Taking the largest disk average without creating too much variation

← All digests

Family 085 · Real and complex analysis
Importance 4/5 · Provisional editorial rating
Sources consulted 2026-10-14

1. The question

Averaging usually smooths a function. But suppose that at every point we try all possible disk sizes and keep the largest average. Different points may choose completely different radii, so the resulting function is no longer an ordinary convolution. Can that selection create substantially more variation than the original function possessed?

This is a regularity question for the Hardy–Littlewood maximal operator, one of the basic tools of analysis. Bounds away from the integrability endpoint have been known for decades. At the endpoint, the usual short argument breaks down, and the geometry of the averaging sets becomes decisive.

Family 085 contains one manuscript. It claims that, for disks centered at the evaluation point in the plane, the integral of the output’s gradient is bounded by an absolute constant times the integral of the input’s gradient. The output itself need not be integrable over the whole plane. Keeping those two conclusions separate is essential to understanding the result.

2. Background and essential definitions

For a locally integrable function $f:\mathbb R^2\to\mathbb R$, its centered disk maximal function is

$$Mf(x)=\sup_{r>0}\frac{1}{\pi r^2} \int_{B(x,r)}|f(y)|\,dy.$$

Here $B(x,r)$ is the open disk of radius $r$ centered at $x$. Every positive radius is allowed, and the absolute value is taken before averaging. The word centered means that the disk’s center must be the output point; allowing arbitrary disks containing that point defines a different operator.

The Sobolev space $W^{1,1}(\mathbb R^2)$ consists of globally integrable functions whose first weak derivatives are also globally integrable. A weak derivative is defined by integration by parts against smooth test functions. It allows a notion of gradient for functions that need not be differentiable everywhere in the classical sense.

The quantity

$$\int_{\mathbb R^2}|\nabla f(x)|\,dx$$

measures total first-order variation, with Euclidean length used for the gradient vector. Membership in $W^{1,1}_{\mathrm{loc}}$ requires these integrability properties only on bounded regions.

There is also a larger space, $BV$, of functions of bounded variation. Their distributional derivatives may be measures with singular parts, such as mass concentrated on a jump curve. A bound on a derivative measure alone is therefore weaker than proving that the derivative is represented by an integrable function. The manuscript addresses both the size estimate and this regularity distinction.

3. Development history and previous partial results

Kinnunen’s 1997 Sobolev theorem established boundedness of the maximal operator on $W^{1,p}$ for $p>1$. A central pointwise comparison bounds the gradient of the maximal function by a maximal function of the input gradient. The strong $L^p$ maximal inequality then completes the argument.

At $p=1$, that last step fails: the maximal operator is not strongly bounded on $L^1$. Hajłasz and Onninen’s 2004 paper asked whether the gradient estimate could nevertheless survive. Failure of the standard proof did not settle the endpoint question.

One dimension supplied important partial solutions. Tanaka treated the uncentered operator, and Aldaz and Pérez Lázaro obtained the sharp variation constant one with additional regularity conclusions. Kurka’s centered one-dimensional theorem controlled variation by an absolute constant. Intervals have a linear ordering that helps organize their extrema; disks in the plane lack that ordering.

Higher-dimensional results imposed structure on the input or changed the averaging geometry. Radial inputs, which depend only on distance from the origin, and block-decreasing inputs admitted results unavailable for arbitrary functions. Weigt proved variation estimates for dyadic maximal operators and for the uncentered cube operator. Axis-parallel cubes, dyadic cubes and centered disks have different overlap patterns.

Fractional maximal operators introduce a positive power of radius. Endpoint gradient estimates for these operators also advanced, but the positive order supplies a geometric advantage that disappears at order zero. Lahti and Weigt’s conditional regularity theorem explained how a suitable variation bound could yield Sobolev regularity for the centered operator. M1 claims the missing planar centered-disk bound and gives its own argument removing singular derivative mass.

4. What the new manuscripts claim

For every real-valued $f\in W^{1,1}(\mathbb R^2)$, M1 states that $Mf$ belongs to $W^{1,1}_{\mathrm{loc}}(\mathbb R^2)$, has a globally integrable weak gradient, and satisfies

$$\int_{\mathbb R^2}|\nabla Mf(x)|\,dx \leq C\int_{\mathbb R^2}|\nabla f(x)|\,dx,$$

where $C$ is an absolute finite constant. No radiality, monotonicity or compact-support hypothesis is imposed. The manuscript does not determine the best constant.

It also states an intermediate signed estimate. For a smooth compactly supported real function $g$, let

$$A_tg(x)=\frac{1}{\pi t^2}\int_{B(x,t)}g(y)\,dy, \qquad S_{a,b}g(x)=\sup_{a\leq t\leq b}A_tg(x).$$

These averages retain the sign of $g$. When $b/a$ is a power of two, the claimed gradient bound for $S_{a,b}$ has a constant independent of the number of radius scales between $a$ and $b$. A single radius is included.

That uniformity is the bridge to all radii. The signed formulation is also useful when a local mean is subtracted from the input: subtracting a constant changes average values but not their gradients. An estimate available only after replacing the localized function by its absolute value would not preserve the same cancellation.

5. Why the advance matters

Importance: 4/5. This would resolve a major planar case of a longstanding endpoint regularity problem for a foundational operator. It concerns arbitrary Sobolev inputs and the usual centered disks, rather than a symmetry class or an auxiliary averaging geometry.

The statement reveals a difference between controlling a function and controlling its variation. For a nonzero nonnegative compactly supported function, large disks centered far away can still collect its total mass. In the plane this gives a tail comparable from below to a constant times $|x|^{-2}$, whose integral diverges logarithmically. Thus one cannot expect $Mf$ itself to lie in global $L^1$, even for very pleasant inputs.

The gradient can nevertheless be globally integrable. The theorem claims precisely that more selective regularity property. It would show that choosing the best disk at each point does not create unlimited first-order variation.

The score also reflects the remaining scope. This is the plane with centered disks. It does not, as stated, resolve all dimensions, all convex averaging bodies, the optimal constant, or continuity of the map taking an input to the gradient of its maximal function.

6. The main idea

Work with finitely many scales while keeping constants uniform

The proof first considers radii in a finite dyadic band. For each fixed radius, differentiating the average is manageable. The difficulty is that the maximizing radius changes with the center, and summing a separate estimate for every radius scale would accumulate an unacceptable factor.

At an interior maximizing radius, the derivative with respect to radius vanishes. Rotational symmetry supplies another moment identity for a disk. Together, these identities allow cancellation in thin boundary shells, especially when much of the relevant mass is concentrated near one shell point.

The manuscript packages this into a directional estimate: if most gradient mass points into a given cone, then components of the output gradient pointing against that cone can be bounded. Localized error pieces with zero total vector mass become cheaper when observed at a much larger scale.

Give cancellation and changing directions separate budgets

The input gradient is distributed over nested spatial grids using nonnegative tent functions. Scalar positive and negative pieces are paired from fine scales upward. Each cancellation is charged at its scale of origin, with a discount when it is used at coarser scales. The resulting cell costs have a total bounded by the original gradient’s integral.

A second construction tracks vector mass along refinement paths. Its purpose is different: it controls the price of changing the direction used in the local estimate. A direction change creates an endpoint between radius bands; variation of the direction choice across space creates extra terms when integrating by parts.

The manuscript uses a finite-state distribution of directional choices to budget both costs. This avoids pretending that there is one globally suitable direction for a general input, or that selecting a new direction at every scale is free.

Preserve cancellation until the final assembly

Consecutive scales with the same chosen direction form a run. Local estimates are applied to each run and averaged over the possible histories of directional choices. Baseline averages pay for run endpoints.

When the spatially varying history weights are differentiated, their normalization gives a cancellation: derivatives of probabilities in a row sum to zero. The argument subtracts a preceding baseline before taking absolute values. That order matters, because an early absolute value would discard the identity that makes the remaining costs summable.

Finally, approximation and expanding radius bands give a bounded derivative measure for the maximal function. Signed localization, subtraction of local means and Poincaré’s inequality show that this measure has no singular part. It is therefore a genuine integrable weak gradient, not merely a finite measure.

7. Scope, limitations, and open questions

The input is real-valued and belongs to the inhomogeneous space $W^{1,1}$: both it and its weak derivatives are globally integrable. The conclusion gives local Sobolev regularity plus a global gradient bound, not global membership of $Mf$ in $W^{1,1}$.

All averaging sets are Euclidean disks centered at the output point in $\mathbb R^2$. Uncentered disks, higher-dimensional balls and other shapes should not be substituted silently. Earlier theorems for cubes or radial inputs remain distinct results.

The estimate asserts boundedness. Continuity of the gradient map under convergence of inputs is an additional question, as illustrated by separate one-dimensional continuity results. Nor does the manuscript claim the sharp numerical value of $C$.

The exposition follows M1’s stated theorem and its multiscale mechanism. It does not assess the proof’s correctness. The main mathematical boundary to retain is the passage from a finite variation measure to an absolutely continuous derivative: both are needed for the claimed endpoint Sobolev conclusion.

8. Sources and further reading

M1’s introduction gives the theorem, the signed finite-band estimate and an unusually detailed account of the two separate multiscale constructions. Those are the sources for sections 4 and 6.

The linked papers by Kinnunen and Hajłasz–Onninen establish the original regularity setting and endpoint question. Kurka’s interval theorem, Weigt’s cube theorem and Lahti–Weigt’s conditional result explain three different partial advances without identifying them with the planar centered-disk claim.

Source manuscripts

• M1. OpenAI, An Endpoint Gradient Bound for the Centered Disk Maximal Operator (September 26, 2026).

Sources consulted on 14 October 2026 at repository revision 1581d4c. Download the manuscript citations.