A note on strong approximation of SDEs with smooth coefficients that have at most linearly growing derivatives

TL;DR

This paper constructs a smooth coefficient SDE with linearly growing derivatives that cannot be approximated at a polynomial rate using fixed observation methods, but can be efficiently approximated adaptively.

Contribution

It presents a novel SDE example with smooth coefficients and linear growth derivatives demonstrating the superiority of adaptive approximation methods over fixed observation approaches.

Findings

Fixed observation methods cannot achieve polynomial convergence rates for the constructed SDE.

Adaptive methods can approximate the SDE with rate 1 relative to the number of evaluations.

This is only the second known example where adaptive methods outperform non-adaptive ones in SDE approximation.

Abstract

Recently, it has been shown in [Jentzen, A., M\"uller-Gronbach, T., and Yaroslavtseva, L., Commun. Math. Sci., 14, 2016] that there exists a system of autonomous stochastic differential equations (SDE) on the time interval $[0,T]$ with infinitely differentiable and bounded coefficients such that no strong approximation method based on evaluation of the driving Brownian motion at finitely many fixed times in $[0,T]$, e.g. on an equidistant grid, can converge in absolute mean to the solution at the final time with a polynomial rate in terms of the number of Brownian motion values that are used. In the literature on strong approximation of SDEs, polynomial error rate results are typically achieved under the assumption that the first order derivatives of the coefficients of the equation satisfy a polynomial growth condition. This assumption is violated for the pathological SDEs from the above mentioned negative result. However, in the present article we construct an SDE with smooth coefficients that have first order derivatives of at most linear growth such that the solution at the final time can not be approximated with a polynomial rate, whatever method based on observations of the driving Brownian motion at finitely many fixed times is used. Most interestingly, it turns out that using a method that adjusts the number of evaluations of the driving Brownian motion to its actual path, the latter SDE can be approximated with rate 1 in terms of the average number of evaluations that are used. To the best of our knowledge, this is only the second example in the literature of an SDE for which there exist adaptive methods that perform superior to non-adaptive ones with respect to the convergence rate.