Riesz transforms and uniform rectifiability

Source and verification

This article explains family 081 of the October 6, 2026 OpenAI collection. The research conclusions are attributed to its manuscripts. The source review covers statements, introductions, and the proof mechanisms described below; it is not an independent verification of the complete proofs.

Overview

Can boundedness of a singular integral force a rough set to contain large pieces of Lipschitz surfaces? Family 081 claims this implication for integer-dimensional Riesz transforms in higher codimension. Its key mechanism converts small transform oscillation into quantitative improvement of geometric flatness.

🏷️ Preliminaries: regular measures and Lipschitz pieces

Let $\mu$ be a Radon measure on ${\mathbb{R}}^d$ with support $E$. It is $n$-Ahlfors–David regular if there is $C\ge 1$ such that

$$\begin{array}{c}{C}^{-1}{r}^n\le \mu (B(x,r))\le C{r}^n,\\ x\in E,0<r\le \mathrm{diam}E.\end{array}$$

The two inequalities say that every relevant ball sees the amount of mass expected of an $n$-dimensional object. They do not say that the object is smooth or even a graph.

The measure is uniformly $n$-rectifiable if there are fixed $M<\infty$ and $\theta >0$ such that every such ball contains at least $\theta r^n$ of its mass in the image of an $M$-Lipschitz map from an $n$-dimensional radius-$r$ ball. The uniformity in position and scale makes this substantially stronger than a qualitative covering by countably many Lipschitz images.

🏷️ The Riesz transform and the claimed implication

For a positive truncation parameter $\epsilon$, define

$$\begin{array}{c}{R}_{\mu ,\epsilon }f(x)\\ ={\int }_{|x-y|>\epsilon }\frac{x-y}{|x-y{|}^{n+1}}f(y)d\mu (y).\end{array}$$

Its vector-valued kernel remembers the direction from $y$ to $x$. A uniform $L^2$ bound means that the norms of these truncated operators are bounded independently of $\epsilon$.

Family 081, as stated in the manuscript

Suppose $d\ge 4$ and $2\le n\le d-2$. If $\mu$ is $n$-Ahlfors–David regular and

$$\begin{array}{c}\Vert R_{\mu ,\epsilon }f{\Vert }_{L^2(\mu ;{\mathbb{R}}^d)}\le C_R\Vert f{\Vert }_{L^2(\mu )},\\ \text{for every }\epsilon >0,\end{array}$$

then $\mu$ is uniformly $n$-rectifiable, with quantitative constants depending only on the dimensions, regularity constant, and operator bound. (OpenAI, 2026)

The hypothesis is about all hard truncations. It does not assume existence of a pointwise principal value or boundedness of a pointwise maximal truncation. Degenerate supports with no admissible positive radii make the geometric conclusion vacuous; the substantive case is a nontrivial regular support.

🏷️ Flatness and its normalization

A typical squared flatness quantity is

$$\begin{array}{c}{\beta }_{\mu ,2}(x,r{)}^2\\ =\frac{\underset{P}{\inf }{\int }_{B(x,r)}\mathrm{dist}(y,P{)}^{2}d\mu (y)}{r^{n+2}},\end{array}$$

where $P$ ranges over affine $n$-planes. The powers of $r$ make the expression invariant under simultaneous rescaling of space and the measure’s natural $n$-dimensional normalization.

The model example is $n$-dimensional measure on a plane. Oddness cancels the Riesz kernel on symmetric annuli, and the transform reduces to the classical $n$-dimensional Riesz transform along the plane. The difficult direction is to infer large Lipschitz pieces when no plane or parametrization is initially given.

🏷️ Normal heights and fractional rigidity

The source first obtains balls where the support is close to a plane. It then examines the displacement normal to that plane. Ordinary compactness can show that the displacement tends to zero; improvement of flatness needs an estimate relative to its initial size. The normal heights are therefore divided by that size before a second limiting argument.

The transform is tested against heights with compact cutoffs. Its antisymmetric kernel generates a positive fractional-energy expression controlling the normalized heights. Pairings with mean-zero Lipschitz tests supply the limiting equation. Here the mean is taken against the actual moving measure, not an already assumed plane measure.

The limiting operator has order one. Together with the specified growth and integrability conditions, its rigidity theorem forces the limiting height to be affine. Tilting the comparison plane by that affine function improves the original geometric excess. Tail control and treatment of the singular diagonal are essential to justify this limit; formal linearization alone would not suffice.

🏷️ From local improvement to uniform rectifiability

The remaining problem is to count scales where flatness fails. The manuscript uses dyadic cells on the support and charges failures of propagation to cells with detectable transform oscillation. The desired Carleson packing estimate has the schematic form

$$\sum _{Q\subset Q_0,Q\text{ bad}}\mu (Q)\le C\mu (Q_0).$$

It says that a point cannot be repeatedly charged at too many scales on average. Combined with the geometric hypotheses and the bilateral flatness information in the source, this enables the David–Semmes criterion for uniform rectifiability. A local affine limit without this packing stage would not yet give the uniform conclusion.

🏷️ Key steps and proof outlines

Mean-zero tests turn operator boundedness into a packing budget

A localized test ${\phi }_Q$ is supported near a dyadic cell $Q$, has controlled Lipschitz constant at scale $\ell (Q)$, and has zero integral against $\mu$. Normalize it by $\ell (Q{)}^{n/2}$. If $P$ is much smaller than $Q$, subtract the value of the larger test at the center of $P$ inside their inner product. The mean-zero condition kills that constant, and the Lipschitz bound supplies a factor proportional to $\ell (P)/\ell (Q)$.

Summing these decaying cross-scale inner products gives an almost-orthogonality estimate for the family of tests. The uniform Riesz-transform bound can then pay for all cells where a corresponding mean-zero pairing is large. This is the analytic oscillation packing lemma. It turns one operator norm bound into a total budget over locations and scales, rather than applying the bound separately and losing the number of scales. (OpenAI, 2026)

Relative normalization retains the shape of a vanishing error

Suppose a sequence of supports becomes close to a plane $P$. Write its normal displacement as $h_j$ and let ${\delta }_j$ measure its initial excess. Ordinary convergence gives only $h_j\to 0$. To prove that a smaller ball is better approximated by some tilted plane, study $w_j=h_j/{\delta }_j$ instead.

The local energy estimate controls a fractional difference energy of the normalized heights, schematically

$$\iint \frac{|w_j(x)-w_j(y){|}^2}{|x-y{|}^{n+1}}d{\mu }_j(x)d{\mu }_j(y),$$

with compact cutoffs and the source’s normalization. The exponent $n+1$ corresponds to a half-derivative energy on an $n$-plane. Oddness of the Riesz kernel, paired with differences of heights, produces the positive squared difference. Cutoff terms and tails must be estimated separately; they have no automatic sign.

The measures move with $j$, so standard compactness for functions on one fixed domain is not immediately applicable. The manuscript constructs finite partitions in projected regions and controls the measures ${\mu }_j$, $w_j{\mu }_j$, and $w_j^2{\mu }_j$ together. This supplies one common subsequence and one limiting height, preserving enough quadratic information to recover strong local convergence.

The limiting equation has an affine rigidity conclusion

The normalized pairings with mean-zero tests converge to an order-one equation on the limiting plane. Its natural model is the half-Laplacian, whose quadratic form is the fractional energy above. A local solution of such an equation need not be affine; the source also uses its specified growth and weighted tail bounds. Those global conditions exclude higher-growth or nontrivial exterior contributions.

The rigidity theorem identifies the limiting height with an affine function $v$. Polarization then converts weak linear and quadratic convergence into strong convergence:

$$\int |w_j-v{|}^{2}d{\mu }_j=\int w_j^{2}d{\mu }_j-2\int w_{j}vd{\mu }_j+\int v^{2}d{\mu }_j.$$

Each term has a controlled limit on the localized region. The affine function corresponds geometrically to tilting and translating $P$. Thus the residual normal excess after changing the comparison plane is smaller than the original normalization. This closes a contradiction proof of relative excess improvement.

Improvement must propagate through a quantitatively large set of scales

The final counting argument marks flat seed cells and follows chains of descendants. A chain that fails to propagate good geometry must encounter a cell with analytically detectable oscillation or another explicitly charged obstruction. The finite seed-counting lemma assigns these charges before choosing a point, preventing repeated arbitrary reassignment along different chains.

Summing over chains and integrating against $\mu$ yields a Carleson estimate for bad cells. Bilateral flatness requires both that the support lies near a plane and that the relevant portion of the plane lies near the support. The manuscript verifies this stronger coefficient and then applies the bilateral weak geometric lemma criterion for uniform rectifiability. The lower Ahlfors–David bound is essential here: upper growth alone would permit large holes and would not give the required big Lipschitz pieces.

This explains the full implication: bounded transforms pay for oscillation; low oscillation yields flat seeds and affine height limits; relative improvement propagates those seeds; packing converts the local geometry into a uniform statement on every ball.

See Also

• Family 073: distance sets starts from a one-sided Frostman condition, whereas this family requires two-sided Ahlfors–David regularity.
• Family 085: maximal-function gradients also turns scale-independent local estimates into a global geometric-analytic conclusion.

🏷️ Manuscripts and verification scope

• Riesz transforms and uniform rectifiability in higher codimension — the pinned manuscript; its directory supplies the source and citation metadata.

The repository provides a formalization-scope page. Its scope description covers quantitative uniform rectifiability; it explicitly excludes the variation and principal-value corollary. These are statements about the repository’s documentation: no Lean build, axiom audit, or comparator-equivalence check was performed for this article.

OpenAI collection gives the reading sequence and source policy; coverage register records which families have articles and which remain to be written.

🐻  OpenAI 2026. Riesz transforms and uniform rectifiability in higher codimension.