Bochner–Riesz multipliers in three dimensions

Source and verification

This article explains family 078 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

Bochner–Riesz summation softens the boundary of a frequency ball. Family 078 claims that in three dimensions every positive smoothing order suffices at $p=3$, and derives the full strict range of orders and exponents from this central estimate.

🏷️ Preliminaries: multipliers and smoothing order

With Fourier transform $\hat{f}(\xi )=\int f(x)e^{-2\pi ix\cdot \xi }dx$, define

$$\begin{array}{c}{T}_{\delta }f(x)\\ ={\int }_{{\mathbb{R}}^3}(1-|\xi {|}^2{)}_{+}^{\delta }\hat{f}(\xi )e^{2\pi ix\cdot \xi }d\xi ,\\ \delta >0.\end{array}$$

Here $a_{+}=\max (a,0)$. The multiplier equals zero outside the unit ball and tapers to zero at its boundary. Larger $\delta$ gives more smoothing there. Replacing $|\xi |$ by $|\xi |/R$ produces the summation operator at radius $R$.

Plancherel immediately gives an $L^2$ bound, since the multiplier has absolute value at most one. For $p\ne 2$, boundedness of the multiplier as a function does not imply boundedness of its operator on $L^p$. Spatial oscillation and cancellation become decisive.

🏷️ The claimed range

Family 078, as stated in the manuscript

For every $\delta >0$, $T_{\delta }$ is bounded on $L^3({\mathbb{R}}^3)$. The resulting strict range is

$$\begin{array}{c}\delta >\max \left\{3\left|\frac{1}{p}-\frac{1}{2}\right|-\frac{1}{2},0\right\},\\ 1\le p\le \infty .\end{array}$$

Constants may depend on the fixed order and exponent. (OpenAI, 2026)

At $p=3$ and $p=3/2$, the threshold is zero, but the assertion still requires $\delta >0$. It does not prove boundedness of the unsmoothed ball multiplier. Nor does a bound uniform in the radius mean boundedness of the pointwise supremum over all radii.

🏷️ Why the kernel predicts this threshold

Stationary phase gives a leading oscillatory kernel at large distance with size about $|x{|}^{-2-\delta }$ in dimension three. Absolute integrability at infinity would require

$${\int }_1^{\infty }{r}^2{r}^{-2-\delta }dr<\infty ,$$

which holds when $\delta >1$. Young’s inequality then supplies the easy $L^1$ and $L^{\infty }$ range. For smaller positive orders, cancellation must compensate for the lack of an integrable kernel.

The central $L^3$ result interpolates with the $L^2$ bound and, by duality, supplies the central interval $3/2\le p\le 3$ at every positive order. Interpolation of the analytic family of complex orders with the integrable-kernel estimates gives the outer slopes. For $p\ge 3$ the boundary is $\delta =1-3/p$; duality gives $\delta =3/p-2$ for $p\le 3/2$. Strict inequalities leave room for the auxiliary positive losses in that argument.

🏷️ The manuscript’s two-stage argument

The geometric stage studies positive masses moving on lines $A+tV$ in two transverse coordinates. It records velocity and position at linked resolutions and penalizes excessively eccentric observation windows. An entropy score compares the resolution gained with the information needed to specify the particle and observation time. Projection and incidence estimates are used to exclude excessive growth of that score.

The oscillatory stage replaces positive particles by wave packets. Squared packet amplitudes are positive, but their superposition is not. Close to the uncertainty limit, an earlier packet cannot be assigned a spatial position as finely as a later observation demands. The paper records this mismatch as additional information about velocity. Hypothetical persistent amplification would force that information to grow beyond the capacity of the available frequency window.

These arguments yield an oscillatory estimate with an arbitrarily small power loss. On a dyadic spatial shell of radius $\lambda$, the kernel amplitude contains the smoothing factor ${\lambda }^{-\delta }$ in addition to its geometric decay. The final shell estimate in the source has a residual norm loss of the schematic form ${\lambda }^{\epsilon -\delta }$. Choosing $\epsilon <\delta$ makes the dyadic sum converge. This is the concrete role of positive order in the argument.

🏷️ Summation and its limits

Dilation transfers a fixed-radius bound to every $R>0$. For finite $p$ in the claimed range, approximation by Schwartz functions then yields convergence of the radius-$R$ means to $f$ in $L^p$ as $R\to \infty$. Almost-everywhere convergence would require additional control, typically a maximal theorem; it is not a consequence of this norm convergence alone.

🏷️ Key steps and proof outlines

Entropy comparisons must use the actual joint law

The positive transport model observes particles with position $A+tV$. A finite label may encode a velocity bin, a position bin, or a time bin. Conditioning on a selected class changes the law of all three. The chain rule for mutual information is the basic accounting identity: the information carried by two successive labels is the first cost plus the conditional second cost.

The source also uses a quantitative comparison with a product law. If $P$ is the actual joint law and $Q$ a comparison law, write $r=dP/dQ$ where this is defined. Large likelihood ratios must be removed or charged before applying an incidence theorem formulated for a product distribution. Small mutual information alone does not say every conditional fiber is well behaved. The finite cutoff and exceptional-mass estimates supply a retained set on which a controlled comparison is valid. See “Finite entropy, conditioning, and passage to profiles.” (OpenAI, 2026)

Extremality converts geometry into rate inequalities

The geometric growth functional penalizes narrow apertures and the information needed to specify the observation time. A hypothetical failure gives nearly extremal positive transport experiments. Splitting them into angular and spatial classes produces profiles of growth and width. At a differentiability point, local improvements become inequalities between profile derivatives.

The argument first excludes isotropic extremizers by counting incidences in slope bins. A line with many retained time children supplies many distinct spatial cells; concentration of their intercepts then imposes a strong constraint on slopes. The remaining anisotropic case is organized around a matched characteristic plane. Trace theorems justify restricting the profile to that plane, rather than assuming that ambient almost-everywhere derivatives automatically have the desired traces.

The rate-transfer section constructs a finite affine-block experiment at matched clocks. Projection estimates apply to that finite experiment, and the resulting inequalities pass to the selected tangent rates. The final sign calculation rules out the proposed positive growth. The important order is finite conditioning and projection first, limiting derivative calculation second.

Packet composition encounters the uncertainty boundary

For oscillatory packets, a spatial window of width $w$ and a frequency window of width $\theta$ cannot both be arbitrarily fine: their product is constrained by the wavelength. Away from that boundary, frame analysis and synthesis reproduce the positive transport tests with controllable tails. Near it, a geometric descendant may ask for a spatial label that the incoming packet does not resolve.

The source’s composition inequality includes this mismatch as additional information about incoming fine velocity. Its repayment argument shows that repeated amplification would require more recoverable fine-velocity information than the original frequency support can contain. Thus the missing spatial resolution is not discarded as an error at every iteration; it is accumulated in a quantity with a finite total budget.

A representative conditioning estimate explains how the proof preserves usable fibers. If an event $A$ has probability $p_A>0$, then

$$P(P(A\mid X)<p_A\eta \mid A)\le \eta .$$

Indeed, integrate $P(A\mid X)$ over the set where it is below $p_A\eta$ and divide by $p_A$. Outside that exceptional set, conditioning on $A$ inflates a conditional probability by at most $(p_A\eta {)}^{-1}$. This elementary lemma lets packet subclasses inherit nonconcentration with an explicitly charged loss.

The radial kernel supplies the summable margin

After the packet theorem, decompose the radial convolution kernel into shells $|x|\asymp 2^j$. Its stationary-phase expansion separates oscillations $e^{\pm 2\pi i|x|}$ from amplitudes and lower-order remainders. On each shell, rescale to a fixed region; the distance phase is then checked against the packet theorem’s phase hypotheses. The amplitude derivatives and rapidly decaying remainders are bounded separately.

The decisive operator bound has a factor $2^{j(\epsilon -\delta )}$. Hence, for $0<\epsilon <\delta$,

$$\sum _{j\ge 0}{2}^{j(\epsilon -\delta )}=\frac{1}{1-2^{\epsilon -\delta }}<\infty .$$

Minkowski’s inequality sums the shell operators on $L^3$. The bounded spatial core is harmless because its kernel is integrable. This explicitly locates the use of positive smoothing order: it dominates the small loss from the packet theorem. No version of this sum is available at $\delta =0$.

Duality uses the real radial multiplier to transfer $L^3$ boundedness to $L^{3/2}$. Analytic interpolation with Plancherel and integrable-kernel orders yields the stated open polygon of exponents. These final operations are standard once the critical shell estimate is established; the new source claims lie in the positive transport and packet-growth arguments.

See Also

• Family 077: Fourier extension studies the surface oscillations associated with the boundary of the frequency ball.
• Family 079: wave local smoothing gives a distinct route to multiplier and maximal-summation consequences, with its own hypotheses.

🏷️ Manuscripts and verification scope

• Bochner–Riesz multipliers in three dimensions — the pinned manuscript; its directory supplies the source and citation metadata.

No family-specific scope page was available at lean/docs/078.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. Bochner–Riesz multipliers in three dimensions.