Ergodic Ramsey Theory started with Furstenberg’s proof of Szemeredi’s theorem in arithmetic progressions in 1977. Through a correspondence principle, Furstenberg realized that Szemeredi’s theorem follows from a dynamical statement:
and it is natural to question whether the limit exists.
Since Furstenberg’s seminal paper, many versions of results on recurrence and convergence of non-conventional ergodic averages were established by several people. In this post I will list the most important developments in this area, including some recent results. I will restrict to -actions and averaging over the classical F\o lner sequence . There are many developments for other group actions which have interesting combinatorial applications, but that’s another story.
— 1. First example: Poincare’s recurrence and von Neumann’s mean ergodic theorem —
In this section I will use the example of Poincare’s recurrence theorem and the mean ergodic theorem to illustrate the relation between convergence and recurrence results. Essentially neither convergence nor recurrence imply the other result, although convergence plus some information about the limit (i.e. if it is an orthogonal projection) is enough to give a proof of the recurrence result. Despite this fact, it is usually (at least historically) harder to establish convergence results than recurrence results. One reason for this is the fact that if a convergence result gives enough information about the limit to deduce the corresponding recurrence result, then one usually obtains additional information about the recurrence result (for instance, the set of return times will be “large” in certain senses).
One of the many versions of Poincare recurrence theorem can be stated as
for any non-negative function such that .
The convergence result that deals with the same ergodic averages (i.e. the convergence statement that contains the expression ) is von Neumann’s mean ergodic theorem. A version of this theorem can be formulated the following way: for any (where again is a -system) the limit
exists in the strong norm. Notice that (because strong convergence implies weak convergence in ) this statement implies that for any , the limit
exists. In particular putting we conclude that in (3) we can replace with . However, this is not enough to conclude that the limit is positive when and . Luckily, the mean ergodic theorem gives some information about the limit in (4). Namely, we have that is an orthogonal projection onto the subspace formed by the functions such that . This is enough to deduce recurrence from convergence (i.e. to deduce equation (3) from equation (4)): In the following equation, the norm and inner product are of the Hilbert space and we use the symbol to denote the constant function equal to in .
where on the first line we use the continuity of the inner-product, in the second line we use the fact that is an orthogonal projection (and hence is orthogonal to the image of under ), in the third line we use the Cauchy-Schwartz inequality with the functions and and in the last line we use again the fact that is a projection and is in its image.
— 2. Non-conventional ergodic averages for a single action —
The first non-trivial recurrence result is the ergodic version of Roth’s theorem. This result is equivalent to Roth’s theorem in arithmetic progressions (Szemeredi’s theorem for progressions of length ). The ergodic version states that whenever is a set of positive measure in a -system, then for some . More generally, Furstenberg’s multiple recurrence theorem, which is equivalent to Szemeredi’s theorem, states that under the same conditions, for every there exists some such that
This was proved by Furstenberg in his seminal paper in 1977. In fact he showed that
exits in the norm. Observe that the case of this result is von Neumann’s mean ergodic theorem. This result was establish when in Furstenberg’s 1977 paper, and the general case was established by Host and Kra in 2005.
Observe that, while the recurrence result in equation (6) was proved in 1977, only in 2005 the convergence result (7) was deduced. Moreover, by itself, the convergence result does not imply recurrence. However, as in the mean ergodic theorem, the convergence result (7) of Host and Kra gives some information about the limit obtained (roughly speaking, the dynamics of the limit in (7) behaves as a function on a nil-system. For comparison, the dynamics of the limit in (4) behaves as a function on a trivial one-point system). This information can be used to deduce the recurrence result (6), it seems this has been done implicitly in many places (and maybe explicitly somewhere), see this post of Terry Tao (and the comments in particular).
There is one more type of results about a single -action, and this concerns polynomial sequences.
Theorem 1 (Bergelson-Leibman) Let be an invertible measure preserving system and let be polynomials such that for all . Let with . Then
This result was later generalized by Bergelson, Leibman and Lesigne to include all families of jointly divisible polynomials, i.e. such that for every there is some (dependent on but not on ) such that for all . It is easy to check that for the polynomial in a -point rotation, no point returns to itself after an iteration of the form . In fact, the condition that the polynomials are jointly divisible is a necessary and sufficient for recurrence to occur.
The convergence result associated with this result also holds (and for any polynomials). This was first established by Host and Kra and independently by Leibman. It also follows from the significantly more general Theorem 3 below.
— 3. Several -actions —
In this section I still deal with more than one measure preserving transformation. For commuting actions, this is in principle the same as studying a single action, but the averages are taken with respect to a parameter living in , so even in the case when the actions commute, our interest is in the individual -actions, rather than the action they induce.
— 3.1. Commuting -actions —
The first result in this category is the multidimensional Szemeredi theorem. This result was deduced by Furstenberg and Katznelson in 1978 through its ergodic version. It states that given any commuting measure preserving transformations of a probability space and any set with positive measure there exists such that
Notice that when this implies Furstenberg’s multiple recurrence theorem (5). This result was established in 1978 and was deduced from the following averaging result
exists in . This proof does not give any information about the limit in (9), and is therefore essentially independent from the recurrence result (8). A second proof that the limit in (9) exists was obtained by Austin in 2008 and gives some information about the limit. Austin then used this approach with some additional ideas to give a new proof of the recurrence result (8).
The ergodic multidimensional polynomial Szemeredi theorem of Bergelson and Leibman is a generalization of Theorem 1 (obtained in the same paper as Theorem 1) to the case of many commuting actions. It states that for any polynomials , with for all , for any commuting measure preserving transformations of a probability space and any non-negative function such that we have
Observe that when this result reduces to (8). There is currently no generalization of this result to more general polynomials (in the spirit of the Bergelson-Leibman-Lesigne result mentioned above), although there have been some generalizations for other functions, culminating in this recent paper by Zorin-Kranich about generalized polynomials.
The convergence of the averages in (10) follows from the more general Theorem 3 by Walsh.
— 3.2. Non-commuting -actions —
To say that are commuting invertible measure preserving transformations is a short way to say that the group generated by is abelian. The next simplest case is when the transformations generate a nilpotent group. We will conveniently refer to this situation by saying that the transformations “nilpot” or “are nilpoting”.
Theorem 2(Leibman, 1998) Let and be nilpoting invertible measure preserving transformations of a probability space . Let be polynomials such that . Then for any with we have
The convergence result associated with this result is due to Walsh:
Theorem 3 (Walsh, 2012)Let and be nilpoting invertible measure preserving transformations of a probability space . Let be polynomials and let . Then the limit
exists in .
Walsh’s result gives no information about the limit, in the same spirit as Tao’s proof of (9). Observe that in Walsh’s result there are no restriction on the polynomials, unlike the recurrence results such as Theorem 1 or even the Bergelson-Leibman-Lesigne result.
Finally if and are measure preserving transformations of the probability space that do not nilpot, both convergence and recurrence may fail, even when the group generated by and is solvable. This was discovered by Bergelson and Leibman. This means that for -actions, the convergence result of Walsh is (in a certain sense) the best possible.