Asymptotic error distributions for the Euler method for stochastic differential equations

Abstract

We are interested in the rate of convergence of the Euler scheme approximation of the solution to a stochastic differential equation driven by a general (possibly discontinuous) semimartingale, and by the asymptotic behavior of the associated normalized error. It is well known that for Ito's equations the rate is $1/\sqrt{n}$ ; we provide a necessary and sufficient condition for this rate to be $1/\sqrt{n}$ when the driving semimartingale is a continuous martingale, or a continuous semimartingale under a mild additional assumption; we also prove that in these cases the normalized error processes converge in law. The rate can also differ from $1/\sqrt{n}$ : this is the case for instance if the driving process is deterministic, or if it is a Levy process without a Brownian component. It is again $1/\sqrt{n}$ when the driving process is Levy with a nonvanishing Brownian component, but then the normalized error processes converge in law in the finite-dimensional sense only, while the discretized normalized error processes converge in law in the Skorohod sense, and the limit is given an explicit form. CR - Copyright © 1998 Institute of Mathematical Statistics