Implementation and Analysis of an Adaptive Multilevel Monte Carlo Algorithm

This work generalizes a multilevel Monte Carlo (MLMC) method in-troduced in [7] for the approximation of expected values of functions depending on the solution to an Ito stochastic differential equation. The work [7] proposed and analyzed a forward Euler MLMC method based on a hierarchy of uniform time discretizations and control variates to reduce the computational effort required by a standard, single level, forward Euler Monte Carlo method from O( TOL^(−3) ) to O( TOL^(−2) log( TOL^(−1))^2 ) for a meansquare error of size 2 . This work uses instead a hierarchy of adaptivelyrefined, non uniform, time discretizations, generated by an adaptive algo-rithm introduced in [20, 19, 5]. Given a prescribed accuracy TOL in theweak error, this adaptive algorithm generates time discretizations basedon a posteriori expansions of the weak error first developed in [24]. Atheoretical analysis gives results on the stopping, the accuracy, and thecomplexity of the resulting adaptive MLMC algorithm. In particular, it isshown that: the adaptive refinements stop after a finite number of steps;the probability of the error being smaller than TOL is under certain as-sumptions controlled by a given confidence parameter, asymptotically asTOL → 0; the complexity is essentially the expected for MLMC methods,but with better control of the constant factors. We also show that themultilevel estimator is asymptotically normal using the Lindeberg-FellerCentral Limit Theorem. These theoretical results are based on previouslydeveloped single level estimates, and results on Monte Carlo stoppingfrom [3]. Our numerical tests include cases, one with singular drift andone with stopped diffusion, where the complexity of uniform single levelmethod is O TOL−4 . In both these cases the results confirm the theoryby exhibiting savings in the computational cost to achieve an accuracy of O(TOL), from O( TOL^(−3) )for the adaptive single level algorithm toessentially O( TOL^(−2) log(TOL−1)^2 ) for the adaptive MLMC.QC 20120508