betaFeedback

WorksMathematics

The Lorentz Threshold in Fourier Restriction

Abstract

Classical Stein–Tomas restriction controls a Fourier transform on a curved surface, but the estimate changes when its values lie in a Banach space. We determine the sharp Lorentz threshold at the critical exponent 4/34/3: for coefficients in ℓD4/3,q\ell_D^{4/3,q}, the optimal Gaussian restriction constant grows like (1+log⁡D)∣3/4−1/q∣(1+\log D)^{|3/4-1/q|}. A multiscale wave-packet construction proves sharpness, including the delicate index q=2q=2. The same threshold holds for fractional Gaussian noise, while Gaussian type and cotype alone do not distinguish the positive and negative cases. As a dispersive consequence, we establish a sharp Lorentz-channel spacetime estimate for linear capillary waves.

2020 Mathematics Subject Classification. Primary 42B10, 46B20; Secondary 46E30, 60G15, 35Q35, 76B15.

Introduction

Why restricting a Fourier transform is difficult

The Fourier transform describes how a function is assembled from different frequencies. Restricting it to a surface sounds like ordinary evaluation, but there is a basic difficulty: a sphere has zero three-dimensional volume, and the Fourier transform of a general LpL^p function is not defined pointwise on that sphere. Fourier restriction theory asks when the shape of the surface makes such a trace meaningful.

A fundamental answer is the Stein–Tomas theorem. For a compact curved hypersurface Σ⊂R3\Sigma\subset\mathbb{R}^3, it gives

∥f^∣Σ∥L2(Σ)≤CΣ∥f∥L4/3(R3).(1)\|\widehat{f}|_\Sigma\|_{L^2(\Sigma)} \le C_\Sigma\|f\|_{L^{4/3}(\mathbb{R}^3)}. \tag*{(1)}

Curvature is essential: oscillations from different points of the surface partly cancel, allowing a Fourier trace that would not follow from a simple pointwise estimate. The restriction exponent 4/34/3 is critical. A narrow frequency cap corresponds to a long, thin physical wave packet. For the packet gRg_R defined in (12) below,

∥gR∥Ls(R3)≍R4/s−3,∥gR^∣Σ∥2≳1.\|g_R\|_{L^s(\mathbb{R}^3)} \asymp R^{4/s-3}, \qquad\|\widehat{g_R}|_\Sigma\|_2 \gtrsim1.

Thus s>4/3s>4/3 cannot replace 4/34/3 uniformly. This elementary scaling already explains why the endpoint is particularly delicate.

The estimate emerged from the classical work of Stein and Tomas [12, 10]; its importance extends to Fourier extension, oscillatory integrals and dispersive-wave estimates. These connections and more recent endpoint developments are reviewed in [11]. Bak and Seeger [1] sharpened the scalar estimate at the same critical exponent by replacing its physical source with a Lorentz space, L4/3,2(R3)L^{4/3,2}(\mathbb{R}^{3}). Our question differs: the physical exponent remains 4/34/3, while the values of the functions belong to a Lorentz space.

Many coefficient channels and Gaussian type

Imagine that f(x)f(x) records not one scalar measurement but a vector of many channels. Applying (1) to each channel separately does not answer whether the joint estimate stays stable as the number of channels grows. The answer depends on the Banach space XX in which the values live. The natural joint output is a γ\gamma-radonifying operator. For a finite-rank operator T:H→XT:H\to X and an orthonormal system (hν)(h_\nu) spanning its Hilbert input, its norm is

∥T∥γ(H,X)=(E∥∑νγνThν∥X2)1/2,(2)\|T\|_{\gamma(H,X)}=\left(\mathbb{E}\left\|\sum_{\nu}\gamma_{\nu}Th_{\nu}\right\|_{X}^{2}\right)^{1/2}, \tag*{(2)}

where the γν\gamma_\nu are independent standard Gaussians. Informally, (2) measures the size of all channels after Gaussian mixing. When XX is Hilbert, it reduces to the usual square-sum of channel energies; for a general Banach space it retains finer geometric information.

Fix a compact C2C^{2} surface Σ\Sigma with nonvanishing Gaussian curvature and a smooth positive density dσd\sigma. We study

RΣ:Lp(R3;X)⟶γ(L2(Σ),X),p=43,(3)R_{\Sigma}:L^{p}(\mathbb{R}^{3};X)\longrightarrow\gamma(L^{2}(\Sigma),X),\qquad p=\frac{4}{3}, \tag*{(3)}

and denote its norm by Sp(X;Σ)S_{p}(X;\Sigma). For finite-rank Schwartz functions, RΣfR_{\Sigma}f is simply the Fourier trace f^∣Σ\widehat{f}|_{\Sigma} interpreted as an operator on L2(Σ)L^{2}(\Sigma).

A natural first prediction comes from Gaussian type. A Banach space has Gaussian type pp if

Tp(X):=sup⁡(xj)≠0(E∥∑jγjxj∥X2)1/2(∑j∥xj∥Xp)1/p<∞.(4)T_{p}(X):=\sup_{(x_j)\ne0}\frac{\left(\mathbb{E}\left\|\sum_j\gamma_jx_j\right\|_{X}^{2}\right)^{1/2}}{\left(\sum_j\|x_j\|_{X}^{p}\right)^{1/p}}<\infty. \tag*{(4)}

Type pp says that independent Gaussian contributions cannot accumulate faster than their ℓp\ell^{p} input sizes. The endpoint restriction estimate forces type pp, but its converse is much less clear: the Fourier traces of many physical packets are correlated, not independent. Work on Banach-valued Gaussian embeddings [7, 8] explains why type is a natural candidate, while leaving the curved endpoint as an additional problem. Other recent vector-valued restriction results can be found in [2].

The sharp Lorentz threshold

Lorentz spaces provide an unusually precise test of this question. Let 1≤q<∞1\le q<\infty, D≥2D\ge2, and put EDp,q=ℓDp,qE_{D}^{p,q}=\ell_{D}^{p,q}, using a Banach lattice norm equivalent to

∥z∥p,q∗=(∑r=1Drq/p−1(z∗(r))q)1/q,(5)\|z\|_{p,q}^{*}=\left(\sum_{r=1}^{D}r^{q/p-1}\bigl(z^{*}(r)\bigr)^{q}\right)^{1/q}, \tag*{(5)}

where z∗z^* denotes the decreasing rearrangement of the absolute values. The parameter pp controls the main integrability scale, while qq measures how values from different levels combine. For q>pq>p the displayed expression may be a quasi-norm; the equivalent Banach norm is used throughout, with constants independent of DD.

Our first theorem gives an exact quantitative answer.

Theorem 1.1 (Sharp Lorentz restriction threshold). Let p=4/3p=4/3, 1≤q<∞1\le q<\infty, and Σ\Sigma be as above. There exist constants 0<cΣ,q≤CΣ,q<∞0<c_{\Sigma,q}\le C_{\Sigma,q}<\infty such that for every D≥2D\ge2,

cΣ,q(1+log⁡D)∣1/p−1/q∣≤Sp(EDp,q;Σ)≤CΣ,q(1+log⁡D)∣1/p−1/q∣.(6)c_{\Sigma,q}(1+\log D)^{\lvert1/p-1/q\rvert}\le S_p(E_D^{p,q};\Sigma)\le C_{\Sigma,q}(1+\log D)^{\lvert1/p-1/q\rvert}. \tag*{(6)}

In particular, for X=Lp,q((0,1)3)X=L^{p,q}((0,1)^3),

Sp(X;Σ)<∞⟺q=p.(7)S_p(X;\Sigma)<\infty\quad\Longleftrightarrow\quad q=p. \tag*{(7)}

For p<q<∞p<q<\infty, the space XX still has Gaussian type pp.

The logarithm in (6) is important. It gives not only a yes-or-no answer for an infinite coefficient space but the exact rate at which a finite collection of channels becomes unstable. At q=2q=2 this rate is

S4/3(ℓD4/3,2;Σ)≍(1+log⁡D)1/4.S_{4/3}(\ell_D^{4/3,2};\Sigma)\asymp(1+\log D)^{1/4}.

This is consistent with the positive scalar physical-domain L4/3,2→L2(Σ)L^{4/3,2}\to L^2(\Sigma) restriction estimate: the two Lorentz spaces play different roles. Theorem 1.1 proves that attaining Gaussian type 4/34/3 does not by itself protect the critical Fourier trace.

The fractional Gaussian comparison

The same question can be asked for a sequence of correlated Gaussian increments. Let H=3/4H=3/4, and define the covariance of fractional Gaussian noise by

rH(m)=12(∣m+1∣2H+∣m−1∣2H−2∣m∣2H),FN(j,k)=rH(j−k).r_H(m)=\frac{1}{2}\left(\lvert m+1\rvert^{2H}+\lvert m-1\rvert^{2H}-2\lvert m\rvert^{2H}\right),\qquad F_N(j,k)=r_H(j-k).

For deterministic coefficient vectors xj∈Xx_j\in X and a centered Gaussian vector (Zj)(Z_j) with covariance FNF_N, set

Fp(X)=sup⁡N≥1sup⁡(xj)≠0(E∥∑j=1NZjxj∥X2)1/2(∑j=1N∥xj∥Xp)1/p.(8)F_p(X)=\sup_{N\ge1}\sup_{(x_j)\ne0}\frac{\left(\mathbb{E}\left\lVert\sum_{j=1}^{N}Z_jx_j\right\rVert_X^2\right)^{1/2}}{\left(\sum_{j=1}^{N}\lVert x_j\rVert_X^p\right)^{1/p}}. \tag*{(8)}

Unlike independent noise, these increments retain correlations over long time intervals. The same Lorentz transition appears.

Theorem 1.2 (Matching fractional Gaussian threshold). For p=4/3p=4/3 and 1≤q<∞1\le q<\infty,

Fp(EDp,q)≍q(1+log⁡D)∣1/p−1/q∣,D≥2.(9)F_p(E_D^{p,q})\asymp_q(1+\log D)^{\lvert1/p-1/q\rvert},\qquad D\ge2. \tag*{(9)}

Hence Fp(Lp,q((0,1)3))<∞F_p(L^{p,q}((0,1)^3))<\infty if and only if q=pq=p.

The matching rates reveal a common difficulty in collecting many correlated channels. They do not prove that spherical restriction and the fractional Gaussian model are equivalent for arbitrary Banach spaces.

Results, a dispersive consequence and proof guide

Theorem 1.1 determines the precise cost of combining many spectral channels at the critical Fourier exponent. Its lower bound uses packets at successively larger parabolic scales: although the input remains inexpensive in a Lorentz norm, each channel retains a nontrivial Fourier trace. Theorem 1.2 identifies the corresponding threshold for correlated Gaussian increments, by a separate interval construction. A further example shows that optimal Gaussian type and cotype need not determine the restriction estimate.

There is also a direct dispersive PDE consequence. The linear capillary-wave propagator has frequency law ω(ξ)=c∣ξ∣3/2\omega(\xi)=c|\xi|^{3/2}, whose graph is curved away from zero frequency. Theorem 6.1 gives the sharp dimension dependence of a spacetime L4L^4 estimate for arrays of these waves, measured in Lorentz sequence norms. This application concerns the linearized capillary-wave equation and requires no matrix-operator model; the water-wave dispersion setting is discussed in [3].

The proof begins with finite-dimensional Lorentz inequalities and the necessary Gaussian-type condition. It then treats the two sides of the critical Lorentz index separately: multiscale Fourier packets for q>4/3q>4/3 and Gaussian type for q<4/3q<4/3. The fractional covariance argument follows independently. Finally, we prove the capillary-wave consequence and record an interpolation stability principle for the supercritical Lorentz range.

Elementary tools: Lorentz norms, Gaussian type and scalarization

Lemma 2.1 (Finite Lorentz distortion). For fixed 1≤q<∞1\leq q<\infty and D≥2D\geq2, the identity between EDp,qE_D^{p,q} and ℓDp\ell_D^p has Banach–Mazur distortion at most

Cp,q(1+log⁡D)∣1/p−1/q∣.C_{p,q}(1+\log D)^{|1/p-1/q|}.

Moreover, ∥1D∥EDp,q≍p,qD1/p\|1_D\|_{E_D^{p,q}}\asymp_{p,q}D^{1/p}, uniformly in DD.

Proof. Let z∗z^* be nonincreasing, and partition the ranks into the dyadic intervals Ij=[2j,2j+1)∩{1,…,D}I_j=[2^j,2^{j+1})\cap\{1,\ldots,D\}. There are J+1≍1+log⁡DJ+1\asymp1+\log D intervals. Because z∗z^* is monotone, the contribution of a full block to ∑r(z∗(r))p\sum_r(z^*(r))^p is comparable to 2j(z∗(2j))p2^j(z^*(2^j))^p. Similarly, the block contribution to the Lorentz quasi-norm is comparable to 2jq/p(z∗(2j))q2^{jq/p}(z^*(2^j))^q. Using a one-sided adjacent-block shift at the final short block makes these comparisons uniform. With

aj=2j/pz∗(2j),a_j=2^{j/p}z^*(2^j),

this gives

∥z∥ℓp≍p(∑j=0Jajp)1/p,∥z∥ℓp,q∗≍p,q(∑j=0Jajq)1/q.\|z\|_{\ell^p}\asymp_p\left(\sum_{j=0}^{J}a_j^p\right)^{1/p},\qquad \|z\|_{\ell^{p,q}}^*\asymp_{p,q}\left(\sum_{j=0}^{J}a_j^q\right)^{1/q}.

There is now no Fourier analysis in the calculation: it is the ordinary comparison between ℓp\ell^p and ℓq\ell^q on a sequence of length J+1J+1. Its optimal loss is at most (J+1)∣1/p−1/q∣(J+1)^{|1/p-1/q|}. Replacing the Lorentz quasi-norm by a uniformly equivalent lattice norm changes only a constant that depends on p,qp,q, not on DD. Finally,

\|1_D\|_{\ell^{p,q}^*}^q=\sum_{r=1}^{D}r^{q/p-1}\asymp_{p,q}D^{q/p},

so ∥1D∥EDp,q≲D1/p\|1_D\|_{E_D^{p,q}}\lesssim D^{1/p}.

Lemma 2.2 (Scalarization and the necessary type bound). For p=4/3p=4/3,

Sp(ℓDp;Σ)≤CΣ,Tp(X)≤CΣSp(X;Σ)S_p(\ell_D^p;\Sigma)\le C_\Sigma,\qquad T_p(X)\le C_\Sigma S_p(X;\Sigma)

for every Banach space XX, whenever the right side is finite. In particular,

Sp(EDp,q;Σ)≲Σ,q(1+log⁡D)∣1/p−1/q∣.(10)S_p(E_D^{p,q};\Sigma)\lesssim_{\Sigma,q}(1+\log D)^{\lvert1/p-1/q\rvert}. \tag*{(10)}

Proof. For X=ℓDpX=\ell_D^p, let f=(f1,…,fD)f=(f_1,\ldots,f_D). The Gaussian L2L^2 norm in the definition of γ\gamma, followed by Gaussian moment comparison and Fubini, gives

∥RΣf∥γ(L2(Σ),ℓDp)∼p(∑a=1D∥RΣfa∥2p)1/p.\lVert R_\Sigma f\rVert_{\gamma(L^2(\Sigma),\ell_D^p)} \sim_p \left(\sum_{a=1}^D\lVert R_\Sigma f_a\rVert_2^p\right)^{1/p}.

Apply scalar Stein–Tomas to every faf_a and note that ∥f∥Lp(R3;ℓDp)p=∑a∥fa∥pp\lVert f\rVert_{L^p(\mathbb{R}^3;\ell_D^p)}^p=\sum_a\lVert f_a\rVert_p^p. The resulting constant does not depend on DD.

For necessity, fix a compactly supported scalar bump φ\varphi whose Fourier transform does not vanish on a positive-measure part of Σ\Sigma. Given x1,…,xm∈Xx_1,\ldots,x_m\in X, choose translations aja_j successively so far apart that the physical supports of the translated bumps are disjoint and the Gram matrix of

hj(ω)=φ^(ω)e−iaj⋅ωh_j(\omega)=\widehat{\varphi}(\omega)e^{-ia_j\cdot\omega}

has a lower spectral bound independent of mm. To justify the latter, observe that each off-diagonal Gram entry is the Fourier transform of the smooth finite measure ∣φ^∣2dσ\lvert\widehat{\varphi}\rvert^2d\sigma evaluated at aj−aka_j-a_k. Those entries tend to zero as the distance grows. By making them small enough that the sum of absolute off-diagonal entries in every row is at most half the common diagonal entry, the Gram matrix is bounded below by a fixed multiple of the identity.

Set f(x)=∑j=1mφ(x−aj)xjf(x)=\sum_{j=1}^m\varphi(x-a_j)x_j. Disjoint physical supports give

∥f∥Lp(R3;X)=∥φ∥p(∑j∥xj∥Xp)1/p.\lVert f\rVert_{L^p(\mathbb{R}^3;X)} =\lVert\varphi\rVert_p\left(\sum_j\lVert x_j\rVert_X^p\right)^{1/p}.

The lower Gram bound means that the Gaussian covariance of the restricted family (hj)(h_j) dominates a fixed multiple of the identity. Gaussian covariance domination therefore yields

∥RΣf∥γ(L2(Σ),X)≳(E∥∑jγjxj∥X2)1/2.\lVert R_\Sigma f\rVert_{\gamma(L^2(\Sigma),X)} \gtrsim \left(\mathbb{E}\left\lVert\sum_j\gamma_jx_j\right\rVert_X^2\right)^{1/2}.

The assumed restriction estimate gives precisely the type-pp inequality. Finally, transfer the positive ℓDp\ell_D^p estimate to EDp,qE_D^{p,q} through Lemma 2.1, paying the product of the norms of the identity and its inverse.

Lemma 2.3 (Gaussian square-function lower bound). Let EE be a finite-dimensional Banach lattice with a lattice norm, and let G:Σ→EG:\Sigma\to E have finitely many scalar coordinates GaG_a. Then

∥G∥γ(L2(Σ),E)≥c∥(∥Ga∥L2(Σ))a∥E(11)\lVert G\rVert_{\gamma(L^2(\Sigma),E)} \ge c\left\lVert\left(\lVert G_a\rVert_{L^2(\Sigma)}\right)_a\right\rVert_E \tag*{(11)}

with a numerical c>0c>0. An equivalent lattice norm changes cc only by the equivalence constant.

Proof. The Gaussian radonification of GG has each coordinate a centered Gaussian of variance ∥Ga∥22\|G_a\|_2^2. Thus, using Jensen and the lattice order,

∥G∥γ≥E∥Z∥E≥∥(E∣Za∣)a∥E=2π∥(∥Ga∥2)a∥E.\|G\|_{\gamma} \geq\mathbb{E}\|Z\|_E \geq\|(\mathbb{E}|Z_a|)_a\|_E = \sqrt{\frac{2}{\pi}} \|(\|G_a\|_2)_a\|_E.

For complex Gaussians the analogous fixed absolute-moment constant is used.

Multiscale wave packets: the case q>pq > p

This section proves the main lower bound for p<q<∞p < q < \infty without using scalar Lorentz sharpness. The proof has three ingredients: (i) one packet at any parabolic scale has bounded LpL^p input and nontrivial restriction output; (ii) the same physical point meets only a controlled number of translated packets at each scale; and (iii) the restriction outputs retain a separate coordinate for every translated packet. The last two facts are what create the sharp logarithmic gap.

Fix orthonormal physical coordinates x=(x⊥,x3)x = (x_{\perp}, x_3) adapted to a patch of Σ\Sigma near ξ0\xi_0; locally Σ\Sigma is a graph ξ3=ξ0,3+O(∣ξ⊥−ξ0,⊥∣2)\xi_3 = \xi_{0,3} + O(|\xi_{\perp} - \xi_{0,\perp}|^2). Choose Ψ∈Cc∞(R3)\Psi\in C^\infty_c(\mathbb{R}^3) such that Ψ^(0)≠0\widehat{\Psi}(0) \ne0, and set for R≥1R \geq1

gR(x)=R−3eix⋅ξ0Ψ(x1R,x2R,x3R2).(12)g_R(x) = R^{-3}e^{ix\cdot\xi_0}\Psi\left(\frac{x_1}{R},\frac{x_2}{R},\frac{x_3}{R^2}\right). \tag*{(12)}

Its LpL^p norm is uniformly bounded, since the physical support volume is O(R4)O(R^4) and R−3pR4=1R^{-3p}R^4 = 1 at p=4/3p = 4/3.

Lemma 3.1 (Uniform scalar packet output). There is cΣ,Ψ>0c_{\Sigma,\Psi} > 0 with

∥RΣgR∥L2(Σ)≥cΣ,Ψ,R≥1.\|R_{\Sigma}g_R\|_{L^2(\Sigma)} \geq c_{\Sigma,\Psi}, \qquad R \geq1.

Proof. The change of variables x1=Ry1x_1 = R y_1, x2=Ry2x_2 = R y_2, x3=R2y3x_3 = R^2 y_3 has Jacobian R4R^4. Hence

gR^(ξ)=RΨ^(R(ξ⊥−ξ0,⊥),R2(ξ3−ξ0,3)).\widehat{g_R}(\xi) = R\widehat{\Psi}\bigl(R(\xi_{\perp} - \xi_{0,\perp}), R^2(\xi_3 - \xi_{0,3})\bigr).

Write the local surface graph as ξ3=ξ0,3+ϕ(ξ⊥−ξ0,⊥)\xi_3 = \xi_{0,3} + \phi(\xi_{\perp} - \xi_{0,\perp}), where ϕ(0)=0\phi(0) = 0, ∇ϕ(0)=0\nabla\phi(0) = 0 and ∣ϕ(y)∣≤C∣y∣2|\phi(y)| \leq C|y|^2 for small yy. Since Ψ^(0)≠0\widehat{\Psi}(0) \ne0, its absolute value is bounded below in a sufficiently small neighborhood of the origin. On the cap ∣y∣≤c0/R|y| \leq c_0/R, both RyRy and R2ϕ(y)R^2\phi(y) lie in this neighborhood. Therefore ∣gR^∣≳R|\widehat{g_R}| \gtrsim R throughout that cap. Its surface measure is comparable to R−2R^{-2}, so

∥RΣgR∥22≥∫∣y∣≤c0/R∣gR^∣2 dσ≳R2R−2≳1.\|R_{\Sigma}g_R\|_2^2 \geq\int_{|y|\leq c_0/R}|\widehat{g_R}|^2\,d\sigma\gtrsim R^2R^{-2} \gtrsim1.

The finitely bounded range of smaller RR is handled by reducing the constant after choosing the patch.

Take M=2hM = 2^h, h≥1h \geq1, and put

N=M4=24h,J=4h,Rj=2j/4(0≤j<J).N = M^4 = 2^{4h}, \qquad J = 4h, \qquad R_j = 2^{j/4}\quad(0 \leq j < J).

At level jj the packet occupies a physical box of volume Rj4=2jR_j^4 = 2^j and has height Rj−3=2−j/pR_j^{-3} = 2^{-j/p}. The choice N=2JN = 2^J allows us to fit all JJ levels in one common large box. This balance is the source of the logarithm: a level contributes about 2j2^j nonzero coordinates, but each is only about 2−j/p2^{-j/p} in magnitude. Choose the lattice of centers

AM={1,…,M}2×{1,…,M2},∣AM∣=N,\mathcal{A}_M = \{1,\ldots,M\}^2 \times\{1,\ldots,M^2\}, \qquad|\mathcal{A}_M| = N,

expressed in the adapted physical coordinates. The coefficient index set is I={0,…,J−1}×AM\mathcal{I} = \{0,\ldots,J-1\} \times\mathcal{A}_{M}, of cardinality D=NJD = NJ. Define G:R3→EDp,q\mathcal{G} : \mathbb{R}^{3} \to E_{D}^{p,q} by

G(x)(j,a)=gRj(x−a).(13)\mathcal{G}(x)_{(j,a)} = g_{R_j}(x-a). \tag*{(13)}

Lemma 3.2 (Pointwise weak-ℓp\ell^{p} envelope). The array (13) satisfies

sup⁡x∈R3∥G(x)∥ℓDp,∞≤CΨ,∣supp⁡G∣≤CΨN,\sup_{x \in\mathbb{R}^{3}} \|\mathcal{G}(x)\|_{\ell_{D}^{p,\infty}} \le C_{\Psi}, \qquad|\operatorname{supp}\mathcal{G}| \le C_{\Psi}N,

and hence, for each fixed finite qq,

∥G∥Lp(R3;EDp,q)≤CΨ,p,qN1/pJ1/q.(14)\|\mathcal{G}\|_{L^{p}(\mathbb{R}^{3};E_{D}^{p,q})} \le C_{\Psi,p,q}N^{1/p}J^{1/q}. \tag*{(14)}

Proof. At scale jj, every nonzero coordinate has absolute value at most CRj−3=C2−j/pCR_{j}^{-3} = C2^{-j/p}. Since Ψ\Psi is compactly supported, at a fixed xx there are at most C(Rj+1)2(Rj2+1)≤CRj4=C2jC(R_{j}+1)^{2}(R_{j}^{2}+1) \le CR_{j}^{4} = C2^{j} centers aa with gRj(x−a)≠0g_{R_{j}}(x-a) \ne0. Consequently, for t>0t > 0,

#{(j,a):∣G(x)(j,a)∣>t}≤C∑2−j/p≳t2j≤C′t−p.\#\{(j,a): |\mathcal{G}(x)_{(j,a)}| > t\} \le C \sum_{2^{-j/p} \gtrsim t} 2^{j} \le C't^{-p}.

The key point is that the count grows geometrically: the entire sum through the last active level is comparable to its final term. Thus the counting bound is Ct−pCt^{-p}, independent of the number of scales. This is the uniform weak-ℓp\ell^{p} estimate. Also Rj≤MR_{j} \le M and Rj2≤M2R_{j}^{2} \le M^{2}, so all the translated supports lie in a fixed enlargement of the box M×M×M2M \times M \times M^{2}, whose volume is O(N)O(N).

To pass from weak-ℓp\ell^{p} to the stronger ℓp,q\ell^{p,q} norm, order the nonzero coordinates in decreasing magnitude. The weak bound says that the rr-th largest coordinate is at most Cr−1/pCr^{-1/p}. Substitution into (5) gives

∥z∥ℓp,qq≲∑r=1Drq/p−1(Cr−1/p)q≲1+log⁡D.\|z\|_{\ell^{p,q}}^{q} \lesssim\sum_{r=1}^{D} r^{q/p-1}(Cr^{-1/p})^{q} \lesssim1+\log D.

Since D=NJD = NJ and N=2JN = 2^{J}, 1+log⁡D≳J1+\log D \gtrsim J. Integration over the union of the supports proves (14). □\square

Proof of the lower bound in Theorem 1.1 for q>pq>p. Every scalar coordinate of RΣGR_{\Sigma}\mathcal{G} is a modulation of g^Rj∣Σ\widehat{g}_{R_{j}|_{\Sigma}} by e−ia⋅ξe^{-ia\cdot\xi}, hence has L2(Σ)L^{2}(\Sigma) norm at least cΣ,Ψc_{\Sigma,\Psi}. Lemma 2.3 therefore gives

∥RΣG∥γ(L2(Σ),EDp,q)≳p,q,Σ∥1D∥EDp,q≳D1/p=(NJ)1/p.\|R_{\Sigma}\mathcal{G}\|_{\gamma(L^{2}(\Sigma),E_{D}^{p,q})} \gtrsim_{p,q,\Sigma} \|\mathbf{1}_{D}\|_{E_{D}^{p,q}} \gtrsim D^{1/p} = (NJ)^{1/p}.

Divide by (14) to obtain

Sp(EDp,q;Σ)≳J1/p−1/q.S_{p}(E_{D}^{p,q};\Sigma) \gtrsim J^{1/p-1/q}.

Here D=2JJD = 2^{J}J, so J≳1+log⁡DJ \gtrsim1+\log D. The restricted sequence of values J=4hJ=4h still gives coefficient dimensions with bounded successive ratios. Given an intermediate DD, choose the largest constructed dimension not exceeding DD and append zero coordinates; the norm of this inclusion is controlled independently of dimension. This extends the lower bound to all large DD, and the remaining finite set of dimensions is covered by rank-one scalar tests. □\square

Remark 3.3 (Where the sharp loss occurs). At every physical point the multiscale array has a controlled weak-ℓp\ell^{p} profile, so the source pays only J1/qJ^{1/q} for its JJ levels. By contrast, after restriction each one of the D=NJD=NJ coordinates has an L2(Σ)L^{2}(\Sigma) norm bounded below; Gaussian mixing then pays J1/pJ^{1/p}. The quotient J1/p−1/qJ^{1/p-1/q} is therefore forced by the difference between physical aggregation and Gaussian output aggregation. It is not a loss in the scalar Stein–Tomas estimate for a single packet.

The same obstruction in higher dimensions

The physical-counting mechanism is not peculiar to the three-dimensional sphere. Its decisive feature is that the Knapp packet support volume and the number of overlapping packet centers have the same parabolic exponent.

Corollary 3.4 (Curved hypersurfaces in every ambient dimension). Let d≥2d \ge2 and pd=2(d+1)/(d+3)p_d = 2(d+1)/(d+3). Suppose Σ⊂Rd\Sigma\subset\mathbb{R}^{d} is a compact smooth hypersurface with nonvanishing Gaussian curvature and a smooth positive density. For every fixed 1≤q<∞1 \le q < \infty and D≥2D \ge2,

∥RΣ:Lpd(Rd;ℓDpd,q)→γ(L2(Σ),ℓDpd,q)∥≃d,q,Σ(1+log⁡D)∣1/pd−1/q∣.\left\|R_{\Sigma}:L^{p_d}(\mathbb{R}^{d};\ell_{D}^{p_d,q})\to\gamma(L^{2}(\Sigma),\ell_{D}^{p_d,q})\right\|\simeq_{d,q,\Sigma}(1+\log D)^{\lvert1/p_d-1/q\rvert}.

Consequently Lpd,qL^{p_d,q} as a coefficient space satisfies the critical γ\gamma-valued spherical restriction bound if and only if q=pdq=p_d.

Proof. The scalar Stein–Tomas theorem and Gaussian Fubini give the upper bound on ℓDpd\ell_{D}^{p_d}, and finite Lorentz distortion gives the stated upper bound. The necessary Gaussian type-pdp_d argument and disjoint-block construction from Lemmas 2.2 and 4.1 work in every dimension, proving the lower bound when q<pdq < p_d.

For q>pdq > p_d, near a point of Σ\Sigma with tangent-normal coordinates x=(x′,xd)x=(x',x_d), use

gR(x)=R−(d+1)/pdeix⋅ξ0Ψ(x′/R,xd/R2).g_R(x)=R^{-(d+1)/p_d}e^{ix\cdot\xi_0}\Psi(x'/R,x_d/R^2).

Its physical support has volume O(Rd+1)O(R^{d+1}), and its Fourier transform has magnitude comparable to R(d+1)(1−1/pd)=R(d−1)/2R^{(d+1)(1-1/p_d)}=R^{(d-1)/2} on a cap of surface measure ≳R−(d−1)\gtrsim R^{-(d-1)}. Thus ∥RΣgR∥2≳1\|R_{\Sigma}g_R\|_2\gtrsim1. Choose M=2hM=2^h, N=Md+1N=M^{d+1}, J=(d+1)hJ=(d+1)h, and Rj=2j/(d+1)R_j=2^{j/(d+1)}. At each scale translate gRjg_{R_j} over a common grid {1,…,M}d−1×{1,…,M2}\{1,\ldots,M\}^{d-1}\times\{1,\ldots,M^2\}, containing NN centers. Every point belongs to at most CRjd+1=C2jCR_j^{d+1}=C2^j of the supports at scale jj, and the amplitude of each such coordinate is at most C2−j/pdC2^{-j/p_d}. The coefficient array therefore has pointwise weak-ℓpd\ell^{p_d} norm O(1)O(1), Lorentz ℓpd,q\ell^{p_d,q} norm O(J1/q)O(J^{1/q}), and spatial support volume O(N)O(N). Its input is at most CN1/pdJ1/qCN^{1/p_d}J^{1/q}, while Lemma 2.3 gives output at least c(NJ)1/pdc(NJ)^{1/p_d}. This proves the required lower rate, and padding finishes the proof. □

The Gaussian-type obstruction when q<pq < p

For q<pq < p the obstruction is already present at the Gaussian-type level; curvature-specific packets are unnecessary.

Lemma 4.1 (Critical type failure below the Lorentz diagonal). Let 1≤q<p1 \le q < p. For D≥2D \ge2,

Tp(EDp,q)≳p,q(1+log⁡D)1/q−1/p.T_p(E_D^{p,q})\gtrsim_{p,q}(1+\log D)^{1/q-1/p}.

Proof. Choose J≥2J \ge2 and pairwise disjoint coordinate blocks BjB_j with ∣Bj∣=2j\lvert B_j\rvert=2^j, 0≤j<J0\le j<J. Put yj=2−j/p1Bjy_j=2^{-j/p}\mathbf{1}_{B_j}. The decreasing rearrangement of yjy_j is constant on its nonzero block, and a direct substitution in (5) shows ∥yj∥p,q≃1\|y_j\|_{p,q}\simeq1 uniformly in jj.

The sum y=∑jyjy=\sum_j y_j has decreasing values approximately r−1/pr^{-1/p} for ranks 1≤r≲2J1\le r\lesssim2^J. Each dyadic block consequently contributes a fixed positive amount to the qq-th power of the Lorentz norm, giving

∥y∥p,q≃J1/q.\|y\|_{p,q}\simeq J^{1/q}.

Because the supports are disjoint, the absolute value of the Gaussian sum is the vector ∑j∣γj∣yj\sum_j |\gamma_j|y_j. Jensen’s inequality for the lattice norm and the positive cone yields

(E∥∑jγjyj∥p,q2)1/2≥E∥∑jγjyj∥p,q≥∥∑j(E∣γj∣)yj∥p,q≍J1/q.\begin{aligned} \left(\mathbb{E}\left\|\sum_j \gamma_j y_j\right\|_{p,q}^{2}\right)^{1/2} \ge\mathbb{E}\left\|\sum_j \gamma_j y_j\right\|_{p,q} \\ &\ge\left\|\sum_j (\mathbb{E}|\gamma_j|)y_j\right\|_{p,q} \asymp J^{1/q}. \end{aligned}

The type-pp denominator is (∑j∥yj∥p,qp)1/p≍J1/p(\sum_j \|y_j\|_{p,q}^{p})^{1/p} \asymp J^{1/p}. Finally, D=∑j<J22j≍22JD=\sum_{j<J}2^{2j}\asymp2^{2J}; appending zero coordinates covers other sufficiently large dimensions.

Lemma 2.2 and Lemma 4.1 give the lower spherical bound for q<pq<p. The upper bound for every qq follows from Lemmas 2.1–2.2. This proves (6).

The infinite-dimensional statement (7) follows because EDp,qE_D^{p,q} embeds (with the standard counting-measure rescaling) as a sublattice of Lp,q((0,1)3)L^{p,q}((0,1)^3) with constants independent of DD. For q=pq=p, Lp,p=LpL^{p,p}=L^p, and Gaussian Fubini plus scalar Stein–Tomas give the positive endpoint. For q>pq>p, the lower bound just proved diverges; for q<pq<p, the type obstruction diverges. Creekmore’s theorem [6] supplies attained Gaussian type pp for the infinite Lorentz spaces Lp,qL^{p,q} with q>pq>p and finite qq. This completes Theorem 1.1.

The fractional comparison model

We record the matching Lorentz result for the covariance (8), keeping its proof independent of the spherical construction.

Lemma 5.1 (Scalar fractional amplification). For p=4/3p=4/3, Fp(ℓDp)≤CF_p(\ell_D^p)\le C independently of DD. Moreover, Tp(X)≲Fp(X)T_p(X)\lesssim F_p(X) for all Banach spaces XX.

Proof. A Taylor expansion of ∣m±1∣2H|m\pm1|^{2H} at large ∣m∣|m| gives

rH(m)=H(2H−1)∣m∣2H−2+O(∣m∣2H−4).r_H(m)=H(2H-1)|m|^{2H-2}+O(|m|^{2H-4}).

At H=3/4H=3/4 this is rH(m)=38∣m∣−1/2+O(∣m∣−5/2)r_H(m)=\frac{3}{8}|m|^{-1/2}+O(|m|^{-5/2}), and in particular ∣rH(m)∣≤C(1+∣m∣)−1/2|r_H(m)|\le C(1+|m|)^{-1/2}. For a finitely supported scalar sequence a=(aj)a=(a_j), split the covariance form into the diagonal and off-diagonal parts. The diagonal is bounded by ∥a∥ℓp2\|a\|_{\ell^p}^2 since p<2p<2. Discrete Hardy–Littlewood–Sobolev controls the off-diagonal part:

∑j≠k∣aj∣∣ak∣∣j−k∣1/2≤C∥a∥ℓ4/32.\sum_{j\ne k}\frac{|a_j||a_k|}{|j-k|^{1/2}}\le C\|a\|_{\ell^{4/3}}^2.

Therefore the scalar Gaussian covariance is bounded by C∥a∥ℓp2C\|a\|_{\ell^p}^2, uniformly in the length of the sequence. Gaussian moment comparison and coordinatewise Fubini now give Fp(ℓDp)≤CF_p(\ell_D^p)\le C independently of DD.

To see necessity of type, take any fixed number mm of coefficient vectors, place them at widely separated indices in one sufficiently long window, and put zero vectors at all other indices. Since rH(n)→0r_H(n)\to0, the resulting m×mm\times m covariance matrix can be made arbitrarily close to the identity in operator norm. Gaussian covariance domination then recovers an independent Gaussian sum, proving Tp(X)≲Fp(X)T_p(X)\lesssim F_p(X).

Proof of Theorem 1.2. We first separate the easy bounds from the multiscale lower bound. The upper estimate for every qq follows by applying Lemma 2.1 to the scalar fractional estimate of Lemma 5.1. For q<pq<p, combine the necessary type inequality in Lemma 5.1 with the disjoint-block example of Lemma 4.1. The case q=pq=p is uniformly bounded by the scalar estimate and Gaussian Fubini.

It remains to prove the lower bound for q>pq>p. Fix J≥2J\ge2 and let N=2JN=2^J. At level j=0,…,J−1j=0,\ldots,J-1, use an interval length Lj=2jL_j=2^j. Give one separate coefficient coordinate to every interval [z,z+Lj−1]⊂{1,…,N}[z,z+L_j-1]\subset\{1,\ldots,N\}. Denote its label by (j,z)(j,z) and let

I={(j,z):0≤j<J, 1≤z≤N−Lj+1}.\mathcal{I}=\{(j,z):0\le j<J,\ 1\le z\le N-L_j+1\}.

The total number of coefficient coordinates is

D=∣I∣=∑j=0J−1(N−Lj+1)≲NJ.D=|\mathcal{I}|=\sum_{j=0}^{J-1}(N-L_j+1)\lesssim NJ.

For each time index 1≤n≤N1\le n\le N, define the coefficient vector xn∈EDp,qx_n\in E_D^{p,q} by

xn(j,z)=Lj−1/p1{z≤n<z+Lj}.(15)x_n(j,z)=L_j^{-1/p}\mathbf{1}_{\{z\le n<z+L_j\}}. \tag*{(15)}

Input counting. At level jj, at most Lj=2jL_j=2^j intervals contain a given nn, and every corresponding coordinate of xnx_n has size Lj−1/p=2−j/pL_j^{-1/p}=2^{-j/p}. Exactly as in the spherical packet-counting lemma, the number of coordinates greater than tt is O(t−p)O(t^{-p}). Therefore ∥xn∥ℓp,∞≤C\|x_n\|_{\ell^{p,\infty}}\le C and ∥xn∥EDp,q≲J1/q\|x_n\|_{E_D^{p,q}}\lesssim J^{1/q}. The total input budget is

(∑n=1N∥xn∥EDp,qp)1/p≲N1/pJ1/q.(16)\left(\sum_{n=1}^{N}\|x_n\|_{E_D^{p,q}}^p\right)^{1/p}\lesssim N^{1/p}J^{1/q}. \tag*{(16)}

Gaussian output. Let (Zn)(Z_n) be fractional Gaussian noise with covariance rH(j−k)r_H(j-k). By definition, there is a fractional Brownian motion BHB_H such that Zn=BH(n)−BH(n−1)Z_n=B_H(n)-B_H(n-1). Hence, for every interval of LL consecutive increments,

∑n=zz+L−1Zn=BH(z+L−1)−BH(z−1),Var⁡(∑n=zz+L−1Zn)=L2H.\sum_{n=z}^{z+L-1}Z_n=B_H(z+L-1)-B_H(z-1),\qquad\operatorname{Var}\left(\sum_{n=z}^{z+L-1}Z_n\right)=L^{2H}.

Since H=1/pH=1/p, multiplying by the coefficient L−1/pL^{-1/p} makes this variance exactly one. Thus each coordinate of the vector ∑nZnxn∈EDp,q\sum_n Z_nx_n\in E_D^{p,q} is a centered Gaussian with unit variance, although the different coordinates are correlated. The Gaussian lattice lower bound of Lemma 2.3 applies and gives

(E∥∑nZnxn∥EDp,q2)1/2≳∥1D∥EDp,q≳D1/p≳(NJ)1/p.(17)\left(\mathbb{E}\left\|\sum_n Z_nx_n\right\|_{E_D^{p,q}}^2\right)^{1/2}\gtrsim\|\mathbf{1}_D\|_{E_D^{p,q}}\gtrsim D^{1/p}\gtrsim(NJ)^{1/p}. \tag*{(17)}

Divide (17) by (16). The quotient is at least J1/p−1/qJ^{1/p-1/q}. Since D≲2JJD\lesssim2^J J, we have J≳1+log⁡DJ\gtrsim1+\log D. Padding by zero coordinates extends the result to all dimensions. This completes the fractional lower bound and the theorem. □\square

A sharp Lorentz threshold for linear capillary waves

The endpoint restriction theorem also gives a precise spacetime estimate for a dispersive partial differential equation. The example is the linearized, infinite-depth capillary-wave model on a two-dimensional water surface. After diagonalizing the linearized system and neglecting gravity, its two time-frequency branches have dispersion relation

ω(ξ)=c∣ξ∣3/2,c>0,ξ∈R2.\omega(\xi)=c|\xi|^{3/2}, \qquad c>0,\qquad\xi\in\mathbb{R}^{2}.

The corresponding second-order equation is

∂t2U+c2(−Δx)3/2U=0,(t,x)∈R×R2.(18)\partial_t^2 U+c^2(-\Delta_x)^{3/2}U=0,\qquad(t,x)\in\mathbb{R}\times\mathbb{R}^{2}. \tag*{(18)}

Here (−Δx)3/2(-\Delta_x)^{3/2} is the Fourier multiplier with symbol ∣ξ∣3|\xi|^3. The fractional power is characteristic of capillary dispersion; see [3] for the broader water-wave context. We study one complex frequency branch, which is sufficient to exhibit the sharp coefficient-space threshold.

Let A={ξ∈R2:1<∣ξ∣<2}A=\{\xi\in\mathbb{R}^{2}:1<|\xi|<2\} and fix 1<s<∞1<s<\infty. For DD initial profiles U0=(u0,1,…,u0,D)U_0=(u_{0,1},\ldots,u_{0,D}), set YDs=ℓD4,sY_D^s=\ell_D^{4,s}, with a uniformly equivalent Banach lattice norm. Write

U(t)=e−itc(−Δx)3/4U0,supp⁡u^0,j⋐A.U(t)=e^{-itc(-\Delta_x)^{3/4}}U_0,\qquad\operatorname{supp}\widehat{u}_{0,j}\Subset A.

Thus UU solves (18), with ∂tU(0)=−ic(−Δx)3/4U0\partial_tU(0)=-ic(-\Delta_x)^{3/4}U_0.

Theorem 6.1 (Sharp capillary-wave spacetime estimate). For each c>0c>0 and 1<s<∞1<s<\infty, there exist constants 0<cs,c≤Cs,c<∞0<c_{s,c}\le C_{s,c}<\infty, independent of D≥2D\ge2, such that

∥U∥L4(Rt×Rx2;ℓD4,s)≤Cs,c(1+log⁡D)∣1/4−1/s∣∥(∥u^0,j∥L2(A))j=1D∥ℓD4,s.(19)\|U\|_{L^4(\mathbb{R}_t\times\mathbb{R}^2_x;\ell_D^{4,s})}\le C_{s,c}(1+\log D)^{|1/4-1/s|}\left\|\left(\|\widehat{u}_{0,j}\|_{L^2(A)}\right)_{j=1}^{D}\right\|_{\ell_D^{4,s}}. \tag*{(19)}

The power of 1+log⁡D1+\log D is sharp: the best constant in (19), over all smooth data whose Fourier transforms are supported in AA, is bounded below by cs,c(1+log⁡D)∣1/4−1/s∣c_{s,c}(1+\log D)^{|1/4-1/s|}. In particular, the estimate is dimension-independent at s=4s=4, whereas at s=2s=2 its optimal loss is (1+log⁡D)1/4(1+\log D)^{1/4}.

The index ss is a coefficient Lorentz exponent: the time-space norm remains the ordinary L4L^4 norm. Thus the theorem is not a Lorentz refinement in time or in physical space. It determines how the same dispersive waves behave when many scalar components are combined with different sequence-space geometries.

We record the two functional-analytic ingredients used to pass from restriction to the wave equation. They also explain why the logarithmic lower bound transfers, rather than giving only an upper estimate.

Lemma 6.2 (Gaussian norm and trace duality). Fix 1<s<∞1<s<\infty and write s′=s/(s−1)s'=s/(s-1), XD=ℓD4/3,s′X_D=\ell_D^{4/3,s'} and YD=ℓD4,sY_D=\ell_D^{4,s}. Let HH be a Hilbert space. Given G1,…,GD∈HG_1,\ldots,G_D\in H, regard G=(Gj)j=1DG=(G_j)_{j=1}^{D} as the finite-rank operator TG:H→YDT_G:H\to Y_D, TGh=(⟨h,Gj⟩H)j=1DT_Gh=(\langle h,G_j\rangle_H)_{j=1}^{D}. Then

∥G∥γ(H,YD)≃s∥(∥Gj∥H)j=1D∥YD.(20)\|G\|_{\gamma(H,Y_D)}\simeq_s\left\|\left(\|G_j\|_H\right)_{j=1}^{D}\right\|_{Y_D}. \tag*{(20)}

Moreover, for a compact curved surface Σ\Sigma with smooth positive density and H=L2(Σ)H=L^2(\Sigma),

∥EΣ:γ(H,YD)→L4(R3;YD)∥≃sS4/3(XD;Σ),(21)\|E_\Sigma:\gamma(H,Y_D)\to L^4(\mathbb{R}^3;Y_D)\|\simeq_s S_{4/3}(X_D;\Sigma), \tag*{(21)}

with constants independent of $D.

Proof. Choose an orthonormal expansion Gj=∑ν=1maν,jhνG_j=\sum_{\nu=1}^{m}a_{\nu,j}h_\nu in HH. The associated γ\gamma-radonifying norm is the second Gaussian moment of the YDY_D-valued vector Z=(∑νγνaν,j)j=1DZ=\left(\sum_\nu\gamma_\nu a_{\nu,j}\right)_{j=1}^{D}. The Gaussian square-function theorem for finite-cotype Banach lattices states that this moment is comparable to the lattice norm of its coordinate standard deviations; see [13]. The standard deviations are (∑ν∣aν,j∣2)1/2=∥Gj∥H\left(\sum_\nu|a_{\nu,j}|^2\right)^{1/2}=\|G_j\|_H. The Lorentz sequence spaces ℓD4,s\ell_D^{4,s} have finite-cotype constants independent of DD for fixed ss, so this proves (20).

For (21), Lorentz duality gives XD∗≃YDX_D^*\simeq Y_D, with equivalence constants independent of DD; see [5, 6]. The spaces XDX_D are uniformly KK-convex. Indeed, for fixed s′s' they arise by real interpolation between ℓDr0\ell_D^{r_0} and ℓDr1\ell_D^{r_1} for suitable 1<r0<4/3<r1<∞1<r_0<4/3<r_1<\infty; these interpolation families have nontrivial type, with constants independent of DD. The γ\gamma-trace duality theorem therefore identifies (γ(H,XD))∗(\gamma(H,X_D))^* with γ(H,XD∗)\gamma(H,X_D^*), up to a constant independent of DD; see [13]. On finite-rank functions, the adjoint of Fourier restriction is the extension operator:

∫Σ⟨RΣf(ω),G(ω)⟩ dσ(ω)=∫R3⟨f(x),EΣG(x)⟩ dx.\int_{\Sigma}\langle R_{\Sigma}f(\omega),G(\omega)\rangle\,d\sigma(\omega)=\int_{\mathbb{R}^3}\langle f(x),E_{\Sigma}G(x)\rangle\,dx.

with the usual harmless Fourier-conjugation convention. Bochner L4/3L^{4/3}–L4L^4 duality and the preceding γ\gamma-duality identify their operator norms within the claimed uniform constants. □

Proof of Theorem 6.1. The solution is the oscillatory integral

U(t,x)=(2π)−2∫Aei(x⋅ξ−tc∣ξ∣3/2)U^0(ξ) dξ.(22)U(t,x)=(2\pi)^{-2}\int_{\mathcal{A}}e^{i(x\cdot\xi-tc|\xi|^{3/2})}\widehat{U}_0(\xi)\,d\xi. \tag*{(22)}

It is a Fourier extension integral for the graph

Σc={(ξ,c∣ξ∣3/2):ξ∈A}⊂R3.\Sigma_c=\{(\xi,c|\xi|^{3/2}):\xi\in\mathcal{A}\}\subset\mathbb{R}^3.

For a radial graph τ=ω(∣ξ∣)\tau=\omega(|\xi|), the Hessian in the frequency plane has eigenvalues ω′′(r)\omega''(r) in the radial direction and ω′(r)/r\omega'(r)/r in the tangential direction. For ω(r)=cr3/2\omega(r)=cr^{3/2} these are

ω′′(r)=3c4r−1/2,ω′(r)r=3c2r−1/2,\omega''(r)=\frac{3c}{4}r^{-1/2},\qquad\frac{\omega'(r)}{r}=\frac{3c}{2}r^{-1/2},

so

det⁡D2ω(ξ)=9c28∣ξ∣>0on A.\det D^2\omega(\xi)=\frac{9c^2}{8|\xi|}>0\quad\text{on }\mathcal{A}.

Thus Σc\Sigma_c has nonvanishing Gaussian curvature on the chosen annulus. The graph measure dξd\xi differs from surface measure by a bounded, smooth positive factor, which changes only constants depending on cc and the annulus.

Apply Theorem 1.1 with Lorentz secondary index q=s′q=s':

S4/3(XD;Σc)≍s,c(1+log⁡D)∣3/4−1/s′∣=(1+log⁡D)∣1/4−1/s∣.\mathcal{S}_{4/3}(X_D;\Sigma_c)\asymp_{s,c}(1+\log D)^{|3/4-1/s'|}=(1+\log D)^{|1/4-1/s|}.

Lemma 6.2 transfers this sharp estimate to EΣc:γ(L2(A),YD)→L4(R1+2;YD)E_{\Sigma_c}:\gamma(L^2(\mathcal{A}),Y_D)\to L^4(\mathbb{R}^{1+2};Y_D). Finally, formula (20) turns the γ\gamma-norm of the frequency data into the coefficient norm on the right of (19). This proves its upper bound. The lower bound is equally important: if a smaller power of 1+log⁡D1+\log D held uniformly for every smooth initial array, (21) would imply a smaller restriction constant for XDX_D, contradicting the lower half of Theorem 1.1. Smooth compactly supported frequency arrays are dense in the finite-rank spaces used in the duality argument, so the sharpness is realized up to a fixed multiplicative constant by such arrays. □

Remark 6.3 (Why the coefficient threshold is relevant). The result distinguishes two notions of stability for a family of linear waves. Each scalar capillary mode obeys the familiar L4L^4 spacetime estimate, but assembling DD modes in ℓ4,s\ell^{4,s} can cost an unavoidable logarithmic power. At s=4s=4 the channels combine uniformly; at s=2s=2 a fourth-root logarithm is necessary. The statement concerns the linearized model, not the nonlinear free-boundary evolution.

Real interpolation beyond the critical Lorentz line

The sharp logarithmic estimate of Theorem 1.1 concerns the critical primary Lorentz exponent p=4/3p=4/3. A complementary positive region for larger primary exponents follows from a general stability principle. The argument is included to distinguish a genuine endpoint obstruction from ordinary interpolation of positive spaces. We use standard real-interpolation conventions as in [4, 5].

Proposition 7.1 (Real-interpolation stability of critical restriction). Let (X0,X1)(X_0,X_1) be a compatible regular Banach couple, let 0<θ<10<\theta<1, and put Xθ,q=(X0,X1)θ,qX_{\theta,q}=(X_0,X_1)_{\theta,q} using the real KK-method. If RΣ:Lp(R3;Xi)→γ(L2(Σ),Xi)R_\Sigma:L^p(\mathbb{R}^3;X_i)\to\gamma(L^2(\Sigma),X_i) is bounded for i=0,1i=0,1 and p=4/3p=4/3, then for every finite p≤q<∞p\le q<\infty,

Sp(Xθ,q;Σ)≤Cp,q,θSp(X0;Σ)1−θSp(X1;Σ)θ.S_p(X_{\theta,q};\Sigma)\le C_{p,q,\theta}S_p(X_0;\Sigma)^{1-\theta}S_p(X_1;\Sigma)^\theta.

Proof. Let H=L2(Σ)H=L^2(\Sigma). For the source, the pp-convexified KK-functional

Kp(t,x;X0,X1)=inf⁡x=x0+x1(∥x0∥X0p+tp∥x1∥X1p)1/pK_p(t,x;X_0,X_1)=\inf_{x=x_0+x_1}\left(\lVert x_0\rVert_{X_0}^p+t^p\lVert x_1\rVert_{X_1}^p\right)^{1/p}

is uniformly equivalent to the usual KK-functional. On finite simple Bochner functions the KpK_p-functional for the couple (Lp(X0),Lp(X1))(L^p(X_0),L^p(X_1)) is the LpL^p norm of the pointwise KpK_p-functional. The standard density and measurability argument extends this comparison to the compatible Bochner couple. Minkowski for q≥pq\ge p therefore gives

Lp(R3;(X0,X1)θ,q)↪(Lp(R3;X0),Lp(R3;X1))θ,q.L^p(\mathbb{R}^3;(X_0,X_1)_{\theta,q})\hookrightarrow(L^p(\mathbb{R}^3;X_0),L^p(\mathbb{R}^3;X_1))_{\theta,q}.

(A)

For the target, start with a finite-rank T:H→X0+X1T:H\to X_0+X_1 and an orthonormal basis of its Hilbert input. Let ZT=∑jγjThjZ_T=\sum_j\gamma_jTh_j. For any decomposition T=T0+T1T=T_0+T_1 and almost every Gaussian outcome,

K(t,ZT;X0,X1)≤∥ZT0∥X0+t∥ZT1∥X1.K(t,Z_T;X_0,X_1)\le\lVert Z_{T_0}\rVert_{X_0}+t\lVert Z_{T_1}\rVert_{X_1}.

Taking LqL^q in the Gaussian variable, then infimizing over operator decompositions and using Gaussian Kahane–Khintchine, yields

∥K(t,ZT;X0,X1)∥Lq(Ω)≤CqK(t,T;γ(H,X0),γ(H,X1)).\lVert K(t,Z_T;X_0,X_1)\rVert_{L^q(\Omega)} \le C_qK(t,T;\gamma(H,X_0),\gamma(H,X_1)).

Integrating its qq-th power against t−θq dt/tt^{-\theta q}\,dt/t, applying Fubini, and again using Kahane–Khintchine gives

(γ(H,X0),γ(H,X1))θ,q↪γ(H,(X0,X1)θ,q).(\gamma(H,X_0),\gamma(H,X_1))_{\theta,q}\hookrightarrow\gamma(H,(X_0,X_1)_{\theta,q}).

(B)

These estimates extend from finite-rank operators by completion. Interpolating the same restriction operator between its two endpoint bounds and composing (A)–(B) proves the assertion. □\square

Corollary 7.2 (Positive Lorentz region above the critical primary exponent). For r>4/3r>4/3 and 4/3≤q<∞4/3\le q<\infty, every classical Lorentz function space Lr,qL^{r,q} satisfies the critical Stein–Tomas restriction estimate

S4/3(Lr,q;Σ)<∞.S_{4/3}(L^{r,q};\Sigma)<\infty.

Proof. Choose 4/3≤r0<r<r1<∞4/3 \le r_0 < r < r_1 < \infty and 0<θ<10 < \theta< 1 with 1/r=(1−θ)/r0+θ/r11/r = (1-\theta)/r_0 + \theta/r_1. The real-interpolation identity (Lr0,Lr1)θ,q=Lr,q(L^{r_0}, L^{r_1})_{\theta,q} = L^{r,q} holds with equivalence constants independent of finite-dimensional approximations. For ri≥4/3r_i \ge4/3, scalar Stein–Tomas and the mixed-norm Minkowski inequality yield the critical estimate on LriL^{r_i}: the L2L^2 norm in surface frequency can be passed inside the coefficient LriL^{r_i} norm, and the physical L4/3L^{4/3} norm controls the resulting mixed integral since ri≥4/3r_i \ge4/3. Apply Proposition 7.1 to conclude.

Corollary 7.3 (Matching type and cotype need not predict restriction). Fix q>2q > 2 and put

X−=L4/3,q(0,1),X+=L4/3((0,1);Lq(0,1)).X_- = L^{4/3,q}(0,1), \qquad X_+ = L^{4/3}((0,1);L^q(0,1)).

Both spaces have optimal attained Gaussian type 4/34/3 and optimal cotype qq, but

S4/3(X−;S2)=∞,S4/3(X+;S2)<∞.S_{4/3}(X_-;S^2)=\infty,\qquad S_{4/3}(X_+;S^2)<\infty.

Proof. The type and cotype exponents follow from the classical Lorentz calculations [6] and the standard mixed-Lebesgue geometry. The negative assertion is Theorem 1.1. For the positive assertion, Lq(0,1)L^q(0,1) has actual Gaussian type 2 and therefore critical restriction; outer L4/3L^{4/3} integration preserves the estimate by Gaussian Fubini and Minkowski.

The index restriction q<∞q < \infty in Proposition 7.1 is essential for the argument: Gaussian LqL^q-moment interchange cannot simply be replaced by an essential supremum at q=∞q = \infty. In the usual infinite-dimensional weak-Lorentz spaces Lr,∞L^{r,\infty} for r>1r > 1, disjoint blocks of sizes 2j2^j, scaled by 2−j/r2^{-j/r}, span uniformly isomorphic copies of ℓ∞\ell^\infty. Hence these weak-Lorentz coefficient spaces fail the necessary nontrivial Gaussian type condition; the same obstruction passes to the diagonal of Sr,∞S^{r,\infty}. These corollaries do not settle the general Banach-space classification at attained type 4/34/3.

Geometric interpretation and limitations

Theorems 1.1 and 1.2 identify a precise boundary between independent Gaussian aggregation and the accumulation of multiscale correlations. At the critical primary Lorentz exponent, a finite-dimensional distortion of logarithmic size can become visible in the Fourier trace even though every scalar component satisfies the same Stein–Tomas estimate.

Gaussian type is related to Besov-to-γ\gamma embeddings by Kalton–van Neerven–Veraar–Weis [7]. Critical Bessel-potential-to-γ\gamma embeddings and positive lattice cases are studied in [8]. These provide useful context, but type and cotype alone do not characterize the curved restriction endpoint: Corollary 7.3 supplies two explicit spaces with matching optimal type and cotype and different restriction behavior.

The matching Lorentz growth laws do not imply either universal Banach implication

S4/3(X)<∞⟹F4/3(X)<∞,F4/3(X)<∞⟹S4/3(X)<∞.S_{4/3}(X)<\infty\Longrightarrow F_{4/3}(X)<\infty,\qquad F_{4/3}(X)<\infty\Longrightarrow S_{4/3}(X)<\infty.

Both would require estimates uniform across all finite-dimensional coefficient geometries, beyond the Lorentz family. The operator/Banach-space bipolar theorem of de la Salle [9] provides an abstract framework for such universal implications; it does not itself supply a comparison for these two operators.

The critical Gaussian-type threshold and open problems

The results above separate Gaussian type from the curvature-sensitive endpoint. At p=4/3p = 4/3, Gaussian type pp is necessary but is not sufficient for either SpS_p or FpF_p, while strict Gaussian type r>pr > p is known to imply the critical fractional γ\gamma-embedding by the type–γ\gamma theory of [8]. This leaves a genuine comparison question at attained type pp. It is important not to conjecture that type pp alone suffices: Theorems 1.1 and 1.2 disprove that statement.

Conjecture 9.1 (Critical spherical-to-fractional transfer). For p=4/3p = 4/3 and every Banach space XX,

Sp(X;S2)<∞⟹Fp(X)<∞.S_p(X; S^2) < \infty\quad\Longrightarrow\quad F_p(X) < \infty.

Equivalently, for each finite KK there exists C(K)<∞C(K) < \infty such that

Fp(E)≤C(K)for every finite-dimensional E with Sp(E;S2)≤K.F_p(E) \le C(K) \qquad\text{for every finite-dimensional } E \text{ with } S_p(E; S^2) \le K.

The quantitative equivalence in Conjecture 9.1 follows by Gaussian pp-moment equivalence and the exact ℓp\ell^p-direct-sum law for both amplification constants: finite-dimensional counterexamples with uniformly bounded spherical constants and divergent fractional constants could be combined in one ℓp\ell^p-sum. The conjecture is a proposed universal implication, not a consequence of the Lorentz results in this paper.

Problem 9.2 (Strict-type spherical sufficiency and the reverse implication). Fix any r∈(4/3,2]r \in(4/3, 2]. Is

sup⁡dim⁡E<∞Tr(E)≤KS4/3(E;S2)<∞for every finite K?\sup_{\substack{\dim E<\infty\\ T_r(E)\le K}} S_{4/3}(E; S^2) < \infty\qquad\text{for every finite } K?

More generally, does F4/3(X)<∞F_{4/3}(X) < \infty imply S4/3(X;S2)<∞S_{4/3}(X; S^2) < \infty for arbitrary Banach spaces XX?

A negative answer to the first question at one fixed r>4/3r > 4/3, with a uniformly type-rr family and unbounded spherical constants, would also answer the second question negatively, because the fractional model is uniformly bounded on such a family by the established strict-type theorem. On the other hand, a positive strict-type theorem would not alone settle the reverse implication at attained type 4/34/3. Neither problem is claimed resolved here.

Conclusion

The Lorentz coefficient family has a sharp critical threshold and a quantitative finite-dimensional growth law. The same logarithmic law holds for long-memory fractional Gaussian increments and yields a sharp spacetime bound for frequency-localized capillary waves. The remaining open problems concern geometric structures beyond Lorentz lattices, where Gaussian type no longer determines how curved Fourier traces recombine.

Use of generative AI and author responsibility

Gemini used for argument checking, and language preparation. The authors are responsible for the mathematical statements, proofs, citations, and conclusions.

References

  1. [1]J.-G. Bak and A. Seeger, Extensions of the Stein–Tomas theorem, Math. Res. Lett. 18 (2011), 767–781, doi:10.4310/MRL.2011.v18.n4.a14.DOI
  2. [2]R. Shahmurov and V. Shahmurov, γ-Valued Stein–Tomas Restriction for Schatten Classes, arXiv:2610.40044 (2026), https://arxiv.org/abs/2610.40044.
  3. [3]Y. Deng, A. D. Ionescu, B. Pausader, and F. Pusateri, Global solutions of the gravity-capillary water-wave system in three dimensions, Acta Math. 219 (2017), 213–402, doi:10.4310/ACTA.2017.v219.n2.a1.DOI
  4. [4]J. Bergh and J. Löfström, Interpolation Spaces: An Introduction, Springer, Berlin, 1976.
  5. [5]C. Bennett and R. Sharpley, Interpolation of Operators, Academic Press, Boston, 1988.
  6. [6]J. Creekmore, Type and cotype in Lorentz L_{p,q} spaces, Indag. Math. (Proc.) 84 (1981), 145–152, doi:10.1016/1385-7258(81)90143-8.DOI
  7. [7]N. Kalton, J. van Neerven, M. Veraar, and L. Weis, Embedding vector-valued Besov spaces into spaces of γ-radonifying operators, Math. Nachr. 281 (2008), 238–252.
  8. [8]M. Veraar, Embedding results for γ-spaces, in Recent Trends in Analysis, Amer. Math. Soc. (2013), 209–219.
  9. [9]M. de la Salle, A duality operators/Banach spaces, Ann. Inst. Fourier 76 (2026), no. 3, 1119–1161, doi:10.5802/aif.3730.DOI
  10. [10]E. M. Stein, Harmonic Analysis: Real-Variable Methods, Orthogonality, and Oscillatory Integrals, Princeton University Press, 1993.
  11. [11]D. Oliveira e Silva, The endpoint Stein–Tomas inequality: old and new, São Paulo J. Math. Sci. 18 (2024), 1015–1024.DOI
  12. [12]P. Tomas, A restriction theorem for the Fourier transform, Bull. Amer. Math. Soc. 81 (1975), 477–478.DOI
  13. [13]J. van Neerven, γ-Radonifying operators—a survey, in The AMSI–ANU Workshop on Spectral Theory and Harmonic Analysis, Proc. Centre Math. Appl. 44 (2010), 1–61.DOI

Paper details

vlnc:W_pt7r9dyq
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 4

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Gemini (multiple versions)

View statement
Looking for related papers…
Contents