Fourier-Laplace Transforms in Polynomial Ornstein-Uhlenbeck Volatility Models

Fourier-Laplace transforms in polynomial
Ornstein-Uhlenbeck volatility models

Eduardo Abi Jaber∗1 , Shaun (Xiaoyuan) Li†2,3 , and Xuyang Lin‡1
1
Ecole Polytechnique, CMAP
2
AXA Investment Managers
3
Université Paris 1 Panthéon-Sorbonne, CES
arXiv:2405.02170v1 [q-fin.MF] 3 May 2024
May 6, 2024
Abstract
We consider the Fourier-Laplace transforms of a broad class of polynomial Ornstein-

Uhlenbeck (OU) volatility models, including the well-known Stein-Stein, Schöbel-Zhu, one-
factor Bergomi, and the recently introduced Quintic OU models motivated by the SPX-VIX
joint calibration problem. We show the connection between the joint Fourier-Laplace func-
tional of the log-price and the integrated variance, and the solution of an infinite dimensional
Riccati equation. Next, under some non-vanishing conditions of the Fourier-Laplace trans-
forms, we establish an existence result for such Riccati equation and we provide a discretized
approximation of the joint characteristic functional that is exponentially entire. On the practi-
cal side, we develop a numerical scheme to solve the stiff infinite dimensional Riccati equations
and demonstrate the efficiency and accuracy of the scheme for pricing SPX options and volatil-
ity swaps using Fourier and Laplace inversions, with specific examples of the Quintic OU and
the one-factor Bergomi models and their calibration to real market data.
JEL Classification: G13, C63, G10.
Keywords: Stochastic volatility, Derivative pricing, Fourier methods, Riccati equations, SPX-
VIX calibration
1 Introduction
Fourier inversion techniques hold a pivotal role in stochastic volatility modeling, particularly in the
context of option pricing and hedging (Andersen and Andreasen [8], Carr and Madan [10], Eberlein,
Glau, and Papapantoleon [16], Fang and Oosterlee [19], Lewis [26], Lipton [27]). They offer the
dual advantage of significantly reducing computational time while maintaining a remarkable degree
of accuracy, especially when compared to standard Monte Carlo methods. When it comes to model
∗ [email protected]. The first author is grateful for the financial support from the Chaires
FiME-FDD, Financial Risks, Deep Finance & Statistics at Ecole Polytechnique.

† [email protected]. The second author acknowledges the financial support from AXA Investment Man-
agers
‡ [email protected].
We would like to thank Alessandro Bondi, Louis-Amand Gérard, Camille Illand, Sergio Pulido and Ning Tang for
fruitful discussions.
1
calibration, the applicability of Fourier methods is paramount, as thousands of derivatives across
various maturities and strikes need to be evaluated simultaneously in real time.
Despite their numerical advantage, Fourier techniques have traditionally been confined to specific
continuous stochastic volatility models where the characteristic function of the log-price is known
in (semi)-closed form. These models usually exhibit Markovian affine structures in the sense
of Duffie, Filipović, and Schachermayer [14] such as the renowned Heston [24], Stein-Stein [32]
and Schöbel-Zhu [30] models, together with some of their non-Markovian Volterra counterparts
(Abi Jaber [1, 2], Abi Jaber, Larsson, and Pulido [5], Cuchiero and Teichmann [11], El Euch and
Rosenbaum [18], Gatheral and Keller-Ressel [22]). The key ingredient in all these models is to
compute the characteristic function of the log-price by solving a specific system of deterministic
Riccati equation.
More recently, Fourier techniques have also found applications in Signature volatility models in
Abi Jaber and Gérard [3] and Cuchiero, Gazzani, Möller, and Svaluto-Ferro [12], where the volatil-
ity process is modeled as a linear functional of the path-signature of semi-martingales (e.g. a
Brownian motion). In such models, certain characteristic functionals have been related to non-
standard infinite dimensional system of Riccati ordinary differential equation (ODE). However,
there exists no general theory regarding the existence of solutions for such equations, except for
the specific result in Cuchiero, Svaluto-Ferro, and Teichmann [13, Proposition 6.2] which provides
the existence of a solution to the infinite dimensional ODE for the case of the characteristic func-
tion of powers of a single Brownian motion modulo a non-vanishing condition of the characteristic
function. The primary challenge comes from the intricate questions on analyticity of the logarithm
of the characteristic function. Furthermore, numerically solving these equations poses challenges
due to their stiffness and complexity. This forms our primary motivation: to establish theoretical
results within a specific framework for a larger class of Riccati equations related to integrated
quantities of power series of Ornstein-Uhlenbeck processes and to develop more suitable numerical
schemes.
We demonstrate that Fourier techniques can be effectively extended to a broad class of flexible
models previously considered infeasible as they fall beyond the conventional class of affine diffusions,
including the celebrated one-factor Bergomi model (Bergomi [9], Dupire [15]) that has been shown
to fit well to the SPX smiles and the recently introduced Quintic Ornstein-Uhlenbeck (OU) model
of Abi Jaber, Illand, and Li [7], which has demonstrated remarkable capabilities in fitting jointly
the SPX-VIX volatility surface for maturities between one week to three months supported by
extensive empirical studies on more than 10 years of data in Abi Jaber and Li [4], Abi Jaber,
Illand, and Li [6].
We refer to the class of models covered in this paper as the class of polynomial OU models, where
the volatility of the log-price, denoted by σ, is defined as a power series of an Ornstein-Uhlenbeck
process. Our three main theoretical results regarding the joint Fourier-Laplace functional of the
log-price and integrated variance process are:
(i) A verification result on the expression of the joint Fourier-Laplace functional in terms of a
solution to an infinite dimensional system of Riccati ODE in Theorem 3.2,
(ii) The existence of a solution to such a system of ODE modulo a non-vanishing condition of
the joint characteristic functional in Theorem 3.5,
(iii) An approximation procedure for the joint Fourier-Laplace functional in terms of exponentially
entire expressions in Theorem 3.13.
On the practical side, we devise a numerical scheme to solve this system of infinite dimensional
(stiff) Riccati equations making the application of Fourier pricing for models such as the one-factor
Bergomi and the quintic OU models usable in practice. We demonstrate the efficiency and accuracy
of the scheme for pricing SPX options and volatility swaps using Fourier and Laplace inversions.
We also successfully calibrate using Fourier techniques the Quintic OU and the one-factor Bergomi
2
models on real market data to highlight the stability and robustness of our numerical method
across a wide range of realistic parameters values. We provide a Python notebook implementation
here: https://colab.research.google.com/drive/1VCVyN1qQmLgOWjOy4fbWftyDqEQdm5n5?us
p=sharing.
The paper is organised as follows: in Section 2, we introduce the class of polynomial OU volatility
models. In Section 3, we present our three main theoretical results. In Section 4, we design a
numerical scheme for solving the infinite dimensional Riccati ODE and test our scheme on pricing
SPX vanilla options and volatility swaps for the one-factor Bergomi and quintic OU volatility
models of various maturities and calibrate both models using real market data. Proofs of theorems
discussed in Section 3 are collected in Section 5 onwards.
P∞
Notations. For a power series q, we denote by qk the coefficientsP of xk , i.e. q(x) = k=0 qk xk .
∞
We denote by |q| the absolute power series of q, i.e. |q|(x) = k=0 |qk |xk . We remark that
|q|k = |qk |. It is well-known that if p and q have an infinite radius of convergence, then p(x)q(x) =
P∞ k
Pk
k=0 (p∗q)k x also has an infinite radius of convergence, with (p∗q)k := l=0 pl qk−l . Unless stated
otherwise, we will assume that all power series in this paper have infinite radius of convergence.
Moreover, we specify that the term continuity means joint continuity when applied to a function
taking (t, x) as variables. For example, a function f : [0, T ] × R → C is C ∞ in x if all its space
derivatives in x exist and are continuous in (t, x). We also adopt the following notations for the
∂2f
partial derivatives: ft = ∂f ∂f
∂t , fx = ∂x , fxx = ∂x2 when they exist. We say that a function f : R → C
is (real)-entire on R, if itPis equal to a power series q (with infinite radius of convergence), that
∞
is for all x ∈ R, f (x) = k=0 qk xk . We say that it is (complex)-entire when R is replaced by C.
Clearly, any real-entire function admits a complex-entire extension. When there is no-ambiguity,
we will simply use the word entire. We say that a function does not vanish if it is non-zero at each
point of its domain.
2 Polynomial Ornstein-Uhlenbeck volatility models
We consider the class of polynomial Ornstein-Uhlenbeck (OU) models for a stock price S with the
stochastic volatility process σ expressed as a power series of an OU process X:
dSt = St σt dBt , S0 > 0,
∞
X
σt = g0 (t)p(Xt ), p(x) = pk xk , (2.1)
k=0
dXt = (a + bXt )dt + cdWt , X0 ∈ R,
p
with a, b, pk ∈ R, c ̸= 0, B = ρW + 1 − ρ2 W ⊥ and ρ ∈ [−1, 1]. Here (W, W ⊥ ) is a two-dimensional
Brownian motion on a risk-neutral filtered probability space (Ω, F, (Ft )t≥0 , Q) satisfying the usual
conditions. The deterministic bounded input curve g0 : [0, T ] → R+ allows the model to match
certain
p term structures of volatility, e.g. the forward variance curve since we have for g0 (t) :=
ξ0 (t)/E[p2 (Xt )]:
Z t Z t
2
E σs ds = ξ0 (s)ds, t ≥ 0.
0 0
The real-valued coefficients (pk )k≥0 are such that the power series p has infinite radius of conver-
P∞ RT
gence, i.e. k=0 |pk ||x|k < ∞ for all x ∈ R and 0 E[p2 (Xs )]ds < ∞ so that the stochastic integral
is well-defined. This is the case for instance when p is a finite polynomial or the exponential
function. These two specifications already provide several interesting models used in practice:
• Stein-Stein [32] and Schöbel-Zhu [30] model: p(x) = x;

• Bergomi model [9, 15]: p(x) = exp(x);
3
• Quintic OU model [7]: p(x) = p0 + p1 x + p3 x3 + p5 x5 , p0 , p1 , p3 , p5 ≥ 0.
RT
We will consider the following class of power series p for which 0 E[p2 (Xs )]ds < ∞.
P∞
Definition 2.1. AP power series p : x 7→ k=0 pk xk is said to be negligible to double factorial
∞
if the power series k=0 (k − 1)! ! pk xk has an infinite radius of convergence, i.e.
1
lim sup(|pk |(k − 1)! ! ) k = 0,
k→∞
with the convention (−1)! ! = 0! ! = 1.
By Definition 2.1, a power series p negligible to double factorial also has infinite radius of conver-
gence. This allows us to show that our class of polynomial Ornstein-Uhlenbeck volatility models
(2.1) is well-posed in Proposition 2.3. For this we need a simple lemma about the absolute power
series which will be useful later.
Lemma 2.2. ˙
(i) for all x ∈ R, |p(x)|≤ |p|(|x|) and |p| is monotonically increasing on R+ ;
(ii) for all x, y, z ≥ 0, |p|(x + y + z) ≤ (|p|(3x) + |p|(3y) + |p|(3z)).
Proof. (i) is obvious; for (ii), it suffices to notice that (x + y + z)k ≤ 13 ((3x)k + (3y)k + (3z)k ) ≤
((3x)k + (3y)k + (3z)k ).
Proposition 2.3. Let g0 : [0, T ] → R be a measurable and bounded function. Let p be a power
series with infinite radius of convergence such that p2 is negligible to double factorial. Then,
Z T
E[p2 (Xt )]dt < ∞, (2.2)
0
R·
so that the stochastic integral 0
σs dBs , with σs = g0 (s)p(Xs ), is well-defined and
Z t Z t
1
St = S0 exp − σs2 ds + σs dBs , t ≥ 0, (2.3)
2 0 0
is the unique strong solution to (2.1).
Proof. Under (2.2), it is straightforward to obtain that the unique strong solution S to (2.1) is
given by (2.3). It suffices to prove (2.2). For this, we note that the explicit solution of X is given
by
Z t Z t
Xt = ebt X0 + a eb(t−s) ds + c eb(t−s) dWs .
0 0
Rt
We set w(t) := a 0 eb(t−s) ds, which is a deterministic continuous function of t, and W ft :=
R t b(t−s) e 2bt
−1
c 0e dWs , which is a Gaussian random variable ∼ N (0, c2 2b ) for t ∈ [0, T ]. Applying
Lemma 2.2 yields
Z T Z T
|p2 | ebs X0 + w(s) + W
fs ds ≤ |p2 | ebs |X0 |+|w(s)|+|W
fs | ds
0 0
Z ∞
T X
≤ |p2 |k (3ebs |X0 |)k + |3w(s)|k +|3W
fs |k ds,
0 k=0
4
and since ebs ≤ e|b|T with w(s) bounded in [0, T ], there exists a constant CT ≥ 0 such that
Z ∞
T X Z ∞
T X
2
|p |k bs k k k
(3e |X0 |) + |3w(s)| +|3Ws | ds ≤
f |p2 |k |3e|b|T |X0 |k +(3CT )k + |3W
fs |k ds.
0 k=0 0 k=0
By assumption, p2 has an infinite radius of convergence, then so does |p2 |, therefore

∞
X ∞
X
|p2 |k |3e|b|T |X0 |k < ∞, |p2 |k (3CT )k < ∞,
k=0 k=0
and thus we only need to prove

Z T
E[|p2 |(|3W
ft |)]dt < ∞.
0
The random variable 3W ft is Gaussian with bounded variance in [0, T ], so there exists a constant
k
C > 0 such that E[|3W ft | ] = E[|CZ|k ] for all t ∈ [0, T ], where Z is a standard normal variable.
Given E[|Z|k ] ≤ (k − 1)! !, hence
∞
X
E[|p2 |(C|Z|)] ≤ |p2 |k (k − 1)! ! C k < ∞,
k=0
since p2 is negligible to double factorial. This completes the proof.

Remark 2.4. In Proposition 2.3, p2 needs to be negligible to double factorial. As we will see later
in Lemma 6.3 below, the square of a power series negligible to double factorial is also negligible to
double factorial. Hence p negligible to double factorial is sufficient for Proposition 2.3 to hold.
Example 2.5. Polynomials of finite degree (e.g. the Stein-Stein and the Quintic OU models) and
the exponential function (e.g. the one-factor Bergomi model) are clearly negligible to double facto-
k 1
rial. For the exponential function p(x) = eεx , ε > 0, it suffices to observe that limk→∞ ( ε (k−1)!!
k! )k =
εk k1 2
limk→∞ ( k!! ) = 0. However, the function p(x) = eεx is not negligible to double factorial. Indeed,
1 √
lim supk→∞ (|pk |(k − 1)! ! ) k = 2ε. In particular, if b > 0, then the expectation of p(Xt ) does not
exist for all t ≥ 2c12 ϵ .
3 The joint Fourier-Laplace transform
Our goal is to compute the joint Fourier-Laplace transform of the log-price and integrated variance
" ! #
Z T Z T
E exp g1 (T − s)d log Ss + g2 (T − s)σs2 ds Ft , t ≤ T,
t t
for some complex-valued functions g1 , g2 : [0, T ] → C.

Remark 3.1. If g1 , g2 are measurable functions [0, T ] → C, such that ℜ(g1 ) = 0, ℜ(g2 ) ≤ 0, and
p2 is negligible to double factorial, then the exponential above exists and with modulus of at most
1, so that the conditional expectation is well-defined.
It follows from the Markov property of X that the above conditional expectation can be reduced
to computing the deterministic measurable function F : [0, T ] × R → C given by
" ! #
Z TZ T
F (t, x) := E exp g1 (T − s)d log Ss + g2 (T − s)σs2 ds Xt = x . (3.1)
t t
5
We will present three main results related to the computation of the characteristic functional F
in (3.1). The first result can be seen as a verification result and uncovers an affine structure in
infinite dimension in terms of the powers (1, X, X 2 , X 3 , . . .) by making a connection with infinite
dimensional Riccati deterministic equations (Theorem 3.2 and Corollary 3.6). The second result
is concerned with the existence of a solution to the infinite dimensional Riccati deterministic
equation (Theorem 3.5). The last result provides an approximation procedure for obtaining the
characteristic functional (Theorem 3.13).
3.1 A verification result
Our first main result uncovers an affine structure in infinite dimension expressed in powers (1, X, X 2 , . . .)
and makes a connection with the following system of infinite dimensional deterministic Riccati
equations:
g1 (t)
ψk′ (t) = g2 (t) + (g1 (t) − 1) g02 (T − t)(p ∗ p)k
2
c2 (k + 2)(k + 1)
+ bkψk (t) + a(k + 1)ψk+1 (t) + ψk+2 (t)
2
c2 e
+ (ψ(t) ∗ ψ(t))
e k + ρg1 (t)g0 (T − t)c(p ∗ ψ(t)) e k , ψek (t) := (k + 1)ψk+1 (t), (3.2)
2
ψk (0) = 0. (3.3)
Theorem 3.2. Let g0 : [0, T ] → R, g1 , g2 : [0, T ] → C be measurable and bounded functions.

Let p be a power series with infinite radius of convergence such that p2 is negligible to double
factorial. Assume that there exists a continuously differentiable solutionP (ψk )≥0 to the system of
infinite dimensional Riccati equations (3.3) such that the power series k≥0 supt∈[0,T ] |ψk (t)|xk has
an infinite radius of convergence. Define the process
X Z t Z t
Ut = ψk (T − t)Xtk + g1 (T − s)d log Ss + g2 (T − s)σs2 ds.
k≥0 0 0
Then the process M := exp(U ) is a local martingale. If in addition M is a true martingale, then
the following expression holds for the joint characteristic functional F given in (3.1):
 
X
F (t, x) = exp  ψk (T − t)xk  , t ≤ T. (3.4)
k≥0
Proof. The proof is given in Section 5.

Remark 3.3. Theorem 3.2 is in the spirit of [13, Theorem 5.5] which provides a similar verification
result for the characteristic function of (real)-entire functions of solutions to stochastic differential
equations with (real)-entire coefficients. Contrary to [13, Theorem 5.5], Theorem 3.2 deals with
time-integrated quantities.
A possible strategy for obtaining the representation (3.4) would involve verifying the assumptions
outlined in Theorem 3.2. These assumptions include ensuring the existence of a solution for the
system of Riccati equations (3.2)-(3.3), such that the power series (ψk )k≥0 has infinite radius of
convergence; together with proving that the local martingale M is a true martingale. The latter
can typically be achieved by arguing, for instance, that k≥0 ℜ(ψk (t))xk ≤ 0 to obtain that M is
P
uniformly bounded by 1 whenever g1 , g2 are purely imaginary for instance.
In the specific case of the Stein-Stein model, i.e. when p is an affine function, these assumptions
are comparatively easier to confirm, as highlighted in the next example.
6
Example 3.4. In the case of the classical Stein-Stein model [32], the volatility process process is
defined as:
σt = g0 (t)Xt .
This is equivalent to (2.1) by setting p1 = 1 and pk = 0 for k ̸= 1. Notice the convolution term
(p ∗ p)k = 1 for k = 2 and zero otherwise. In this case, the infinite dimensional Riccati equations
(3.3)-(3.2) reduce to the following:
c2 2
ψ0′ (t) = aψ1 (t) + c2 ψ2 (t) + ψ (t),
2 1
′ 2
ψ1 (t) = bψ1 (t) + 2aψ2 (t) + 2c ψ1 (t)ψ2 (t) + ρg1 (t)g0 (T − t)cψ1 (t),
g1 (t)
ψ2′ (t) = g2 (t) + (g1 (t) − 1) g02 (T − t)
2
+ 2bψ2 (t) + 2c2 ψ22 (t) + 2ρg1 (t)g0 (T − t)cψ2 (t),
ψk (0) = 0, k ∈ {0, 1, 2},

ψk ≡ 0, k ≥ 3,
which is a system of standard finite-dimensional Riccati equations whose existence is well-known

whenever g1 , g2 are such that
g1 (t)
ℜ g2 (t) + (g1 (t) − 1) ≤ 0, t ≤ T.
2
Hence the characteristic functional of the classical Stein-Stein model is affine in (1, X, X 2 ):
F (t, x) = exp ψ0 (T − t) + ψ1 (T − t)x + ψ2 (T − t)x2 , t ≤ T,

for which ψ0 , ψ1 and ψ2 can even be solved explicitly when g0 , g1 and g2 are constants, see [28].
Obtaining the representation (3.4) under the Stein-Stein model can be attributed to the finite
number of terms, i.e. (ψ0 , ψ1 , ψ2 ) of the Riccati equations, involved in the sum. However, when
dealing with an infinite sum, a notable challenge arises in proving that the log of the characteristic
functional log F is entire in the variable x on R. In Section 3.2, we show how to generate a solution
for the system of infinite-dimensional Riccati equations modulo a non-vanishing condition of log F .
In Section 3.3, we provide an expression for the characteristic functional using approximation
arguments where the approximations are entire functions thanks to the Gaussianity of the process
X.
In Section 4, we numerically illustrate the validity of Fourier pricing in the context of the Bergomi
and Quintic OU models using the representation (3.4). Readers interested in the numerical imple-
mentation can jump directly to Section 4.
3.2 Existence for the Riccati equations
Our second main result generates a solution to the infinite dimensional system of Riccati equations
(3.2)-(3.3) by differentiating the logarithm of the characteristic functional F :
1 k
ψk (t) := ∂ log F (T − t, x) , t ≥ 0.
k! x x=0
This requires the logarithm of a complex-valued function, see Appendix B for its precise definition.
Theorem 3.5. Fix g0 : [0, T ] → R, g1 , g2 : [0, T ] → C continuously differentiable such that
ℜ(g1 ) = 0 and ℜ(g2 ) ≤ 0. Let p be a power series negligible to double factorial. Then, F in (3.1)
is well-defined for any t ≤ T , x ∈ R and (t, x) 7→ F (t, x) is continuous. If in addition F does not
7
vanish on [0, T ] × R, then log F can be defined as in Definition B.5. Furthermore, log F is C ∞ in
x and C 1 in t. In particular, the family of functions
1 k
ψk (t) = ∂ log F (T − t, x) , t ≥ 0, k ≥ 0, (3.5)
k! x x=0
solves the system of Riccati ODEs (3.2)-(3.3).
Theorem 3.5 establishes the existence of a solution to the Riccati ODEs (3.2)-(3.3) when the
coefficients are real as shown in the following corollary.
Corollary 3.6. Fix g0 : [0, T ] → R, g1 = 0 and g2 : [0, T ] → R− continuously differentiable. Let p
be a power series negligible to double factorial. Then, log F is C ∞ in x and C 1 in t and the family
of functions (ψk )k≥0 given by (3.5) solves the system of Riccati ODEs (3.2)-(3.3).
Proof. In this case, F reduces to the Laplace transform of the integrated variance:
" Z T ! #
2
F (t, x) = E exp g2 (T − s)σs ds Xt = x ,
t
which shows that F > 0 and that it trivially satisfies the non-vanishing condition on [0, T ] × R in
Theorem 3.5, so that an application of Theorem 3.5 yields the result.
Remark 3.7. Theorem 3.5 is in the spirit of [13, Proposition 6.2], which provides the existence of
a solution to the infinite dimensional ODE for the case of the characteristic function of powers of
a single Brownian motion modulo similar assumptions. In contrast, Theorem 3.5 deals with more
involved time-integrated quantities which requires a more delicate analysis.
Although Theorem 3.5 does not establish that log F (t, ·) is entire, its analyticity can be inferred
from the properties of solutions to parabolic partial differential equations (PDEs).
Remark 3.8. In order to prove Theorem 3.5, we establish the following PDE for f (t, x) =
F (t, x) exp(v(t)q(x)) in Theorem of 6.11:
( P3
ft (t, x) + fx (t, x)(a + bx) + 12 c2 fxx (t, x) + f (t, x) i=1 ui (t)pi (x) = 0,
f (T, x) = exp(v(T )q(x)).
where v, ui are continuous and pi are analytic functions. Following [17, Theorem 6.2 in Section
II.2.], given that functions ui (t) are continuous in [0, T ], a + bx, pi (x) are bounded, Hölder contin-
uous and analytic in variable x within any open bounded set of R, one would obtain that f (t, ·) is
analytic. Notice that exp(v(t)q(·)) is also analytic, then so is F (t, ·). Since analyticity is a local
property, if F does not vanish, then log F (t, ·) is also analytic. Therefore, we can deduce a local
representation: for any t ≤ T , there exists rt > 0 such that
∞
!
X
k
F (t, x) = exp ψk (T − t)x , x ∈ (−rt , rt ),
k=0
where the (ψk )k≥0 are defined by (3.5). In our case, both pi and q are not only analytic but also
entire. Extending the proof of analyticity using PDE techniques to establish that F (t, ·) is entire
might be possible, and that will lead to a global representation as in (3.4). However, this needs a
more delicate analysis at the PDE level, diverging from our probabilistic approach that allowed us
to obtain a global representation in terms of approximations by exponentially entire functions, see
Theorem 3.13 below. Such a question holds independent interest for PDEs in its own right.
8
3.3 The Fourier-Laplace transform by approximation
Our third main result provides an approximation procedure for the characteristic functional F .
The main idea is to approximate the Riemann sum on the sample paths of the Ornstein-Uhlenbeck
process X to exploit an underlying Gaussian density in finite dimension.
For this, we need first to get rid of the stochastic integrals and express them in terms of functions
of XT and Lebesgue’s integrals on (Xs )s≤T . This is done by combining Itô’s Lemma and the
Romano-Touzi conditioning trick [29] in the next Lemma.
Lemma 3.9. For p negligible to double factorial, g0 : [0, T ] → R, g1 , g2 : [0, T ] → C continuously
differentiable with ℜ(g1 ) = 0, ℜ(g2 ) ≤ 0, there exists power series q, pi , i = 1, 2, 3 with infinite
radius of convergence, and functions v, ui , i = 1, 2, 3 such that:
" Z TX 3
! #
F (t, x) = E exp ui (s)pi (Xs )ds + v(T )q(XT ) − v(t)q(Xt ) Xt = x .
t i=1
In addition, ui , v are continuous on [0, T ], ℜ(u1 ) ≤ 0, ℜ(u2 ), ℜ(u3 ), ℜ(v) = 0, p1 = p2 such that
pi , q are all negligible to double factorial.
Proof. The proof and the precise expressions of ui , v, pi , q are given in Appendix A.
We now introduce Fn the discretized version of F .

Pn−1
Definition 3.10. For µn = Tn i=0 δ iT , where δt is the Dirac mass at a point t, we define the
n
function
" Z TX3
! #
Fn (t, x) := E exp ui (s)pi (Xs )µn (ds) + v(T )q(XT ) − v(t)q(Xt ) Xt = x .
t i=1
Definition 3.11. Given an entire function g on R, we define the extension of g to the complex
P∞ (k)
plane by gc (z) = 0 g k!(0) z k , z ∈ C.
P∞ g(k) (0) k
Remark 3.12. Given an entire function g on R, the power series k=0 k! x has infinite
P∞ g(k) (0) k
radius of convergence, so does k=0 k! z , thus gc (z) is well-defined in C, and is entire on C.
Theorem 3.13. Suppose that p is negligible to double factorial, g0 : [0, T ] → R, g1 , g2 : [0, T ] → C

continuously differentiable such that ℜ(g1 ) = 0, ℜ(g2 ) ≤ 0, then Fn (t, ·) as specified in Defini-
tion 3.10 is well-defined and entire in x ∈ R. If in addition, the extension of Fn (t, ·) to the
complex plane (Fn (t, ·))c does not vanish for all n ∈ N and t ≤ T , then, log Fn is entire in x ∈ R
such that  
X
Fn (t, x) = exp  ψn,k (T − t)xk  , t ≤ T,
k≥0
and
F (t, x) = lim Fn (t, x),
n→∞
where for all n ∈ N, the family of functions ψn,k is defined by

pk−1
ψn,k (t) = ϕn,k (t) − ρg1 (t)g0 (T − t) , k ≥ 1,
ck
ψn,0 (t) = ϕn,0 (t),
9
where the family ϕn,k solves a (step-wise) ‘discretized’ system of Riccati ODEs:
c2 (k + 2)(k + 1)
ϕ′n,k (t) = bkϕn,k (t) + a(k + 1)ϕn,k+1 (t) + ϕn,k+2 (t)
2
c2 e nt
+ (ϕn (t) ∗ ϕen (t))k , ϕen,k = (k + 1)ϕn,k+1 , ∈
/ {0, 1, 2, . . . , n}
2 T
3
T X nt
ϕn,k (t) = lim− ϕn,k (s) + ui (T − t)(pi )k , ∈ {1, 2, . . . , n}
s→t n i=1 T
ϕn,k (0) = v(T )qk

Remark 3.14. One of the features of Theorem 3.13 is to obtain that the functions Fn (t, ·) are
entire in x ∈ R with no additional assumptions. Unfortunately, log Fn (t, ·) inherits this property
only when the extension of Fn (t, ·) to the complex plane (Fn (t, ·))c does not vanish, see Lemma B.3.
The requirement of Fn being non-vanishing on the real line is not enough, see Remark B.4. For
this reason, we could not obtain a Corollary of Theorem 3.13 without the non-vanishing condition
to deal with Laplace transforms as we did in Corollary 3.6.
4 Numerical illustration
4.1 Numerical scheme for high-dimensional Riccati equations
In this section, we show how to solve the infinite dimensional Riccati equations ψ in (3.2)-(3.3)
numerically for derivative pricing and hedging. The idea is to solve a truncated version of the ODE
by assuming ψk ≈ 0, k > M for some integer M .
Even the truncated version of the infinite dimensional system of Riccati equations can be extremely
difficult to solve. Standard techniques such as the Runge–Kutta methods usually fail to solve the
ODE, especially when the parameters b and c in (2.1) are large. For example, see [13, Figures 2
and 3] for other attempts to a similar problem.
We present a customized algorithm that combines variation of constants and the implicit Euler
scheme. For this section alone, all matrix and vector indices will start from 0. With some abuse
of notation, we re-write the Riccati ODE (3.3) in the following matrix form:
c2 e
ψ ′ (t) = P (t) + Aψ(t) + Qψ(t) + (ψ(t) ∗ ψ(t))
e + l(t)N ψ(t),
2
with ψ(t) denoting the vector (ψ0 (t), ψ1 (t), . . .)⊤ and ′
ψ (t) the vector of element-wise
derivative of
ψ(t). P (t) is a vector with it’s k th element being g2 (t) + g12(t) (g1 (t) − 1) g02 (T − t)(α ∗ α)k for
k ∈ {0, 1, . . .}. A is a diagonal matrix with its diagonal Ak,k = bk. Q is an upper triangular matrix
with Qk,k+1 = a(k + 1), Qk,k+2 = c2 (k + 1)(k + 2)/2 and zero elsewhere. The term (ψ(t) e ∗ ψ(t))
e
th
denotes the discrete convolution of the vectors ψ(t), with its k element given by (ψ(t) ∗ ψ(t))k .
e e e
Finally, N is a matrix with Nj,k = kpj+1−k where pm denotes the power series coefficient such
that pm = 0 whenever m < 0, with l(t) collecting the term cρg1 (t)g0 (T − t).
Variation of constants gives:

Z ti+1 Z ti+1
ψ(ti+1 ) = eA∆t ψ(ti ) + eA(ti+1 −s) P (s)ds + eA(ti+1 −s) Qψ(s)ds
ti ti
ti+1 ti+1
(4.1)
c2
Z Z
+ eA(ti+1 −s) (ψ(s)
e ∗ ψ(s))ds
e + eA(ti+1 −s) l(s)N ψ(s)ds,
2 ti ti
10
with ∆ = ti+1 − ti . By truncating the vector ψ up to some level M and assuming (ψk )k>M = 0,
we approximate the solution of ψ(ti+1 ) in (4.1) by the following quasi-implicit scheme:
ψ(ti+1 ) ≈ eA∆t ψ(ti ) + G∆ P (ti ) + G∆ Qψ(ti+1 )

c2 e i ) ∗ ψ(t
+ G∆ (ψ(t e i+1 )) + G∆ l(ti )N ψ(ti+1 ),
2
Rt
where G∆ = tii+1 eA(ti+1 −s) ds is a diagonal matrix with Gk,k = (eb(k−1)∆ − 1)/(b(k − 1)), 1 < k ≤
M + 1 and G1,1 = ∆. Similarly, Q and N are now (M + 1) × (M + 1) matrices as defined above,
with ψ and P (ti ) both vectors of dimension M + 1 .
e i ) ∗ ψ(t
For the quadratic term (ψ(t e i+1 )), we define
R(i) ψ(t e i ) ∗ ψ(t

e i+1 ) := (ψ(t e i+1 )),
(i)
where R(i) is a (M + 1) × (M + 1) matrix with Rj,k = (j + 2 − k)kψj+2−k (ti ), (ψz (ti ))z<0 = 0. We
can now solve ψ iteratively by:

ψ(ti+1 ) ≈ J −1 eA∆t ψ(ti ) + G∆ P (ti ) ,
2
where J = I − G∆ c2 R(i) + l(ti )N + Q . Detailed implementation of our algorithm can be
found in a Python notebook here: https://colab.research.google.com/drive/1VCVyN1qQmLg
OWjOy4fbWftyDqEQdm5n5?usp=sharing.
Lemma 4.1. Assume that ℜ(ψ) ≤ 0, the matrix J is invertible.

c2 (i)
Proof. Since ℜ(ψ) ≤ 0, the matrix G∆ 2 R + l(ti )N + Q is a upper triangular matrix, where
the real part of the its diagonal is less than zero. Thus J is also an upper triangular matrix with
non-zero diagonal elements, thus completes the proof.

Matrix Q contains very large coefficients resulting from the term c2 (k + 2)(k + 1)/2 when k is
2
large. This term introduces numerical instability so we capped (k + 2)(k + 1) to some level kmax
to ensure the scheme does not blown up.
W4
We first test our algorithm on the term E[exp(− 4!t )] considered in [13, Figures 2 and 3], which
can be expressed by the solution of a particular case of the Riccati ODE in (3.2)-(3.3) with initial
condition ψ4 (0) = −1/4!, (a, b, g1 , g2 ) = 0 and c = 1. The reference value can be computed via a
numerical integration quadrature with respect to the Gaussian density, allowing us to evaluate our
algorithm’s performance. Figure 1 illustrate numerical convergence in terms of M , with step size
n = 100 and kmax = 15 fixed. We clearly observe the instability of the Runge-Kutta scheme with
increasing truncation M , whereas the scheme we propose remains stable with the increase of M .
11
W4
Figure 1: Numerical solution of E[exp(− 4!t )] with different values of M using our algorithm vs.
the exact solution. For comparison the explicit Runge-Kutta 4 solution with M = 20 are added in
red.
4.2 Pricing SPX derivatives via Fourier
Our numerical scheme gives us direct access to the characteristic function of log S to price deriva-
tives via Fourier inversion techniques, using the expression (3.4) with g2 (t) = 0 and g1 (t) = iu, for
u ∈ R. Specifically, we show how one can price European options via Fourier techniques for the
Quintic Ornstein-Uhlenbeck Model [7] and the one-factor Bergomi model [9, 15]. This also serves
as a numerical validation of the representation (3.4).
There are several Fourier inversion techniques available in the literature to price European-style
Vanilla call and put options. Here we adapt the pricing formula suggested in [26] which involves
only one integral to evaluate:
K ∞
Z
+
(iu+ 21 )kt i du
Ct (St , K, T ) := E (ST − K) |Ft = St − ℜ e φ u− ,
π 0 2 u2 + 14
where Ct (St , K, T ) denotes the European call option price with strike K and maturity T − t, kt :=
log(K/St ) is the log-moneyness and φ(u) := F (t, x) the Fourier-Laplace transform of log(ST /St )
by fixing g1 ≡ iu and g2 ≡ 0. We use the the representation of F (t, x) in (3.4) and compute the
improper integral numerically via the Gauss-Laguerre quadrature, which has been demonstrated
to be efficient, see [3, 25].
To speed up the computation of Ct , we add a control variate to reduce the number of evaluation
of φ(u) for different u:
K ∞
Z
(iu+ 21 )kt i i du
Ct (St , K, T ) = Ct (St , K, T ) −
b ℜ e φ u− −φ
b u− ,
π 0 2 2 u + 14
2
where Cbt (St , K, T ) is the appropriate call price of the control variate and φ
b (u) is the Fourier-
Laplace transform of log(ST /St ) of the control variate.
In this section, we choose the Heston model as the control variate, which admits a closed form char-
acteristic function that is also affine in its state variables, see [24]. The Heston model parameters
can be efficiently selected
via a standard optimization algorithm such that the difference between
φ u − 2i and φ b u − 2i is minimized. Of course, other control variates with explicit characteristic

functions such as the Black & Scholes model can also be used.
4.2.1 Quintic Ornstein-Uhlenbeck volatility model
The volatility process σt under the Quintic OU Model takes the form of:
12
s
ξ0 (t)
σt = p(Xt ), p(x) = p0 + p1 x + p3 x3 + p5 x5 ,
E [p(Xt )2 ]
Z t (4.2)
−1
Xt = εα eαε (t−s) dWs ,
0
with ε > 0 and α ≤ 0. The non-negative coefficients p0 , p1 , p3 , p5 ≥ 0 (p2 = p4 = 0) ensure a
negative leverage effect as well as the martingality of S whenever ρ ≤ 0, see [7] . The particular
parametrization with X means a = 0, b = αε−1 and c = εα from (2.1). This model has shown to
produce remarkable joint fits to both SPX and VIX implied volatility surfaces [6, 7].
Since X is an OU process which can be simulated exactly, one could be tempted to use Monte
Carlo to estimate the SPX derivatives with appropriate control variates. However, the calibrated
values of ε are usually very small and α is negative, pushing the model effectively into a fast
regime with large mean reversion of order αε−1 and large vol of vol εα . It is known the standard
Euler-scheme for pricing SPX derivatives can reproduce large estimation bias for longer maturities
T due to the highly erratic paths of σ, requiring finer step size or other asymptotic approximation
techniques [20, 21]. Pricing via Fourier methods hence presents an attractive alternative given the
increased efficiency and accuracy. To demonstrate, we choose the following parameters: ρ = −0.65,
α = −0.6, (p0 , p1 , p3 , p5 ) = (0.01, 1, 0.214, 0.227), ξ0 (t) = 0.025, ε = 1/52 which are typical values
one can expect from calibrating the model to SPX and VIX smiles from [7]. Figure 2 shows the
convergence of SPX implied volatility of different maturities as the truncation level M increases:
Figure 2: SPX implied volatility of different maturities in the quintic OU model, computed with
our algorithm with different level M . Dotted red lines are Monte-Carlo 95% interval computed
with 500,000 simulations and n = 10, 000 number of steps per maturity slice.
4.2.2 One-factor Bergomi model
The one-factor Bergomi model [9, 15] assumes σt to be log-normal:

p 1 1
σt = ξ0 (t) exp ηXt − η 2 E(Xt ) ,
2 4
Z t (4.3)
−1
Xt = εα eαε (t−s) dWs ,
0
13
with ε > 0 and α ≤ 0. Similar to quintic OU model before, we have a = 0, b = αε−1 and c = εα .
We now approximate the exponential as a truncated sum up to level N :
s
N
ξ0 (t) X
k ηk
σ
et = p(Xt ), p(Xt ) = p k Xt , pk = ,
E[p(Xt )2 ] 2k k!
k=0
where σet in converges to σt in (4.3) when sending N → ∞. We now fix ρ = −0.7, ε = 1/52, α =
−0.7, η = 1.2, ξ0 (t) = 0.025 for the numerical experiment and set N = 8. Figure 3 shows that our
numerical scheme quickly converges to Monte-Carlo estimates of the original one-factor Bergomi
model:
Figure 3: SPX implied volatility of different maturities in the one-factor Bergomi model, computed
with our algorithm with different level M . Dotted red lines are Monte-Carlo 95% interval computed
with 500,000 simulations and n = 10, 000 number of steps per maturity slice.
4.3 Pricing q-volatility swaps via Laplace inversion
The q-volatility swap rate Kq is defined by:

" !q #
1 T 2
Z
Kq := E σ ds , q ∈ [0, 1].
T 0 s
For the case of a standard volatility swap (i.e. q = 1/2), one can price the swap rate K q via the
following inverse Laplace transform [31]:
 !1/2 
Z T Z ∞
1 1 1 − F̃ (u)
K 21 = E  σs2 ds = √ du,
T 0 2 π 0 u3/2
where F̃ (u) = F (t, x) from (3.1) by setting g1 ≡ 0 and g2 (t) ≡ −u/T using the presentation (3.4).
Again, we can accelerate the computation via a control variate, for example the Black & Scholes
control variate: Z ∞ e
1 FBS (u) − F̃ (u)
K 12 = σBS + √ du,
2 π 0 u3/2
2
with FBS (u) = exp(−uσBS ) where σBS is an arbitrary level of volatility that can be fixed upfront.
14
Compared to the previous section, we are even more confident of our numerical scheme thanks
to Corollary 3.6. For demonstration purposes, we use the same model parameters for the Quintic
Ornstein-Uhlenbeck and the one-factor Bergomi model as per the previous section, and adopt a
parametric forward variance curve in the form of ξ0 (t) = V0 e−kt +V∞ (1−e−kt ) with V0 = 0.025, k =
5 and V∞ = 0.06.
Figure 4 shows the volatility swaps of the two models computed by inverse Laplace transform with
truncation level M = 32 vs. that computed by Monte-Carlo with 400,000 simulations and 10,000
steps:
Figure 4: Volatility swaps under the Quintic OU model (left) and the one-factor Bergomi model
(right) for maturities up to 2 years, using the same parameter values as in the previous section.
4.4 Model calibration to market data via Fourier
The family of polynomial OU volatility models allows fast pricing of VIX futures and VIX options
via numerical integration of the payoff with respect to the standard Gaussian density, see [7]. To-
gether with fast Fourier pricing, this opens doors to joint calibration to SPX and VIX smiles. This
section also serves the purpose of highlighting the stability of our numerical discretization scheme
combined with Fourier inversion in a calibration procedure where a large number of evaluations of
the characteristic function are needed for a wide range of realistic model parameters and Fourier
variables.
In a nutshell, the calibration of a model involves minimizing the mean square error between prices
coming from the model vs. that from market data. Without going too much into the details, we
first demonstrate the capability of the quintic OU model (4.2) to jointly calibrate two slices of
maturities of SPX smiles together with one slice of maturity of the VIX smile, with calibrated
parameters ρ = −0.6763, α = −0.6821, (p0 , p1 , p3 , p5 ) = (0.0202, 1.3332, 0.0578, 0.0071) and fixed
ε = 1/52: with ξ0 (t) coming directly from market data:
15
Figure 5: Quintic OU model (green lines) jointly calibrated to the SPX and VIX smiles (bid/ask
in blue/red) on 23 October 2017 via Fourier using the Nelder-Mead optimization algorithm. The
truncation level of the Riccati solver is set at M = 32.
Next, we showcase the calibration results of the one-factor Bergomi model (4.3) on four slices
of maturities of SPX smiles between around 1 week to 3 months, with calibrated parameters
η = 1.1416, ρ = −0.6744, α = −0.7377 and fixed ε = 1/52:
Figure 6: One-factor Bergomi model (green lines) calibrated to the SPX smiles (bid/ask in
blue/red) on 23 October 2017 via Fourier using the Nelder-Mead optimization algorithm. The
truncation level of the Riccati solver is set at M = 32.
The market data of SPX and VIX volatility surface is purchased from the CBOE website https:
//datashop.cboe.com/. For more calibration examples under the quintic OU and the one-factor
Bergomi models, please refer to the the Appendix C.
5 Proof of Theorem 3.2
We first introduce a lemma to justify the use of Itô’s formula to the series k≥0 ψk (T − t)Xtk .
P
Lemma 5.1. Under the condition of Theorem 3.2, h(t, x) := k≥0 ψk (T − t)xk is C 2 in x and
P
C 1 in t, with hx (t, x) = k≥0 (k + 1)ψk+1 (T − t)xk , hxx (t, x) = k≥0 (k + 2)(k + 1)ψk+2 (T − t)xk ,
P P
ht (t, x) = k≥0 −ψk′ (T − t)xk .

P
Proof. First, |ψk (T − t)xk |≤ supt∈[0,T ] |ψk (t)||x|k , and supt∈[0,T ] |ψk (t)||x|k < ∞. So h(t, x) =
P
k≤0
16
ψk (T − t)xk is well defined. In addition, if we restrict x in a bounded set of R, then
P
k≥0 Pn
h (t, x) := k=0 ψk (T − t)xk is continuous and converges uniformly to h(t, x) when n → ∞, so
n
h(t, x) is continuous.
Notice that the domain of convergence for k≤0 supt∈[0,T ] |ψk (t)|xk < ∞ is R, then so is k≥0 ψk (T −
P P
t)xk . So hx (t, x), hxx (t, x) are well-defined, continuous and have the expression as the statement
for the same reason as for h(t, x).
To treat ht (t, x), we should at first prove that k supt∈[0,T ] |ψk′ (t)|xk has also an infinite radius of
P
convergence.
Indeed, using the Riccati expression (3.2), we only have to check that we have an infinite radius
of convergence for X
sup |at,k |xk , (5.1)
k≥0 t∈[0,T ]
where at,k is among (p ∗ p)k , kψk (t), (k + 1)ψk+1 , (k+1)(k+1)

2
e ∗ ψ(t))
ψk+1 , (ψ(t) e k , (p ∗ ψ(t))
e k.
1
By assumption and Cauchy–Hadamard Theorem, we know that lim supn→∞ (supt∈[0,T ] |ψn (t)|) n =
0 and thus we obtain that when at,x = kψk (t), (k + 1)ψk+1 = ψ(t)e k , (k+1)(k+2) ψk+2 , (5.1) has an
2
infinite radius of convergence.
P∞ P∞
Also, for two power series n=0 an xn and n=0 bn xn both with infinite radius of convergence, their
P∞
product n=0 (a ∗ b)n xn also has an infinite radius of convergence. And notice that (p ∗ p)k , (ψ(t)
e ∗
ψ(t))
e k , (p∗ψ(t))
e k are dominated by (|p|∗|p|)k , (sup |ψ(t)|∗
e
t∈[0,T ] sup |ψ(t)|)
e k , (|p|∗ sup
t∈[0,T ] |ψ(t)|)
e k. t∈[0,T ]
Thus when at,k is among (p P ∗ p)k , (ψ(t) ∗′ ψ(t))kk , (p ∗ ψ(t))k , (5.1) has also an infinite radius
e e e
of convergence. Therefore k supt∈[0,T ] |ψk (t)|x has an infinite radius of convergence, so that
′ k
P
k≥0 −ψ k (T − t)x also has infinite radius of convergence and converges uniformly when x is in
a bounded subset of R and also on t ∈ [0, T ]. So k≥0 −ψk′ (T − t)xk is well-defined and also
P
continuous.
Pn ′
Next, Pnotice that hnt (t, x) = P k
k=0 −ψk (T − t)x and given the the uniform Pconvergence of both
′ k k
series k≥0 −ψk (T −t)x and k≥0 ψk (T −t)x , we deduce that ht (t, x) = ( k≥0 ψk (T −t)xk )t =
′ k
P
k≥0 −ψk (T − t)x .
k
P
Proof of Theorem 3.2. We first notice that since the power series k |ψk (t)|x has an infinite radius
of convergence, Ut is well-defined for all t ≤ T .
With Lemma 5.1, we can now apply Itô’s formula on the semimartingale M :
dMt 1
= dUt + ⟨U ⟩t ,
Mt 2
X
dUt = g1 (T − t)d log St + g2 (T − t)σt2 dt − ψk′ (T − t)Xtk dt
(5.2)
k
X X1
+ (k + 1)ψk (T − t)Xtk dXt + (k + 2)(k + 1)ψk (T − t)Xtk d⟨X⟩t .
2
k k
Plugging (2.1) into (5.2), we have:
17

g1 (T − t) 2 X ′
dUt = (g2 (T − t) − σt − ψk (T − t)Xtk
2
k
!
2 (k + 2)(k + 1)
X X X
+ a(k + 1)ψk (T − t)Xtk + bkψk (T − t)Xtk + c k
ψk (T − t)Xt dt
2
k k k
X p
+ c(k + 1)ψk (T − t)Xtk + ρg1 (T − t)σt dWt + 1 − ρ2 g1 (T − t)σt dWt⊥ ,
k
X 2
d⟨U ⟩t = ckψk (T − t)Xtk + ρg1 (T − t)σt dt + (1 − ρ2 )g12 (T − t)σt2 dt.
k
pk Xtk leads to
P
Applying the Cauchy product on the power series k
dMt Xh g1 (T − t) 2
= (g2 (T − t) − )g0 (t)(p ∗ p)k − ψk′ (T − t) + a(k + 1)ψk+1 (T − t)
Mt 2
k
c2 (k + 2)(k + 1)
+ bkψk (T − t) + ψk+2 (T − t)
2
c2 e 2 2
e − t))k + g1 (T − t)g0 (t) (p ∗ p)k
+ (ψ(T − t) ∗ ψ(T
2 i 2
+ ρg1 (T − t)g0 (t)c(p ∗ ψ(T − t))k Xtk dt
e
X p
+ ckψk (T − t)Xtk + ρg1 (T − t)σt dWt + 1 − ρ2 g1 (T − t)σt dWt⊥ .
k
M is a local martingale if and only if it has zero drift (i.e. dt part is zero a.s.). This is true for all
values of X by assumption from (3.2), so that M is a local martingale. Moreover, if M is a true
martingale, we have
" ! #
Z T Z T
E exp g1 (T − s)d log Ss + g2 (T − s)σs2 ds Ft
t t
 
X
= exp  ψk (T − t)Xtk  , t ≤ T.
k≥0
which follows from the martingality of M and the assumption of the initial condition of ψk from
(3.3). This completes the proof.
To prove Theorem 3.5, we adopt the representation of F as mentioned in Lemma 3.9:

" Z TX3
! #
F (t, x) = E exp ui (s)pi (Xs )ds + v(T )q(XT ) − v(t)q(Xt ) Xt = x ,
t i=1
2
where the functions p1 = p , p2 , p3 , q are all negligible to double factorial and (u1 , u2 , u3 , v) are
continuous such that ℜ(u1 ) ≤ 0, ℜ(u2 ) = ℜ(u3 ) = ℜ(v) = 0. Consider the following function f
defined by:
" Z TX 3
! #
f (t, x) : = E exp ui (s)pi (Xs )ds + v(T )q(XT ) Xt = x
t i=1
= F (t, x) exp(v(t)q(x)), (6.1)
18
for all t ≤ T and x ∈ R. We recall that f is well-defined, since
Z TX 3
! Z
T
ℜ u1 (s)p2 (Xs ) ds ≤ 0.

ℜ ui (s)pi (Xs )ds + v(T )q(XT ) =
t i=1 t
Our proof is composed of five parts:
• In Subsection 6.1, we start by deriving some properties of power series negligible to double
factorial which will be used later.
• In Subsection 6.2, we prove the regularity in x of f given by (6.1), i.e. f is C ∞ in x such
k
that the partial derivatives ∂∂xfk are bounded in a sense in preparation for the Feynman-Kac
formula in Subsection 6.3.
• In Subsection 6.3, we prove that f solves the associated PDE coming from Feynman-Kac
formula. In particular, we prove that ∂f
∂t indeed exists.
• In Subsection 6.4, we introduce a system of ODE and obtain a solution by comparing the
derivatives of all orders of both sides of Feynman-Kac formula at x = 0.
• In Subsection 6.5, we prove that the system of ODE in Theorem 3.5 has a solution, via a
system of ODE introduced in Subsection 6.4.
6.1 Properties of power series negligible to double factorial
We collect some properties of power series that are negligible to double factorial as defined per
Definition 2.1.
P∞
Given a power series p(x) = k=0 pk xk , we define the set AD (p) as the R-algebra generated by all
higher-order derivatives {p, p′ , p′′ . . .} of p:
( l )
X
′ ′′
AD (p) := cs fs,1 fs,2 · · · fs,ms : l ∈ N, ms ∈ N, cs ∈ R, fs,i ∈ {p, p , p . . .} for every i and s .
s=0
Notice that if ms = 0, then by convention the product fs,1 fs,2 · · · fs,ms = 1.
Remark 6.1. This definition comes from the calculations of the successive partial derivatives
Fx , F in the simple case when g0 = 1, g1 = 0, g2 = −1, by setting Z =
xx ,R. . . For example,
T −t 2
exp − 0 p (Xs ) ds , we have F (t, x) = E [Z|X0 = x] by the Markov property of X. Since
we can write Xs = ebs X0 + Ys , where Ys does not depend on X0 , this means formally, Fx =
R T −t bs ′
E [Zx |X0 = x], where Zx = −Z 0 2e pp (Xs ) ds and pp′ ∈ AD (p) appears. Higher order
derivatives like Zxx , Zxxx will also formally generate the elements in AD (p).
Now we prove some properties of power series which are negligible to double factorial. For the
lemma below, we recall the convention (−1)! ! = (0)! ! = 1. Notice that although the proof is a little
ℓ Pℓ
cumbersome, the essential idea is the observation that ℓ! ! behaves like (⌊ 2ℓ ⌋)! 2 2 and k=0 kℓ = 2ℓ .

Lemma 6.2. There exists a constant C > 0, such that for all ℓ ∈ N, we have that
ℓ ℓ
X 1 Cℓ3 2 2
≤ .
(k − 1)! ! (ℓ − k − 1)! ! (ℓ − 1)! !
k=0
19
1 1
Proof. For k = 0, ℓ, we have (k−1)!!(ℓ−k−1)!! = (ℓ−1)!! . So we need only to prove that for a constant
C > 0,
ℓ−1 ℓ−2 ℓ
X 1 X 1 Cℓ3 2 2
= ≤ .
(k − 1)! ! (ℓ − k − 1)! ! k! ! (ℓ − 2 − k)! ! (ℓ − 1)! !
k=1 k=0
Case 1: ℓ = 2ℓ1 is an even number. If k = 2k1 is an even number, k! ! (ℓ − 2 − k)! ! = 2ℓ1 −1 k1 ! (ℓ1 −
1 − k1 )!. If k = 2k1 + 1 is an odd number, k! ! (ℓ − 2 − k)! ! = (2k1 + 1)! ! (ℓ − 3 − 2k1 )! ! ≥
ℓ1 −1
(2k1 +1)!!(ℓ−1−2k1 )!! k1 !(ℓ1 −1−k1 )!
ℓ−1−2k1 ≥ (2k1 )!!(ℓ−2−2k
ℓ
1 )!!
=2 ℓ . Therefore,
ℓ−2 1 −1
ℓX
X 1 ℓ+1 ℓ+1 (ℓ + 1)2ℓ1 −1
≤ = =
k! ! (ℓ − 2 − k)! ! 2ℓ1 −1 k1 ! (ℓ1 − 1 − k1 )! (ℓ1 − 1)! (ℓ − 2)! !
k=0 k1 =0
ℓ
(ℓ + 1)(ℓ − 1)2ℓ1 −1 ℓ2 2 2
≤ ≤ .
(ℓ − 1)! ! (ℓ − 1)! !
Case 2: ℓ = 2ℓ1 + 1 is an odd number, then

ℓ−2 ℓ−3
X 1 X 1 1
≤ + ,
k! ! (ℓ − 2 − k)! ! k! ! (ℓ − 3 − k)! ! (ℓ − 2)! !
k=0 k=0
and since ℓ − 3 is even, as the Case 1, we have, for a constant C > 0:

ℓ−3 ℓ−1 ℓ
X 1 1 (ℓ − 1)2 2 2 1 Cℓ3 2 2
+ ≤ + ≤ .
k! ! (ℓ − 3 − k)! ! (ℓ − 2)! ! (ℓ − 2)! ! (ℓ − 2)! ! (ℓ − 1)! !
k=0
ℓ
Cℓ3 2 2
Pℓ−2 1
Therefore there exists constant C such that for all ℓ, k=0 k!!(ℓ−2−k)!! ≤ (ℓ−1)!! .
Lemma 6.3. The following statements are true.
Rx
(i) If p is negligible to double factorial, then so is q(x) = 0
p(y)dy.
(ii) The set of power series which are negligible to double factorial is closed under addition,
differentiation, scalar multiplication, and multiplication. Thus if p is negligible to double
factorial, then for any q ∈ AD (p), q is also negligible to double factorial.
pk xk is negligible to double factorial, then so is q(x) = pk ck xk , where c is a

P P
(iii) If p(x) =
constant.
(iv) If p is negligible to double factorial, then so is |p|.
Proof.
P∞ pk−1 k
(i) Note that q(x) = k=1 k x . It follows that
1 (k − 1)! ! k−1
1 k−1
lim sup(|qk |(k − 1)! ! ) k = lim sup((|pk−1 | ) ) k
k→∞ k→∞ k
1 k−1
≤ lim sup((|pk−1 |(k − 2)! ! ) k−1 ) k = 0,
k→∞
(k−1)!! k!!
using that k ≤ k = (k − 2)! !.
20
P∞
(ii) Let p, q be negligible to double factorial and note that q ′ (x) = k=0 (k + 1)qk+1 xk . Then,
1 1 k+1
lim sup(|qk′ |(k − 1)! ! ) k = lim sup((|qk+1 |(k + 1)! ! ) k+1 ) k ,
k→∞ k→∞
then by (k + 1)! ! ≤ (k + 1)k! !,

1 1 k+1 1
lim sup(|qk′ |(k − 1)! ! ) k ≤ lim sup((|qk+1 |k! ! ) k+1 ) k (k + 1) k = 0.
k→∞ k→∞
1 1 1 1
By |a + b| k ≤ |a| k +|b| k , lim supk→∞ (|qk + pk |(k − 1)! ! ) k = 0, and lim supk→∞ (|cqk |(k −
1
1)! ! ) k = 0 for any constant c.
Next, we will show that pq is again negligible to double factorial. For any ϵ > 0, there exists
N , such that for all k > N , we have that (|qk |(k − 1)! ! ) ≤ ϵk , (|pk |(k − 1)! ! ) ≤ ϵk . Therefore
there exists constant Cϵ such that (|qk |(k − 1)! ! ) ≤ Cϵ ϵk , (|pk |(k − 1)! ! ) ≤ Cϵ ϵk for all k.
Thus by Lemma 6.2,
l
X
|(pq)l |= | pk ql−k |
k=0
l l √
X X Cϵ 2 ϵl Cϵ 2 C( 2ϵ)l l3
≤ |pk ||ql−k |≤ ≤ ,
(k − 1)! ! (l − k − 1)! ! (l − 1)! !
k=0 k=0
√ 1
thus lim supl→∞ (|(pq)l |(l − 1)! ! ) ≤ 2ϵ. Since ϵ can be any arbitrary positive number, we
l
obtain
1
lim sup(|(pq)l |(l − 1)! ! ) l = 0.
l→∞
Therefore, pq is also negligible to double factorial.

(iii) Notice that
1 1
lim sup(|qk |(k − 1)! ! ) k = |c|lim sup(|pk |(k − 1)! ! ) k = 0.
k→∞ k→∞
1
(iv) It is obvious, since the modulus of p has no impact on the lim supk→∞ (|pk |(k − 1)! ! ) k .
6.2 Regularity in space
In this subsection, we prove the regularity of f in x given by (6.1), i.e. f is C ∞ in x such that
k
the partial derivatives ∂∂xfk are bounded in a sense in preparation for the Feynman-Kac formula in
Subsection 6.3.
k
Theorem 6.4. The function f given by (6.1) is C ∞ in x, i.e. the partial derivatives ∂∂xfk exist
and are continuous in (t, x), for k ∈ N. Furthermore, for each k ∈ N, there exists a power series
qk negligible to double factorial and a constant Ck independent of (t, x), such that
∂kf
(t, x) ≤ |qk |(|x|) + Ck , t ∈ [0, T ], x ∈ R. (6.2)
∂xk
Before proving Theorem 6.4, we first simplify the expression of f in (6.1) to get rid of the condi-
tioning. Thanks to the Markovianity of X, we can write f after a change of variable as:
f (t, x) = E [Z(t, x)] , (6.3)
21
with
Z T −t 3
X
Z(t, x) := exp ui (t + s)pi ebs x + w(s) + W
fs ds
0 i=1

+ v(T )q(eb(T −t) x + w(T − t) + W
fT −t ) . (6.4)
ft when X0 = x, with w(t) := a t eb(t−u) du and W

Notice Xt = ebt x + w(t) + W ft := c t eb(t−u) dWu .
R R
0 0
From now on, we will consider mainly the representation (6.3) for f (t, x).
The main idea for proving Theorem 6.4 consists in taking successive derivatives in x inside the
expectation above and applying the dominated convergence theorem. To illustrate this, we first
calculate formally ∂f
∂x . Notice that the derivative in x for the terms inside the exponential of Z(t, x)
is:
Z T −t 3
X
h(x) := ui (t + s)ebs p′i ebs x + w(s) + W
fs ds
0 i=1

b(T −t) ′
+ v(T )e q eb(T −t) x + w(T − t) + W
fT −t , (6.5)
so that one expects

∂f
(t, x) = E [h(x)Z(t, x)] . (6.6)
∂x
Notice that |Z(t, x)|≤ 1 since the terms inside the exponential of Z(t, x) have a non-positive real
part as a result of Lemma 3.9. Therefore we only need to bound h(x + θ∆x) for θ ∈ [0, 1] and small
enough ∆x ∈ R. Notice that ui , v are continuous and thus bounded on [0, T ], ebs is also bounded
in [0, T ], p′i , q ′ are negligible to double factorial by the statement (ii) of Lemma 6.3. So we only
need to bound the expressions of the following form
Z t
q ebs x + w(s) + W
fs ds and q ebt x + w(t) + W ft , (6.7)
0
where q is any power series negligible to double factorial.
First of all, we introduce a new definition and a lemma to help bound in a certain sense the
ft∗ := sups≤t |W
quantities in (6.7). We also introduce the notation W fs |, which will be useful for
applying Doob’s inequality.
Definition 6.5. We say that a family of processes (Mt (x))t≤T,x∈R is estimable if there exists a
power series q negligible to double factorial such that for all fixed t ∈ [0, T ] and x ∈ R,
fT∗ ),
|Mt (x)|≤ |q|(|x|) + |q|(W a.s.
We will prove that the expressions in (6.7) are estimable in Proposition 6.7 below. For now, let us
f ∗ ) is integrable via the following lemma.
first prove that |q|(|x|) + |q|(W T
Lemma 6.6. The set of estimable family of processes is closed under h addition
i and multiplication.
∗
In addition, for all power series q negligible to double factorial, E |q|(WT ) < ∞. Therefore, if
f
Mt (x) is estimable, then there exists a power series q negligible to double factorial and a constant
C, such that for all t ∈ [0, T ] and x ∈ R, E[|Mt (x)|] ≤ |q|(|x|) + C.
Proof. For the first part of this lemma, recall that by the statements (ii) and (iv) of Lemma 6.3, the
property of being negligible to double factorial is closed under addition, multiplication, and taking
22
absolute power series. So the set of estimable family of process is obviously closed under addition.
For the multiplication, it suffices to use the basic inequality (a + b)(c + d) ≤ a2 + b2 + c2 + d2 .
h i
f ∗ ) < ∞, recall W ft = ebt t e−bs cdWs , where t e−bs cdWs is a true mar-
R R
To show that E |q|(W T 0 0
tingale. There is a positive constant CT such that C1T < ebt < CT when 0 ≤ t ≤ T . We use W t
Rt ∗
to denote 0 e−bs cdWs and W t to denote its maximal process sups≤t |W s |. By Doob’s maximal
inequality for k ≥ 2,
k
h
∗ k
i
k
h ∗ i
k k k h i
E |W T |k ≤ 4CTk E |W T |k ≤ 4CT2k E |W
fT |k .

E (WT ) ≤ CT E (W T ) ≤ CT
f
k−1
For k = 1,
fT∗ ≤ 1 1 + E (W
h i h i h i
E W fT∗ )2 ≤ 1 + 4CT4 E (W
fT )2 .
2
Notice that W
fT is a centred normal distribution and that the constant CT can be taken large
h i
fT |k ≤ (C 2 ) k2 (k − 1)! !.
fT is smaller than C 2 . It follows that E |W
enough so that the variance of W T T
|q|0 , |q|1 , |q|2 in the power series |q|, which does not influence
By modifying a little the coefficients P
∞
the conclusion, we obtain that C = 4 k=0 |q|k (k − 1)! ! CT3k satisfies the requirements and is finite,
since q is negligible to double factorial.
Proposition 6.7. If q is negligible to double factorial, then the expressions in (6.7) are estimable,
and
Z t h i
E q e x + w(s) + Ws ds + E q ebt x + w(t) + W
bs f ft ≤ |q̃|(|x|) + C, t ∈ [0, T ], x ∈ R,
0
where q̃ is negligible to double factorial and only depends on T, q, and C is a constant.

Proof. Fix t ∈ [0, T ] and x ∈ R. We first bound q ebt x + w(t) + W
ft :
X∞
q ebt x + w(t) + W
ft ≤ |q| ebt |x|+|w(t)|+|W
ft | ≤ |q|k (3ebt |x|)k + |3w(t)|k +|3W
ft |k
k=0
by Lemma 2.2. Since ebt and w(t) are both bounded in [0, T ], there exists a constant CT ≥ 0 such
that
∞
X ∞
X
bt k k k
|q|k (3e |x|) + |3w(t)| +|3Wt | ≤
f |q|k |3CT x|k +(3CT )k + |3W
ft |k
k=0 k=0
fT∗ ).
≤|q|(|3CT x|) + |q|(|3CT |) + |q|(3W

Setting q̃(x) = |q|(3(CT + 1)x) + |q|(|3CT |), we get q ebt x + w(t) + W ft ≤ q̃(|x|) + q̃(Wf∗ ) =
T
f ∗ ). By Lemma 6.3, |q| is also negligible to double factorial, and q̃ is still negligible
|q̃|(|x|) + |q̃|(W T
to double factorial. So q ebt x + w(t) + W ft is estimable. An application of Lemma 6.6 yields the
Rt
bound for the expectation. Furthermore, the term 0 q ebs x + w(s) + W fs ds is clearly bounded
f ∗ ). So we can update q̃ to q̃(x) = T ∨ 1|q|(3(CT + 1)x) + T ∨ 1|q|(|3CT |) to
by T |q̃|(|x|) + T |q̃|(W T
obtain the claimed bound and end the proof.
Rt
We have now that | 0 q ebs (x + θ∆x) + w(s) + W f ∗ ) ≤ |q̃|(|x|+1) +
fs ds|≤ |q̃|(|x + θ∆x|) + |q̃|(W
T
f ∗ ) since we can choose that |∆x|≤ 1, θ ∈ [0, 1] with coefficient of |q̃| all positive. The
|q̃|(W T
f ∗ ) is integrable. For the term q ebt (x + θ∆x) + w(t) + W
dominating function |q̃|(|x|+1)+|q̃|(W ft ,
T
we can build a dominating function in a similar way.
23
Going back to (6.6), we now have all the ingredients to apply the dominated convergence theorem
when ∆x → 0 on

f (t, x + ∆x) − f (t, x) Z(t, x + ∆x) − Z(t, x)
=E = E [h(x + θ∆x)Z(t, x + θ∆x)] ,
∆x ∆x
where θ ∈ [0, 1]. Recall h(x) from (6.5), and that |Z(t, x)|≤ 1 and ui , v are bounded, ebs is bounded
in [0, T ], p′i , q ′ are negligible to double factorial. Therefore, (6.6) holds. Of course, we would like
to prove by induction that
∂kf
(t, x) = E [Hk (t, x)Z(t, x)]
∂xk
holds for all k ∈ N to show that f is C ∞ in x , where
∂Hk (t, x)
Hk+1 (t, x) = + Hk (t, x)h(x), H0 (t, x) = 1, H1 (t, x) = h(x). (6.8)
∂x
To achieve this, we need to prove first that Hk are well-defined. Then, similarly to before, we will
bound Hk (x) so that we can apply by induction the dominated convergence theorem to
∂k f k
∂xk
+ ∆x) − ∂∂xfk (t, x)
(t, x
lim
∆x→0
∆x
Hk (t, x + ∆x)Z(t, x + ∆x) − Hk (t, x)Z(t, x)
= lim E
∆x→0 ∆x
= lim E [Hk+1 (t, x + θ∆x)Z(t, x + θ∆x)] .
∆x→0
For H2 (t, x), we need to compute ∂h(x)

∂x , recall that h(x) is just a sum of the Riemann integral of
the continuously differentiable function and a differentiable function with respect to x for a fixed
ω outside a null-set. Therefore,
Z T −t 3
∂h(x) X
= ui (t + s)e2bs p′′i ebs x + w(s) + W
fs ds
∂x 0 i=1

2b(T −t) ′′
+ v(T )e q eb(T −t) x + w(T − t) + W
fT −t .
Thus H2 can be obtained explicitly, and similarly for H3 , H4 , . . . , Hk , . . .. This procedure will
differentiate many times the function h(x), so it is useful to define:
Z T −t 3
(k)
X
hk (x) : = ui (t + s)ekbs pi ebs x + w(s) + W
fs ds
0 i=1

+ v(T )ekb(T −t) q (k) eb(T −t) x + w(T − t) + W
fT −t .
Definition 6.8. We define the set Ax (h):

( l )
X
Ax (h) := cs hs,1 hs,2 · · · hs,ms : l ∈ N, ms ∈ N, cs ∈ R, hs,i ∈ {h}k for every i and s .
s=0
as the R-algebra generated by higher order derivatives of the function h(x). We call hk generating
elements of Ax (h).
Notice that if ms = 0, then by convention the product hs,1 hs,2 · · · hs,ms = 1.
In the next Lemma 6.9, we will characterize Hk defined in (6.8).
24
Lemma 6.9. For every k ∈ N, the function Hk in (6.8) is well-defined. In addition, Hk is
differentiable in x, continuous in (t, x) and estimable.
Proof. First, note the following three facts:
(i) h(x) = h1 (x) is a generating element of Ax (h),
(ii) H0 = 1 ∈ Ax (h),
(iii) All generating elements hk (x) of Ax (h) are differentiable in x by applying the Leibniz’s rule,
and ∂h∂x
k (x) ∂g
= hk+1 (x). So if g ∈ Ax (h), then ∂x exists and ∈ Ax (h).
By induction and noticing that Ax (h) is closed under linear sum and multiplication, Hk are well-
defined and in Ax (h). For fixed ω ∈ Ω outside a null-set, the generating elements hk are continuous
in (t, x) and differentiable in x. This implies that Hk is also continuous in (t, x) and differentiable
in x.
(k)
We know ui , v are continuous (and thus bounded), ekbs is bounded in [0, T ], pi , q (k) are negligible
to double factorial by the statement 2 of Lemma 6.3. By Proposition 6.7, generating elements hk
of Ax (h) are estimable. Since Hk ∈ Ax (h) is a linear sum of finite product of generating elements
of Ax (h), by Lemma 6.6 Hk is also estimable.
We are now ready to prove Theorem 6.4.
Proof of Theorem 6.4. Take Hk as defined in (6.8) which are estimable by Lemma 6.9. By applying
Lemma 6.6,the term Hk (t, x)Z(t, x) is estimable with its expectation bounded by |qk |(|x|) + Ck .
∂k f
Using induction, we first notice that for the case k = 0, ∂x = |f (t, x)|≤ E[|Z(t, x)|] ≤ 1 which
trivially satisfies the inequality in (6.2). Next, suppose that it is true for up to case k, recall the
definition of Z from (6.4) and that Zx = hZ, where h is defined in (6.5). Thus (Hk Z)x = Hk+1 Z.
Since both Z and Hk+1 are differentiable in x as per 6.9, by the Mean value theorem:
1
Hk (t, x + ∆x)Z(t, x + ∆x) − Hk (t, x)Z(t, x) = Hk+1 (t, x + θ∆x)Z(t, x + θ∆x),
∆x
where θ is ∈ [0, 1]. Strictly speaking, since Hk Z is a complex-valued functions, we need to apply
the Mean value on both the real part and imaginary part separately with different θ. However,
this does not change the proof at all, so to simplify the notation and discussion, only θ is used.
By Lemma 2.2, Lemma 6.9 and the fact that |Z|≤ 1, and Hk+1 (t, x + θ∆x) is dominated by
f ∗ ) with bounded expectation by Lemma 6.6 if we choose that |∆x|≤ 1,
|qk+1 |(|x|+1) + |qk+1 |(W T
where qk+1 is negligible to double factorial. Applying dominated convergence theorem, for ∆x → 0
we have:
" #
∂ k+1 f
(t, x) = E Hk+1 (t, x)Z(t, x) .
∂xk+1
Therefore, for all k ∈ N, where qk negligible to double factorial and a constant Ck , we have that
k
(6.2) holds. Finally, for the continuity of ∂∂xfk (t, x), it suffices to notice that before taking the
expectation, the random variable Hk (t, x)Z(t, x) is continuous with respect to (t, x), Then fixing
(t0 , x0 ), again by Proposition 6.7, Lemma 6.9, its expectation is uniformly bounded with t ≤ T ,
and |x|≤ |x0 |+1 bounded. So again by the dominated convergence theorem and taking the limit
at (t0 , x0 ), the continuity holds.
25
6.3 Feynman-Kac
In this subsection, we derive the Feynman-Kac formula in Theorem 6.11. Since we do not have
that ft exists a priori in our setting, we shall prove its existence and obtain the Feynman-Kac
formula at the same time.
First, we introduce a lemma which will be useful later.

Lemma 6.10. Let z ∈ C such that ℜ(z) ≤ 0, then |exp(z) − 1|≤ 3|z|.
Proof. We write z = x+yi, x ≤ 0 and y ∈ R. Then |exp(z)−1|= |exp(x+yi)−1|= |exp(yi)(exp(x)−

1) + exp(yi) − 1|≤ |exp(x) − 1|+|cos(y) − 1|+|sin(y)|.
If x = 0, then |exp(x) − 1|= 0. If x < 0, | exp(x)−1

z |≤ | exp(x)−1
x |≤ 1.
If y = 0, then |cos(y) − 1|= 0, |sin(y)|= 0. If y ̸= 0, | cos(y)−1

z |≤ | cos(y)−1
y |≤ 1, | sin(y) sin(y)
z |≤ | y |≤
1.
Theorem 6.11. The function f given by (6.1) and equivalently (6.3) is C ∞ in x, C 1 in t, and
solves the following partial differential equation (PDE):
( P3
ft (t, x) + fx (t, x)(a + bx) + 12 c2 fxx (t, x) + f (t, x) i=1 ui (t)pi (x) = 0,
(6.9)
f (T, x) = exp(v(T )q(x)).
Proof. By Theorem 6.4, f is C ∞ in x. We will compute the infinitesimal generator

1 x
lim (E [f (t0 , Xr ) − f (t0 , x)])
r→0+ r
in two ways to obtain the PDE and the existence of ft at the same time. Here Ex means the
conditional expectation with X0 = x.
First way: fix t0 , we compute the quantity using the classical definition of generator. By Itô’s
formula:

1 1 2
df (t0 , Xt ) = fx dXt + fxx d⟨X⟩t = cfx dWt + c fxx + (a + bXt )fx dt.
2 2
Note that fx , fxx are evaluated at (t0 , Xt ) here. By Theorem 6.4, statement (ii) of Lemma 6.3,
fx (t, x), (a + bx)fx (t, x), fxx (t, x), fx2 (t, x) are all bounded
R r by |q|(|x|) + C, where q is negligible to
double factorial. Combining with Proposition 6.7, Ex 0 fx2 (t0 , Xt )dt is bounded by |q̃|(|x|) + C
Rr
for r ∈ [0, T ], where q̃ is negligible to double factorial. Thus 0 cfx (t0 , Xt )dWt is in L2 and is a true
Rr 1 2
martingale for t ∈ [0, T ]. Similarly, the Riemann integral 0 ( 2 c fxx (t0 , Xt )+(a+bXt )fx (t0 , Xt ))dt
is estimable with finite expectation. In addition, we have
1 r 1 2
Z
1
( c fxx (t0 , Xt ) + (a + bXt )fx (t0 , Xt )dt ≤ sup c2 |fxx (t0 , Xs )|+|(a + bXs )fx (t0 , Xs )|,
r 0 2 s∈[0,r] 2
where the right hand side is still estimable and dominated by a random variable with finite expec-
tation. By dominated convergence theorem,
1 x
lim (E [f (t0 , Xr ) − f (t0 , x)])
r
r→0+
Z r
1 1
= lim+ Ex ( c2 fxx (t0 , Xt ) + (a + bXt )fx (t0 , Xt ))dt
r→0 r 0 2
1 2
= c fxx (t0 , x) + (a + bx)fx (t0 , x)). (6.10)
2
26
Second way: compute the generator limr→0+ 1r (Ex [f (t0 , Xr ) − f (t0 , x)]) directly by applying the
R t P3
Markov property of Xt . Define Zt := exp( 0 i=1 ui (T − t + s)pi (Xs ) ds) and Yt = v(T )q(Xt ), we
apply Markov property of Xt to the representation of f (t, x) in (6.1), i.e.
" 3
Z T −t X ! #
f (t, x) = E exp ui (t + s)pi (Xs )ds + v(T )q(XT −t ) X0 = x ,
0 i=1
and by Markov property for r ∈ [0, t],

" Z T −t+r X 3
! #
f (t, Xr ) =E exp ui (t − r + s)pi (Xs )ds + v(T )q(XT −t+r ) Xr .
r i=1
By the tower property of conditional expectation, we have

" " Z T −t+r X3
! ##
x x
E [f (t, Xr )] =E E exp ui (t − r + s)pi (Xs )ds + v(T )q(XT −t+r ) Xr
r i=1
" 3
!#
Z T −t+r X
x
=E exp ui (t − r + s)pi (Xs )ds + v(T )q(XT −t+r )
r i=1
" 3
! #
Z r X
=Ex ZT −t+r · exp − ui (t − r + s)pi (Xs ) ds exp (YT −t+r ) .
0 i=1
Therefore,
1 x
(E [f (t0 , Xr ) − f (t0 , x)])
r " ! #
Z rX 3
1 x
= E ZT −t0 +r exp − ui (t0 − r + s)pi (Xs ) ds exp (YT −t0 +r ) − ZT −t0 exp(YT −t0 )
r 0 i=1
" #
1 x
= E exp (YT −t0 +r ) ZT −t0 +r − exp (YT −t0 ) ZT −t0
r
" Z rX 3
! !#
1 x
+ E exp (YT −t0 +r ) ZT −t0 +r exp − ui (t0 − r + s)pi (Xs ) ds − 1
r 0 i=1
1
= (f (t0 − r, x) − f (t0 , x))
r " ! !#
Z r 3
1 X
+ Ex exp (YT −t0 +r ) ZT −t0 +r exp − ui (t0 − r + s)pi (Xs ) ds −1 .
r 0 i=1
Now we want to apply dominated convergence theorem when r → 0+ to the term

" Z rX3
! !#
1 x
E exp (YT −t0 +r ) ZT −t0 +r exp − ui (t0 − r + s)pi (Xs ) ds − 1 .
r 0 i=1
To dominate the term inside the expectation, notice that at first ℜ(Yt ) = ℜ(v(T )q(Xt )) = 0, thus
|exp(YT −t0 +r )|= 1. In addition,
Z rX 3
! !
1
ZT −t0 +r exp − ui (t0 − r + s)pi (Xs ) ds − 1
r 0 i=1
Z T −t0 +r X3
! Z rX3
!!
1
= exp ui (t0 − r + s)pi (Xs ) ds 1 − exp ui (t0 − r + s)pi (Xs ) ds .
r r i=1 0 i=1
27
Since ℜ(u1 ) ≤ 0, ℜ(u2 ), ℜ(u3 ) = 0, p1 = p2 , therefore
Z T −t0 +r 3
X Z T −t0 +r
ℜ ui (t0 − r + s)pi (Xs ) ds = ℜ u1 (t0 − r + s)p1 (Xs ) ds ≤ 0
r i=1 r
R
T −t0 +r P3
and thus exp r i=1 ui (t 0 − r + s)p i (Xs ) ds ≤ 1. Therefore we only need to dominate
the term ! !
Z r 3
1 X
exp ui (t0 − r + s)pi (Xs ) ds −1 .
r 0 i=1
R r P3
Since ℜ( 0 i=1 ui (t0 − r + s)pi (Xs ) ds) ≤ 0, by Lemma 6.10 we just need to dominate the term
Z r 3 3
1 X X
ui (t0 − r + s)pi (Xs ) ds ≤ ||ui ||∞ sup |pi (Xs )|.
r 0 i=1 i=1 s∈[0,r]
Notice |ui (s)| are bounded, pi are negligible to double factorial so |pi (Xs )|, s ∈ [0, T ] is dominated
when conditioned X0 = x by Proposition 6.7. Thus by the dominated convergence theorem :
" Z rX 3
! !#
1 x
lim E exp (YT −t0 +r ) ZT −t0 +r exp − ui (t0 − r + s)pi (Xs ) ds − 1
r→0+ r 0 i=1
3
X
=− ui (t0 )pi (x)f (t0 , x)
i=1
which is finite. Equating with (6.10), the term limr→0+ 1r (f (t0 −r, x)−f (t0 , x)) is thus well-defined.
Of course, one can also replace t0 by t0 + r and perform similar computations as per the above
two methods on the quantity
1
lim+ (Ex [f (t0 + r, Xr ) − f (t0 + r, x)])
r→0 r
to obtain the existence of ft and also the PDE in (6.9). A subtle point to note is that for the
first method, we do not apply Itô’s formula to f (t0 + r, Xr ) for the variable r, which requires the
existence of the partial derivative ft a priori. Instead, we obtain the existence of ft by applying
Itô’s formula to Xr in f (t0 + u, Xr ) and evaluate at u = r:
1 x
E [f (t0 + r, Xr ) − f (t0 + r, x)]
r Z
r
1 x 1
= E ( c2 fxx (t0 + r, Xt ) + (a + bXt )fx (t0 + r, Xt ))dt ,
r 0 2
and taking the limit for r → 0+ as above. The continuity of ft is directly obtained from the
continuity of the other terms in this equation, with boundary condition at T comes from the
continuity of f and its definition.
6.4 Infinite dimensional ODE
In this subsection, we obtain a solution of the system of ODE by comparing the derivatives of both
sides of the PDE (6.9) when x = 0.
Since f is continuous by Theorem 6.4, in addition if f does not vanish, we can define log f such
that exp(log f ) = f and log f (T, 0) = v(T )q(0), see Lemma B.1. Note log f is also continuous in
(t, x), and exp(log f )(T, x) = f (T, x) = exp(v(T )q(x)) with log f (T, x) = v(T )q(x).
28
Theorem 6.12. If f does not vanish, then
1 k
ϕk (t) := ∂ log f (T − t, x) , t≥0
k! x x=0
solves the system of ODE

3
X
ϕ′k (t) = ui (T − t)(pi )k
i=1
c2 (k + 2)(k + 1)
+ bkϕk (t) + a(k + 1)ϕk+1 (t) + ϕk+2 (t)
2
c2 e
+(ϕ(t) ∗ ϕ(t))
e k, ϕek (t) = (k + 1)ϕk+1 (t),
2
ϕk (0) = v(T )qk . (6.11)
In order to prove the Theorem, we first link the PDE (6.9) to the system of ODE (6.11). Suppose
f does not vanish and set g(t, x) = log f (T − t, x), then the following PDE holds:
( P3
gt (t, x) = i=1 ui (T − t)pi (x) + gx (t, x) (a + bx) + 21 c2 gxx (t, x) + 12 c2 gx2 (t, x),
g(0, x) = v(T )q(x).
Since by Theorem 6.11, f is C 1 in t and C ∞ in x, by Lemma B.2, log f is also C 1 in t and C ∞ in

x. Then so is g.
To prove this theorem, we will need the following lemma.

∂2h ∂2h
Lemma 6.13. If a function h is C 1 in (t, x), and one of the partial derivatives ∂t∂x and ∂x∂t
exists and is continuous, then they both exist and are equal.
Now we can prove Theorem 6.12.
Proof of Theorem 6.12. Since F (t, x) does not vanish, we can define log F as per Definition B.5,
i.e.
exp(log F (t, x)) = F (t, x), log F (T, 0) = 0.
We can now define log f (t, x) = log F (t, x) + v(t)q(x) such that exp(log f (t, x)) = f (t, x). By
Theorem 6.11, f is C ∞ in x and C 1 in t, then since log f is continuous, by Lemma B.2, log f is
also C ∞ in x and C 1 in t. In addition, the PDE below can also be deduced:
3
X 1 1
−(log f )t (t, x) = ui (t)pi (x) + (log f )x (t, x) (a + bx) + c2 (log f )xx (t, x) + c2 (log f )2x (t, x)
i=1
2 2
log f (T, x) = v(T )q(x).
The boundary condition comes from the fact log f (T, 0) = v(T )q(0), with the continuity of log f
and the fact that exp(log f ))(T, x) = f (T, x) = exp(v(T )q(x)).
Now substitute g(t, x) = log f (T − t, x):

3
X 1 1
gt (t, x) = ui (T − t)pi (x) + gx (t, x) (a + bx) + c2 gxx (t, x) + c2 gx2 (t, x) (6.12)
i=1
2 2
g(0, x) = v(T )q(x).
29
k
Since the right side of (6.12) is C ∞ in x, so is gt . Notice that by definition ∂x
∂
k g(t, x)|x=0 = k! ϕk (t).
th
We aim to take the k partial derivative with respect to x on both sides of (6.12) at x = 0 to
deduce the equation in (6.11). For the left side of (6.11), Lemma 6.13 allows the interchangeability
of ∂t, ∂x and the following equality can be proven by induction:
∂k ∂ ∂ ∂k
g(t, x)| x=0 = g(t, x)|x=0 = k! ϕ′k (t).
∂xk ∂t ∂t ∂xk
k
∂
The right side of (6.11) can be deduced by noticing that ∂x k g(t, x)|x=0 = k! ϕk (t) for all k. To
obtain that
∂k 2 e ∗ ϕ(t))
g (t, x)|x=0 = k! (ϕ(t) e k , ϕek (t) = (k + 1)ϕk+1 (t),
∂xk x
∂k ∂k
first notice that ∂xk g(t, x)|x=0 = k! ϕk (t) implies ∂xk gx (t, x)|x=0 = k! ϕk (t). The convolution follows
e
naturally by applying the general Leibniz rule.
Lastly, the initial condition of (6.11) can be easily deduced from the fact that g(0, x) = v(T )q(x).
6.5 Putting everything together
Proof of Theorem 3.5. Since p is negligible to double factorial by assumption, and g0 : [0, T ] → R,
g1 , g2 : [0, T ] → C are continuously differentiable such that ℜ(g1 ) = 0, ℜ(g2 ) ≤ 0, then F (t, x) in
(3.1) is well-defined. We recall from Lemma 3.9 that
" Z TX3
! #
F (t, x) = E exp ui (s)pi (Xs )ds + v(T )q(XT ) − v(t)q(Xt ) Xt = x ,
t i=1
with ui , v are continuous in [0, T ], ℜ(u1 ) ≤ 0, ℜ(u2 ), ℜ(u3 ), ℜ(v) = 0, p1 = p2 , and pi , q are
negligible to double factorial. Notice that F (t, x) = f (t, x) exp(−v(t)q(x)). Theorem 6.4 yields
that F is continuous.
If in addition, F does not vanish, then so does f . By Lemma B.1, we can define log f, log F such
that log F = log f − v(t)q(x). By Theorems 6.4 and 6.11, f is C ∞ in x and C 1 in t, and so is
F . By Lemma B.2, log f, log F are also C ∞ in x and C 1 in t. Then, we can take the k th partial
derivative with respect to x for log F = log f − v(t)q(x) around x = 0. Then, the system of ODE
(6.11) in Theorem 6.12 induces the system of ODE (3.3) by setting ψk := ϕk − v(T − t)qk with
ϕk defined in (6.11), with the coefficients of the ODE (3.3) coming from the precise definition of
(ui , pi )i∈{1,2,3} as defined in Appendix A.
Before proving Theorem 3.13, we introduce the following definition of the function fn :
Pn−1
Definition 7.1. For µn = Tn i=0 δ iT , with δt the Dirac measure at point t, we define
n
" 3
! #
Z T X
fn (t, x) = E exp ui (s)pi (Xs )µn (ds) + v(T )q(XT ) Xt = x
t i=1
= Fn (t, x) exp(v(t)q(x)), t ≤ T, x ∈ R,
with Fn as defined in Definition 3.10.
30
By the Markov property of X, we can remove the conditioning and write fn (t, x) equivalently as
" 3
Z T −t X
fn (t, x) = E exp ui (s + t)pi (ebs x + w(s) + W
fs )µn (ds + t)
0 i=1
!#
T −t
+ v(T )q(e x + w(T − t) + WT −t ) .
f (7.1)
Notice that Xt = ebt x + w(t) + W ft for X0 = x, w(t) := a t eb(t−u) du and W ft := c t eb(t−u) dWu .
R R
0 0
R T −t
The function fn is well-defined, since 0 ℜ u1 (s)p2 (ebs x + w(s) + W
fs ) µ(ds + t) ≤ 0 given the
condition of ui , pi , v and q from Lemma 3.9.
The main advantage of considering functionals (fn , Fn ) over (f, F ) is their entire property in x
shown in Theorem 7.3 below. As we shall see, the results of (f, F ) in Section 6 can be easily
extended to to (fn , Fn ). We now layout the following steps to prove Theorem 3.13:
• In Subsection 7.1, we prove that fn is entire in x ∈ R, without any additional conditions

imposed on fn or Fn .
• In Subsection 7.2, we prove discretized versions of the results in Section 6, in the case when
fn (t, ·) (or equivalently Fn (t, ·)) does not vanish on R.
• In Subsection 7.3, we prove Theorem 3.13 under the assumption that the extension of
(Fn (t, ·)) to the complex plane (Fn (t, ·))c does not vanish.
Lemma 7.2. fn (t, x) → f (t, x), when n → ∞, for all t ≤ T and x ∈ R.
Proof. This follows from a direct application of the bounded convergence theorem: the Riemann
sum converges to the Riemann integral, and the exponential has a module at most 1.
7.1 Entire property
In this subsection, we do not need any additional condition imposed on fn or Fn about their zeros.
Theorem 7.3. For fixed t, fn (t, ·) is entire in x ∈ R.
We first introduce a lemma that will be useful in proving Theorem 7.3. The idea of the proof
is standard and comes from exploiting the fact that the Gaussian density is entire. A similar
argument has been used for single marginals in [13, Lemma 6.1].
Lemma 7.4. Suppose that (Wt )t≥0 is a centered Gaussian Process with independent increments
such that Wt − Ws has non-zero variance when t > s. Let 0 = t0 < t1 < . . . < tn = T . Take g a
bounded measurable function from Rn to C, then E [g(x + Wt1 , x + Wt2 , . . . , x + Wtn )] is entire in
x.
Proof. Denote the variance of Wti − Wti−1 as vi > 0, we have
E [g(x + Wt1 , x + Wt2 , . . . , x + Wtn )]

n n
!
2
(y1 − x) yi2
Z
Y 1 X
= √ g(y1 , y1 + y2 , . . . , y1 + . . . + yn ) exp − − dy1 dy2 . . . dyn .
i=1
2πvi Rn 2v1 i=2
2vi
31
Denote g(y) = g(y1 , y1 + y2 , . . . , y1 + . . . + yn ), we have
E [g(x + Wt1 , x + Wt2 , . . . , x + Wtn )]
2Y n n
!
yi2

−x
Z
1 X xy1
= exp √ g(y) exp − exp dy1 dy2 . . . dyn
2v1 i=1
2πvi Rn i=1
2vi v1
2Y n n
! ∞ k
yi2 X 1 xy1

−x
Z
1 X
= exp √ g(y) exp − dy1 dy2 . . . dyn
2v1 i=1
2πvi Rn i=1
2vi k! v1
k=0
∞
n n
! !
2Y k
−x yi2
Z
1 X 1 X y1
= exp √ g(y) exp − dy1 dy2 . . . dyn xk ,
2v1 i=1
2πv i k! R n
i=1
2v i v 1
k=0
where we applied Fubini in the last step since

∞ n
! k
yi2
Z X X 1 xy1
|g(y)| exp − dy1 dy2 . . . dyn
Rn k=0 i=1
2vi k! v1
n
!
yi2 |xy1 |
Z X
≤||g||∞ exp − + dy1 dy2 . . . dyn < ∞.
Rn i=1
2vi v1
This implies an infinite radius of convergence for

∞
n n
! !
2 Y k
−x yi2
Z
1 X 1 X y1
exp √ g(y) exp − dy1 dy2 . . . dyn xk
2v1 i=1
2πv i k! Rn
i=1
2v i v 1
k=0
i.e. E [g(x + Wt1 , x + Wt2 , . . . , x + Wtn )] is equal to a power series with an infinite radius of con-
vergence, so it is entire in x ∈ R.
Proof of Theorem 7.3. We define tj = jT

n , j = 0, 1, . . . , n − 1, then
" 3
X X
fn (t, x) = E exp ui (tj )pi (eb(tj −t) x + w(tj − t) + W
ft −t )
j
i=1 t≤tj ≤T
!#
b(T −t)
+ v(T )q(e x + w(T − t) + W
fT −t ) .
The exponential is bounded by 1. And notice that eb(tj −t) x + W ft −t = eb(tj −t) (x + e−b(tj −t) W
j
ft −t ).
j
R t b(t−u) −bt f
R t −bu
Recall that Wft =
0
e cdWu . Therefore, e W t = c 0
e dW u is a Gaussian Process sat-
isfying the condition of Lemma 7.4, since c ̸= 0. Then, an application of Lemma 7.4 yields the
result.
7.2 Discretized versions of the theorems in Section 6
In this subsection, we prove discretized versions of the results in Section 6, in the case when fn (t, ·)
(or equivalently Fn (t, ·)) does not vanish on R.
Remark 7.5. Recalling Corollary 3.6, the conditions that g1 = 0 and g2 R-valued with g2 ≤ 0
guarantee that fn (t, ·) and Fn (t, ·) do not vanish on R.
Theorem 7.6. (Discretized version of Theorem 6.4) The function fn defined in (7.1) is C ∞ in x
k
(j+1)T
with ∂∂xfkn continuous in (t, x) in the region ( jT
n , n ) × R for j = 0, 1, . . . , n − 1. In addition,
there exists a power series qk negligible to double factorial and constant Ck independent of (t, x),
k
such that | ∂∂xfkn |≤ |qk |(|x|) + Ck .
32
Proof. Notice that all the lemmas and proofs used in Section 6 dealing with f being C ∞ in x and
k
the bound of ∂∂xfkn can be easily adopted to fn by replacing the integral with a finite sum.
Theorem 7.7. (Discretized version of Theorem 6.11) For j = 0, 1, . . . , n − 1, the function fn
(j+1)T
(i) is continuous in the region ( jT
n , n ] × R. In addition, fn is C 1 in t and C ∞ in x in the
jT (j+1)T
region ( n , n ) × R, j = 0, 1, . . . , n − 1,
(ii) solves the following PDE
1
ft + fx (a + bx) + c2 fxx = 0, t∈/ supp(µn )∪{T },
2 !
3
T X
lim f (s, x) = f (t, x) exp − ui (t)pi (x) , t ∈ supp(µn ),
s→t+ n i=1
f (T, x) = exp(v(T )q(x)),
(n−1)T
with supp(µn ) = {0, Tn , 2T
n ,..., n }.
Proof. The proof follows along the same lines as the proof of Theorem 6.11. By Theorem 7.6 and
Lemma 6.3, (fn )x (t, x) and (fn )2x (t, x) are bounded by |q|(|x|) +
R ·C, where q is negligible to double
factorial. Therefore by Proposition 6.7, the stochastic integral 0 (fn )x (s, x)dWs ∈ L2 and is a true
martingale. The Riemann integral is still dominated, allowing the interchange between limit and
the expectation and thus computing the the infinitesimal generator limr→0+ 1r (Ex [fn (t0 , Xr )] −
fn (t0 , x)) as per the first way in Theorem 6.11.
The second way of computing the generator in Theorem 6.11 can also be adopted to compute
limr→0+ 1r (Ex [fn (t0 , Xr )] − fn (t0 , x)) thanks to the Markov property of Xt . Indeed, define
3
Z tX !
Ztn = exp ui (T − t + s)pi (Xs ) µn (T − t + ds) , Yt = v(T )q(Xt ).
0 i=1
we have
1 x
(E [fn (t0 , Xr )] − fn (t0 , x))
r " ! #
Z rX 3
1 x n n
= E ZT −t0 +r · exp − ui (t0 − r + s)pi (Xs ) µn (t0 − r + ds) exp (YT −t0 +r ) − ZT −t0 exp(YT −t0 )
r 0 i=1
1
= Ex exp (YT −t0 +r ) ZTn−t0 +r − exp (YT −t0 ) ZTn−t0

r " ! !#
Z rX 3
1 x n
+ E exp (YT −t0 +r ) ZT −t0 +r · exp − ui (t0 − r + s)pi (Xs ) µn (t0 − r + ds) − 1
r 0 i=1
1
= (f (t0 − r, x) − f (t0 , x))
r " ! !#
Z r 3
1 X
+ Ex exp (YT −t0 +r ) ZTn−t0 +r · exp − ui (t0 − r + s)pi (Xs ) µn (t0 − r + ds) −1 .
r 0 i=1
Notice that, if t0 ∈
/ supp(µn ) ∪ {T }, where supp(µn ) ∪ {T } is a finite set, then there exists r0 > 0
such that r0 < |t0 − t| for all t ∈ supp(µn ) ∪ {T }, thus for 0 < r < r0 ,
h R P i
r 3
Ex 1r exp (YT −t0 +r ) ZTn−t0 +r · exp − 0 i=1 ui (t0 − r + s)pi (Xs ) µn (t0 − r + ds) − 1 = 0.
33
So similar the proof of Theorem 6.11, (fn )t exists and 0 = (fn )t + (fn )x (a + bx) + 12 c2 (fn )xx . The
boundary condition when t ∈ / supp(µn ) ∪ {T } can be deduced by applying the definition of µn .
Remark 7.8. Theorem 7.7 shows that for fixed x, fn (t, x) is left continuous for t ∈ [0, T ] and is
discontinuous at t ∈ supp(µn ).
We now define log fn in the following Lemma 7.9.

Lemma 7.9. If fn (t, ·) does not vanish for x ∈ R, then there exists log f n such that exp(log fn ) =
(j+1)T
fn , and in the regions ( jT
n , n ] × R, j = 0, 1, . . . , n − 1, log fn (t, x) is continuous in (t, x), and
3
T X
lim log fn (s, x) = log fn (t, x) − ui (t)pi (x), t ∈ supp(µn ),
s→t+ n i=1
log fn (T, x) = v(T )q(x).
Proof. We want to use Lemma B.1 to define log fn from fn . However, Theorem 7.7 shows that fn
is discontinuous when t ∈ supp(µn ), with
3
!
T X
lim fn (s, x) = fn (t, x) exp − ui (t)pi (x) , t ∈ supp(µn ),
s→t+ n i=1
we need to first define

 
3
 T X X
fñ (t, x) = fn (t, x) exp 

− n ui (s)pi (x)
,
s∈supp(µn ) i=1
s≥t
where fñ is continuous on [0, T ] × R and fñ (T, x) = fn (T, x) = exp(v(T )q(x)). And since fn is non
zero, so is fñ . We can now apply Lemma B.1 to yield the existence of g̃ continuous on [0, T ] × R,
such that exp(g̃) = fñ , and to specify g̃, we choose that g̃(T, 0) = v(T )q(0). We can now define:
3
T X X
log fn (t, x) = g̃(t, x) + ui (s)pi (x). (7.2)
n
s∈supp(µn ) i=1
s≥t
P T
P3 jT (j+1)T
Since s∈supp(µn ),s≥t n i=1 ui (s)pi (x) is continuous in the regions ( n , n ] × R, for j =
0, 1, . . . , n − 1,the definition of log fn in (7.2) yields that
1. exp(log fn ) = fn .
(j+1)T
2. In the regions ( jT
n , n ] × R, j = 0, 1, . . . , n − 1, log fn (t, x) is continuous in (t, x).
P3
3. lims→t+ log fn (s, x) = log fn (t, x) − Tn i=1 ui (t)pi (x), t ∈ supp(µn ).
Furhtermore, since exp(log fn (T, x)) = exp(g̃(T, x)) = fñ (T, x) = fn (T, x) = exp(v(T )q(x)),
log fn (T, x) − v(T )q(x) ∈ {2kπi, k ∈ Z}. Noticing that log fn (T, x) = g̃(T, x) is continuous in x,
and log fn (T, 0) − v(T )q(0) = g̃(T, 0) − v(T )q(0) = 0, we deduce that log fn (T, x) = v(T )q(x).
We now introduce the discretized version of Theorem 6.12.
34
Theorem 7.10. Suppose that fn (t, ·) does not vanish on R, then
1 k
ϕn,k (t) = ∂ log fn (T − t, x) , t≥0
k! x x=0
solves the following system of ODE:
c2 (k + 2)(k + 1)
ϕ′n,k (t) = bkϕn,k (t) + a(k + 1)ϕn,k+1 (t) + ϕn,k+2 (t)
2
c2 e
+ (ϕn (t) ∗ ϕen (t))k , ϕen,k = ϕn,k+1 (k + 1), t ∈ {T − u : u ∈ / supp(µn )∪{T }}
2
3
T X
ϕn,k (t) = lim− ϕn,k (s) + ui (T − t)(pi )k , t ∈ {T − u : u ∈ supp(µn )}
s→t n i=1
ϕn,k (0) = v(T )qk . (7.3)
(j+1)T
Proof. By Theorem 7.7, in the regions ( jT n , n ) × R, for j = 0, 1, . . . , n − 1, fn is C 1 in t and
C in x. Since fn ̸= 0, thus by Lemma B.2, log fn in 7.9 is C 1 in t and C ∞ in x in these regions.
∞
Set gn (t, x) = log fn (T − t, x), gn then satisfies the following PDE:

1 1
gt (t, x) = gx (t, x) (a + bx) + c2 gxx (t, x) + c2 gx2 , t ∈ {T − u : u ∈/ supp(µn )∪{T }}
2 2
3
T X
lim− g(s, x) = g(t, x) − ui (T − t)pi (x), t ∈ {T − u : u ∈ supp(µn )}
s→t n i=1
g(0, x) = v(T )q(x).
Next, we apply Lemma 6.13 as in the proof of Theorem 6.12, and then comparing the derivatives
of both sides in x yields the system of ODE. Lastly, the initial condition of (7.3) can be easily
deduced from the fact that gn (0, x) = v(T )q(x).
7.3 Putting everything together
In this subsection, we will use the condition that the extension of fn in the the complex plane
(fn (t, ·))c , defined in 3.11 does not vanish for x ∈ C as opposed to ∈ R. This condition is both
necessary and sufficient to guarantee that log fn is entire in x ∈ R, see Lemma B.3 and Remark
B.4.
P
k
Lemma 7.11. If (fn )c (t, ·) does not vanish on C, then fn (t, x) = exp ϕ
k≥0 n,k (T − t)x ,
where ϕn,k are defined as in Theorem 7.10.
1 k
Proof. By Lemma B.3, log fn is entire on R, and notice that ϕn,k (t) = k! ∂x log fn (T − t, x) x=0
.
P
Proof of Theorem 3.13. By Lemma 7.11, fn (t, x) = exp k≥0 ϕn,k (T − t)xk , where ϕn,k are
defined as in Theorem 7.10.
We set ψn,k (t) = ϕn,k (t) − v(T − t)qk . By Lemma 3.9 and Lemma A.1, this is equivalent to
pk−1
ψn,k (t) = ϕn,k (t) − ρg1 (t)g0 (T − t) , k ≥ 1,
ck
ψn,0 (t) = ϕn,0 (t).
35
Therefore, by Lemma 7.2,
F (t, x) =f (t, x) exp(−v(t)q(x))

= lim fn (t, x) exp(−v(t)q(x))
n→∞
and, by Lemma 7.11,

 
X
fn (t, x) exp(−v(t)q(x)) = exp  ϕn,k (T − t)xk  exp(−v(t)q(x))
k≥0
 
X
= exp  (ϕn,k (T − t) − v(t)qk )xk 
k≥0
 
X
= exp  ψn,k (T − t)xk  .
k≥0
Recalling that Fn (t, x) = fn (t, x) exp(−v(t)q(x)) by Definition 7.1 ends the proof.
A Proof of Lemma 3.9
Lemma A.1. Let p be a power series with an infinite radius of convergence. Define the power
series r1 , r2 by
1 x
Z
1 1 2 ′
r1 (x) = − ((a + bx)p(x) + c p (x)), r2 (x) = p(y)dy, x ∈ R,
c 2 c 0
which again have an infinite radius of convergence. Then, for any continuously differentiable
function h : [0, T ] → C, it holds that
Z T Z T
h(s)p(Xs )dWs = (h(s)r1 (Xs ) − h′ (s)r2 (Xs )) ds + h(T )r2 (XT ) − h(t)r2 (Xt ).
t t
Proof. This follows from a straightforward application of Itô’s Lemma on the process (h(s)r2 (Xs ))s≤T
between t and T . Recall the dynamics of X in (2.1).
σt2
Proof of Lemma 3.9. Using d log St = − 2 dt + σt dBt , we have:
Z T Z T
1
g1 (T − s)d log Ss = − g1 (T − s)g02 (s)p2 (Xs )ds + ρg1 (T − s)g0 (s)p(Xs )dWs
t t 2
p
+ 1 − ρ2 g1 (T − s)g0 (s)p(Xs )dWs⊥ .
RT p
Conditioning on Ft ∨ FTW ,the integral t 1 − ρ2 g1 (T − s)g0 (s)p(Xs )dWs⊥ is Gaussian with con-
RT 2 2 2 2
ditional variance t (1 − ρ )g1 (T − s)g0 (s)p (Xs )ds. Therefore,
" Z Tp ! #
2 ⊥ W
E exp 1 − ρ g1 (T − s)g0 (s)p(Xs )dWs Ft ∨ FT
t
!
Z T
1
= exp (1 − ρ2 )g12 (T − s)g02 (s)p2 (Xs )ds ,
2 t
36
so that
" ! #
Z T Z T
F (t, Xt ) = E exp g1 (T − s)d log Ss + g2 (T − s)σs2 ds Ft
t t
" Z T
1 1
= E exp (1 − ρ )g1 (T − s) − g1 (T − s) + g2 (T − s) g02 (s)p2 (Xs )ds
2 2
t 2 2
! #
Z T
+ ρg1 (T − s)g0 (s)p(Xs )dWs Ft .
t
By applying Lemma A.1 and choosing h(t) = ρg1 (T − t)g0 (t), we can rewrite the following:
Z T Z T
ρg1 (T − s)g0 (s)p(Xs )dWs = h(s)r1 (Xs ) − h′ (s)r2 (Xs )ds + h(T )r2 (XT ) − h(t)r2 (Xt ).
t t
Next, define u1 (s) = ( 21 (1 − ρ2 )g12 (T − s) − 21 g1 (T − s) + g2 (T − s))g02 (s), p1 = p2 , u2 (s) = h(s), p2 =

r1 , u3 (s) = h′ (s), p3 = −r2 , v(s) = h(s), q = r2 . Since p is negligible to double factorial, g0 , g1 , g2
are continuously differentiable, and ℜ(g1 ) = ℑ(g0 ) = 0, ℜ(g2 ) ≤ 0, we deduce that ui , v are
continuous in [0, T ], ℜ(u1 ) ≤ 0, ℜ(u2 ), ℜ(u3 ), ℜ(v) = 0, p1 = p2 , pi , q are negligible to double
factorial by Lemma 6.3, and
" Z TX 3
! #
F (t, x) = E exp ui (s)pi (Xs )ds + v(T )q(XT ) − v(t)q(Xt ) Xt = x .
t i=1
B A small remark on the complex logarithm
For complex-valued functions, such as F in (3.1), the definition of log F is not trivial, especially if
the range of F is not simply connected. We use the lifting property in Algebraic Topology.
Lemma B.1. If f : I × R → C \ {0} is a continuous function where I is an interval of R. Then,
there exists a continuous function g : I × R → C such that exp(g) = f . And if the value of g at
one point is specified, then g is entirely specified, i.e. there exists only one g that satisfies these
conditions in this case.
Proof. Notice that p = exp(z) is a covering projection from C to C\{0}, i.e. a local homeomorphism.
And I × R is path-connected, locally path-connected, and with a trivial fundamental group. Then
by Proposition 1.33 of [23], the lifting criterion, f as a continuous function from I × R to C \ {0},
can be lifted to a continuous function g from I ×R to C, i.e. f = p◦g = exp(g). And by Proposition
of [23, 1.34], the unique lifting property, and the fact that I × R is connected, if the value of g at
one point is given, then g is specified.
Lemma B.2. Let f : I × R → C \ {0} be a continuous function, where I is an interval of R.

Assume that f is C 1 in the first variable t and C ∞ in the second variable x. Then, the function g
as defined in Lemma B.1 is also C 1 in t and C ∞ in x.
Proof. Suppose that we define g as in the Lemma B.1. We need only to prove that for all fixed point
(t0 , x0 ) ∈ I × R, g is C 1 in t, C ∞ in x in neighborhood of (t0 , x0 ). Since f (t0 , x0 ) ̸= 0, there exists a
δ > 0, for all (t, x) ∈ I×R such that |t−t0 |+|x−x0 |< δ, we have that |f (t, x)−f (t0 , x0 )|< |f (t0 , x0 )|.
Therefore, for all (t, x) ∈ I × R such that |t − t0 |+|x − x0 |< δ, f (t, x) is contained in an open ball
B, with center f (t0 , x0 ) and radius |f (t0 , x0 )|. Particularly, B is simply connected and does not
contain 0.
37
We define h from B to C as h(z) = lf (t ,x ),z w1 dw + g(t0 , x0 ), where lf (t0 ,x0 ),z is the line segment
R
0 0
between 0 and z. We aim to prove that g(t, x) = h(f (t, x)) for all (t, x) ∈ I × R such that
|t − t0 |+|x − x0 |< δ.
By the definition of h, h′ (z) = z1 , thus ( exp(h(z))

z )′ = 0, so exp(h(z))
z is a constant on B. What is
exp(h(f (t0 ,x0 )) exp(g(t0 ,x0 ))
more, f (t0 ,x0 ) = f (t0 ,x0 ) = 1. So exp(h(z)) = z, and thus exp(h(f (t, x))) = f (t, x) =
exp(g(t, x)) for all (t, x) ∈ I × R such that |t − t0 |+|x − x0 |< δ. Thus h(f (t, x)) − g(t, x) ∈
{2kπi, k ∈ Z}. Notice again that h(f (t0 , x0 )) − g(t0 , x0 ) = 0, and by continuity, we have that
g(t, x) = h(f (t, x)) for all (t, x) ∈ I × R such that |t − t0 |+|x − x0 |< δ. Notice that h′ (z) = z1 , so
h is holomorphic on B. Therefore h is C ∞ on B when regarding B as a subset of R2 . Then since
g is a composition of f and h, g is also C 1 in t, C ∞ in x.
Lemma B.3. Let f : R → C \ {0} be an entire function on R and g : R → C a continuous function
such that exp(g(x)) = f (x). Then, g is entire on R if and only if the extension fc of f to the
complex plane, as defined in Definition 3.11, does not vanish on C.
Proof. ⇐ Assume that fc : C → C does not vanish. Then, for z ∈ C, we can define h(z) =
fc′ (w) f ′ (z)
dw + g(0), where l0,z is the line segment between 0 and z. It follows that h′ (z) = fcc (z) ,
R
l0,z fc (w)
showing that h is holomorphic on C. What is more, consider the function F (z) = fc (z) exp(−h(z)),
we know that F (0) = fc (0) exp(−h(0)) = f (0) exp(−g(0)) = 1, F ′ (z) = 0, thus F (z) = 1, for all
z ∈ C. Thus exp(g(x) − h(x)) = exp(g(x)) exp(−h(x)) = f (x) exp(−h(x)) = F (x) = 1 for all
x ∈ R. Thus there exists a k ∈ Z, such that g(x) − h(x) = 2kπi. And notice that h(0) = g(0), we
have that g(x) = h(x) for x ∈ R. Particularly, h is holomorphic on C, so it is entire on C and thus
on R, then so is g.
P∞ n
⇒ Assume g : R → C is entire on R. Then, it can be written in the form P∞g(x) =n n=0 an x ,
which is a power series with infinite radius of convergence. Then gc (z) = n=0 an z is entire on
C, and so is exp(gc (z)). Notice that exp(gc (z)) and fc (z) are both entire on C, and exp(gc (x)) =
exp(g(x)) = f (x) = fc (x) for all x ∈ R. Since the zeros of an entire function are isolated, except
for the zero function, we have that fc (z) = exp(gc (z)), and hence fc (z) ̸= 0 for all z ∈ C.
Remark B.4. We point out that for a real entire function f : R → C that does not vanish
on R, there exists a continuous function log f : R → C such that exp(log f ) = f , (choosing
I = [0, 0] = {0} in Lemma B.1). However log f is not necessarily entire on R. For instance the
function f (x) = 1 + ix, is entire on R and does not vanish on R. However, log f = 12 log(1 + x2 ) +
arctan(x)i + 2kπi, k ∈ Z is no longer entire on R. In fact, the Taylor series at x = 0 for both
log(1 + x2 ) and arctan(x) have a finite convergence radius of 1 (can easily be checked) and not ∞.
Definition B.5. If the joint characteristic functional F (t, x) is continuous on [0, T ] × R and
F (t, x) ̸= 0, for all (t, x) ∈ [0, T ] × R, we define log F by Lemma B.1, as the function g such that
exp(g) = F and g(T, 0) = 0.
Remark B.6. Notice that by the definition of F in 3.1, F (T, x) = 1, so the condition g(T, 0) = 0
is satisfied.
38
C Another example of model calibration via Fourier
Figure 7: Quintic OU model (green lines) jointly calibrated to the SPX and VIX smiles (bid/ask
in blue/red) on 9 November 2021 via Fourier using the Nelder-Mead optimization algorithm. The
truncation level of the Riccati solver is set at M = 32, with calibrated parameters ρ = −0.6838, α =
−0.3914, (p0 , p1 , p3 , p5 ) = (0.0062, 0.4964, 0.0939, 0.0654). ε is fixed upfront at 1/52 without cali-
bration.
Figure 8: One-factor Bergomi model (green lines) calibrated to the SPX smiles (bid/ask in
blue/red) on 9 November 2021 via Fourier using the Nelder-Mead optimization algorithm. The
truncation level of the Riccati solver is set at M = 32, with calibrated parameters η = 1.6002, ρ =
−0.7214, α = −0.5992. ε is fixed upfront at 1/52 without calibration.
References
[1] Eduardo Abi Jaber. Lifting the heston model. Quantitative finance, 19(12):1995–2013, 2019.
[2] Eduardo Abi Jaber. The characteristic function of gaussian stochastic volatility models: an
analytic expression. Finance and Stochastics, 26(4):733–769, 2022.
[3] Eduardo Abi Jaber and Louis-Amand Gérard. Signature volatility models: pricing and hedging
with fourier. Available at SSRN 4714535, 2024.
[4] Eduardo Abi Jaber and Shaun Xiaoyuan Li. Volatility models in practice: Rough, path-
dependent or markovian? Available at SSRN 4684016, 2024.
[5] Eduardo Abi Jaber, Martin Larsson, and Sergio Pulido. Affine volterra processes. The Annals
of Applied Probability, 29(5):3155–3200, 2019.
39
[6] Eduardo Abi Jaber, Camille Illand, and Shaun Xiaoyuan Li. Joint SPX–VIX calibration
with gaussian polynomial volatility models: deep pricing with quantization hints. Available
at SSRN 4292544, 2022.
[7] Eduardo Abi Jaber, Camille Illand, and Shaun Xiaoyuan Li. The quintic ornstein-uhlenbeck
volatility model that jointly calibrates SPX & VIX smiles. Risk Magazine, Cutting Edge
Section, 2023.
[8] Leif Andersen and Jesper Andreasen. Jump-diffusion processes: Volatility smile fitting and
numerical methods for option pricing. Review of derivatives research, 4:231–262, 2000.
[9] L Bergomi. Smile dynamics II. Risk Magazine, 2005.
[10] Peter Carr and Dilip Madan. Option valuation using the fast fourier transform. Journal of
computational finance, 2(4):61–73, 1999.
[11] Christa Cuchiero and Josef Teichmann. Generalized feller processes and markovian lifts of
stochastic volterra processes: the affine case. Journal of evolution equations, 20(4):1301–1348,
2020.
[12] Christa Cuchiero, Guido Gazzani, Janka Möller, and Sara Svaluto-Ferro. Joint calibration to
spx and vix options with signature-based models. arXiv preprint arXiv:2301.13235, 2023.
[13] Christa Cuchiero, Sara Svaluto-Ferro, and Josef Teichmann. Signature sdes from an affine
and polynomial perspective. arXiv preprint arXiv:2302.01362, 2023.
[14] Darrell Duffie, Damir Filipović, and Walter Schachermayer. Affine processes and applications
in finance. The Annals of Applied Probability, 13(3):984–1053, 2003.
[15] Bruno Dupire. Arbitrage pricing with stochastic volatility. Société Générale, 1992.
[16] Ernst Eberlein, Kathrin Glau, and Antonis Papapantoleon. Analysis of fourier transform
valuation formulas and applications. Applied Mathematical Finance, 17(3):211–240, 2010.
[17] S. D. Eidelman. Parabolic Systems. North-Holland Publishing Co., Amsterdam, 1969. Trans-
lated from the Russian by Scripta Technica, London.
[18] Omar El Euch and Mathieu Rosenbaum. The characteristic function of rough heston models.
Mathematical Finance, 29(1):3–38, 2019.
[19] Fang Fang and Cornelis W Oosterlee. A novel pricing method for european options based on
fourier-cosine series expansions. SIAM Journal on Scientific Computing, 31(2):826–848, 2009.
[20] Jean-Pierre Fouque, George Papanicolaou, Ronnie Sircar, and Knut Solna. Multiscale stochas-
tic volatility asymptotics. Multiscale Modeling & Simulation, 2(1):22–42, 2003.
[21] Jean-Pierre Fouque, Matthew Lorig, and Ronnie Sircar. Second order multiscale stochas-
tic volatility asymptotics: stochastic terminal layer analysis and calibration. Finance and
Stochastics, 20(3):543–588, 2016.
[22] Jim Gatheral and Martin Keller-Ressel. Affine forward variance models. Finance and Stochas-
tics, 23:501–533, 2019.
[23] Allen Hatcher. Algebraic Topology. Cambridge University Press, 2002.
[24] Steven L Heston. A closed-form solution for options with stochastic volatility with applications
to bond and currency options. The review of financial studies, 6(2):327–343, 1993.
[25] Thomas R Hurd and Zhuowei Zhou. A fourier transform method for spread option pricing.
SIAM Journal on Financial Mathematics, 1(1):142–157, 2010.
[26] Alan L Lewis. A simple option formula for general jump-diffusion and other exponential lévy
processes. Available at SSRN 282110, 2001.
40
[27] Alexander Lipton. Mathematical methods for foreign exchange: A financial engineer’s ap-
proach. World Scientific, 2001.
[28] Roger Lord and Christian Kahl. Why the rotation count algorithm works. Tinbergen Institute
Discussion Paper, 2006.
[29] Marc Romano and Nizar Touzi. Contingent claims and market completeness in a stochastic
volatility model. Mathematical Finance, 7(4):399–412, 1997.
[30] Rainer Schöbel and Jianwei Zhu. Stochastic volatility with an ornstein–uhlenbeck process: an
extension. Review of Finance, 3(1):23–46, 1999.
[31] Klaus Schürger. Laplace transforms and suprema of stochastic processes. Advances in finance
and stochastics: essays in honour of Dieter Sondermann, pages 285–294, 2002.
[32] Elias M Stein and Jeremy C Stein. Stock price distributions with stochastic volatility: an
analytic approach. The review of financial studies, 4(4):727–752, 1991.
41

Fourier-Laplace Transforms in Polynomial Ornstein-Uhlenbeck Volatility Models

Uploaded by

Copyright:

Available Formats

Fourier-Laplace Transforms in Polynomial Ornstein-Uhlenbeck Volatility Models

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Fourier-Laplace Transforms in Polynomial Ornstein-Uhlenbeck Volatility Models

Uploaded by

Copyright:

Available Formats

Fourier-Laplace transforms in polynomial

Ornstein-Uhlenbeck volatility models

We consider the Fourier-Laplace transforms of a broad class of polynomial Ornstein-

JEL Classification: G13, C63, G10.

FiME-FDD, Financial Risks, Deep Finance & Statistics at Ecole Polytechnique.

2 Polynomial Ornstein-Uhlenbeck volatility models

• Stein-Stein [32] and Schöbel-Zhu [30] model: p(x) = x;

with the convention (−1)! ! = 0! ! = 1.

is the unique strong solution to (2.1).

By assumption, p2 has an infinite radius of convergence, then so does |p2 |, therefore

and thus we only need to prove

since p2 is negligible to double factorial. This completes the proof.

3 The joint Fourier-Laplace transform

for some complex-valued functions g1 , g2 : [0, T ] → C.

3.1 A verification result

Theorem 3.2. Let g0 : [0, T ] → R, g1 , g2 : [0, T ] → C be measurable and bounded functions.

Proof. The proof is given in Section 5.

ψk (0) = 0, k ∈ {0, 1, 2},

which is a system of standard finite-dimensional Riccati equations whose existence is well-known

F (t, x) = exp ψ0 (T − t) + ψ1 (T − t)x + ψ2 (T − t)x2 , t ≤ T,

3.2 Existence for the Riccati equations

solves the system of Riccati ODEs (3.2)-(3.3).

Proof. The proof is given in Section 6.

We now introduce Fn the discretized version of F .

Theorem 3.13. Suppose that p is negligible to double factorial, g0 : [0, T ] → R, g1 , g2 : [0, T ] → C

where for all n ∈ N, the family of functions ψn,k is defined by

Proof. The proof is given in Section 7.

4.1 Numerical scheme for high-dimensional Riccati equations

Variation of constants gives:

ψ(ti+1 ) ≈ eA∆t ψ(ti ) + G∆ P (ti ) + G∆ Qψ(ti+1 )

R(i) ψ(t e i ) ∗ ψ(t

Lemma 4.1. Assume that ℜ(ψ) ≤ 0, the matrix J is invertible.

4.2 Pricing SPX derivatives via Fourier

4.2.1 Quintic Ornstein-Uhlenbeck volatility model

4.2.2 One-factor Bergomi model

The one-factor Bergomi model [9, 15] assumes σt to be log-normal:

4.3 Pricing q-volatility swaps via Laplace inversion

The q-volatility swap rate Kq is defined by:

4.4 Model calibration to market data via Fourier

5 Proof of Theorem 3.2

ht (t, x) = k≥0 −ψk′ (T − t)xk .

where at,k is among (p ∗ p)k , kψk (t), (k + 1)ψk+1 , (k+1)(k+1)

Plugging (2.1) into (5.2), we have:

6 Proof of Theorem 3.5

To prove Theorem 3.5, we adopt the representation of F as mentioned in Lemma 3.9:

Our proof is composed of five parts:

6.1 Properties of power series negligible to double factorial

Notice that if ms = 0, then by convention the product fs,1 fs,2 · · · fs,ms = 1.

Case 2: ℓ = 2ℓ1 + 1 is an odd number, then

and since ℓ − 3 is even, as the Case 1, we have, for a constant C > 0:

Lemma 6.3. The following statements are true.

pk xk is negligible to double factorial, then so is q(x) = pk ck xk , where c is a

then by (k + 1)! ! ≤ (k + 1)k! !,

Therefore, pq is also negligible to double factorial.

6.2 Regularity in space

f (t, x) = E [Z(t, x)] , (6.3)

ft when X0 = x, with w(t) := a t eb(t−u) du and W

so that one expects

where q is any power series negligible to double factorial.