )
 
σ1 (T )σ2 (T )ρ1,2
C=
σ1 (T )σ2 (T )ρ1,2 σ22 (T )
and A = 21 C, which implies ν12 (T ) = ν21 (T ).
Assume we have the prices for European basket call options quoted in the market:
Cw T = 0.15 T = 0.20
w = (0.30, 0.65) C1,1 = 1.50 C1,2 = 2.20
w = (0.35, 0.65) C2,1 = 2.50 C2,2 = 3.20
w = (0.40, 0.60) C3,1 = 3.00 C3,2 = 3.50
w = (0.45, 0.55) C4,1 = 3.20 C4,2 = 3.00
w = (0.50, 0.50) C5,1 = 3.10 C5,2 = 2.90
From 4.6 we can write
2 X
2 2
X ∂ 2 Cw ∂Cw X ∂Cw
νij (T )wi wj = − wi γi (T ) .
i=1 j=1
∂wi ∂wj ∂T i=1
∂wi
We use difference approximations to find the partial derivatives
∂Cw Ci,j+1 − Ci,j
= ,
∂T 0.05
∂Cw ∂Cw Ci+1,j − Ci,j
=− = ,
∂w1 ∂w2 0.05
∂ 2 Cw ∂ 2 Cw Ci−1,j − 2Ci,j + Ci+1,j
2 =− = ,
∂w1 ∂w22 0.0025
∂ 2 Cw ∂ 2 Cw −Ci−1,j + 2Ci,j − Ci+1,j
=− = .
∂w1 ∂w2 ∂w2 ∂w1 0.0025
2 2
∂ Cw ∂ Cw
Since ν12 (T ) = ν21 (T ) and ∂w 1 ∂w2
= − ∂w 2 ∂w1
, by applying 4.6 to C2,1 and C3,1
and simplifying we get the following system of equations:
 2 2
 ν11 (T )w2 ∂ C2,1 2 ∂ C2,1 ∂C2,1 ∂C2,1
1 ∂w2 + ν22 (T )w2 ∂w2 = ∂T + (w2 − w1 )γ ∂w1
2
1
2
2 .
 ν11 (T )w12 ∂ C3,1
2 + ν22 (T )w22
∂ C3,1
2
∂C
= 3,1 + (w2 − w1 )γ 3,1
∂C
∂w1 ∂w2 ∂T ∂w1

Let γ = 5%. Inserting the value of the partial derivatives and parameters in the
system of equations we get:

ν11 (T )(0.1225)(−120) + ν22 (T )(0.4225)(120) = (14) + (0.65 − 0.35)(0.05)(10)
,
ν11 (T )(0.16)(−120) + ν22 (T )(0.36)(120) = (10) + (0.60 − 0.40)(0.05)(4)
which simplifies to

−14.70ν11 (T ) + 50.70ν22 (T ) = 14.15
.
−19.20ν11 (T ) + 43.20ν22 (T ) = 10.04
Since both options have the same expiration T = 0.15, the system can be solved
to find
ν11 (0.15) ≈ 30.21%
.
ν22 (0.15) ≈ 36.67%
34 DANIEL MONTENEGRO DA FONSECA LIMA

This gives
σ1 (0.15) ≈ 60.42%
.
σ2 (0.15) ≈ 73.34%
We only required 2 options with the same expiration T = 0.15 (C2,1 and C3,1 )
because, in this specific example, w2 = 1 − w1 which resulted in partial derivatives
cancellation, removed the impact of correlation, and yielded a system of equations
with 2 unkowns. For higher dimensions, one would need n2 options with the same
expiration to recover the volatilities and correlations.
 LOCAL VOLATILITY MODEL AND MULTI-DIMENSIONALITY 35

5. Conclusion
This text has taken the reader on a journey from stochastic processes and the
basics of derivatives pricing frameworks to more sophisticated models that are cur-
rently used in the industry.
The two main problems addressed throughout this project are multi-dimensionality
and the constant volatility assumption. This was done firstly by addressing each
problem separately going through the multi-dimensional Black & Scholes model
and the Local Volatility Model, and finally, by detailing recent developments in the
field that took an important step in dealing with both problems combined.
As suggested by the authors in [6], possible further developments would include
relaxing the assumption that the volatility is time-dependent only and include also
a dependency on the underlying asset price.
36 DANIEL MONTENEGRO DA FONSECA LIMA

References
[1] Capiński, M., Kopp, E., & Traple, J. (2012). Stochastic Calculus for Finance (Mas-
tering Mathematical Finance). Cambridge: Cambridge University Press.
[2] Capiński, M., & Kopp, E. (2012). The Black–Scholes Model (Mastering Mathematical
Finance). Cambridge: Cambridge University Press.
[3] Risken, H., Haken, H. (1989). The Fokker-Planck Equation: Methods of Solution
and Applications Second Edition. Springer.
[4] B. Dupire. (1994) “Pricing with a smile”, Risk, 7, 18–20.
[5] Derman, Emanuel & Kani, Iraj. (1994). “Riding on a Smile”. Risk, 7.
[6] P. Amster, P. De Nápoli, J.P. Zubelli (2009). “Towards a generalization of Dupire’s
equation for several assets”. Journal of Mathematical Analysis and Applications 355
170-179