The approximate Euler method for Lévy driven stochastic differential equations (Q2485324)

Consider \(d\)-dimensional ordinary stochastic differential equations (SDEs) \[ X_t = x + \int^t_0 f(X_{s-}) dY_s \] driven by a \(d\)-dimensional Lévy process \(Y\), and its equidistant forward Euler approximation \[ X^n_0, \quad X^n_{(i+1)/n} = X^n_{i/n} + f(X^n_{i/n}) (Y_{(i+1)/n}-Y_{i/n}) . \] The authors suggest to replace this scheme by the approximate Euler scheme \[ X^n_0, \quad X^n_{(i+1)/n} = X^n_{i/n} + f(X^n_{i/n}) \zeta^n_{i+1} \] where \(\zeta^n_{i+1}\) are appropriate i.i.d. random variables which are both close enough to increments \(\Delta Y_i\) and exactly simulatable. As the main objective, they study the error \[ \delta_{n,i} (g(x)) = E_x [g(X^n_{i/n})] - E_x [g(X_{i/n})] \] for a sufficiently smooth class of test functions \(g \in C^4_0(\mathbb R^d)\), particularly if \(g\) is unbounded. Error expansions of first and higher order based on Romberg-type schemes are discussed and verified. One can recover estimates of \textit{P. Protter} and \textit{D. Talay} [Ann. Probab. 25, No. 1, 393-423 (1997; Zbl 0876.60030)]. Examples such as the compound Poisson process as driving process \(Y\) are also presented as a good introduction to the more general case. The discussed issues are of high practical value for related simulation procedures due to the difficulty rising with the simulation of Lévy processes.