Let $M$ be a compact, finite-dimensional Riemannian manifold, let $T: M \rightarrow M$ be an Anosov diffeomorphism, and let $\mu$ be a Sinai-Ruelle-Bowen (probability) measure. Write $\mathcal{R} = \{ R_1,\dots,R_n \}$ for a Markov partition; write $p_j^{(\mathcal{R})} := \mu(R_j)$.


Does there ever/always exist $\mathcal{R}$ s.t. $p^{(\mathcal{R})}$ is a nontrivial uniform measure on $\{1,\dots,n\}$? If not, does there ever/always exist a sequence of partitions $\mathcal{R}_m$ s.t. $p^{(\mathcal{R}_m)}$ converges to a uniform measure in some nontrivial sense?

  • $\begingroup$ Steve, are you still interested in an answer? Probably (not at all sure) I can say that it is "yes" sometimes. Only for $\mathbb T^2$ though. It's not going to work for geodesic flows. $\endgroup$ – Andrey Gogolev Jun 15 '10 at 17:17
  • $\begingroup$ I'd definitely be interested. Not sure if I can actually accept a new answer as there was a bounty that expired though... $\endgroup$ – Steve Huntsman Jun 15 '10 at 20:02
  • $\begingroup$ OTOH if you answer mathoverflow.net/questions/26834 I could accept there, and for $\mathbb{T}^2$ it's the same thing anyway. $\endgroup$ – Steve Huntsman Jun 15 '10 at 20:13
  • $\begingroup$ Ok, I'll think if the approach I have in mind really works. It may take some time. And again, it's not going to work for flows and for diffeos in higher dimension. Don't worry about accepting, I don't care much about it. $\endgroup$ – Andrey Gogolev Jun 15 '10 at 20:58

Since (p_1....,p_n) is an eigenvector of the transition matrix, your uniformity assumption is satisfied iff the transition matrix associated to the partition is bistochastic. I guess this is rarely the case.

  • $\begingroup$ The transition matrix has entries which are either 0 or 1... such a matrix is bistochastic if and only if it's a permutation matrix. $\endgroup$ – Vaughn Climenhaga May 18 '10 at 17:09
  • $\begingroup$ Sorry, I meant the matrix of transition probability. $\endgroup$ – coudy May 18 '10 at 17:10
  • $\begingroup$ Ah... understood. $\endgroup$ – Vaughn Climenhaga May 18 '10 at 17:13
  • $\begingroup$ @coudy--thanks for your answer: however, several of your statements are unclear to me. For instance, I don't see why the Markov chain associated to the partition has to be Bernoulli. There are non-Bernoulli chains that have the uniform distribution as their invariant distribution (e.g., take rows of the transition matrix to be permutations of a nondegenerate tuple summing to unity). $\endgroup$ – Steve Huntsman May 18 '10 at 17:15
  • $\begingroup$ Also, there is more than one Markov partition for the cat map with fewer than N rectangles. Consider the entropy on the set of such partitions, and choose a partition with the maximal entropy S(N). It seems plausible to me that S(N)/log(N) could tend to unity. $\endgroup$ – Steve Huntsman May 18 '10 at 17:21

I wanted to elaborate on coudy's original answer, but I also think something is wrong with this argument. In the absence of any other answers I will make the bounty available for an explanation of what goes wrong here...

If $\mathcal{P}$ is a partition of $M$, write $\mathcal{P}_m^\vee := \bigvee_{j=0}^m T^j \mathcal{P}$. (Because $T$ preserves the SRB measure $\mu$ we don't have to consider, e.g., $j<0$ terms.)

A Markov partition $\mathcal{R}$ is generating, so the supremum over partitions in the Kolmogorov-Sinai entropy is realized by it: the KS entropy is $h_\mu^{KS}(T) = -\lim_m m^{-1}\sum_{R \in \mathcal{R}_m^\vee} \mu(R) \log \mu(R)$. Meanwhile the topological entropy is $h(T) = \lim_m m^{-1} \log \# \mathcal{R}_m^\vee$.

So if $\mu$ is also the measure of maximal entropy (as is the case, e.g. when $T$ is a hyperbolic toral automorphism or the Poincaré map for a geodesic flow on a surface of negative curvature), then $h_\mu^{KS}(T) = h(T)$ and in the limit we have that $-\sum_{R \in \mathcal{R}_m^\vee} \mu(R) \log \mu(R) \sim \log \# \mathcal{R}_m^\vee$, so that $\mu$ is asymptotically uniform on $\mathcal{R}_m^\vee$.

  • 1
    $\begingroup$ AFAIU, your argument is perfectly correct. The point is, you don't really tell us what you mean by "asymptotically uniform". I think that what you get with your computation is a particular case of the Shannon-McMillan Breiman theorem, valid for any ergodic transform and any partition: for a.e. x, we have 1/n mu(log R_n(x)) -> h. This implies that the ratio of the log of the measures of two elements in R_n tends to 1. But be careful, it does not imply that the ratio of the measures of two elements in the partition R_n goes to 1. It just means that this ratio is subexponential. $\endgroup$ – coudy May 21 '10 at 20:41
  • $\begingroup$ OK, so the argument is less relevant, though not wrong. I've been looking at some examples with hyperbolic toral automorphisms and trying to square the behavior with what I see numerically for geodesic flow. Your point about subexponential ratios makes perfect sense now...but as you said earlier, it leaves the original question still standing. $\endgroup$ – Steve Huntsman May 21 '10 at 21:46
  • $\begingroup$ BTW, my statement about "asymptotically uniform" isn't quite kosher unless it's interpreted more weakly than I'd intended, i.e., more weakly than in the spirit of the original question. $\endgroup$ – Steve Huntsman May 21 '10 at 21:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.