Fourier restriction theory

Fourier restriction theory studies estimates for Fourier transforms restricted to curved submanifolds and for the corresponding extension operators.

The discrete paraboloid is the graph of the sum-of-squares quadratic form in a vector space over a finite field. Its surface measure is normalized counting measure, of total mass one. A Fourier extension operator takes a function on this graph to its character sum in the ambient finite vector space. Ambient counting measure and normalized surface measure must be distinguished when defining a restriction constant.

If $i^2=-1$ in the prime field, the line $\ell =\{(t,it,0):t\in {\mathbb{F}}_p\}$ lies in the three-dimensional discrete paraboloid. Its indicator has normalized surface $L^2$ norm $p^{-1/2}$. Its extension is $p^{-1}$ on the two-dimensional annihilator plane and zero elsewhere. Thus its ambient ${\ell }^q$ norm is $p^{2/q-1}$, proving the displayed lower bound, which diverges for $q<4$.

For the three-dimensional discrete paraboloid over an odd prime field, the surface transform away from zero has magnitude at most $p^{-1}$, by completing squares and evaluating Quadratic Gauss sums. The adjoint-composition convolution kernel is the transform of the surface measure. Subtract its unit mass at the origin. Convolution by the remainder has ${\ell }^1\to {\ell }^{\infty }$ norm at most $p^{-1}$ and ${\ell }^2\to {\ell }^2$ norm at most $p$. The Riesz-Thorin interpolation theorem gives ${\ell }^{4/3}\to {\ell }^4$ norm at most one; adding the identity kernel gives at most two. The adjoint-composition identity and duality therefore give the displayed extension constant for normalized surface measure and ambient counting measure. This argument works for either quadratic character of minus one.

Assume a finite diagonal Fourier extension estimate on the unit circle. For direction-separated rectangles of dimensions ${\delta }^{-2}\times {\delta }^{-1}$, choose disjoint frequency caps and translate their transforms to the rectangles by the Fourier modulation and translation identity. Randomize their signs. Disjointness gives an input $p$th-power norm at most $CM\delta$; the Khintchine inequality converts the averaged output norm into the square function. The circle cap Fourier lower bound gives ${\delta }^p\int ({\sum }_R{1}_R{)}^{p/2}{\lesssim }_{p}M\delta$. Each rectangle has area ${\delta }^{-3}$, giving the displayed form. Input cap disjointness is compatible with arbitrary spatial rectangle overlap.

A Fourier extension operator forms an oscillatory integral from a density on a frequency surface with measure $\sigma$. The opposite sign in the phase merely reflects the output variable, so it gives identical Lp norms. In the usual dual formulation it is the adjoint of restricting the Fourier transform to that surface.

Let $\mu$ be a finite positive surface measure and $|f|\le 1$. Expand $|\widehat{f\mu }(x){|}^{2k}$ and integrate against $e^{-\pi |x{|}^2/R^2}$. The Fourier transform of a Gaussian turns the oscillatory kernel into $R^d{e}^{-\pi R^2|{\sum }_{j\le k}{\omega }_j-{\sum }_{j>k}{\omega }_j{|}^2}$, which is nonnegative. Taking absolute values of the bounded amplitudes gives an upper bound by the same moment with $f=1$. This weighted positivity is valid even when $f$ is merely measurable; it transfers a scalar surface-decay estimate to a local bounded-density moment estimate.

For normalized circle measure and a nonnegative cutoff near $(1,0)$, equal to one on $|\theta |\le \delta /C$ and supported on $|\theta |\le 2\delta /C$, remove the constant phase $e^{-i{\xi }_1}$. On $|{\xi }_1|\le {\delta }^{-2}$, $|{\xi }_2|\le {\delta }^{-1}$, the remaining phase has magnitude at most $2/C^2+2/C$. Taking $C$ sufficiently large keeps its real part positive and comparable to one, so the Fourier transform has magnitude at least a constant times the cap mass, which is at least $\delta /(\pi C)$. No upper bound on the cutoff is required for this lower bound.

An extension estimate is a norm bound for the Fourier extension operator, uniform over its input. The exponents encode the curvature and scale of the frequency surface. For normalized circle measure, a cap of angular width $\delta$ produces size comparable to $\delta$ on a rectangle of dimensions ${\delta }^{-2}\times {\delta }^{-1}$. This single-cap example forces $p\ge 4$ for a finite diagonal estimate with the same exponent $p$ on both sides.

On the unit circle, the arclength-measure transform satisfies $|\hat{\sigma }(x)|\lesssim (1+|x|{)}^{-1/2}$. This follows by removing angular neighborhoods of width $|x{|}^{-1/2}$ around the two stationary points and integrating by parts on the remainder. Gaussian positivity for even extension moments bounds the weighted fourth moment for any bounded $f$ by that for $f=1$. Radial integration of $(1+r{)}^{-2}{e}^{-\pi r^2/R^2}$ with area element $rdr$ costs only $\log R$, proving the displayed local estimate. Modulated disjoint caps and random signs convert it into the planar Kakeya neighborhood lower bound.

A decoupling inequality controls an $L^p$ norm of a function with Fourier support near a curved set by an ${\ell }^2$ sum of norms of pieces supported near smaller caps.

For Fourier supports in $R^{-1/2}\times R^{-1}$ caps along the parabola,

Functions with pairwise disjoint, or uniformly finitely overlapping, Fourier supports are almost orthogonal in $L^2$ by the Plancherel theorem.

A function whose Fourier support has widths $N$ in spatial-frequency directions and $N^2$ in a time-frequency direction varies only on the dual scales $N^{-1}$ and $N^{-2}$. Convolution with an adapted Schwartz kernel makes this quantitative.

Constructive interference occurs when many oscillatory phases nearly agree, making their exponential terms add with comparable arguments and producing a large value of the sum.

Kakeya inequalities bound overlaps of long thin tubes with controlled directions.

These tube estimates are used to study Kakeya sets and their dimension.

For a direction $e$, the Kakeya maximal function is the supremum over translations of normalized averages of $|f|$ over unit-length Kakeya tubes in direction $e$. It measures how large a function can be on at least one thin tube in each direction. Normalizing by tube volume is essential when comparing different thicknesses.

The maximal estimate used in the classical Kakeya set problem asks, for every $\epsilon >0$, for a bound from $L^n({\mathbb{R}}^n)$ to $L^n(S^{n-1})$ with loss at most ${\delta }^{-\epsilon }$. Constants may depend on dimension and $\epsilon$, but not tube thickness or the function. Applied to an indicator function of a neighborhood of a Kakeya set, it gives the Kakeya Minkowski dimension conjecture.

For a separated angular net, two translated unit rectangles of width $\delta$ intersect in area at most a constant times ${\delta }^2/(\delta +\alpha )$, where $\alpha$ is their unoriented angular distance. Each row of their intersection matrix therefore has sum at most $C\delta \log (2/\delta )$. Applying the Schur test to the adjoint averaging operator gives the displayed bound. A wider-tube comparison extends it from the net to all directions. Since $\sqrt{\log (2/\delta )}{\lesssim }_{\epsilon }{\delta }^{-\epsilon }$, this proves the planar version of the Kakeya maximal conjecture.

Choose $M\asymp {\delta }^{-1}$ angularly separated unit segments in a bounded Kakeya set, and let $F$ sum the indicators of their Kakeya tubes. Then $\int F\asymp 1$. Two tubes at angular separation $\alpha$ intersect in area at most $C{\delta }^2/(\delta +\alpha )$, so summing each harmonic angular row gives $\int F^2\lesssim \log (2/\delta )$. The Cauchy-Schwarz inequality on their union proves the displayed neighborhood lower bound. Since $|E_{\delta }|\lesssim {\delta }^2{N}_{\delta }(E)$, both planar Minkowski dimensions equal two.

For three transverse families of $R^{1/2}\times R^{1/2}\times R$ tubes in ${\mathbb{R}}^3$,

The degree-$d$ moment curve is $\gamma (t)=(t,t^2,\dots ,t^d)$. Tangent or radial directions from separated parameter intervals are quantitatively transverse by the Vandermonde determinant.

The cubic moment curve is ${\mathcal{M}}^3=\{(t,t^2,t^3):0\le t\le 1\}\subset {\mathbb{R}}^3$.