This document summarizes a method for approximating the density function of a Lévy process using Fourier transforms and fundamental sk-splines. The method represents the density function as a sum of Fourier transforms of the characteristic function weighted by fundamental sk-splines. An approximant for the density function is proposed as a finite sum of these terms. An example applying the method to a Gaussian density function is also provided to illustrate the approach.
This document summarizes a method for approximating the density function of a Lévy process using Fourier transforms and fundamental sk-splines. The method represents the density function as a sum of Fourier transforms of the characteristic function weighted by fundamental sk-splines. An approximant for the density function is proposed as a finite sum of these terms. An example applying the method to a Gaussian density function is also provided to illustrate the approach.
This document summarizes a method for approximating the density function of a Lévy process using Fourier transforms and fundamental sk-splines. The method represents the density function as a sum of Fourier transforms of the characteristic function weighted by fundamental sk-splines. An approximant for the density function is proposed as a finite sum of these terms. An example applying the method to a Gaussian density function is also provided to illustrate the approach.
This document summarizes a method for approximating the density function of a Lévy process using Fourier transforms and fundamental sk-splines. The method represents the density function as a sum of Fourier transforms of the characteristic function weighted by fundamental sk-splines. An approximant for the density function is proposed as a finite sum of these terms. An example applying the method to a Gaussian density function is also provided to illustrate the approach.
A. Kushpel Department of Mathematics Çankaya University Ankara, TURKEY
AMS Subject Classification: 41A05, 42B05
Key Words: Fourier transform, PDE, sk-spline, Lévy process, density function
Received: November 10, 2018 Revised: March 20, 2019
Published: March 23, 2019 doi: 10.12732/ijdea.v18i1.3 Academic Publications, Ltd. https://acadpubl.eu
1. INTRODUCTION
The problem of inversion of Fourier transforms is a frequently discussed topic in the
theory of PDEs, Stochastic Processes and many other branches of Analysis. We consider here in more details an application of a method proposed in Financial Modeling. As a motivating example consider a frictionless market with no arbitrage opportunities and a constant riskless interest rate r > 0. Assuming the existence of a risk-neutral equivalent martingale measure Q, we get the option value V = e−rT EQ [ϕ] at time 0 and maturity T > 0, where ϕ is a reward function and the expectation EQ is taken with respect to the equivalent martingale measure Q. Usually, the reward function ϕ has a simple structure. Hence, the main problem is to approximate properly the respective density function and then to approximate EQ [ϕ]. Here we offer an approximant for the density function without proof of any convergence results. These problems will be considered in details in our future publications. 26 A. Kushpel
2. THE RESULTS
P Let x and y be two vectors in Rn , hx, yi := nk=1 xk yk be the usual scalar product and |x| := hx, xi1/2 . For f (x) ∈ L1 (Rn ) define its Fourier transform Z Ff (y) = exp (−i hx, yi) f (x)dx Rn
and its formal inverse as
Z −1 −n F f (x) = (2π) exp (i hx, yi) f (y)dy. Rn
We will need the following well-known result (see e.g. [7]).
Theorem 1. (Plancherel’s theorem) The Fourier transform is a linear continu- ous operator from L2 (Rn ) onto L2 (Rn ) . The inverse Fourier transform, F−1 , can be obtained by letting
F−1 g (y) = (2π)−n (Fg) (−y)
for any g ∈ L2 (Rn ) .
The density function pQ t of any Lévy process X = {Xt }t∈R+ can be expressed in
terms of the characteristic function ΦQ (x,t) = exp −tψ Q (y) of the distribution of X −n as pQ t = (2π) F ΦQ (x,t) , where ψ Q (y) is the characteristic exponent. According to the Khintchine-Lévy formula, for any Lévy process X, the characteristic exponent ψ admits the representation Z ψ (y) = hLx, xi − i hh, xi − (1 − exp (i hx, zi) − i hx, zi χD (x)) Π (dx) (1) Rn
where χD (y) is the characteristic function of the unit ball in Rn , h ∈ Rn , L is a
symmetric nonnegative-definite matrix and Π (dy) is a measure such that Z min {1, hx, xi} Π (dy) < ∞, Π ({0}) = 0. Rn
See [6] for more details. For simplicity we assume absolute convergence of multiple series under consideration which is sufficient for our applications. Let K : Rn → R be a fixed kernel function and A be a nonsingular n × n matrix. Definition 2. We say that K ∈ K if the series X T Υ−1 (y, z) := F (K) z + 2π y−1 m (2) m∈Zn
converges absolutely and Υ (y, z) F (K) (z) ∈ L1 (Rn ).
A METHOD OF INVERSION OF FOURIER... 27
The set K is sufficiently large for our applications. In particular, if K (z) > 0, ∀z ∈ Rn then instead of Υ (z) F (K) (z) ∈ L1 (Rn ) we may just clime that F (K) (z) ∈ L1 (R n ). A typical example of K ∈ K is given by a Gaussian of the form K (y) = 2 exp − |By| , where B is a nonsingular matrix. In this case K (y) > 0 and the condition (2) is easily verifiable. Fix a kernel function K ∈ K. Consider the function Z f det (A) sk (y) := Υ (y, z) F (K) (z) exp (i hy, zi) dz (2π)n Rn which we call the fundamental sk-spline. It is possible to show that ( f (Am) = 1, mk = 0, 1 ≤ k ≤ n, sk . 0, otherwise. f (y) are analogs of periodic fundamental for any m ∈ Zn if K ∈ K [5] . The functions sk splines introduced in [3, 4] n−1 ˜ 1 1 X Reλj (x)Reλj (y) + Imλj (x)Imλj (y) sk(x) = + , n n j=1 |λj (y)|2 P where λj (y) = nν=1 exp 2πiνj n K y− 2πν n 6= 0, 1 ≤ j ≤ n − 1 and K ∈ C T1 is a fixed kernel function. Observe that ( ˜ + 2πj/n) = 1, j = 0mod(n), sk(y 0 otherwise.
Fundamental sk-splines on parallelepipedal grids in Td (see [2]) were considered in [1].
To construct an approximant q (·) for the density function pQ Q t (·) defined by Φ (x, t) in (1) we assume that K ∈ K ∩ L2 (Rn ). Consequently, by Plancherel theorem we get
Q F ΦQ (x, t) (·) pt (·) = (2π)n ! X ΦQ (Am,t)sk(xf − Am) ≈F (·) (2π)n m∈Z n X Q det (A) Φ (Am,t) F (K) (z) =F F−1 (x − Am) (·) nP −1 )T s m∈Z n (2π) s∈Zn F (K) z+2π (y P det (A) F (K) (·) m∈Zn ΦQ (−Am,t) exp (i h·, Ami) = P . (2π)n s∈Zn F (K) · + 2π (y−1 )T s
Assume that for some domain Ω ⊂ Rn , Ω ∋ 0 the sum
X T F (K) · + 2π y−1 s s∈Zn \0 28 A. Kushpel
is ”relatively small” for any z ∈ Ω. Then
X T F (K) · + 2π y−1 s ≈ F (K) (·). s∈Zn
Hence det (A) X Q pQ t (·) ≈ q(·) = Φ (−Am,t) exp (i h·, Ami) . (2π)n m∈Zn
Let mk , k ∈ N corresponds to the nonincreasing rearrangement of ΦQ (−Am,t) ,
m ∈ Zn . Hence for a fixed N = N ΦQ , y, t, we get Theorem 3. In our notations the approximant q(z) for the density function pQ t (z) has the form N det (A) X Q q(z) = Φ (−Amk , t) exp (i hz, Amk i) . (2π)n k=1
Example 4. Let n = 2 and p(y) = p(x1 , x2 ) = π −1 exp −x21 − x22 be Gaus-
sian density. Then Φ (y) = Fp (−y) = exp − x21 + x22 /4 . Let P and M be fixed parameters. In the case of the square grid (2πk/P, 2πs/P ), (k, s) ∈ Z2 we get y = diag (2π/P, 2π/P ), det(y) = (2π/P )2 . Hence the approximant q(y) takes the form
q(y) =q(x1 , x2 )
(2π/P )2 X X 2πk 2πs 2πki 2πsi = Φ − ,− exp x1 + x2 (2π)2 |k|≤M |s|≤M P P P P 2 2 ! 1 X X 2π k + s2 2πikx1 2πisx2 = 2 exp − exp + . P P 4 P P |k|≤M |s|≤M
Let d (P, M, a) := max {x ∈ [−a/2, a/2] × [−a/2, a/2] | |p(y) − q(y)|}. Numerical ex- amples show that d (5, 4, 1) = 2.36 × 10−5 , d (5, 6, 1) = 1.8 × 10−8 .
REFERENCES
[1] S.M. Gomes, A.K. Kushpel, J. Levesley, D.L. Ragozin, Interpolation on the Torus using sk-splines with Number Theoretic Knots, J. of Approximation Theory, 98 (1999), 56-71. [2] N.M. Korobov, Exponential Sums and their Applications, Boston, Kluwer Aca- demic Publishers, 1992. [3] A.K. Kushpel, Extremal properties of splines and diameters of classes of periodic functions in the space C2π , Preprint, 84.15, Kiev, Inst. Math. Acad. Nauk Ukrain. SSR (1984), 1-44. A METHOD OF INVERSION OF FOURIER... 29
[4] A.K. Kushpel, Sharp estimates of the widths of convolution classes, Math. USSR Izvestiya, American Mathematical Society, 33, No. 3 (1989), 631-649. [5] A. Kushpel, sk-Spline interpolation on Rn , ArXiv: 1809.08618v1 [math. NA] 23 Sep 2018. [6] K. Sato, Lévy Processes and Infinitely Divisible Distributions, Cambridge Univer- sity Press, 1999. [7] E. Stein, G. Weiss, Fourier Analysis on Euclidean Spaces, Princeton, 1990. 30
