Select a passage or formula to leave a suggestion here. Saved suggestions remain private until reviewed.

When frequency is not enough

A state may occur half the time in a long record, yet that number says little about how its occurrences are arranged. It might appear at every other step, or in runs separated by long absences. The overall proportion can be the same while the relations between different times are quite different. Studying a dynamical system means understanding these relations as well as its long-term frequencies: what patterns does its evolution preserve?

One important question concerns equally spaced occurrences. Mark the times when a target state appears on the integer line. Can three, four, or more marks be found at equal spacings, such as $a,a+n,a+2n$ ? The gap $n$ may vary, but the occurrences must fit together. The question has moved from how often a state occurs to the patterns formed by several of its occurrences.

Multiple ergodic averages study observations related in this way. They connect arithmetic progressions in the integers with the internal structure of dynamical systems. A small example makes the statistic concrete.

The same frequency, different pattern counts

Write $1$ when the state appears and $0$ when it does not. Consider two idealized periodic records, both starting at time $0$ on the left:

\begin{aligned} \text{A:}&\quad 1,0,1,0,1,0,1,0,\ldots\\ \text{B:}&\quad 1,1,0,0,1,1,0,0,\ldots \end{aligned}

In both records, the long-term proportion of $1$ is $1/2$ . Keep the starting point fixed and try gaps $n=1,2,3,\ldots$ , checking whether the entries at $n$ , $2n$ , and $3n$ are all $1$ .

For A, precisely the even gaps $n$ qualify, so their long-term proportion is $1/2$ . For B, a gap $n$ qualifies only when it is a multiple of $4$ , giving proportion $1/4$ . If its remainder is $1$ , the entry at $2n$ is $0$ . If the remainder is $2$ or $3$ , the entry at $n$ is already $0$ .

The ordinary frequencies agree, but the frequencies of this three-time pattern differ. Multiple averages can distinguish arrangements that a single overall proportion does not capture.

These are periodic models used to identify the statistic. They do not satisfy the mixing assumption of the preprints discussed below.

From equally spaced returns to arithmetic progressions

If a set of integers occupies enough positions, must it contain long arithmetic progressions? Szemerédi's theorem says yes: every set of positive upper density contains finite arithmetic progressions of every length. Positive upper density means that there is a sequence of initial intervals, with lengths tending to infinity, on which the proportion belonging to the set stays above one fixed positive number. The proportion need not converge.

Furstenberg found a way to study this question through dynamics. Shifting an integer record provides the evolution rule; a suitable invariant probability distribution connects arithmetic progressions with multiple returns in the resulting system. His 1977 proof made multiple recurrence an important connection between ergodic theory and combinatorial number theory.

A probability-preserving system has a distribution of starting points that remains unchanged after one step of the evolution. The evolution itself may be entirely deterministic. Probability describes the overall statistics; it does not make observations at different times independent samples.

Multiple recurrence asks whether a region of positive probability can be visited simultaneously at several equally spaced times. The classical result also gives a positive lower bound for the limiting inferior of the average return probabilities. Convergence of that average, and convergence for almost every fixed starting point, are further questions. A positive lower limit does not force convergence; convergence alone does not make the limit positive.

Convergence of multiple averages therefore asks for more than the existence of a pattern. It asks whether these necessarily recurring configurations have stable long-term statistics.

Averaging over the gap, with the starting point fixed

Let $T$ be one step of the system, $x$ a fixed starting point, and $A$ the region of interest. For each positive integer $n$ , check

T^nx\in A,\qquad T^{2n}x\in A,\qquad T^{3n}x\in A.

Assign $1$ when all three conditions hold and $0$ otherwise. Averaging the results for the first $N$ gaps gives

A_N(x)=\frac1N\sum_{n=1}^{N} \mathbf1_A(T^nx)\mathbf1_A(T^{2n}x)\mathbf1_A(T^{3n}x).

The usual occurrence frequency counts visits during the first $N$ steps. This average instead counts how many of the first $N$ gaps produce the specified three-time pattern, using orbit positions as far as step $3N$ . The three positions may also be tested against different regions, or the indicator functions replaced by fixed bounded numerical observations.

There are two different operations here. Moving the starting point while keeping the gap fixed measures the frequency of one fixed pattern. Keeping the starting point fixed while increasing the gap tests patterns whose positions move farther apart. In the first case, the whole pattern can be treated as a new observation function in an ordinary ergodic average. The second case, which is the subject here, does not follow directly from that construction.

In the integer problem, the initial position $a$ may also vary. Starting at time $a$ means considering $T^a x$ . But when asking whether $A_N(x)$ converges, $x$ must remain fixed. The existence of a suitable pair $a,n$ in an integer set is different from the existence of a limit for almost every fixed starting point.

The number of times being checked also matters. The displayed average tests the three later times $n,2n,3n$ . Requiring the starting point to belong to $A$ as well adds a factor $\mathbf1_A(x)$ , making four times in total: $0,n,2n,3n$ . Formulas for multiple recurrence often include this initial condition.

The relevant structure goes beyond visible periods

In the opening example, knowing the current position modulo the period is enough to calculate the pattern frequency. Correlations in a general system need not be explained by such an obvious period.

This leads to a deeper question. Which information in the original system determines the long-term behavior of a multiple average? Can that information be retained while other details are set aside?

Characteristic-factor theory makes this precise. It extracts a factor system that retains its own evolution rule. Replacing each observation by its conditional average given this information changes the multiple average by an error tending to zero in $L^2$ . For that average, the limit problem can therefore be studied on the factor.

The work of Host–Kra and Ziegler reveals algebraic structures richer than simple periodicity, described through nilsystems and their inverse limits. The structures governing multiple correlations can then be analyzed precisely.

This is important background to norm convergence in general systems. It explains why system structure matters in this subject. The mixing-system preprints discussed below, however, use their own proof constructions; their claims do not follow by directly applying this theory.

Small overall error does not control every later length at one point

For general invertible probability-preserving systems, bounded multiple averages at consecutive integer multiples of time already have a full $L^2$ convergence theorem. Convergence in $L^2$ means that the squared error, averaged over all starting points, tends to zero.

But $A_N$ is evaluated at a fixed $x$ . To interpret its limit as the long-term pattern frequency along that orbit, another question remains: as $N$ increases, does the average at this point settle down?

Small overall error says that, at each sufficiently late fixed length, the set of starting points with a large error has small probability. It does not say that these are always the same points. The exceptional sets can move with the length, leaving a fixed point exposed to repeated deviations. This is why mean-square convergence cannot simply stand in for pointwise convergence.

A pointwise theorem must rule out a set of positive probability on which deviations of some fixed size keep recurring, however far out the lengths are taken. “Almost everywhere” allows a set of probability zero to be excluded. The statement is relative to the specified probability distribution and fixed observations; it does not cover every starting point or give a rate uniform over all of them.

What mixing says, and what it leaves to prove

The periodic records at the start retain strong temporal relations. Mixing requires the joint probability of two fixed events to tend to the product of their probabilities as their time separation tends to infinity.

This controls asymptotic correlation between two times. The joint behavior at three or more times still requires analysis. Treating all the observations as independent samples would assume away that question.

For a mixing system, classical norm arguments already identify the $L^2$ limit of multiple averages as the product of the observation means. The remaining difficulty is to obtain this limit for almost every fixed starting point along all lengths. This is why mixing systems are a meaningful setting for the problem: the overall statistical answer is known, while control along a single orbit at every sufficiently late length is still needed. No mixing rate is assumed here.

A further distinction is essential. Higher-order mixing studies joint probabilities directly as several times move apart. Multiple ergodic averages first average the correlated observations over $n$ , then ask about a limit. These questions are related, but one cannot replace the other.

Where the four preprints aim to contribute

The four OpenAI preprints discussed here all carry the date October 4, 2026 on their title pages. The following describes their claims and their place in the subject. The companion close readings give the assumptions and scope of the individual proof arguments.

All consider invertible probability-preserving mixing systems and bounded measurable observations fixed in advance. They claim almost-everywhere convergence, along all positive integer lengths, to the product of the observation means. The triple result allows any three fixed, pairwise distinct nonzero integer coefficients, including a mixture of positive and negative coefficients. The main arbitrary-order statement uses $n,2n,\ldots,dn$ for each fixed finite order. None assumes a mixing rate.

The triple argument starts with oscillation control. If the averages fail to converge on a set of positive probability, many deterministic intervals of lengths can be chosen, each contributing a fixed amount of total oscillation. The key estimate in the preprints bounds the growth of that total more sharply, ruling out such a family of intervals. Once a pointwise limit exists, the norm theorem identifies it as the product of the means.

The fourfold and arbitrary-order arguments use a different construction. Assuming pointwise convergence fails, they work on a set of positive probability and choose, for each starting point in that set, a late length with a substantial deviation. They retain the joint distributions formed from these orbit segments. Selecting lengths changes the statistics. The proof must therefore establish which conditional laws survive. The subsequent structural argument seeks a contradiction between the retained nonzero correlation and these constraints.

Triple oscillation depends on a core estimate from the companion harmonic-analysis work; see the reading guide to H. The complete analytic and geometric proof has not yet been developed throughout. The fourfold and arbitrary-order arguments also use another preprint's claim that ordinary mixing implies mixing of every finite order, a form of Rokhlin's multiple-mixing problem; see the reading guide to R. That claim is not a direct consequence of the definition of mixing, and the companion readings do not constitute formal verification or external certification.

The questions form a coherent line of inquiry: which patterns can repeated events form, which parts of a system determine their statistics, and whether those statistics hold for almost every fixed starting point. The four preprints belong in this broader setting. Each step they claim to establish needs its corresponding proof.

Related examples and details

Moving intervals: overall error and a fixed point

Imagine a row of positions, with only a short interval lit at a time. At the first level, divide the row into two intervals and light them in turn. At the next level, use four intervals, then eight, then sixteen. The lit interval keeps getting shorter.

The illuminated region shrinks toward zero in size. Yet every fixed position is visited once at every level, so it keeps alternating between light and darkness.

Interactive example · Two notions of convergence

The lit interval shrinks, but every fixed point keeps lighting up

Divide $[0,1)$ into $2^k$ equal half-open intervals. Light one interval at a time, moving to the next at each step. After completing a level, halve every interval and continue at the next level. The function is 1 in the blue interval and 0 elsewhere.

Level 2 · Interval 2 / 4

Choose a level

k

Fixed point

x=

0.370 (orange line)

The slider sets $x$ between 0 and 0.999; the mathematical definition covers all of $[0,1)$ . Once adjusted, $x$ stays at its new position until the slider is moved again.

0 (included)1 (excluded)

Current interval (left endpoint included, right excluded)

\left[\frac14,\frac24\right)

Current value at the fixed point

x

is inside the lit interval

Integral of the square

\int|u_n|^2\,\mathrm dx

\frac14=0.25

L^2

norm

2^{-2/2}=0.5

Currently $n=5$ . The demonstration covers levels 1–6.

This distinguishes two notions of convergence; it is not a counterexample to the ergodic-average claims.

The finite animation illustrates the rule. The argument below, rather than the animation alone, proves the different limiting behavior of the infinite sequence.

Proof for the infinite sequence

Set $u_1=0$ . For $k\ge1$ and $0\le j<2^k$ , define

u_{2^k+j}(x)=\mathbf1_{[j/2^k,(j+1)/2^k)}(x)

Overall error:The function equals 1 only on an interval of length $2^{-k}$ . Thus $\int|u_n|^2\,\mathrm dx=2^{-k}$ , and $\|u_n\|_2=2^{-k/2}\to0$ . As $n\to\infty$ , its level $k$ also tends to infinity.

At each fixed point:The half-open intervals in each level are disjoint and cover $[0,1)$ exactly. For every $k\ge1$ , each $x$ therefore has value 1 exactly once during that level and value 0 at all its other steps. There are infinitely many levels, so every $x$ satisfies $\limsup_{n\to\infty}u_n(x)=1$ and $\liminf_{n\to\infty}u_n(x)=0$ . No pointwise limit exists.

Thus $L^2$ convergence alone does not imply almost-everywhere convergence of the entire sequence. Pointwise convergence of particular ergodic averages requires further structure and estimates specific to those averages.

At level $k$ , each interval has length $2^{-k}$ , so the integrated squared error is $2^{-k}$ , which tends to zero. Every fixed position is lit exactly once at that level and is dark at the other times. Its readings therefore never converge.

This is an example of a general sequence of functions, illustrating the distinction between two kinds of convergence. It has not been constructed as a sequence of ergodic averages.

What if we observe only at selected times?

A sequence can settle down along a sparse set of times while still making large excursions between them. The four papers seek convergence through all positive integer lengths $N$ , rather than along a well-chosen subsequence.

Four states: pairwise and joint independence

Toss two independent fair coins, writing their outcomes as 1 and $-1$ . Call the resulting signs $U$ and $V$ . For the third reading, do not toss another coin: set $W=UV$ .

Four equally likely outcomes, each with probability $\frac{1}{4}$
$U$	$V$	$W$
1	1	1
1	$-1$	$-1$
$-1$	1	$-1$
$-1$	$-1$	1

The answer

Choose any two columns. Their four possible pairs occur exactly once each, so those two readings are independent. But the product $UVW$ is always 1, whereas the product of three independent fair signs would be equally likely to be positive or negative.

Pairwise independence therefore does not imply joint independence.

The fourfold and arbitrary-order papers likewise need to preserve certain independence relations while allowing other correlations to remain. This table is only a finite probability example; the auxiliary systems in the papers must also satisfy dynamical conditions.

How an oscillation bound rules out repeated excursions

Divide the observation scales into successive blocks. In each block, measure the largest departure of the average from its value at the beginning of the block. Average this quantity over all starting points to obtain the block's oscillation.

The triple-average paper claims that, for any $w$ blocks chosen in advance, the sum of these oscillations is at most $C\sqrt{w}$ , where $C$ is a fixed constant independent of the choice of blocks.

The contradiction argument

Suppose convergence fails on a set of starting points of positive probability. A positive-probability subset can then be chosen on which every point continues to make excursions larger than one fixed amount, however late the observations begin.

From any starting scale, a sufficiently late endpoint captures such an excursion for at least half of this subset. The overall oscillation of that block therefore has a fixed positive lower bound. The points contributing an excursion may differ from block to block; a block need not capture the entire subset.

Repeating the choice from each new starting scale produces $w$ blocks whose total oscillation is bounded below by a fixed positive multiple of $w$ .

This contradicts the upper bound $C\sqrt{w}$ : for large enough $w$ , a fixed positive multiple of $w$ exceeds $C\sqrt{w}$ . Thus failure to converge on a set of positive probability is ruled out.

The detailed notes prove this implication in full. The uniform $\sqrt{w}$ bound itself is a central harmonic-analytic estimate in the original paper. It depends on a companion manuscript and has not been independently verified here.

The paper first treats smoothed averages on multiplicative scales, then uses approximation and finer grids to recover all lengths.

Point-dependent lengths and joint distributions

In the moving-light example, each position has a time at which it is lit. At any one fixed time, only a short interval is illuminated; after choosing the time separately for each position, every position is illuminated.

The fourfold paper starts from a similar difficulty. If some starting points never settle down, choose for each one a late length at which its average is still far from the proposed limit. Combine the resulting orbit segments into a joint distribution. The failure of convergence then leaves a detectable correlation.

The authors must show that this distribution also satisfies many constraints: single coordinates and smaller collections of coordinates must have the right distributions, and specified auxiliary groups must retain their independence. They then extend the distribution, arrange the variables in a square array, and regroup them to seek a contradiction with the original choice.

The arbitrary-finite-order paper first chooses a nonzero bad correlation involving as few positions as possible, then one with as many auxiliary output positions as possible. A regrouping that preserves the first number while increasing the second would contradict this choice.

Several points require separate proofs: which independence relations survive the choice of lengths, whether the conditional distributions in the extension are compatible, and whether the correlation remains nonzero after regrouping.

Video explanation

From occurrence frequencies to patterns of recurrence

About 3 min 42 s · AI narration in English · English captions

Full transcript

How often an event occurs does not determine how its occurrences are arranged in time.

Two records can have the same overall frequency and very different patterns of repeated occurrence.

Consider two idealized records.

The first alternates one, zero, one, zero.

The second repeats one, one, zero, zero.

In both, half the entries are ones.

Keep the starting point fixed at position zero and try different gaps.

For each gap, check three later positions: one gap from the start, two gaps from the start, and three gaps from the start.

Are all three entries ones?

In the first record, every even gap works.

In the second, only gaps divisible by four work.

The proportions of qualifying gaps are therefore one half and one quarter.

The ordinary frequencies agree, but the statistics of this pattern differ.

Multiple ergodic averages make this finer kind of observation.

The starting point stays fixed while the gap varies, and several related times are considered together.

These periodic records are simply a way to specify the question.

Why study such patterns?

In the integers, equally spaced marks form an arithmetic progression.

Szemerédi's theorem says that a set of positive upper density contains finite arithmetic progressions of every length.

Furstenberg connected this problem with multiple returns in a dynamical system and gave a proof using ergodic theory.

The systems here preserve a probability distribution: after one step, the overall distribution of starting points is unchanged.

The evolution itself may be completely deterministic.

Each step need not involve a random choice.

Classical multiple recurrence says that, for a target region of positive probability, the averaged return probabilities stay bounded away from zero in the long run.

But does the average converge?

And with a starting point fixed, do the pattern counts settle down as more and more gaps are tried?

The combinatorial question allows a search over starting points.

Pointwise convergence keeps one starting point fixed and considers every later averaging length.

For these bounded observations, a general theory of mean-square convergence is already available in invertible probability-preserving systems.

It has also led to methods for understanding which structures govern multiple correlations.

Yet a shrinking overall error can mean that only a small set of starting points has a large error at each length, while that set keeps changing.

A pointwise theorem must also control repeated deviations at the same starting point.

Mixing says that two fixed events, separated far enough in time, have a joint probability approaching the product of their probabilities.

It does not immediately turn three or four observations into independent samples.

For mixing systems, the product limit is already known in the mean-square sense.

The difficult question is whether that same answer holds along all lengths for almost every fixed starting point.

The preprints discussed in this topic aim to address that step.

The triple arguments control oscillation across scales.

The higher-order arguments study which joint-distribution constraints survive when lengths are selected separately for different starting points.

Their assumptions, companion inputs, and proofs need to be read individually.

From how often an event occurs, to the patterns formed by repeated occurrences, to stable statistics along an orbit: these are the questions behind multiple ergodic averages.