# Under what conditions do time averages of ergodic transformations satisfy a central limit theorem?

Let $(X, \mu)$ be a probability space and $T:X\rightarrow X$ an ergodic transformation, i.e. $T$ is measure preserving and the only $T$ invariant subspaces have either measure $0$ or measure $1$ (measure preserving means that $\mu(T^{-1}(A)) = \mu(A)$ for every measurable set $A$). Let $f: X \rightarrow \mathbb{R}$ be a measurable function that is in $L^1(X, \mu)$. Given an $x \in X$, define the following quantities $$I := \int_X f d\mu, \qquad I_n(x) := \frac{\sum_{k=0}^{n-1} f(T^kx)}{n}.$$ By the ergodic theorem, we know that for almost all $x\in X$, $I_n(x)$ converges to $I$.

My question is now the following: under what hypothesis on $f$, can one claim that $$\lim_{n \rightarrow \infty} \mu\{x \in X: a\leq \sqrt{n}(I-I_n(x)) \leq b)\} = \int_{a}^b G_{\sigma}(y) dy,$$ where $G_{\sigma}(y)$ is the Gaussian centered around $0$ with standard deviation $\sigma$? And moreover what will be that $\sigma$ (I would imagine there should be some formula for $\sigma$ in terms of $f$)?

To keep things simple, assume that $\int_X f^2 d\mu$ is finite (but ideally I would also like to know what is known about the rate of convergence when the integral of $f^2$ is not finite).

$\textbf{EDIT:}$ It has been pointed out that it is not realistic to expect an answer to this general question. I am therefore looking for references that address this question for specific examples of $X$ and $T$. Ideally, I am looking for a comprehensive survey article (that includes examples, counter examples and open questions) on this topic.

$\textbf{EDIT:}$ Examples involving $X:= [0,1]$ and $T$ being multiplication by some number modulo one are also fine (I had written earlier that I am not looking for that particular example; ignore that remark if you saw it).

• This is far too much to ask in this level of generality. Definite counterexamples exist even for quite nice situations like an irrational rotation and $f$ a difference of characteristic functions of intervals. The kind of condition where theorems like this are known is: $T$ hyperbolic and $f$ smooth. – Anthony Quas Jan 18 '15 at 16:24
• The sequence of random variables $(f\circ T^i)$ is strictly stationary, and there exists a vast litterature about the central limit theorem for stationary sequences. In this context, we can try (for example) a martingale approximation. In any case, I think you have to be more specific about the conditions you are looking for. – Davide Giraudo Jan 18 '15 at 16:25
• @Anthony and Davide: I see; I wasn't aware of that. I have made a small edit; I am basically looking for some references on this subject. – Ritwik Jan 18 '15 at 16:38
• @Anthony: I want to understand that last comment of yours: suppose X was the torus and T:X->X was the arnold cat map; are you saying there is a general theorem saying that the CLT for any smooth f would be satisfied? Here is the definition of arnold's cat map en.wikipedia.org/wiki/Arnold%27s_cat_map. – Ritwik Jan 18 '15 at 17:37
• The key word if you want to look for yourself is Dynamical Central Limit Theorem. A survey (from 2005) by Melbourne and Nicol can be found at math.uh.edu/%7Enicol/psfiles/refineCLT.ps – Anthony Quas Jan 18 '15 at 18:00