Critical local smoothing for the three-dimensional wave equation

Source and verification

This article explains family 079 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

A wave can be poorly behaved at an individual time while becoming more integrable after averaging in time. Family 079 claims the sharp strict local-smoothing range for the Euclidean wave equation in three spatial dimensions. Its critical estimate permits every positive Sobolev loss at $p=3$.

🏷️ Preliminaries: half-waves and Sobolev regularity

Use $\hat{f}(\xi )=\int e^{-ix\cdot \xi }f(x)dx$ and define

$$\begin{array}{c}Uf(x,t)=e^{it|D|}f(x)\\ =\frac{1}{(2\pi {)}^3}\int e^{i(x\cdot \xi +t|\xi |)}\hat{f}(\xi )d\xi .\end{array}$$

The half-wave solves ${\partial }_t^{2}u-\Delta u=0$ with compatible initial velocity. The multiplier $J^{\alpha }=(1-\Delta {)}^{\alpha /2}$ measures $\alpha$ derivatives; the norm $\Vert J^{\alpha }f{\Vert }_p$ is a Bessel-potential Sobolev norm.

At a fixed nonzero time, the relevant three-dimensional $L^p$ derivative loss for $p>2$ is $1-2/p$. Local smoothing asks how much of that loss can be recovered by integrating over a bounded time interval. The estimate below uses $t\in [1,2]$ and integrates over all of space.

🏷️ The claimed critical and full ranges

Family 079, as stated in the manuscript

For every $\epsilon >0$,

$$\begin{array}{c}\Vert Uf{\Vert }_{L^3({\mathbb{R}}^3\times [1,2])}\\ \le C_{\epsilon }\Vert J^{\epsilon }f{\Vert }_{L^3({\mathbb{R}}^3)}.\end{array}$$

Consequently, for $2<p<\infty$ and $\alpha >\max \{0,1-3/p\}$,

$$\Vert Uf{\Vert }_{L^p({\mathbb{R}}^3\times [1,2])}\le C_{p,\alpha }\Vert J^{\alpha }f{\Vert }_p.$$

These are strict-regularity assertions. (OpenAI, 2026)

At $p=3$, fixed-time propagation costs $1/3$ of a derivative, whereas the stated spacetime estimate costs arbitrarily little. This does not assert the lossless $\alpha =0$ estimate.

🏷️ Frequency localization explains the exponent

Suppose $\hat{f}$ is supported where $|\xi |\asymp N$. The critical estimate becomes $\Vert Uf{\Vert }_3{\lesssim }_{\epsilon }{N}^{\epsilon }\Vert f{\Vert }_3$. The $L^2$ energy estimate has no loss, while the corresponding annular $L^{\infty }$ kernel estimate has cost of order $N$.

Interpolating the $p=3$ and $p=\infty$ estimates, the interpolation parameter satisfies $1/p=(1-\theta )/3$, so $\theta =1-3/p$. The derivative cost is therefore $1-3/p$ plus an arbitrarily small positive amount. Interpolating with $p=2$ gives an arbitrarily small positive loss for $2<p<3$. Frequency summation then uses the strict gap in $\alpha$ to absorb the remaining dyadic losses.

This calculation explains the shape of the theorem. Its difficult input is the critical annular estimate, not the interpolation.

🏷️ Packet energy and the multiscale obstruction

The source changes coordinates so that the relevant rays have velocities of the form $(A,|A{|}^2)$ with $A\in {\mathbb{R}}^2$. It decomposes the wave into localized packets and assigns their squared $L^2$ norms as positive energy masses. At the final scale, a synthesis estimate reduces the cubic spacetime norm to a weighted sum of masses raised to the power $3/2$.

Intermediate windows can be anisotropic: one width follows a curved direction and another follows a nearly flat direction. The proof uses separate round and cylindrical models with different weights. These weights express the amount of energy permitted in a window at its position, angle, and time scales.

A subtle point is that propagation copies energy into successive time intervals. Those copies are not disjoint portions of the original input. Summing them as if they were orthogonal would erase the principal difficulty. The manuscript’s normalized cubic functional includes the time-scale factor explicitly.

The contradiction argument selects nearly extremal mass and width profiles. It records finite descriptions of the relevant phase variables and subtracts the entropy dictated by the regularity bounds. The resulting entropy excess is constrained both by projection estimates and by recoverability from the selected windows. The source uses the incompatibility of these requirements to bound the cubic packet functional.

🏷️ Time averages versus time maxima

An $L^p$ norm in time does not control the supremum in time for arbitrary functions. A Sobolev embedding in the time variable can supply such control, but consumes extra derivatives. For the half-wave, time differentiation corresponds to a spatial factor $|\xi |$, so this cost becomes additional spatial regularity.

The source accordingly states a maximal half-wave consequence for $p\ge 3$ with $s>1-2/p$, stronger than the smoothing hypothesis $\alpha >1-3/p$. At $p=3$, this means $s>1/3$ for the maximal conclusion. One must not replace that condition by $s>0$ merely because the spacetime theorem allows it.

🏷️ Key steps and proof outlines

Null coordinates reveal the curved ray geometry

The half-wave cone has one radial and two angular variables. In a suitable patch, the source uses null coordinates that express propagation through a quadratic angular phase. Its physical kernel contains

$$(b-z_{\perp })\cdot \eta -(y+z_3)\rho -(T-z_3)\frac{|\eta {|}^2}{4\rho },$$

where $z=(z_{\perp },z_3)$ is the input point, $(b,y,T)$ are the transformed output coordinates, and $(\eta ,\rho )$ are dual variables with $\rho$ bounded away from zero on the patch. Differentiating in $\eta$ identifies the moving packet center. This is the source of the parabolic angular velocity description; it is a coordinate identity for the cone, not an assertion that the wave equation is globally the Schrödinger equation. (OpenAI, 2026)

The input on the physical initial slice must then be analyzed in these new coordinates. Kernel decay and the Jacobian of the frequency change control its packet energy. Without this initial-state calculation, a bound for an abstract propagation experiment would not yet imply a bound in terms of $\Vert f{\Vert }_3$.

Cubic synthesis requires a time-normalized energy budget

Let $m_{\alpha }$ denote the nonnegative energy assigned to a packet window. A cubic synthesis estimate involves $m_{\alpha }^{3/2}$ because amplitude scales as the square root of energy. Concentrating a fixed total energy in fewer windows therefore increases the cubic functional. The geometric problem is to prevent that concentration from persisting along too many time intervals.

The same packet can visit many intervals. Its energy at two times is not two disjoint pieces of initial energy. The manuscript’s cylinder and round-cone experiments include explicit interval lengths and weights in their functionals. Ordinary splitting compares those normalized quantities, with copying in time accounted for. This is why a bare estimate $\sum m_{\alpha }\le \Vert f{\Vert }_2^2$ cannot simply be applied to every temporal generation simultaneously.

Canonical labels distinguish linear and quadratic errors

A canonical name records a time cell, a velocity cell, and transported position cells. Suppose an angular velocity is known only up to error $s^v$, and time is known up to $s^a$. Backflow creates a first-order position uncertainty of size $s^{a+v}$. After subtracting the linear tilt, the remaining curved-coordinate error is quadratic, of size $s^{a+2v}$. These are different resolutions and generate different entropy contributions.

The source checks that an observation in one frame determines the canonical name in another up to a controlled number of choices. Entropy chain rules convert this finite recoverability into inequalities. The regularity bounds provide a baseline information cost; the proof tracks the excess above that baseline. At a tangent profile, the cost of new time bits and the cost of refining angular or spatial bins give lower bounds for entropy derivatives.

Projection amplification supplies competing lower bounds when directions remain genuinely spread. If those bounds fail, the directions concentrate near a plane, and a lower-dimensional estimate applies. Thus the geometric dichotomy is used quantitatively: both alternatives must improve the same normalized functional with compatible weights.

Cylinder bounds transfer to thin round-cone configurations

The round-cone argument splits according to the shape of its occupied windows. Nearly round configurations are handled directly by round labels at the uncertainty scale. Thin configurations are put into charts where a cylindrical experiment describes their dynamics. The transfer must preserve both phase-space volume and packet overlap estimates.

For example, the overlap kernel is estimated after scaling angular and radial frequency variables to a fixed compact set. The quadratic propagation phase remains bounded in those coordinates. Integration by parts then gives decay in the two spatial-position differences, with the normalization determined by the frequency Jacobian. Summing this kernel by rows and columns controls analysis and synthesis through the chart change. This verifies that the cylinder estimate can be used on the actual oscillatory arrays, not only on their geometric supports.

Returning to physical frequencies

The resulting annular estimate is uniform over the finite cone patches covering $|\xi |\asymp N$. Summing those finitely many charts gives the critical $N^{\epsilon }$ bound. For a target regularity $\alpha >0$ at $p=3$, choose $\epsilon <\alpha$ and write each dyadic contribution as $2^{-j(\alpha -\epsilon )}$ times its $\alpha$-derivative norm. The geometric factor is summable. Interpolation then gives the wider range already computed above.

The proof therefore has three separate interfaces: geometric concentration controls normalized packet energies; frame estimates convert that control into an oscillatory norm; the physical initial-state and frequency calculations express the result in the original Sobolev norm. Each interface is required before the local smoothing claim follows.

See Also

• Family 080: Schrödinger convergence concerns a different evolution and claims an exact Sobolev endpoint for its spatial maximal problem.
• Family 078: Bochner–Riesz explains why a fixed-radius multiplier estimate and a maximal summation estimate are distinct.

🏷️ Manuscripts and verification scope

• Critical local smoothing for the three-dimensional wave equation — the pinned manuscript; its directory supplies the source and citation metadata.

No family-specific scope page was available at lean/docs/079.md in this snapshot. This records the catalogue lookup, not a claim that no related Lean code exists. No Lean build or axiom audit was performed.

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

🐻  OpenAI 2026. Critical local smoothing for the three-dimensional wave equation.