Pointwise multiple ergodic averages under mixing

Source and verification

This article explains family 154 of the October 6, 2026 OpenAI collection. Research conclusions are attributed to its manuscripts. The review covers statements, introductions, and selected proof mechanisms; it is not an independent verification of the complete proofs.

🏷️ Multiple observations along one orbit

Let $T$ be an invertible, bimeasurable probability-preserving transformation. For bounded functions $f_1,\dots ,f_n$, the consecutive multiple averages are

$$A_N(x)=\frac{1}{N}\sum _{k=1}^N\prod _{j=1}^n{f}_j(T^{jk}x).$$

The observations occur along one orbit at arithmetic-progression times. The question is convergence for almost every starting point along all positive integer lengths $N$.

Mixing means $\mu (A\cap T^{-r}B)\to \mu (A)\mu (B)$ as $|r|\to \infty$. It gives norm-convergence tools, but convergence in $L^2$ alone does not give pointwise convergence along every length. A subsequence with summable errors is weaker than the full claim.

🏷️ Four companions and their stated ranges

The family contains a progression of results:

• The triple paper treats slopes $1,2,3$ under mixing. (OpenAI, 2026)
• The distinct-slope paper treats every fixed triple of pairwise distinct nonzero integer slopes, including negative slopes, under mixing. Constants may depend on the slopes. (OpenAI, 2026b)
• The fourfold paper treats slopes $1,2,3,4$ under mixing. (OpenAI, 2026a)
• The general paper treats every finite consecutive length $n\ge 2$ under mixing. (OpenAI, 2026c)

Each claimed limit is the product of the integrals. The original probability space need not be standard, and no mixing rate is assumed.

The exceptional null set may depend on the fixed functions, and in the distinct-slope statement also on the slopes. None of these formulations requires one common conull set working for all bounded functions at once.

The all-length manuscript does not state the corresponding all-length result for arbitrary unrelated slopes. The three-slope extension must not be extrapolated to that larger assertion.

🏷️ Oscillation rather than one maximal bound

For the triple problem, replace the interval average by a smooth profile $\phi$ of integral one. At dyadic scale $2^{\ell }$, write

$$\begin{array}{rl}& V_{\ell }^{\phi }(x)\\ & =\sum _{k\in \mathbb{Z}}{2}^{-\ell }\phi (2^{-\ell }k)\prod _{j=1}^3{f}_j(T^{jk}x).\end{array}$$

For every fixed list of endpoints ${\ell }_1<\cdots <{\ell }_{W+1}$, the source proves an oscillation estimate of the form

$$\begin{array}{c}{O}_i=\underset{{\ell }_i\le \ell \le {\ell }_{i+1}}{\max }|V_{\ell }^{\phi }-V_{{\ell }_i}^{\phi }|,\\ \sum _{i=1}^W\int O_{i}d\mu \le C_{\phi }\sqrt{W}.\end{array}$$

If the sequence failed to be Cauchy on a set of positive measure, endpoints could be chosen so that each block contributed a fixed positive amount. The left side would grow linearly in $W$, contradicting the square-root bound. (OpenAI, 2026b)

The analytic oscillation bound applies without mixing. Mixing is used separately to identify the limit through norm convergence. Approximation of interval profiles and rational meshes then recovers all integer lengths, rather than only powers of two.

🏷️ Why the Hilbert-transform bound needs strengthening

The triple proof imports local estimates from the collection’s trilinear Hilbert-transform manuscript. A fixed-input operator bound alone does not immediately control pointwise oscillation.

At each $x$, the maximizing scale in a block may differ. Linearizing that maximum creates an input depending on depth and truncated at a point-dependent stopping index. Its amplitude and internal variation are normalized by $W^{-1/2}$.

The source extends the local estimates to these depth-dependent inputs. A Bessel inequality uses the same input in every row; tested adjoint partial sums and summation by parts control the changing input. Cauchy–Schwarz across blocks cancels the normalization. Restoring the scale factor gives the $O(\sqrt{W})$ oscillation estimate. (OpenAI, 2026b)

For arbitrary triple slopes, rescaling and permuting do not turn the configuration into $0,1,2,3$. The distinct-slope companion adapts the full integer configuration, its quadratic cancellation relation, and the divisibility in the frequency-counting argument. It must retain polynomial rather than exponential dependence in the relevant dimension parameter. (OpenAI, 2026d)

🏷️ Selected lengths change joint laws

The fourfold and general proofs use a different extension argument. Failure of pointwise convergence allows a length $N(x)$ to be selected at each point so that a centered average remains separated from zero.

Averaging the selected configurations gives a joining-like law with a nonzero correlation. Because the chosen length depends on the starting point, conditioning on it can change independence relations among auxiliary variables. Qualitative norm convergence cannot simply be substituted for a uniform bound on all such selections.

A finite model shows the issue. If $U,V,W$ are independent symmetric signs, $UVW$ is independent of any pair. Conditioning on $UVW=1$ preserves those pair marginals while constraining the triple.

The manuscripts maintain three auxiliary groups whose joint spatial law is a product, with the original coordinate independent of any two groups. Independence from all three together is deliberately not assumed. (OpenAI, 2026a) (OpenAI, 2026c)

🏷️ Packing tensors and preserving the allowed independence

A tensor-packing estimate controls projections of selected averages onto a leading covariance subspace. Roughly, when bounded vector families have small pairwise inner products outside a bounded-degree exceptional graph, few tensor products can have large projections onto a slowly growing finite-dimensional subspace.

The remaining orthogonal component has square-summable error across possible lengths. Together with higher-order mixing from the separate Rokhlin manuscript, this preserves the precise auxiliary independence conditions. No quantitative mixing rate is inserted.

A constant-roof suspension supplies a flow language for different coordinate speeds. It is a technical reorganization of the original discrete-time problem, not an extra dynamical assumption.

🏷️ A square construction closes the contradiction

The fourfold argument chooses a nonzero centered correlation maximizing the number of slots that use auxiliary outputs. Independence and an iterated-array Bessel argument exclude the largest possible counts.

An extension is then made stable for selected conditional expectations: further admissible information cannot improve their prediction. Applying the selected-law construction once more yields a square array. Its distinguished row and column meet at a corner, and the stability identities determine their conditional laws.

Regrouping the outputs preserves a nonzero correlation while increasing the number of auxiliary slots, contradicting maximality. (OpenAI, 2026a)

For arbitrary $n$, the source first minimizes the number of slots carrying a nonzero centered correlation, then maximizes the auxiliary-slot count within that minimum. It retains three auxiliary groups even as the number of slots grows. A slot may access appropriate proper subsets of these groups; it does not assume independence of arbitrary whole states. (OpenAI, 2026c)

The proof imports pointwise control for at most three slots on arbitrary systems in its specified form, plus the multiple-mixing theorem. These dependencies are part of the claim’s proof architecture and are not reproduced or independently certified here.

🏷️ Key steps and proof outlines

Oscillation estimates give convergence along all scales

Suppose a bounded scalar sequence of smooth dyadic averages is not Cauchy on a positive-measure set. There are a fixed threshold $\epsilon >0$ and positive mass $a$ of points with arbitrarily late oscillations larger than $\epsilon$. Starting from an endpoint, increase the next endpoint until a set of mass at least a fixed fraction of $a$ witnesses that oscillation within the block. Repeat this deterministic choice. Every block then contributes at least a fixed positive amount to its integrated maximum oscillation, so the sum grows linearly in the number of blocks $W$.

The triple manuscript’s bound $C\sqrt{W}$ contradicts that growth. This explains why its estimate is stronger than one maximal bound or an $L^2$ convergence theorem. Approximation of the interval indicator by smooth profiles controls the difference uniformly for bounded inputs, while meshes with successive length ratios close to one control the remaining integer lengths: changing $N$ to $N^{\prime }\ge N$ changes a bounded normalized average by at most $2(N^{\prime }-N)/N^{\prime }$. Taking finer meshes closes the all-length conclusion. Mixing separately identifies the limit through norm convergence. (OpenAI, 2026b)

Tensor packing separates a low-dimensional exceptional part

The later companions need a uniform estimate that survives point-dependent selection of averaging lengths. Their deterministic packing theorem starts with unit-ball families $x_j,y_j$ whose large off-diagonal inner products lie in graphs of maximum degree $L$. Outside those graphs the inner products have size at most $\eta \le {\epsilon }^2/10^4$. For a rank-$d$ orthogonal projection $P$ on the tensor product, the source bounds the number of indices with $\Vert P(x_j\otimes y_j)\Vert \ge \epsilon$ by

$$C(1+L){\epsilon }^{-2}\exp (C{\epsilon }^{-2}\sqrt{d}\log (2+d)).$$

Here is the proof mechanism. Choose a Hilbert–Schmidt orthonormal matrix basis $A_1,\dots ,A_r$ of the testing space, $r\le d$. The covariance matrices $R_1=\sum A_a{A}_a^{\ast }$ and $R_2=\sum A_a^{\ast }{A}_a$ each have trace at most $d$. Their spectral subspaces above eigenvalue $\sqrt{d}$ therefore have dimension at most $\sqrt{d}$.

A covering argument controls vectors with a substantial projection onto these exceptional spaces. For any index set $J$, the graph hypothesis gives

$${\Vert \frac{1}{|J|}\sum _{j\in J}{x}_j\Vert }^2\le \frac{1+L}{|J|}+\eta .$$

Vectors whose projections lie in one small ball away from zero have an average bounded away from zero, so this inequality bounds the size of that cluster. A finite net of the at-most-$\sqrt{d}$ dimensional ball bounds all discarded indices. (OpenAI, 2026c; OpenAI, 2026a)

Gaussian comparison controls the remaining testing directions

Compress the $A_a$ to the low-covariance complements, obtaining $B_a$. The Gaussian matrix $G={\sum }_a{g}_a{B}_a$ has expected operator norm at most $C{d}^{1/4}\sqrt{\log (2+d)}$. The source proves this without ambient-dimension dependence: self-adjoint dilation gives covariance operator norm at most $\sqrt{d}$ and trace at most $2d$, and Gaussian integration by parts yields

$$\mathbb{E}\mathrm{tr}{H}^{2q}\le (2q-1)!!\Vert S{\Vert }^{q-1}\mathrm{tr}S.$$

Taking $q$ comparable to $\log (2+d)$ gives the stated norm bound. The remaining large-projection coefficient vectors have norm at least $\epsilon /2$. Extract a separated subset of size $l$. Gaussian comparison gives $c\epsilon \sqrt{\log l}\le \mathbb{E}\Vert G\Vert$, hence $\log l\le C{\epsilon }^{-2}\sqrt{d}\log (2+d)$. Combining this with the cluster bound proves tensor packing.

The selected-law proof preserves only the independence it needs

Packing shows that a slowly growing covariance subspace captures few of the relevant tensor terms; square-summable control handles its orthogonal remainder. These estimates support the selected-length extension while preserving the three auxiliary groups’ specified product laws. They do not assert independence of an entire state from all three groups at once.

The final square argument uses a correlation chosen with minimal support and then maximal auxiliary-slot count. Its stable conditional-expectation identities allow regrouping the square into a new nonzero correlation with more auxiliary slots, contradicting maximality. The fourfold and all-length manuscripts supply that closure theorem separately from packing. The distinct-triple companion adapts the integer configuration and divisibility estimates for arbitrary distinct nonzero slopes; it is not obtained by rescaling every triple to $1,2,3$. These distinctions preserve the exact ranges of the four attributed results. (OpenAI, 2026d)

See Also

Rokhlin’s multiple-mixing problem is the source of the higher-order mixing input. The trilinear Hilbert transform supplies the local harmonic-analysis estimates that the triple companions strengthen for oscillation.

🏷️ Manuscripts and verification scope

• Pointwise Multiple Ergodic Averages for Mixing Transformations — pinned manuscript and companion source.
• Pointwise convergence of fourfold ergodic averages for mixing transformations — pinned manuscript and companion source.
• Triple ergodic averages with distinct integer slopes — pinned manuscript and companion source.
• Pointwise convergence of triple ergodic averages for mixing transformations — pinned manuscript and companion source.

No family scope page was present at lean/docs/154.md in the pinned snapshot. This does not establish that all related formalization code is absent. No Lean build, axiom audit, or manuscript-to-formal-statement equivalence check was performed for these notes.

OpenAI collection provides prerequisites and the reading sequence; coverage register records completed and pending families.

🐻  OpenAI 2026a. Pointwise convergence of fourfold ergodic averages for mixing transformations.

🐻  OpenAI 2026b. Pointwise convergence of triple ergodic averages for mixing transformations.

🐻  OpenAI 2026c. Pointwise Multiple Ergodic Averages for Mixing Transformations.

🐻  OpenAI 2026d. Triple ergodic averages with distinct integer slopes.