Riesz transforms and rectifiability in higher codimension

The explanation

Picture a flat two-dimensional sheet sitting inside four-dimensional space. Every point of the sheet sends an arrow to every other point. The arrow lies along the line between them. Its length is 1 divided by the squared distance. A point 1 unit away gives an arrow of length 1. A point 2 units away gives an arrow of length 1/4.

Now add the arrows ring by ring. The ring between distance 1 and 2 has area 3π, about 9.42. Its arrows have lengths between 1/4 and 1. The ring between 10 and 20 has area 300π, about 942. Its arrows are 100 times shorter. So every ring adds a total length between about 2.36 and 9.42, near or far. Distant rings matter as much as close ones. On a flat sheet this does no harm. Each arrow has a partner from the opposite side, and the two cancel.

This adding-up of arrows is the Riesz transform. It can also weight each point by a function f, a number attached to each point. A hard truncation skips all points closer than some cutoff ε. The transform is bounded when its output is never more than a fixed multiple of its input. Size here means the root-mean-square, the square root of the average squared value. The paper asks for one multiple, called C_R, that works for every cutoff at once.

The question is what a bounded transform says about shape. The paper fixes two notions. A measure is Ahlfors–David regular of dimension n when balls centered on it hold mass comparable to r^n. That is, the mass sits between r^n/C and C·r^n for a fixed C. On the flat sheet, a ball of radius r holds area πr². That is about 3.14 at radius 1 and about 314 at radius 10. So C = π works. Such a measure is uniformly rectifiable when it has big flat-like pieces at every place and scale. In every ball, a fixed share of the mass lies on a mildly bent copy of a flat disc. Mildly bent means distances are stretched by at most a fixed factor M. The share and M must not change from ball to ball.

David and Semmes built uniform rectifiability for singular integrals like this one. Their question asks whether a bounded Riesz transform alone forces this geometry. Mattila, Melnikov and Verdera settled the planar case in 1996. Curvature methods also handled one-dimensional sets in higher dimensions. Nazarov, Tolsa and Volberg settled codimension one in 2014. Codimension means how many dimensions the space has beyond the set. Mas and Tolsa covered every dimension in 2014, but under a stronger hypothesis about changes across scales. That left sets of dimension 2 up to d−2, in a space of dimension d. Tolsa stated this gap explicitly in his 2026 survey. Our sheet in four-dimensional space is the smallest such case.

Nazarov, Tolsa and Volberg named their main obstacle. Their proof relied on a tool called a maximum principle. They reported no known version of it in higher codimension. The paper takes another road. Its core difficulty is to carry flatness down to smaller scalesSource 1. It measures bending by the excess, the scaled average squared distance from the best flat plane.

Here is the idea in plain words. Zoom in on a nearly flat piece. Measure how far it rises off its plane, and blow those heights up by the excessSource 2. The rescaled heights obey a limiting equation. That equation forces them to be flatSource 3. So the zoomed-in piece is flatter than before. A counting argument then limits how often this can failSource 4. A theorem of David and Semmes turns that count into the bent discsSource 4.

So the paper proves the full answer for this range, with a Lean proof listed. For integers d ≥ 4 and 2 ≤ n ≤ d−2, a regular measure with bounded transform is uniformly rectifiableSource 5. The share and the stretch factor depend only on d, n, C and C_R. Now picture a regular set scattered like dust at every scale, with no big flat-like pieces. Read in reverse, the theorem means its transform cannot have one bound C_R for every cutoff. The theorem needs regularity at the start. It assumes no principal values, the limits as the cutoff shrinks to zeroSource 6. It does not treat sets of fractional dimension.

How sure is this? A computer proof checker called Lean has checked a formal version of the main theorem. That version includes unbounded sets and works at every point and admissible radius. The paper's corollary on variation and principal values is outside that checkSource 7. No specialists have published a review yet.

What it doesn't show. This does not cover sets of fractional dimension, and the variation and principal-value corollary is outside the Lean check.