Local Smoothing Conjecture
Recommended reading: 【Modern Mathematics】 Table of Contents for Modern Mathematics, 【Calculus】 Chapter 14. Differential Geometry, 【Calculus】 Chapter 21. Fourier Series
1. Overview
2. Proof
3. Applications
1. Overview
⑴ Problem: Given an initial wave $f(x)$, how many derivatives of the initial data are required to control the spacetime size of the subsequent wave $w(T,x)$? The goal is to find the smallest possible derivative order $s$.
⑵ Waves arriving from different directions can form a large peak when their phases align. However, measuring not only the height of the peak but also the space it occupies and how long it lasts may yield a better upper bound than examining concentration at a single instant.
⑶ Local smoothing conjecture: The conjecture that integrating the size of a wave over a finite time interval reduces the amount of differentiation required of the initial data compared with controlling the wave at a single instant. Here, “local” means observing a finite time interval.
⑷ For example, in three dimensions with $p=3$, controlling the size at a fixed instant requires $1/3$ of a derivative of the initial data. In contrast, the conjecture states that including integration over time requires only an arbitrarily small positive derivative order $\varepsilon>0$.
⑸ On October 26, 2026, OpenAI announced a proof of the local smoothing conjecture for the Euclidean wave equation in every spatial dimension $n\ge2$ (ref, ref).
⑹ The proof strategy is to construct an auxiliary quantity that detects wave concentration: curvature and determinants make this quantity detect the desired power $\left|w\right|^{p_c}$, while conservation laws and comparison principles relate its size to the initial data.
2. Proof
⑴ Step 1. Fourier Series
① The original wave is represented as a continuous, infinite superposition.
② Collecting the frequency coefficients expresses their evolution over time as a diagonal matrix.
③ A question about wave concentration:
○ Since the multipliers above have absolute value $1$, the $L^2$ norm is conserved.
○ Nevertheless, a wave can concentrate into a narrower region and form a higher peak while maintaining the same $L^2$ norm.
○ For $p>2$, the $L^p$ norm assigns greater weight to high values, so it can increase when the phases of multiple components align and the wave concentrates.
○ How long can such strong wave concentration persist?
⑵ Step 2. Controlling Concentration at the Critical Exponent: First control a wave in a single frequency band at the most important exponent, $p=p_c=\frac{2n}{n-1}>2$.
① First restrict to a single frequency band on a periodic space, so that the wave becomes a finite sum.
② Use an imaginary translation to analyze waves traveling in different directions: $(T,x)\mapsto(T-i\rho,x-i\rho\omega)$, where $\left|\omega\right|=1$.
③ Substitution into a single plane wave has the effect of applying a Gaussian filter: $\exp\left(-\frac{\rho\left|\xi\right|}{2}\left|\omega-\frac{\xi}{\left|\xi\right|}\right|^2\right)$. Directions close to $\omega$ receive greater weight → used in ⑪.
④ Combine the filtered waves with direction vectors to construct a vector $\chi$: $\chi=\sum_j(\text{direction vector})(\text{wave amplitude})(\text{Gaussian weight})$.
⑤ Since the wave equation can be written as a symmetric first-order system, it has a conserved quantity (∵ Noether’s theorem).
○ This conserved quantity turns out to satisfy $\chi\chi^*=0,$ where * denotes complex conjugation.
○ If $A$ is the normalized and corrected matrix obtained from $\chi\chi^*$, then $\operatorname{div}A=0$ still holds.
⑥ Define the following background function using a holomorphic function $\Phi$ whose value remains constant along the current.
○ $B_0=\lambda\left|\Phi\right|^2$, where $\lambda>0$.
○ The spatial quaternionic Hessian of this function has rank $1$ and can be expressed as follows in a suitable basis.
○ The Hessian is connected to curvature: the following formula holds generally.
○ In other words, the spatial quaternionic Hessian has a structure in which the background provides curvature in only one principal direction, while the curvature in the remaining principal directions is zero.
○ For $0<\alpha\le\lambda$, $\det\left(H_{\mathrm{fib}}(B_0)+\alpha I\right)=(\beta+\alpha)\alpha^{n-1}\asymp\lambda\alpha^{n-1}$, where $\alpha$ represents curvature.
○ Here, $\alpha I$ is the Hessian of a comparison paraboloid used to estimate this depth: the Hessian of the actual solution is not simply prescribed to be $B+\alpha I$.
⑦ Interpret the density $\mu$, which represents wave concentration, in terms of how deeply an auxiliary function $v$ descends.
○ $\mu$: a density representing wave concentration.
○ Set $\mu=A_{TT}>0$ and $B_0=\lambda\left|\Phi\right|^2$, and define $v\le0$ by solving the following:
⑧ Consider a ball-shaped region of radius $r$.
○ If the wave density in the ball is approximately $m$, then $\lambda\alpha^{n-1}\sim\lambda m\iff\alpha\sim m^{1/(n-1)}$, so curvature of this order is required.
○ In a ball of radius $r$, the depth is approximately $r^2\alpha$. Thus, multiplying depth by density gives the scale detected by $(-v)\mu$ as $r^2m^{1+1/(n-1)}=r^2m^{n/(n-1)}$.
○ Combining this with the fact that the density detects the square of the wave gives $\left(\left|w\right|^2\right)^{n/(n-1)}=\left|w\right|^{2n/(n-1)}=\left|w\right|^{p_c}$.
○ Summary: density $m$ → required curvature $m^{1/(n-1)}$ → depth × density $\sim m^{n/(n-1)}$ → $\left|w\right|^{p_c}$.
○ Orthogonality: Even after introducing this curvature, $A:H(B_0)=0$. Thus, the background curvature contributes to the determinant but disappears without producing an additional term when paired with the current (∵ geometrically, the current moves along level sets, while the background curvature lies in a direction invisible to that current).
⑨ Complete the concentration measure by integrating a potential associated with the determinant.
○ $E=\int(-v)\mu$ detects the desired concentration because it multiplies the wave density by the depth of the auxiliary function.
○ The spatial Hessian is positive definite, but the full Hessian, including time, is not guaranteed to be positive definite.
○ Use the Schur complement to collect the sign-indefinite part into a single scalar $\sigma$.
○ Using the auxiliary equation $\det D=\lambda\mu$ gives $A:H(v)=A:\Sigma+\lambda^{-1}\det M$, where $A:\Sigma\ge0$, while $\lambda^{-1}\det M$ has no definite sign.
○ Integrate the spatial determinant from the background $B_0$ to $B_0+v$ to construct a correction whose differentiation exactly cancels the sign-indefinite term $\det M$.
○ This integration constructs a potential associated with the determinant. The differentiation formula for the determinant and the conservation property of the cofactor establish, through an exact identity, that differentiating the correction produces the required cancellation term.
○ Also prove that the correction does not erase the original concentration information: $\frac{n-1}{n+1}E_0\le\mathcal F\le\frac{n}{n+1}E_0$.
○ Thus, $\mathcal F$ detects concentration at a scale comparable to $E_0$, while its differentiated expression consists of a term with a favorable sign and a boundary term.
⑪ Divide all directions into small ranges in advance and apply the same estimate to each range.
○ First cover the entire sphere of directions with finitely many small ranges of width $\gamma$, and decompose the wave accordingly.
○ Fix this partition before considering the initial data or the frequency scale $N$.
○ Place the current-carrying part of the boundary away from the principal wave directions so that the Gaussian filter makes the boundary current exponentially small, of order $e^{-cs/h}$ (∵ ③).
○ Recovering the wave in the interior costs approximately $e^{C\gamma^2s/h}$. Fixing $\gamma$ sufficiently small leaves decay as follows.
○ This applies to every wave belonging to that directional range.
○ After estimating each component, recombine them into the original wave: even if interference between the components increases the size, it can be controlled as follows when the number of components is $L$.
○ The smooth frequency partition of the initial data also controls its $L^p$ size up to a constant factor.
○ Since the directional width $\gamma$ is fixed, the number of components $L$ is independent of $N$.
○ Consequently, the cost of recombination is absorbed into a constant without increasing the frequency loss, yielding an estimate for the original wave containing all directions.
⑫ As a result, the complicated calculations involving matrices and the auxiliary equation reduce to the following scalar inequality.
○ Since the auxiliary variable $\tau$ is two-dimensional, a harmonic function depending only on the radius $s$ in this plane satisfies $\psi’‘(s)+\frac1s\psi’(s)=0$.
○ Solving this differential equation gives $\psi(s)=A\log s+B$.
⑭ More specifically, compare the concentration observed on a small auxiliary circle at the wavelength scale $a$ with that at radius approximately $1$, connect it to the initial time there, and then return to a small auxiliary circle at the initial time → each of the two annulus comparisons incurs a logarithmic cost.
⑶ Step 3. Combining All Directions and Extending to Euclidean Space
① Cover all directions with small directional ranges and sum the estimates for the components: because the width of each directional range is fixed, the number of components is independent of the frequency scale $N$.
② Divide Euclidean space into small observation regions, truncate the initial data around each region, and periodically replicate the truncated data.
③ Apply the existing estimate on periodic space, and control the effects of distant copies and the discarded data using the rapid decay of the kernel.
④ Sum the estimates over the small observation regions to obtain the logarithmic estimate for a single frequency band in $\mathbb R^n$.
⑷ Step 4. Extending to General $p$ and All Frequencies
① Interpolating the estimate at $p_c$ with two known estimates gives the required frequency exponent for general $p$ as
○ $p=2$: conservation of the $L^2$ norm.
○ $p=\infty$: an operator bound of order $N^{(n-1)/2}$ on a single frequency band.
② Decompose the initial data into frequency bands as $f=\sum_{j\ge0}f_{2^j}$ and sum the estimates for the bands.
○ The question is whether the infinite sum over frequency bands converges in the presence of logarithmic losses.
○ For fixed $A>0$ and every $\varepsilon>0$, $[\log(2+N)]^A\le C_{\varepsilon,A}N^\varepsilon$. Since a fixed power of a logarithm grows more slowly than any positive power, however small, this resolves the convergence issue.
③ Finally, for every $\varepsilon>0$:
○ The quantifier $\forall\varepsilon>0$ means that the required amount of differentiation can be brought arbitrarily close to the conjectured threshold; it does not mean that $\varepsilon=0$ is allowed.
3. Applications
⑴ The Kakeya Problem
① Decomposing a wave into packets localized in frequency and space concentrates each packet in a thin tube in physical space. Thus, understanding the concentration of many waves (= the local smoothing problem) is connected to studying how much tubes pointing in different directions can overlap (= the Kakeya problem).
② In joint work, 2026 Fields Medalist Hong Wang and Joshua Zahl proved the three-dimensional Kakeya set conjecture in 2025: any set in three-dimensional space containing a unit line segment in every direction must have both Hausdorff and Minkowski dimension $3$.
Posted: 2026.11.12 23:51