Approximating Lévy processes with completely monotone jumps (Q259581)

Very often processes with completely monotone jumps are used as appropriate models in modeling phenomena arising in mathematical finance. This is due to the fact that they admit many desirable features such as finite/infinite jumps activity, finite/infinite variation, etc., and as such provide a good fit to market data. Furthermore, in some situations they also enjoy certain analytical properties (for example, the transition density function of the underlying process is given explicitly). However, there are situations where the analysis of these models becomes very complicated. In the paper under review, the authors develop an approximation method of processes with completely monotone jumps by processes with hyperexponential jumps. The latter processes admit an explicit Wiener-Hopf factorization and as such form the most convenient class for performing numerical calculations. The method is based on the Padé approximation. The Padé approximation \(f^{[m/n]}(x)\) of a function \(f(x)=\sum_{k\geq 0}a_kx^{k}\) is defined as the rational function \(P_m(x)/Q_n(x)\) which matches the first \(n+m+1\) Taylor coefficients of \(f(x)\), where \(P_m(x)\) and \(Q_n(x)\) are polynomials satisfying \(\operatorname{deg}(P_m)\leq m\) and \(\operatorname{deg}(Q_n)\leq n\). The idea is to approximate the Laplace exponent \(\psi(z)\) of the underlying process (this is equivalent to approximating the process). As the first result, which can be proved very easily, the authors show that \(\psi^{[m/n]}(z)\) can be the Laplace exponent of some Lévy process if and only if \(m\in\{n,n+1,n+2\}\). The much more challenging task consists of proving that \(\psi^{[n/n]}(z)\), \(\psi^{[n+1/n]}(z)\) and \(\psi^{[n+2/n]}(z)\) are indeed Laplace exponents of some hyperexponential Lévy processes which, as \(n\to \infty\), converge (in distribution) to the underlying processes with completely monotone jumps. The second main result states that, if the starting process \(X\), with Laplace exponent \(\psi(z)\), has completely monotone jumps, then the Padé approximation \(\psi^{[n+1/n]}(z)\) is a Laplace exponent of a hyperexponential process \(X^{n}\) (whose Lévy triplet is computed explicitly) and \(\psi^{[n+1/n]}(z)\) converges to \(\psi(z)\). Hence, \(X^{n}\) converges to \(X\) in distribution. Further, if \(X\) has only positive (negative) jumps, then the same result holds true for \(\psi^{[n+2/n]}(z)\) and \(\psi^{[n/n]}(z)\) provided that \(X\) has jumps of bounded variation. At the end, as an application of the main results, the authors discuss the approximation of Gamma subordinators, one-sided tempered stable processes and VG, CGMY and NIG processes.