1: \begin{abstract}

2: \noindent

3: Let $\rho_0$ be an invariant probability density of a deterministic dynamical

4: system $f$ and $\rho_\eps$ the invariant probability density of a random

5: perturbation of $f$ by additive noise of amplitude $\eps.$ Suppose $\rho_0$ is

6: stochastically stable in the sense that $\rho_\eps\to\rho_0$ as $\eps\to 0.$

7: Through a systematic numerical study of concrete examples, I show that:

8: \begin{enumerate}

9: 

10: \item

11: The rate of convergence of $\rho_\eps$ to $\rho_0$ as $\eps\to 0$ is frequently

12: governed by power laws: $\onenorm{\rho_\eps-\rho_0}\sim\eps^\gamma$ for some

13: $\gamma>0.$

14: 

15: \item

16: When the deterministic system $f$ exhibits exponential decay of correlations, a

17: simple heuristic can correctly predict the exponent $\gamma$ based on the

18: structure of $\rho_0.$

19: 

20: \item

21: The heuristic fails for systems with some ``intermittency,'' {\em i.e.} systems

22: which do not exhibit exponential decay of correlations.  For these examples,

23: the convergence of $\rho_\eps$ to $\rho_0$ as $\eps\to 0$ continues to be

24: governed by power laws but the heuristic provides only an upper bound on the

25: power law exponent $\gamma.$

26: 

27: \end{enumerate}

28: Furthermore, this numerical study requires the computation of

29: $\onenorm{\rho_\eps-\rho_0}$ for $1.5 - 2.5$ decades of $\eps$ and provides an

30: opportunity to discuss and compare standard numerical methods for computing

31: invariant probability densities in some depth.

32: \end{abstract}

33: