1: \begin{abstract}

2: The Euler--Maruyama scheme

3: is known to diverge strongly

4: and numerically weakly

5: when applied to nonlinear

6: stochastic differential 

7: equations (SDEs) with

8: superlinearly growing and

9: globally one-sided Lipschitz

10: continuous drift coefficients.

11: Classical Monte Carlo simulations

12: do, however, not suffer from this

13: divergence behavior of Euler's

14: method because this divergence

15: behavior happens on 

16: {\it rare events}.

17: Indeed, 

18: for such nonlinear SDEs

19: the classical

20: Monte Carlo Euler method has been

21: shown to converge by exploiting that the Euler approximations

22: diverge only on events whose probabilities

23: decay to zero very rapidly.

24: Significantly more efficient 

25: than the classical Monte Carlo Euler

26: method is the recently introduced

27: multilevel Monte Carlo Euler method.

28: The main observation of this article is 

29: that 

30: this multilevel Monte Carlo Euler

31: method does

32: -- in contrast to classical Monte

33: Carlo methods --

34: not converge in general.

35: More precisely, we establish divergence of the

36: multilevel Monte Carlo Euler method 

37: for a family of SDEs with superlinearly

38: growing and globally one-sided Lipschitz 

39: continuous drift coefficients.

40: In particular, the multilevel Monte

41: Carlo Euler method diverges

42: for these nonlinear SDEs

43: on an event that is not at all

44: rare but has

45: {\it probability one}.

46: As a consequence for applications, we recommend

47: not to use the multilevel Monte Carlo Euler method

48: for SDEs with superlinearly

49: growing nonlinearities. Instead we 

50: propose to combine

51: the multilevel Monte Carlo method with a

52: slightly modified Euler method.

53: More precisely,

54: we show that 

55: the multilevel Monte

56: Carlo method 

57: combined with

58: a

59: tamed Euler method 

60: converges

61: for nonlinear SDEs with globally one-sided Lipschitz continuous

62: drift coefficients

63: and preserves

64: its strikingly higher order convergence

65: rate from the Lipschitz case.

66: \end{abstract}