You are on page 1of 59

Stochastic Volatility I:

Heston Model: Analytics


(Lecture script version April 17, 2010)

Master of Advanced Studies in Finance

ETH Zurich / University of Zurich

Spring Term 2010

Paolo Vanini
Miret Padovani

Zürcher Kantonalbank and Swiss Finance Institute


E-mail address: paolo.vanini@zkb.ch

University of Zurich

Electronic copy available at: http://ssrn.com/abstract=1579593


We consider in these lecture notes of the Master of Advanced Studies in Fi-
nance at ETH Zurich and the University of Zurich analytical topics of the Heston
model. Mathematical methods such as Fourier theory, complex analysis, gener-
alized functions and partial differential equations are discussed. To make these
theories accessible to students with a standard economic or finance background we
stress intuition and explicit calculations.

Keywords: Stochastic Volatility, Heston Model, Fourier Theory, Complex Anal-


ysis, Pricing Equations, Generalized Functions.

These are rough Lecture Notes which we use in the MAS in Finance at the
University of Zurich and ETH Zurich. Feedback is welcome.

Electronic copy available at: http://ssrn.com/abstract=1579593


Contents

Chapter 1. The Heston model 1


1.1. References 1
1.2. Model dynamics 1
1.3. Pricing equation 2
1.4. Construction of Heston’s Fundamental Solution 11
1.5. Singularities 13
1.6. Synthesis of Option Pricing 18
Chapter 2. Mathematical tools 27
2.1. The Fourier transform 27
2.2. Complex analysis 35
2.3. Generalized functions 49
Index 55

iii
CHAPTER 1

The Heston model

1.1. References
• P. Carr and D. Madan. Option valuation using the fast Fourier transform.
Journal of Computational Finance, 2:61-73, 1999. (available on-line)
• E. Derman. Laughter in the Dark: An Introduction to the Volatility
Smile. Lecture Notes, 2008. ()
• J. Gatheral. The Volatility Surface: A Practitioner’s Guide. John Wiley
& Sons, 2006. [Amazon]
• S. Heston. A closed-form solution for options with stochastic volatil-
ity with application to bond and currency options. Review of Financial
Studies, 6:327-343, 1993. (abstract)
• C. Kahl and P. Jäckel. Not-so-complex logarithms in the Heston model.
Wilmott Magazine, 2005. (available on-line)
• R. Lee. Option pricing by transform methods: Extensions, unification,
and error control. Journal of Computational Finance, 7, 2004. (available on-line)
• A. Lewis. Option Valuation Under Stochastic Volatility. Finance Press,
2000. [Amazon]
• R. Lord and C. Kahl. Why the rotation count algorithm works. Tinbergen
Institute Discussion Paper, 2006. (available on-line)
• R. Lord and C. Kahl. Optimal Fourier inversion in semi-analytical option
pricing. Journal of Computational Finance, 10:1-30, 2007. (available on-line)
• V. Lucic, On Singularities in the Heston Model, see ()
• R. Lord, R. Koekkoek, and D. van Dijk. A comparison of biased simula-
tion schemes for stochastic volatility models. Tinbergen Institute Discus-
sion Paper, 2008. (available at SSRN)

1.2. Model dynamics


The dynamics in the Heston model are:
p
dSt = µt St dt + Vt St dZ1
and
p
dVt = −λ(Vt − V̄ ) dt + η Vt dZ2 ,
so that the variance V follows a square-root process. The parameter η is the
vol of vol parameter. The dynamics of V is mean-reverting. The choice of the
variance process is due to two facts. First, variance cannot become negative and
second, this CIR-type diffusion leads to quasi-analytic option prices, i.e. option
1
2 1. THE HESTON MODEL

price formula which only need a singe simulation. The volatility process has the
following properties:
(i) 0 is an attainable boundary when η 2 > 2λV̄ and the boundary is strongly
reflecting. Hence, volatility cannot become negative.
(ii) ∞ is an unattainable boundary. Hence, volatility cannot explode.
There are two standard Brownian motions Z1 , Z2 . If they are assumed to be uncor-
related, then there is no skew. Therefore, in equity markets for example one assumes
that ρ is a typically time-dependent correlation between the two Brownian motions.

1.3. Pricing equation


Given the model, the goal is to derive the pricing equation for options. Let
F (S, V, t) be the price of an option. Then we apply the usual replicating portfolio
approach to derive the pricing equation. Since there are two state variables - S and
V - the replicating portfolio π is build of three independent assets: Two options
F 0 , F 1 and the underlying process S, i.e.
π = F 1 − a1 F 0 − a2 S
with the weights a1 , a2 . Proceeding in the same way as in the derivation of the
Black and Scholes pricing equation means that:
• Write the stochastic differential dπ.
• Apply to each option dF j , j = 0, 1, Itô’s Formula and insert then Heston’s
dynamics.
• Use the no-arbitrage condition dπ = rπ dt. This determines the weights
FV1 0 FV1
a0 = FS1 − F , a1 =
FV0 S FV0
where lower indices denote (partial) derivatives and leads to the equa-
tion:

−rF 0 + Ft0 + AHe F 0 −rF 1 + Ft1 + AHe F 1


(1.1) =
FV0 FV1
with
∂ 1 ∂2 1 ∂2 ∂2 ∂
AHe = rS + η2 V 2
+ V S 2 2 + SηρV + λ(Vt − V̄ )
∂S 2 ∂V 2 ∂S ∂S∂V ∂V
(1.2)
the generator of Heston’s model.
Since the RHS (LHS) of the equation (1.1) is a function only of F 0 ( F 1 ), each
side of the equation has to be equal to a common constant1 function Φ(S, V, t) -
the so-called volatility risk premium. It is then is sufficient to consider only one
equation where we denote the option price simply by F . Hence, we have
(1.3) FV Φ(S, V, t) = −rF + Ft + AHe F .
This is the pricing equation for a general claim F where the volatility risk premium
is yet to be specified. We recall that the expression −rF represents the funding
costs of a market maker which issues an option F , that Ft reflects the time decay
1i.e. independent on the option price functions F .
1.3. PRICING EQUATION 3

(Theta) of the option and that AHe F represents the changes or sensitivities of the
option due to the dynamics of the underlying and the volatility process. Contrary
to the Black and Scholes model the volatility risk premium or market price of risk
is not specified within the model. This reflects the market incompleteness, i.e. we
have two risk sources but only one tradeable asset S. Therefore, the method is in
this respect similar to the short rate interest rate models where the short rate is
also not tradeable.

We specify the further analysis to the case of a call option, i.e.


F (S, T ) = max(S(T ) − K, 0).
To solve (1.3), we have to specify Φ(S, V, t). Heston assznes that the market price
of risk is linear in the instantaneous variance, i.e.
Φ = θV,
which is an economically sound assumption. We recall that volatility is not trade-
able like interest rates. Therefore options or specific volatility like products such
as variance swaps are needed to span the space, i.e. products which are volatility
sensitive to hedge the volatility variability.
So far, the pricing equation is the following partial differential equation (PDE)

(1.4)
1 1 
Ft − rF + rSFS + V S 2 FSS + η 2 V FV V + ρηV SFSV − λ(Vt − V̄ ) − θVt FV = 0.
2 2
Equation (1.4) looks similar than the Black and Scholes pricing equation - apart
that the equation is more complicated due presence of the second stochastic pro-
cess. If one steps back to constant volatility all expressions with a V -derivative
drop and the Black and Scholes equation follows.

The first simplification of equation (1.4) arises if we exploit the homogeneity


property of the equation w.r.t. the underlying value. That is, we may transform
the variable S in such a way that the S-dependent coefficients in the PDE become
constant. Indeed the change of variable x = ln S exploits the homogeneity of (1.4)
in S and removes the S-coefficients in the equation.2 A second transform is to use
forward values and dimensionless variables. We define
 
S(t) F (t, T )
x = ln = ln ,
Ke−rτ K
where
2We illustrate the procedure of how to exploit the homogeneity relation. Consider the ex-
pression SFS . The chain rule reads for x(S)
∂ ∂x ∂
= .
∂S ∂S ∂x
Using x(S) = ln S we get
∂x 1
= .
∂S S
Hence,
∂x ∂F 1 ∂F ∂F
SFS = S =S = .
∂S ∂x S ∂x ∂x
4 1. THE HESTON MODEL

Figure 1.1. Solution approach to Heston’s model using Fourier transform.

• F (t, T ) is the T -time forward price of the stock index;


• K is the strike price of the option, T is the maturity date;
• τ = T − t.
Taking the forward rather than spot price removes the −rF -term in the PDE:
Since the cost to enter a Forward are zero, there can be no funding cost term in
the pricing equation.
We then get the PDE
1 1 1 
(1.5) − Fτ + V Fxx + V Fx + η 2 V FV V + ρηV FxV − λ(Vt − V̄ ) − θV FV = 0,
2 2 2
which is an equation in (x, V, τ ). The boundary condition for the call transforms
to
F (x, V, 0) = max(ex − K, 0).
The next step is to use Fourier transformation: This reduces the number of deriva-
tives in the PDE from 3 to 2, i.e. the transformed equation will be simpler to be
solved. The original pricing function is then expressed by the inverse transforma-
tion which in case of the Heston model can only be calculated numerically. But
the advantage is that we end up with a analytical formula which needs only one
numerical integration at the end.
1.3. PRICING EQUATION 5

We do the Fourier transformation in the x-variable, i.e. we define:


Definition 1.1. The Fourier transform F̂ (k) of a function F (x) is defined by
Z ∞
1
(1.6) F̂ (k, V, t) = √ eikx F (x, V, t) dx
2π −∞
if the integral is well-defined. The inverse Fourier transform is
Z ∞
1
(1.7) F̌ (k, V, t) = √ e−ikx F (x, V, t) dx .
2π −∞
We sometimes use the notation F and F −1 to denote the Fourier transform F̂ and
its inverse F̌ , respectively.
The Fourier transform and its inverse are linear transformations, i.e. they
can be applied to each term separately in the pricing PDE. Before we apply the
transform to the pricing PDE, we first check that this transform indeed transforms
derivatives into powers. We calculate:
Z
1
fbx (k) = √ eikx ∂x f (x)dx
2π R
Z
1 1
= √ eikx f (x)dx|∞ −∞ − √ ∂x (eikx )f (x)dx
2π 2π R
Z
1 ikx ∞ ik
= √ e f (x)dx|−∞ − √ eikx f (x)dx
2π 2π R
(1.8) = −ik fb(k) ,
where we assumed that f → 0 for x → ±∞. Summarizing, we get
(1.9) fbx (k) = −ipfb(k) ,
Using Fourier theory, we fix the overall logic of the approach, see Figure 1.1:
• We start with a pricing PDE of F (x, V, τ ) which we cannot solve.
• We Fourier transform the PDE into an equation for a function F̂ (k, V, τ ).
• We solve the transformed equation, i.e. we obtain explicitly the Fourier
transformed pricing function F̂ (k, V, τ ).
• We apply the inverse Fourier transform to the function F̂ (k, V, τ ) which
gives us an expression for the pricing function F (x, V, τ ).
In the Black and Scholes model all steps lead to an explicit expression, therefore we
end up with the known Black and Scholes pricing equation. In the Heston model,
the last step can not be done explicitly, i.e. the integral which defines the inverse
Fourier transform can not be calculated explicitly but needs to be treated numeri-
cally.

We next Fourier transform the PDE (1.5) by applying the following rule:

A derivative with respect to x becomes a multiplication by −ik.

We get:
(1.10)
1 1 1 
− F̂τ − k 2 V F̂ − ikV F̂ + η 2 V F̂V V − ikρηV F̂V − λ(Vt − V̄ ) − θV F̂V = 0 .
2 2 2
6 1. THE HESTON MODEL

Equation (1.10) is structural similar to the Black and Scholes equation since there
are only derivatives with respect to time and one state variable.We also need to
Fourier transform the boundary condition, i.e. to transform the Call payoff
Z ∞
1
F̂ (k, V, 0) = √ eikx F (x, V, 0) dx
2π −∞
which is the Fourier transform of the option payoff expressed in log of the asset
price. We get:
Z ∞
1
F̂ (k, V, 0) = √ eikx max(ex − K, 0) dx
2π −∞
Z ∞
1
= √ eikx (ex − K)Iex ≥K dx
2π −∞
Z ∞
1
= √ eikx (ex − K) dx
2π ln K
 (ik+1)x  x=∞
1 e eikx
= √ −K
2π ik + 1 ik x=ln K
1 K 1+ik
(1.11) = √ , =k > 1.
2π ik − k 2
Note that the complex function eikx with k, x real numbers defines the unit circle
in the complex plain. Therefore, if say x tends to infinity, the complex function
eikx is not converging - the value is spinning on the unit circle with an increasing
velocity. This problem shows up in the second last line of the above calculation. In
order to obtain convergence and hence existence of the call Fourier transform, we
have to assume that k is a complex number. Writing k in the representation of its
real and imaginary part, i.e. k = <k + i=k the product ikx reads
ix<k − x=k .
Therefore, if the imaginary part is positive, the exponential eikx converges. In sum-
mary, the above calculation show that the Fourier transform of the call only exists
if k is a complex number with imaginary part =k > 1. So Fourier-transforming
the payoff of a call option (and of most options) requires the use of the complex
Fourier integral or equivalently the Laplace transform, see the Section ”Mathemat-
ical Tools” for details.
The following Table summarizes the Fourier transform and the regularity do-
mains for some payoff functions:
Claim Payoff Function Transform Strip of Regularity
x + K iz+1
Call Option (e − K) − z2 −iz =(z) > 1
x + K iz+1
Put Option (K − e ) − z2 −iz =(z) < 0
K iz+1
Covered Call (ex , K)− z 2 −iz 0 < =(z) < 1
iz
Arrow-Debreu Claim δ(x − ln K) K Entire C-plane
Zero Coupon Bond 1 2πδ(z) =(z) = 0
exp(−iz ln(K))
Digital Call Option χ{[K,+∞} (x) iz =(z) ∈ (0, β)
We finally consider the solution of the PDE (1.10) with the transformed bound-
ary condition (1.11) for the call. The following mathematical fact holds for such
kind of equations:
1.3. PRICING EQUATION 7

Definition 1.2. The solution F̂ (k, V, τ ) of the PDE (1.10) with the boundary
condition
F̂ (k, V, 0) = 1
is called a Fundamental SolutionFundamental Solutions are also called Green’s
functions.
Theorem 1.3. The solution of the PDE (1.10) with the transformed option
payoff (1.11) for a call is equal to the product of the transformed option payoff
and the Fundamental Solution of the PDE.
We first discuss the content of this theorem and construct in the next section
the Fundamental Solution for the Heston model.

1.3.1. Fundamental Solution of Black and Scholes Equation. To un-


derstand the intuition of theorem 1.3, it suffices to consider the case of Black and
Scholes. The PDE for a contingent for an call option with price function F (t, S)
reads
1
Ft + rSFS + σ 2 S 2 FSS − rF = 0.
2
Substituting
1
S = ex , τ = σ 2 (T − t)
2
and defining the function
r
H(x, t) = eαx+βτ F (x, τ ), m = 1 2

where
1−m (m + 1)2
α= , β=−
2 4
and using the chain rule, H satisfies the heat equation
(1.12) Hτ (x, τ ) = Hxx (x, τ )
1 1
H(x, 0) = e 2 (m−1)x (ex − K)+ =: e 2 (m−1)x h(ex ).
where the last line represents the call contract. Therefore, the Black-Scholes equa-
tion is equivalent to the heat equation with the initial condition given in (1.12)
representing the transformed contracting terms: If we know H, we know F .

The Heston model is of the same type, i.e. the pricing equation is also a para-
bolic differential equation with an initial condition.

We solve first the equation (1.12) with the initial condition


H 0 (x, 0) = δ(x) ,
i.e. the delta function as an initial condition. The solution of the heat equation
H 0 (x, τ ) with this initial condition is called the Fundamental Solution. Why
do we call this solution also the Fundamental Solution as we did in Theorem 1.3?
The reason is due to the Fourier transform, i.e. the Fourier transform of the delta
function is up to a constant value equal to the unit function 1. Hence, the notion
8 1. THE HESTON MODEL

of a Fundamental Solution in theorem 1.3 is the same than above but only applied
to the Fourier case.3
To solve (1.12) with the delta function condition, we Fourier transform G w.r.t.
to x. The heat equation reads
1
(1.13) Ĥτ0 (k, τ ) = −k 2 Ĥ 0 (k, τ ), Ĥ 0 (k, 0) = √ .

The first order ordinary differential equation has the solution
2 1
(1.14) Ĥ 0 (k, τ ) = e−τ k √ .

Applying the inverse Fourier transform we get
1 2 1 2
(1.15) H 0 (x, τ ) = F −1 Ĥ 0 (k, τ ) = √ F −1 e−τ k = √ e−x /(4τ )
2π 4πτ
where we used the fact that
2 1 2
F −1 e−τ k = √ e−x /(4τ ) .
2πτ
Note that the Fundamental Solution is singular at (x, 0), i.e. the function
1 2
H 0 (x, τ ) = √ e−x /(4τ )
4πτ
becomes singular if τ → 0. Intuitively, the function becomes more an more peaked
about x, i.e. the delta function initial condition follows.

If H 0 (x, τ ) solve the heat equation, then for any fixed y the function H 0 (x−y, τ )
is also a solution of the heat equation. This is the main observation to obtain a
solution to the heat equation with a general function as its boundary condition,
i.e. a call contact for example and not the delta function. Since for any fixed y a
solution of the heat equation is obtained, we claim that
Z
(1.16) H(x, τ ) = H 0 (x − y, τ )H(y, 0)dy
R
solve the original Black and Scholes problem. If we can prove this, we illustrated
theorem 1.3 using the Black and Scholes model: Equation (1.16) states that the
3We check the claim about the Fourier transform of the delta function, see the Appendix
for details about the delta function. We have for an arbitrary infinitely differentiable function φ
which decays fast at infinity:
Z
hFδ, φi(k) = φ(x)Fδ(x) dx
ZR
:= F φ(x)δ(x) dx
R
= Fφ(0)
Z
1
= √ ei0x φ(x) dx
2π R
Z
1
= √ 1φ(x) dx.
2π R
Hence,
1
Fδ(k) = √ 1.

This shows that the Fourier transform maps a maximally peaked function into a maximally diluted
one.
1.3. PRICING EQUATION 9

solution of an initial-value boundary heat equation problem is equal to the convo-


lution of the Fundamental Solution with the boundary condition. Then the Fourier
transform of this equation implies that the solution of the Fourier transformed
prices - the content of theorem 1.3 - is the Fourier transform of a convolution which
is the multiplication of the individual Fourier transforms.

We check that (1.16) solves the Black and Scholes pricing problem. We have:
Z
∂ ∂ 0
H(x, τ ) = H (x − y, τ )H(y, 0)dy .
∂τ R ∂τ
Z
∂2 ∂2 0
H(x, τ ) = H (x − y, τ )H(y, 0)dy .
∂x2 R ∂x
2

Hence,
(1.17) Hτ (x, τ ) − Hxx (x, τ )
Z  
∂ 0 ∂2
= H (x − y, τ ) − 2 H 0 (x − y, τ ) H(y, 0)dy
R ∂τ ∂x
(1.18) = 0
where we used that H 0 is a Fundamental Solution. We verify next the boundary
condition, i.e. Z
lim H(x, τ ) = lim H 0 (x − y, τ )H(y, 0)dy .
τ →0 τ →0 R
But H 0 becomes the delta function and we get
Z
lim H(x, τ ) = δ(x − y, 0)H(y, 0)dy = H(x, 0) .
τ →0 R
This shows that the general solution is a convolution of the Fundamental Solution
and the boundary condition.
Using these facts we finally compute the pricing function in Black and Scholes
model:
 2
 1 
H(x, τ ) = F −1 e−τ k F e 2 (m−1)x h(ex )
1  2
  1 
= √ F −1 e−τ k ? F −1 F e 2 (m−1)x h(ex )

1 2 1
= √ e−x /(4τ ) ? e 2 (m−1)x h(ex )
4πτ
Z
1 2 1
= √ e−(x−y) /(4τ ) e 2 (m−1)y h(ey ) dy
4πτ R
Z
1 2 1
= √ e−(x−y) /(4τ ) e 2 (m−1)y h(ey ) dy
4πτ R
Z
1 2 1
= √ e−(x−y) /(4τ ) e 2 (m−1)y (ey − K)+ dy
4πτ R
Z ∞
1 2 1
= √ e−(x−y) /(4τ ) e 2 (m−1)y (ey − K) dy
4πτ ln K
(1.19) → Black-Scholes,
where we inserted a call option in the last steps.
10 1. THE HESTON MODEL

1.3.2. Properties of the Fundamental Solution. We discuss:


• The Fundamental Solution (or the Green function) is both norm- and
martingale preserving.
• The Fundamental Solution is a characteristic function.
To explain the meaning of these statements, we consider an Arrow-Debreu claim,
i.e. the option payoff reads
F (S, V, 0) = δ(S(T ) − K) .
The option price is given in the S-variable by:

Z ia+∞
Fδ (S, V, K, τ ) = e−rτ e−ik ln S F(δ(S(T ) − K))F̂ (k, V, τ )dk,
ia−−∞

with Fδ the Fundamental Solution or the Green function, i.e the solution of a
PDE with a delta function as its boundary condition: The Green’s function is the
price of an Arrow-Debreu security. This function is closely related to the risk-
neutral transition density function of the martingale process which corresponds to
the PDE, i.e. to p(S, V, S(T ), τ ) for a stochastic process with initial value S and V
will after the elapse of time τ reach the point S(T ). Hence, p(S, V, S(T ), τ )dS(T )
is the probability that after τ the process ends up in (S(T ), S(T ) + dS(T )). Since
the stock must end somewhere,
Z ∞
p(S, V, S(T ), τ )dS(T ) = 1.
0

Indeed, if the Green function is also norm preserving as the transition density is,
then the two functions are the same. There is a simple test to check whether the
Green function is norm preserving. Basically, we have to invert the option price
formula. That is, we have to express F̂ as function of the Green’s function.//
A calculation leads to the following inversion result:
Z ∞
(1.20) F̂ (k, V, τ ) = eiky Fδ (S, V, S(T ), τ ) dS(T )
0

with y = ln(S/S(T )) + rτ . It then follows that if if k = 0 and F̂ (0, V, τ ) = 1, then


the Green function is norm preserving.
Furthermore, if k = i and F̂ (i, V, τ ) = 1 hold, then (1.20) becomes
Z ∞
(1.21) S = e−rτ S(T )Fδ (S, V, S(T )τ ) dS(T )
0

i.e. the martingale property of the underlying process follows. In summary, the
Fundamental solution F̂ can be norm- and martingale property preserving if the
cited two conditions hold.

Example 1.4. Consider F̂ and its solution in the Heston model, i.e. (1.22)
and (1.24). k = 0 implies that d = β. But this in turn implies r− = 0. This leads
to g =. But this finally gives D = C = 0 and hence F̂ (k, V, τ ) = e0 = 1, i.e. norm
preservation. Therefore, the Green function in the Heston model is equal to the
risk-neutral transition density function.
1.4. CONSTRUCTION OF HESTON’S FUNDAMENTAL SOLUTION 11

We finally claim that the Fundamental Solution F̂ (k, V, τ ) is a characteristic


function, i.e.
Z ∞
 
F̂ (k, V, τ ) = eikx dG(x) = E eikX = ψ(−k)
−∞

with G the cumulative distribution function, x real number and X the underlying
state variable, i.e. for example the variable S in the Heston model. To prove this,
one has to show that (i) G is non-decreasing and (ii) that it attains the values
0 and 1 at ±∞. The proof is an exercise. The Green function (or Fundamental
Solution) in the Heston model is identical to the risk-neutral transition density of
the underlying stock process. Furthermore, the Fundamental Solution is norm -
and martingale preserving.

The Fundamental Solution (Green function) is equal to the risk-neutral transition


density of the underlying stock process if the model in norm-preserving. The Fun-
damental Solution is a characteristic function, i.e.
Z ∞
 
F̂ (k, V, τ ) = eikx dG(x) = E eikX = ψ(−k) .
−∞
The characteristic exponent φ is defined by
ψ(k, t) = eφ(k,t) .

1.3.3. Option Price - Inverse Fourier Transform. The last step in the
calculation of the option price is given by the inverse Fourier transform:

The option price F (x, V, τ ) is given by the inverse Fourier transform of the Fourier
transformed payoff F̂ (k, V, 0) times the Fundamental Solution F̂ (k, V, τ ), i.e.
Z ia+∞
F (x, V, τ ) = e−ikx F̂ (k, V, 0)F̂ (k, V, τ )dk , =k > 1, a− < a < a+ .
ia−∞

First we recall that the integration cannot be done explicitly in the Heston case,
i.e. we use the numerical Fast Fourier Transform to solve it. This is considered
for the Variance Gamma model in the next Chapter. Second, we still do not know
the explicit form of the Fundamental Solution in Heston’s model in Fourier space.
Third we need to discuss the second restriction a− < a < a+ . This restriction
is necessary that a Fundamental Solution exists. This is a strip in the complex
number plane and is called the strip of regularity.

1.4. Construction of Heston’s Fundamental Solution


We consider the construction of the Fundamental Solution of Heston’s model
in Fourier space. By definition, the Fundamental Solution in Fourier space is the
solution of Heston’s PDE (1.10) with a unit function initial condition. To obtain
the Fundamental Solution, we note that the dynamics of the Heston state variables
are of the affine type in the variables log S and V . Hence, the solution of the
PDE which is given by an expected value due to the Feynman-Kac Theorem is an
exponential of the affine type, i.e. we set for the Fundamental Solution:
12 1. THE HESTON MODEL

(1.22) F̂ (k, V, τ ) = exp {C(k, τ ) + D(k, τ )V }


and
C(k, 0) = D(k, 0) = 0
such that F̂ = 1 is satisfied at maturity. Inserting this Ansatz into the PDE
(1.10) implies

Ċ + V Ḋ = V −k 2 /2 − ik/2 + 1/2η 2 D2 − ikρηD − λ(1 − Θ)D + V̄ D
with Ċ the time derivative. This equation has to hold for all values of V . Therefore,
comparing powers in V we get that the following system of ODE has to hold:
Ċ = V̄ D
Ḋ = −k 2 /2 − ik/2 + 1/2η 2 D2 − ikρηD − λ(1 − Θ)D =: α − βD + γD2 .
(1.23)
The first equation is trivial, the second non-linear one is a Riccati equation. Such a
system of equation is standard in affine finance models, see the short rate interest
rate models.
Theorem 1.5. The functions
1 − e−dτ
(1.24) D(k, τ ) = r− ,
1 − g e−dτ
  
2 1 − g e−dτ
(1.25) C(k, τ ) = λ0 r− τ − 2 log ,
η 1−g
with
r−
g = ,
r+
β±d
r± =
η2
p
d = (k 2 − ik)η 2 + β 2
λ0 = λ−Θ
β = λ0 − ρη − ρηik
solve (1.23).
Putting all pieces together we finally arrived at a first semi-analytical solution
of the Heston model.
Proof. The trick to solve the Riccati ODE in (1.23) is to set

D=λ .
u
Then
ü u̇2 u̇ u̇2
Ḋ = λ − λ 2 = α − βλ + γλ2 2 .
u u u u
By choosing λ = − γ1 , what remains is a second order linear equation in u:
ü u̇
λ = α − βλ ,
u u
1.5. SINGULARITIES 13

or
(1.26) λü + βλu̇ − αu = 0,
which is a second-order linear ODE with constant coefficients. The solution to
(1.26) is a linear superposition of
u = eµτ .
Inserting this we get a quadratic equation in the unknown parameter µ,

(1.27) λµ2 + βλµ − α eµτ = 0, ∀τ.
| {z }
=0

The general solution to (1.27) is


p
µ1 τ µ2 τ −βλ ± β 2 λ2 + 4λα
u = c1 e + c2 e , µ1,2 = .

We then obtain
u̇ c1 µ1 eµ1 τ + c2 µ2 eµ2 τ
D=λ =λ .
u c1 eµ1 τ + c2 eµ2 τ
Factorizing out c1 and setting c = c2 /c1 , we are left with a single constant of
integration c
µ1 eµ1 τ + c µ2 eµ2 τ
D=λ .
eµ1 τ + c eµ2 τ
From the initial condition D(0) = 0, we get
µ1 + c µ2
λ = 0;
1+c
hence, the constant of integration is
µ1
c=− .
µ2
Finally, we get
1 eµ1 τ − eµ2 τ
D(τ ) = − µ1 µ2 ,
γ µ2 eµ1 τ − µ1 eµ2 τ
and, from Ċ = aD,
Z τ
C(τ ) = a D(s)ds + c3 ,
0
with initial condition C(0) = 0, i.e. c3 = 0. 

1.5. Singularities
The Heston model and also most other models face the problem that some
functions might be singular for some specific values. If we recall the pricing formula
in Heston’s model
Z ia+∞
F (x, V, τ ) = e−ikx F̂ (k, V, 0)F̂ (k, V, τ )dk , =k > 1, a− < a < a+ ,
ia−∞

we see that singularities can occur in the Fourier transform


• of the Fundamentals Solution F̂ (k, V, τ ) and
• of the option payoff F̂ (k, V, 0) .
14 1. THE HESTON MODEL

We have seen that the Fourier transform of a call option exists only for =k > 1.
Furthermore, the transformed payoff has two singularities: Two simple poles at 0, i,
see the denominator in (1.11).

We consider the singularities of the Fundamental Solution in the sequel.

Since the dynamics are affine in ln S and V in the Heston model, the Fundamen-
tal Solution (which is equal to the characteristic function) is of the exponentially-
affine form
F̂ (k, V, τ ) = exp {C(k, τ ) + D(k, τ )V } ,
see (1.24) for the explicit formulae. There are two solutions in the literature of
the corresponding ODE system for the functions C and D. The solutions are
algebraically equivalent but the singularity structure is different.
The first formulation of the solution F̂ = eC+DV reads:
 
β − d 1 − e−dτ
D(k, τ ) =
η2 1 − g e−dτ
g e−dτ − 1
(1.28) C(k, τ ) = λV̄ η −2 ((β − d)τ − 2 ln Ψ) , Ψ = ,
g−1
with
p
• g = β−d
β+d , d = β 2 − 4δγ,
• δ ≡ − 2 k(i + k), β = λ − ρηki, γ = 21 η 2 .
1

An algebraically equivalent solution is


 
β + d eDτ − 1
D̃(k, τ ) =
η2 c edτ − 1
  c edτ − 1
(1.29) C̃(k, τ ) = λV̄ η −2 (β − d)τ − 2 ln Ψ̃ , Ψ̃ = ,
c−1
with
β+d 1
c= = .
β−d g
C̃ is the formulation originally used by Heston (93) and which we call “for-
mulation 1”; the alternative formulation, which we call “formulation 2”, has been
suggested by Duffie, Pan and Singleton (00), Schouten et al. (04) and Gatheral
(06). Again, the two formulations are algebraically equivalent. We consider the
singularity4 structure of the two formulations. Clearly, any singularity of the Fun-
damental Solution is a singularity of the exponent above and vice versa. There
are three possible sources of singularities: (i) the square root in d, (ii) the complex
logarithm, ln Ψ or ln Ψ̃ and (iii) the denominator 1 − e−dτ .

We consider only (ii) and (iii) in the sequel.


Singularity of Log-Term

First, we recall some basic facts about the complex logarithm. The solution of
the equation ew = z, w, z ∈ C, is called the complex logarithm and is defined as
4We are not very precise at this stage, i.e. we use the word singularity for discontinuities and
true singularities.
1.5. SINGULARITIES 15

log z = log |z| + i(arg(z) + 2πn) , n ∈ Z .


w+2πin
Since e = 1 for different integer n, there is no unique inverse function, i.e.
logarithm.
The principal branch is defined for n = 0. For this branch the logarithm is
a continuous function expect on the negative real axis and at zero. In this case
the argument satisfies arg(z) ∈ (−π, π). For a complex number z = x + iy, the
argument gives the angle between the line joining the number to the origin and the
positive real axis, see the Technical Appendix
p Complex Analysis for details. The
argument is calculated as follows (r = x2 + y 2 ):

 arccos x/r, for y ≥ 0
arg(z) = − arccos x/r, , for y < 0;

undef ined, r = 0.
To understand the problem of discontinuities, consider the straight line in the
complex plane
fn (z) = 1 − x − n + i(x − 1/2) , n ∈ Z .
If you draw the lines f0 , f−1 , f1 , f2 , then the first two lines do not cross the 2nd and
3rd quadrant, whereas the two other functions do. Since the logarithm is continuous
except on the negative real axis, i.e. the argument is a continuous function, one
expects that the first two functions are continuous, the other are discontinuous. To
see this, we consider the argument around the critical value x = 1/2. We get for
x ↓ 1/2:
1−x 1/2
arg(f0 (z)) = arccos p → arccos =0.
(1 − x)2 + (x − 1/2)2 1/2
and for x ↑ 1/2:
1−x 1/2
arg(f0 (z)) = − arccos p → − arccos =0.
(1 − x)2 + (x − 1/2)2 1/2
Hence, the argument of f0 is a continuous function. We do the same calculations
for f1 . We get for x ↓ 1/2:
−x −1/2
arg(f1 (z)) = arccos p → arccos =π .
x2 + (x − 1/2)2 1/2
and for x ↑ 1/2:
−x −1/2
arg(f1 (z)) = − arccos p → − arccos = −π .
x2 + (x − 1/2)2 1/2
Hence, the principal argument of f1 is not continuous.

Therefore for f0 , f−1 the principal argument is a continuous function whereas


it is discontinuous for the two other functions.

Since Ψ̃(k, τ ) crosses the negative real axis for k ∈ (0, ∞), ln Ψ̃ faces at each
crossing a singularity. Singular values of Ψ̃ lead to discontinuous option prices
and, therefore, need to be avoided. Contrary, ln Ψ does not faces any singularities.
Therefore, one should work with this formulation.
16 1. THE HESTON MODEL

If one insist to work with representation of Heston, there is an algorithm of


Kahl and Jaeckel (05), the Rotation Count Algorithm, which guarantees Ψ̃ to be a
continuous function of k. The idea of the Algorithm is to choose the phase of the
complex number such that the argument - which is not the principal one - becomes
a continuous function.

In summary:
• Using the principal branch of the logarithm, formulation 2 of the Heston
model has a continuous characteristic function.
• Using formulation 1, the characteristic function is not continuous using the
principal branch of the logarithm. Using the Rotation Count Algorithm
continuity is restored.
• Since formulation 2 is numerically more stable than formulation 1, use
formulation 2.
Singularity of Denominator

The Fundamental Solution given in (1.24) shows that further singularities arise
if in the function
1 − e−dτ
D(k, τ ) = r−
1 − g e−dτ
the denominator becomes zero, i.e. for
1 − g e−dτ = 0 .
The following result of Lukacs (1970) is basic for the analysis:
Theorem 1.6. If a characteristic function (Fundamental Solution) ψ(p) is an-
alytic in the neighborhood of p = 0, then it is also analytic in a horizontal strip and
it can be represented by a Fourier integral. The strip is either the whole plane or
has one or two boundary lines. The characteristic function ψ(p) is singular at the
boundary of the strip.
Therefore, for the Heston model we consider the strip of regularity ΛHe , i.e. a
strip in the complex plain such that within the strip the Fundamental Solution is
analytic. Since the strip is possibly bounded by two imaginary numbers −ia− <
0, 1 < ia+ with a−/+ ∈ R, the boundaries are given by the solution of the equation
1 − g(−ia) e−d(−ia)τ = 0 .
Then, a− is the largest solution close to 0 in the interval (−∞, y− ) and a+ is
smallest solution in the interval (y+ , ∞) with
p
η − 2ρ(λ − Θ) ± η 2 − 4ρη(λ − Θ) + 4(λ − Θ)2
y± = .
2η(1 − ρ2 )
Are there other singularities except those on the imaginary axis and are there
singularities outside of the strip of convergence? Lucic (07) proves:
Theorem 1.7. All singularities of (1.24) are pure imaginary.
Example 1.8. The characteristic exponent φ(k) , ψt (k) = etφ(k) of the variance
gamma model is given by
 
MG
φ (u) = C log .
(M − ik) (G + ik)
1.5. SINGULARITIES 17

ψ (u) has poles at


a− = −iM and a+ = iG
5
and that there exists a regular analytic extension on the line
{ia : a ∈ (−M, G)}
i.e. there are no poles or branching points in this interval - this is Lukacs Theorem.
Note that singularities of φ are the singularities of ψ.
Example 1.9. We try to find the strip of regularity for the pricing of a call
option in the simplified stochastic volatility model
p
(1.30) dV (t) = ξ V (t)dW (t)
where we further assume that the S-dynamics is the same as in Heston model but
that ρ = 0. The Fundamental Solution of this model satisfies:
Lemma 1.10. Using the PDE (1.10), the Fundamental Solution F̂ (τ, V, k) in
the reduced model satisfies the PDE
1
(1.31) F̂τ = ξ 2 V F̂V V − (ik − k 2 )V F̂ , F̂ (k, V, 0) = 1
2
with solution
r !!
Vp ik − k 2
(1.32) F̂ (τ ) = exp − ik − k 2 tanh ξτ .
ξ 2
Proof. The proof is left as an exercise for the reader. 
2
We define c(k) = ik − k . Since c(0) = c(i) = 0 at the two poles of the payoff
function and tanh(z) is analytic for |z| ≤ π/2, F̂ is analytic around k = 0, i. Since
we are interested in a horizontal strip in the complex plain, we search singularities
for the imaginary numbers
k = ia , a ∈ R.
Then
1
c(ia) = (a − a2 ) < 0 ⇐⇒ a < 0 or a > 1.
2
In either of the two cases for a,
r !
ik − k 2
< ξτ = 0
2
q
2
Hence, ik−k 2 ξτ is then purely imaginary and we set
r
ik − k 2
ξτ = iφ , φ ∈ R.
2
But the identity
tanh(iφ) = i tan φ
implies:
{k = ia|F̂ is singular}
(2n + 1)π
(1.33) = {φ ∈ R| φ = , n ∈ Z .}
2
5A complex function that is differentiable at every point of a region H is called a regular
analytic function in H.
18 1. THE HESTON MODEL

For n = 0, we get for example


s
1 1 π2
a± (n = 0) = ± + 2 2 .
2 4 ξ τ

For ξ 2 = τ = 1, we get
s ( (
1 1 π2 3.68 a+ (n = 0)
a± (n = 0) = ± + 2 2 = = .
2 4 ξ τ −2.68 a− (n = 0)

Hence we found two bounds for a strip of regularity:

AX = (−2.68, 3.68) .

These bounds depends on the model parameters. Therefore, the strip where ψ̂ is
regular intersects the set =k > 1 where the call Fourier transform exist. Hence,
there exists a strip in the complex domain where both the characteristic function
is analytic and the Fourier transform of the call exists. Note that for ξ 2 → 0 or
τ → 0, i.e. either volatility is constant or time to maturity is zero, a± tends to ±∞.
In this case, the function F̂ is regular on all of C. Contrary, if ξ 2 → ∞, a± → 0, 1.
Hence, if volatility explodes, the strip which guarantees regularity of the function
F̂ and the domain where the Fourier transform payoff exist become disjoint.

1.6. Synthesis of Option Pricing


The missing piece in the Heston model is the discussion of the inverse Fourier
transform, i.e. the expression of the option price as the inverse Fourier transform
of the Fundamental Solution times the option payoff - both Fourier transformed,
see equation (1.3.3). In particular, this raises the question when does the inverse
transform exists. The answer depends not only on the decay properties of the Fun-
damental Solution and the option payoff at infinity but also on the region in the
complex plain where these functions are well-defined, i.e. free of singularities.

We answer these question by considering a larger and more general setup than
the Heston model itself. The Heston model will be used as an example to illustrate
the general approach. The presentation very closely follows the work of Lee (04).
Lee’s approach assumes that the characteristic function of the underlying vari-
able is known. We work with an unspecified state variable or equivalently process.
It can be a diffusion, a jump-diffusion or a Lévy process. Furthermore, interest
rates can be stochastic too. We will specify the state variable in a few examples
below.
Contrary to the former approach in the last sections, the transform is now taken
with respect to the strike k, and not the underlying variable S. The underlying
abstract state variable is X is a P-martingale. This variable could have a SDE spec-
ification (S in Heston model, Bakshi-Madan, Duffie-Pan-Singleton, Lewis, Zhu) or
could follow a Lévy process (such as in Carr-Wu, Barndorff-Nielsen and Sheppard,
Eberlein-Prause, Madan-Chang-Carr, Carr-Madan-Geman-Yor).
1.6. SYNTHESIS OF OPTION PRICING 19

äG GX

LX

AX
-20 -10 10 20

-2 ä

-4 ä
-ä M
Figure 1.2. The sets ΛX , ΓX and AX for M = 4, G = 6

1.6.1. Characteristic Function. We assume for simplicity that the state


variable X is one-dimensional and that we face deterministic interest rates. We
define the following sets:
AX is the interior of the set
AX = interior{ ν ∈ R : E(eνX ) < ∞}.
The set ΛX is a strip defined by:
ΛX = {k̂ ∈ C| − =(k̂) ∈ AX } .
Definition 1.11. The function ψ : ΛX = (a− , a+ ) → C
ψ(k̂, t) =: ψ(k̂) = EP̃ [e−rt+ik̂X ]
is called the discounted characteristic function.
Example 1.12. (Variance Gamma model, continued) Figure 1.2 shows the
different sets for the Variance Gamma model. Lukasc’ theorem implies that the
characteristic function ψ exists for all z in the strip of regularity
ΛX = {x + iy : x ∈ R, y ∈ (−M, G)} .
i(iy)X
In particular, E(e ) = E(e−yX ) < ∞ for all y ∈ (−M, G), i.e. the characteristic
function exists. Then
AX = (−M, G) .
Note that ΛX and AX are independent of t - this holds for VG model, but not for
Heston. The set ΓX in the Figure 1.2 is defined below.
1.6.2. Fourier Theory of Option Pricing. The price of an option is written
V (k) = V (0, X, k) we get for a call option
 
V (k) = e−rT E P [f (X, k)] = e−rT EP (X − k)+ .
Note that k is not denoting the Fourier transformed variable of x of last sections
but the strike! Hence, the synthesis is done by Fourier transforming then below the
20 1. THE HESTON MODEL

strike and not the state variable X. The major advantage of this approach is that
one can use numerical methods to price options with multiple strikes in this new
approach, whereas this is not possible in the X-variable transformation approach.

We know that V (k) ∈ / L1 and therefore the ordinary Fourier transform is not
defined. Therefore, we considered the complex Fourier transform. Another possi-
bility is to enforce convergence defining
Va (k) = eak V (k), a>0,
as Carr-Madan (99) proposed. The damping factor - i.e. the exponential function
- will then ensure that option prices and their Fourier transforms exist if a > 0.
But the value of a can not be chosen arbitrary, i.e. the value of a is related to the
set AX . Hence, it might be that a < 0 is a necessary choice. But then, the Fourier
transformed damped option prices do not exist. In this case we have to consider
the complex Fourier transform and the shift of integration contour technique to get
option prices for arbitrary values of a. We consider the details of these ideas next.
1.6.2.1. a > 0. For a > 0, we denote by
Z ∞
1
V̂a (k̂) = √ eik̂k Va (k) dk
2π −∞
the real Fourier transform, i.e. k, k̂ are real numbers.
Theorem 1.13. Assume that 0 ∈ AX . Then there exists a > 0 with a ∈ AX and
for any such a the Fourier transform of the damped option price exists, V̂a (k) ∈ L1
and V̂a (k) is given by:
1 ψ(k̂ − ia)
(1.34) V̂a (k) = √
2π (ik̂ + a)2
Proof. We claim that for all p > 0 and all x ∈ R
epx
(1.35) x − k ≤ pk+1
pe
holds. To prove this, we fix x0 = k + 1/p and define the functions
epx
m(x) = x − k , n(x) = pk+1 .
pe
If follows that
m(x0 ) = n(x0 ) , mx (x0 ) = nx (x0 )
and while m is zero for all derivatives greater than 1, the second derivative of n is
positive. This proves the claim. Next, we substitute x → X and take expectation
in (1.35). Finally, multiplying by e−rT implies the bound
EP [epX ]
(1.36) V (k) ≤
pepk+1
and
EP [epX ]eak
(1.37) Va (k) = V (k)eak ≤ .
pepk+1
Hence, if we assume a > 0, a necessary condition p > a > 0 follows. Under this
condition, Va (k) is exponentially decaying.
We prove the explicit formula for the Fourier transformed option price.
1.6. SYNTHESIS OF OPTION PRICING 21

Z ∞
1
(1.38) V̂a (k̂) = √ eik̂k Va (k) dk
2π −∞
Z ∞
1
= √ eik̂k eak EP [f (X, k)] dk
2π −∞
Z ∞ 
1
= √ EP eik̂k eak f (X, k) dk
2π −∞
Z ∞ 
1
= √ EP eik̂k eak (X − k)+ dk
2π −∞
1
= √ EP

Z X (ik̂+a)k !
(ik̂+a)k X (ik̂+a)k X e
× e X|−∞ − e k|−∞ + dk .
−∞ (ik̂ + a)

If a > 0, the Fourier transform exists in the last line. Continuing, we get:
Z X (ik̂+a)k !
1 e
V̂a (k̂) = √ EP dk
2π −∞ (ik̂ + a)

1 EP [ e(ik̂+a)X ]
= √
2π (ik̂ + a)2
1 ψ(k̂ − ia)
(1.39) = √
2π (ik̂ + a)2

Given this representation, we bound the transform prices as follows:

1 EP [ e(ik̂+a)X ]
|V̂a (k̂)| = |√ |
2π (ik̂ + a)2
|EP [ e(ik̂+a)X ]|

|(ik̂ + a)2 |
|ψ(−ia)|
≤ .
|(k̂ + a)2 |

Therefore, if a ∈ AX , the Fourier transformed damped option prices are in L1 .




Given the damped Fourier prices V̂a (k̂) in last Theorem, which are in L1 , we
can recover for a > 0 the undamped option prices by simply reverting the Fourier
transform and undamping with the function e−ak . This gives at once the undamped
option price formula for a > 0:

Z ∞ Z ∞  
(1.40) V (k) = e−ak e−ik̂k V̂a (k̂)dk̂ = 2e−ak < e−ik̂k V̂a (k̂) dk̂ .
−∞ 0
22 1. THE HESTON MODEL

1.6.3. General a case. For a < 0, the theory of last section breaks down. To
obtain a theory for general values of a, we define the set
Γ = {z ∈ C | − =(z) ∈ AX }
and ĝ : Γ → C
−ψ(z)
ĝ(z) = .
z2
The set ΓX is illustrated in the Figure 1.2 for the Variance Gamma model. We see,
that this set is a shift of the regularity strip ΛX parallel to the real axis. Hence, this
new set contains singularities, i.e. iG = i5 in the example. Defining this set, we can
if necessary shift the integration of the necessary complex Fourier transform below
to an axis parallel to the real axis where convergence of the Fourier prices follows.
If it is necessary to cross a singularity to obtain convergence we have to use the
residue theorem, i.e. the option price formula picks up factors if such singularities
are crossed.

If a > 0 and a ∈ AX then the formula (1.34) in Theorem 1.13 implies that
V̂a (k̂) = ĝ(k̂ − ia).
But V̂a (k̂) exists is in L1 and the option price can be expressed as in (1.40). There-
fore, in the region where z ∈ Γ and −=(z) > 0, ĝ(z) is represented by the complex
Fourier transform of the undamped option price, i.e.
Z ∞
1
(1.41) ĝ(z) = √ eizk V (k)dk .
2π −∞
Applying Theorem 1.13, we can invert the last formula along a contour −=(z) =
a > 0 i.e.
Z ∞−ia Z ∞−ia

V (k) = e−izk ĝ(z)dz = 2 < e−izk ĝ(z) dz
−∞−ia 0−ia

for a > 0, a ∈ AX . In summary, we rewrote the results for a > 0 by using the new
domain Γ and the function ĝ.

We consider now a < 0. Then the Fourier transform V̂a (k̂) does not exist or
equivalently, the integral (1.41) does not exist for −=(z) < 0: The complex Fourier
transform is not well-defined in this region. But the definition of ĝ(z) do make sense
and the integrals in (1.42) do exist for a < 0. But their value does not give us the
option values we are looking for, since the integrations paths has shifted across the
pole at z = 0. Hence, the option value is the integral (1.42) less the contribution
picked up at the pole z = 0. This contribution is easily calculated using the Residue
theorem which gives us a new option formula. Finally, for a = 0, the integral (1.42)
passes through a pole - the contribution of the residue is cut in half.
The following theorem summarizes these facts.
Theorem 1.14. Let a ∈ R {0} be any real number, such that a ∈ AX . Then
the price V (k) of a European call option is given as
Z ∞−iα

(1.42) C (k) = Rα + 2 < ĝ (z) e−ikz dz,
0−iα
1.6. SYNTHESIS OF OPTION PRICING 23

where 

0, if α > 0,
Rα = 21 (−ψ 0 (0) − kψ(0)), if α = 0,


−iψ 0 (0) − kψ(0), if α < 0,
Proof. We prove a < 0. Fist, ĝ is analytic in a strip around z = 0. Choosing
a rectangular contour γ which encloses 0. The vertical integrals of the rectangle
vanish if the horizontal sides of the rectangle tend to ±∞ since the integrand decays
exponentially in the real axis direction. Therefore, we only have to calculate the
residuum at z = 0. By the general formula
d 
res(z = 0) = −ψ(z)e−ikz |z=0 .
dz
Doing the derivative the a < 0 results follows. The a > 0 case is yet proved and for
the a = 0 case, we only note that if a contour crosses a pole, half of the residuum is
attributed to the interior of the contour and the other half to the exterior part. 
As an example, we consider the Variance-Gamma model proposed by Madan-
Carr-Chang. The state vector X = log S has a Fourier transform
exp (−rT + iz(log S0 + µT )
ψX (z) = T /ν
1 − iνΘz + 21 νσ 2 z 2
with µ a function of r, ν, Θ and σ. The regularity bounds in AX = (a− , a+ ) are
given by the roots of the denominator, i.e.
r
Θ 2 Θ2
a± = − 2 ± 2
+ 4 .
σ νσ σ
If volatility explodes, the domain AX shrinks to zero and also the analyticity do-
main of the Fundamental Solution vanishes. The function ψX decays as ∼ 2T1/ν if
z0
we set z = z0 + iz1 . Therefore, a power decay follows if 2T /ν > 0 holds.

In summary, why is there a restriction α > 0 in Madan-Carr’s approach, but


not in Lee’s approach and what happens for α & 0?

The answer is: as α & 0, the function Vα (k) converges (pointwise) to the
function V (k) which is non-integrable. On the other hand, the Fourier transform
V̂α (k) converges (pointwise) to the well-defined function ĝ (k). We thus see a case,
where taking limits and integration are not interchangeable:
Z ∞
ĝ (z) = lim eizk eαk V (k) dk
a&0 0
Z ∞
6= eizk lim eαk V (k) dk = undefined.
0 a&0

Since Z
V (k) = e−αk e−ikk̂ V̂a (k̂)dk̂
for all α > 0, it follows that
Z ∞  
V (k) = lim e−ikk̂ V̂a k̂ dk̂,
a&0 0
24 1. THE HESTON MODEL

but as we have seen,

Z ∞  
e−ikk̂ lim V̂a k̂ dk̂ = V (k) − Residuum Corrections,
0 a&0

Again, taking limits and integration are not interchangeable. Using the Fourier
symbol we can state the result more neatly as follows:

ĝ = lim F(Va ) 6= F( lim Va ) = undefined


a&0 a&0

and

V = lim F 1− (FVa ) 6= F −1 ( lim FVa ) = V − Residuum .


a&0 a&0

Which analyticity strip should one best choose when calculating option prices?
Since done correctly, the pricing is independent of the strip, other criteria matter
for this question. Because in most models, the Fourier inversion cannot be carried
out explicitly, one has to rely on numerical methods and hence, their error analy-
sis. This analysis determines the best strip for calculating the option prices. We
consider this issue for the Variance Gamma model in detail.

We conclude this section by deriving different option pricing formula for a call
option in Heston’s model.

Example 1.15. The price of a call option C(S, T ) can be written as:

Z ia+∞
K e−rT 1
(1.43) C(S, T ) = − √ e−ikX ψ(−k) dk, a ∈ (1, b),
2π ia−∞ k 2 − ik

where X = log(S/K) + (r − q)T is the dimensionless moneyness.


The restriction a ∈ (1, b) considers both analyticity of the characteristic func-
tion and that of the payoff function. The call has a singularity at 0 and i in Fourier
space. We we want to shift the integration contour for the inverse Fourier transform
to the domain a ∈ (0, 1), we cross the singularity at i and we have to consider the
residuum at k = i. Similar, if the contour also crosses 0. We therefore calculate
the residua. At k = 0 we have:

 
K e−rT −ikX 1
res(0) = k − e ψ(−k) 2
2π k − ik k=0
 
Ke−rT 1
= − e−i0X ψ(−0)
2π 0−i
−rT
 
Ke 1
= .
2π i
1.6. SYNTHESIS OF OPTION PRICING 25

At k = i we have:
 
K e−rT −ikX 1
res(i) = (k − i) − e ψ(−k) 2
2π k − ik k=i
 
K e−rT 1
= − e−iik ψ(−i)
2π i
−rT
 
Ke 1
= e(log(S/K)+(r−q)T ) ψ(−i)
2π i
−rT
 
Ke 1
= S/Ke(r−q)T ψ(−i)
2π i
−rT
 
Se 1
= .
2π i
Therefore, if we shift α → α̃ ∈ (0, 1), we pick up the i-pole. From the residue
theorem, this means that 2πi times the residue is added to the option price, i.e.
Z
K e−rT iã+∞ −ikX 1
(1.44) C(S, T ) = S e−rT − e ψ(−k) 2 dk, ã ∈ (0, 1).
2π iã−∞ k − ik
We then shift a → ã ∈ (−c, 0), thus picking up the i- and 0-poles. Since a < 0
must hold for the Fourier transform of a put’s payoff to exist, we obtain the put-call
parity,
C = P + S e−rT − K e−rT .
CHAPTER 2

Mathematical tools

2.1. The Fourier transform


The Fourier transform is formally defined by

Definition 2.1. Let f : Rn → R be a function. The operator


Z
1
F : f → fˆ = 1
n
ei(p,x) f (x)dn x
(2π) 2 R n

is the Fourier transformation operator and the operator


Z
1
F −1 : f → f˜ = 1 e−i(p,x) f (x)dn x
(2π) 2 n Rn

is the inverse Fourier transformation operator whenever the integrals exist. (p, x)
denotes the Euclidian scalar product in Rn .

Given this formal definition, we have (i) to analyze the function space for f
and (ii) to derive properties of the Fourier transform in the chosen function space.

2.1.1. L1 -Theory. We consider L1 and in parts the L2 Fourier transform


p
theory.
R The set L , 1 ≥ p < ∞, consists of all p-Lebesgue integrable functions,
i.e. R |f (x)|p µ(dx) < ∞. The norms are denoted || • ||p . We consider often the
one-dimensional case below for simplicity.

Lemma 2.2. If f ∈ L1 (Rn ) then fˆ and f˜ exist and are uniformly continuous.
Furthermore the bound

|f˙(p)| ≤ (2π)− 2 n ||f ||1


1
(2.1)

holds where f˙(p) represents fˆ or f˜.

Proof. Let BR the ball in Rn with radius R and χBR (x) its characteristic func-
tion. We set fR (x) := f (x)χBR (x). We have |fR (x)| ≤ |f (x)| and fR (x) converges
R
Rto f (x) pointwise. By the dominated convergence theorem we get limR→∞ fR (x)dx =
f (x)dx and for every  > 0 there exists an R such that
Z

|f (x)|(1 − χBR (x))dx < .
4
27
28 2. MATHEMATICAL TOOLS

We bound
Z
1
|fˆ(p) − fˆ(q)| = 1 | (ei(p,x) − e−i(q,x) )f (x)dx|
(2π) 2 n Rn
Z
1
≤ 1 |ei(p,x) − ei(q,x) | |f (x)|dx
(2π) 2 n Rn
Z
1  (x, p − q)
= 1 2| sin( )||f (x)|dx
(2π) 2 n
BR 2
Z 
+ |ei(p,x) − ei(q,x) | |f (x)|dx
Rn \BR
Z Z
1  
≤ 1 2|p − q| |x| |f (x)|dx + 2|f (x)|dx
(2π) 2 n BR Rn \BR
1    1
≤ 1 |p − q| R ||f (x)||1 + ≤√ 
(2π) 2 n 2 2π

if |p−q| ≤ 2R||f ˆ ˜
||1 . This proves uniform convergence of f . For f the same argument
applies. The bound (2.1) is trivial. 

Lemma 2.2 shows that for f ∈ L1 (Rn ) the Fourier transform is in L∞ (Rn ) and
that it is uniformly continuous. Unfortunately the operators F and F −1 do not
map L1 into itself and therefore
f = F −1 Ff
does not hold generally in L1 .

Example 2.3. Consider the indicator function f = χ[−a,a] , a > 0, on R.


Although f ∈ L1 ,
r
2 sin(ap)
fˆ(p) = / L1 .

π p
Example 2.4. (Scaling) The Fourier transform of f (λx), with λ ∈ R+ , is
 
1 k
F (f (λx)) = fˆ .
λ λ

Proof. Set the change of variable λx = s → dx = λ1 ds, so that


Z ∞
1
F (f (λx)) = √ eikx f (λx) dx
2π −∞
Z ∞
1 1 iks
= √ e λ f (s) ds
λ 2π −∞
 
1ˆ k
= f .
λ λ


Example 2.5. (Shifting) The Fourier transform of f (x + c), with c ∈ R, is


F (f (x + c)) = e−ikc fˆ(k).
2.1. THE FOURIER TRANSFORM 29

Proof. Set the change of variable y = x + c → dy = dx, so that


Z ∞
1
F (f (x + c)) = √ eikx f (x + c) dx
2π −∞
Z ∞
1
= √ eik(y−c) f (y) dy
2π −∞
= e−ikc fˆ(k).


Example 2.6. We consider the Fourier transform of Gaussian functions in R,


1 2
i.e. we consider f (x) = e− 2 x . To calculate the Fourier transform, we differentiate
Ff (p) w.r.t. p, partial integrate the result and obtain the ODE1
dfˆ(p)
= −k fˆ(p) .
dp
Since Ff (0) = 1, the solution of the ODE with this initial condition follows
1 2
(2.2) fˆ(p) = e− 2 p .
Therefore, the Gaussian function is an eigenfunction of the Fourier transform with
1
eigenvalue +1. The general case of a Gaussian function f (x) = e− 2 (x,Ax) in Rn
with A a symmetric, positive n × n matrix follows from the usual factorization
calculation where the matrix A is diagonal. The result is
 1  1 1
F e− 2 (x,Ax) (p) = (det A)− 2 e− 2 (p,Ap) .

Example 2.7. (Symmetry) The Fourier transform of a real, even function is


itself a real, even function; whereas, the Fourier transform of a real, odd function
is a complex, odd function.

Proof. Consider an even function fe (x)2 and an odd function fo (x)3, for which
Z a Z a Z a
fe (x) dx = 2 fe (x) dx and fo (x) dx = 0, a ∈ R,
−a 0 −a

respectively. Rewrite the Fourier transform as


Z ∞ Z ∞
ˆ
f (k) = cos(kx)f (x) dx + i sin(kx)f (x) dx .
−∞ −∞
| {z } | {z }
A B

For f (x) = fe (x) the imaginary part B is zero, so fˆ(k) is real and even; for f (x) =
fo (x) the real part A is zero, so fˆ(k) is complex and odd. 

Although the Fourier transform in L1 is not bijective, nevertheless the following


inversion theorem holds:

1The dominated convergence theorem of Lebesgue allows us to interchange the derivation


limit with the integral of the Fourier transform.
2A function is even when f (x) = f (−x).
3A function is odd when −f (x) = f (−x).
30 2. MATHEMATICAL TOOLS

Theorem 2.8. If fˆ, f in L1 , then


F −1 (Ff ) = f
and similar if F −1 f, f in L1 , then
F(F −1 f ) = f .
We immediately get the corollary,
Corollary 2.9. (1) If fˆ = 0, f ∈ L1 , then f = 0.
(2) If f , f ∈ L , then fˆ, f ∈ L2 and the Plancherel formula holds
ˆ 1

(2.3) ||f ||2 = ||fˆ||2


with || • ||2 theR L2 -norm.
(3) If f ? g(x) = R f (x − y)g(y)dy denotes the convolution of f and g with
f, g ∈ L1 , we get
(2.4) F(f ? g) = (2π)1/2 fˆĝ .
Proof. We prove the convolution property. For f, g ∈ L1 , we define their
convolution f ∗ g by
Z ∞
h(x) = (f ∗ g)(x) = f (x − y)g(y) dy.
−∞

We then have
Z ∞ Z ∞
1
ĥ(k) = √ eikx f (x − y )g(y) dy dx
2π −∞ −∞ | {z }
z
Z ∞ Z ∞
1
= √ eik(z+y) f (z)g(y) dy dz
2π −∞ −∞
Z ∞ Z ∞
1
= √ eikz f (z) dz eiky g(y) dy
2π −∞ −∞

= 2π fˆ(k)ĝ(k),
by Fubini’s theorem. 
The Fourier transform maps decay properties into regularity properties of func-
tions and vice versa. The following theorem makes the ideas precise. We define
C n (X) the space of functions which are n times differentiable with continuous
derivatives. We set f (k) for the k-th derivative.
Theorem 2.10. Let f ∈ C n (R) and f (k) ∈ L1 (R) for k = 0, 1, . . . , n. Then for
all k, we have
 
d
(2.5) F (i )k f (p) = pk fˆ(p)
dx
and lim|p|→∞ |pn fˆ(p)| = 0. Conversely, let f be continuous and xn f ∈ L1 (R).
Then, the Fourier transform fˆ is n times continuously differentiable and for all
k = 0, 1, . . . , n:
d  
(2.6) (i )k fˆ(p) = F xk fˆ (p)
dp
2.1. THE FOURIER TRANSFORM 31

Proof. We only proof the first part for n = 1. We write


Z x
f (x) = f (0) + f 0 (y)dy .
0
Since f 0 ∈ L1 , the right hand side of above equation has a limit for x → ±∞ which
has to be 0 since f ∈ L1 . From the definition of the Fourier transform follows
Z
fˆ0 (p) = (2π)− 2
1
eipx f 0 (x)dx
R
Z
− 12
= −(2π) ip eipx f (x)dx
R
= −ipfˆ(p)
where we did a partial integration and used the fact, that the boundary terms are
zero. 
2.1.2. Complex Fourier and Laplace Transform. We next consider prop-
erties of the Fourier transform in the complex domain.
Theorem 2.11. Let f (z), z = x + iy be analytic in the strip
S := {z ∈ C| y− < y < y+ , y− < 0 < y+ } .
If for any strip S̃ ⊂ S
(
aeτ− x for x → ∞, τ− < 0,
|f (z)| →
beτ+ x for x → −∞, τ+ > 0
then fˆ(p) with p = σ + iτ will be analytic in the strip
F := {p ∈ C| τ− < τ < τ+ , τ− < 0 < τ+ }
and in any strip F̃ ⊂ F
(
cey− σ for σ → ∞,
|fˆ(p)| → .
dey+ σ for σ → −∞
a, b, c, d are real constants.
The theorem states that if a function is analytic in a strip and decays expo-
nentially in both direction on the real line, then there exists an analytic Fourier
transform in a new strip with the same exponential decay behavior. The decay rate
of the transformed function is given by the strip width of the original one: The
wider the original strip is the greater is the decay rate for the Fourier transformed
function.

Proof. Analyticity of fˆ is determined by the convergence of the defining in-


tegral:
Z ∞ Z ∞
ˆ
f (p) = (2π)− 12 ipx
f (x)e dx = (2π) − 12
f (x)e−τ x eiσx dx .
−∞ −∞

Uniform convergence at the upper limit of integration requires that eτ− x−τ x decays
exponentially with x so that τ− < τ and vice versa at the integration limit −∞.
For the behavior of fˆ(p) in the strip τ− < τ < τ+ we consider the convergence of
the integral defining fˆ. Since f is analytic, we may deform according to Cauchy’s
32 2. MATHEMATICAL TOOLS

Theorem - see the Appendix on Complex Analysis - the path of integration parallel
to the x axis as a long as we remain within the strip of analyticity, i.e. as long as
y− < y < y+ . We get
Z ∞ Z ∞
fˆ(p) = (2π)− 2
1 1
f (x)eipx dx = (2π)− 2 e−σy f (x)e−τ x ei(σx−τ y) dx .
−∞ −∞

The region of analyticity of f then implies


|fˆ(p)| → constant × e−σy (|σ| → ∞) .
If y is close to y+ , convergence only occurs for σ > 0 and vice versa for y close to
y− . 

It can arise in applications that the function f which has to be transformed does
not belongs to a space where the transformation is defined. For example, the payoff
of a call option is not in L1 and we therefore had to extend the Fourier transform to
the complex numbers. Another possibility, which also leads to complex integrals,
is the insertion of a convergence enforcing factor. This approach was chosen in the
papers of Carr and Madan, see the Section ”Synthesis of Option Pricing” for details.

/ L2 , like a payoff of a European call option. Then we


Assume that, say f ∈
consider the function
(2.7) g(x) = f (x)e−τ0 x τ0 > 0 ,
which we assume to be square integrable. Hence, the Fourier theorems may be
applied to the function g. The Fourier inversion then reads
Z ∞
−τ0 x 1
(2.8) g(x) = f (x)e =√ ĝ(p)e−ipx dp ,
2π −∞
or equivalently
Z ∞
1
f (x) = √ ĝ(p)e−i(p+iτ0 )x dp .
2π −∞

Setting p + iτ0 = ξ it follows that p = ξ − iτ0 and the integration over the real axis
becomes an integration over a complex line parallel to the real axis crossing the
imaginary axis at iτ0 , i.e. we get
Z ∞+iτ0
1
f (x) = √ ĝ(ξ − iτ0 )e−iξx dξ .
2π −∞+iτ0
Since also
Z ∞
1
ĝ(ξ − iτ0 ) = fˆ(ξ) , ĝ(p) = √ f (x)e−τ0 x+ipx dx
2π −∞

hold, we finally get the simple inversion formula


Z ∞+iτ0
1
(2.9) f (x) = √ fˆ(p)e−ipx dp .
2π −∞+iτ0
Therefore, calculating non integrable Fourier transforms of options implies shifting
of the real axis integral into the complex plane. Sometimes, a single convergence
2.1. THE FOURIER TRANSFORM 33

enforcing factor can not enforce convergence on the whole x domain. If this happens,
we proceed as follows. We define
 (

f (x)e−τ0 x for x > 0 ,
f+ (x) = τ0 > 0


0 for x < 0.

0
 for x
(> 0
(2.10) f− (x) = −τ x x<0 ,

f (x)e 1 for
τ1 < 0
where τ0 (τ1 ) is the minimum (maximum) possible value such that f+ (f− ) converge.
The corresponding Fourier transforms then read
Z ∞ Z ∞
1 1
fˆ+ (p) = √ f (x)eipx dx = √ f (x)eiσx−τ x dx , τ > τ0
2π 0 2π 0
Z 0 Z 0
1 1
(2.11) fˆ− (p) = √ f (x)eipx dx = √ f (x)eiσx−τ x dx , τ < τ1
2π −∞ 2π −∞
with p = σ + iτ . Applying the Fourier inversion formula and adding both resulting
expressions we obtain finally,
Z ∞+iτ0 Z ∞+iτ1
1 ˆ −ipx 1
(2.12) f (x) = √ f+ (p)e dp + √ f− (p)e−ipx dp .
2π −∞+iτ0 2π −∞+iτ1
With the same discussion as above, the analyticity of fˆ+ , fˆ− follows: For example
fˆ+ is analytic in a region in the upper half complex plane of p = σ + iτ , where the
region starts at iτ0 . fˆ− is analytic in the lower half plan τ < τ1 .

Example 2.12. Consider the function f (x) = e|x| . Clearly, this function is not
integrable. Then
 (  (

e|x| e−τ0 x for x > 0 , 
ex−τ0 x for x > 0 ,
f+ (x) = τ0 > 0 = τ0 > 0

 

0 for x < 0. 0 for x < 0.
Therefore τ0 = 1+,  > 0 small, defines τ0 . Similar, τ− = −1− follows. Therefore,
fˆ+ (fˆ− ) should be analytic in the region τ > 1 + (τ < −1 − ). To see this we
calculate Z ∞
1 1 1
fˆ+ (p) = √ ex eipx dx = − √ , τ >1
2π 0 2π 1 + ip
and similar for fˆ− (p). Therefore, f+ has a singularity at p = i and is analytic
above this singularity. When we consider option pricing under time changed Lévy
processes, the Laplace transform become important.
Definition 2.13. Let f be a real value piece wise continuous function. The
Laplace transform Lf is defined by
Z ∞
(2.13) Lf (p) = f (x)e−px dx .
0
34 2. MATHEMATICAL TOOLS

We show that the Laplace transform is a special case of the Fourier transform.
Consider the Fourier transform of
(
f (x) for x > 0,
(2.14) f+ (x) = .
0 otherwise

This transform is
Z ∞
1
(2.15) fˆ+ (p) = √ f (x)eipx dx .
2π 0

Therefore

(2.16) 2π fˆ+ (ip) = Lf (p) .

Hence all properties of fˆ+ can be used to obtain the properties of the Laplace
transform, since the Laplace transform can be seen as a complex Fourier transform.
The inversion formula for fˆ+ for example implies for the inversion of the Laplace
transform:
Z i∞+τ0
1
(2.17) L−1 f (x) = f+ (x) = Lf (p)epx dp , <x > 0 .
2π −i∞+τ0

Example 2.14. Consider the function f (x) = eiqx . The Laplace transform
then reads
1
Lf (p) = .
p − iq
The inversion transform is
Z i∞+τ0
−1 1 epx
L f (x) = f+ (x) = dp , <x > 0 .
2π −i∞+τ0 p − iq

The evaluation is an exercise in residue calculus. First, there is a unique pole at


p = iq of order 1. Second, the integration is on a straight line parallel to the
imaginary axis which crosses the real axis at τ0 > 0. We apply the residue theorem
to this situation, where we choose as a closed contour semi circles which for x > 0
is in the left half plane. For x < 0, the semi circle is in the right half plane. The
reason to choose the semi circle in this way is due to the fact, that for the semi circle
radii tending to infinity the integrals vanishes and we are only left with the integral
along the complex line which we care about. Carry out the details! Next, the left
plane circle enclose the pole while the right hand one is free of any singularities.
Therefore, for x < 0 Cauchy’s theorem implies

L−1 f (x) = f+ (x) = 0 , x < 0 .

For the other semi circle the residuum is equal to eiqx and applying the residue
theorem we get

L−1 f (x) = f+ (x) = eiqx , x > 0

which prove that we get back the original function.


2.2. COMPLEX ANALYSIS 35

2.2. Complex analysis


2.2.1. Complex numbers & definitions. Complex analysis is the analysis
on the complex numbers field. Basic operations and identities for complex numbers
are given next, where z ∈ C, x, y ∈ R:
z = x + iy, Cartesian representation of complex numbers.
z̄ = x − iy, conjugate complex number.
1
<z = (z + z̄), real part.
2
z − z̄
=z = , imaginary part.
2i
|z|2 = z z̄, norm.
If,
p instead of the Cartesian coordinates x, y, we use the polar coordinates r =
x1 + y 2 and the angle φ, we can represent the complex number z = x + iy in the
polar form
z = r eiφ .
Using Euler’s formula
eiφ = cos φ + i sin φ,
a trigonometric representation of z follows. The angle φ is often called the argument
of z, written φ = arg z. Since φ and φ + 2π lead to the same complex number, the
polar form is not unique. To obtain uniqueness, the angle or argument is often
restricted. For example, φ ∈ (−π, π] leads to the so-called principal value of z
(where z 6= 0). We can calculate arg z as follows:
1 − sign(x)
arg z = arctan(y/x) + π .
2
Figure 2.1 illustrates the argument function (left).
2.2.2. Power functions and complex logarithms. Crucial functions in
option pricing are the functions z n with n a natural number and log z, i.e. the
complex logarithm. These functions are not well-defined and continuous on the
whole complex domain but only on a subset. If we do not understand the properties
of these functions and apply them carefully in option pricing, their discontinuities
will lead to meaningless option prices; i.e. we would observe jumps and oscillations
in the pricing formulas. √
To gain some intuition, let us consider the function f (z) = z. Using polar
coordinates z = r eiφ and f (z) = R eiθ , it follows that

R = r, θ = φ/2.
The phases φ and φ + 2π are the same. What happens for the square root function
f ? We get
√ 1 √ 1
f1 ((r, φ)) = r e 2 iφ , f2 ((r, φ + 2π)) = − r e 2 iφ .
Hence, the square root function is multivalued and not continuous, i.e. for z = r,
we get √ √
f1 (r) = i r, f2 (r) = −i r.
This is well-known since we know that the square root of a number is a plus or a
1
minus value. The function f (z) = z n is evidently n-valued, as is also is illustrated
in figure 2.2; the z-plane (left) and the f (z)-plane (right).
36 2. MATHEMATICAL TOOLS

Figure 2.1. Illustration of arg z

Consider an almost closed circle in the z-plane with arbitrary radius r. Then
starting at (−r, π+), the value f (−r, π+) is considered in the image plane. Moving
in the indicated direction along the almost closed circle, the almost semicircle in
the f (z)-plane follows. Having arrived at (−r, π + 2π − ) we next move on and
cross the negative horizontal axis, i.e. we consider the point (−r, π + 2π + ). But
the value of this point under f is not equal to the point under (−r, π + ) but
mirrored at the imaginary axis. Continuing, the second almost semicircle follows.
Having arrived at (−r, π + 4π − ) and crossing the real axis again, i.e. considering
(−r, π + 4π + ), the value of this point equals the value of (−r, π +√).
Therefore, each time we cross the real axis, the function f (z) = z is discontin-
uous and changes to one of the two possible almost semicircles. So the negative real
axis has to be excluded if we want f to be a continuous and analytic function. This
implies that −π < φ < π is a possible range for f to be single-valued and the other
one is π < φ < 3φ.√ The two independent values of f on these two sets are called
the functions of z. The line along which discontinuities occur, φ = π is called a
branch line. Since the values of (−r, π + ) and (−r, π + ) for arbitrary small  are
not the same under the square root function, a singularity must be enclosed. Also,
since the radius of the semicircles was arbitrary, the singularity must be at z = 0.
Such a singularity is called a branch point.
This intuition generalizes to
z n = w = R(cos ψ + i sin ψ)
2.2. COMPLEX ANALYSIS 37


Figure 2.2. Illustration of f (z) = z

with n ∈ N. We set
z = r(cos φ + i sin φ).
From
rn (cos nφ + i sin nφ) = R(cos ψ + i sin ψ)
we obtain the two equations
z n = R, nφ = ψ (mod2π).
The first equation has the unique solution z = R1/n . The second equation has n
solutions
ψ 2πm
φ= + (mod2π), m = 0, . . . , n − 1.
n n
Therefore the n solutions of the original equation are
  
ψ 2πm ψ 2πm
z = R1/n cos( + ) + i sin + m = 0, . . . , n − 1.
n n n n
Hence, if ψ passes through an interval (a, a + 2π) and if m is fixed number in
0, . . . , n − 1, then the corresponding solution of z n = w is called the m-th branch
of the n-th root of all complex numbers w with
w = R1/n (cos ψ + i sin ψ) , R > 0, a < ψ < a + 2π.
See figure 2.3.
38 2. MATHEMATICAL TOOLS

Figure 2.3. Illustration of z n = w for m = 1, a = 3π/4


To summarize, there is in fact no way to choose a so that z is continuous for
all complex values of z. There has to be a ’branch cut’ line in the complex plane
across which the function is discontinuous.
Example 2.15. Consider z = r eφ , where t ∈ [−π, π), i.e. we cut the complex
plane along the negative axis and take the power
z a = ra eaφ .
Since 
eiaπ 6= e−iaπ , a ∈
/ Z;
eiπ = e−iπ ⇒
eiaπ = e−iaπ , else.
So when we cross the negative real axis, a jump follows.
The next equation is
ez = w = R(cos ψ + i sin ψ).
If z, w were real numbers, there would be a unique solution: the real logarithm.
But for z, w complex numbers, uniqueness is basically lost since
e2πin = 1.
Using ez = ex+iy = ex eiy , one obtains the two equations
ex = R , y = ψ(mod2π).
2.2. COMPLEX ANALYSIS 39

Figure 2.4. Illustration of ez = w for m = −1, a = 3π/4

The unique solution of the first equation is x = ln R. The second equation possess
an infinite number of solutions
y = ψ + 2πm, m ∈ Z.
So the infinite solutions are
z = ln R + i(ψ + 2π m), m ∈ Z.
Hence, if ψ passes through an interval (a, a + 2π) and if m is fixed integer, then
z = log w = ln R + i(ψ + 2π m)
is the m-th branch of the logarithm of all complex numbers w with
w = R(cos ψ) + i sin(ψ)), R > 0, a < ψ < a + 2π.
The principal branch occurs at m = 0. This means that the exponential func-
tion w = ez has an inverse function on the domain
{z ∈ C : −π < =z < π}.
On this set, the exponential function is injective and the image is
{w ∈ C : =w 6= 0, <w > 0}.
The inverse function, the principal value of the logarithm, is given by
ln z = ln |z| + i arg(z), −π < arg(z) < π.
40 2. MATHEMATICAL TOOLS

The solution for the branch m is then


ln z = ln |z| + i(arg(z) + 2πm), −π < arg(z) < π.
We need a few definitions from topology at this stage. We write B(ξ, ) for the disk
in the complex plain C with center ξ and radius . If M is a set, the closure of
M is denoted M̄ and the boundary ∂M , respectively. Curves γ(t) are continuous
functions from [a, b] → C where t runs through [a, b]. A set M is connected as
subsets of C, if any two points z, w ∈ M can be connected by a curve γ(t) ∈ M . A
set M is simply connected if any closed contour in M can be shrinked to a point.
2.2.3. Holomorphic functions. The first object we analyze are holomorphic
functions
f : C → C, f (z) = w.
We decompose
z = x + iy, w = u + iv, x, y, u, v ∈ R.
and assume that the real functions
u = <f (z), v = =f (z)
are once continuously differentiable, i.e. C 1 . With this assumption, the partial
derivatives in the differentials
du = ux , dx + uy dy, dv = vx dx + vy dy
are continuous functions. Defining the complex differential
dz = dx + i dy,
we get lemma 2.16:
Lemma 2.16. Let dz = dx + i dy and w = u + iv = f (z). Then the complex
differential dw can be written as:
1 1
dw = (ux + vy + ivx − iuy ) dz + (ux − vy + ivx + iuy ) dz̄.
2 2
Proof. We leave the proof as an exercise for the reader. 
Definition 2.17. A function f : C → C, f (z) = w is holomorphic if dw
depends only on dz.
Therefore, a function is holomorphic if the Cauchy-Riemann differential equa-
tions
(2.18) ux = vy , vx = −uy
hold. If the derivatives in (2.18) are continuous, then the Cauchy-Riemann equa-
tions are also a sufficient condition for holomorphicity. The notion of holomorphicity
makes complex analysis an independent mathematical discipline, since this notion
is unknown in real analysis but is a basic one in complex analysis. For holomorphic
functions,
dw = (ux + ivx ) dz = (vy − iuy ) dz;
the quotient
dw
= f 0 (z)
dz
is called the complex derivative of f .
2.2. COMPLEX ANALYSIS 41

These preparations allow us to consider the main theorems of basic complex


analysis. A major property of complex analysis is the possibility to calculate inte-
grals which cannot be calculated over the reals or where a calculation is difficult.
We consider integrals along a curve γ(t) ∈ C
Z Z Z
f (z) dz = (u dx − v dy) + i (v dx + u dy) .
γ γ γ
The most basic theorem is the integral theorem of Cauchy.
Theorem 2.18 (Integral theorem of Cauchy, 1825). Let E ⊂ C be simply
connected with a closed contour γ and f an arbitrary holomorphic function defined
on Ē. Then I
f (z) dz = 0
γ
H
where indicates that integration is along a closed curve.
Proof. The proof follows from Green’s theorem in the plane. Let P (x, y), Q(x, y)
be two continuously differentiable function on a simply connected domain E with
γ a closed curve defining the boundary of E. Then
I Z Z
P (x, y) dx + Q(x, y) dy = (Qx (x, y) − Py (x, y)) dx dy.
γ E

Using that f is holomorphic in Ē, Green’s formula, and the Cauchy-Riemann dif-
ferential equations (2.18), we get
I Z Z
(u dx − v dy) = − (vx (x, y) + uy (x, y)) dx dy = 0
γ E
I Z Z
(v dx + u dy) = (ux (x, y) − vy (x, y)) dx dy = 0.
γ E

R
Corollary 2.19. One may deform a contour γ in the γ f (z dz without chang-
ing the value of the integral provided the contour variation does not cross any sin-
gularities of the function f .
Proof. We leave the proof as an exercise for the reader. 
In Cauchy’s theorem ??Hthe simple connectivity of the domain is essential. If
this assumption is not met, γ f (z) dz is in general different from 0. To apply the
theorem in these cases, one proceeds as shown in figure 2.5. In panel B,R the contour
is in a simply-connected domain and Cauchy’s theorem applies, i.e. γ1 f (z) dz +
R
γ2
f (z) dz = 0 since the two integral over the straight lines are the same in absolute
value but they have opposite direction, thus canceling each other out.
The other two main theorem considers the case where f is holomorphic except
at the point ξ1 ∈ C, which is called a singularity. See figure 2.6, which considers
the disk with radius  and center ξ and a closed curve γ = P SP QRQP .
We get from Cauchy’s theorem
I I
f (z) dz = f (z) dz
γ ∂B(ξ,−)
H
since 0 = γ f (z) dz (there are no enclosed singularities) and since the two integrals
P Q and QP are the opposite of each other and hence add to zero. Note that
42 2. MATHEMATICAL TOOLS

Figure 2.5. Application of Cauchy’s theorem on non-simply con-


nected domains

contours are always oriented counterclockwise; so we write −. The generalization


to N singularities ξ1 , . . . , ξN with the associated disks B(ξj , −j ), j = 1, . . . , N is
obvious, i.e.
I N I
X
(2.19) f (z) dz = f (z) dz,
γ j=1 ∂B(ξj ,−j )

if all  have the same orientation as γ and all disks are disjoint. Therefore, to
calculate an integral enclosing singularities, (i) choose a disk around each singularity
such that the disks are disjoint and (ii) integrate around the circles circumference.
Next assume that f has only one singularity at ξ1 = 0. Then
f (z)
w=
z−ξ
can have singularities at most at z = ξ and z = 0. Consider two non intersecting
disks B(ξ, ) ∩ B(ξ1 , 1 ) = ∅ with γ, , 1 being counter clockwise oriented. Then
(2.19) implies
I I I
f (z) f (z) f (z)
(z) dz = dz + dz.
γ z−ξ ∂B(ξ,) z − ξ ∂B(ξ1 ,1 ) z − ξ

The integral over the first disk is parameterized as follows:


(2.20) z(t) = ξ + eit , 0 ≤ t ≤ 2π.
2.2. COMPLEX ANALYSIS 43

Figure 2.6. The second main theorem in complex analysis

This implies
I I 2π I 2π
f (z) f (ξ +  eit )
dz = i eit dt = i f (ξ +  eit ) dt.
∂B(ξ,) z−ξ 0 ξ +  eit − ξ 0

For  → 0, f (ξ +  eit ) → f (ξ) uniformly. The limits in the following calculation


can be interchanged:
I I 2π I 2π
f (z) it
dz = i lim f (ξ +  e ) dt = i f (ξ) dt = if (ξ)2π,
∂B(ξ,) z − ξ →0 0 0

so that
I I !
1 f (z) f (z)
(2.21) f (ξ) = (z) dz − dz .
2πi γ z−ξ ∂B(ξ1 ,1 ) z−ξ
This result allows us to consider two cases. First, assume that f has no singu-
larity in z = ξ1 = 0. Then (2.21) proves the second theorem of Cauchy:
Theorem 2.20 (Integral formula of Cauchy, 1831). Let E ⊂ C be simply con-
nected with a closed and positive oriented contour γ and f a holomorphic function
defined on Ē. Then, for all ξ ∈ E
I
1 f (z)
(2.22) f (ξ) = dz.
2π γ z − ξ
44 2. MATHEMATICAL TOOLS

Suppose that f has also a singularity at z = 0 and assume positive oriented


circles all centered at 0 with radii r, R such that
r < r 1 < R1 < R
where r1 < |ξ| < R1 . We distinguish two cases.
The first case is when z is on the R-circle. Then, |ξ/z| < 1 for all z on the
outer circle with radius R. Therefore, we get for the series
 N +1
X∞  n 1 − ξ
ξ z 1
1/z = lim =
n=0
z N →∞ z − ξ z − ξ
where the convergence is uniform. Hence,
I X∞ I
f (z) f (z) n
dz = ξ dz, z ∈ ∂B(0, R), r1 ≤ |ξ| ≤ R1
∂B(0,R) z − ξ n=0 ∂B(0,R)
z n+1
again with uniform convergence for all r1 ≤ |ξ| ≤ R1 .
Secondly, consider the case where z is on the r-circle. Similarly, for z in the inner
P∞  z n
disk with radius r, |z/ξ| < 1 and the series 1/ξ n=1 ξ converges uniformly to
−(z − ξ)−1 . This implies
I ∞ I
X X∞ I
f (z) 1
− dz = f (z)z n dz n+1 = f (z)z n−1 dzξ −n .
∂B(0,r) z − ξ n=0 ∂B(0,r) ξ n=0 ∂B(0,r)

We arrive at the following theorem of Laurent where ξ1 is arbitrary, i.e. not


equal to zero as in the above calculations.
Theorem 2.21 (Laurent series theorem). Let f be holomorphic on B(ξ1 , R),
with the possible exception z = ξ1 , then f possesses for all ξ 6= ξ1 a Laurent series
expression
X∞
(2.23) f (ξ) = an (ξ − ξ1 )n
n=−∞

in B(ξ1 , R). The coefficients an are given by the formulae


I
1 f (z)
(2.24) an = dz, n = 0, 1, 2, . . . .
2πi ∂B(ξ1 ,R) (z − ξ1 )n+1
I
1 n−1
a−n = f (z) (z − ξ1 ) dz, n = 0, 1, 2, . . . .
2πi ∂B(ξ1 ,r)
The Laurent series converges uniformly for all ξ : r1 ≤ |ξ − ξ1 | ≤ R1 .
A Laurent series has therefore two parts: one part with powers ∼ z n , which
is the Taylor series one and is free of singularities, and another part with powers
∼ z −n , which is called the principal part of f around ξ1 . More explicitly, a Laurent
series around a singularity at ξ1 = 0 of a function f reads
a−n a−1
f (z) = . . . + n + . . . 1 +a0 + a1 z + . . . an z n + . . . .
| z {z z }
Principal part

We state and derive some useful facts from these theorems. First, one often uses
the notion of analytic functions, i.e. functions which can be represented in each
2.2. COMPLEX ANALYSIS 45

point of their domain of definition by a power series which converges uniformly.


One can then prove that
f analytic ⇐⇒ f holomorphic.
We continue to use the term ‘analytic’ since it is commonly used in the finance
literature.
If the principal part of a Laurent series consists only of finite many terms, ξ1
is called a pole and the degree of the pole, P , is the largest number n such that
a−n 6= 0. For pole-singularities, the powerful residue theorem follows and allows
us to calculate integrals. If the Laurent series’ principal part consists of an infinite
number of non-zero terms, ξ1 is called an essential singularity.
The last theorem we consider is the residue theorem. Let ξ1 be a pole of degree
P of an otherwise analytic function f defined on a domain Ē ⊂ C. We then have
P
X
(2.25) f (z) = a−n (z − ξ1 )−n + g(z)
n=1

with g the analytic part of f . We define a positively-oriented disk B(ξ1 , ) and,


using the parametrization as in (2.20), we get
I P
X I I
f (z) dz = a−n (z − ξ1 )−n dz + g(z) dz
∂B(ξ1 ,) n=1 ∂B(ξ1 ,) ∂B(ξ1 ,)
P
X I 2π
= a−n i1−n ei(1−n)t dt
n=1 0

= a−1 2πi,
since all integrals with n > 1 vanish. Only the coefficient a−1 gives a non-zero
contribution to the integral. This coefficient is called the residue res(f, ξ1 ) of the
function f at the pole ξ1 . We have just proved the residue theorem.
Theorem 2.22 (Residue theorem of Cauchy, 1825). Let f be defined on the
simply connected domain Ē with a positive contour γ and with poles ξ1 , . . . , ξN as
the only singularities. Then
I N
X
(2.26) f (z) dz = 2πi res(f, ξn ).
γ n=1

Therefore, calculation of the integral is reduced to calculation of the residues.


These are obtained either by expanding f in a Laurent series or by computing the
following derivatives for a pole of order m:
1 dm−1
(2.27) a−1 = res(f, ξn ) = ((z − ξn )m f (z)) |z=ξn .
(m − 1)! dz m−1
In words, to calculate an integral inclosing poles,
(1) determine the poles;
(2) determine the order of the poles;
(3) calculate the residue of each pole;
(4) sum all the residues.
46 2. MATHEMATICAL TOOLS

As a first example, consider the Fourier integral


Z ∞
1 P (x)
(2.28) I=√ eipx dx , x ∈ R,
2π −∞ Q(x)
with P, Q two polynomials in x, p > 0, the degree of Q exceeding the degree of
P by at least 1 and where the roots of Q are assumed not to be on the real axis.
Using the residue theorem, we can write
Z R Z ! X
1 P (z)
(2.29) √ + eipz dxz = res ,
2π −R γR Q(z)
upper half plane

where γR is a semicircle with origin (0, 0) and radius R, such that all possible poles
of Q are surrounded by the semicircle. Hence,
(1) we write the integral over the reals as a complex integral, where
(2) we introduce a closed contour in the complex plain.
(3) We then apply the residue theorem to the closed contour and
(4) make sure that the part of the contour which is not on the real axis, i.e.
the semicircle in this example, has zero contribution to the integral.
(5) This provides us then with the value of the real integral which we are
interested in.
If we can prove that the semicircle integral has zero contribution to the integral
if R → ∞, then the integral calculation is reduced to a calculation of residues.
The proof that the semicircle integral has zero contribution follows from Jordan’s
lemma.
Lemma 2.23 (Lemma of Jordan). Let f be an analytic function except at a
singularity at zero and assume that the condition
lim f (R eiφ ) = 0
R→∞
holds uniformly for all 0 ≤ φ ≤ π. Then
I
lim eiz f (z) dz = 0
R→∞ γR

with γR the same semicircle as above.


Example 2.24. We apply Jordan’s lemma to the Fourier integral
Z ∞
1 1 1
F( 2 2
)(k) = √ eikx 2 dx, a 6= 0, k > 0.
x +a 2π −∞ x + a2
1
The integrand x2 +a 2 clearly satisfies the condition for Jordan’s lemma, and the

semi circle integral contribution around the poles has zero contribution. There are
two poles: ±|a|i. For the one in the upper half plane, the residue is
e−k|a|
a−1 = .
2i|a|
Hence,
Z ∞ √ −k|a|
1 1 1 πe
F( )(k) = √ eikx dx = √ .
x2 + a2 2π −∞ x2 + a2 2|a|
A similar calculation can be carried out for p < 0. This then gives us the result for
k ∈ R.
2.2. COMPLEX ANALYSIS 47

Example 2.25. Consider the integral


I
f (z) dz, z ∈ C,
γ

where
z 2 − 2z
f (z) =
(z + 1)2 (z 2 + 4)
and γ is a circle with radius 10 and center (0, 0). The poles are z = −1 of order 2,
z = ±2i of order 1. Hence, the circle encloses all poles. Calculating the reside we
get
res(f, −1) = −14/25, res(f, 2i) = (7 + i)/25, res(f, −2i) = (7 − i)/25.
The residue theorem then implies
I X
f (z) dz = 2πi res(f, k) = 0.
γ k

Example 2.26. Consider the integral


Z ∞
sin x
dx.
0 x
We replace the sin function using
=eiz = sin x, z = x + iy.
Since the pole x = 0 is on the real axis, we choose the contour as follows: Two semi
circles γ2 , γ1 with radius r and R respectively, both with center 0, are connected
by two straight lines on the real line such that the whole contour is closed. The
theorem of Cauchy implies
I
0 = f (z) dz
closed contour
Z −r ix Z Z R Z
e eiz eix eiz
= dx + dx + dx + dx .
−R x γ1 z r x γ2 z
Since, |eiz /z| ≤ 1/R, the γ1 -integral converges to zero for R → ∞. For the γ2 -
integral, we get
Z
lim f (z) dz
r→0 γ2
Z
1 + iz + 21 (iz)2 + . . .
= lim dz
r→0 γ
2
z
Z  
1
= lim dz + g(z) dz .
r→0 γ
2
z
The integral over the analytic function g converges to zero and the first integral
gives
Z Z Z
1 ireirφ
lim f (z) dz = lim dz = lim dφ = iπ.
r→0 γ
2
r→0 γ z
2
r→0 γ
2
reiφ
Summarizing,
Z 0 Z ∞
eix eix
dx + dx = iπ.
−∞ x 0 x
48 2. MATHEMATICAL TOOLS

R∞ xα
Figure 2.7. Contour for the integral 0 x+1 dx, 0<α<1

Taking the imaginary part, we get


Z ∞
sin x 1
dx = π.
0 x 2
Example 2.27. Consider the integral
Z ∞ α−1
x
I= dx, 0 < α < 1.
0 x +1
The integral has a so called branch, i.e. the equation z α = w has a non-unique
solution in the complex numbers. Restricting to one solution, i.e. choosing a
branch, defines a set in the complex plain where the solution is defined. We choose
as branch cut, i.e. the set where the solution is not defined, the positive real axis.
Therefore, we choose the contour as shown in figure 2.7.
The function has a pole at z = −1 with corresponding residue
zα α−1
res(f, −1) = lim (z + 1) = 2πi eiπ .
z→−1 z + 1

Furthermore, at z = 0, the function has a branch point type singularity. The


theorem of Cauchy implies
I
f (z) dz = eiπ(α−1) .
γ
2.3. GENERALIZED FUNCTIONS 49

Splitting the contour integral in four integrals, on the other hand, implies
Z Z 2π Z 0
(Reiφ )α−1 iφ (r eiφ )α−1
f (z) dz = iRe dφ + ir eiφ dφ
γ 0 R eiφ + 1 iφ
2π r e + 1
Z R α−1 Z r
x (x e2πi )α−1
+ dx + 2πi + 1
dx = 2πi eiπ(α−1) .
r x+1 R xe
For r → 0, R → ∞, the two circle integrals vanish and we get
Z ∞ α−1 Z 0
x (x e2πi )α−1
dx + dx
0 x+1 ∞ x+1
= eiπ(α−1) .
This implies
Z ∞
xα−1 2πi eiπ(α−1) π
dx = iπ(α−1)
= .
0 x + 1 1 − e sin απ
Unlike the case when the two integrals in the opposite direction cancel out, this is
not the case here where the two integrals follow from the multi-valuedness of the
integrand.

2.3. Generalized functions


2.3.1. Introduction. A generalized function f is a map X → R with X
a ‘nice’ function space such that f is continuous and linear. A powerful theory
arises if φ ∈ X is infinitely continuous differentiable and if φ, together with all
its derivatives, vanishes faster at infinity than any power function. If φ is such a
function, the action of f is written as
f (φ) = hf, φi.
Three questions matter:
(1) How can we interpret f (φ) = hf, φi?
(2) How can we define and intuitively understand generalized functions?
(3) Is there a calculus for generalized functions?
We start with the Dirac delta function f = δ defined by
hδ, φi = φ(0), φ ∈ X.
Hence, the delta function applied to a rapid decreasing gives the value of this
function at zero. This underlies a physicist’s intuition, for which the delta function
is a peak-type operation: A function which takes a very large value at a single point
x = 0 and is zero else. This is a model for a mass point in mechanics or a point
charge in electrostatics. In financial economics the same intuition applies and a
delta-function type payoff is an Arrow-Debreu security payoff in continuous state
space. It is a continuous analogue of the discrete Kronecker delta. Although the
following physicists’ integral description makes no mathematical sense, it is very
suggestive and can hardly lead to any errors. Physicists write
Z
hδ, φi = δ(x)φ(x) dx = φ(0),
R
although, here again, there is no such integrable function delta. The general idea
about the integral is as follows: A point mass is a mathematical point. Let φ(x) be
50 2. MATHEMATICAL TOOLS

a mass density. Then


Z
φ(0) = φ(x)δ(x) dx,
R

states that a point mass is the mass concentrated at a single point. The same anal-
ogy applies to Arrow-Debreu securities. Find economic examples for the function
φ.
Besides the delta function, other prominent generalized functions are functions
with jumps or kinks, such as the maximum (as for a call option payoff) or the
heaviside function.
Why are these generalized functions meaningful? Since φ are assumed to be
nice functions, then Dirac’s delta, the max-function, the step function turn out to
be differentiable in a generalized sense. For short, we can take the derivative of a
kink! How is this possible?

Definition 2.28. Let f be a generalized function and φ ∈ X. For the derivative



D = −i ∂x , one defines
hDf, φi = −hf, Dφi.
The Fourier transform F satisfies

hFf, φi = hf, Fφi.

Hence, we exploit the fact that operations - differentiation, Fourier transform -


on generalized functions f are always shifted to the functions φ ∈ X which can be
differentiated, Fourier transformed, etc.
Why is the definition meaningful? If the generalized function f can be repre-
sented by an integral, the definition of the derivative can be motivated using partial
integration, i.e.
Z Z
hDf, φi = Df (x)φ(x) dx = − w(x)Dφ(x) dx + boundary term
R R

where the boundary term vanishes since φ is a rapidly decreasing function. The
derivative formula generalizes to arbitrary higher order derivatives.

2.3.2. Examples.

Example 2.29. Consider the call payoff function

f (x) = max(x − k, 0).

We want to calculate
d d
f (x) = max(x − k, 0),
dx dx
which does not exist as a classical derivative at the kink x = k. We may interpret
f as a generalized function and calculate
Z
d
Q= φ(x) f (x) dx.
R dx
2.3. GENERALIZED FUNCTIONS 51

By definition,
Z
d
Q = φ(x) f (x) dx
R dx
Z
d
:= − φ(x)f (x) dx
dx
ZR
d
= − φ(x) max(x − k, 0) dx
R dx
Z ∞
d
= − φ(x)(x − k) dx
k dx
Z ∞
d
= −φ(x)(x − k)|±∞ + φ(x) (x − k) dx
k dx
Z ∞
d
= 0+ φ(x) (x − k) dx
dx
Z ∞ k
= φ(x)1 dx
k
Z
= φ(x)H(k) dx,
R

where the Heaviside or step function H is defined by



1, x ≥ k;
H(k) =
0, else.
So the derivative of the call payoff - viewed as a generalized function - is the heaviside
function at the strike k, i.e.
d d
f (x) = max(x − k, 0) = H(k)
dx dx
as generalized functions. This is exactly what one gets when drawing a figure.
Example 2.30. Consider the derivative of H(k) i.e. the second derivative of
the call payoff. We get
  Z
d d
H, φ (x) = φ(x) H(k) dx
dx R dx
Z
d
:= − φ(x)H(k) dx
dx
ZR∞
d
= − φ(x) dx
k dx
= −φ(x)|∞ k
= φ(k)
Z
= φ(x)δ(x − k) dx,
R

i.e.
H 0 (k) = δ(x − k).
You may also draw a figure to visualize this result.
52 2. MATHEMATICAL TOOLS

Example 2.31. Let us calculate the Fourier transform of f (x) = eiwx . Note
that f is not integrable, i.e. it is not in L1 , L2 . The proof is left as an exercise.
Z
iwx
hF e , φi(p) = φ(x)F eiwx dx
R
Z
:= Fφ(x) eiwx dx
R
Z
= F φ(x) eiwx dx
R
Z Z
1
= √ eiwx eipx φ(p) dp dx
2π R R
Z Z
1
= √ eiwx+ipx φ(p) dp dx
2π R R
Z Z 
1
= √ φ(p) eiwx+ipx dx φp)
2π R R
Z
1
= √ δ(w + k)φ(x) dx,
2π R
i.e. the Fourier transform is the Dirac delta function. The expression
Z
eiwx+ipx × 1 × dx
R
is the Fourier transform of 1 for the shifted value w + p; but the Fourier transform
of delta is 1 and, similarly, it follows that the Fourier transform of 1 is delta - hence
the claim follows.
Lemma 2.32.
1
Fδ = √ 1


F1 = 2πδ
ipx 1
Fe = √ δ(p + ·).

The last property of generalized functions we want to discuss is the convolution
property, defined by Z
f ? φ(x) = f (y)φ(x − y) dy.
R
This definition is extended to the case where f is a generalized function and φ ∈ X.
In particular, the Fourier transform property

F(f ? φ) = 2πFf Fφ.
This is a basic property used when solving PDEs.

We finally apply the method to generalize Itô’s Formula. For S(t) a diffusion,
i.e.
dS(t) = S(t)µ(t)dt + σ(t)S(t)dW (t) , S(0) = s ,
and f a twice differentiable function Itô’s Formula reads:
1 2
df (S, t) = ∂t f (S, t)dt + ∂S f (S, t)dS + ∂SS f (S, t)(dS)2 . .
2
2.3. GENERALIZED FUNCTIONS 53

If f is not differentiable, the above result is meaningless. But if we interpret the


derivatives in this case as generalized derivatives, then we can use the above formal
expression of Itô, which is then called Tanaka Formula. We illustrate this for a call
option C = f (S) = (S(T ) − K)+ . Using Tanaka’s Formula we get
1 2
dC = ∂S C(S)dS + ∂SS C(S)(dS)2 .
2
But ∂S C(S) = H(K) = χS(T )>K equals the step function and the second derivative
2
∂SS C(S) = δ(S(T ) − K) is a delta function. Therefore, representing Tanaka’s
Formula in integrated form we get
Z T Z
1 T
(S(T ) − K)+ = (S(0) − K)+ + χS(T )>K dS(t) + δ(S(s) − K)S(s)2 σ 2 (s)ds .
0 2 0
The first term represents the intrinsic value of the option at issuance, the second
one is a self-financing strategy where one holds one unit of the stock if it is in-the-
money and zero else and the third term measures the ”local time at the strike”. √
This term corrects for the scaling behavior of Brownian motion. Since dW ∼ dt
for small time intervals the Brownian motion changes faster than linear time does.
Hence, if the stock is below the strike and above strike the change in price and
the change in time are not on the same scale. Therefore, there is a hedging error
due to this fact which is measured as the time spend at the strike weighted by the
variance. We next consider the expected value of the local time expression, i.e.
"Z #
T
1 2 2
E δ(S(s) − K)S(s) σ (s)ds .
2 0

We write:
"Z # "Z #
T T  2 
2 2 2
A=E δ(S(s) − K)S(s) σ (s)ds = E δ(S(s) − K)S(s) E σ (s)|S(s) ds .
0 0

Using φ(s, x) for the risk-neutral density of the stock process at time s and stock
value x, we get
"Z #
T  2 
2
A = E δ(S(s) − K)S(s) E σ (s)|S(s) ds
0
"Z Z #
T ∞  
= E δ(S(s) − K)S(s) E σ 2 (s)|S(s) = x φ(s, x)dxds
2
0 −∞
"Z #
T  2 
2
(2.30) = E K E σ (s)|S(s) = K φ(s, K)ds .
0

Therefore,
"Z # "Z #
T T  
1 1
E δ(S(s) − K)S(s)2 σ 2 (s)ds = K 2 E E σ 2 (s)|S(s) = K φ(s, K)ds .
2 0 2 0

If σ is constant, then we get


"Z # "Z #
T T
1 2 2 1 2 2
E δ(S(s) − K)S(s) σ (s)ds = K σ E φ(s, K)ds .
2 0 2 0
Index

L1 Fourier theory, 27 Heston dynamics, 1


Heston model, 1, 10
Argument of a complex number, 15 holomorphic function, 40
Argument or a complex number, 35 Homogeneity property of Heston’s
equation, 3
Boundary condition, 6
Inverse Fourier transform, 5
Carr-Madan, 20
Cauchy Jordan
integral formula of, 43 Lemma of, 46
integral theorem of, 41
residue theorem of, 45 Kahl, Jaeckel, 16
Cauchy-Riemann differential equations, 40
Laurent series theorem, 44
Characteristic exponent, 16
Lebesgue integrable functions, 27
Characteristic function, 11, 19
Lee, 18
Complex analysis, 35
Lucic, 16
Complex Fourier and Laplace Transform,
Lukacs Theorem, 16
31
Complex Fourier integral, 6 Market Price of Risk, 3
Complex logarithm, 14 Martingale preserving, 10
complex logarithm, 35–40
Complex numbers & definitions, 35 Norm preserving, 10

Damping option prices, 20 Ordinary differential equation


Discounted characteristic function, 19 Riccati, 12
Duffie, Pan, Singleton, 14
Payoffs and their Fourier transform, 6
Euler’s formula, 35 Power functions and complex logarithms,
35
Fourier transform, 4, 5, 27 Principal branch, 15
convolution, 30
Gaussian functions, 29 Residue Theorem, 22, 25
Plancherel formula, 30 Riccati equation, 12
scaling property, 28 Rotation Count Algorithm, 16
shifting property, 28
Schouten et al., 14
symmetry property, 29
Singularities, 14
Fourier transform of a call option payoff, 6
Strip of regularity, 11, 19
Fourier transform, decay and regularity
properties, 30 Variance Gamma model, 16
Fundamental solution, 7 Volatility risk premium, 2
Gatheral, 14
Generalized Derivative, 53
Green function, 10

Heat equation, 8

55

You might also like