Pointwise multiple ergodic averages for mixing transformations

The explanation

Flip a fair coin once for every whole number, negative ones included. Call the record of flips a starting point x. The system T shifts the record one place each step. Let f(x) be 1 if the flip at position 0 is heads, and 0 if tails. Then f at time k reads the flip at position k. Now multiply f at times k, 2k and 3k. The product is 1 only when flips k, 2k and 3k are all heads. Average this product over k from 1 to N. For almost every record, the average tends to 1/2 × 1/2 × 1/2 = 1/8 as N grows.

Here is a system where that fails. Spin a point around a circle by half a turn each step. Let f be 1 on the first half of the circle and 0 on the second half. Look at times k and 2k. After 2k steps, the point is back where it started. If it starts in the first half, the product is 1 for even k and 0 for odd k. The average tends to 1/2. If it starts in the second half, the average is 0. Neither equals 1/2 × 1/2 = 1/4. The answer depends on the starting point.

The difference is mixing. A system is mixing when, after many steps, any two events become independent. For the coin flips, heads at position 0 and heads at position r have probability exactly 1/4 together, for any r other than 0. The half-turn spin never forgets where it started, so it is not mixing. The papers also need T to be invertible, meaning each step can be undone. T must also preserve probability, so every region keeps its chance under T.

The averages here are multiple ergodic averages. They became central after Furstenberg's 1977 proof of Szemerédi's theorem on arithmetic progressions, runs of evenly spaced numbers. There are two ways an average can settle. Norm convergence means it settles when errors are averaged over all starting points. Pointwise convergence means it settles at almost every single starting point. Host and Kra proved norm convergence in 2005, for every length and every such system. Pointwise convergence is harder. Bourgain proved it for two times, k and 2k, in 1990. Longer cases were known only with extra structure, such as Huang, Shao and Ye's 2019 work on a structured class called distal systems.

The main paper claims pointwise convergence for every length n of at least 2Source 1. The system must be invertible and mixing, and the measurements bounded. The limit is the product of the separate averages. No rate of mixing is neededSource 2. The underlying probability space need not be of the standard kind. Three companion papers carry smaller cases. One treats times k, 2k and 3k. One treats four timesSource 3. One treats three distinct whole-number speeds, positive or negativeSource 4.

The hard part is a single starting point that misbehaves. Suppose the averages fail to settle. Then for each starting point one can pick bad averaging lengthsSource 5. The proof turns those choices into a new probability law on several output streamsSource 6. That law keeps the right separate behaviour in each stream. The proof then uses mixing of every orderSource 7. That means any finite number of events become independent when spread far apart in time. It forces enough independence among the streams. An argument that pushes a certain count of streams as high as possible then gives a contradictionSource 8.

The proof leans on other work. It uses the claim that ordinary mixing implies mixing of every orderSource 7. That is Rokhlin's multiple-mixing problem, claimed in a separate openai/math paper. It also uses the paper on distinct speeds. Those proofs are not repeated in the main paperSource 9. The exceptional starting points may depend on the chosen measurementsSource 10.

How sure is this? No Lean formalization is listed for these papers in the openai/math release. So no computer proof checker has checked them. The result also depends on the separate multiple-mixing paper, which is itself unreviewed. No specialist review has been published.

What it doesn't show. This does not cover systems that are not mixing, such as rotations, or give a rate of convergence.