Introduction
A central problem in Fourier restriction theory is to estimate ∥ f ∥ L p \lVert f\rVert_{L^p} ∥ f ∥ L p when the Fourier support of f f f lies in a small neighborhood of a manifold M M M . Wave packet decomposition is a standard tool for estimating ∥ f ∥ L p \lVert f\rVert_{L^p} ∥ f ∥ L p . We briefly recall the decomposition. First, we partition the Fourier support of f f f into rectangular boxes. To each frequency box we associate a dual box in physical space with reciprocal side lengths, as suggested by the uncertainty principle. We then write the corresponding part of f f f as a sum of functions, each essentially supported on a translate of the dual box. These functions are called wave packets. This leads to the following informal question.
Question 1.1. Let M ⊂ R n M \subset\mathbb{R}^{n} M ⊂ R n be a manifold , and let f f f be a Schwartz function with Fourier support in a small neighborhood of M M M . How sharply can ∥ f ∥ L p \|f\|_{L^{p}} ∥ f ∥ L p be bounded in terms of the geometry and amplitudes of its wave packets?
Guth-Wang-Zhang [14 ] first provided a satisfactory answer for the parabola in R 2 \mathbb{R}^{2} R 2 and the cone in R 3 \mathbb{R}^{3} R 3 by introducing the notion of wave envelopes and proving an L 4 L^{4} L 4 wave envelope estimate. Their L 4 L^{4} L 4 wave envelope estimate led to a resolution of the local smoothing conjecture for the wave equation in 2 + 1 2+1 2 + 1 dimensions. Guth-Maldague [9 ] subsequently considered the following informal question, which also takes the amplitude of f f f into account.
Question 1.2. Let M ⊂ R n M \subset\mathbb{R}^{n} M ⊂ R n be a manifold , and let f f f be a Schwartz function with Fourier support in a small neighborhood of M M M . For α > 0 \alpha> 0 α > 0 , how sharply can the measure
{ x ∈ R n : α ≤ ∣ f ( x ) ∣ < 2 α } \left\{x \in\mathbb{R}^{n}: \alpha\leq|f(x)| < 2\alpha\right\} { x ∈ R n : α ≤ ∣ f ( x ) ∣ < 2 α } be bounded in terms of the geometry and amplitudes of its wave packets?
Guth-Maldague [9 ] gave satisfactory answers for the parabola in R 2 \mathbb{R}^{2} R 2 and the cone in R 3 \mathbb{R}^{3} R 3 through amplitude dependent wave envelope estimates. Their estimates provide a unified proof of several results in harmonic analysis that had previously been established by different methods based on wave packet decomposition. They yield Bochner–Riesz estimates, square function estimates, decoupling inequalities, small cap decoupling inequalities, and restriction estimates for both the parabola in R 2 \mathbb{R}^{2} R 2 and the cone in R 3 \mathbb{R}^{3} R 3 . Guth-Maldague subsequently used their small cap decoupling for the cone in R 3 \mathbb{R}^{3} R 3 to prove sharp small cap decoupling for the moment curve in R 3 \mathbb{R}^{3} R 3 [11 ] . The same inequality for the cone was also used by Li-Yang [17 ] , and more recently by Cen-Gao-Zhang-Zhang [CGZZ26], to obtain improved bounds in the Gauss circle problem.
In this paper, we answer Question 1.2 for the moment curve in R n \mathbb{R}^{n} R n by proving an amplitude dependent wave envelope type estimate. A key ingredient in our proof is the sharp square function estimate for the moment curve, proved by Maldague [18 ] for n = 3 n = 3 n = 3 and by Guth-Maldague [10 ] for n ≥ 4 n \geq4 n ≥ 4 . Our amplitude dependent wave envelope type estimate implies the following known results in harmonic analysis for the moment curve in R n \mathbb{R}^{n} R n :
square function estimates [18 , 10 ] ;
the wave envelope estimate of Gan-Maldague-Oh [12 ] ;
Bochner–Riesz type estimates [12 ] ;
restriction estimates [7 ] .
Moreover, our amplitude dependent wave envelope type estimate implies the following new result:
sharp small cap decoupling inequalities for the moment curve in every dimension n ≥ 4 n \geq4 n ≥ 4 , thereby giving an affirmative answer to the conjecture of Demeter-Guth-Wang [6 Conjecture 2.5] . The applications are proved in Part III.
Small cap decoupling inequalities
Our main result is the following exponential sum estimate.
Theorem 1.3. Let n ≥ 4 n \geq4 n ≥ 4 , let N ≥ 1 N \geq1 N ≥ 1 be an integer, and let 2 ≤ p < ∞ 2 \leq p < \infty 2 ≤ p < ∞ . For 0 ≤ σ ≤ n 0 \leq\sigma\leq n 0 ≤ σ ≤ n and an interval H ⊂ R H \subset\mathbb{R} H ⊂ R of length N − σ N^{-\sigma} N − σ ,
∫ [ 0 , 1 ] n − 1 × H ∣ ∑ k = 1 N a k e 2 π i ( k x 1 + k 2 x 2 + ⋯ + k n x n ) ∣ p d x ≲ N ϵ ( N p / 2 − σ + N p − n ( n + 1 ) / 2 ) (1) \int_{[0,1]^{n-1} \times H} \left|\sum_{k=1}^{N} a_{k}e^{2\pi i(kx_{1}+k^{2}x_{2}+\cdots+k^{n}x_{n})}\right|^{p}\,dx \lesssim N^{\epsilon}\left(N^{p/2-\sigma}+N^{p-n(n+1)/2}\right) \tag*{(1)} ∫ [ 0 , 1 ] n − 1 × H k = 1 ∑ N a k e 2 π i ( k x 1 + k 2 x 2 + ⋯ + k n x n ) p d x ≲ N ϵ ( N p /2 − σ + N p − n ( n + 1 ) /2 ) (1) for every complex sequence ( a k ) k = 1 N (a_{k})_{k=1}^{N} ( a k ) k = 1 N satisfying ∣ a k ∣ ≤ 1 |a_{k}| \leq1 ∣ a k ∣ ≤ 1 .
Theorem 1.3 is sharp up to N ϵ N^{\epsilon} N ϵ losses: the two terms on the right hand side are obtained by square root cancellation and constructive interference.
The case σ = 0 \sigma= 0 σ = 0 is the Vinogradov mean value estimate, proved for n = 3 n = 3 n = 3 by Wooley [22 ] , and in general by Bourgain-Demeter-Guth [3 ] and Wooley [23 ] . The same bound at σ = n \sigma= n σ = n follows from the ( n − 1 ) (n-1) ( n − 1 ) dimensional estimate, applied for each fixed x n x_{n} x n . For 0 < σ ≤ n − 1 0 < \sigma\leq n-1 0 < σ ≤ n − 1 , taking a k = 1 a_{k} = 1 a k = 1 and H = [ 0 , N − σ ] H = [0,N^{-\sigma}] H = [ 0 , N − σ ] resolves the conjecture of Demeter-Guth-Wang [6 Conjecture 2.5] in all dimensions. For n − 1 < σ ≤ n n-1 < \sigma\leq n n − 1 < σ ≤ n , the estimate with a k = 1 a_{k} = 1 a k = 1 is predicted by Wooley’s extended main conjecture in Vinogradov’s mean value theorem [Woo23, Conjecture 8.2].
Demeter-Guth-Wang [6 ] proved the conjecture for n = 3 n=3 n = 3 in the range 0 ≤ σ ≤ 3 / 2 0\leq\sigma\leq3/2 0 ≤ σ ≤ 3/2 . Jung [Jun23, Theorem 1.2] proved sharp L 10 L^{10} L 10 small cap decoupling for the moment curve in R 3 \mathbb{R}^{3} R 3 at cap width R − 1 / 2 R^{-1/2} R − 1/2 , which recovers the tenth moment bound in (1) for cubic exponential sums when σ = 1 \sigma=1 σ = 1 . Guth-Maldague [11 ] subsequently proved (1) for n = 3 n=3 n = 3 and 0 ≤ σ ≤ 2 0\leq\sigma\leq2 0 ≤ σ ≤ 2 , resolving the full cubic case of the conjecture. Maldague-Oh [MO26, Theorem 1.1] considered the remaining range 2 < σ < 3 2<\sigma<3 2 < σ < 3 . For 0 ≤ σ < 3 0\leq\sigma<3 0 ≤ σ < 3 , 2 ≤ p < ∞ 2\leq p<\infty 2 ≤ p < ∞ , ∣ a k ∣ ≤ 1 |a_k|\leq1 ∣ a k ∣ ≤ 1 , and an interval H H H of length N − σ N^{-\sigma} N − σ , their result gives
∫ [ 0 , 1 ] 2 × H ∣ ∑ k = 1 N a k e 2 π i ( k x 1 + k 2 x 2 + k 3 x 3 ) ∣ p d x ≲ ϵ N ϵ ( N p / 2 − σ + N 1 + σ ( p − 7 ) / 3 + N p − 6 ) . \int_{[0,1]^2\times H}\left|\sum_{k=1}^{N}a_k e^{2\pi i(kx_1+k^2x_2+k^3x_3)}\right|^p\,dx\lesssim_{\epsilon}N^\epsilon\left(N^{p/2-\sigma}+N^{1+\sigma(p-7)/3}+N^{p-6}\right). ∫ [ 0 , 1 ] 2 × H k = 1 ∑ N a k e 2 π i ( k x 1 + k 2 x 2 + k 3 x 3 ) p d x ≲ ϵ N ϵ ( N p /2 − σ + N 1 + σ ( p − 7 ) /3 + N p − 6 ) .
The middle term is dominated by the other two for σ ≤ 5 / 2 \sigma\leq5/2 σ ≤ 5/2 , but is necessary when 5 / 2 < σ < 3 5/2<\sigma<3 5/2 < σ < 3 . Note that Theorem 1.3 is stated only for n ≥ 4 n\geq4 n ≥ 4 , where no analogous middle term is needed. In dimension four, the range 3 ≤ σ < 4 3\leq\sigma<4 3 ≤ σ < 4 of (1) follows from [8 Corollary 1.4] .
Theorem 1.3 follows from a small cap decoupling estimate which we now describe. Let γ n ( t ) = ( t , t 2 , … , t n ) \gamma_n(t)=(t,t^2,\ldots,t^n) γ n ( t ) = ( t , t 2 , … , t n ) for 0 ≤ t ≤ 1 0\leq t\leq1 0 ≤ t ≤ 1 . For n ≥ 2 n\geq2 n ≥ 2 , R ≥ 1 R\geq1 R ≥ 1 , and β ≥ 1 / n \beta\geq1/n β ≥ 1/ n , consider the anisotropic neighborhood
M n ( R β , R ) = { γ n ( t ) + ∑ j = 2 n a j j ! γ n ( j ) ( t ) : t ∈ [ 0 , 1 ] , ∣ a j ∣ ≤ R − j β ( 2 ≤ j ≤ n − 1 ) , ∣ a n ∣ ≤ R − 1 } . \mathcal{M}^n(R^\beta,R)=\left\{\gamma_n(t)+\sum_{j=2}^{n}\frac{a_j}{j!}\gamma_n^{(j)}(t):\ t\in[0,1],\ |a_j|\leq R^{-j\beta}\ (2\leq j\leq n-1),\ |a_n|\leq R^{-1}\right\}. M n ( R β , R ) = { γ n ( t ) + j = 2 ∑ n j ! a j γ n ( j ) ( t ) : t ∈ [ 0 , 1 ] , ∣ a j ∣ ≤ R − j β ( 2 ≤ j ≤ n − 1 ) , ∣ a n ∣ ≤ R − 1 } .
For n = 1 n=1 n = 1 , set M 1 ( R ) = [ 0 , 1 ] \mathcal{M}^1(R)=[0,1] M 1 ( R ) = [ 0 , 1 ] and use the canonical partition. For 0 < δ ≤ 1 0<\delta\leq1 0 < δ ≤ 1 , let I ( δ ) \mathcal{I}(\delta) I ( δ ) be the dyadic partition of [ 0 , 1 ] [0,1] [ 0 , 1 ] with interval lengths in ( δ / 2 , δ ] (\delta/2,\delta] ( δ /2 , δ ] . For I ∈ I ( R − β ) I\in\mathcal{I}(R^{-\beta}) I ∈ I ( R − β ) , define the small cap
γ I : = { γ n ( t ) + ∑ j = 2 n a j j ! γ n ( j ) ( t ) : t ∈ I , ∣ a j ∣ ≤ R − j β ( 2 ≤ j < n ) , ∣ a n ∣ ≤ R − 1 } . \gamma_I:=\left\{\gamma_n(t)+\sum_{j=2}^{n}\frac{a_j}{j!}\gamma_n^{(j)}(t):\ t\in I,\ |a_j|\leq R^{-j\beta}\ (2\leq j<n),\ |a_n|\leq R^{-1}\right\}. γ I := { γ n ( t ) + j = 2 ∑ n j ! a j γ n ( j ) ( t ) : t ∈ I , ∣ a j ∣ ≤ R − j β ( 2 ≤ j < n ) , ∣ a n ∣ ≤ R − 1 } .
Since the first coordinate is t t t , these caps form a disjoint partition of M n ( R β , R ) \mathcal{M}^n(R^\beta,R) M n ( R β , R ) . Each γ I \gamma_I γ I is contained in a fixed dilate of an R − β × ⋯ × R − ( n − 1 ) β × R − 1 R^{-\beta}\times\cdots\times R^{-(n-1)\beta}\times R^{-1} R − β × ⋯ × R − ( n − 1 ) β × R − 1 box. When β = 1 / n \beta=1/n β = 1/ n , this is the canonical partition of M n ( R ) : = M n ( R 1 / n , R ) \mathcal{M}^n(R):=\mathcal{M}^n(R^{1/n},R) M n ( R ) := M n ( R 1/ n , R ) , whose caps have dimensions R − 1 / n × R − 2 / n × ⋯ × R − 1 R^{-1/n}\times R^{-2/n}\times\cdots\times R^{-1} R − 1/ n × R − 2/ n × ⋯ × R − 1 . For n = 3 n=3 n = 3 , this is the small cap considered by Guth-Maldague [11 ] . For each I ∈ I ( R − β ) I\in\mathcal{I}(R^{-\beta}) I ∈ I ( R − β ) , define f γ I f_{\gamma_I} f γ I by f γ I ^ = 1 γ I f ^ \widehat{f_{\gamma_I}}=1_{\gamma_I}\widehat f f γ I = 1 γ I f .
Theorem 1.4 (Small cap decoupling). Let n ≥ 4 n\geq4 n ≥ 4 , β ≥ 1 / n \beta\geq1/n β ≥ 1/ n , and 2 ≤ p < ∞ 2\leq p<\infty 2 ≤ p < ∞ . For every ϵ > 0 \epsilon>0 ϵ > 0 ,
∥ f ∥ p ≲ ϵ R ϵ ( R β ( p / 2 − 1 ) + R β ( p − n ( n − 1 ) / 2 − 1 ) − 1 ) ( ∑ I ∈ I ( R − β ) ∥ f γ I ∥ p p ) 1 / p (2) \|f\|_p\lesssim_\epsilon R^\epsilon\left(R^{\beta(p/2-1)}+R^{\beta(p-n(n-1)/2-1)-1}\right)\left(\sum_{I\in\mathcal{I}(R^{-\beta})}\|f_{\gamma_I}\|_p^p\right)^{1/p} \tag*{(2)} ∥ f ∥ p ≲ ϵ R ϵ ( R β ( p /2 − 1 ) + R β ( p − n ( n − 1 ) /2 − 1 ) − 1 ) I ∈ I ( R − β ) ∑ ∥ f γ I ∥ p p 1/ p (2)
for every Schwartz function f f f with supp f ^ ⊂ M n ( R β , R ) \operatorname{supp}\widehat f\subset\mathcal{M}^n(R^\beta,R) supp f ⊂ M n ( R β , R ) .
Theorem 1.4 implies Theorem 1.3 for 0 ≤ σ < n 0\leq\sigma<n 0 ≤ σ < n by the argument of Guth-Maldague [11 Section 1.1] . Both powers of R R R are sharp. The first comes from square root cancellation and the second from constructive interference of wave packets centered at the origin. The two powers agree at the critical exponent p = n ( n − 1 ) + 2 β p=n(n-1)+\frac{2}{\beta} p = n ( n − 1 ) + β 2 . At β = 1 / n \beta=1/n β = 1/ n , Theorem 1.4 recovers the canonical ℓ p ( L p ) \ell^p(L^p) ℓ p ( L p ) decoupling estimate with critical exponent n ( n + 1 ) n(n+1) n ( n + 1 ) . As β \beta β increases, the critical exponent decreases to n ( n − 1 ) n(n-1) n ( n − 1 ) , the canonical critical exponent for the moment curve projected onto the first n − 1 n-1 n − 1 coordinates.
Amplitude dependent wave envelope type estimates
The key ingredient in the proof of Theorem 1.4 is an amplitude dependent wave envelope type estimate for the moment curve in R n \mathbb{R}^n R n . Since its formulation is involved, we first introduce the ordinary wave envelope type estimate. As mentioned above, Guth-Wang-Zhang [14 ] introduced the wave envelope estimate for the cone in R 3 \mathbb{R}^3 R 3 , which is equivalent to the L 4 L^4 L 4 square function estimate. For the moment curve, the sharp square function estimates were proved by Maldague [18 ] in R 3 \mathbb{R}^3 R 3 and by Guth-Maldague [10 ] in higher dimensions. We formulate a wave envelope type estimate, which is equivalent to their square function estimate. This “wave envelope type” estimate is different from wave envelope estimates in the literature. We explain more after introducing Proposition 1.5 .
For 0 < s ≤ 1 0<s\leq1 0 < s ≤ 1 , let Θ ( s ) \Theta(s) Θ ( s ) denote the canonical partition of M n ( s − n ) \mathcal{M}^n(s^{-n}) M n ( s − n ) . For each cap τ = γ I \tau=\gamma_I τ = γ I with I ∈ I ( s ) I\in\mathcal{I}(s) I ∈ I ( s ) , write ∣ τ ∣ = s |\tau|=s ∣ τ ∣ = s and let t τ t_\tau t τ be the left endpoint of I I I . We write θ \theta θ for a cap in Θ ( R − 1 / n ) \Theta(R^{-1/n}) Θ ( R − 1/ n ) . For cones, use
ϕ d ( t ) = ( 1 , t , t 2 / 2 ! , … , t d / d ! ) . \phi_d(t)=(1,t,t^2/2!,\ldots,t^d/d!). ϕ d ( t ) = ( 1 , t , t 2 /2 ! , … , t d / d !) .
and the neighborhood of the cone over the moment curve
Γ d + 1 ( R ) = { ∑ i = 0 d a i ϕ d ( i ) ( t ) : t ∈ [ 0 , 1 ] , 1 ≤ a 0 ≤ 2 , ∣ a i ∣ ≤ R − i / d ( 1 ≤ i ≤ d ) } . \Gamma^{d+1}(R)=\left\{\sum_{i=0}^{d}a_i\phi_d^{(i)}(t):t\in[0,1],\ 1\le a_0\le2,\ |a_i|\le R^{-i/d}\ (1\le i\le d)\right\}. Γ d + 1 ( R ) = { i = 0 ∑ d a i ϕ d ( i ) ( t ) : t ∈ [ 0 , 1 ] , 1 ≤ a 0 ≤ 2 , ∣ a i ∣ ≤ R − i / d ( 1 ≤ i ≤ d ) } .
For 0 < s ≤ 1 0<s\le1 0 < s ≤ 1 and ξ = ( ξ 0 , … , ξ d ) ∈ Γ d + 1 ( s − d ) \xi=(\xi_0,\ldots,\xi_d)\in\Gamma^{d+1}(s^{-d}) ξ = ( ξ 0 , … , ξ d ) ∈ Γ d + 1 ( s − d ) , let u ( ξ ) u(\xi) u ( ξ ) be the point of [ 0 , 1 ] [0,1] [ 0 , 1 ] nearest to ξ 1 / ξ 0 \xi_1/\xi_0 ξ 1 / ξ 0 . Define
τ I = { ξ ∈ Γ d + 1 ( s − d ) : u ( ξ ) ∈ I } , Ξ ( s ) = { τ I : I ∈ I ( s ) } . \tau_I=\{\xi\in\Gamma^{d+1}(s^{-d}):u(\xi)\in I\},\qquad\Xi(s)=\{\tau_I:I\in\mathcal{I}(s)\}. τ I = { ξ ∈ Γ d + 1 ( s − d ) : u ( ξ ) ∈ I } , Ξ ( s ) = { τ I : I ∈ I ( s )} .
These caps partition Γ d + 1 ( s − d ) \Gamma^{d+1}(s^{-d}) Γ d + 1 ( s − d ) . Each τ I \tau_I τ I is contained in a fixed dilate of a 1 × s × ⋯ × s d 1\times s\times\cdots\times s^d 1 × s × ⋯ × s d box. For each cap τ = τ I ∈ Ξ ( s ) \tau=\tau_I\in\Xi(s) τ = τ I ∈ Ξ ( s ) , write ∣ τ ∣ = s |\tau|=s ∣ τ ∣ = s and let t τ t_\tau t τ be the left endpoint of I I I . In the wave envelope type estimates below, θ \theta θ ranges over Θ ( R − 1 / n ) \Theta(R^{-1/n}) Θ ( R − 1/ n ) for the curve and Ξ ( R − 1 / d ) \Xi(R^{-1/d}) Ξ ( R − 1/ d ) for the cone. In either case, f θ f_\theta f θ denotes the sharp Fourier projection defined by f θ ^ = 1 θ f ^ \widehat{f_\theta}=1_\theta\widehat f f θ = 1 θ f , so that f = ∑ θ f θ f=\sum_\theta f_\theta f = ∑ θ f θ . 1
For R − 1 / n ≤ s ≤ 1 R^{-1/n}\le s\le1 R − 1/ n ≤ s ≤ 1 and τ ∈ Θ ( s ) \tau\in\Theta(s) τ ∈ Θ ( s ) , define the centered box
V τ , R : = { x ∈ R n : ∣ x ⋅ γ n ( j ) ( t τ ) ∣ ≤ R 2 / n s 2 − j for 1 ≤ j ≤ n } . (3) V_{\tau,R}:=\left\{x\in\mathbb{R}^n:|x\cdot\gamma_n^{(j)}(t_\tau)|\le R^{2/n}s^{2-j}\text{ for }1\le j\le n\right\}. \tag*{(3)} V τ , R := { x ∈ R n : ∣ x ⋅ γ n ( j ) ( t τ ) ∣ ≤ R 2/ n s 2 − j for 1 ≤ j ≤ n } . (3)
For R − 1 / d ≤ s ≤ 1 R^{-1/d}\le s\le1 R − 1/ d ≤ s ≤ 1 and τ ∈ Ξ ( s ) \tau\in\Xi(s) τ ∈ Ξ ( s ) , set ϕ d ( 0 ) = ϕ d \phi_d^{(0)}=\phi_d ϕ d ( 0 ) = ϕ d and define
U τ , R : = { x ∈ R d + 1 : ∣ x ⋅ ϕ d ( j ) ( t τ ) ∣ ≤ R 2 / d s 2 − j for 0 ≤ j ≤ d } . (4) U_{\tau,R}:=\left\{x\in\mathbb{R}^{d+1}:|x\cdot\phi_d^{(j)}(t_\tau)|\le R^{2/d}s^{2-j}\text{ for }0\le j\le d\right\}. \tag*{(4)} U τ , R := { x ∈ R d + 1 : ∣ x ⋅ ϕ d ( j ) ( t τ ) ∣ ≤ R 2/ d s 2 − j for 0 ≤ j ≤ d } . (4)
We call these the averaging boxes for the estimates below. For curves with n ≥ 3 n\ge3 n ≥ 3 and cones with d ≥ 3 d\ge3 d ≥ 3 , the averaging boxes at s = 1 s=1 s = 1 need not contain a full wave packet box and hence need not be wave envelopes in the usual sense.
For a spatial box B ⊂ R n B\subset\mathbb{R}^n B ⊂ R n , W B W_B W B denotes a nonnegative smooth weight adapted to B B B , normalized by ∫ W B = ∣ B ∣ \int W_B=|B| ∫ W B = ∣ B ∣ . Write V ∥ V τ , R V\parallel V_{\tau,R} V ∥ V τ , R and U ∥ U τ , R U\parallel U_{\tau,R} U ∥ U τ , R for the translates in fixed tilings of the ambient space. We now state the ordinary wave envelope type estimates.
Proposition 1.5 (Wave envelope type estimate for curves). Let n ≥ 1 n\ge1 n ≥ 1 and p n = n ( n + 1 ) 2 + 1 p_n=\frac{n(n+1)}{2}+1 p n = 2 n ( n + 1 ) + 1 . For any ϵ > 0 \epsilon>0 ϵ > 0 ,
∫ R n ∣ f ∣ p n ≲ ϵ R ϵ ∑ R − 1 / n ≤ s ≤ 1 s d y a d i c ∑ τ ∈ Θ ( s ) ∑ V ∥ V τ , R ∣ V ∣ [ 1 ∣ V ∣ ∫ V ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p n / 4 W V ] 2 . (5) \int_{\mathbb{R}^n}|f|^{p_n}\lesssim_\epsilon R^\epsilon\sum_{\substack{R^{-1/n}\le s\le1\\s\ \mathrm{dyadic}}}\sum_{\tau\in\Theta(s)}\sum_{V\parallel V_{\tau,R}}|V|\left[\frac{1}{|V|}\int_V\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{p_n/4}W_V\right]^2. \tag*{(5)} ∫ R n ∣ f ∣ p n ≲ ϵ R ϵ R − 1/ n ≤ s ≤ 1 s dyadic ∑ τ ∈ Θ ( s ) ∑ V ∥ V τ , R ∑ ∣ V ∣ ∣ V ∣ 1 ∫ V ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p n /4 W V 2 . (5) for every Schwartz function f f f with supp f ^ ⊂ M n ( R ) \operatorname{supp}\widehat f\subset\mathcal{M}^n(R) supp f ⊂ M n ( R ) .
Proposition 1.6 (Wave envelope type estimate for cones over moment curves). Let d ≥ 1 d\ge1 d ≥ 1 and p d = d ( d + 1 ) 2 + 1 p_d=\frac{d(d+1)}{2}+1 p d = 2 d ( d + 1 ) + 1 . For any ϵ > 0 \epsilon>0 ϵ > 0 ,
∫ R d + 1 ∣ f ∣ p d ≲ ϵ R ϵ ∑ R − 1 / d ≤ s ≤ 1 s d y a d i c ∑ τ ∈ Ξ ( s ) ∑ U ∥ U τ , R ∣ U ∣ [ 1 ∣ U ∣ ∫ U ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p d / 4 W U ] 2 . (6) \int_{\mathbb{R}^{d+1}}|f|^{p_d}\lesssim_\epsilon R^\epsilon\sum_{\substack{R^{-1/d}\le s\le1\\s\ \mathrm{dyadic}}}\sum_{\tau\in\Xi(s)}\sum_{U\parallel U_{\tau,R}}|U|\left[\frac{1}{|U|}\int_U\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{p_d/4}W_U\right]^2. \tag*{(6)} ∫ R d + 1 ∣ f ∣ p d ≲ ϵ R ϵ R − 1/ d ≤ s ≤ 1 s dyadic ∑ τ ∈ Ξ ( s ) ∑ U ∥ U τ , R ∑ ∣ U ∣ ∣ U ∣ 1 ∫ U ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p d /4 W U 2 . (6) for every Schwartz function f f f with supp f ^ ⊂ Γ d + 1 ( R ) \operatorname{supp}\widehat f\subset\Gamma^{d+1}(R) supp f ⊂ Γ d + 1 ( R ) .
To answer Question 1.1 for the moment curve, we examined all the examples known to us of functions f f f with Fourier support near the curve, and this suggested the formulation of Proposition 1.5. Since the estimate is different from a wave envelope estimate in the sense that V τ , R V_{\tau,R} V τ , R is comparable to a ball of radius R 2 / n R^{2/n} R 2/ n when s = 1 s=1 s = 1 , it was not so clear if this estimate is useful to the study of small cap decouplings. Interestingly, the amplitude dependent form of this estimate (Theorem 1.7) turned out to be precisely what we needed to prove sharp small cap decoupling for the moment curve.
Proposition 1.5 and the sharp square function estimate for the moment curve [18 , 10 ] are equivalent up to R ϵ R^\epsilon R ϵ losses. To obtain the square function estimate from (1.5), note that Hölder’s inequality bounds the contribution at each scale by a constant times ∫ ( ∑ θ ∣ f θ ∣ 2 ) p n / 2 \int(\sum_{\theta}|f_\theta|^2)^{p_n/2} ∫ ( ∑ θ ∣ f θ ∣ 2 ) p n /2 . For the reverse implication, it suffices to prove
∫ ( ∑ θ ∣ f θ ∣ 2 ) p n / 2 ≲ ϵ R ϵ ∑ R − 1 / n ≤ s ≤ 1 s d y a d i c ∑ τ ∈ Θ ( s ) ∑ V ∥ V τ , R ∣ V ∣ [ 1 ∣ V ∣ ∫ ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p n / 4 W V ] 2 . (7) \int\left(\sum_{\theta}|f_\theta|^2\right)^{p_n/2} \lesssim_\epsilon R^\epsilon\sum_{\substack{R^{-1/n}\le s\le1\\ s\ \mathrm{dyadic}}}\sum_{\tau\in\Theta(s)}\sum_{V\parallel V_{\tau,R}} |V|\left[\frac{1}{|V|}\int\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{p_n/4}W_V\right]^2. \tag*{(7)} ∫ ( θ ∑ ∣ f θ ∣ 2 ) p n /2 ≲ ϵ R ϵ R − 1/ n ≤ s ≤ 1 s dyadic ∑ τ ∈ Θ ( s ) ∑ V ∥ V τ , R ∑ ∣ V ∣ ∣ V ∣ 1 ∫ ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p n /4 W V 2 . (7)
The cases n = 1 , 2 n=1,2 n = 1 , 2 are obvious. For n ≥ 3 n\ge3 n ≥ 3 , apply Lemma 2.8 to h θ = ∣ f θ ∣ 2 h_\theta=|f_\theta|^2 h θ = ∣ f θ ∣ 2 , whose Fourier transform is supported in a fixed dilate of θ − θ \theta-\theta θ − θ . Enlarging the sum to all V ∥ V τ , R V\parallel V_{\tau,R} V ∥ V τ , R and letting α ↓ 0 \alpha\downarrow0 α ↓ 0 gives (7) .
For the application to the small cap decoupling inequality, we need the amplitude dependent form of the estimate. For each dyadic s ∈ [ R − 1 / n , 1 ] s\in[R^{-1/n},1] s ∈ [ R − 1/ n , 1 ] , let
# τ = # { τ ′ ∈ Θ ( s ) : f θ ≢ 0 for some θ ⊂ τ ′ } . \#\tau=\#\{\tau'\in\Theta(s): f_\theta\not\equiv0\text{ for some }\theta\subset\tau'\}. # τ = # { τ ′ ∈ Θ ( s ) : f θ ≡ 0 for some θ ⊂ τ ′ } .
For the cone, use the same definition with Ξ ( s ) \Xi(s) Ξ ( s ) in place of Θ ( s ) \Theta(s) Θ ( s ) for each dyadic s ∈ [ R − 1 / d , 1 ] s\in[R^{-1/d},1] s ∈ [ R − 1/ d , 1 ] . For caps define
G τ M ( α ) = { V ∥ V τ , R : 1 ∣ V ∣ ∫ ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p n / 4 W V ≥ ( α C ϵ R ϵ # τ ) p n / 2 } . (8) \mathcal{G}^{M}_{\tau}(\alpha)=\left\{V\parallel V_{\tau,R}:\frac{1}{|V|}\int\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{p_n/4}W_V\ge\left(\frac{\alpha}{C_\epsilon R^\epsilon\#\tau}\right)^{p_n/2}\right\}. \tag*{(8)} G τ M ( α ) = ⎩ ⎨ ⎧ V ∥ V τ , R : ∣ V ∣ 1 ∫ ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p n /4 W V ≥ ( C ϵ R ϵ # τ α ) p n /2 ⎭ ⎬ ⎫ . (8)
and for caps in R d + 1 \mathbb{R}^{d+1} R d + 1 define
G τ C ( α ) = { U ∥ U τ , R : 1 ∣ U ∣ ∫ ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p d / 4 W U ≥ ( α C ϵ R ϵ # τ ) p d / 2 } . \mathcal{G}^{C}_{\tau}(\alpha)=\left\{U\parallel U_{\tau,R}:\frac{1}{|U|}\int\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{p_d/4}W_U\ge\left(\frac{\alpha}{C_\epsilon R^\epsilon\#\tau}\right)^{p_d/2}\right\}. G τ C ( α ) = ⎩ ⎨ ⎧ U ∥ U τ , R : ∣ U ∣ 1 ∫ ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p d /4 W U ≥ ( C ϵ R ϵ # τ α ) p d /2 ⎭ ⎬ ⎫ .
We now state the amplitude dependent wave envelope type estimates for the moment curve and the cone over the moment curve.
Theorem 1.7. Let n ≥ 1 n\ge1 n ≥ 1 and p n = n ( n + 1 ) 2 + 1 p_n=\frac{n(n+1)}{2}+1 p n = 2 n ( n + 1 ) + 1 . For every ϵ > 0 \epsilon>0 ϵ > 0 and α > 0 \alpha>0 α > 0 ,
∫ { ∣ f ∣ > α } ∣ f ∣ p n ≤ C ϵ R ϵ ∑ R − 1 / n ≤ s ≤ 1 s d y a d i c ∑ τ ∈ Θ ( s ) ∑ V ∈ G τ M ( α ) ∣ V ∣ [ 1 ∣ V ∣ ∫ ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p n / 4 W V ] 2 . (9) \int_{\{|f|>\alpha\}}|f|^{p_n}\le C_\epsilon R^\epsilon\sum_{\substack{R^{-1/n}\le s\le1\\ s\ \mathrm{dyadic}}}\sum_{\tau\in\Theta(s)}\sum_{V\in\mathcal{G}^{M}_{\tau}(\alpha)}|V|\left[\frac{1}{|V|}\int\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{p_n/4}W_V\right]^2. \tag*{(9)} ∫ { ∣ f ∣ > α } ∣ f ∣ p n ≤ C ϵ R ϵ R − 1/ n ≤ s ≤ 1 s dyadic ∑ τ ∈ Θ ( s ) ∑ V ∈ G τ M ( α ) ∑ ∣ V ∣ ∣ V ∣ 1 ∫ ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p n /4 W V 2 . (9) for every Schwartz function f f f with supp f ^ ⊂ M n ( R ) \operatorname{supp}\widehat{f}\subset\mathcal{M}^n(R) supp f ⊂ M n ( R ) .
Theorem 1.8. Let d ≥ 1 d\ge1 d ≥ 1 and p d = d ( d + 1 ) 2 + 1 p_d=\frac{d(d+1)}{2}+1 p d = 2 d ( d + 1 ) + 1 . For every ϵ > 0 \epsilon>0 ϵ > 0 and α > 0 \alpha>0 α > 0 ,
∫ { ∣ f ∣ > α } ∣ f ∣ p d ≤ C ϵ R ϵ ∑ R − 1 / d ≤ s ≤ 1 s d y a d i c ∑ τ ∈ Ξ ( s ) ∑ U ∈ G τ C ( α ) ∣ U ∣ [ 1 ∣ U ∣ ∫ ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p d / 4 W U ] 2 . (10) \int_{\{|f|>\alpha\}}|f|^{p_d}\le C_\epsilon R^\epsilon\sum_{\substack{R^{-1/d}\le s\le1\\ s\ \mathrm{dyadic}}}\sum_{\tau\in\Xi(s)}\sum_{U\in\mathcal{G}^{C}_{\tau}(\alpha)}|U|\left[\frac{1}{|U|}\int\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{p_d/4}W_U\right]^2. \tag*{(10)} ∫ { ∣ f ∣ > α } ∣ f ∣ p d ≤ C ϵ R ϵ R − 1/ d ≤ s ≤ 1 s dyadic ∑ τ ∈ Ξ ( s ) ∑ U ∈ G τ C ( α ) ∑ ∣ U ∣ ∣ U ∣ 1 ∫ ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p d /4 W U 2 . (10) for every Schwartz function f f f with supp f ^ ⊂ Γ d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\Gamma^{d+1}(R) supp f ⊂ Γ d + 1 ( R ) .
The cases n = 1 n=1 n = 1 and d = 1 d=1 d = 1 follow from the locally constant property and L 2 L^2 L 2 orthogonality at the finest scale. The case n = 2 n=2 n = 2 for curves and the case d = 2 d=2 d = 2 for cones follow from [9 Theorems 4 and 2] , respectively. Actually, these theorems estimate α 4 ∣ { ∣ f ∣ > α } ∣ \alpha^4|\{|f|>\alpha\}| α 4 ∣ { ∣ f ∣ > α } ∣ ; the argument in Appendix B replaces this left hand side by ∫ ∣ f ∣ > α ∣ f ∣ 4 \int_{|f|>\alpha}|f|^4 ∫ ∣ f ∣ > α ∣ f ∣ 4 . Since p 2 / 4 = 1 p_2/4=1 p 2 /4 = 1 , the convolutions can be absorbed into the weighted averages over V V V and U U U by (2.26). Theorems 1.7 and 1.8 are new for n ≥ 3 n\ge3 n ≥ 3 and d ≥ 3 d\ge3 d ≥ 3 , respectively.
Proof strategy and organization
We prove the amplitude dependent wave envelope type estimates for moment curves and cones over moment curves in Theorems 1.7 and 1.8 by induction on the ambient dimension. Building on the wave envelope estimate of Guth-Wang-Zhang [14 Theorem 1.3] for the cone in R 3 \mathbb{R}^3 R 3 , Guth-Maldague proved an amplitude dependent wave envelope estimate for this cone [9 Theorem 2] . Together with their amplitude dependent estimate for the parabola in R 2 \mathbb{R}^2 R 2 [9 Theorem 4] , this gives the base cases of the induction, as explained after Theorem 1.8 .
We divide the induction step into two parts. In Part I (Sections 2–7), we use the estimate for the cone over the moment curve in R n \mathbb{R}^n R n to prove the estimate for the moment curve in the same ambient dimension. In Part II (Sections 8–11), we use the estimates for moment curves in R k \mathbb{R}^k R k , 2 ≤ k ≤ d 2\le k\le d 2 ≤ k ≤ d , to prove the estimate for the cone over the moment curve in R d + 1 \mathbb{R}^{d+1} R d + 1 . Starting with the cone in R 3 \mathbb{R}^3 R 3 , we alternate between these two steps to obtain
cone in R 3 ⟹ curve in R 3 ⟹ cone in R 4 ⟹ curve in R 4 ⟹ ⋯ \text{cone in }\mathbb{R}^3 \Longrightarrow\text{curve in }\mathbb{R}^3 \Longrightarrow\text{cone in }\mathbb{R}^4 \Longrightarrow\text{curve in }\mathbb{R}^4 \Longrightarrow\cdots cone in R 3 ⟹ curve in R 3 ⟹ cone in R 4 ⟹ curve in R 4 ⟹ ⋯
§ 2 and § 9 outline the proofs of the two steps, respectively. Readers familiar with wave envelope estimates can follow the main ideas of the proof from these two sections alone.
In Part I, we split the superlevel set { x ∈ R n : ∣ f ( x ) ∣ > α } \{x \in\mathbb{R}^n : |f(x)| > \alpha\} { x ∈ R n : ∣ f ( x ) ∣ > α } into a narrow part, which we estimate by rescaling a single cap, and a broad part, where several separated caps contribute. To estimate the broad part, we use a high low decomposition of ∑ τ ∣ ∑ θ ⊂ τ f θ ∣ 2 ∗ w τ \sum_{\tau}|\sum_{\theta\subset\tau}f_\theta|^2 * w_\tau ∑ τ ∣ ∑ θ ⊂ τ f θ ∣ 2 ∗ w τ and their pruned versions, where w τ w_\tau w τ is an L 1 L^1 L 1 normalized weight adapted to the dual box of τ \tau τ . The low frequency part is controlled by the sum over the averaging boxes at the coarsest angular scale s = 1 s=1 s = 1 , while the high frequency pieces, after grouping and rescaling, have Fourier support in caps of the cone over the moment curve, as in the cubic case [11 Proposition 21] . To estimate these high frequency pieces, we prune wave packets as in [11 Section 2.1] , thereby obtaining improved L ∞ L^\infty L ∞ bounds for the pruned functions f τ j f_\tau^j f τ j . Together with the hypothesis of Proposition 2.1 on the cone over the moment curve, this bounds the broad contribution at the given amplitude by a sum over averaging boxes for the pruned functions. To turn this into a wave envelope type estimate for the original function, we adapt Maldague’s unpruning algorithm [18 proof of Proposition 4.6] , which bounds the sum for the pruned functions by the corresponding sum for the original functions f θ f_\theta f θ . We group caps of width σ \sigma σ inside caps of width s s s . At each step, we either obtain a bound in terms of the original functions f θ f_\theta f θ , pass to a finer scale, or advance the pruning index. Since there are only finitely many scales and pruning stages, the iteration terminates.
Part II uses the estimates for moment curves obtained in this way to prove the estimate for the cone over the moment curve in the next dimension. The three dimensional precedent is the proof of Guth-Maldague [9 Section 8] , who apply the estimate for the parabola to slices of a cone. In higher dimensions we again use a high low decomposition, but the Fourier support of the rescaled high frequency part need not lie in the neighborhood of the cone over the moment curve. Indeed, the radial coefficient can be small while the coefficient of the tangent vector to the generating curve has full size at its scale. Following Guth-Maldague [10 ] , we include Taylor cones in the induction and use their square function estimates for these sets. More precisely, the high frequency part at Taylor order m < d − 2 m<d-2 m < d − 2 lies, after rescaling, in a Taylor cone of order m + 1 m+1 m + 1 . We begin at m = d − 2 m=d-2 m = d − 2 , where L 2 L^2 L 2 orthogonality controls the high frequency part, and descend to m = 0 m=0 m = 0 , the cone over the moment curve. At each order m m m , the estimate for the moment curve in R d − m \mathbb{R}^{d-m} R d − m gives a bound in terms of the sums ∑ θ ⊂ τ f θ \sum_{\theta\subset\tau}f_\theta ∑ θ ⊂ τ f θ over caps τ \tau τ at a coarser angular scale, so we still need to bound these averages in terms of the original functions f θ f_\theta f θ . In their proof for the cone in R 3 \mathbb{R}^3 R 3 , Guth-Maldague [9 Section 7] prune the fine pieces at successive scales, keeping their contributions on envelopes that meet the amplitude condition. In place of this pruning, we exploit the multiscale structure of the wave envelope type estimate itself to argue by induction on the product σ s \sigma s σ s . The idea comes from the unpruning in Part I, where we reduce either σ \sigma σ or s s s to bound the averages of the pruned pieces by those of the original functions f θ f_\theta f θ .
For the cone in R 3 \mathbb{R}^3 R 3 , the wave envelope associated with τ ∈ Ξ ( s ) \tau\in\Xi(s) τ ∈ Ξ ( s ) in the work of Guth-Wang-Zhang [14 ] and Guth-Maldague [9 ] is comparable to the convex hull of the centered wave packet boxes for all θ ∈ Ξ ( R − 1 / 2 ) \theta\in\Xi(R^{-1/2}) θ ∈ Ξ ( R − 1/2 ) with θ ⊂ τ \theta\subset\tau θ ⊂ τ . In higher dimensions, our estimates for moment curves and for cones over moment curves require smaller averaging boxes at coarser angular scales. The difficulty in both parts of the proof is that the estimate for the high frequency part, when brought back to the original coordinates, can give averages over boxes larger than these averaging boxes. To recover the required scales, we start with the low frequency cutoff that gives the average over the coarsest box. For a Taylor cone of order m m m , this cutoff restricts the first m + 2 m+2 m + 2 Taylor coordinates, so we take dyadic differences in those coordinates to isolate the high frequency pieces. We track the spatial widths of the resulting kernels under rescaling and use the locally constant property at those widths. When an averaging box is still too large, we subdivide it at the required scale and select the smaller averaging boxes on which the amplitude condition holds.
In Section 12, we prove sharp small cap decoupling for the moment curve in R n \mathbb{R}^n R n for n ≥ 4 n\geq4 n ≥ 4 . Motivated by Guth-Maldague’s work on the cone in R 3 \mathbb{R}^3 R 3 [9 §1.1] , we use the amplitude dependent wave envelope type estimate to bound the level set { x : ∥ f ∥ > α } \{x: \|f\|>\alpha\} { x : ∥ f ∥ > α } . Roughly speaking, when estimating the integral over each box V V V on the right hand side of (9) , we need only consider points x ∈ V x\in V x ∈ V for which ∑ θ ⊂ τ ∣ f θ ( x ) ∣ 2 > c ε R − ε ( α s ) 2 \sum_{\theta\subset\tau}|f_\theta(x)|^2>c_\varepsilon R^{-\varepsilon}(\alpha s)^2 ∑ θ ⊂ τ ∣ f θ ( x ) ∣ 2 > c ε R − ε ( α s ) 2 , where s s s is the width of τ \tau τ . When this threshold is sufficiently high, we first apply decoupling for the cone over the moment curve in R n − 1 \mathbb{R}^{n-1} R n − 1 and then decoupling for the moment curve in the same dimension. Together these estimates give a stronger bound for the square function than decoupling for the curve alone, and this gain controls the contribution of these scales to the small cap decoupling inequality. We give two further applications of the amplitude dependent wave envelope type estimate for moment curves. In Section 13 we recover the wave envelope estimate of Gan-Maldague-Oh [12 Theorem 2.1] for moment curves in dimensions n ≥ 3 n \ge3 n ≥ 3 , and in Section 14 we derive restriction estimates for the moment curve. Appendix A gives the interpolation for the cone over the moment curve and Taylor cones. Appendix B gives the passage from weak type to strong type for the moment curve.
Notations
Throughout the paper, dimensions and exponents, including β \beta β when it occurs, are fixed. The notation A ≲ B A \lesssim B A ≲ B means A ≤ C B A \le CB A ≤ C B for an absolute constant C C C ; A ∼ B A \sim B A ∼ B means both inequalities hold. A subscript records an additional dependence, and ϵ \epsilon ϵ denotes an arbitrarily small positive loss exponent. We write A ⪅ B A \lessapprox B A ⪅ B if A ≲ ϵ R ϵ B A \lesssim_{\epsilon} R^{\epsilon}B A ≲ ϵ R ϵ B for every ϵ > 0 \epsilon> 0 ϵ > 0 . We also use A ≪ B A \ll B A ≪ B to mean that A A A is sufficiently small relative to B B B , with the required smallness depending on the context.
For a spatial box B ⊂ R n B \subset\mathbb{R}^{n} B ⊂ R n , let ϕ B : R n → R n \phi_B : \mathbb{R}^{n} \to\mathbb{R}^{n} ϕ B : R n → R n be an affine map sending B B B onto [ − 1 / 2 , 1 / 2 ] n [-1/2,1/2]^n [ − 1/2 , 1/2 ] n . Define W B ( x ) = w ( ϕ B ( x ) ) W_B(x) = w(\phi_B(x)) W B ( x ) = w ( ϕ B ( x )) and w B = ∣ B ∣ − 1 W B w_B = |B|^{-1}W_B w B = ∣ B ∣ − 1 W B , where w ≥ 0 w \ge0 w ≥ 0 , ∫ w = 1 \int w = 1 ∫ w = 1 , and w ( z ) ∼ ( 1 + ∣ z ∣ ) − M w(z) \sim(1+|z|)^{-M} w ( z ) ∼ ( 1 + ∣ z ∣ ) − M with M M M sufficiently large. Thus ∫ W B = ∣ B ∣ \int W_B = |B| ∫ W B = ∣ B ∣ and ∫ w B = 1 \int w_B = 1 ∫ w B = 1 . When necessary, we write w B = w B , E w_B = w_{B,E} w B = w B , E to indicate the decay exponent E E E . For a spatial box V ⊂ R n V \subset\mathbb{R}^{n} V ⊂ R n and a function F : R n → C F : \mathbb{R}^{n} \to\mathbb{C} F : R n → C , write
\fint_V F := \frac{1}{|V|}\int_{\mathbb{R}^{n}} F(x)W_V(x)\,dx,\qquad\|F\|_{L_{\#}^{p}(V)} := \left(\fint_V |F|^p\right)^{1/p}.
Acknowledgements
Changkeun Oh was supported by the POSCO Science Fellowship of POSCO TJ Park Foundation, and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) RS-2024-00341891. The second author learned the high-low method through conversations with Larry Guth and Dominique Maldague. He is grateful for his past collaborations with them.
Use of AI
The authors used ChatGPT-6 for discussions during the development of the arguments and for assistance in preparing the manuscript. The authors take full responsibility for the mathematical claims and the final text.
From cones over moment curves to moment curves: Overview of the proof
We prove Theorem 1.7 for the curve in R n \mathbb{R}^{n} R n assuming Theorem 1.8 for the cone over the moment curve in R n \mathbb{R}^{n} R n . In this section, we give overview of the proof.
Proposition 2.1. Let n ≥ 3 n \ge3 n ≥ 3 . Assume the amplitude dependent wave envelope type estimate of Theorem 1.8 for the cone over the moment curve in R n \mathbb{R}^{n} R n , at exponent p n − 1 p_{n-1} p n − 1 . Then the amplitude dependent wave envelope type estimate of Theorem 1.7 for the moment curve in R n \mathbb{R}^{n} R n holds at exponent p n p_n p n .
The proof of Proposition 2.1 is involved and consisted of several steps. Our proof relies on the induction on the dimension n n n , and the base case is an amplitude dependent estimate for the cone in R 3 \mathbb{R}^{3} R 3 , proved in [9 Theorem 2] .
Our goal is to prove the following
∫ { ∣ f ∣ > α } ∣ f ∣ p n ≤ C ϵ R ϵ ∑ R − 1 / n ≤ s ≤ 1 ∑ τ ∈ Θ ( s ) ∑ V ∈ G τ ( α ) ∣ V ∣ [ 1 ∣ V ∣ ∫ R n ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p n / 4 W V ] 2 . (11) \int_{\{|f|>\alpha\}} |f|^{p_n} \le C_{\epsilon}R^{\epsilon}\sum_{R^{-1/n}\le s\le1}\sum_{\tau\in\Theta(s)}\sum_{V\in\mathcal{G}_{\tau}(\alpha)} |V|\left[\frac{1}{|V|}\int_{\mathbb{R}^{n}}\left(\sum_{\theta\subset\tau}|f_{\theta}|^2\right)^{p_n/4}W_V\right]^2. \tag*{(11)} ∫ { ∣ f ∣ > α } ∣ f ∣ p n ≤ C ϵ R ϵ R − 1/ n ≤ s ≤ 1 ∑ τ ∈ Θ ( s ) ∑ V ∈ G τ ( α ) ∑ ∣ V ∣ ∣ V ∣ 1 ∫ R n ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p n /4 W V 2 . (11)
for every Schwartz function f f f with supp f ⊂ M n ( R ) \operatorname{supp} f \subset\mathcal{M}^{n}(\mathbb{R}) supp f ⊂ M n ( R ) . Recall that for τ ∈ Θ ( s ) \tau\in\Theta(s) τ ∈ Θ ( s ) , V τ , R V_{\tau,R} V τ , R is the box centered at the origin whose side lengths in the dual derivative coordinates at τ \tau τ are
R 2 / n s × R 2 / n s × R 2 / n s − 1 × ⋯ × R 2 / n s 2 − n . (12) R^{2/n}s\times R^{2/n}s\times R^{2/n}s^{-1}\times\cdots\times R^{2/n}s^{2-n}. \tag*{(12)} R 2/ n s × R 2/ n s × R 2/ n s − 1 × ⋯ × R 2/ n s 2 − n . (12)
Recall also that
G τ M ( α ) = { V ∥ V τ , R : 1 ∣ V ∣ ∫ R n ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p n / 4 W V ≥ ( α C ϵ R ϵ # T ) p n / 2 } . (13) \mathcal{G}_{\tau}^{\mathcal{M}}(\alpha)=\left\{V\parallel V_{\tau,R}:\frac{1}{|V|}\int_{\mathbb{R}^{n}}\left(\sum_{\theta\subset\tau}|f_{\theta}|^2\right)^{p_n/4}W_V\ge\left(\frac{\alpha}{C_{\epsilon}R^{\epsilon}\#\mathcal{T}}\right)^{p_n/2}\right\}. \tag*{(13)} G τ M ( α ) = ⎩ ⎨ ⎧ V ∥ V τ , R : ∣ V ∣ 1 ∫ R n ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p n /4 W V ≥ ( C ϵ R ϵ # T α ) p n /2 ⎭ ⎬ ⎫ . (13)
For a cap τ \tau τ , write w τ = w τ ∗ = ∣ τ ∗ ∣ − 1 W τ ∗ w_{\tau}=w_{\tau^*}=|\tau^*|^{-1}W_{\tau^*} w τ = w τ ∗ = ∣ τ ∗ ∣ − 1 W τ ∗ , where τ ∗ = τ − τ \tau^*=\tau-\tau τ ∗ = τ − τ is its dual box centred at the origin. To prove our goal, we first estimate the function f f f in terms of weighted averages of ∑ θ ∣ f θ ∣ 2 ∗ w θ \sum_{\theta}|f_{\theta}|^2*w_{\theta} ∑ θ ∣ f θ ∣ 2 ∗ w θ instead of ∑ θ ∣ f θ ∣ 2 \sum_{\theta}|f_{\theta}|^2 ∑ θ ∣ f θ ∣ 2 . Since ∣ f θ ∣ |f_{\theta}| ∣ f θ ∣ is “essentially constant” on a dual set θ \theta θ , naively speaking, ∣ f θ ∣ 2 ∗ w θ |f_{\theta}|^2 * w_{\theta} ∣ f θ ∣ 2 ∗ w θ behaves similarly to w θ w_{\theta} w θ . To this end, we introduce the following definitions.
Definition 2.2. Let 1 ≤ r ≤ R 1 \le r \le R 1 ≤ r ≤ R and R − 1 / n ≤ s ≤ 1 R^{-1/n} \le s \le1 R − 1/ n ≤ s ≤ 1 . For a Schwartz function f : R n → C f : \mathbb{R}^n \to\mathbb{C} f : R n → C with supp f ^ ⊂ M n ( r ) \operatorname{supp}\widehat{f} \subset\mathcal{M}^n(r) supp f ⊂ M n ( r ) , call τ ∈ Θ ( s ) \tau\in\Theta(s) τ ∈ Θ ( s ) active if f θ ≢ 0 f_{\theta} \not\equiv0 f θ ≡ 0 for some θ ⊂ τ \theta\subset\tau θ ⊂ τ , and write
N ( s ) = # { τ ∈ Θ ( s ) : τ is active } , N ρ ( s ) = # { τ ∈ Θ ( s ) : τ ⊂ ρ , τ is active } . N(s) = \#\{\tau\in\Theta(s) : \tau\text{ is active}\}, \qquad N_{\rho}(s) = \#\{\tau\in\Theta(s) : \tau\subset\rho,\ \tau\text{ is active}\}. N ( s ) = # { τ ∈ Θ ( s ) : τ is active } , N ρ ( s ) = # { τ ∈ Θ ( s ) : τ ⊂ ρ , τ is active } . For cones over moment curves, we use the same notation with Ξ ( s ) \Xi(s) Ξ ( s ) in place of Θ ( s ) \Theta(s) Θ ( s ) . The counts are always taken from the finest scale family under consideration, and sums over parents omit inactive caps. Define the set of good averaging boxes corresponding to τ \tau τ with ∣ τ ∣ = s |\tau|=s ∣ τ ∣ = s by
G τ M , w ( α ; f ) : = { V ∥ V τ , r : ∥ ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ∗ w θ ) 1 / 2 ∥ L # p n / 2 ( V ) ≥ α N ( s ) } . \mathcal{G}^{\mathcal{M},w}_{\tau}(\alpha;f):=\left\{V\parallel V_{\tau,r}:\left\|\left(\sum_{\theta\subset\tau}|f_{\theta}|^2*w_{\theta}\right)^{1/2}\right\|_{L_{\#}^{p_n/2}(V)}\ge\frac{\alpha}{N(s)}\right\}. G τ M , w ( α ; f ) := ⎩ ⎨ ⎧ V ∥ V τ , r : ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ∗ w θ ) 1/2 L # p n /2 ( V ) ≥ N ( s ) α ⎭ ⎬ ⎫ . The superscript M \mathcal{M} M denotes the moment curve case, as in (1.8).
Definition 2.3. Define S 1 ( R ) ≥ 1 S_1(R)\ge1 S 1 ( R ) ≥ 1 to be the least constant for which
∫ { ∣ f ∣ > α } ∣ f ∣ p n ≤ S 1 ( R ) p n ∑ R − 1 / n ≤ s ≤ 1 ∑ τ ∈ Θ ( s ) ∑ V ∈ G τ M , w ( α / S 1 ( R ) ; f ) ∣ V ∣ ∥ ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ∗ w θ ) 1 / 2 ∥ L # p n / 2 ( V ) p n (14) \int_{\{|f|>\alpha\}}|f|^{p_n}\le S_1(R)^{p_n}\sum_{R^{-1/n}\le s\le1}\sum_{\tau\in\Theta(s)}\sum_{V\in\mathcal{G}^{\mathcal{M},w}_{\tau}(\alpha/S_1(R);f)}|V|\left\|\left(\sum_{\theta\subset\tau}|f_{\theta}|^2*w_{\theta}\right)^{1/2}\right\|_{L_{\#}^{p_n/2}(V)}^{p_n} \tag*{(14)} ∫ { ∣ f ∣ > α } ∣ f ∣ p n ≤ S 1 ( R ) p n R − 1/ n ≤ s ≤ 1 ∑ τ ∈ Θ ( s ) ∑ V ∈ G τ M , w ( α / S 1 ( R ) ; f ) ∑ ∣ V ∣ ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ∗ w θ ) 1/2 L # p n /2 ( V ) p n (14) holds for every 1 ≤ r ≤ R 1\le r\le R 1 ≤ r ≤ R , every such Schwartz function with supp f ^ ⊂ M n ( r ) \operatorname{supp}\widehat{f}\subset\mathcal{M}^n(r) supp f ⊂ M n ( r ) , and every α > 0 \alpha>0 α > 0 .
Define S 2 ( R ) ≥ 1 S_2(R)\ge1 S 2 ( R ) ≥ 1 analogously, replacing G τ M , w \mathcal{G}^{\mathcal{M},w}_{\tau} G τ M , w with G τ M \mathcal{G}^{\mathcal{M}}_{\tau} G τ M and ∣ f θ ∣ 2 ∗ w θ |f_{\theta}|^2*w_{\theta} ∣ f θ ∣ 2 ∗ w θ with ∣ f θ ∣ 2 |f_{\theta}|^2 ∣ f θ ∣ 2 in (2.4).
Our proof has a lot of similarities to those of [9 , 18 , 10 ] . By the nature of the problems, the definitions and statements of [9 ] are involved. Unfortunately, ours are more involved than theirs as we are combining and developing the tools in [18 , 11 ] . To simplify the notations and follow the key ideas, we use the following convention.
Convention 2.4. When the scale and family are clear, we abbreviate G τ M , w ( α ; f ) \mathcal{G}^{\mathcal{M},w}_{\tau}(\alpha;f) G τ M , w ( α ; f ) to G τ ( α ) \mathcal{G}_{\tau}(\alpha) G τ ( α ) . More generally, in a sum over averaging boxes of the form
∑ B ∥ B 0 B ∈ G τ ( α ) ∣ B ∣ ∥ F τ ∥ L # p n / 2 ( B ) p n / 2 , \sum_{\substack{B\parallel B_0\\B\in\mathcal{G}_{\tau}(\alpha)}}|B|\|F_{\tau}\|_{L_{\#}^{p_n/2}(B)}^{p_n/2}, B ∥ B 0 B ∈ G τ ( α ) ∑ ∣ B ∣∥ F τ ∥ L # p n /2 ( B ) p n /2 , the choice of B B B is always understood to use the same function F τ F_{\tau} F τ as the summand:
G τ ( α ) = { B ∥ B 0 : ∥ F τ ∥ L # p n / 2 ( B ) ≥ α N ( s ) } , ∣ τ ∣ = s . \mathcal{G}_{\tau}(\alpha)=\left\{B\parallel B_0:\|F_{\tau}\|_{L_{\#}^{p_n/2}(B)}\ge\frac{\alpha}{N(s)}\right\},\qquad|\tau|=s. G τ ( α ) = { B ∥ B 0 : ∥ F τ ∥ L # p n /2 ( B ) ≥ N ( s ) α } , ∣ τ ∣ = s . The count N ( s ) N(s) N ( s ) is always computed from the underlying finest scale family specified above.
Our goal is to prove S 2 ( R ) ≲ 1 S_2(R)\lesssim1 S 2 ( R ) ≲ 1 . However, as in [18 Sections 1.2 and 6.3] , we first prove the estimate with these weights and then remove them by applying it at a smaller scale. The following proposition can be proved by following the argument of [18 Proposition 6.8] .
Proposition 2.5. Under the hypothesis of Proposition 2.1 on the cone over the moment curve,
S 2 ( R ) ≲ S 1 ( R 1 / n ) S 2 ( C R 1 − 1 / n ) . (15) S_2(R)\lesssim S_1(R^{1/n})S_2(CR^{1-1/n}). \tag*{(15)} S 2 ( R ) ≲ S 1 ( R 1/ n ) S 2 ( C R 1 − 1/ n ) . (15)
By the proposition, the inequality S 2 ( R ) ≲ 1 S_2(R)\lesssim1 S 2 ( R ) ≲ 1 immediately follows from S 1 ( R ) ≲ 1 S_1(R)\lesssim1 S 1 ( R ) ≲ 1 . So it suffices to prove S 1 ( R ) ≲ 1 S_1(R)\lesssim1 S 1 ( R ) ≲ 1 . The following is the main step of the proof of Proposition 2.1 .
Proposition 2.6. Under the hypothesis of Proposition 2.1 on the cone over the moment curve, the following holds true. Let η > 0 \eta>0 η > 0 be sufficiently small, with N = η − 1 ∈ N N=\eta^{-1}\in\mathbb{N} N = η − 1 ∈ N , let 1 ≤ N 0 < N 1\le N_0<N 1 ≤ N 0 < N , and let 0 < κ ≪ η / ( 1000 n 2 ) 0<\kappa\ll\eta/(1000n^2) 0 < κ ≪ η / ( 1000 n 2 ) . Then
S 1 ( R ) ≲ η max { R C η N 0 , R C η S 1 ( R ′ ) N − N 0 + C , R C η 10 S 1 ( R 1 − η κ ) } . (16) S_1(R)\lesssim_{\eta}\max\left\{R^{C\eta N_0},\ R^{C\eta}S_1(R')^{N-N_0+C},\ R^{C\eta^{10}}S_1(R^{1-\eta\kappa})\right\}. \tag*{(16)} S 1 ( R ) ≲ η max { R C η N 0 , R C η S 1 ( R ′ ) N − N 0 + C , R C η 10 S 1 ( R 1 − η κ ) } . (16)
The above proposition easily implies the following corollary.
Corollary 2.7. Under the hypothesis of Proposition 2.1 on the cone over the moment curve, S 1 ( R ) ≲ 1 S_1(R)\lesssim1 S 1 ( R ) ≲ 1 .
Proposition 2.6 and Proposition 2.5 are proved in Sections 6 and 7, respectively. As mentioned above, the proof of Proposition 2.5 is similar to that for Proposition 6.8 of Maldague [18 ] . So the main task is to prove Proposition 2.6.
Proof of Corollary 2.7. Set d 1 = inf { d ≥ 0 : sup R ≥ 1 S 1 ( R ) R d < ∞ } d_1=\inf\{d\ge0:\sup_{R\ge1}\frac{S_1(R)}{R^d}<\infty\} d 1 = inf { d ≥ 0 : sup R ≥ 1 R d S 1 ( R ) < ∞ } , and suppose that d 1 > 0 d_1>0 d 1 > 0 . Then for every δ > 0 \delta>0 δ > 0 , S 1 ( u ) ≲ δ u d 1 + δ S_1(u)\lesssim_\delta u^{d_1+\delta} S 1 ( u ) ≲ δ u d 1 + δ , u ≥ 1 u\ge1 u ≥ 1 . Put N 0 = η − 1 / 2 N_0=\eta^{-1/2} N 0 = η − 1/2 and κ = η 2 \kappa=\eta^2 κ = η 2 . Substituting into (2.6), we obtain
S 1 ( R ) ≲ η , δ max { R C η , R ( d 1 + δ ) ( 1 − η + C η ) + C η , R ( d 1 + δ ) ( 1 − n η 2 ) + C η 10 } . S_1(R)\lesssim_{\eta,\delta}\max\left\{
R^{C\sqrt{\eta}},
R^{(d_1+\delta)(1-\sqrt{\eta}+C\eta)+C\eta},
R^{(d_1+\delta)(1-n\eta^2)+C\eta^{10}}
\right\}. S 1 ( R ) ≲ η , δ max { R C η , R ( d 1 + δ ) ( 1 − η + C η ) + C η , R ( d 1 + δ ) ( 1 − n η 2 ) + C η 10 } .
Taking η \eta η sufficiently small in terms of d 1 d_1 d 1 , and then δ > 0 \delta>0 δ > 0 sufficiently small in terms of d 1 d_1 d 1 and η \eta η , makes all three exponents strictly less than d 1 d_1 d 1 , contradicting its definition. Hence d 1 = 0 d_1=0 d 1 = 0 , and S 1 ( R ) ≲ 1 S_1(R)\lesssim1 S 1 ( R ) ≲ 1 . □
Roadmap of Proposition 2.6
Throughout Sections 2–7, we assume the hypothesis of Proposition 2.1 on the cone over the moment curve. Here we outline the proof of Proposition 2.6 and explain how it follows from the lemmas and propositions stated below. We give the detailed proof in Sections 3–6. Fix a Schwartz function f : R n → C f:\mathbb{R}^n\to\mathbb{C} f : R n → C with supp f ^ ⊂ M n ( R ) \operatorname{supp}\widehat{f}\subset\mathcal{M}^n(R) supp f ⊂ M n ( R ) , and let α > 0 \alpha>0 α > 0 . By standard reductions, it suffices to bound
α p n ∣ { ∣ f ∣ > α } ∣ \alpha^{p_n}\left|\{|f|>\alpha\}\right| α p n ∣ { ∣ f ∣ > α } ∣
in place of the integral on the left hand side of (2.4), at the cost of a logarithmic factor. We give a proof of the reduction in details in Appendix B for the completeness of the paper. We now describe the estimates needed for this bound. We split the contribution into two cases:
α p n ∣ { ∣ f ∣ > α , ∑ θ ∣ f θ ∣ 2 ∗ w θ > α 2 } ∣ or α p n ∣ { ∣ f ∣ > α , ∑ θ ∣ f θ ∣ 2 ∗ w θ ≤ α 2 } ∣ . \alpha^{p_n}\left|\left\{|f|>\alpha,\ \sum_\theta|f_\theta|^2*w_\theta>\alpha^2\right\}\right|
\quad\text{or}\quad
\alpha^{p_n}\left|\left\{|f|>\alpha,\ \sum_\theta|f_\theta|^2*w_\theta\le\alpha^2\right\}\right|. α p n { ∣ f ∣ > α , θ ∑ ∣ f θ ∣ 2 ∗ w θ > α 2 } or α p n { ∣ f ∣ > α , θ ∑ ∣ f θ ∣ 2 ∗ w θ ≤ α 2 } .
We first treat the first case, where α p n ≤ ( ∑ θ ∣ f θ ∣ 2 ∗ w θ ) p n / 2 \alpha^{p_n}\le\left(\sum_\theta|f_\theta|^2*w_\theta\right)^{p_n/2} α p n ≤ ( ∑ θ ∣ f θ ∣ 2 ∗ w θ ) p n /2 . The high frequency pieces of this sum have Fourier supports in caps after rescaling. Since the hypothesis on the cone is at exponent p n − 1 > p n / 2 p_{n-1}>p_n/2 p n − 1 > p n /2 , we can lower the exponent through interpolation and apply the estimate for the cone over the moment curve in a high low decomposition, without introducing pruning.
Case 1: Bounding α p n ∣ { ∣ f ∣ > α , ∑ θ ∣ f θ ∣ 2 ∗ w θ > α 2 } ∣ \alpha^{p_n}\left|\left\{|f|>\alpha,\ \sum_\theta|f_\theta|^2*w_\theta>\alpha^2\right\}\right| α p n { ∣ f ∣ > α , ∑ θ ∣ f θ ∣ 2 ∗ w θ > α 2 } .
By Lemma 2.8, this case already satisfies the required bound. We prove Lemma 2.8 in Section 3.
Lemma 2.8. Let { h θ } θ ∈ Θ ( R − 1 / n ) \{h_\theta\}_{\theta\in\Theta(R^{-1/n})} { h θ } θ ∈ Θ ( R − 1/ n ) be a family of nonnegative functions in L 1 ( R n ) L^1(\mathbb{R}^n) L 1 ( R n ) satisfying supp h θ ^ ⊂ C ( θ − θ ) \operatorname{supp}\widehat{h_\theta}\subset C(\theta-\theta) supp h θ ⊂ C ( θ − θ ) . For every ϵ > 0 \epsilon>0 ϵ > 0 and α > 0 \alpha>0 α > 0 ,
∫ ∑ θ h θ > α 2 ( ∑ θ h θ ) p n / 2 ≲ R ϵ ∑ R − 1 / n ≤ s ≤ 1 ∑ τ ∈ Θ ( s ) ∑ V ∈ G τ ( R ϵ α ) ∣ V ∣ ∥ ( ∑ θ h θ ) 1 / 2 ∥ L # p n / 2 ( V ) p n . (17) \int_{\sum_\theta h_\theta>\alpha^2}\left(\sum_\theta h_\theta\right)^{p_n/2}
\lesssim R^\epsilon
\sum_{R^{-1/n}\le s\le1}
\sum_{\tau\in\Theta(s)}
\sum_{V\in\mathcal{G}_\tau\left(\frac{R^\epsilon}{\alpha}\right)}
|V|
\left\|
\left(\sum_\theta h_\theta\right)^{1/2}
\right\|_{L_\#^{p_n/2}(V)}^{p_n}. \tag*{(17)} ∫ ∑ θ h θ > α 2 ( θ ∑ h θ ) p n /2 ≲ R ϵ R − 1/ n ≤ s ≤ 1 ∑ τ ∈ Θ ( s ) ∑ V ∈ G τ ( α R ϵ ) ∑ ∣ V ∣ ( θ ∑ h θ ) 1/2 L # p n /2 ( V ) p n . (17)
Here N ( s ) N(s) N ( s ) in the selection condition is computed from the family { h θ } \{h_\theta\} { h θ } .
We use this lemma repeatedly in Sections 4–7.
Let us assume the lemma and finish the proof for the case. Taking h θ = ∣ f θ ∣ 2 ∗ w θ h_\theta=|f_\theta|^2*w_\theta h θ = ∣ f θ ∣ 2 ∗ w θ in Lemma 2.8 gives
α p n ∣ { ∣ f ∣ > α , ∑ θ ∣ f θ ∣ 2 ∗ w θ > α 2 } ∣ ≤ ∫ ∑ θ ∣ f θ ∣ 2 ∗ w θ > α 2 ( ∑ θ ∣ f θ ∣ 2 ∗ w θ ) p n / 2 \alpha^{p_n}\left|\left\{|f|>\alpha,\ \sum_\theta|f_\theta|^2*w_\theta>\alpha^2\right\}\right|
\le
\int_{\sum_\theta|f_\theta|^2*w_\theta>\alpha^2}
\left(\sum_\theta|f_\theta|^2*w_\theta\right)^{p_n/2} α p n { ∣ f ∣ > α , θ ∑ ∣ f θ ∣ 2 ∗ w θ > α 2 } ≤ ∫ ∑ θ ∣ f θ ∣ 2 ∗ w θ > α 2 ( θ ∑ ∣ f θ ∣ 2 ∗ w θ ) p n /2
≲ ϵ R ϵ ∑ R − 1 / n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∈ G ρ ( R ϵ α ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w θ ) 1 / 2 ∥ L # p n / 2 ( V ) p n . (18) \lesssim_\epsilon R^\epsilon
\sum_{R^{-1/n}\le s\le1}
\sum_{\rho\in\Theta(s)}
\sum_{V\in\mathcal{G}_\rho\left(\frac{R^\epsilon}{\alpha}\right)}
|V|
\left\|
\left(\sum_{\theta\subset\rho}|f_\theta|^2*w_\theta\right)^{1/2}
\right\|_{L_\#^{p_n/2}(V)}^{p_n}. \tag*{(18)} ≲ ϵ R ϵ R − 1/ n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∈ G ρ ( α R ϵ ) ∑ ∣ V ∣ θ ⊂ ρ ∑ ∣ f θ ∣ 2 ∗ w θ 1/2 L # p n /2 ( V ) p n . (18)
This finishes the proof.
Case 2: Bounding α p n ∣ { ∣ f ∣ > α , ∑ θ ∣ f θ ∣ 2 ∗ w θ ≤ α 2 } ∣ \alpha^{p_n}\left|\left\{|f|>\alpha,\ \sum_\theta|f_\theta|^2*w_\theta\le\alpha^2\right\}\right| α p n { ∣ f ∣ > α , ∑ θ ∣ f θ ∣ 2 ∗ w θ ≤ α 2 } .
On { ∣ f ∣ > α } \{|f|>\alpha\} { ∣ f ∣ > α } , by Cauchy–Schwarz inequality and the locally constant property, we obtain
α 2 < ∣ f ( x ) ∣ 2 ≲ N ( R − 1 / n ) ∑ θ ∣ f θ ( x ) ∣ 2 ≲ N ( R − 1 / n ) ∑ θ ∣ f θ ∣ 2 ∗ w θ ( x ) . (19) \alpha^2<|f(x)|^2\lesssim N(R^{-1/n})\sum_\theta|f_\theta(x)|^2\lesssim N(R^{-1/n})\sum_\theta|f_\theta|^2*w_\theta(x). \tag*{(19)} α 2 < ∣ f ( x ) ∣ 2 ≲ N ( R − 1/ n ) θ ∑ ∣ f θ ( x ) ∣ 2 ≲ N ( R − 1/ n ) θ ∑ ∣ f θ ∣ 2 ∗ w θ ( x ) . (19)
Hence, in case 2, we have
α 2 N ( R − 1 / n ) ≲ ∑ θ ∣ f θ ∣ 2 ∗ w θ ( x ) ≤ α 2 . \frac{\alpha^2}{N\left(R^{-1/n}\right)} \lesssim\sum_{\theta} |f_\theta|^2 * w_\theta(x) \leq\alpha^2. N ( R − 1/ n ) α 2 ≲ θ ∑ ∣ f θ ∣ 2 ∗ w θ ( x ) ≤ α 2 .
By applying the dyadic pigeonholing to the above inequality, it suffices to bound α p n ∣ E α , β ∣ \alpha^{p_n}|E_{\alpha,\beta}| α p n ∣ E α , β ∣ , where
E α , β = { x : ∣ f ( x ) ∣ > α , β ≤ ( ∑ θ ∣ f θ ∣ 2 ∗ w θ ) ( x ) ≤ 2 β } , α 2 N ( R − 1 / n ) ≲ β ≤ α 2 . (20) E_{\alpha,\beta}=\left\{x: |f(x)|>\alpha,\quad\beta\leq\left(\sum_{\theta}|f_\theta|^2*w_\theta\right)(x)\leq2\beta\right\},\qquad\frac{\alpha^2}{N\left(R^{-1/n}\right)}\lesssim\beta\leq\alpha^2. \tag*{(20)} E α , β = { x : ∣ f ( x ) ∣ > α , β ≤ ( θ ∑ ∣ f θ ∣ 2 ∗ w θ ) ( x ) ≤ 2 β } , N ( R − 1/ n ) α 2 ≲ β ≤ α 2 . (20)
To bound this contribution, we prune the wave packets of f f f , with square root cancellation guiding the choice of threshold. The pruned functions satisfy the improved L ∞ L^\infty L ∞ bound (25) . We then decompose E α , β E_{\alpha,\beta} E α , β into the sets Ω j \Omega_j Ω j and a remaining low set L L L , according to the first scale at which the square function exceeds its threshold as we pass from fine to coarse scales. We describe this construction below.
Pruning wave packets
We first introduce cutoff functions for the wave packet decomposition. Choose φ ∈ C c ∞ ( R n ) \varphi\in C^\infty_c(\mathbb{R}^n) φ ∈ C c ∞ ( R n ) with ∥ φ ∥ 2 = 1 \|\varphi\|_2=1 ∥ φ ∥ 2 = 1 and support in a sufficiently small neighborhood of the origin, and put
ψ 0 ( x ) = ∫ [ − 1 / 2 , 1 / 2 ] n ∣ φ ~ ( x − y ) ∣ 2 d y . \psi_0(x)=\int_{[-1/2,1/2]^n}|\widetilde{\varphi}(x-y)|^2\,dy. ψ 0 ( x ) = ∫ [ − 1/2 , 1/2 ] n ∣ φ ( x − y ) ∣ 2 d y .
For a cap τ ∈ Θ ( s ) \tau\in\Theta(s) τ ∈ Θ ( s ) , let A τ A_\tau A τ be an invertible linear map such that τ ∗ = A τ ( [ − 1 / 2 , 1 / 2 ] n ) \tau^*=A_\tau([-1/2,1/2]^n) τ ∗ = A τ ([ − 1/2 , 1/2 ] n ) . For T ∥ τ ∗ T\parallel\tau^* T ∥ τ ∗ , set
ψ τ , T ( x ) = ψ 0 ( A τ − 1 ( x − c T ) ) . \psi_{\tau,T}(x)=\psi_0\left(A_\tau^{-1}(x-c_T)\right). ψ τ , T ( x ) = ψ 0 ( A τ − 1 ( x − c T ) ) .
Then this function satisfies the following.
∑ T ψ τ , T = 1 , ∑ T ψ τ , T 1 / 2 ≲ 1 , ∫ ψ τ , T 1 / 2 ≲ ∣ T ∣ , supp ψ ^ τ , T ⊂ ( τ − τ ) . \sum_T\psi_{\tau,T}=1,\qquad\sum_T\psi_{\tau,T}^{1/2}\lesssim1,\qquad\int\psi_{\tau,T}^{1/2}\lesssim|T|,\qquad\operatorname{supp}\widehat{\psi}_{\tau,T}\subset(\tau-\tau). T ∑ ψ τ , T = 1 , T ∑ ψ τ , T 1/2 ≲ 1 , ∫ ψ τ , T 1/2 ≲ ∣ T ∣ , supp ψ τ , T ⊂ ( τ − τ ) .
By the locally constant property, we obtain
∥ ψ τ , T 1 / 2 g τ ∥ ∞ r ≲ r ∫ T ∣ g τ ∣ r , r ≥ 1 , supp g τ ^ ⊂ C τ . (21) \|\psi_{\tau,T}^{1/2}g_\tau\|_\infty^r\lesssim_r\int_T|g_\tau|^r,\qquad r\geq1,\quad\operatorname{supp}\widehat{g_\tau}\subset C\tau. \tag*{(21)} ∥ ψ τ , T 1/2 g τ ∥ ∞ r ≲ r ∫ T ∣ g τ ∣ r , r ≥ 1 , supp g τ ⊂ C τ . (21)
Choose η > 0 \eta>0 η > 0 sufficiently small, with N = η − 1 ∈ N N=\eta^{-1}\in\mathbb{N} N = η − 1 ∈ N . We have a sequence of numbers.
R 0 ≤ R η ≤ R 2 η ≤ R 3 η ≤ ⋯ ≤ R N η = R . (22) R^0\leq R^\eta\leq R^{2\eta}\leq R^{3\eta}\leq\cdots\leq R^{N\eta}=R. \tag*{(22)} R 0 ≤ R η ≤ R 2 η ≤ R 3 η ≤ ⋯ ≤ R N η = R . (22)
We use the notations
R j = R j η R_j=R^{j\eta} R j = R j η for 0 ≤ j ≤ N 0\leq j\leq N 0 ≤ j ≤ N
K = R κ K=R^\kappa K = R κ , where 0 < κ ≪ η / ( 1000 n 2 ) 0<\kappa\ll\eta/(1000n^2) 0 < κ ≪ η / ( 1000 n 2 ) , and choose a sufficiently large A = A ( η ) A=A(\eta) A = A ( η ) .
τ j ∈ Θ ( R j − 1 / n ) \tau_j\in\Theta(R_j^{-1/n}) τ j ∈ Θ ( R j − 1/ n ) .
Set f τ N N + 1 = f θ N + 1 = f θ f_{\tau_N}^{N+1}=f_\theta^{N+1}=f_\theta f τ N N + 1 = f θ N + 1 = f θ . For j = N , … , 1 j=N,\ldots,1 j = N , … , 1 , define good planks
T τ j g = { T ∥ τ j ∗ : ∥ ψ τ j , T 1 / 2 f τ j j + 1 ∥ ∞ ≤ K 100 n 2 A N − j + 1 β α } . (23) T_{\tau_j}^g=\left\{T\parallel\tau_j^*: \|\psi_{\tau_j,T}^{1/2}f_{\tau_j}^{j+1}\|_\infty\leq K^{100n^2}A^{N-j+1}\frac{\beta}{\alpha}\right\}. \tag*{(23)} T τ j g = { T ∥ τ j ∗ : ∥ ψ τ j , T 1/2 f τ j j + 1 ∥ ∞ ≤ K 100 n 2 A N − j + 1 α β } . (23)
For each cap τ \tau τ , write f τ = ∑ θ ⊂ τ f θ f_\tau=\sum_{\theta\subset\tau}f_\theta f τ = ∑ θ ⊂ τ f θ . Define the pruned functions by
f τ j j = f τ j j + 1 ∑ T ∈ T τ j g ψ τ j , T , f τ j − 1 j = ∑ τ j ⊂ τ j − 1 f τ j j . (24) f_{\tau_j}^j=f_{\tau_j}^{j+1}\sum_{T\in T_{\tau_j}^g}\psi_{\tau_j,T},\qquad f_{\tau_{j-1}}^j=\sum_{\tau_j\subset\tau_{j-1}}f_{\tau_j}^j. \tag*{(24)} f τ j j = f τ j j + 1 T ∈ T τ j g ∑ ψ τ j , T , f τ j − 1 j = τ j ⊂ τ j − 1 ∑ f τ j j . (24)
For ∣ τ ∣ ≥ R j − 1 / n |\tau|\geq R_j^{-1/n} ∣ τ ∣ ≥ R j − 1/ n , write f τ j = ∑ τ j ⊂ τ f τ j j f_\tau^j=\sum_{\tau_j\subset\tau}f_{\tau_j}^j f τ j = ∑ τ j ⊂ τ f τ j j .
Lemma 2.9. [GM23, Lemma 4.1], [Mal24, Lemma 2.1] For R R R sufficiently large in terms of η \eta η , the pruned functions satisfy
supp f τ j j ^ ⊂ 3 τ j , ∣ f θ N ∣ ≤ ∣ f θ ∣ , ∣ f τ j j ∣ ≤ ∣ f τ j j + 1 ∣ , ∥ f τ j j ∥ ∞ ≲ K 100 n 2 A N − j + 1 β α . (25) \operatorname{supp}\widehat{f_{\tau_j}^j}\subset3\tau_j,\qquad|f_\theta^N|\leq|f_\theta|,\qquad|f_{\tau_j}^j|\leq|f_{\tau_j}^{j+1}|,\qquad\|f_{\tau_j}^j\|_\infty\lesssim K^{100n^2}A^{N-j+1}\frac{\beta}{\alpha}. \tag*{(25)} supp f τ j j ⊂ 3 τ j , ∣ f θ N ∣ ≤ ∣ f θ ∣ , ∣ f τ j j ∣ ≤ ∣ f τ j j + 1 ∣ , ∥ f τ j j ∥ ∞ ≲ K 100 n 2 A N − j + 1 α β . (25)
Choose 1 ≤ N 0 < N 1 \le N_0 < N 1 ≤ N 0 < N , as in Proposition 2.6 , and define for
Ω j = { x ∈ E α , β : ( ∑ τ j ∣ f τ j j + 1 ∣ 2 ∗ w τ j ) ( x ) ≥ A N − j β , ( ∑ τ l ∣ f τ l l + 1 ∣ 2 ∗ w τ l ) ( x ) < A N − l β for all j < l < N } , N 0 ≤ j < N . \Omega_j=\left\{x\in E_{\alpha,\beta}:\left(\sum_{\tau_j}|f_{\tau_j}^{j+1}|^2*w_{\tau_j}\right)(x)\ge A^{N-j}\beta,\quad\left(\sum_{\tau_l}|f_{\tau_l}^{l+1}|^2*w_{\tau_l}\right)(x)<A^{N-l}\beta\ \text{for all }j<l<N\right\},\qquad N_0\le j<N. Ω j = ⎩ ⎨ ⎧ x ∈ E α , β : τ j ∑ ∣ f τ j j + 1 ∣ 2 ∗ w τ j ( x ) ≥ A N − j β , ( τ l ∑ ∣ f τ l l + 1 ∣ 2 ∗ w τ l ) ( x ) < A N − l β for all j < l < N ⎭ ⎬ ⎫ , N 0 ≤ j < N .
L = E α , β ∖ ⋃ j = N 0 N − 1 Ω j . L=E_{\alpha,\beta}\setminus\bigcup_{j=N_0}^{N-1}\Omega_j. L = E α , β ∖ j = N 0 ⋃ N − 1 Ω j .
Let η ∈ C c ∞ ( R n ) \eta\in C^\infty_c(\mathbb{R}^n) η ∈ C c ∞ ( R n ) be a radial bump function with η = 1 \eta=1 η = 1 on B ( 0 , 1 ) B(0,1) B ( 0 , 1 ) and supp η ⊂ B ( 0 , 2 ) \operatorname{supp}\eta\subset B(0,2) supp η ⊂ B ( 0 , 2 ) , and
η ≤ r ( ξ ) = η ( ξ / r ) , η > r = 1 − η ≤ r , η r = η ≤ r − η ≤ r / 2 . \eta_{\le r}(\xi)=\eta(\xi/r),\qquad\eta_{>r}=1-\eta_{\le r},\qquad\eta_r=\eta_{\le r}-\eta_{\le r/2}. η ≤ r ( ξ ) = η ( ξ / r ) , η > r = 1 − η ≤ r , η r = η ≤ r − η ≤ r /2 .
The following estimates, for N 0 ≤ j < N N_0\le j<N N 0 ≤ j < N , give high frequency dominance on Ω j \Omega_j Ω j , control the square function on L L L , and bound the pruning error.
Lemma 2.10. [10 Lemma 4.4, Corollary 4.5, and Lemma 4.6]
∣ ( ∑ τ j ∣ f τ j j + 1 ∣ 2 ∗ w τ j ) ∗ η ~ ≤ R j + 1 − 1 / n ∣ ≲ ∑ τ j + 1 ∣ f τ j + 1 j + 2 ∣ 2 ∗ w τ j + 1 . (26) \left|\left(\sum_{\tau_j}|f_{\tau_j}^{j+1}|^2*w_{\tau_j}\right)*\widetilde{\eta}_{\le R_{j+1}^{-1/n}}\right|\lesssim\sum_{\tau_{j+1}}|f_{\tau_{j+1}}^{j+2}|^2*w_{\tau_{j+1}}. \tag*{(26)} τ j ∑ ∣ f τ j j + 1 ∣ 2 ∗ w τ j ∗ η ≤ R j + 1 − 1/ n ≲ τ j + 1 ∑ ∣ f τ j + 1 j + 2 ∣ 2 ∗ w τ j + 1 . (26) ∑ τ j ∣ f τ j j + 1 ∣ 2 ∗ w τ j ≲ ∣ ( ∑ τ j ∣ f τ j j + 1 ∣ 2 ∗ w τ j ) ∗ η ~ > R j + 1 − 1 / n ∣ on Ω j . \sum_{\tau_j}|f_{\tau_j}^{j+1}|^2*w_{\tau_j}\lesssim\left|\left(\sum_{\tau_j}|f_{\tau_j}^{j+1}|^2*w_{\tau_j}\right)*\widetilde{\eta}_{>R_{j+1}^{-1/n}}\right|\quad\text{on }\Omega_j. τ j ∑ ∣ f τ j j + 1 ∣ 2 ∗ w τ j ≲ τ j ∑ ∣ f τ j j + 1 ∣ 2 ∗ w τ j ∗ η > R j + 1 − 1/ n on Ω j . ∑ τ N 0 ∣ f τ N 0 N 0 + 1 ∣ 2 ∗ w τ N 0 ≲ A N − N 0 β on L . (27) \sum_{\tau_{N_0}}|f_{\tau_{N_0}}^{N_0+1}|^2*w_{\tau_{N_0}}\lesssim A^{N-N_0}\beta\quad\text{on }L. \tag*{(27)} τ N 0 ∑ ∣ f τ N 0 N 0 + 1 ∣ 2 ∗ w τ N 0 ≲ A N − N 0 β on L . (27) ∣ f τ − f τ j + 1 ∣ ≤ α A 1 / 2 K 100 n 2 on Ω j , |f_\tau-f_\tau^{j+1}|\le\frac{\alpha}{A^{1/2}K^{100n^2}}\quad\text{on }\Omega_j, ∣ f τ − f τ j + 1 ∣ ≤ A 1/2 K 100 n 2 α on Ω j , ∣ f τ − f τ N 0 ∣ ≤ α A 1 / 2 K 100 n 2 on L , ∣ τ ∣ ≥ R 1 − 1 / n . (28) |f_\tau-f_\tau^{N_0}|\le\frac{\alpha}{A^{1/2}K^{100n^2}}\quad\text{on }L,\qquad|\tau|\ge R_1^{-1/n}. \tag*{(28)} ∣ f τ − f τ N 0 ∣ ≤ A 1/2 K 100 n 2 α on L , ∣ τ ∣ ≥ R 1 − 1/ n . (28)
Since the proof is already given in [10 Lemma 4.4, Corollary 4.5, and Lemma 4.6] , we only give a sketch of the proof.
Sketch of the proof of Lemma 2.10. We briefly outline the argument in [10 ] . Expand each f τ j j + 1 f_{\tau_j}^{j+1} f τ j j + 1 into its R j + 1 − 1 / n R_{j+1}^{-1/n} R j + 1 − 1/ n children. For τ j + 1 ⊂ τ j \tau_{j+1}\subset\tau_j τ j + 1 ⊂ τ j , we have τ j ∗ ⊂ C τ j + 1 ∗ \tau_j^*\subset C\tau_{j+1}^* τ j ∗ ⊂ C τ j + 1 ∗ , and the shortest side of τ j + 1 ∗ \tau_{j+1}^* τ j + 1 ∗ is comparable to R j + 1 1 / n R_{j+1}^{1/n} R j + 1 1/ n . The rapid decay of η ~ ≤ R j + 1 − 1 / n \widetilde{\eta}_{\le R_{j+1}^{-1/n}} η ≤ R j + 1 − 1/ n outside the ball of this radius therefore gives
w τ j + 1 ∗ w τ j ∗ ∗ ∣ η ~ ≤ R j + 1 − 1 / n ∣ ≲ w τ j + 1 . w_{\tau_{j+1}}*w_{\tau_j^*}*\left|\widetilde{\eta}_{\le R_{j+1}^{-1/n}}\right|\lesssim w_{\tau_{j+1}}. w τ j + 1 ∗ w τ j ∗ ∗ η ≤ R j + 1 − 1/ n ≲ w τ j + 1 .
Together with the locally constant property and ∣ f τ j + 1 j + 1 ∣ ≤ ∣ f τ j + 1 j + 2 ∣ |f_{\tau_{j+1}}^{j+1}|\le|f_{\tau_{j+1}}^{j+2}| ∣ f τ j + 1 j + 1 ∣ ≤ ∣ f τ j + 1 j + 2 ∣ , this proves the first inequality in (2.15). The definitions of Ω j \Omega_j Ω j and L L L , together with ∑ θ ∣ f θ ∣ 2 ∗ w θ ≤ 2 β \sum_\theta|f_\theta|^2*w_\theta\le2\beta ∑ θ ∣ f θ ∣ 2 ∗ w θ ≤ 2 β on E α , β E_{\alpha,\beta} E α , β , then give the second inequality in (2.15) and (2.16) for sufficiently large A A A .
For (2.17), (2.10) with r = 2 r=2 r = 2 and the pruning threshold give, on Ω j \Omega_j Ω j for j < l ≤ N j<l\le N j < l ≤ N , and on L L L for N 0 ≤ l ≤ N N_0\le l\le N N 0 ≤ l ≤ N ,
∑ τ l ⊂ τ ∣ f τ l l + 1 − f τ l l ∣ ≲ α K 100 n 2 A N − l + 1 β ∑ τ l ⊂ τ ∣ f τ l l + 1 ∣ 2 ∗ w τ l ≲ α A K 100 n 2 . \sum_{\tau_l\subset\tau}|f_{\tau_l}^{l+1}-f_{\tau_l}^l|\lesssim\frac{\alpha}{K^{100n^2}A^{N-l+1}\beta}\sum_{\tau_l\subset\tau}|f_{\tau_l}^{l+1}|^2*w_{\tau_l}\lesssim\frac{\alpha}{AK^{100n^2}}. τ l ⊂ τ ∑ ∣ f τ l l + 1 − f τ l l ∣ ≲ K 100 n 2 A N − l + 1 β α τ l ⊂ τ ∑ ∣ f τ l l + 1 ∣ 2 ∗ w τ l ≲ A K 100 n 2 α .
Summing over l = j + 1 , … , N l=j+1,\ldots,N l = j + 1 , … , N on Ω j \Omega_j Ω j , and over l = N 0 , … , N l=N_0,\ldots,N l = N 0 , … , N on L L L , bounds both errors in (2.17) by ≲ N α / ( A K 100 n 2 ) \lesssim N\alpha/(AK^{100n^2}) ≲ N α / ( A K 100 n 2 ) . Taking A A A sufficiently large in terms of N N N proves (2.17).
Roadmap of Proposition 2.6 continued
We return to Case 2 of the roadmap, where we have ∑ θ ∣ f θ ∣ 2 ∗ w θ ≤ α 2 \sum_\theta|f_\theta|^2*w_\theta\le\alpha^2 ∑ θ ∣ f θ ∣ 2 ∗ w θ ≤ α 2 . This subsection is intended to give brief ideas of the proof. So I’ll omit some details here.
We need to bound the set α p n ∣ E α , β ∣ \alpha^{pn}|E_{\alpha,\beta}| α p n ∣ E α , β ∣ . By the standard broad narrow analysis, we may assume that all the points are “broad”. The narrow case can be dealt by using rescaling as always in literature. By the error estimate (2.17), we may replace f τ f_\tau f τ by the pruned functions f τ j + 1 f_\tau^{j+1} f τ j + 1 . On Ω j \Omega_j Ω j , by computations, the square function is dominated by its high frequency part. After dyadic decomposition on the frequency side and making suitable affine changes of variables, the Fourier supports of these pieces lie in a neighborhood of the cone over the moment curve in R n \mathbb{R}^{n} R n . We can therefore apply the hypothesis of Proposition 2.1 on the cone over the moment curve. More precisely, after the standard broad narrow decomposition, (28) allows us to apply this argument to the broad set defined below.
For K − 1 ≥ R 1 − 1 / n K^{-1} \ge R_{1}^{-1/n} K − 1 ≥ R 1 − 1/ n , call a tuple of caps in Θ ( K − 1 ) \Theta(K^{-1}) Θ ( K − 1 ) K − 1 K^{-1} K − 1 -separated if the corresponding intervals in [ 0 , 1 ] [0,1] [ 0 , 1 ] have pairwise distance at least K − 1 K^{-1} K − 1 . Fix a K − 1 K^{-1} K − 1 -separated tuple τ 1 , … , τ n ∈ Θ ( K − 1 ) \tau_{1},\ldots,\tau_{n}\in\Theta(K^{-1}) τ 1 , … , τ n ∈ Θ ( K − 1 ) , and put
Br α , β K = { x ∈ E α , β : max τ ∈ Θ ( K − 1 ) ∣ f τ ( x ) ∣ ≤ α , ∣ f τ 1 ( x ) ⋯ f τ n ( x ) ∣ 1 / n ≥ α K 100 n } . \operatorname{Br}^{K}_{\alpha,\beta}=\left\{x\in E_{\alpha,\beta}:\max_{\tau\in\Theta(K^{-1})}|f_{\tau}(x)|\le\alpha,\ |f_{\tau_{1}}(x)\cdots f_{\tau_{n}}(x)|^{1/n}\ge\frac{\alpha}{K^{100n}}\right\}. Br α , β K = { x ∈ E α , β : τ ∈ Θ ( K − 1 ) max ∣ f τ ( x ) ∣ ≤ α , ∣ f τ 1 ( x ) ⋯ f τ n ( x ) ∣ 1/ n ≥ K 100 n α } .
The hypothesis on the cone in the same ambient dimension gives the following proposition below; see (69) and (70) .
Proposition 2.11. Let 1 ≤ N 0 < N 1\le N_{0}<N 1 ≤ N 0 < N , and let f : R n → C f:\mathbb{R}^{n}\to\mathbb{C} f : R n → C be a Schwartz function with supp f ^ ⊂ M n ( R ) \operatorname{supp}\widehat{f}\subset\mathcal{M}^{n}(\mathbb{R}) supp f ⊂ M n ( R ) . For each N 0 ≤ k < N N_{0}\le k<N N 0 ≤ k < N , there is a dyadic σ \sigma σ with R k + 1 − 1 / n < σ ≤ R k − 1 / n R_{k+1}^{-1/n}<\sigma\le R_{k}^{-1/n} R k + 1 − 1/ n < σ ≤ R k − 1/ n such that
α p n ∣ Br α , β K ∩ Ω k ∣ ≲ η R C η ∑ σ ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ B ∈ G ρ ( R C η α ) ∣ B ∣ = ( s / σ ) V ρ , σ − n ∣ B ∣ ∥ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ ) 1 / 2 ∥ L # p n / 2 ( B ) p n . (29) \alpha^{p_{n}}|\operatorname{Br}^{K}_{\alpha,\beta}\cap\Omega_{k}|\lesssim_{\eta}R^{C\eta}\sum_{\sigma\le s\le1}\sum_{\rho\in\Theta(s)}\sum_{\substack{B\in\mathcal{G}_{\rho}\left(\frac{R^{C\eta}}{\alpha}\right)\\ |B|=(s/\sigma)V_{\rho,\sigma^{-n}}}}|B|\left\|\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_{\tau}|^{2}*w_{\tau}\right)^{1/2}\right\|_{L_{\#}^{p_{n}/2}(B)}^{p_{n}}. \tag*{(29)} α p n ∣ Br α , β K ∩ Ω k ∣ ≲ η R C η σ ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ B ∈ G ρ ( α R C η ) ∣ B ∣ = ( s / σ ) V ρ , σ − n ∑ ∣ B ∣ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ 1/2 L # p n /2 ( B ) p n . (29) \alpha^{p_{n}}|L|\lesssim_{\eta}R^{C\eta N_{0}}\sum_{R^{-1/n}\le s\le1}\sum_{\rho\in\Theta(s)}\sum_{V\in\mathcal{G}_{\rho}\left(\frac{\alpha}{R^{C\eta N_{0}}}\right)}|V|\left|\left(\sum_{\theta\subset\rho}|f_{\theta}|^{2}*w_{\theta}\right)^{1/2}\right|_{L_{#}^{p_{n}/2}(V)}^{p_{n}}.
Here the averaging boxes ( s / σ ) V ρ , σ − n (s/\sigma)V_{\rho,\sigma^{-n}} ( s / σ ) V ρ , σ − n , ∣ ρ ∣ = s |\rho|=s ∣ ρ ∣ = s , are obtained by pulling the averaging boxes U ρ , σ − ( n − 1 ) U_{\rho,\sigma^{-(n-1)}} U ρ , σ − ( n − 1 ) back to the original coordinates. The second inequality is always of the desired form. On the other hand, the right hand side of the first inequality is still far from what we are hoping for. Proposition 2.12 below unprunes the right hand side of (29) , expressing it in terms of the original pieces f θ f_{\theta} f θ .
Proposition 2.12. Let 1 ≤ j ≤ N 1\le j\le N 1 ≤ j ≤ N and R j − 1 / n ≤ σ ≤ R j − 1 − 1 / n R_{j}^{-1/n}\le\sigma\le R_{j-1}^{-1/n} R j − 1/ n ≤ σ ≤ R j − 1 − 1/ n . For s ∈ [ σ , 1 ] s\in[\sigma,1] s ∈ [ σ , 1 ] and α > 0 \alpha>0 α > 0 ,
∑ ρ ∈ Θ ( s ) ∑ B ∈ G ρ ( α ) ∣ B ∣ = ( s / σ ) V ρ , σ − n ∣ B ∣ ∥ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ ) 1 / 2 ∥ L # p n / 2 ( B ) p n ≲ η ( R C η S 1 ( R η ) N − j + C ) p n ∑ R − 1 / n ≤ s ′ ≤ s ∑ ρ ′ ∈ Θ ( s ′ ) ∑ V ∈ G ρ ′ ( α R C η S 1 ( R η ) N − j + C ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ′ ∣ f θ ∣ 2 ∗ w θ ) 1 / 2 ∥ L # p n / 2 ( V ) p n . (30) \sum_{\rho\in\Theta(s)}\sum_{\substack{B\in\mathcal{G}_{\rho}(\alpha)\\ |B|=(s/\sigma)V_{\rho,\sigma^{-n}}}}\lvert B\rvert\left\|\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_{\tau}|^{2}*w_{\tau}\right)^{1/2}\right\|_{L_{\#}^{p_{n}/2}(B)}^{p_{n}}
\lesssim_{\eta}(R^{C\eta}S_{1}(R^{\eta})^{N-j+C})^{p_{n}}
\sum_{R^{-1/n}\le s'\le s}\sum_{\rho'\in\Theta(s')}\sum_{V\in\mathcal{G}_{\rho'}\left(\frac{\alpha}{R^{C\eta}S_{1}(R^{\eta})^{N-j+C}}\right)}|V|\left\|\left(\sum_{\theta\subset\rho'}|f_{\theta}|^{2}*w_{\theta}\right)^{1/2}\right\|_{L_{\#}^{p_{n}/2}(V)}^{p_{n}}. \tag*{(30)} ρ ∈ Θ ( s ) ∑ B ∈ G ρ ( α ) ∣ B ∣ = ( s / σ ) V ρ , σ − n ∑ ∣ B ∣ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ 1/2 L # p n /2 ( B ) p n ≲ η ( R C η S 1 ( R η ) N − j + C ) p n R − 1/ n ≤ s ′ ≤ s ∑ ρ ′ ∈ Θ ( s ′ ) ∑ V ∈ G ρ ′ ( R C η S 1 ( R η ) N − j + C α ) ∑ ∣ V ∣ θ ⊂ ρ ′ ∑ ∣ f θ ∣ 2 ∗ w θ 1/2 L # p n /2 ( V ) p n . (30) The counts $N(s)$ and $N(s')$ in the selection conditions are computed from the fine scale $\{f_{\theta}\}_{\theta\in\Theta(R^{-1/n})}$.
Combining Propositions 2.11 and 2.12 gives (2.20)
α p n ∣ Br α , β K ∣ ≲ η R C η N 0 ∑ R − 1 / n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∈ G ρ ( α R C η N 0 ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w θ ) 1 / 2 ∥ L # p n / 2 ( V ) p n + ( R C η S 1 ( R η ) N − N 0 + C ) p n ∑ R − 1 / n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∈ G ρ ( α R C η S 1 ( R η ) N − N 0 + C ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w θ ) 1 / 2 ∥ L # p n / 2 ( V ) p n . \begin{aligned}
\alpha^{p_{n}}|\operatorname{Br}^{K}_{\alpha,\beta}|\lesssim_{\eta}R^{C\eta N_{0}}\sum_{R^{-1/n}\le s\le1}\sum_{\rho\in\Theta(s)}\sum_{V\in\mathcal{G}_{\rho}\left(\frac{\alpha}{R^{C\eta N_{0}}}\right)}|V|\left\|\left(\sum_{\theta\subset\rho}|f_{\theta}|^{2}*w_{\theta}\right)^{1/2}\right\|_{L_{\#}^{p_{n}/2}(V)}^{p_{n}} \\
&+\left(R^{C\eta}S_{1}(R^{\eta})^{N-N_{0}+C}\right)^{p_{n}}\sum_{R^{-1/n}\le s\le1}\sum_{\rho\in\Theta(s)}\sum_{V\in\mathcal{G}_{\rho}\left(\frac{\alpha}{R^{C\eta}S_{1}(R^{\eta})^{N-N_{0}+C}}\right)}|V|\left\|\left(\sum_{\theta\subset\rho}|f_{\theta}|^{2}*w_{\theta}\right)^{1/2}\right\|_{L_{\#}^{p_{n}/2}(V)}^{p_{n}}.
\end{aligned} α p n ∣ Br α , β K ∣ ≲ η R C η N 0 R − 1/ n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∈ G ρ ( R C η N 0 α ) ∑ ∣ V ∣ θ ⊂ ρ ∑ ∣ f θ ∣ 2 ∗ w θ 1/2 L # p n /2 ( V ) p n + ( R C η S 1 ( R η ) N − N 0 + C ) p n R − 1/ n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∈ G ρ ( R C η S 1 ( R η ) N − N 0 + C α ) ∑ ∣ V ∣ θ ⊂ ρ ∑ ∣ f θ ∣ 2 ∗ w θ 1/2 L # p n /2 ( V ) p n .
Recall Proposition 2.6 :
S 1 ( R ) ≲ η max { R C η N 0 , R C η S 1 ( R η ) N − N 0 + C , R C η 10 S 1 ( R 1 − n κ ) } . (31) S_{1}(R)\lesssim_{\eta}\max\left\{R^{C\eta N_{0}},\ R^{C\eta}S_{1}(R^{\eta})^{N-N_{0}+C},\ R^{C\eta^{10}}S_{1}(R^{1-n\kappa})\right\}. \tag*{(31)} S 1 ( R ) ≲ η max { R C η N 0 , R C η S 1 ( R η ) N − N 0 + C , R C η 10 S 1 ( R 1 − nκ ) } . (31)
The two sums in (2.20) give the first two terms in
max { R C η N 0 , R C η S 1 ( R η ) N − N 0 + C , R C η 10 S 1 ( R 1 − n κ ) } \max\left\{R^{C\eta N_{0}},\ R^{C\eta}S_{1}(R^{\eta})^{N-N_{0}+C},\ R^{C\eta^{10}}S_{1}(R^{1-n\kappa})\right\} max { R C η N 0 , R C η S 1 ( R η ) N − N 0 + C , R C η 10 S 1 ( R 1 − nκ ) }
in (16) , where K = R κ K=R^{\kappa} K = R κ . The last term comes from the narrow contribution by induction, via rescaling. This is an outline of the proof of Proposition 2.6.
We finish this subsection by explaining what we prove in § 4 –§ 7 . § 4 develops the geometric lemmas needed for Proposition 2.12 . § 5 then iterates these estimates to prove Proposition 2.12 . In § 6 , we first prove Proposition 2.11 and then use (2.20) in a broad narrow analysis to prove Proposition 2.6 . This finishes the proof of Proposition 2.6 . § 7 proves Proposition 2.5 , following the strategy of [18 Section 6.3] .
Preliminaries
We record the rescaling maps and auxiliary lemmas used in the subsequent sections.
Rescaling
We begin with the affine rescaling for the moment curve; see [18 proof of Lemma 4.3] . For 0 < s 0 ≤ 1 0 < s_0 \le1 0 < s 0 ≤ 1 and 0 ≤ t 0 ≤ 1 − s 0 0 \le t_0 \le1-s_0 0 ≤ t 0 ≤ 1 − s 0 , set
L t 0 , s 0 = ( s 0 γ n ′ ( t 0 ) , s 0 2 2 ! γ n ′ ′ ( t 0 ) , … , s 0 n n ! γ n ( n ) ( t 0 ) ) . L_{t_0,s_0}=\left(s_0\gamma_n'(t_0),\frac{s_0^2}{2!}\gamma_n''(t_0),\ldots,\frac{s_0^n}{n!}\gamma_n^{(n)}(t_0)\right). L t 0 , s 0 = ( s 0 γ n ′ ( t 0 ) , 2 ! s 0 2 γ n ′′ ( t 0 ) , … , n ! s 0 n γ n ( n ) ( t 0 ) ) .
Taylor’s formula gives
L t 0 , s 0 − 1 ( γ n ( t 0 + s 0 u ) − γ n ( t 0 ) ) = γ n ( u ) , 0 ≤ u ≤ 1. L_{t_0,s_0}^{-1}\left(\gamma_n(t_0+s_0u)-\gamma_n(t_0)\right)=\gamma_n(u),\qquad0\le u\le1. L t 0 , s 0 − 1 ( γ n ( t 0 + s 0 u ) − γ n ( t 0 ) ) = γ n ( u ) , 0 ≤ u ≤ 1.
Thus the affine frequency map ξ ↦ L t 0 , s 0 − 1 ( ξ − γ n ( t 0 ) ) \xi\mapsto L_{t_0,s_0}^{-1}\left(\xi-\gamma_n(t_0)\right) ξ ↦ L t 0 , s 0 − 1 ( ξ − γ n ( t 0 ) ) sends a cap of parameter length σ ≤ s 0 \sigma\le s_0 σ ≤ s 0 , contained in the parent cap over [ t 0 , t 0 + s 0 ] [t_0,t_0+s_0] [ t 0 , t 0 + s 0 ] , to a cap of length σ / s 0 \sigma/s_0 σ / s 0 . In particular, the fine scale R − 1 / n R^{-1/n} R − 1/ n becomes ( s 0 n R ) − 1 / n (s_0^nR)^{-1/n} ( s 0 n R ) − 1/ n . Difference sets θ − θ \theta-\theta θ − θ are transformed by the linear map L t 0 , s 0 − 1 L_{t_0,s_0}^{-1} L t 0 , s 0 − 1 . The corresponding change of spatial variables is x = L t 0 , s 0 − T y x=L_{t_0,s_0}^{-T}y x = L t 0 , s 0 − T y ; returning to the original coordinates multiplies the i i i th dual derivative length by s 0 − i s_0^{-i} s 0 − i .
For the cone generated by γ n ′ \gamma_n' γ n ′ , the linear frequency map is ξ ↦ s 0 L t 0 , s 0 − 1 ξ \xi\mapsto s_0L_{t_0,s_0}^{-1}\xi ξ ↦ s 0 L t 0 , s 0 − 1 ξ , since
s 0 L t 0 , s 0 − 1 γ n ′ ( t 0 + s 0 u ) = γ n ′ ( u ) . s_0L_{t_0,s_0}^{-1}\gamma_n'(t_0+s_0u)=\gamma_n'(u). s 0 L t 0 , s 0 − 1 γ n ′ ( t 0 + s 0 u ) = γ n ′ ( u ) .
This preserves the radial parameter and sends aperture σ \sigma σ to σ / s 0 \sigma/s_0 σ / s 0 . The spatial change is x = s 0 L t 0 , s 0 − T y x=s_0L_{t_0,s_0}^{-T}y x = s 0 L t 0 , s 0 − T y , so the corresponding factors for the dual derivative lengths are s 0 1 − i s_0^{1-i} s 0 1 − i , 1 ≤ i ≤ n 1\le i\le n 1 ≤ i ≤ n .
Auxiliary lemmas
Lemma 2.13. For a measurable F : R n → C F:\mathbb{R}^n\to\mathbb{C} F : R n → C , t > 0 t>0 t > 0 , and a tiling B ∥ B 0 B\parallel B_0 B ∥ B 0 , we have
\sum_{\substack{B\parallel B_0\\ \int_B|F|>t}}|B|\left(\fint_B|F|\right)^2\lesssim\int_{|F|>t/2}|F|^2. \tag*{(32)}
Proof. If ∫ B ∣ F ∣ > t \int_B|F|>t ∫ B ∣ F ∣ > t , then
∫ ∣ F ∣ ≤ t / 2 ∣ F ∣ W B ≤ t 2 ∣ B ∣ < 1 2 ∫ ∣ F ∣ W B . \int_{|F|\le t/2}|F|W_B\le\frac{t}{2}|B|<\frac{1}{2}\int|F|W_B. ∫ ∣ F ∣ ≤ t /2 ∣ F ∣ W B ≤ 2 t ∣ B ∣ < 2 1 ∫ ∣ F ∣ W B . Since ∫ W B = ∣ B ∣ \int W_B=|B| ∫ W B = ∣ B ∣ , Cauchy–Schwarz gives
|B|\left(\fint_B|F|\right)^2\lesssim\frac{1}{|B|}\left(\int_{|F|>t/2}|F|W_B\right)^2\lesssim\int_{|F|>t/2}|F|^2W_B.
Summing over B B B proves (2.22). □ \square □
The first lemma allows us to replace sums over larger averaging boxes by sums over smaller averaging boxes while preserving the amplitude condition.
Lemma 2.14 (Changing averaging boxes). Let F ≥ 0 F\ge0 F ≥ 0 and t > 0 t>0 t > 0 . Fix a tiling by translates B ∥ B 0 B\parallel B_0 B ∥ B 0 of a parallelepiped B 0 B_0 B 0 . If U 0 ⊂ C B 0 U_0\subset CB_0 U 0 ⊂ C B 0 has the same axes as B 0 B_0 B 0 , then
\sum_{\substack{B\parallel B_0:\,\fint_B F>t}}|B|\left(\fint_BF\right)^2\lesssim\sum_{\substack{U\parallel U_0:\,\fint_UF>ct}}|U|\left(\fint_UF\right)^2. \tag*{(33)}
Here c > 0 c>0 c > 0 is fixed independently of F F F , B 0 B_0 B 0 , and t t t .
Proof. Set
G ( x ) = ∑ U ∥ U 0 ( ∫ U F ) 1 U ( x ) . G(x)=\sum_{U\parallel U_0}\left(\int_U F\right)\mathbf{1}_U(x). G ( x ) = U ∥ U 0 ∑ ( ∫ U F ) 1 U ( x ) .
Since U 0 ⊂ C B 0 U_0\subset CB_0 U 0 ⊂ C B 0 has the same axes as B 0 B_0 B 0 , on each U U U we have W B ∼ W B ( c U ) W_B\sim W_B(cU) W B ∼ W B ( c U ) and W U ≳ 1 W_U\gtrsim1 W U ≳ 1 . Hence
∫ B F = 1 ∣ B ∣ ∑ U ∥ U 0 ∫ U F W B ≲ ∑ U ∥ U 0 W B ( c U ) ∣ U ∣ ∣ B ∣ ∫ U F ∼ ∫ B G . \int_B F=\frac{1}{|B|}\sum_{U\parallel U_0}\int_U FW_B\lesssim\sum_{U\parallel U_0}\frac{W_B(cU)|U|}{|B|}\int_U F\sim\int_B G. ∫ B F = ∣ B ∣ 1 U ∥ U 0 ∑ ∫ U F W B ≲ U ∥ U 0 ∑ ∣ B ∣ W B ( c U ) ∣ U ∣ ∫ U F ∼ ∫ B G .
Applying (2.22) to G G G , for a sufficiently small fixed c > 0 c>0 c > 0 , gives
∑ B ∥ B 0 : ∫ B F > t ∣ B ∣ ( ∫ B F ) 2 ≲ ∑ B ∥ B 0 : ∫ B G > 2 c t ∣ B ∣ ( ∫ B G ) 2 \sum_{B\parallel B_0:\int_B F>t}|B|\left(\int_B F\right)^2
\lesssim\sum_{B\parallel B_0:\int_B G>2ct}|B|\left(\int_B G\right)^2 B ∥ B 0 : ∫ B F > t ∑ ∣ B ∣ ( ∫ B F ) 2 ≲ B ∥ B 0 : ∫ B G > 2 c t ∑ ∣ B ∣ ( ∫ B G ) 2
≲ ∫ G > c t G 2 = ∑ U ∥ U 0 : ∫ U F > c t ∣ U ∣ ( ∫ U F ) 2 . \lesssim\int_{G>ct}G^2=\sum_{U\parallel U_0:\int_U F>ct}|U|\left(\int_U F\right)^2. ≲ ∫ G > c t G 2 = U ∥ U 0 : ∫ U F > c t ∑ ∣ U ∣ ( ∫ U F ) 2 .
This proves (2.23). □
Remark 2.15. Let us briefly explain our notations. We use V , U , B V,U,B V , U , B to denote averaging boxes. We write V ∥ V τ , r V\parallel V_{\tau,r} V ∥ V τ , r for the averaging boxes in the tiling for the moment curve and U ∥ U τ , r U\parallel U_{\tau,r} U ∥ U τ , r for those in the tiling for the cone over the moment curve. To estimate the high frequency part, we will apply the estimate for the cone over the moment curve after rescaling. The boxes obtained from U U U on returning to the original coordinates will be denoted by B B B .
The next lemma allows us to remove convolution kernels.
Lemma 2.16. Let F : R n → C F:\mathbb{R}^n\to\mathbb{C} F : R n → C be measurable, t > 0 t>0 t > 0 , r ≥ 1 r\geq1 r ≥ 1 , and let K K K be an integrable kernel with ∥ K ∥ 1 ≲ 1 \|K\|_1\lesssim1 ∥ K ∥ 1 ≲ 1 . Fix a tiling B ∥ B 0 B\parallel B_0 B ∥ B 0 . Then
∫ ∣ K ∗ F ∣ > t ∣ K ∗ F ∣ r ≲ r ∫ ∣ F ∣ > c t ∣ F ∣ r . (34) \int_{\lvert K*F\rvert>t}\lvert K*F\rvert^r\lesssim_r\int_{\lvert F\rvert>ct}\lvert F\rvert^r. \tag*{(34)} ∫ ∣ K ∗ F ∣ > t ∣ K ∗ F ∣ r ≲ r ∫ ∣ F ∣ > c t ∣ F ∣ r . (34) \sum_{B\parallel B_0}\lvert B\rvert\left(\fint_B\lvert K*F\rvert\right)^2_{\fint_B\lvert K*F\rvert>t}\lesssim\sum_{B\parallel B_0}\lvert B\rvert\left(\fint_B\lvert F\rvert\right)^2_{\fint_B\lvert F\rvert>ct}. \tag*{(35)}
If ∣ K ∣ ∗ W B ≲ W B \lvert K\rvert*W_B\lesssim W_B ∣ K ∣ ∗ W B ≲ W B , then
\fint_B\lvert K*F\rvert^r\lesssim_r\fint_B\lvert F\rvert^r. \tag*{(36)}
Proof. Choosing c > 0 c>0 c > 0 sufficiently small, on { ∣ K ∗ F ∣ > t } \{\lvert K*F\rvert>t\} {∣ K ∗ F ∣ > t } ,
∣ K ∗ F ∣ ≲ ∣ K ∣ ∗ ( ∣ F ∣ 1 ∣ F ∣ > c t ) . \lvert K*F\rvert\lesssim\lvert K\rvert*(\lvert F\rvert\mathbf{1}_{\lvert F\rvert>ct}). ∣ K ∗ F ∣ ≲ ∣ K ∣ ∗ (∣ F ∣ 1 ∣ F ∣ > c t ) .
Young’s inequality and ∥ K ∥ 1 ≲ 1 \|K\|_1\lesssim1 ∥ K ∥ 1 ≲ 1 give (2.24).
For y ∈ B y\in B y ∈ B , the definition of the weights gives
W B 0 ( x − y + c B 0 ) ∼ W B ( x ) , ∫ ∣ F ( x ) ∣ w B 0 ( x − y + c B 0 ) d x ∼ ∫ B ∣ F ∣ . W_{B_0}(x-y+cB_0)\sim W_B(x),\qquad\int\lvert F(x)\rvert w_{B_0}(x-y+cB_0)\,dx\sim\int_B\lvert F\rvert. W B 0 ( x − y + c B 0 ) ∼ W B ( x ) , ∫ ∣ F ( x )∣ w B 0 ( x − y + c B 0 ) d x ∼ ∫ B ∣ F ∣ .
The comparison is uniform in x x x , y y y , and B B B . It also holds with ∣ F ∣ \lvert F\rvert ∣ F ∣ replaced by ∣ K ∣ ∗ ∣ F ∣ \lvert K\rvert*\lvert F\rvert ∣ K ∣ ∗ ∣ F ∣ . Since ∣ K ∗ F ∣ ≤ ∣ K ∣ ∗ ∣ F ∣ \lvert K*F\rvert\leq\lvert K\rvert*\lvert F\rvert ∣ K ∗ F ∣ ≤ ∣ K ∣ ∗ ∣ F ∣ , the left side of (2.25) is bounded by
C ∫ ∫ ( ∣ K ∣ ∗ ∣ F ∣ ) ( x ) w B 0 ( x − y + c B 0 ) d x > c t [ ∫ ( ∣ K ∣ ∗ ∣ F ∣ ) ( x ) w B 0 ( x − y + c B 0 ) d x ] 2 d y C\int_{\int(\lvert K\rvert*\lvert F\rvert)(x)w_{B_0}(x-y+cB_0)\,dx>ct}
\left[\int(\lvert K\rvert*\lvert F\rvert)(x)w_{B_0}(x-y+cB_0)\,dx\right]^2dy C ∫ ∫ (∣ K ∣ ∗ ∣ F ∣) ( x ) w B 0 ( x − y + c B 0 ) d x > c t [ ∫ (∣ K ∣ ∗ ∣ F ∣) ( x ) w B 0 ( x − y + c B 0 ) d x ] 2 d y
≲ ∫ ∫ ∣ F ( x ) ∣ w B 0 ( x − y + c B 0 ) d x > c t [ ∫ ∣ F ( x ) ∣ w B 0 ( x − y + c B 0 ) d x ] 2 d y \lesssim\int_{\int\lvert F(x)\rvert w_{B_0}(x-y+cB_0)\,dx>ct}
\left[\int\lvert F(x)\rvert w_{B_0}(x-y+cB_0)\,dx\right]^2dy ≲ ∫ ∫ ∣ F ( x )∣ w B 0 ( x − y + c B 0 ) d x > c t [ ∫ ∣ F ( x )∣ w B 0 ( x − y + c B 0 ) d x ] 2 d y
\lesssim\sum_{\fint_B\lvert F\rvert>ct}\lvert B\rvert\left(\fint_B\lvert F\rvert\right)^2.
The middle inequality is (34) with r = 2 r=2 r = 2 , since convolution with ∣ K ∣ |K| ∣ K ∣ commutes with the displayed averaging integral. The last inequality follows by integrating over the tiles B B B . The fixed constant c c c may decrease in these inequalities.
Finally, Hölder’s inequality gives
∣ K ∗ F ∣ r r ≤ ∥ K ∥ r r − 1 ( ∣ K ∣ ∗ ∣ F ∣ r ) ≲ r ∣ K ∣ ∗ ∣ F ∣ r . \lvert K * F\rvert_r^r \leq\lVert K\rVert_r^{r-1}(\lvert K\rvert* \lvert F\rvert^r) \lesssim_r \lvert K\rvert* \lvert F\rvert^r. ∣ K ∗ F ∣ r r ≤ ∥ K ∥ r r − 1 (∣ K ∣ ∗ ∣ F ∣ r ) ≲ r ∣ K ∣ ∗ ∣ F ∣ r .
By Fubini’s theorem, using W B ( 2 c B − x ) ∼ W B ( x ) W_B(2c_B-x)\sim W_B(x) W B ( 2 c B − x ) ∼ W B ( x ) in the last inequality,
∫ B ∣ K ∗ F ∣ r ≲ r 1 ∣ B ∣ ∫ B ( ∣ K ∣ ∗ ∣ F ∣ r ) W B = 1 ∣ B ∣ ∫ ∣ F ( x ) ∣ r ( ∫ ∣ K ( y ) ∣ W B ( x + y ) d y ) d x ≲ r ∫ B ∣ F ∣ r . \int_B \lvert K*F\rvert^r \lesssim_r \frac{1}{\lvert B\rvert}\int_B(\lvert K\rvert*\lvert F\rvert^r)W_B=\frac{1}{\lvert B\rvert}\int\lvert F(x)\rvert^r\left(\int\lvert K(y)\rvert W_B(x+y)\,dy\right)dx\lesssim_r\int_B\lvert F\rvert^r. ∫ B ∣ K ∗ F ∣ r ≲ r ∣ B ∣ 1 ∫ B (∣ K ∣ ∗ ∣ F ∣ r ) W B = ∣ B ∣ 1 ∫ ∣ F ( x ) ∣ r ( ∫ ∣ K ( y )∣ W B ( x + y ) d y ) d x ≲ r ∫ B ∣ F ∣ r .
This finishes the proof. □
Lemma 2.17. Let F : R n → C F:\mathbb{R}^n\to\mathbb{C} F : R n → C be measurable, p ≥ 2 p\geq2 p ≥ 2 , and let Q ∥ Q 0 Q\parallel Q_0 Q ∥ Q 0 tile R n \mathbb{R}^n R n by translates of a fixed parallelepiped. Suppose that, uniformly in Q Q Q ,
∣ F ( x ) ∣ ≲ ∥ F ∥ L # p / 2 ( Q ) for a.e. x ∈ Q . \lvert F(x)\rvert\lesssim\lVert F\rVert_{L_{\#}^{p/2}(Q)}\qquad\text{for a.e. }x\in Q. ∣ F ( x )∣ ≲ ∥ F ∥ L # p /2 ( Q ) for a.e. x ∈ Q . Then, for every a > 0 a>0 a > 0 ,
∫ ∣ F ∣ > C a ∣ F ∣ p ≤ ∑ Q ∥ Q 0 ∥ F ∥ L # p / 2 ( Q ) ≥ a ∣ Q ∣ ∥ F ∥ L # p / 2 ( Q ) p ≲ ∫ ∣ F ∣ > c a ∣ F ∣ p . (37) \int_{\lvert F\rvert>Ca}\lvert F\rvert^p\leq\sum_{\substack{Q\parallel Q_0\\ \lVert F\rVert_{L_{\#}^{p/2}(Q)}\geq a}}\lvert Q\rvert\lVert F\rVert_{L_{\#}^{p/2}(Q)}^p\lesssim\int_{\lvert F\rvert>ca}\lvert F\rvert^p. \tag*{(37)} ∫ ∣ F ∣ > C a ∣ F ∣ p ≤ Q ∥ Q 0 ∥ F ∥ L # p /2 ( Q ) ≥ a ∑ ∣ Q ∣ ∥ F ∥ L # p /2 ( Q ) p ≲ ∫ ∣ F ∣ > c a ∣ F ∣ p . (37) for some constants C > 1 C>1 C > 1 and 0 < c < 1 0<c<1 0 < c < 1 .
Proof. The second inequality follows from (32) applied to ∣ F ∣ p / 2 \lvert F\rvert^{p/2} ∣ F ∣ p /2 . For the first, integrate the pointwise bound over Q ∩ { ∣ F ∣ > C a } Q\cap\{\lvert F\rvert>Ca\} Q ∩ {∣ F ∣ > C a } and sum over Q Q Q ; this set is null unless ∥ F ∥ L # p / 2 ( Q ) ≥ a \lVert F\rVert_{L_{\#}^{p/2}(Q)}\geq a ∥ F ∥ L # p /2 ( Q ) ≥ a . □
Finally, in proving Propositions 2.5 and 2.6 , we may assume the counting inequalities in the following uniformization lemma. Recall Definition 2.2 .
Lemma 2.18 (Uniformization). A subcollection of { f θ } θ ∈ Θ ( R − 1 / n ) \{f_\theta\}_{\theta\in\Theta(R^{-1/n})} { f θ } θ ∈ Θ ( R − 1/ n ) can be partitioned into O ϵ ( R ϵ ) O_\epsilon(R^\epsilon) O ϵ ( R ϵ ) classes such that, in each class,
N ( s ) N τ ( σ ) ≲ ϵ N ( σ ) , τ ∈ Θ ( s ) , R − 1 / n ≤ σ ≤ s ≤ 1. (38) N(s)N_\tau(\sigma)\lesssim_\epsilon N(\sigma),\qquad\tau\in\Theta(s),\qquad R^{-1/n}\leq\sigma\leq s\leq1. \tag*{(38)} N ( s ) N τ ( σ ) ≲ ϵ N ( σ ) , τ ∈ Θ ( s ) , R − 1/ n ≤ σ ≤ s ≤ 1. (38) The same inequality holds hold inside any dyadic parent ρ \rho ρ , with N ρ ( s ) N_\rho(s) N ρ ( s ) and N ρ ( σ ) N_\rho(\sigma) N ρ ( σ ) in place of N ( s ) N(s) N ( s ) and N ( σ ) N(\sigma) N ( σ ) , for τ ⊂ ρ \tau\subset\rho τ ⊂ ρ and σ ≤ s ≤ ∣ ρ ∣ \sigma\leq s\leq\lvert\rho\rvert σ ≤ s ≤ ∣ ρ ∣ .
Proof. Let A = { θ ∈ Θ ( R − 1 / n ) : f θ ≢ 0 } \mathcal{A}=\{\theta\in\Theta(R^{-1/n}):f_\theta\not\equiv0\} A = { θ ∈ Θ ( R − 1/ n ) : f θ ≡ 0 } and choose an integer b b b such that log 2 ( b + 1 ) / b < ϵ \log_2(b+1)/b<\epsilon log 2 ( b + 1 ) / b < ϵ . Starting from scale 1 1 1 , take every b b b th dyadic scale and include the finest scale, obtaining the block scales
1 = s 0 > s 1 > ⋯ > s L = R − 1 / n , L = ⌈ log 2 R n b ⌉ . 1=s_0>s_1>\cdots>s_L=R^{-1/n},\qquad L=\left\lceil\frac{\log_2 R}{nb}\right\rceil. 1 = s 0 > s 1 > ⋯ > s L = R − 1/ n , L = ⌈ nb log 2 R ⌉ . Each cap at scale s j s_j s j has at most 2 b 2^b 2 b children at scale s j + 1 s_{j+1} s j + 1 . First consider the scales s L − 1 , s L s_{L-1},s_L s L − 1 , s L . For each τ ∈ Θ ( s L − 1 ) \tau\in\Theta(s_{L-1}) τ ∈ Θ ( s L − 1 ) , write
# { θ ∈ A : θ ⊂ τ } = ∑ k 1 ∈ E τ 2 k 1 , E τ ⊂ { 0 , … , b } . \#\{\theta\in\mathcal{A}:\theta\subset\tau\}=\sum_{k_1\in E_\tau}2^{k_1},\qquad E_\tau\subset\{0,\ldots,b\}. # { θ ∈ A : θ ⊂ τ } = k 1 ∈ E τ ∑ 2 k 1 , E τ ⊂ { 0 , … , b } . Partition these caps into sets C τ , k 1 C_{\tau,k_1} C τ , k 1 of size 2 k 1 2^{k_1} 2 k 1 , and put
A k 1 = ⋃ τ : k 1 ∈ E τ C τ , k 1 . \mathcal{A}_{k_1}=\bigcup_{\tau:k_1\in E_\tau}C_{\tau,k_1}. A k 1 = τ : k 1 ∈ E τ ⋃ C τ , k 1 . The nonempty sets A k 1 \mathcal{A}_{k_1} A k 1 partition A \mathcal{A} A . In each A k 1 \mathcal{A}_{k_1} A k 1 , every active s L − 1 s_{L-1} s L − 1 cap contains exactly 2 k 1 2^{k_1} 2 k 1 caps at scale s L = R − 1 / n s_L=R^{-1/n} s L = R − 1/ n .
Proceed upwards one block at a time. After r r r steps, fix a class A k 1 , … , k r \mathcal{A}_{k_1,\ldots,k_r} A k 1 , … , k r . For each parent τ ∈ Θ ( s L − r − 1 ) \tau\in\Theta(s_{L-r-1}) τ ∈ Θ ( s L − r − 1 ) , count its children ν ∈ Θ ( s L − r ) \nu\in\Theta(s_{L-r}) ν ∈ Θ ( s L − r ) containing a member of this class. Using the binary expansion of this count, partition these children into groups C τ , k r + 1 C_{\tau,k_{r+1}} C τ , k r + 1 of size 2 k r + 1 2^{k_{r+1}} 2 k r + 1 , with k r + 1 ∈ E τ ⊂ { 0 , … , b } k_{r+1}\in E_\tau\subset\{0,\ldots,b\} k r + 1 ∈ E τ ⊂ { 0 , … , b } , and define
A k 1 , … , k r , k r + 1 = ⋃ τ : k r + 1 ∈ E τ ⋃ ν ∈ C τ , k r + 1 { θ ∈ A k 1 , … , k r : θ ⊂ ν } . \mathcal{A}_{k_1,\ldots,k_r,k_{r+1}}
=\bigcup_{\tau:k_{r+1}\in E_\tau}\ \bigcup_{\nu\in C_{\tau,k_{r+1}}}
\{\theta\in\mathcal{A}_{k_1,\ldots,k_r}:\theta\subset\nu\}. A k 1 , … , k r , k r + 1 = τ : k r + 1 ∈ E τ ⋃ ν ∈ C τ , k r + 1 ⋃ { θ ∈ A k 1 , … , k r : θ ⊂ ν } .
In each new class, every active s L − r − 1 s_{L-r-1} s L − r − 1 cap has exactly 2 k r + 1 2^{k_{r+1}} 2 k r + 1 active s L − r s_{L-r} s L − r children. Each selected child keeps all its fine descendants in A k 1 , … , k r A_{k_1,\ldots,k_r} A k 1 , … , k r , so the branching numbers already fixed at finer block scales are preserved. Each step splits a class into at most b + 1 b+1 b + 1 classes. Hence the number of final classes is at most
( b + 1 ) L = ( b + 1 ) ⌈ log 2 R / ( n b ) ⌉ ≲ ϵ R ϵ . (b+1)^L=(b+1)^{\lceil\log_2 R/(nb)\rceil}\lesssim_\epsilon R^\epsilon. ( b + 1 ) L = ( b + 1 ) ⌈ l o g 2 R / ( nb )⌉ ≲ ϵ R ϵ .
In each final class, all active parents at a block scale have the same number of active children at the next block scale. Multiplying these branching numbers gives N ( s ) N τ ( σ ) = N ( σ ) N(s)N_\tau(\sigma)=N(\sigma) N ( s ) N τ ( σ ) = N ( σ ) at block scales. Rounding s , σ s,\sigma s , σ to adjacent block scales changes the counts by a factor depending only on b b b , which proves (2.28). Applying the same argument inside ρ \rho ρ gives the corresponding counting inequalities. □
In proving the bounds for S 1 S_1 S 1 and S 2 S_2 S 2 , we assume from now on that the original fine family { f θ } \{f_\theta\} { f θ } satisfies the counting inequalities of Lemma 2.18. To justify this reduction, write f = ∑ i = 1 J f ( i ) f=\sum_{i=1}^J f^{(i)} f = ∑ i = 1 J f ( i ) , where each f ( i ) f^{(i)} f ( i ) is the sum over one of these classes. We have
∫ { ∣ f ∣ > α } ∣ f ∣ p n ≤ J p n ∑ i = 1 J ∫ { ∣ f ( i ) ∣ > α / J } ∣ f ( i ) ∣ p n . \int_{\{|f|>\alpha\}} |f|^{p_n}\leq J^{p_n}\sum_{i=1}^J\int_{\{|f^{(i)}|>\alpha/J\}} |f^{(i)}|^{p_n}. ∫ { ∣ f ∣ > α } ∣ f ∣ p n ≤ J p n i = 1 ∑ J ∫ { ∣ f ( i ) ∣ > α / J } ∣ f ( i ) ∣ p n .
and for every class A \mathcal A A ,
N A ( s ) ≤ N ( s ) , ∑ θ ⊂ τ : θ ∈ A ∣ f θ ∣ 2 ∗ w θ ≤ ∑ θ ⊂ τ ∣ f θ ∣ 2 ∗ w θ . N_{\mathcal A}(s)\leq N(s),\qquad
\sum_{\theta\subset\tau:\theta\in\mathcal A}|f_\theta|^2*w_\theta\leq\sum_{\theta\subset\tau}|f_\theta|^2*w_\theta. N A ( s ) ≤ N ( s ) , θ ⊂ τ : θ ∈ A ∑ ∣ f θ ∣ 2 ∗ w θ ≤ θ ⊂ τ ∑ ∣ f θ ∣ 2 ∗ w θ .
The same inequality holds without the convolutions. Hence the boxes entering the sums for a class at amplitude α / J \alpha/J α / J also enter those for the full family after multiplying the selection constant by J J J . Taking the uniformization parameter sufficiently small absorbs these factors into R ϵ R^\epsilon R ϵ .
From cones over moment curves to moment curves: Proof of Lemma 2.8
In this section, we prove Lemma 2.8. Here is what we need to prove: for any α > 0 \alpha>0 α > 0
∫ ∑ θ h θ > α 2 ( ∑ θ h θ ) p n / 2 ≲ ϵ R ϵ ∑ R − 1 / n ≤ s ≤ 1 ∑ τ ∈ Θ ( s ) ∑ V ∈ G τ ( R ϵ ) ∣ V ∣ ∥ ( ∑ θ ⊂ τ h θ ) 1 / 2 ∥ L # p n / 2 ( V ) p n (39) \int_{\sum_\theta h_\theta>\alpha^2}\left(\sum_\theta h_\theta\right)^{p_n/2}
\lesssim_\epsilon R^\epsilon
\sum_{R^{-1/n}\leq s\leq1}\sum_{\tau\in\Theta(s)}\sum_{V\in\mathcal G_\tau\left(R^\epsilon\right)}
|V|\left\|\left(\sum_{\theta\subset\tau}h_\theta\right)^{1/2}\right\|_{L_\#^{p_n/2}(V)}^{p_n} \tag*{(39)} ∫ ∑ θ h θ > α 2 ( θ ∑ h θ ) p n /2 ≲ ϵ R ϵ R − 1/ n ≤ s ≤ 1 ∑ τ ∈ Θ ( s ) ∑ V ∈ G τ ( R ϵ ) ∑ ∣ V ∣ ( θ ⊂ τ ∑ h θ ) 1/2 L # p n /2 ( V ) p n (39)
for any nonnegative function { h θ } θ ∈ Θ ( R − 1 / n ) \{h_\theta\}_{\theta\in\Theta(R^{-1/n})} { h θ } θ ∈ Θ ( R − 1/ n ) satisfying supp h ^ θ ⊂ C ( θ − θ ) \operatorname{supp}\widehat h_\theta\subset C(\theta-\theta) supp h θ ⊂ C ( θ − θ ) . Recall that we are assuming Theorem 1.8 for the cone over the moment curve in R n \mathbb R^n R n at exponent p n − 1 p_{n-1} p n − 1 .
Since the high low analysis applies it to quadratic expressions in the Fourier projections f θ f_\theta f θ , we need the following estimate at exponent p n / 2 p_n/2 p n /2 , as in [18 Proposition 4.4] . A fixed affine map takes ϕ n − 1 \phi_{n-1} ϕ n − 1 to γ n ′ \gamma'_n γ n ′ . This change of normalization affects only implicit constants and introduces no additional difficulty in adapting Maldague’s argument under our hypothesis on the cone. After the corresponding change of spatial variables, U ρ , σ − ( n − 1 ) U_{\rho,\sigma^{-(n-1)}} U ρ , σ − ( n − 1 ) , for ρ ∈ Ξ ( s ) \rho\in\Xi(s) ρ ∈ Ξ ( s ) , agrees up to a fixed dilation with
{ x ∈ R n : ∣ x ⋅ γ n ( i ) ( t ρ ) ∣ ≤ σ − 2 s 3 − i for 1 ≤ i ≤ n } . \{x\in\mathbb R^n:|x\cdot\gamma_n^{(i)}(t_\rho)|\leq\sigma^{-2}s^{3-i}\text{ for }1\leq i\leq n\}. { x ∈ R n : ∣ x ⋅ γ n ( i ) ( t ρ ) ∣ ≤ σ − 2 s 3 − i for 1 ≤ i ≤ n } .
For the cone families on Γ n \Gamma_n Γ n , the notation for good averaging boxes G ρ \mathcal G_\rho G ρ follows the convention 2.4. The following estimate is very similar to (1.10). The main difference is that we raise p n / 2 p_n/2 p n /2 th power than p n − 1 p_{n-1} p n − 1 (note that we are using d = n − 1 d=n-1 d = n − 1 as hypothesis). Since 2 ≤ p n / 2 ≤ p n − 1 2\leq p_n/2\leq p_{n-1} 2 ≤ p n /2 ≤ p n − 1 , by interpolating the exponents p n − 1 p_{n-1} p n − 1 and 2, one may expect that the estimate is true for p n / 2 p_n/2 p n /2 as well.
Lemma 3.1. Assume Theorem 1.8 for the cone over the moment curve in R n \mathbb R^n R n . Let α > 0 \alpha>0 α > 0 . We have
∫ ∣ g ∣ > α 2 ∣ g ∣ p n / 2 ≲ ϵ σ − ϵ ∑ σ ≤ s ≤ 1 ∑ ρ ∈ Ξ ( s ) ∑ U ∈ G ρ ( c σ − ϵ C σ − ϵ ) ∣ U ∣ ∥ ( ∑ τ ⊂ ρ ∣ g τ ∣ ) 1 / 2 ∥ L # p n / 2 ( U ) p n \int_{|g|>\alpha^2}|g|^{p_n/2}\lesssim_\epsilon\sigma^{-\epsilon}\sum_{\sigma\leq s\leq1}\sum_{\rho\in\Xi(s)}\sum_{U\in\mathcal G_\rho\left(\frac{c\sigma^{-\epsilon}}{C_{\sigma-\epsilon}}\right)}|U|\left\|\left(\sum_{\tau\subset\rho}|g_\tau|\right)^{1/2}\right\|_{L_\#^{p_n/2}(U)}^{p_n} ∫ ∣ g ∣ > α 2 ∣ g ∣ p n /2 ≲ ϵ σ − ϵ σ ≤ s ≤ 1 ∑ ρ ∈ Ξ ( s ) ∑ U ∈ G ρ ( C σ − ϵ c σ − ϵ ) ∑ ∣ U ∣ ( τ ⊂ ρ ∑ ∣ g τ ∣ ) 1/2 L # p n /2 ( U ) p n for a Schwartz function g g g with supp g ^ ⊂ Γ n ( σ − ( n − 1 ) ) \operatorname{supp}\widehat g\subset\Gamma^n(\sigma^{-(n-1)}) supp g ⊂ Γ n ( σ − ( n − 1 ) ) . Here N ( s ) N(s) N ( s ) counts the active parents of { g τ } τ ∈ Ξ ( σ ) \{g_\tau\}_{\tau\in\Xi(\sigma)} { g τ } τ ∈ Ξ ( σ ) .
Since the proof of the lemma is rather standard, we postpone the proof to Appendix A.1. Let us begin the proof of Lemma 2.8.
Proof of Lemma 2.8. By standard approximation, we may assume that the h θ h_\theta h θ are Schwartz functions. Using the ball cutoffs from § 2.2 , dyadic pigeonholing, we have one of two cases:
∫ ∑ θ h θ > α 2 ( ∑ θ h θ ) p n / 2 ≲ ∫ ∣ ( ∑ θ h θ ) ∗ η ˇ λ ∣ ≳ α 2 ∣ ( ∑ θ h θ ) ∗ η ˇ λ ∣ p n / 2 (40) \int_{\sum_\theta h_\theta>\alpha^2}\left(\sum_\theta h_\theta\right)^{p_n/2}\lesssim\int_{\left|(\sum_\theta h_\theta)*\check{\eta}_\lambda\right|\gtrsim\alpha^2}\left|(\sum_\theta h_\theta)*\check{\eta}_\lambda\right|^{p_n/2} \tag*{(40)} ∫ ∑ θ h θ > α 2 ( θ ∑ h θ ) p n /2 ≲ ∫ ∣ ( ∑ θ h θ ) ∗ η ˇ λ ∣ ≳ α 2 ( θ ∑ h θ ) ∗ η ˇ λ p n /2 (40)
for some R − 2 / n ≪ λ < R − 1 / n R^{-2/n}\ll\lambda<R^{-1/n} R − 2/ n ≪ λ < R − 1/ n , or
∫ ∑ θ h θ > α 2 ( ∑ θ h θ ) p n / 2 ≲ ∫ ∣ ( ∑ θ h θ ) ∗ η ˇ ≤ R − 2 / n ∣ ≳ α 2 ∣ ( ∑ θ h θ ) ∗ η ˇ ≤ R − 2 / n ∣ p n / 2 . (41) \int_{\sum_\theta h_\theta>\alpha^2}\left(\sum_\theta h_\theta\right)^{p_n/2}\lesssim\int_{\left|(\sum_\theta h_\theta)*\check{\eta}_{\leq R^{-2/n}}\right|\gtrsim\alpha^2}\left|(\sum_\theta h_\theta)*\check{\eta}_{\leq R^{-2/n}}\right|^{p_n/2}. \tag*{(41)} ∫ ∑ θ h θ > α 2 ( θ ∑ h θ ) p n /2 ≲ ∫ ( ∑ θ h θ ) ∗ η ˇ ≤ R − 2/ n ≳ α 2 ( θ ∑ h θ ) ∗ η ˇ ≤ R − 2/ n p n /2 . (41)
The first case and the second case are called the high frequency case and the low frequency case, respectively. The high frequency case is treated by Lemma 3.1 after rescaling; the low frequency case is treated by averaging over balls of radius R 2 / n R^{2/n} R 2/ n as usual.
Case 1: High frequency case. Every ξ ∈ supp ( h θ η λ ^ ) \xi\in\operatorname{supp}(\widehat{h_\theta\eta_\lambda}) ξ ∈ supp ( h θ η λ ) satisfies ∣ ξ ∣ ∼ λ |\xi|\sim\lambda ∣ ξ ∣ ∼ λ and has the expansion
ξ = ∑ i = 1 n a i γ n ( i ) ( t θ ) , ∣ a 1 ∣ ≲ λ , ∣ a i ∣ ≲ R − i / n ( 2 ≤ i ≤ n ) . \xi=\sum_{i=1}^{n}a_i\gamma^{(i)}_n(t_\theta),\qquad|a_1|\lesssim\lambda,\qquad|a_i|\lesssim R^{-i/n}\quad(2\leq i\leq n). ξ = i = 1 ∑ n a i γ n ( i ) ( t θ ) , ∣ a 1 ∣ ≲ λ , ∣ a i ∣ ≲ R − i / n ( 2 ≤ i ≤ n ) .
Under λ − 1 \lambda^{-1} λ − 1 dilation, the dimensions of supp ( h θ η λ ^ ) \operatorname{supp}(\widehat{h_\theta\eta_\lambda}) supp ( h θ η λ ) change as follows:
( λ × R − 2 / n × ⋯ × R − 1 ) → λ − 1 d i l a t i o n ( 1 × R − 2 / n λ × ⋯ × R − 1 λ ) ⊂ ( 1 × R − 2 / n λ × ⋯ × ( R − 2 / n λ ) n − 1 ) . \left(\lambda\times R^{-2/n}\times\cdots\times R^{-1}\right)\xrightarrow{\lambda^{-1}\ \mathrm{dilation}}\left(1\times\frac{R^{-2/n}}{\lambda}\times\cdots\times\frac{R^{-1}}{\lambda}\right)\subset\left(1\times\frac{R^{-2/n}}{\lambda}\times\cdots\times\left(\frac{R^{-2/n}}{\lambda}\right)^{n-1}\right). ( λ × R − 2/ n × ⋯ × R − 1 ) λ − 1 dilation ( 1 × λ R − 2/ n × ⋯ × λ R − 1 ) ⊂ ( 1 × λ R − 2/ n × ⋯ × ( λ R − 2/ n ) n − 1 ) .
Put σ = R − 2 / n / λ \sigma=R^{-2/n}/\lambda σ = R − 2/ n / λ and group the caps θ ∈ Θ ( R − 1 / n ) \theta\in\Theta(R^{-1/n}) θ ∈ Θ ( R − 1/ n ) according to their parents τ ∈ Θ ( σ ) \tau\in\Theta(\sigma) τ ∈ Θ ( σ ) . Since ∣ t θ − t τ ∣ ≤ σ |t_\theta-t_\tau|\leq\sigma ∣ t θ − t τ ∣ ≤ σ , Taylor expansion from t θ t_\theta t θ to t τ t_\tau t τ , as in the proof of Lemma 12.4 , gives
λ − 1 supp [ ( ∑ θ ⊂ τ h θ ) ∗ η ˇ λ ] ⊂ { ∑ i = 1 n b i γ n ( i ) ( t τ ) : ∣ b 1 ∣ ∼ 1 , ∣ b i ∣ ≲ σ i − 1 ( i ≥ 2 ) } . \lambda^{-1}\operatorname{supp}\left[\left(\sum_{\theta\subset\tau}h_\theta\right)*\check{\eta}_\lambda\right]\subset\left\{\sum_{i=1}^{n}b_i\gamma_n^{(i)}(t_\tau):|b_1|\sim1,\ |b_i|\lesssim\sigma^{i-1}\ (i\geq2)\right\}. λ − 1 supp [ ( θ ⊂ τ ∑ h θ ) ∗ η ˇ λ ] ⊂ { i = 1 ∑ n b i γ n ( i ) ( t τ ) : ∣ b 1 ∣ ∼ 1 , ∣ b i ∣ ≲ σ i − 1 ( i ≥ 2 ) } .
Thus the Fourier supports of the dilated functions lie in caps of aperture σ \sigma σ , whose active count at scale s s s is at most N ( s ) N(s) N ( s ) . For g ( x ) = [ ( ∑ θ h θ ) ∗ η ˇ λ ] ( x / λ ) g(x)=[(\sum_\theta h_\theta)*\check{\eta}_\lambda](x/\lambda) g ( x ) = [( ∑ θ h θ ) ∗ η ˇ λ ] ( x / λ ) , apply Lemma 3.1 to the family [ ( ∑ θ ⊂ τ h θ ) ∗ η ˇ λ ] ( x / λ ) τ ∈ Θ ( σ ) [(\sum_{\theta\subset\tau}h_\theta)*\check{\eta}_\lambda](x/\lambda)_{\tau\in\Theta(\sigma)} [( ∑ θ ⊂ τ h θ ) ∗ η ˇ λ ] ( x / λ ) τ ∈ Θ ( σ ) with σ = R − 2 / n / λ \sigma=R^{-2/n}/\lambda σ = R − 2/ n / λ and α ↦ α R ϵ \alpha\mapsto\frac{\alpha}{R^\epsilon} α ↦ R ϵ α . On returning to the original coordinates, the averaging boxes U ∥ U ρ , σ − ( n − 1 ) U\parallel U_{\rho,\sigma^{-(n-1)}} U ∥ U ρ , σ − ( n − 1 ) become
B ∥ λ − 1 U ρ , σ − ( n − 1 ) = λ s R 2 / n V ρ , R , σ ≤ s ≤ 1. B\parallel\lambda^{-1}U_{\rho,\sigma^{-(n-1)}}=\lambda sR^{2/n}V_{\rho,R},\qquad\sigma\leq s\leq1. B ∥ λ − 1 U ρ , σ − ( n − 1 ) = λ s R 2/ n V ρ , R , σ ≤ s ≤ 1.
Their dimensions are λ R 4 / n s 2 × λ R 4 / n s × ⋯ × λ R 4 / n s 3 − n \lambda R^{4/n}s^2\times\lambda R^{4/n}s\times\cdots\times\lambda R^{4/n}s^{3-n} λ R 4/ n s 2 × λ R 4/ n s × ⋯ × λ R 4/ n s 3 − n . The same Jacobian occurs in the integral in (40) and in ∣ U ∣ |U| ∣ U ∣ . Thus the right hand side of (40) is bounded by
R ϵ ∑ σ ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ B ∥ λ s R 2 / n V ρ , R B ∈ G ρ ( R ϵ ) ∣ B ∣ ∥ ( ∑ τ ⊂ ρ ( ∑ θ ⊂ τ h θ ) ∗ η ˇ λ ) 1 / 2 ∥ L # p n / 2 ( B ) p n . (42) R^\epsilon\sum_{\sigma\leq s\leq1}\sum_{\rho\in\Theta(s)}\sum_{\substack{B\parallel\lambda sR^{2/n}V_{\rho,R}\\B\in\mathcal{G}_\rho(R^\epsilon)}}|B|\left\|\left(\sum_{\tau\subset\rho}\left(\sum_{\theta\subset\tau}h_\theta\right)*\check{\eta}_\lambda\right)^{1/2}\right\|_{L_\#^{p_n/2}(B)}^{p_n}. \tag*{(42)} R ϵ σ ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ B ∥ λ s R 2/ n V ρ , R B ∈ G ρ ( R ϵ ) ∑ ∣ B ∣ ( τ ⊂ ρ ∑ ( θ ⊂ τ ∑ h θ ) ∗ η ˇ λ ) 1/2 L # p n /2 ( B ) p n . (42)
Applying Hölder and (35) in Lemma 2.16 to each sum over B B B in (42) bounds (3.4) by
R ϵ ∑ σ ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ B ∥ λ s R 2 / n V ρ , R B ∈ G ρ ( R ϵ ) ∣ B ∣ ∥ ( ∑ θ ⊂ ρ h θ ) 1 / 2 ∥ L # p n / 2 ( B ) p n . (43) R^\epsilon\sum_{\sigma\leq s\leq1}\sum_{\rho\in\Theta(s)}\sum_{\substack{B\parallel\lambda sR^{2/n}V_{\rho,R}\\B\in\mathcal{G}_\rho(R^\epsilon)}}|B|\left\|\left(\sum_{\theta\subset\rho}h_\theta\right)^{1/2}\right\|_{L_\#^{p_n/2}(B)}^{p_n}. \tag*{(43)} R ϵ σ ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ B ∥ λ s R 2/ n V ρ , R B ∈ G ρ ( R ϵ ) ∑ ∣ B ∣ θ ⊂ ρ ∑ h θ 1/2 L # p n /2 ( B ) p n . (43)
Now apply Lemma 2.14 to each sum over B B B in (43) , with F = ( ∑ θ ⊂ ρ h θ ) p n / 4 F=(\sum_{\theta\subset\rho}h_\theta)^{p_n/4} F = ( ∑ θ ⊂ ρ h θ ) p n /4 , B 0 = λ s R 2 / n V ρ , R B_0=\lambda sR^{2/n}V_{\rho,R} B 0 = λ s R 2/ n V ρ , R , and U 0 = V ρ , R U_0=V_{\rho,R} U 0 = V ρ , R . Since λ s R 2 / n ≥ 1 \lambda sR^{2/n}\geq1 λ s R 2/ n ≥ 1 , this bounds (3.5) by
R ϵ ∑ R − 1 / n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , R V ∈ G ρ ( R ϵ ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ h θ ) 1 / 2 ∥ L # p n / 2 ( V ) p n . (44) R^\epsilon\sum_{R^{-1/n}\leq s\leq1}\sum_{\rho\in\Theta(s)}\sum_{\substack{V\parallel V_{\rho,R}\\V\in\mathcal{G}_\rho(R^\epsilon)}}|V|\left\|\left(\sum_{\theta\subset\rho}h_\theta\right)^{1/2}\right\|_{L_\#^{p_n/2}(V)}^{p_n}. \tag*{(44)} R ϵ R − 1/ n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , R V ∈ G ρ ( R ϵ ) ∑ ∣ V ∣ θ ⊂ ρ ∑ h θ 1/2 L # p n /2 ( V ) p n . (44)
Case 2: Low frequency case. Tile by Q ∥ V ρ , R Q \parallel V_{\rho,R} Q ∥ V ρ , R , ∣ ρ ∣ = 1 |\rho|=1 ∣ ρ ∣ = 1 . These boxes are comparable to balls of radius R 2 / n R^{2/n} R 2/ n . The decay of η ˇ ≤ R − 2 / n \check{\eta}_{\le R^{-2/n}} η ˇ ≤ R − 2/ n and Hölder give
\sup_{x\in Q}\left|\left(\sum_{\theta}h_{\theta}\right)*\check{\eta}_{\le R^{-2/n}}(x)\right|^{p_n/4}\lesssim\fint_Q\left(\sum_{\theta}h_{\theta}\right)^{p_n/4}.
Consequently, the right hand side of (3.3) is bounded by the s = 1 s=1 s = 1 sum in (3.6). This proves (2.7). □ \square □
From cones over moment curves to moment curves: Auxiliary estimates for unpruning
In this section and the next, we prove the unpruning estimate in Proposition 2.12. Here we establish two auxiliary estimates for the pruned square functions, which we combine in Section 5 to complete the proof.
Wave envelope type estimates for the pruned functions
The following lemma bounds the sum over Q Q Q in (4.1) by sums over V V V and B B B , involving the original functions f θ f_\theta f θ and finer pruned groups, respectively.
Lemma 4.1. Let 1 ≤ j ≤ N 1\le j\le N 1 ≤ j ≤ N and R j − 1 / n ≤ σ ≤ R j − 1 − 1 / n R_j^{-1/n}\le\sigma\le R_{j-1}^{-1/n} R j − 1/ n ≤ σ ≤ R j − 1 − 1/ n . For every s ∈ [ σ , 1 ] s\in[\sigma,1] s ∈ [ σ , 1 ] and α > 0 \alpha>0 α > 0 , there exists a dyadic σ ′ ∈ [ R − 1 / n , σ ] \sigma'\in[R^{-1/n},\sigma] σ ′ ∈ [ R − 1/ n , σ ] such that
∑ ρ ∈ Θ ( s ) ∑ Q ∈ G ρ ( α ) ∣ τ ∣ = σ ∥ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ ) 1 / 2 ∥ L # p n / 2 ( Q ) p n ≲ η , ε R ε ∑ R − 1 / n ≤ s ′ ≤ s ρ ′ ∈ Θ ( s ′ ) ∑ V ∥ V ρ ′ , R V ∈ G ρ ′ ( R ε ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ′ ∣ f θ ∣ 2 ∗ w θ ) 1 / 2 ∥ L # p n / 2 ( V ) p n + R ε ∑ σ ′ ≤ s ′ ≤ s ρ ′ ∈ Θ ( s ′ ) ∑ B ∥ ( s ′ / σ ′ ) V ρ ′ , ( σ ′ ) − n B ∈ G ρ ′ ( R ε ) ∣ B ∣ ∥ ( ∑ τ ′ ⊂ ρ ′ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ′ ∣ 2 ∗ w τ ′ ) 1 / 2 ∥ L # p n / 2 ( B ) p n . (45) \begin{aligned}
&\sum_{\rho\in\Theta(s)}\sum_{\substack{Q\in\mathcal{G}_{\rho}(\alpha)\\|\tau|=\sigma}}
\left\|\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_\tau|^2*w_\tau\right)^{1/2}\right\|_{L_{\#}^{p_n/2}(Q)}^{p_n}\\
&\lesssim_{\eta,\varepsilon} R^\varepsilon
\sum_{\substack{R^{-1/n}\le s'\le s\\\rho'\in\Theta(s')}}
\sum_{\substack{V\parallel V_{\rho',R}\\V\in\mathcal{G}_{\rho'}(R^\varepsilon)}}
|V|\left\|\left(\sum_{\theta\subset\rho'}|f_\theta|^2*w_\theta\right)^{1/2}\right\|_{L_{\#}^{p_n/2}(V)}^{p_n}\\
&\quad+R^\varepsilon
\sum_{\substack{\sigma'\le s'\le s\\\rho'\in\Theta(s')}}
\sum_{\substack{B\parallel(s'/\sigma')V_{\rho',(\sigma')^{-n}}\\B\in\mathcal{G}_{\rho'}(R^\varepsilon)}}
|B|\left\|\left(\sum_{\substack{\tau'\subset\rho'\\|\tau'|=\sigma'}}|f_{\tau'}'|^2*w_{\tau'}\right)^{1/2}\right\|_{L_{\#}^{p_n/2}(B)}^{p_n}. \tag*{(45)}
\end{aligned} ρ ∈ Θ ( s ) ∑ Q ∈ G ρ ( α ) ∣ τ ∣ = σ ∑ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ 1/2 L # p n /2 ( Q ) p n ≲ η , ε R ε R − 1/ n ≤ s ′ ≤ s ρ ′ ∈ Θ ( s ′ ) ∑ V ∥ V ρ ′ , R V ∈ G ρ ′ ( R ε ) ∑ ∣ V ∣ θ ⊂ ρ ′ ∑ ∣ f θ ∣ 2 ∗ w θ 1/2 L # p n /2 ( V ) p n + R ε σ ′ ≤ s ′ ≤ s ρ ′ ∈ Θ ( s ′ ) ∑ B ∥ ( s ′ / σ ′ ) V ρ ′ , ( σ ′ ) − n B ∈ G ρ ′ ( R ε ) ∑ ∣ B ∣ τ ′ ⊂ ρ ′ ∣ τ ′ ∣ = σ ′ ∑ ∣ f τ ′ ′ ∣ 2 ∗ w τ ′ 1/2 L # p n /2 ( B ) p n . (45)
Here j ≤ j ′ ≤ N j\le j'\le N j ≤ j ′ ≤ N satisfies R j ′ − 1 / n ≤ σ ′ ≤ R j ′ − 1 − 1 / n R_{j'}^{-1/n}\le\sigma'\le R_{j'-1}^{-1/n} R j ′ − 1/ n ≤ σ ′ ≤ R j ′ − 1 − 1/ n .
For the original functions f τ f_\tau f τ , the same inequality holds for R − 1 / n ≤ σ ≤ s ≤ 1 R^{-1/n}\le\sigma\le s\le1 R − 1/ n ≤ σ ≤ s ≤ 1 , with j , j ′ j,j' j , j ′ omitted.
By Lemma 2.17, for some fixed 0 < c < 1 0<c<1 0 < c < 1 ,
LHS of (4.1) ≲ ∑ ρ ∈ Θ ( s ) ∫ ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ > ( c α / N ( s ) ) 2 ( ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ ) p n / 2 . \operatorname{LHS\ of\ (4.1)}\lesssim
\sum_{\rho\in\Theta(s)}
\int_{\sum_{|\tau|=\sigma}|f_\tau|^2*w_\tau>(c\alpha/N(s))^2}
\left(\sum_{|\tau|=\sigma}|f_\tau|^2*w_\tau\right)^{p_n/2}. LHS of (4.1) ≲ ρ ∈ Θ ( s ) ∑ ∫ ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ > ( c α / N ( s ) ) 2 ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ p n /2 .
We use a high low decomposition to the square function. The high frequency part is estimated by Lemma 3.1 after rescaling; in the low frequency case we increase the pruning index and repeat the decomposition, applying Lemma 2.8 at j = N j=N j = N .
Proof. For each ρ ∈ Θ ( s ) \rho\in\Theta(s) ρ ∈ Θ ( s ) , split its superlevel set in the preceding display. The high frequency region consists of the points where
∣ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ ) ∗ η ˇ > R j − 1 / n ( x ) ∣ ≥ 1 2 ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ ) ( x ) . \left|\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_\tau|^2*w_\tau\right)*\check{\eta}_{>R_j^{-1/n}}(x)\right|
\ge\frac{1}{2}\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_\tau|^2*w_\tau\right)(x). τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ ∗ η ˇ > R j − 1/ n ( x ) ≥ 2 1 τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ ( x ) .
On the low frequency region, we have
( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ ) ( x ) ≲ ∣ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ ) ∗ η ˇ ≤ R j − 1 / n ( x ) ∣ . \left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_\tau|^2*w_\tau\right)(x)
\lesssim
\left|\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_\tau|^2*w_\tau\right)*\check{\eta}_{\le R_j^{-1/n}}(x)\right|. τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ ( x ) ≲ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ ∗ η ˇ ≤ R j − 1/ n ( x ) .
We distinguish which region contributes at least half of the sum of integrals. In the high frequency case, decompose this part into dyadic pieces using η σ ′ \eta_{\sigma'} η σ ′ , R j − 1 / n < σ ′ ≤ σ R_j^{-1/n}<\sigma'\le\sigma R j − 1/ n < σ ′ ≤ σ , and pigeonhole σ ′ \sigma' σ ′ after summing over ρ \rho ρ . Let us first consider the case that the high frequency part dominates.
Case 1: High frequency part
For each ρ ∈ Θ ( s ) \rho\in\Theta(s) ρ ∈ Θ ( s ) , expand every f τ j f_{\tau}^{j} f τ j into its σ ′ \sigma' σ ′ pieces and write
F = ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ ) ∗ η ~ σ ′ = ∑ τ ′ ⊂ ρ ∣ τ ′ ∣ = σ ′ h τ ′ . F=\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_{\tau}^{j}|^{2}*w_{\tau}\right)*\widetilde{\eta}_{\sigma'}=\sum_{\substack{\tau'\subset\rho\\|\tau'|=\sigma'}}h_{\tau'}. F = τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ j ∣ 2 ∗ w τ ∗ η σ ′ = τ ′ ⊂ ρ ∣ τ ′ ∣ = σ ′ ∑ h τ ′ .
Here, for τ ′ ⊂ τ \tau'\subset\tau τ ′ ⊂ τ with ∣ τ ∣ = σ |\tau|=\sigma ∣ τ ∣ = σ ,
h τ ′ = ∑ τ ′ ⊂ τ , ∣ τ ′ ∣ = σ ′ ∣ t τ ′ − t τ ~ ′ ∣ ≲ σ ′ ( f τ j f τ ~ j ‾ ) ∗ w τ ∗ η ~ σ ′ . h_{\tau'}=\sum_{\substack{\tau'\subset\tau,\ |\tau'|=\sigma'\\|t_{\tau'}-t_{\widetilde{\tau}'}|\lesssim\sigma'}}(f_{\tau}^{j}\overline{f_{\widetilde{\tau}}^{j}})*w_{\tau}*\widetilde{\eta}_{\sigma'}. h τ ′ = τ ′ ⊂ τ , ∣ τ ′ ∣ = σ ′ ∣ t τ ′ − t τ ′ ∣ ≲ σ ′ ∑ ( f τ j f τ j ) ∗ w τ ∗ η σ ′ .
The ball cutoff selects only these neighbouring pairs. Dyadic pigeonholing after summing over ρ \rho ρ bounds the left hand side of (45) by ≲ ∑ ρ ∈ Θ ( s ) ∣ F ∣ ≳ ( α / N ( s ) ) 2 p n / 2 \lesssim\sum_{\rho\in\Theta(s)}|F|_{\gtrsim(\alpha/N(s))^{2}}^{pn/2} ≲ ∑ ρ ∈ Θ ( s ) ∣ F ∣ ≳ ( α / N ( s ) ) 2 p n /2 for some R j − 1 / n < σ ′ ≤ σ R_{j}^{-1/n}<\sigma'\leq\sigma R j − 1/ n < σ ′ ≤ σ .
Fix ρ ∈ Θ ( s ) \rho\in\Theta(s) ρ ∈ Θ ( s ) in the preceding sum and let t ρ t_{\rho} t ρ be the left endpoint of the interval indexing ρ \rho ρ . As in the proof of Lemma 2.8, every ξ ∈ supp h ^ τ ′ \xi\in\operatorname{supp}\widehat{h}_{\tau'} ξ ∈ supp h τ ′ satisfies ∣ ξ ∣ ∼ σ ′ |\xi|\sim\sigma' ∣ ξ ∣ ∼ σ ′ and can be written as
ξ = ∑ i = 1 n a i γ ( i ) ( t τ ′ ) , ∣ a i ∣ ≲ ( σ ′ ) i ( 1 ≤ i ≤ n ) . \xi=\sum_{i=1}^{n}a_{i}\gamma^{(i)}(t_{\tau'}),\qquad|a_{i}|\lesssim(\sigma')^{i}\quad(1\leq i\leq n). ξ = i = 1 ∑ n a i γ ( i ) ( t τ ′ ) , ∣ a i ∣ ≲ ( σ ′ ) i ( 1 ≤ i ≤ n ) .
Since F F F is the sum over τ ′ ⊂ ρ \tau'\subset\rho τ ′ ⊂ ρ , we follow dilation by ( σ ′ ) − 1 (\sigma')^{-1} ( σ ′ ) − 1 with the rescaling s L t ρ , s − 1 sL_{t_{\rho},s}^{-1} s L t ρ , s − 1 that sends [ t ρ , t ρ + s ] [t_{\rho},t_{\rho}+s] [ t ρ , t ρ + s ] to [ 0 , 1 ] [0,1] [ 0 , 1 ] . The widths of each piece then change as follows:
( σ ′ × ( σ ′ ) 2 × ⋯ × ( σ ′ ) n ) → ( σ ′ ) − 1 d i l a t i o n ( 1 × σ ′ × ⋯ × ( σ ′ ) n − 1 ) → s L t ρ , s − 1 ( 1 × σ ′ s × ⋯ × ( σ ′ s ) n − 1 ) . (\sigma'\times(\sigma')^{2}\times\cdots\times(\sigma')^{n})\xrightarrow{(\sigma')^{-1}\ \mathrm{dilation}}(1\times\sigma'\times\cdots\times(\sigma')^{n-1})\xrightarrow{sL_{t_{\rho},s}^{-1}}\left(1\times\frac{\sigma'}{s}\times\cdots\times\left(\frac{\sigma'}{s}\right)^{n-1}\right). ( σ ′ × ( σ ′ ) 2 × ⋯ × ( σ ′ ) n ) ( σ ′ ) − 1 dilation ( 1 × σ ′ × ⋯ × ( σ ′ ) n − 1 ) s L t ρ , s − 1 ( 1 × s σ ′ × ⋯ × ( s σ ′ ) n − 1 ) .
Thus the rescaled pieces have Fourier support in caps of aperture σ ′ / s \sigma'/s σ ′ / s . For g ( x ) = F ( ( s / σ ′ ) L t ρ , s − T x ) g(x)=F((s/\sigma')L_{t_{\rho},s}^{-T}x) g ( x ) = F (( s / σ ′ ) L t ρ , s − T x ) , apply Lemma 3.1 to the family { h τ ′ ( ( s / σ ′ ) L t ρ , s − T x ) } τ ′ ⊂ ρ , ∣ τ ′ ∣ = σ ′ \{h_{\tau'}((s/\sigma')L_{t_{\rho},s}^{-T}x)\}_{\tau'\subset\rho,\ |\tau'|=\sigma'} { h τ ′ (( s / σ ′ ) L t ρ , s − T x ) } τ ′ ⊂ ρ , ∣ τ ′ ∣ = σ ′ at amplitude α / ( R ϵ N ( s ) ) \alpha/(R^{\epsilon}N(s)) α / ( R ϵ N ( s )) .
In the sum given by Lemma 3.1 , let ρ ′ ′ \rho'' ρ ′′ be a parent of width s ′ / s s'/s s ′ / s and let U ∥ U ρ ′ ′ , ( σ ′ / s ) − ( n − 1 ) U\parallel U_{\rho'',(\sigma'/s)^{-(n-1)}} U ∥ U ρ ′′ , ( σ ′ / s ) − ( n − 1 ) . Returning to the original coordinates sends ρ ′ ′ \rho'' ρ ′′ to a cap ρ ′ ⊂ ρ \rho'\subset\rho ρ ′ ⊂ ρ of width s ′ s' s ′ and U U U to B = ( s / σ ′ ) L t ρ , s − T U B=(s/\sigma')L_{t_{\rho},s}^{-T}U B = ( s / σ ′ ) L t ρ , s − T U . The i i i th side length of B B B is
s 1 − i σ ′ ( σ ′ s ) − 2 ( s ′ s ) 3 − i = ( σ ′ ) − 3 ( s ′ ) 3 − i , 1 ≤ i ≤ n . \frac{s^{1-i}}{\sigma'}\left(\frac{\sigma'}{s}\right)^{-2}\left(\frac{s'}{s}\right)^{3-i}=(\sigma')^{-3}(s')^{3-i},\qquad1\leq i\leq n. σ ′ s 1 − i ( s σ ′ ) − 2 ( s s ′ ) 3 − i = ( σ ′ ) − 3 ( s ′ ) 3 − i , 1 ≤ i ≤ n .
Thus B ∥ ( s ′ / σ ′ ) V ρ ′ , ( σ ′ ) − n B\parallel(s'/\sigma')V_{\rho',(\sigma')^{-n}} B ∥ ( s ′ / σ ′ ) V ρ ′ , ( σ ′ ) − n .
The active parent count at width s ′ / s s'/s s ′ / s is at most N ρ ( s ′ ) N_{\rho}(s') N ρ ( s ′ ) , since each nonzero h τ ′ h_{\tau'} h τ ′ comes from a cap τ ′ \tau' τ ′ containing a cap θ \theta θ for which f θ ≢ 0 f_{\theta}\not\equiv0 f θ ≡ 0 . Using N ( s ) N ρ ( s ′ ) ≲ ϵ N ( s ′ ) N(s)N_{\rho}(s')\lesssim_{\epsilon}N(s') N ( s ) N ρ ( s ′ ) ≲ ϵ N ( s ′ ) from Lemma 2.18 in the selection condition and summing over ρ \rho ρ , we obtain
∑ ρ ∈ Θ ( s ) ∣ F ∣ ≳ ( α / N ( s ) ) 2 p n / 2 ≲ ϵ R ϵ ∑ σ ′ ≤ s ′ ≤ s ∑ ρ ′ ∈ Θ ( s ′ ) ∑ B ∥ ( s ′ / σ ′ ) V ρ ′ , ( σ ′ ) − n B ∈ G ρ ′ ( R ϵ ) ∣ B ∣ ( ∑ τ ′ ⊂ ρ ′ ∣ τ ′ ∣ = σ ′ ∣ h τ ′ ∣ ) p n / 2 ∥ 1 B ∥ L # p n / 2 p n / 2 . (46) \sum_{\rho\in\Theta(s)}|F|_{\gtrsim(\alpha/N(s))^{2}}^{pn/2}\lesssim_{\epsilon}R^{\epsilon}\sum_{\sigma'\leq s'\leq s}\sum_{\rho'\in\Theta(s')}\sum_{\substack{B\parallel(s'/\sigma')V_{\rho',(\sigma')^{-n}}\\B\in\mathcal{G}_{\rho'}(R^{\epsilon})}}|B|\left(\sum_{\substack{\tau'\subset\rho'\\|\tau'|=\sigma'}}|h_{\tau'}|\right)^{pn/2}\|1_{B}\|_{L_{\#}^{pn/2}}^{pn/2}. \tag*{(46)} ρ ∈ Θ ( s ) ∑ ∣ F ∣ ≳ ( α / N ( s ) ) 2 p n /2 ≲ ϵ R ϵ σ ′ ≤ s ′ ≤ s ∑ ρ ′ ∈ Θ ( s ′ ) ∑ B ∥ ( s ′ / σ ′ ) V ρ ′ , ( σ ′ ) − n B ∈ G ρ ′ ( R ϵ ) ∑ ∣ B ∣ τ ′ ⊂ ρ ′ ∣ τ ′ ∣ = σ ′ ∑ ∣ h τ ′ ∣ p n /2 ∥ 1 B ∥ L # p n /2 p n /2 . (46)
Since τ ′ ⊂ τ \tau'\subset\tau τ ′ ⊂ τ , we have ∣ f τ ′ j ∣ 2 ∗ w τ ′ ≲ ∣ f τ j ∣ 2 ∗ w τ |f_{\tau'}^{j}|^{2}*w_{\tau'}\lesssim|f_{\tau}^{j}|^{2}*w_{\tau} ∣ f τ ′ j ∣ 2 ∗ w τ ′ ≲ ∣ f τ j ∣ 2 ∗ w τ . Bounding neighbouring products by separate squares and absorbing neighbouring parents by bounded overlap, we may replace ∑ τ ′ ⊂ ρ ′ , ∣ τ ′ ∣ = σ ′ ∣ h τ ′ ∣ \sum_{\tau'\subset\rho',\ |\tau'|=\sigma'}|h_{\tau'}| ∑ τ ′ ⊂ ρ ′ , ∣ τ ′ ∣ = σ ′ ∣ h τ ′ ∣ on the right of (46) by ∣ η ~ σ ′ ∣ ∗ ∑ τ ′ ⊂ ρ ′ , ∣ τ ′ ∣ = σ ′ ∣ f τ ′ j ∣ 2 ∗ w τ ′ |\widetilde{\eta}_{\sigma'}|*\sum_{\tau'\subset\rho',\ |\tau'|=\sigma'}|f_{\tau'}^{j}|^{2}*w_{\tau'} ∣ η σ ′ ∣ ∗ ∑ τ ′ ⊂ ρ ′ , ∣ τ ′ ∣ = σ ′ ∣ f τ ′ j ∣ 2 ∗ w τ ′ . Apply Hölder and (35) to each sum over B B B on the right of (46) , with F = ( ∑ τ ′ ⊂ ρ ′ , ∣ τ ′ ∣ = σ ′ ∣ f τ ′ j ∣ 2 ∗ w τ ′ ) p n / 4 F=(\sum_{\tau'\subset\rho',\ |\tau'|=\sigma'}|f_{\tau'}^{j}|^{2}*w_{\tau'})^{pn/4} F = ( ∑ τ ′ ⊂ ρ ′ , ∣ τ ′ ∣ = σ ′ ∣ f τ ′ j ∣ 2 ∗ w τ ′ ) p n /4 and K = ∣ η ~ σ ′ ∣ K=|\widetilde{\eta}_{\sigma'}| K = ∣ η σ ′ ∣ . This bounds the right hand side of (46) by the second sum on the right of (45) , with j ′ = j j'=j j ′ = j .
Case 2: The low frequency part dominates
Using (26) for the σ \sigma σ pieces inside ρ \rho ρ gives
∣ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ ) ∗ η ~ ≤ R j − 1 / n ∣ ≲ ∑ τ j ⊂ ρ ∣ f τ j j ∣ 2 ∗ w τ j . \left|\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_{\tau}^{j}|^{2}*w_{\tau}\right)*\widetilde{\eta}_{\leq R_{j}^{-1/n}}\right|\lesssim\sum_{\tau_{j}\subset\rho}|f_{\tau_{j}}^{j}|^{2}*w_{\tau_{j}}. τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ j ∣ 2 ∗ w τ ∗ η ≤ R j − 1/ n ≲ τ j ⊂ ρ ∑ ∣ f τ j j ∣ 2 ∗ w τ j .
For j < N j<N j < N , applying ∣ f τ j j ∣ ≤ ∣ f τ j j + 1 ∣ |f_{\tau_{j}}^{j}|\leq|f_{\tau_{j}}^{j+1}| ∣ f τ j j ∣ ≤ ∣ f τ j j + 1 ∣ yields
LHS of x r e f ‘ ( 4.1 ) ‘ ≲ η , ϵ R ϵ ∑ ρ ∈ Θ ( s ) ∑ Q ∈ G ρ ( R ϵ ) Q ∥ ρ ∗ ∣ Q ∣ ( ∑ τ j ⊂ ρ ∣ f τ j j + 1 ∣ 2 ∗ w τ j ) 1 / 2 p n ∥ 1 Q ∥ L # p n / 2 ( Q ) p n / 2 . (47) \operatorname{LHS\ of\ }{xref}`(4.1)`\lesssim_{\eta,\epsilon}R^{\epsilon}\sum_{\rho\in\Theta(s)}\sum_{\substack{Q\in\mathcal{G}_{\rho}(R^{\epsilon})\\Q\parallel\rho^{*}}}|Q|\left(\sum_{\tau_{j}\subset\rho}|f_{\tau_{j}}^{j+1}|^{2}*w_{\tau_{j}}\right)^{1/2\;pn}\|1_{Q}\|_{L_{\#}^{pn/2}(Q)}^{pn/2}. \tag*{(47)} LHS of x r e f ‘ ( 4.1 ) ‘ ≲ η , ϵ R ϵ ρ ∈ Θ ( s ) ∑ Q ∈ G ρ ( R ϵ ) Q ∥ ρ ∗ ∑ ∣ Q ∣ τ j ⊂ ρ ∑ ∣ f τ j j + 1 ∣ 2 ∗ w τ j 1/2 p n ∥ 1 Q ∥ L # p n /2 ( Q ) p n /2 . (47)
The sum over Q Q Q on the right of (4.3) is the left hand side of (4.1) with ( j , σ , α ) (j,\sigma,\alpha) ( j , σ , α ) replaced by ( j + 1 , R j − 1 / n , α / R ε ) (j+1,R_j^{-1/n},\alpha/R^\varepsilon) ( j + 1 , R j − 1/ n , α / R ε ) , while s s s is unchanged. Repeat the decomposition into Cases 1 and 2 on this sum over Q Q Q with these parameters. If the high frequency part in Case 1 dominates at index j ′ j' j ′ , repeat the argument leading to (4.2) at that index and remove ∣ η ~ σ ′ ∣ |\widetilde{\eta}_{\sigma'}| ∣ η σ ′ ∣ as above. This gives the second sum on the right of (4.1), with R j ′ − 1 / n < σ ′ ≤ R j ′ − 1 − 1 / n R_{j'}^{-1/n}<\sigma'\le R_{j'-1}^{-1/n} R j ′ − 1/ n < σ ′ ≤ R j ′ − 1 − 1/ n .
If the low frequency part dominates at every step, we repeat Case 2 until the pruning index reaches N N N . At j = N j=N j = N , we use ∣ f θ N ∣ ≤ ∣ f θ ∣ |f_\theta^N|\le|f_\theta| ∣ f θ N ∣ ≤ ∣ f θ ∣ to obtain
LHS of (4.1) ≲ η , ε R ε ∑ ρ ∈ Θ ( s ) ∑ Q ∈ G ρ ( α R ε ) ∣ Q ∣ ∥ ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w θ ) 1 / 2 ∥ L # p n / 2 ( Q ) p n . (48) \operatorname{LHS}\text{ of (4.1)}\lesssim_{\eta,\varepsilon}R^\varepsilon\sum_{\rho\in\Theta(s)}\sum_{Q\in\mathcal{G}_\rho\left(\frac{\alpha}{R^\varepsilon}\right)}|Q|\left\|\left(\sum_{\theta\subset\rho}|f_\theta|^2*w_\theta\right)^{1/2}\right\|_{L_\#^{p_n/2}(Q)}^{p_n}. \tag*{(48)} LHS of (4.1) ≲ η , ε R ε ρ ∈ Θ ( s ) ∑ Q ∈ G ρ ( R ε α ) ∑ ∣ Q ∣ θ ⊂ ρ ∑ ∣ f θ ∣ 2 ∗ w θ 1/2 L # p n /2 ( Q ) p n . (48)
It remains to bound the right hand side of (4.4) by the first sum on the right of (4.1). For each ρ ∈ Θ ( s ) \rho\in\Theta(s) ρ ∈ Θ ( s ) , (2.22) bounds the sum over Q Q Q in (4.4) by
∫ ∑ Q ⊂ ρ ∣ f θ ∣ 2 ∗ w θ ≳ ( α R ε N ( s ) ) 2 ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w θ ) p n / 2 . \int_{\sum_{Q\subset\rho}|f_\theta|^2*w_\theta}\gtrsim\left(\frac{\alpha}{R^\varepsilon N(s)}\right)^2\left(\sum_{\theta\subset\rho}|f_\theta|^2*w_\theta\right)^{p_n/2}. ∫ ∑ Q ⊂ ρ ∣ f θ ∣ 2 ∗ w θ ≳ ( R ε N ( s ) α ) 2 θ ⊂ ρ ∑ ∣ f θ ∣ 2 ∗ w θ p n /2 .
The Taylor rescaling taking ρ \rho ρ to the unit cap sends each θ ⊂ ρ \theta\subset\rho θ ⊂ ρ to a cap of M n ( s n R ) \mathcal{M}^n(s^nR) M n ( s n R ) at scale R − 1 / n / s R^{-1/n}/s R − 1/ n / s . We therefore apply Lemma 2.8 to { ( ∣ f θ ∣ 2 ∗ w θ ) ( L t ρ , s ⋅ ) } θ ⊂ ρ \{(|f_\theta|^2*w_\theta)(L_{t_\rho,s}\cdot)\}_{\theta\subset\rho} {( ∣ f θ ∣ 2 ∗ w θ ) ( L t ρ , s ⋅ ) } θ ⊂ ρ , with parameter s n R s^nR s n R and amplitude α / ( R ε N ( s ) ) \alpha/(R^\varepsilon N(s)) α / ( R ε N ( s )) .
Returning to the original coordinates sends a parent of width s ′ / s s'/s s ′ / s to a cap ρ ′ ⊂ ρ \rho'\subset\rho ρ ′ ⊂ ρ of width s ′ s' s ′ , where R − 1 / n ≤ s ′ ≤ s R^{-1/n}\le s'\le s R − 1/ n ≤ s ′ ≤ s . Its averaging box returns to V ∥ V ρ ′ , R V\parallel V_{\rho',R} V ∥ V ρ ′ , R , since its side length is
s − i ( s n R ) 2 / n ( s ′ / s ) 2 − i = R 2 / n ( s ′ ) 2 − i , 1 ≤ i ≤ n . s^{-i}(s^nR)^{2/n}(s'/s)^{2-i}=R^{2/n}(s')^{2-i},\qquad1\le i\le n. s − i ( s n R ) 2/ n ( s ′ / s ) 2 − i = R 2/ n ( s ′ ) 2 − i , 1 ≤ i ≤ n .
Using N ( s ) N ρ ( s ′ ) ≲ ε N ( s ′ ) N(s)N_\rho(s')\lesssim_\varepsilon N(s') N ( s ) N ρ ( s ′ ) ≲ ε N ( s ′ ) from Lemma 2.18 in the selection condition and summing over ρ \rho ρ gives the first sum on the right of (4.1).
There are at most N − j + 1 ≤ η − 1 N-j+1\le\eta^{-1} N − j + 1 ≤ η − 1 steps. Pigeonholing in j ′ j' j ′ and σ ′ \sigma' σ ′ , with a sufficiently small loss at each step, gives (4.1). For the original functions, repeat the arguments of Cases 1 and 2 with f τ f_\tau f τ in place of f τ j f_{\tau}^{j} f τ j , taking the low frequency cutoff at R − 1 / n R^{-1/n} R − 1/ n . □
Refining either σ \sigma σ or s s s
Since the first sum on the right of (4.1) already involves the original functions f θ f_\theta f θ , it remains to estimate the second sum, which involves the pruned pieces. Lemma 4.2 bounds this sum over B B B by two Q Q Q sums of the form on the left of (4.1), one with σ \sigma σ reduced to σ ′ \sigma' σ ′ and the other with s s s reduced to s ′ s' s ′ . In the proof of Proposition 2.12, we can therefore apply Lemma 4.1 again at the smaller scales. Since σ , s ≥ R − 1 / n \sigma,s\ge R^{-1/n} σ , s ≥ R − 1/ n , neither width can decrease indefinitely. In Section 5 we use this fact to advance the pruning index j j j to N N N and bound the sum over B B B in terms of the original functions f θ f_\theta f θ .
Lemma 4.2. Let 1 ≤ j ≤ N 1\le j\le N 1 ≤ j ≤ N and R j − 1 / n ≤ σ ≤ R j − 1 − 1 / n R_j^{-1/n}\le\sigma\le R_{j-1}^{-1/n} R j − 1/ n ≤ σ ≤ R j − 1 − 1/ n . For every s ∈ [ σ , 1 ] s\in[\sigma,1] s ∈ [ σ , 1 ] and α > 0 \alpha>0 α > 0 ,
∑ ρ ∈ Θ ( s ) ∑ B ∈ G ρ ( α ) ∣ B ∣ ∥ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ ) 1 / 2 ∥ L # p n / 2 ( B ) p n ≲ ε R ε ∑ ρ ∈ Θ ( s ) ∑ Q ∈ G ρ ( α R ε ) ∣ Q ∣ ∥ ( ∑ τ ′ ⊂ ρ ∣ τ ′ ∣ = σ ∣ f τ ′ j ∣ 2 ∗ w τ ′ ) 1 / 2 ∥ L # p n / 2 ( Q ) p n + R ε ∑ ρ ′ ∈ Θ ( s ′ ) ∑ Q ∈ G ρ ′ ( α R ε ) ∣ Q ∣ ∥ ( ∑ τ ′ ⊂ ρ ′ ∣ τ ′ ∣ = σ ∣ f τ ′ j ∣ 2 ∗ w τ ′ ) 1 / 2 ∥ L # p n / 2 ( Q ) p n . (49) \begin{aligned}
&\sum_{\rho\in\Theta(s)}\sum_{B\in\mathcal{G}_\rho(\alpha)}|B|\left\|\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_\tau^j|^2*w_\tau\right)^{1/2}\right\|_{L_\#^{p_n/2}(B)}^{p_n} \\
&\lesssim_\varepsilon R^\varepsilon\sum_{\rho\in\Theta(s)}\sum_{Q\in\mathcal{G}_\rho\left(\frac{\alpha}{R^\varepsilon}\right)}|Q|\left\|\left(\sum_{\substack{\tau'\subset\rho\\|\tau'|=\sigma}}|f_{\tau'}^j|^2*w_{\tau'}\right)^{1/2}\right\|_{L_\#^{p_n/2}(Q)}^{p_n} \\
&\quad+R^\varepsilon\sum_{\rho'\in\Theta(s')}\sum_{Q\in\mathcal{G}_{\rho'}\left(\frac{\alpha}{R^\varepsilon}\right)}|Q|\left\|\left(\sum_{\substack{\tau'\subset\rho'\\|\tau'|=\sigma}}|f_{\tau'}^j|^2*w_{\tau'}\right)^{1/2}\right\|_{L_\#^{p_n/2}(Q)}^{p_n}. \tag*{(49)}
\end{aligned} ρ ∈ Θ ( s ) ∑ B ∈ G ρ ( α ) ∑ ∣ B ∣ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ j ∣ 2 ∗ w τ 1/2 L # p n /2 ( B ) p n ≲ ε R ε ρ ∈ Θ ( s ) ∑ Q ∈ G ρ ( R ε α ) ∑ ∣ Q ∣ τ ′ ⊂ ρ ∣ τ ′ ∣ = σ ∑ ∣ f τ ′ j ∣ 2 ∗ w τ ′ 1/2 L # p n /2 ( Q ) p n + R ε ρ ′ ∈ Θ ( s ′ ) ∑ Q ∈ G ρ ′ ( R ε α ) ∑ ∣ Q ∣ τ ′ ⊂ ρ ′ ∣ τ ′ ∣ = σ ∑ ∣ f τ ′ j ∣ 2 ∗ w τ ′ 1/2 L # p n /2 ( Q ) p n . (49)
where σ ′ = max { R j − 1 / n , σ 2 / s } \sigma'=\max\{R_j^{-1/n},\sigma^2/s\} σ ′ = max { R j − 1/ n , σ 2 / s } and s ′ = s ( σ / s ) 1 / ( n − 1 ) ∈ [ σ , s ] s'=s(\sigma/s)^{1/(n-1)}\in[\sigma,s] s ′ = s ( σ / s ) 1/ ( n − 1 ) ∈ [ σ , s ] . For the original grouped functions f τ f_\tau f τ , the same estimate holds for R − 1 / n ≤ σ ≤ s ≤ 1 R^{-1/n}\le\sigma\le s\le1 R − 1/ n ≤ σ ≤ s ≤ 1 , with j j j omitted and σ ′ = max { R − 1 / n , σ 2 / s } \sigma'=\max\{R^{-1/n},\sigma^2/s\} σ ′ = max { R − 1/ n , σ 2 / s } .
The high frequency part of the proof uses the following square function estimate for the cone over the moment curve.
Proposition 4.3 ([10 Theorem 1 and Proposition 1.1] ). For a Schwartz function g g g with supp g ^ ⊂ Γ n ( σ − ( n − 1 ) ) \operatorname{supp}\widehat{g} \subset\Gamma^{n}(\sigma^{-(n-1)}) supp g ⊂ Γ n ( σ − ( n − 1 ) ) and 2 ≤ p ≤ p n − 1 2 \le p \le p_{n-1} 2 ≤ p ≤ p n − 1 , we have
∫ R n ∣ g ∣ p ≲ ϵ σ − ϵ ∫ R n ( ∑ τ ∈ Ξ ( σ ) ∣ g τ ∣ 2 ) p / 2 . \int_{\mathbb{R}^{n}} |g|^{p} \lesssim_{\epsilon} \sigma^{-\epsilon} \int_{\mathbb{R}^{n}} \left(\sum_{\tau\in\Xi(\sigma)} |g_{\tau}|^{2}\right)^{p/2}. ∫ R n ∣ g ∣ p ≲ ϵ σ − ϵ ∫ R n τ ∈ Ξ ( σ ) ∑ ∣ g τ ∣ 2 p /2 .
Proof of Lemma 4.2. For each ρ ∈ Θ ( s ) \rho\in\Theta(s) ρ ∈ Θ ( s ) , decompose ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ \sum_{\tau\subset\rho,\ |\tau|=\sigma} |f_{\tau}^{j}|^{2} * w_{\tau} ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ using the cutoffs η ≤ σ ′ \eta_{\leq\sigma'} η ≤ σ ′ and η λ \eta_{\lambda} η λ , where σ ′ < λ ≲ σ \sigma' < \lambda\lesssim\sigma σ ′ < λ ≲ σ . We split the sum over B B B on the left of (49) according to whether the low frequency part controls the average over B B B .
Case 1: Low frequency part
Consider the boxes B ∈ G ρ ( α ) B \in\mathcal{G}_{\rho}(\alpha) B ∈ G ρ ( α ) , B ∥ ( s / σ ) V ρ , σ − n B \parallel(s/\sigma)V_{\rho,\sigma^{-n}} B ∥ ( s / σ ) V ρ , σ − n , for which
∫ B ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ ) p n / 4 ≲ ∫ B ∣ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ ) ∗ η ˇ ≤ σ ′ ∣ p n / 4 . \int_{B}\left(\sum_{\substack{\tau\subset\rho\\ |\tau|=\sigma}} |f_{\tau}^{j}|^{2} * w_{\tau}\right)^{p_{n}/4}
\lesssim
\int_{B}\left|\left(\sum_{\substack{\tau\subset\rho\\ |\tau|=\sigma}} |f_{\tau}^{j}|^{2} * w_{\tau}\right) * \check{\eta}_{\leq\sigma'}\right|^{p_{n}/4}. ∫ B τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ j ∣ 2 ∗ w τ p n /4 ≲ ∫ B τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ j ∣ 2 ∗ w τ ∗ η ˇ ≤ σ ′ p n /4 .
When we expand each f τ j f_{\tau}^{j} f τ j into its σ ′ \sigma' σ ′ pieces, the cutoff η ≤ σ ′ \eta_{\leq\sigma'} η ≤ σ ′ selects only products of neighbouring pieces. As in Case 2 of the proof of Lemma 4.1, this gives
∣ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ ) ∗ η ≤ σ ′ ∣ ≲ ∑ τ ′ ⊂ ρ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ j ∣ 2 ∗ w τ ′ . \left|\left(\sum_{\substack{\tau\subset\rho\\ |\tau|=\sigma}} |f_{\tau}^{j}|^{2} * w_{\tau}\right) * \eta_{\leq\sigma'}\right|
\lesssim
\sum_{\substack{\tau' \subset\rho\\ |\tau'|=\sigma'}} |f_{\tau'}^{j}|^{2} * w_{\tau'}. τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ j ∣ 2 ∗ w τ ∗ η ≤ σ ′ ≲ τ ′ ⊂ ρ ∣ τ ′ ∣ = σ ′ ∑ ∣ f τ ′ j ∣ 2 ∗ w τ ′ .
Since B 0 = ( s / σ ) V ρ , σ − n = ( s / σ ) 3 ρ ∗ B_{0}=(s/\sigma)V_{\rho,\sigma^{-n}}=(s/\sigma)^{3}\rho^{*} B 0 = ( s / σ ) V ρ , σ − n = ( s / σ ) 3 ρ ∗ and s ≥ σ s\geq\sigma s ≥ σ , we apply Lemma 2.14 to the sum over these B B B , with F = ( ∑ τ ′ ⊂ ρ , ∣ τ ′ ∣ = σ ′ ∣ f τ ′ j ∣ 2 ∗ w τ ′ ) p n / 4 F=(\sum_{\tau'\subset\rho,\ |\tau'|=\sigma'} |f_{\tau'}^{j}|^{2}*w_{\tau'})^{p_{n}/4} F = ( ∑ τ ′ ⊂ ρ , ∣ τ ′ ∣ = σ ′ ∣ f τ ′ j ∣ 2 ∗ w τ ′ ) p n /4 and U 0 = ρ ∗ U_{0}=\rho^{*} U 0 = ρ ∗ . Summing over ρ \rho ρ gives the first sum over Q Q Q in (49) .
Case 2: High frequency part
Fix ρ ∈ Θ ( s ) \rho\in\Theta(s) ρ ∈ Θ ( s ) and let t ρ t_{\rho} t ρ be the left endpoint of the interval indexing ρ \rho ρ . For each box B ∈ G ρ ( α ) B \in\mathcal{G}_{\rho}(\alpha) B ∈ G ρ ( α ) not covered by Case 1, dyadic pigeonholing gives a width λ \lambda λ , σ ′ < λ ≲ σ \sigma' < \lambda\lesssim\sigma σ ′ < λ ≲ σ , for which
∫ B ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ ) p n / 4 ≲ ∫ B ∣ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ ) ∗ η ˇ λ ∣ p n / 4 . (50) \int_{B}\left(\sum_{\substack{\tau\subset\rho\\ |\tau|=\sigma}} |f_{\tau}^{j}|^{2} * w_{\tau}\right)^{p_{n}/4}
\lesssim
\int_{B}\left|\left(\sum_{\substack{\tau\subset\rho\\ |\tau|=\sigma}} |f_{\tau}^{j}|^{2} * w_{\tau}\right) * \check{\eta}_{\lambda}\right|^{p_{n}/4}. \tag*{(50)} ∫ B τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ j ∣ 2 ∗ w τ p n /4 ≲ ∫ B τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ j ∣ 2 ∗ w τ ∗ η ˇ λ p n /4 . (50)
Fix λ \lambda λ and consider the boxes B B B satisfying (50) . To estimate its right hand side, we group the σ \sigma σ caps in ρ \rho ρ into parents ρ ′ \rho' ρ ′ of width s ′ = s ( σ / s ) 1 / ( n − 1 ) s'=s(\sigma/s)^{1/(n-1)} s ′ = s ( σ / s ) 1/ ( n − 1 ) and write
( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ ) ∗ η ˇ λ = ∑ ρ ′ ⊂ ρ ∣ ρ ′ ∣ = s ′ h ρ ′ , h ρ ′ = ( ∑ τ ⊂ ρ ′ ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ ) ∗ η ˇ λ . \left(\sum_{\substack{\tau\subset\rho\\ |\tau|=\sigma}} |f_{\tau}^{j}|^{2} * w_{\tau}\right) * \check{\eta}_{\lambda}
=
\sum_{\substack{\rho' \subset\rho\\ |\rho'|=s'}} h_{\rho'},
\qquad
h_{\rho'}=
\left(\sum_{\substack{\tau\subset\rho'\\ |\tau|=\sigma}} |f_{\tau}^{j}|^{2} * w_{\tau}\right) * \check{\eta}_{\lambda}. τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ j ∣ 2 ∗ w τ ∗ η ˇ λ = ρ ′ ⊂ ρ ∣ ρ ′ ∣ = s ′ ∑ h ρ ′ , h ρ ′ = τ ⊂ ρ ′ ∣ τ ∣ = σ ∑ ∣ f τ j ∣ 2 ∗ w τ ∗ η ˇ λ .
When we expand each f τ j f_{\tau}^{j} f τ j into its R j − 1 / n R_{j}^{-1/n} R j − 1/ n pieces, the cutoff η λ \eta_{\lambda} η λ selects only pairs indexed by intervals at distance O ( λ ) O(\lambda) O ( λ ) from each other. Since R j − 1 / n ≤ λ R_{j}^{-1/n}\leq\lambda R j − 1/ n ≤ λ , the Taylor expansion in the proof of Lemma 4.1 gives, for every ξ \xi ξ in the Fourier support of ∣ f τ j ∣ 2 ∗ w τ ∗ η ˇ λ |f_{\tau}^{j}|^{2}*w_{\tau}*\check{\eta}_{\lambda} ∣ f τ j ∣ 2 ∗ w τ ∗ η ˇ λ ,
∣ ξ ∣ ∼ λ , ξ = ∑ i = 1 n a i γ ( i ) ( t τ ) , ∣ a i ∣ ≲ λ σ i − 1 ( 1 ≤ i ≤ n ) . |\xi|\sim\lambda,\qquad
\xi=\sum_{i=1}^{n}a_{i}\gamma^{(i)}(t_{\tau}),\qquad
|a_{i}|\lesssim\lambda\sigma^{i-1}\quad(1\leq i\leq n). ∣ ξ ∣ ∼ λ , ξ = i = 1 ∑ n a i γ ( i ) ( t τ ) , ∣ a i ∣ ≲ λ σ i − 1 ( 1 ≤ i ≤ n ) .
For τ ⊂ ρ ′ \tau\subset\rho' τ ⊂ ρ ′ , we have σ ≤ s ′ \sigma\leq s' σ ≤ s ′ and ∣ t τ − t ρ ′ ∣ ≤ s ′ |t_{\tau}-t_{\rho'}|\leq s' ∣ t τ − t ρ ′ ∣ ≤ s ′ . Expanding at t ρ ′ t_{\rho'} t ρ ′ as in Case 1 of the proof of Lemma 2.8, we obtain
ξ ∈ supp h ^ ρ ′ ⟹ ξ = ∑ i = 1 n b i γ ( i ) ( t ρ ′ ) , ∣ b i ∣ ≲ λ ( s ′ ) i − 1 ( 1 ≤ i ≤ n ) . \xi\in\operatorname{supp}\widehat{h}_{\rho'}
\Longrightarrow
\xi=\sum_{i=1}^{n}b_{i}\gamma^{(i)}(t_{\rho'}),\qquad
|b_{i}|\lesssim\lambda(s')^{i-1}\quad(1\leq i\leq n). ξ ∈ supp h ρ ′ ⟹ ξ = i = 1 ∑ n b i γ ( i ) ( t ρ ′ ) , ∣ b i ∣ ≲ λ ( s ′ ) i − 1 ( 1 ≤ i ≤ n ) .
As in Lemma 4.1, we follow dilation by λ − 1 \lambda^{-1} λ − 1 with the rescaling s L t ρ , s − 1 sL^{-1}_{t_{\rho,s}} s L t ρ , s − 1 that sends [ t ρ , t ρ + s ] [t_{\rho},t_{\rho}+s] [ t ρ , t ρ + s ] to [ 0 , 1 ] [0,1] [ 0 , 1 ] . The widths of each piece change as follows:
( λ × λ s ′ × ⋯ × λ ( s ′ ) n − 1 ) → λ − 1 d i l a t i o n ( 1 × s ′ × ⋯ × ( s ′ ) n − 1 ) → s L t ρ , s − 1 ( 1 × s ′ s × ⋯ × ( s ′ s ) n − 1 ) . \left(\lambda\times\lambda s'\times\cdots\times\lambda(s')^{n-1}\right)
\xrightarrow{\lambda^{-1}\ \mathrm{dilation}}
\left(1\times s'\times\cdots\times(s')^{n-1}\right)
\xrightarrow{sL^{-1}_{t_{\rho,s}}}
\left(1\times\frac{s'}{s}\times\cdots\times\left(\frac{s'}{s}\right)^{n-1}\right). ( λ × λ s ′ × ⋯ × λ ( s ′ ) n − 1 ) λ − 1 dilation ( 1 × s ′ × ⋯ × ( s ′ ) n − 1 ) s L t ρ , s − 1 ( 1 × s s ′ × ⋯ × ( s s ′ ) n − 1 ) .
Thus the rescaled h ρ ′ h_{\rho'} h ρ ′ have Fourier support in caps of aperture s ′ / s s'/s s ′ / s .
For each ρ ′ ⊂ ρ \rho' \subset\rho ρ ′ ⊂ ρ of width s ′ s' s ′ , let a ρ ′ = ∣ η λ ˇ ∣ ∗ ∑ τ ⊂ ρ ′ , ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ a_{\rho'} = |\check{\eta_\lambda}| * \sum_{\tau\subset\rho',\,|\tau|=\sigma} |f_\tau^j|^2 * w_\tau a ρ ′ = ∣ η λ ˇ ∣ ∗ ∑ τ ⊂ ρ ′ , ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ , so that ∣ h ρ ′ ∣ ≤ a ρ ′ |h_{\rho'}| \le a_{\rho'} ∣ h ρ ′ ∣ ≤ a ρ ′ . To obtain the second sum on the right of (49) , we estimate the high frequency average on the right of (50) . We claim that, for each B B B satisfying (50) ,
\left[\fint_B \left|\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}} |f_\tau^j|^2 * w_\tau\right) * \check{\eta_\lambda}\right|^{p_n/4}\right]^2 \lesssim_{\epsilon} \frac{R^\epsilon}{|B|} \sum_{\substack{\rho'\subset\rho\\|\rho'|=s'}} \int_{a_{\rho'}>\left[\alpha/(R^\epsilon N(s))\right]^2} a_{\rho'}^{p_n/2} W_B. \tag*{(51)}
Once we prove (51) , we combine it with (50) , multiply by ∣ B ∣ |B| ∣ B ∣ , and sum over the boxes for which this λ \lambda λ was chosen. Since ∑ B W B ≲ 1 \sum_B W_B \lesssim1 ∑ B W B ≲ 1 , we apply (34) with F = ∑ τ ⊂ ρ ′ , ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ F=\sum_{\tau\subset\rho',\,|\tau|=\sigma} |f_\tau^j|^2 * w_\tau F = ∑ τ ⊂ ρ ′ , ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ , K = ∣ η λ ˇ ∣ K=|\check{\eta_\lambda}| K = ∣ η λ ˇ ∣ , and exponent p n / 2 p_n/2 p n /2 to remove the convolution from a ρ ′ a_{\rho'} a ρ ′ . The locally constant property on Q ∥ ( ρ ′ ) ∗ Q \parallel(\rho')^* Q ∥ ( ρ ′ ) ∗ and N ( s ′ ) ≥ N ( s ) N(s')\ge N(s) N ( s ′ ) ≥ N ( s ) then give the second sum on the right of (49) after summing over ρ \rho ρ and λ \lambda λ .
It remains to prove (51) . When n ≥ 4 n\ge4 n ≥ 4 , we use Proposition 4.3 at p = p n / 4 p=p_n/4 p = p n /4 , since 2 ≤ p n / 4 ≤ p n − 1 2\le p_n/4\le p_{n-1} 2 ≤ p n /4 ≤ p n − 1 . When n = 3 n=3 n = 3 , Proposition 4.3 requires p ≥ 2 p\ge2 p ≥ 2 , so we first use Hölder on B B B to pass from p 3 / 4 = 7 / 4 p_3/4=7/4 p 3 /4 = 7/4 to p = 7 / 3 p=7/3 p = 7/3 .
First suppose n ≥ 4 n\ge4 n ≥ 4 . Choose χ B ∼ W B 4 / p n \chi_B\sim W_B^{4/p_n} χ B ∼ W B 4/ p n with supp χ B ⊂ B ∗ \operatorname{supp}\chi_B\subset B^* supp χ B ⊂ B ∗ . By the definition of h ρ ′ h_{\rho'} h ρ ′ , the left hand side of (51) satisfies
\left[\fint_B \left|\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}} |f_\tau^j|^2 * w_\tau\right) * \check{\eta_\lambda}\right|^{p_n/4}\right]^2 \sim\left[\frac{1}{|B|}\int\chi_B\sum_{\substack{\rho'\subset\rho\\|\rho'|=s'}} |h_{\rho'}|^{p_n/4}\right]^2. \tag*{(52)}
To apply Proposition 4.3 to the integral on the right of (52) , we check the Fourier support of each χ B h ρ ′ \chi_Bh_{\rho'} χ B h ρ ′ after rescaling. Multiplication by χ B \chi_B χ B enlarges the Fourier support of h ρ ′ h_{\rho'} h ρ ′ by at most the widths of B ∗ B^* B ∗ , namely σ 3 s i − 3 \sigma^3s^{i-3} σ 3 s i − 3 in direction i i i . Under the frequency map ξ ↦ ( s / λ ) L t ρ , s T ξ \xi\mapsto(s/\lambda)L_{t_{\rho,s}}^T\xi ξ ↦ ( s / λ ) L t ρ , s T ξ , the added width in each direction is at most
σ 3 s i − 3 λ s i − 1 = σ 3 λ s 2 ≤ σ s = ( s ′ s ) n − 1 , \frac{\sigma^3s^{i-3}}{\lambda s^{i-1}}=\frac{\sigma^3}{\lambda s^2}\le\frac{\sigma}{s}=\left(\frac{s'}{s}\right)^{n-1}, λ s i − 1 σ 3 s i − 3 = λ s 2 σ 3 ≤ s σ = ( s s ′ ) n − 1 ,
where we used λ > σ ′ ≥ σ 2 / s \lambda>\sigma'\ge\sigma^2/s λ > σ ′ ≥ σ 2 / s . Since ( s ′ / s ) n − 1 (s'/s)^{n-1} ( s ′ / s ) n − 1 is the smallest width of a cap of aperture s ′ / s s'/s s ′ / s , the Fourier support of ( χ B h ρ ′ ) ( ( s / λ ) L t ρ , s T ⋅ ) (\chi_Bh_{\rho'})((s/\lambda)L_{t_{\rho,s}}^T\cdot) ( χ B h ρ ′ ) (( s / λ ) L t ρ , s T ⋅ ) remains in the corresponding caps for Γ n ( ( s / s ′ ) n − 1 ) \Gamma^n((s/s')^{n-1}) Γ n (( s / s ′ ) n − 1 ) , up to a fixed enlargement.
We may therefore apply Proposition 4.3 at p = p n / 4 p=p_n/4 p = p n /4 to the sum in the integral on the right of (52) after rescaling, that is, to x ↦ [ χ B ∑ ρ ′ ⊂ ρ , ∣ ρ ′ ∣ = s ′ h ρ ′ ] ( ( s / λ ) L t ρ , s T x ) x\mapsto\left[\chi_B\sum_{\rho'\subset\rho,\,|\rho'|=s'}h_{\rho'}\right]((s/\lambda)L_{t_{\rho,s}}^Tx) x ↦ [ χ B ∑ ρ ′ ⊂ ρ , ∣ ρ ′ ∣ = s ′ h ρ ′ ] (( s / λ ) L t ρ , s T x ) , with aperture s ′ / s s'/s s ′ / s . On returning to the original coordinates, the square function integrand is ∣ χ B ∣ p n / 4 ( ∑ ρ ′ ⊂ ρ , ∣ ρ ′ ∣ = s ′ ∣ h ρ ′ ∣ 2 ) p n / 8 |\chi_B|^{p_n/4}\left(\sum_{\rho'\subset\rho,\,|\rho'|=s'}|h_{\rho'}|^2\right)^{p_n/8} ∣ χ B ∣ p n /4 ( ∑ ρ ′ ⊂ ρ , ∣ ρ ′ ∣ = s ′ ∣ h ρ ′ ∣ 2 ) p n /8 . Using ∣ χ B ∣ p n / 4 ∼ W B |\chi_B|^{p_n/4}\sim W_B ∣ χ B ∣ p n /4 ∼ W B in (52) , we obtain
\left[\fint_B \left|\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}} |f_\tau^j|^2 * w_\tau\right) * \check{\eta_\lambda}\right|^{p_n/4}\right]^2 \lesssim_{\epsilon} R^\epsilon\left[\fint_B\left(\sum_{\substack{\rho'\subset\rho\\|\rho'|=s'}}|h_{\rho'}|^2\right)^{p_n/8}\right]^2. \tag*{(53)}
On the right of (53) , use ∣ h ρ ′ ∣ ≤ a ρ ′ |h_{\rho'}|\le a_{\rho'} ∣ h ρ ′ ∣ ≤ a ρ ′ and split ∑ ρ ′ a ρ ′ 2 \sum_{\rho'}a_{\rho'}^2 ∑ ρ ′ a ρ ′ 2 at the threshold [ α / ( R ϵ N ( s ) ) ] 2 [\alpha/(R^\epsilon N(s))]^2 [ α / ( R ϵ N ( s )) ] 2 in (51) . The sum below the threshold satisfies
∑ a ρ ′ ≤ [ α / ( R ϵ N ( s ) ) ] 2 a ρ ′ 2 ≤ ( α R ϵ N ( s ) ) 2 ∑ ρ ′ a ρ ′ . \sum_{a_{\rho'}\le[\alpha/(R^\epsilon N(s))]^2}a_{\rho'}^2\le\left(\frac{\alpha}{R^\epsilon N(s)}\right)^2\sum_{\rho'}a_{\rho'}. a ρ ′ ≤ [ α / ( R ϵ N ( s )) ] 2 ∑ a ρ ′ 2 ≤ ( R ϵ N ( s ) α ) 2 ρ ′ ∑ a ρ ′ .
Raising this inequality to the power p n / 8 p_n/8 p n /8 , averaging over B B B , and applying Hölder gives
∫ B ( ∑ a ρ ′ ≤ [ α / ( R ε N ( s ) ) ] 2 a ρ ′ 2 ) p n / 8 ≲ ( α R ε N ( s ) ) p n / 4 [ ∫ B ( ∑ ρ ′ a ρ ′ ) p n / 4 ] 1 / 2 = ( α R ε N ( s ) ) p n / 4 [ ∫ B ( ∣ λ ~ ∣ ∗ ∑ τ ∈ C ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ ) p n / 4 ] 1 / 2 \begin{aligned}
\int_B\left(\sum_{a_{\rho'}\leq[\alpha/(R^{\varepsilon}N(s))]^2}a_{\rho'}^2\right)^{p_n/8}\lesssim\left(\frac{\alpha}{R^{\varepsilon}N(s)}\right)^{p_n/4}\left[\int_B\left(\sum_{\rho'}a_{\rho'}\right)^{p_n/4}\right]^{1/2} \\
&=\left(\frac{\alpha}{R^{\varepsilon}N(s)}\right)^{p_n/4}\left[\int_B\left(|\widetilde{\lambda}|*\sum_{\substack{\tau\in\mathcal{C}_{\rho}\\|\tau|=\sigma}}|f_\tau|^2*w_\tau\right)^{p_n/4}\right]^{1/2}
\end{aligned} ∫ B a ρ ′ ≤ [ α / ( R ε N ( s )) ] 2 ∑ a ρ ′ 2 p n /8 ≲ ( R ε N ( s ) α ) p n /4 ∫ B ρ ′ ∑ a ρ ′ p n /4 1/2 = ( R ε N ( s ) α ) p n /4 ∫ B ∣ λ ∣ ∗ τ ∈ C ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ p n /4 1/2
≲ ( α R ε N ( s ) ) p n / 4 [ ∫ B ( ∑ τ ∈ C ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ ) p n / 4 ] 1 / 2 . \lesssim\left(\frac{\alpha}{R^{\varepsilon}N(s)}\right)^{p_n/4}\left[\int_B\left(\sum_{\substack{\tau\in\mathcal{C}_{\rho}\\|\tau|=\sigma}}|f_\tau|^2*w_\tau\right)^{p_n/4}\right]^{1/2}. ≲ ( R ε N ( s ) α ) p n /4 ∫ B τ ∈ C ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ p n /4 1/2 .
In the last inequality, we used λ − 1 ≤ σ − 3 s 2 \lambda^{-1}\leq\sigma^{-3}s^2 λ − 1 ≤ σ − 3 s 2 , the shortest side of B B B , so that ∣ λ ~ ∣ ∗ W B ≲ W B |\widetilde{\lambda}|*W_B\lesssim W_B ∣ λ ∣ ∗ W B ≲ W B , and applied (2.26) with F = ∑ τ ∈ C ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ F=\sum_{\tau\in\mathcal{C}_{\rho},|\tau|=\sigma}|f_\tau|^2*w_\tau F = ∑ τ ∈ C ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , K = ∣ λ ~ ∣ K=|\widetilde{\lambda}| K = ∣ λ ∣ , and r = p n / 4 r=p_n/4 r = p n /4 .
Since B ∈ B ρ ( α ) B\in\mathcal{B}_{\rho}(\alpha) B ∈ B ρ ( α ) , we have ∫ B ( ∑ τ ∈ C ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ ) p n / 4 ≥ ( α / N ( s ) ) p n / 2 \int_B(\sum_{\tau\in\mathcal{C}_{\rho},|\tau|=\sigma}|f_\tau|^2*w_\tau)^{p_n/4}\geq(\alpha/N(s))^{p_n/2} ∫ B ( ∑ τ ∈ C ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ ) p n /4 ≥ ( α / N ( s ) ) p n /2 . Squaring the preceding bound, we obtain
[ ∫ B ( ∑ a ρ ′ ≤ [ α / ( R ε N ( s ) ) ] 2 a ρ ′ 2 ) p n / 8 ] 2 ≲ R − ε p n / 2 [ ∫ B ( ∑ τ ∈ C ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ ) p n / 4 ] 2 \left[\int_B\left(\sum_{a_{\rho'}\leq[\alpha/(R^{\varepsilon}N(s))]^2}a_{\rho'}^2\right)^{p_n/8}\right]^2\lesssim R^{-\varepsilon p_n/2}\left[\int_B\left(\sum_{\substack{\tau\in\mathcal{C}_{\rho}\\|\tau|=\sigma}}|f_\tau|^2*w_\tau\right)^{p_n/4}\right]^2 ∫ B a ρ ′ ≤ [ α / ( R ε N ( s )) ] 2 ∑ a ρ ′ 2 p n /8 2 ≲ R − ε p n /2 ∫ B τ ∈ C ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ p n /4 2
≲ R − ( p n / 2 − 3 ) ε [ ∫ B ( ∑ ρ ′ ∣ h ρ ′ ∣ 2 ) p n / 8 ] 2 , \lesssim R^{-(p_n/2-3)\varepsilon}\left[\int_B\left(\sum_{\rho'}|h_{\rho'}|^2\right)^{p_n/8}\right]^2, ≲ R − ( p n /2 − 3 ) ε ∫ B ρ ′ ∑ ∣ h ρ ′ ∣ 2 p n /8 2 ,
where we used the square of (4.6) and then (4.9) in the last inequality. Since p n / 2 − 3 > 0 p_n/2-3>0 p n /2 − 3 > 0 , for sufficiently large R R R we can absorb the terms below the threshold and replace ∑ ρ ′ ∣ h ρ ′ ∣ 2 \sum_{\rho'}|h_{\rho'}|^2 ∑ ρ ′ ∣ h ρ ′ ∣ 2 on the right of (4.9) by ∑ a ρ ′ > [ α / ( R ε N ( s ) ) ] 2 a ρ ′ 2 \sum_{a_{\rho'}>[\alpha/(R^{\varepsilon}N(s))]^2}a_{\rho'}^2 ∑ a ρ ′ > [ α / ( R ε N ( s )) ] 2 a ρ ′ 2 .
For the remaining terms, we use
∑ a ρ ′ > [ α / ( R ε N ( s ) ) ] 2 a ρ ′ 2 ≤ ( ∑ ρ ′ a ρ ′ ) ( ∑ a ρ ′ > [ α / ( R ε N ( s ) ) ] 2 a ρ ′ p n / 2 ) 2 / p n . \sum_{a_{\rho'}>[\alpha/(R^{\varepsilon}N(s))]^2}a_{\rho'}^2\leq\left(\sum_{\rho'}a_{\rho'}\right)\left(\sum_{a_{\rho'}>[\alpha/(R^{\varepsilon}N(s))]^2}a_{\rho'}^{p_n/2}\right)^{2/p_n}. a ρ ′ > [ α / ( R ε N ( s )) ] 2 ∑ a ρ ′ 2 ≤ ρ ′ ∑ a ρ ′ a ρ ′ > [ α / ( R ε N ( s )) ] 2 ∑ a ρ ′ p n /2 2/ p n .
Raising this inequality to the power p n / 8 p_n/8 p n /8 , applying Hölder in the B B B average, and squaring gives
[ ∫ B ( ∑ a ρ ′ > [ α / ( R ε N ( s ) ) ] 2 a ρ ′ 2 ) p n / 8 ] 2 ≲ ∫ B ( ∑ ρ ′ a ρ ′ ) p n / 4 [ ∫ B ∑ a ρ ′ > [ α / ( R ε N ( s ) ) ] 2 a ρ ′ p n / 2 ] 1 / 2 . \left[\int_B\left(\sum_{a_{\rho'}>[\alpha/(R^{\varepsilon}N(s))]^2}a_{\rho'}^2\right)^{p_n/8}\right]^2\lesssim\int_B\left(\sum_{\rho'}a_{\rho'}\right)^{p_n/4}\left[\int_B\sum_{a_{\rho'}>[\alpha/(R^{\varepsilon}N(s))]^2}a_{\rho'}^{p_n/2}\right]^{1/2}. ∫ B a ρ ′ > [ α / ( R ε N ( s )) ] 2 ∑ a ρ ′ 2 p n /8 2 ≲ ∫ B ρ ′ ∑ a ρ ′ p n /4 ∫ B a ρ ′ > [ α / ( R ε N ( s )) ] 2 ∑ a ρ ′ p n /2 1/2 .
As above, (2.26) bounds ∫ B ( ∑ ρ ′ a ρ ′ ) p n / 4 \int_B(\sum_{\rho'}a_{\rho'})^{p_n/4} ∫ B ( ∑ ρ ′ a ρ ′ ) p n /4 by ∫ B ( ∑ τ ∈ C ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ ) p n / 4 \int_B(\sum_{\tau\in\mathcal{C}_{\rho},|\tau|=\sigma}|f_\tau|^2*w_\tau)^{p_n/4} ∫ B ( ∑ τ ∈ C ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ ) p n /4 . By (4.6), this last average is ≲ ∫ B ∣ ∑ ρ ′ h ρ ′ ∣ p n / 4 \lesssim\int_B|\sum_{\rho'}h_{\rho'}|^{p_n/4} ≲ ∫ B ∣ ∑ ρ ′ h ρ ′ ∣ p n /4 . Substituting the resulting bound for the remaining terms on the right of (4.9), we obtain
[ ∫ B ∣ ∑ ρ ′ h ρ ′ ∣ p n / 4 ] 2 ≲ ∫ B ∣ ∑ ρ ′ h ρ ′ ∣ p n / 4 [ ∫ B ∑ a ρ ′ > [ α / ( R ε N ( s ) ) ] 2 a ρ ′ p n / 2 ] 1 / 2 . \left[\int_B\left|\sum_{\rho'}h_{\rho'}\right|^{p_n/4}\right]^2\lesssim\int_B\left|\sum_{\rho'}h_{\rho'}\right|^{p_n/4}\left[\int_B\sum_{a_{\rho'}>[\alpha/(R^{\varepsilon}N(s))]^2}a_{\rho'}^{p_n/2}\right]^{1/2}. ∫ B ρ ′ ∑ h ρ ′ p n /4 2 ≲ ∫ B ρ ′ ∑ h ρ ′ p n /4 ∫ B a ρ ′ > [ α / ( R ε N ( s )) ] 2 ∑ a ρ ′ p n /2 1/2 .
Dividing by ∫ B ∣ ∑ ρ ′ h ρ ′ ∣ p n / 4 \int_B|\sum_{\rho'}h_{\rho'}|^{p_n/4} ∫ B ∣ ∑ ρ ′ h ρ ′ ∣ p n /4 and squaring gives (4.7) for n ≥ 4 n\geq4 n ≥ 4 .
Now suppose n = 3 n=3 n = 3 . Choose χ B ∼ W B 3 / 7 \chi_B\sim W_B^{3/7} χ B ∼ W B 3/7 with supp χ B ^ ⊂ B ∗ \operatorname{supp}\widehat{\chi_B}\subset B^* supp χ B ⊂ B ∗ . Hölder on B B B , followed by Proposition 4.3 at p = 7 / 3 p=7/3 p = 7/3 with the same rescaling as in (4.9), gives
[ ∫ B ∣ ∑ ρ ′ h ρ ′ ∣ 7 / 4 ] 4 / 3 ≲ 1 ∣ B ∣ ∫ ∣ χ B ∑ ρ ′ h ρ ′ ∣ 7 / 3 ≲ ∫ B ( ∑ ρ ′ ∣ h ρ ′ ∣ 2 ) 7 / 6 . \left[\int_B\left|\sum_{\rho'}h_{\rho'}\right|^{7/4}\right]^{4/3}\lesssim\frac{1}{|B|}\int\left|\chi_B\sum_{\rho'}h_{\rho'}\right|^{7/3}\lesssim\int_B\left(\sum_{\rho'}|h_{\rho'}|^2\right)^{7/6}. ∫ B ρ ′ ∑ h ρ ′ 7/4 4/3 ≲ ∣ B ∣ 1 ∫ χ B ρ ′ ∑ h ρ ′ 7/3 ≲ ∫ B ρ ′ ∑ ∣ h ρ ′ ∣ 2 7/6 .
As in the n ≥ 4 n\geq4 n ≥ 4 case, we use (2.26), the selection condition on B B B , and (4.6) to absorb the terms with a ρ ′ ≤ [ α / ( R ε N ( s ) ) ] 2 a_{\rho'}\leq[\alpha/(R^{\varepsilon}N(s))]^2 a ρ ′ ≤ [ α / ( R ε N ( s )) ] 2 into the square function average on the right. For the terms above this threshold, the same pointwise bound followed by Hölder with exponents 3/2 and 3 gives
[ ∫ B ∣ ∑ ρ ′ h ρ ′ ∣ 7 / 4 ] 4 / 3 ≲ [ ∫ B ∣ ∑ ρ ′ h ρ ′ ∣ 7 / 4 ] 2 / 3 [ ∫ B , a ρ ′ ≥ [ α / ( R ϵ N ( s ) ) ] 2 ∑ ρ ′ a ρ ′ 7 / 2 ] 1 / 3 . \left[\int_B\left|\sum_{\rho'}h_{\rho'}\right|^{7/4}\right]^{4/3}
\lesssim
\left[\int_B\left|\sum_{\rho'}h_{\rho'}\right|^{7/4}\right]^{2/3}
\left[\int_{B,\ a_{\rho'}\geq[\alpha/(R^\epsilon N(s))]^2}\sum_{\rho'}a_{\rho'}^{7/2}\right]^{1/3}. ∫ B ρ ′ ∑ h ρ ′ 7/4 4/3 ≲ ∫ B ρ ′ ∑ h ρ ′ 7/4 2/3 ∫ B , a ρ ′ ≥ [ α / ( R ϵ N ( s )) ] 2 ρ ′ ∑ a ρ ′ 7/2 1/3 .
Dividing by the 2/3 power on the right and cubing gives (4.7) for n = 3 n=3 n = 3 . Together with Case 1, this proves (4.5). For the original functions, the same argument applies with R − 1 / n R^{-1/n} R − 1/ n in place of R j − 1 / n R_j^{-1/n} R j − 1/ n and f τ f_\tau f τ in place of f τ j f_\tau^j f τ j throughout. This completes the proof of the lemma. □ \square □
From cones over moment curves to moment curves: Proof of Proposition 2.12
In this section, we prove Proposition 2.12. We follow the algorithm in [18 proof of Proposition 4.6] to bound the left hand side of (2.19), namely
∑ ρ ∈ Θ ( s ) ∑ B ∈ G ρ ( α ) ∣ B ∣ ∥ ( ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ ) 1 / 2 ∥ L # p n / 2 ( B ) p n . (54) \sum_{\rho\in\Theta(s)}\sum_{B\in\mathcal{G}_\rho(\alpha)}|B|\left\|\left(\sum_{|\tau|=\sigma}|f_\tau|^2*w_\tau\right)^{1/2}\right\|_{L_{\#}^{p_n/2}(B)}^{p_n}. \tag*{(54)} ρ ∈ Θ ( s ) ∑ B ∈ G ρ ( α ) ∑ ∣ B ∣ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ 1/2 L # p n /2 ( B ) p n . (54)
At each step, j j j is the pruning index, α \alpha α is the amplitude, and caps of width σ \sigma σ are grouped inside caps of width s s s .
Let us briefly explain the idea. To obtain the right hand side of (2.19), we seek a bound involving the original functions f θ f_\theta f θ , of the form
∑ R − 1 / n ≤ u ≤ s ∑ ρ ′ ∈ Θ ( u ) ∑ V ∈ G ρ ′ ( α ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ′ ∣ f θ ∣ 2 ∗ w θ ) 1 / 2 ∥ L # p n / 2 ( V ) p n . (55) \sum_{R^{-1/n}\leq u\leq s}\sum_{\rho'\in\Theta(u)}\sum_{V\in\mathcal{G}_{\rho'}(\alpha)}|V|\left\|\left(\sum_{\theta\subset\rho'}|f_\theta|^2*w_\theta\right)^{1/2}\right\|_{L_{\#}^{p_n/2}(V)}^{p_n}. \tag*{(55)} R − 1/ n ≤ u ≤ s ∑ ρ ′ ∈ Θ ( u ) ∑ V ∈ G ρ ′ ( α ) ∑ ∣ V ∣ θ ⊂ ρ ′ ∑ ∣ f θ ∣ 2 ∗ w θ 1/2 L # p n /2 ( V ) p n . (55)
Using Lemmas 4.1 and 4.2 with 0 < ϵ ≪ η 0<\epsilon\ll\eta 0 < ϵ ≪ η , together with (2.4) after Taylor rescaling, we either obtain a sum of the form (5.2) and stop, or return to (5.1) with
( j , σ , s , α ) ⟼ ( j ′ , σ ′ , s ′ , α ′ ) , j ′ ≥ j , σ ′ ≤ σ , s ′ ≤ s , (j,\sigma,s,\alpha)\longmapsto(j',\sigma',s',\alpha'),\qquad j'\geq j,\quad\sigma'\leq\sigma,\quad s'\leq s, ( j , σ , s , α ) ⟼ ( j ′ , σ ′ , s ′ , α ′ ) , j ′ ≥ j , σ ′ ≤ σ , s ′ ≤ s ,
where j ′ > j j'>j j ′ > j , σ ′ ≪ σ \sigma'\ll\sigma σ ′ ≪ σ , or s ′ ≪ s s'\ll s s ′ ≪ s . Since σ , s ≥ R − 1 / n \sigma,s\geq R^{-1/n} σ , s ≥ R − 1/ n , the decreases in σ \sigma σ and s s s cannot continue indefinitely. Continuing the iteration therefore advances the pruning index to j = N j=N j = N . The final step then uses ∣ f θ N ∣ ≤ ∣ f θ ∣ |f_\theta^N|\leq|f_\theta| ∣ f θ N ∣ ≤ ∣ f θ ∣ to obtain the averages of the original functions f θ f_\theta f θ in (5.2).
The factor outside the sum and the new amplitude α ′ \alpha' α ′ will be specified in each case. We will bound the accumulated factor by ( R C η S 1 ( R η ) N − j + C ) p n (R^{C\eta}S_1(R^\eta)^{N-j+C})^{p_n} ( R C η S 1 ( R η ) N − j + C ) p n and keep the amplitude in (5.2) at least α / ( R C η S 1 ( R η ) N − j + C ) \alpha/(R^{C\eta}S_1(R^\eta)^{N-j+C}) α / ( R C η S 1 ( R η ) N − j + C ) , with the initial values of j j j and α \alpha α . Let 0 < ϵ ≪ η 0<\epsilon\ll\eta 0 < ϵ ≪ η be sufficiently small. We will show that the iteration has O ( η − 3 ) O(\eta^{-3}) O ( η − 3 ) steps, so all the R ϵ R^\epsilon R ϵ losses arising during the iteration can be absorbed into R C η R^{C\eta} R C η , both outside the sum and in the amplitude denominator.
Let us now give a detailed proof. First consider the case that σ = R − 1 / n \sigma=R^{-1/n} σ = R − 1/ n . In this case, we have j = N j=N j = N . For each ρ \rho ρ , (2.22), with F = ( ∑ θ ⊂ ρ ∣ f θ N ∣ 2 ∗ w θ ) p n / 4 F=\left(\sum_{\theta\subset\rho}|f_\theta^N|^2*w_\theta\right)^{p_n/4} F = ( ∑ θ ⊂ ρ ∣ f θ N ∣ 2 ∗ w θ ) p n /4 and t = ( α / N ( s ) ) p n / 2 t=(\alpha/N(s))^{p_n/2} t = ( α / N ( s ) ) p n /2 , together with ∣ f θ N ∣ ≤ ∣ f θ ∣ |f_\theta^N|\leq|f_\theta| ∣ f θ N ∣ ≤ ∣ f θ ∣ , bounds its sum over B B B in (5.1) by
∫ ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w θ ≳ ( α / N ( s ) ) 2 ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w θ ) p n / 2 . (56) \int_{\sum_{\theta\subset\rho}|f_\theta|^2*w_\theta\gtrsim(\alpha/N(s))^2}
\left(\sum_{\theta\subset\rho}|f_\theta|^2*w_\theta\right)^{p_n/2}. \tag*{(56)} ∫ ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w θ ≳ ( α / N ( s ) ) 2 θ ⊂ ρ ∑ ∣ f θ ∣ 2 ∗ w θ p n /2 . (56)
The Taylor rescaling taking ρ \rho ρ to the unit cap sends each θ ⊂ ρ \theta\subset\rho θ ⊂ ρ to a cap of M n ( s n R ) \mathcal{M}^n(s^nR) M n ( s n R ) at scale R − 1 / n / s R^{-1/n}/s R − 1/ n / s . Applying Lemma 2.8 to { ( ∣ f θ ∣ 2 ∗ w θ ) ( L τ ρ , s T ⋅ ) } θ ⊂ ρ \{(|f_\theta|^2*w_\theta)(L_{\tau_{\rho,s}}^T\cdot)\}_{\theta\subset\rho} {( ∣ f θ ∣ 2 ∗ w θ ) ( L τ ρ , s T ⋅ ) } θ ⊂ ρ , with parameter s n R s^nR s n R and amplitude α / N ( s ) \alpha/N(s) α / N ( s ) , then returning to the original coordinates and summing over ρ \rho ρ , bounds (5.1) by
≲ ϵ , η R ϵ ∑ R − 1 / n ≤ u ≤ s ∑ ρ ′ ∈ Θ ( u ) ∑ V ∈ G ρ ′ ( α / R ϵ ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ′ ∣ f θ ∣ 2 ∗ w θ ) 1 / 2 ∥ L # p n / 2 ( V ) p n , \lesssim_{\epsilon,\eta}R^\epsilon
\sum_{R^{-1/n}\leq u\leq s}\sum_{\rho'\in\Theta(u)}\sum_{V\in\mathcal{G}_{\rho'}(\alpha/R^\epsilon)}|V|
\left\|\left(\sum_{\theta\subset\rho'}|f_\theta|^2*w_\theta\right)^{1/2}\right\|_{L_{\#}^{p_n/2}(V)}^{p_n}, ≲ ϵ , η R ϵ R − 1/ n ≤ u ≤ s ∑ ρ ′ ∈ Θ ( u ) ∑ V ∈ G ρ ′ ( α / R ϵ ) ∑ ∣ V ∣ θ ⊂ ρ ′ ∑ ∣ f θ ∣ 2 ∗ w θ 1/2 L # p n /2 ( V ) p n ,
where we used α ( s n R ) ϵ N ρ ( u ) ≳ α R ϵ N ( u ) \dfrac{\alpha}{(s^nR)^\epsilon N_\rho(u)}\gtrsim\dfrac{\alpha}{R^\epsilon N(u)} ( s n R ) ϵ N ρ ( u ) α ≳ R ϵ N ( u ) α by Lemma 2.18. This is R ϵ R^\epsilon R ϵ times (5.2), with amplitude α / R ϵ \alpha/R^\epsilon α / R ϵ , and the iteration stops.
We may therefore assume σ > R − 1 / n \sigma> R^{-1/n} σ > R − 1/ n . We estimate (5.1) separately for s ≤ R η 3 σ s \le R^{\eta^3}\sigma s ≤ R η 3 σ and s > R η 3 σ s > R^{\eta^3}\sigma s > R η 3 σ . In the first range, each cap of width s s s contains at most R η 3 R^{\eta^3} R η 3 caps of width σ \sigma σ , and we use (2.4) after Taylor rescaling. In the second range, we use Lemma 4.2 to reduce either σ \sigma σ or s s s .
Case 1: s ≤ R η 3 σ s \le R^{\eta^3}\sigma s ≤ R η 3 σ
If σ = R j − 1 / n \sigma=R_j^{-1/n} σ = R j − 1/ n , first use ∣ f τ j ∣ ≤ ∣ f τ j + 1 ∣ |f_{\tau}^{j}| \le|f_{\tau}^{j+1}| ∣ f τ j ∣ ≤ ∣ f τ j + 1 ∣ in (5.1) and replace j j j by j + 1 j+1 j + 1 . We may therefore assume σ > R j − 1 / n \sigma>R_j^{-1/n} σ > R j − 1/ n .
As in the derivation of (5.3), (2.22) bounds the sum over B B B in (5.1), for each ρ \rho ρ , by the integral on the left of (5.4) below. Since each ρ \rho ρ contains at most R η 3 R^{\eta^3} R η 3 caps of width σ \sigma σ , we use Hölder to move the power p n / 2 p_n/2 p n /2 inside the sum over τ \tau τ , and then apply (2.24) with F = ∣ f τ j ∣ 2 F=|f_{\tau}^{j}|^2 F = ∣ f τ j ∣ 2 and K = w τ K=w_{\tau} K = w τ to remove the convolution, obtaining
∫ ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ ≳ ( α / N ( s ) ) 2 ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ ) p n / 2 ≲ R η 3 ( p n / 2 − 1 ) ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ j ∣ ≳ α / ( R η 3 / 2 N ( s ) ) ∫ ∣ f τ j ∣ p n . (57) \int_{\sum_{\tau\subset\rho,\ |\tau|=\sigma}|f_{\tau}^{j}|^2*w_{\tau}\gtrsim(\alpha/N(s))^2}
\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_{\tau}^{j}|^2*w_{\tau}\right)^{p_n/2}
\lesssim R^{\eta^3(p_n/2-1)}
\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma\\|f_{\tau}^{j}| \gtrsim\alpha/(R^{\eta^3/2}N(s))}}
\int|f_{\tau}^{j}|^{p_n}. \tag*{(57)} ∫ ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ ≳ ( α / N ( s ) ) 2 τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ j ∣ 2 ∗ w τ p n /2 ≲ R η 3 ( p n /2 − 1 ) τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ j ∣ ≳ α / ( R η 3 /2 N ( s )) ∑ ∫ ∣ f τ j ∣ p n . (57)
Thus (5.1) is bounded by the sum over ρ \rho ρ of the right hand side of (5.4).
The Taylor rescaling taking each τ \tau τ to the unit cap sends the caps at scale R j − 1 / n R_j^{-1/n} R j − 1/ n inside τ \tau τ to caps of M n ( σ n R j ) \mathcal{M}^{n}(\sigma^{n}R_j) M n ( σ n R j ) at scale R j − 1 / n / σ R_j^{-1/n}/\sigma R j − 1/ n / σ , where 1 < σ n R j ≤ R η 3 1<\sigma^{n}R_j\le R^{\eta^3} 1 < σ n R j ≤ R η 3 . Applying (2.4) to each rescaled f τ j f_{\tau}^{j} f τ j , with parameter σ n R j \sigma^{n}R_j σ n R j , constant S 1 ( R η 3 ) S_1(R^{\eta^3}) S 1 ( R η 3 ) , and amplitude α / ( R C η 3 N ( s ) ) \alpha/(R^{C\eta^3}N(s)) α / ( R C η 3 N ( s )) , then returning to the original coordinates and summing over ρ , τ \rho,\tau ρ , τ , bounds (5.1) by
( R C η 3 S 1 ( R η 3 ) ) p n ∑ R j − 1 / n ≤ u ≤ σ ∑ ρ ′ ∈ Θ ( u ) ∑ V ∈ G ρ ′ ∥ V ∥ ∥ V ρ ′ , R j α R C η 3 S 1 ( R η 3 ) ∣ V ∣ ( ∑ τ j ⊂ ρ ∣ τ j ∣ = R j − 1 / n ∣ f τ j j ∣ 2 ∗ w τ j ) 1 / 2 ∥ L # p n / 2 ( V ) p n . (58) \left(R^{C\eta^3}S_1(R^{\eta^3})\right)^{p_n}
\sum_{R_j^{-1/n}\le u\le\sigma}
\sum_{\rho'\in\Theta(u)}
\sum_{\substack{V\in\mathcal G_{\rho'}\\
\|V\|\parallel V_{\rho',R_j}\\
\frac{\alpha}{R^{C\eta^3}S_1(R^{\eta^3})}}}
|V|
\left(
\sum_{\substack{\tau_j\subset\rho\\|\tau_j|=R_j^{-1/n}}}
|f_{\tau_j}^{j}|^2*w_{\tau_j}
\right)^{1/2}
\bigg\|_{L_{\#}^{p_n/2}(V)}^{p_n}. \tag*{(58)} ( R C η 3 S 1 ( R η 3 ) ) p n R j − 1/ n ≤ u ≤ σ ∑ ρ ′ ∈ Θ ( u ) ∑ V ∈ G ρ ′ ∥ V ∥ ∥ V ρ ′ , R j R C η 3 S 1 ( R η 3 ) α ∑ ∣ V ∣ τ j ⊂ ρ ∣ τ j ∣ = R j − 1/ n ∑ ∣ f τ j j ∣ 2 ∗ w τ j 1/2 L # p n /2 ( V ) p n . (58)
where we used
α R C η 3 S 1 ( R η 3 ) N ( s ) N τ ( u ) ≳ α R C η 3 S 1 ( R η 3 ) N ( u ) \frac{\alpha}{R^{C\eta^3}S_1(R^{\eta^3})N(s)N_{\tau}(u)}
\gtrsim
\frac{\alpha}{R^{C\eta^3}S_1(R^{\eta^3})N(u)} R C η 3 S 1 ( R η 3 ) N ( s ) N τ ( u ) α ≳ R C η 3 S 1 ( R η 3 ) N ( u ) α
by Lemma 2.18.
We next estimate (5.5): for j = N j=N j = N we return to the original functions f θ f_\theta f θ , while for j < N j<N j < N we first increase the current pruning index.
Case 1.1: j = N j=N j = N
The sums over V ∥ V ρ ′ , R V\parallel V_{\rho',R} V ∥ V ρ ′ , R in (5.5) already involve the fine pieces f θ N f_{\theta}^{N} f θ N . Using ∣ f θ N ∣ ≤ ∣ f θ ∣ |f_{\theta}^{N}|\le|f_{\theta}| ∣ f θ N ∣ ≤ ∣ f θ ∣ and u ≤ σ ≤ s u\le\sigma\le s u ≤ σ ≤ s therefore bounds (5.1) by ( R C η 3 S 1 ( R η 3 ) ) p n \left(R^{C\eta^3}S_1(R^{\eta^3})\right)^{p_n} ( R C η 3 S 1 ( R η 3 ) ) p n times (5.2), with α ↦ α / ( R C η 3 S 1 ( R η 3 ) ) \alpha\mapsto\alpha/(R^{C\eta^3}S_1(R^{\eta^3})) α ↦ α / ( R C η 3 S 1 ( R η 3 )) , and the iteration stops.
Case 1.2: j < N j<N j < N
After using ∣ f τ j ∣ ≤ ∣ f τ j + 1 ∣ |f_{\tau}^{j}|\le|f_{\tau}^{j+1}| ∣ f τ j ∣ ≤ ∣ f τ j + 1 ∣ in (5.5), dyadic pigeonholing reduces the sum over u u u in (5.5) to one scale u ∈ [ R j − 1 / n , σ ] u\in[R_j^{-1/n},\sigma] u ∈ [ R j − 1/ n , σ ] , with the logarithmic loss absorbed into R C η 3 R^{C\eta^3} R C η 3 . Since V ρ ′ , R j = ( u R j 1 / n ) 2 ( ρ ′ ) ∗ V_{\rho',R_j}=(uR_j^{1/n})^2(\rho')^* V ρ ′ , R j = ( u R j 1/ n ) 2 ( ρ ′ ) ∗ and u R j 1 / n ≥ 1 uR_j^{1/n}\ge1 u R j 1/ n ≥ 1 , we apply Lemma 2.14 to the sum over V V V for each ρ ′ \rho' ρ ′ , with F = ( ∑ τ j ⊂ ρ ′ ∣ f τ j j + 1 ∣ 2 ∗ w τ j ) p n / 4 F=\left(\sum_{\tau_j\subset\rho'}|f_{\tau_j}^{j+1}|^2*w_{\tau_j}\right)^{p_n/4} F = ( ∑ τ j ⊂ ρ ′ ∣ f τ j j + 1 ∣ 2 ∗ w τ j ) p n /4 , B 0 = V ρ ′ , R j B_0=V_{\rho',R_j} B 0 = V ρ ′ , R j , and U 0 = ( ρ ′ ) ∗ U_0=(\rho')^* U 0 = ( ρ ′ ) ∗ , to replace it by a sum over Q ∥ ( ρ ′ ) ∗ Q\parallel(\rho')^* Q ∥ ( ρ ′ ) ∗ . Combining this with (5.5), we bound (5.1) by
( R C η 3 S 1 ( R η 3 ) ) p n ∑ ρ ′ ∈ Θ ( u ) ∑ Q ∈ G ρ ′ Q ∥ ( ρ ′ ) ∗ α R C η 3 S 1 ( R η 3 ) ∣ Q ∣ ( ∑ τ j ⊂ ρ ′ ∣ τ j ∣ = R j − 1 / n ∣ f τ j j + 1 ∣ 2 ∗ w τ j ) 1 / 2 ∥ L # p n / 2 ( Q ) p n . (59) \left(R^{C\eta^3}S_1(R^{\eta^3})\right)^{p_n}
\sum_{\rho'\in\Theta(u)}
\sum_{\substack{Q\in\mathcal G_{\rho'}\\
Q\parallel(\rho')^*\\
\frac{\alpha}{R^{C\eta^3}S_1(R^{\eta^3})}}}
|Q|
\left(
\sum_{\substack{\tau_j\subset\rho'\\|\tau_j|=R_j^{-1/n}}}
|f_{\tau_j}^{j+1}|^2*w_{\tau_j}
\right)^{1/2}
\bigg\|_{L_{\#}^{p_n/2}(Q)}^{p_n}. \tag*{(59)} ( R C η 3 S 1 ( R η 3 ) ) p n ρ ′ ∈ Θ ( u ) ∑ Q ∈ G ρ ′ Q ∥ ( ρ ′ ) ∗ R C η 3 S 1 ( R η 3 ) α ∑ ∣ Q ∣ τ j ⊂ ρ ′ ∣ τ j ∣ = R j − 1/ n ∑ ∣ f τ j j + 1 ∣ 2 ∗ w τ j 1/2 L # p n /2 ( Q ) p n . (59)
We now apply Lemma 4.1 to (5.6), with ( j , σ , s , α ) (j,\sigma,s,\alpha) ( j , σ , s , α ) replaced by ( j + 1 , R j − 1 / n , u , α / ( R C η 3 S 1 ( R η 3 ) ) ) (j+1,R_j^{-1/n},u,\alpha/(R^{C\eta^3}S_1(R^{\eta^3}))) ( j + 1 , R j − 1/ n , u , α / ( R C η 3 S 1 ( R η 3 ))) . If the first sum over V V V on the RHS of Lemma 4.1 dominates, absorbing the lemma’s R ϵ R^\epsilon R ϵ loss into R C η 3 R^{C\eta^3} R C η 3 and using u ≤ σ ≤ s u\le\sigma\le s u ≤ σ ≤ s bounds (5.1) by ( R C η 3 S 1 ( R η 3 ) ) p n \left(R^{C\eta^3}S_1(R^{\eta^3})\right)^{p_n} ( R C η 3 S 1 ( R η 3 ) ) p n times (5.2), with α ↦ α / ( R C η 3 S 1 ( R η 3 ) ) \alpha\mapsto\alpha/(R^{C\eta^3}S_1(R^{\eta^3})) α ↦ α / ( R C η 3 S 1 ( R η 3 )) , and the iteration stops.
If instead the second sum over B B B on the RHS of Lemma 4.1 dominates, let σ ′ \sigma' σ ′ and j ′ j' j ′ be its cap width and pruning index, and use dyadic pigeonholing to fix s ′ ∈ [ σ ′ , u ] s'\in[\sigma',u] s ′ ∈ [ σ ′ , u ] . Then (5.1) is bounded by
( R C η 3 S 1 ( R η 3 ) ) p n ∑ ρ ′ ′ ∈ Θ ( s ′ ) ∑ B ∈ G ρ ′ ′ ∣ B ∣ ≲ ( s ′ / σ ′ ) V ρ ′ ′ , ( σ ′ ) − n α R C η 3 S 1 ( R η 3 ) ∣ B ∣ ( ∑ τ ′ ⊂ ρ ′ ′ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ j ′ ∣ 2 ∗ w τ ′ ) 1 / 2 ∥ L # p n / 2 ( B ) p n . (60) \left(R^{C\eta^3}S_1(R^{\eta^3})\right)^{p_n}
\sum_{\rho''\in\Theta(s')}
\sum_{\substack{B\in\mathcal G_{\rho''}\\
|B|\lesssim(s'/\sigma')V_{\rho'',(\sigma')^{-n}}\\
\frac{\alpha}{R^{C\eta^3}S_1(R^{\eta^3})}}}
|B|
\left(
\sum_{\substack{\tau'\subset\rho''\\|\tau'|=\sigma'}}
|f_{\tau'}^{j'}|^2*w_{\tau'}
\right)^{1/2}
\bigg\|_{L_{\#}^{p_n/2}(B)}^{p_n}. \tag*{(60)} ( R C η 3 S 1 ( R η 3 ) ) p n ρ ′′ ∈ Θ ( s ′ ) ∑ B ∈ G ρ ′′ ∣ B ∣ ≲ ( s ′ / σ ′ ) V ρ ′′ , ( σ ′ ) − n R C η 3 S 1 ( R η 3 ) α ∑ ∣ B ∣ τ ′ ⊂ ρ ′′ ∣ τ ′ ∣ = σ ′ ∑ ∣ f τ ′ j ′ ∣ 2 ∗ w τ ′ 1/2 L # p n /2 ( B ) p n . (60)
The sum over B B B in (5.7) has exactly the form (5.1), with ( j , σ , s , α ) (j,\sigma,s,\alpha) ( j , σ , s , α ) replaced by ( j ′ , σ ′ , s ′ , α / ( R C η 3 S 1 ( R η 3 ) ) ) (j',\sigma',s',\alpha/(R^{C\eta^3}S_1(R^{\eta^3}))) ( j ′ , σ ′ , s ′ , α / ( R C η 3 S 1 ( R η 3 ))) . Since j ′ ≥ j + 1 j'\ge j+1 j ′ ≥ j + 1 , σ ′ ≤ σ \sigma'\le\sigma σ ′ ≤ σ , and s ′ ≤ s s'\le s s ′ ≤ s , we return to (5.1) with these new parameters and repeat the algorithm.
Case 2: s > R η σ s > R^\eta\sigma s > R η σ .
Applying Lemma 4.2 to (54) , with the current parameters j , σ , s , α j,\sigma,s,\alpha j , σ , s , α , bounds (54) by
R ϵ ∑ ρ ∈ Θ ( s ) ∑ Q ∈ G ρ ( R ϵ R ) Q ∈ G ρ ∗ ∣ Q ∣ ∥ ( ∑ τ ′ ⊂ C ρ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ j ∣ 2 ∗ w τ ′ ) 1 / 2 ∥ L # p n / 2 ( Q ) + R ϵ ∑ ρ ′ ∈ Θ ( s ′ ) ∑ Q ∈ G ρ ′ ( R ϵ R ) Q ∈ G ρ ′ ∗ ∣ Q ∣ ∥ ( ∑ τ ⊂ C ρ ′ ∣ τ ∣ = σ ∣ f τ j ∣ 2 ∗ w τ ) 1 / 2 ∥ L # p n / 2 ( Q ) , (61) R^\epsilon\sum_{\rho\in\Theta(s)}\sum_{\substack{Q\in\mathcal{G}_{\rho}\left(\frac{R^\epsilon}{R}\right)\\Q\in\mathcal{G}_{\rho}^{*}}}|Q|\left\|\left(\sum_{\substack{\tau'\subset C_\rho\\|\tau'|=\sigma'}}|f_{\tau'}^j|^2*w_{\tau'}\right)^{1/2}\right\|_{L_{\#}^{p_n/2}(Q)}
+R^\epsilon\sum_{\rho'\in\Theta(s')}\sum_{\substack{Q\in\mathcal{G}_{\rho'}\left(\frac{R^\epsilon}{R}\right)\\Q\in\mathcal{G}_{\rho'}^{*}}}|Q|\left\|\left(\sum_{\substack{\tau\subset C_{\rho'}\\|\tau|=\sigma}}|f_{\tau}^j|^2*w_{\tau}\right)^{1/2}\right\|_{L_{\#}^{p_n/2}(Q)}, \tag*{(61)} R ϵ ρ ∈ Θ ( s ) ∑ Q ∈ G ρ ( R R ϵ ) Q ∈ G ρ ∗ ∑ ∣ Q ∣ τ ′ ⊂ C ρ ∣ τ ′ ∣ = σ ′ ∑ ∣ f τ ′ j ∣ 2 ∗ w τ ′ 1/2 L # p n /2 ( Q ) + R ϵ ρ ′ ∈ Θ ( s ′ ) ∑ Q ∈ G ρ ′ ( R R ϵ ) Q ∈ G ρ ′ ∗ ∑ ∣ Q ∣ τ ⊂ C ρ ′ ∣ τ ∣ = σ ∑ ∣ f τ j ∣ 2 ∗ w τ 1/2 L # p n /2 ( Q ) , (61)
where σ ′ = max { R j − 1 / n , σ 2 / s } \sigma'=\max\{R_j^{-1/n},\sigma^2/s\} σ ′ = max { R j − 1/ n , σ 2 / s } and s ′ = s ( σ / s ) 1 / ( n − 1 ) s'=s(\sigma/s)^{1/(n-1)} s ′ = s ( σ / s ) 1/ ( n − 1 ) . We distinguish two cases according to which sum in (61) dominates. Case 2.1 treats the second sum, at scales ( σ , s ′ ) (\sigma,s') ( σ , s ′ ) , while Case 2.2 treats the first sum, at scales ( σ ′ , s ) (\sigma',s) ( σ ′ , s ) .
Smaller parents.
Suppose that the second sum in (61) dominates. We apply Lemma 4.1 to this sum, with ( j , σ , s , α ) (j,\sigma,s,\alpha) ( j , σ , s , α ) replaced by ( j , σ , s ′ , α / R ϵ ) (j,\sigma,s',\alpha/R^\epsilon) ( j , σ , s ′ , α / R ϵ ) . If the first sum over V V V on the RHS of Lemma 4.1 dominates, this bounds (54) by R ϵ R^\epsilon R ϵ times (55) , with α ↦ α / R ϵ \alpha\mapsto\alpha/R^\epsilon α ↦ α / R ϵ , and the iteration stops.
If instead the second sum over B B B on the RHS of Lemma 4.1 dominates, let σ ′ ′ \sigma'' σ ′′ and j ′ j' j ′ be its cap width and pruning index, and apply dyadic pigeonholing to select one u ∈ [ σ ′ ′ , s ′ ] u\in[\sigma'',s'] u ∈ [ σ ′′ , s ′ ] . This bounds (54) by R ϵ R^\epsilon R ϵ times (54) , with ( j , σ , s , α ) (j,\sigma,s,\alpha) ( j , σ , s , α ) replaced by ( j ′ , σ ′ ′ , u , α / R ϵ ) (j',\sigma'',u,\alpha/R^\epsilon) ( j ′ , σ ′′ , u , α / R ϵ ) . Since j ′ ≥ j j'\ge j j ′ ≥ j , σ ′ ′ ≤ σ \sigma''\le\sigma σ ′′ ≤ σ , and u ≤ s ′ ≤ s R − 3 η / ( n − 1 ) < s u\le s'\le sR^{-3\eta/(n-1)}<s u ≤ s ′ ≤ s R − 3 η / ( n − 1 ) < s , we return to (54) with these new parameters and repeat the algorithm.
Finer caps.
Now suppose that the first sum in (61) dominates.
First suppose that σ ′ > R j − 1 / n \sigma'>R_j^{-1/n} σ ′ > R j − 1/ n . We apply Lemma 4.1 to the first sum in (61) , with ( j , σ , s , α ) (j,\sigma,s,\alpha) ( j , σ , s , α ) replaced by ( j , σ ′ , s , α / R ϵ ) (j,\sigma',s,\alpha/R^\epsilon) ( j , σ ′ , s , α / R ϵ ) . If the first sum over V V V on the RHS of Lemma 4.1 dominates, this bounds (54) by R ϵ R^\epsilon R ϵ times (55) , with α ↦ α / R ϵ \alpha\mapsto\alpha/R^\epsilon α ↦ α / R ϵ , and the iteration stops.
If instead the second sum over B B B on the RHS of Lemma 4.1 dominates, let σ ′ ′ \sigma'' σ ′′ and j ′ j' j ′ be its cap width and pruning index, and apply dyadic pigeonholing to select one u ∈ [ σ ′ ′ , s ] u\in[\sigma'',s] u ∈ [ σ ′′ , s ] . This bounds (54) by R ϵ R^\epsilon R ϵ times (54) , with ( j , σ , s , α ) (j,\sigma,s,\alpha) ( j , σ , s , α ) replaced by ( j ′ , σ ′ ′ , u , α / R ϵ ) (j',\sigma'',u,\alpha/R^\epsilon) ( j ′ , σ ′′ , u , α / R ϵ ) . Since j ′ ≥ j j'\ge j j ′ ≥ j , σ ′ ′ ≤ σ ′ = σ 2 / s ≤ σ R − η < σ \sigma''\le\sigma'=\sigma^2/s\le\sigma R^{-\eta}<\sigma σ ′′ ≤ σ ′ = σ 2 / s ≤ σ R − η < σ , and u ≤ s u\le s u ≤ s , we return to (54) with these new parameters and repeat the algorithm.
Next suppose that σ ′ = R j − 1 / n \sigma'=R_j^{-1/n} σ ′ = R j − 1/ n and j < N j<N j < N . Using ∣ f τ j ∣ ≤ ∣ f τ j + 1 ∣ |f_{\tau}^{j}|\le|f_{\tau}^{j+1}| ∣ f τ j ∣ ≤ ∣ f τ j + 1 ∣ in the first sum of (61) , we apply Lemma 4.1 with ( j , σ , s , α ) (j,\sigma,s,\alpha) ( j , σ , s , α ) replaced by ( j + 1 , R − 1 / n , s , α / R ϵ ) (j+1,R^{-1/n},s,\alpha/R^\epsilon) ( j + 1 , R − 1/ n , s , α / R ϵ ) . If the first sum over V V V on the RHS of Lemma 4.1 dominates, this bounds (54) by R ϵ R^\epsilon R ϵ times (55) , with α ↦ α / R ϵ \alpha\mapsto\alpha/R^\epsilon α ↦ α / R ϵ , and the iteration stops.
If instead the second sum over B B B on the RHS of Lemma 4.1 dominates, let σ ′ ′ \sigma'' σ ′′ and j ′ j' j ′ be its cap width and pruning index, and apply dyadic pigeonholing to select one u ∈ [ σ ′ ′ , s ] u\in[\sigma'',s] u ∈ [ σ ′′ , s ] . This bounds (54) by R ϵ R^\epsilon R ϵ times (54) , with ( j , σ , s , α ) (j,\sigma,s,\alpha) ( j , σ , s , α ) replaced by ( j ′ , σ ′ ′ , u , α / R ϵ ) (j',\sigma'',u,\alpha/R^\epsilon) ( j ′ , σ ′′ , u , α / R ϵ ) . Since j ′ ≥ j + 1 j'\ge j+1 j ′ ≥ j + 1 , σ ′ ′ ≤ σ \sigma''\le\sigma σ ′′ ≤ σ , and u ≤ s u\le s u ≤ s , we return to (54) with these new parameters and repeat the algorithm.
Finally, suppose that σ ′ = R j − 1 / n = R − 1 / n \sigma'=R_j^{-1/n}=R^{-1/n} σ ′ = R j − 1/ n = R − 1/ n . We use ∣ f θ N ∣ ≤ ∣ f θ ∣ |f_{\theta}^{N}|\le|f_{\theta}| ∣ f θ N ∣ ≤ ∣ f θ ∣ in the first sum of (61) . For each ρ \rho ρ , (32) then bounds its sum over Q Q Q by (56) , with α ↦ α / R ϵ \alpha\mapsto\alpha/R^\epsilon α ↦ α / R ϵ . Repeating the Taylor rescaling and application of Lemma 2.8 used to estimate (56) bounds (54) by R ϵ R^\epsilon R ϵ times (55) , with α ↦ α / R ϵ \alpha\mapsto\alpha/R^\epsilon α ↦ α / R ϵ , and the iteration stops.
Termination of the iteration.
At every continued step, s s s and σ \sigma σ do not increase and j j j does not decrease. Case 2.1 decreases s s s by a fixed power of R η 3 R^{\eta^3} R η 3 , while Case 2.2 decreases σ \sigma σ by such a power or increases j j j . Case 1 increases j j j after applying S 1 ( R η ) S_1(R^\eta) S 1 ( R η ) , or terminates. Since both widths lie in [ R − 1 / n , 1 ] [R^{-1/n},1] [ R − 1/ n , 1 ] , there are O n ( η − 3 ) O_n(\eta^{-3}) O n ( η − 3 ) steps in all and at most N − j + 1 N-j+1 N − j + 1 applications of S 1 ( R η ) S_1(R^\eta) S 1 ( R η ) , where j j j is the initial index in the statement. The factors R ϵ R^\epsilon R ϵ from the refinement steps and R C η R^{C\eta} R C η from the S 1 S_1 S 1 steps are bounded together by R C η n R^{C\eta n} R C η n , after choosing ϵ \epsilon ϵ sufficiently small in terms of n n n and using N = η − 1 N=\eta^{-1} N = η − 1 . Thus the accumulated coefficient is at most ( R C η S 1 ( R η ) N − j + C ) p n (R^{C\eta}S_1(R^\eta)^{N-j+C})^{p_n} ( R C η S 1 ( R η ) N − j + C ) p n , and the amplitude in (55) is at least α / ( R C η S 1 ( R η ) N − j + C ) \alpha/(R^{C\eta}S_1(R^\eta)^{N-j+C}) α / ( R C η S 1 ( R η ) N − j + C ) , with the initial α \alpha α . This proves (30) .
From cones over moment curves to moment curves: Proof of Proposition 2.6.
As outlined in § 2 , we first prove Proposition 2.11. We then combine this with Proposition 2.12 and apply a standard broad narrow analysis to prove Proposition 2.6.
Proof of Proposition 2.11
We split B r α , β K \mathrm{Br}^{K}_{\alpha,\beta} Br α , β K into its intersections with Ω k \Omega_k Ω k , N 0 ≤ k < N N_0 \le k < N N 0 ≤ k < N , and the remaining part contained in L L L . On each Ω k \Omega_k Ω k , we use the multilinear estimate and the amplitude dependent wave envelope type estimate for the cone over the moment curve in R n \mathbb{R}^n R n (Theorem 1.8 ); on L L L , we use Lemma 2.8.
Case 1: Bounding B r α , β K ∩ Ω k \mathrm{Br}^{K}_{\alpha,\beta}\cap\Omega_k Br α , β K ∩ Ω k . Fix N 0 ≤ k < N N_0 \le k < N N 0 ≤ k < N and estimate B r α , β K ∩ Ω k \mathrm{Br}^{K}_{\alpha,\beta}\cap\Omega_k Br α , β K ∩ Ω k . By the two defining inequalities of B r α , β K \mathrm{Br}^{K}_{\alpha,\beta} Br α , β K , we have ∣ f τ i ( x ) ∣ ≥ α / ( 100 K 100 n 2 ) |f_{\tau_i}(x)| \ge\alpha/(100K^{100n^2}) ∣ f τ i ( x ) ∣ ≥ α / ( 100 K 100 n 2 ) , 1 ≤ i ≤ n 1 \le i \le n 1 ≤ i ≤ n . Since K − 1 ≥ R 1 − 1 / n K^{-1} \ge R_1^{-1/n} K − 1 ≥ R 1 − 1/ n , after choosing A A A sufficiently large, (2.17) gives ∣ f τ i k + 1 ∣ ≥ ∣ f τ i ∣ / 2 |f_{\tau_i}^{k+1}| \ge|f_{\tau_i}|/2 ∣ f τ i k + 1 ∣ ≥ ∣ f τ i ∣/2 on B r α , β K ∩ Ω k \mathrm{Br}^{K}_{\alpha,\beta}\cap\Omega_k Br α , β K ∩ Ω k , and hence ∣ f τ 1 k + 1 ⋯ f τ n k + 1 ∣ ≥ α / ( 200 K 100 n ) |f_{\tau_1}^{k+1}\cdots f_{\tau_n}^{k+1}| \ge\alpha/(200K^{100n}) ∣ f τ 1 k + 1 ⋯ f τ n k + 1 ∣ ≥ α / ( 200 K 100 n ) there.
Let B B B be a ball of radius R k 1 / n R_k^{1/n} R k 1/ n , and fix x B ∈ B r α , β K ∩ Ω k ∩ B x_B\in\mathrm{Br}^{K}_{\alpha,\beta}\cap\Omega_k\cap B x B ∈ Br α , β K ∩ Ω k ∩ B . By the multilinear restriction estimate used in [10 proof of Proposition 6.1] , we obtain
α 2 n ∣ B r α , β K ∩ Ω k ∩ B ∣ ≲ η R C n ∣ B ∣ [ ( ∑ τ k ∣ f τ k k + 1 ∣ 2 ∗ w τ k ) ( x B ) ] n . (62) \alpha^{2n}|\mathrm{Br}^{K}_{\alpha,\beta}\cap\Omega_k\cap B|\lesssim_\eta R^{Cn}|B|\left[\left(\sum_{\tau_k}|f_{\tau_k}^{k+1}|^2*w_{\tau_k}\right)(x_B)\right]^n . \tag*{(62)} α 2 n ∣ Br α , β K ∩ Ω k ∩ B ∣ ≲ η R C n ∣ B ∣ [ ( τ k ∑ ∣ f τ k k + 1 ∣ 2 ∗ w τ k ) ( x B ) ] n . (62)
On Ω k \Omega_k Ω k , the high frequency estimate (2.15) from the pruning construction gives
∣ ( ∑ τ k ∣ f τ k k + 1 ∣ 2 ∗ w τ k ) ∗ η ~ > R k + 1 − 1 / n ∣ ≳ ∣ ∑ τ k ∣ f τ k k + 1 ∣ 2 ∗ w τ k ∣ ≥ A N − k β . \left|\left(\sum_{\tau_k}|f_{\tau_k}^{k+1}|^2*w_{\tau_k}\right)*\widetilde{\eta}_{>R_{k+1}^{-1/n}}\right|\gtrsim\left|\sum_{\tau_k}|f_{\tau_k}^{k+1}|^2*w_{\tau_k}\right|\ge A^{N-k}\beta. ( τ k ∑ ∣ f τ k k + 1 ∣ 2 ∗ w τ k ) ∗ η > R k + 1 − 1/ n ≳ τ k ∑ ∣ f τ k k + 1 ∣ 2 ∗ w τ k ≥ A N − k β .
Summing (6.1) over the balls B B B intersecting B r α , β K ∩ Ω k \mathrm{Br}^{K}_{\alpha,\beta}\cap\Omega_k Br α , β K ∩ Ω k and using the locally constant property, we obtain
α 2 n ∣ B r α , β K ∩ Ω k ∣ ≲ η R C n ∑ B ∣ B ∣ [ ( ∑ τ k ∣ f τ k k + 1 ∣ 2 ∗ w τ k ) ( x B ) ] n \alpha^{2n}|\mathrm{Br}^{K}_{\alpha,\beta}\cap\Omega_k|\lesssim_\eta R^{Cn}\sum_B |B|\left[\left(\sum_{\tau_k}|f_{\tau_k}^{k+1}|^2*w_{\tau_k}\right)(x_B)\right]^n α 2 n ∣ Br α , β K ∩ Ω k ∣ ≲ η R C n B ∑ ∣ B ∣ [ ( τ k ∑ ∣ f τ k k + 1 ∣ 2 ∗ w τ k ) ( x B ) ] n
≲ η R C n ( A N − k β ) n − p n − 1 ∑ B ∣ B ∣ [ ∣ f B ∣ ( ( ∑ τ k ∣ f τ k k + 1 ∣ 2 ∗ w τ k ) ∗ η ~ > R k + 1 − 1 / n ) ∣ p n − 1 / 2 ] 2 \lesssim_\eta R^{Cn}(A^{N-k}\beta)^{n-p_{n-1}}\sum_B |B|[|f_B|((\sum_{\tau_k}|f_{\tau_k}^{k+1}|^2*w_{\tau_k})*\widetilde{\eta}_{>R_{k+1}^{-1/n}})|^{p_{n-1}/2}]^2 ≲ η R C n ( A N − k β ) n − p n − 1 B ∑ ∣ B ∣ [ ∣ f B ∣ (( τ k ∑ ∣ f τ k k + 1 ∣ 2 ∗ w τ k ) ∗ η > R k + 1 − 1/ n ) ∣ p n − 1 /2 ] 2
≲ η R C n ( A N − k β ) n − p n − 1 ∫ ∣ ( ∑ τ k ∣ f τ k k + 1 ∣ 2 ∗ w τ k ) ∗ η ~ > R k + 1 − 1 / n ∣ p n − 1 \lesssim_\eta R^{Cn}(A^{N-k}\beta)^{n-p_{n-1}}\int\left|\left(\sum_{\tau_k}|f_{\tau_k}^{k+1}|^2*w_{\tau_k}\right)*\widetilde{\eta}_{>R_{k+1}^{-1/n}}\right|^{p_{n-1}} ≲ η R C n ( A N − k β ) n − p n − 1 ∫ ( τ k ∑ ∣ f τ k k + 1 ∣ 2 ∗ w τ k ) ∗ η > R k + 1 − 1/ n p n − 1
where we used p n − 1 ≥ n p_{n-1}\ge n p n − 1 ≥ n , and the last inequality follows from (2.22).
We now estimate the last integral in (6.2) by decomposing its high frequency cutoff into the dyadic cutoffs η σ \eta_\sigma η σ . Dyadic pigeonholing gives a fixed width R k + 1 − 1 / n < σ ≤ R k − 1 / n R_{k+1}^{-1/n}<\sigma\le R_k^{-1/n} R k + 1 − 1/ n < σ ≤ R k − 1/ n for which the RHS of (6.2) is bounded by
R C n ( A N − k β ) n − p n − 1 ∫ ∣ ∑ τ h τ ∣ > A N − k β R C n ∣ ∑ τ h τ ∣ p n − 1 , (63) R^{Cn}(A^{N-k}\beta)^{n-p_{n-1}}\int_{\left|\sum_\tau h_\tau\right|>\frac{A^{N-k}\beta}{R^{Cn}}}\left|\sum_\tau h_\tau\right|^{p_{n-1}}, \tag*{(63)} R C n ( A N − k β ) n − p n − 1 ∫ ∣ ∑ τ h τ ∣ > R C n A N − k β τ ∑ h τ p n − 1 , (63)
where, for τ ∈ Θ ( σ ) \tau\in\Theta(\sigma) τ ∈ Θ ( σ ) ,
h τ = ( ∑ τ ⊂ τ k , ∣ τ ∣ = σ ∣ τ − τ k ∣ ≲ σ f τ k + 1 f τ k + 1 ‾ ) ∗ w τ k ∗ η ~ σ , h_\tau=\left(\sum_{\substack{\tau\subset\tau_k,\ |\tau|=\sigma\\ |\tau-\tau_k|\lesssim\sigma}}f_{\tau}^{k+1}\overline{f_{\tau}^{k+1}}\right)*w_{\tau_k}*\widetilde{\eta}_\sigma, h τ = τ ⊂ τ k , ∣ τ ∣ = σ ∣ τ − τ k ∣ ≲ σ ∑ f τ k + 1 f τ k + 1 ∗ w τ k ∗ η σ ,
and τ k \tau_k τ k is the unique cap in Θ ( R k − 1 / n ) \Theta(R_k^{-1/n}) Θ ( R k − 1/ n ) containing τ \tau τ .
As in the proof of Lemma 4.1, dilation by σ − 1 \sigma^{-1} σ − 1 sends the Fourier supports of h τ h_\tau h τ into caps of aperture σ \sigma σ in R n ( σ − ( n − 1 ) ) \mathbb{R}^n(\sigma^{-(n-1)}) R n ( σ − ( n − 1 ) ) . We therefore apply the amplitude dependent wave envelope type estimate for the cone over the moment curve in R n \mathbb{R}^n R n (Theorem 1.8 ) to ∑ τ h τ ( ⋅ / σ ) \sum_\tau h_\tau(\cdot/\sigma) ∑ τ h τ ( ⋅ / σ ) , with parameter σ − ( n − 1 ) \sigma^{-(n-1)} σ − ( n − 1 ) , exponent p n − 1 p_{n-1} p n − 1 , and amplitude A N − k β / R C n A^{N-k}\beta/R^{Cn} A N − k β / R C n . Returning to the original coordinates gives
∫ ∣ ∑ τ h τ ∣ > A N − k β R C n ∣ ∑ τ h τ ∣ p n − 1 ≲ η R C n ∑ σ ≲ σ 1 ∑ ρ ∈ Θ ( σ ) ∑ B ∥ ( σ / σ 1 ) V ρ , σ − ( n − 1 ) ∣ B ∣ [ ∫ B ( ∑ τ ⊂ ρ ∣ h τ ∣ 2 ) p n − 1 / 4 ] 2 , (64) \int_{\left|\sum_\tau h_\tau\right|>\frac{A^{N-k}\beta}{R^{Cn}}}\left|\sum_\tau h_\tau\right|^{p_{n-1}}\lesssim_\eta R^{Cn}\sum_{\sigma\lesssim\sigma_1}\sum_{\rho\in\Theta(\sigma)}\sum_{B\parallel(\sigma/\sigma_1)V_{\rho,\sigma^{-(n-1)}}}|B|\left[\int_B\left(\sum_{\tau\subset\rho}|h_\tau|^2\right)^{p_{n-1}/4}\right]^2 , \tag*{(64)} ∫ ∣ ∑ τ h τ ∣ > R C n A N − k β τ ∑ h τ p n − 1 ≲ η R C n σ ≲ σ 1 ∑ ρ ∈ Θ ( σ ) ∑ B ∥ ( σ / σ 1 ) V ρ , σ − ( n − 1 ) ∑ ∣ B ∣ ∫ B ( τ ⊂ ρ ∑ ∣ h τ ∣ 2 ) p n − 1 /4 2 , (64)
f B ( ∑ τ ⊂ ρ ∣ h τ ∣ 2 ) p n − 1 / 4 ≥ [ A N − k β R C n N ( σ ) ] p n − 1 / 2 f_B\left(\sum_{\tau\subset\rho}|h_\tau|^2\right)^{p_{n-1}/4}\ge\left[\frac{A^{N-k}\beta}{R^{Cn}N(\sigma)}\right]^{p_{n-1}/2} f B ( τ ⊂ ρ ∑ ∣ h τ ∣ 2 ) p n − 1 /4 ≥ [ R C n N ( σ ) A N − k β ] p n − 1 /2
where s = ∣ ρ ∣ s=|\rho| s = ∣ ρ ∣ is the dyadic scale in the estimate for the cone over the moment curve, and the averaging boxes return to B ∥ σ − 1 U ρ , σ − ( n − 1 ) = ( s / σ ) V ρ , σ − n B\parallel\sigma^{-1}U_{\rho,\sigma^{-(n-1)}}=(s/\sigma)V_{\rho,\sigma^{-n}} B ∥ σ − 1 U ρ , σ − ( n − 1 ) = ( s / σ ) V ρ , σ − n , with dimensions σ − 3 s 2 × σ − 3 s × ⋯ × σ − 3 s 3 − n \sigma^{-3}s^2\times\sigma^{-3}s\times\cdots\times\sigma^{-3}s^{3-n} σ − 3 s 2 × σ − 3 s × ⋯ × σ − 3 s 3 − n .
To bound the RHS of (6.4) in terms of f τ k + 1 f_\tau^{k+1} f τ k + 1 , we use the pruning L ∞ L^\infty L ∞ bound (2.14). Each f τ k + 1 f_\tau^{k+1} f τ k + 1 is a sum of O ( R n / n ) R k + 1 − 1 / n O(R^{n/n})R_{k+1}^{-1/n} O ( R n / n ) R k + 1 − 1/ n children, so ∥ f τ k + 1 ∥ ∞ ≲ η R C n A N − k β / α \|f_\tau^{k+1}\|_\infty\lesssim_\eta R^{Cn}A^{N-k}\beta/\alpha ∥ f τ k + 1 ∥ ∞ ≲ η R C n A N − k β / α . Since the cutoff η σ \eta_\sigma η σ leaves only products f τ k + 1 f τ k + 1 ‾ f_\tau^{k+1}\overline{f_\tau^{k+1}} f τ k + 1 f τ k + 1 … with ∣ t τ − t τ ~ ∣ ≲ σ ≤ s |t_\tau-t_{\widetilde{\tau}}|\lesssim\sigma\le s ∣ t τ − t τ ∣ ≲ σ ≤ s , the terms with τ ⊂ ρ \tau\subset\rho τ ⊂ ρ , ρ ∈ Θ ( s ) \rho\in\Theta(s) ρ ∈ Θ ( s ) , involve only τ ~ \widetilde{\tau} τ within distance O ( σ ) O(\sigma) O ( σ ) of ρ \rho ρ . Writing ρ ~ ∼ ρ \widetilde{\rho}\sim\rho ρ ∼ ρ for the O ( 1 ) O(1) O ( 1 ) neighbouring parents ρ ~ ∈ Θ ( s ) \widetilde{\rho}\in\Theta(s) ρ ∈ Θ ( s ) with ∣ t ρ ~ − t ρ ∣ ≲ s |t_{\widetilde{\rho}}-t_\rho|\lesssim s ∣ t ρ − t ρ ∣ ≲ s , we have
∑ τ ⊂ ρ ∣ h τ ∣ 2 ≲ η R C η ( A N − k β α ) 2 [ ∣ η ~ σ ∣ ∗ ∑ ρ ~ ∼ ρ ∑ τ ⊂ ρ ~ ∣ f τ k + 1 ∣ 2 ∗ w τ ] , \sum_{\tau\subset\rho}|h_\tau|^2\lesssim_\eta R^{C\eta}\left(\frac{A^{N-k}\beta}{\alpha}\right)^2\left[|\widetilde{\eta}_\sigma|*\sum_{\widetilde{\rho}\sim\rho}\sum_{\tau\subset\widetilde{\rho}}|f_\tau^{k+1}|^2*w_\tau\right], τ ⊂ ρ ∑ ∣ h τ ∣ 2 ≲ η R C η ( α A N − k β ) 2 ∣ η σ ∣ ∗ ρ ∼ ρ ∑ τ ⊂ ρ ∑ ∣ f τ k + 1 ∣ 2 ∗ w τ ,
∑ τ ⊂ ρ ∣ h τ ∣ 2 ≲ [ ∣ η ~ σ ∣ ∗ ∑ ρ ~ ∼ ρ ∑ τ ⊂ ρ ~ ∣ f τ k + 1 ∣ 2 ∗ w τ ] 2 . (65) \sum_{\tau\subset\rho}|h_\tau|^2\lesssim\left[|\widetilde{\eta}_\sigma|*\sum_{\widetilde{\rho}\sim\rho}\sum_{\tau\subset\widetilde{\rho}}|f_\tau^{k+1}|^2*w_\tau\right]^2. \tag*{(65)} τ ⊂ ρ ∑ ∣ h τ ∣ 2 ≲ ∣ η σ ∣ ∗ ρ ∼ ρ ∑ τ ⊂ ρ ∑ ∣ f τ k + 1 ∣ 2 ∗ w τ 2 . (65)
where we used ∣ f τ k + 1 ∣ 2 ∗ w τ k ≲ ∣ f τ k + 1 ∣ 2 ∗ w τ |f_\tau^{k+1}|^2*w_{\tau_k}\lesssim|f_\tau^{k+1}|^2*w_\tau ∣ f τ k + 1 ∣ 2 ∗ w τ k ≲ ∣ f τ k + 1 ∣ 2 ∗ w τ . Raising the first inequality in (6.5) to the power ( p n − 1 − n ) / 4 (p_{n-1}-n)/4 ( p n − 1 − n ) /4 and the second to n / 4 n/4 n /4 , then multiplying, gives
( ∑ τ ⊂ ρ ∣ h τ ∣ 2 ) p n − 1 / 4 ≲ η R C η ( A N − k β α ) ( p n − 1 − n ) / 2 [ ∣ η ~ σ ∣ ∗ ∑ ρ ~ ∼ ρ ∑ τ ⊂ ρ ~ ∣ f τ k + 1 ∣ 2 ∗ w τ ] p n / 4 , (66) \left(\sum_{\tau\subset\rho}|h_\tau|^2\right)^{p_{n-1}/4}\lesssim_\eta R^{C\eta}\left(\frac{A^{N-k}\beta}{\alpha}\right)^{(p_{n-1}-n)/2}\left[|\widetilde{\eta}_\sigma|*\sum_{\widetilde{\rho}\sim\rho}\sum_{\tau\subset\widetilde{\rho}}|f_\tau^{k+1}|^2*w_\tau\right]^{p_n/4}, \tag*{(66)} ( τ ⊂ ρ ∑ ∣ h τ ∣ 2 ) p n − 1 /4 ≲ η R C η ( α A N − k β ) ( p n − 1 − n ) /2 ∣ η σ ∣ ∗ ρ ∼ ρ ∑ τ ⊂ ρ ∑ ∣ f τ k + 1 ∣ 2 ∗ w τ p n /4 , (66)
where we used p n − 1 + n = p n p_{n-1}+n=p_n p n − 1 + n = p n . The first inequality in (6.5) and Hölder also show that each B B B on the RHS of (6.4) satisfies
∥ ( ∣ η ~ σ ∣ ∗ ∑ ρ ~ ∼ ρ ∑ τ ⊂ ρ ~ ∣ f τ k + 1 ∣ 2 ∗ w τ ) 1 / 2 ∥ L # p n / 2 ( B ) ≥ α R C η N ( s ) . (67) \left\|\left(|\widetilde{\eta}_\sigma|*\sum_{\widetilde{\rho}\sim\rho}\sum_{\tau\subset\widetilde{\rho}}|f_\tau^{k+1}|^2*w_\tau\right)^{1/2}\right\|_{L_{\#}^{p_n/2}(B)}\ge\frac{\alpha}{R^{C\eta}\mathcal N(s)}. \tag*{(67)} ∣ η σ ∣ ∗ ρ ∼ ρ ∑ τ ⊂ ρ ∑ ∣ f τ k + 1 ∣ 2 ∗ w τ 1/2 L # p n /2 ( B ) ≥ R C η N ( s ) α . (67)
Combining (6.2)–(6.7) and using bounded overlap of the neighbouring parents, we obtain
α p n ∣ Br α , β K ∩ Ω k ∣ ≲ η R C η ∑ σ ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ B ∥ ( s / σ ) V ρ , σ − n B ∈ G ρ ( α R C η ) ∣ B ∣ ∥ ( ∣ η ~ σ ∣ ∗ ∑ τ ⊂ ρ ∣ f τ k + 1 ∣ 2 ∗ w τ ) 1 / 2 ∥ L # p n / 2 ( B ) p n , (68) \alpha^{p_n}|\operatorname{Br}_{\alpha,\beta}^{K}\cap\Omega_k|\lesssim_\eta R^{C\eta}\sum_{\sigma\le s\le1}\sum_{\rho\in\Theta(s)}\sum_{\substack{B\parallel(s/\sigma)V_{\rho,\sigma^{-n}}\\B\in\mathcal G_\rho\left(\frac{\alpha}{R^{C\eta}}\right)}}|B|\left\|\left(|\widetilde{\eta}_\sigma|*\sum_{\tau\subset\rho}|f_\tau^{k+1}|^2*w_\tau\right)^{1/2}\right\|_{L_{\#}^{p_n/2}(B)}^{p_n}, \tag*{(68)} α p n ∣ Br α , β K ∩ Ω k ∣ ≲ η R C η σ ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ B ∥ ( s / σ ) V ρ , σ − n B ∈ G ρ ( R C η α ) ∑ ∣ B ∣ ( ∣ η σ ∣ ∗ τ ⊂ ρ ∑ ∣ f τ k + 1 ∣ 2 ∗ w τ ) 1/2 L # p n /2 ( B ) p n , (68)
where we used α p n − 2 n ( A N − k β ) n − p n − 1 ( A N − k β / α ) p n − 1 − n = 1 \alpha^{p_n-2n}(A^{N-k}\beta)^{n-p_{n-1}}(A^{N-k}\beta/\alpha)^{p_{n-1}-n}=1 α p n − 2 n ( A N − k β ) n − p n − 1 ( A N − k β / α ) p n − 1 − n = 1 .
To remove the remaining convolution with ∣ η ~ σ ∣ |\widetilde{\eta}_\sigma| ∣ η σ ∣ from the RHS of (6.8), we apply Hölder and (2.25), with F = ( ∑ τ ⊂ ρ ∣ f τ k + 1 ∣ 2 ∗ w τ ) p n / 4 F=(\sum_{\tau\subset\rho}|f_\tau^{k+1}|^2*w_\tau)^{p_n/4} F = ( ∑ τ ⊂ ρ ∣ f τ k + 1 ∣ 2 ∗ w τ ) p n /4 and K = ∣ η ~ σ ∣ K=|\widetilde{\eta}_\sigma| K = ∣ η σ ∣ . This gives
α p n ∣ Br α , β K ∩ Ω k ∣ ≲ η R C η ∑ σ ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ B ∥ ( s / σ ) V ρ , σ − n B ∈ G ρ ( α R C η ) ∣ B ∣ ∥ ( ∑ τ ⊂ ρ ∣ f τ k + 1 ∣ 2 ∗ w τ ) 1 / 2 ∥ L # p n / 2 ( B ) p n . (69) \alpha^{p_n}|\operatorname{Br}_{\alpha,\beta}^{K}\cap\Omega_k|\lesssim_\eta R^{C\eta}\sum_{\sigma\le s\le1}\sum_{\rho\in\Theta(s)}\sum_{\substack{B\parallel(s/\sigma)V_{\rho,\sigma^{-n}}\\B\in\mathcal G_\rho\left(\frac{\alpha}{R^{C\eta}}\right)}}|B|\left\|\left(\sum_{\tau\subset\rho}|f_\tau^{k+1}|^2*w_\tau\right)^{1/2}\right\|_{L_{\#}^{p_n/2}(B)}^{p_n}. \tag*{(69)} α p n ∣ Br α , β K ∩ Ω k ∣ ≲ η R C η σ ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ B ∥ ( s / σ ) V ρ , σ − n B ∈ G ρ ( R C η α ) ∑ ∣ B ∣ ( τ ⊂ ρ ∑ ∣ f τ k + 1 ∣ 2 ∗ w τ ) 1/2 L # p n /2 ( B ) p n . (69)
This proves the first inequality in Proposition 2.11.
Case 2: Bounding L L L . The pruning error estimate (2.17), followed by Cauchy–Schwarz over τ N 0 ∈ Θ ( R − 1 / n ) \tau_{N_0}\in\Theta(R^{-1/n}) τ N 0 ∈ Θ ( R − 1/ n ) and (2.16), gives
α 2 ≲ R N 0 1 / n ( ∑ τ N 0 ∣ f τ N 0 N 0 + 1 ∣ 2 ∗ w τ N 0 ) ≲ η R C η N 0 β on L . \alpha^2\lesssim R_{N_0}^{1/n}\left(\sum_{\tau_{N_0}}|f_{\tau_{N_0}}^{N_0+1}|^2*w_{\tau_{N_0}}\right)\lesssim_\eta R^{C\eta}N_0\beta\quad\text{on }L. α 2 ≲ R N 0 1/ n τ N 0 ∑ ∣ f τ N 0 N 0 + 1 ∣ 2 ∗ w τ N 0 ≲ η R C η N 0 β on L .
Since L ⊂ E α , β L\subset E_{\alpha,\beta} L ⊂ E α , β , we have β ≤ ∑ θ ∣ f θ ∣ 2 ∗ w θ \beta\le\sum_\theta|f_\theta|^2*w_\theta β ≤ ∑ θ ∣ f θ ∣ 2 ∗ w θ on L L L , and hence
α p n ∣ L ∣ ≲ η R C η N 0 ∫ ∑ θ ∣ f θ ∣ 2 ∗ w θ > α R C η N 0 ( ∑ θ ∣ f θ ∣ 2 ∗ w θ ) p n / 2 . \alpha^{p_n}|L|\lesssim_\eta R^{C\eta N_0}\int_{\sum_\theta|f_\theta|^2*w_\theta>\frac{\alpha}{R^{C\eta N_0}}}\left(\sum_\theta|f_\theta|^2*w_\theta\right)^{p_n/2}. α p n ∣ L ∣ ≲ η R C η N 0 ∫ ∑ θ ∣ f θ ∣ 2 ∗ w θ > R C η N 0 α ( θ ∑ ∣ f θ ∣ 2 ∗ w θ ) p n /2 .
Applying Lemma 2.8 to { ∣ f θ ∣ 2 ∗ w θ } θ ∈ Θ ( R − 1 / n ) \{|f_\theta|^2*w_\theta\}_{\theta\in\Theta(R^{-1/n})} { ∣ f θ ∣ 2 ∗ w θ } θ ∈ Θ ( R − 1/ n ) , with parameter R R R , amplitude α / R C η N 0 \alpha/R^{C\eta N_0} α / R C η N 0 , and loss η \eta η , bounds this integral and gives
α p n ∣ L ∣ ≲ η R C η N 0 ∑ R − 1 / n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , R V ∈ G ρ ( α R C η N 0 ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w θ ) 1 / 2 ∥ L # p n / 2 ( V ) p n . (70) \alpha^{p_n}|L|\lesssim_\eta R^{C\eta N_0}\sum_{R^{-1/n}\le s\le1}\sum_{\rho\in\Theta(s)}\sum_{\substack{V\parallel V_{\rho,R}\\V\in\mathcal G_\rho\left(\frac{\alpha}{R^{C\eta N_0}}\right)}}|V|\left\|\left(\sum_{\theta\subset\rho}|f_\theta|^2*w_\theta\right)^{1/2}\right\|_{L_{\#}^{p_n/2}(V)}^{p_n}. \tag*{(70)} α p n ∣ L ∣ ≲ η R C η N 0 R − 1/ n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , R V ∈ G ρ ( R C η N 0 α ) ∑ ∣ V ∣ θ ⊂ ρ ∑ ∣ f θ ∣ 2 ∗ w θ 1/2 L # p n /2 ( V ) p n . (70)
This proves the second inequality in (2.18), completing the proof of Proposition 2.11. □
Lemma 6.1. Let Λ ( R ) ≥ 1 \Lambda(R) \ge1 Λ ( R ) ≥ 1 be a constant such that, for every 1 ≤ r ≤ R 1 \le r \le R 1 ≤ r ≤ R , every Schwartz function g g g with supp g ^ ⊂ M n ( r ) \operatorname{supp}\widehat{g} \subset\mathcal{M}^n(r) supp g ⊂ M n ( r ) , and every α > 0 \alpha> 0 α > 0 ,
α p n ∣ { ∣ g ∣ > α } ∣ ≤ Λ ( R ) p n ∑ r − 1 / n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , r V ∈ G ρ ( Λ ( R ) α ) ∣ V ∣ ∥ ( ∑ τ ⊂ ρ ∣ g τ ∣ 2 ∗ w τ ) 1 / 2 ∥ L # p n / 2 ( V ) p n . \alpha^{p_n} \lvert\{ \lvert g\rvert> \alpha\}\rvert\le\Lambda(R)^{p_n} \sum_{r^{-1/n}\le s\le1}\sum_{\rho\in\Theta(s)}\sum_{\substack{V\parallel V_{\rho,r}\\ V\in\mathcal{G}_{\rho}\left(\frac{\Lambda(R)}{\alpha}\right)}} \lvert V\rvert\left\lVert\left(\sum_{\tau\subset\rho}\lvert g_{\tau}\rvert^2*w_{\tau}\right)^{1/2}\right\rVert_{L_{\#}^{p_n/2}(V)}^{p_n}. α p n ∣{∣ g ∣ > α }∣ ≤ Λ ( R ) p n r − 1/ n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , r V ∈ G ρ ( α Λ ( R ) ) ∑ ∣ V ∣ ( τ ⊂ ρ ∑ ∣ g τ ∣ 2 ∗ w τ ) 1/2 L # p n /2 ( V ) p n . Then the same functions satisfy
∫ ∣ g ∣ > α ∣ g ∣ p n ≲ ϵ R ϵ Λ ( R ) p n ∑ r − 1 / n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , r V ∈ G ρ ( R ϵ Λ ( R ) α ) ∣ V ∣ ∥ ( ∑ τ ⊂ ρ ∣ g τ ∣ 2 ∗ w τ ) 1 / 2 ∥ L # p n / 2 ( V ) p n . \int_{\lvert g\rvert>\alpha}\lvert g\rvert^{p_n}\lesssim_{\epsilon}R^{\epsilon}\Lambda(R)^{p_n}\sum_{r^{-1/n}\le s\le1}\sum_{\rho\in\Theta(s)}\sum_{\substack{V\parallel V_{\rho,r}\\ V\in\mathcal{G}_{\rho}\left(\frac{R^{\epsilon}\Lambda(R)}{\alpha}\right)}}\lvert V\rvert\left\lVert\left(\sum_{\tau\subset\rho}\lvert g_{\tau}\rvert^2*w_{\tau}\right)^{1/2}\right\rVert_{L_{\#}^{p_n/2}(V)}^{p_n}. ∫ ∣ g ∣ > α ∣ g ∣ p n ≲ ϵ R ϵ Λ ( R ) p n r − 1/ n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , r V ∈ G ρ ( α R ϵ Λ ( R ) ) ∑ ∣ V ∣ ( τ ⊂ ρ ∑ ∣ g τ ∣ 2 ∗ w τ ) 1/2 L # p n /2 ( V ) p n . Since the proof uses standard dyadic pigeonholing, we postpone it to Appendix B.
Proof of Proposition 2.6
We first bound α p n ∣ { ∣ f ∣ > α } ∣ \alpha^{p_n}\lvert\{\lvert f\rvert>\alpha\}\rvert α p n ∣{∣ f ∣ > α }∣ , then deduce the estimate for ∫ ∣ f ∣ > α ∣ f ∣ p n \int_{\lvert f\rvert>\alpha}\lvert f\rvert^{p_n} ∫ ∣ f ∣ > α ∣ f ∣ p n in (14) . For the region where ∑ θ ∣ f θ ∣ 2 ∗ w θ > α 2 \sum_{\theta}\lvert f_{\theta}\rvert^2*w_{\theta}>\alpha^2 ∑ θ ∣ f θ ∣ 2 ∗ w θ > α 2 , applying Lemma 2.8 to { ∣ f θ ∣ 2 ∗ w θ } θ ∈ Θ ( R − 1 / n ) \{\lvert f_{\theta}\rvert^2*w_{\theta}\}_{\theta\in\Theta(R^{-1/n})} {∣ f θ ∣ 2 ∗ w θ } θ ∈ Θ ( R − 1/ n ) , with parameter R R R , amplitude α \alpha α , and loss η 10 \eta^{10} η 10 , gives
α p n ∣ { ∣ f ∣ > α , ∑ θ ∣ f θ ∣ 2 ∗ w θ > α 2 } ∣ ≤ ∫ ∑ θ ∣ f θ ∣ 2 ∗ w θ > α 2 ( ∑ θ ∣ f θ ∣ 2 ∗ w θ ) p n / 2 ≲ η R C η 10 ∑ R − 1 / n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , R V ∈ G ρ ( R C η 10 α ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w θ ) 1 / 2 ∥ L # p n / 2 ( V ) p n . (71) \begin{aligned}
\alpha^{p_n}\left\lvert\left\{\lvert f\rvert>\alpha,\ \sum_{\theta}\lvert f_{\theta}\rvert^2*w_{\theta}>\alpha^2\right\}\right\rvert
&\le\int_{\sum_{\theta}\lvert f_{\theta}\rvert^2*w_{\theta}>\alpha^2}\left(\sum_{\theta}\lvert f_{\theta}\rvert^2*w_{\theta}\right)^{p_n/2}\\
&\lesssim_{\eta}R^{C\eta^{10}}\sum_{R^{-1/n}\le s\le1}\sum_{\rho\in\Theta(s)}\sum_{\substack{V\parallel V_{\rho,R}\\ V\in\mathcal{G}_{\rho}\left(\frac{R^{C\eta^{10}}}{\alpha}\right)}}\lvert V\rvert\left\lVert\left(\sum_{\theta\subset\rho}\lvert f_{\theta}\rvert^2*w_{\theta}\right)^{1/2}\right\rVert_{L_{\#}^{p_n/2}(V)}^{p_n}. \tag*{(71)}
\end{aligned} α p n { ∣ f ∣ > α , θ ∑ ∣ f θ ∣ 2 ∗ w θ > α 2 } ≤ ∫ ∑ θ ∣ f θ ∣ 2 ∗ w θ > α 2 ( θ ∑ ∣ f θ ∣ 2 ∗ w θ ) p n /2 ≲ η R C η 10 R − 1/ n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , R V ∈ G ρ ( α R C η 10 ) ∑ ∣ V ∣ θ ⊂ ρ ∑ ∣ f θ ∣ 2 ∗ w θ 1/2 L # p n /2 ( V ) p n . (71)
On the remaining region { ∣ f ∣ > α , ∑ θ ∣ f θ ∣ 2 ∗ w θ ≤ α 2 } \{\lvert f\rvert>\alpha,\ \sum_{\theta}\lvert f_{\theta}\rvert^2*w_{\theta}\le\alpha^2\} {∣ f ∣ > α , ∑ θ ∣ f θ ∣ 2 ∗ w θ ≤ α 2 } , the dyadic decomposition leading to (2.9) reduces us to bounding α p n ∣ E α , β ∣ \alpha^{p_n}\lvert E_{\alpha,\beta}\rvert α p n ∣ E α , β ∣ for each dyadic β \beta β . Fix such a β \beta β , and choose a dyadic K ∼ R κ K\sim R^{\kappa} K ∼ R κ , where 0 < κ ≪ η / ( 1000 n 2 ) 0<\kappa\ll\eta/(1000n^2) 0 < κ ≪ η / ( 1000 n 2 ) . The standard broad narrow decomposition gives
∣ f ∣ ≲ max τ ∈ Θ ( K − 1 ) ∣ f τ ∣ + K 100 n max τ 1 , … , τ n K − 1 s e p a r a t e d ∣ f τ 1 ⋯ f τ n ∣ 1 / n . (72) \lvert f\rvert\lesssim\max_{\tau\in\Theta(K^{-1})}\lvert f_{\tau}\rvert+K^{100n}\max_{\substack{\tau_1,\ldots,\tau_n\\K^{-1}\ \mathrm{separated}}}\lvert f_{\tau_1}\cdots f_{\tau_n}\rvert^{1/n}. \tag*{(72)} ∣ f ∣ ≲ τ ∈ Θ ( K − 1 ) max ∣ f τ ∣ + K 100 n τ 1 , … , τ n K − 1 separated max ∣ f τ 1 ⋯ f τ n ∣ 1/ n . (72)
Together with the definition of Br α , β K \operatorname{Br}_{\alpha,\beta}^{K} Br α , β K , this gives
E α , β ⊂ ⋃ τ 0 ∈ Θ ( K − 1 ) { ∣ f τ 0 ∣ > α 100 n } ∪ ⋃ τ 1 , … , τ n K − 1 s e p a r a t e d Br α , β K . E_{\alpha,\beta}\subset\bigcup_{\tau_0\in\Theta(K^{-1})}\left\{\lvert f_{\tau_0}\rvert>\frac{\alpha}{100n}\right\}\cup\bigcup_{\substack{\tau_1,\ldots,\tau_n\\K^{-1}\ \mathrm{separated}}}\operatorname{Br}_{\alpha,\beta}^{K}. E α , β ⊂ τ 0 ∈ Θ ( K − 1 ) ⋃ { ∣ f τ 0 ∣ > 100 n α } ∪ τ 1 , … , τ n K − 1 separated ⋃ Br α , β K .
We estimate α p n ∣ E α , β ∣ \alpha^{p_n}\lvert E_{\alpha,\beta}\rvert α p n ∣ E α , β ∣ by separating the narrow part, where max τ 0 ∈ Θ ( K − 1 ) ∣ f τ 0 ∣ > α / ( 100 n ) \max_{\tau_0\in\Theta(K^{-1})}\lvert f_{\tau_0}\rvert>\alpha/(100n) max τ 0 ∈ Θ ( K − 1 ) ∣ f τ 0 ∣ > α / ( 100 n ) , from its complement, the broad part. In the narrow part, we apply (14) to each f τ 0 f_{\tau_0} f τ 0 after Taylor rescaling, with constant S 1 ( R / K n ) S_1(R/K^n) S 1 ( R / K n ) ; in the broad part, we apply (2.20) to each Br α , β K \operatorname{Br}_{\alpha,\beta}^{K} Br α , β K .
Case 1: Narrow part. For each τ 0 ∈ Θ ( K − 1 ) \tau_0\in\Theta(K^{-1}) τ 0 ∈ Θ ( K − 1 ) , Taylor rescaling to the unit cap sends each θ ⊂ τ 0 \theta\subset\tau_0 θ ⊂ τ 0 to a cap of M n ( R / K n ) \mathcal{M}^n(R/K^n) M n ( R / K n ) at scale K R − 1 / n KR^{-1/n} K R − 1/ n . Applying (14) to the rescaled f τ 0 f_{\tau_0} f τ 0 , with parameter R / K n R/K^n R / K n and amplitude α / ( 100 n ) \alpha/(100n) α / ( 100 n ) , then returning to the original coordinates gives
α p n ∣ { ∣ f τ 0 ∣ > α 100 n } ∣ ≲ S 1 ( R / K n ) p n ∑ R − 1 / n ≤ s ≤ K − 1 ∑ ρ ∈ Θ ( s ) ρ ⊂ τ 0 ∑ V ∥ V ρ , R V ∈ G ρ ( R C η 10 S 1 ( R / K n ) α ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w θ ) 1 / 2 ∥ L # p n / 2 ( V ) p n . (73) \alpha^{p_n}\left\lvert\left\{\lvert f_{\tau_0}\rvert>\frac{\alpha}{100n}\right\}\right\rvert
\lesssim S_1(R/K^n)^{p_n}\sum_{R^{-1/n}\le s\le K^{-1}}\sum_{\substack{\rho\in\Theta(s)\\\rho\subset\tau_0}}\sum_{\substack{V\parallel V_{\rho,R}\\ V\in\mathcal{G}_{\rho}\left(\frac{R^{C\eta^{10}}S_1(R/K^n)}{\alpha}\right)}}\lvert V\rvert\left\lVert\left(\sum_{\theta\subset\rho}\lvert f_{\theta}\rvert^2*w_{\theta}\right)^{1/2}\right\rVert_{L_{\#}^{p_n/2}(V)}^{p_n}. \tag*{(73)} α p n { ∣ f τ 0 ∣ > 100 n α } ≲ S 1 ( R / K n ) p n R − 1/ n ≤ s ≤ K − 1 ∑ ρ ∈ Θ ( s ) ρ ⊂ τ 0 ∑ V ∥ V ρ , R V ∈ G ρ ( α R C η 10 S 1 ( R / K n ) ) ∑ ∣ V ∣ θ ⊂ ρ ∑ ∣ f θ ∣ 2 ∗ w θ 1/2 L # p n /2 ( V ) p n . (73)
Here the averaging box corresponding to ρ ⊂ τ 0 \rho\subset\tau_0 ρ ⊂ τ 0 , ∣ ρ ∣ = s \lvert\rho\rvert=s ∣ ρ ∣ = s , returns with i i i th width K i ( R / K n ) 2 / n ( K s ) 2 − i = R 2 / n s 2 − i K^i(R/K^n)^{2/n}(Ks)^{2-i}=R^{2/n}s^{2-i} K i ( R / K n ) 2/ n ( K s ) 2 − i = R 2/ n s 2 − i , as required for V ρ , R V_{\rho,R} V ρ , R , and we used N τ 0 ( s ) ≤ N ( s ) N_{\tau_0}(s)\le N(s) N τ 0 ( s ) ≤ N ( s ) in the selection condition.
Summing (6.13) over τ 0 \tau_0 τ 0 and using S 1 ( R / K κ ) ≤ S 1 ( R 1 − n κ ) S_1(R/K^\kappa) \le S_1(R^{1-n\kappa}) S 1 ( R / K κ ) ≤ S 1 ( R 1 − nκ ) , we obtain
α p n ∣ { τ 0 ∈ Θ ( K − 1 ) : ∣ f τ 0 ∣ > α 100 n } ∣ ≲ n ( R C η 10 S 1 ( R 1 − n κ ) ) p n ∑ R − 1 / n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , R V ∈ G ρ ( α R C η 10 S 1 ( R 1 − n κ ) ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w θ ) 1 / 2 ∥ L # p n / 2 ( V ) p n . (74) \alpha^{p_n}\left|\left\{\tau_0\in\Theta(K^{-1}): |f_{\tau_0}|>\frac{\alpha}{100n}\right\}\right|
\lesssim_n \left(R^{C\eta^{10}}S_1(R^{1-n\kappa})\right)^{p_n}
\sum_{R^{-1/n}\le s\le1}\sum_{\rho\in\Theta(s)}
\sum_{\substack{V\parallel V_{\rho,R}\\V\in\mathcal{G}_\rho\left(\frac{\alpha}{R^{C\eta^{10}}S_1(R^{1-n\kappa})}\right)}}
|V|\left\|\left(\sum_{\theta\subset\rho}|f_\theta|^2*w_\theta\right)^{1/2}\right\|_{L_\#^{p_n/2}(V)}^{p_n}. \tag*{(74)} α p n { τ 0 ∈ Θ ( K − 1 ) : ∣ f τ 0 ∣ > 100 n α } ≲ n ( R C η 10 S 1 ( R 1 − nκ ) ) p n R − 1/ n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , R V ∈ G ρ ( R C η 10 S 1 ( R 1 − nκ ) α ) ∑ ∣ V ∣ θ ⊂ ρ ∑ ∣ f θ ∣ 2 ∗ w θ 1/2 L # p n /2 ( V ) p n . (74)
Case 2: Broad part. Apply (2.20) to bound α p n ∣ Br α , β K ∣ \alpha^{p_n}| \operatorname{Br}_{\alpha,\beta}^{K}| α p n ∣ Br α , β K ∣ for each separated tuple in (6.12). There are at most K n K^n K n tuples and logarithmically many β \beta β in (2.9). Summing these bounds over the tuples and dyadic β \beta β bounds α p n ∣ { ∣ f ∣ > α , ∑ θ ∣ f θ ∣ 2 ∗ w θ ≤ α 2 , max τ 0 ∣ f τ 0 ∣ ≤ α / ( 100 n ) } ∣ \alpha^{p_n}|\{|f|>\alpha,\sum_\theta|f_\theta|^2*w_\theta\le\alpha^2,\max_{\tau_0}|f_{\tau_0}|\le\alpha/(100n)\}| α p n ∣ { ∣ f ∣ > α , ∑ θ ∣ f θ ∣ 2 ∗ w θ ≤ α 2 , max τ 0 ∣ f τ 0 ∣ ≤ α / ( 100 n )} ∣ by
R C η N 0 ∑ R − 1 / n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , R V ∈ G ρ ( α R C η N 0 ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w θ ) 1 / 2 ∥ L # p n / 2 ( V ) p n R^{C\eta N_0}
\sum_{R^{-1/n}\le s\le1}\sum_{\rho\in\Theta(s)}
\sum_{\substack{V\parallel V_{\rho,R}\\V\in\mathcal{G}_\rho\left(\frac{\alpha}{R^{C\eta N_0}}\right)}}
|V|\left\|\left(\sum_{\theta\subset\rho}|f_\theta|^2*w_\theta\right)^{1/2}\right\|_{L_\#^{p_n/2}(V)}^{p_n} R C η N 0 R − 1/ n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , R V ∈ G ρ ( R C η N 0 α ) ∑ ∣ V ∣ θ ⊂ ρ ∑ ∣ f θ ∣ 2 ∗ w θ 1/2 L # p n /2 ( V ) p n
+ ( R C η S 1 ( R n ) N − N 0 + C ) p n ∑ R − 1 / n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , R V ∈ G ρ ( α R C η S 1 ( R n ) N − N 0 + C ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w θ ) 1 / 2 ∥ L # p n / 2 ( V ) p n . (75) \qquad+\left(R^{C\eta}S_1(R^n)^{N-N_0+C}\right)^{p_n}
\sum_{R^{-1/n}\le s\le1}\sum_{\rho\in\Theta(s)}
\sum_{\substack{V\parallel V_{\rho,R}\\V\in\mathcal{G}_\rho\left(\frac{\alpha}{R^{C\eta}S_1(R^n)^{N-N_0+C}}\right)}}
|V|\left\|\left(\sum_{\theta\subset\rho}|f_\theta|^2*w_\theta\right)^{1/2}\right\|_{L_\#^{p_n/2}(V)}^{p_n}. \tag*{(75)} + ( R C η S 1 ( R n ) N − N 0 + C ) p n R − 1/ n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , R V ∈ G ρ ( R C η S 1 ( R n ) N − N 0 + C α ) ∑ ∣ V ∣ θ ⊂ ρ ∑ ∣ f θ ∣ 2 ∗ w θ 1/2 L # p n /2 ( V ) p n . (75)
Here K n K^n K n is absorbed using κ ≪ η \kappa\ll\eta κ ≪ η .
Completion of the strong type estimate. To combine (71) , (74) , and (75) , put
Λ ( R ) = max { R C η N 0 , R C η S 1 ( R n ) N − N 0 + C , R C η 10 S 1 ( R 1 − n κ ) } . \Lambda(R)=\max\left\{R^{C\eta N_0},R^{C\eta}S_1(R^n)^{N-N_0+C},R^{C\eta^{10}}S_1(R^{1-n\kappa})\right\}. Λ ( R ) = max { R C η N 0 , R C η S 1 ( R n ) N − N 0 + C , R C η 10 S 1 ( R 1 − nκ ) } .
The uniformization reduction following Lemma 2.18 gives, for every Schwartz function g g g with supp g ^ ⊂ M n ( R ) \operatorname{supp}\widehat{g}\subset\mathcal{M}^n(R) supp g ⊂ M n ( R ) ,
α p n ∣ { ∣ g ∣ > α } ∣ ≲ η Λ ( R ) p n ∑ R − 1 / n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , R V ∈ G ρ ( α Λ ( R ) ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ∣ g θ ∣ 2 ∗ w θ ) 1 / 2 ∥ L # p n / 2 ( V ) p n . (76) \alpha^{p_n}|\{|g|>\alpha\}|\lesssim_\eta\Lambda(R)^{p_n}
\sum_{R^{-1/n}\le s\le1}\sum_{\rho\in\Theta(s)}
\sum_{\substack{V\parallel V_{\rho,R}\\V\in\mathcal{G}_\rho\left(\frac{\alpha}{\Lambda(R)}\right)}}
|V|\left\|\left(\sum_{\theta\subset\rho}|g_\theta|^2*w_\theta\right)^{1/2}\right\|_{L_\#^{p_n/2}(V)}^{p_n}. \tag*{(76)} α p n ∣ { ∣ g ∣ > α } ∣ ≲ η Λ ( R ) p n R − 1/ n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , R V ∈ G ρ ( Λ ( R ) α ) ∑ ∣ V ∣ θ ⊂ ρ ∑ ∣ g θ ∣ 2 ∗ w θ 1/2 L # p n /2 ( V ) p n . (76)
The same bound holds for every r ≤ R r\le R r ≤ R , with Λ ( r ) ≤ Λ ( R ) \Lambda(r)\le\Lambda(R) Λ ( r ) ≤ Λ ( R ) . Applying Lemma 6.1 to this estimate proves Proposition 2.6 . □
From cones over moment curves to moment curves: Completion of the proof of Proposition 2.1
In this section, we complete the proof of Proposition 2.1 by establishing Proposition 2.5 . Under the hypothesis of Proposition 2.1 on the cone over the moment curve, this proposition asserts
S 2 ( R ) ≲ S 1 ( R 1 / n ) S 2 ( C R 1 − 1 / n ) . S_2(R)\lesssim S_1(R^{1/n})S_2(CR^{1-1/n}). S 2 ( R ) ≲ S 1 ( R 1/ n ) S 2 ( C R 1 − 1/ n ) .
Combined with Corollary 2.7 , this recurrence yields S 2 ( R ) ≲ 1 S_2(R)\lesssim1 S 2 ( R ) ≲ 1 by iteration.
To prove this recurrence, we first apply S 1 S_1 S 1 at parameter R 1 / n R^{1/n} R 1/ n . Throughout this section, write σ 0 = R − 1 / n 2 \sigma_0=R^{-1/n^2} σ 0 = R − 1/ n 2 , the width of the caps in this application. We keep the convolution weights attached to these initial caps as the caps are subdivided. Their dual boxes have longest side R 1 / n R^{1/n} R 1/ n , so the locally constant property of the original functions f θ f_\theta f θ allows us to remove the weights at the fine scale. For a cap of width σ ≤ σ 0 \sigma\le\sigma_0 σ ≤ σ 0 , rescaling gives the smaller parameter σ n R ≤ R 1 − 1 / n \sigma^nR\le R^{1-1/n} σ n R ≤ R 1 − 1/ n , at which we can apply S 2 S_2 S 2 .
Auxiliary estimates
We establish three auxiliary estimates for the proof of Proposition 2.5 . For each cap τ \tau τ with ∣ τ ∣ ≤ σ 0 |\tau|\le\sigma_0 ∣ τ ∣ ≤ σ 0 , let par σ 0 ( τ ) \operatorname{par}_{\sigma_0}(\tau) par σ 0 ( τ ) denote the cap in Θ ( σ 0 ) \Theta(\sigma_0) Θ ( σ 0 ) containing τ \tau τ . The proofs of the last two lemmas follow those of Lemmas 4.1 and 4.2; we give the details needed to preserve the weights attached to the initial σ 0 \sigma_0 σ 0 caps.
Lemma 7.1. Let f f f be a Schwartz function with supp f ^ ⊂ M n ( R ) \operatorname{supp}\widehat{f}\subset\mathcal{M}^{n}(\mathbb{R}) supp f ⊂ M n ( R ) . For every ϵ > 0 \epsilon>0 ϵ > 0 , α > 0 \alpha>0 α > 0 , and ρ ∈ Θ ( s ) \rho\in\Theta(s) ρ ∈ Θ ( s ) with R − 1 / n ≤ s ≤ 1 R^{-1/n}\leq s\leq1 R − 1/ n ≤ s ≤ 1 , we have
∑ V ∥ V ρ , R V ∈ G ρ ( α ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w par σ 0 ( θ ) ) 1 / 2 ∥ L # p n / 2 ( V ) p n ≲ ϵ R ϵ ∑ V ∥ V ρ , R V ∈ G ρ ( R ϵ ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ) 1 / 2 ∥ L # p n / 2 ( V ) p n . (77) \sum_{\substack{V\parallel V_{\rho,R}\\ V\in\mathcal{G}_{\rho}(\alpha)}} |V|\left\|\left(\sum_{\theta\subset\rho}|f_{\theta}|^{2}*w_{\operatorname{par}_{\sigma_{0}}(\theta)}\right)^{1/2}\right\|_{L_{\#}^{p_{n}/2}(V)}^{p_{n}}
\lesssim_{\epsilon}R^{\epsilon}\sum_{\substack{V\parallel V_{\rho,R}\\ V\in\mathcal{G}_{\rho}(R^{\epsilon})}} |V|\left\|\left(\sum_{\theta\subset\rho}|f_{\theta}|^{2}\right)^{1/2}\right\|_{L_{\#}^{p_{n}/2}(V)}^{p_{n}}. \tag*{(77)} V ∥ V ρ , R V ∈ G ρ ( α ) ∑ ∣ V ∣ θ ⊂ ρ ∑ ∣ f θ ∣ 2 ∗ w par σ 0 ( θ ) 1/2 L # p n /2 ( V ) p n ≲ ϵ R ϵ V ∥ V ρ , R V ∈ G ρ ( R ϵ ) ∑ ∣ V ∣ θ ⊂ ρ ∑ ∣ f θ ∣ 2 1/2 L # p n /2 ( V ) p n . (77)
Proof. Since supp ∣ f θ ^ ∣ 2 ⊂ B ( 0 , C R − 1 / n ) \operatorname{supp}|\widehat{f_{\theta}}|^{2}\subset B(0,CR^{-1/n}) supp ∣ f θ ∣ 2 ⊂ B ( 0 , C R − 1/ n ) and the dual boxes of the initial σ 0 \sigma_{0} σ 0 caps have longest side R 1 / n R^{1/n} R 1/ n , the locally constant property gives ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w par σ 0 ( θ ) ≲ w B R 1 / n ∗ ∑ θ ⊂ ρ ∣ f θ ∣ 2 \sum_{\theta\subset\rho}|f_{\theta}|^{2}*w_{\operatorname{par}_{\sigma_{0}}(\theta)}\lesssim w_{B_{R^{1/n}}}*\sum_{\theta\subset\rho}|f_{\theta}|^{2} ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w par σ 0 ( θ ) ≲ w B R 1/ n ∗ ∑ θ ⊂ ρ ∣ f θ ∣ 2 . Hölder followed by (35) , with kernel w B R 1 / n w_{B_{R^{1/n}}} w B R 1/ n , F = ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ) p n / 4 F=(\sum_{\theta\subset\rho}|f_{\theta}|^{2})^{p_{n}/4} F = ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ) p n /4 and t = ( α / N ( s ) ) p n / 2 t=(\alpha/N(s))^{p_{n}/2} t = ( α / N ( s ) ) p n /2 , gives (7.1). □
Lemma 7.2. Let f f f be a Schwartz function with supp f ^ ⊂ M n ( R ) \operatorname{supp}\widehat{f}\subset\mathcal{M}^{n}(\mathbb{R}) supp f ⊂ M n ( R ) , and let R − 1 / n ≤ σ ≤ σ 0 R^{-1/n}\leq\sigma\leq\sigma_{0} R − 1/ n ≤ σ ≤ σ 0 . For every s ∈ [ σ , 1 ] s\in[\sigma,1] s ∈ [ σ , 1 ] , ϵ > 0 \epsilon>0 ϵ > 0 , and α > 0 \alpha>0 α > 0 , there exists a dyadic σ ′ ∈ [ R − 1 / n , σ ] \sigma'\in[R^{-1/n},\sigma] σ ′ ∈ [ R − 1/ n , σ ] such that
∑ ρ ∈ Θ ( s ) ∑ Q ∈ G ρ ( α ) ∣ τ ∣ = σ ∣ Q ∣ ∥ ( ∑ τ ⊂ ρ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) ) 1 / 2 ∥ L # p n / 2 ( Q ) p n \sum_{\rho\in\Theta(s)}\sum_{\substack{Q\in\mathcal{G}_{\rho}(\alpha)\\ |\tau|=\sigma}}|Q|\left\|\left(\sum_{\tau\subset\rho}|f_{\tau}|^{2}*w_{\operatorname{par}_{\sigma_{0}}(\tau)}\right)^{1/2}\right\|_{L_{\#}^{p_{n}/2}(Q)}^{p_{n}} ρ ∈ Θ ( s ) ∑ Q ∈ G ρ ( α ) ∣ τ ∣ = σ ∑ ∣ Q ∣ ( τ ⊂ ρ ∑ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) ) 1/2 L # p n /2 ( Q ) p n
≲ ϵ R ϵ ∑ R − 1 / n ≤ u ≤ s ∑ ρ ′ ∈ Θ ( u ) ∑ V ∥ V ρ ′ , R V ∈ G ρ ′ ( R ϵ ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ′ ∣ f θ ∣ 2 ∗ w par σ 0 ( θ ) ) 1 / 2 ∥ L # p n / 2 ( V ) p n + R ϵ ∑ σ ′ ≤ u ≤ s ∑ ρ ′ ∈ Θ ( u ) ∑ B ∥ ( u / σ ′ ) V ρ ′ , ( σ ′ ) − n B ∈ G ρ ′ ( R ϵ ) ∣ B ∣ ∥ ( ∑ τ ′ ⊂ ρ ′ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ∗ w par σ 0 ( τ ′ ) ) 1 / 2 ∥ L # p n / 2 ( B ) p n . (78) \begin{aligned}
\lesssim_{\epsilon}R^{\epsilon}\sum_{R^{-1/n}\leq u\leq s}\sum_{\rho'\in\Theta(u)}\sum_{\substack{V\parallel V_{\rho',R}\\ V\in\mathcal{G}_{\rho'}(R^{\epsilon})}}|V|\left\|\left(\sum_{\theta\subset\rho'}|f_{\theta}|^{2}*w_{\operatorname{par}_{\sigma_{0}}(\theta)}\right)^{1/2}\right\|_{L_{\#}^{p_{n}/2}(V)}^{p_{n}} \\
&+R^{\epsilon}\sum_{\sigma'\leq u\leq s}\sum_{\rho'\in\Theta(u)}\sum_{\substack{B\parallel(u/\sigma')V_{\rho',(\sigma')^{-n}}\\ B\in\mathcal{G}_{\rho'}(R^{\epsilon})}}|B|\left\|\left(\sum_{\substack{\tau'\subset\rho'\\ |\tau'|=\sigma'}}|f_{\tau'}|^{2}*w_{\operatorname{par}_{\sigma_{0}}(\tau')}\right)^{1/2}\right\|_{L_{\#}^{p_{n}/2}(B)}^{p_{n}}. \tag*{(78)}
\end{aligned} ≲ ϵ R ϵ R − 1/ n ≤ u ≤ s ∑ ρ ′ ∈ Θ ( u ) ∑ V ∥ V ρ ′ , R V ∈ G ρ ′ ( R ϵ ) ∑ ∣ V ∣ θ ⊂ ρ ′ ∑ ∣ f θ ∣ 2 ∗ w par σ 0 ( θ ) 1/2 L # p n /2 ( V ) p n + R ϵ σ ′ ≤ u ≤ s ∑ ρ ′ ∈ Θ ( u ) ∑ B ∥ ( u / σ ′ ) V ρ ′ , ( σ ′ ) − n B ∈ G ρ ′ ( R ϵ ) ∑ ∣ B ∣ τ ′ ⊂ ρ ′ ∣ τ ′ ∣ = σ ′ ∑ ∣ f τ ′ ∣ 2 ∗ w par σ 0 ( τ ′ ) 1/2 L # p n /2 ( B ) p n . (78)
Proof. As in the proof of Lemma 4.1, for each ρ ∈ Θ ( s ) \rho\in\Theta(s) ρ ∈ Θ ( s ) , decompose ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) \sum_{\tau\subset\rho,\ |\tau|=\sigma}|f_{\tau}|^{2}*w_{\operatorname{par}_{\sigma_{0}}(\tau)} ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) using η ≤ R − 1 / n \eta_{\leq R^{-1/n}} η ≤ R − 1/ n and η σ ′ , R − 1 / n < σ ′ ≤ σ \eta_{\sigma'}, R^{-1/n}<\sigma'\leq\sigma η σ ′ , R − 1/ n < σ ′ ≤ σ . We will bound the high frequency terms by the second sum on the RHS of (7.2), and the low frequency term by the first.
Case 1: The high frequency part dominates. After summing the high frequency terms over ρ \rho ρ , dyadic pigeonholing gives a single width σ ′ \sigma' σ ′ whose contribution bounds the total high frequency contribution up to an R ϵ R^{\epsilon} R ϵ loss. We follow Case 1 of the proof of Lemma 4.1, with f τ f_{\tau} f τ and w par σ 0 ( τ ) w_{\operatorname{par}_{\sigma_{0}}(\tau)} w par σ 0 ( τ ) in place of f τ ′ f_{\tau'} f τ ′ and w τ ′ w_{\tau'} w τ ′ .
Convolution with this weight does not enlarge Fourier support. Moreover, neighbouring σ ′ \sigma' σ ′ caps τ ′ , τ ′ ′ ⊂ τ \tau',\tau''\subset\tau τ ′ , τ ′′ ⊂ τ satisfy par σ 0 ( τ ′ ) = par σ 0 ( τ ′ ′ ) = par σ 0 ( τ ) \operatorname{par}_{\sigma_{0}}(\tau')=\operatorname{par}_{\sigma_{0}}(\tau'')=\operatorname{par}_{\sigma_{0}}(\tau) par σ 0 ( τ ′ ) = par σ 0 ( τ ′′ ) = par σ 0 ( τ ) , so Cauchy–Schwarz gives
∣ ( f τ ′ f τ ′ ′ ‾ ) ∗ w par σ 0 ( τ ) ∗ η ~ σ ′ ∣ ≲ ∣ η ~ σ ′ ∣ ∗ [ ( ∣ f τ ′ ∣ 2 + ∣ f τ ′ ′ ∣ 2 ) ∗ w par σ 0 ( τ ) ] . \left|(f_{\tau'}\overline{f_{\tau''}})*w_{\operatorname{par}_{\sigma_{0}}(\tau)}*\widetilde{\eta}_{\sigma'}\right|
\lesssim|\widetilde{\eta}_{\sigma'}|*\left[(|f_{\tau'}|^{2}+|f_{\tau''}|^{2})*w_{\operatorname{par}_{\sigma_{0}}(\tau)}\right]. ( f τ ′ f τ ′′ ) ∗ w par σ 0 ( τ ) ∗ η σ ′ ≲ ∣ η σ ′ ∣ ∗ [ ( ∣ f τ ′ ∣ 2 + ∣ f τ ′′ ∣ 2 ) ∗ w par σ 0 ( τ ) ] .
Thus the rescaling and subsequent estimates in Case 1 of the proof of Lemma 4.1 apply with these weights. Removing ∣ η ~ σ ′ ∣ |\widetilde{\eta}_{\sigma'}| ∣ η σ ′ ∣ by Hölder and (35) , as in that case, gives the second sum on the RHS of (7.2).
Case 2: The low frequency part dominates. Expand f τ = ∑ θ ⊂ τ f θ f_{\tau}=\sum_{\theta\subset\tau}f_{\theta} f τ = ∑ θ ⊂ τ f θ . The cutoff η ≤ R − 1 / n \eta_{\leq R^{-1/n}} η ≤ R − 1/ n selects only products of neighbouring fine pieces. Since w par σ 0 ( θ ) = w par σ 0 ( τ ) w_{\operatorname{par}_{\sigma_{0}}(\theta)}=w_{\operatorname{par}_{\sigma_{0}}(\tau)} w par σ 0 ( θ ) = w par σ 0 ( τ ) for θ ⊂ τ \theta\subset\tau θ ⊂ τ , we obtain
∣ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) ) ∗ η ~ ≤ R − 1 / n ∣ ≲ ∣ η ~ ≤ R − 1 / n ∣ ∗ ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w par σ 0 ( θ ) . \left|\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_{\tau}|^{2}*w_{\operatorname{par}_{\sigma_{0}}(\tau)}\right)*\widetilde{\eta}_{\leq R^{-1/n}}\right|
\lesssim|\widetilde{\eta}_{\leq R^{-1/n}}|*\sum_{\theta\subset\rho}|f_{\theta}|^{2}*w_{\operatorname{par}_{\sigma_{0}}(\theta)}. τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) ∗ η ≤ R − 1/ n ≲ ∣ η ≤ R − 1/ n ∣ ∗ θ ⊂ ρ ∑ ∣ f θ ∣ 2 ∗ w par σ 0 ( θ ) .
Applying (32) to the sum over Q Q Q on the LHS of (7.2), followed by (34) with F = ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w par σ 0 ( θ ) F=\sum_{\theta\subset\rho}|f_{\theta}|^{2}*w_{\operatorname{par}_{\sigma_{0}}(\theta)} F = ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w par σ 0 ( θ ) , kernel ∣ η ~ ≤ R − 1 / n ∣ |\widetilde{\eta}_{\leq R^{-1/n}}| ∣ η ≤ R − 1/ n ∣ , and exponent p n / 2 p_{n}/2 p n /2 , gives
LHS of (7.2) ≲ ∑ ρ ∈ Θ ( s ) ∫ ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w par σ 0 ( θ ) ≳ [ α / ( R ϵ N ( s ) ) ] 2 ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w par σ 0 ( θ ) ) p n / 2 . (79) \text{LHS of (7.2)}\lesssim\sum_{\rho\in\Theta(s)}\int_{\sum_{\theta\subset\rho}|f_{\theta}|^{2}*w_{\operatorname{par}_{\sigma_{0}}(\theta)}\gtrsim[\alpha/(R^{\epsilon}N(s))]^{2}}\left(\sum_{\theta\subset\rho}|f_{\theta}|^{2}*w_{\operatorname{par}_{\sigma_{0}}(\theta)}\right)^{p_{n}/2}. \tag*{(79)} LHS of (7.2) ≲ ρ ∈ Θ ( s ) ∑ ∫ ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∗ w par σ 0 ( θ ) ≳ [ α / ( R ϵ N ( s )) ] 2 θ ⊂ ρ ∑ ∣ f θ ∣ 2 ∗ w par σ 0 ( θ ) p n /2 . (79)
To estimate each integral in (7.3), we repeat the Taylor rescaling and application of Lemma 2.8 following (5.3). For each ρ \rho ρ , let t ρ t_{\rho} t ρ be the left endpoint of the interval indexing ρ \rho ρ . The rescaled family { ( ∣ f θ ∣ 2 ∗ w par σ 0 ( θ ) ) ( L t ρ , s ⋅ ) } θ ⊂ ρ \{(|f_{\theta}|^{2}*w_{\operatorname{par}_{\sigma_{0}}(\theta)})(L_{t_{\rho},s}\cdot)\}_{\theta\subset\rho} {( ∣ f θ ∣ 2 ∗ w par σ 0 ( θ ) ) ( L t ρ , s ⋅ ) } θ ⊂ ρ satisfies the Fourier support hypothesis of Lemma 2.8 at parameter s n R s^{n}R s n R . Applying that lemma with amplitude α / ( R ϵ N ( s ) ) \alpha/(R^{\epsilon}N(s)) α / ( R ϵ N ( s )) , returning to the original coordinates, and summing over ρ \rho ρ bounds (7.3) by the first sum on the RHS of (7.2), where we used N ( s ) N ρ ( u ) ≲ ϵ N ( u ) N(s)N_{\rho}(u) \lesssim_{\epsilon} N(u) N ( s ) N ρ ( u ) ≲ ϵ N ( u ) for ρ ′ ⊂ ρ \rho' \subset\rho ρ ′ ⊂ ρ , ∣ ρ ′ ∣ = u |\rho'|=u ∣ ρ ′ ∣ = u . This proves (7.2). □ \square □
Lemma 7.3. Let f f f be a Schwartz function with supp f ^ ⊂ M n ( R ) \operatorname{supp}\widehat{f} \subset\mathcal{M}^{n}(\mathbb{R}) supp f ⊂ M n ( R ) , and let R − 1 / n ≤ σ ≤ σ 0 R^{-1/n} \leq\sigma\leq\sigma_{0} R − 1/ n ≤ σ ≤ σ 0 . For every s ∈ [ σ , 1 ] s \in[\sigma,1] s ∈ [ σ , 1 ] , ϵ > 0 \epsilon> 0 ϵ > 0 , and α > 0 \alpha> 0 α > 0 , we have
∑ ρ ∈ Θ ( s ) ∑ B ∈ G ρ ( α ) ∣ B ∣ = ( s / σ ) V ρ , σ − n ∣ B ∣ ∥ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) ) 1 / 2 ∥ L # p n ( B ) p n ≲ ϵ R ϵ ∑ ρ ∈ Θ ( s ) ∑ Q ∈ G ρ ′ ( R ϵ ) Q ∥ ( ρ ′ ) ∗ ∣ Q ∣ ∥ ( ∑ τ ′ ⊂ ρ ′ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ∗ w par σ 0 ( τ ′ ) ) 1 / 2 ∥ L # p n / 2 ( Q ) p n / 2 \sum_{\rho\in\Theta(s)}\sum_{\substack{B \in\mathcal{G}_{\rho}(\alpha)\\ |B|=(s/\sigma)V_{\rho,\sigma^{-n}}}} |B|\left\|\left(\sum_{\substack{\tau\subset\rho\\ |\tau|=\sigma}} |f_{\tau}|^{2} * w_{\operatorname{par}_{\sigma_{0}}(\tau)}\right)^{1/2}\right\|_{L_{\#}^{p_{n}}(B)}^{p_{n}}
\lesssim_{\epsilon} R^{\epsilon}\sum_{\rho\in\Theta(s)}\sum_{\substack{Q \in\mathcal{G}_{\rho'}(R^{\epsilon})\\ Q \parallel(\rho')^{*}}} |Q|\left\|\left(\sum_{\substack{\tau' \subset\rho'\\ |\tau'|=\sigma'}} |f_{\tau'}|^{2} * w_{\operatorname{par}_{\sigma_{0}}(\tau')}\right)^{1/2}\right\|_{L_{\#}^{p_{n}/2}(Q)}^{p_{n}/2} ρ ∈ Θ ( s ) ∑ B ∈ G ρ ( α ) ∣ B ∣ = ( s / σ ) V ρ , σ − n ∑ ∣ B ∣ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) 1/2 L # p n ( B ) p n ≲ ϵ R ϵ ρ ∈ Θ ( s ) ∑ Q ∈ G ρ ′ ( R ϵ ) Q ∥ ( ρ ′ ) ∗ ∑ ∣ Q ∣ τ ′ ⊂ ρ ′ ∣ τ ′ ∣ = σ ′ ∑ ∣ f τ ′ ∣ 2 ∗ w par σ 0 ( τ ′ ) 1/2 L # p n /2 ( Q ) p n /2 + R ϵ ∑ ρ ′ ∈ Θ ( s ′ ) ∑ Q ∈ G ρ ′ ( R ϵ ) Q ∥ ( ρ ′ ) ∗ ∣ Q ∣ ∥ ( ∑ τ ⊂ ρ ′ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) ) 1 / 2 ∥ L # p n / 2 ( Q ) p n / 2 , (80) {}+R^{\epsilon}\sum_{\rho' \in\Theta(s')}\sum_{\substack{Q \in\mathcal{G}_{\rho'}(R^{\epsilon})\\ Q \parallel(\rho')^{*}}} |Q|\left\|\left(\sum_{\substack{\tau\subset\rho'\\ |\tau|=\sigma}} |f_{\tau}|^{2} * w_{\operatorname{par}_{\sigma_{0}}(\tau)}\right)^{1/2}\right\|_{L_{\#}^{p_{n}/2}(Q)}^{p_{n}/2}, \tag*{(80)} + R ϵ ρ ′ ∈ Θ ( s ′ ) ∑ Q ∈ G ρ ′ ( R ϵ ) Q ∥ ( ρ ′ ) ∗ ∑ ∣ Q ∣ τ ⊂ ρ ′ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) 1/2 L # p n /2 ( Q ) p n /2 , (80) where σ ′ = max { R − 1 / n , σ 2 / s } \sigma'=\max\{R^{-1/n},\sigma^{2}/s\} σ ′ = max { R − 1/ n , σ 2 / s } and s ′ = s ( σ / s ) 1 / ( n − 1 ) ∈ [ σ , s ] s'=s(\sigma/s)^{1/(n-1)} \in[\sigma,s] s ′ = s ( σ / s ) 1/ ( n − 1 ) ∈ [ σ , s ] .
Proof. As in the proof of Lemma 4.2, for each ρ ∈ Θ ( s ) \rho\in\Theta(s) ρ ∈ Θ ( s ) , decompose ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) \sum_{\tau\subset\rho,\,|\tau|=\sigma}|f_{\tau}|^{2} * w_{\operatorname{par}_{\sigma_{0}}(\tau)} ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) using η ≤ σ ′ \eta_{\leq\sigma'} η ≤ σ ′ and η λ \eta_{\lambda} η λ , where σ ′ < λ ≲ σ \sigma' < \lambda\lesssim\sigma σ ′ < λ ≲ σ , and separate the boxes B ∈ G ρ ( α ) B \in\mathcal{G}_{\rho}(\alpha) B ∈ G ρ ( α ) according to whether their averages are controlled by the low frequency part or by a high frequency term. We will bound these contributions by the first and second sums on the RHS of (7.4), respectively.
Case 1: The low frequency part dominates. As in Case 2 of the proof of Lemma 7.2, expand each f τ f_{\tau} f τ into its σ ′ \sigma' σ ′ pieces. Since par σ 0 ( τ ′ ) = par σ 0 ( τ ) \operatorname{par}_{\sigma_{0}}(\tau')=\operatorname{par}_{\sigma_{0}}(\tau) par σ 0 ( τ ′ ) = par σ 0 ( τ ) for τ ′ ⊂ τ \tau' \subset\tau τ ′ ⊂ τ , we obtain
∣ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) ) ∗ η ˇ ≤ σ ′ ∣ ≲ ∣ η ˇ ≤ σ ′ ∣ ∗ ∑ τ ′ ⊂ ρ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ∗ w par σ 0 ( τ ′ ) . \left|\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}} |f_{\tau}|^{2} * w_{\operatorname{par}_{\sigma_{0}}(\tau)}\right) * \check{\eta}_{\leq\sigma'}\right|
\lesssim|\check{\eta}_{\leq\sigma'}| * \sum_{\substack{\tau'\subset\rho\\|\tau'|=\sigma'}} |f_{\tau'}|^{2} * w_{\operatorname{par}_{\sigma_{0}}(\tau')}. τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) ∗ η ˇ ≤ σ ′ ≲ ∣ η ˇ ≤ σ ′ ∣ ∗ τ ′ ⊂ ρ ∣ τ ′ ∣ = σ ′ ∑ ∣ f τ ′ ∣ 2 ∗ w par σ 0 ( τ ′ ) .
Applying Hölder and (2.25) to the sum over these B B B , with kernel ∣ η ˇ ≤ σ ′ ∣ |\check{\eta}_{\leq\sigma'}| ∣ η ˇ ≤ σ ′ ∣ and F = ( ∑ τ ′ ⊂ ρ , ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ∗ w par σ 0 ( τ ′ ) ) p n / 4 F=\left(\sum_{\tau'\subset\rho,\,|\tau'|=\sigma'}|f_{\tau'}|^{2}*w_{\operatorname{par}_{\sigma_{0}}(\tau')}\right)^{p_{n}/4} F = ( ∑ τ ′ ⊂ ρ , ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ∗ w par σ 0 ( τ ′ ) ) p n /4 , removes the convolution with ∣ η ˇ ≤ σ ′ ∣ |\check{\eta}_{\leq\sigma'}| ∣ η ˇ ≤ σ ′ ∣ . Since B 0 = ( s / σ ) V ρ , σ − n = ( s / σ ) 3 ρ ∗ B_{0}=(s/\sigma)V_{\rho,\sigma^{-n}}=(s/\sigma)^{3}\rho^{*} B 0 = ( s / σ ) V ρ , σ − n = ( s / σ ) 3 ρ ∗ and s ≥ σ s\geq\sigma s ≥ σ , Lemma 2.14, applied with this F F F , B 0 B_{0} B 0 , and U 0 = ρ ∗ U_{0}=\rho^{*} U 0 = ρ ∗ , bounds the contribution of these boxes by the first sum on the RHS of (7.4).
Case 2: The high frequency part dominates. For each remaining B B B , dyadic pigeonholing gives a width λ \lambda λ , σ ′ < λ ≲ σ \sigma'<\lambda\lesssim\sigma σ ′ < λ ≲ σ , for which the term with η λ \eta_{\lambda} η λ controls its average. Fix λ \lambda λ and consider the boxes with this choice. We follow Case 2 of the proof of Lemma 4.2, replacing ∣ f τ ∣ 2 ∗ w τ |f_{\tau}|^{2}*w_{\tau} ∣ f τ ∣ 2 ∗ w τ throughout by ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) |f_{\tau}|^{2}*w_{\operatorname{par}_{\sigma_{0}}(\tau)} ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) . Convolution with the inherited weight does not enlarge Fourier support, so the rescaling and the Fourier support calculation for multiplication by χ B \chi_{B} χ B remain valid. We use exponent p n / 4 p_{n}/4 p n /4 when n ≥ 4 n\geq4 n ≥ 4 and 7 / 3 7/3 7/3 when n = 3 n=3 n = 3 , as in that proof. The calculation leading to (4.7) therefore gives that inequality with the same replacement.
Multiplying this inequality by ∣ B ∣ |B| ∣ B ∣ and summing over the boxes with this λ \lambda λ , we apply (2.24) with F = ∑ τ ⊂ ρ ′ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) F=\sum_{\tau\subset\rho',\,|\tau|=\sigma}|f_{\tau}|^{2}*w_{\operatorname{par}_{\sigma_{0}}(\tau)} F = ∑ τ ⊂ ρ ′ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) , kernel ∣ η ˇ λ ∣ |\check{\eta}_{\lambda}| ∣ η ˇ λ ∣ , and exponent p n / 2 p_{n}/2 p n /2 . The locally constant property on Q ∥ ( ρ ′ ) ∗ Q\parallel(\rho')^{*} Q ∥ ( ρ ′ ) ∗ , together with N ( s ′ ) ≥ N ( s ) N(s')\geq N(s) N ( s ′ ) ≥ N ( s ) , then bounds the contribution of these boxes to the LHS of (7.4) by
R ϵ ∑ τ ⊂ ρ ′ , ∣ τ ∣ = σ ∑ Q ∥ ( ρ ′ ) ∗ Q ∈ G ρ ′ ( R ϵ ) ∣ Q ∣ ∥ ( ∑ τ ⊂ ρ ′ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) ) 1 / 2 ∥ L # p n / 2 ( Q ) p n / 2 . R^{\epsilon}\sum_{\tau\subset\rho',\,|\tau|=\sigma}\sum_{\substack{Q\parallel(\rho')^{*}\\Q\in\mathcal{G}_{\rho'}(R^{\epsilon})}} |Q|\left\|\left(\sum_{\substack{\tau\subset\rho'\\|\tau|=\sigma}} |f_{\tau}|^{2} * w_{\operatorname{par}_{\sigma_{0}}(\tau)}\right)^{1/2}\right\|_{L_{\#}^{p_{n}/2}(Q)}^{p_{n}/2}. R ϵ τ ⊂ ρ ′ , ∣ τ ∣ = σ ∑ Q ∥ ( ρ ′ ) ∗ Q ∈ G ρ ′ ( R ϵ ) ∑ ∣ Q ∣ τ ⊂ ρ ′ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) 1/2 L # p n /2 ( Q ) p n /2 .
Summing over ρ \rho ρ and λ \lambda λ gives the second sum on the RHS of (7.4). Together with Case 1, this proves (7.4). □ \square □
Proof of Proposition 2.5
We combine the three preceding lemmas in an iteration analogous to that in Section 5. Fix η > 0 \eta>0 η > 0 sufficiently small and choose 0 < ϵ ≪ η 0<\epsilon\ll\eta 0 < ϵ ≪ η . Let f f f be a Schwartz function with supp f ^ ⊂ M n ( R ) \operatorname{supp}\widehat{f}\subset\mathcal{M}^{n}(\mathbb{R}) supp f ⊂ M n ( R ) , and let α > 0 \alpha>0 α > 0 . Applying (2.4) to f f f at parameter R 1 / n R^{1/n} R 1/ n , with fine pieces { f τ } τ ∈ Θ ( σ 0 ) \{f_\tau\}_{\tau\in\Theta(\sigma_0)} { f τ } τ ∈ Θ ( σ 0 ) , and then Lemma 2.14 to the sums over V ∥ V ρ , R 1 / n V\parallel V_{\rho,R^{1/n}} V ∥ V ρ , R 1/ n , gives
∫ ∣ f ∣ > α ∣ f ∣ p n ≲ S 1 ( R 1 / n ) p n ∑ σ 0 ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ Q ∈ G ρ ( Q ∥ f ∥ ∗ S 1 ( R 1 / n ) ) ∣ Q ∣ ∥ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ 0 ∣ f τ ∣ 2 ∗ w τ ) 1 / 2 ∥ L # p n / 2 ( Q ) p n . (81) \int_{\lvert f\rvert>\alpha}\lvert f\rvert^{p_n}\lesssim S_1(R^{1/n})^{p_n}\sum_{\sigma_0\leq s\leq1}\sum_{\rho\in\Theta(s)}\sum_{Q\in\mathcal{G}_\rho\left(\frac{Q\lVert f\rVert_*}{S_1(R^{1/n})}\right)}\lvert Q\rvert\left\lVert\left(\sum_{\substack{\tau\subset\rho\\\lvert\tau\rvert=\sigma_0}}\lvert f_\tau\rvert^2*w_\tau\right)^{1/2}\right\rVert_{L_{\#}^{p_n/2}(Q)}^{p_n}. \tag*{(81)} ∫ ∣ f ∣ > α ∣ f ∣ p n ≲ S 1 ( R 1/ n ) p n σ 0 ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ Q ∈ G ρ ( S 1 ( R 1/ n ) Q ∥ f ∥ ∗ ) ∑ ∣ Q ∣ τ ⊂ ρ ∣ τ ∣ = σ 0 ∑ ∣ f τ ∣ 2 ∗ w τ 1/2 L # p n /2 ( Q ) p n . (81)
Here we used V ρ , R 1 / n = ( s / σ 0 ) ρ ∗ V_{\rho,R^{1/n}}=(s/\sigma_0)\rho^* V ρ , R 1/ n = ( s / σ 0 ) ρ ∗ and s / σ 0 ≥ 1 s/\sigma_0\geq1 s / σ 0 ≥ 1 .
Dyadic pigeonholing reduces the sum over s s s in (81) to one scale s ∈ [ σ 0 , 1 ] s\in[\sigma_0,1] s ∈ [ σ 0 , 1 ] . For this s s s , we apply Lemma 7.2 to the corresponding sum over ρ \rho ρ and Q Q Q , with σ = σ 0 \sigma=\sigma_0 σ = σ 0 and amplitude α / S 1 ( R 1 / n ) \alpha/S_1(R^{1/n}) α / S 1 ( R 1/ n ) .
If the first sum over V V V on the RHS of Lemma 7.2 dominates, applying Lemma 7.1 to that sum gives
∫ ∣ f ∣ > α ∣ f ∣ p n ≲ ϵ R C ϵ S 1 ( R 1 / n ) p n ∑ R − 1 / n ≤ u ≤ 1 ∑ ρ ′ ∈ Θ ( u ) ∑ V ∈ G ρ ′ ( V ∥ V ρ ′ , R R C ϵ S 1 ( R 1 / n ) ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ′ ∣ f θ ∣ 2 ) 1 / 2 ∥ L # p n / 2 ( V ) p n . (82) \int_{\lvert f\rvert>\alpha}\lvert f\rvert^{p_n}\lesssim_\epsilon R^{C\epsilon}S_1(R^{1/n})^{p_n}\sum_{R^{-1/n}\leq u\leq1}\sum_{\rho'\in\Theta(u)}\sum_{V\in\mathcal{G}_{\rho'}\left(\frac{V\lVert V_{\rho',R}}{R^{C\epsilon}S_1(R^{1/n})}\right)}\lvert V\rvert\left\lVert\left(\sum_{\theta\subset\rho'}\lvert f_\theta\rvert^2\right)^{1/2}\right\rVert_{L_{\#}^{p_n/2}(V)}^{p_n}. \tag*{(82)} ∫ ∣ f ∣ > α ∣ f ∣ p n ≲ ϵ R C ϵ S 1 ( R 1/ n ) p n R − 1/ n ≤ u ≤ 1 ∑ ρ ′ ∈ Θ ( u ) ∑ V ∈ G ρ ′ ( R C ϵ S 1 ( R 1/ n ) V ∥ V ρ ′ , R ) ∑ ∣ V ∣ θ ⊂ ρ ′ ∑ ∣ f θ ∣ 2 1/2 L # p n /2 ( V ) p n . (82)
Since S 2 ( R 1 − 1 / n ) ≥ 1 S_2(R^{1-1/n})\geq1 S 2 ( R 1 − 1/ n ) ≥ 1 , this gives the bound required for Proposition 2.5, and the iteration stops.
If instead the second sum over B B B on the RHS of Lemma 7.2 dominates, apply dyadic pigeonholing to fix s s s . Writing σ \sigma σ for the width of the caps grouped at this scale, we obtain
∫ ∣ f ∣ > α ∣ f ∣ p n ≲ ϵ R C ϵ S 1 ( R 1 / n ) p n ∑ ρ ∈ Θ ( s ) ∑ B ∈ G ρ ( B ∥ ( s / σ ) V ρ , σ − n ∥ R C ϵ S 1 ( R 1 / n ) ) ∣ B ∣ ∥ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) ) 1 / 2 ∥ L # p n / 2 ( B ) p n . (83) \int_{\lvert f\rvert>\alpha}\lvert f\rvert^{p_n}\lesssim_\epsilon R^{C\epsilon}S_1(R^{1/n})^{p_n}\sum_{\rho\in\Theta(s)}\sum_{B\in\mathcal{G}_\rho\left(\frac{B\lVert(s/\sigma)V_{\rho,\sigma^{-n}}\rVert}{R^{C\epsilon}S_1(R^{1/n})}\right)}\lvert B\rvert\left\lVert\left(\sum_{\substack{\tau\subset\rho\\\lvert\tau\rvert=\sigma}}\lvert f_\tau\rvert^2*w_{\operatorname{par}_{\sigma_0}(\tau)}\right)^{1/2}\right\rVert_{L_{\#}^{p_n/2}(B)}^{p_n}. \tag*{(83)} ∫ ∣ f ∣ > α ∣ f ∣ p n ≲ ϵ R C ϵ S 1 ( R 1/ n ) p n ρ ∈ Θ ( s ) ∑ B ∈ G ρ ( R C ϵ S 1 ( R 1/ n ) B ∥( s / σ ) V ρ , σ − n ∥ ) ∑ ∣ B ∣ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) 1/2 L # p n /2 ( B ) p n . (83)
where R − 1 / n ≤ σ ≤ σ 0 R^{-1/n}\leq\sigma\leq\sigma_0 R − 1/ n ≤ σ ≤ σ 0 and σ ≤ s ≤ 1 \sigma\leq s\leq1 σ ≤ s ≤ 1 .
As in Section 5, caps of width σ \sigma σ are grouped inside caps of width s s s . Using Lemmas 7.2 and 7.3, together with the defining estimate for S 2 S_2 S 2 after Taylor rescaling, we either replace the sum over B B B in (83) by sums over V ∥ V ρ ′ , R V\parallel V_{\rho',R} V ∥ V ρ ′ , R involving the original functions f θ f_\theta f θ , as in (82) , and stop, or return to a bound of the form (83) with ( σ , s ) ↦ ( σ ′ , s ′ ) (\sigma,s)\mapsto(\sigma',s') ( σ , s ) ↦ ( σ ′ , s ′ ) , σ ′ ≤ σ \sigma'\leq\sigma σ ′ ≤ σ , and s ′ ≤ s s'\leq s s ′ ≤ s , where σ ′ ≪ σ \sigma'\ll\sigma σ ′ ≪ σ or s ′ ≪ s s'\ll s s ′ ≪ s . At each step, we also update the factor outside the sum and the denominator in the selection condition. We will bound the accumulated factor by ( R C η S 1 ( R 1 / n ) S 2 ( R 1 − 1 / n ) ) p n (R^{C\eta}S_1(R^{1/n})S_2(R^{1-1/n}))^{p_n} ( R C η S 1 ( R 1/ n ) S 2 ( R 1 − 1/ n ) ) p n and keep the final amplitude at least α / ( R C η S 1 ( R 1 / n ) S 2 ( R 1 − 1 / n ) ) \alpha/(R^{C\eta}S_1(R^{1/n})S_2(R^{1-1/n})) α / ( R C η S 1 ( R 1/ n ) S 2 ( R 1 − 1/ n )) , with the initial value of α \alpha α .
We first deal with the case σ = R − 1 / n \sigma=R^{-1/n} σ = R − 1/ n . In this case, the derivation of (5.2) from (5.1) in Section 5, together with Lemma 7.1, bounds the RHS of (83) by the RHS of (82) , and the iteration stops.
We may therefore assume σ > R − 1 / n \sigma>R^{-1/n} σ > R − 1/ n . We distinguish the cases s ≤ R η 3 σ s\leq R^{\eta^3}\sigma s ≤ R η 3 σ and s > R η 3 σ s>R^{\eta^3}\sigma s > R η 3 σ .
Case 1: s ≤ R η 3 σ s\leq R^{\eta^3}\sigma s ≤ R η 3 σ
Each ρ \rho ρ contains at most R η 3 R^{\eta^3} R η 3 caps of width σ \sigma σ . Applying the argument leading to (5.4) to the RHS of (83) , with f τ f_\tau f τ and w par σ 0 ( τ ) w_{\operatorname{par}_{\sigma_0}(\tau)} w par σ 0 ( τ ) in place of f τ j f_\tau^j f τ j and w τ w_\tau w τ , gives
∫ ∣ f ∣ > α ∣ f ∣ p n ≲ η R C η 3 S 1 ( R 1 / n ) p n ∑ ρ ∈ Θ ( s ) ∑ τ ⊂ ρ ∣ τ ∣ = σ ∫ ∣ f τ ∣ ≳ α R C η 3 S 1 ( R 1 / n ) N ( s ) ∣ f τ ∣ p n . (84) \int_{\lvert f\rvert>\alpha}\lvert f\rvert^{p_n}\lesssim_\eta R^{C\eta^3}S_1(R^{1/n})^{p_n}\sum_{\rho\in\Theta(s)}\sum_{\substack{\tau\subset\rho\\\lvert\tau\rvert=\sigma}}\int_{\lvert f_\tau\rvert\gtrsim\frac{\alpha}{R^{C\eta^3}S_1(R^{1/n})N(s)}}\lvert f_\tau\rvert^{p_n}. \tag*{(84)} ∫ ∣ f ∣ > α ∣ f ∣ p n ≲ η R C η 3 S 1 ( R 1/ n ) p n ρ ∈ Θ ( s ) ∑ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∫ ∣ f τ ∣ ≳ R C η 3 S 1 ( R 1/ n ) N ( s ) α ∣ f τ ∣ p n . (84)
The Taylor rescaling taking each τ \tau τ to the unit cap sends each θ ⊂ τ \theta\subset\tau θ ⊂ τ to a cap of M n ( σ n R ) \mathcal{M}^{n}(\sigma^nR) M n ( σ n R ) at scale R − 1 / n / σ R^{-1/n}/\sigma R − 1/ n / σ , where 1 < σ n R ≤ R 1 − 1 / n 1<\sigma^nR\leq R^{1-1/n} 1 < σ n R ≤ R 1 − 1/ n . To bound each integral on the RHS of (84) , we apply the defining estimate for S 2 S_2 S 2 to the rescaled f τ f_\tau f τ , with parameter σ n R \sigma^nR σ n R and amplitude α / [ R C η 3 S 1 ( R 1 / n ) N ( s ) ] \alpha/[R^{C\eta^3}S_1(R^{1/n})N(s)] α / [ R C η 3 S 1 ( R 1/ n ) N ( s )] . Returning to the original coordinates as in the derivation of (5.5), and summing over ρ , τ \rho,\tau ρ , τ , bounds the RHS of (84) by
≲ η R C η 3 ( S 1 ( R 1 / n ) S 2 ( σ n R ) ) p n ∑ R − 1 / n ≤ u ≤ σ ∑ ρ ′ ∈ Θ ( u ) ∑ V ∈ G ρ ′ ( V ∥ V ρ ′ , R ∥ R C η 3 S 1 ( R 1 / n ) S 2 ( σ n R ) ) ∣ V ∣ ∥ ( ∑ θ ⊂ ρ ′ ∣ f θ ∣ 2 ) 1 / 2 ∥ L # p n / 2 ( V ) p n . \lesssim_\eta R^{C\eta^3}\bigl(S_1(R^{1/n})S_2(\sigma^nR)\bigr)^{p_n}\sum_{R^{-1/n}\leq u\leq\sigma}\sum_{\rho'\in\Theta(u)}\sum_{V\in\mathcal{G}_{\rho'}\left(\frac{V\lVert V_{\rho',R}\rVert}{R^{C\eta^3}S_1(R^{1/n})S_2(\sigma^nR)}\right)}\lvert V\rvert\left\lVert\left(\sum_{\theta\subset\rho'}\lvert f_\theta\rvert^2\right)^{1/2}\right\rVert_{L_{\#}^{p_n/2}(V)}^{p_n}. ≲ η R C η 3 ( S 1 ( R 1/ n ) S 2 ( σ n R ) ) p n R − 1/ n ≤ u ≤ σ ∑ ρ ′ ∈ Θ ( u ) ∑ V ∈ G ρ ′ ( R C η 3 S 1 ( R 1/ n ) S 2 ( σ n R ) V ∥ V ρ ′ , R ∥ ) ∑ ∣ V ∣ θ ⊂ ρ ′ ∑ ∣ f θ ∣ 2 1/2 L # p n /2 ( V ) p n .
where we used N ( s ) N τ ( u ) ≤ N ( σ ) N τ ( u ) ≲ η N ( u ) N(s)N_\tau(u)\leq N(\sigma)N_\tau(u)\lesssim_\eta N(u) N ( s ) N τ ( u ) ≤ N ( σ ) N τ ( u ) ≲ η N ( u ) by Lemma 2.18.
By monotonicity, we may replace S 2 ( σ n R ) S_2(\sigma^nR) S 2 ( σ n R ) by S 2 ( R 1 − 1 / n ) S_2(R^{1-1/n}) S 2 ( R 1 − 1/ n ) in both the coefficient and the amplitude denominator. Extending the sum over u u u to R − 1 / n ≤ u ≤ 1 R^{-1/n}\leq u\leq1 R − 1/ n ≤ u ≤ 1 gives the sums over the original functions f θ f_\theta f θ in (82) , and the iteration stops.
Case 2: s > R η / 3 σ s > R^{\eta/3}\sigma s > R η /3 σ .
Applying Lemma 7.3 to the sum over ρ \rho ρ and B B B on the RHS of (83) , with amplitude α / [ R C ϵ S 1 ( R 1 / n ) ] \alpha/[R^{C\epsilon}S_1(R^{1/n})] α / [ R C ϵ S 1 ( R 1/ n )] , gives
∥ f ∥ L p n p n ≲ R C ϵ S 1 ( R 1 / n ) p n ∑ ρ ∈ Θ ( s ) ∑ Q ∈ G ρ ( ∥ Q ∥ ∗ R C ϵ S 1 ( R 1 / n ) ) ∥ Q ∥ ( ∑ τ ′ ⊂ ρ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ∗ w par σ 0 ( τ ′ ) ) 1 / 2 ∥ ∥ L # p n / 2 ( Q ) p n + R C ϵ S 1 ( R 1 / n ) p n ∑ ρ ′ ∈ Θ ( s ′ ) ∑ Q ∈ G ρ ′ ( ∥ Q ∥ ∗ R C ϵ S 1 ( R 1 / n ) ) ∥ Q ∥ ( ∑ τ ⊂ ρ ′ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) ) 1 / 2 ∥ ∥ L # p n / 2 ( Q ) p n , (85) \begin{aligned}
\|f\|_{L^{p_n}}^{p_n}
&\lesssim R^{C\epsilon}S_1(R^{1/n})^{p_n}
\sum_{\rho\in\Theta(s)}
\sum_{Q\in\mathcal{G}_{\rho}\left(\frac{\|Q\|_*}{R^{C\epsilon}S_1(R^{1/n})}\right)}
\|Q\|
\left(\sum_{\substack{\tau'\subset\rho\\|\tau'|=\sigma'}}
|f_{\tau'}|^2*w_{\operatorname{par}_{\sigma_0}(\tau')}\right)^{1/2}
\biggl\| \biggr\|_{L_{\#}^{p_n/2}(Q)}^{p_n}\\
&\quad+
R^{C\epsilon}S_1(R^{1/n})^{p_n}
\sum_{\rho'\in\Theta(s')}
\sum_{Q\in\mathcal{G}_{\rho'}\left(\frac{\|Q\|_*}{R^{C\epsilon}S_1(R^{1/n})}\right)}
\|Q\|
\left(\sum_{\substack{\tau\subset\rho'\\|\tau|=\sigma}}
|f_{\tau}|^2*w_{\operatorname{par}_{\sigma_0}(\tau)}\right)^{1/2}
\biggl\| \biggr\|_{L_{\#}^{p_n/2}(Q)}^{p_n}, \tag*{(85)}
\end{aligned} ∥ f ∥ L p n p n ≲ R C ϵ S 1 ( R 1/ n ) p n ρ ∈ Θ ( s ) ∑ Q ∈ G ρ ( R C ϵ S 1 ( R 1/ n ) ∥ Q ∥ ∗ ) ∑ ∥ Q ∥ τ ′ ⊂ ρ ∣ τ ′ ∣ = σ ′ ∑ ∣ f τ ′ ∣ 2 ∗ w par σ 0 ( τ ′ ) 1/2 L # p n /2 ( Q ) p n + R C ϵ S 1 ( R 1/ n ) p n ρ ′ ∈ Θ ( s ′ ) ∑ Q ∈ G ρ ′ ( R C ϵ S 1 ( R 1/ n ) ∥ Q ∥ ∗ ) ∑ ∥ Q ∥ τ ⊂ ρ ′ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w par σ 0 ( τ ) 1/2 L # p n /2 ( Q ) p n , (85)
where σ ′ = max { R − 1 / n , σ 2 / s } \sigma'=\max\{R^{-1/n},\sigma^2/s\} σ ′ = max { R − 1/ n , σ 2 / s } and s ′ = s ( σ / s ) 1 / ( n − 1 ) s'=s(\sigma/s)^{1/(n-1)} s ′ = s ( σ / s ) 1/ ( n − 1 ) . We distinguish two cases according to which sum in (85) dominates. Case 2.1 treats the second sum, where s s s decreases to s ′ s' s ′ , and Case 2.2 treats the first sum, where σ \sigma σ decreases to σ ′ \sigma' σ ′ .
Case 2.1. Suppose that the second sum in (85) dominates. We apply Lemma 7.2 to its sum over ρ ′ \rho' ρ ′ and Q Q Q , at scales ( σ , s ′ ) (\sigma,s') ( σ , s ′ ) and amplitude α / [ R C ϵ S 1 ( R 1 / n ) ] \alpha/[R^{C\epsilon}S_1(R^{1/n})] α / [ R C ϵ S 1 ( R 1/ n )] .
If the first sum over V V V on the RHS of Lemma 7.2 dominates, applying Lemma 7.1 to that sum gives (82) , and the iteration stops.
If instead the second sum over B B B on the RHS of Lemma 7.2 dominates, let σ ′ ′ \sigma'' σ ′′ be its cap width. Dyadic pigeonholing reduces the sum over u u u to one scale u ∈ [ σ ′ ′ , s ′ ] u\in[\sigma'',s'] u ∈ [ σ ′′ , s ′ ] . We obtain (83) with ( σ , s ) (\sigma,s) ( σ , s ) replaced by ( σ ′ ′ , u ) (\sigma'',u) ( σ ′′ , u ) . Since σ ′ ′ ≤ σ \sigma''\leq\sigma σ ′′ ≤ σ and u ≤ s ′ ≤ s R − η / 3 / ( n − 1 ) u\leq s'\leq sR^{-\eta/3/(n-1)} u ≤ s ′ ≤ s R − η /3/ ( n − 1 ) , we return to (83) with these new parameters and repeat the algorithm.
Case 2.2. Suppose that the first sum in (85) dominates.
If σ ′ > R − 1 / n \sigma'>R^{-1/n} σ ′ > R − 1/ n , we apply Lemma 7.2 to its sum over ρ \rho ρ and Q Q Q , at scales ( σ ′ , s ) (\sigma',s) ( σ ′ , s ) and amplitude α / [ R C ϵ S 1 ( R 1 / n ) ] \alpha/[R^{C\epsilon}S_1(R^{1/n})] α / [ R C ϵ S 1 ( R 1/ n )] .
If the first sum over V V V on the RHS of Lemma 7.2 dominates, applying Lemma 7.1 to that sum gives (82) , and the iteration stops.
If instead the second sum over B B B on the RHS of Lemma 7.2 dominates, let σ ′ ′ \sigma'' σ ′′ be its cap width. Dyadic pigeonholing reduces the sum over u u u to one scale u ∈ [ σ ′ ′ , s ] u\in[\sigma'',s] u ∈ [ σ ′′ , s ] . We obtain (83) with ( σ , s ) (\sigma,s) ( σ , s ) replaced by ( σ ′ ′ , u ) (\sigma'',u) ( σ ′′ , u ) . Since σ ′ ′ ≤ σ ′ = σ 2 / s ≤ σ R − η \sigma''\leq\sigma'=\sigma^2/s\leq\sigma R^{-\eta} σ ′′ ≤ σ ′ = σ 2 / s ≤ σ R − η and u ≤ s u\leq s u ≤ s , we return to (83) with these new parameters and repeat the algorithm.
If σ ′ = R − 1 / n \sigma'=R^{-1/n} σ ′ = R − 1/ n , the first sum over Q Q Q in (85) involves the original functions f θ f_\theta f θ . The derivation of (55) from (54) for σ = R − 1 / n \sigma=R^{-1/n} σ = R − 1/ n in Section 5, together with Lemma 7.1, applies to this sum and gives (82) , and the iteration stops.
Termination criterion. At each return to (83) , neither σ \sigma σ nor s s s increases, and at least one decreases by a factor of at least R η / 3 / ( n − 1 ) R^{\eta/3/(n-1)} R η /3/ ( n − 1 ) . Since both lie in [ R − 1 / n , 1 ] [R^{-1/n},1] [ R − 1/ n , 1 ] , there are O ( η − 3 ) O(\eta^{-3}) O ( η − 3 ) steps. The factor S 2 ( σ n R ) S_2(\sigma^nR) S 2 ( σ n R ) occurs at most once, when Case 1 terminates the iteration.
The R C ϵ R^{C\epsilon} R C ϵ losses in the coefficient and the amplitude denominator accumulate to at most R C ϵ η − 3 R^{C\epsilon\eta^{-3}} R C ϵ η − 3 . Taking ϵ \epsilon ϵ sufficiently small in terms of η \eta η , we absorb these losses into R C η R^{C\eta} R C η . Combining the estimates above, we bound ∥ f ∥ L p n p n \|f\|_{L^{p_n}}^{p_n} ∥ f ∥ L p n p n by
( R C η S 1 ( R 1 / n ) S 2 ( R 1 − 1 / n ) ) p n ∑ R − 1 / n ≤ u ≤ 1 ∑ ρ ′ ∈ Θ ( u ) ∑ V ∈ G ρ ′ ( α R C η S 1 ( R 1 / n ) S 2 ( R 1 − 1 / n ) ) V ∈ G ρ ′ , R ∥ V ∥ ( ∑ θ ⊂ ρ ′ ∣ f θ ∣ 2 ) 1 / 2 ∥ ∥ L # p n / 2 ( V ) p n . \left(R^{C\eta}S_1(R^{1/n})S_2(R^{1-1/n})\right)^{p_n}
\sum_{R^{-1/n}\leq u\leq1}
\sum_{\rho'\in\Theta(u)}
\sum_{\substack{V\in\mathcal{G}_{\rho'}\left(\frac{\alpha}{R^{C\eta}S_1(R^{1/n})S_2(R^{1-1/n})}\right)\\V\in\mathcal{G}_{\rho',R}}}
\|V\|
\left(\sum_{\theta\subset\rho'}|f_\theta|^2\right)^{1/2}
\biggl\|\biggr\|_{L_{\#}^{p_n/2}(V)}^{p_n}. ( R C η S 1 ( R 1/ n ) S 2 ( R 1 − 1/ n ) ) p n R − 1/ n ≤ u ≤ 1 ∑ ρ ′ ∈ Θ ( u ) ∑ V ∈ G ρ ′ ( R C η S 1 ( R 1/ n ) S 2 ( R 1 − 1/ n ) α ) V ∈ G ρ ′ , R ∑ ∥ V ∥ θ ⊂ ρ ′ ∑ ∣ f θ ∣ 2 1/2 L # p n /2 ( V ) p n .
The same proof with R R R replaced by any 1 ≤ r ≤ R 1\leq r\leq R 1 ≤ r ≤ R , together with the monotonicity of S 1 , S 2 S_1,S_2 S 1 , S 2 , gives S 2 ( R ) ≲ R C η S 1 ( R 1 / n ) S 2 ( R 1 − 1 / n ) S_2(R)\lesssim R^{C\eta}S_1(R^{1/n})S_2(R^{1-1/n}) S 2 ( R ) ≲ R C η S 1 ( R 1/ n ) S 2 ( R 1 − 1/ n ) by the definition of S 2 S_2 S 2 . Since η \eta η can be chosen arbitrarily small, this proves Proposition 2.5. □ \square □
From curves to higher dimensional cones over moment curves: Overview of the proof
We prove the amplitude dependent wave envelope type estimate for the cone over the moment curve in R d + 1 \mathbb{R}^{d+1} R d + 1 (Theorem 1.8 ), assuming the corresponding estimates for moment curves in R k \mathbb{R}^k R k , 2 ≤ k ≤ d 2\leq k\leq d 2 ≤ k ≤ d (Theorem 1.7). Here is the statement.
Proposition 8.1. Let d ≥ 3 d \ge3 d ≥ 3 . Assume that the amplitude dependent wave envelope type estimate in Theorem 1.7 holds for the moment curve in R k \mathbb{R}^{k} R k at exponent p k p_k p k for every 2 ≤ k ≤ d 2 \le k \le d 2 ≤ k ≤ d . Then the corresponding estimate in Theorem 1.8 holds for the cone over the moment curve in R d + 1 \mathbb{R}^{d+1} R d + 1 at exponent p d p_d p d , with the averaging boxes defined in (1.4).
Recall that p k = k ( k + 1 ) / 2 + 1 p_k = k(k+1)/2+1 p k = k ( k + 1 ) /2 + 1 , ϕ d ( t ) = ( 1 , t , t 2 / 2 ! , … , t d / d ! ) \phi_d(t)=(1,t,t^2/2!,\ldots,t^d/d!) ϕ d ( t ) = ( 1 , t , t 2 /2 ! , … , t d / d !) , and
Γ d + 1 ( R ) = { ∑ i = 0 d λ i ϕ d ( i ) ( t ) : 0 ≤ t ≤ 1 , 1 ≤ λ 0 ≤ 2 , ∣ λ i ∣ ≤ R − i / d ( 1 ≤ i ≤ d ) } . \Gamma^{d+1}(R)=\left\{\sum_{i=0}^{d}\lambda_i\phi_d^{(i)}(t):0\le t\le1,\quad1\le\lambda_0\le2,\quad|\lambda_i|\le R^{-i/d}\ (1\le i\le d)\right\}. Γ d + 1 ( R ) = { i = 0 ∑ d λ i ϕ d ( i ) ( t ) : 0 ≤ t ≤ 1 , 1 ≤ λ 0 ≤ 2 , ∣ λ i ∣ ≤ R − i / d ( 1 ≤ i ≤ d ) } .
For a Schwartz function f : R d + 1 → C f:\mathbb{R}^{d+1}\to\mathbb{C} f : R d + 1 → C with supp f ^ ⊂ Γ d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\Gamma^{d+1}(R) supp f ⊂ Γ d + 1 ( R ) , we count the caps that contain at least one θ ∈ Ξ ( R − 1 / d ) \theta\in\Xi(R^{-1/d}) θ ∈ Ξ ( R − 1/ d ) with f θ ≢ 0 f_\theta\not\equiv0 f θ ≡ 0 :
N ( s ) = # { τ ∈ Ξ ( s ) : f θ ≢ 0 for some θ ⊂ τ } , N ρ ( u ) = # { τ ∈ Ξ ( u ) : τ ⊂ ρ , f θ ≢ 0 for some θ ⊂ τ } . N(s)=\#\{\tau\in\Xi(s):f_\theta\not\equiv0\text{ for some }\theta\subset\tau\},\qquad N_\rho(u)=\#\{\tau\in\Xi(u):\tau\subset\rho,\ f_\theta\not\equiv0\text{ for some }\theta\subset\tau\}. N ( s ) = # { τ ∈ Ξ ( s ) : f θ ≡ 0 for some θ ⊂ τ } , N ρ ( u ) = # { τ ∈ Ξ ( u ) : τ ⊂ ρ , f θ ≡ 0 for some θ ⊂ τ } .
Let t ρ t_\rho t ρ be the left endpoint of the interval corresponding to ρ \rho ρ . Recall from (1.3) and (1.4) that
U ρ , R = { x ∈ R d + 1 : ∣ x ⋅ ϕ d ( j ) ( t ρ ) ∣ ≤ R 2 / d s 2 − j for 0 ≤ j ≤ d } , ρ ∈ Ξ ( s ) , U_{\rho,R}=\{x\in\mathbb{R}^{d+1}:|x\cdot\phi_d^{(j)}(t_\rho)|\le R^{2/d}s^{2-j}\text{ for }0\le j\le d\},\qquad\rho\in\Xi(s), U ρ , R = { x ∈ R d + 1 : ∣ x ⋅ ϕ d ( j ) ( t ρ ) ∣ ≤ R 2/ d s 2 − j for 0 ≤ j ≤ d } , ρ ∈ Ξ ( s ) ,
V ρ , R = { x ∈ R k : ∣ x ⋅ γ k ( j ) ( t ρ ) ∣ ≤ R 2 / k s 2 − j for 1 ≤ j ≤ k } , ρ ∈ Θ ( s ) , V_{\rho,R}=\{x\in\mathbb{R}^{k}:|x\cdot\gamma_k^{(j)}(t_\rho)|\le R^{2/k}s^{2-j}\text{ for }1\le j\le k\},\qquad\rho\in\Theta(s), V ρ , R = { x ∈ R k : ∣ x ⋅ γ k ( j ) ( t ρ ) ∣ ≤ R 2/ k s 2 − j for 1 ≤ j ≤ k } , ρ ∈ Θ ( s ) ,
where ϕ d ( 0 ) = ϕ d \phi_d^{(0)}=\phi_d ϕ d ( 0 ) = ϕ d and 2 ≤ k ≤ d 2\le k\le d 2 ≤ k ≤ d .
Convention 8.2. For a parent ρ \rho ρ over an interval of length s s s , in a sum over boxes of the form
∑ B ∥ B 0 B ∈ G ρ ( α ) ∣ B ∣ ∥ F ρ ∥ L # p / 2 ( B ) p \sum_{\substack{B\parallel B_0\\B\in\mathcal{G}_\rho(\alpha)}}|B|\|F_\rho\|_{L_\#^{p/2}(B)}^p B ∥ B 0 B ∈ G ρ ( α ) ∑ ∣ B ∣∥ F ρ ∥ L # p /2 ( B ) p the selection uses the same function, exponent and tiling as the summand:
G ρ ( α ) = { B ∥ B 0 : ∥ F ρ ∥ L # p / 2 ( B ) ≥ α N ( s ) } . \mathcal{G}_\rho(\alpha)=\left\{B\parallel B_0:\|F_\rho\|_{L_\#^{p/2}(B)}\ge\frac{\alpha}{N(s)}\right\}. G ρ ( α ) = { B ∥ B 0 : ∥ F ρ ∥ L # p /2 ( B ) ≥ N ( s ) α } .
We now write the hypothesis and conclusion of Proposition 8.1 in this notation. For curves in R k \mathbb{R}^{k} R k , 2 ≤ k ≤ d 2\le k\le d 2 ≤ k ≤ d , recall that γ k ( t ) = ( t , t 2 , … , t k ) \gamma_k(t)=(t,t^2,\ldots,t^k) γ k ( t ) = ( t , t 2 , … , t k ) and
M k ( R ) = { γ k ( t ) + ∑ i = 2 k a i i ! γ k ( i ) ( t ) : 0 ≤ t ≤ 1 , ∣ a i ∣ ≤ R − i / k } . \mathcal{M}^{k}(R)=\left\{\gamma_k(t)+\sum_{i=2}^{k}\frac{a_i}{i!}\gamma_k^{(i)}(t):0\le t\le1,\quad|a_i|\le R^{-i/k}\right\}. M k ( R ) = { γ k ( t ) + i = 2 ∑ k i ! a i γ k ( i ) ( t ) : 0 ≤ t ≤ 1 , ∣ a i ∣ ≤ R − i / k } .
For each 2 ≤ k ≤ d 2\le k\le d 2 ≤ k ≤ d , we apply the amplitude dependent wave envelope type estimate for the moment curve in R k \mathbb{R}^{k} R k , assumed in Proposition 8.1. For every Schwartz function g : R k → C g:\mathbb{R}^{k}\to\mathbb{C} g : R k → C with supp g ^ ⊂ M k ( R ) \operatorname{supp}\widehat{g}\subset\mathcal{M}^{k}(R) supp g ⊂ M k ( R ) and every α > 0 \alpha>0 α > 0 , this gives
∫ ∣ g ∣ > α ∣ g ∣ p k ≲ ϵ R ϵ ∑ R − 1 / k ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , R V ∈ G ρ ( R ϵ ) ∣ V ∣ ( ∑ θ ⊂ ρ ∣ g θ ∣ 2 ) p k / 2 ∥ g ∥ L # p k / 2 ( V ) p k . (86) \int_{|g|>\alpha}|g|^{p_k}\lesssim_\epsilon R^\epsilon\sum_{R^{-1/k}\le s\le1}\sum_{\rho\in\Theta(s)}\sum_{\substack{V\parallel V_{\rho,R}\\V\in\mathcal{G}_\rho(R^\epsilon)}}|V|\left(\sum_{\theta\subset\rho}|g_\theta|^2\right)^{p_k/2}\|g\|_{L_\#^{p_k/2}(V)}^{p_k}. \tag*{(86)} ∫ ∣ g ∣ > α ∣ g ∣ p k ≲ ϵ R ϵ R − 1/ k ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∥ V ρ , R V ∈ G ρ ( R ϵ ) ∑ ∣ V ∣ θ ⊂ ρ ∑ ∣ g θ ∣ 2 p k /2 ∥ g ∥ L # p k /2 ( V ) p k . (86)
The conclusion is that every Schwartz function f : R d + 1 → C f:\mathbb{R}^{d+1}\to\mathbb{C} f : R d + 1 → C with supp f ^ ⊂ Γ d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\Gamma^{d+1}(R) supp f ⊂ Γ d + 1 ( R ) satisfies
∫ ∣ f ∣ > α ∣ f ∣ p d ≲ ϵ R ϵ ∑ R − 1 / d ≤ s ≤ 1 ∑ ρ ∈ Ξ ( s ) ∑ U ∥ U ρ , R U ∈ G ρ ( R ϵ ) ∣ U ∣ ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ) p d / 2 ∥ f ∥ L # p d / 2 ( U ) p d . \int_{|f|>\alpha}|f|^{p_d}\lesssim_\epsilon R^\epsilon\sum_{R^{-1/d}\le s\le1}\sum_{\rho\in\Xi(s)}\sum_{\substack{U\parallel U_{\rho,R}\\U\in\mathcal{G}_\rho(R^\epsilon)}}|U|\left(\sum_{\theta\subset\rho}|f_\theta|^2\right)^{p_d/2}\|f\|_{L_\#^{p_d/2}(U)}^{p_d}. ∫ ∣ f ∣ > α ∣ f ∣ p d ≲ ϵ R ϵ R − 1/ d ≤ s ≤ 1 ∑ ρ ∈ Ξ ( s ) ∑ U ∥ U ρ , R U ∈ G ρ ( R ϵ ) ∑ ∣ U ∣ θ ⊂ ρ ∑ ∣ f θ ∣ 2 p d /2 ∥ f ∥ L # p d /2 ( U ) p d .
Throughout Sections 8–10, we assume the hypothesis of Proposition 8.1 on moment curves. The high frequency decomposition produces Fourier supports that need not satisfy the condition on the cone over the moment curve ∣ λ 0 ∣ ∼ 1 |\lambda_0|\sim1 ∣ λ 0 ∣ ∼ 1 , even after rescaling. Following [10 ] , we therefore introduce the following neighborhoods of Taylor cones. For R ≥ 1 R\ge1 R ≥ 1 and 0 ≤ m ≤ d − 2 0\le m\le d-2 0 ≤ m ≤ d − 2 , define
Γ m d + 1 ( R ) = { ∑ i = 0 d λ i ϕ d ( i ) ( t ) : t ∈ [ 0 , 1 ] , ∣ λ i ∣ ≤ R − i / d ( 0 ≤ i ≤ d ) , max 0 ≤ i ≤ m 2 m + 1 − i R i / d ∣ λ i ∣ ≥ 1 } = ⋃ θ ∈ Ξ m d + 1 ( R ) θ . (87) \Gamma_m^{d+1}(R)=\left\{\sum_{i=0}^{d}\lambda_i\phi_d^{(i)}(t):\ t\in[0,1],\ |\lambda_i|\le R^{-i/d}\ (0\le i\le d),\ \max_{0\le i\le m}2^{m+1-i}R^{i/d}|\lambda_i|\ge1\right\}=\bigcup_{\theta\in\Xi_m^{d+1}(R)}\theta. \tag*{(87)} Γ m d + 1 ( R ) = { i = 0 ∑ d λ i ϕ d ( i ) ( t ) : t ∈ [ 0 , 1 ] , ∣ λ i ∣ ≤ R − i / d ( 0 ≤ i ≤ d ) , 0 ≤ i ≤ m max 2 m + 1 − i R i / d ∣ λ i ∣ ≥ 1 } = θ ∈ Ξ m d + 1 ( R ) ⋃ θ . (87)
where Ξ m d + 1 ( R ) \Xi^{d+1}_{m}(R) Ξ m d + 1 ( R ) is the family of caps indexed by the standard partition of [ 0 , 1 ] [0,1] [ 0 , 1 ] into intervals I I I of length R − 1 / d R^{-1/d} R − 1/ d , with
θ I = { ∑ i = 0 d λ i ϕ d ( i ) ( t ) : t ∈ I , ∣ λ i ∣ ≤ R − i / d ( 0 ≤ i ≤ d ) , max 0 ≤ i ≤ m 2 m + 1 − i R i / d ∣ λ i ∣ ≥ 1 } . (88) \theta_I=\left\{\sum_{i=0}^{d}\lambda_i\phi_d^{(i)}(t):\ t\in I,\ |\lambda_i|\le R^{-i/d}\ (0\le i\le d),\ \max_{0\le i\le m}2^{m+1-i}R^{i/d}|\lambda_i|\ge1\right\}. \tag*{(88)} θ I = { i = 0 ∑ d λ i ϕ d ( i ) ( t ) : t ∈ I , ∣ λ i ∣ ≤ R − i / d ( 0 ≤ i ≤ d ) , 0 ≤ i ≤ m max 2 m + 1 − i R i / d ∣ λ i ∣ ≥ 1 } . (88)
The maximum in (87) requires at least one of λ 0 , R 1 / d λ 1 , … , R m / d λ m \lambda_0,R^{1/d}\lambda_1,\ldots,R^{m/d}\lambda_m λ 0 , R 1/ d λ 1 , … , R m / d λ m to stay away from zero. Section 8.3 explains why the high frequency decomposition raises this order. For f f f with Fourier support in Γ m d + 1 ( R ) \Gamma^{d+1}_{m}(R) Γ m d + 1 ( R ) , we use a smooth partition of unity adapted to these caps to decompose f f f into pieces f θ I f_{\theta_I} f θ I , each with Fourier support in a fixed enlargement of θ I \theta_I θ I .
For dyadic 0 < s ≤ 1 0<s\le1 0 < s ≤ 1 , let J ( s ) \mathcal{J}(s) J ( s ) denote the standard partition of [ 0 , 1 ] [0,1] [ 0 , 1 ] into intervals of length s s s . We identify each interval in J ( s ) \mathcal{J}(s) J ( s ) with the cap over it and use J ( s ) \mathcal{J}(s) J ( s ) for both families.
Fix ϵ > 0 \epsilon>0 ϵ > 0 . We choose A = A ( ϵ ) A=A(\epsilon) A = A ( ϵ ) large, then K = K ( A , ϵ ) K=K(A,\epsilon) K = K ( A , ϵ ) large, and take R R R sufficiently large. To apply the hypothesis on moment curves (86) , we first restrict the lower Taylor coefficients to short intervals. For 0 ≤ j ≤ m 0\le j\le m 0 ≤ j ≤ m and I ∈ J ( R − 1 / d ) I\in\mathcal{J}(R^{-1/d}) I ∈ J ( R − 1/ d ) , let t I t_I t I be the left endpoint of I I I . Define
Γ j , 1 / K d + 1 ( R ) : = ⋃ I ∈ J ( R − 1 / d ) { ∑ i = 0 d λ i ϕ d ( i ) ( t ) : ∣ λ i ∣ ≤ K − 1 R − j / d , 0 ≤ i < j , ∣ λ j − R − j / d ∣ ≤ K − 1 R − j / d , ∣ λ i ∣ ≲ R − i / d , j < i ≤ d } . (89) \Gamma^{d+1}_{j,1/K}(R):=\bigcup_{I\in\mathcal{J}(R^{-1/d})}\left\{\sum_{i=0}^{d}\lambda_i\phi_d^{(i)}(t):\ |\lambda_i|\le K^{-1}R^{-j/d},\ 0\le i<j,\ |\lambda_j-R^{-j/d}|\le K^{-1}R^{-j/d},\ |\lambda_i|\lesssim R^{-i/d},\ j<i\le d\right\}. \tag*{(89)} Γ j , 1/ K d + 1 ( R ) := I ∈ J ( R − 1/ d ) ⋃ { i = 0 ∑ d λ i ϕ d ( i ) ( t ) : ∣ λ i ∣ ≤ K − 1 R − j / d , 0 ≤ i < j , ∣ λ j − R − j / d ∣ ≤ K − 1 R − j / d , ∣ λ i ∣ ≲ R − i / d , j < i ≤ d } . (89)
Geometrically, Γ j , 1 / K d + 1 ( R ) \Gamma^{d+1}_{j,1/K}(R) Γ j , 1/ K d + 1 ( R ) is obtained by thickening the curve R − j / d ϕ d ( j ) ( t ) R^{-j/d}\phi_d^{(j)}(t) R − j / d ϕ d ( j ) ( t ) , 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 , by K − 1 R − j / d K^{-1}R^{-j/d} K − 1 R − j / d in the directions ϕ d ( t ) , … , ϕ d ( j ) ( t ) \phi_d(t),\ldots,\phi_d^{(j)}(t) ϕ d ( t ) , … , ϕ d ( j ) ( t ) , and by R − i / d R^{-i/d} R − i / d in the directions ϕ d ( i ) ( t ) \phi_d^{(i)}(t) ϕ d ( i ) ( t ) , j < i ≤ d j<i\le d j < i ≤ d .
In particular, Γ 0 , 1 / K d + 1 ( R ) \Gamma^{d+1}_{0,1/K}(R) Γ 0 , 1/ K d + 1 ( R ) can be viewed as the anisotropic neighborhood of the moment curve ( 1 , t , t 2 / 2 ! , … , t d / d ! ) (1,t,t^2/2!,\ldots,t^d/d!) ( 1 , t , t 2 /2 ! , … , t d / d !) , 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 , thickened by K − 1 K^{-1} K − 1 in the ϕ d ( t ) \phi_d(t) ϕ d ( t ) direction. Note that, up to fixed enlargement, Γ j , 1 / K d + 1 ( R ) ⊂ Γ m d + 1 ( R ) \Gamma^{d+1}_{j,1/K}(R)\subset\Gamma^{d+1}_{m}(R) Γ j , 1/ K d + 1 ( R ) ⊂ Γ m d + 1 ( R ) , 0 ≤ j ≤ m 0\le j\le m 0 ≤ j ≤ m . As before, we write f = ∑ I f θ I f=\sum_I f_{\theta_I} f = ∑ I f θ I , allowing any decomposition with the cap supports specified above. When supp f ^ ⊂ Γ j , 1 / K d + 1 ( R ) \operatorname{supp}\widehat f\subset\Gamma^{d+1}_{j,1/K}(R) supp f ⊂ Γ j , 1/ K d + 1 ( R ) , we further require each f θ I f_{\theta_I} f θ I to be supported in a fixed enlargement of the I I I th set in (8.4). We now introduce the constants needed to prove Proposition 8.1 by descending induction on m m m .
Definition 8.3. Let C m ( R ) C^m(R) C m ( R ) be the least constant such that, for every Schwartz function f f f with supp f ^ ⊂ Γ m d + 1 ( R ) \operatorname{supp}\widehat f\subset\Gamma^{d+1}_{m}(R) supp f ⊂ Γ m d + 1 ( R ) and every α > 0 \alpha>0 α > 0 ,
∫ ∣ f ∣ > α ∣ f ∣ p d − m ≤ C m ( R ) ∑ R − 1 / d ≤ s ≤ 1 ∑ ρ ∈ J ( s ) ∑ U ∈ G ρ ( α C m ( R ) ) ∣ U ∣ ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ) p d − m / 2 ∥ ⋅ ∥ L p d − m / 2 ( U ) . (90) \int_{|f|>\alpha}|f|^{p_{d-m}}\le C^m(R)\sum_{R^{-1/d}\le s\le1}\sum_{\rho\in\mathcal{J}(s)}\sum_{U\in\mathcal{G}_{\rho}\left(\frac{\alpha}{C^m(R)}\right)}|U|\left(\sum_{\theta\subset\rho}|f_\theta|^2\right)^{p_{d-m}/2}\|\,\cdot\,\|_{L^{p_{d-m}/2}(U)}. \tag*{(90)} ∫ ∣ f ∣ > α ∣ f ∣ p d − m ≤ C m ( R ) R − 1/ d ≤ s ≤ 1 ∑ ρ ∈ J ( s ) ∑ U ∈ G ρ ( C m ( R ) α ) ∑ ∣ U ∣ θ ⊂ ρ ∑ ∣ f θ ∣ 2 p d − m /2 ∥ ⋅ ∥ L p d − m /2 ( U ) . (90)
Definition 8.4. Let C K m ( R ) C^m_K(R) C K m ( R ) be the least constant such that, for every 0 ≤ j ≤ m 0\le j\le m 0 ≤ j ≤ m , every Schwartz function f f f with supp f ^ ⊂ Γ j , 1 / K d + 1 ( R ) \operatorname{supp}\widehat f\subset\Gamma^{d+1}_{j,1/K}(R) supp f ⊂ Γ j , 1/ K d + 1 ( R ) , and every α > 0 \alpha>0 α > 0 ,
∫ ∣ f ∣ > α ∣ f ∣ p d − m ≤ C K m ( R ) ∑ R − 1 / d ≤ s ≤ 1 ∑ ρ ∈ J ( s ) ∑ U ∈ G ρ ( α C K m ( R ) ) ∣ U ∣ ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ) p d − m / 2 ∥ ⋅ ∥ L p d − m / 2 ( U ) . (91) \int_{|f|>\alpha}|f|^{p_{d-m}}\le C^m_K(R)\sum_{R^{-1/d}\le s\le1}\sum_{\rho\in\mathcal{J}(s)}\sum_{U\in\mathcal{G}_{\rho}\left(\frac{\alpha}{C^m_K(R)}\right)}|U|\left(\sum_{\theta\subset\rho}|f_\theta|^2\right)^{p_{d-m}/2}\|\,\cdot\,\|_{L^{p_{d-m}/2}(U)}. \tag*{(91)} ∫ ∣ f ∣ > α ∣ f ∣ p d − m ≤ C K m ( R ) R − 1/ d ≤ s ≤ 1 ∑ ρ ∈ J ( s ) ∑ U ∈ G ρ ( C K m ( R ) α ) ∑ ∣ U ∣ θ ⊂ ρ ∑ ∣ f θ ∣ 2 p d − m /2 ∥ ⋅ ∥ L p d − m /2 ( U ) . (91)
Our goal is C 0 ( R ) ≲ 1 C^0(R)\lesssim1 C 0 ( R ) ≲ 1 . We establish the bounds C K m ( R ) ≲ 1 C^m_K(R)\lesssim1 C K m ( R ) ≲ 1 and C m ( R ) ≲ 1 C^m(R)\lesssim1 C m ( R ) ≲ 1 successively in the order
C K d − 2 ( R ) ⟶ C d − 2 ( R ) ⟶ C K d − 3 ( R ) ⟶ C d − 3 ( R ) ⟶ ⋯ ⟶ C K 0 ( R ) ⟶ C 0 ( R ) . C^{d-2}_K(R)\longrightarrow C^{d-2}(R)\longrightarrow C^{d-3}_K(R)\longrightarrow C^{d-3}(R)\longrightarrow\cdots\longrightarrow C^0_K(R)\longrightarrow C^0(R). C K d − 2 ( R ) ⟶ C d − 2 ( R ) ⟶ C K d − 3 ( R ) ⟶ C d − 3 ( R ) ⟶ ⋯ ⟶ C K 0 ( R ) ⟶ C 0 ( R ) .
We proceed in decreasing m m m because the argument for the high frequency part uses estimates at exponent p d − m − 1 p_{d-m-1} p d − m − 1 to obtain estimates at exponent p d − m p_{d-m} p d − m . For technical reasons, applying the induction hypothesis to the high frequency terms requires keeping the convolutions between successive powers of the square functions. We now formulate these estimates.
For m < d − 2 m<d-2 m < d − 2 , consider angular widths R − 1 / d ≤ q 1 ≤ ⋯ ≤ q d − m − 2 ≤ u ≤ 1 R^{-1/d}\le q_1\le\cdots\le q_{d-m-2}\le u\le1 R − 1/ d ≤ q 1 ≤ ⋯ ≤ q d − m − 2 ≤ u ≤ 1 and caps θ ⊂ τ 1 ⊂ ⋯ ⊂ τ d − m − 2 ⊂ ρ \theta\subset\tau_1\subset\cdots\subset\tau_{d-m-2}\subset\rho θ ⊂ τ 1 ⊂ ⋯ ⊂ τ d − m − 2 ⊂ ρ , where ∣ τ l ∣ = q l |\tau_l|=q_l ∣ τ l ∣ = q l and ∣ ρ ∣ = u |\rho|=u ∣ ρ ∣ = u . Let B τ l B_{\tau_l} B τ l consist of all x ∈ R d + 1 x\in\mathbb{R}^{d+1} x ∈ R d + 1 satisfying
∣ x ⋅ ϕ d ( i ) ( t τ l ) ∣ ≤ R ( m + 2 ) / d q 1 ⋯ q l − 1 { q l m + l + 1 − i , 0 ≤ i ≤ m + l + 1 , u m + l + 1 − i , m + l + 1 < i ≤ d , 1 ≤ l ≤ d − m − 2. (92) |x\cdot\phi_d^{(i)}(t_{\tau_l})|\le\frac{R^{(m+2)/d}}{q_1\cdots q_{l-1}}\begin{cases}q_l^{m+l+1-i},&0\le i\le m+l+1,\\u^{m+l+1-i},&m+l+1<i\le d,\end{cases}\qquad1\le l\le d-m-2. \tag*{(92)} ∣ x ⋅ ϕ d ( i ) ( t τ l ) ∣ ≤ q 1 ⋯ q l − 1 R ( m + 2 ) / d { q l m + l + 1 − i , u m + l + 1 − i , 0 ≤ i ≤ m + l + 1 , m + l + 1 < i ≤ d , 1 ≤ l ≤ d − m − 2. (92)
The boxes in (8.7) are chosen so that (i) their form is preserved under Taylor rescaling, and (ii) the high frequency terms produced by the cutoffs in § 9.3 can be bounded by convolutions with w B τ l w_{B_{\tau_l}} w B τ l , using the locally constant property and Hölder. The role of this choice will become explicit in § 10 ; see (128) and (147) . For notational convenience, we suppress the dependence of B τ l B_{\tau_l} B τ l on q 1 , … , q l − 1 q_1,\ldots,q_{l-1} q 1 , … , q l − 1 and on the width u u u of ρ \rho ρ .
In proving C m ( R ) ≲ 1 C^m(R) \lesssim1 C m ( R ) ≲ 1 , we will prove the following stronger estimate that for every Schwartz function f = ∑ θ ∈ Γ m + 1 d + 1 ( R ) f θ f=\sum_{\theta\in\Gamma^{d+1}_{m+1}(R)}f_\theta f = ∑ θ ∈ Γ m + 1 d + 1 ( R ) f θ with supp f ^ ⊂ Γ m + 1 d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\Gamma^{d+1}_{m+1}(R) supp f ⊂ Γ m + 1 d + 1 ( R ) ,
∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ ϵ R ϵ ∑ R − 1 / d ≤ q 1 ≤ ⋯ ≤ q d − m − 2 ≤ u ≤ 1 ∑ ρ ∈ J ( u ) ∑ B ∥ u d − m − 2 q 1 ⋯ q d − m − 2 ( u R 1 / d ) m U ρ , R ∣ B ∣ B ∈ G ρ ( α / R ϵ ) [ ∫ B ∑ τ d − m − 2 ⊂ ⋯ ⊂ τ 1 ⊂ ρ ( ( ⋯ ( ∑ τ 1 ⊂ τ 2 ( ∑ θ ⊂ τ 1 ∣ f θ ∣ 2 ) ∗ w B τ 1 ) p d − m / p d − m − 1 ⋯ ) ∗ w B τ d − m − 2 ) 7 / 4 ] 2 . (93) \int_{|f|>\alpha}|f|^{p_{d-m}}\lesssim_\epsilon R^\epsilon
\sum_{R^{-1/d}\leq q_1\leq\cdots\leq q_{d-m-2}\leq u\leq1}
\sum_{\rho\in\mathcal{J}(u)}
\sum_{B\parallel\frac{u^{d-m-2}}{q_1\cdots q_{d-m-2}}(uR^{1/d})^mU_{\rho,R}}
\frac{|B|}{B\in\mathcal{G}_\rho(\alpha/R^\epsilon)}
\left[
\int_B\sum_{\tau_{d-m-2}\subset\cdots\subset\tau_1\subset\rho}
\left(\left(\cdots\left(\sum_{\tau_1\subset\tau_2}\left(\sum_{\theta\subset\tau_1}|f_\theta|^2\right)*w_{B_{\tau_1}}\right)^{p_{d-m}/p_{d-m-1}}\cdots\right)*w_{B_{\tau_{d-m-2}}}\right)^{7/4}
\right]^2. \tag*{(93)} ∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ ϵ R ϵ R − 1/ d ≤ q 1 ≤ ⋯ ≤ q d − m − 2 ≤ u ≤ 1 ∑ ρ ∈ J ( u ) ∑ B ∥ q 1 ⋯ q d − m − 2 u d − m − 2 ( u R 1/ d ) m U ρ , R ∑ B ∈ G ρ ( α / R ϵ ) ∣ B ∣ ∫ B τ d − m − 2 ⊂ ⋯ ⊂ τ 1 ⊂ ρ ∑ ⋯ ( τ 1 ⊂ τ 2 ∑ ( θ ⊂ τ 1 ∑ ∣ f θ ∣ 2 ) ∗ w B τ 1 ) p d − m / p d − m − 1 ⋯ ∗ w B τ d − m − 2 7/4 2 . (93)
where B ∈ G ρ ( α / R ϵ ) B\in\mathcal{G}_\rho(\alpha/R^\epsilon) B ∈ G ρ ( α / R ϵ ) means that the weighted average over B B B inside the outermost square brackets is at least [ α / ( R ϵ N ( u ) ) ] p d − m / 2 [\alpha/(R^\epsilon N(u))]^{p_{d-m}/2} [ α / ( R ϵ N ( u )) ] p d − m /2 . The boxes B ∥ u d − m − 2 q 1 ⋯ q d − m − 2 ( u R 1 / d ) m U ρ , R B\parallel\frac{u^{d-m-2}}{q_1\cdots q_{d-m-2}}(uR^{1/d})^mU_{\rho,R} B ∥ q 1 ⋯ q d − m − 2 u d − m − 2 ( u R 1/ d ) m U ρ , R arise by returning the averaging boxes obtained from successive applications of the induction hypothesis to the high frequency terms to the original coordinates. For example, (8.8) with m = d − 4 m=d-4 m = d − 4 states that every Schwartz function f f f with supp f ^ ⊂ Γ d − 4 d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\Gamma^{d+1}_{d-4}(R) supp f ⊂ Γ d − 4 d + 1 ( R ) satisfies
∫ ∣ f ∣ > α ∣ f ∣ 11 ≲ ϵ R ϵ ∑ R − 1 / d ≤ q 1 ≤ q 2 ≤ u ≤ 1 ∑ ρ ∈ J ( u ) ∑ B ∥ u 2 q 1 q 2 ( u R 1 / d ) d − 4 U ρ , R ∣ B ∣ B ∈ G ρ ( α / R ϵ ) [ ∫ B ∑ τ 2 ⊂ ρ ∣ τ 2 ∣ = q 2 ( ∑ τ 1 ⊂ τ 2 ∣ τ 1 ∣ = q 1 ( ∑ θ ⊂ τ 1 ∣ f θ ∣ 2 ) ∗ w B τ 1 ) 11 / 7 ∗ w B τ 2 ] 7 / 4 . \int_{|f|>\alpha}|f|^{11}\lesssim_\epsilon R^\epsilon
\sum_{R^{-1/d}\leq q_1\leq q_2\leq u\leq1}
\sum_{\rho\in\mathcal{J}(u)}
\sum_{B\parallel\frac{u^2}{q_1q_2}(uR^{1/d})^{d-4}U_{\rho,R}}
\frac{|B|}{B\in\mathcal{G}_\rho(\alpha/R^\epsilon)}
\left[
\int_B\sum_{\tau_2\subset\rho\atop|\tau_2|=q_2}
\left(\sum_{\tau_1\subset\tau_2\atop|\tau_1|=q_1}
\left(\sum_{\theta\subset\tau_1}|f_\theta|^2\right)*w_{B_{\tau_1}}\right)^{11/7}
*w_{B_{\tau_2}}
\right]^{7/4}
. ∫ ∣ f ∣ > α ∣ f ∣ 11 ≲ ϵ R ϵ R − 1/ d ≤ q 1 ≤ q 2 ≤ u ≤ 1 ∑ ρ ∈ J ( u ) ∑ B ∥ q 1 q 2 u 2 ( u R 1/ d ) d − 4 U ρ , R ∑ B ∈ G ρ ( α / R ϵ ) ∣ B ∣ ∫ B ∣ τ 2 ∣ = q 2 τ 2 ⊂ ρ ∑ ∣ τ 1 ∣ = q 1 τ 1 ⊂ τ 2 ∑ ( θ ⊂ τ 1 ∑ ∣ f θ ∣ 2 ) ∗ w B τ 1 11/7 ∗ w B τ 2 7/4 .
In proving C K m ( R ) ≲ 1 C^m_K(R)\lesssim1 C K m ( R ) ≲ 1 , we will also prove the corresponding stronger estimate that every Schwartz function f f f with supp f ^ ⊂ Γ j , 1 / K d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\Gamma^{d+1}_{j,1/K}(R) supp f ⊂ Γ j , 1/ K d + 1 ( R ) ,
∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ ϵ RHS of (8.8) , (94) \int_{|f|>\alpha}|f|^{p_{d-m}}\lesssim_\epsilon\text{RHS of (8.8)}, \tag*{(94)} ∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ ϵ RHS of (8.8) , (94)
uniformly in 0 ≤ j ≤ m 0\leq j\leq m 0 ≤ j ≤ m . Indeed, the following lemma shows that (8.8) and (8.9) imply the estimates (8.5) and (8.6), respectively, with C m ( R ) ≲ 1 C^m(R)\lesssim1 C m ( R ) ≲ 1 and C K m ( R ) ≲ 1 C^m_K(R)\lesssim1 C K m ( R ) ≲ 1 .
Lemma 8.5. Let 0 ≤ m < d − 2 0\leq m<d-2 0 ≤ m < d − 2 and R ≥ 1 R\geq1 R ≥ 1 , and let f f f be a Schwartz function with supp f ^ ⊂ Γ m d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\Gamma^{d+1}_m(R) supp f ⊂ Γ m d + 1 ( R ) . For every ϵ > 0 \epsilon>0 ϵ > 0 and α > 0 \alpha>0 α > 0 ,
RHS of (8.8) ≲ ∑ R − 1 / d ≤ u ≤ 1 ∑ ρ ∈ J ( u ) ∑ U ∈ G ρ ( α / R ϵ ) ∥ U ∥ ( u R 1 / d ) m U ρ , R ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ) 1 / 2 ∥ ∑ θ ⊂ ρ ∣ f θ ∣ 2 ∥ L # p d − m / 2 ( U ) p d − m . \operatorname{RHS\ of\ (8.8)}
\lesssim
\sum_{R^{-1/d}\leq u\leq1}
\sum_{\rho\in\mathcal{J}(u)}
\sum_{U\in\mathcal{G}_\rho(\alpha/R^\epsilon)}
\|U\|_{(uR^{1/d})^mU_{\rho,R}}
\left(\sum_{\theta\subset\rho}|f_\theta|^2\right)^{1/2}
\biggl\|\sum_{\theta\subset\rho}|f_\theta|^2\biggr\|_{L^{p_{d-m}/2}_\#(U)}^{p_{d-m}}. RHS of (8.8) ≲ R − 1/ d ≤ u ≤ 1 ∑ ρ ∈ J ( u ) ∑ U ∈ G ρ ( α / R ϵ ) ∑ ∥ U ∥ ( u R 1/ d ) m U ρ , R θ ⊂ ρ ∑ ∣ f θ ∣ 2 1/2 θ ⊂ ρ ∑ ∣ f θ ∣ 2 L # p d − m /2 ( U ) p d − m .
We defer the proof to the end of this section. The induction is organized by the following propositions.
Proposition 8.6 (Base case). We have C K d − 2 ( R ) ≲ 1 C^{d-2}_K(R)\lesssim1 C K d − 2 ( R ) ≲ 1 and C d − 2 ( R ) ≲ 1 C^{d-2}(R)\lesssim1 C d − 2 ( R ) ≲ 1 .
Proposition 8.7. Let 0 ≤ m ≤ d − 3 0\leq m\leq d-3 0 ≤ m ≤ d − 3 . Assume C m + 1 ( R ) ≲ 1 C^{m+1}(R)\lesssim1 C m + 1 ( R ) ≲ 1 and that (8.8) holds with m + 1 m+1 m + 1 in place of m m m . Then C m ( R ) ≲ 1 C^m(R)\lesssim1 C m ( R ) ≲ 1 , and (8.8) holds for this m m m . When m = d − 3 m=d-3 m = d − 3 , the assumption on (8.8) is omitted.
We prove Proposition 8.7 in two steps.
Step 1. Under the hypotheses of Proposition 8.7, we establish C K m ( R ) ≲ 1 C^m_K(R)\lesssim1 C K m ( R ) ≲ 1 together with (8.9) for this $m.
Step 2. Keeping the hypotheses of Proposition 8.7, we use C K m ( R ) ≲ 1 C^m_K(R)\lesssim1 C K m ( R ) ≲ 1 and (8.9) from Step 1 to prove C m ( R ) ≲ 1 C^m(R)\lesssim1 C m ( R ) ≲ 1 and (8.8) for this $m.
We carry out Steps 1 and 2 in Sections 9 and 10, respectively. Section 11 proves Proposition 8.6, the base case m = d − 2 m=d-2 m = d − 2 of the induction on m m m .
Roadmap: Proof of Proposition 8.7, Step 1
To make the induction easier to follow, we first outline the proof of (8.6) with C K m ( R ) ≲ 1 C_K^m(R) \lesssim1 C K m ( R ) ≲ 1 , rather than the stronger estimate involving the convolutions in (8.8). The argument for the stronger estimate (8.9) is discussed in the remark at the end of Section 8.2.
We first use the hypothesis on moment curves to bound the integral of ∣ f ∣ p d − m |f|^{p_{d-m}} ∣ f ∣ p d − m over { ∣ f ∣ > α } \{|f|>\alpha\} { ∣ f ∣ > α } in terms of the pieces f τ f_\tau f τ , ∣ τ ∣ = A − 1 |\tau|=A^{-1} ∣ τ ∣ = A − 1 .
By Lemma 2.18, we split the active caps into J ≲ κ R κ J\lesssim_\kappa R^\kappa J ≲ κ R κ regular classes and apply Lemma 8.8 to each class at amplitude α / J \alpha/J α / J . The loss from summing over the classes is absorbed into R ϵ R^\epsilon R ϵ . In Proposition 8.9 and Corollary 9.3, all counts refer to a fixed class, so N ( s ) N ρ ( v ) ≲ κ N ( v ) N(s)N_\rho(v)\lesssim_\kappa N(v) N ( s ) N ρ ( v ) ≲ κ N ( v ) for v ≤ s = ∣ ρ ∣ v\leq s=|\rho| v ≤ s = ∣ ρ ∣ . For Proposition 8.15, we apply Lemma 2.18 to the family { f θ j , ν , ℓ } θ \{f_\theta^{j,\nu,\ell}\}_\theta { f θ j , ν , ℓ } θ after fixing j , ν , ℓ j,\nu,\ell j , ν , ℓ in Section 8.4.
Lemma 8.8. Let f f f be a Schwartz function with supp f ^ ⊂ Γ j , 1 / K d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\Gamma_{j,1/K}^{d+1}(R) supp f ⊂ Γ j , 1/ K d + 1 ( R ) , 0 ≤ j ≤ m 0\leq j\leq m 0 ≤ j ≤ m . Then
∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ ϵ A ϵ ∑ A − 1 ≤ s ≤ 1 ∑ ρ ∈ J ( s ) ∑ Q ∈ G ρ ( A ϵ ) Q ∈ G ρ ( α ) ∣ Q ∣ ∥ ( ∑ τ ∈ J ( A − 1 ) τ ⊂ ρ ∣ f τ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( Q ) p d − m \int_{|f|>\alpha}|f|^{p_{d-m}}\lesssim_\epsilon A^\epsilon\sum_{A^{-1}\leq s\leq1}\sum_{\rho\in\mathcal{J}(s)}\sum_{\substack{Q\in\mathcal{G}_\rho(A^\epsilon)\\ Q\in\mathcal{G}_\rho(\alpha)}}|Q|\left\|\left(\sum_{\substack{\tau\in\mathcal{J}(A^{-1})\\ \tau\subset\rho}}|f_\tau|^2\right)^{1/2}\right\|_{L_\#^{p_{d-m}/2}(Q)}^{p_{d-m}} ∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ ϵ A ϵ A − 1 ≤ s ≤ 1 ∑ ρ ∈ J ( s ) ∑ Q ∈ G ρ ( A ϵ ) Q ∈ G ρ ( α ) ∑ ∣ Q ∣ τ ∈ J ( A − 1 ) τ ⊂ ρ ∑ ∣ f τ ∣ 2 1/2 L # p d − m /2 ( Q ) p d − m
Proof. Frequency dilation by R j / d R^{j/d} R j / d sends the curve R − j / d ϕ d ( j ) ( t ) R^{-j/d}\phi_d^{(j)}(t) R − j / d ϕ d ( j ) ( t ) in (8.4) to ϕ d ( j ) ( t ) \phi_d^{(j)}(t) ϕ d ( j ) ( t ) . Projection onto coordinates j + 1 , … , j + d − m j+1,\ldots,j+d-m j + 1 , … , j + d − m , followed by a linear change, sends ϕ d ( j ) ( t ) \phi_d^{(j)}(t) ϕ d ( j ) ( t ) to γ d − m ( t ) = ( t , t 2 , … , t d − m ) \gamma_{d-m}(t)=(t,t^2,\ldots,t^{d-m}) γ d − m ( t ) = ( t , t 2 , … , t d − m ) . Taking K K K and then R R R sufficiently large, the Fourier supports of f τ ( R j / d ⋅ ) f_\tau(R^{j/d}\cdot) f τ ( R j / d ⋅ ) , τ ∈ J ( A − 1 ) \tau\in\mathcal{J}(A^{-1}) τ ∈ J ( A − 1 ) , project into the caps of M d − m ( A d − m ) \mathcal{M}^{d-m}(A^{d-m}) M d − m ( A d − m ) in these coordinates. Apply (8.1) to f ( R j / d x ) f(R^{j/d}x) f ( R j / d x ) in the physical variables x j + 1 , … , x j + d − m x_{j+1},\ldots,x_{j+d-m} x j + 1 , … , x j + d − m , with parameter A d − m A^{d-m} A d − m .
Apply (2.22) to the sum over V V V on the right hand side of (8.1), with F = ( ∑ τ ∈ J ( A − 1 ) , τ ⊂ ρ ∣ f τ ( R j / d x ) ∣ 2 ) p d − m / 4 F=\left(\sum_{\substack{\tau\in\mathcal{J}(A^{-1}),\ \tau\subset\rho}}|f_\tau(R^{j/d}x)|^2\right)^{p_{d-m}/4} F = ( ∑ τ ∈ J ( A − 1 ) , τ ⊂ ρ ∣ f τ ( R j / d x ) ∣ 2 ) p d − m /4 and t = [ α / ( A ϵ N ( s ) ) ] p d − m / 2 t=[\alpha/(A^\epsilon N(s))]^{p_{d-m}/2} t = [ α / ( A ϵ N ( s )) ] p d − m /2 . Integrating over the remaining physical variables and returning to the original coordinates gives the inequality in Lemma 8.8. □
The RHS of Lemma 8.8 still involves f τ f_\tau f τ , ∣ τ ∣ = A − 1 |\tau|=A^{-1} ∣ τ ∣ = A − 1 . To obtain C K m ( R ) ≲ 1 C_K^m(R)\lesssim1 C K m ( R ) ≲ 1 , we use the following proposition to replace these pieces by the original f θ f_\theta f θ . The factor ( σ s ) ϵ (\sigma s)^\epsilon ( σ s ) ϵ in the following estimate provides a gain when either σ \sigma σ or s s s decreases, allowing us to absorb the refinement losses and close the induction.
Proposition 8.9. Assume the hypotheses of Proposition 8.7 when m < d − 2 m<d-2 m < d − 2 . Let f f f be a Schwartz function with supp f ^ ⊂ Γ j , 1 / K d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\Gamma_{j,1/K}^{d+1}(R) supp f ⊂ Γ j , 1/ K d + 1 ( R ) , 0 ≤ j ≤ m 0\leq j\leq m 0 ≤ j ≤ m . If R − 1 / d ≤ σ ≤ A − 1 R^{-1/d}\leq\sigma\leq A^{-1} R − 1/ d ≤ σ ≤ A − 1 and σ ≤ s ≤ 1 \sigma\leq s\leq1 σ ≤ s ≤ 1 , then
∑ ρ ∈ J ( s ) ∑ Q ∈ G ρ ( α ) Q ∈ G ρ ( α ) ∣ Q ∣ ∥ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( Q ) p d − m ≲ ϵ R ϵ ( σ s ) ϵ ∑ R − 1 / d ≤ v ≤ 1 ∑ ρ ′ ∈ J ( v ) ∑ U ∈ G ρ ′ ( R ϵ ( σ s ) ϵ ) U ⊂ ( v R 1 / d ) m U ρ ′ , R ∣ U ∣ ∥ ( ∑ θ ⊂ ρ ′ ∣ f θ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( U ) p d − m . \sum_{\rho\in\mathcal{J}(s)}\sum_{\substack{Q\in\mathcal{G}_\rho(\alpha)\\ Q\in\mathcal{G}_\rho(\alpha)}}|Q|\left\|\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_\tau|^2\right)^{1/2}\right\|_{L_\#^{p_{d-m}/2}(Q)}^{p_{d-m}}
\lesssim_\epsilon R^\epsilon(\sigma s)^\epsilon
\sum_{R^{-1/d}\leq v\leq1}\sum_{\rho'\in\mathcal{J}(v)}\sum_{\substack{U\in\mathcal{G}_{\rho'}(R^\epsilon(\sigma s)^\epsilon)\\ U\subset(vR^{1/d})^mU_{\rho',R}}}|U|\left\|\left(\sum_{\theta\subset\rho'}|f_\theta|^2\right)^{1/2}\right\|_{L_\#^{p_{d-m}/2}(U)}^{p_{d-m}}. ρ ∈ J ( s ) ∑ Q ∈ G ρ ( α ) Q ∈ G ρ ( α ) ∑ ∣ Q ∣ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L # p d − m /2 ( Q ) p d − m ≲ ϵ R ϵ ( σ s ) ϵ R − 1/ d ≤ v ≤ 1 ∑ ρ ′ ∈ J ( v ) ∑ U ∈ G ρ ′ ( R ϵ ( σ s ) ϵ ) U ⊂ ( v R 1/ d ) m U ρ ′ , R ∑ ∣ U ∣ θ ⊂ ρ ′ ∑ ∣ f θ ∣ 2 1/2 L # p d − m /2 ( U ) p d − m .
For any parameter R R R at which Proposition 8.9 holds and any f f f satisfying its support hypothesis, apply Lemma 8.8 with ϵ / 2 \epsilon/2 ϵ /2 in place of ϵ \epsilon ϵ , and then Proposition 8.9 with σ = A − 1 \sigma=A^{-1} σ = A − 1 to the sum over ρ \rho ρ and Q Q Q on the right hand side of Lemma 8.8, for each fixed s s s . The resulting sum over R ϵ R^\epsilon R ϵ has outer coefficient and amplitude denominator bounded by R ϵ A − ϵ / 2 s ϵ R^\epsilon A^{-\epsilon/2}s^\epsilon R ϵ A − ϵ /2 s ϵ . Since ∑ A − 1 ≤ s ≤ 1 s ϵ ≲ 1 \sum_{A^{-1}\leq s\leq1}s^\epsilon\lesssim1 ∑ A − 1 ≤ s ≤ 1 s ϵ ≲ 1 , this gives
∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ ϵ R ϵ ∑ R − 1 / d ≤ v ≤ 1 ∑ ρ ′ ∈ J ( v ) ∑ U ∈ G ρ ′ ( R ϵ ) U ⊂ ( v R 1 / d ) m U ρ ′ , R ∣ U ∣ ∥ ( ∑ θ ⊂ ρ ′ ∣ f θ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( U ) p d − m . (95) \int_{|f|>\alpha}|f|^{p_{d-m}}\lesssim_\epsilon R^\epsilon\sum_{R^{-1/d}\leq v\leq1}\sum_{\rho'\in\mathcal{J}(v)}\sum_{\substack{U\in\mathcal{G}_{\rho'}(R^\epsilon)\\ U\subset(vR^{1/d})^mU_{\rho',R}}}|U|\left\|\left(\sum_{\theta\subset\rho'}|f_\theta|^2\right)^{1/2}\right\|_{L_\#^{p_{d-m}/2}(U)}^{p_{d-m}}. \tag*{(95)} ∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ ϵ R ϵ R − 1/ d ≤ v ≤ 1 ∑ ρ ′ ∈ J ( v ) ∑ U ∈ G ρ ′ ( R ϵ ) U ⊂ ( v R 1/ d ) m U ρ ′ , R ∑ ∣ U ∣ θ ⊂ ρ ′ ∑ ∣ f θ ∣ 2 1/2 L # p d − m /2 ( U ) p d − m . (95)
We will prove Proposition 8.9 by induction on R R R and, for each fixed R R R , on σ \sigma σ . Assume that the proposition holds at every parameter R ′ < R / 2 R'<R/2 R ′ < R /2 . The deduction above then gives that (8.10) holds at every such R ′ R' R ′ . With R R R fixed, we now prove Proposition 8.9 by induction on σ \sigma σ . For the pair ( σ , s ) (\sigma,s) ( σ , s ) , assume that Proposition 8.9 holds for every pair ( σ ′ , s ′ ) (\sigma',s') ( σ ′ , s ′ ) satisfying R − 1 / d ≤ σ ′ ≤ A − 1 R^{-1/d}\leq\sigma'\leq A^{-1} R − 1/ d ≤ σ ′ ≤ A − 1 , σ ′ ≤ s ′ ≤ 1 \sigma'\leq s'\leq1 σ ′ ≤ s ′ ≤ 1 , and σ ′ s ′ < σ s \sigma's'<\sigma s σ ′ s ′ < σ s . In this argument, Taylor rescaling of f τ f_\tau f τ , ∣ τ ∣ = σ |\tau|=\sigma ∣ τ ∣ = σ , gives the parameter R ′ = R σ d ≤ R A − d ≤ R / 2 R'=R\sigma^d\leq RA^{-d}\leq R/2 R ′ = R σ d ≤ R A − d ≤ R /2 , so (8.10) is available at R ′ R' R ′ . The following subsection describes this induction on σ s \sigma s σ s .
Once Proposition 8.9 is proved at R R R , applying Lemma 8.8 and summing over s s s as above gives (8.10) at R R R . In particular, this gives C K m ( R ) ≲ 1 C_K^m(R)\lesssim1 C K m ( R ) ≲ 1 .
Roadmap continued: induction on σ s \sigma s σ s
We now show how the induction on σ s \sigma s σ s at fixed R R R reduces the proof of Proposition 8.9 to Lemmas 8.10 and 8.11 below. For the argument below, we will choose the parameters such that
Specifically, we will take δ ≪ ϵ η 5 \delta\ll\epsilon\eta^5 δ ≪ ϵ η 5 . Let f f f be a Schwartz function with supp f ^ ⊂ Γ j , 1 / K ( R ) \operatorname{supp}\widehat f\subset\Gamma_{j,1/K}(R) supp f ⊂ Γ j , 1/ K ( R ) , 0 ≤ j ≤ m 0\leq j\leq m 0 ≤ j ≤ m . To estimate the left hand side of Proposition 8.9, we first dilate the Fourier supports of f τ f_\tau f τ , ∣ τ ∣ = σ |\tau|=\sigma ∣ τ ∣ = σ , into the caps of a Taylor cone. For τ ∈ J ( σ ) \tau\in\mathcal{J}(\sigma) τ ∈ J ( σ ) , write ξ = ∑ i = 0 d λ i ϕ d ( i ) ( t τ ) \xi=\sum_{i=0}^{d}\lambda_i\phi_d^{(i)}(t_\tau) ξ = ∑ i = 0 d λ i ϕ d ( i ) ( t τ ) for ξ ∈ supp f τ ^ \xi\in\operatorname{supp}\widehat{f_\tau} ξ ∈ supp f τ . Taylor expansion and (89) give
∣ λ i ∣ ≲ { R − j / d , 0 ≤ i ≤ j , R − j / d σ i − j , j < i ≤ d , ∣ λ j ∣ ∼ R − j / d . |\lambda_i|\lesssim
\begin{cases}
R^{-j/d}, & 0\leq i\leq j,\\
R^{-j/d}\sigma^{i-j}, & j<i\leq d,
\end{cases}
\qquad|\lambda_j|\sim R^{-j/d}. ∣ λ i ∣ ≲ { R − j / d , R − j / d σ i − j , 0 ≤ i ≤ j , j < i ≤ d , ∣ λ j ∣ ∼ R − j / d .
Here we used R − 1 / d ≤ σ R^{-1/d}\leq\sigma R − 1/ d ≤ σ ; the last comparison follows from λ j = R − j / d ( 1 + O ( K − 1 + σ ) ) \lambda_j=R^{-j/d}(1+O(K^{-1}+\sigma)) λ j = R − j / d ( 1 + O ( K − 1 + σ )) and σ ≤ A − 1 \sigma\leq A^{-1} σ ≤ A − 1 .
Set g ( x ) = f ( R j / d σ j x ) g(x)=f(R^{j/d}\sigma^j x) g ( x ) = f ( R j / d σ j x ) . This dilation multiplies each coefficient λ i \lambda_i λ i by R j / d σ j R^{j/d}\sigma^j R j / d σ j , so
∣ R j / d σ j λ i ∣ ≲ σ i ( 0 ≤ i ≤ d ) , ∣ R j / d σ j λ j ∣ ∼ σ j . |R^{j/d}\sigma^j\lambda_i|\lesssim\sigma^i\quad(0\leq i\leq d),
\qquad|R^{j/d}\sigma^j\lambda_j|\sim\sigma^j. ∣ R j / d σ j λ i ∣ ≲ σ i ( 0 ≤ i ≤ d ) , ∣ R j / d σ j λ j ∣ ∼ σ j .
Since σ i = ( σ − d ) − i / d \sigma^i=(\sigma^{-d})^{-i/d} σ i = ( σ − d ) − i / d and j ≤ m j\leq m j ≤ m , this gives
supp f τ ^ ( R j / d σ j ⋅ ) ⊂ θ τ , θ τ ∈ Ξ m d + 1 ( σ − d ) . \operatorname{supp}\widehat{f_\tau}(R^{j/d}\sigma^j\cdot)\subset\theta_\tau,
\qquad\theta_\tau\in\Xi_m^{d+1}(\sigma^{-d}). supp f τ ( R j / d σ j ⋅ ) ⊂ θ τ , θ τ ∈ Ξ m d + 1 ( σ − d ) .
We therefore use the following lemma with g g g and r = σ − d r=\sigma^{-d} r = σ − d .
Lemma 8.10. Assume the hypotheses of Proposition 8.7 for a fixed 0 ≤ m ≤ d − 3 0\leq m\leq d-3 0 ≤ m ≤ d − 3 . Let g g g be a Schwartz function with supp g ^ ⊂ Γ m d + 1 ( r ) \operatorname{supp}\widehat g\subset\Gamma_m^{d+1}(r) supp g ⊂ Γ m d + 1 ( r ) . Then
∫ ∑ θ ∣ g θ ∣ 2 ≳ α 2 ( ∑ θ ∣ g θ ∣ 2 ) p d − m / 2 ≲ ϵ r ϵ ∑ r − 1 / d ≤ u ≤ 1 ∑ ρ ∈ J ( u ) ∑ U ∈ G ρ ( r ϵ ) ∥ U ∥ ∥ ( ∑ θ ⊂ ρ ∣ g θ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( U ) p d − m . \int_{\sum_{\theta}|g_\theta|^2\gtrsim\alpha^2}
\left(\sum_\theta|g_\theta|^2\right)^{p_{d-m}/2}
\lesssim_\epsilon r^\epsilon
\sum_{r^{-1/d}\leq u\leq1}
\sum_{\rho\in\mathcal{J}(u)}
\sum_{U\in\mathcal{G}_\rho(r^\epsilon)}
\|U\|
\left\|\left(\sum_{\theta\subset\rho}|g_\theta|^2\right)^{1/2}\right\|_{L_\#^{p_{d-m}/2}(U)}^{p_{d-m}}. ∫ ∑ θ ∣ g θ ∣ 2 ≳ α 2 ( θ ∑ ∣ g θ ∣ 2 ) p d − m /2 ≲ ϵ r ϵ r − 1/ d ≤ u ≤ 1 ∑ ρ ∈ J ( u ) ∑ U ∈ G ρ ( r ϵ ) ∑ ∥ U ∥ θ ⊂ ρ ∑ ∣ g θ ∣ 2 1/2 L # p d − m /2 ( U ) p d − m .
Apply Lemma 8.10 to g ( x ) = f ( R j / d σ j x ) g(x)=f(R^{j/d}\sigma^j x) g ( x ) = f ( R j / d σ j x ) , with r = σ − d r=\sigma^{-d} r = σ − d and pieces f τ ( R j / d σ j x ) f_\tau(R^{j/d}\sigma^j x) f τ ( R j / d σ j x ) , ∣ τ ∣ = σ |\tau|=\sigma ∣ τ ∣ = σ . On returning to the original coordinates, an averaging box of angular width u u u has dimensions
R j / d σ j − m − 2 u m + 2 × R j / d σ j − m − 2 u m + 1 × ⋯ × R j / d σ j − m − 2 u m + 2 − d . R^{j/d}\sigma^{j-m-2}u^{m+2}\times R^{j/d}\sigma^{j-m-2}u^{m+1}\times\cdots\times R^{j/d}\sigma^{j-m-2}u^{m+2-d}. R j / d σ j − m − 2 u m + 2 × R j / d σ j − m − 2 u m + 1 × ⋯ × R j / d σ j − m − 2 u m + 2 − d .
These boxes are ( u / σ ) m − j × ( u R 1 / d σ R 1 / d ) j U ρ , R (u/\sigma)^{m-j}\times\left(\frac{uR^{1/d}}{\sigma R^{1/d}}\right)^jU_{\rho,R} ( u / σ ) m − j × ( σ R 1/ d u R 1/ d ) j U ρ , R . Since u ≥ σ u\geq\sigma u ≥ σ and j ≤ m j\leq m j ≤ m , we can apply Lemma 2.14 , with
F = ( ∑ τ ⊂ ρ ∣ f τ ∣ 2 ) p d − m / 4 , F=\left(\sum_{\tau\subset\rho}|f_\tau|^2\right)^{p_{d-m}/4}, F = ( τ ⊂ ρ ∑ ∣ f τ ∣ 2 ) p d − m /4 ,
from these larger boxes to B ∥ ( u R 1 / d σ R 1 / d ) j U ρ , R B\parallel\left(\frac{uR^{1/d}}{\sigma R^{1/d}}\right)^jU_{\rho,R} B ∥ ( σ R 1/ d u R 1/ d ) j U ρ , R . The output, with an arbitrarily small loss δ > 0 \delta>0 δ > 0 , is
∫ ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ≳ α 2 ( ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) p d − m / 2 ≲ δ σ − δ ∑ σ ≤ u ≤ 1 ∑ ρ ∈ J ( u ) ∑ B ∈ G ρ ( σ − δ ) B ∥ ( u R 1 / d σ R 1 / d ) j U ρ , R ∣ B ∣ ∥ ( ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( B ) p d − m . (96) \int_{\sum_{|\tau|=\sigma}|f_\tau|^2\gtrsim\alpha^2}
\left(\sum_{|\tau|=\sigma}|f_\tau|^2\right)^{p_{d-m}/2}
\lesssim_\delta\sigma^{-\delta}
\sum_{\sigma\leq u\leq1}
\sum_{\rho\in\mathcal{J}(u)}
\sum_{\substack{B\in\mathcal{G}_\rho(\sigma^{-\delta})\\
B\parallel\left(\frac{uR^{1/d}}{\sigma R^{1/d}}\right)^jU_{\rho,R}}}
|B|
\left\|\left(\sum_{\tau\subset\rho,\ |\tau|=\sigma}|f_\tau|^2\right)^{1/2}\right\|_{L_\#^{p_{d-m}/2}(B)}^{p_{d-m}}. \tag*{(96)} ∫ ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ≳ α 2 ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 p d − m /2 ≲ δ σ − δ σ ≤ u ≤ 1 ∑ ρ ∈ J ( u ) ∑ B ∈ G ρ ( σ − δ ) B ∥ ( σ R 1/ d u R 1/ d ) j U ρ , R ∑ ∣ B ∣ τ ⊂ ρ , ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L # p d − m /2 ( B ) p d − m . (96)
For a parent ρ = [ t ρ , t ρ + s ] \rho=[t_\rho,t_\rho+s] ρ = [ t ρ , t ρ + s ] on the left hand side of Proposition 8.9, use (96) when s ≫ σ s\gg\sigma s ≫ σ , and the locally constant property when s ∼ σ s\sim\sigma s ∼ σ . In the first case, apply (96) to f ρ ( L t ρ , s x ) f_\rho(L_{t_\rho,s}x) f ρ ( L t ρ , s x ) , with pieces f τ ( L t ρ , s x ) f_\tau(L_{t_\rho,s}x) f τ ( L t ρ , s x ) , τ ⊂ ρ \tau\subset\rho τ ⊂ ρ , where L t ρ , s = ( ϕ d ( t ρ ) , s ϕ d ′ ( t ρ ) , … , s d ϕ d ( d ) ( t ρ ) ) L_{t_\rho,s}=(\phi_d(t_\rho),s\phi_d'(t_\rho),\ldots,s^d\phi_d^{(d)}(t_\rho)) L t ρ , s = ( ϕ d ( t ρ ) , s ϕ d ′ ( t ρ ) , … , s d ϕ d ( d ) ( t ρ )) . The parameter is R s d Rs^d R s d , the cap width is σ / s \sigma/s σ / s , and the amplitude is α / N ( s ) \alpha/N(s) α / N ( s ) . When s ∼ σ s\sim\sigma s ∼ σ , the locally constant property on the dual of the Fourier support box, whose dimensions are R − j / d σ − j × R − j / d σ 1 − j × ⋯ × R − j / d σ d − j R^{-j/d}\sigma^{-j}\times R^{-j/d}\sigma^{1-j}\times\cdots\times R^{-j/d}\sigma^{d-j} R − j / d σ − j × R − j / d σ 1 − j × ⋯ × R − j / d σ d − j , gives directly
∑ Q ∈ G ρ ( α ) ∣ Q ∣ ∥ ( ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( Q ) p d − m ≲ ∑ B ∈ G ρ ( α ) B ∥ ( s R 1 / d σ R 1 / d ) 2 U ρ , R ∣ B ∣ ∥ ( ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( B ) p d − m . (97) \sum_{Q\in\mathcal{G}_\rho(\alpha)}
|Q|
\left\|\left(\sum_{\tau\subset\rho,\ |\tau|=\sigma}|f_\tau|^2\right)^{1/2}\right\|_{L_\#^{p_{d-m}/2}(Q)}^{p_{d-m}}
\lesssim
\sum_{\substack{B\in\mathcal{G}_\rho(\alpha)\\
B\parallel\left(\frac{sR^{1/d}}{\sigma R^{1/d}}\right)^2U_{\rho,R}}}
|B|
\left\|\left(\sum_{\tau\subset\rho,\ |\tau|=\sigma}|f_\tau|^2\right)^{1/2}\right\|_{L_\#^{p_{d-m}/2}(B)}^{p_{d-m}}. \tag*{(97)} Q ∈ G ρ ( α ) ∑ ∣ Q ∣ τ ⊂ ρ , ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L # p d − m /2 ( Q ) p d − m ≲ B ∈ G ρ ( α ) B ∥ ( σ R 1/ d s R 1/ d ) 2 U ρ , R ∑ ∣ B ∣ τ ⊂ ρ , ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L # p d − m /2 ( B ) p d − m . (97)
Returning to the original coordinates in (8.11), we combine it with (8.12) and use N ( s ) N ρ ( u ) ≲ N ( u ) N(s)N_{\rho}(u)\lesssim N(u) N ( s ) N ρ ( u ) ≲ N ( u ) from Lemma 2.18 to bound the left hand side of Proposition 8.9 by
≲ δ σ − δ ∑ σ ≤ u ≤ s ∑ ρ ∈ J ( u ) ∑ B ∈ G ρ ( ( u R 1 / d ) j ( σ R 1 / d ) 2 U ρ , R ) ∣ B ∣ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ ⋅ ∥ L # p d − m / 2 ( B ) p d − m (98) \lesssim_{\delta}\sigma^{-\delta}\sum_{\sigma\leq u\leq s}\sum_{\rho\in\mathcal{J}(u)}\sum_{B\in\mathcal{G}_{\rho}\left(\frac{(uR^{1/d})^{j}}{(\sigma R^{1/d})^{2}}U_{\rho,R}\right)}|B|\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_{\tau}|^{2}\right)^{1/2}\left\|\,\cdot\,\right\|_{L_{\#}^{p_{d}-m/2}(B)}^{p_{d}-m} \tag*{(98)} ≲ δ σ − δ σ ≤ u ≤ s ∑ ρ ∈ J ( u ) ∑ B ∈ G ρ ( ( σ R 1/ d ) 2 ( u R 1/ d ) j U ρ , R ) ∑ ∣ B ∣ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 ∥ ⋅ ∥ L # p d − m /2 ( B ) p d − m (98)
The averages in (8.13) still contain f τ f_{\tau} f τ , whereas Proposition 8.9 requires the original f θ f_{\theta} f θ . We make this replacement by rescaling when a parent contains few σ \sigma σ intervals, and by refining when it contains many. Fix u u u . We estimate the corresponding term on the RHS of (8.13).
Case 1: u < σ 1 − η u<\sigma^{1-\eta} u < σ 1 − η . There are at most σ − η \sigma^{-\eta} σ − η intervals τ ⊂ ρ \tau\subset\rho τ ⊂ ρ of width σ \sigma σ . Apply (2.22) to the fixed u u u sum in (8.13), with F = ( ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ) p d − m / 4 F=\left(\sum_{\tau\subset\rho,\ |\tau|=\sigma}|f_{\tau}|^{2}\right)^{p_{d}-m/4} F = ( ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ) p d − m /4 and t = [ α σ δ / N ( u ) ] p d − m / 2 t=[\alpha\sigma^{\delta}/N(u)]^{p_{d}-m/2} t = [ α σ δ / N ( u ) ] p d − m /2 . The fixed u u u term is therefore bounded by
≲ σ − δ ∑ ρ ∈ J ( u ) ∫ ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ≳ [ α σ δ / N ( u ) ] 2 ( ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ) p d − m / 2 . \lesssim\sigma^{-\delta}\sum_{\rho\in\mathcal{J}(u)}\int_{\sum_{\tau\subset\rho,\ |\tau|=\sigma}|f_{\tau}|^{2}\gtrsim[\alpha\sigma^{\delta}/N(u)]^{2}}\left(\sum_{\tau\subset\rho,\ |\tau|=\sigma}|f_{\tau}|^{2}\right)^{p_{d}-m/2}. ≲ σ − δ ρ ∈ J ( u ) ∑ ∫ ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ≳ [ α σ δ / N ( u ) ] 2 τ ⊂ ρ , ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 p d − m /2 .
In this integral, discard the terms ∣ f τ ∣ ≪ α σ δ / ( N ( u ) N ρ ( σ ) ) |f_{\tau}|\ll\alpha\sigma^{\delta}/(N(u)N_{\rho}(\sigma)) ∣ f τ ∣ ≪ α σ δ / ( N ( u ) N ρ ( σ )) and apply Hölder. The fixed u u u term is bounded by
≲ σ − C η ∑ ρ ∈ J ( u ) ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ ≳ α σ δ / ( N ( u ) N ρ ( σ ) ) ∫ ∣ f τ ∣ p d − m . (99) \lesssim\sigma^{-C\eta}\sum_{\rho\in\mathcal{J}(u)}\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma\\|f_{\tau}|\gtrsim\alpha\sigma^{\delta}/(N(u)N_{\rho}(\sigma))}}\int|f_{\tau}|^{p_{d}-m}. \tag*{(99)} ≲ σ − C η ρ ∈ J ( u ) ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ ≳ α σ δ / ( N ( u ) N ρ ( σ )) ∑ ∫ ∣ f τ ∣ p d − m . (99)
We use (8.10) inductively at parameters smaller than R R R . For τ = [ t τ , t τ + σ ] \tau=[t_{\tau},t_{\tau}+\sigma] τ = [ t τ , t τ + σ ] , the function f τ ( L t τ , σ T x ) f_{\tau}(L_{t_{\tau},\sigma}^{T}x) f τ ( L t τ , σ T x ) has parameter R σ d < R R\sigma^{d}<R R σ d < R . The left endpoint t I t_I t I is sent to ( t I − t τ ) / σ (t_I-t_{\tau})/\sigma ( t I − t τ ) / σ , and its coefficients become σ − i λ i \sigma^{-i}\lambda_i σ − i λ i . For i < j i<j i < j , ∣ σ − i λ i ∣ ≤ K − 1 ( R σ d ) − j / d |\sigma^{-i}\lambda_i|\leq K^{-1}(R\sigma^{d})^{-j/d} ∣ σ − i λ i ∣ ≤ K − 1 ( R σ d ) − j / d , while ∣ σ − j λ j − ( R σ d ) − j / d ∣ ≤ K − 1 ( R σ d ) − j / d |\sigma^{-j}\lambda_j-(R\sigma^{d})^{-j/d}|\leq K^{-1}(R\sigma^{d})^{-j/d} ∣ σ − j λ j − ( R σ d ) − j / d ∣ ≤ K − 1 ( R σ d ) − j / d . Thus the Fourier support of f τ ( L t τ , σ T ⋅ ) f_{\tau}(L_{t_{\tau},\sigma}^{T}\cdot) f τ ( L t τ , σ T ⋅ ) lies in Γ j , 1 / K d + 1 ( R σ d ) \Gamma_{j,1/K}^{d+1}(R\sigma^{d}) Γ j , 1/ K d + 1 ( R σ d ) with the same K K K . Apply (8.10) to this function with amplitude α σ δ / ( N ( u ) N ρ ( σ ) ) \alpha\sigma^{\delta}/(N(u)N_{\rho}(\sigma)) α σ δ / ( N ( u ) N ρ ( σ )) . The averaging boxes return to translates of ( v R 1 / d ) m U ρ ′ , R (vR^{1/d})^{m}U_{\rho',R} ( v R 1/ d ) m U ρ ′ , R , where ρ ′ ⊂ τ \rho'\subset\tau ρ ′ ⊂ τ and ∣ ρ ′ ∣ = v |\rho'|=v ∣ ρ ′ ∣ = v , since
σ − i ( R σ d ) ( m + 2 ) / d ( v / σ ) m + 2 − i = R ( m + 2 ) / d v m + 2 − i , 0 ≤ i ≤ d . (100) \sigma^{-i}(R\sigma^{d})^{(m+2)/d}(v/\sigma)^{m+2-i}=R^{(m+2)/d}v^{m+2-i},\qquad0\leq i\leq d. \tag*{(100)} σ − i ( R σ d ) ( m + 2 ) / d ( v / σ ) m + 2 − i = R ( m + 2 ) / d v m + 2 − i , 0 ≤ i ≤ d . (100)
Moreover, N ( u ) N ρ ( σ ) N τ ( v ) ≲ N ( v ) N(u)N_{\rho}(\sigma)N_{\tau}(v)\lesssim N(v) N ( u ) N ρ ( σ ) N τ ( v ) ≲ N ( v ) by Lemma 2.18 for the uniformized family. Consequently (8.14) is bounded by
≲ ϵ σ − C η ( R σ d ) ϵ ∑ R − 1 / d ≤ v ≤ σ ∑ ρ ′ ∈ J ( v ) ∑ U ∈ G ρ ′ ( σ − C η ( R σ d ) ϵ ) ∣ U ∣ ( ∑ θ ⊂ ρ ′ ∣ f θ ∣ 2 ) 1 / 2 ∥ ⋅ ∥ L # p d − m / 2 ( U ) p d − m . (101) \lesssim_{\epsilon}\sigma^{-C\eta}(R\sigma^{d})^{\epsilon}\sum_{R^{-1/d}\leq v\leq\sigma}\sum_{\rho'\in\mathcal{J}(v)}\sum_{U\in\mathcal{G}_{\rho'}\left(\frac{\sigma^{-C\eta}(R\sigma^{d})^{\epsilon}}{}\right)}|U|\left(\sum_{\theta\subset\rho'}|f_{\theta}|^{2}\right)^{1/2}\|\,\cdot\,\|_{L_{\#}^{p_{d}-m/2}(U)}^{p_{d}-m}. \tag*{(101)} ≲ ϵ σ − C η ( R σ d ) ϵ R − 1/ d ≤ v ≤ σ ∑ ρ ′ ∈ J ( v ) ∑ U ∈ G ρ ′ ( σ − C η ( R σ d ) ϵ ) ∑ ∣ U ∣ θ ⊂ ρ ′ ∑ ∣ f θ ∣ 2 1/2 ∥ ⋅ ∥ L # p d − m /2 ( U ) p d − m . (101)
Since s ≥ σ s\geq\sigma s ≥ σ ,
σ − C η ( R σ d ) ϵ R ϵ ( σ s ) ϵ ≤ σ ( d − 2 ) ϵ − C η . \frac{\sigma^{-C\eta}(R\sigma^{d})^{\epsilon}}{R^{\epsilon}(\sigma s)^{\epsilon}}\leq\sigma^{(d-2)\epsilon-C\eta}. R ϵ ( σ s ) ϵ σ − C η ( R σ d ) ϵ ≤ σ ( d − 2 ) ϵ − C η .
Since (8.16) already contains the original f θ f_{\theta} f θ , the iteration stops in this case. We keep this factor for (8.18), where the bounds from both cases are combined.
Case 2: u ≥ σ 1 − η u\geq\sigma^{1-\eta} u ≥ σ 1 − η . For each ρ ∈ J ( u ) \rho\in\mathcal{J}(u) ρ ∈ J ( u ) , Taylor rescaling sends the boxes B B B in (8.13) to cubes of side length ( R u d ) j / d ( σ / u ) − 2 (Ru^{d})^{j/d}(\sigma/u)^{-2} ( R u d ) j / d ( σ / u ) − 2 , with parameter R u d Ru^{d} R u d and cap width σ / u \sigma/u σ / u . We therefore use the following lemma to reduce either σ \sigma σ or u u u , allowing us to continue the induction on σ s \sigma s σ s .
Lemma 8.11. Let f f f be a Schwartz function with supp f ^ ⊂ Γ j , 1 / K d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\Gamma_{j,1/K}^{d+1}(R) supp f ⊂ Γ j , 1/ K d + 1 ( R ) . For η > 0 \eta>0 η > 0 sufficiently small and R − 1 / d < σ R^{-1/d}<\sigma R − 1/ d < σ sufficiently small in terms of η \eta η , there are dyadic σ ′ , s ′ \sigma',s' σ ′ , s ′ with R − 1 / d ≤ σ ′ ≤ max { R − 1 / d , σ 3 / 2 } R^{-1/d}\leq\sigma'\leq\max\{R^{-1/d},\sigma^{3/2}\} R − 1/ d ≤ σ ′ ≤ max { R − 1/ d , σ 3/2 } , and
σ ≤ s ′ ≲ σ η 4 \sigma\le s' \lesssim\sigma^{\eta^4} σ ≤ s ′ ≲ σ η 4 such that
∑ B ∥ [ 0 , R j / d σ − 2 ] × R d + 1 B ∈ G [ 0 , 1 ] ( α ) ∥ ∣ B ∣ ∥ ∥ ( ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( B ) p d − m ≲ δ , η σ − C η 4 ∑ Q ∥ [ 0 , 1 ] d + 1 Q ∈ G [ 0 , 1 ] ( α σ − C η 4 ) ∣ Q ∣ ∥ ( ∑ ∣ τ ∣ = σ ′ ∣ f τ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( Q ) p d − m + σ − δ ∑ ρ ∈ J ( s ′ ) ∑ Q ∈ G ρ ( α σ − δ ) Q ∥ ρ ∗ ∣ Q ∣ ∥ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( Q ) p d − m . \begin{aligned}
\sum_{\substack{B\parallel[0,R^{j/d}\sigma^{-2}]\times\mathbb{R}^{d+1}\\ B\in\mathcal{G}_{[0,1]}(\alpha)}} \lVert|B|\rVert\left\lVert\left(\sum_{|\tau|=\sigma}|f_\tau|^2\right)^{1/2}\right\rVert_{L_{\#}^{p_d-m/2}(B)}^{p_d-m}
\lesssim_{\delta,\eta} \sigma^{-C\eta^4}\sum_{\substack{Q\parallel[0,1]^{d+1}\\ Q\in\mathcal{G}_{[0,1]}\left(\frac{\alpha}{\sigma-C\eta^4}\right)}} |Q|\left\lVert\left(\sum_{|\tau|=\sigma'}|f_\tau|^2\right)^{1/2}\right\rVert_{L_{\#}^{p_d-m/2}(Q)}^{p_d-m} \\
&+\sigma^{-\delta}\sum_{\rho\in\mathcal{J}(s')}\sum_{\substack{Q\in\mathcal{G}_{\rho}\left(\frac{\alpha}{\sigma^{-\delta}}\right)\\ Q\parallel\rho^*}} |Q|\left\lVert\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_\tau|^2\right)^{1/2}\right\rVert_{L_{\#}^{p_d-m/2}(Q)}^{p_d-m}.
\end{aligned} B ∥ [ 0 , R j / d σ − 2 ] × R d + 1 B ∈ G [ 0 , 1 ] ( α ) ∑ ∥ ∣ B ∣ ∥ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L # p d − m /2 ( B ) p d − m ≲ δ , η σ − C η 4 Q ∥ [ 0 , 1 ] d + 1 Q ∈ G [ 0 , 1 ] ( σ − C η 4 α ) ∑ ∣ Q ∣ ∣ τ ∣ = σ ′ ∑ ∣ f τ ∣ 2 1/2 L # p d − m /2 ( Q ) p d − m + σ − δ ρ ∈ J ( s ′ ) ∑ Q ∈ G ρ ( σ − δ α ) Q ∥ ρ ∗ ∑ ∣ Q ∣ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L # p d − m /2 ( Q ) p d − m .
For p d − m = 4 p_d-m=4 p d − m = 4 , only the first sum over Q Q Q is needed.
Assuming Lemma 8.11, we now complete Case 2. For ρ = [ t ρ , t ρ + u ] \rho=[t_\rho,t_\rho+u] ρ = [ t ρ , t ρ + u ] , the function f ρ ( L t ρ , u T ⋅ ) f_\rho(L_{t_\rho,u}^{T}\cdot) f ρ ( L t ρ , u T ⋅ ) has Fourier support in Γ j , 1 / K d + 1 ( R u d ) \Gamma_{j,1/K}^{d+1}(Ru^d) Γ j , 1/ K d + 1 ( R u d ) , by the calculation in Case 1 with u u u in place of σ \sigma σ . Since σ / u ≤ σ η ≤ A − η \sigma/u\le\sigma^\eta\le A^{-\eta} σ / u ≤ σ η ≤ A − η , taking A A A sufficiently large in terms of η \eta η allows us to apply Lemma 8.11 to this function with parameter R u d Ru^d R u d , cap width σ / u \sigma/u σ / u , and amplitude α σ δ / N ( u ) \alpha\sigma^\delta/N(u) α σ δ / N ( u ) . Sum over ρ \rho ρ before choosing common dyadic scales; the losses from these choices are absorbed into σ − C η 4 \sigma^{-C\eta^4} σ − C η 4 in the first term and σ − δ \sigma^{-\delta} σ − δ in the second. Returning to the original coordinates and using σ / u ≤ σ η \sigma/u\le\sigma^\eta σ / u ≤ σ η and δ ≪ η 4 \delta\ll\eta^4 δ ≪ η 4 , we bound the fixed u u u term on the RHS of (8.13) by
≲ σ − C η 4 ∑ ρ ∈ J ( u ) ∑ Q ∥ ρ ∗ Q ∈ G ρ ( α σ − C η 4 ) ∣ Q ∣ ∥ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( Q ) p d − m + σ − 2 δ ∑ ρ ′ ∈ J ( u ′ ) ∑ Q ∥ ( ρ ′ ) ∗ Q ∈ G ρ ′ ( α σ − 2 δ ) ∣ Q ∣ ∥ ( ∑ τ ⊂ ρ ′ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( Q ) p d − m . (102) \begin{aligned}
\lesssim\sigma^{-C\eta^4}\sum_{\rho\in\mathcal{J}(u)}\sum_{\substack{Q\parallel\rho^*\\ Q\in\mathcal{G}_{\rho}\left(\frac{\alpha}{\sigma^{-C\eta^4}}\right)}} |Q|\left\lVert\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_\tau|^2\right)^{1/2}\right\rVert_{L_{\#}^{p_d-m/2}(Q)}^{p_d-m} \\
&+\sigma^{-2\delta}\sum_{\rho'\in\mathcal{J}(u')}\sum_{\substack{Q\parallel(\rho')^*\\ Q\in\mathcal{G}_{\rho'}\left(\frac{\alpha}{\sigma^{-2\delta}}\right)}} |Q|\left\lVert\left(\sum_{\substack{\tau\subset\rho'\\|\tau|=\sigma}}|f_\tau|^2\right)^{1/2}\right\rVert_{L_{\#}^{p_d-m/2}(Q)}^{p_d-m}. \tag*{(102)}
\end{aligned} ≲ σ − C η 4 ρ ∈ J ( u ) ∑ Q ∥ ρ ∗ Q ∈ G ρ ( σ − C η 4 α ) ∑ ∣ Q ∣ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L # p d − m /2 ( Q ) p d − m + σ − 2 δ ρ ′ ∈ J ( u ′ ) ∑ Q ∥ ( ρ ′ ) ∗ Q ∈ G ρ ′ ( σ − 2 δ α ) ∑ ∣ Q ∣ τ ⊂ ρ ′ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L # p d − m /2 ( Q ) p d − m . (102)
where σ ′ ≤ max { R − 1 / d , σ 1 + η / 4 } \sigma'\le\max\{R^{-1/d},\sigma^{1+\eta/4}\} σ ′ ≤ max { R − 1/ d , σ 1 + η /4 } and u ′ ≤ u σ η 5 / 2 u'\le u\sigma^{\eta^5/2} u ′ ≤ u σ η 5 /2 . The first sum over Q Q Q is comparable to the left hand side of Proposition 8.9 at ( σ ′ , u ) (\sigma',u) ( σ ′ , u ) ; the second is comparable to that left hand side at ( σ , u ′ ) (\sigma,u') ( σ , u ′ ) , with the amplitudes in (8.17). In both cases the new product is smaller than σ u \sigma u σ u .
Suppose first that σ ′ > R − 1 / d \sigma'>R^{-1/d} σ ′ > R − 1/ d . Applying Proposition 8.9 inductively at ( σ ′ , u ) (\sigma',u) ( σ ′ , u ) , with amplitude α / σ − C η 4 \alpha/\sigma^{-C\eta^4} α / σ − C η 4 , bounds the first term on the right hand side of (8.17) by
≲ ϵ σ − C η 4 R ϵ ( σ ′ u ) ϵ ∑ R − 1 / d ≤ v ≤ 1 ∑ ρ ′ ∈ J ( v ) ∑ U ∥ ( v R 1 / d ) m U ρ ′ , R U ∈ G ρ ′ ( α σ − C η 4 R ϵ ( σ ′ u ) ϵ ) ∣ U ∣ ∥ ( ∑ θ ⊂ ρ ′ ∣ f θ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( U ) p d − m . \lesssim_\epsilon\sigma^{-C\eta^4}R^\epsilon(\sigma'u)^\epsilon
\sum_{R^{-1/d}\le v\le1}\sum_{\rho'\in\mathcal{J}(v)}
\sum_{\substack{U\parallel(vR^{1/d})^mU_{\rho',R}\\ U\in\mathcal{G}_{\rho'}\left(\frac{\alpha}{\sigma^{-C\eta^4}R^\epsilon(\sigma'u)^\epsilon}\right)}}
|U|\left\lVert\left(\sum_{\theta\subset\rho'}|f_\theta|^2\right)^{1/2}\right\rVert_{L_{\#}^{p_d-m/2}(U)}^{p_d-m}. ≲ ϵ σ − C η 4 R ϵ ( σ ′ u ) ϵ R − 1/ d ≤ v ≤ 1 ∑ ρ ′ ∈ J ( v ) ∑ U ∥ ( v R 1/ d ) m U ρ ′ , R U ∈ G ρ ′ ( σ − C η 4 R ϵ ( σ ′ u ) ϵ α ) ∑ ∣ U ∣ θ ⊂ ρ ′ ∑ ∣ f θ ∣ 2 1/2 L # p d − m /2 ( U ) p d − m .
Since σ ′ ≤ σ 1 + η / 4 \sigma'\le\sigma^{1+\eta/4} σ ′ ≤ σ 1 + η /4 and u ≤ s u\le s u ≤ s ,
σ − C η 4 R ϵ ( σ ′ u ) ϵ R ϵ ( σ s ) ϵ = σ − C η 4 ( σ ′ σ ) ϵ ( u s ) ϵ ≤ σ ϵ η / 4 − C η 4 . \frac{\sigma^{-C\eta^4}R^\epsilon(\sigma'u)^\epsilon}{R^\epsilon(\sigma s)^\epsilon}
=\sigma^{-C\eta^4}\left(\frac{\sigma'}{\sigma}\right)^\epsilon\left(\frac{u}{s}\right)^\epsilon
\le\sigma^{\epsilon\eta/4-C\eta^4}. R ϵ ( σ s ) ϵ σ − C η 4 R ϵ ( σ ′ u ) ϵ = σ − C η 4 ( σ σ ′ ) ϵ ( s u ) ϵ ≤ σ ϵη /4 − C η 4 .
If instead σ ′ = R − 1 / d \sigma'=R^{-1/d} σ ′ = R − 1/ d , the first sum over Q Q Q on the right hand side of (8.17) already contains f θ f_\theta f θ . Apply Lemma 8.10 with η \eta η in place of ϵ \epsilon ϵ to each function f ρ ( L t ρ , u T x ) f_\rho(L_{t_\rho,u}^{T}x) f ρ ( L t ρ , u T x ) , ρ ∈ J ( u ) \rho\in\mathcal{J}(u) ρ ∈ J ( u ) , with parameter R u d Ru^d R u d and amplitude α / ( σ − C η 4 N ( u ) ) \alpha/(\sigma^{-C\eta^4}N(u)) α / ( σ − C η 4 N ( u )) . For each returned parent ρ ′ ⊂ ρ \rho'\subset\rho ρ ′ ⊂ ρ of width v v v , Lemma 2.18 gives N ( u ) N ρ ( v ) ≲ N ( v ) N(u)N_{\rho}(v)\lesssim N(v) N ( u ) N ρ ( v ) ≲ N ( v ) . Using this bound in the selection condition and (8.15) for the averaging boxes, we sum over ρ \rho ρ to obtain the right hand side of (8.10) with R ϵ R^\epsilon R ϵ replaced, both outside the sum and in G ρ ′ \mathcal{G}_{\rho'} G ρ ′ , by σ − C η 4 R η ≤ R C η \sigma^{-C\eta^4}R^\eta\le R^{C\eta} σ − C η 4 R η ≤ R C η .
For the second term on the right hand side of (8.17), regardless of σ ′ \sigma' σ ′ , applying Proposition 8.9 inductively at ( σ , u ′ ) (\sigma,u') ( σ , u ′ ) , with amplitude α / σ − 2 δ \alpha/\sigma^{-2\delta} α / σ − 2 δ , gives
≲ ϵ σ − 2 δ R ϵ ( σ u ′ ) ϵ ∑ R − 1 / d ≤ v ≤ 1 ∑ ρ ′ ∈ J ( v ) ∑ U ∥ ( v R 1 / d ) m U ρ ′ , R U ∈ G ρ ′ ( α σ − 2 δ R ϵ ( σ u ′ ) ϵ ) ∣ U ∣ ∥ ( ∑ θ ⊂ ρ ′ ∣ f θ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( U ) p d − m . \lesssim_\epsilon\sigma^{-2\delta}R^\epsilon(\sigma u')^\epsilon
\sum_{R^{-1/d}\le v\le1}\sum_{\rho'\in\mathcal{J}(v)}
\sum_{\substack{U\parallel(vR^{1/d})^mU_{\rho',R}\\ U\in\mathcal{G}_{\rho'}\left(\frac{\alpha}{\sigma^{-2\delta}R^\epsilon(\sigma u')^\epsilon}\right)}}
|U|\left\lVert\left(\sum_{\theta\subset\rho'}|f_\theta|^2\right)^{1/2}\right\rVert_{L_{\#}^{p_d-m/2}(U)}^{p_d-m}. ≲ ϵ σ − 2 δ R ϵ ( σ u ′ ) ϵ R − 1/ d ≤ v ≤ 1 ∑ ρ ′ ∈ J ( v ) ∑ U ∥ ( v R 1/ d ) m U ρ ′ , R U ∈ G ρ ′ ( σ − 2 δ R ϵ ( σ u ′ ) ϵ α ) ∑ ∣ U ∣ θ ⊂ ρ ′ ∑ ∣ f θ ∣ 2 1/2 L # p d − m /2 ( U ) p d − m .
Since u ′ ≤ u σ η 5 / 2 u'\le u\sigma^{\eta^5/2} u ′ ≤ u σ η 5 /2 and u ≤ s u\le s u ≤ s ,
σ − 2 δ R ϵ ( σ u ′ ) ϵ R ϵ ( σ s ) ϵ = σ − 2 δ ( u ′ s ) ϵ ≤ σ ϵ η 5 / 2 − 2 δ . \frac{\sigma^{-2\delta}R^\epsilon(\sigma u')^\epsilon}{R^\epsilon(\sigma s)^\epsilon}
=\sigma^{-2\delta}\left(\frac{u'}{s}\right)^\epsilon
\le\sigma^{\epsilon\eta^5/2-2\delta}. R ϵ ( σ s ) ϵ σ − 2 δ R ϵ ( σ u ′ ) ϵ = σ − 2 δ ( s u ′ ) ϵ ≤ σ ϵ η 5 /2 − 2 δ .
For the initial case σ = R − 1 / d \sigma=R^{-1/d} σ = R − 1/ d , apply Lemma 8.10 with η \eta η in place of ϵ \epsilon ϵ to f ρ ( L τ ρ , s T x ) f_\rho(L^T_{\tau_{\rho,s}}x) f ρ ( L τ ρ , s T x ) , ρ ∈ J ( s ) \rho\in\mathcal{J}(s) ρ ∈ J ( s ) , with parameter R s d Rs^d R s d and amplitude α / N ( s ) \alpha/N(s) α / N ( s ) . Its output is the right hand side of (8.10), with both occurrences of R ϵ R^\epsilon R ϵ replaced by R η R^\eta R η .
The estimates for the two terms in (8.17) (with σ ′ > R − 1 / d \sigma'>R^{-1/d} σ ′ > R − 1/ d for the first term) and the bound (8.16) give, respectively, the following bounds for the ratios of the outer coefficients and amplitude denominators to R ϵ ( σ s ) ϵ R^\epsilon(\sigma s)^\epsilon R ϵ ( σ s ) ϵ :
σ ϵ η / 4 − C η 4 , σ ϵ η 5 / 2 − 2 δ , σ ( d − 2 ) ϵ − C η . (103) \sigma^{\epsilon\eta/4-C\eta^4},\qquad\sigma^{\epsilon\eta^{5/2}-2\delta},\qquad\sigma^{(d-2)\epsilon-C\eta}. \tag*{(103)} σ ϵη /4 − C η 4 , σ ϵ η 5/2 − 2 δ , σ ( d − 2 ) ϵ − C η . (103)
Our choices of η , δ \eta,\delta η , δ make all three exponents in (8.18) positive. Since σ ≤ A − 1 \sigma\leq A^{-1} σ ≤ A − 1 , taking A A A sufficiently large therefore allows these factors to absorb the implicit constants and the dyadic sum over u u u in (8.13). For σ ′ = R − 1 / d \sigma'=R^{-1/d} σ ′ = R − 1/ d or σ = R − 1 / d \sigma=R^{-1/d} σ = R − 1/ d , the inequalities s ≥ σ ≥ R − 1 / d s\geq\sigma\geq R^{-1/d} s ≥ σ ≥ R − 1/ d give R C η R ϵ ( σ s ) ϵ ≤ R − ϵ / C \frac{R^{C\eta}}{R^\epsilon(\sigma s)^\epsilon}\leq R^{-\epsilon/C} R ϵ ( σ s ) ϵ R C η ≤ R − ϵ / C , which also absorbs the constants and the dyadic sum over u u u for sufficiently large R R R .
Summing the estimates for the terms in (8.13) therefore proves Proposition 8.9 at the current R R R . Applying Lemma 8.8 as in the derivation of (8.10) then gives (8.10) at the same R R R , completing the induction on R R R .
We prove Lemmas 8.10 and 8.11 for m < d − 2 m<d-2 m < d − 2 in Section 9.
Remark 8.12. To complete Step 1 of Proposition 8.7, we also need (8.9) for the current m m m . In Section 9, we first bound the left hand side of Lemma 8.10 by the RHS of (8.8) with R = r R=r R = r and f = g f=g f = g . Lemma 8.5 then gives Lemma 8.10 as stated.
Following the induction in Section 8.2, we use this strengthened form of Lemma 8.10 together with Lemma 8.11 to strengthen Proposition 8.9; we bound its left hand side by the RHS of (8.8) for the current m m m , with both occurrences of R ϵ R^\epsilon R ϵ replaced by R ϵ ( σ s ) R^\epsilon(\sigma s) R ϵ ( σ s ) . Applying this bound at σ = A − 1 \sigma=A^{-1} σ = A − 1 to the sum in Lemma 8.8 gives (8.9) at the current R R R , as in Section 8.1. Together with C K m ( R ) ≲ 1 C^m_K(R)\lesssim1 C K m ( R ) ≲ 1 from (8.10), this completes Step 1.
Why the Taylor order induction is needed
The proof of Proposition 8.9 has been reduced to Lemmas 8.10 and 8.11. We explain the Taylor order induction in Lemma 8.10, first for m = 0 m=0 m = 0 . Let g g g be a Schwartz function with supp g ^ ⊂ Γ d + 1 ( r ) \operatorname{supp}\widehat{g}\subset\Gamma^{d+1}(r) supp g ⊂ Γ d + 1 ( r ) . The integral to be bounded is
∫ ∑ θ ∣ g θ ∣ 2 ≳ α 2 ( ∑ θ ∣ g θ ∣ 2 ) p d / 2 , θ ∈ Ξ ( r − 1 / d ) . (104) \int_{\sum_\theta|g_\theta|^2\gtrsim\alpha^2}\left(\sum_\theta|g_\theta|^2\right)^{p_d/2},\qquad\theta\in\Xi(r^{-1/d}). \tag*{(104)} ∫ ∑ θ ∣ g θ ∣ 2 ≳ α 2 ( θ ∑ ∣ g θ ∣ 2 ) p d /2 , θ ∈ Ξ ( r − 1/ d ) . (104)
We split the Fourier transform of ∑ θ ∣ g θ ∣ 2 \sum_\theta|g_\theta|^2 ∑ θ ∣ g θ ∣ 2 at scales adapted to the averaging boxes on the right hand side of Lemma 8.10. The lowest frequencies will give its u = 1 u=1 u = 1 term. To see which coordinates must be cut off, compare the dimensions
θ ∗ : 1 × r 1 / d × ⋯ × r , U [ 0 , 1 ] , r : r 2 / d × ⋯ × r 2 / d . \theta^*:1\times r^{1/d}\times\cdots\times r,\qquad U_{[0,1],r}:r^{2/d}\times\cdots\times r^{2/d}. θ ∗ : 1 × r 1/ d × ⋯ × r , U [ 0 , 1 ] , r : r 2/ d × ⋯ × r 2/ d .
The locally constant property already gives averaging on scale r 2 / d r^{2/d} r 2/ d in directions 2 , … , d 2,\ldots,d 2 , … , d . To obtain the U [ 0 , 1 ] , r U_{[0,1],r} U [ 0 , 1 ] , r average, we therefore restrict the first two Taylor coordinates to width r − 2 / d r^{-2/d} r − 2/ d .
For an interval τ \tau τ of length u u u , the corresponding averaging box is
U τ , r : r 2 / d u 2 × r 2 / d u × ⋯ × r 2 / d u 2 − d . U_{\tau,r}:r^{2/d}u^2\times r^{2/d}u\times\cdots\times r^{2/d}u^{2-d}. U τ , r : r 2/ d u 2 × r 2/ d u × ⋯ × r 2/ d u 2 − d .
Write ξ = ∑ i λ i ϕ d ( i ) ( t θ ) = ∑ i λ ~ i ϕ d ( i ) ( t τ ) \xi=\sum_i\lambda_i\phi_d^{(i)}(t_\theta)=\sum_i\widetilde{\lambda}_i\phi_d^{(i)}(t_\tau) ξ = ∑ i λ i ϕ d ( i ) ( t θ ) = ∑ i λ i ϕ d ( i ) ( t τ ) . For θ ⊂ τ \theta\subset\tau θ ⊂ τ , put h = t θ − t τ h=t_\theta-t_\tau h = t θ − t τ . On supp ∣ g θ ^ ∣ 2 \operatorname{supp}|\widehat{g_\theta}|^2 supp ∣ g θ ∣ 2 , ∣ λ i ∣ ≲ r − i / d |\lambda_i|\lesssim r^{-i/d} ∣ λ i ∣ ≲ r − i / d , and Taylor expansion gives
λ ~ i − λ i = h i i ! λ 0 + h i − 1 ( i − 1 ) ! λ 1 + O ( r − 2 / d u i − 2 ) , 2 ≤ i ≤ d . \widetilde{\lambda}_i-\lambda_i=\frac{h^i}{i!}\lambda_0+\frac{h^{i-1}}{(i-1)!}\lambda_1+O(r^{-2/d}u^{i-2}),\qquad2\leq i\leq d. λ i − λ i = i ! h i λ 0 + ( i − 1 )! h i − 1 λ 1 + O ( r − 2/ d u i − 2 ) , 2 ≤ i ≤ d .
Thus the restrictions ∣ λ 0 ∣ ≲ r − 2 / d u − 2 |\lambda_0|\lesssim r^{-2/d}u^{-2} ∣ λ 0 ∣ ≲ r − 2/ d u − 2 and ∣ λ 1 ∣ ≲ r − 2 / d u − 1 |\lambda_1|\lesssim r^{-2/d}u^{-1} ∣ λ 1 ∣ ≲ r − 2/ d u − 1 keep the change in the i i i th coefficient within the i i i th width of U τ , r U_{\tau,r} U τ , r . Put σ = r − 1 / d / u \sigma=r^{-1/d}/u σ = r − 1/ d / u . Choose χ ∈ C c ∞ ( R ) \chi\in C^\infty_c(\mathbb{R}) χ ∈ C c ∞ ( R ) , 0 ≤ χ ≤ 1 0\leq\chi\leq1 0 ≤ χ ≤ 1 , with χ = 1 \chi=1 χ = 1 on [ − C , C ] [-C,C] [ − C , C ] , and set
η θ , ≤ σ 2 ( ξ ) = χ ( λ 0 σ 2 ) χ ( λ 1 r − 1 / d σ ) ∏ i = 2 d χ ( r i / d λ i ) , η θ , σ 2 = η θ , ≤ σ 2 − η θ , ≤ σ / 2 2 . \eta^2_{\theta,\leq\sigma}(\xi)=\chi\left(\frac{\lambda_0}{\sigma^2}\right)\chi\left(\frac{\lambda_1}{r^{-1/d}\sigma}\right)\prod_{i=2}^{d}\chi(r^{i/d}\lambda_i),\qquad\eta^2_{\theta,\sigma}=\eta^2_{\theta,\leq\sigma}-\eta^2_{\theta,\leq\sigma/2}. η θ , ≤ σ 2 ( ξ ) = χ ( σ 2 λ 0 ) χ ( r − 1/ d σ λ 1 ) i = 2 ∏ d χ ( r i / d λ i ) , η θ , σ 2 = η θ , ≤ σ 2 − η θ , ≤ σ /2 2 .
On supp ∣ g θ ^ ∣ 2 \operatorname{supp}|\widehat{g_\theta}|^2 supp ∣ g θ ∣ 2 , we have ∏ i = 2 d χ ( r i / d λ i ) = 1 \prod_{i=2}^{d}\chi(r^{i/d}\lambda_i)=1 ∏ i = 2 d χ ( r i / d λ i ) = 1 . Thus the factors with 2 ≤ i ≤ d 2\leq i\leq d 2 ≤ i ≤ d do not change ∣ g θ ∣ 2 ∗ η θ , ≤ σ 2 |g_\theta|^2*\eta^2_{\theta,\leq\sigma} ∣ g θ ∣ 2 ∗ η θ , ≤ σ 2 .
At σ = r − 1 / d \sigma=r^{-1/d} σ = r − 1/ d , the cutoff widths in directions 0 , 1 0,1 0 , 1 are both r − 2 / d r^{-2/d} r − 2/ d , and η θ , ≤ 1 2 = 1 \eta^2_{\theta,\leq1}=1 η θ , ≤ 1 2 = 1 on supp ∣ g θ ^ ∣ 2 \operatorname{supp}|\widehat{g_\theta}|^2 supp ∣ g θ ∣ 2 .
Dyadic pigeonholing now gives r − 1 / d < σ ≤ 1 r^{-1/d}<\sigma\leq1 r − 1/ d < σ ≤ 1 such that, with
z l o w = ∑ θ ∣ g θ ∣ 2 ∗ η ~ θ , ≤ r − 1 / d , z h i g h = ∑ θ ∣ g θ ∣ 2 ∗ η ~ θ , σ , z_{\mathrm{low}}=\sum_{\theta}|g_{\theta}|^{2}*\widetilde{\eta}_{\theta,\leq r^{-1/d}},\qquad z_{\mathrm{high}}=\sum_{\theta}|g_{\theta}|^{2}*\widetilde{\eta}_{\theta,\sigma}, z low = θ ∑ ∣ g θ ∣ 2 ∗ η θ , ≤ r − 1/ d , z high = θ ∑ ∣ g θ ∣ 2 ∗ η θ , σ ,
the integral (104) is bounded by
≲ ∫ ∣ z l o w ∣ ≳ α 2 ∣ z l o w ∣ p d / 2 + ∫ ∣ z h i g h ∣ ≳ α 2 ∣ z h i g h ∣ p d / 2 . (105) \lesssim\int_{|z_{\mathrm{low}}|\gtrsim\alpha^{2}}|z_{\mathrm{low}}|^{p_d/2}+\int_{|z_{\mathrm{high}}|\gtrsim\alpha^{2}}|z_{\mathrm{high}}|^{p_d/2}. \tag*{(105)} ≲ ∫ ∣ z low ∣ ≳ α 2 ∣ z low ∣ p d /2 + ∫ ∣ z high ∣ ≳ α 2 ∣ z high ∣ p d /2 . (105)
For the first integral, the locally constant property, Hölder’s inequality, and (36) give
∫ ∣ z l o w ∣ ≳ α 2 ∣ z l o w ∣ p d / 2 ≲ ∑ U ∥ U [ 0 , 1 ] , r U ∈ G [ 0 , 1 ] ( α ) ∣ U ∣ ∥ ( ∑ θ ∣ g θ ∣ 2 ) 1 / 2 ∥ L # p d / 2 ( U ) p d . (106) \int_{|z_{\mathrm{low}}|\gtrsim\alpha^{2}}|z_{\mathrm{low}}|^{p_d/2}\lesssim\sum_{\substack{U\parallel U_{[0,1],r}\\ U\in\mathcal{G}_{[0,1]}(\alpha)}}|U|\left\|\left(\sum_{\theta}|g_{\theta}|^{2}\right)^{1/2}\right\|_{L_{\#}^{p_d/2}(U)}^{p_d}. \tag*{(106)} ∫ ∣ z low ∣ ≳ α 2 ∣ z low ∣ p d /2 ≲ U ∥ U [ 0 , 1 ] , r U ∈ G [ 0 , 1 ] ( α ) ∑ ∣ U ∣ ( θ ∑ ∣ g θ ∣ 2 ) 1/2 L # p d /2 ( U ) p d . (106)
This is the u = 1 u=1 u = 1 term in Lemma 8.10.
We next show why estimating the z h i g h z_{\mathrm{high}} z high term in (105) leads to the Taylor cone of order one. On supp ∣ g θ ∣ 2 ∗ η ~ θ , σ \operatorname{supp}|g_{\theta}|^{2}*\widetilde{\eta}_{\theta,\sigma} supp ∣ g θ ∣ 2 ∗ η θ , σ , the coefficients satisfy
∣ λ 0 ∣ ≲ σ 2 , ∣ λ 1 ∣ ≲ r − 1 / d σ , max { σ − 2 ∣ λ 0 ∣ , r 1 / d σ − 1 ∣ λ 1 ∣ } ∼ 1. (107) |\lambda_{0}|\lesssim\sigma^{2},\qquad|\lambda_{1}|\lesssim r^{-1/d}\sigma,\qquad\max\{\sigma^{-2}|\lambda_{0}|,r^{1/d}\sigma^{-1}|\lambda_{1}|\}\sim1. \tag*{(107)} ∣ λ 0 ∣ ≲ σ 2 , ∣ λ 1 ∣ ≲ r − 1/ d σ , max { σ − 2 ∣ λ 0 ∣ , r 1/ d σ − 1 ∣ λ 1 ∣ } ∼ 1. (107)
Group the θ \theta θ ’s in intervals τ \tau τ of length u = r − 1 / d σ − 1 u=r^{-1/d}\sigma^{-1} u = r − 1/ d σ − 1 , so that z h i g h = ∑ τ ∑ θ ⊂ τ ∣ g θ ∣ 2 ∗ η ~ θ , σ z_{\mathrm{high}}=\sum_{\tau}\sum_{\theta\subset\tau}|g_{\theta}|^{2}*\widetilde{\eta}_{\theta,\sigma} z high = ∑ τ ∑ θ ⊂ τ ∣ g θ ∣ 2 ∗ η θ , σ . By (107) , frequency dilation by σ − 2 \sigma^{-2} σ − 2 sends each inner sum to a cap with coefficients
∣ λ i ∣ ≲ ( r σ d ) − i / d ( 0 ≤ i ≤ d ) , max { ∣ λ 0 ∣ , ( r σ d ) 1 / d ∣ λ 1 ∣ } ∼ 1. (108) |\lambda_i|\lesssim(r\sigma^d)^{-i/d}\ (0\leq i\leq d),\qquad\max\{|\lambda_0|,(r\sigma^d)^{1/d}|\lambda_1|\}\sim1. \tag*{(108)} ∣ λ i ∣ ≲ ( r σ d ) − i / d ( 0 ≤ i ≤ d ) , max { ∣ λ 0 ∣ , ( r σ d ) 1/ d ∣ λ 1 ∣ } ∼ 1. (108)
On the cone over the moment curve in Proposition 8.1 , the radial coefficient satisfies ∣ λ 0 ∣ ∼ 1 |\lambda_0|\sim1 ∣ λ 0 ∣ ∼ 1 . But in (108) , ∣ λ 0 ∣ |\lambda_0| ∣ λ 0 ∣ can be small, in which case ( r σ d ) 1 / d ∣ λ 1 ∣ ∼ 1 (r\sigma^d)^{1/d}|\lambda_1|\sim1 ( r σ d ) 1/ d ∣ λ 1 ∣ ∼ 1 . This is precisely the additional support allowed by Ξ 1 d + 1 ( r σ d ) \Xi_1^{d+1}(r\sigma^d) Ξ 1 d + 1 ( r σ d ) in (8.2). More explicitly,
σ − 2 supp ( ∑ θ ⊂ τ ∣ g θ ∣ 2 ∗ η ~ θ , σ ) ⊂ ω τ ∈ Ξ 1 d + 1 ( r σ d ) . (109) \sigma^{-2}\operatorname{supp}\left(\sum_{\theta\subset\tau}|g_{\theta}|^{2}*\widetilde{\eta}_{\theta,\sigma}\right)\subset\omega_{\tau}\in\Xi_1^{d+1}(r\sigma^d). \tag*{(109)} σ − 2 supp ( θ ⊂ τ ∑ ∣ g θ ∣ 2 ∗ η θ , σ ) ⊂ ω τ ∈ Ξ 1 d + 1 ( r σ d ) . (109)
where ω τ \omega_\tau ω τ is the cap corresponding to τ \tau τ .
In general, for supp g ^ ⊂ Γ m d ( r ) \operatorname{supp}\widehat{g}\subset\Gamma_m^d(r) supp g ⊂ Γ m d ( r ) , the low frequency integral must give the average over r m / d U [ 0 , 1 ] , r r^{m/d}U_{[0,1],r} r m / d U [ 0 , 1 ] , r , while, for m < d − 2 m<d-2 m < d − 2 , the high frequency support must lie at Taylor order m + 1 m+1 m + 1 . We therefore define
η θ , ≤ σ m + 2 ( ξ ) = ∏ i = 0 m + 1 χ ( λ i r − i / d σ m + 2 − i ) ∏ i = m + 2 d χ ( r i / d λ i ) , η θ , σ m + 2 = η θ , ≤ σ m + 2 − η θ , ≤ σ / 2 m + 2 . \eta_{\theta,\leq\sigma}^{m+2}(\xi)=\prod_{i=0}^{m+1}\chi\left(\frac{\lambda_i}{r^{-i/d}\sigma^{m+2-i}}\right)\prod_{i=m+2}^{d}\chi(r^{i/d}\lambda_i),\qquad\eta_{\theta,\sigma}^{m+2}=\eta_{\theta,\leq\sigma}^{m+2}-\eta_{\theta,\leq\sigma/2}^{m+2}. η θ , ≤ σ m + 2 ( ξ ) = i = 0 ∏ m + 1 χ ( r − i / d σ m + 2 − i λ i ) i = m + 2 ∏ d χ ( r i / d λ i ) , η θ , σ m + 2 = η θ , ≤ σ m + 2 − η θ , ≤ σ /2 m + 2 .
On supp ∣ g θ ∣ 2 \operatorname{supp}|g_{\theta}|^{2} supp ∣ g θ ∣ 2 , we have ∏ i = m + 2 d χ ( r i / d λ i ) = 1 \prod_{i=m+2}^{d}\chi(r^{i/d}\lambda_i)=1 ∏ i = m + 2 d χ ( r i / d λ i ) = 1 . Thus the factors with m + 2 ≤ i ≤ d m+2\leq i\leq d m + 2 ≤ i ≤ d do not change ∣ g θ ∣ 2 ∗ η ~ θ , ≤ σ m + 2 |g_{\theta}|^{2}*\widetilde{\eta}_{\theta,\leq\sigma}^{m+2} ∣ g θ ∣ 2 ∗ η θ , ≤ σ m + 2 . For dyadic r − 1 / d < σ ≤ 1 r^{-1/d}<\sigma\leq1 r − 1/ d < σ ≤ 1 , set
z l o w m = ∑ θ ∈ Ξ m d + 1 ( r ) ∣ g θ ∣ 2 ∗ η ~ θ , ≤ r − 1 / d m + 2 , z h i g h m = ∑ θ ∈ Ξ m d + 1 ( r ) ∣ g θ ∣ 2 ∗ η ~ θ , σ m + 2 . z_{\mathrm{low}}^{m}=\sum_{\theta\in\Xi_m^{d+1}(r)}|g_{\theta}|^{2}*\widetilde{\eta}_{\theta,\leq r^{-1/d}}^{m+2},\qquad z_{\mathrm{high}}^{m}=\sum_{\theta\in\Xi_m^{d+1}(r)}|g_{\theta}|^{2}*\widetilde{\eta}_{\theta,\sigma}^{m+2}. z low m = θ ∈ Ξ m d + 1 ( r ) ∑ ∣ g θ ∣ 2 ∗ η θ , ≤ r − 1/ d m + 2 , z high m = θ ∈ Ξ m d + 1 ( r ) ∑ ∣ g θ ∣ 2 ∗ η θ , σ m + 2 .
As in (105) , dyadic pigeonholing reduces the estimate in Lemma 8.10 to two integrals with exponent p d − m / 2 p_{d-m}/2 p d − m /2 , one for z l o w m z_{\mathrm{low}}^{m} z low m and one for z h i g h m z_{\mathrm{high}}^{m} z high m at a single dyadic scale σ \sigma σ .
For the low frequency integral, the cutoff at σ = r − 1 / d \sigma=r^{-1/d} σ = r − 1/ d has width r − ( m + 2 ) / d r^{-(m+2)/d} r − ( m + 2 ) / d in directions 0 , … , m + 1 0,\ldots,m+1 0 , … , m + 1 . By the locally constant property at scale r ( m + 2 ) / d r^{(m+2)/d} r ( m + 2 ) / d in the remaining directions, we can apply the argument of (106) to z l o w m z_{\mathrm{low}}^{m} z low m , with p d − m p_{d-m} p d − m in place of p d p_d p d and boxes U ∥ r m / d U [ 0 , 1 ] , r U\parallel r^{m/d}U_{[0,1],r} U ∥ r m / d U [ 0 , 1 ] , r . This gives the u = 1 u=1 u = 1 term on the right hand side of Lemma 8.10.
For the high frequency integral, the same Taylor calculation as in (109) gives
σ − ( m + 2 ) supp ( ∑ θ ⊂ τ ∣ g θ ∣ 2 ∗ η ~ θ , σ m + 2 ) ⊂ ω τ ∈ Ξ m + 1 d + 1 ( r σ d ) , 0 ≤ m ≤ d − 3. (110) \sigma^{-(m+2)}\operatorname{supp}\left(\sum_{\theta\subset\tau}|g_{\theta}|^{2}*\widetilde{\eta}_{\theta,\sigma}^{m+2}\right)\subset\omega_{\tau}\in\Xi_{m+1}^{d+1}(r\sigma^d),\qquad0\leq m\leq d-3. \tag*{(110)} σ − ( m + 2 ) supp ( θ ⊂ τ ∑ ∣ g θ ∣ 2 ∗ η θ , σ m + 2 ) ⊂ ω τ ∈ Ξ m + 1 d + 1 ( r σ d ) , 0 ≤ m ≤ d − 3. (110)
Thus (93) , with m + 1 m+1 m + 1 in place of m m m , applies to the sum of ( ∑ θ ⊂ τ ∣ g θ ∣ 2 ∗ η ~ θ , σ m + 2 ) ( σ − ( m + 2 ) x ) \left(\sum_{\theta\subset\tau}|g_{\theta}|^{2}*\widetilde{\eta}_{\theta,\sigma}^{m+2}\right)(\sigma^{-(m+2)}x) ( ∑ θ ⊂ τ ∣ g θ ∣ 2 ∗ η θ , σ m + 2 ) ( σ − ( m + 2 ) x ) , with parameter r σ d r\sigma^d r σ d . Its exponent is p d − m − 1 p_{d-m-1} p d − m − 1 , whereas the high frequency integral has exponent p d − m / 2 p_{d-m}/2 p d − m /2 . Section 9 performs the wave packet interpolation between these exponents and first bounds the integral in Lemma 8.10 by the right hand side of (93) for the function g g g at parameter r r r , before passing to the sum over U U U in Lemma 8.10.
At m = d − 2 m=d-2 m = d − 2 , the exponent of the high frequency integral is p 2 / 2 = 2 p_2/2=2 p 2 /2 = 2 , and its proof uses L 2 L^2 L 2 orthogonality instead.
The averaging box ( u r 1 / d ) m U τ , r (ur^{1/d})^mU_{\tau,r} ( u r 1/ d ) m U τ , r also follows from rescaling the u = 1 u=1 u = 1 averaging box. For τ = [ t τ , t τ + u ] \tau=[t_\tau,t_\tau+u] τ = [ t τ , t τ + u ] , rescale g τ g_\tau g τ to g τ ( L t τ , u − T x ) g_\tau(L_{t_\tau,u}^{-T}x) g τ ( L t τ , u − T x ) , with parameter r u d ru^d r u d . Returning the box ( r u d ) m / d U [ 0 , 1 ] , r u d (ru^d)^{m/d}U_{[0,1],ru^d} ( r u d ) m / d U [ 0 , 1 ] , r u d multiplies its i i i th width by u − i u^{-i} u − i , giving
r ( m + 2 ) / d u m + 2 × r ( m + 2 ) / d u m + 1 × ⋯ × r ( m + 2 ) / d u m + 2 − d . r^{(m+2)/d}u^{m+2}\times r^{(m+2)/d}u^{m+1}\times\cdots\times r^{(m+2)/d}u^{m+2-d}. r ( m + 2 ) / d u m + 2 × r ( m + 2 ) / d u m + 1 × ⋯ × r ( m + 2 ) / d u m + 2 − d .
These are the dimensions of ( u r 1 / d ) m U τ , r (ur^{1/d})^mU_{\tau,r} ( u r 1/ d ) m U τ , r .
Roadmap: Proof of Proposition 8.7, Step 2
In this subsection, we reduce Step 2 of Proposition 8.7 to Proposition 8.15, which we prove in Section 10 by adapting the induction on σ s \sigma_s σ s from Step 1.
Fix 0 ≤ m < d − 2 0\le m<d-2 0 ≤ m < d − 2 . Under the hypotheses of Proposition 8.7, Step 1 gives C K m ( R ) ≲ 1 C_K^m(R)\lesssim1 C K m ( R ) ≲ 1 and (8.9) for this m m m . We will prove that for R ≥ 1 R\ge1 R ≥ 1 , ϵ > 0 \epsilon>0 ϵ > 0 , and α > 0 \alpha>0 α > 0 , every Schwartz function f f f with supp f ^ ⊂ Γ m d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\Gamma_m^{d+1}(R) supp f ⊂ Γ m d + 1 ( R ) satisfies
∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ ϵ RHS of (8.8) ≲ RHS of (8.10) . (111) \int_{\lvert f\rvert>\alpha}\lvert f\rvert^{p_{d-m}}\lesssim_\epsilon\operatorname{RHS\ of\ (8.8)}\lesssim\operatorname{RHS\ of\ (8.10)}. \tag*{(111)} ∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ ϵ RHS of (8.8) ≲ RHS of (8.10) . (111)
The last inequality follows from Lemma 8.5, so it suffices to prove the first inequality. We do this successively in the order j = 0 , … , m j=0,\ldots,m j = 0 , … , m , keeping the exponent p d − m p_{d-m} p d − m . The case j = 0 j=0 j = 0 follows directly from Step 1 after the coefficient decomposition and dilation below. Taking j = m j=m j = m will then give C m ( R ) ≲ 1 C^m(R)\lesssim1 C m ( R ) ≲ 1 and (8.8) for the current m m m .
Let f f f be a Schwartz function with supp f ^ ⊂ Γ m d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\Gamma_m^{d+1}(R) supp f ⊂ Γ m d + 1 ( R ) . For ξ ∈ supp f θ I ^ \xi\in\operatorname{supp}\widehat{f_{\theta_I}} ξ ∈ supp f θ I , write ξ = ∑ i = 0 d λ i ϕ d ( i ) ( t I ) \xi=\sum_{i=0}^{d}\lambda_i\phi_d^{(i)}(t_I) ξ = ∑ i = 0 d λ i ϕ d ( i ) ( t I ) . The condition defining Γ m d + 1 ( R ) \Gamma_m^{d+1}(R) Γ m d + 1 ( R ) gives ∣ λ j ∣ ∼ R − j / d \lvert\lambda_j\rvert\sim R^{-j/d} ∣ λ j ∣ ∼ R − j / d for some j ≤ m j\le m j ≤ m . Using a smooth partition of unity in frequency space, we obtain
f θ = ∑ j = 0 m ∑ ν ∈ Z ∣ ν ∣ ∼ K ∑ R − 1 / d ≤ ℓ ≤ 1 f θ j , ν , ℓ , f_\theta=\sum_{j=0}^{m}\sum_{\nu\in\mathbb{Z}\atop\lvert\nu\rvert\sim K}\sum_{R^{-1/d}\le\ell\le1}f_\theta^{j,\nu,\ell}, f θ = j = 0 ∑ m ∣ ν ∣ ∼ K ν ∈ Z ∑ R − 1/ d ≤ ℓ ≤ 1 ∑ f θ j , ν , ℓ ,
where ℓ \ell ℓ is dyadic and
supp f θ j , ν , ℓ ^ ⊂ { ∑ i = 0 d λ i ϕ d ( i ) ( t I ) : ∣ λ i ∣ ≲ R − i / d ( 0 ≤ i ≤ d ) , ∣ R j / d λ j − ν / K ∣ ≲ K − 1 , max { 1 , ( R j / d ∣ λ i ∣ ) 1 / ( j − i ) : 0 ≤ i < j } ∼ ℓ − 1 } . \operatorname{supp}\widehat{f_\theta^{j,\nu,\ell}}\subset\left\{\sum_{i=0}^{d}\lambda_i\phi_d^{(i)}(t_I):
\begin{array}{l}
\lvert\lambda_i\rvert\lesssim R^{-i/d}\quad(0\le i\le d),\\
\lvert R^{j/d}\lambda_j-\nu/K\rvert\lesssim K^{-1},\\
\max\{1,(R^{j/d}\lvert\lambda_i\rvert)^{1/(j-i)}:0\le i<j\}\sim\ell^{-1}
\end{array}
\right\}. supp f θ j , ν , ℓ ⊂ ⎩ ⎨ ⎧ i = 0 ∑ d λ i ϕ d ( i ) ( t I ) : ∣ λ i ∣ ≲ R − i / d ( 0 ≤ i ≤ d ) , ∣ R j / d λ j − ν / K ∣ ≲ K − 1 , max { 1 , ( R j / d ∣ λ i ∣ ) 1/ ( j − i ) : 0 ≤ i < j } ∼ ℓ − 1 ⎭ ⎬ ⎫ .
By Lemma 2.18, we may assume that each family { f θ j , ν , ℓ } θ ∈ Ξ m d + 1 ( R ) \{f_\theta^{j,\nu,\ell}\}_{\theta\in\Xi_m^{d+1}(R)} { f θ j , ν , ℓ } θ ∈ Ξ m d + 1 ( R ) satisfies the counting inequalities (2.28). Dyadic pigeonholing then gives fixed j , ν , ℓ j,\nu,\ell j , ν , ℓ such that, putting f ~ θ = f θ j , ν , ℓ \widetilde f_\theta=f_\theta^{j,\nu,\ell} f θ = f θ j , ν , ℓ and f ~ = ∑ θ ∈ Ξ m d + 1 ( R ) f ~ θ \widetilde f=\sum_{\theta\in\Xi_m^{d+1}(R)}\widetilde f_\theta f = ∑ θ ∈ Ξ m d + 1 ( R ) f θ ,
∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ ∫ ∣ f ~ ∣ ≳ α ∣ f ~ ∣ p d − m . (112) \int_{\lvert f\rvert>\alpha}\lvert f\rvert^{p_{d-m}}\lesssim\int_{\lvert\widetilde f\rvert\gtrsim\alpha}\lvert\widetilde f\rvert^{p_{d-m}}. \tag*{(112)} ∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ ∫ ∣ f ∣ ≳ α ∣ f ∣ p d − m . (112)
Since ∣ ν ∣ ∼ K \lvert\nu\rvert\sim K ∣ ν ∣ ∼ K , replacing each f ~ θ ( x ) \widetilde f_\theta(x) f θ ( x ) by f ~ θ ( ( K / ν ) x ) \widetilde f_\theta((K/\nu)x) f θ (( K / ν ) x ) , we may assume ν = K \nu=K ν = K . Thus supp f ~ ^ ⊂ Γ j , 1 / K , ℓ d + 1 ( R ) \operatorname{supp}\widehat{\widetilde f}\subset\Gamma_{j,1/K,\ell}^{d+1}(R) supp f ⊂ Γ j , 1/ K , ℓ d + 1 ( R ) , where
Γ j , 1 / K , ℓ d + 1 ( R ) : = ⋃ I ∈ J ( R − 1 / d ) { ∑ i = 0 d λ i ϕ d ( i ) ( t I ) : ∣ λ i ∣ ≲ R − i / d ( 0 ≤ i ≤ d ) , ∣ λ j − R − j / d ∣ ≲ K − 1 R − j / d , max { 1 , ( R j / d ∣ λ i ∣ ) 1 / ( j − i ) : 0 ≤ i < j } ∼ ℓ − 1 } . (113) \Gamma_{j,1/K,\ell}^{d+1}(R):=\bigcup_{I\in\mathcal{J}(R^{-1/d})}\left\{\sum_{i=0}^{d}\lambda_i\phi_d^{(i)}(t_I):
\begin{array}{l}
\lvert\lambda_i\rvert\lesssim R^{-i/d}\quad(0\le i\le d),\\
\lvert\lambda_j-R^{-j/d}\rvert\lesssim K^{-1}R^{-j/d},\\
\max\{1,(R^{j/d}\lvert\lambda_i\rvert)^{1/(j-i)}:0\le i<j\}\sim\ell^{-1}
\end{array}
\right\}. \tag*{(113)} Γ j , 1/ K , ℓ d + 1 ( R ) := I ∈ J ( R − 1/ d ) ⋃ ⎩ ⎨ ⎧ i = 0 ∑ d λ i ϕ d ( i ) ( t I ) : ∣ λ i ∣ ≲ R − i / d ( 0 ≤ i ≤ d ) , ∣ λ j − R − j / d ∣ ≲ K − 1 R − j / d , max { 1 , ( R j / d ∣ λ i ∣ ) 1/ ( j − i ) : 0 ≤ i < j } ∼ ℓ − 1 ⎭ ⎬ ⎫ . (113)
We will use induction for this j j j since, after dilation, supp f ~ ^ ( R j / d ℓ d ⋅ ) ⊂ Γ j − 1 d + 1 ( ℓ − d ) \operatorname{supp}\widehat{\widetilde f}(R^{j/d}\ell^{\,d}\,\cdot)\subset\Gamma_{j-1}^{d+1}(\ell^{-d}) supp f ( R j / d ℓ d ⋅ ) ⊂ Γ j − 1 d + 1 ( ℓ − d ) , as we verify below. By the following lemma and (8.27), to prove (8.26) for the original f f f , it suffices to show that, for ϵ > 0 \epsilon>0 ϵ > 0 and α > 0 \alpha>0 α > 0 ,
∫ ∣ f ~ ∣ > α ∣ f ~ ∣ p d − m ≲ ϵ RHS of (8.8) with f ~ in place of f . \int_{\lvert\widetilde f\rvert>\alpha}\lvert\widetilde f\rvert^{p_{d-m}}\lesssim_\epsilon\operatorname{RHS\ of\ (8.8)\ with\ }\widetilde f\operatorname{\ in\ place\ of\ }f. ∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ ϵ RHS of (8.8) with f in place of f .
Lemma 8.13. Let 0 ≤ m < d − 2 0\le m<d-2 0 ≤ m < d − 2 , R ≥ 1 R\ge1 R ≥ 1 , and let f f f be a Schwartz function with supp f ^ ⊂ Γ m d + 1 ( R ) \operatorname{supp}\widehat f\subset\Gamma_m^{d+1}(R) supp f ⊂ Γ m d + 1 ( R ) . Let f ~ \widetilde f f be obtained from f f f by the coefficient decomposition, uniformization, and dilation above. For ϵ > 0 \epsilon>0 ϵ > 0 and α > 0 \alpha>0 α > 0 ,
RHS of (8.8) for f ~ ≲ ϵ RHS of (8.8) for f . (114) \operatorname{RHS\ of\ (8.8)\ for\ }\widetilde f\lesssim_\epsilon\operatorname{RHS\ of\ (8.8)\ for\ }f. \tag*{(114)} RHS of (8.8) for f ≲ ϵ RHS of (8.8) for f . (114)
We defer the proof to the end of this section.
In summary, assuming (8.26) for every j ′ < j j'<j j ′ < j , it suffices to prove the following proposition.
Proposition 8.14. Fix 0 ≤ m < d − 2 0 \le m < d-2 0 ≤ m < d − 2 and 0 ≤ j ≤ m 0 \le j \le m 0 ≤ j ≤ m . Assume the hypotheses of Proposition 8.7, together with C K m ( R ) ≈ 1 C_{\mathcal{K}}^{m}(R) \approx1 C K m ( R ) ≈ 1 and (8.9) for this m m m , and suppose that (8.26) holds for every j ′ < j j' < j j ′ < j . Let R ≥ 1 R \ge1 R ≥ 1 , let R − 1 / d ≤ ℓ ≤ 1 R^{-1/d} \le\ell\le1 R − 1/ d ≤ ℓ ≤ 1 be dyadic, and let f f f be a Schwartz function with supp f ^ ⊂ Γ j , 1 / K , ℓ d + 1 ( R ) \operatorname{supp} \widehat{f} \subset\Gamma_{j,1/K,\ell}^{d+1}(R) supp f ⊂ Γ j , 1/ K , ℓ d + 1 ( R ) . Then, for ϵ > 0 \epsilon> 0 ϵ > 0 and α > 0 \alpha> 0 α > 0 ,
∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ ϵ RHS of ( 8.8 ) . \int_{\lvert f\rvert>\alpha} \lvert f\rvert^{p_{d-m}} \lesssim_{\epsilon} \operatorname{RHS\ of\ }(8.8). ∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ ϵ RHS of ( 8.8 ) .
We now prove Proposition 8.14. Let f f f be a Schwartz function with supp f ^ ⊂ Γ j , 1 / K , ℓ d + 1 ( R ) \operatorname{supp} \widehat{f} \subset\Gamma_{j,1/K,\ell}^{d+1}(R) supp f ⊂ Γ j , 1/ K , ℓ d + 1 ( R ) . We first consider ℓ ∼ 1 \ell\sim1 ℓ ∼ 1 . In this case, the lower coefficients satisfy ∣ λ i ∣ ≲ R − j / d \lvert\lambda_i\rvert\lesssim R^{-j/d} ∣ λ i ∣ ≲ R − j / d for i < j i<j i < j , whereas the support hypothesis in (8.9) requires an additional factor K − 1 K^{-1} K − 1 . We obtain this factor by Taylor rescaling on sufficiently short caps.
Choose a ≤ A − 1 a \le A^{-1} a ≤ A − 1 sufficiently small in terms of K K K . For τ ∈ J ( a ) \tau\in\mathcal{J}(a) τ ∈ J ( a ) , let t τ t_\tau t τ denote the left endpoint of τ \tau τ . Taylor rescaling by L t τ , a − T L_{t_\tau,a}^{-T} L t τ , a − T , which sends τ \tau τ to a cap over the unit interval, gives parameter R a d Ra^d R a d and
∣ a − i λ i ∣ ( R a d ) − j / d = a j − i R j / d ∣ λ i ∣ ≲ a j − i ≤ K − 1 , i < j . \frac{\lvert a^{-i}\lambda_i\rvert}{(Ra^d)^{-j/d}}=a^{j-i}R^{j/d}\lvert\lambda_i\rvert\lesssim a^{j-i}\le K^{-1},\qquad i<j. ( R a d ) − j / d ∣ a − i λ i ∣ = a j − i R j / d ∣ λ i ∣ ≲ a j − i ≤ K − 1 , i < j .
Thus supp f τ ( L t τ , a − T ⋅ ) ⊂ Γ j , 1 / K d + 1 ( R a d ) \operatorname{supp} f_\tau(L_{t_\tau,a}^{-T}\cdot)\subset\Gamma_{j,1/K}^{d+1}(Ra^d) supp f τ ( L t τ , a − T ⋅ ) ⊂ Γ j , 1/ K d + 1 ( R a d ) . Hölder and (8.9), applied to these functions at parameter R a d Ra^d R a d and amplitude α / N ( a ) \alpha/N(a) α / N ( a ) , give, on returning to the original coordinates,
∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ ∑ τ ∈ J ( a ) ∫ ∣ f τ ∣ ≳ α / N ( a ) ∣ f τ ∣ p d − m ≲ ϵ RHS of ( 8.8 ) for f . (115) \int_{\lvert f\rvert>\alpha}\lvert f\rvert^{p_{d-m}}\lesssim\sum_{\tau\in\mathcal{J}(a)}\int_{\lvert f_\tau\rvert\gtrsim\alpha/N(a)}\lvert f_\tau\rvert^{p_{d-m}}\lesssim_{\epsilon}\operatorname{RHS\ of\ }(8.8)\text{ for }f. \tag*{(115)} ∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ τ ∈ J ( a ) ∑ ∫ ∣ f τ ∣ ≳ α / N ( a ) ∣ f τ ∣ p d − m ≲ ϵ RHS of ( 8.8 ) for f . (115)
This also treats the initial index j = 0 j=0 j = 0 , since only ℓ ∼ 1 \ell\sim1 ℓ ∼ 1 occurs in that case.
Suppose now that ℓ ≪ 1 \ell\ll1 ℓ ≪ 1 . For τ ∈ J ( ℓ ) \tau\in\mathcal{J}(\ell) τ ∈ J ( ℓ ) , Taylor expansion at t τ t_\tau t τ and the conditions defining Γ j , 1 / K , ℓ d + 1 ( R ) \Gamma_{j,1/K,\ell}^{d+1}(R) Γ j , 1/ K , ℓ d + 1 ( R ) give
supp f τ ( R j / d ℓ − 1 ⋅ ) ⊂ { ∑ i = 0 d λ i Φ d ( i ) ( t τ ) : ∣ λ i ∣ ≲ ℓ i , 0 ≤ i ≤ d , max i < j ℓ − i ∣ λ i ∣ ≳ 1 } . (116) \operatorname{supp} f_\tau(R^{j/d}\ell^{-1}\cdot)\subset\left\{\sum_{i=0}^{d}\lambda_i\Phi_d^{(i)}(t_\tau):\ \lvert\lambda_i\rvert\lesssim\ell^i,\ 0\le i\le d,\ \max_{i<j}\ell^{-i}\lvert\lambda_i\rvert\gtrsim1\right\}. \tag*{(116)} supp f τ ( R j / d ℓ − 1 ⋅ ) ⊂ { i = 0 ∑ d λ i Φ d ( i ) ( t τ ) : ∣ λ i ∣ ≲ ℓ i , 0 ≤ i ≤ d , i < j max ℓ − i ∣ λ i ∣ ≳ 1 } . (116)
Thus supp f ( R j / d ℓ − 1 ⋅ ) ⊂ Γ j − 1 d + 1 ( ℓ − d ) \operatorname{supp} f(R^{j/d}\ell^{-1}\cdot)\subset\Gamma_{j-1}^{d+1}(\ell^{-d}) supp f ( R j / d ℓ − 1 ⋅ ) ⊂ Γ j − 1 d + 1 ( ℓ − d ) . Apply (8.10), available by the induction hypothesis at j − 1 j-1 j − 1 , to f ( R j / d ℓ − 1 x ) f(R^{j/d}\ell^{-1}x) f ( R j / d ℓ − 1 x ) , with parameter ℓ − d \ell^{-d} ℓ − d and amplitude α \alpha α . Returning to the original coordinates and applying (2.22) to the sums over averaging boxes on the RHS of (8.10), with F = ( ∑ τ ⊂ ρ , ∣ τ ∣ = ℓ ∣ f τ ∣ 2 ) p d − m / 4 F=\left(\sum_{\tau\subset\rho,\ \lvert\tau\rvert=\ell}\lvert f_\tau\rvert^2\right)^{p_{d-m}/4} F = ( ∑ τ ⊂ ρ , ∣ τ ∣ = ℓ ∣ f τ ∣ 2 ) p d − m /4 and t = [ α / ( R ϵ N ( s ) ) ] p d − m / 2 t=[\alpha/(R^\epsilon N(s))]^{p_{d-m}/2} t = [ α / ( R ϵ N ( s )) ] p d − m /2 , followed by the locally constant property on Q ∥ ρ Q\parallel\rho Q ∥ ρ ,
∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ R ϵ ∑ ℓ ≤ s ≤ 1 ∑ ρ ∈ J ( s ) ∑ Q ∈ G ρ ( α / R ϵ ) ∣ Q ∣ ( ∑ τ ⊂ ρ ∣ τ ∣ = ℓ ∣ f τ ∣ 2 ) 1 / 2 ∣ p d − m ∥ F ∥ L # p d − m / 2 ( Q ) . (117) \int_{\lvert f\rvert>\alpha}\lvert f\rvert^{p_{d-m}}\lesssim R^\epsilon\sum_{\ell\le s\le1}\sum_{\rho\in\mathcal{J}(s)}\sum_{Q\in\mathcal{G}_\rho(\alpha/R^\epsilon)}\lvert Q\rvert\left(\sum_{\substack{\tau\subset\rho\\\lvert\tau\rvert=\ell}}\lvert f_\tau\rvert^2\right)^{1/2\lvert p_{d-m}}\lVert F\rVert_{L_\#^{p_{d-m}/2}(Q)}. \tag*{(117)} ∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ R ϵ ℓ ≤ s ≤ 1 ∑ ρ ∈ J ( s ) ∑ Q ∈ G ρ ( α / R ϵ ) ∑ ∣ Q ∣ τ ⊂ ρ ∣ τ ∣ = ℓ ∑ ∣ f τ ∣ 2 1/2 ∣ p d − m ∥ F ∥ L # p d − m /2 ( Q ) . (117)
We next subdivide the ℓ \ell ℓ caps so that either the pieces are the original f θ f_\theta f θ , or the hypothesis (8.9) for the current m m m applies after Taylor rescaling. For this purpose, set σ = max { R − 1 / d , a ℓ } \sigma=\max\{R^{-1/d},a\ell\} σ = max { R − 1/ d , a ℓ } . When σ > R − 1 / d \sigma>R^{-1/d} σ > R − 1/ d , the support condition needed to apply (8.9) is
supp f τ ( L t τ , σ − T ⋅ ) ⊂ Γ j , 1 / K d + 1 ( R σ d ) , τ ∈ J ( σ ) . (118) \operatorname{supp} f_\tau(L_{t_\tau,\sigma}^{-T}\cdot)\subset\Gamma_{j,1/K}^{d+1}(R\sigma^d),\qquad\tau\in\mathcal{J}(\sigma). \tag*{(118)} supp f τ ( L t τ , σ − T ⋅ ) ⊂ Γ j , 1/ K d + 1 ( R σ d ) , τ ∈ J ( σ ) . (118)
This follows from the j j j th coefficient condition in the definition of Γ j , 1 / K , ℓ d + 1 ( R ) \Gamma_{j,1/K,\ell}^{d+1}(R) Γ j , 1/ K , ℓ d + 1 ( R ) and
∣ σ − i λ i ∣ ( R σ d ) − j / d ≲ ( σ / ℓ ) j − i ≤ K − 1 \frac{\lvert\sigma^{-i}\lambda_i\rvert}{(R\sigma^d)^{-j/d}}\lesssim(\sigma/\ell)^{j-i}\le K^{-1} ( R σ d ) − j / d ∣ σ − i λ i ∣ ≲ ( σ / ℓ ) j − i ≤ K − 1
for i < j i<j i < j .
In either case, each τ ′ ∈ J ( ℓ ) \tau'\in\mathcal{J}(\ell) τ ′ ∈ J ( ℓ ) contains ℓ / σ ≤ a − 1 \ell/\sigma\le a^{-1} ℓ / σ ≤ a − 1 caps of width σ \sigma σ . Cauchy–Schwarz therefore gives
∣ f τ ′ ∣ 2 ≲ a ∑ τ ⊂ τ ′ ∣ τ ∣ = σ ∣ f τ ∣ 2 , \lvert f_{\tau'}\rvert^2\lesssim_a\sum_{\substack{\tau\subset\tau'\\\lvert\tau\rvert=\sigma}}\lvert f_\tau\rvert^2, ∣ f τ ′ ∣ 2 ≲ a τ ⊂ τ ′ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ,
so the RHS of (8.32) is bounded by
≲ R ϵ ∑ ℓ ≤ s ≤ 1 ∑ ρ ∈ J ( s ) ∑ Q ∈ G ρ ( α / R ϵ ) ∣ Q ∣ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∣ p d − m ∥ F ∥ L # p d − m / 2 ( Q ) . (119) \lesssim R^\epsilon\sum_{\ell\le s\le1}\sum_{\rho\in\mathcal{J}(s)}\sum_{Q\in\mathcal{G}_\rho(\alpha/R^\epsilon)}\lvert Q\rvert\left(\sum_{\substack{\tau\subset\rho\\\lvert\tau\rvert=\sigma}}\lvert f_\tau\rvert^2\right)^{1/2\lvert p_{d-m}}\lVert F\rVert_{L_\#^{p_{d-m}/2}(Q)}. \tag*{(119)} ≲ R ϵ ℓ ≤ s ≤ 1 ∑ ρ ∈ J ( s ) ∑ Q ∈ G ρ ( α / R ϵ ) ∑ ∣ Q ∣ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 ∣ p d − m ∥ F ∥ L # p d − m /2 ( Q ) . (119)
We estimate the sum over ρ \rho ρ and Q Q Q in (8.34), for each fixed s s s , by the following extension of Proposition 8.9. In its proof in Section 10, we use (8.9) for the current m m m , since the range σ ≤ a ℓ \sigma\le a\ell σ ≤ a ℓ gives the support inclusion (8.33). We state a stronger estimate, with the additional factor ( σ s ) ϵ (\sigma s)^\epsilon ( σ s ) ϵ , to close the induction on σ s \sigma s σ s in its proof.
Proposition 8.15. Assume the hypotheses of Proposition 8.7 for a fixed 0 ≤ m < d − 2 0 \le m < d-2 0 ≤ m < d − 2 , together with C K m ( R ) ≲ 1 C_K^{m}(R) \lesssim1 C K m ( R ) ≲ 1 and (8.9) for this m m m . Let 1 ≤ j ≤ m 1 \le j \le m 1 ≤ j ≤ m , R − 1 / d ≤ ℓ ≪ 1 R^{-1/d} \le\ell\ll1 R − 1/ d ≤ ℓ ≪ 1 , and let f f f be a Schwartz function with supp f ^ ⊂ Γ j , 1 / K , ℓ d + 1 ( R ) \operatorname{supp}\widehat{f} \subset\Gamma_{j,1/K,\ell}^{d+1}(R) supp f ⊂ Γ j , 1/ K , ℓ d + 1 ( R ) . For R − 1 / d ≤ σ ≤ max { R − 1 / d , a ℓ } R^{-1/d} \le\sigma\le\max\{R^{-1/d},a\ell\} R − 1/ d ≤ σ ≤ max { R − 1/ d , a ℓ } , σ ≤ s ≤ 1 \sigma\le s \le1 σ ≤ s ≤ 1 , and α > 0 \alpha>0 α > 0 , we have
∑ ρ ∈ J ( s ) ∑ Q ∈ G ρ ( α ) Q ⊂ Q ∗ ∣ Q ∣ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∥ L # p d − m / 2 ( Q ) p d − m ≲ ϵ RHS of (8.8) , \sum_{\rho\in\mathcal{J}(s)}\sum_{\substack{Q\in\mathcal{G}_{\rho}(\alpha)\\Q\subset Q^*}}\lvert Q\rvert\left(\sum_{\substack{\tau\subset\rho\\\lvert\tau\rvert=\sigma}}\lvert f_\tau\rvert^2\right)^{1/2}\left\|\sum_{\substack{\tau\subset\rho\\\lvert\tau\rvert=\sigma}}\lvert f_\tau\rvert^2\right\|_{L_\#^{p_{d-m/2}}(Q)}^{p_{d-m}}
\lesssim_\epsilon\operatorname{RHS\ of\ (8.8)}, ρ ∈ J ( s ) ∑ Q ∈ G ρ ( α ) Q ⊂ Q ∗ ∑ ∣ Q ∣ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 L # p d − m /2 ( Q ) p d − m ≲ ϵ RHS of (8.8) , with both occurrences of R ϵ R^\epsilon R ϵ replaced by R ϵ ( σ s ) ϵ R^\epsilon(\sigma s)^\epsilon R ϵ ( σ s ) ϵ .
Assuming Proposition 8.15, use (8.34) with ϵ / 2 \epsilon/2 ϵ /2 in place of ϵ \epsilon ϵ . For each fixed s s s , apply the proposition to the sum over ρ \rho ρ and Q Q Q in (8.34), with ϵ / 2 \epsilon/2 ϵ /2 in place of ϵ \epsilon ϵ and amplitude α / R ϵ / 2 \alpha/R^{\epsilon/2} α / R ϵ /2 . The outer coefficient and amplitude denominator are bounded by R ϵ / 2 R ϵ / 2 ( σ s ) ϵ / 2 ≤ R ϵ s ϵ / 2 ≤ R ϵ R^{\epsilon/2}R^{\epsilon/2}(\sigma s)^{\epsilon/2}\le R^\epsilon s^{\epsilon/2}\le R^\epsilon R ϵ /2 R ϵ /2 ( σ s ) ϵ /2 ≤ R ϵ s ϵ /2 ≤ R ϵ . Summing over s s s gives
∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ ϵ ℓ ϵ / 2 ∑ ℓ ≤ s ≤ 1 s ϵ / 2 × RHS of (8.8) ≲ ϵ RHS of (8.8) . (120) \int_{\lvert f\rvert>\alpha}\lvert f\rvert^{p_{d-m}}\lesssim_\epsilon\ell^{\epsilon/2}\sum_{\ell\le s\le1}s^{\epsilon/2}\times\operatorname{RHS\ of\ (8.8)}\lesssim_\epsilon\operatorname{RHS\ of\ (8.8)}. \tag*{(120)} ∫ ∣ f ∣ > α ∣ f ∣ p d − m ≲ ϵ ℓ ϵ /2 ℓ ≤ s ≤ 1 ∑ s ϵ /2 × RHS of (8.8) ≲ ϵ RHS of (8.8) . (120)
Together with (8.30) for ℓ ∼ 1 \ell\sim1 ℓ ∼ 1 , this shows that Proposition 8.15 implies Proposition 8.14.
The reduction (8.27) and comparison (8.29) then give (8.8) for every Schwartz function f f f with supp f ^ ⊂ Γ j d + 1 ( R ) \operatorname{supp}\widehat f\subset\Gamma_j^{d+1}(R) supp f ⊂ Γ j d + 1 ( R ) . Lemma 8.5 gives (8.10), so both estimates are established at index j j j . Proceeding successively through j = 0 , … , m j=0,\ldots,m j = 0 , … , m completes Step 2.
Remark 8.16. Section 10 proves Proposition 8.15 by induction on σ s \sigma s σ s , with R , j , ℓ R,j,\ell R , j , ℓ fixed. Section 11 proves Proposition 8.6, the base case m = d − 2 m=d-2 m = d − 2 of the induction on m m m . The proof follows Steps 1 and 2 in the proof of Proposition 8.7, with p d − m = 4 p_{d-m}=4 p d − m = 4 .
We close this section with the deferred proofs of Lemmas 8.13 and 8.5.
Proof of Lemma 8.13. For ξ = ∑ i = 0 d λ i ϕ d ( i ) ( t θ ) ∈ supp f ^ θ \xi=\sum_{i=0}^{d}\lambda_i\phi_d^{(i)}(t_\theta)\in\operatorname{supp}\widehat f_\theta ξ = ∑ i = 0 d λ i ϕ d ( i ) ( t θ ) ∈ supp f θ , the bounds in (8.28) give
∣ λ i ∣ ≲ { R − j / d ℓ ϱ − j , 0 ≤ i ≤ j , R − j / d ℓ ϱ − j , j < i ≤ d . \lvert\lambda_i\rvert\lesssim
\begin{cases}
R^{-j/d}\ell^{\varrho-j},&0\le i\le j,\\
R^{-j/d}\ell^{\varrho-j},&j<i\le d.
\end{cases} ∣ λ i ∣ ≲ { R − j / d ℓ ϱ − j , R − j / d ℓ ϱ − j , 0 ≤ i ≤ j , j < i ≤ d .
For i ≤ k ≤ d i\le k\le d i ≤ k ≤ d , the coefficient of x ⋅ ϕ d ( k ) ( t θ ) x\cdot\phi_d^{(k)}(t_\theta) x ⋅ ϕ d ( k ) ( t θ ) in the Taylor expansion of x ⋅ ϕ d ( i ) ( t τ ) x\cdot\phi_d^{(i)}(t_\tau) x ⋅ ϕ d ( i ) ( t τ ) is ( t τ − t θ ) k − i / ( k − i ) ! (t_\tau-t_\theta)^{k-i}/(k-i)! ( t τ − t θ ) k − i / ( k − i )! . Since θ ⊂ τ \theta\subset\tau θ ⊂ τ and ∣ τ ∣ = ℓ \lvert\tau\rvert=\ell ∣ τ ∣ = ℓ , we have ∣ t τ − t θ ∣ ≲ ℓ \lvert t_\tau-t_\theta\rvert\lesssim\ell ∣ t τ − t θ ∣ ≲ ℓ , so
R j / d ℓ j − k ∣ t θ − t τ ∣ k − i ≲ R j / d ℓ j − i , 0 ≤ i ≤ k ≤ j , R^{j/d}\ell^{j-k}\lvert t_\theta-t_\tau\rvert^{k-i}\lesssim R^{j/d}\ell^{j-i},\qquad0\le i\le k\le j, R j / d ℓ j − k ∣ t θ − t τ ∣ k − i ≲ R j / d ℓ j − i , 0 ≤ i ≤ k ≤ j ,
R j / d ℓ j ∣ t θ − t τ ∣ k − i ≲ R j / d ℓ j , 0 ≤ i ≤ k , j < k ≤ d . R^{j/d}\ell^{j}\lvert t_\theta-t_\tau\rvert^{k-i}\lesssim R^{j/d}\ell^{j},\qquad0\le i\le k,\quad j<k\le d. R j / d ℓ j ∣ t θ − t τ ∣ k − i ≲ R j / d ℓ j , 0 ≤ i ≤ k , j < k ≤ d .
Thus, for τ ∈ J ( ℓ ) \tau\in\mathcal{J}(\ell) τ ∈ J ( ℓ ) , let D τ D_\tau D τ consist of all x ∈ R d + 1 x\in\mathbb{R}^{d+1} x ∈ R d + 1 satisfying
∣ x ⋅ ϕ d ( i ) ( t τ ) ∣ ≤ { R j / d ℓ ϱ − j , 0 ≤ i ≤ j , R j / d ℓ ϱ − j , j < i ≤ d . \lvert x\cdot\phi_d^{(i)}(t_\tau)\rvert\le
\begin{cases}
R^{j/d}\ell^{\varrho-j},&0\le i\le j,\\
R^{j/d}\ell^{\varrho-j},&j<i\le d.
\end{cases} ∣ x ⋅ ϕ d ( i ) ( t τ )∣ ≤ { R j / d ℓ ϱ − j , R j / d ℓ ϱ − j , 0 ≤ i ≤ j , j < i ≤ d .
The locally constant property then gives ∣ f ^ θ ∣ 2 ≲ ∣ f θ ∣ 2 ∗ w D τ \lvert\widehat f_\theta\rvert^2\lesssim\lvert f_\theta\rvert^2*w_{D_\tau} ∣ f θ ∣ 2 ≲ ∣ f θ ∣ 2 ∗ w D τ for θ ⊂ τ \theta\subset\tau θ ⊂ τ .
Fix the scales q 1 ≤ ⋯ ≤ q d − m − 2 ≤ u q_1\le\cdots\le q_{d-m-2}\le u q 1 ≤ ⋯ ≤ q d − m − 2 ≤ u and ρ ∈ J ( u ) \rho\in\mathcal{J}(u) ρ ∈ J ( u ) on the RHS of (8.8). Its weighted average over B B B contains sums over θ ⊂ τ 1 ⊂ ⋯ ⊂ τ d − m − 2 ⊂ ρ \theta\subset\tau_1\subset\cdots\subset\tau_{d-m-2}\subset\rho θ ⊂ τ 1 ⊂ ⋯ ⊂ τ d − m − 2 ⊂ ρ , where ∣ τ i ∣ = q i \lvert\tau_i\rvert=q_i ∣ τ i ∣ = q i , with convolution weights w B τ i w_{B_{\tau_i}} w B τ i defined in (8.7). We divide the argument into three cases according to the size of ℓ \ell ℓ relative to q d − m − 2 q_{d-m-2} q d − m − 2 and u u u .
*Case 1: ℓ ≤ q d − m − 2 \ell\le q_{d-m-2} ℓ ≤ q d − m − 2 . Let l l l be the smallest index such that q l ≥ ℓ q_l\ge\ell q l ≥ ℓ . If l > 1 l>1 l > 1 , then each τ 1 \tau_1 τ 1 lies in one τ ∈ J ( ℓ ) \tau\in\mathcal{J}(\ell) τ ∈ J ( ℓ ) , and the preceding pointwise bound and Hölder give
[ ( ∑ θ ⊂ τ 1 ∣ f ^ θ ∣ 2 ) ∗ w B τ 1 ] p d − m / p d − m − 1 ≲ [ ( ∑ θ ⊂ τ 1 ∣ f θ ∣ 2 ) ∗ w B τ 1 ] p d − m / p d − m − 1 ∗ w D τ . (121) \left[\left(\sum_{\theta\subset\tau_1}\lvert\widehat f_\theta\rvert^2\right)*w_{B_{\tau_1}}\right]^{p_{d-m}/p_{d-m-1}}
\lesssim
\left[\left(\sum_{\theta\subset\tau_1}\lvert f_\theta\rvert^2\right)*w_{B_{\tau_1}}\right]^{p_{d-m}/p_{d-m-1}}*w_{D_\tau}. \tag*{(121)} [ ( θ ⊂ τ 1 ∑ ∣ f θ ∣ 2 ) ∗ w B τ 1 ] p d − m / p d − m − 1 ≲ [ ( θ ⊂ τ 1 ∑ ∣ f θ ∣ 2 ) ∗ w B τ 1 ] p d − m / p d − m − 1 ∗ w D τ . (121)
The same application of Hölder is valid at each subsequent index i < l i<l i < l , since τ i \tau_i τ i is still contained in one ℓ \ell ℓ cap.
At the l l l th convolution, we show w D τ ∗ w B τ l ≲ w B τ l w_{D_\tau}*w_{B_{\tau_l}}\lesssim w_{B_{\tau_l}} w D τ ∗ w B τ l ≲ w B τ l for τ ⊂ τ l \tau\subset\tau_l τ ⊂ τ l . Once this is proved, for θ ⊂ τ \theta\subset\tau θ ⊂ τ we have ∣ f ^ θ ∣ 2 ∗ w B τ l ≲ ∣ f θ ∣ 2 ∗ ( w D τ ∗ w B τ l ) ≲ ∣ f θ ∣ 2 ∗ w B τ l \lvert\widehat f_\theta\rvert^2*w_{B_{\tau_l}}\lesssim\lvert f_\theta\rvert^2*(w_{D_\tau}*w_{B_{\tau_l}})\lesssim\lvert f_\theta\rvert^2*w_{B_{\tau_l}} ∣ f θ ∣ 2 ∗ w B τ l ≲ ∣ f θ ∣ 2 ∗ ( w D τ ∗ w B τ l ) ≲ ∣ f θ ∣ 2 ∗ w B τ l . Together with the preceding applications of Hölder, this bounds the contribution of these fixed scales to the RHS of (8.8) for f ~ \widetilde{f} f by the corresponding contribution for f f f . To prove the kernel bound, we verify that D τ ⊂ C B τ l D_{\tau} \subset CB_{\tau_l} D τ ⊂ C B τ l .
For x ∈ D τ x \in D_{\tau} x ∈ D τ and 0 ≤ i ≤ d 0 \le i \le d 0 ≤ i ≤ d ,
∣ x ⋅ ϕ d ( i ) ( t τ ) ∣ ≲ { R j / d ℓ j − i , 0 ≤ i ≤ j , R j / d ℓ j , j < i ≤ d ≲ R ( m + 2 ) / d q 1 ⋯ q l − 1 { q l m + l + 1 − i , 0 ≤ i ≤ m + l + 1 , u m + l + 1 − i , m + l + 1 < i ≤ d . \left|x \cdot\phi_d^{(i)}(t_{\tau})\right| \lesssim\begin{cases} R^{j/d}\ell^{j-i}, & 0 \le i \le j, \\ R^{j/d}\ell^j, & j < i \le d \end{cases} \lesssim\frac{R^{(m+2)/d}}{q_1 \cdots q_{l-1}} \begin{cases} q_l^{m+l+1-i}, & 0 \le i \le m+l+1, \\ u^{m+l+1-i}, & m+l+1 < i \le d. \end{cases} x ⋅ ϕ d ( i ) ( t τ ) ≲ { R j / d ℓ j − i , R j / d ℓ j , 0 ≤ i ≤ j , j < i ≤ d ≲ q 1 ⋯ q l − 1 R ( m + 2 ) / d { q l m + l + 1 − i , u m + l + 1 − i , 0 ≤ i ≤ m + l + 1 , m + l + 1 < i ≤ d .
where we used q 1 ⋯ q l − 1 ≤ q l l − 1 q_1 \cdots q_{l-1} \le q_l^{l-1} q 1 ⋯ q l − 1 ≤ q l l − 1 , R − 1 / d ≤ ℓ ≤ q l ≤ u ≤ 1 R^{-1/d} \le\ell\le q_l \le u \le1 R − 1/ d ≤ ℓ ≤ q l ≤ u ≤ 1 , and j ≤ m j \le m j ≤ m . To pass from t τ t_{\tau} t τ to t τ l t_{\tau_l} t τ l , use the identity
x ⋅ ϕ d ( i ) ( b ) = ∑ k = i d ( b − a ) k − i ( k − i ) ! x ⋅ ϕ d ( k ) ( a ) . x \cdot\phi_d^{(i)}(b) = \sum_{k=i}^{d} \frac{(b-a)^{k-i}}{(k-i)!}x \cdot\phi_d^{(k)}(a). x ⋅ ϕ d ( i ) ( b ) = k = i ∑ d ( k − i )! ( b − a ) k − i x ⋅ ϕ d ( k ) ( a ) .
Since ∣ t τ l − t τ ∣ ≲ q l ≤ u |t_{\tau_l}-t_{\tau}| \lesssim q_l \le u ∣ t τ l − t τ ∣ ≲ q l ≤ u , the same upper bound holds for ∣ x ⋅ ϕ d ( i ) ( t τ l ) ∣ \left|x \cdot\phi_d^{(i)}(t_{\tau_l})\right| x ⋅ ϕ d ( i ) ( t τ l ) . By (8.7), D τ ⊂ C B τ l D_{\tau} \subset CB_{\tau_l} D τ ⊂ C B τ l , and hence w D τ ∗ w B τ l ≲ w B τ l w_{D_{\tau}} * w_{B_{\tau_l}} \lesssim w_{B_{\tau_l}} w D τ ∗ w B τ l ≲ w B τ l .
Case 2: q d − m − 2 < ℓ ≤ u q_{d-m-2} < \ell\le u q d − m − 2 < ℓ ≤ u . After applying Hölder at every index in (8.8), we absorb w D τ w_{D_{\tau}} w D τ into the weight W B W_B W B of the outer weighted average over B B B . To do this, we show w D τ ∗ W B ≲ W B w_{D_{\tau}} * W_B \lesssim W_B w D τ ∗ W B ≲ W B , so that (2.26) bounds the weighted average over B B B for f ~ \widetilde{f} f by that for f f f .
For x ∈ D τ x \in D_{\tau} x ∈ D τ and 0 ≤ i ≤ d 0 \le i \le d 0 ≤ i ≤ d ,
∣ x ⋅ ϕ d ( i ) ( t τ ) ∣ ≲ { R j / d ℓ j − i , 0 ≤ i ≤ j , R j / d ℓ j , j < i ≤ d ≲ R ( m + 2 ) / d u d − i q 1 ⋯ q d − m − 2 . \left|x \cdot\phi_d^{(i)}(t_{\tau})\right| \lesssim\begin{cases} R^{j/d}\ell^{j-i}, & 0 \le i \le j, \\ R^{j/d}\ell^j, & j < i \le d \end{cases} \lesssim\frac{R^{(m+2)/d}u^{d-i}}{q_1 \cdots q_{d-m-2}}. x ⋅ ϕ d ( i ) ( t τ ) ≲ { R j / d ℓ j − i , R j / d ℓ j , 0 ≤ i ≤ j , j < i ≤ d ≲ q 1 ⋯ q d − m − 2 R ( m + 2 ) / d u d − i .
where we used q 1 ⋯ q d − m − 2 ≤ u d − m − 2 q_1 \cdots q_{d-m-2} \le u^{d-m-2} q 1 ⋯ q d − m − 2 ≤ u d − m − 2 , R − 1 / d ≤ ℓ ≤ u ≤ 1 R^{-1/d} \le\ell\le u \le1 R − 1/ d ≤ ℓ ≤ u ≤ 1 , and j ≤ m j \le m j ≤ m . Since ∣ t ρ − t τ ∣ ≲ u |t_{\rho}-t_{\tau}| \lesssim u ∣ t ρ − t τ ∣ ≲ u , Taylor’s formula gives the same upper bound for ∣ x ⋅ ϕ d ( i ) ( t ρ ) ∣ \left|x \cdot\phi_d^{(i)}(t_{\rho})\right| x ⋅ ϕ d ( i ) ( t ρ ) . Thus D τ ⊂ C ( B − c B ) D_{\tau} \subset C(B-c_B) D τ ⊂ C ( B − c B ) , where c B c_B c B is the center of B B B . Hence w D τ ∗ W B ≲ W B w_{D_{\tau}} * W_B \lesssim W_B w D τ ∗ W B ≲ W B , and (2.26) bounds the weighted average over B B B in (8.8) for f ~ \widetilde{f} f by that for f f f .
Case 3: u < ℓ u < \ell u < ℓ . Then ρ \rho ρ is contained in one τ ∈ J ( ℓ ) \tau\in\mathcal{J}(\ell) τ ∈ J ( ℓ ) . In this case, w D τ ∗ W B ≲ W B w_{D_{\tau}} * W_B \lesssim W_B w D τ ∗ W B ≲ W B need not hold, so we use (2.25) to remove w D τ w_{D_{\tau}} w D τ from the sum over B B B in (8.8).
To apply (2.25), let F F F be the integrand inside the weighted average over B B B in (8.8) for f f f , and let F ~ \widetilde{F} F be the same expression with f θ f_{\theta} f θ replaced by f ~ θ \widetilde{f}_{\theta} f θ . Since every θ ⊂ ρ \theta\subset\rho θ ⊂ ρ lies in this single τ \tau τ , applying Hölder as in the first case gives F ~ ≲ F ∗ w D τ \widetilde{F} \lesssim F * w_{D_{\tau}} F ≲ F ∗ w D τ . With t = [ α / ( R ε N ( u ) ) ] p d − m / 2 t = [\alpha/(R\varepsilon N(u))]^{p_{d-m/2}} t = [ α / ( R εN ( u )) ] p d − m /2 , where N ( u ) N(u) N ( u ) is computed from { f ~ θ } \{\widetilde{f}_{\theta}\} { f θ } , we therefore obtain
∑ B ∥ F ~ ∥ L # 1 ( B ) ≳ t ∣ B ∣ ∥ F ~ ∥ L # 1 ( B ) 2 ≲ ∑ B ∥ F ∗ w D τ ∥ L # 1 ( B ) ≳ t ∣ B ∣ ∥ F ∗ w D τ ∥ L # 1 ( B ) 2 ≲ ∑ B ∥ F ∥ L # 1 ( B ) ≳ t ∣ B ∣ ∥ F ∥ L # 1 ( B ) 2 . \sum_{\substack{B\\ \|\widetilde{F}\|_{L^1_{\#}(B)} \gtrsim t}} |B|\|\widetilde{F}\|_{L^1_{\#}(B)}^2 \lesssim\sum_{\substack{B\\ \|F*w_{D_{\tau}}\|_{L^1_{\#}(B)} \gtrsim t}} |B|\|F*w_{D_{\tau}}\|_{L^1_{\#}(B)}^2 \lesssim\sum_{\substack{B\\ \|F\|_{L^1_{\#}(B)} \gtrsim t}} |B|\|F\|_{L^1_{\#}(B)}^2. B ∥ F ∥ L # 1 ( B ) ≳ t ∑ ∣ B ∣∥ F ∥ L # 1 ( B ) 2 ≲ B ∥ F ∗ w D τ ∥ L # 1 ( B ) ≳ t ∑ ∣ B ∣∥ F ∗ w D τ ∥ L # 1 ( B ) 2 ≲ B ∥ F ∥ L # 1 ( B ) ≳ t ∑ ∣ B ∣∥ F ∥ L # 1 ( B ) 2 .
where all sums use the B B B tiling in (8.8), and the last inequality follows from (2.25). The final sum is bounded by the sum over B B B on the RHS of (8.8) for f f f .
The counts N ( u ) N(u) N ( u ) computed from { f ~ θ } \{\widetilde{f}_{\theta}\} { f θ } are no larger than those from { f θ } \{f_{\theta}\} { f θ } , so using the latter only enlarges the sums obtained in all three cases. This proves the lemma. □
Proof of Lemma 8.5. Starting with the innermost convolution in (8.8), we move each power inside its convolution by Hölder and absorb the kernel into the next convolution. At the last layer, we absorb the kernel into the outer weight W B W_B W B .
To absorb w B τ l w_{B_{\tau_l}} w B τ l into the next convolution, we show that B τ l ⊂ C B τ l + 1 B_{\tau_l} \subset CB_{\tau_{l+1}} B τ l ⊂ C B τ l + 1 . For x ∈ B τ l x \in B_{\tau_l} x ∈ B τ l and 0 ≤ i ≤ d 0 \le i \le d 0 ≤ i ≤ d , (8.7) gives
∣ x ⋅ ϕ d ( i ) ( t τ l ) ∣ ≲ R ( m + 2 ) / d q 1 ⋯ q l − 1 { q l m + l + 1 − i , 0 ≤ i ≤ m + l + 1 , u m + l + 1 − i , m + l + 1 < i ≤ d ≲ R ( m + 2 ) / d q 1 ⋯ q l { q l + 1 m + l + 2 − i , 0 ≤ i ≤ m + l + 2 , u m + l + 2 − i , m + l + 2 < i ≤ d . \left|x \cdot\phi_d^{(i)}(t_{\tau_l})\right| \lesssim\frac{R^{(m+2)/d}}{q_1 \cdots q_{l-1}} \begin{cases} q_l^{m+l+1-i}, & 0 \le i \le m+l+1, \\ u^{m+l+1-i}, & m+l+1 < i \le d \end{cases} \lesssim\frac{R^{(m+2)/d}}{q_1 \cdots q_l} \begin{cases} q_{l+1}^{m+l+2-i}, & 0 \le i \le m+l+2, \\ u^{m+l+2-i}, & m+l+2 < i \le d. \end{cases} x ⋅ ϕ d ( i ) ( t τ l ) ≲ q 1 ⋯ q l − 1 R ( m + 2 ) / d { q l m + l + 1 − i , u m + l + 1 − i , 0 ≤ i ≤ m + l + 1 , m + l + 1 < i ≤ d ≲ q 1 ⋯ q l R ( m + 2 ) / d { q l + 1 m + l + 2 − i , u m + l + 2 − i , 0 ≤ i ≤ m + l + 2 , m + l + 2 < i ≤ d .
where we used q l ≤ q l + 1 ≤ u q_l \le q_{l+1} \le u q l ≤ q l + 1 ≤ u . Since ∣ t τ l + 1 − t τ l ∣ ≲ q l + 1 ≤ u |t_{\tau_{l+1}}-t_{\tau_l}| \lesssim q_{l+1} \le u ∣ t τ l + 1 − t τ l ∣ ≲ q l + 1 ≤ u , Taylor’s formula gives the same upper bound for ∣ x ⋅ ϕ d ( i ) ( t τ l + 1 ) ∣ \left|x \cdot\phi_d^{(i)}(t_{\tau_{l+1}})\right| x ⋅ ϕ d ( i ) ( t τ l + 1 ) . Hence B τ l ⊂ C B τ l + 1 B_{\tau_l} \subset CB_{\tau_{l+1}} B τ l ⊂ C B τ l + 1 .
At the last convolution, we compare B τ d − m − 2 B_{\tau_{d-m-2}} B τ d − m − 2 with the outer box B B B . For x ∈ B τ d − m − 2 x \in B_{\tau_{d-m-2}} x ∈ B τ d − m − 2 and 0 ≤ i ≤ d 0 \le i \le d 0 ≤ i ≤ d , (8.7) gives
∣ x ⋅ ϕ d ( i ) ( t τ d − m − 2 ) ∣ ≲ R ( m + 2 ) / d q 1 ⋯ q d − m − 3 { q d − m − 2 d − 1 − i , 0 ≤ i ≤ d − 1 , u − 1 , i = d ≲ R ( m + 2 ) / d u d − i q 1 ⋯ q d − m − 2 \left|x \cdot\phi_d^{(i)}(t_{\tau_{d-m-2}})\right| \lesssim\frac{R^{(m+2)/d}}{q_1 \cdots q_{d-m-3}} \begin{cases} q_{d-m-2}^{d-1-i}, & 0 \le i \le d-1, \\ u^{-1}, & i=d \end{cases} \lesssim\frac{R^{(m+2)/d}u^{d-i}}{q_1 \cdots q_{d-m-2}} x ⋅ ϕ d ( i ) ( t τ d − m − 2 ) ≲ q 1 ⋯ q d − m − 3 R ( m + 2 ) / d { q d − m − 2 d − 1 − i , u − 1 , 0 ≤ i ≤ d − 1 , i = d ≲ q 1 ⋯ q d − m − 2 R ( m + 2 ) / d u d − i
where we used q d − m − 2 ≤ u q_{d-m-2} \le u q d − m − 2 ≤ u . Since ∣ t τ d − m − 2 − t ρ ∣ ≲ u \lvert t_{\tau_{d-m-2}}-t_{\rho}\rvert\lesssim u ∣ t τ d − m − 2 − t ρ ∣ ≲ u , Taylor’s formula gives the same upper bound for ∣ x ⋅ ϕ d ′ ( t ρ ) ∣ \lvert x\cdot\phi_d'(t_{\rho})\rvert ∣ x ⋅ ϕ d ′ ( t ρ )∣ . Thus B τ d − m − 2 ⊂ C ( B − c B ) B_{\tau_{d-m-2}}\subset C(B-c_B) B τ d − m − 2 ⊂ C ( B − c B ) , where c B c_B c B is the center of B B B . These comparisons give w B τ l ∗ w B τ l + 1 ≲ w B τ l + 1 w_{B_{\tau_l}}*w_{B_{\tau_{l+1}}}\lesssim w_{B_{\tau_{l+1}}} w B τ l ∗ w B τ l + 1 ≲ w B τ l + 1 , w B τ d − m − 2 ∗ w B ≲ w B w_{B_{\tau_{d-m-2}}}*w_B\lesssim w_B w B τ d − m − 2 ∗ w B ≲ w B .
For example, when m = d − 4 m=d-4 m = d − 4 , Hölder and the first kernel bound give
( ∑ τ 1 ⊂ τ 2 [ ( ∑ θ ⊂ τ 1 ∣ f θ ∣ 2 ) ∗ w B τ 1 ] 11 / 7 ) ∗ w B τ 2 ≲ ( ∑ θ ⊂ τ 2 ∣ f θ ∣ 2 ) 11 / 7 ∗ w B τ 2 . \left(\sum_{\tau_1\subset\tau_2}\left[\left(\sum_{\theta\subset\tau_1}|f_\theta|^2\right)*w_{B_{\tau_1}}\right]^{11/7}\right)*w_{B_{\tau_2}}\lesssim\left(\sum_{\theta\subset\tau_2}|f_\theta|^2\right)^{11/7}*w_{B_{\tau_2}}. τ 1 ⊂ τ 2 ∑ [ ( θ ⊂ τ 1 ∑ ∣ f θ ∣ 2 ) ∗ w B τ 1 ] 11/7 ∗ w B τ 2 ≲ ( θ ⊂ τ 2 ∑ ∣ f θ ∣ 2 ) 11/7 ∗ w B τ 2 .
Applying (2.26) with exponent 7 / 4 7/4 7/4 and summing over τ 2 ⊂ ρ \tau_2\subset\rho τ 2 ⊂ ρ gives
∫ B ∑ τ 2 ⊂ ρ [ ( ∑ τ 1 ⊂ τ 2 [ ( ∑ θ ⊂ τ 1 ∣ f θ ∣ 2 ) ∗ w B τ 1 ] 11 / 7 ) ∗ w B τ 2 ] 7 / 4 ≲ ∫ B ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ) 11 / 4 . \int_B\sum_{\tau_2\subset\rho}\left[\left(\sum_{\tau_1\subset\tau_2}\left[\left(\sum_{\theta\subset\tau_1}|f_\theta|^2\right)*w_{B_{\tau_1}}\right]^{11/7}\right)*w_{B_{\tau_2}}\right]^{7/4}\lesssim\int_B\left(\sum_{\theta\subset\rho}|f_\theta|^2\right)^{11/4}. ∫ B τ 2 ⊂ ρ ∑ τ 1 ⊂ τ 2 ∑ [ ( θ ⊂ τ 1 ∑ ∣ f θ ∣ 2 ) ∗ w B τ 1 ] 11/7 ∗ w B τ 2 7/4 ≲ ∫ B θ ⊂ ρ ∑ ∣ f θ ∣ 2 11/4 .
The powers in (8.8) have product p d − m / 4 p_{d-m}/4 p d − m /4 . Repeating the same calculation bounds its weighted average over B B B by \mathop{\fint}_B\left(\sum_{\theta\subset\rho}|f_\theta|^2\right)^{p_{d-m}/4} .
For fixed q 1 , … , q d − m − 2 , u , ρ q_1,\ldots,q_{d-m-2},u,\rho q 1 , … , q d − m − 2 , u , ρ , put F = ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ) p d − m / 4 F=\left(\sum_{\theta\subset\rho}|f_\theta|^2\right)^{p_{d-m}/4} F = ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ) p d − m /4 and t = [ α / ( R ϵ N ( u ) ) ] p d − m / 2 t=\left[\alpha/(R^\epsilon N(u))\right]^{p_{d-m}/2} t = [ α / ( R ϵ N ( u )) ] p d − m /2 . The comparison of the weighted averages over B B B shows that every B B B counted in (8.8) satisfies \fint_B F\gtrsim t . Since u d − m − 2 / ( q 1 ⋯ q d − m − 2 ) ≥ 1 u^{d-m-2}/(q_1\cdots q_{d-m-2})\ge1 u d − m − 2 / ( q 1 ⋯ q d − m − 2 ) ≥ 1 , Lemma 2.14 bounds this sum over B B B by the sum over U ∥ ( u R 1 / d ) m U ρ , R U\parallel(uR^{1/d})^mU_{\rho,R} U ∥ ( u R 1/ d ) m U ρ , R , giving
\sum_{\substack{U\parallel(uR^{1/d})^mU_{\rho,R}\\ U\in\mathcal{G}_{\rho}(R^\epsilon)}}|U|\left[\fint_U\left(\sum_{\theta\subset\rho}|f_\theta|^2\right)^{p_{d-m}/4}\right]^2. \tag*{(122)}
Summing (8.37) over u , ρ u,\rho u , ρ gives the right hand side of (8.10), with the original function f f f . Summing over the dyadic scales q 1 , … , q d − m − 2 q_1,\ldots,q_{d-m-2} q 1 , … , q d − m − 2 proves the lemma. □
From curves to higher dimensional cones over moment curves: Proof of Proposition 8.7, Step 1
Assume the hypotheses of Proposition 8.7 for 0 ≤ m ≤ d − 3 0\le m\le d-3 0 ≤ m ≤ d − 3 . We first prove Lemma 8.10 and then Lemma 8.11. These give the four branches as in Section 8.2. We then use those branches to prove Proposition 8.9 together with (8.9), completing Step 1 in the proof of Proposition 8.7. The case m = d − 2 m=d-2 m = d − 2 is treated in Section 11.
Proof of Lemma 8.10
The argument for the high frequency part applies the estimate at exponent p d − m − 1 p_{d-m-1} p d − m − 1 to a sum of quadratic expressions. We first record the resulting estimate at exponent p d − m / 2 p_{d-m}/2 p d − m /2 , keeping the convolutions needed for the induction.
Lemma 9.1. Assume the hypotheses of Proposition 8.7 for a fixed 0 ≤ m ≤ d − 3 0\le m\le d-3 0 ≤ m ≤ d − 3 . Let h = ∑ θ ∈ Ξ m + 1 d + 1 ( r ) h θ h=\sum_{\theta\in\Xi_{m+1}^{d+1}(r)}h_\theta h = ∑ θ ∈ Ξ m + 1 d + 1 ( r ) h θ , where each h θ h_\theta h θ is a Schwartz function with supp h θ ^ ⊂ θ \operatorname{supp}\widehat{h_\theta}\subset\theta supp h θ ⊂ θ . Then
\begin{aligned}
\int_{|h|>\alpha^2}|h|^{p_{d-m}/2}
&\lesssim_\epsilon r^\epsilon
\sum_{r^{-1/d}\le q_1\le\cdots\le q_{d-m-3}\le u\le1}
\sum_{\rho\in\mathcal{J}(u)}
\frac{u^{d-m-3}}{q_1\cdots q_{d-m-3}}
\sum_{B\in\mathcal{G}_\rho(r^\epsilon)}
\frac{|B|}{(ur^{1/d})^{m+1}U_{\rho,r}}\\
&\quad\times\left[\fint_B
\sum_{\tau_{d-m-3}\subset\rho}\left(\cdots\sum_{\tau_1\subset\tau_2}
\left[\left(\sum_{\theta\subset\tau_1}|h_\theta|^{p_{d-m}/p_{d-m-1}}\right)*w_{B_{\tau_1}}\right]^{p_{d-m-1}/p_{d-m-2}}
\cdots*w_{B_{\tau_{d-m-3}}}\right)^{7/4}\right]^2\\
&\quad+r^\epsilon\sum_\theta\sum_{\substack{Q\parallel\theta^*\\Q\in\mathcal{G}_\theta(r^\epsilon)}}|Q|
\left[\fint_Q|h_\theta|^{p_{d-m}/4}\right]^2.
\end{aligned}
In both sums, the condition on B B B or Q Q Q means that the average inside the outer square is at least [ α r ϵ N ( u ) ] p d − m / 2 \left[\frac{\alpha}{r^\epsilon N(u)}\right]^{p_{d-m}/2} [ r ϵ N ( u ) α ] p d − m /2 , with u = r − 1 / d u=r^{-1/d} u = r − 1/ d in the second sum.
Since the proof of the lemma is rather standard, we postpone the proof to Appendix § 16 .§ 2 .
In the following lemma, we take q 1 = ⋯ = q d − m − 2 = u q_1=\cdots=q_{d-m-2}=u q 1 = ⋯ = q d − m − 2 = u and τ 1 = ⋯ = τ d − m − 2 = ρ \tau_1=\cdots=\tau_{d-m-2}=\rho τ 1 = ⋯ = τ d − m − 2 = ρ in (92) , with parameter r r r . Then B τ l = ( u r 1 / d ) m U ρ , r B_{\tau_l}=(ur^{1/d})^mU_{\rho,r} B τ l = ( u r 1/ d ) m U ρ , r for 1 ≤ l ≤ d − m − 2 1\leq l\leq d-m-2 1 ≤ l ≤ d − m − 2 . The convolutions and powers follow the order in (93) .
Lemma 9.2. Let 0 ≤ m ≤ d − 3 0\leq m\leq d-3 0 ≤ m ≤ d − 3 , and let g g g be a Schwartz function with supp g ^ ⊂ Γ m d + 1 ( r ) \operatorname{supp}\widehat{g}\subset\Gamma_m^{d+1}(r) supp g ⊂ Γ m d + 1 ( r ) . For r − 1 / d ≤ u ≤ 1 r^{-1/d}\leq u\leq1 r − 1/ d ≤ u ≤ 1 , ρ ∈ J ( u ) \rho\in\mathcal{J}(u) ρ ∈ J ( u ) , and B ∥ B 0 = ( u r 1 / d ) m U ρ , r B\parallel B_0=(ur^{1/d})^mU_{\rho,r} B ∥ B 0 = ( u r 1/ d ) m U ρ , r , we have
∫ B [ ( ∑ θ ∈ P ∣ g θ ∣ 2 ) ∗ w B 0 ] p d − m / 4 ≲ ∫ B ( ( [ ( ∑ θ ∈ P ∣ g θ ∣ 2 ) ∗ w B 0 ] p d − m / p d − m − 1 ∗ w B 0 ) p d − m − 1 / p d − m − 2 ⋯ ∗ w B 0 ) 7 / 4 . \int_B\left[\left(\sum_{\theta\in\mathcal{P}}|g_\theta|^2\right)*w_{B_0}\right]^{p_{d-m}/4}
\lesssim
\int_B\left(\left(\left[\left(\sum_{\theta\in\mathcal{P}}|g_\theta|^2\right)*w_{B_0}\right]^{p_{d-m}/p_{d-m-1}}*w_{B_0}\right)^{p_{d-m-1}/p_{d-m-2}}\cdots*w_{B_0}\right)^{7/4}. ∫ B [ ( θ ∈ P ∑ ∣ g θ ∣ 2 ) ∗ w B 0 ] p d − m /4 ≲ ∫ B [ ( θ ∈ P ∑ ∣ g θ ∣ 2 ) ∗ w B 0 ] p d − m / p d − m − 1 ∗ w B 0 p d − m − 1 / p d − m − 2 ⋯ ∗ w B 0 7/4 . When m = d − 3 m=d-3 m = d − 3 , the two averages agree.
Proof. Since w B 0 ≲ w B 0 ∗ w B 0 w_{B_0}\lesssim w_{B_0}*w_{B_0} w B 0 ≲ w B 0 ∗ w B 0 , Hölder’s inequality gives, for β ≥ 1 \beta\geq1 β ≥ 1 ,
[ ( ∑ θ ∈ P ∣ g θ ∣ 2 ) ∗ w B 0 ] β ≲ β [ ( ∑ θ ∈ P ∣ g θ ∣ 2 ) ∗ w B 0 ] β ∗ w B 0 . (123) \left[\left(\sum_{\theta\in\mathcal{P}}|g_\theta|^2\right)*w_{B_0}\right]^\beta
\lesssim_\beta
\left[\left(\sum_{\theta\in\mathcal{P}}|g_\theta|^2\right)*w_{B_0}\right]^\beta*w_{B_0}. \tag*{(123)} [ ( θ ∈ P ∑ ∣ g θ ∣ 2 ) ∗ w B 0 ] β ≲ β [ ( θ ∈ P ∑ ∣ g θ ∣ 2 ) ∗ w B 0 ] β ∗ w B 0 . (123)
Apply (123) successively for l = 2 , … , d − m − 2 l=2,\ldots,d-m-2 l = 2 , … , d − m − 2 , with β = p d − m / p d − m − l + 1 \beta=p_{d-m}/p_{d-m-l+1} β = p d − m / p d − m − l + 1 , and take the power p d − m − l + 1 / p d − m − l p_{d-m-l+1}/p_{d-m-l} p d − m − l + 1 / p d − m − l after each convolution. The successive powers have product p d − m / 4 p_{d-m}/4 p d − m /4 , so averaging over B B B proves the lemma. □
Proof of Lemma 8.10 for m < d − 2 m<d-2 m < d − 2 . Let g g g be a Schwartz function with supp g ^ ⊂ Γ m d + 1 ( r ) \operatorname{supp}\widehat{g}\subset\Gamma_m^{d+1}(r) supp g ⊂ Γ m d + 1 ( r ) . We prove the stronger estimate
∫ ∑ θ ∣ g θ ∣ 2 ≳ α 2 ( ∑ θ ∣ g θ ∣ 2 ) p d − m / 2 ≲ ϵ RHS of (8.8) with R = r , f = g . (124) \int_{\sum_{\theta}|g_\theta|^2\gtrsim\alpha^2}\left(\sum_\theta|g_\theta|^2\right)^{p_{d-m}/2}\lesssim_\epsilon\operatorname{RHS\ of\ (8.8)\ with\ }R=r,\ f=g. \tag*{(124)} ∫ ∑ θ ∣ g θ ∣ 2 ≳ α 2 ( θ ∑ ∣ g θ ∣ 2 ) p d − m /2 ≲ ϵ RHS of (8.8) with R = r , f = g . (124)
Lemma 8.5 will then give Lemma 8.10. The estimate (124) is also the input at the fine endpoint in the proof of Proposition 8.9 below.
We use the cutoffs η θ , ≤ σ m + 2 \eta_{\theta,\leq\sigma}^{m+2} η θ , ≤ σ m + 2 and η θ , σ m + 2 \eta_{\theta,\sigma}^{m+2} η θ , σ m + 2 from § 9.3 . For ξ = ∑ i = 0 m + 2 λ i ϕ d ( i ) ( t θ ) ∈ supp η θ , σ m + 2 \xi=\sum_{i=0}^{m+2}\lambda_i\phi_d^{(i)}(t_\theta)\in\operatorname{supp}\eta_{\theta,\sigma}^{m+2} ξ = ∑ i = 0 m + 2 λ i ϕ d ( i ) ( t θ ) ∈ supp η θ , σ m + 2 , we have
∣ λ i ∣ ≲ r − i / d σ m + 2 − i ( 0 ≤ i ≤ m + 1 ) , max 0 ≤ i ≤ m + 1 r i / d σ i − m − 2 ∣ λ i ∣ ∼ 1. (125) |\lambda_i|\lesssim r^{-i/d}\sigma^{m+2-i}\quad(0\leq i\leq m+1),\qquad\max_{0\leq i\leq m+1}r^{i/d}\sigma^{i-m-2}|\lambda_i|\sim1. \tag*{(125)} ∣ λ i ∣ ≲ r − i / d σ m + 2 − i ( 0 ≤ i ≤ m + 1 ) , 0 ≤ i ≤ m + 1 max r i / d σ i − m − 2 ∣ λ i ∣ ∼ 1. (125)
Since η θ , ≤ 1 m + 2 = 1 \eta_{\theta,\leq1}^{m+2}=1 η θ , ≤ 1 m + 2 = 1 on supp ∣ g θ ^ ∣ 2 \operatorname{supp}|\widehat{g_\theta}|^2 supp ∣ g θ ∣ 2 , dyadic decomposition gives
∫ ∑ θ ∣ g θ ∣ 2 ≳ α 2 ( ∑ θ ∣ g θ ∣ 2 ) p d − m / 2 ≲ ∫ ∣ ∑ θ ∣ g θ ∣ 2 ∗ η ~ θ , ≤ r − 1 / d m + 2 ∣ ≳ α 2 ∣ ∑ θ ∣ g θ ∣ 2 ∗ η ~ θ , ≤ r − 1 / d m + 2 ∣ p d − m / 2 + ∑ r − 1 / d < σ ≤ 1 ∫ ∣ ∑ θ ∣ g θ ∣ 2 ∗ η ~ θ , σ m + 2 ∣ ≳ α 2 ∣ ∑ θ ∣ g θ ∣ 2 ∗ η ~ θ , σ m + 2 ∣ p d − m / 2 . (126) \int_{\sum_\theta|g_\theta|^2\gtrsim\alpha^2}\left(\sum_\theta|g_\theta|^2\right)^{p_{d-m}/2}
\lesssim
\int_{\left|\sum_\theta|g_\theta|^2*\widetilde{\eta}_{\theta,\leq r^{-1/d}}^{m+2}\right|\gtrsim\alpha^2}\left|\sum_\theta|g_\theta|^2*\widetilde{\eta}_{\theta,\leq r^{-1/d}}^{m+2}\right|^{p_{d-m}/2}
+\sum_{r^{-1/d}<\sigma\leq1}\int_{\left|\sum_\theta|g_\theta|^2*\widetilde{\eta}_{\theta,\sigma}^{m+2}\right|\gtrsim\alpha^2}\left|\sum_\theta|g_\theta|^2*\widetilde{\eta}_{\theta,\sigma}^{m+2}\right|^{p_{d-m}/2}. \tag*{(126)} ∫ ∑ θ ∣ g θ ∣ 2 ≳ α 2 ( θ ∑ ∣ g θ ∣ 2 ) p d − m /2 ≲ ∫ ∑ θ ∣ g θ ∣ 2 ∗ η θ , ≤ r − 1/ d m + 2 ≳ α 2 θ ∑ ∣ g θ ∣ 2 ∗ η θ , ≤ r − 1/ d m + 2 p d − m /2 + r − 1/ d < σ ≤ 1 ∑ ∫ ∣ ∑ θ ∣ g θ ∣ 2 ∗ η θ , σ m + 2 ∣ ≳ α 2 θ ∑ ∣ g θ ∣ 2 ∗ η θ , σ m + 2 p d − m /2 . (126)
We will bound the first integral on the RHS of (126) by the terms with q 1 = ⋯ = q d − m − 2 = u = 1 q_1=\cdots=q_{d-m-2}=u=1 q 1 = ⋯ = q d − m − 2 = u = 1 in (93) . For each σ > r − 1 / d \sigma>r^{-1/d} σ > r − 1/ d , we will group the θ \theta θ caps and apply Lemma 9.1 after dilation to bound the corresponding integral in the second term of (126) .
For the first integral on the RHS of (126) , take B 0 = r m / d U [ 0 , 1 ] , r B_0=r^{m/d}U_{[0,1],r} B 0 = r m / d U [ 0 , 1 ] , r , as in Lemma 9.2 with u = 1 u=1 u = 1 and ρ = [ 0 , 1 ] \rho=[0,1] ρ = [ 0 , 1 ] . This box has width r ( m + 2 ) / d r^{(m+2)/d} r ( m + 2 ) / d in every dual Taylor direction.
At σ = r − 1 / d \sigma=r^{-1/d} σ = r − 1/ d , the first m + 2 m+2 m + 2 factors of η θ , ≤ r − 1 / d m + 2 \eta_{\theta,\leq r^{-1/d}}^{m+2} η θ , ≤ r − 1/ d m + 2 cut off the frequency coordinates at widths r − i / d ( r − 1 / d ) m + 2 − i = r − ( m + 2 ) / d r^{-i/d}(r^{-1/d})^{m+2-i}=r^{-(m+2)/d} r − i / d ( r − 1/ d ) m + 2 − i = r − ( m + 2 ) / d , 0 ≤ i ≤ m + 1 0\leq i\leq m+1 0 ≤ i ≤ m + 1 . The remaining factors equal 1 1 1 on supp ∣ g θ ^ ∣ 2 \operatorname{supp}|\widehat{g_\theta}|^2 supp ∣ g θ ∣ 2 , and the locally constant property allows averaging over width r ( m + 2 ) / d r^{(m+2)/d} r ( m + 2 ) / d in directions m + 2 , … , d m+2,\ldots,d m + 2 , … , d , since r ( m + 2 ) / d ≤ r i / d r^{(m+2)/d}\leq r^{i/d} r ( m + 2 ) / d ≤ r i / d .
Since θ ⊂ [ 0 , 1 ] \theta\subset[0,1] θ ⊂ [ 0 , 1 ] , Taylor expansion shows that boxes with width r ( m + 2 ) / d r^{(m+2)/d} r ( m + 2 ) / d in every dual Taylor direction at t θ t_\theta t θ are comparable to B 0 B_0 B 0 . The locally constant property then gives
∣ ∑ θ ∣ g θ ∣ 2 ∗ η ~ θ , ≤ r − 1 / d m + 2 ∣ ≲ ( ∑ θ ∣ g θ ∣ 2 ) ∗ w B 0 . \left|\sum_\theta|g_\theta|^2*\widetilde{\eta}_{\theta,\leq r^{-1/d}}^{m+2}\right|
\lesssim
\left(\sum_\theta|g_\theta|^2\right)*w_{B_0}. θ ∑ ∣ g θ ∣ 2 ∗ η θ , ≤ r − 1/ d m + 2 ≲ ( θ ∑ ∣ g θ ∣ 2 ) ∗ w B 0 .
and bound the first integral on the RHS of (126) by
∑ U ∥ η [ 0 , 1 ] , r m / d U ∈ G [ 0 , 1 ] ( r ϵ σ ) ∣ U ∣ [ ∫ U ( ( ∑ θ ∣ g θ ∣ 2 ) ∗ w B 0 ) p d − m / 4 ] 2 . (127) \sum_{U\parallel\eta^{m/d}_{[0,1],r}\atop U\in\mathcal{G}_{[0,1]}(\frac{r\epsilon}{\sigma})} |U|\left[\int_U\left(\left(\sum_{\theta}|g_\theta|^2\right)*w_{B_0}\right)^{p_{d-m}/4}\right]^2. \tag*{(127)} U ∈ G [ 0 , 1 ] ( σ r ϵ ) U ∥ η [ 0 , 1 ] , r m / d ∑ ∣ U ∣ ∫ U ( ( θ ∑ ∣ g θ ∣ 2 ) ∗ w B 0 ) p d − m /4 2 . (127)
Applying Lemma 9.2 to g g g , with u = 1 u=1 u = 1 and ρ = [ 0 , 1 ] \rho=[0,1] ρ = [ 0 , 1 ] , bounds (9.5) by the terms of (93) with q 1 = ⋯ = q d − m − 2 = u = 1 q_1=\cdots=q_{d-m-2}=u=1 q 1 = ⋯ = q d − m − 2 = u = 1 , including the amplitude condition.
We next estimate the integral indexed by σ > r − 1 / d \sigma>r^{-1/d} σ > r − 1/ d in the second term of (126) . Put q 1 = r − 1 / d σ − 1 q_1=r^{-1/d}\sigma^{-1} q 1 = r − 1/ d σ − 1 , and group the θ \theta θ caps into τ 1 ∈ J ( q 1 ) \tau_1\in\mathcal{J}(q_1) τ 1 ∈ J ( q 1 ) . With B τ 1 B_{\tau_1} B τ 1 as in (8.7), where R = r R=r R = r and l = 1 l=1 l = 1 , we have
h τ 1 = ∑ θ ⊂ τ 1 ∣ g θ ∣ 2 ∗ η ~ θ , σ m + 2 , ∣ h τ 1 ∣ ≲ ( ∑ θ ⊂ τ 1 ∣ g θ ∣ 2 ) ∗ w B τ 1 , (128) h_{\tau_1}=\sum_{\theta\subset\tau_1}|g_\theta|^2*\widetilde{\eta}^{m+2}_{\theta,\sigma},\qquad|h_{\tau_1}|\lesssim\left(\sum_{\theta\subset\tau_1}|g_\theta|^2\right)*w_{B_{\tau_1}}, \tag*{(128)} h τ 1 = θ ⊂ τ 1 ∑ ∣ g θ ∣ 2 ∗ η θ , σ m + 2 , ∣ h τ 1 ∣ ≲ ( θ ⊂ τ 1 ∑ ∣ g θ ∣ 2 ) ∗ w B τ 1 , (128)
uniformly for q 1 ≤ u ≤ 1 q_1\leq u\leq1 q 1 ≤ u ≤ 1 .
Since θ ⊂ τ 1 \theta\subset\tau_1 θ ⊂ τ 1 , we have ∣ t θ − t τ 1 ∣ ≲ q 1 ≤ u |t_\theta-t_{\tau_1}|\lesssim q_1\leq u ∣ t θ − t τ 1 ∣ ≲ q 1 ≤ u . By Taylor expansion, B τ 1 B_{\tau_1} B τ 1 agrees up to a fixed dilation with the box defined by
∣ x ⋅ ϕ d ( i ) ( t θ ) ∣ ≤ r ( m + 2 ) / d { q 1 m + 2 − i , 0 ≤ i ≤ m + 2 , u m + 2 − i , m + 2 < i ≤ d . |x\cdot\phi_d^{(i)}(t_\theta)|\leq r^{(m+2)/d}
\begin{cases}
q_1^{m+2-i}, & 0\leq i\leq m+2,\\
u^{m+2-i}, & m+2<i\leq d.
\end{cases} ∣ x ⋅ ϕ d ( i ) ( t θ ) ∣ ≤ r ( m + 2 ) / d { q 1 m + 2 − i , u m + 2 − i , 0 ≤ i ≤ m + 2 , m + 2 < i ≤ d .
In directions 0 ≤ i ≤ m + 1 0\leq i\leq m+1 0 ≤ i ≤ m + 1 , the convolution kernel η ~ θ , σ m + 2 \widetilde{\eta}^{m+2}_{\theta,\sigma} η θ , σ m + 2 decays at spatial scales r i / d σ i − m − 2 = r ( m + 2 ) / d q 1 m + 2 − i r^{i/d}\sigma^{i-m-2}=r^{(m+2)/d}q_1^{m+2-i} r i / d σ i − m − 2 = r ( m + 2 ) / d q 1 m + 2 − i , which are the corresponding widths of B τ 1 B_{\tau_1} B τ 1 . In directions m + 2 ≤ i ≤ d m+2\leq i\leq d m + 2 ≤ i ≤ d , the cutoff factors equal one on supp ∣ g θ ∣ 2 \operatorname{supp}|g_\theta|^2 supp ∣ g θ ∣ 2 , and the widths of B τ 1 B_{\tau_1} B τ 1 satisfy r ( m + 2 ) / d u m + 2 − i ≤ r i / d r^{(m+2)/d}u^{m+2-i}\leq r^{i/d} r ( m + 2 ) / d u m + 2 − i ≤ r i / d . The locally constant property of g θ g_\theta g θ therefore gives ∣ g θ ∣ 2 ∗ η ~ θ , σ m + 2 ≲ ∣ g θ ∣ 2 ∗ w B τ 1 |g_\theta|^2*\widetilde{\eta}^{m+2}_{\theta,\sigma}\lesssim|g_\theta|^2*w_{B_{\tau_1}} ∣ g θ ∣ 2 ∗ η θ , σ m + 2 ≲ ∣ g θ ∣ 2 ∗ w B τ 1 for each θ ⊂ τ 1 \theta\subset\tau_1 θ ⊂ τ 1 . Summing over θ \theta θ proves the pointwise bound in (128) .
Taylor expansion as in (8.25) also gives σ − ( m + 2 ) supp h τ 1 ^ ⊂ ω τ 1 \sigma^{-(m+2)}\operatorname{supp}\widehat{h_{\tau_1}}\subset\omega_{\tau_1} σ − ( m + 2 ) supp h τ 1 ⊂ ω τ 1 for a cap ω τ 1 ∈ Ξ m + 1 d + 1 ( r σ d ) \omega_{\tau_1}\in\Xi^{d+1}_{m+1}(r\sigma^d) ω τ 1 ∈ Ξ m + 1 d + 1 ( r σ d ) . We therefore apply Lemma 9.1 to h ( x ) = ∑ τ 1 h τ 1 ( σ − ( m + 2 ) x ) h(x)=\sum_{\tau_1}h_{\tau_1}(\sigma^{-(m+2)}x) h ( x ) = ∑ τ 1 h τ 1 ( σ − ( m + 2 ) x ) , with parameter r σ d r\sigma^d r σ d , pieces h ω τ 1 ( x ) = h τ 1 ( σ − ( m + 2 ) x ) h_{\omega_{\tau_1}}(x)=h_{\tau_1}(\sigma^{-(m+2)}x) h ω τ 1 ( x ) = h τ 1 ( σ − ( m + 2 ) x ) , and amplitude α / r ϵ \alpha/r^\epsilon α / r ϵ . Returning to the original coordinates by y = σ − ( m + 2 ) x y=\sigma^{-(m+2)}x y = σ − ( m + 2 ) x multiplies all spatial widths by σ − ( m + 2 ) \sigma^{-(m+2)} σ − ( m + 2 ) .
In the first sum of Lemma 9.1 at parameter r σ d r\sigma^d r σ d , denote the intermediate caps by τ l ′ \tau'_l τ l ′ , the convolution weights by w B τ l ′ w_{B_{\tau'_l}} w B τ l ′ , and the averaging boxes by B ′ B' B ′ . Since (9.6) gives the first convolution with w B τ 1 w_{B_{\tau_1}} w B τ 1 , returning by y = σ − ( m + 2 ) x y=\sigma^{-(m+2)}x y = σ − ( m + 2 ) x , we write τ l = σ m + 2 τ l − 1 ′ \tau_l=\sigma^{m+2}\tau'_{l-1} τ l = σ m + 2 τ l − 1 ′ and B = σ − ( m + 2 ) B ′ B=\sigma^{-(m+2)}B' B = σ − ( m + 2 ) B ′ , where 2 ≤ l ≤ d − m − 2 2\leq l\leq d-m-2 2 ≤ l ≤ d − m − 2 and ∣ τ l − 1 ′ ∣ = q l |\tau'_{l-1}|=q_l ∣ τ l − 1 ′ ∣ = q l .
The i i i th widths of σ − ( m + 2 ) B τ l − 1 ′ \sigma^{-(m+2)}B_{\tau'_{l-1}} σ − ( m + 2 ) B τ l − 1 ′ satisfy
σ − ( m + 2 ) ( r σ d ) ( m + 3 ) / d q l m + l + 1 − i q 2 ⋯ q l − 1 = r ( m + 2 ) / d q l m + l + 1 − i q 1 ⋯ q l − 1 , 0 ≤ i ≤ m + l + 1 , \sigma^{-(m+2)}\frac{(r\sigma^d)^{(m+3)/d}q_l^{m+l+1-i}}{q_2\cdots q_{l-1}}
=\frac{r^{(m+2)/d}q_l^{m+l+1-i}}{q_1\cdots q_{l-1}},
\qquad0\leq i\leq m+l+1, σ − ( m + 2 ) q 2 ⋯ q l − 1 ( r σ d ) ( m + 3 ) / d q l m + l + 1 − i = q 1 ⋯ q l − 1 r ( m + 2 ) / d q l m + l + 1 − i , 0 ≤ i ≤ m + l + 1 ,
σ − ( m + 2 ) ( r σ d ) ( m + 3 ) / d u m + l + 1 − i q 2 ⋯ q l − 1 = r ( m + 2 ) / d u m + l + 1 − i q 1 ⋯ q l − 1 , m + l + 1 < i ≤ d . \sigma^{-(m+2)}\frac{(r\sigma^d)^{(m+3)/d}u^{m+l+1-i}}{q_2\cdots q_{l-1}}
=\frac{r^{(m+2)/d}u^{m+l+1-i}}{q_1\cdots q_{l-1}},
\qquad m+l+1<i\leq d. σ − ( m + 2 ) q 2 ⋯ q l − 1 ( r σ d ) ( m + 3 ) / d u m + l + 1 − i = q 1 ⋯ q l − 1 r ( m + 2 ) / d u m + l + 1 − i , m + l + 1 < i ≤ d .
These are exactly the widths of B τ l B_{\tau_l} B τ l in (8.7), with R = r R=r R = r .
The i i i th widths of the averaging boxes B = σ − ( m + 2 ) B ′ B=\sigma^{-(m+2)}B' B = σ − ( m + 2 ) B ′ satisfy
σ − ( m + 2 ) ( r σ d ) ( m + 3 ) / d u d − i q 2 ⋯ q d − m − 2 = r ( m + 2 ) / d u d − i q 1 ⋯ q d − m − 2 , 0 ≤ i ≤ d . \sigma^{-(m+2)}\frac{(r\sigma^d)^{(m+3)/d}u^{d-i}}{q_2\cdots q_{d-m-2}}
=\frac{r^{(m+2)/d}u^{d-i}}{q_1\cdots q_{d-m-2}},
\qquad0\leq i\leq d. σ − ( m + 2 ) q 2 ⋯ q d − m − 2 ( r σ d ) ( m + 3 ) / d u d − i = q 1 ⋯ q d − m − 2 r ( m + 2 ) / d u d − i , 0 ≤ i ≤ d .
Hence B ∥ u d − m − 2 q 1 ⋯ q d − m − 2 ( u r 1 / d ) m U ρ , r B\parallel\dfrac{u^{d-m-2}}{q_1\cdots q_{d-m-2}}(ur^{1/d})^mU_{\rho,r} B ∥ q 1 ⋯ q d − m − 2 u d − m − 2 ( u r 1/ d ) m U ρ , r , exactly as in (8.8) with R = r R=r R = r .
Using (9.6) with this u u u , the innermost expressions satisfy
∣ h τ 1 ∣ p d − m / p d − m − 1 ≲ [ ( ∑ θ ⊂ τ 1 ∣ g θ ∣ 2 ) ∗ w B τ 1 ] p d − m / p d − m − 1 . \left|h_{\tau_1}\right|_{p_{d-m}/p_{d-m-1}}
\lesssim
\left[\left(\sum_{\theta\subset\tau_1}|g_\theta|^2\right)*w_{B_{\tau_1}}\right]_{p_{d-m}/p_{d-m-1}}. ∣ h τ 1 ∣ p d − m / p d − m − 1 ≲ [ ( θ ⊂ τ 1 ∑ ∣ g θ ∣ 2 ) ∗ w B τ 1 ] p d − m / p d − m − 1 .
Substituting this bound into the first sum on the RHS of Lemma 9.1, after returning to the original coordinates, gives the convolution weights w B τ l w_{B_{\tau_l}} w B τ l and successive powers p d − m / p d − m − 1 , … , p 3 / p 2 p_{d-m}/p_{d-m-1},\ldots,p_3/p_2 p d − m / p d − m − 1 , … , p 3 / p 2 appearing in (8.8). The counts computed from { h τ 1 } \{h_{\tau_1}\} { h τ 1 } are no larger than those from { g θ } \{g_\theta\} { g θ } , so using the latter N ( u ) N(u) N ( u ) only lowers the selection thresholds. Hence the first sum in Lemma 9.1, after returning to the original coordinates, is bounded by the terms on the RHS of (8.8) with R = r R=r R = r , f = g f=g f = g , and q 1 = r − 1 / d σ − 1 q_1=r^{-1/d}\sigma^{-1} q 1 = r − 1/ d σ − 1 .
For the second sum in Lemma 9.1, use (9.6) with u = q 1 u=q_1 u = q 1 . After the same change of coordinates, writing B = σ − ( m + 2 ) Q B=\sigma^{-(m+2)}Q B = σ − ( m + 2 ) Q , this sum is bounded by
≲ ∑ τ 1 ∈ J ( q 1 ) ∑ B ∈ G τ 1 ( α / r ϵ ) B ∥ σ − ( m + 2 ) ω τ 1 ∗ ∣ B ∣ [ ∫ B ( ∑ θ ⊂ τ 1 ∣ g θ ∣ 2 ) ∗ w B τ 1 ] p d − m / 4 (129) \lesssim\sum_{\tau_1\in\mathcal{J}(q_1)}\sum_{\substack{B\in\mathcal{G}_{\tau_1}(\alpha/r^\epsilon)\\ B\parallel\sigma^{-(m+2)}\omega^*_{\tau_1}}}|B|\left[\int_B\left(\sum_{\theta\subset\tau_1}|g_\theta|^2\right)*w_{B_{\tau_1}}\right]^{p_{d-m}/4} \tag*{(129)} ≲ τ 1 ∈ J ( q 1 ) ∑ B ∈ G τ 1 ( α / r ϵ ) B ∥ σ − ( m + 2 ) ω τ 1 ∗ ∑ ∣ B ∣ [ ∫ B ( θ ⊂ τ 1 ∑ ∣ g θ ∣ 2 ) ∗ w B τ 1 ] p d − m /4 (129)
Both B τ 1 B_{\tau_1} B τ 1 and the averaging boxes in (129) have dimensions σ − m − 2 × r 1 / d σ − m − 1 × ⋯ × r σ d − m − 2 \sigma^{-m-2}\times r^{1/d}\sigma^{-m-1}\times\cdots\times r^{\sigma d-m-2} σ − m − 2 × r 1/ d σ − m − 1 × ⋯ × r σ d − m − 2 . Thus these boxes are parallel to ( u r 1 / d ) m U τ 1 , r (ur^{1/d})^mU_{\tau_1,r} ( u r 1/ d ) m U τ 1 , r , where u = q 1 = r − 1 / d σ − 1 u=q_1=r^{-1/d}\sigma^{-1} u = q 1 = r − 1/ d σ − 1 . Lemma 9.2, applied to g g g with this u u u and ρ = τ 1 \rho=\tau_1 ρ = τ 1 , bounds (129) by the terms of (8.8) with q 1 = ⋯ = q d − m − 2 = u = r − 1 / d σ − 1 q_1=\cdots=q_{d-m-2}=u=r^{-1/d}\sigma^{-1} q 1 = ⋯ = q d − m − 2 = u = r − 1/ d σ − 1 , including the amplitude condition.
Summing these bounds over r − 1 / d < σ ≤ 1 r^{-1/d}<\sigma\leq1 r − 1/ d < σ ≤ 1 and using the bound for (9.5) in (9.4) proves (9.2). Lemma 8.5, applied to g g g with parameter r r r , bounds the RHS of (9.2) by the RHS of Lemma 8.10 and completes the proof.
The following consequence of Lemma 8.10 will be used in both Steps 1 and 2. We first introduce a common notation for the Fourier support conditions in these two steps. For R ≥ 1 R\geq1 R ≥ 1 , 0 ≤ j ≤ d − 2 0\leq j\leq d-2 0 ≤ j ≤ d − 2 , and R − 1 / d ≤ ℓ ≤ 1 R^{-1/d}\leq\ell\leq1 R − 1/ d ≤ ℓ ≤ 1 , define
Γ ~ j , ℓ d + 1 ( R ) : = ⋃ I ∈ J ( R − 1 / d ) { ∑ i = 0 d λ i ϕ d ( i ) ( t I ) : ∣ λ i ∣ ≲ R − j / d ℓ i − j , 0 ≤ i < j , ∣ λ j ∣ ∼ R − j / d , ∣ λ i ∣ ≲ R − i / d , j < i ≤ d } . (130) \widetilde{\Gamma}^{d+1}_{j,\ell}(R):=\bigcup_{I\in\mathcal{J}(R^{-1/d})}\left\{\sum_{i=0}^{d}\lambda_i\phi_d^{(i)}(t_I):\begin{array}{ll}|\lambda_i|\lesssim R^{-j/d}\ell^{i-j},&0\leq i<j,\\ |\lambda_j|\sim R^{-j/d},&\\ |\lambda_i|\lesssim R^{-i/d},&j<i\leq d\end{array}\right\}. \tag*{(130)} Γ j , ℓ d + 1 ( R ) := I ∈ J ( R − 1/ d ) ⋃ ⎩ ⎨ ⎧ i = 0 ∑ d λ i ϕ d ( i ) ( t I ) : ∣ λ i ∣ ≲ R − j / d ℓ i − j , ∣ λ j ∣ ∼ R − j / d , ∣ λ i ∣ ≲ R − i / d , 0 ≤ i < j , j < i ≤ d ⎭ ⎬ ⎫ . (130)
The supports Γ j , 1 / K d + 1 ( R ) \Gamma^{d+1}_{j,1/K}(R) Γ j , 1/ K d + 1 ( R ) in Step 1 and Γ j , 1 / K , ℓ d + 1 ( R ) \Gamma^{d+1}_{j,1/K,\ell}(R) Γ j , 1/ K , ℓ d + 1 ( R ) in Step 2 are contained in Γ ~ j , 1 d + 1 ( R ) \widetilde{\Gamma}^{d+1}_{j,1}(R) Γ j , 1 d + 1 ( R ) and Γ ~ j , ℓ d + 1 ( R ) \widetilde{\Gamma}^{d+1}_{j,\ell}(R) Γ j , ℓ d + 1 ( R ) , respectively.
Corollary 9.3. Assume the hypotheses of Proposition 8.7 for a fixed 0 ≤ m ≤ d − 3 0\leq m\leq d-3 0 ≤ m ≤ d − 3 . Let 0 ≤ j ≤ m 0\leq j\leq m 0 ≤ j ≤ m , R − 1 / d ≤ ℓ ≤ 1 R^{-1/d}\leq\ell\leq1 R − 1/ d ≤ ℓ ≤ 1 , and let f f f be a Schwartz function with supp f ^ ⊂ Γ ~ j , ℓ d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\widetilde{\Gamma}^{d+1}_{j,\ell}(R) supp f ⊂ Γ j , ℓ d + 1 ( R ) . For R − 1 / d ≤ σ ≪ ℓ R^{-1/d}\leq\sigma\ll\ell R − 1/ d ≤ σ ≪ ℓ , σ ≤ s ≤ 1 \sigma\leq s\leq1 σ ≤ s ≤ 1 , δ > 0 \delta>0 δ > 0 , and α > 0 \alpha>0 α > 0 , we have
∑ ρ ∈ J ( s ) ∑ Q ∈ G ρ ( α ) ∣ Q ∣ = σ ∣ Q ∣ ∥ ( ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L p d − m / 2 ( Q ) p d − m ≲ δ σ − δ ∑ σ ≤ u ≤ s ∑ ρ ′ ∈ J ( u ) ∑ B ∈ G ρ ′ ( α σ − δ ) ∣ B ∣ = σ ∣ B ∣ ∥ ( ∑ τ ⊂ ρ ′ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L p d − m / 2 ( B ) p d − m . \sum_{\rho\in\mathcal{J}(s)}\sum_{\substack{Q\in\mathcal{G}_{\rho}(\alpha)\\ |Q|=\sigma}}\lvert Q\rvert\left\|\left(\sum_{\tau\subset\rho,\,|\tau|=\sigma}|f_\tau|^2\right)^{1/2}\right\|_{L^{p_{d-m}/2}(Q)}^{p_{d-m}}
\lesssim_\delta\sigma^{-\delta}\sum_{\sigma\leq u\leq s}\sum_{\rho'\in\mathcal{J}(u)}\sum_{\substack{B\in\mathcal{G}_{\rho'}\left(\frac{\alpha}{\sigma^{-\delta}}\right)\\ |B|=\sigma}}\lvert B\rvert\left\|\left(\sum_{\tau\subset\rho',\,|\tau|=\sigma}|f_\tau|^2\right)^{1/2}\right\|_{L^{p_{d-m}/2}(B)}^{p_{d-m}}. ρ ∈ J ( s ) ∑ Q ∈ G ρ ( α ) ∣ Q ∣ = σ ∑ ∣ Q ∣ τ ⊂ ρ , ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L p d − m /2 ( Q ) p d − m ≲ δ σ − δ σ ≤ u ≤ s ∑ ρ ′ ∈ J ( u ) ∑ B ∈ G ρ ′ ( σ − δ α ) ∣ B ∣ = σ ∑ ∣ B ∣ τ ⊂ ρ ′ , ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L p d − m /2 ( B ) p d − m .
Proof. We first prove the integral estimate in (8.11) under the support hypothesis of this corollary, that is,
∫ ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ≳ α 2 ( ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) p d − m / 2 ≲ δ σ − δ ∑ σ ≤ u ≤ 1 ∑ ρ ∈ J ( u ) ∑ B ∈ G ρ ( α / σ − δ ) ∣ B ∣ = σ ∣ B ∣ ∥ ( ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L p d − m / 2 ( B ) p d − m . (131) \int_{\sum_{|\tau|=\sigma}|f_\tau|^2\gtrsim\alpha^2}\left(\sum_{|\tau|=\sigma}|f_\tau|^2\right)^{p_{d-m}/2}\lesssim_\delta\sigma^{-\delta}\sum_{\sigma\leq u\leq1}\sum_{\rho\in\mathcal{J}(u)}\sum_{\substack{B\in\mathcal{G}_{\rho}(\alpha/\sigma^{-\delta})\\ |B|=\sigma}}|B|\left\|\left(\sum_{\tau\subset\rho,\,|\tau|=\sigma}|f_\tau|^2\right)^{1/2}\right\|_{L^{p_{d-m}/2}(B)}^{p_{d-m}}. \tag*{(131)} ∫ ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ≳ α 2 ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 p d − m /2 ≲ δ σ − δ σ ≤ u ≤ 1 ∑ ρ ∈ J ( u ) ∑ B ∈ G ρ ( α / σ − δ ) ∣ B ∣ = σ ∑ ∣ B ∣ τ ⊂ ρ , ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L p d − m /2 ( B ) p d − m . (131)
For τ ∈ J ( σ ) \tau\in\mathcal{J}(\sigma) τ ∈ J ( σ ) , write ξ = ∑ i = 0 d λ i ϕ d ( i ) ( t τ ) ∈ supp f τ ^ \xi=\sum_{i=0}^{d}\lambda_i\phi_d^{(i)}(t_\tau)\in\operatorname{supp}\widehat{f_\tau} ξ = ∑ i = 0 d λ i ϕ d ( i ) ( t τ ) ∈ supp f τ . Since f τ = ∑ θ ⊂ τ f θ f_\tau=\sum_{\theta\subset\tau}f_\theta f τ = ∑ θ ⊂ τ f θ and ∣ t θ − t τ ∣ ≲ σ |t_\theta-t_\tau|\lesssim\sigma ∣ t θ − t τ ∣ ≲ σ , Taylor expansion and (9.8) give
∣ R j / d σ j λ i ∣ ≲ σ i ( 0 ≤ i < d ) , ∣ R j / d σ j λ j ∣ ∼ σ j . |R^{j/d}\sigma^j\lambda_i|\lesssim\sigma^i\quad(0\leq i<d),\qquad|R^{j/d}\sigma^j\lambda_j|\sim\sigma^j. ∣ R j / d σ j λ i ∣ ≲ σ i ( 0 ≤ i < d ) , ∣ R j / d σ j λ j ∣ ∼ σ j .
The j j j th coefficient differs from its value at t θ t_\theta t θ by O ( R − j / d σ / ℓ ) O(R^{-j/d}\sigma/\ell) O ( R − j / d σ / ℓ ) , so σ ≪ ℓ \sigma\ll\ell σ ≪ ℓ gives the last comparison. Thus
supp f τ ^ ( R j / d σ − j ⋅ ) ⊂ Θ τ , Θ τ ∈ Ξ m d + 1 ( σ − d ) . \operatorname{supp}\widehat{f_\tau}(R^{j/d}\sigma^{-j}\cdot)\subset\Theta_\tau,\qquad\Theta_\tau\in\Xi_m^{d+1}(\sigma^{-d}). supp f τ ( R j / d σ − j ⋅ ) ⊂ Θ τ , Θ τ ∈ Ξ m d + 1 ( σ − d ) .
We therefore apply Lemma 8.10 to g ( x ) = f ( R j / d σ − j x ) g(x)=f(R^{j/d}\sigma^{-j}x) g ( x ) = f ( R j / d σ − j x ) , with parameter σ − d \sigma^{-d} σ − d , pieces f τ ( R j / d σ − j x ) f_\tau(R^{j/d}\sigma^{-j}x) f τ ( R j / d σ − j x ) , and amplitude α \alpha α .
As in the derivation of (8.11), returning to the original coordinates gives boxes parallel to ( u / σ ) m − j ( u R 1 / d ) j ( σ R 1 / d ) 2 U ρ , R (u/\sigma)^{m-j}\frac{(uR^{1/d})^j}{(\sigma R^{1/d})^2}U_{\rho,R} ( u / σ ) m − j ( σ R 1/ d ) 2 ( u R 1/ d ) j U ρ , R . Since ( u / σ ) m − j ≥ 1 (u/\sigma)^{m-j}\geq1 ( u / σ ) m − j ≥ 1 , Lemma 2.14, applied with F = ( ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ) p d − m / 4 F=\left(\sum_{\tau\subset\rho,\,|\tau|=\sigma}|f_\tau|^2\right)^{p_{d-m}/4} F = ( ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ) p d − m /4 , replaces these boxes by ( u R 1 / d ) j ( σ R 1 / d ) 2 U ρ , R \frac{(uR^{1/d})^j}{(\sigma R^{1/d})^2}U_{\rho,R} ( σ R 1/ d ) 2 ( u R 1/ d ) j U ρ , R while preserving the amplitude condition. This proves (9.9).
To prove the corollary, we first use the locally constant property when σ ∼ s \sigma\sim s σ ∼ s and then apply (9.9) when σ ≪ s \sigma\ll s σ ≪ s . For each ρ ∈ J ( s ) \rho\in\mathcal{J}(s) ρ ∈ J ( s ) and τ ⊂ ρ \tau\subset\rho τ ⊂ ρ of width σ \sigma σ , the Fourier support of f τ f_\tau f τ lies in a box of dimensions R − j / d σ − j × R − j / d σ 1 − j × ⋯ × R − j / d σ d − j R^{-j/d}\sigma^{-j} \times R^{-j/d}\sigma^{1-j} \times\cdots\times R^{-j/d}\sigma^{d-j} R − j / d σ − j × R − j / d σ 1 − j × ⋯ × R − j / d σ d − j . As in (8.12), the locally constant property bounds the left hand side of the corollary by the terms with u = s u=s u = s on its right hand side.
Now suppose that σ ≪ s \sigma\ll s σ ≪ s . By (2.22) with F = ( ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ) p d − m / 4 F=\left(\sum_{\tau\subset\rho,\lvert\tau\rvert=\sigma}\lvert f_\tau\rvert^2\right)^{p_{d-m}/4} F = ( ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ) p d − m /4 and t = [ α / N ( s ) ] p d − m / 2 t=[\alpha/N(s)]^{p_{d-m}/2} t = [ α / N ( s ) ] p d − m /2 , it suffices to bound
∑ ρ ∈ J ( s ) ∫ ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ≥ [ α / N ( s ) ] 2 ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) p d − m / 2 . \sum_{\rho\in\mathcal{J}(s)}\int_{\sum_{\tau\subset\rho,\lvert\tau\rvert=\sigma}\lvert f_\tau\rvert^2\geq[\alpha/N(s)]^2}\left(\sum_{\substack{\tau\subset\rho\\\lvert\tau\rvert=\sigma}}\lvert f_\tau\rvert^2\right)^{p_{d-m}/2}. ρ ∈ J ( s ) ∑ ∫ ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ≥ [ α / N ( s ) ] 2 τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 p d − m /2 .
For ρ = [ t ρ , t ρ + s ] ∈ J ( s ) \rho=[t_\rho,t_\rho+s]\in\mathcal{J}(s) ρ = [ t ρ , t ρ + s ] ∈ J ( s ) and each cap θ ⊂ ρ \theta\subset\rho θ ⊂ ρ at scale R − 1 / d R^{-1/d} R − 1/ d , the Fourier change of variables is
ξ = ∑ i = 0 d λ i ϕ d ( i ) ( t θ ) ⟹ L t ρ , s − 1 ξ = ∑ i = 0 d s − i λ i ϕ d ( i ) ( t θ − t ρ s ) . \xi=\sum_{i=0}^{d}\lambda_i\phi_d^{(i)}(t_\theta)
\quad\Longrightarrow\quad
L_{t_\rho,s}^{-1}\xi=\sum_{i=0}^{d}s^{-i}\lambda_i\phi_d^{(i)}\left(\frac{t_\theta-t_\rho}{s}\right). ξ = i = 0 ∑ d λ i ϕ d ( i ) ( t θ ) ⟹ L t ρ , s − 1 ξ = i = 0 ∑ d s − i λ i ϕ d ( i ) ( s t θ − t ρ ) .
Hence supp f ρ ( L t ρ , s − T ⋅ ) ^ ⊂ Γ ~ j , min { ℓ / s , 1 } d + 1 ( R s d ) \operatorname{supp}\widehat{f_\rho(L_{t_\rho,s}^{-T}\cdot)}\subset\widetilde{\Gamma}_{j,\min\{\ell/s,1\}}^{d+1}(R s^d) supp f ρ ( L t ρ , s − T ⋅ ) ⊂ Γ j , m i n { ℓ / s , 1 } d + 1 ( R s d ) . Since σ / s ≪ min { ℓ / s , 1 } \sigma/s\ll\min\{\ell/s,1\} σ / s ≪ min { ℓ / s , 1 } , we apply (9.9) to this function with parameter R s d R s^d R s d , pieces f τ ( L t ρ , s − T ⋅ ) f_\tau(L_{t_\rho,s}^{-T}\cdot) f τ ( L t ρ , s − T ⋅ ) of width σ / s \sigma/s σ / s , and amplitude α / N ( s ) \alpha/N(s) α / N ( s ) .
In the resulting sum, denote the parent of width u / s u/s u / s by ρ ′ ′ \rho^{\prime\prime} ρ ′′ and the averaging boxes by B ′ B' B ′ . Returning by x = L t ρ , s y x=L_{t_\rho,s}y x = L t ρ , s y , write ρ ′ = L t ρ , s ρ ′ ′ ⊂ ρ \rho'=L_{t_\rho,s}\rho^{\prime\prime}\subset\rho ρ ′ = L t ρ , s ρ ′′ ⊂ ρ , with ∣ ρ ′ ∣ = u \lvert\rho'\rvert=u ∣ ρ ′ ∣ = u , and B = L t ρ , s B ′ B=L_{t_\rho,s}B' B = L t ρ , s B ′ . The i i i th width of B B B is s − i ( R s d ) j / d ( σ / s ) − 2 ( u / s ) j + 2 − i = R j / d σ − 2 u j + 2 − i s^{-i}(R s^d)^{j/d}(\sigma/s)^{-2}(u/s)^{j+2-i}=R^{j/d}\sigma^{-2}u^{j+2-i} s − i ( R s d ) j / d ( σ / s ) − 2 ( u / s ) j + 2 − i = R j / d σ − 2 u j + 2 − i for 0 ≤ i ≤ d 0\leq i\leq d 0 ≤ i ≤ d . Thus B ∥ ( u R 1 / d σ R 1 / d ) 2 U ρ ′ , R B\parallel\left(\frac{uR^{1/d}}{\sigma R^{1/d}}\right)^2U_{\rho',R} B ∥ ( σ R 1/ d u R 1/ d ) 2 U ρ ′ , R , as in the corollary.
The active parent count at rescaled width u / s u/s u / s is N ρ ( u ) N_\rho(u) N ρ ( u ) . Using N ( s ) N ρ ( u ) ≲ N ( u ) N(s)N_\rho(u)\lesssim N(u) N ( s ) N ρ ( u ) ≲ N ( u ) from Lemma 2.18 and ( σ / s ) − δ ≤ σ − δ (\sigma/s)^{-\delta}\leq\sigma^{-\delta} ( σ / s ) − δ ≤ σ − δ , we obtain α ( σ / s ) δ N ( s ) N ρ ( u ) ≳ α σ δ N ( u ) \frac{\alpha}{(\sigma/s)^\delta N(s)N_\rho(u)}\gtrsim\frac{\alpha}{\sigma^\delta N(u)} ( σ / s ) δ N ( s ) N ρ ( u ) α ≳ σ δ N ( u ) α in the selection condition. Summing over ρ \rho ρ proves the corollary. □
Proof of Lemma 8.11
We now prove the estimate in Lemma 8.11. We allow Fourier support in Γ ~ j , ℓ d + 1 ( R ) \widetilde{\Gamma}_{j,\ell}^{d+1}(R) Γ j , ℓ d + 1 ( R ) , so that the same result applies to the coefficient shells in Step 2. Taking ℓ = 1 \ell=1 ℓ = 1 gives Lemma 8.11. The proof uses the coefficient bounds and the following square function estimate for Taylor cones in R d + 1 \mathbb{R}^{d+1} R d + 1 .
Proposition 9.4 ((GM23, Theorem 1, Proposition 1.1)). Let 0 ≤ k ≤ d − 2 0\leq k\leq d-2 0 ≤ k ≤ d − 2 and 2 ≤ p ≤ p d − k 2\leq p\leq p_{d-k} 2 ≤ p ≤ p d − k . For a Schwartz function g g g with supp g ^ ⊂ Γ ~ k d + 1 ( r ) \operatorname{supp}\widehat{g}\subset\widetilde{\Gamma}_{k}^{d+1}(r) supp g ⊂ Γ k d + 1 ( r ) ,
∫ ∣ g ∣ p ≲ ϵ r ϵ ∫ ( ∑ θ ∈ Γ ~ k d + 1 ( r ) ∣ g θ ∣ 2 ) p / 2 . \int\lvert g\rvert^p\lesssim_\epsilon r^\epsilon\int\left(\sum_{\theta\in\widetilde{\Gamma}_{k}^{d+1}(r)}\lvert g_\theta\rvert^2\right)^{p/2}. ∫ ∣ g ∣ p ≲ ϵ r ϵ ∫ θ ∈ Γ k d + 1 ( r ) ∑ ∣ g θ ∣ 2 p /2 .
Lemma 9.5. Let 0 ≤ j ≤ m ≤ d − 3 0\leq j\leq m\leq d-3 0 ≤ j ≤ m ≤ d − 3 and R − 1 / d ≤ ℓ ≤ 1 R^{-1/d}\leq\ell\leq1 R − 1/ d ≤ ℓ ≤ 1 . Let f f f be a Schwartz function with supp f ^ ⊂ Γ ~ j , ℓ d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\widetilde{\Gamma}_{j,\ell}^{d+1}(R) supp f ⊂ Γ j , ℓ d + 1 ( R ) . Let R − 1 / d < σ ≪ ℓ R^{-1/d}<\sigma\ll\ell R − 1/ d < σ ≪ ℓ and α , δ > 0 \alpha,\delta>0 α , δ > 0 . For η > 0 \eta>0 η > 0 sufficiently small and σ \sigma σ sufficiently small in terms of η \eta η , there are dyadic scales σ ′ , s ′ \sigma',s' σ ′ , s ′ with R − 1 / d ≤ σ ′ ≤ max { R − 1 / d , σ 3 / 2 } R^{-1/d}\leq\sigma'\leq\max\{R^{-1/d},\sigma^{3/2}\} R − 1/ d ≤ σ ′ ≤ max { R − 1/ d , σ 3/2 } and σ ≤ s ′ ≲ σ η \sigma\leq s'\lesssim\sigma^\eta σ ≤ s ′ ≲ σ η such that
∑ B ∥ [ 0 , R j / d σ − 2 ] d + 1 B ∈ G [ 0 , 1 ] ( α ) ∣ B ∣ ∥ ( ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L # p d − m ( B ) p d − m ≲ δ , η σ − C η 4 ∑ Q ∈ G [ 0 , 1 ] d + 1 ( α σ C η 4 ) ∣ Q ∣ = σ ′ ∣ Q ∣ ∥ ( ∑ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ) 1 / 2 ∥ L # p d − m ( Q ) p d − m + σ − δ ∑ ρ ∈ J ( s ′ ) ∑ Q ∈ G ρ ( α σ − δ ) Q ⊂ ρ ∣ Q ∣ ∥ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L # p d − m ( Q ) p d − m . \sum_{\substack{B\parallel[0,R^{j/d}\sigma^{-2}]^{d+1}\\B\in\mathcal{G}_{[0,1]}(\alpha)}}\lvert B\rvert\left\|\left(\sum_{\lvert\tau\rvert=\sigma}\lvert f_\tau\rvert^2\right)^{1/2}\right\|_{L_{\#}^{p_{d-m}}(B)}^{p_{d-m}}
\lesssim_{\delta,\eta}\sigma^{-C\eta^4}
\sum_{\substack{Q\in\mathcal{G}_{[0,1]^{d+1}}\left(\frac{\alpha}{\sigma^{C\eta^4}}\right)\\\lvert Q\rvert=\sigma'}}
\lvert Q\rvert\left\|\left(\sum_{\lvert\tau'\rvert=\sigma'}\lvert f_{\tau'}\rvert^2\right)^{1/2}\right\|_{L_{\#}^{p_{d-m}}(Q)}^{p_{d-m}}
+\sigma^{-\delta}\sum_{\rho\in\mathcal{J}(s')}\sum_{\substack{Q\in\mathcal{G}_\rho\left(\frac{\alpha}{\sigma^{-\delta}}\right)\\Q\subset\rho}}
\lvert Q\rvert\left\|\left(\sum_{\substack{\tau\subset\rho\\\lvert\tau\rvert=\sigma}}\lvert f_\tau\rvert^2\right)^{1/2}\right\|_{L_{\#}^{p_{d-m}}(Q)}^{p_{d-m}}. B ∥ [ 0 , R j / d σ − 2 ] d + 1 B ∈ G [ 0 , 1 ] ( α ) ∑ ∣ B ∣ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L # p d − m ( B ) p d − m ≲ δ , η σ − C η 4 Q ∈ G [ 0 , 1 ] d + 1 ( σ C η 4 α ) ∣ Q ∣ = σ ′ ∑ ∣ Q ∣ ∣ τ ′ ∣ = σ ′ ∑ ∣ f τ ′ ∣ 2 1/2 L # p d − m ( Q ) p d − m + σ − δ ρ ∈ J ( s ′ ) ∑ Q ∈ G ρ ( σ − δ α ) Q ⊂ ρ ∑ ∣ Q ∣ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L # p d − m ( Q ) p d − m .
Proof. For τ ∈ J ( σ ) \tau\in\mathcal{J}(\sigma) τ ∈ J ( σ ) , write ξ = ∑ i = 0 d λ i ϕ d ( i ) ( t τ ) ∈ supp f τ ^ \xi=\sum_{i=0}^{d}\lambda_i\phi_d^{(i)}(t_\tau)\in\operatorname{supp}\widehat{f_\tau} ξ = ∑ i = 0 d λ i ϕ d ( i ) ( t τ ) ∈ supp f τ . Taylor expansion using the coefficient bounds in (9.8), together with σ ≪ ℓ \sigma\ll\ell σ ≪ ℓ , gives
∣ λ i ∣ ≲ { R − j / d ℓ i − j , 0 ≤ i < j , R − j / d σ i − j , j ≤ i ≤ d , ∣ λ j ∣ ∼ R − j / d . \lvert\lambda_i\rvert\lesssim
\begin{cases}
R^{-j/d}\ell^{i-j}, & 0\leq i<j,\\
R^{-j/d}\sigma^{i-j}, & j\leq i\leq d,
\end{cases}
\qquad
\lvert\lambda_j\rvert\sim R^{-j/d}. ∣ λ i ∣ ≲ { R − j / d ℓ i − j , R − j / d σ i − j , 0 ≤ i < j , j ≤ i ≤ d , ∣ λ j ∣ ∼ R − j / d .
For σ ≤ u ≤ 1 \sigma\leq u\leq1 σ ≤ u ≤ 1 , choose a nonnegative, L 1 L^1 L 1 normalized kernel w σ , u w_{\sigma,u} w σ , u adapted to the box of all x ∈ R d + 1 x\in\mathbb{R}^{d+1} x ∈ R d + 1 satisfying
∣ x ⋅ ϕ d ( i ) ( t τ ) ∣ ≲ { R j / d σ − 2 u d − i , 0 ≤ i ≤ j + 1 , R j / d σ − 2 u d , j + 2 ≤ i ≤ d . \lvert x\cdot\phi_d^{(i)}(t_\tau)\rvert\lesssim
\begin{cases}
R^{j/d}\sigma^{-2}u^{d-i}, & 0\leq i\leq j+1,\\
R^{j/d}\sigma^{-2}u^{d}, & j+2\leq i\leq d.
\end{cases} ∣ x ⋅ ϕ d ( i ) ( t τ )∣ ≲ { R j / d σ − 2 u d − i , R j / d σ − 2 u d , 0 ≤ i ≤ j + 1 , j + 2 ≤ i ≤ d .
Then w τ , u ≲ w τ , 2 u w_{\tau,u} \lesssim w_{\tau,2u} w τ , u ≲ w τ , 2 u for dyadic σ ≤ u < 1 \sigma\le u < 1 σ ≤ u < 1 , and
supp w ^ τ , u ⊂ { ∑ i = 0 d λ i ϕ d ( i ) ( t τ ) : ∣ λ i ∣ ≲ { R − j / d σ 2 u i − d , 0 ≤ i ≤ j + 1 , R − j / d σ 2 u − d , j + 2 ≤ i ≤ d } . \operatorname{supp}\widehat{w}_{\tau,u} \subset\left\{\sum_{i=0}^{d}\lambda_i\phi_d^{(i)}(t_\tau): |\lambda_i| \lesssim\begin{cases} R^{-j/d}\sigma^2u^{i-d}, & 0 \le i \le j+1, \\ R^{-j/d}\sigma^2u^{-d}, & j+2 \le i \le d \end{cases}\right\}. supp w τ , u ⊂ { i = 0 ∑ d λ i ϕ d ( i ) ( t τ ) : ∣ λ i ∣ ≲ { R − j / d σ 2 u i − d , R − j / d σ 2 u − d , 0 ≤ i ≤ j + 1 , j + 2 ≤ i ≤ d } .
At u = σ u=\sigma u = σ , the kernel widths satisfy R j / d σ d − i − 2 ≤ R j / d σ j − i R^{j/d}\sigma^{d-i-2} \le R^{j/d}\sigma^{j-i} R j / d σ d − i − 2 ≤ R j / d σ j − i for 0 ≤ i ≤ j + 1 0 \le i \le j+1 0 ≤ i ≤ j + 1 , and R j / d σ d − 2 ≤ R j / d σ j − i R^{j/d}\sigma^{d-2} \le R^{j/d}\sigma^{j-i} R j / d σ d − 2 ≤ R j / d σ j − i for j + 2 ≤ i ≤ d j+2 \le i \le d j + 2 ≤ i ≤ d . The locally constant property therefore gives
∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ≲ ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , σ . (132) \sum_{|\tau|=\sigma}|f_\tau|^2 \lesssim\sum_{|\tau|=\sigma}|f_\tau|^2*w_{\tau,\sigma}. \tag*{(132)} ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ≲ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ , σ . (132)
We divide the boxes B ∈ G [ 0 , 1 ] ( α ) B\in\mathcal{G}_{[0,1]}(\alpha) B ∈ G [ 0 , 1 ] ( α ) , B ∥ [ 0 , R j / d σ − 2 d + 1 ] B\parallel[0,R^{j/d}\sigma^{-2}d+1] B ∥ [ 0 , R j / d σ − 2 d + 1 ] , according to whether
∥ ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∥ L # p d − m / 4 ( B ) ≤ σ − η 4 ∥ ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , u ∥ L # p d − m / 4 ( B ) (133) \left\|\sum_{|\tau|=\sigma}|f_\tau|^2\right\|_{L_{\#}^{p_d-m/4}(B)}
\le\sigma^{-\eta^4}\left\|\sum_{|\tau|=\sigma}|f_\tau|^2*w_{\tau,u}\right\|_{L_{\#}^{p_d-m/4}(B)} \tag*{(133)} ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 L # p d − m /4 ( B ) ≤ σ − η 4 ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ , u L # p d − m /4 ( B ) (133)
holds for some dyadic σ η 4 ≤ u ≤ 1 \sigma^{\eta^4}\le u\le1 σ η 4 ≤ u ≤ 1 . If this holds, we will bound the RHS using finer pieces f τ ′ f_{\tau'} f τ ′ , with ∣ τ ′ ∣ = σ ′ < σ |\tau'|=\sigma'<\sigma ∣ τ ′ ∣ = σ ′ < σ , to obtain the first term in Lemma 9.5. For the remaining boxes, we will use a further high frequency decomposition and Proposition 9.4 to obtain the second term, with parents of width s ′ ≲ σ η 4 s'\lesssim\sigma^{\eta^4} s ′ ≲ σ η 4 . After summing over B B B , dyadic pigeonholing will give common scales σ ′ \sigma' σ ′ and s ′ s' s ′ .
Case 1. Suppose that (9.11) holds for some dyadic u ∈ [ σ η 4 , 1 ] u\in[\sigma^{\eta^4},1] u ∈ [ σ η 4 , 1 ] . We first determine which products f θ f θ ′ ‾ f_\theta\overline{f_{\theta'}} f θ f θ ′ , θ , θ ′ ⊂ τ \theta,\theta'\subset\tau θ , θ ′ ⊂ τ , survive convolution with w τ , u w_{\tau,u} w τ , u . Take ξ ∈ supp f θ ^ \xi\in\operatorname{supp}\widehat{f_\theta} ξ ∈ supp f θ and ξ ′ ∈ supp f θ ′ ^ \xi'\in\operatorname{supp}\widehat{f_{\theta'}} ξ ′ ∈ supp f θ ′ with ξ − ξ ′ ∈ supp w ^ τ , u \xi-\xi'\in\operatorname{supp}\widehat{w}_{\tau,u} ξ − ξ ′ ∈ supp w τ , u . Let t θ , t θ ′ t_\theta,t_{\theta'} t θ , t θ ′ be the left endpoints of the intervals indexing θ , θ ′ \theta,\theta' θ , θ ′ , respectively, and write
ξ = ∑ i = 0 d λ i ( t θ ) ϕ d ( i ) ( t θ ) , ξ ′ = ∑ i = 0 d λ i ( t θ ′ ) ϕ d ( i ) ( t θ ′ ) . \xi=\sum_{i=0}^{d}\lambda_i(t_\theta)\phi_d^{(i)}(t_\theta),\qquad
\xi'=\sum_{i=0}^{d}\lambda_i(t_{\theta'})\phi_d^{(i)}(t_{\theta'}). ξ = i = 0 ∑ d λ i ( t θ ) ϕ d ( i ) ( t θ ) , ξ ′ = i = 0 ∑ d λ i ( t θ ′ ) ϕ d ( i ) ( t θ ′ ) .
Since ∣ t θ ′ − t τ ∣ ≤ σ ≤ u |t_{\theta'}-t_\tau|\le\sigma\le u ∣ t θ ′ − t τ ∣ ≤ σ ≤ u , applying the support bounds for w ^ τ , u \widehat{w}_{\tau,u} w τ , u to the coefficient of ϕ d ( j + 1 ) ( t θ ′ ) \phi_d^{(j+1)}(t_{\theta'}) ϕ d ( j + 1 ) ( t θ ′ ) in the Taylor expansion of ξ − ξ ′ \xi-\xi' ξ − ξ ′ gives
∣ λ j + 1 ( t θ ) − λ j + 1 ( t θ ′ ) + λ j ( t θ ) ( t θ − t θ ′ ) + ∑ i < j λ i ( t θ ) ( t θ − t θ ′ ) j + 1 − i ( j + 1 − i ) ! ∣ ≲ R − j / d σ 2 u j + 1 − d . \left|\lambda_{j+1}(t_\theta)-\lambda_{j+1}(t_{\theta'})+\lambda_j(t_\theta)(t_\theta-t_{\theta'})+\sum_{i<j}\lambda_i(t_\theta)\frac{(t_\theta-t_{\theta'})^{j+1-i}}{(j+1-i)!}\right|\lesssim R^{-j/d}\sigma^2u^{j+1-d}. λ j + 1 ( t θ ) − λ j + 1 ( t θ ′ ) + λ j ( t θ ) ( t θ − t θ ′ ) + i < j ∑ λ i ( t θ ) ( j + 1 − i )! ( t θ − t θ ′ ) j + 1 − i ≲ R − j / d σ 2 u j + 1 − d .
By (9.8) and ∣ t θ − t θ ′ ∣ ≤ σ |t_\theta-t_{\theta'}|\le\sigma ∣ t θ − t θ ′ ∣ ≤ σ ,
∣ ∑ i < j λ i ( t θ ) ( t θ − t θ ′ ) j + 1 − i ( j + 1 − i ) ! ∣ ≲ σ ℓ R − j / d ∣ t θ − t θ ′ ∣ ≪ R − j / d ∣ t θ − t θ ′ ∣ . \left|\sum_{i<j}\lambda_i(t_\theta)\frac{(t_\theta-t_{\theta'})^{j+1-i}}{(j+1-i)!}\right|\lesssim\frac{\sigma}{\ell}R^{-j/d}|t_\theta-t_{\theta'}|\ll R^{-j/d}|t_\theta-t_{\theta'}|. i < j ∑ λ i ( t θ ) ( j + 1 − i )! ( t θ − t θ ′ ) j + 1 − i ≲ ℓ σ R − j / d ∣ t θ − t θ ′ ∣ ≪ R − j / d ∣ t θ − t θ ′ ∣.
Together with ∣ λ j ( t θ ) ∣ ∼ R − j / d |\lambda_j(t_\theta)|\sim R^{-j/d} ∣ λ j ( t θ ) ∣ ∼ R − j / d and ∣ λ j + 1 ( t θ ) ∣ + ∣ λ j + 1 ( t θ ′ ) ∣ ≲ R − ( j + 1 ) / d |\lambda_{j+1}(t_\theta)|+|\lambda_{j+1}(t_{\theta'})|\lesssim R^{-(j+1)/d} ∣ λ j + 1 ( t θ ) ∣ + ∣ λ j + 1 ( t θ ′ ) ∣ ≲ R − ( j + 1 ) / d , this gives
∣ t θ − t θ ′ ∣ ≲ max { R − 1 / d , σ 2 u − ( d − j − 1 ) } . (134) |t_\theta-t_{\theta'}|\lesssim\max\{R^{-1/d},\sigma^2u^{-(d-j-1)}\}. \tag*{(134)} ∣ t θ − t θ ′ ∣ ≲ max { R − 1/ d , σ 2 u − ( d − j − 1 ) } . (134)
Since u ≥ σ η 4 u\ge\sigma^{\eta^4} u ≥ σ η 4 and η \eta η is sufficiently small, we may choose a dyadic σ ′ ∼ max { R − 1 / d , σ 2 u − ( d − j − 1 ) } \sigma'\sim\max\{R^{-1/d},\sigma^2u^{-(d-j-1)}\} σ ′ ∼ max { R − 1/ d , σ 2 u − ( d − j − 1 ) } with R − 1 / d ≤ σ ′ ≤ max { R − 1 / d , σ 3 / 2 } < σ R^{-1/d}\le\sigma'\le\max\{R^{-1/d},\sigma^{3/2}\}<\sigma R − 1/ d ≤ σ ′ ≤ max { R − 1/ d , σ 3/2 } < σ . Only neighbouring σ ′ \sigma' σ ′ intervals interact, so Cauchy–Schwarz gives
∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , u ≲ ∑ ∣ τ ∣ = σ ∑ ∣ τ ′ ∣ = σ ′ τ ′ ∼ τ ∣ f τ ′ ∣ 2 ∗ w τ ′ , u . \sum_{|\tau|=\sigma}|f_\tau|^2*w_{\tau,u}\lesssim\sum_{|\tau|=\sigma}\sum_{\substack{|\tau'|=\sigma'\\ \tau'\sim\tau}}|f_{\tau'}|^2*w_{\tau',u}. ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ , u ≲ ∣ τ ∣ = σ ∑ ∣ τ ′ ∣ = σ ′ τ ′ ∼ τ ∑ ∣ f τ ′ ∣ 2 ∗ w τ ′ , u .
To combine these terms into one convolution, let w w w be a nonnegative, L 1 L^1 L 1 normalized radial kernel with sufficiently rapid decay at scale R j / d σ − 2 u d − j − 1 R^{j/d}\sigma^{-2}u^{d-j-1} R j / d σ − 2 u d − j − 1 , so that every width of w τ , u w_{\tau,u} w τ , u is at most this radius. The ratio of the ball volume to the kernel box volume is at most u − C u^{-C} u − C , so choosing w w w to decay more slowly than w τ , u w_{\tau,u} w τ , u gives w τ , u ≲ u − C w ≲ σ − C η 4 w w_{\tau,u}\lesssim u^{-C}w\lesssim\sigma^{-C\eta^4}w w τ , u ≲ u − C w ≲ σ − C η 4 w . Substituting into (9.11) gives
∥ ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∥ L # p d − m / 4 ( B ) ≲ σ − C η 4 ∥ ( ∑ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ) ∗ w ∥ L # p d − m / 4 ( B ) . (135) \left\|\sum_{|\tau|=\sigma}|f_\tau|^2\right\|_{L_{\#}^{p_d-m/4}(B)}
\lesssim\sigma^{-C\eta^4}\left\|\left(\sum_{|\tau'|=\sigma'}|f_{\tau'}|^2\right)*w\right\|_{L_{\#}^{p_d-m/4}(B)}. \tag*{(135)} ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 L # p d − m /4 ( B ) ≲ σ − C η 4 ∣ τ ′ ∣ = σ ′ ∑ ∣ f τ ′ ∣ 2 ∗ w L # p d − m /4 ( B ) . (135)
Fix u u u , σ ′ \sigma' σ ′ , raise (135) to the power p d − m / 2 p_d-m/2 p d − m /2 , multiply by ∣ B ∣ |B| ∣ B ∣ , and sum over the boxes B B B satisfying (133) with this u u u . Applying (32) with F = [ ( ∑ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ) ∗ w ] p d − m / 4 F = [(\sum_{|\tau'|=\sigma'} |f_{\tau'}|^2) * w]^{p_d-m/4} F = [( ∑ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ) ∗ w ] p d − m /4 and t = [ α / σ − C η 4 ] p d − m / 2 t = [\alpha/\sigma-C\eta^4]^{p_d-m/2} t = [ α / σ − C η 4 ] p d − m /2 bounds their contribution to the LHS of Lemma 9.5 by
≲ η σ − C η 4 ∫ ( ∑ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ) ∗ w ≳ [ α / σ − C η 4 ] 2 [ ( ∑ ∣ τ ∣ = σ ′ ∣ f τ ∣ 2 ) ∗ w ] p d − m / 2 . (136) \lesssim_{\eta} \sigma^{-C\eta^4}\int_{(\sum_{|\tau'|=\sigma'} |f_{\tau'}|^2)*w\gtrsim[\alpha/\sigma-C\eta^4]^2}\left[\left(\sum_{|\tau|=\sigma'}|f_\tau|^2\right)*w\right]^{p_d-m/2}. \tag*{(136)} ≲ η σ − C η 4 ∫ ( ∑ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ) ∗ w ≳ [ α / σ − C η 4 ] 2 ∣ τ ∣ = σ ′ ∑ ∣ f τ ∣ 2 ∗ w p d − m /2 . (136)
Apply (34) with F = ∑ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 F=\sum_{|\tau'|=\sigma'}|f_{\tau'}|^2 F = ∑ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 , K = w K=w K = w , and exponent p d − m / 2 p_d-m/2 p d − m /2 . The locally constant property on unit cubes then bounds (136) by
≲ η σ − C η 4 ∑ Q ∥ [ 0 , 1 ] d + 1 Q ∈ G [ 0 , 1 ] ( α σ − C η 4 ) ∣ Q ∣ ∥ ( ∑ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ) 1 / 2 ∥ L # p d − m ( Q ) p d − m ∥ ( ∑ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( Q ) p d − m / 2 . (137) \lesssim_{\eta}\sigma^{-C\eta^4}\sum_{\substack{Q\parallel[0,1]^{d+1}\\Q\in\mathcal{G}_{[0,1]\left(\frac{\alpha}{\sigma-C\eta^4}\right)}}}|Q|\left\|\left(\sum_{|\tau'|=\sigma'}|f_{\tau'}|^2\right)^{1/2}\right\|_{L_{\#}^{p_d-m}(Q)}^{p_d-m}\left\|\left(\sum_{|\tau'|=\sigma'}|f_{\tau'}|^2\right)^{1/2}\right\|_{L_{\#}^{p_d-m/2}(Q)}^{p_d-m/2}. \tag*{(137)} ≲ η σ − C η 4 Q ∥ [ 0 , 1 ] d + 1 Q ∈ G [ 0 , 1 ] ( σ − C η 4 α ) ∑ ∣ Q ∣ ∣ τ ′ ∣ = σ ′ ∑ ∣ f τ ′ ∣ 2 1/2 L # p d − m ( Q ) p d − m ∣ τ ′ ∣ = σ ′ ∑ ∣ f τ ′ ∣ 2 1/2 L # p d − m /2 ( Q ) p d − m /2 . (137)
This gives the first sum in Lemma 9.5.
Case 2. Suppose that (133) fails for every dyadic u ∈ [ σ η 4 , 1 ] u\in[\sigma^{\eta^4},1] u ∈ [ σ η 4 , 1 ] , and fix such a box B B B . Starting at u = σ η 4 u=\sigma^{\eta^4} u = σ η 4 , decrease u u u dyadically and let u ∗ u_* u ∗ be the first scale for which
∥ ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , u ∗ ∥ L # p d − m / 4 ( B ) ≳ ∥ ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∥ L # p d − m / 4 ( B ) . \left\|\sum_{|\tau|=\sigma}|f_\tau|^2*w_{\tau,u_*}\right\|_{L_{\#}^{p_d-m/4}(B)}\gtrsim\left\|\sum_{|\tau|=\sigma}|f_\tau|^2\right\|_{L_{\#}^{p_d-m/4}(B)}. ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ , u ∗ L # p d − m /4 ( B ) ≳ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 L # p d − m /4 ( B ) .
At u = σ u=\sigma u = σ , the locally constant property gives (132) , so such a scale exists. The failure of (133) gives u ∗ < σ η 4 u_*<\sigma^{\eta^4} u ∗ < σ η 4 . By the choice of u ∗ u_* u ∗ and comparison of the kernels at comparable scales,
∫ B ( ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , u ) p d − m / 4 ≲ ∫ B ( ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) p d − m / 4 , σ ≤ u ≤ σ η 4 , u ∗ ≲ u . (138) \int_B\left(\sum_{|\tau|=\sigma}|f_\tau|^2*w_{\tau,u}\right)^{p_d-m/4}\lesssim\int_B\left(\sum_{|\tau|=\sigma}|f_\tau|^2\right)^{p_d-m/4},\qquad\sigma\leq u\leq\sigma^{\eta^4},\quad u_*\lesssim u. \tag*{(138)} ∫ B ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ , u p d − m /4 ≲ ∫ B ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 p d − m /4 , σ ≤ u ≤ σ η 4 , u ∗ ≲ u . (138)
We now decompose ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , u ∗ \sum_{|\tau|=\sigma}|f_\tau|^2*w_{\tau,u_*} ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , u ∗ into low and high frequencies. As in § 9.3 , choose χ ∈ C c ∞ ( R ) \chi\in C^\infty_c(\mathbb{R}) χ ∈ C c ∞ ( R ) , 0 ≤ χ ≤ 1 0\leq\chi\leq1 0 ≤ χ ≤ 1 , with χ = 1 \chi=1 χ = 1 on [ − C , C ] [-C,C] [ − C , C ] and supp χ ⊂ [ − 2 C , 2 C ] \operatorname{supp}\chi\subset[-2C,2C] supp χ ⊂ [ − 2 C , 2 C ] . For ξ = ∑ i λ i ϕ d ( i ) ( t τ ) \xi=\sum_i\lambda_i\phi_d^{(i)}(t_\tau) ξ = ∑ i λ i ϕ d ( i ) ( t τ ) , put
χ τ , ≤ u ( ξ ) = ∏ i = 0 j + 1 χ ( R j / d σ − 2 u d − i λ i ) ∏ i = j + 2 d χ ( R j / d σ − 2 u d λ i ) , χ τ , u = χ τ , ≤ u − χ τ , ≤ 2 u . \chi_{\tau,\leq u}(\xi)=\prod_{i=0}^{j+1}\chi(R^{j/d}\sigma^{-2}u^{d-i}\lambda_i)\prod_{i=j+2}^{d}\chi(R^{j/d}\sigma^{-2}u^d\lambda_i),\qquad\chi_{\tau,u}=\chi_{\tau,\leq u}-\chi_{\tau,\leq2u}. χ τ , ≤ u ( ξ ) = i = 0 ∏ j + 1 χ ( R j / d σ − 2 u d − i λ i ) i = j + 2 ∏ d χ ( R j / d σ − 2 u d λ i ) , χ τ , u = χ τ , ≤ u − χ τ , ≤ 2 u .
On supp ∣ f τ ∣ 2 ^ \operatorname{supp}\widehat{|f_\tau|^2} supp ∣ f τ ∣ 2 , we have ∣ λ i ∣ ≲ R − j / d σ i − j |\lambda_i|\lesssim R^{-j/d}\sigma^{i-j} ∣ λ i ∣ ≲ R − j / d σ i − j . Thus χ τ , ≤ σ = 1 \chi_{\tau,\leq\sigma}=1 χ τ , ≤ σ = 1 there, and the factors with i ≥ j + 2 i\geq j+2 i ≥ j + 2 equal one for σ ≤ u ≤ 1 \sigma\leq u\leq1 σ ≤ u ≤ 1 . Consequently,
∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , u ∗ = ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , u ∗ ∗ χ ~ τ , ≤ σ η 4 + ∑ σ ≤ s ′ < σ η 4 ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , u ∗ ∗ χ ~ τ , s ′ . \sum_{|\tau|=\sigma}|f_\tau|^2*w_{\tau,u_*}=\sum_{|\tau|=\sigma}|f_\tau|^2*w_{\tau,u_*}*\widetilde{\chi}_{\tau,\leq\sigma^{\eta^4}}+\sum_{\sigma\leq s'<\sigma^{\eta^4}}\sum_{|\tau|=\sigma}|f_\tau|^2*w_{\tau,u_*}*\widetilde{\chi}_{\tau,s'}. ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ , u ∗ = ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ , u ∗ ∗ χ τ , ≤ σ η 4 + σ ≤ s ′ < σ η 4 ∑ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ , u ∗ ∗ χ τ , s ′ .
The first term is small: since u ∗ ≤ σ η 4 u_*\leq\sigma^{\eta^4} u ∗ ≤ σ η 4 , the widths of w τ , u ∗ w_{\tau,u_*} w τ , u ∗ are at most the corresponding widths of χ ~ τ , ≤ σ η 4 \widetilde{\chi}_{\tau,\leq\sigma^{\eta^4}} χ τ , ≤ σ η 4 , so w τ , u ∗ ∗ ∣ χ ~ τ , ≤ σ η 4 ∣ ≲ w τ , σ η 4 w_{\tau,u_*}*|\widetilde{\chi}_{\tau,\leq\sigma^{\eta^4}}|\lesssim w_{\tau,\sigma^{\eta^4}} w τ , u ∗ ∗ ∣ χ τ , ≤ σ η 4 ∣ ≲ w τ , σ η 4 . The failure of (133) at u = σ η 4 u=\sigma^{\eta^4} u = σ η 4 therefore gives
∥ ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , u ∗ ∗ χ ~ τ , ≤ σ η 4 ∥ L # p d − m / 4 ( B ) ≲ σ η 4 ∥ ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∥ L # p d − m / 4 ( B ) . \left\|\sum_{|\tau|=\sigma}|f_\tau|^2*w_{\tau,u_*}*\widetilde{\chi}_{\tau,\leq\sigma^{\eta^4}}\right\|_{L_{\#}^{p_d-m/4}(B)}\lesssim\sigma^{\eta^4}\left\|\sum_{|\tau|=\sigma}|f_\tau|^2\right\|_{L_{\#}^{p_d-m/4}(B)}. ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ , u ∗ ∗ χ τ , ≤ σ η 4 L # p d − m /4 ( B ) ≲ σ η 4 ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 L # p d − m /4 ( B ) .
Moreover, the Fourier support bounds give w ^ τ , u ∗ χ τ , s ′ = 0 \widehat{w}_{\tau,u_*}\chi_{\tau,s'}=0 w τ , u ∗ χ τ , s ′ = 0 unless s ′ ≳ u ∗ s'\gtrsim u_* s ′ ≳ u ∗ . Dyadic pigeonholing therefore gives σ ≤ s ′ < σ η 4 \sigma\leq s'<\sigma^{\eta^4} σ ≤ s ′ < σ η 4 , with u ∗ ≲ s ′ u_*\lesssim s' u ∗ ≲ s ′ , such that the functions
g ρ = ∑ τ ∈ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , u ∗ ∗ χ ~ τ , s ′ , ρ ∈ J ( s ′ ) . g_\rho=\sum_{\tau\in\rho\atop|\tau|=\sigma}|f_\tau|^2*w_{\tau,u_*}*\widetilde{\chi}_{\tau,s'},\qquad\rho\in\mathcal{J}(s'). g ρ = ∣ τ ∣ = σ τ ∈ ρ ∑ ∣ f τ ∣ 2 ∗ w τ , u ∗ ∗ χ τ , s ′ , ρ ∈ J ( s ′ ) .
satisfy
∫ B ( ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) p d − m / 4 ≲ δ σ − δ ∫ B ∣ ∑ ρ ∈ J ( s ′ ) g ρ ∣ p d − m / 4 . (139) \int_B\left(\sum_{|\tau|=\sigma}|f_\tau|^2\right)^{p_d-m/4}\lesssim_{\delta}\sigma^{-\delta}\int_B\left|\sum_{\rho\in\mathcal{J}(s')}g_\rho\right|^{p_d-m/4}. \tag*{(139)} ∫ B ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 p d − m /4 ≲ δ σ − δ ∫ B ρ ∈ J ( s ′ ) ∑ g ρ p d − m /4 . (139)
To estimate the RHS of (139) , we first bound each g ρ g_\rho g ρ in terms of the original pieces f τ f_\tau f τ . Since u ∗ ≲ s ′ u_* \lesssim s' u ∗ ≲ s ′ , comparing the kernel widths gives w τ , u ∗ ∗ ∣ χ ˇ τ , s ′ ∣ ≲ w τ , s ′ w_{\tau,u_*} * |\check{\chi}_{\tau,s'}| \lesssim w_{\tau,s'} w τ , u ∗ ∗ ∣ χ ˇ τ , s ′ ∣ ≲ w τ , s ′ , and hence
∣ g ρ ∣ ≲ ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , s ′ . (140) |g_\rho| \lesssim\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}} |f_\tau|^2 * w_{\tau,s'}. \tag*{(140)} ∣ g ρ ∣ ≲ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ , s ′ . (140)
We now show that dilation places the Fourier support of each g ρ g_\rho g ρ in a cap of Γ j + 1 d + 1 ( ( s ′ ) − d ) \Gamma^{d+1}_{j+1}((s')^{-d}) Γ j + 1 d + 1 (( s ′ ) − d ) . This will allow us to apply Proposition 9.4. Take ξ = ∑ i λ i ϕ d ( i ) ( t τ ) \xi=\sum_i\lambda_i\phi_d^{(i)}(t_\tau) ξ = ∑ i λ i ϕ d ( i ) ( t τ ) in the Fourier support of the τ \tau τ summand defining g ρ g_\rho g ρ . The cutoff χ τ , s ′ \chi_{\tau,s'} χ τ , s ′ gives
∣ λ i ∣ ≲ R − j / d σ − 2 ( s ′ ) i − d ( 0 ≤ i ≤ j + 1 ) , max 0 ≤ i ≤ j + 1 R j / d σ − 2 ( s ′ ) d − i ∣ λ i ∣ ∼ 1. |\lambda_i|\lesssim R^{-j/d}\sigma^{-2}(s')^{i-d}\quad(0\le i\le j+1),\qquad\max_{0\le i\le j+1}R^{j/d}\sigma^{-2}(s')^{d-i}|\lambda_i|\sim1. ∣ λ i ∣ ≲ R − j / d σ − 2 ( s ′ ) i − d ( 0 ≤ i ≤ j + 1 ) , 0 ≤ i ≤ j + 1 max R j / d σ − 2 ( s ′ ) d − i ∣ λ i ∣ ∼ 1.
For j + 2 ≤ i ≤ d j+2\le i\le d j + 2 ≤ i ≤ d , the bound ∣ λ i ∣ ≲ R − j / d σ i − j |\lambda_i|\lesssim R^{-j/d}\sigma^{i-j} ∣ λ i ∣ ≲ R − j / d σ i − j on supp ∣ f τ ^ ∣ 2 \operatorname{supp}|\widehat{f_\tau}|^2 supp ∣ f τ ∣ 2 gives
∣ λ i ∣ R − j / d σ 2 ( s ′ ) i − d ≲ σ i − j − 2 ( s ′ ) d − i ≤ 1. (141) \frac{|\lambda_i|}{R^{-j/d}\sigma^2(s')^{i-d}}\lesssim\sigma^{i-j-2}(s')^{d-i}\le1. \tag*{(141)} R − j / d σ 2 ( s ′ ) i − d ∣ λ i ∣ ≲ σ i − j − 2 ( s ′ ) d − i ≤ 1. (141)
Thus dilation by R j / d σ − 2 ( s ′ ) d R^{j/d}\sigma^{-2}(s')^d R j / d σ − 2 ( s ′ ) d gives
∣ R j / d σ − 2 ( s ′ ) d λ i ∣ ≲ ( s ′ ) i ( 0 ≤ i ≤ d ) , max 0 ≤ i ≤ j + 1 ( s ′ ) − i ∣ R j / d σ − 2 ( s ′ ) d λ i ∣ ∼ 1. |R^{j/d}\sigma^{-2}(s')^d\lambda_i|\lesssim(s')^i\quad(0\le i\le d),\qquad\max_{0\le i\le j+1}(s')^{-i}|R^{j/d}\sigma^{-2}(s')^d\lambda_i|\sim1. ∣ R j / d σ − 2 ( s ′ ) d λ i ∣ ≲ ( s ′ ) i ( 0 ≤ i ≤ d ) , 0 ≤ i ≤ j + 1 max ( s ′ ) − i ∣ R j / d σ − 2 ( s ′ ) d λ i ∣ ∼ 1.
Hence
supp g ρ ^ ( R j / d σ − 2 ( s ′ ) d ⋅ ) ⊂ θ ρ , θ ρ ∈ Ξ j + 1 d + 1 ( ( s ′ ) − d ) . \operatorname{supp}\widehat{g_\rho}(R^{j/d}\sigma^{-2}(s')^d\cdot)\subset\theta_\rho,\qquad\theta_\rho\in\Xi^{d+1}_{j+1}((s')^{-d}). supp g ρ ( R j / d σ − 2 ( s ′ ) d ⋅ ) ⊂ θ ρ , θ ρ ∈ Ξ j + 1 d + 1 (( s ′ ) − d ) .
We now estimate the RHS of (139) using Proposition 9.4. As in the proof of Lemma 4.2, we use exponent p d − m / 4 p_{d-m}/4 p d − m /4 when p d − m ≥ 11 p_{d-m}\ge11 p d − m ≥ 11 , and first apply Hölder to pass from exponent 7 / 4 7/4 7/4 to 7 / 3 7/3 7/3 when p d − m = 7 p_{d-m}=7 p d − m = 7 .
First suppose p d − m ≥ 11 p_{d-m}\ge11 p d − m ≥ 11 . Choose χ B ∼ W B 4 / p d − m \chi_B\sim W_B^{4/p_{d-m}} χ B ∼ W B 4/ p d − m with supp χ B ^ ⊂ B ∗ \operatorname{supp}\widehat{\chi_B}\subset B^* supp χ B ⊂ B ∗ , as in the proof of Lemma 4.2. Then
∫ B ∣ ∑ ρ ∈ J ( s ′ ) g ρ ∣ p d − m / 4 ≲ 1 ∣ B ∣ ∫ ∣ χ B ∣ ∣ ∑ ρ ∈ J ( s ′ ) g ρ ∣ p d − m / 4 . \int_B\left|\sum_{\rho\in\mathcal{J}(s')}g_\rho\right|^{p_{d-m}/4}\lesssim\frac{1}{|B|}\int|\chi_B| \left|\sum_{\rho\in\mathcal{J}(s')}g_\rho\right|^{p_{d-m}/4}. ∫ B ρ ∈ J ( s ′ ) ∑ g ρ p d − m /4 ≲ ∣ B ∣ 1 ∫ ∣ χ B ∣ ρ ∈ J ( s ′ ) ∑ g ρ p d − m /4 .
Since B B B has side length R j / d σ − 2 R^{j/d}\sigma^{-2} R j / d σ − 2 , multiplication by χ B \chi_B χ B adds Fourier widths R − j / d σ 2 R^{-j/d}\sigma^2 R − j / d σ 2 . Their ratios to the coefficient widths R − j / d σ 2 ( s ′ ) i − d R^{-j/d}\sigma^2(s')^{i-d} R − j / d σ 2 ( s ′ ) i − d are ( s ′ ) d − i ≤ 1 (s')^{d-i}\le1 ( s ′ ) d − i ≤ 1 , and for i ≤ j + 1 i\le j+1 i ≤ j + 1 they are at most ( s ′ ) d − j − 1 ≪ 1 (s')^{d-j-1}\ll1 ( s ′ ) d − j − 1 ≪ 1 . Hence
supp ( χ B g ρ ) ( R j / d σ − 2 ( s ′ ) d ⋅ ) ⊂ θ ρ , θ ρ ∈ Ξ j + 1 d + 1 ( ( s ′ ) − d ) . \operatorname{supp}(\chi_Bg_\rho)(R^{j/d}\sigma^{-2}(s')^d\cdot)\subset\theta_\rho,\qquad\theta_\rho\in\Xi^{d+1}_{j+1}((s')^{-d}). supp ( χ B g ρ ) ( R j / d σ − 2 ( s ′ ) d ⋅ ) ⊂ θ ρ , θ ρ ∈ Ξ j + 1 d + 1 (( s ′ ) − d ) .
Apply Proposition 9.4 at exponent p d − m / 4 ∈ [ 2 , p d − j − 1 ] p_{d-m}/4\in[2,p_{d-j-1}] p d − m /4 ∈ [ 2 , p d − j − 1 ] and parameter ( s ′ ) − d (s')^{-d} ( s ′ ) − d to ( χ B ∑ ρ ∈ J ( s ′ ) g ρ ) ( R j / d σ − 2 ( s ′ ) d ⋅ ) \left(\chi_B\sum_{\rho\in\mathcal{J}(s')}g_\rho\right)(R^{j/d}\sigma^{-2}(s')^d\cdot) ( χ B ∑ ρ ∈ J ( s ′ ) g ρ ) ( R j / d σ − 2 ( s ′ ) d ⋅ ) , with fine pieces ( χ B g ρ ) ( R j / d σ − 2 ( s ′ ) d ⋅ ) (\chi_Bg_\rho)(R^{j/d}\sigma^{-2}(s')^d\cdot) ( χ B g ρ ) ( R j / d σ − 2 ( s ′ ) d ⋅ ) . Returning to the original coordinates and using (139) and (140) , we obtain
∫ B ( ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) p d − m / 4 ≲ δ σ − δ ∫ B [ ∑ ρ ∈ J ( s ′ ) ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , s ′ ) 2 ] p d − m / 8 . (142) \int_B\left(\sum_{|\tau|=\sigma}|f_\tau|^2\right)^{p_{d-m}/4}
\lesssim_\delta\sigma^{-\delta}\int_B\left[\sum_{\rho\in\mathcal{J}(s')}\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_\tau|^2*w_{\tau,s'}\right)^2\right]^{p_{d-m}/8}. \tag*{(142)} ∫ B ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 p d − m /4 ≲ δ σ − δ ∫ B ρ ∈ J ( s ′ ) ∑ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ , s ′ 2 p d − m /8 . (142)
For p d − m = 7 p_{d-m}=7 p d − m = 7 , choose χ B 7 / 3 ∼ W B \chi_B^{7/3}\sim W_B χ B 7/3 ∼ W B with supp χ B ^ ⊂ B ∗ \operatorname{supp}\widehat{\chi_B}\subset B^* supp χ B ⊂ B ∗ . As in the n = 3 n=3 n = 3 case of the proof of Lemma 4.2, Hölder gives
( ∫ B ∣ ∑ ρ ∈ J ( s ′ ) g ρ ∣ 7 / 4 ) 4 / 3 ≲ 1 ∣ B ∣ ∫ ∣ χ B ∣ ∣ ∑ ρ ∈ J ( s ′ ) g ρ ∣ 7 / 3 . \left(\int_B\left|\sum_{\rho\in\mathcal{J}(s')}g_\rho\right|^{7/4}\right)^{4/3}\lesssim\frac{1}{|B|}\int|\chi_B|\left|\sum_{\rho\in\mathcal{J}(s')}g_\rho\right|^{7/3}. ∫ B ρ ∈ J ( s ′ ) ∑ g ρ 7/4 4/3 ≲ ∣ B ∣ 1 ∫ ∣ χ B ∣ ρ ∈ J ( s ′ ) ∑ g ρ 7/3 .
With this choice of χ B \chi_B χ B , apply Proposition 9.4 to ∑ ρ ∈ J ( s ′ ) ( χ B g ρ ) ( R j / d σ − 2 ( s ′ ) d ⋅ ) \sum_{\rho\in\mathcal{J}(s')}(\chi_Bg_\rho)(R^{j/d}\sigma^{-2}(s')^d\cdot) ∑ ρ ∈ J ( s ′ ) ( χ B g ρ ) ( R j / d σ − 2 ( s ′ ) d ⋅ ) at exponent 7 / 3 ∈ [ 2 , p d − j − 1 ] 7/3\in[2,p_{d-j-1}] 7/3 ∈ [ 2 , p d − j − 1 ] and parameter ( s ′ ) − d (s')^{-d} ( s ′ ) − d . Returning to the original coordinates and using (139) and (140) gives
[ ∫ B ( ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 7 / 4 ] 4 / 3 ≲ δ σ − δ ∫ B [ ∑ ρ ∈ J ( s ′ ) ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , s ′ ) 2 ] 7 / 6 . (143) \left[\int_B\left(\sum_{|\tau|=\sigma}|f_\tau|^2\right)^{7/4}\right]^{4/3}
\lesssim_\delta\sigma^{-\delta}\int_B\left[\sum_{\rho\in\mathcal{J}(s')}\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_\tau|^2*w_{\tau,s'}\right)^2\right]^{7/6}. \tag*{(143)} ∫ B ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 7/4 4/3 ≲ δ σ − δ ∫ B ρ ∈ J ( s ′ ) ∑ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ , s ′ 2 7/6 . (143)
To obtain the second sum in Lemma 9.5, we first bound (142) and (143) by integrals over the set where ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , s ′ \sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_\tau|^2*w_{\tau,s'} ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , s ′ exceeds the threshold [ α / σ − δ ] 2 [\alpha/\sigma^{-\delta}]^2 [ α / σ − δ ] 2 . Since B ∈ G [ 0 , 1 ] ( α ) B\in\mathcal{G}_{[0,1]}(\alpha) B ∈ G [ 0 , 1 ] ( α ) , we have $\int_B(\sum_{|\tau|=\sigma}|f_\tau|^2)^{p_{d-m}/4}\ge α p d − m / 2 \alpha^{p_d-m/2} α p d − m /2 . Together with (9.16) at u = s ′ u=s' u = s ′ , this allows us to repeat the argument leading to (4.7), with ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , s ′ \sum_{\tau\subset\rho,\ |\tau|=\sigma}|f_\tau|^2*w_{\tau,s'} ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , s ′ in place of a ρ ′ a_{\rho'} a ρ ′ . Applying that argument to (9.20) and (9.21) gives
\left[\mathop{\fint}_{B}\left(\sum_{|\tau|=\sigma}|f_\tau|^2\right)^{p_d-m/4}\right]^2 \lesssim_\delta\frac{\sigma^{-\delta}}{|B|}\sum_{\rho\in\mathcal{J}(s')}\int_{\sum_{\tau\subset\rho,\ |\tau|=\sigma}|f_\tau|^2*w_{\tau,s'}\gtrsim[\alpha/\sigma^{-\delta}]^2}\left(\sum_{\tau\subset\rho\atop|\tau|=\sigma}|f_\tau|^2*w_{\tau,s'}\right)^{p_d-m/2}W_B. \tag*{(144)}
To remove the convolutions in (9.22) after summing over B B B , we first compare the kernels w τ , s ′ w_{\tau,s'} w τ , s ′ , τ ⊂ ρ \tau\subset\rho τ ⊂ ρ , with a single kernel for each ρ \rho ρ . For w ρ , s ′ w_{\rho,s'} w ρ , s ′ , take a nonnegative, L 1 L^1 L 1 normalized kernel with sufficiently rapid decay, adapted to the box in the definition of w τ , u w_{\tau,u} w τ , u with u = s ′ u=s' u = s ′ and ∣ x ⋅ ϕ d ( i ) ( t ρ ) ∣ |x\cdot\phi_d^{(i)}(t_\rho)| ∣ x ⋅ ϕ d ( i ) ( t ρ ) ∣ in place of ∣ x ⋅ ϕ d ( i ) ( t τ ) ∣ |x\cdot\phi_d^{(i)}(t_\tau)| ∣ x ⋅ ϕ d ( i ) ( t τ ) ∣ .
Indeed, since τ ⊂ ρ \tau\subset\rho τ ⊂ ρ , we have ∣ t τ − t ρ ∣ ≲ s ′ |t_\tau-t_\rho|\lesssim s' ∣ t τ − t ρ ∣ ≲ s ′ . Taylor expansion gives
R j / d σ − 2 ( s ′ ) d − k ∣ t τ − t ρ ∣ k − i ≲ R j / d σ − 2 ( s ′ ) d − i , 0 ≤ i ≤ k ≤ j + 1 , R^{j/d}\sigma^{-2}(s')^{d-k}|t_\tau-t_\rho|^{k-i}\lesssim R^{j/d}\sigma^{-2}(s')^{d-i},\qquad0\le i\le k\le j+1, R j / d σ − 2 ( s ′ ) d − k ∣ t τ − t ρ ∣ k − i ≲ R j / d σ − 2 ( s ′ ) d − i , 0 ≤ i ≤ k ≤ j + 1 ,
R j / d σ − 2 ( s ′ ) d ∣ t τ − t ρ ∣ k − i ≲ R j / d σ − 2 ( s ′ ) d , 0 ≤ i ≤ k , j + 2 ≤ k ≤ d . R^{j/d}\sigma^{-2}(s')^{d}|t_\tau-t_\rho|^{k-i}\lesssim R^{j/d}\sigma^{-2}(s')^d,\qquad0\le i\le k,\quad j+2\le k\le d. R j / d σ − 2 ( s ′ ) d ∣ t τ − t ρ ∣ k − i ≲ R j / d σ − 2 ( s ′ ) d , 0 ≤ i ≤ k , j + 2 ≤ k ≤ d .
These inequalities also hold with t τ t_\tau t τ and t ρ t_\rho t ρ interchanged. Thus the bounds on ∣ x ⋅ ϕ d ( i ) ( t τ ) ∣ |x\cdot\phi_d^{(i)}(t_\tau)| ∣ x ⋅ ϕ d ( i ) ( t τ ) ∣ and on ∣ x ⋅ ϕ d ( i ) ( t ρ ) ∣ |x\cdot\phi_d^{(i)}(t_\rho)| ∣ x ⋅ ϕ d ( i ) ( t ρ ) ∣ define boxes comparable up to a fixed dilation, uniformly in τ ⊂ ρ \tau\subset\rho τ ⊂ ρ . We may therefore choose w ρ , s ′ w_{\rho,s'} w ρ , s ′ so that w τ , s ′ ≲ w ρ , s ′ w_{\tau,s'}\lesssim w_{\rho,s'} w τ , s ′ ≲ w ρ , s ′ for all such τ \tau τ , giving
∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ∗ w τ , s ′ ≲ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) ∗ w ρ , s ′ . \sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_\tau|^2*w_{\tau,s'}\lesssim\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_\tau|^2\right)*w_{\rho,s'}. τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w τ , s ′ ≲ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w ρ , s ′ .
Fix s ′ s' s ′ , multiply (9.22) by ∣ B ∣ |B| ∣ B ∣ , and sum over the boxes B B B in Case 2 for which this s ′ s' s ′ was chosen. Using ∑ B W B ≲ 1 \sum_B W_B\lesssim1 ∑ B W B ≲ 1 and lowering the threshold to [ α / ( σ − δ N ( s ′ ) ) ] 2 [\alpha/(\sigma^{-\delta}N(s'))]^2 [ α / ( σ − δ N ( s ′ )) ] 2 , we bound the contribution of these boxes to the LHS of Lemma 9.5 by
≲ δ σ − δ ∑ ρ ∈ J ( s ′ ) ∫ ( ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ) ∗ w ρ , s ′ ≳ [ α / ( σ − δ N ( s ′ ) ) ] 2 [ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) ∗ w ρ , s ′ ] p d − m / 2 . \lesssim_\delta\sigma^{-\delta}\sum_{\rho\in\mathcal{J}(s')}\int_{\left(\sum_{\tau\subset\rho,\ |\tau|=\sigma}|f_\tau|^2\right)*w_{\rho,s'}\gtrsim[\alpha/(\sigma^{-\delta}N(s'))]^2}\left[\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_\tau|^2\right)*w_{\rho,s'}\right]^{p_d-m/2}. ≲ δ σ − δ ρ ∈ J ( s ′ ) ∑ ∫ ( ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ 2 ) ∗ w ρ , s ′ ≳ [ α / ( σ − δ N ( s ′ )) ] 2 τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 ∗ w ρ , s ′ p d − m /2 .
We now apply (2.24) to each integral, with F = ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ F=\sum_{\tau\subset\rho,\ |\tau|=\sigma}|f_\tau| F = ∑ τ ⊂ ρ , ∣ τ ∣ = σ ∣ f τ ∣ , K = w ρ , s ′ K=w_{\rho,s'} K = w ρ , s ′ , exponent p d − m / 2 p_d-m/2 p d − m /2 , and threshold [ α / ( σ − δ N ( s ′ ) ) ] 2 [\alpha/(\sigma^{-\delta}N(s'))]^2 [ α / ( σ − δ N ( s ′ )) ] 2 . This removes the convolution, and the locally constant property on ρ ∗ \rho^* ρ ∗ then bounds these integrals by
≲ δ σ − δ ∑ ρ ∈ J ( s ′ ) ∑ Q ∈ G ρ ∗ ( α σ − δ ) ∣ Q ∣ ∥ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( Q ) p d − m . (145) \lesssim_\delta\sigma^{-\delta}\sum_{\rho\in\mathcal{J}(s')}\sum_{Q\in\mathcal{G}_{\rho^*}\left(\frac{\alpha}{\sigma^{-\delta}}\right)}|Q|\left\|\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_\tau|^2\right)^{1/2}\right\|_{L_{\#}^{p_d-m/2}(Q)}^{p_d-m}. \tag*{(145)} ≲ δ σ − δ ρ ∈ J ( s ′ ) ∑ Q ∈ G ρ ∗ ( σ − δ α ) ∑ ∣ Q ∣ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L # p d − m /2 ( Q ) p d − m . (145)
Summing (9.15) over ( u , σ ′ ) (u,\sigma') ( u , σ ′ ) from Case 1 and (9.23) over s ′ s' s ′ from Case 2 covers every B ∈ G [ 0 , 1 ] ( α ) B\in\mathcal{G}_{[0,1]}(\alpha) B ∈ G [ 0 , 1 ] ( α ) . Dyadic pigeonholing gives one common σ ′ \sigma' σ ′ in the first sum and one common s ′ s' s ′ in the second. The logarithmic losses are absorbed into σ − C η 4 \sigma^{-C\eta_4} σ − C η 4 and σ − δ \sigma^{-\delta} σ − δ , respectively, both outside the sums and in the selection conditions. This proves Lemma 9.5. □
Proof of Lemma 8.11. By definition, Γ j , 1 / K d + 1 ( R ) ⊂ Γ ~ j , 1 d + 1 ( R ) \Gamma^{d+1}_{j,1/K}(R)\subset\widetilde{\Gamma}^{d+1}_{j,1}(R) Γ j , 1/ K d + 1 ( R ) ⊂ Γ j , 1 d + 1 ( R ) for 0 ≤ j ≤ m 0\le j\le m 0 ≤ j ≤ m . Applying Lemma 9.5 with ℓ = 1 \ell=1 ℓ = 1 therefore proves the claim. □
Proof of Proposition 8.9 and (8.9).
Proof of Proposition 8.9 and (8.9). Under the hypotheses of Proposition 8.7, fix m < d − 2 m<d-2 m < d − 2 and let f f f be a Schwartz function with supp f ^ ⊂ Γ j , 1 / K d + 1 ( R ) \operatorname{supp}\widehat f\subset\Gamma^{d+1}_{j,1/K}(R) supp f ⊂ Γ j , 1/ K d + 1 ( R ) , 0 ≤ j ≤ m 0\le j\le m 0 ≤ j ≤ m . We choose small parameters 0 < δ ≪ η ≪ ϵ 0<\delta\ll\eta\ll\epsilon 0 < δ ≪ η ≪ ϵ as in Section 8.2. For R − 1 / d ≤ σ ≤ A − 1 R^{-1/d}\le\sigma\le A^{-1} R − 1/ d ≤ σ ≤ A − 1 and σ ≤ s ≤ 1 \sigma\le s\le1 σ ≤ s ≤ 1 , we prove the stronger bound
∑ ρ ∈ J ( s ) ∑ Q ∈ G ρ ∗ ( α ) ∣ Q ∣ ∥ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( Q ) p d − m ≲ ϵ RHS of (8.8) . (146) \sum_{\rho\in\mathcal{J}(s)}\sum_{Q\in\mathcal{G}_{\rho^*}(\alpha)}|Q|\left\|\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_\tau|^2\right)^{1/2}\right\|_{L_{\#}^{p_d-m/2}(Q)}^{p_d-m}\lesssim_\epsilon\operatorname{RHS\ of\ (8.8)}. \tag*{(146)} ρ ∈ J ( s ) ∑ Q ∈ G ρ ∗ ( α ) ∑ ∣ Q ∣ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L # p d − m /2 ( Q ) p d − m ≲ ϵ RHS of (8.8) . (146)
where both occurrences of R ϵ R^\epsilon R ϵ on the right are replaced by R ϵ ( σ s ) ϵ R^\epsilon(\sigma s)^\epsilon R ϵ ( σ s ) ϵ . We prove (9.24) and (8.9) together by induction on R R R and, at fixed R R R , on the dyadic parameter σ s \sigma s σ s . For bounded R R R , both bounds follow from Hölder and the locally constant property, followed by Lemma 9.2 applied to f f f with parameter R R R and u = R − 1 / d u=R^{-1/d} u = R − 1/ d , using the terms q 1 = ⋯ = q d − m − 2 = u = R − 1 / d q_1=\cdots=q_{d-m-2}=u=R^{-1/d} q 1 = ⋯ = q d − m − 2 = u = R − 1/ d in (93) .
Assume that (146) and (94) hold at every parameter R ′ < R / 2 R'<R/2 R ′ < R /2 . With R R R fixed, assume also that (146) holds for pairs ( σ ′ , s ′ ) (\sigma',s') ( σ ′ , s ′ ) satisfying σ ′ s ′ < σ s / 2 \sigma's'<\sigma s/2 σ ′ s ′ < σ s /2 .
We first deal with the case σ = R − 1 / d \sigma=R^{-1/d} σ = R − 1/ d . By (32) ,
LHS of (9.24) ≲ ∑ ρ ∈ J ( s ) ∫ ⋃ θ ⊂ ρ ∣ f θ ∣ 2 ≳ [ α / N ( s ) ] 2 ( ∑ θ ⊂ ρ ∣ f θ ∣ 2 ) p d − m / 2 . \text{LHS of (9.24)}\lesssim\sum_{\rho\in\mathcal{J}(s)}\int_{\bigcup_{\theta\subset\rho}|f_\theta|^2\gtrsim[\alpha/N(s)]^2}\left(\sum_{\theta\subset\rho}|f_\theta|^2\right)^{p_{d-m}/2}. LHS of (9.24) ≲ ρ ∈ J ( s ) ∑ ∫ ⋃ θ ⊂ ρ ∣ f θ ∣ 2 ≳ [ α / N ( s ) ] 2 θ ⊂ ρ ∑ ∣ f θ ∣ 2 p d − m /2 .
For each interval ρ = [ t ρ , t ρ + s ] ∈ J ( s ) \rho=[t_\rho,t_\rho+s]\in\mathcal{J}(s) ρ = [ t ρ , t ρ + s ] ∈ J ( s ) , set g ( y ) = f ρ ( L t ρ , s − T y ) g(y)=f_\rho(L_{t_\rho,s}^{-T}y) g ( y ) = f ρ ( L t ρ , s − T y ) . Then supp g ^ ⊂ Γ m d + 1 ( R s d ) \operatorname{supp}\widehat{g}\subset\Gamma_m^{d+1}(Rs^d) supp g ⊂ Γ m d + 1 ( R s d ) , with fine pieces f θ ( L t ρ , s − T y ) f_\theta(L_{t_\rho,s}^{-T}y) f θ ( L t ρ , s − T y ) , θ ⊂ ρ \theta\subset\rho θ ⊂ ρ , of angular width ( R s d ) − 1 / d (Rs^d)^{-1/d} ( R s d ) − 1/ d .
We therefore apply (124) to g g g with parameter R s d Rs^d R s d , amplitude α / N ( s ) \alpha/N(s) α / N ( s ) , and loss η \eta η . This further bounds the integral for this ρ \rho ρ in the sum above, after rescaling, by the RHS of (93) for g g g at parameter R s d Rs^d R s d . In this sum, denote the intermediate caps by τ l ′ \tau_l' τ l ′ , with ∣ τ l ′ ∣ = q l / s |\tau_l'|=q_l/s ∣ τ l ′ ∣ = q l / s , their parent by ρ ′ \rho' ρ ′ , with ∣ ρ ′ ∣ = u / s |\rho'|=u/s ∣ ρ ′ ∣ = u / s , the convolution weights by w B τ l ′ ′ w_{B_{\tau_l'}'} w B τ l ′ ′ , and the averaging boxes by B ′ B' B ′ . Returning by x = L t ρ , s − T y x=L_{t_\rho,s}^{-T}y x = L t ρ , s − T y , write τ l = L t ρ , s τ l ′ ⊂ ρ \tau_l=L_{t_\rho,s}\tau_l'\subset\rho τ l = L t ρ , s τ l ′ ⊂ ρ , with ∣ τ l ∣ = q l |\tau_l|=q_l ∣ τ l ∣ = q l , and B = L t ρ , s B ′ B=L_{t_\rho,s}B' B = L t ρ , s B ′ . The i i i th widths of L t ρ , s − T B τ l ′ ′ L_{t_\rho,s}^{-T}B_{\tau_l'}' L t ρ , s − T B τ l ′ ′ and B B B satisfy
L t ρ , s − T B τ l ′ ′ : s − i ( R s d ) ( m + 2 ) / d ( q l / s ) m + l + 1 − i ∏ b < l ( q b / s ) = R ( m + 2 ) / d q l m + l + 1 − i ∏ b < l q b , 0 ≤ i ≤ m + l + 1 , s − i ( R s d ) ( m + 2 ) / d ( u / s ) m + l + 1 − i ∏ b < l ( q b / s ) = R ( m + 2 ) / d u m + l + 1 − i ∏ b < l q b , m + l + 1 < i ≤ d , B : s − i ( R s d ) ( m + 2 ) / d ( u / s ) d − i ∏ b = 1 d − m − 2 ( q b / s ) = R ( m + 2 ) / d u d − i ∏ b = 1 d − m − 2 q b , 0 ≤ i ≤ d . (147) \begin{aligned}
L_{t_\rho,s}^{-T}B_{\tau_l'}':\quad
&s^{-i}\frac{(Rs^d)^{(m+2)/d}(q_l/s)^{m+l+1-i}}{\prod_{b<l}(q_b/s)}
=\frac{R^{(m+2)/d}q_l^{m+l+1-i}}{\prod_{b<l}q_b},
&&0\le i\le m+l+1,\\
&s^{-i}\frac{(Rs^d)^{(m+2)/d}(u/s)^{m+l+1-i}}{\prod_{b<l}(q_b/s)}
=\frac{R^{(m+2)/d}u^{m+l+1-i}}{\prod_{b<l}q_b},
&&m+l+1<i\le d,\\
B:\quad
&s^{-i}\frac{(Rs^d)^{(m+2)/d}(u/s)^{d-i}}{\prod_{b=1}^{d-m-2}(q_b/s)}
=\frac{R^{(m+2)/d}u^{d-i}}{\prod_{b=1}^{d-m-2}q_b},
&&0\le i\le d. \tag*{(147)}
\end{aligned} L t ρ , s − T B τ l ′ ′ : B : s − i ∏ b < l ( q b / s ) ( R s d ) ( m + 2 ) / d ( q l / s ) m + l + 1 − i = ∏ b < l q b R ( m + 2 ) / d q l m + l + 1 − i , s − i ∏ b < l ( q b / s ) ( R s d ) ( m + 2 ) / d ( u / s ) m + l + 1 − i = ∏ b < l q b R ( m + 2 ) / d u m + l + 1 − i , s − i ∏ b = 1 d − m − 2 ( q b / s ) ( R s d ) ( m + 2 ) / d ( u / s ) d − i = ∏ b = 1 d − m − 2 q b R ( m + 2 ) / d u d − i , 0 ≤ i ≤ m + l + 1 , m + l + 1 < i ≤ d , 0 ≤ i ≤ d . (147)
These are exactly the widths of B τ l B_{\tau_l} B τ l in (92) and of the averaging boxes in (93) , with parent L t ρ , s ρ ′ ⊂ ρ L_{t_\rho,s}\rho'\subset\rho L t ρ , s ρ ′ ⊂ ρ of width u u u . Here R − 1 / d ≤ q 1 ≤ ⋯ ≤ q d − m − 2 ≤ u ≤ s R^{-1/d}\le q_1\le\cdots\le q_{d-m-2}\le u\le s R − 1/ d ≤ q 1 ≤ ⋯ ≤ q d − m − 2 ≤ u ≤ s .
The active parent count at rescaled width u / s u/s u / s is N ρ ( u ) N_\rho(u) N ρ ( u ) . Since ( R s d ) η ≤ R η (Rs^d)^\eta\le R^\eta ( R s d ) η ≤ R η , the bound N ( s ) N ρ ( u ) ≲ N ( u ) N(s)N_\rho(u)\lesssim N(u) N ( s ) N ρ ( u ) ≲ N ( u ) from Lemma 2.18 gives
α ( R s d ) η N ( s ) N ρ ( u ) ≳ α R η N ( u ) . \frac{\alpha}{(Rs^d)^\eta N(s)N_\rho(u)}\gtrsim\frac{\alpha}{R^\eta N(u)}. ( R s d ) η N ( s ) N ρ ( u ) α ≳ R η N ( u ) α .
Consequently, summing over ρ \rho ρ and extending u ≤ s u\le s u ≤ s to u ≤ 1 u\le1 u ≤ 1 bounds the LHS of (146) by the RHS of (93) , with both occurrences of R ϵ R^\epsilon R ϵ replaced by R η R^\eta R η . Since σ = R − 1 / d \sigma=R^{-1/d} σ = R − 1/ d and s ≥ σ s\ge\sigma s ≥ σ ,
R η R ϵ ( σ s ) ϵ ≤ R − ϵ / C \frac{R^\eta}{R^\epsilon(\sigma s)^\epsilon}\le R^{-\epsilon/C} R ϵ ( σ s ) ϵ R η ≤ R − ϵ / C
for η \eta η sufficiently small in terms of ϵ \epsilon ϵ . This proves (146) at σ = R − 1 / d \sigma=R^{-1/d} σ = R − 1/ d .
We now consider the case σ > R − 1 / d \sigma>R^{-1/d} σ > R − 1/ d . Corollary 9.3, applied to f f f with ℓ = 1 \ell=1 ℓ = 1 and loss δ \delta δ , bounds the LHS of (146) by the RHS of (98) . Fix σ ≤ u ≤ s \sigma\le u\le s σ ≤ u ≤ s . We estimate the corresponding term on the RHS of (98) , namely
σ − δ ∑ ρ ∈ J ( u ) ∑ B ∈ G ρ ( α σ δ ) ( u R 1 / d σ R 1 / d ) 2 ∣ B ∣ ∥ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( B ) p d − m . \sigma^{-\delta}\sum_{\rho\in\mathcal{J}(u)}\sum_{B\in\mathcal{G}_\rho(\alpha\sigma^\delta)}
\left(\frac{uR^{1/d}}{\sigma R^{1/d}}\right)^2
|B|\left\|\left(\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}|f_\tau|^2\right)^{1/2}\right\|_{L_\#^{p_{d-m}/2}(B)}^{p_{d-m}}. σ − δ ρ ∈ J ( u ) ∑ B ∈ G ρ ( α σ δ ) ∑ ( σ R 1/ d u R 1/ d ) 2 ∣ B ∣ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L # p d − m /2 ( B ) p d − m .
For u < σ 1 − η u<\sigma^{1-\eta} u < σ 1 − η , we use (99) and then apply (94) at parameter R σ d < R / 2 R\sigma^d<R/2 R σ d < R /2 after Taylor rescaling f τ f_\tau f τ . For u ≥ σ 1 − η u\ge\sigma^{1-\eta} u ≥ σ 1 − η , Lemma 8.11 gives the scales ( σ ′ , u ) (\sigma',u) ( σ ′ , u ) with σ ′ < σ \sigma'<\sigma σ ′ < σ , or ( σ , u ′ ) (\sigma,u') ( σ , u ′ ) with u ′ < u u'<u u ′ < u .
*Case 1: u < σ 1 − η u<\sigma^{1-\eta} u < σ 1 − η . As in (99) , we can bound the fixed u u u term on the RHS of (98) by
≲ σ − C η ∑ ρ ∈ J ( u ) ∑ τ ⊂ ρ ∣ τ ∣ = σ ∫ ∣ f τ ∣ ≳ α σ δ / [ N ( u ) N ρ ( σ ) ] ∣ f τ ∣ p d − m . \lesssim\sigma^{-C\eta}\sum_{\rho\in\mathcal{J}(u)}\sum_{\substack{\tau\subset\rho\\|\tau|=\sigma}}\int_{|f_\tau|\gtrsim\alpha\sigma^\delta/[N(u)N_\rho(\sigma)]}|f_\tau|^{p_{d-m}}. ≲ σ − C η ρ ∈ J ( u ) ∑ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∫ ∣ f τ ∣ ≳ α σ δ / [ N ( u ) N ρ ( σ )] ∣ f τ ∣ p d − m .
Since supp f τ ^ ( L t τ , σ − T ⋅ ) ⊂ Γ j , j + 1 d + 1 ( R σ d ) \operatorname{supp}\widehat{f_\tau}(L_{t_\tau,\sigma}^{-T}\cdot)\subset\Gamma_{j,j+1}^{d+1}(R\sigma^d) supp f τ ( L t τ , σ − T ⋅ ) ⊂ Γ j , j + 1 d + 1 ( R σ d ) for τ ∈ J ( σ ) \tau\in\mathcal{J}(\sigma) τ ∈ J ( σ ) , we can apply (94) inductively to this function with parameter R σ d < R / 2 R\sigma^d<R/2 R σ d < R /2 and amplitude α σ δ / [ N ( u ) N ρ ( σ ) ] \alpha\sigma^\delta/[N(u)N_\rho(\sigma)] α σ δ / [ N ( u ) N ρ ( σ )] .
Returning to the original coordinates, (147) , with σ \sigma σ in place of s s s and v v v in place of u u u , gives the convolution weights w B τ l w_{B_{\tau_l}} w B τ l and boxes B B B in (93) , with R − 1 / d ≤ q 1 ≤ ⋯ ≤ q d − m − 2 ≤ v ≤ σ R^{-1/d}\le q_1\le\cdots\le q_{d-m-2}\le v\le\sigma R − 1/ d ≤ q 1 ≤ ⋯ ≤ q d − m − 2 ≤ v ≤ σ . For a parent ρ ′ ⊂ τ \rho'\subset\tau ρ ′ ⊂ τ with ∣ ρ ′ ∣ = v |\rho'|=v ∣ ρ ′ ∣ = v , applying Lemma 2.18 twice gives N ( u ) N ρ ( σ ) N τ ( v ) ≲ N ( v ) N(u)N_\rho(\sigma)N_\tau(v)\lesssim N(v) N ( u ) N ρ ( σ ) N τ ( v ) ≲ N ( v ) . Hence the amplitude in the selection condition satisfies
α σ δ ( R σ d ) ϵ N ( u ) N ρ ( σ ) N τ ( v ) ≳ α σ − C η ( R σ d ) ϵ N ( v ) . \frac{\alpha\sigma^\delta}{(R\sigma^d)^\epsilon N(u)N_\rho(\sigma)N_\tau(v)}\gtrsim\frac{\alpha}{\sigma^{-C\eta}(R\sigma^d)^\epsilon N(v)}. ( R σ d ) ϵ N ( u ) N ρ ( σ ) N τ ( v ) α σ δ ≳ σ − C η ( R σ d ) ϵ N ( v ) α .
Together with the factor σ − C η \sigma^{-C\eta} σ − C η in (8.14), summing over ρ , τ \rho,\tau ρ , τ and extending v ≤ σ v\leq\sigma v ≤ σ to v ≤ 1 v\leq1 v ≤ 1 therefore bounds the fixed u u u term on the RHS of (8.13) by the RHS of (8.8), with both occurrences of R ϵ R^\epsilon R ϵ replaced by σ − C η ( R σ d ) ϵ \sigma^{-C\eta}(R\sigma^d)^\epsilon σ − C η ( R σ d ) ϵ .
Since s ≥ σ s\geq\sigma s ≥ σ ,
σ − C η ( R σ d ) ϵ R ϵ ( σ s ) ϵ ≤ σ ( d − 2 ) ϵ − C η . \frac{\sigma^{-C\eta}(R\sigma^d)^\epsilon}{R^\epsilon(\sigma s)^\epsilon}\leq\sigma^{(d-2)\epsilon-C\eta}. R ϵ ( σ s ) ϵ σ − C η ( R σ d ) ϵ ≤ σ ( d − 2 ) ϵ − C η .
This is the third factor in (8.18) and finishes the fixed u u u estimate in Case 1.
Case 2: u ≥ σ 1 − η u\geq\sigma^{1-\eta} u ≥ σ 1 − η . For each ρ ∈ J ( u ) \rho\in\mathcal{J}(u) ρ ∈ J ( u ) , apply Lemma 8.11 to f ρ ( L t ρ , u T ⋅ ) f_\rho(L_{t_\rho,u}^{T}\cdot) f ρ ( L t ρ , u T ⋅ ) , with parameter R u d Ru^d R u d , cap width σ / u \sigma/u σ / u , and amplitude α σ δ / N ( u ) \alpha\sigma^\delta/N(u) α σ δ / N ( u ) . After applying the lemma and returning to the original coordinates, denote the new scales by ( σ ′ , u ′ ) (\sigma',u') ( σ ′ , u ′ ) . Summing over ρ \rho ρ and applying dyadic pigeonholing as in (8.17), we can bound the fixed u u u term on the RHS of (8.13) by
≲ σ − C η 4 ∑ ρ ∈ J ( u ) ∑ Q ∈ G ρ Q ∈ G ρ ( α / σ − C η 4 ) ∣ Q ∣ ( ∑ τ ′ ⊂ ρ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ) 1 / 2 ∥ ∑ τ ′ ⊂ ρ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ∣ ∥ L # p d − m 2 ( Q ) p d − m p d − m + σ − 2 δ ∑ ρ ′ ∈ J ( u ′ ) ∑ Q ∈ G ρ ′ Q ∈ G ρ ′ ( α / σ − 2 δ ) ∣ Q ∣ ( ∑ τ ′ ⊂ ρ ′ ∣ τ ′ ∣ = σ ∣ f τ ′ ∣ 2 ) 1 / 2 ∥ ∑ τ ′ ⊂ ρ ′ ∣ τ ′ ∣ = σ ∣ f τ ′ ∣ 2 ∣ ∥ L # p d − m 2 ( Q ) p d − m p d − m , \begin{aligned}
&\lesssim\sigma^{-C\eta^4}\sum_{\rho\in\mathcal{J}(u)}
\sum_{\substack{Q\in\mathcal{G}_\rho\\ Q\in\mathcal{G}_\rho(\alpha/\sigma^{-C\eta^4})}}
|Q|\left(\sum_{\substack{\tau'\subset\rho\\|\tau'|=\sigma'}}|f_{\tau'}|^2\right)^{1/2}
\left\| \vphantom{\sum_{\substack{\tau'\subset\rho\\|\tau'|=\sigma'}}|f_{\tau'}|^2}^{\vphantom{\big|}}\right\|_{L_\#^{\frac{pd-m}{2}}(Q)}^{\frac{pd-m}{pd-m}}\\
&\quad+\sigma^{-2\delta}\sum_{\rho'\in\mathcal{J}(u')}
\sum_{\substack{Q\in\mathcal{G}_{\rho'}\\ Q\in\mathcal{G}_{\rho'}(\alpha/\sigma^{-2\delta})}}
|Q|\left(\sum_{\substack{\tau'\subset\rho'\\|\tau'|=\sigma}}|f_{\tau'}|^2\right)^{1/2}
\left\| \vphantom{\sum_{\substack{\tau'\subset\rho'\\|\tau'|=\sigma}}|f_{\tau'}|^2}^{\vphantom{\big|}}\right\|_{L_\#^{\frac{pd-m}{2}}(Q)}^{\frac{pd-m}{pd-m}},
\end{aligned} ≲ σ − C η 4 ρ ∈ J ( u ) ∑ Q ∈ G ρ Q ∈ G ρ ( α / σ − C η 4 ) ∑ ∣ Q ∣ τ ′ ⊂ ρ ∣ τ ′ ∣ = σ ′ ∑ ∣ f τ ′ ∣ 2 1/2 τ ′ ⊂ ρ ∣ τ ′ ∣ = σ ′ ∑ ∣ f τ ′ ∣ 2 L # 2 p d − m ( Q ) p d − m p d − m + σ − 2 δ ρ ′ ∈ J ( u ′ ) ∑ Q ∈ G ρ ′ Q ∈ G ρ ′ ( α / σ − 2 δ ) ∑ ∣ Q ∣ τ ′ ⊂ ρ ′ ∣ τ ′ ∣ = σ ∑ ∣ f τ ′ ∣ 2 1/2 τ ′ ⊂ ρ ′ ∣ τ ′ ∣ = σ ∑ ∣ f τ ′ ∣ 2 L # 2 p d − m ( Q ) p d − m p d − m ,
where R − 1 / d ≤ σ ′ ≤ max { R − 1 / d , σ 1 + η / 4 } R^{-1/d}\leq\sigma'\leq\max\{R^{-1/d},\sigma^{1+\eta/4}\} R − 1/ d ≤ σ ′ ≤ max { R − 1/ d , σ 1 + η /4 } and σ ≤ u ′ ≤ u σ η / 2 \sigma\leq u'\leq u\sigma^{\eta/2} σ ≤ u ′ ≤ u σ η /2 .
For the first sum over Q Q Q in (8.17), suppose that σ ′ > R − 1 / d \sigma'>R^{-1/d} σ ′ > R − 1/ d . Since σ ≤ A − 1 \sigma\leq A^{-1} σ ≤ A − 1 , we have σ ′ u ≤ σ η / 4 σ s < σ s / 2 \sigma'u\leq\sigma^{\eta/4}\sigma s<\sigma s/2 σ ′ u ≤ σ η /4 σ s < σ s /2 for sufficiently large A A A . Apply (9.24) inductively at the fixed parameter R R R , at scales ( σ ′ , u ) (\sigma',u) ( σ ′ , u ) and amplitude α / σ − C η 4 \alpha/\sigma^{-C\eta^4} α / σ − C η 4 . This bounds the first term in (8.17) by the RHS of (8.8), with both occurrences of R ϵ R^\epsilon R ϵ replaced by σ − C η 4 R ϵ ( σ ′ u ) ϵ \sigma^{-C\eta^4}R^\epsilon(\sigma'u)^\epsilon σ − C η 4 R ϵ ( σ ′ u ) ϵ , and
σ − C η 4 R ϵ ( σ ′ u ) ϵ R ϵ ( σ s ) ϵ ≤ σ ϵ η / 4 − C η 4 . \frac{\sigma^{-C\eta^4}R^\epsilon(\sigma'u)^\epsilon}{R^\epsilon(\sigma s)^\epsilon}\leq\sigma^{\epsilon\eta/4-C\eta^4}. R ϵ ( σ s ) ϵ σ − C η 4 R ϵ ( σ ′ u ) ϵ ≤ σ ϵη /4 − C η 4 .
This is the first factor in (8.18).
If σ ′ = R − 1 / d \sigma'=R^{-1/d} σ ′ = R − 1/ d , the first sum over Q Q Q in (8.17) contains the original functions f θ f_\theta f θ . Use the estimate proved for σ = R − 1 / d \sigma=R^{-1/d} σ = R − 1/ d before Case 1, with s = u s=u s = u and amplitude α / σ − C η 4 \alpha/\sigma^{-C\eta^4} α / σ − C η 4 . It bounds the first term in (8.17) by the RHS of (8.8), with both occurrences of R ϵ R^\epsilon R ϵ replaced by σ − C η 4 R η ≤ R C η \sigma^{-C\eta^4}R^\eta\leq R^{C\eta} σ − C η 4 R η ≤ R C η . Since s ≥ σ ≥ R − 1 / d s\geq\sigma\geq R^{-1/d} s ≥ σ ≥ R − 1/ d and C η ≪ ϵ C\eta\ll\epsilon C η ≪ ϵ , we have
R C η R ϵ ( σ s ) ϵ ≤ R − ϵ / C . \frac{R^{C\eta}}{R^\epsilon(\sigma s)^\epsilon}\leq R^{-\epsilon/C}. R ϵ ( σ s ) ϵ R C η ≤ R − ϵ / C .
For the second sum over Q Q Q in (8.17), regardless of σ ′ \sigma' σ ′ , we have σ u ′ ≤ σ η / 2 σ s < σ s / 2 \sigma u'\leq\sigma^{\eta/2}\sigma s<\sigma s/2 σ u ′ ≤ σ η /2 σ s < σ s /2 . Apply (9.24) inductively at scales ( σ , u ′ ) (\sigma,u') ( σ , u ′ ) and amplitude α / σ − 2 δ \alpha/\sigma^{-2\delta} α / σ − 2 δ . This bounds the second term in (8.17) by the RHS of (8.8), with both occurrences of R ϵ R^\epsilon R ϵ replaced by σ − 2 δ R ϵ ( σ u ′ ) ϵ \sigma^{-2\delta}R^\epsilon(\sigma u')^\epsilon σ − 2 δ R ϵ ( σ u ′ ) ϵ , and
σ − 2 δ R ϵ ( σ u ′ ) ϵ R ϵ ( σ s ) ϵ ≤ σ ϵ η / 2 − 2 δ . \frac{\sigma^{-2\delta}R^\epsilon(\sigma u')^\epsilon}{R^\epsilon(\sigma s)^\epsilon}\leq\sigma^{\epsilon\eta/2-2\delta}. R ϵ ( σ s ) ϵ σ − 2 δ R ϵ ( σ u ′ ) ϵ ≤ σ ϵη /2 − 2 δ .
This is the second factor in (8.18).
We now sum the estimates of Cases 1 and 2 over σ ≤ u ≤ s \sigma\leq u\leq s σ ≤ u ≤ s to prove (9.24) at the current σ s \sigma s σ s . Our choices of η , δ \eta,\delta η , δ make all three exponents in (8.18) positive. Since σ ≤ A − 1 \sigma\leq A^{-1} σ ≤ A − 1 , the positive powers of σ \sigma σ in (8.18) and the gain R − ϵ / C R^{-\epsilon/C} R − ϵ / C absorb the dyadic sum over u u u and the implicit constants when A A A and then R R R are sufficiently large. The sum of the outer coefficients and each amplitude denominator are therefore bounded by R ϵ ( σ s ) ϵ R^\epsilon(\sigma s)^\epsilon R ϵ ( σ s ) ϵ . This proves (9.24) at the current R R R by induction on σ s \sigma s σ s .
To prove (8.9) at the same R R R , apply Lemma 8.8 with ϵ / 2 \epsilon/2 ϵ /2 in place of ϵ \epsilon ϵ . For each s s s , apply (9.24) to the sum over ρ , Q \rho,Q ρ , Q on the RHS of the lemma, with σ = A − 1 \sigma=A^{-1} σ = A − 1 and amplitude α / A ϵ / 2 \alpha/A^{\epsilon/2} α / A ϵ /2 . Including the factor A ϵ / 2 A^{\epsilon/2} A ϵ /2 from Lemma 8.8, this bounds the fixed s s s term by the RHS of (8.8), with both occurrences of R ϵ R^\epsilon R ϵ replaced by
A ϵ / 2 R ϵ ( A − 1 s ) ϵ = R ϵ A − ϵ / 2 s ϵ . A^{\epsilon/2}R^\epsilon(A^{-1}s)^\epsilon=R^\epsilon A^{-\epsilon/2}s^\epsilon. A ϵ /2 R ϵ ( A − 1 s ) ϵ = R ϵ A − ϵ /2 s ϵ .
Since A − ϵ / 2 s ϵ ≤ 1 A^{-\epsilon/2}s^{\epsilon}\le1 A − ϵ /2 s ϵ ≤ 1 , each averaging box in these sums satisfies the selection condition in (8.9). Summing over s s s and using A − ϵ / 2 ∑ A − 1 ≤ s ≤ 1 s ϵ ≲ ϵ A − ϵ / 2 A^{-\epsilon/2}\sum_{A^{-1}\le s\le1}s^{\epsilon}\lesssim_{\epsilon} A^{-\epsilon/2} A − ϵ /2 ∑ A − 1 ≤ s ≤ 1 s ϵ ≲ ϵ A − ϵ /2 therefore proves (8.9) at R R R . This completes the induction on R R R , proving (9.24) and (8.9) for the current m m m .
To obtain Proposition 8.9 and (8.10), use (9.24) and (8.9) with ϵ / 2 \epsilon/2 ϵ /2 in place of ϵ \epsilon ϵ , and apply Lemma 8.5 to their right hand sides. The additional loss, including fixed constants, may be taken to be R η R^{\eta} R η in both the outer coefficient and the amplitude denominator. Since R σ ≥ R 1 − 2 / d R\sigma\ge R^{1-2/d} R σ ≥ R 1 − 2/ d and η ≪ ϵ \eta\ll\epsilon η ≪ ϵ ,
R η ( R σ ) ϵ / 2 ≤ ( R σ ) ϵ , R ϵ / 2 + η ≤ R ϵ . R^{\eta}(R\sigma)^{\epsilon/2}\le(R\sigma)^{\epsilon},\qquad R^{\epsilon/2+\eta}\le R^{\epsilon}. R η ( R σ ) ϵ /2 ≤ ( R σ ) ϵ , R ϵ /2 + η ≤ R ϵ .
The first inequality gives Proposition 8.9; the second gives (8.10) for supp f ^ ⊂ Γ j , 1 / K d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\Gamma^{d+1}_{j,1/K}(R) supp f ⊂ Γ j , 1/ K d + 1 ( R ) , 0 ≤ j ≤ m 0\le j\le m 0 ≤ j ≤ m . Hence C K m ( R ) ≲ 1 C^m_K(R)\lesssim1 C K m ( R ) ≲ 1 , and together with (8.9) this completes Step 1 in the proof of Proposition 8.7. □
From curves to higher dimensional cones over moment curves: Proof of Proposition 8.7, Step 2
Proof of Proposition 8.15. Let f f f be a Schwartz function with supp f ^ ⊂ Γ j , 1 / K , ℓ d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\Gamma^{d+1}_{j,1/K,\ell}(R) supp f ⊂ Γ j , 1/ K , ℓ d + 1 ( R ) . We choose small parameters 0 < δ ≪ η ≪ ϵ 0<\delta\ll\eta\ll\epsilon 0 < δ ≪ η ≪ ϵ as in Section 8.2 and induct on the dyadic product σ s \sigma s σ s , keeping R , j , ℓ R,j,\ell R , j , ℓ fixed.
We first deal with the case σ = R − 1 / d \sigma=R^{-1/d} σ = R − 1/ d , where the pieces f τ f_{\tau} f τ are the fine pieces f θ f_{\theta} f θ . By (2.22),
LHS of Proposition 8.15 ≲ ∑ ρ ∈ J ( s ) ∫ ∑ θ ⊂ ρ ∣ f θ ∣ 2 ≳ [ α / N ( s ) ] 2 ( ∑ θ ⊂ ρ ∣ f θ ∣ ) p d − m / 2 . \operatorname{LHS}\text{ of Proposition 8.15}\lesssim\sum_{\rho\in\mathcal{J}(s)}\int_{\sum_{\theta\subset\rho}|f_{\theta}|^2\gtrsim[\alpha/N(s)]^2}\left(\sum_{\theta\subset\rho}|f_{\theta}|\right)^{p_{d-m}/2}. LHS of Proposition 8.15 ≲ ρ ∈ J ( s ) ∑ ∫ ∑ θ ⊂ ρ ∣ f θ ∣ 2 ≳ [ α / N ( s ) ] 2 θ ⊂ ρ ∑ ∣ f θ ∣ p d − m /2 .
For each interval ρ = [ t ρ , t ρ + s ] ∈ J ( s ) \rho=[t_{\rho},t_{\rho}+s]\in\mathcal{J}(s) ρ = [ t ρ , t ρ + s ] ∈ J ( s ) , set g ( x ) = f ρ ( L t ρ , s − T x ) g(x)=f_{\rho}(L_{t_{\rho},s}^{-T}x) g ( x ) = f ρ ( L t ρ , s − T x ) . The coefficient bounds in (8.28) give supp g ^ ⊂ Γ m d + 1 ( R s d ) \operatorname{supp}\widehat{g}\subset\Gamma^{d+1}_{m}(R s^d) supp g ⊂ Γ m d + 1 ( R s d ) , up to fixed enlargement, with fine pieces f θ ( L t ρ , s − T x ) f_{\theta}(L_{t_{\rho},s}^{-T}x) f θ ( L t ρ , s − T x ) , θ ⊂ ρ \theta\subset\rho θ ⊂ ρ . Apply (9.2) to g g g with parameter R s d R s^d R s d , amplitude α / N ( s ) \alpha/N(s) α / N ( s ) , and loss η \eta η .
Returning to the original coordinates, (9.25) gives the convolution weights w B τ w_{B_{\tau}} w B τ and boxes B B B in (8.8), with R − 1 / d ≤ q 1 ≤ ⋯ ≤ q d − m − 2 ≤ u ≤ s R^{-1/d}\le q_1\le\cdots\le q_{d-m-2}\le u\le s R − 1/ d ≤ q 1 ≤ ⋯ ≤ q d − m − 2 ≤ u ≤ s . For a returned parent ρ ′ ⊂ ρ \rho'\subset\rho ρ ′ ⊂ ρ of width u u u , Lemma 2.18 gives N ( s ) N ρ ( u ) ≲ N ( u ) N(s)N_{\rho}(u)\lesssim N(u) N ( s ) N ρ ( u ) ≲ N ( u ) . Since ( R s d ) η ≤ R η (R s^d)^{\eta}\le R^{\eta} ( R s d ) η ≤ R η , it follows that α ( R s d ) η N ( s ) N ρ ( u ) ≳ α R η N ( u ) \frac{\alpha}{(R s^d)^{\eta}N(s)N_{\rho}(u)}\gtrsim\frac{\alpha}{R^{\eta}N(u)} ( R s d ) η N ( s ) N ρ ( u ) α ≳ R η N ( u ) α . Thus summing over ρ \rho ρ and extending u ≤ s u\le s u ≤ s to u ≤ 1 u\le1 u ≤ 1 bounds the LHS of Proposition 8.15 by the RHS of (8.8), with both occurrences of R ϵ R^{\epsilon} R ϵ replaced by R η R^{\eta} R η . Since σ = R − 1 / d \sigma=R^{-1/d} σ = R − 1/ d and s ≥ σ s\ge\sigma s ≥ σ ,
R η R ϵ ( σ s ) − ϵ / C ≤ R − ϵ / C . \frac{R^{\eta}}{R^{\epsilon}(\sigma s)^{-\epsilon/C}}\le R^{-\epsilon/C}. R ϵ ( σ s ) − ϵ / C R η ≤ R − ϵ / C .
for η \eta η sufficiently small in terms of ϵ \epsilon ϵ . This proves Proposition 8.15 at σ = R − 1 / d \sigma=R^{-1/d} σ = R − 1/ d .
We now consider the case σ > R − 1 / d \sigma>R^{-1/d} σ > R − 1/ d , so that σ ≤ a ℓ ≪ ℓ \sigma\le a_{\ell}\ll\ell σ ≤ a ℓ ≪ ℓ . Since Γ j , 1 / K , ℓ d + 1 ( R ) ⊂ Γ ~ j , ℓ d + 1 ( R ) \Gamma^{d+1}_{j,1/K,\ell}(R)\subset\widetilde{\Gamma}^{d+1}_{j,\ell}(R) Γ j , 1/ K , ℓ d + 1 ( R ) ⊂ Γ j , ℓ d + 1 ( R ) , Corollary 9.3, applied to f f f with loss δ \delta δ , bounds the LHS of Proposition 8.15 by the RHS of (8.13). Fix σ ≤ u ≤ s \sigma\le u\le s σ ≤ u ≤ s . We estimate the corresponding term on the RHS of (8.13), namely
σ − δ ∑ ρ ∈ J ( u ) ∑ B ∈ G ρ ( α σ δ ) B ⊂ ( u R 1 / d ) d − j / ( σ R 1 / d ) d − j U ρ , R ∣ B ∣ ( ∑ τ ⊂ ρ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 1 L # p d − m / 2 ( B ) . (148) \sigma^{-\delta}\sum_{\rho\in\mathcal{J}(u)}\sum_{\substack{B\in\mathcal{G}_{\rho}(\alpha\sigma^{\delta})\\ B\subset(uR^{1/d})^{d-j}/(\sigma R^{1/d})^{d-j}U_{\rho,R}}}|B|\left(\sum_{\substack{\tau\subset\rho\\ |\tau|=\sigma}}|f_{\tau}|^2\right)^{1/2} \frac{1}{L_{\#}^{p_{d-m}/2}(B)}. \tag*{(148)} σ − δ ρ ∈ J ( u ) ∑ B ∈ G ρ ( α σ δ ) B ⊂ ( u R 1/ d ) d − j / ( σ R 1/ d ) d − j U ρ , R ∑ ∣ B ∣ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L # p d − m /2 ( B ) 1 . (148)
As in the proof of (9.24), we distinguish u < σ 1 − η u<\sigma^{1-\eta} u < σ 1 − η , where we use (8.14) and then apply (8.9) after Taylor rescaling f τ f_{\tau} f τ , from u ≥ σ 1 − η u\ge\sigma^{1-\eta} u ≥ σ 1 − η , where Lemma 9.5 gives either ( σ ′ , u ) (\sigma',u) ( σ ′ , u ) with σ ′ < σ \sigma'<\sigma σ ′ < σ or ( σ , u ′ ) (\sigma,u') ( σ , u ′ ) with u ′ < u u'<u u ′ < u .
Case 1: u < σ 1 − η u<\sigma^{1-\eta} u < σ 1 − η . As in (8.14), we can bound (10.1) by
≲ σ − C η ∑ ρ ∈ J ( u ) ∑ τ ⊂ ρ ∣ τ ∣ = σ ∫ ∣ f τ ∣ 2 ≳ α σ δ / [ N ( u ) N ρ ( σ ) ] ∣ f τ ∣ p d − m . \lesssim\sigma^{-C\eta}\sum_{\rho\in\mathcal{J}(u)}\sum_{\substack{\tau\subset\rho\\ |\tau|=\sigma}}\int_{|f_{\tau}|^2\gtrsim\alpha\sigma^{\delta}/[N(u)N_{\rho}(\sigma)]}|f_{\tau}|^{p_{d-m}}. ≲ σ − C η ρ ∈ J ( u ) ∑ τ ⊂ ρ ∣ τ ∣ = σ ∑ ∫ ∣ f τ ∣ 2 ≳ α σ δ / [ N ( u ) N ρ ( σ )] ∣ f τ ∣ p d − m .
Since supp f τ ( L t τ , σ − T ⋅ ) ⊂ Γ j , 1 / K d + 1 ( R σ d ) \operatorname{supp}f_{\tau}(L_{t_{\tau},\sigma}^{-T}\cdot)\subset\Gamma^{d+1}_{j,1/K}(R\sigma^d) supp f τ ( L t τ , σ − T ⋅ ) ⊂ Γ j , 1/ K d + 1 ( R σ d ) for τ ∈ J ( σ ) \tau\in\mathcal{J}(\sigma) τ ∈ J ( σ ) , we can apply (8.9), proved in Step 1 for the current m m m , to this function with parameter R σ d R\sigma^d R σ d and amplitude α σ δ / [ N ( u ) N ρ ( σ ) ] \alpha\sigma^{\delta}/[N(u)N_{\rho}(\sigma)] α σ δ / [ N ( u ) N ρ ( σ )] .
Returning to the original coordinates, (9.25), with σ \sigma σ in place of s s s and v v v in place of u u u , gives the convolution weights w B τ w_{B_{\tau}} w B τ and boxes B B B in (8.8), with R − 1 / d ≤ q 1 ≤ ⋯ ≤ q d − m − 2 ≤ v ≤ σ R^{-1/d}\le q_1\le\cdots\le q_{d-m-2}\le v\le\sigma R − 1/ d ≤ q 1 ≤ ⋯ ≤ q d − m − 2 ≤ v ≤ σ . For a parent ρ ′ ⊂ τ \rho'\subset\tau ρ ′ ⊂ τ with લ аҭыԥ ∣ ρ ′ ∣ = v |\rho'|=v ∣ ρ ′ ∣ = v , applying Lemma 2.18 twice gives N ( u ) N ρ ( σ ) N τ ( v ) ≲ N ( v ) N(u)N_\rho(\sigma)N_\tau(v)\lesssim N(v) N ( u ) N ρ ( σ ) N τ ( v ) ≲ N ( v ) . Hence the amplitude in the selection condition satisfies
α σ δ ( R σ d ) ϵ N ( u ) N ρ ( σ ) N τ ( v ) ≳ α σ − C η ( R σ d ) ϵ N ( v ) . \frac{\alpha\sigma^\delta}{(R\sigma^d)^\epsilon N(u)N_\rho(\sigma)N_\tau(v)}
\gtrsim\frac{\alpha}{\sigma^{-C\eta}(R\sigma^d)^\epsilon N(v)}. ( R σ d ) ϵ N ( u ) N ρ ( σ ) N τ ( v ) α σ δ ≳ σ − C η ( R σ d ) ϵ N ( v ) α .
Summing over ρ \rho ρ , τ \tau τ and extending v ≤ σ v\leq\sigma v ≤ σ to v ≤ 1 v\leq1 v ≤ 1 therefore bounds (10.1) by the RHS of (8.8), with both occurrences of R ϵ R^\epsilon R ϵ replaced by σ − C η ( R σ d ) ϵ \sigma^{-C\eta}(R\sigma^d)^\epsilon σ − C η ( R σ d ) ϵ .
Since s ≥ σ s\geq\sigma s ≥ σ ,
σ − C η ( R σ d ) ϵ R ϵ ( σ s ) ϵ ≤ σ ( d − 2 ) ϵ − C η . \frac{\sigma^{-C\eta}(R\sigma^d)^\epsilon}{R^\epsilon(\sigma s)^\epsilon}\leq\sigma^{(d-2)\epsilon-C\eta}. R ϵ ( σ s ) ϵ σ − C η ( R σ d ) ϵ ≤ σ ( d − 2 ) ϵ − C η .
This is the third factor in (8.18) and finishes the fixed u u u estimate in Case 1.
Case 2: u ≥ σ 1 − η u\geq\sigma^{1-\eta} u ≥ σ 1 − η . For each interval ρ = [ t ρ , t ρ + u ] ∈ J ( u ) \rho=[t_\rho,t_\rho+u]\in\mathcal J(u) ρ = [ t ρ , t ρ + u ] ∈ J ( u ) , we have supp f ρ ( L ~ t ρ , u T ⋅ ) ⊂ Γ ~ j , min { ℓ / u , 1 } d + 1 ( R u d ) \operatorname{supp} f_\rho(\widetilde L_{t_\rho,u}^{T}\cdot)\subset\widetilde\Gamma_{j,\min\{\ell/u,1\}}^{d+1}(Ru^d) supp f ρ ( L t ρ , u T ⋅ ) ⊂ Γ j , m i n { ℓ / u , 1 } d + 1 ( R u d ) , as in the proof of Corollary 9.3. Since σ / u ≤ σ η \sigma/u\leq\sigma^\eta σ / u ≤ σ η and σ / u ≪ min { ℓ / u , 1 } \sigma/u\ll\min\{\ell/u,1\} σ / u ≪ min { ℓ / u , 1 } , we can apply Lemma 9.5 to this function with cap width σ / u \sigma/u σ / u and amplitude α σ δ / N ( u ) \alpha\sigma^\delta/N(u) α σ δ / N ( u ) . After applying Lemma 9.5 and returning to the original coordinates, we denote the new scales by ( σ ′ , u ′ ) (\sigma',u') ( σ ′ , u ′ ) . Summing over ρ \rho ρ and applying dyadic pigeonholing as in (8.17), we can bound (10.1) by
≲ σ − C η 4 ∑ ρ ∈ J ( u ) ∑ Q ∈ G ρ ( σ − C η 4 ) ∣ Q ∣ ∥ ( ∑ τ ′ ⊂ ρ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( Q ) + σ − 2 δ ∑ ρ ′ ∈ J ( u ′ ) ∑ Q ∈ G ρ ′ ( σ − 2 δ ) ∣ Q ∣ ∥ ( ∑ τ ⊂ ρ ′ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L # p d − m / 2 ( Q ) . (149) \begin{aligned}
&\lesssim\sigma^{-C\eta^4}\sum_{\rho\in\mathcal J(u)}
\sum_{Q\in\mathcal G_\rho(\sigma^{-C\eta^4})}
|Q|\left\|\left(\sum_{\substack{\tau'\subset\rho\\|\tau'|=\sigma'}}|f_{\tau'}|^2\right)^{1/2}\right\|_{L_\#^{pd-m/2}(Q)}
\\
&\quad+\sigma^{-2\delta}\sum_{\rho'\in\mathcal J(u')}
\sum_{Q\in\mathcal G_{\rho'}(\sigma^{-2\delta})}
|Q|\left\|\left(\sum_{\substack{\tau\subset\rho'\\|\tau|=\sigma}}|f_\tau|^2\right)^{1/2}\right\|_{L_\#^{pd-m/2}(Q)}. \tag*{(149)}
\end{aligned} ≲ σ − C η 4 ρ ∈ J ( u ) ∑ Q ∈ G ρ ( σ − C η 4 ) ∑ ∣ Q ∣ τ ′ ⊂ ρ ∣ τ ′ ∣ = σ ′ ∑ ∣ f τ ′ ∣ 2 1/2 L # p d − m /2 ( Q ) + σ − 2 δ ρ ′ ∈ J ( u ′ ) ∑ Q ∈ G ρ ′ ( σ − 2 δ ) ∑ ∣ Q ∣ τ ⊂ ρ ′ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L # p d − m /2 ( Q ) . (149)
where R − 1 / d ≤ σ ′ ≤ max { R − 1 / d , σ 1 + η / 4 } R^{-1/d}\leq\sigma'\leq\max\{R^{-1/d},\sigma^{1+\eta/4}\} R − 1/ d ≤ σ ′ ≤ max { R − 1/ d , σ 1 + η /4 } and σ ≤ u ′ ≤ u σ η / 2 \sigma\leq u'\leq u\sigma^{\eta/2} σ ≤ u ′ ≤ u σ η /2 . In particular, σ ′ ≤ σ ≤ a ℓ \sigma'\leq\sigma\leq a\ell σ ′ ≤ σ ≤ a ℓ , so both pairs ( σ ′ , u ) (\sigma',u) ( σ ′ , u ) and ( σ , u ′ ) (\sigma,u') ( σ , u ′ ) lie in the scale range of Proposition 8.15, with R R R , j j j , ℓ \ell ℓ fixed.
For the first sum over Q Q Q in (149) , suppose that σ ′ > R − 1 / d \sigma'>R^{-1/d} σ ′ > R − 1/ d . Then σ ′ u ≤ σ η / 4 σ s < σ s \sigma'u\leq\sigma^{\eta/4}\sigma s<\sigma s σ ′ u ≤ σ η /4 σ s < σ s . Apply the induction hypothesis of Proposition 8.15 at scales ( σ ′ , u ) (\sigma',u) ( σ ′ , u ) and amplitude α / σ − C η 4 \alpha/\sigma^{-C\eta^4} α / σ − C η 4 . This bounds the first term in (149) by the RHS of (8.8), with both occurrences of R ϵ R^\epsilon R ϵ replaced by σ − C η 4 R ϵ ( σ ′ u ) ϵ \sigma^{-C\eta^4}R^\epsilon(\sigma'u)^\epsilon σ − C η 4 R ϵ ( σ ′ u ) ϵ , and
σ − C η 4 R ϵ ( σ ′ u ) ϵ R ϵ ( σ s ) ϵ ≤ σ ϵ η / 4 − C η 4 . \frac{\sigma^{-C\eta^4}R^\epsilon(\sigma'u)^\epsilon}{R^\epsilon(\sigma s)^\epsilon}\leq\sigma^{\epsilon\eta/4-C\eta^4}. R ϵ ( σ s ) ϵ σ − C η 4 R ϵ ( σ ′ u ) ϵ ≤ σ ϵη /4 − C η 4 .
This is the first factor in (8.18).
If σ ′ = R − 1 / d \sigma'=R^{-1/d} σ ′ = R − 1/ d , the first sum over Q Q Q in (149) contains the fine pieces f θ f_\theta f θ . Use the σ = R − 1 / d \sigma=R^{-1/d} σ = R − 1/ d calculation at the beginning of this proof, with s = u s=u s = u and amplitude α / σ − C η 4 \alpha/\sigma^{-C\eta^4} α / σ − C η 4 . It bounds the first term in (149) by the RHS of (8.8), with both occurrences of R ϵ R^\epsilon R ϵ replaced by σ − C η 4 R η ≤ R C η \sigma^{-C\eta^4}R^\eta\leq R^{C\eta} σ − C η 4 R η ≤ R C η . Since s ≥ σ ≥ R − 1 / d s\geq\sigma\geq R^{-1/d} s ≥ σ ≥ R − 1/ d and C η ≪ ϵ C\eta\ll\epsilon C η ≪ ϵ , we have
R C η R ϵ ( σ s ) ϵ ≤ R − ϵ / C . \frac{R^{C\eta}}{R^\epsilon(\sigma s)^\epsilon}\leq R^{-\epsilon/C}. R ϵ ( σ s ) ϵ R C η ≤ R − ϵ / C .
For the second sum over Q Q Q in (149) , regardless of σ ′ \sigma' σ ′ , we have σ u ′ ≤ σ η / 2 σ s < σ s \sigma u'\leq\sigma^{\eta/2}\sigma s<\sigma s σ u ′ ≤ σ η /2 σ s < σ s . Apply the induction hypothesis of Proposition 8.15 at scales ( σ , u ′ ) (\sigma,u') ( σ , u ′ ) and amplitude α / σ − 2 δ \alpha/\sigma^{-2\delta} α / σ − 2 δ . This bounds the second term in (149) by the RHS of (8.8), with both occurrences of R ϵ R^\epsilon R ϵ replaced by σ − 2 δ R ϵ ( σ u ′ ) ϵ \sigma^{-2\delta}R^\epsilon(\sigma u')^\epsilon σ − 2 δ R ϵ ( σ u ′ ) ϵ , and
σ − 2 δ R ϵ ( σ u ′ ) ϵ R ϵ ( σ s ) ϵ ≤ σ ϵ η / 2 − 2 δ . \frac{\sigma^{-2\delta}R^\epsilon(\sigma u')^\epsilon}{R^\epsilon(\sigma s)^\epsilon}\leq\sigma^{\epsilon\eta/2-2\delta}. R ϵ ( σ s ) ϵ σ − 2 δ R ϵ ( σ u ′ ) ϵ ≤ σ ϵη /2 − 2 δ .
This is the second factor in (8.18).
We now sum the estimates of Cases 1 and 2 over σ ≤ u ≤ s \sigma\leq u\leq s σ ≤ u ≤ s to prove Proposition 8.15 at the current σ s \sigma s σ s . Our choices of η , δ \eta,\delta η , δ make all three exponents in (8.18) positive. Since σ ≤ a ℓ ≤ A − 1 \sigma\leq a\ell\leq A^{-1} σ ≤ a ℓ ≤ A − 1 , the positive powers of σ \sigma σ in (8.18) and the gain R − ϵ / C R^{-\epsilon/C} R − ϵ / C absorb the dyadic sum over u u u and the implicit constants when A A A and then R R R are sufficiently large. The amplitude denominators in (8.8) and the sum of the outer coefficients are therefore bounded by R ϵ ( σ s ) ϵ R^\epsilon(\sigma s)^\epsilon R ϵ ( σ s ) ϵ . This closes the induction on σ s \sigma s σ s and proves Proposition 8.15. □ \square □
Completion of Proposition 8.7, Step 2. Proposition 8.15 completes the proof of Proposition 8.14 as described in Section 8.4. Combining Proposition 8.14 with (8.27) and (8.29) gives (8.8) for functions with Fourier support in R j d + 1 ( R ) \mathbb{R}^{d+1}_{j}(R) R j d + 1 ( R ) . Lemma 8.5 then gives (95) , completing the induction step in j j j . Proceeding through j = 0 , … , m j=0,\ldots,m j = 0 , … , m , we obtain (93) and C m ( R ) ≲ 1 C^{m}(R)\lesssim1 C m ( R ) ≲ 1 , proving Step 2. □
The base case m = d − 2 m=d-2 m = d − 2
We prove Proposition 8.6. Here m = d − 2 m=d-2 m = d − 2 , so p d − m = 4 p_{d-m}=4 p d − m = 4 . We first establish the inequalities in Lemmas 8.10 and 9.5 for m = d − 2 m=d-2 m = d − 2 , and then follow Steps 1 and 2 in the proof of Proposition 8.7 as in Sections 9 and 10.
Proof of Proposition 8.6. To prove C K d − 2 ( R ) ≲ 1 C_{K}^{d-2}(R)\lesssim1 C K d − 2 ( R ) ≲ 1 , it suffices to establish the estimates of Lemmas 8.10 and 9.5 for m = d − 2 m=d-2 m = d − 2 . Taking ℓ = 1 \ell=1 ℓ = 1 in Lemma 9.5 gives Lemma 8.11, after which the argument in Sections 8.1–8.2 applies.
We first prove the estimate of Lemma 8.10 for m = d − 2 m=d-2 m = d − 2 . Let g g g be a Schwartz function with supp g ^ ⊂ R d − 2 d + 1 ( r ) \operatorname{supp}\widehat{g}\subset\mathbb{R}^{d+1}_{d-2}(r) supp g ⊂ R d − 2 d + 1 ( r ) . For each r − 1 / d < σ ≤ 1 r^{-1/d}<\sigma\leq1 r − 1/ d < σ ≤ 1 , put u = r − 1 / d σ − 1 u=r^{-1/d}\sigma^{-1} u = r − 1/ d σ − 1 and define
h τ = ∑ θ ⊂ τ ∣ g ^ θ ∣ 2 ∗ η ~ θ , σ d , τ ∈ J ( u ) . h_{\tau}=\sum_{\theta\subset\tau}|\widehat{g}_{\theta}|^{2}*\widetilde{\eta}^{d}_{\theta,\sigma},\qquad\tau\in\mathcal{J}(u). h τ = θ ⊂ τ ∑ ∣ g θ ∣ 2 ∗ η θ , σ d , τ ∈ J ( u ) .
Then (126) , with m = d − 2 m=d-2 m = d − 2 , gives
∫ ∑ θ ∣ g ^ θ ∣ 2 ≳ α 2 ( ∑ θ ∣ g ^ θ ∣ 2 ) 2 ≲ ε r ε ∫ ∣ ∑ θ ∣ g ^ θ ∣ 2 ∗ η ~ θ , ≤ r − 1 / d d ∣ ≳ ( α / r ε ) 2 ∣ ∑ θ ∣ g ^ θ ∣ 2 ∗ η ~ θ , ≤ r − 1 / d d ∣ 2 + r ε ∑ r − 1 / d < σ ≤ 1 ∫ ∣ ∑ τ ∈ J ( u ) h τ ∣ ≳ ( α / r ε ) 2 ∣ ∑ τ ∈ J ( u ) h τ ∣ 2 . \int_{\sum_{\theta}|\widehat{g}_{\theta}|^{2}\gtrsim\alpha^{2}}\left(\sum_{\theta}|\widehat{g}_{\theta}|^{2}\right)^{2}\lesssim_{\varepsilon}r^{\varepsilon}\int_{\left|\sum_{\theta}|\widehat{g}_{\theta}|^{2}*\widetilde{\eta}^{d}_{\theta,\leq r^{-1/d}}\right|\gtrsim(\alpha/r^{\varepsilon})^{2}}\left|\sum_{\theta}|\widehat{g}_{\theta}|^{2}*\widetilde{\eta}^{d}_{\theta,\leq r^{-1/d}}\right|^{2}
+r^{\varepsilon}\sum_{r^{-1/d}<\sigma\leq1}\int_{\left|\sum_{\tau\in\mathcal{J}(u)}h_{\tau}\right|\gtrsim(\alpha/r^{\varepsilon})^{2}}\left|\sum_{\tau\in\mathcal{J}(u)}h_{\tau}\right|^{2}. ∫ ∑ θ ∣ g θ ∣ 2 ≳ α 2 ( θ ∑ ∣ g θ ∣ 2 ) 2 ≲ ε r ε ∫ ∑ θ ∣ g θ ∣ 2 ∗ η θ , ≤ r − 1/ d d ≳ ( α / r ε ) 2 θ ∑ ∣ g θ ∣ 2 ∗ η θ , ≤ r − 1/ d d 2 + r ε r − 1/ d < σ ≤ 1 ∑ ∫ ∣ ∑ τ ∈ J ( u ) h τ ∣ ≳ ( α / r ε ) 2 τ ∈ J ( u ) ∑ h τ 2 .
As in (127) , the first term is bounded by the u = 1 u=1 u = 1 term on the RHS of Lemma 8.10. We therefore estimate each integral in the σ \sigma σ sum.
Fix σ \sigma σ . As in the proof of (128) , ∣ h τ ∣ ≲ ( ∑ θ ⊂ τ ∣ g ^ θ ∣ 2 ) ∗ w B τ |h_{\tau}|\lesssim\left(\sum_{\theta\subset\tau}|\widehat{g}_{\theta}|^{2}\right)*w_{B_{\tau}} ∣ h τ ∣ ≲ ( ∑ θ ⊂ τ ∣ g θ ∣ 2 ) ∗ w B τ , where, since r i / d σ i − d = r u d − i r^{i/d}\sigma^{i-d}=ru^{d-i} r i / d σ i − d = r u d − i ,
B τ = { x ∈ R d + 1 : ∣ x ⋅ φ d ( i ) ( t τ ) ∣ ≤ r u d − i for 0 ≤ i ≤ d } . B_{\tau}=\left\{x\in\mathbb{R}^{d+1}:|x\cdot\varphi_{d}^{(i)}(t_{\tau})|\leq ru^{d-i}\ \text{for }0\leq i\leq d\right\}. B τ = { x ∈ R d + 1 : ∣ x ⋅ φ d ( i ) ( t τ ) ∣ ≤ r u d − i for 0 ≤ i ≤ d } .
To apply L 2 L^{2} L 2 orthogonality to ∑ τ ∈ J ( u ) h τ \sum_{\tau\in\mathcal{J}(u)}h_{\tau} ∑ τ ∈ J ( u ) h τ , we first show that the Fourier supports of the h τ h_{\tau} h τ have bounded overlap.
Write ξ = ∑ i = 0 d λ i ( t ; ξ ) φ d ( i ) ( t ) \xi=\sum_{i=0}^{d}\lambda_{i}(t;\xi)\varphi_{d}^{(i)}(t) ξ = ∑ i = 0 d λ i ( t ; ξ ) φ d ( i ) ( t ) . Taylor expansion at t τ t_{\tau} t τ , using (125) , gives, for ξ ∈ supp h ^ τ \xi\in\operatorname{supp}\widehat{h}_{\tau} ξ ∈ supp h τ ,
∣ λ i ( t τ ; ξ ) ∣ ≲ r − 1 u i − d ( 0 ≤ i ≤ d ) , max 0 ≤ i < d r u d − i ∣ λ i ( t τ ; ξ ) ∣ ∼ 1. |\lambda_{i}(t_{\tau};\xi)|\lesssim r^{-1}u^{i-d}\quad(0\leq i\leq d),\qquad\max_{0\leq i<d}ru^{d-i}|\lambda_{i}(t_{\tau};\xi)|\sim1. ∣ λ i ( t τ ; ξ ) ∣ ≲ r − 1 u i − d ( 0 ≤ i ≤ d ) , 0 ≤ i < d max r u d − i ∣ λ i ( t τ ; ξ ) ∣ ∼ 1.
For fixed ξ \xi ξ , the coefficient λ d ( t ; ξ ) \lambda_{d}(t;\xi) λ d ( t ; ξ ) is a polynomial of degree at most d d d , whose derivatives are the preceding Taylor coefficients up to signs. The set { t : ∣ λ d ( t ; ξ ) ∣ ≲ r − 1 } \{t:|\lambda_{d}(t;\xi)|\lesssim r^{-1}\} { t : ∣ λ d ( t ; ξ ) ∣ ≲ r − 1 } has O ( 1 ) O(1) O ( 1 ) components. If a component J J J contains t τ t_{\tau} t τ with ξ ∈ supp h ^ τ \xi\in\operatorname{supp}\widehat{h}_{\tau} ξ ∈ supp h τ , the polynomial derivative bounds give ∣ λ i ( t τ ; ξ ) ∣ ≲ r − 1 ∣ J ∣ i − d |\lambda_{i}(t_{\tau};\xi)|\lesssim r^{-1}|J|^{i-d} ∣ λ i ( t τ ; ξ ) ∣ ≲ r − 1 ∣ J ∣ i − d , i < d i<d i < d . The annular condition therefore forces ∣ J ∣ ≲ u |J|\lesssim u ∣ J ∣ ≲ u . Each component contains only O ( 1 ) O(1) O ( 1 ) left endpoints t τ t_{\tau} t τ , so the Fourier supports of the h τ h_{\tau} h τ have bounded overlap.
To obtain the condition selecting the averaging boxes in Lemma 8.10, we decompose h τ h_{\tau} h τ using the cutoffs ψ τ , T \psi_{\tau,T} ψ τ , T constructed before (21) :
h τ = ∑ T ∥ ( u r 1 / d ) d − 2 U τ , r ψ τ , T h τ , ∑ T ψ τ , T = 1. h_{\tau}=\sum_{T\parallel(ur^{1/d})^{d-2}U_{\tau,r}}\psi_{\tau,T}h_{\tau},\qquad\sum_{T}\psi_{\tau,T}=1. h τ = T ∥ ( u r 1/ d ) d − 2 U τ , r ∑ ψ τ , T h τ , T ∑ ψ τ , T = 1.
Discard the wave packets satisfying ∥ ψ τ , T 1 / 2 h τ ∥ ∞ ≪ ( α / r ε ) 2 / N ( u ) \|\psi_{\tau,T}^{1/2}h_{\tau}\|_{\infty}\ll(\alpha/r^{\varepsilon})^{2}/N(u) ∥ ψ τ , T 1/2 h τ ∥ ∞ ≪ ( α / r ε ) 2 / N ( u ) , and denote the boxes of the remaining wave packets by T τ \mathcal{T}_{\tau} T τ . There are at most N ( u ) N(u) N ( u ) nonzero functions h τ h_{\tau} h τ , so ∑ τ , T ψ τ , T 1 / 2 ≤ 1 \sum_{\tau,T}\psi_{\tau,T}^{1/2}\leq1 ∑ τ , T ψ τ , T 1/2 ≤ 1 gives
∣ ∑ τ ∑ T ∉ T τ ψ τ , T h τ ∣ ≤ ∑ τ ∑ T ∉ T τ ψ τ , T 1 / 2 ∥ ψ τ , T 1 / 2 h τ ∥ ∞ ≤ 1 2 ( α / r ε ) 2 . \left|\sum_{\tau}\sum_{T\notin\mathcal{T}_{\tau}}\psi_{\tau,T}h_{\tau}\right|\leq\sum_{\tau}\sum_{T\notin\mathcal{T}_{\tau}}\psi_{\tau,T}^{1/2}\|\psi_{\tau,T}^{1/2}h_{\tau}\|_{\infty}\leq\frac{1}{2}(\alpha/r^{\varepsilon})^{2}. τ ∑ T ∈ / T τ ∑ ψ τ , T h τ ≤ τ ∑ T ∈ / T τ ∑ ψ τ , T 1/2 ∥ ψ τ , T 1/2 h τ ∥ ∞ ≤ 2 1 ( α / r ε ) 2 .
Since T T T has width r u d − i ru^{d-i} r u d − i in direction i i i , the wave packet cutoff ψ τ , T \psi_{\tau,T} ψ τ , T has Fourier width ≪ r − 1 u i − d \ll r^{-1}u^{i-d} ≪ r − 1 u i − d in that direction. The Fourier supports of ∑ T ∈ T τ ψ τ , T h τ \sum_{T\in\mathcal{T}_{\tau}}\psi_{\tau,T}h_{\tau} ∑ T ∈ T τ ψ τ , T h τ therefore also have bounded overlap as τ \tau τ varies. By this bounded overlapping property and Cauchy–Schwarz in T T T , we obtain
∫ ∣ ∑ τ h τ ∣ ≳ ( α / r ϵ ) 2 ∣ ∑ τ h τ ∣ 2 ≲ ∑ τ ∣ ∫ T ∈ T τ ψ τ , T h τ ∣ 2 ≲ ∑ τ ∑ T ∈ T τ ∫ ψ τ , T 3 / 2 ∣ h τ ∣ 2 ≲ ∑ τ ∑ T ∈ T τ ∣ T ∣ ∥ ψ τ , T 1 / 2 h τ ∥ ∞ 2 . \int_{\left|\sum_{\tau} h_{\tau}\right|\gtrsim(\alpha/r^{\epsilon})^{2}}\left|\sum_{\tau}h_{\tau}\right|^{2}
\lesssim\sum_{\tau}\left|\int_{T\in\mathcal{T}_{\tau}}\psi_{\tau,T}h_{\tau}\right|^{2}
\lesssim\sum_{\tau}\sum_{T\in\mathcal{T}_{\tau}}\int\psi_{\tau,T}^{3/2}|h_{\tau}|^{2}
\lesssim\sum_{\tau}\sum_{T\in\mathcal{T}_{\tau}}|T|\|\psi_{\tau,T}^{1/2}h_{\tau}\|_{\infty}^{2}. ∫ ∣ ∑ τ h τ ∣ ≳ ( α / r ϵ ) 2 τ ∑ h τ 2 ≲ τ ∑ ∫ T ∈ T τ ψ τ , T h τ 2 ≲ τ ∑ T ∈ T τ ∑ ∫ ψ τ , T 3/2 ∣ h τ ∣ 2 ≲ τ ∑ T ∈ T τ ∑ ∣ T ∣∥ ψ τ , T 1/2 h τ ∥ ∞ 2 .
Here Cauchy–Schwarz uses ∑ T ∈ T τ ψ τ , T 1 / 2 ≲ 1 \sum_{T\in\mathcal{T}_{\tau}}\psi_{\tau,T}^{1/2}\lesssim1 ∑ T ∈ T τ ψ τ , T 1/2 ≲ 1 , and the last inequality uses ∫ ψ τ , T 1 / 2 ≲ ∣ T ∣ \int\psi_{\tau,T}^{1/2}\lesssim|T| ∫ ψ τ , T 1/2 ≲ ∣ T ∣ .
Applying (21) to h τ h_{\tau} h τ with exponent one and using the pointwise bound for h τ h_{\tau} h τ , we obtain
∥ ψ τ , T 1 / 2 h τ ∥ ∞ ≲ ∥ ( ∑ θ ⊂ τ ∣ g θ ∣ 2 ) ∗ w B τ ∥ L # 1 ( T ) 1 / 2 ≲ ∥ ( ∑ θ ⊂ τ ∣ g θ ∣ 2 ) ∥ L # 2 ( T ) 1 / 2 . \|\psi_{\tau,T}^{1/2}h_{\tau}\|_{\infty}\lesssim\left\|\left(\sum_{\theta\subset\tau}|g_{\theta}|^{2}\right)*w_{B_{\tau}}\right\|_{L^{1}_{\#}(T)}^{1/2}\lesssim\left\|\left(\sum_{\theta\subset\tau}|g_{\theta}|^{2}\right)\right\|_{L^{2}_{\#}(T)}^{1/2}. ∥ ψ τ , T 1/2 h τ ∥ ∞ ≲ ( θ ⊂ τ ∑ ∣ g θ ∣ 2 ) ∗ w B τ L # 1 ( T ) 1/2 ≲ ( θ ⊂ τ ∑ ∣ g θ ∣ 2 ) L # 2 ( T ) 1/2 .
Since T T T is a translate of B τ B_{\tau} B τ , we have w B τ ∗ w T ≲ w T w_{B_{\tau}}*w_{T}\lesssim w_{T} w B τ ∗ w T ≲ w T . The last inequality follows from (36) .
Since ( α / r ϵ ) 2 / N ( u ) ≥ [ α / ( r ϵ N ( u ) ) ] 2 (\alpha/r^{\epsilon})^{2}/N(u)\geq[\alpha/(r^{\epsilon}N(u))]^{2} ( α / r ϵ ) 2 / N ( u ) ≥ [ α / ( r ϵ N ( u )) ] 2 , every T ∈ T τ T\in\mathcal{T}_{\tau} T ∈ T τ belongs to G τ ( α / r ϵ ) \mathcal{G}_{\tau}(\alpha/r^{\epsilon}) G τ ( α / r ϵ ) . Substituting the wave packet bound into the preceding L 2 L^{2} L 2 estimate gives
∫ ∣ ∑ τ h τ ∣ ≳ ( α / r ϵ ) 2 ∣ ∑ τ h τ ∣ 2 ≲ ∑ τ ∈ J ( u ) ∑ T ∈ G τ ( α / r ϵ ) ∣ T ∣ ∥ ( ∑ θ ⊂ τ ∣ g θ ∣ 2 ) 1 / 2 ∥ L # 2 ( T ) 4 ( u r 1 / d ) d − 2 U τ , r . (150) \int_{\left|\sum_{\tau} h_{\tau}\right|\gtrsim(\alpha/r^{\epsilon})^{2}}\left|\sum_{\tau}h_{\tau}\right|^{2}
\lesssim\sum_{\tau\in\mathcal{J}(u)}\sum_{\substack{T\in\mathcal{G}_{\tau}(\alpha/r^{\epsilon})}}\left|T\right|\left\|\left(\sum_{\theta\subset\tau}|g_{\theta}|^{2}\right)^{1/2}\right\|_{L^{2}_{\#}(T)}^{4}(ur^{1/d})^{d-2}U_{\tau,r}. \tag*{(150)} ∫ ∣ ∑ τ h τ ∣ ≳ ( α / r ϵ ) 2 τ ∑ h τ 2 ≲ τ ∈ J ( u ) ∑ T ∈ G τ ( α / r ϵ ) ∑ ∣ T ∣ ( θ ⊂ τ ∑ ∣ g θ ∣ 2 ) 1/2 L # 2 ( T ) 4 ( u r 1/ d ) d − 2 U τ , r . (150)
Summing (150) over σ \sigma σ , with u = r − 1 / d σ − 1 u=r^{-1/d}\sigma^{-1} u = r − 1/ d σ − 1 , and combining this with the u = 1 u=1 u = 1 bound for the first term in (126) proves the estimate of Lemma 8.10 for m = d − 2 m=d-2 m = d − 2 .
We next prove the estimate of Lemma 9.5 for m = d − 2 m=d-2 m = d − 2 . Fix 0 ≤ j ≤ d − 2 0\leq j\leq d-2 0 ≤ j ≤ d − 2 and R − 1 / d < σ ≪ ℓ ≤ 1 R^{-1/d}<\sigma\ll\ell\leq1 R − 1/ d < σ ≪ ℓ ≤ 1 . Let f f f be a Schwartz function with supp f ^ ⊂ Γ j , ℓ d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\Gamma_{j,\ell}^{d+1}(R) supp f ⊂ Γ j , ℓ d + 1 ( R ) . Since p d − m / 4 = 1 p_{d-m/4}=1 p d − m /4 = 1 , we now refine the caps directly in the weighted average over B B B of ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 \sum_{|\tau|=\sigma}|f_{\tau}|^{2} ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 .
For B ∥ [ 0 , R j / d σ − 2 ] d + 1 B\parallel[0,R^{j/d}\sigma^{-2}]^{d+1} B ∥ [ 0 , R j / d σ − 2 ] d + 1 , replace W B W_{B} W B by a comparable weight whose Fourier transform is supported in C B ∗ CB^{*} C B ∗ , and expand the weighted average over B B B of ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 \sum_{|\tau|=\sigma}|f_{\tau}|^{2} ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 . The calculation leading to (134) , with u = 1 u=1 u = 1 , shows that only pairs of caps at scale R − 1 / d R^{-1/d} R − 1/ d whose intervals have left endpoints satisfying ∣ t − t ′ ∣ ≲ R − 1 / d + σ 2 |t-t'|\lesssim R^{-1/d}+\sigma^{2} ∣ t − t ′ ∣ ≲ R − 1/ d + σ 2 contribute to the average. Choose a dyadic σ ′ ∼ max { R − 1 / d , σ 2 } \sigma'\sim\max\{R^{-1/d},\sigma^{2}\} σ ′ ∼ max { R − 1/ d , σ 2 } . Only neighbouring σ ′ \sigma' σ ′ intervals interact, so Cauchy–Schwarz gives
∥ ( ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L # 2 ( B ) 2 ≲ ∥ ( ∑ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ) 1 / 2 ∥ L # 2 ( B ) 2 . (151) \left\|\left(\sum_{|\tau|=\sigma}|f_{\tau}|^{2}\right)^{1/2}\right\|_{L^{2}_{\#}(B)}^{2}
\lesssim
\left\|\left(\sum_{|\tau'|=\sigma'}|f_{\tau'}|^{2}\right)^{1/2}\right\|_{L^{2}_{\#}(B)}^{2}. \tag*{(151)} ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L # 2 ( B ) 2 ≲ ∣ τ ′ ∣ = σ ′ ∑ ∣ f τ ′ ∣ 2 1/2 L # 2 ( B ) 2 . (151)
For B ∈ G [ 0 , 1 ] ( α ) B\in\mathcal{G}_{[0,1]}(\alpha) B ∈ G [ 0 , 1 ] ( α ) , the LHS of (151) is at least α 2 \alpha^{2} α 2 . Apply (32) with F = ∑ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 F=\sum_{|\tau'|=\sigma'}|f_{\tau'}|^{2} F = ∑ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 and t = α 2 t=\alpha^{2} t = α 2 . Together with (151) and the locally constant property on unit cubes, this gives
∑ B ∥ [ 0 , R j / d σ − 2 ] d + 1 B ∈ G [ 0 , 1 ] ( α ) ∣ B ∣ ∥ ( ∑ ∣ τ ∣ = σ ∣ f τ ∣ 2 ) 1 / 2 ∥ L # 2 ( B ) 4 ≲ ∑ Q ∥ [ 0 , 1 ] d + 1 Q ∈ G [ 0 , 1 ] ( α ) ∣ Q ∣ ∥ ( ∑ ∣ τ ′ ∣ = σ ′ ∣ f τ ′ ∣ 2 ) 1 / 2 ∥ L # 2 ( Q ) 4 . (152) \sum_{\substack{B\parallel[0,R^{j/d}\sigma^{-2}]^{d+1}\\B\in\mathcal{G}_{[0,1]}(\alpha)}}|B|
\left\|\left(\sum_{|\tau|=\sigma}|f_{\tau}|^{2}\right)^{1/2}\right\|_{L^{2}_{\#}(B)}^{4}
\lesssim
\sum_{\substack{Q\parallel[0,1]^{d+1}\\Q\in\mathcal{G}_{[0,1]}(\alpha)}}|Q|
\left\|\left(\sum_{|\tau'|=\sigma'}|f_{\tau'}|^{2}\right)^{1/2}\right\|_{L^{2}_{\#}(Q)}^{4}. \tag*{(152)} B ∥ [ 0 , R j / d σ − 2 ] d + 1 B ∈ G [ 0 , 1 ] ( α ) ∑ ∣ B ∣ ∣ τ ∣ = σ ∑ ∣ f τ ∣ 2 1/2 L # 2 ( B ) 4 ≲ Q ∥ [ 0 , 1 ] d + 1 Q ∈ G [ 0 , 1 ] ( α ) ∑ ∣ Q ∣ ∣ τ ′ ∣ = σ ′ ∑ ∣ f τ ′ ∣ 2 1/2 L # 2 ( Q ) 4 . (152)
This proves the estimate of Lemma 9.5 for m = d − 2 m=d-2 m = d − 2 . In particular, for supp f ^ ⊂ Γ j , 1 / K d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\Gamma_{j,1/K}^{d+1}(R) supp f ⊂ Γ j , 1/ K d + 1 ( R ) , taking ℓ = 1 \ell=1 ℓ = 1 gives Lemma 8.11 for m = d − 2 m=d-2 m = d − 2 .
The argument in Sections 8.1–8.2 therefore gives Proposition 8.9 and (95) for m = d − 2 m=d-2 m = d − 2 , and hence C K d − 2 ( R ) ≲ 1 C_{K}^{d-2}(R)\lesssim1 C K d − 2 ( R ) ≲ 1 , completing Step 1.
We now prove C K d − 2 ( R ) ≲ 1 C_{K}^{d-2}(R)\lesssim1 C K d − 2 ( R ) ≲ 1 . As in the reduction in Section 8.4, it suffices to establish Lemma 8.13 and Proposition 8.14 for m = d − 2 m=d-2 m = d − 2 , replacing the RHS of (93) by the RHS of (95) in both statements. Lemma 8.13 holds also for m = d − 2 m=d-2 m = d − 2 , since its proof uses no inductive hypothesis.
In proving Proposition 8.14 for m = d − 2 m=d-2 m = d − 2 , we use (95) in place of the hypothesis (94) for functions with supp f ^ ⊂ Γ j , 1 / K d + 1 ( R ) \operatorname{supp}\widehat{f}\subset\Gamma_{j,1/K}^{d+1}(R) supp f ⊂ Γ j , 1/ K d + 1 ( R ) . This follows from the estimates of Lemmas 8.10 and 9.5 proved in Step 1, by the argument in Sections 8.1–8.2. Following the proof of Proposition 8.14, it therefore remains to prove the estimate of Proposition 8.15 for m = d − 2 m=d-2 m = d − 2 , with the RHS of (93) replaced by the RHS of (95) .
Using (95) at m = d − 2 m=d-2 m = d − 2 in place of the hypothesis (94) , we follow the proof in Section 10 with the RHS of (93) replaced throughout by the RHS of (95) . The estimates of Lemmas 8.10 and 9.5 established above, together with the rescaling and counting arguments in Section 10, give the same bounds for the outer coefficients and amplitude denominators. The comparisons involving the convolution weights in (8.8) are therefore unnecessary, and the induction on σ s \sigma_s σ s closes as before. This proves the estimate of Proposition 8.15 for m = d − 2 m=d-2 m = d − 2 . The reduction in Section 8.4 then gives C d − 2 ( R ) ≲ 1 C^{d-2}(R)\lesssim1 C d − 2 ( R ) ≲ 1 , completing the proof of Proposition 8.6.
Starting with Proposition 8.6, Proposition 8.7 proves (8.8) and C m ( R ) ≲ 1 C^m(R)\lesssim1 C m ( R ) ≲ 1 successively for m = d − 3 , … , 0 m=d-3,\ldots,0 m = d − 3 , … , 0 . Taking m = 0 m=0 m = 0 proves Proposition 8.1. The induction on dimension starts with the cone over the moment curve in R 3 \mathbb{R}^3 R 3 , the d = 2 d=2 d = 2 case of Theorem 1.8. Proposition 2.1 then gives the corresponding estimate for the moment curve in R 3 \mathbb{R}^3 R 3 . Now suppose the estimates for moment curves hold in dimensions 2 ≤ k ≤ d 2\le k\le d 2 ≤ k ≤ d . Proposition 8.1 gives the estimate for the cone over the moment curve in R d + 1 \mathbb{R}^{d+1} R d + 1 , and Proposition 2.1 gives the estimate for the moment curve in the same dimension. This completes the induction and proves Theorems 1.7 and 1.8.
Small cap decoupling inequalities
In this section, we prove Theorem 1.4 using the amplitude dependent wave envelope type estimate of Theorem 1.7. Throughout this section, γ \gamma γ denotes a small cap in the decomposition of M n ( R β , R ) \mathcal{M}^n(\mathbb{R}^\beta,R) M n ( R β , R ) from the introduction, and supp f γ ^ ⊂ γ \operatorname{supp}\widehat{f_\gamma}\subset\gamma supp f γ ⊂ γ . For each dyadic scale R − β ≤ s ≤ 1 R^{-\beta}\le s\le1 R − β ≤ s ≤ 1 , we use ν \nu ν to denote the cap obtained by adding the vertical segment { 0 } n − 1 × [ − R − 1 , R − 1 ] \{0\}^{n-1}\times[-R^{-1},R^{-1}] { 0 } n − 1 × [ − R − 1 , R − 1 ] to the caps in Θ ( s ) \Theta(s) Θ ( s ) . Write ∣ ν ∣ = s |\nu|=s ∣ ν ∣ = s in such a case, and t ν t_\nu t ν for the left endpoint of the interval corresponding to the cap. The cap is contained in a rectangle with dimension
s × s 2 × ⋯ × s n − 1 × max { s n , R − 1 } . s\times s^2\times\cdots\times s^{n-1}\times\max\{s^n,R^{-1}\}. s × s 2 × ⋯ × s n − 1 × max { s n , R − 1 } .
Let us state the main proposition in this section.
Proposition 12.1. Let n ≥ 4 n\ge4 n ≥ 4 , R ≥ 1 R\ge1 R ≥ 1 , β ≥ 1 / n \beta\ge1/n β ≥ 1/ n , and p = n ( n − 1 ) + 2 / β p=n(n-1)+2/\beta p = n ( n − 1 ) + 2/ β . If ∥ f γ ∥ ∞ ≤ 1 \|f_\gamma\|_\infty\le1 ∥ f γ ∥ ∞ ≤ 1 for every small cap γ \gamma γ , then for every ϵ > 0 \epsilon>0 ϵ > 0 and α > 0 \alpha>0 α > 0 ,
α p ∣ { ∣ ∑ γ f γ ∣ > α } ∣ ≲ ϵ R 1 + β ( n ( n − 1 ) / 2 − 1 ) + β ϵ ∑ γ ∥ f γ ∥ 2 2 (153) \alpha^p\left|\left\{\left|\sum_\gamma f_\gamma\right|>\alpha\right\}\right|\lesssim_\epsilon R^{1+\beta(n(n-1)/2-1)+\beta\epsilon}\sum_\gamma\|f_\gamma\|_2^2 \tag*{(153)} α p { γ ∑ f γ > α } ≲ ϵ R 1 + β ( n ( n − 1 ) /2 − 1 ) + β ϵ γ ∑ ∥ f γ ∥ 2 2 (153)
After the standard reduction, one may see that this proposition implies Theorem 1.4. For the completeness, we give the details at the end of this section.
To prove Proposition 12.1, we consider two cases. The first case n ≥ 5 n\ge5 n ≥ 5 or n = 4 n=4 n = 4 , β ≤ 1 \beta\le1 β ≤ 1 is proved in Subsection 12.2, and the remaining case n = 4 n=4 n = 4 , β > 1 \beta>1 β > 1 is proved in Subsection 12.3.
Key inequalities
We state three lemmas to prove the proposition. Recall from Subsection 1.1 that Θ ( s ) \Theta(s) Θ ( s ) denotes the family of caps in R n \mathbb{R}^n R n of dimensions s × s 2 × ⋯ × s n s\times s^2\times\cdots\times s^n s × s 2 × ⋯ × s n , and Ξ ( s ) \Xi(s) Ξ ( s ) the family of caps in R d + 1 \mathbb{R}^{d+1} R d + 1 of dimensions 1 × s × ⋯ × s d 1\times s\times\cdots\times s^d 1 × s × ⋯ × s d .
Lemma 12.2 (Canonical decoupling). Let n , d ≥ 2 n,d\ge2 n , d ≥ 2 . For 2 ≤ p ≤ n ( n + 1 ) 2\le p\le n(n+1) 2 ≤ p ≤ n ( n + 1 ) we have
∥ ∑ τ ∈ Θ ( s ) f τ ∥ p ≲ ϵ s − ϵ ( ∑ τ ∈ Θ ( s ) ∥ f τ ∥ p 2 ) 1 / 2 (154) \left\|\sum_{\tau\in\Theta(s)}f_\tau\right\|_p\lesssim_\epsilon s^{-\epsilon}\left(\sum_{\tau\in\Theta(s)}\|f_\tau\|_p^2\right)^{1/2} \tag*{(154)} τ ∈ Θ ( s ) ∑ f τ p ≲ ϵ s − ϵ τ ∈ Θ ( s ) ∑ ∥ f τ ∥ p 2 1/2 (154)
for any function f τ f_\tau f τ with supp f τ ^ ⊂ τ \operatorname{supp}\widehat{f_\tau}\subset\tau supp f τ ⊂ τ . For 2 ≤ p ≤ d ( d + 1 ) 2\le p\le d(d+1) 2 ≤ p ≤ d ( d + 1 ) ,
∥ ∑ τ ∈ Ξ ( s ) f τ ∥ p ≲ ϵ s − ϵ ( ∑ τ ∈ Ξ ( s ) ∥ f τ ∥ p 2 ) 1 / 2 (155) \left\|\sum_{\tau\in\Xi(s)}f_\tau\right\|_p\lesssim_\epsilon s^{-\epsilon}\left(\sum_{\tau\in\Xi(s)}\|f_\tau\|_p^2\right)^{1/2} \tag*{(155)} τ ∈ Ξ ( s ) ∑ f τ p ≲ ϵ s − ϵ τ ∈ Ξ ( s ) ∑ ∥ f τ ∥ p 2 1/2 (155)
for any function f τ f_\tau f τ with supp f τ ^ ⊂ τ \operatorname{supp}\widehat{f_\tau}\subset\tau supp f τ ⊂ τ .
The first estimate is the decoupling of Bourgain-Demeter-Guth [3 ] : decoupling for the cone over the moment curve in R d + 1 \mathbb{R}^{d+1} R d + 1 follows from [KLÖ23, Corollary 2.13]. Recall that p n = n ( n + 1 ) 2 + 1 p_n=\frac{n(n+1)}{2}+1 p n = 2 n ( n + 1 ) + 1 . The following is a consequence of Theorem 1.7.
Lemma 12.3. Let n ≥ 3 n\ge3 n ≥ 3 and let supp f θ ^ ⊂ θ \operatorname{supp}\widehat{f_\theta}\subset\theta supp f θ ⊂ θ for θ ∈ Θ ( R − 1 / n ) \theta\in\Theta(R^{-1/n}) θ ∈ Θ ( R − 1/ n ) . For every ϵ > 0 \epsilon>0 ϵ > 0 and α > 0 \alpha>0 α > 0 ,
α p n ∣ { ∣ ∑ θ f θ ∣ > α } ∣ ≤ C ϵ R ϵ ∑ R − 1 / n ≤ s ≤ 1 ∑ τ ∈ Θ ( s ) ∫ { ∑ θ ⊂ τ ∣ f θ ∣ 2 > C ϵ − 1 R − ϵ α 2 s 2 } ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p n / 2 . (156) \alpha^{p_n}\left|\left\{\left|\sum_\theta f_\theta\right|>\alpha\right\}\right|
\le C_\epsilon R^\epsilon\sum_{R^{-1/n}\le s\le1}\sum_{\tau\in\Theta(s)}
\int_{\left\{\sum_{\theta\subset\tau}|f_\theta|^2>C_\epsilon^{-1}R^{-\epsilon}\alpha^2s^2\right\}}
\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{p_n/2}. \tag*{(156)} α p n { θ ∑ f θ > α } ≤ C ϵ R ϵ R − 1/ n ≤ s ≤ 1 ∑ τ ∈ Θ ( s ) ∑ ∫ { ∑ θ ⊂ τ ∣ f θ ∣ 2 > C ϵ − 1 R − ϵ α 2 s 2 } ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p n /2 . (156)
Proof. Since # τ ≲ s − 1 \#\tau\lesssim s^{-1} # τ ≲ s − 1 , every good averaging box in Theorem 1.7 , applied with loss ϵ / 2 \epsilon/2 ϵ /2 , satisfies
1 ∣ V ∣ ∫ ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p n / 4 W V ≳ ϵ ( R − ϵ α 2 s ) p n / 4 . \frac{1}{|V|}\int\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{p_n/4}W_V \gtrsim_\epsilon\left(R^{-\epsilon}\alpha^2s\right)^{p_n/4}. ∣ V ∣ 1 ∫ ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p n /4 W V ≳ ϵ ( R − ϵ α 2 s ) p n /4 .
For C ϵ C_\epsilon C ϵ sufficiently large, the part where ∑ θ ⊂ τ ∣ f θ ∣ 2 ≤ C ϵ − 1 R − ϵ α 2 s 2 \sum_{\theta\subset\tau}|f_\theta|^2\le C_\epsilon^{-1}R^{-\epsilon}\alpha^2s^2 ∑ θ ⊂ τ ∣ f θ ∣ 2 ≤ C ϵ − 1 R − ϵ α 2 s 2 contributes at most half this average. In other words,
1 ∣ V ∣ ∫ { ∑ θ ⊂ τ ∣ f θ ∣ 2 ≤ C ϵ − 1 R − ϵ α 2 s 2 } ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p n / 4 W V ≤ C ϵ − p n / 4 ( R − ϵ α 2 s 2 ) p n / 4 . \frac{1}{|V|}\int_{\{\sum_{\theta\subset\tau}|f_\theta|^2\le C_\epsilon^{-1}R^{-\epsilon}\alpha^2s^2\}}\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{p_n/4}W_V\le C_\epsilon^{-p_n/4}\left(R^{-\epsilon}\alpha^2s^2\right)^{p_n/4}. ∣ V ∣ 1 ∫ { ∑ θ ⊂ τ ∣ f θ ∣ 2 ≤ C ϵ − 1 R − ϵ α 2 s 2 } ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p n /4 W V ≤ C ϵ − p n /4 ( R − ϵ α 2 s 2 ) p n /4 .
So we can ignore this term. For the good averaging box, Cauchy–Schwarz on the remaining set therefore gives
1 ∣ V ∣ [ ∫ ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p n / 4 W V ] 2 ≤ 4 ∫ { ∑ θ ⊂ τ ∣ f θ ∣ 2 > C ϵ − 1 R − ϵ α 2 s 2 } ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p n / 2 W V . \frac{1}{|V|}\left[\int\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{p_n/4}W_V\right]^2\le4\int_{\{\sum_{\theta\subset\tau}|f_\theta|^2>C_\epsilon^{-1}R^{-\epsilon}\alpha^2s^2\}}\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{p_n/2}W_V. ∣ V ∣ 1 ∫ ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p n /4 W V 2 ≤ 4 ∫ { ∑ θ ⊂ τ ∣ f θ ∣ 2 > C ϵ − 1 R − ϵ α 2 s 2 } ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p n /2 W V .
Summing over the translates of V τ , R V_{\tau,R} V τ , R and using Theorem 1.7 proves the lemma. □
Lemma 12.4 (Key lemma). Let n ≥ 4 n\ge4 n ≥ 4 , R − β ≤ σ ≤ s ≤ 1 R^{-\beta}\le\sigma\le s\le1 R − β ≤ σ ≤ s ≤ 1 , and ν \nu ν satisfy ∣ ν ∣ = s |\nu|=s ∣ ν ∣ = s . Suppose ∥ f γ ∥ ∞ ≤ 1 \|f_\gamma\|_\infty\le1 ∥ f γ ∥ ∞ ≤ 1 . For ν ′ ⊂ ν \nu'\subset\nu ν ′ ⊂ ν with ∣ ν ′ ∣ = σ |\nu'|=\sigma ∣ ν ′ ∣ = σ , we have the following estimates.
(i) Pointwise bound.
∑ ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∣ ∑ γ ⊂ ν ′ f γ ∣ 2 ≲ R 2 β σ s . (157) \sum_{\substack{\nu'\subset\nu\\|\nu'|=\sigma}}\left|\sum_{\gamma\subset\nu'}f_\gamma\right|^2\lesssim R^{2\beta}\sigma s. \tag*{(157)} ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∑ γ ⊂ ν ′ ∑ f γ 2 ≲ R 2 β σ s . (157) (ii) Moment bound.
∫ ( ∑ ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∣ ∑ γ ⊂ ν ′ f γ ∣ 2 ) n ( n − 1 ) / 2 ≲ ϵ R β ϵ ( R β s ) n ( n − 1 ) / 2 − 1 ∑ γ ⊂ ν ∥ f γ ∥ 2 2 . (158) \int\left(\sum_{\substack{\nu'\subset\nu\\|\nu'|=\sigma}}\left|\sum_{\gamma\subset\nu'}f_\gamma\right|^2\right)^{n(n-1)/2}\lesssim_\epsilon R^{\beta\epsilon}(R^\beta s)^{n(n-1)/2-1}\sum_{\gamma\subset\nu}\|f_\gamma\|_2^2. \tag*{(158)} ∫ ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∑ γ ⊂ ν ′ ∑ f γ 2 n ( n − 1 ) /2 ≲ ϵ R β ϵ ( R β s ) n ( n − 1 ) /2 − 1 γ ⊂ ν ∑ ∥ f γ ∥ 2 2 . (158) (iii) Restricted moment bound. For a sufficiently large constant C C C ,
∫ { ∑ ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∑ γ ⊂ ν ′ ∣ f γ ∣ 2 > C R β s } ( ∑ ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∣ ∑ γ ⊂ ν ′ f γ ∣ 2 ) n ( n − 1 ) / 2 \int_{\left\{\sum_{\substack{\nu'\subset\nu\\|\nu'|=\sigma}}\sum_{\gamma\subset\nu'}|f_\gamma|^2>CR^\beta s\right\}}\left(\sum_{\substack{\nu'\subset\nu\\|\nu'|=\sigma}}\left|\sum_{\gamma\subset\nu'}f_\gamma\right|^2\right)^{n(n-1)/2} ∫ { ∑ ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∑ γ ⊂ ν ′ ∣ f γ ∣ 2 > C R β s } ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∑ γ ⊂ ν ′ ∑ f γ 2 n ( n − 1 ) /2 ≲ ϵ R β ϵ ( R β s ) n ( n − 1 ) / 2 − 1 ( σ s ) n ( n − 1 ) / 4 ∑ γ ⊂ ν ∥ f γ ∥ 2 2 . (159) \lesssim_\epsilon R^{\beta\epsilon}(R^\beta s)^{n(n-1)/2-1}\left(\frac{\sigma}{s}\right)^{n(n-1)/4}\sum_{\gamma\subset\nu}\|f_\gamma\|_2^2. \tag*{(159)} ≲ ϵ R β ϵ ( R β s ) n ( n − 1 ) /2 − 1 ( s σ ) n ( n − 1 ) /4 γ ⊂ ν ∑ ∥ f γ ∥ 2 2 . (159) The first two items are straightforward, while the third is the key to Proposition 12.1. We implement the iterated broad narrow decomposition of each ∣ ∑ γ ⊂ ν ′ f γ ∣ 2 \left|\sum_{\gamma\subset\nu'}f_\gamma\right|^2 ∑ γ ⊂ ν ′ f γ 2 through the Whitney decomposition. The separated pairs form the broad terms, while E ν ′ E_{\nu'} E ν ′ below is the terminal narrow remainder at scale R − β R^{-\beta} R − β . For each fixed x n x_n x n , a suitable change of variables in the first n − 1 n-1 n − 1 coordinates places the Fourier supports of the broad terms in a small neighborhood of a cone generated by a moment curve. We estimate these terms by decoupling for the cone generated by ϕ n − 2 \phi_{n-2} ϕ n − 2 in R n − 1 \mathbb{R}^{n-1} R n − 1 , as in [8 Lemma 7.1] . Related bilinear changes of variables appear in [4 , 1 , 2 ] and [20 Section 2] . The terminal narrow contribution, summed over ν ′ \nu' ν ′ , is O ( R β s ) O(R^\beta s) O ( R β s ) and can be absorbed on the restricted set in (159) . Combining these estimates proves the third item.
Proof. For item (i), the triangle inequality and ∥ f γ ∥ ∞ ≤ 1 \|f_\gamma\|_\infty\le1 ∥ f γ ∥ ∞ ≤ 1 give ∣ ∑ γ ⊂ ν ′ f γ ∣ ≤ ∑ γ ⊂ ν ′ ∣ f γ ∣ ≲ R β σ \left|\sum_{\gamma\subset\nu'}f_\gamma\right|\le\sum_{\gamma\subset\nu'}|f_\gamma|\lesssim R^{\beta}\sigma ∑ γ ⊂ ν ′ f γ ≤ ∑ γ ⊂ ν ′ ∣ f γ ∣ ≲ R β σ . There are O ( s / σ ) O(s/\sigma) O ( s / σ ) such caps ν ′ ⊂ ν \nu'\subset\nu ν ′ ⊂ ν , so
∑ ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∣ ∑ γ ⊂ ν ′ f γ ∣ 2 ≲ s σ ( R β σ ) 2 = R 2 β σ s . \sum_{\substack{\nu'\subset\nu\\|\nu'|=\sigma}}\left|\sum_{\gamma\subset\nu'}f_\gamma\right|^2\lesssim\frac{s}{\sigma}(R^\beta\sigma)^2=R^{2\beta}\sigma s. ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∑ γ ⊂ ν ′ ∑ f γ 2 ≲ σ s ( R β σ ) 2 = R 2 β σ s . To prove item (ii), let r ν ′ r_{\nu'} r ν ′ be independent random signs taking the values ± 1 \pm1 ± 1 with equal probability. Khintchine’s inequality gives
∫ ( ∑ ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∣ ∑ γ ⊂ ν ′ f γ ∣ 2 ) n ( n − 1 ) / 2 ≲ E ∥ ∑ ν ′ ⊂ ν ∣ ν ′ ∣ = σ r ν ′ ∑ γ ⊂ ν ′ f γ ∥ n ( n − 1 ) n ( n − 1 ) \int\left(\sum_{\substack{\nu'\subset\nu\\|\nu'|=\sigma}}\left|\sum_{\gamma\subset\nu'}f_\gamma\right|^2\right)^{n(n-1)/2}\lesssim\mathbb{E}\left\|\sum_{\substack{\nu'\subset\nu\\|\nu'|=\sigma}}r_{\nu'}\sum_{\gamma\subset\nu'}f_\gamma\right\|_{n(n-1)}^{n(n-1)} ∫ ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∑ γ ⊂ ν ′ ∑ f γ 2 n ( n − 1 ) /2 ≲ E ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∑ r ν ′ γ ⊂ ν ′ ∑ f γ n ( n − 1 ) n ( n − 1 )
We next apply Lemma 12.2 to the first n − 1 n-1 n − 1 coordinates, after rescaling the cap ν \nu ν . Then applying Hölder’s inequality gives
∥ ∑ ν ′ ⊂ ν ∣ ν ′ ∣ = σ r ν ′ ∑ γ ⊂ ν ′ f γ ∥ n ( n − 1 ) ≲ R β ε ( ∑ γ ⊂ ν ∥ f γ ∥ n ( n − 1 ) n ( n − 1 ) ) n ( n − 1 ) / 2 ≲ R β ε ( R β s ) n ( n − 1 ) / 2 − 1 ∑ γ ⊂ ν ∥ f γ ∥ n ( n − 1 ) . \left\| \sum_{\substack{\nu'\subset\nu\\|\nu'|=\sigma}} r_{\nu'}\sum_{\gamma\subset\nu'}f_\gamma\right\|_{n(n-1)}
\lesssim R^{\beta\varepsilon}\left(\sum_{\gamma\subset\nu}\|f_\gamma\|_{n(n-1)}^{n(n-1)}\right)^{n(n-1)/2}
\lesssim R^{\beta\varepsilon}(R^\beta s)^{n(n-1)/2-1}\sum_{\gamma\subset\nu}\|f_\gamma\|_{n(n-1)}. ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∑ r ν ′ γ ⊂ ν ′ ∑ f γ n ( n − 1 ) ≲ R β ε ( γ ⊂ ν ∑ ∥ f γ ∥ n ( n − 1 ) n ( n − 1 ) ) n ( n − 1 ) /2 ≲ R β ε ( R β s ) n ( n − 1 ) /2 − 1 γ ⊂ ν ∑ ∥ f γ ∥ n ( n − 1 ) .
Since ∥ f γ ∥ ∞ ≤ 1 \|f_\gamma\|_\infty\le1 ∥ f γ ∥ ∞ ≤ 1 , this proves (12.6).
To prove item (iii), use a disjoint Whitney decomposition, and let W ω ( ν ′ ) \mathcal{W}_\omega(\nu') W ω ( ν ′ ) denote its pairs of subcaps of ν ′ \nu' ν ′ satisfying ∣ ν 1 ∣ = ∣ ν 2 ∣ = ω |\nu_1|=|\nu_2|=\omega ∣ ν 1 ∣ = ∣ ν 2 ∣ = ω and t ν 1 − t ν 2 ∼ ω t_{\nu_1}-t_{\nu_2}\sim\omega t ν 1 − t ν 2 ∼ ω . For a sufficiently large fixed constant C 0 C_0 C 0 , summing over dyadic ω \omega ω gives
∣ ∑ γ ⊂ ν ′ f γ ∣ 2 = 2 Re ∑ C 0 R − β < ω ≤ σ ∑ ( ν 1 , ν 2 ) ∈ W ω ( ν ′ ) ( ∑ γ ⊂ ν 1 f γ ) ( ∑ γ ⊂ ν 2 f γ ) ‾ + E ν ′ . \left|\sum_{\gamma\subset\nu'}f_\gamma\right|^2
=2\operatorname{Re}\sum_{C_0R^{-\beta}<\omega\le\sigma}
\sum_{(\nu_1,\nu_2)\in\mathcal{W}_\omega(\nu')}
\left(\sum_{\gamma\subset\nu_1}f_\gamma\right)
\overline{\left(\sum_{\gamma\subset\nu_2}f_\gamma\right)}
+E_{\nu'}. γ ⊂ ν ′ ∑ f γ 2 = 2 Re C 0 R − β < ω ≤ σ ∑ ( ν 1 , ν 2 ) ∈ W ω ( ν ′ ) ∑ ( γ ⊂ ν 1 ∑ f γ ) ( γ ⊂ ν 2 ∑ f γ ) + E ν ′ .
Here the remainder is defined by
E ν ′ : = ∑ γ ⊂ ν ′ ∣ f γ ∣ 2 + 2 Re ∑ R − β ≤ ω ≤ C 0 R − β ∑ ( ν 1 , ν 2 ) ∈ W ω ( ν ′ ) ( ∑ γ ⊂ ν 1 f γ ) ( ∑ γ ⊂ ν 2 f γ ) ‾ . E_{\nu'}:=\sum_{\gamma\subset\nu'}|f_\gamma|^2
+2\operatorname{Re}\sum_{R^{-\beta}\le\omega\le C_0R^{-\beta}}
\sum_{(\nu_1,\nu_2)\in\mathcal{W}_\omega(\nu')}
\left(\sum_{\gamma\subset\nu_1}f_\gamma\right)
\overline{\left(\sum_{\gamma\subset\nu_2}f_\gamma\right)}. E ν ′ := γ ⊂ ν ′ ∑ ∣ f γ ∣ 2 + 2 Re R − β ≤ ω ≤ C 0 R − β ∑ ( ν 1 , ν 2 ) ∈ W ω ( ν ′ ) ∑ ( γ ⊂ ν 1 ∑ f γ ) ( γ ⊂ ν 2 ∑ f γ ) .
At the finest scale, W R − β ( ν ′ ) \mathcal{W}_{R^{-\beta}}(\nu') W R − β ( ν ′ ) also includes the remaining pairs of distinct neighbouring small caps, each counted once with t ν 1 > t ν 2 t_{\nu_1}>t_{\nu_2} t ν 1 > t ν 2 . Writing t γ t_\gamma t γ for the left endpoint of the interval indexing γ \gamma γ , each γ \gamma γ has only O ( C 0 ) O(C_0) O ( C 0 ) such neighbours, so
∑ ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∣ E ν ′ ∣ ≲ ∑ γ , γ ′ ⊂ ν ∣ t γ − t γ ′ ∣ ≲ C 0 R − β ∣ f γ ∣ ∣ f γ ′ ∣ ≲ ∑ γ ⊂ ν ∣ f γ ∣ 2 ≲ R β s . \sum_{\substack{\nu'\subset\nu\\|\nu'|=\sigma}}|E_{\nu'}|
\lesssim
\sum_{\substack{\gamma,\gamma'\subset\nu\\|t_\gamma-t_{\gamma'}|\lesssim C_0R^{-\beta}}}
|f_\gamma||f_{\gamma'}|
\lesssim\sum_{\gamma\subset\nu}|f_\gamma|^2
\lesssim R^\beta s. ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∑ ∣ E ν ′ ∣ ≲ γ , γ ′ ⊂ ν ∣ t γ − t γ ′ ∣ ≲ C 0 R − β ∑ ∣ f γ ∣∣ f γ ′ ∣ ≲ γ ⊂ ν ∑ ∣ f γ ∣ 2 ≲ R β s .
Write W ω ( ν ) \mathcal{W}_\omega(\nu) W ω ( ν ) for the union of W ω ( ν ′ ) \mathcal{W}_\omega(\nu') W ω ( ν ′ ) over all ν ′ ⊂ ν \nu'\subset\nu ν ′ ⊂ ν with ∣ ν ′ ∣ = σ |\nu'|=\sigma ∣ ν ′ ∣ = σ . On the set in (12.7), we consequently have
∑ ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∣ ∑ γ ⊂ ν ′ f γ ∣ 2 ≲ ∣ ∑ C 0 R − β < ω ≤ σ ∑ ( ν 1 , ν 2 ) ∈ W ω ( ν ) ( ∑ γ ⊂ ν 1 f γ ) ( ∑ γ ⊂ ν 2 f γ ) ‾ ∣ . \sum_{\substack{\nu'\subset\nu\\|\nu'|=\sigma}}
\left|\sum_{\gamma\subset\nu'}f_\gamma\right|^2
\lesssim
\left|
\sum_{C_0R^{-\beta}<\omega\le\sigma}
\sum_{(\nu_1,\nu_2)\in\mathcal{W}_\omega(\nu)}
\left(\sum_{\gamma\subset\nu_1}f_\gamma\right)
\overline{\left(\sum_{\gamma\subset\nu_2}f_\gamma\right)}
\right|. ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∑ γ ⊂ ν ′ ∑ f γ 2 ≲ C 0 R − β < ω ≤ σ ∑ ( ν 1 , ν 2 ) ∈ W ω ( ν ) ∑ ( γ ⊂ ν 1 ∑ f γ ) ( γ ⊂ ν 2 ∑ f γ ) .
Applying this bound to the left hand side of (12.7), and then using the triangle inequality, we obtain
∫ { ∑ ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∣ ∑ γ ⊂ ν ′ f γ ∣ 2 > C R β s } ( ∑ ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∣ ∑ γ ⊂ ν ′ f γ ∣ 2 ) n ( n − 1 ) / 2 ≲ ( ∑ C 0 R − β < ω ≤ σ ∥ ∑ ( ν 1 , ν 2 ) ∈ W ω ( ν ) ( ∑ γ ⊂ ν 1 f γ ) ( ∑ γ ⊂ ν 2 f γ ) ‾ ∥ n ( n − 1 ) / 2 ) n ( n − 1 ) / 2 . \int_{\left\{\sum_{\substack{\nu'\subset\nu\\|\nu'|=\sigma}}\left|\sum_{\gamma\subset\nu'}f_\gamma\right|^2>CR^{\beta}s\right\}}
\left(\sum_{\substack{\nu'\subset\nu\\|\nu'|=\sigma}}
\left|\sum_{\gamma\subset\nu'}f_\gamma\right|^2\right)^{n(n-1)/2}
\lesssim
\left(
\sum_{C_0R^{-\beta}<\omega\le\sigma}
\left\|
\sum_{(\nu_1,\nu_2)\in\mathcal{W}_\omega(\nu)}
\left(\sum_{\gamma\subset\nu_1}f_\gamma\right)
\overline{\left(\sum_{\gamma\subset\nu_2}f_\gamma\right)}
\right\|_{n(n-1)/2}
\right)^{n(n-1)/2}. ∫ { ∑ ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∣ ∑ γ ⊂ ν ′ f γ ∣ 2 > C R β s } ν ′ ⊂ ν ∣ ν ′ ∣ = σ ∑ γ ⊂ ν ′ ∑ f γ 2 n ( n − 1 ) /2 ≲ C 0 R − β < ω ≤ σ ∑ ( ν 1 , ν 2 ) ∈ W ω ( ν ) ∑ ( γ ⊂ ν 1 ∑ f γ ) ( γ ⊂ ν 2 ∑ f γ ) n ( n − 1 ) /2 n ( n − 1 ) /2 .
We now estimate each summand. At each ω \omega ω , we may split these pairs into finitely many collections in which the points ( t ν 1 + t ν 2 ) / 2 (t_{\nu_1}+t_{\nu_2})/2 ( t ν 1 + t ν 2 ) /2 are ω \omega ω separated. Each cap occurs in boundedly many pairs.
We check the Fourier support needed to apply decoupling for the cone generated by ϕ n − 2 \phi_{n-2} ϕ n − 2 in R n − 1 \mathbb{R}^{n-1} R n − 1 . If t , t ′ t,t' t , t ′ are the left endpoints of the intervals indexing small caps grouped in ν 1 , ν 2 \nu_1,\nu_2 ν 1 , ν 2 , respectively, and t 0 = ( t + t ′ ) / 2 t_0=(t+t')/2 t 0 = ( t + t ′ ) /2 , then
γ n − 1 ( t ) − γ n − 1 ( t ′ ) = ( t − t ′ ) γ n − 1 ′ ( t 0 ) + ∑ k ≥ 1 ( t − t ′ ) 2 k + 1 2 2 k ( 2 k + 1 ) ! γ n − 1 ( 2 k + 1 ) ( t 0 ) . (160) \gamma_{n-1}(t)-\gamma_{n-1}(t')
=(t-t')\gamma_{n-1}'(t_0)
+\sum_{k\ge1}\frac{(t-t')^{2k+1}}{2^{2k}(2k+1)!}\gamma_{n-1}^{(2k+1)}(t_0). \tag*{(160)} γ n − 1 ( t ) − γ n − 1 ( t ′ ) = ( t − t ′ ) γ n − 1 ′ ( t 0 ) + k ≥ 1 ∑ 2 2 k ( 2 k + 1 )! ( t − t ′ ) 2 k + 1 γ n − 1 ( 2 k + 1 ) ( t 0 ) . (160)
Since t − t ′ ∼ ω t-t'\sim\omega t − t ′ ∼ ω , division by ω \omega ω gives a radial coefficient comparable to ∼ 1 \sim1 ∼ 1 . In the ( ℓ + 1 ) (\ell+1) ( ℓ + 1 ) st derivative coordinate, the errors from the widths of the two small caps satisfy
ω − 1 ∑ j = 1 ℓ + 1 R − j β ( C ω ) ℓ + 1 − j ≲ ω ℓ , 0 ≤ ℓ ≤ n − 2. \omega^{-1}\sum_{j=1}^{\ell+1}R^{-j\beta}(C\omega)^{\ell+1-j}\lesssim\omega^\ell,\qquad0\le\ell\le n-2. ω − 1 j = 1 ∑ ℓ + 1 R − j β ( C ω ) ℓ + 1 − j ≲ ω ℓ , 0 ≤ ℓ ≤ n − 2.
In the radial coordinate this error is O ( ( R β ω ) − 1 ) O((R^{\beta}\omega)^{-1}) O (( R β ω ) − 1 ) , so our choice of C 0 C_0 C 0 keeps the radial coefficient comparable to ∼ 1 \sim1 ∼ 1 . Recentering at ( t ν 1 + t ν 2 ) / 2 (t_{\nu_1}+t_{\nu_2})/2 ( t ν 1 + t ν 2 ) /2 preserves the widths 1 , ω , … , ω n − 2 1,\omega,\ldots,\omega^{n-2} 1 , ω , … , ω n − 2 , and note that
γ n − 1 ′ ( t ) = diag ( 1 ! , 2 ! , … , ( n − 1 ) ! ) ϕ n − 2 ( t ) . \gamma'_{n-1}(t)=\operatorname{diag}(1!,2!,\ldots,(n-1)!)\phi_{n-2}(t). γ n − 1 ′ ( t ) = diag ( 1 ! , 2 ! , … , ( n − 1 )!) ϕ n − 2 ( t ) .
Thus, after this fixed diagonal change of coordinates and division by ω \omega ω , the projected Fourier supports of the products
( ∑ γ ⊂ ν 1 f γ ) ( ∑ γ ⊂ ν 2 f γ ) ‾ \left(\sum_{\gamma\subset\nu_1}f_\gamma\right)\overline{\left(\sum_{\gamma\subset\nu_2}f_\gamma\right)} ( γ ⊂ ν 1 ∑ f γ ) ( γ ⊂ ν 2 ∑ f γ )
lie in caps of aperture comparable to ω \omega ω in R n − 1 \mathbb{R}^{n-1} R n − 1 .
The exponent n ( n − 1 ) / 2 n(n-1)/2 n ( n − 1 ) /2 is admissible in (155) , since n ( n − 1 ) / 2 ≤ ( n − 2 ) ( n − 1 ) n(n-1)/2\le(n-2)(n-1) n ( n − 1 ) /2 ≤ ( n − 2 ) ( n − 1 ) for n ≥ 4 n\ge4 n ≥ 4 . Applying the cylindrical version of the decoupling estimate for the cone generated by ϕ n − 2 \phi_{n-2} ϕ n − 2 in R n − 1 \mathbb{R}^{n-1} R n − 1 (see (155) ) and then Hölder’s inequality over the O ( s / ω ) O(s/\omega) O ( s / ω ) pairs ( ν 1 , ν 2 ) ∈ W ω ( ν ) (\nu_1,\nu_2)\in\mathcal{W}_{\omega}(\nu) ( ν 1 , ν 2 ) ∈ W ω ( ν ) gives
∥ ∑ ( ν 1 , ν 2 ) ∈ W ω ( ν ) ( ∑ γ ⊂ ν 1 f γ ) ( ∑ γ ⊂ ν 2 f γ ) ‾ ∥ n ( n − 1 ) / 2 n ( n − 1 ) / 2 \left\|\sum_{(\nu_1,\nu_2)\in\mathcal{W}_{\omega}(\nu)}\left(\sum_{\gamma\subset\nu_1}f_\gamma\right)\overline{\left(\sum_{\gamma\subset\nu_2}f_\gamma\right)}\right\|_{n(n-1)/2}^{n(n-1)/2} ( ν 1 , ν 2 ) ∈ W ω ( ν ) ∑ ( γ ⊂ ν 1 ∑ f γ ) ( γ ⊂ ν 2 ∑ f γ ) n ( n − 1 ) /2 n ( n − 1 ) /2
≲ ϵ R β ϵ ( s / ω ) n ( n − 1 ) / 4 − 1 ∑ ( ν 1 , ν 2 ) ∈ W ω ( ν ) ∥ ( ∑ γ ⊂ ν 1 f γ ) ( ∑ γ ⊂ ν 2 f γ ) ‾ ∥ n ( n − 1 ) / 2 n ( n − 1 ) / 2 . \lesssim_{\epsilon}R^{\beta\epsilon}(s/\omega)^{n(n-1)/4-1}\sum_{(\nu_1,\nu_2)\in\mathcal{W}_{\omega}(\nu)}\left\|\left(\sum_{\gamma\subset\nu_1}f_\gamma\right)\overline{\left(\sum_{\gamma\subset\nu_2}f_\gamma\right)}\right\|_{n(n-1)/2}^{n(n-1)/2}. ≲ ϵ R β ϵ ( s / ω ) n ( n − 1 ) /4 − 1 ( ν 1 , ν 2 ) ∈ W ω ( ν ) ∑ ( γ ⊂ ν 1 ∑ f γ ) ( γ ⊂ ν 2 ∑ f γ ) n ( n − 1 ) /2 n ( n − 1 ) /2 .
The cylindrical version of the decoupling estimate for the moment curve γ n − 1 \gamma_{n-1} γ n − 1 (see equation 12.2), now applied at exponent n ( n − 1 ) n(n-1) n ( n − 1 ) on each cap of parameter length ω \omega ω , also gives
∥ ( ∑ γ ⊂ ν 1 f γ ) ( ∑ γ ⊂ ν 2 f γ ) ‾ ∥ n ( n − 1 ) / 2 n ( n − 1 ) / 2 ≲ ϵ R β ϵ ( R β ω ) n ( n − 1 ) / 2 − 1 ∑ γ ⊂ ν 1 ∪ ν 2 ∥ f γ ∥ 2 2 . \left\|\left(\sum_{\gamma\subset\nu_1}f_\gamma\right)\overline{\left(\sum_{\gamma\subset\nu_2}f_\gamma\right)}\right\|_{n(n-1)/2}^{n(n-1)/2}\lesssim_{\epsilon}R^{\beta\epsilon}(R^{\beta}\omega)^{n(n-1)/2-1}\sum_{\gamma\subset\nu_1\cup\nu_2}\|f_\gamma\|_2^2. ( γ ⊂ ν 1 ∑ f γ ) ( γ ⊂ ν 2 ∑ f γ ) n ( n − 1 ) /2 n ( n − 1 ) /2 ≲ ϵ R β ϵ ( R β ω ) n ( n − 1 ) /2 − 1 γ ⊂ ν 1 ∪ ν 2 ∑ ∥ f γ ∥ 2 2 .
Each small cap occurs in boundedly many pairs. Combining the last two estimates, taking roots, and summing in ω \omega ω gives
∑ C 0 R − β < ω ≤ σ ∥ ∑ ( ν 1 , ν 2 ) ∈ W ω ( ν ) ( ∑ γ ⊂ ν 1 f γ ) ( ∑ γ ⊂ ν 2 f γ ) ‾ ∥ n ( n − 1 ) / 2 \sum_{C_0R^{-\beta}<\omega\le\sigma}\left\|\sum_{(\nu_1,\nu_2)\in\mathcal{W}_{\omega}(\nu)}\left(\sum_{\gamma\subset\nu_1}f_\gamma\right)\overline{\left(\sum_{\gamma\subset\nu_2}f_\gamma\right)}\right\|_{n(n-1)/2} C 0 R − β < ω ≤ σ ∑ ( ν 1 , ν 2 ) ∈ W ω ( ν ) ∑ ( γ ⊂ ν 1 ∑ f γ ) ( γ ⊂ ν 2 ∑ f γ ) n ( n − 1 ) /2
≲ ϵ R β ( n ( n − 1 ) − 2 + 2 ϵ ) n ( n − 1 ) s 1 2 − 2 n ( n − 1 ) σ 1 / 2 ( ∑ γ ⊂ ν ∥ f γ ∥ 2 2 ) 2 n ( n − 1 ) . \lesssim_{\epsilon}R^{\frac{\beta(n(n-1)-2+2\epsilon)}{n(n-1)}}s^{\frac12-\frac{2}{n(n-1)}}\sigma^{1/2}\left(\sum_{\gamma\subset\nu}\|f_\gamma\|_2^2\right)^{\frac{2}{n(n-1)}}. ≲ ϵ R n ( n − 1 ) β ( n ( n − 1 ) − 2 + 2 ϵ ) s 2 1 − n ( n − 1 ) 2 σ 1/2 ( γ ⊂ ν ∑ ∥ f γ ∥ 2 2 ) n ( n − 1 ) 2 .
where we used ∑ ω ≤ σ ω 1 / 2 ≲ σ 1 / 2 \sum_{\omega\le\sigma}\omega^{1/2}\lesssim\sigma^{1/2} ∑ ω ≤ σ ω 1/2 ≲ σ 1/2 . By the reduction above, raising this bound to the power n ( n − 1 ) / 2 n(n-1)/2 n ( n − 1 ) /2 proves (159) . □
Let us now prove Proposition 12.1. Consider two cases:
n ≥ 5 , β ≥ 1 / n , or n = 4 , 1 / 4 ≤ β ≤ 1 , or n = 4 , β > 1. (161) n\ge5,\ \beta\ge1/n,\quad\text{or}\quad n=4,\ 1/4\le\beta\le1,\quad\text{or}\quad n=4,\ \beta>1. \tag*{(161)} n ≥ 5 , β ≥ 1/ n , or n = 4 , 1/4 ≤ β ≤ 1 , or n = 4 , β > 1. (161)
Case 1: n ≥ 5 n\ge5 n ≥ 5 , β ≥ 1 / n \beta\ge1/n β ≥ 1/ n , or n = 4 n=4 n = 4 , $1/4\le\beta\le1
We prove Proposition 12.1 for such n , β n,\beta n , β . Fix the final loss ϵ > 0 \epsilon>0 ϵ > 0 . In the estimates below we use an auxiliary parameter δ = δ ( ϵ ) > 0 \delta=\delta(\epsilon)>0 δ = δ ( ϵ ) > 0 , to be chosen sufficiently small at the end. Group the small caps into θ ∈ Θ ( R − 1 / n ) \theta\in\Theta(R^{-1/n}) θ ∈ Θ ( R − 1/ n ) , writing f θ = ∑ γ ⊂ θ f γ f_\theta=\sum_{\gamma\subset\theta}f_\gamma f θ = ∑ γ ⊂ θ f γ . Applying Lemma 12.3 with δ / n \delta/n δ / n , and using β ≥ 1 / n \beta\ge1/n β ≥ 1/ n ,
α p ∣ { ∣ ∑ γ f γ ∣ > α } ∣ ≲ δ R β δ α p − p n ∑ R − 1 / n ≤ s ≤ 1 ∑ τ ∈ Θ ( s ) ∫ { ∑ θ ⊂ τ ∣ f θ ∣ 2 > C δ − 1 R − β δ α 2 s 2 } ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p n / 2 . (162) \alpha^p\left|\left\{\left|\sum_\gamma f_\gamma\right|>\alpha\right\}\right|\lesssim_\delta R^{\beta\delta}\alpha^{p-p_n}\sum_{R^{-1/n}\le s\le1}\sum_{\tau\in\Theta(s)}\int_{\left\{\sum_{\theta\subset\tau}|f_\theta|^2>C_\delta^{-1}R^{-\beta\delta}\alpha^2s^2\right\}}\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{p_n/2}. \tag*{(162)} α p { γ ∑ f γ > α } ≲ δ R β δ α p − p n R − 1/ n ≤ s ≤ 1 ∑ τ ∈ Θ ( s ) ∑ ∫ { ∑ θ ⊂ τ ∣ f θ ∣ 2 > C δ − 1 R − β δ α 2 s 2 } ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p n /2 . (162)
Since p n / 2 < n ( n − 1 ) / 2 p_n/2<n(n-1)/2 p n /2 < n ( n − 1 ) /2 , on the indicated set
( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p n / 2 ≲ δ ( R − β δ α 2 s 2 ) p n / 2 − n ( n − 1 ) / 2 ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) n ( n − 1 ) / 2 . \left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{p_n/2}\lesssim_\delta\left(R^{-\beta\delta}\alpha^2s^2\right)^{p_n/2-n(n-1)/2}\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{n(n-1)/2}. ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p n /2 ≲ δ ( R − β δ α 2 s 2 ) p n /2 − n ( n − 1 ) /2 ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) n ( n − 1 ) /2 .
Substituting this bound into (162) , we obtain
α p ∣ { ∣ ∑ γ f γ ∣ > α } ∣ ≲ δ R C β δ α 2 / β ∑ R − 1 / n ≤ s ≤ 1 s n + 1 − n ( n − 1 ) / 2 \alpha^p\left|\left\{\left|\sum_\gamma f_\gamma\right|>\alpha\right\}\right|\lesssim_\delta R^{C\beta\delta}\alpha^{2/\beta}\sum_{R^{-1/n}\le s\le1}s^{n+1-n(n-1)/2} α p { γ ∑ f γ > α } ≲ δ R C β δ α 2/ β R − 1/ n ≤ s ≤ 1 ∑ s n + 1 − n ( n − 1 ) /2
× ∑ τ ∈ Θ ( s ) ∫ { ∑ θ ⊂ τ ∣ f θ ∣ 2 > C δ − 1 R − β δ α 2 s 2 } ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) n ( n − 1 ) / 2 . {}\times\sum_{\tau\in\Theta(s)}\int_{\left\{\sum_{\theta\subset\tau}|f_\theta|^2>C_\delta^{-1}R^{-\beta\delta}\alpha^2s^2\right\}}\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{n(n-1)/2}. × τ ∈ Θ ( s ) ∑ ∫ { ∑ θ ⊂ τ ∣ f θ ∣ 2 > C δ − 1 R − β δ α 2 s 2 } ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) n ( n − 1 ) /2 .
We estimate the integral using Lemma 12.4 with ν = τ \nu=\tau ν = τ , ν ′ = θ \nu'=\theta ν ′ = θ , and σ = R − 1 / n \sigma=R^{-1/n} σ = R − 1/ n , so that ∑ γ ⊂ ν ′ f γ = f θ \sum_{\gamma\subset\nu'}f_\gamma=f_\theta ∑ γ ⊂ ν ′ f γ = f θ , distinguishing two ranges of s s s . Consider two ranges
s ≤ C R β ( 1 + δ ) α − 2 , s > C R β ( 1 + δ ) α − 2 . (163) s\le CR^{\beta(1+\delta)}\alpha^{-2},\qquad s>CR^{\beta(1+\delta)}\alpha^{-2}. \tag*{(163)} s ≤ C R β ( 1 + δ ) α − 2 , s > C R β ( 1 + δ ) α − 2 . (163)
This split compares the threshold C δ − 1 R − β δ α 2 s 2 C_\delta^{-1}R^{-\beta\delta}\alpha^2s^2 C δ − 1 R − β δ α 2 s 2 in (162) with the threshold C R β s CR^\beta s C R β s in (159) . We use (158) in the first range and (159) in the second. We split the entire right hand side of the preceding level set estimate into the two terms
I : = R C β δ α 2 / β ∑ R − 1 / n ≤ s ≤ 1 s ≤ C R β ( 1 + δ ) α − 2 s n + 1 − n ( n − 1 ) / 2 ∑ τ ∈ Θ ( s ) ∫ { ∑ θ ⊂ τ ∣ f θ ∣ 2 > C δ − 1 R − β δ α 2 s 2 } ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) n ( n − 1 ) / 2 , \mathrm{I}:=R^{C\beta\delta}\alpha^{2/\beta}
\sum_{\substack{R^{-1/n}\le s\le1\\s\le CR^{\beta(1+\delta)}\alpha^{-2}}}
s^{n+1-n(n-1)/2}
\sum_{\tau\in\Theta(s)}
\int_{\{\sum_{\theta\subset\tau}|f_\theta|^2>C_\delta^{-1}R^{-\beta\delta}\alpha^2s^2\}}
\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{n(n-1)/2}, I := R C β δ α 2/ β R − 1/ n ≤ s ≤ 1 s ≤ C R β ( 1 + δ ) α − 2 ∑ s n + 1 − n ( n − 1 ) /2 τ ∈ Θ ( s ) ∑ ∫ { ∑ θ ⊂ τ ∣ f θ ∣ 2 > C δ − 1 R − β δ α 2 s 2 } ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) n ( n − 1 ) /2 ,
I I : = R C β δ α 2 / β ∑ R − 1 / n ≤ s ≤ 1 s > C R β ( 1 + δ ) α − 2 s n + 1 − n ( n − 1 ) / 2 ∑ τ ∈ Θ ( s ) ∫ { ∑ θ ⊂ τ ∣ f θ ∣ 2 > C δ − 1 R − β δ α 2 s 2 } ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) n ( n − 1 ) / 2 . \mathrm{II}:=R^{C\beta\delta}\alpha^{2/\beta}
\sum_{\substack{R^{-1/n}\le s\le1\\s>CR^{\beta(1+\delta)}\alpha^{-2}}}
s^{n+1-n(n-1)/2}
\sum_{\tau\in\Theta(s)}
\int_{\{\sum_{\theta\subset\tau}|f_\theta|^2>C_\delta^{-1}R^{-\beta\delta}\alpha^2s^2\}}
\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{n(n-1)/2}. II := R C β δ α 2/ β R − 1/ n ≤ s ≤ 1 s > C R β ( 1 + δ ) α − 2 ∑ s n + 1 − n ( n − 1 ) /2 τ ∈ Θ ( s ) ∑ ∫ { ∑ θ ⊂ τ ∣ f θ ∣ 2 > C δ − 1 R − β δ α 2 s 2 } ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) n ( n − 1 ) /2 .
It suffices to prove the following two bounds, with δ \delta δ chosen sufficiently small in terms of ϵ \epsilon ϵ :
I + I I ≲ ϵ R 1 + β ( n ( n − 1 ) / 2 − 1 ) + β ϵ ∑ γ ∥ f γ ∥ 2 2 . \mathrm{I}+\mathrm{II}\lesssim_\epsilon R^{1+\beta(n(n-1)/2-1)+\beta\epsilon}\sum_\gamma\|f_\gamma\|_2^2. I + II ≲ ϵ R 1 + β ( n ( n − 1 ) /2 − 1 ) + β ϵ γ ∑ ∥ f γ ∥ 2 2 .
Indeed, the preceding level set estimate gives
α p ∣ { ∣ ∑ γ f γ ∣ > α } ∣ ≲ δ I + I I . \alpha^p\left|\left\{\left|\sum_\gamma f_\gamma\right|>\alpha\right\}\right|\lesssim_\delta\mathrm{I}+\mathrm{II}. α p { γ ∑ f γ > α } ≲ δ I + II .
Assuming the two bounds above, adding them therefore yields
α p ∣ { ∣ ∑ γ f γ ∣ > α } ∣ ≲ ϵ R 1 + β ( n ( n − 1 ) / 2 − 1 ) + β ϵ ∑ γ ∥ f γ ∥ 2 2 , \alpha^p\left|\left\{\left|\sum_\gamma f_\gamma\right|>\alpha\right\}\right|
\lesssim_\epsilon R^{1+\beta(n(n-1)/2-1)+\beta\epsilon}\sum_\gamma\|f_\gamma\|_2^2, α p { γ ∑ f γ > α } ≲ ϵ R 1 + β ( n ( n − 1 ) /2 − 1 ) + β ϵ γ ∑ ∥ f γ ∥ 2 2 ,
which is (153) . We now estimate I and II separately.
Estimate for I
Fix a dyadic scale s ≤ C R β ( 1 + δ ) α − 2 s\le CR^{\beta(1+\delta)}\alpha^{-2} s ≤ C R β ( 1 + δ ) α − 2 in its defining sum. For every τ ∈ Θ ( s ) \tau\in\Theta(s) τ ∈ Θ ( s ) , enlarge the integration region to R n \mathbb{R}^n R n and apply (158) with ν = τ \nu=\tau ν = τ , ν ′ = θ \nu'=\theta ν ′ = θ , and σ = R − 1 / n \sigma=R^{-1/n} σ = R − 1/ n to obtain
∫ R n ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) n ( n − 1 ) / 2 ≲ δ R β δ ( R β s ) n ( n − 1 ) / 2 − 1 ∑ γ ⊂ τ ∥ f γ ∥ 2 2 . \int_{\mathbb{R}^n}
\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{n(n-1)/2}
\lesssim_\delta R^{\beta\delta}(R^\beta s)^{n(n-1)/2-1}\sum_{\gamma\subset\tau}\|f_\gamma\|_2^2. ∫ R n ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) n ( n − 1 ) /2 ≲ δ R β δ ( R β s ) n ( n − 1 ) /2 − 1 γ ⊂ τ ∑ ∥ f γ ∥ 2 2 .
Substituting this estimate directly into the definition of I, we obtain
I ≲ δ R β ( n ( n − 1 ) / 2 − 1 ) + C β δ ∑ R − 1 / n ≤ s ≤ 1 s ≤ C R β ( 1 + δ ) α − 2 α 2 / β s n ∑ γ ∥ f γ ∥ 2 2 . \mathrm{I}\lesssim_\delta R^{\beta(n(n-1)/2-1)+C\beta\delta}
\sum_{\substack{R^{-1/n}\le s\le1\\s\le CR^{\beta(1+\delta)}\alpha^{-2}}}
\alpha^{2/\beta}s^n\sum_\gamma\|f_\gamma\|_2^2. I ≲ δ R β ( n ( n − 1 ) /2 − 1 ) + C β δ R − 1/ n ≤ s ≤ 1 s ≤ C R β ( 1 + δ ) α − 2 ∑ α 2/ β s n γ ∑ ∥ f γ ∥ 2 2 .
For every scale in this sum, the condition s ≤ C R β ( 1 + δ ) α − 2 s\le CR^{\beta(1+\delta)}\alpha^{-2} s ≤ C R β ( 1 + δ ) α − 2 gives α 2 s ≲ R β ( 1 + δ ) \alpha^2s\lesssim R^{\beta(1+\delta)} α 2 s ≲ R β ( 1 + δ ) . Hence
α 2 / β s n = ( α 2 s ) 1 / β s n − 1 / β ≲ R 1 + δ s n − 1 / β ≤ R 1 + δ , \alpha^{2/\beta}s^n=(\alpha^2s)^{1/\beta}s^{n-1/\beta}\lesssim R^{1+\delta}s^{n-1/\beta}\le R^{1+\delta}, α 2/ β s n = ( α 2 s ) 1/ β s n − 1/ β ≲ R 1 + δ s n − 1/ β ≤ R 1 + δ ,
where the last inequality uses s ≤ 1 s\le1 s ≤ 1 and β ≥ 1 / n \beta\ge1/n β ≥ 1/ n . Substituting this bound into the preceding estimate for I and summing over the O ( log ( 2 R ) ) O(\log(2R)) O ( log ( 2 R )) dyadic scales yields
I ≲ δ R 1 + β ( n ( n − 1 ) / 2 − 1 ) + C β δ ∑ R − 1 / n ≤ s ≤ 1 s ≤ C R β ( 1 + δ ) α − 2 ∑ γ ∥ f γ ∥ 2 2 \mathrm{I}\lesssim_\delta R^{1+\beta(n(n-1)/2-1)+C\beta\delta}
\sum_{\substack{R^{-1/n}\le s\le1\\s\le CR^{\beta(1+\delta)}\alpha^{-2}}}
\sum_\gamma\|f_\gamma\|_2^2 I ≲ δ R 1 + β ( n ( n − 1 ) /2 − 1 ) + C β δ R − 1/ n ≤ s ≤ 1 s ≤ C R β ( 1 + δ ) α − 2 ∑ γ ∑ ∥ f γ ∥ 2 2
≲ ϵ R 1 + β ( n ( n − 1 ) / 2 − 1 ) + β ϵ ∑ γ ∥ f γ ∥ 2 2 . \lesssim_\epsilon R^{1+\beta(n(n-1)/2-1)+\beta\epsilon}
\sum_\gamma\|f_\gamma\|_2^2. ≲ ϵ R 1 + β ( n ( n − 1 ) /2 − 1 ) + β ϵ γ ∑ ∥ f γ ∥ 2 2 .
This proves the desired bound for I.
Estimate for II
Fix a dyadic scale s > C R β ( 1 + δ ) α − 2 s > CR^{\beta(1+\delta)}\alpha^{-2} s > C R β ( 1 + δ ) α − 2 in its defining sum. With the constant in this range condition chosen sufficiently large compared with the threshold constant in (159) , we have
{ ∑ θ ⊂ τ ∣ f θ ∣ 2 > C δ − 1 R − β δ α 2 s 2 } ⊂ { ∑ θ ⊂ τ ∣ f θ ∣ 2 > C R β s } . \left\{\sum_{\theta\subset\tau}|f_\theta|^2>C_\delta^{-1}R^{-\beta\delta}\alpha^2s^2\right\}\subset\left\{\sum_{\theta\subset\tau}|f_\theta|^2>CR^\beta s\right\}. { θ ⊂ τ ∑ ∣ f θ ∣ 2 > C δ − 1 R − β δ α 2 s 2 } ⊂ { θ ⊂ τ ∑ ∣ f θ ∣ 2 > C R β s } .
For every τ ∈ Θ ( s ) \tau\in\Theta(s) τ ∈ Θ ( s ) , we may therefore apply (159) with ν = τ \nu=\tau ν = τ , ν ′ = θ \nu'=\theta ν ′ = θ , and σ = R − 1 / n \sigma=R^{-1/n} σ = R − 1/ n to obtain
∫ { ∑ θ ⊂ τ ∣ f θ ∣ 2 > C δ − 1 R − β δ α 2 s 2 } ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) n ( n − 1 ) / 2 ≲ R β ( n ( n − 1 ) / 2 − 1 ) + β δ R − ( n − 1 ) / 4 s n ( n − 1 ) / 4 − 1 ∑ γ ⊂ τ ∥ f γ ∥ 2 2 . \int_{\{\sum_{\theta\subset\tau}|f_\theta|^2>C_\delta^{-1}R^{-\beta\delta}\alpha^2s^2\}}\left(\sum_{\theta\subset\tau}|f_\theta|^2\right)^{n(n-1)/2}\lesssim R^{\beta(n(n-1)/2-1)+\beta\delta}R^{-(n-1)/4}s^{n(n-1)/4-1}\sum_{\gamma\subset\tau}\|f_\gamma\|_2^2. ∫ { ∑ θ ⊂ τ ∣ f θ ∣ 2 > C δ − 1 R − β δ α 2 s 2 } ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) n ( n − 1 ) /2 ≲ R β ( n ( n − 1 ) /2 − 1 ) + β δ R − ( n − 1 ) /4 s n ( n − 1 ) /4 − 1 γ ⊂ τ ∑ ∥ f γ ∥ 2 2 .
Discard scales for which all integration regions in II are empty; the sums below run only over the remaining scales. Substituting this estimate directly into the definition of II, we obtain
I I ≲ δ R β ( n ( n − 1 ) / 2 − 1 ) + C β δ ∑ R − 1 / n ≤ s ≤ 1 s > C R β ( 1 + δ ) α − 2 α 2 / β R − ( n − 1 ) / 4 s n − n ( n − 1 ) / 4 ∑ γ ∥ f γ ∥ 2 2 . \mathrm{II}\lesssim_\delta R^{\beta(n(n-1)/2-1)+C\beta\delta}\sum_{\substack{R^{-1/n}\le s\le1\\s>CR^{\beta(1+\delta)}\alpha^{-2}}}\alpha^{2/\beta}R^{-(n-1)/4}s^{n-n(n-1)/4}\sum_\gamma\|f_\gamma\|_2^2. II ≲ δ R β ( n ( n − 1 ) /2 − 1 ) + C β δ R − 1/ n ≤ s ≤ 1 s > C R β ( 1 + δ ) α − 2 ∑ α 2/ β R − ( n − 1 ) /4 s n − n ( n − 1 ) /4 γ ∑ ∥ f γ ∥ 2 2 .
The inequality (157) gives
C δ − 1 R − β δ α 2 s 2 ≲ ∑ θ ⊂ τ ∣ f θ ∣ 2 ≲ R 2 β R − 1 / n s . C_\delta^{-1}R^{-\beta\delta}\alpha^2s^2\lesssim\sum_{\theta\subset\tau}|f_\theta|^2\lesssim R^{2\beta}R^{-1/n}s. C δ − 1 R − β δ α 2 s 2 ≲ θ ⊂ τ ∑ ∣ f θ ∣ 2 ≲ R 2 β R − 1/ n s .
It follows that
α 2 / β ≲ δ R 2 + δ ( R − 1 / n s ) 1 / β . \alpha^{2/\beta}\lesssim_\delta R^{2+\delta}\left(\frac{R^{-1/n}}{s}\right)^{1/\beta}. α 2/ β ≲ δ R 2 + δ ( s R − 1/ n ) 1/ β .
Consequently, the remaining factor satisfies
α 2 / β R − ( n − 1 ) / 4 s n − n ( n − 1 ) / 4 ≲ R 1 + δ ( R − 1 / n s ) 1 / β + n ( n − 5 ) / 4 ≤ R 1 + δ . \alpha^{2/\beta}R^{-(n-1)/4}s^{n-n(n-1)/4}\lesssim R^{1+\delta}\left(\frac{R^{-1/n}}{s}\right)^{1/\beta+n(n-5)/4}\le R^{1+\delta}. α 2/ β R − ( n − 1 ) /4 s n − n ( n − 1 ) /4 ≲ R 1 + δ ( s R − 1/ n ) 1/ β + n ( n − 5 ) /4 ≤ R 1 + δ .
For the last inequality, R − 1 / n / s ≤ 1 R^{-1/n}/s\le1 R − 1/ n / s ≤ 1 , and the exponent is nonnegative: this is immediate for n ≥ 5 n\ge5 n ≥ 5 , while for n = 4 n=4 n = 4 it is 1 / β − 1 ≥ 0 1/\beta-1\ge0 1/ β − 1 ≥ 0 because β ≤ 1 \beta\le1 β ≤ 1 . Summing over the O ( log ( 2 R ) ) O(\log(2R)) O ( log ( 2 R )) dyadic scales gives the required bound
I I ≲ ϵ R 1 + β ( n ( n − 1 ) / 2 − 1 ) + β ϵ ∑ γ ∥ f γ ∥ 2 2 . \mathrm{II}\lesssim_\epsilon R^{1+\beta(n(n-1)/2-1)+\beta\epsilon}\sum_\gamma\|f_\gamma\|_2^2. II ≲ ϵ R 1 + β ( n ( n − 1 ) /2 − 1 ) + β ϵ γ ∑ ∥ f γ ∥ 2 2 .
Here we choose δ > 0 \delta>0 δ > 0 sufficiently small that C δ ≤ ϵ / 2 C\delta\le\epsilon/2 C δ ≤ ϵ /2 and use log ( 2 R ) ≲ ϵ R β ϵ / 2 \log(2R)\lesssim_\epsilon R^{\beta\epsilon/2} log ( 2 R ) ≲ ϵ R β ϵ /2 .
Case 2: n = 4 n=4 n = 4 , β > 1 \beta>1 β > 1
It remains to prove Proposition 12.1 for n = 4 n=4 n = 4 and β > 1 \beta>1 β > 1 . We first establish the following broad estimate, then pass to the general case by a broad narrow analysis.
Proposition 12.5 (Broad estimate). Let n = 4 n=4 n = 4 , β ≥ 1 \beta\ge1 β ≥ 1 , and let K K K be a sufficiently large number. Let ν 1 , … , ν 4 \nu_1,\ldots,\nu_4 ν 1 , … , ν 4 be caps with ∣ ν i ∣ = K − 1 |\nu_i|=K^{-1} ∣ ν i ∣ = K − 1 such that dist ( ν i , ν j ) > K − 1 \operatorname{dist}(\nu_i,\nu_j)>K^{-1} dist ( ν i , ν j ) > K − 1 for any i ≠ j i\ne j i = j . Suppose ∥ f γ ∥ ∞ ≤ 1 \|f_\gamma\|_\infty\le1 ∥ f γ ∥ ∞ ≤ 1 for every small cap γ \gamma γ . For α > 0 \alpha>0 α > 0 , define the broad set
U α = { x : min 1 ≤ i ≤ 4 ∣ ∑ γ ⊂ ν i f γ ( x ) ∣ > α } . U_\alpha=\left\{x:\min_{1\le i\le4}\left|\sum_{\gamma\subset\nu_i}f_\gamma(x)\right|>\alpha\right\}. U α = { x : 1 ≤ i ≤ 4 min γ ⊂ ν i ∑ f γ ( x ) > α } . Then, for every ϵ > 0 \epsilon>0 ϵ > 0 ,
α 12 + 2 / β ∣ U α ∣ ≲ ϵ , K R 5 β + 1 + β ϵ ∑ γ ∥ f γ ∥ 2 2 . \alpha^{12+2/\beta}|U_\alpha|\lesssim_{\epsilon,K}R^{5\beta+1+\beta\epsilon}\sum_\gamma\|f_\gamma\|_2^2. α 12 + 2/ β ∣ U α ∣ ≲ ϵ , K R 5 β + 1 + β ϵ γ ∑ ∥ f γ ∥ 2 2 .
Assuming Proposition 12.5, the remaining case of Proposition 12.1 follows by a standard broad narrow analysis. For completeness, we give this deduction at the end of this subsection. We prepare the proof of Proposition 12.5, keeping the notation in its statement.
Let B s − 1 , R B_{s^{-1},R} B s − 1 , R be the box centered at the origin, with sides parallel to the coordinate axes and dimensions
s − 1 × s − 1 × s − 1 × min { s − 1 , R } . s^{-1}\times s^{-1}\times s^{-1}\times\min\{s^{-1},R\}. s − 1 × s − 1 × s − 1 × min { s − 1 , R } .
Let us start with the observation.
Lemma 12.6. For every x ∈ R 4 x \in\mathbb{R}^{4} x ∈ R 4 ,
1 ∣ B R β , R ∣ ∫ ∑ γ ∣ f γ ( y ) ∣ 2 W B R β , R ( x − y ) d y ≲ R β . (164) \frac{1}{\lvert B_{R^{\beta},R}\rvert}\int\sum_{\gamma} \lvert f_{\gamma}(y)\rvert^{2}W_{B_{R^{\beta},R}}(x-y)\,dy \lesssim R^{\beta}. \tag*{(164)} ∣ B R β , R ∣ 1 ∫ γ ∑ ∣ f γ ( y ) ∣ 2 W B R β , R ( x − y ) d y ≲ R β . (164) For every x ∈ U α x \in U_{\alpha} x ∈ U α ,
α 2 ≲ 1 ∣ B K , R ∣ ∫ ∑ ∣ ν ′ ∣ = K − 1 ∣ ∑ γ ⊂ ν ′ f γ ( y ) ∣ 2 W B K , R ( x − y ) d y . \alpha^{2}\lesssim\frac{1}{\lvert B_{K,R}\rvert}\int\sum_{\lvert\nu'\rvert=K^{-1}} \left\lvert\sum_{\gamma\subset\nu'} f_{\gamma}(y)\right\rvert^{2}W_{B_{K,R}}(x-y)\,dy. α 2 ≲ ∣ B K , R ∣ 1 ∫ ∣ ν ′ ∣ = K − 1 ∑ γ ⊂ ν ′ ∑ f γ ( y ) 2 W B K , R ( x − y ) d y . Proof. The first bound follows from ∥ f γ ∥ ∞ ≤ 1 \lVert f_{\gamma}\rVert_{\infty}\leq1 ∥ f γ ∥ ∞ ≤ 1 , # { γ } ≲ R β \#\{\gamma\}\lesssim R^{\beta} # { γ } ≲ R β , and ∫ W B R β , R = ∣ B R β , R ∣ \int W_{B_{R^{\beta},R}}=\lvert B_{R^{\beta},R}\rvert ∫ W B R β , R = ∣ B R β , R ∣ . For R − β ≤ s ≤ K − 1 R^{-\beta}\leq s\leq K^{-1} R − β ≤ s ≤ K − 1 and ∣ ν ′ ∣ = s \lvert\nu'\rvert=s ∣ ν ′ ∣ = s , we have
supp ∑ γ ⊂ ν ′ f γ ⊂ γ 4 ( t ν ′ ) + C B s − 1 , R ∗ . \operatorname{supp}\sum_{\gamma\subset\nu'} f_{\gamma}\subset\gamma_{4}(t_{\nu'})+CB^{*}_{s^{-1},R}. supp γ ⊂ ν ′ ∑ f γ ⊂ γ 4 ( t ν ′ ) + C B s − 1 , R ∗ . The locally constant property, as in [9 Lemma 2] , therefore gives
∑ ∣ ν ′ ∣ = s ∣ ∑ γ ⊂ ν ′ f γ ( x ) ∣ 2 ≲ 1 ∣ B s − 1 , R ∣ ∫ ∑ ∣ ν ′ ∣ = s ∣ ∑ γ ⊂ ν ′ f γ ( y ) ∣ 2 W B s − 1 , R ( x − y ) d y . (165) \sum_{\lvert\nu'\rvert=s}\left\lvert\sum_{\gamma\subset\nu'}f_{\gamma}(x)\right\rvert^{2}\lesssim\frac{1}{\lvert B_{s^{-1},R}\rvert}\int\sum_{\lvert\nu'\rvert=s}\left\lvert\sum_{\gamma\subset\nu'}f_{\gamma}(y)\right\rvert^{2}W_{B_{s^{-1},R}}(x-y)\,dy. \tag*{(165)} ∣ ν ′ ∣ = s ∑ γ ⊂ ν ′ ∑ f γ ( x ) 2 ≲ ∣ B s − 1 , R ∣ 1 ∫ ∣ ν ′ ∣ = s ∑ γ ⊂ ν ′ ∑ f γ ( y ) 2 W B s − 1 , R ( x − y ) d y . (165) Taking s = K − 1 s=K^{-1} s = K − 1 proves the second bound, since ∣ ∑ γ ⊂ ν f γ ( x ) ∣ > α \left\lvert\sum_{\gamma\subset\nu}f_{\gamma}(x)\right\rvert>\alpha ∑ γ ⊂ ν f γ ( x ) > α on U α U_{\alpha} U α .
Choose C ∗ C_{*} C ∗ sufficiently large, and then C K C_{K} C K sufficiently large depending on C ∗ C_{*} C ∗ . Suppose that α > C K R β / 2 \alpha>C_{K}R^{\beta/2} α > C K R β /2 . By Lemma 12.6, the average at scale K − 1 K^{-1} K − 1 exceeds C ∗ R β C_{*}R^{\beta} C ∗ R β , whereas the average at scale R − β R^{-\beta} R − β is at most C ∗ R β C_{*}R^{\beta} C ∗ R β . For dyadic R − β ≤ s < K − 1 R^{-\beta}\leq s<K^{-1} R − β ≤ s < K − 1 , define
U α , s : = { x ∈ U α : 1 ∣ B s − 1 , R ∣ ∫ ∑ ∣ ν ′ ∣ = s ∣ ∑ γ ⊂ ν ′ f γ ( y ) ∣ 2 W B s − 1 , R ( x − y ) d y ≤ C ∗ R β , U_{\alpha,s}:=\{x\in U_{\alpha}:\frac{1}{\lvert B_{s^{-1},R}\rvert}\int\sum_{\lvert\nu'\rvert=s}\lvert\sum_{\gamma\subset\nu'}f_{\gamma}(y)\rvert^{2}W_{B_{s^{-1},R}}(x-y)\,dy\leq C_{*}R^{\beta}, U α , s := { x ∈ U α : ∣ B s − 1 , R ∣ 1 ∫ ∣ ν ′ ∣ = s ∑ ∣ γ ⊂ ν ′ ∑ f γ ( y ) ∣ 2 W B s − 1 , R ( x − y ) d y ≤ C ∗ R β ,
1 ∣ B ( s ′ ) − 1 , R ∣ ∫ ∑ ∣ ν ′ ∣ = s ′ ∣ ∑ γ ⊂ ν ′ f γ ( y ) ∣ 2 W B ( s ′ ) − 1 , R ( x − y ) d y > C ∗ R β for every dyadic s < s ′ ≤ K − 1 } . \frac{1}{\lvert B_{(s')^{-1},R}\rvert}\int\sum_{\lvert\nu'\rvert=s'}\lvert\sum_{\gamma\subset\nu'}f_{\gamma}(y)\rvert^{2}W_{B_{(s')^{-1},R}}(x-y)\,dy>C_{*}R^{\beta}\text{ for every dyadic }s<s'\leq K^{-1}\}. ∣ B ( s ′ ) − 1 , R ∣ 1 ∫ ∣ ν ′ ∣ = s ′ ∑ ∣ γ ⊂ ν ′ ∑ f γ ( y ) ∣ 2 W B ( s ′ ) − 1 , R ( x − y ) d y > C ∗ R β for every dyadic s < s ′ ≤ K − 1 } .
Thus, U α , s U_{\alpha,s} U α , s consists of the points where the weighted average first drops to at most C ∗ R β C_{*}R^{\beta} C ∗ R β at scale s s s , as we pass from coarse to fine scales. These sets partition U α U_{\alpha} U α :
U α = ⨆ R − β ≤ s < K − 1 s d y a d i c U α , s . U_{\alpha}=\bigsqcup_{\substack{R^{-\beta}\leq s<K^{-1}\\ s\ \mathrm{dyadic}}}U_{\alpha,s}. U α = R − β ≤ s < K − 1 s dyadic ⨆ U α , s .
Using a fixed tiling by translates B ∥ B s − 1 , R B\parallel B_{s^{-1},R} B ∥ B s − 1 , R , define
Λ s : = ⋃ B ∥ B s − 1 , R B ∩ U α , s ≠ ∅ B . \Lambda_{s}:=\bigcup_{\substack{B\parallel B_{s^{-1},R}\\ B\cap U_{\alpha,s}\neq\varnothing}}B. Λ s := B ∥ B s − 1 , R B ∩ U α , s = ∅ ⋃ B .
For R − β ≤ s < K − 1 R^{-\beta}\leq s<K^{-1} R − β ≤ s < K − 1 and x ∈ Λ s x\in\Lambda_{s} x ∈ Λ s ,
1 ∣ B s − 1 , R ∣ ∫ ∑ ∣ ν ′ ∣ = s ∣ ∑ γ ⊂ ν ′ f γ ( y ) ∣ 2 W B s − 1 , R ( x − y ) d y ≲ C ∗ R β , \frac{1}{\lvert B_{s^{-1},R}\rvert}\int\sum_{\lvert\nu'\rvert=s}\left\lvert\sum_{\gamma\subset\nu'}f_{\gamma}(y)\right\rvert^{2}W_{B_{s^{-1},R}}(x-y)\,dy\lesssim C_{*}R^{\beta}, ∣ B s − 1 , R ∣ 1 ∫ ∣ ν ′ ∣ = s ∑ γ ⊂ ν ′ ∑ f γ ( y ) 2 W B s − 1 , R ( x − y ) d y ≲ C ∗ R β ,
1 ∣ B ( 2 s ) − 1 , R ∣ ∫ ∑ ∣ ν ′ ∣ = 2 s ∣ ∑ γ ⊂ ν ′ f γ ( y ) ∣ 2 W B ( 2 s ) − 1 , R ( x − y ) d y ≳ C ∗ R β . (166) \frac{1}{\lvert B_{(2s)^{-1},R}\rvert}\int\sum_{\lvert\nu'\rvert=2s}\left\lvert\sum_{\gamma\subset\nu'}f_{\gamma}(y)\right\rvert^{2}W_{B_{(2s)^{-1},R}}(x-y)\,dy\gtrsim C_{*}R^{\beta}. \tag*{(166)} ∣ B ( 2 s ) − 1 , R ∣ 1 ∫ ∣ ν ′ ∣ = 2 s ∑ γ ⊂ ν ′ ∑ f γ ( y ) 2 W B ( 2 s ) − 1 , R ( x − y ) d y ≳ C ∗ R β . (166)
Indeed, for x ∈ Λ s x\in\Lambda_{s} x ∈ Λ s , choose x 0 ∈ U α , s x_{0}\in U_{\alpha,s} x 0 ∈ U α , s in the same tile B ∥ B s − 1 , R B\parallel B_{s^{-1},R} B ∥ B s − 1 , R . Since W B t − 1 , R ( x − y ) ∼ W B t − 1 , R ( x 0 − y ) W_{B_{t^{-1},R}}(x-y)\sim W_{B_{t^{-1},R}}(x_{0}-y) W B t − 1 , R ( x − y ) ∼ W B t − 1 , R ( x 0 − y ) for t = s , 2 s t=s,2s t = s , 2 s , the defining inequalities at x 0 x_{0} x 0 give (12.15) at x x x .
Lemma 12.7. For R − β ≤ s < K − 1 R^{-\beta}\leq s<K^{-1} R − β ≤ s < K − 1 and every ϵ > 0 \epsilon>0 ϵ > 0 ,
∣ Λ s ∣ ≲ ϵ R β ( − 1 + ϵ ) s 3 ∑ γ ∥ f γ ∥ 2 2 , α 2 ≲ R β s − 1 if U α , s ≠ ∅ . (167) \lvert\Lambda_{s}\rvert\lesssim_{\epsilon} R^{\beta(-1+\epsilon)}s^{3}\sum_{\gamma}\lVert f_{\gamma}\rVert_{2}^{2},\qquad\alpha^{2}\lesssim R^{\beta}s^{-1}\ \text{if }U_{\alpha,s}\neq\varnothing. \tag*{(167)} ∣ Λ s ∣ ≲ ϵ R β ( − 1 + ϵ ) s 3 γ ∑ ∥ f γ ∥ 2 2 , α 2 ≲ R β s − 1 if U α , s = ∅ . (167)
Lemma 12.8 (Multilinear restriction estimate). Let 0 < δ ≤ 1 0 < \delta\le1 0 < δ ≤ 1 , and let ν 1 , … , ν 4 ⊂ [ 0 , 1 ] \nu_1,\ldots,\nu_4 \subset[0,1] ν 1 , … , ν 4 ⊂ [ 0 , 1 ] be intervals satisfying dist ( ν i , ν j ) > K − 1 \operatorname{dist}(\nu_i,\nu_j) > K^{-1} dist ( ν i , ν j ) > K − 1 for i ≠ j i \ne j i = j . Suppose that f 1 , … , f 4 f_1,\ldots,f_4 f 1 , … , f 4 are Schwartz functions with
supp f i ^ ⊂ γ 4 ( ν i ) + C [ − δ , δ ] 4 . \operatorname{supp}\widehat{f_i} \subset\gamma_4(\nu_i) + C[-\delta,\delta]^4. supp f i ⊂ γ 4 ( ν i ) + C [ − δ , δ ] 4 . Then
∫ R 4 ∏ i = 1 4 ∣ f i ( x ) ∣ 2 d x ≲ K δ 12 ∏ i = 1 4 ∥ f i ∥ 2 2 . (168) \int_{\mathbb{R}^4} \prod_{i=1}^{4} |f_i(x)|^2\,\mathrm{d}x \lesssim_K \delta^{12} \prod_{i=1}^{4} \|f_i\|_2^2. \tag*{(168)} ∫ R 4 i = 1 ∏ 4 ∣ f i ( x ) ∣ 2 d x ≲ K δ 12 i = 1 ∏ 4 ∥ f i ∥ 2 2 . (168)
Assuming these two lemmas, we first prove Proposition 12.5, and then return to their proofs.
Proof of Proposition 12.5. The case α ≤ C K R β / 2 \alpha\le C_K R^{\beta/2} α ≤ C K R β /2 is easy. By L 2 L^2 L 2 orthogonality,
α 12 + 2 / β ∣ U α ∣ ≲ α 10 + 2 / β ∑ γ ∥ f γ ∥ 2 2 ≲ K R 5 β + 1 ∑ γ ∥ f γ ∥ 2 2 . \alpha^{12+2/\beta}|U_\alpha| \lesssim\alpha^{10+2/\beta}\sum_\gamma\|f_\gamma\|_2^2 \lesssim_K R^{5\beta+1}\sum_\gamma\|f_\gamma\|_2^2. α 12 + 2/ β ∣ U α ∣ ≲ α 10 + 2/ β γ ∑ ∥ f γ ∥ 2 2 ≲ K R 5 β + 1 γ ∑ ∥ f γ ∥ 2 2 .
We may therefore assume α > C K R β / 2 \alpha> C_KR^{\beta/2} α > C K R β /2 . Fix R − β ≤ s < K − 1 R^{-\beta} \le s < K^{-1} R − β ≤ s < K − 1 and B ∥ B s − 1 , R B \parallel B_{s^{-1},R} B ∥ B s − 1 , R meeting U α , s U_{\alpha,s} U α , s . Choose a Schwartz cutoff χ B \chi_B χ B with ∣ χ B ∣ ≥ 1 |\chi_B| \ge1 ∣ χ B ∣ ≥ 1 on B B B , supp χ B ^ ⊂ C B s − 1 , R ∗ \operatorname{supp}\widehat{\chi_B} \subset CB^*_{s^{-1},R} supp χ B ⊂ C B s − 1 , R ∗ , and ∣ χ B ∣ 2 ≲ W B |\chi_B|^2 \lesssim W_B ∣ χ B ∣ 2 ≲ W B . The pieces χ B ∑ γ ⊂ ν ′ f γ \chi_B\sum_{\gamma\subset\nu'}f_\gamma χ B ∑ γ ⊂ ν ′ f γ , ∣ ν ′ ∣ = s |\nu'|=s ∣ ν ′ ∣ = s , have bounded Fourier overlap in the first coordinate. Hence almost orthogonality and the upper bound in (12.15) give
∥ χ B ∑ γ ⊂ ν i f γ ∥ 2 2 ≲ ∑ ν ′ ⊂ ν i ∣ ν ′ ∣ = s ∥ χ B ∑ γ ⊂ ν ′ f γ ∥ 2 2 ≲ R β ∣ B ∣ . \left\|\chi_B\sum_{\gamma\subset\nu_i}f_\gamma\right\|_2^2 \lesssim\sum_{\substack{\nu'\subset\nu_i\\|\nu'|=s}}\left\|\chi_B\sum_{\gamma\subset\nu'}f_\gamma\right\|_2^2 \lesssim R^\beta|B|. χ B γ ⊂ ν i ∑ f γ 2 2 ≲ ν ′ ⊂ ν i ∣ ν ′ ∣ = s ∑ χ B γ ⊂ ν ′ ∑ f γ 2 2 ≲ R β ∣ B ∣.
On the Fourier supports of these localized functions, ∣ ξ 4 − ξ 4 1 ∣ ≲ max { s , R − 1 } |\xi_4-\xi_4^1|\lesssim\max\{s,R^{-1}\} ∣ ξ 4 − ξ 4 1 ∣ ≲ max { s , R − 1 } . To apply Lemma 12.8, partition ξ 4 − ξ 4 1 \xi_4-\xi_4^1 ξ 4 − ξ 4 1 smoothly into intervals of length s s s and write χ B ∑ γ ⊂ ν i f γ = ∑ j g i , j \chi_B\sum_{\gamma\subset\nu_i}f_\gamma=\sum_jg_{i,j} χ B ∑ γ ⊂ ν i f γ = ∑ j g i , j . There are O ( max { 1 , ( R s ) − 1 } ) O(\max\{1,(Rs)^{-1}\}) O ( max { 1 , ( R s ) − 1 }) pieces, with bounded Fourier overlap, so
∑ j ∥ g i , j ∥ 2 2 ≲ ∥ χ B ∑ γ ⊂ ν i f γ ∥ 2 2 . \sum_j\|g_{i,j}\|_2^2\lesssim\left\|\chi_B\sum_{\gamma\subset\nu_i}f_\gamma\right\|_2^2. j ∑ ∥ g i , j ∥ 2 2 ≲ χ B γ ⊂ ν i ∑ f γ 2 2 .
After modulation in x 4 x_4 x 4 , each g i , j g_{i,j} g i , j has Fourier support in γ 4 ( ν i ) + C [ − s , s ] 4 \gamma_4(\nu_i)+C[-s,s]^4 γ 4 ( ν i ) + C [ − s , s ] 4 . The triangle inequality, Lemma 12.8 with δ = s \delta=s δ = s , and Cauchy–Schwarz in j j j therefore give
∥ ∏ i = 1 4 ( χ B ∑ γ ⊂ ν i f γ ) ∥ 2 ≤ ∑ j 1 , … , j 4 ∥ ∏ i = 1 4 g i , j i ∥ 2 \left\|\prod_{i=1}^{4}\left(\chi_B\sum_{\gamma\subset\nu_i}f_\gamma\right)\right\|_2
\le\sum_{j_1,\ldots,j_4}\left\|\prod_{i=1}^{4}g_{i,j_i}\right\|_2 i = 1 ∏ 4 ( χ B γ ⊂ ν i ∑ f γ ) 2 ≤ j 1 , … , j 4 ∑ i = 1 ∏ 4 g i , j i 2
≲ K s 6 ∏ i = 1 4 ∑ j ∥ g i , j ∥ 2 \lesssim_K s^6\prod_{i=1}^{4}\sum_j\|g_{i,j}\|_2 ≲ K s 6 i = 1 ∏ 4 j ∑ ∥ g i , j ∥ 2
≲ K s 6 max { 1 , ( R s ) − 1 } 2 ∏ i = 1 4 ∥ χ B ∑ γ ⊂ ν i f γ ∥ 2 . \lesssim_K s^6\max\{1,(Rs)^{-1}\}^{2}\prod_{i=1}^{4}\left\|\chi_B\sum_{\gamma\subset\nu_i}f_\gamma\right\|_2. ≲ K s 6 max { 1 , ( R s ) − 1 } 2 i = 1 ∏ 4 χ B γ ⊂ ν i ∑ f γ 2 .
Squaring and using the preceding L 2 L^2 L 2 bound together with ∣ B ∣ = s − 3 max { s , R − 1 } − 1 |B|=s^{-3}\max\{s,R^{-1}\}^{-1} ∣ B ∣ = s − 3 max { s , R − 1 } − 1 , we obtain
∫ B ∏ i = 1 4 ∣ ∑ γ ⊂ ν i f γ ∣ 2 ≲ K s 12 max { 1 , ( R s ) − 1 } 4 R 4 β ∣ B ∣ (169) \int_B\prod_{i=1}^{4}\left|\sum_{\gamma\subset\nu_i}f_\gamma\right|^2\lesssim_K s^{12}\max\{1,(Rs)^{-1}\}^{4}R^{4\beta}|B| \tag*{(169)} ∫ B i = 1 ∏ 4 γ ⊂ ν i ∑ f γ 2 ≲ K s 12 max { 1 , ( R s ) − 1 } 4 R 4 β ∣ B ∣ (169)
=R^{4\beta}\max{1,(Rs)^{-1}}|B|.
Summing (12.18) over the tiles in $\Lambda_s$ and using (12.16), we obtain
\alpha^8|U_{\alpha,s}|\lesssim_K R^{4\beta}\max{1,(Rs)^{-1}}|\Lambda_s|
\lesssim_{\epsilon,K}R^{\beta(3+\epsilon)}s^3\max{1,(Rs)^{-1}}\sum_\gamma|f_\gamma|_2^2.
Multiplying by $\alpha^{4+2/\beta}\lesssim(R^\beta/s)^{2+1/\beta}$, we have
\alpha^{12+2/\beta}|U_{\alpha,s}|\lesssim_{\epsilon,K}R^{5\beta+1+\beta\epsilon}s^{1-1/\beta}\max{1,(Rs)^{-1}}\sum_\gamma|f_\gamma|_2^2.
The scale factor satisfies
s 1 − 1 / β max { 1 , ( R s ) − 1 } = { s 1 − 1 / β ≤ 1 , s ≥ R − 1 , R − 1 s − 1 / β ≤ R − 1 ( R − β ) − 1 / β = 1 , R − β ≤ s < R − 1 . s^{1-1/\beta}\max\{1,(Rs)^{-1}\}=
\begin{cases}
s^{1-1/\beta}\le1, & s\ge R^{-1},\\
R^{-1}s^{-1/\beta}\le R^{-1}(R^{-\beta})^{-1/\beta}=1, & R^{-\beta}\le s<R^{-1}.
\end{cases} s 1 − 1/ β max { 1 , ( R s ) − 1 } = { s 1 − 1/ β ≤ 1 , R − 1 s − 1/ β ≤ R − 1 ( R − β ) − 1/ β = 1 , s ≥ R − 1 , R − β ≤ s < R − 1 .
Summing over the dyadic scales and relabeling ϵ \epsilon ϵ gives (12.12). □
Proof of Lemma 12.7. Applying (159) with n = 4 n=4 n = 4 , ∣ ν ∣ = 1 |\nu|=1 ∣ ν ∣ = 1 , and σ = s \sigma=s σ = s gives
∫ { ∑ ∣ ν ′ ∣ = s ∣ ∑ γ ⊂ ν ′ f γ ∣ 2 > C R β } ( ∑ ∣ ν ′ ∣ = s ∣ ∑ γ ⊂ ν ′ f γ ∣ 2 ) 6 ≲ ϵ R β ( 5 + ϵ ) s 3 ∑ γ ∥ f γ ∥ 2 2 . (170) \int_{\left\{\sum_{|\nu'|=s}\left|\sum_{\gamma\subset\nu'}f_\gamma\right|^2>CR^\beta\right\}}
\left(\sum_{|\nu'|=s}\left|\sum_{\gamma\subset\nu'}f_\gamma\right|^2\right)^6
\lesssim_\epsilon R^{\beta(5+\epsilon)}s^3\sum_\gamma\lVert f_\gamma\rVert_2^2. \tag*{(170)} ∫ { ∑ ∣ ν ′ ∣ = s ∣ ∑ γ ⊂ ν ′ f γ ∣ 2 > C R β } ∣ ν ′ ∣ = s ∑ γ ⊂ ν ′ ∑ f γ 2 6 ≲ ϵ R β ( 5 + ϵ ) s 3 γ ∑ ∥ f γ ∥ 2 2 . (170)
for R − β ≤ s ≤ 1 R^{-\beta}\le s\le1 R − β ≤ s ≤ 1 . If C C C is the constant in (170) , then
1 ∣ B ( 2 s ) − 1 , R ∣ ∫ { ∑ ∣ ν ′ ∣ = 2 s ∣ ∑ γ ⊂ ν ′ f γ ( y ) ∣ 2 ≤ C R β } ∑ ∣ ν ′ ∣ = 2 s ∣ ∑ γ ⊂ ν ′ f γ ( y ) ∣ 2 W B ( 2 s ) − 1 , R ( x − y ) d y ≤ C R β . \frac{1}{|B(2s)^{-1},R|}\int_{\left\{\sum_{|\nu'|=2s}\left|\sum_{\gamma\subset\nu'}f_\gamma(y)\right|^2\le CR^\beta\right\}}\sum_{|\nu'|=2s}\left|\sum_{\gamma\subset\nu'}f_\gamma(y)\right|^2W_{B(2s)^{-1},R}(x-y)\,dy\le CR^\beta. ∣ B ( 2 s ) − 1 , R ∣ 1 ∫ { ∑ ∣ ν ′ ∣ = 2 s ∣ ∑ γ ⊂ ν ′ f γ ( y ) ∣ 2 ≤ C R β } ∣ ν ′ ∣ = 2 s ∑ γ ⊂ ν ′ ∑ f γ ( y ) 2 W B ( 2 s ) − 1 , R ( x − y ) d y ≤ C R β .
By the choice of C ∗ C_* C ∗ , subtracting this contribution from the second average in (166) gives
R β ≲ 1 ∣ B ( 2 s ) − 1 , R ∣ ∫ { ∑ ∣ ν ′ ∣ = 2 s ∣ ∑ γ ⊂ ν ′ f γ ( y ) ∣ 2 > C R β } ∑ ∣ ν ′ ∣ = 2 s ∣ ∑ γ ⊂ ν ′ f γ ( y ) ∣ 2 W B ( 2 s ) − 1 , R ( x − y ) d y , x ∈ Λ s . R^\beta\lesssim\frac{1}{|B(2s)^{-1},R|}\int_{\left\{\sum_{|\nu'|=2s}\left|\sum_{\gamma\subset\nu'}f_\gamma(y)\right|^2>CR^\beta\right\}}\sum_{|\nu'|=2s}\left|\sum_{\gamma\subset\nu'}f_\gamma(y)\right|^2W_{B(2s)^{-1},R}(x-y)\,dy,\qquad x\in\Lambda_s. R β ≲ ∣ B ( 2 s ) − 1 , R ∣ 1 ∫ { ∑ ∣ ν ′ ∣ = 2 s ∣ ∑ γ ⊂ ν ′ f γ ( y ) ∣ 2 > C R β } ∣ ν ′ ∣ = 2 s ∑ γ ⊂ ν ′ ∑ f γ ( y ) 2 W B ( 2 s ) − 1 , R ( x − y ) d y , x ∈ Λ s .
Integrating over Λ s \Lambda_s Λ s , and then applying Young’s inequality and (170) , yields
R 6 β ∣ Λ s ∣ ≲ ∫ Λ s [ 1 ∣ B ( 2 s ) − 1 , R ∣ ∫ { ∑ ∣ ν ′ ∣ = 2 s ∣ ∑ γ ⊂ ν ′ f γ ( y ) ∣ 2 > C R β } ∑ ∣ ν ′ ∣ = 2 s ∣ ∑ γ ⊂ ν ′ f γ ( y ) ∣ 2 W B ( 2 s ) − 1 , R ( x − y ) d y ] 6 d x R^{6\beta}|\Lambda_s|\lesssim\int_{\Lambda_s}\left[\frac{1}{|B(2s)^{-1},R|}\int_{\left\{\sum_{|\nu'|=2s}\left|\sum_{\gamma\subset\nu'}f_\gamma(y)\right|^2>CR^\beta\right\}}\sum_{|\nu'|=2s}\left|\sum_{\gamma\subset\nu'}f_\gamma(y)\right|^2W_{B(2s)^{-1},R}(x-y)\,dy\right]^6dx R 6 β ∣ Λ s ∣ ≲ ∫ Λ s ∣ B ( 2 s ) − 1 , R ∣ 1 ∫ { ∑ ∣ ν ′ ∣ = 2 s ∣ ∑ γ ⊂ ν ′ f γ ( y ) ∣ 2 > C R β } ∣ ν ′ ∣ = 2 s ∑ γ ⊂ ν ′ ∑ f γ ( y ) 2 W B ( 2 s ) − 1 , R ( x − y ) d y 6 d x
≲ ∫ { ∑ ∣ ν ′ ∣ = 2 s ∣ ∑ γ ⊂ ν ′ f γ ∣ 2 > C R β } ( ∑ ∣ ν ′ ∣ = 2 s ∣ ∑ γ ⊂ ν ′ f γ ∣ 2 ) 6 ≲ ϵ R β ( 5 + ϵ ) s 3 ∑ γ ∥ f γ ∥ 2 2 . \lesssim\int_{\left\{\sum_{|\nu'|=2s}\left|\sum_{\gamma\subset\nu'}f_\gamma\right|^2>CR^\beta\right\}}
\left(\sum_{|\nu'|=2s}\left|\sum_{\gamma\subset\nu'}f_\gamma\right|^2\right)^6
\lesssim_\epsilon R^{\beta(5+\epsilon)}s^3\sum_\gamma\lVert f_\gamma\rVert_2^2. ≲ ∫ { ∑ ∣ ν ′ ∣ = 2 s ∣ ∑ γ ⊂ ν ′ f γ ∣ 2 > C R β } ∣ ν ′ ∣ = 2 s ∑ γ ⊂ ν ′ ∑ f γ 2 6 ≲ ϵ R β ( 5 + ϵ ) s 3 γ ∑ ∥ f γ ∥ 2 2 .
Finally, for x ∈ U α , s x\in U_{\alpha,s} x ∈ U α , s , Cauchy–Schwarz over the subcaps of ν 1 \nu_1 ν 1 , followed by (165) , gives
α 2 < ∣ ∑ ν ′ ⊂ ν 1 ∣ ν ′ ∣ = s ∑ γ ⊂ ν ′ f γ ( x ) ∣ 2 ≲ s − 1 ∑ ∣ ν ′ ∣ = s ∣ ∑ γ ⊂ ν ′ f γ ( x ) ∣ 2 ≲ s − 1 R β . \alpha^2<\left|\sum_{\substack{\nu'\subset\nu_1\\|\nu'|=s}}\sum_{\gamma\subset\nu'}f_\gamma(x)\right|^2\lesssim s^{-1}\sum_{|\nu'|=s}\left|\sum_{\gamma\subset\nu'}f_\gamma(x)\right|^2\lesssim s^{-1}R^\beta. α 2 < ν ′ ⊂ ν 1 ∣ ν ′ ∣ = s ∑ γ ⊂ ν ′ ∑ f γ ( x ) 2 ≲ s − 1 ∣ ν ′ ∣ = s ∑ γ ⊂ ν ′ ∑ f γ ( x ) 2 ≲ s − 1 R β .
□
Proof of Lemma 12.8. By Bernstein’s inequality, we may assume δ < c / K \delta<c/K δ < c / K for a sufficiently small absolute constant c c c . Write t = ξ 1 t=\xi_1 t = ξ 1 and v j = ξ j − t j v_j=\xi_j-t^j v j = ξ j − t j for j = 2 , 3 , 4 j=2,3,4 j = 2 , 3 , 4 , and set
g i , v ( t ) = f ^ i ( t , t 2 + v 2 , t 3 + v 3 , t 4 + v 4 ) , E g ( x ) = ∫ g ( t ) e 2 π i x ⋅ γ 4 ( t ) d t . g_{i,v}(t)=\widehat f_i(t,t^2+v_2,t^3+v_3,t^4+v_4),\qquad Eg(x)=\int g(t)e^{2\pi i x\cdot\gamma_4(t)}\,dt. g i , v ( t ) = f i ( t , t 2 + v 2 , t 3 + v 3 , t 4 + v 4 ) , E g ( x ) = ∫ g ( t ) e 2 π i x ⋅ γ 4 ( t ) d t .
Then g i , v g_{i,v} g i , v is supported in ν i + [ − C δ , C δ ] \nu_i+[-C\delta,C\delta] ν i + [ − C δ , C δ ] , these enlarged intervals remain separated, and the v v v support has volume O ( δ 3 ) O(\delta^3) O ( δ 3 ) . Fourier inversion and Plancherel give
f i ( x ) = ∫ e 2 π i ( x 2 v 2 + x 3 v 3 + x 4 v 4 ) E g i , v ( x ) d v , ∫ ∥ g i , v ∥ 2 2 d v = ∥ f i ∥ 2 2 . f_i(x)=\int e^{2\pi i(x_2v_2+x_3v_3+x_4v_4)}Eg_{i,v}(x)\,dv,\qquad\int\lVert g_{i,v}\rVert_2^2\,dv=\lVert f_i\rVert_2^2. f i ( x ) = ∫ e 2 π i ( x 2 v 2 + x 3 v 3 + x 4 v 4 ) E g i , v ( x ) d v , ∫ ∥ g i , v ∥ 2 2 d v = ∥ f i ∥ 2 2 .
For fixed shifts v ( 1 ) , … , v ( 4 ) v^{(1)},\ldots,v^{(4)} v ( 1 ) , … , v ( 4 ) , the change of variables Φ ( t 1 , … , t 4 ) = ∑ i = 1 4 γ 4 ( t i ) \Phi(t_1,\ldots,t_4)=\sum_{i=1}^4\gamma_4(t_i) Φ ( t 1 , … , t 4 ) = ∑ i = 1 4 γ 4 ( t i ) is injective on the product of these four enlarged intervals and satisfies ∣ det D Φ ∣ = 24 ∏ i < j ∣ t i − t j ∣ ≳ K 1 |\det D\Phi|=24\prod_{i<j}|t_i-t_j|\gtrsim_K1 ∣ det D Φ∣ = 24 ∏ i < j ∣ t i − t j ∣ ≳ K 1 . Thus Plancherel yields
∥ ∏ i = 1 4 E g i , v ( i ) ∥ 2 ≲ K ∏ i = 1 4 ∥ g i , v ( i ) ∥ 2 . \left\lVert\prod_{i=1}^4Eg_{i,v^{(i)}}\right\rVert_2\lesssim_K\prod_{i=1}^4\lVert g_{i,v^{(i)}}\rVert_2. i = 1 ∏ 4 E g i , v ( i ) 2 ≲ K i = 1 ∏ 4 ∥ g i , v ( i ) ∥ 2 .
Minkowski and Cauchy–Schwarz in v v v now imply
∥ ∏ i = 1 4 f i ∥ 2 ≲ K ∏ i = 1 4 ∫ ∥ g i , v ∥ 2 d v ≲ K δ 6 ∏ i = 1 4 ∥ f i ∥ 2 . \left\lVert\prod_{i=1}^4f_i\right\rVert_2\lesssim_K\prod_{i=1}^4\int\lVert g_{i,v}\rVert_2\,dv\lesssim_K\delta^6\prod_{i=1}^4\lVert f_i\rVert_2. i = 1 ∏ 4 f i 2 ≲ K i = 1 ∏ 4 ∫ ∥ g i , v ∥ 2 d v ≲ K δ 6 i = 1 ∏ 4 ∥ f i ∥ 2 .
Squaring proves (168) . □
Proof of Proposition 12.1 for case 2. Our proof uses the induction on R β R^{\beta} R β . In the standard broad narrow analysis, we apply the rescaling to the narrow case. After applying the rescaling, we use the induction hypothesis with different β \beta β . To do so, we prove the stronger statement
α p ∣ { ∣ ∑ γ f γ ∣ > α } ∣ ≤ C ε R − 1 + β ( p − 7 ) + β ε ∑ γ ∥ f γ ∥ 2 2 (171) \alpha^p\left|\left\{\left|\sum_{\gamma}f_{\gamma}\right|>\alpha\right\}\right|\leq C_{\varepsilon}R^{-1+\beta(p-7)+\beta\varepsilon}\sum_{\gamma}\|f_{\gamma}\|_2^2 \tag*{(171)} α p { γ ∑ f γ > α } ≤ C ε R − 1 + β ( p − 7 ) + β ε γ ∑ ∥ f γ ∥ 2 2 (171)
for 12 ≤ p ≤ 14 12\leq p\leq14 12 ≤ p ≤ 14 and 1 ≤ β ( p − 12 ) / 2 1\leq\beta(p-12)/2 1 ≤ β ( p − 12 ) /2 , by induction on R β R^{\beta} R β . The bounded scale case follows from L 2 L^2 L 2 orthogonality. The standard broad narrow decomposition gives
{ ∣ ∑ γ f γ ∣ > α } ⊂ ⋃ ∣ ν ∣ = K − 1 { ∣ ∑ γ ⊂ ν f γ ∣ > c α } ∪ ⋃ ν 1 , … , ν 4 separated ∣ ν i ∣ = K − 1 { min i ∣ ∑ γ ⊂ ν i f γ ∣ > α / ( C K ) } . \left\{\left|\sum_{\gamma}f_{\gamma}\right|>\alpha\right\}\subset\bigcup_{|\nu|=K^{-1}}\left\{\left|\sum_{\gamma\subset\nu}f_{\gamma}\right|>c\alpha\right\}\cup\bigcup_{\substack{\nu_1,\ldots,\nu_4\ \text{separated}\\|\nu_i|=K^{-1}}}\left\{\min_i\left|\sum_{\gamma\subset\nu_i}f_{\gamma}\right|>\alpha/(CK)\right\}. { γ ∑ f γ > α } ⊂ ∣ ν ∣ = K − 1 ⋃ { γ ⊂ ν ∑ f γ > c α } ∪ ν 1 , … , ν 4 separated ∣ ν i ∣ = K − 1 ⋃ { i min γ ⊂ ν i ∑ f γ > α / ( C K ) } .
where c > 0 c>0 c > 0 is independent of K K K . Since p ≥ 12 + 2 / β p\geq12+2/\beta p ≥ 12 + 2/ β and ∥ ∑ γ ⊂ ν i f γ ∥ ∞ ≲ R β \left\|\sum_{\gamma\subset\nu_i}f_{\gamma}\right\|_{\infty}\lesssim R^{\beta} ∑ γ ⊂ ν i f γ ∞ ≲ R β , Proposition 12.5 bounds the broad contribution by
C ε , K R β ( p − 12 − 2 / β ) R 5 β + 1 + β ε ∑ γ ∥ f γ ∥ 2 2 = C ε , K R − 1 + β ( p − 7 ) + β ε ∑ γ ∥ f γ ∥ 2 2 . C_{\varepsilon,K}R^{\beta(p-12-2/\beta)}R^{5\beta+1+\beta\varepsilon}\sum_{\gamma}\|f_{\gamma}\|_2^2=C_{\varepsilon,K}R^{-1+\beta(p-7)+\beta\varepsilon}\sum_{\gamma}\|f_{\gamma}\|_2^2. C ε , K R β ( p − 12 − 2/ β ) R 5 β + 1 + β ε γ ∑ ∥ f γ ∥ 2 2 = C ε , K R − 1 + β ( p − 7 ) + β ε γ ∑ ∥ f γ ∥ 2 2 .
For the narrow contribution, rescale the interval of length K − 1 K^{-1} K − 1 corresponding to ν \nu ν to [ 0 , 1 ] [0,1] [ 0 , 1 ] . The small cap dimensions become
K R − β × K 2 R − 2 β × K 3 R − 3 β × K 4 R − 1 . KR^{-\beta}\times K^2R^{-2\beta}\times K^3R^{-3\beta}\times K^4R^{-1}. K R − β × K 2 R − 2 β × K 3 R − 3 β × K 4 R − 1 .
If R / K 4 > 1 R/K^4>1 R / K 4 > 1 , the inductive range is preserved because
R K 4 ≤ R β ( p − 12 ) / 2 K 4 ≤ ( R β K ) ( p − 12 ) / 2 . \frac{R}{K^4}\leq\frac{R^{\beta(p-12)/2}}{K^4}\leq\left(\frac{R^{\beta}}{K}\right)^{(p-12)/2}. K 4 R ≤ K 4 R β ( p − 12 ) /2 ≤ ( K R β ) ( p − 12 ) /2 .
Applying the inductive hypothesis at fine scale K R − β KR^{-\beta} K R − β and last width K 4 / R K^4/R K 4 / R , and returning to the original variables, gives the coefficient
( R / K 4 ) − 1 ( R β / K ) p − 7 + ε = K 11 − p − ε R − 1 + β ( p − 7 ) + β ε . (R/K^4)^{-1}(R^{\beta}/K)^{p-7+\varepsilon}=K^{11-p-\varepsilon}R^{-1+\beta(p-7)+\beta\varepsilon}. ( R / K 4 ) − 1 ( R β / K ) p − 7 + ε = K 11 − p − ε R − 1 + β ( p − 7 ) + β ε .
The Jacobian occurs on both sides of (171) . Since p ≥ 12 p\geq12 p ≥ 12 , summing over ν \nu ν bounds the narrow contribution by
C 0 C ε K − 1 − ε R − 1 + β ( p − 7 ) + β ε ∑ γ ∥ f γ ∥ 2 2 , C_0C_{\varepsilon}K^{-1-\varepsilon}R^{-1+\beta(p-7)+\beta\varepsilon}\sum_{\gamma}\|f_{\gamma}\|_2^2, C 0 C ε K − 1 − ε R − 1 + β ( p − 7 ) + β ε γ ∑ ∥ f γ ∥ 2 2 ,
where C 0 C_0 C 0 is independent of K K K and p p p . For the remaining case R / K 4 ≤ 1 R/K^4\leq1 R / K 4 ≤ 1 , cylindrical decoupling for the cubic moment curve, followed by Hölder’s inequality over the R β ∣ ν ∣ R^{\beta}|\nu| R β ∣ ν ∣ small caps and the normalization ∥ f γ ∥ ∞ ≤ 1 \|f_{\gamma}\|_{\infty}\leq1 ∥ f γ ∥ ∞ ≤ 1 , gives
∥ ∑ γ ⊂ ν f γ ∥ 12 12 ≲ ε ( R β ∣ ν ∣ ) ε ( ∑ γ ⊂ ν ∥ f γ ∥ 12 2 ) 6 ≤ ( R β ∣ ν ∣ ) 5 + ε ∑ γ ⊂ ν ∥ f γ ∥ 12 12 ≤ ( R β ∣ ν ∣ ) 5 + ε ∑ γ ⊂ ν ∥ f γ ∥ 2 2 . (172) \begin{aligned}
\left\|\sum_{\gamma\subset\nu}f_{\gamma}\right\|_{12}^{12}\lesssim_{\varepsilon}(R^{\beta}|\nu|)^{\varepsilon}\left(\sum_{\gamma\subset\nu}\|f_{\gamma}\|_{12}^2\right)^6 \\
&\leq(R^{\beta}|\nu|)^{5+\varepsilon}\sum_{\gamma\subset\nu}\|f_{\gamma}\|_{12}^{12}\leq(R^{\beta}|\nu|)^{5+\varepsilon}\sum_{\gamma\subset\nu}\|f_{\gamma}\|_2^2. \tag*{(172)}
\end{aligned} γ ⊂ ν ∑ f γ 12 12 ≲ ε ( R β ∣ ν ∣ ) ε ( γ ⊂ ν ∑ ∥ f γ ∥ 12 2 ) 6 ≤ ( R β ∣ ν ∣ ) 5 + ε γ ⊂ ν ∑ ∥ f γ ∥ 12 12 ≤ ( R β ∣ ν ∣ ) 5 + ε γ ⊂ ν ∑ ∥ f γ ∥ 2 2 . (172)
Together with ∥ ∑ γ ⊂ ν f γ ∥ ∞ ≲ R β / K \left\|\sum_{\gamma\subset\nu}f_{\gamma}\right\|_{\infty}\lesssim R^{\beta}/K ∑ γ ⊂ ν f γ ∞ ≲ R β / K , this gives the required bound directly:
∥ ∑ γ ⊂ ν f γ ∥ p p ≲ ε ( R β / K ) p − 7 + ε ∑ γ ⊂ ν ∥ f γ ∥ 2 2 ≤ ( R / K 4 ) − 1 ( R β / K ) p − 7 + ε ∑ γ ⊂ ν ∥ f γ ∥ 2 2 . \left\|\sum_{\gamma\subset\nu}f_{\gamma}\right\|_p^p\lesssim_{\varepsilon}(R^{\beta}/K)^{p-7+\varepsilon}\sum_{\gamma\subset\nu}\|f_{\gamma}\|_2^2\leq(R/K^4)^{-1}(R^{\beta}/K)^{p-7+\varepsilon}\sum_{\gamma\subset\nu}\|f_{\gamma}\|_2^2. γ ⊂ ν ∑ f γ p p ≲ ε ( R β / K ) p − 7 + ε γ ⊂ ν ∑ ∥ f γ ∥ 2 2 ≤ ( R / K 4 ) − 1 ( R β / K ) p − 7 + ε γ ⊂ ν ∑ ∥ f γ ∥ 2 2 .
Choose K K K so that C 0 K − 1 − ε < 1 / 2 C_0K^{-1-\varepsilon}<1/2 C 0 K − 1 − ε < 1/2 , and then choose C ε C_{\varepsilon} C ε large enough to close the induction. Setting p = 12 + 2 / β p=12+2/\beta p = 12 + 2/ β gives − 1 + β ( p − 7 ) = 1 + 5 β -1+\beta(p-7)=1+5\beta − 1 + β ( p − 7 ) = 1 + 5 β , which proves (12.1) for n = 4 n=4 n = 4 and β ≥ 1 \beta\geq1 β ≥ 1 . This completes the proof of Proposition 12.1. □
Proof of Theorem 1.4
Set p 0 = n ( n − 1 ) + 2 / β p_0=n(n-1)+2/\beta p 0 = n ( n − 1 ) + 2/ β . Choose smooth cutoffs η γ = 1 \eta_\gamma=1 η γ = 1 on γ \gamma γ , supported in 2 γ 2\gamma 2 γ , with ∥ η γ ^ ∥ 1 ≲ 1 \|\widehat{\eta_\gamma}\|_1\lesssim1 ∥ η γ ∥ 1 ≲ 1 , and define T g = ∑ γ g γ ∗ η γ ^ Tg=\sum_\gamma g_\gamma*\widehat{\eta_\gamma} T g = ∑ γ g γ ∗ η γ . We view the input as a function on the disjoint union of one copy of R n \mathbb{R}^n R n for each γ \gamma γ . Orthogonality and Young’s inequality give
∥ T g ∥ 2 2 ≲ ∑ γ ∥ g γ ∥ 2 2 , ∥ T g ∥ ∞ ≲ R β sup γ ∥ g γ ∥ ∞ . \|Tg\|_2^2\lesssim\sum_\gamma\|g_\gamma\|_2^2,\qquad\|Tg\|_\infty\lesssim R^\beta\sup_\gamma\|g_\gamma\|_\infty. ∥ T g ∥ 2 2 ≲ γ ∑ ∥ g γ ∥ 2 2 , ∥ T g ∥ ∞ ≲ R β γ sup ∥ g γ ∥ ∞ .
The proof of Proposition 12.1 also applies to fixed enlargements of the caps. Thus, for ∣ g γ ∣ ≤ 1 E γ |g_\gamma|\leq\mathbf{1}_{E_\gamma} ∣ g γ ∣ ≤ 1 E γ with ∣ E γ ∣ < ∞ |E_\gamma|<\infty ∣ E γ ∣ < ∞ , Young’s inequality and that proposition imply
α p 0 ∣ { ∣ T g ∣ > α } ∣ ≲ δ R β ( p 0 / 2 − 1 + δ ) ∑ γ ∣ E γ ∣ . \alpha^{p_0}|\{|Tg|>\alpha\}|\lesssim_\delta R^{\beta(p_0/2-1+\delta)}\sum_\gamma|E_\gamma|. α p 0 ∣ { ∣ T g ∣ > α } ∣ ≲ δ R β ( p 0 /2 − 1 + δ ) γ ∑ ∣ E γ ∣.
For 0 < t < p 0 − 2 0<t<p_0-2 0 < t < p 0 − 2 , the L 2 L^2 L 2 and L ∞ L^\infty L ∞ bounds above give restricted weak type estimates at p 0 − t p_0-t p 0 − t and p 0 + t p_0+t p 0 + t , with the common coefficient R β ( p 0 / 2 − 1 + δ + t ) R^{\beta(p_0/2-1+\delta+t)} R β ( p 0 /2 − 1 + δ + t ) : use the L 2 L^2 L 2 bound for α ≤ 1 \alpha\leq1 α ≤ 1 and the preceding estimate for 1 < α ≲ R β 1<\alpha\lesssim R^\beta 1 < α ≲ R β . Taking δ + t < ϵ \delta+t<\epsilon δ + t < ϵ and applying restricted weak type interpolation [21 Theorem 8.5] , we obtain
∥ T g ∥ p 0 p 0 ≲ ϵ R β ( p 0 / 2 − 1 + ϵ ) ∑ γ ∥ g γ ∥ p 0 p 0 . \|Tg\|_{p_0}^{p_0}\lesssim_\epsilon R^{\beta(p_0/2-1+\epsilon)}\sum_\gamma\|g_\gamma\|_{p_0}^{p_0}. ∥ T g ∥ p 0 p 0 ≲ ϵ R β ( p 0 /2 − 1 + ϵ ) γ ∑ ∥ g γ ∥ p 0 p 0 .
Interpolating with the L 2 L^2 L 2 bound for 2 ≤ p ≤ p 0 2\leq p\leq p_0 2 ≤ p ≤ p 0 and the L ∞ L^\infty L ∞ bound for p ≥ p 0 p\geq p_0 p ≥ p 0 , and using T ( ( f γ ) γ ) = ∑ γ f γ T((f_\gamma)_\gamma)=\sum_\gamma f_\gamma T (( f γ ) γ ) = ∑ γ f γ , gives
∥ ∑ γ f γ ∥ p p ≲ ϵ , p R β ϵ ( R β ( p / 2 − 1 ) + R β ( p − n ( n − 1 ) / 2 − 1 ) − 1 ) ∑ γ ∥ f γ ∥ p p . (173) \left\|\sum_\gamma f_\gamma\right\|_p^p\lesssim_{\epsilon,p}R^{\beta\epsilon}\left(R^{\beta(p/2-1)}+R^{\beta(p-n(n-1)/2-1)-1}\right)\sum_\gamma\|f_\gamma\|_p^p. \tag*{(173)} γ ∑ f γ p p ≲ ϵ , p R β ϵ ( R β ( p /2 − 1 ) + R β ( p − n ( n − 1 ) /2 − 1 ) − 1 ) γ ∑ ∥ f γ ∥ p p . (173)
This finishes the proof.
Wave envelope estimate by Gan-Maldague-Oh
Let n ≥ 2 n\geq2 n ≥ 2 and let γ : [ 0 , 1 ] → R n \gamma:[0,1]\to\mathbb{R}^n γ : [ 0 , 1 ] → R n be a smooth nondegenerate curve, so that
inf t ∈ [ 0 , 1 ] ∣ det ( γ ′ ( t ) , … , γ ( n ) ( t ) ) ∣ > 0. \inf_{t\in[0,1]}\left|\det(\gamma'(t),\ldots,\gamma^{(n)}(t))\right|>0. t ∈ [ 0 , 1 ] inf det ( γ ′ ( t ) , … , γ ( n ) ( t )) > 0.
For 0 < s ≤ 1 0<s\leq1 0 < s ≤ 1 , let Θ γ ( s ) \Theta_\gamma(s) Θ γ ( s ) be the canonical cap cover indexed by I = [ k s , ( k + 2 ) s ] ∩ [ 0 , 1 ] I=[ks,(k+2)s]\cap[0,1] I = [ k s , ( k + 2 ) s ] ∩ [ 0 , 1 ] , k ∈ Z k\in\mathbb{Z} k ∈ Z , 0 ≤ k s < 1 0\leq ks<1 0 ≤ k s < 1 , with
τ I = { γ ( t ) + ∑ j = 1 n λ j γ ( j ) ( t ) : t ∈ I , ∣ λ j ∣ ≤ s j } . \tau_I=\left\{\gamma(t)+\sum_{j=1}^n\lambda_j\gamma^{(j)}(t):t\in I,\ |\lambda_j|\leq s^j\right\}. τ I = { γ ( t ) + j = 1 ∑ n λ j γ ( j ) ( t ) : t ∈ I , ∣ λ j ∣ ≤ s j } .
Thus a cap at scale s s s has dimensions s × s 2 × ⋯ × s n s\times s^2\times\cdots\times s^n s × s 2 × ⋯ × s n . For R ≥ 1 R\geq1 R ≥ 1 ,
M γ ( R ) = ⋃ θ ∈ Θ γ ( R − 1 / n ) θ . \mathcal{M}_\gamma(R)=\bigcup_{\theta\in\Theta_\gamma(R^{-1/n})}\theta. M γ ( R ) = θ ∈ Θ γ ( R − 1/ n ) ⋃ θ .
For τ ∈ Θ γ ( s ) \tau\in\Theta_\gamma(s) τ ∈ Θ γ ( s ) , let t τ t_\tau t τ be the left endpoint of the interval I I I corresponding to τ \tau τ , and define the dual box and the wave envelope by
τ ∗ = { x ∈ R n : ∣ x ⋅ γ ( j ) ( t τ ) ∣ ≤ s − j , 1 ≤ j ≤ n } , V ~ τ , R = R s n τ ∗ . \tau^*=\{x\in\mathbb{R}^n:|x\cdot\gamma^{(j)}(t_\tau)|\leq s^{-j},\ 1\leq j\leq n\},\qquad\widetilde{V}_{\tau,R}=Rs^n\tau^*. τ ∗ = { x ∈ R n : ∣ x ⋅ γ ( j ) ( t τ ) ∣ ≤ s − j , 1 ≤ j ≤ n } , V τ , R = R s n τ ∗ .
The wave envelope V ~ τ , R \widetilde{V}_{\tau,R} V τ , R has coordinate half widths R s n − 1 , R s n − 2 , … , R Rs^{n-1},Rs^{n-2},\ldots,R R s n − 1 , R s n − 2 , … , R . For τ ∈ Ξ ( s ) \tau\in\Xi(s) τ ∈ Ξ ( s ) , the corresponding wave envelope U ~ τ , R \widetilde{U}_{\tau,R} U τ , R of Gan-Maldague-Oh [12 ] is centered at the origin and has half widths R s d , R s d − 1 , … , R Rs^d,Rs^{d-1},\ldots,R R s d , R s d − 1 , … , R in the coordinates of (1.4). We write V ∥ V ~ τ , R V\parallel\widetilde{V}_{\tau,R} V ∥ V τ , R and U ∥ U ~ τ , R U\parallel\widetilde{U}_{\tau,R} U ∥ U τ , R for the tiles in fixed tilings by translates of these boxes. Up to a fixed dilation, each envelope contains the centered dual boxes of all caps θ ⊂ τ \theta\subset\tau θ ⊂ τ at the finest scale. For n = 3 n=3 n = 3 , the wave envelope V ~ τ , R \widetilde{V}_{\tau,R} V τ , R is comparable to the one used by Maldague [18 ] .
For the moment curve γ = γ n \gamma=\gamma_n γ = γ n and the cone over the moment curve, comparison with (1.3) and (1.4) gives
V ~ τ , R = ( s R 1 / n ) n − 2 V τ , R , U ~ τ , R = ( s R 1 / d ) d − 2 U τ , R . \widetilde{V}_{\tau,R}=(sR^{1/n})^{n-2}V_{\tau,R},\qquad\widetilde{U}_{\tau,R}=(sR^{1/d})^{d-2}U_{\tau,R}. V τ , R = ( s R 1/ n ) n − 2 V τ , R , U τ , R = ( s R 1/ d ) d − 2 U τ , R .
Thus V ~ τ , R = V τ , R \widetilde{V}_{\tau,R}=V_{\tau,R} V τ , R = V τ , R at s = R − 1 / n s=R^{-1/n} s = R − 1/ n and U ~ τ , R = U τ , R \widetilde{U}_{\tau,R}=U_{\tau,R} U τ , R = U τ , R at s = R − 1 / d s=R^{-1/d} s = R − 1/ d . For d = 2 d=2 d = 2 , the latter equality holds at every scale, as in [14 , 9 ] . At coarser scales, our averaging boxes are smaller for curves with n ≥ 3 n\geq3 n ≥ 3 and cones with d ≥ 3 d\geq3 d ≥ 3 . At s = 1 s=1 s = 1 , our boxes are comparable to balls of radii R 2 / n R^{2/n} R 2/ n and R 2 / d R^{2/d} R 2/ d , respectively, while both envelopes of Gan-Maldague-Oh are comparable to balls of radius R R R .
Theorem 13.1 ([12 Theorem 2.1] ). Let n ≥ 2 n \ge2 n ≥ 2 , 4 ≤ p ≤ n ( n + 1 ) − 2 4 \le p \le n(n+1)-2 4 ≤ p ≤ n ( n + 1 ) − 2 , and ϵ > 0 \epsilon> 0 ϵ > 0 . For every Schwartz function f f f with supp f ^ ⊂ M γ ( R ) \operatorname{supp}\widehat{f} \subset\mathcal{M}_{\gamma}(\mathbb{R}) supp f ⊂ M γ ( R ) ,
∫ R n ∣ f ∣ p ≲ R ϵ + p − 4 2 n ∑ R − 1 / n ≤ s ≤ 1 ∑ τ ∈ Θ γ ( s ) ∑ V ∥ V ~ τ , R ∣ V ∣ − 1 ( ∫ R n ∑ θ ⊂ τ ∣ f θ ∣ p / 2 W V ) 2 . (174) \int_{\mathbb{R}^{n}} |f|^{p} \lesssim R^{\epsilon+\frac{p-4}{2n}} \sum_{R^{-1/n}\le s\le1} \sum_{\tau\in\Theta_{\gamma}(s)} \sum_{V\parallel\widetilde{V}_{\tau,R}} |V|^{-1}\left(\int_{\mathbb{R}^{n}}\sum_{\theta\subset\tau}|f_{\theta}|^{p/2}W_{V}\right)^{2}. \tag*{(174)} ∫ R n ∣ f ∣ p ≲ R ϵ + 2 n p − 4 R − 1/ n ≤ s ≤ 1 ∑ τ ∈ Θ γ ( s ) ∑ V ∥ V τ , R ∑ ∣ V ∣ − 1 ( ∫ R n θ ⊂ τ ∑ ∣ f θ ∣ p /2 W V ) 2 . (174)
Proposition 13.2. Let n ≥ 3 n \ge3 n ≥ 3 . Theorem 1.7 implies Theorem 13.1 for the moment curve γ ( t ) = ( t , t 2 , … , t n ) \gamma(t)=(t,t^{2},\ldots,t^{n}) γ ( t ) = ( t , t 2 , … , t n ) .
Proof. It suffices to prove
∫ ∣ f ∣ p ≲ R ϵ / 2 + p − 4 2 n ∫ ( ∑ θ ∣ f θ ∣ p / 2 ) 2 , 4 ≤ p ≤ n ( n + 1 ) − 2. (175) \int|f|^{p}\lesssim R^{\epsilon/2+\frac{p-4}{2n}}\int\left(\sum_{\theta}|f_{\theta}|^{p/2}\right)^{2}, \qquad4\le p\le n(n+1)-2. \tag*{(175)} ∫ ∣ f ∣ p ≲ R ϵ /2 + 2 n p − 4 ∫ ( θ ∑ ∣ f θ ∣ p /2 ) 2 , 4 ≤ p ≤ n ( n + 1 ) − 2. (175)
Indeed, by the essentially constant property, we have
∣ f θ ∣ p / 2 ≲ ∣ f θ ∣ p / 2 ∗ ( W θ ∗ ∣ θ ∗ ∣ ) . |f_{\theta}|^{p/2}\lesssim|f_{\theta}|^{p/2}*\left(\frac{W_{\theta^{*}}}{|\theta^{*}|}\right). ∣ f θ ∣ p /2 ≲ ∣ f θ ∣ p /2 ∗ ( ∣ θ ∗ ∣ W θ ∗ ) .
Assuming (175) , [12 Proposition 4.6] , applied with exponent p / 2 ≥ 2 p/2\ge2 p /2 ≥ 2 , gives
∫ ∣ f ∣ p ≲ R ϵ / 2 + p − 4 2 n ∫ ( ∑ θ ∣ f θ ∣ p / 2 ∗ W θ ∗ ∣ θ ∗ ∣ ) 2 \int|f|^{p}\lesssim R^{\epsilon/2+\frac{p-4}{2n}}\int\left(\sum_{\theta}|f_{\theta}|^{p/2}*\frac{W_{\theta^{*}}}{|\theta^{*}|}\right)^{2} ∫ ∣ f ∣ p ≲ R ϵ /2 + 2 n p − 4 ∫ ( θ ∑ ∣ f θ ∣ p /2 ∗ ∣ θ ∗ ∣ W θ ∗ ) 2
≲ R ϵ + p − 4 2 n ∑ R − 1 / n ≤ s ≤ 1 ∑ τ ∈ Θ γ ( s ) ∑ V ∥ V ~ τ , R ∣ V ∣ − 1 ( ∫ ∑ θ ⊂ τ ∣ f θ ∣ p / 2 W V ) 2 . (176) \lesssim R^{\epsilon+\frac{p-4}{2n}}\sum_{R^{-1/n}\le s\le1}\sum_{\tau\in\Theta_{\gamma}(s)}\sum_{V\parallel\widetilde{V}_{\tau,R}}|V|^{-1}\left(\int\sum_{\theta\subset\tau}|f_{\theta}|^{p/2}W_{V}\right)^{2}. \tag*{(176)} ≲ R ϵ + 2 n p − 4 R − 1/ n ≤ s ≤ 1 ∑ τ ∈ Θ γ ( s ) ∑ V ∥ V τ , R ∑ ∣ V ∣ − 1 ( ∫ θ ⊂ τ ∑ ∣ f θ ∣ p /2 W V ) 2 . (176)
This is the desired result.
We now prove (175) . The case 4 ≤ p ≤ p n 4\le p\le p_{n} 4 ≤ p ≤ p n is immediate by the square function estimate [10 ] . By the estimate and Hölder’s inequality,
∫ ∣ f ∣ p ≲ R ϵ / 2 ( ∫ ( ∑ θ ∣ f θ ∣ 2 ) p / 2 ) ≲ R ϵ / 2 + p − 4 2 n ∫ ( ∑ θ ∣ f θ ∣ p / 2 ) 2 . \int|f|^{p}\lesssim R^{\epsilon/2}\left(\int\left(\sum_{\theta}|f_{\theta}|^{2}\right)^{p/2}\right)\lesssim R^{\epsilon/2+\frac{p-4}{2n}}\int\left(\sum_{\theta}|f_{\theta}|^{p/2}\right)^{2}. ∫ ∣ f ∣ p ≲ R ϵ /2 ∫ ( θ ∑ ∣ f θ ∣ 2 ) p /2 ≲ R ϵ /2 + 2 n p − 4 ∫ ( θ ∑ ∣ f θ ∣ p /2 ) 2 .
So we may assume p n < p ≤ n ( n + 1 ) − 2 p_{n}<p\le n(n+1)-2 p n < p ≤ n ( n + 1 ) − 2 . We’ll apply Theorem 1.7. For the averaging box V V V in the theorem, its amplitude condition gives
α ≲ R ϵ # τ [ 1 ∣ V ∣ ∫ ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p n / 4 W V ] 2 / p n . (177) \alpha\lesssim R^{\epsilon}\#\tau\left[\frac{1}{|V|}\int\left(\sum_{\theta\subset\tau}|f_{\theta}|^{2}\right)^{p_{n}/4}W_{V}\right]^{2/p_{n}}. \tag*{(177)} α ≲ R ϵ # τ ∣ V ∣ 1 ∫ ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p n /4 W V 2/ p n . (177)
To apply the inequality, we use the following identity.
∫ ∣ f ∣ p = p ∫ 0 ∞ α p − p n − 1 [ α p n ∣ { ∣ f ∣ > α } ∣ ] d α . \int|f|^{p}=p\int_{0}^{\infty}\alpha^{p-p_{n}-1}\left[\alpha^{p_{n}}|\{|f|>\alpha\}|\right]\,d\alpha. ∫ ∣ f ∣ p = p ∫ 0 ∞ α p − p n − 1 [ α p n ∣ { ∣ f ∣ > α } ∣ ] d α .
We now apply Theorem 1.7 together with (177) , and obtain
∫ ∣ f ∣ p ≲ R ϵ / 4 ∑ R − 1 / n ≤ s ≤ 1 ∑ τ ∈ Θ γ ( s ) ( # τ ) p − p n ∑ V ∥ V ~ τ , R ∣ V ∣ [ 1 ∣ V ∣ ∫ ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p n / 4 W V ] 2 p / p n \int|f|^{p}\lesssim R^{\epsilon/4}\sum_{R^{-1/n}\le s\le1}\sum_{\tau\in\Theta_{\gamma}(s)}(\#\tau)^{p-p_{n}}\sum_{V\parallel\widetilde{V}_{\tau,R}}|V|\left[\frac{1}{|V|}\int\left(\sum_{\theta\subset\tau}|f_{\theta}|^{2}\right)^{p_{n}/4}W_{V}\right]^{2p/p_{n}} ∫ ∣ f ∣ p ≲ R ϵ /4 R − 1/ n ≤ s ≤ 1 ∑ τ ∈ Θ γ ( s ) ∑ ( # τ ) p − p n V ∥ V τ , R ∑ ∣ V ∣ ∣ V ∣ 1 ∫ ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p n /4 W V 2 p / p n
≲ R ϵ / 4 ∑ R − 1 / n ≤ s ≤ 1 ∑ τ ∈ Θ γ ( s ) ( # τ ) p − p n ∫ ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p / 2 . (178) \lesssim R^{\epsilon/4}\sum_{R^{-1/n}\le s\le1}\sum_{\tau\in\Theta_{\gamma}(s)}(\#\tau)^{p-p_{n}}\int\left(\sum_{\theta\subset\tau}|f_{\theta}|^{2}\right)^{p/2}. \tag*{(178)} ≲ R ϵ /4 R − 1/ n ≤ s ≤ 1 ∑ τ ∈ Θ γ ( s ) ∑ ( # τ ) p − p n ∫ ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p /2 . (178)
The second inequality is by Hölder’s inequality. Since # τ ≲ s − 1 \#\tau\lesssim s^{-1} # τ ≲ s − 1 and τ \tau τ contains O ( R 1 / n s ) O(R^{1/n}s) O ( R 1/ n s ) caps at scale R − 1 / n R^{-1/n} R − 1/ n , Hölder gives
( # τ ) p − p n ( ∑ θ ⊂ τ ∣ f θ ∣ 2 ) p / 2 ≲ s p n − p ( R 1 / n s ) p / 2 − 2 ( ∑ θ ⊂ τ ∣ f θ ∣ p / 2 ) 2 = R p − 4 2 n s n ( n + 1 ) − 2 − p 2 ( ∑ θ ⊂ τ ∣ f θ ∣ p / 2 ) 2 . \begin{aligned}
(\#\tau)^{p-p_{n}}\left(\sum_{\theta\subset\tau}|f_{\theta}|^{2}\right)^{p/2}\lesssim s^{p_{n}-p}(R^{1/n}s)^{p/2-2}\left(\sum_{\theta\subset\tau}|f_{\theta}|^{p/2}\right)^{2} \\
&=R^{\frac{p-4}{2n}}s^{\frac{n(n+1)-2-p}{2}}\left(\sum_{\theta\subset\tau}|f_{\theta}|^{p/2}\right)^{2}.
\end{aligned} ( # τ ) p − p n ( θ ⊂ τ ∑ ∣ f θ ∣ 2 ) p /2 ≲ s p n − p ( R 1/ n s ) p /2 − 2 ( θ ⊂ τ ∑ ∣ f θ ∣ p /2 ) 2 = R 2 n p − 4 s 2 n ( n + 1 ) − 2 − p ( θ ⊂ τ ∑ ∣ f θ ∣ p /2 ) 2 .
The power of s s s is nonnegative. Hence, summing up the parameter s s s and τ \tau τ , we obtain the desired result. □ \square □
Restriction estimates
Write
γ n ( t ) = ( t , t 2 , … , t n ) , E g ( x ) = ∫ 0 1 g ( t ) e 2 π i x ⋅ γ n ( t ) d t , p n = n ( n + 1 ) 2 + 1. \gamma_n(t)=(t,t^2,\ldots,t^n),\qquad Eg(x)=\int_0^1 g(t)e^{2\pi i x\cdot\gamma_n(t)}\,dt,\qquad p_n=\frac{n(n+1)}{2}+1. γ n ( t ) = ( t , t 2 , … , t n ) , E g ( x ) = ∫ 0 1 g ( t ) e 2 π i x ⋅ γ n ( t ) d t , p n = 2 n ( n + 1 ) + 1.
Proposition 14.1. Let n ≥ 3 n\geq3 n ≥ 3 . The amplitude dependent wave envelope type estimate for the moment curve in R n \mathbb{R}^n R n (Theorem 1.7 ) implies that, for every ϵ > 0 \epsilon>0 ϵ > 0 , every ball B R ⊂ R n B_R\subset\mathbb{R}^n B R ⊂ R n of radius R ≥ 2 R\geq2 R ≥ 2 , and every g ∈ L p n ( [ 0 , 1 ] ) g\in L^{p_n}([0,1]) g ∈ L p n ([ 0 , 1 ]) , we have
∥ E g ∥ L p n ( B R ) ≲ ϵ , n R ϵ ∥ g ∥ L p n ( [ 0 , 1 ] ) . (179) \|Eg\|_{L^{p_n}(B_R)}\lesssim_{\epsilon,n} R^\epsilon\|g\|_{L^{p_n}([0,1])}. \tag*{(179)} ∥ E g ∥ L p n ( B R ) ≲ ϵ , n R ϵ ∥ g ∥ L p n ([ 0 , 1 ]) . (179) Consequently, for every q > p n q>p_n q > p n ,
∥ E g ∥ L q ( R n ) ≲ q , n ∥ g ∥ L p n ( [ 0 , 1 ] ) . (180) \|Eg\|_{L^q(\mathbb{R}^n)}\lesssim_{q,n}\|g\|_{L^{p_n}([0,1])}. \tag*{(180)} ∥ E g ∥ L q ( R n ) ≲ q , n ∥ g ∥ L p n ([ 0 , 1 ]) . (180)
Proof. The square function consequence. Let 0 < δ < 1 0<\delta<1 0 < δ < 1 be dyadic, and apply Theorem 1.7 at scale δ − n \delta^{-n} δ − n . Enlarging the collection of selected averaging boxes to the full tiling and letting α ↓ 0 \alpha\downarrow0 α ↓ 0 gives the ordinary wave envelope type estimate. For each parent τ \tau τ , Cauchy–Schwarz and ∫ W V = ∣ V ∣ \int W_V=|V| ∫ W V = ∣ V ∣ give
∣ V ∣ [ 1 ∣ V ∣ ∫ ( ∑ θ ⊂ τ ∣ F θ ∣ 2 ) p n / 4 W V ] 2 ≤ ∫ ( ∑ θ ⊂ τ ∣ F θ ∣ 2 ) p n / 2 W V . |V|\left[\frac{1}{|V|}\int\left(\sum_{\theta\subset\tau}|F_\theta|^2\right)^{p_n/4}W_V\right]^2\leq\int\left(\sum_{\theta\subset\tau}|F_\theta|^2\right)^{p_n/2}W_V. ∣ V ∣ ∣ V ∣ 1 ∫ ( θ ⊂ τ ∑ ∣ F θ ∣ 2 ) p n /4 W V 2 ≤ ∫ ( θ ⊂ τ ∑ ∣ F θ ∣ 2 ) p n /2 W V .
Sum over V V V , using ∑ V ∥ V τ , δ − n W V ≲ n 1 \sum_V\|V_{\tau,\delta^{-n}}W_V\lesssim_n 1 ∑ V ∥ V τ , δ − n W V ≲ n 1 , and then over τ \tau τ . The O ( log δ − 1 ) O(\log\delta^{-1}) O ( log δ − 1 ) angular scales are absorbed into an arbitrarily small power loss. Thus, for every η > 0 \eta>0 η > 0 ,
∫ R n ∣ ∑ θ ∈ Θ ( δ ) F θ ∣ p n ≲ η , n δ − η ∫ R n ( ∑ θ ∈ Θ ( δ ) ∣ F θ ∣ 2 ) p n / 2 , supp F θ ^ ⊂ θ . (181) \int_{\mathbb{R}^n}\left|\sum_{\theta\in\Theta(\delta)}F_\theta\right|^{p_n}\lesssim_{\eta,n}\delta^{-\eta}\int_{\mathbb{R}^n}\left(\sum_{\theta\in\Theta(\delta)}|F_\theta|^2\right)^{p_n/2},\qquad\operatorname{supp}\widehat{F_\theta}\subset\theta. \tag*{(181)} ∫ R n θ ∈ Θ ( δ ) ∑ F θ p n ≲ η , n δ − η ∫ R n θ ∈ Θ ( δ ) ∑ ∣ F θ ∣ 2 p n /2 , supp F θ ⊂ θ . (181)
Overlap of centered dual boxes. Partition [ 0 , 1 ] [0,1] [ 0 , 1 ] into intervals I I I of length δ \delta δ , and let t I t_I t I be their midpoints. Define
T I = { x ∈ R n : ∣ δ j j ! x ⋅ γ n ( j ) ( t I ) ∣ ≤ 1 ( 1 ≤ j ≤ n ) } , μ = ∑ I 1 T I . T_I=\left\{x\in\mathbb{R}^n:\left|\frac{\delta^j}{j!}x\cdot\gamma_n^{(j)}(t_I)\right|\leq1\ (1\leq j\leq n)\right\},\qquad\mu=\sum_I\mathbf{1}_{T_I}. T I = { x ∈ R n : j ! δ j x ⋅ γ n ( j ) ( t I ) ≤ 1 ( 1 ≤ j ≤ n ) } , μ = I ∑ 1 T I .
For each fixed I 0 I_0 I 0 , we claim
∣ T I 0 ∩ { μ ≥ m } ∣ ≲ n δ − n ( n + 1 ) / 2 m − n ( n − 1 ) / 2 ( m ≥ 1 ) . (182) |T_{I_0}\cap\{\mu\geq m\}|\lesssim_n\delta^{-n(n+1)/2}m^{-n(n-1)/2}\qquad(m\geq1). \tag*{(182)} ∣ T I 0 ∩ { μ ≥ m } ∣ ≲ n δ − n ( n + 1 ) /2 m − n ( n − 1 ) /2 ( m ≥ 1 ) . (182)
Put t 0 = t I 0 t_0=t_{I_0} t 0 = t I 0 and use coordinates
y j = δ j j ! x ⋅ γ n ( j ) ( t 0 ) ( 1 ≤ j ≤ n ) , d x = δ − n ( n + 1 ) / 2 d y . y_j=\frac{\delta^j}{j!}x\cdot\gamma_n^{(j)}(t_0)\ (1\leq j\leq n),\qquad dx=\delta^{-n(n+1)/2}\,dy. y j = j ! δ j x ⋅ γ n ( j ) ( t 0 ) ( 1 ≤ j ≤ n ) , d x = δ − n ( n + 1 ) /2 d y .
Then T I 0 T_{I_0} T I 0 becomes [ − 1 , 1 ] n [-1,1]^n [ − 1 , 1 ] n . If t I = t 0 + k δ t_I=t_0+k\delta t I = t 0 + k δ , the first defining inequality for T I T_I T I gives
∣ P y ( k ) ∣ ≤ 1 , P y ( u ) = ∑ j = 1 n j y j u j − 1 . |P_y(k)|\leq1,\qquad P_y(u)=\sum_{j=1}^n j y_j u^{j-1}. ∣ P y ( k ) ∣ ≤ 1 , P y ( u ) = j = 1 ∑ n j y j u j − 1 .
The set { u : ∣ P y ( u ) ∣ ≤ 1 } \{u:|P_y(u)|\leq1\} { u : ∣ P y ( u ) ∣ ≤ 1 } has O n ( 1 ) O_n(1) O n ( 1 ) interval components. Thus, if μ ( x ) ≥ m \mu(x)\geq m μ ( x ) ≥ m and m m m exceeds a fixed dimensional constant, one component contains an interval of length ≳ n m \gtrsim_n m ≳ n m . Fix h = c n m h=c_nm h = c n m , where c n > 0 c_n>0 c n > 0 is a sufficiently small dimensional constant independent of x x x . This interval then contains a cell [ k h , ( k + 1 ) h ] [kh,(k+1)h] [ k h , ( k + 1 ) h ] of the fixed h h h grid. Polynomial derivative bounds on this cell give
z j = h j j ! P y ( j ) ( k h ) , ∣ z j ∣ ≲ n 1 ( 0 ≤ j ≤ n − 1 ) , d y = h − n ( n − 1 ) / 2 n ! d z . z_j=\frac{h^j}{j!}P_y^{(j)}(kh),\qquad|z_j|\lesssim_n 1\ (0\leq j\leq n-1),\qquad dy=\frac{h^{-n(n-1)/2}}{n!}\,dz. z j = j ! h j P y ( j ) ( k h ) , ∣ z j ∣ ≲ n 1 ( 0 ≤ j ≤ n − 1 ) , d y = n ! h − n ( n − 1 ) /2 d z .
We also have ∣ P y ( 0 ) ∣ = ∣ y 1 ∣ ≤ 1 |P_y(0)|=|y_1|\leq1 ∣ P y ( 0 ) ∣ = ∣ y 1 ∣ ≤ 1 , since x ∈ T I 0 x\in T_{I_0} x ∈ T I 0 . In the z z z coordinates this is the slab condition
∣ ∑ j = 0 n − 1 ( − k ) j z j ∣ ≤ 1. \left|\sum_{j=0}^{n-1}(-k)^jz_j\right|\leq1. j = 0 ∑ n − 1 ( − k ) j z j ≤ 1.
Inside the bounded z z z box, its volume is O n ( ( 1 + ∣ k ∣ ) − ( n − 1 ) ) O_n((1+|k|)^{-(n-1)}) O n (( 1 + ∣ k ∣ ) − ( n − 1 ) ) ; for ∣ k ∣ ≥ 1 |k|\geq1 ∣ k ∣ ≥ 1 , integrate first in z n − 1 z_{n-1} z n − 1 . Summing over the grid cells therefore gives
∣ T I 0 ∩ { μ ≥ m } ∣ ≲ n δ − n ( n + 1 ) / 2 h − n ( n − 1 ) / 2 ∑ k ∈ Z ( 1 + ∣ k ∣ ) − ( n − 1 ) ≲ n δ − n ( n + 1 ) / 2 m − n ( n − 1 ) / 2 . |T_{I_0}\cap\{\mu\geq m\}|\lesssim_n\delta^{-n(n+1)/2}h^{-n(n-1)/2}\sum_{k\in\mathbb{Z}}(1+|k|)^{-(n-1)}\lesssim_n\delta^{-n(n+1)/2}m^{-n(n-1)/2}. ∣ T I 0 ∩ { μ ≥ m } ∣ ≲ n δ − n ( n + 1 ) /2 h − n ( n − 1 ) /2 k ∈ Z ∑ ( 1 + ∣ k ∣ ) − ( n − 1 ) ≲ n δ − n ( n + 1 ) /2 m − n ( n − 1 ) /2 .
The series converges because n ≥ 3 n \ge3 n ≥ 3 ; bounded m m m is immediate. This proves (182) . Since
n ( n − 1 ) 2 − ( p n 2 − 1 ) = ( n − 1 ) ( n − 2 ) 4 > 0 , \frac{n(n-1)}{2}-\left(\frac{p_n}{2}-1\right)=\frac{(n-1)(n-2)}{4}>0, 2 n ( n − 1 ) − ( 2 p n − 1 ) = 4 ( n − 1 ) ( n − 2 ) > 0 ,
layer cake yields ∫ T I μ p n / 2 − 1 ≲ n δ − n ( n + 1 ) / 2 \int_{T_I}\mu^{p_n/2-1}\lesssim_n\delta^{-n(n+1)/2} ∫ T I μ p n /2 − 1 ≲ n δ − n ( n + 1 ) /2 . Consequently, Hölder’s inequality gives, for arbitrary coefficients c I c_I c I ,
( ∫ ( ∑ I ∣ c I ∣ 2 1 T I ) p n / 2 ) 2 / p n ≤ ∑ I ∣ c I ∣ p n ∫ T I μ p n / 2 − 1 ≲ n δ − n ( n + 1 ) / 2 ∑ I ∣ c I ∣ p n . (183) \left(\int\left(\sum_I|c_I|^2\mathbf{1}_{T_I}\right)^{p_n/2}\right)^{2/p_n}\leq\sum_I|c_I|^{p_n}\int_{T_I}\mu^{p_n/2-1}\lesssim_n\delta^{-n(n+1)/2}\sum_I|c_I|^{p_n}. \tag*{(183)} ∫ ( I ∑ ∣ c I ∣ 2 1 T I ) p n /2 2/ p n ≤ I ∑ ∣ c I ∣ p n ∫ T I μ p n /2 − 1 ≲ n δ − n ( n + 1 ) /2 I ∑ ∣ c I ∣ p n . (183)
Approximation of the extension operator. Choose a fixed Schwartz function φ \varphi φ on R n \mathbb{R}^n R n with φ ( 0 ) = 1 \varphi(0)=1 φ ( 0 ) = 1 and Fourier support in a sufficiently small cube about the origin, and set
Φ I ( x ) = e 2 π i x ⋅ γ n ( t I ) φ ( ( δ j x ⋅ γ n ( j ) ( t I ) / j ! ) j = 1 n ) . \Phi_I(x)=e^{2\pi i x\cdot\gamma_n(t_I)}\varphi\left(\left(\delta^j x\cdot\gamma_n^{(j)}(t_I)/j!\right)_{j=1}^n\right). Φ I ( x ) = e 2 π i x ⋅ γ n ( t I ) φ ( ( δ j x ⋅ γ n ( j ) ( t I ) / j ! ) j = 1 n ) .
The exact Taylor formula for γ n \gamma_n γ n shows that supp Φ ^ I ⊂ θ I ∈ Θ ( δ ) \operatorname{supp}\widehat{\Phi}_I\subset\theta_I\in\Theta(\delta) supp Φ I ⊂ θ I ∈ Θ ( δ ) , where θ I \theta_I θ I is the cap corresponding to I I I . Schwartz decay gives
∣ Φ I ∣ 2 ≲ n ∑ ℓ ≥ 0 2 − 10 n ℓ 1 2 ℓ T I . |\Phi_I|^2\lesssim_n\sum_{\ell\geq0}2^{-10n\ell}\mathbf{1}_{2^\ell T_I}. ∣ Φ I ∣ 2 ≲ n ℓ ≥ 0 ∑ 2 − 10 n ℓ 1 2 ℓ T I .
Apply (183) to the common dilations 2 ℓ T I 2^\ell T_I 2 ℓ T I . Their L p n / 2 L^{p_n/2} L p n /2 norms acquire the factor 2 2 n ℓ / p n 2^{2n\ell/p_n} 2 2 n ℓ / p n , so Minkowski and the rapid decay imply
( ∫ ( ∑ I ∣ c I Φ I ∣ 2 ) p n / 2 ) 2 / p n ≲ n δ − n ( n + 1 ) / 2 ∑ I ∣ c I ∣ p n . (184) \left(\int\left(\sum_I|c_I\Phi_I|^2\right)^{p_n/2}\right)^{2/p_n}\lesssim_n\delta^{-n(n+1)/2}\sum_I|c_I|^{p_n}. \tag*{(184)} ∫ ( I ∑ ∣ c I Φ I ∣ 2 ) p n /2 2/ p n ≲ n δ − n ( n + 1 ) /2 I ∑ ∣ c I ∣ p n . (184)
It suffices to consider B R = B ( 0 , R ) B_R=B(0,R) B R = B ( 0 , R ) ; a translated ball is handled by modulation of g g g . Choose dyadic δ ∼ R − 2 \delta\sim R^{-2} δ ∼ R − 2 , and put
c I = ∫ I g ( t ) d t , F = ∑ I c I Φ I . c_I=\int_Ig(t)\,dt,\qquad F=\sum_Ic_I\Phi_I. c I = ∫ I g ( t ) d t , F = I ∑ c I Φ I .
The functions c I Φ I c_I\Phi_I c I Φ I are auxiliary Fourier bumps. On B R B_R B R , the smoothness of γ n \gamma_n γ n , φ ( 0 ) = 1 \varphi(0)=1 φ ( 0 ) = 1 , and R δ ≲ 1 R\delta\lesssim1 R δ ≲ 1 give
∣ E g ( x ) − ∑ I c I e 2 π i x ⋅ γ n ( t I ) ∣ + ∣ F ( x ) − ∑ I c I e 2 π i x ⋅ γ n ( t I ) ∣ ≲ n R δ ∥ g ∥ L 1 ( [ 0 , 1 ] ) . \left|Eg(x)-\sum_Ic_Ie^{2\pi i x\cdot\gamma_n(t_I)}\right|+\left|F(x)-\sum_Ic_Ie^{2\pi i x\cdot\gamma_n(t_I)}\right|\lesssim_n R\delta\|g\|_{L^1([0,1])}. E g ( x ) − I ∑ c I e 2 π i x ⋅ γ n ( t I ) + F ( x ) − I ∑ c I e 2 π i x ⋅ γ n ( t I ) ≲ n R δ ∥ g ∥ L 1 ([ 0 , 1 ]) .
Consequently, since p n > n p_n>n p n > n ,
∥ E g − F ∥ L p n ( B R ) ≲ n R 1 + n / p n δ ∥ g ∥ L 1 ( [ 0 , 1 ] ) ≲ n R − 1 + n / p n ∥ g ∥ L p n ( [ 0 , 1 ] ) . \|Eg-F\|_{L^{p_n}(B_R)}\lesssim_n R^{1+n/p_n}\delta\|g\|_{L^1([0,1])}\lesssim_n R^{-1+n/p_n}\|g\|_{L^{p_n}([0,1])}. ∥ E g − F ∥ L p n ( B R ) ≲ n R 1 + n / p n δ ∥ g ∥ L 1 ([ 0 , 1 ]) ≲ n R − 1 + n / p n ∥ g ∥ L p n ([ 0 , 1 ]) .
On the other hand, (181) , (184) , and Hölder on each I I I give
∥ F ∥ p n p n ≲ η , n δ − η − n ( n + 1 ) / 2 ∑ I ∫ I ∣ g ∣ p n ≲ η , n δ − η ∑ I ∫ I ∣ g ∣ p n ≲ η , n R 2 η ∥ g ∥ p n p n . \|F\|_{p_n}^{p_n}\lesssim_{\eta,n}\delta^{-\eta-n(n+1)/2}\sum_I\int_I|g|^{p_n}\lesssim_{\eta,n}\delta^{-\eta}\sum_I\int_I|g|^{p_n}\lesssim_{\eta,n}R^{2\eta}\|g\|_{p_n}^{p_n}. ∥ F ∥ p n p n ≲ η , n δ − η − n ( n + 1 ) /2 I ∑ ∫ I ∣ g ∣ p n ≲ η , n δ − η I ∑ ∫ I ∣ g ∣ p n ≲ η , n R 2 η ∥ g ∥ p n p n .
Here the cancellation of powers of δ \delta δ is exactly p n − 1 = n ( n + 1 ) / 2 p_n-1=n(n+1)/2 p n − 1 = n ( n + 1 ) /2 . Taking η > 0 \eta>0 η > 0 sufficiently small in terms of ϵ \epsilon ϵ proves (14.1).
Finally, a smoothly truncated moment curve in R n \mathbb{R}^n R n has Fourier decay O n ( ( 1 + ∣ x ∣ ) − 1 / n ) O_n((1+|x|)^{-1/n}) O n (( 1 + ∣ x ∣ ) − 1/ n ) by van der Corput’s lemma. The local estimate holds on every ball and, by a fixed affine change of parameter, on any fixed interval slightly larger than [ 0 , 1 ] [0,1] [ 0 , 1 ] . The epsilon removal theorem [13 Theorem 4.1] , with both local exponents equal to p n p_n p n , therefore gives (14.2). □ \square □
This argument uses only the ordinary wave envelope type estimate that follows from Theorem 1.7 . The local endpoint estimate (14.1) has an R ϵ R^\epsilon R ϵ loss; the global conclusion (14.2) is for q > p n q>p_n q > p n .
Standard interpolation argument
The estimate for the cone over the moment curve.
Proof of Lemma 3.1. By the layer cake formula, ∫ ∣ g ∣ > α 2 ∣ g ∣ p n / 2 ≲ ∫ α 2 / 2 ∞ t p n / 2 − 1 ∣ { ∣ g ∣ > t } ∣ d t \int_{|g|>\alpha^2} |g|^{p_n/2} \lesssim\int_{\alpha^2/2}^{\infty} t^{p_n/2-1} |\{|g|>t\}|\,dt ∫ ∣ g ∣ > α 2 ∣ g ∣ p n /2 ≲ ∫ α 2 /2 ∞ t p n /2 − 1 ∣ { ∣ g ∣ > t } ∣ d t . For each τ ∈ Ξ ( σ ) \tau\in\Xi(\sigma) τ ∈ Ξ ( σ ) , take the partition of unity { ψ τ , B } B ∥ τ ∗ \{\psi_{\tau,B}\}_{B\parallel\tau^*} { ψ τ , B } B ∥ τ ∗ constructed before (21) and write g τ = ∑ B ∥ τ ∗ ψ τ , B g τ g_\tau=\sum_{B\parallel\tau^*}\psi_{\tau,B}g_\tau g τ = ∑ B ∥ τ ∗ ψ τ , B g τ . Thus if ξ = ∑ i = 0 n − 1 λ i ϕ n − 1 ( i ) ( t τ ) ∈ supp ψ τ , B g τ \xi=\sum_{i=0}^{n-1}\lambda_i\phi_{n-1}^{(i)}(t_\tau)\in\operatorname{supp}\psi_{\tau,B}g_\tau ξ = ∑ i = 0 n − 1 λ i ϕ n − 1 ( i ) ( t τ ) ∈ supp ψ τ , B g τ , then 1 2 ≤ ∣ λ 0 ∣ ≤ 3 \frac{1}{2}\leq|\lambda_0|\leq3 2 1 ≤ ∣ λ 0 ∣ ≤ 3 and ∣ λ i ∣ ≲ σ i |\lambda_i|\lesssim\sigma^i ∣ λ i ∣ ≲ σ i for 1 ≤ i ≤ n − 1 1\leq i\leq n-1 1 ≤ i ≤ n − 1 . For t > 0 t>0 t > 0 , put
g τ , t = g τ ∑ B ∥ τ ∗ ∥ ψ τ , B 1 / 2 g τ ∥ ∞ ≤ t ψ τ , B . g_{\tau,t}=g_\tau\sum_{\substack{B\parallel\tau^*\\ \|\psi_{\tau,B}^{1/2}g_\tau\|_\infty\leq t}}\psi_{\tau,B}. g τ , t = g τ B ∥ τ ∗ ∥ ψ τ , B 1/2 g τ ∥ ∞ ≤ t ∑ ψ τ , B .
Then ∣ g τ , t ∣ ≲ min { ∣ g τ ∣ , t } |g_{\tau,t}|\lesssim\min\{|g_\tau|,t\} ∣ g τ , t ∣ ≲ min { ∣ g τ ∣ , t } , and the same Fourier support bounds hold for g τ , t g_{\tau,t} g τ , t . Since g τ , t = 0 g_{\tau,t}=0 g τ , t = 0 whenever g τ = 0 g_\tau=0 g τ = 0 , every active parent for { g τ , t } τ \{g_{\tau,t}\}_\tau { g τ , t } τ is also active for { g τ } τ \{g_\tau\}_\tau { g τ } τ ; hence its count is at most N ( s ) N(s) N ( s ) .
Splitting g = ∑ τ ( g τ − g τ , t ) + ∑ τ g τ , t g=\sum_\tau(g_\tau-g_{\tau,t})+\sum_\tau g_{\tau,t} g = ∑ τ ( g τ − g τ , t ) + ∑ τ g τ , t , we will estimate the level sets of the first sum by Chebyshev’s inequality and those of the second by the hypothesis on the cone, each with threshold t / 4 t/4 t /4 .
We first estimate the level set integral for ∑ τ ( g τ − g τ , t ) \sum_\tau(g_\tau-g_{\tau,t}) ∑ τ ( g τ − g τ , t ) . The bound ∥ ψ τ , B g τ ∥ 1 ≲ ∣ B ∣ ∥ ψ τ , B 1 / 2 g τ ∥ ∞ \|\psi_{\tau,B}g_\tau\|_1\lesssim|B|\|\psi_{\tau,B}^{1/2}g_\tau\|_\infty ∥ ψ τ , B g τ ∥ 1 ≲ ∣ B ∣∥ ψ τ , B 1/2 g τ ∥ ∞ and Chebyshev’s inequality give
∫ α 2 / 2 ∞ t p n / 2 − 1 ∣ { ∣ ∑ τ ( g τ − g τ , t ) ∣ > t / 4 } ∣ d t ≲ ∫ α 2 / 2 ∞ t p n / 2 − 2 ∑ τ ∑ B ∥ τ ∗ ∥ ψ τ , B 1 / 2 g τ ∥ ∞ > t ∥ ψ τ , B g τ ∥ 1 d t \int_{\alpha^2/2}^{\infty}t^{p_n/2-1}\left|\left\{\left|\sum_\tau(g_\tau-g_{\tau,t})\right|>t/4\right\}\right|\,dt
\lesssim
\int_{\alpha^2/2}^{\infty}t^{p_n/2-2}\sum_\tau\sum_{\substack{B\parallel\tau^*\\\|\psi_{\tau,B}^{1/2}g_\tau\|_\infty>t}}\|\psi_{\tau,B}g_\tau\|_1\,dt ∫ α 2 /2 ∞ t p n /2 − 1 { τ ∑ ( g τ − g τ , t ) > t /4 } d t ≲ ∫ α 2 /2 ∞ t p n /2 − 2 τ ∑ B ∥ τ ∗ ∥ ψ τ , B 1/2 g τ ∥ ∞ > t ∑ ∥ ψ τ , B g τ ∥ 1 d t
≲ ∑ τ ∑ B ∥ τ ∗ ∥ ψ τ , B 1 / 2 g τ ∥ ∞ > α 2 / 2 ∣ B ∣ ∥ ψ τ , B 1 / 2 g τ ∥ ∞ p n / 2 . \lesssim\sum_\tau\sum_{\substack{B\parallel\tau^*\\\|\psi_{\tau,B}^{1/2}g_\tau\|_\infty>\alpha^2/2}}|B|\|\psi_{\tau,B}^{1/2}g_\tau\|_\infty^{p_n/2}. ≲ τ ∑ B ∥ τ ∗ ∥ ψ τ , B 1/2 g τ ∥ ∞ > α 2 /2 ∑ ∣ B ∣∥ ψ τ , B 1/2 g τ ∥ ∞ p n /2 .
Apply (21) to each g τ g_\tau g τ , with exponent p n / 4 p_n/4 p n /4 . Since the averaging box U τ , σ − ( n − 1 ) U_{\tau,\sigma^{-(n-1)}} U τ , σ − ( n − 1 ) at s = σ s=\sigma s = σ is the dual box τ ∗ \tau^* τ ∗ , the last sum is bounded by
∑ τ ∑ U ∈ G τ ( α C ϵ σ − ϵ ) U ∥ τ ∗ ∣ U ∣ ∥ g τ 1 / 2 ∥ L # p n / 2 ( U ) p n . \sum_\tau\sum_{\substack{U\in\mathcal G_\tau\left(\frac{\alpha}{C_\epsilon\sigma^{-\epsilon}}\right)\\U\parallel\tau^*}}|U|\|g_\tau^{1/2}\|_{L_\#^{p_n/2}(U)}^{p_n}. τ ∑ U ∈ G τ ( C ϵ σ − ϵ α ) U ∥ τ ∗ ∑ ∣ U ∣∥ g τ 1/2 ∥ L # p n /2 ( U ) p n .
We next estimate ∫ α 2 / 2 ∞ t p n / 2 − 1 ∣ { ∣ ∑ τ g τ , t ∣ > t / 4 } ∣ d t \int_{\alpha^2/2}^{\infty}t^{p_n/2-1}|\{|\sum_\tau g_{\tau,t}|>t/4\}|\,dt ∫ α 2 /2 ∞ t p n /2 − 1 ∣ { ∣ ∑ τ g τ , t ∣ > t /4 } ∣ d t . Apply the amplitude dependent wave envelope type estimate for the cone over the moment curve in R n \mathbb R^n R n (Theorem 1.8 with d = n − 1 d=n-1 d = n − 1 ) to ∑ τ g τ , t \sum_\tau g_{\tau,t} ∑ τ g τ , t at scale R = σ − ( n − 1 ) R=\sigma^{-(n-1)} R = σ − ( n − 1 ) and amplitude t / 4 t/4 t /4 , with g τ , t g_{\tau,t} g τ , t as the fine pieces. Using the identity
∥ ( ∑ τ ⊂ ρ ∣ g τ , t ∣ 2 ) 1 / 2 ∥ L # p n − 1 / 2 ( U ) p n − 1 = ∥ ( ∑ τ ⊂ ρ ∣ g τ , t ∣ 2 ) p n − 1 2 p n ∥ L # p n / 2 ( U ) p n \left\|\left(\sum_{\tau\subset\rho}|g_{\tau,t}|^2\right)^{1/2}\right\|_{L_\#^{p_n-1/2}(U)}^{p_n-1}
=
\left\|\left(\sum_{\tau\subset\rho}|g_{\tau,t}|^2\right)^{\frac{p_n-1}{2p_n}}\right\|_{L_\#^{p_n/2}(U)}^{p_n} ( τ ⊂ ρ ∑ ∣ g τ , t ∣ 2 ) 1/2 L # p n − 1/2 ( U ) p n − 1 = ( τ ⊂ ρ ∑ ∣ g τ , t ∣ 2 ) 2 p n p n − 1 L # p n /2 ( U ) p n
and N ( s ) p n − 1 / p n ≤ N ( s ) N(s)^{p_n-1/p_n}\leq N(s) N ( s ) p n − 1/ p n ≤ N ( s ) in the selection condition, we obtain
t p n − 1 ∣ { ∣ ∑ τ g τ , t ∣ > t / 4 } ∣ ≲ ϵ σ − ϵ ∑ σ ≤ s ≤ 1 ∑ ρ ∈ Ξ ( s ) ∑ U ∥ U ρ , σ − ( n − 1 ) U ∈ G ρ ( ( t C ϵ σ − ϵ ) p n − 1 / p n ) ∣ U ∣ ∥ ( ∑ τ ⊂ ρ ∣ g τ , t ∣ 2 ) p n − 1 2 p n ∥ L # p n / 2 ( U ) p n . t^{p_n-1}\left|\left\{\left|\sum_\tau g_{\tau,t}\right|>t/4\right\}\right|
\lesssim_\epsilon\sigma^{-\epsilon}\sum_{\sigma\leq s\leq1}\sum_{\rho\in\Xi(s)}
\sum_{\substack{U\parallel U_{\rho,\sigma^{-(n-1)}}\\U\in\mathcal G_\rho\left(\left(\frac{t}{C_\epsilon\sigma^{-\epsilon}}\right)^{p_n-1/p_n}\right)}}|U|
\left\|\left(\sum_{\tau\subset\rho}|g_{\tau,t}|^2\right)^{\frac{p_n-1}{2p_n}}\right\|_{L_\#^{p_n/2}(U)}^{p_n}. t p n − 1 { τ ∑ g τ , t > t /4 } ≲ ϵ σ − ϵ σ ≤ s ≤ 1 ∑ ρ ∈ Ξ ( s ) ∑ U ∥ U ρ , σ − ( n − 1 ) U ∈ G ρ ( ( C ϵ σ − ϵ t ) p n − 1/ p n ) ∑ ∣ U ∣ ( τ ⊂ ρ ∑ ∣ g τ , t ∣ 2 ) 2 p n p n − 1 L # p n /2 ( U ) p n .
We next verify the selection condition in Lemma 3.1. For t ≥ α 2 / 2 t\geq\alpha^2/2 t ≥ α 2 /2 , the selection condition in the estimate for the cone over the moment curve and min { a , t } 2 ≤ t 2 − p n / p n − 1 a p n / p n − 1 \min\{a,t\}^2\leq t^{2-p_n/p_{n-1}}a^{p_n/p_{n-1}} min { a , t } 2 ≤ t 2 − p n / p n − 1 a p n / p n − 1 give
∥ ( ∑ τ ⊂ ρ ∣ g τ ∣ 1 / 2 ) ∥ L # p n / 2 ( U ) ≥ ∥ ( ∑ τ ⊂ ρ ∣ g τ ∣ p n / p n − 1 ) p n − 1 / ( 2 p n ) ∥ L # p n / 2 ( U ) ≳ ϵ t σ − ϵ p n − 1 / p n N ( s ) . \left\|\left(\sum_{\tau\subset\rho}|g_\tau|^{1/2}\right)\right\|_{L_\#^{p_n/2}(U)}
\geq
\left\|\left(\sum_{\tau\subset\rho}|g_\tau|^{p_n/p_{n-1}}\right)^{p_{n-1}/(2p_n)}\right\|_{L_\#^{p_n/2}(U)}
\gtrsim_\epsilon\frac{\sqrt{t}}{\sigma^{-\epsilon p_{n-1}/p_n}N(s)}. ( τ ⊂ ρ ∑ ∣ g τ ∣ 1/2 ) L # p n /2 ( U ) ≥ ( τ ⊂ ρ ∑ ∣ g τ ∣ p n / p n − 1 ) p n − 1 / ( 2 p n ) L # p n /2 ( U ) ≳ ϵ σ − ϵ p n − 1 / p n N ( s ) t .
Since t ≥ α 2 / 2 t \ge\alpha^2/2 t ≥ α 2 /2 and p n − 1 < p n p_{n-1}<p_n p n − 1 < p n , we have U ∈ G ρ ( α / ( C ϵ σ − ϵ ) ) U \in G_\rho(\alpha/(C\epsilon\sigma^{-\epsilon})) U ∈ G ρ ( α / ( C ϵ σ − ϵ )) . Multiplying the estimate for this cone by t p n / 2 − p n − 1 − 1 t^{p_n/2-p_{n-1}-1} t p n /2 − p n − 1 − 1 and integrating over t ≥ α 2 / 2 t \ge\alpha^2/2 t ≥ α 2 /2 , Tonelli’s theorem gives
∫ α 2 / 2 ∞ t p n / 2 − 1 ∣ { ∣ ∑ τ g τ , t ∣ > t / 4 } ∣ d t ≲ ϵ σ − ϵ ∑ σ ≤ s ≤ 1 ∑ ρ ∈ Ξ ( s ) ∑ U ∈ G ρ U ∈ G ρ ( α C ϵ σ − ϵ ) ∣ U ∣ ∫ 0 ∞ ∥ ( ∑ τ ⊂ ρ ∣ g τ , t ∣ 2 ) 1 / 2 ∥ L # p n / 2 ( U ) p n t p n / 2 − p n − 1 − 1 d t . (185) \int_{\alpha^2/2}^{\infty} t^{p_n/2-1}\left|\left\{\left|\sum_\tau g_{\tau,t}\right|>t/4\right\}\right|\,dt
\lesssim_\epsilon\sigma^{-\epsilon}\sum_{\sigma\le s\le1}\sum_{\rho\in\Xi(s)}\sum_{\substack{U\in G_\rho\\ U\in G_\rho\left(\frac{\alpha}{C\epsilon\sigma^{-\epsilon}}\right)}} |U|\int_0^\infty\left\|\left(\sum_{\tau\subset\rho}|g_{\tau,t}|^2\right)^{1/2}\right\|_{L_\#^{p_n/2}(U)}^{p_n}t^{p_n/2-p_{n-1}-1}\,dt. \tag*{(185)} ∫ α 2 /2 ∞ t p n /2 − 1 { τ ∑ g τ , t > t /4 } d t ≲ ϵ σ − ϵ σ ≤ s ≤ 1 ∑ ρ ∈ Ξ ( s ) ∑ U ∈ G ρ U ∈ G ρ ( C ϵ σ − ϵ α ) ∑ ∣ U ∣ ∫ 0 ∞ ( τ ⊂ ρ ∑ ∣ g τ , t ∣ 2 ) 1/2 L # p n /2 ( U ) p n t p n /2 − p n − 1 − 1 d t . (185)
We now bound the t t t integral in each summand of (A.1) by ∥ ( ∑ τ ⊂ ρ ∣ g τ ∣ 2 ) 1 / 2 ∥ L # p n / 2 ( U ) p n \left\|\left(\sum_{\tau\subset\rho}|g_\tau|^2\right)^{1/2}\right\|_{L_\#^{p_n/2}(U)}^{p_n} ( ∑ τ ⊂ ρ ∣ g τ ∣ 2 ) 1/2 L # p n /2 ( U ) p n . For a > 0 a>0 a > 0 , splitting at t = a t=a t = a gives ∫ 0 ∞ min { a , t } p n − 1 t p n / 2 − p n − 1 − 1 d t ∼ a p n / 2 \int_0^\infty\min\{a,t\}^{p_{n-1}}t^{p_n/2-p_{n-1}-1}\,dt\sim a^{p_n/2} ∫ 0 ∞ min { a , t } p n − 1 t p n /2 − p n − 1 − 1 d t ∼ a p n /2 , since p n / 2 < p n − 1 p_n/2<p_{n-1} p n /2 < p n − 1 . Using ∣ g τ , t ∣ ≲ min { ∣ g τ ∣ , t } |g_{\tau,t}|\lesssim\min\{|g_\tau|,t\} ∣ g τ , t ∣ ≲ min { ∣ g τ ∣ , t } , Minkowski first in the weighted average over U U U and then in the sum over τ \tau τ gives
[ ∫ 0 ∞ ∥ ( ∑ τ ⊂ ρ ∣ g τ , t ∣ 2 ) 1 / 2 ∥ L # p n / 2 ( U ) p n t p n / 2 − p n − 1 − 1 d t ] 1 / 2 ≤ ∫ U [ ∫ 0 ∞ ( ∑ τ ⊂ ρ ∣ g τ , t ( x ) ∣ 2 ) p n / 2 t p n / 2 − p n − 1 − 1 d t ] 1 / 2 \begin{aligned}
\left[\int_0^\infty\left\|\left(\sum_{\tau\subset\rho}|g_{\tau,t}|^2\right)^{1/2}\right\|_{L_\#^{p_n/2}(U)}^{p_n}t^{p_n/2-p_{n-1}-1}\,dt\right]^{1/2} \\
&\le\int_U\left[\int_0^\infty\left(\sum_{\tau\subset\rho}|g_{\tau,t}(x)|^2\right)^{p_n/2}t^{p_n/2-p_{n-1}-1}\,dt\right]^{1/2}
\end{aligned} ∫ 0 ∞ ( τ ⊂ ρ ∑ ∣ g τ , t ∣ 2 ) 1/2 L # p n /2 ( U ) p n t p n /2 − p n − 1 − 1 d t 1/2 ≤ ∫ U ∫ 0 ∞ ( τ ⊂ ρ ∑ ∣ g τ , t ( x ) ∣ 2 ) p n /2 t p n /2 − p n − 1 − 1 d t 1/2
≲ ∥ ( ∑ τ ⊂ ρ ∣ g τ ∣ p n / p n − 1 ) p n − 1 / ( 2 p n ) ∥ L # p n / 2 ( U ) p n / 2 ≤ ∥ ( ∑ τ ⊂ ρ ∣ g τ ∣ 2 ) 1 / 2 ∥ L # p n / 2 ( U ) p n / 2 . (186) \lesssim\left\|\left(\sum_{\tau\subset\rho}|g_\tau|^{p_n/p_{n-1}}\right)^{p_{n-1}/(2p_n)}\right\|_{L_\#^{p_n/2}(U)}^{p_n/2}
\le\left\|\left(\sum_{\tau\subset\rho}|g_\tau|^2\right)^{1/2}\right\|_{L_\#^{p_n/2}(U)}^{p_n/2}. \tag*{(186)} ≲ ( τ ⊂ ρ ∑ ∣ g τ ∣ p n / p n − 1 ) p n − 1 / ( 2 p n ) L # p n /2 ( U ) p n /2 ≤ ( τ ⊂ ρ ∑ ∣ g τ ∣ 2 ) 1/2 L # p n /2 ( U ) p n /2 . (186)
The last inequality follows from ( ∑ τ ⊂ ρ ∣ g τ ∣ p n / p n − 1 ) p n − 1 / p n ≤ ∑ τ ⊂ ρ ∣ g τ ∣ \left(\sum_{\tau\subset\rho}|g_\tau|^{p_n/p_{n-1}}\right)^{p_{n-1}/p_n}\le\sum_{\tau\subset\rho}|g_\tau| ( ∑ τ ⊂ ρ ∣ g τ ∣ p n / p n − 1 ) p n − 1 / p n ≤ ∑ τ ⊂ ρ ∣ g τ ∣ . Squaring (A.2) and substituting into (A.1), we obtain
∫ α 2 / 2 ∞ t p n / 2 − 1 ∣ { ∣ ∑ τ g τ , t ∣ > t / 4 } ∣ d t ≲ ϵ σ − ϵ ∑ σ ≤ s ≤ 1 ∑ ρ ∈ Ξ ( s ) ∑ U ∈ G ρ U ∈ G ρ ( α C ϵ σ − ϵ ) ∣ U ∣ ∥ ( ∑ τ ⊂ ρ ∣ g τ ∣ 2 ) 1 / 2 ∥ L # p n / 2 ( U ) p n . \int_{\alpha^2/2}^{\infty}t^{p_n/2-1}\left|\left\{\left|\sum_\tau g_{\tau,t}\right|>t/4\right\}\right|\,dt
\lesssim_\epsilon\sigma^{-\epsilon}\sum_{\sigma\le s\le1}\sum_{\rho\in\Xi(s)}\sum_{\substack{U\in G_\rho\\ U\in G_\rho\left(\frac{\alpha}{C\epsilon\sigma^{-\epsilon}}\right)}}|U|\left\|\left(\sum_{\tau\subset\rho}|g_\tau|^2\right)^{1/2}\right\|_{L_\#^{p_n/2}(U)}^{p_n}. ∫ α 2 /2 ∞ t p n /2 − 1 { τ ∑ g τ , t > t /4 } d t ≲ ϵ σ − ϵ σ ≤ s ≤ 1 ∑ ρ ∈ Ξ ( s ) ∑ U ∈ G ρ U ∈ G ρ ( C ϵ σ − ϵ α ) ∑ ∣ U ∣ ( τ ⊂ ρ ∑ ∣ g τ ∣ 2 ) 1/2 L # p n /2 ( U ) p n .
Together with the estimate for ∑ τ ( g τ − g τ , t ) \sum_\tau(g_\tau-g_{\tau,t}) ∑ τ ( g τ − g τ , t ) , this proves the lemma. □
The estimate for Taylor cones.
Proof of Lemma 9.1. Use the wave packet decomposition of Appendix § 16 .§ 1 , writing h θ , t h_{\theta,t} h θ , t for the sum of the wave packets of h θ h_\theta h θ with ∥ ψ θ , T 1 / 2 h θ ∥ ∞ ≤ t \|\psi_{\theta,T}^{1/2}h_\theta\|_\infty\le t ∥ ψ θ , T 1/2 h θ ∥ ∞ ≤ t . The wave packet cutoffs preserve the cap bounds up to fixed constants. In particular, every ξ ∈ supp h ^ θ , t \xi\in\operatorname{supp}\widehat{h}_{\theta,t} ξ ∈ supp h θ , t satisfies
ξ = ∑ i = 0 d λ i ϕ d ( i ) ( t θ ) , max 0 ≤ i ≤ m + 1 r i / d ∣ λ i ∣ ≥ 1. \xi=\sum_{i=0}^{d}\lambda_i\phi_{d}^{(i)}(t\theta),\qquad\max_{0\le i\le m+1}r^{i/d}|\lambda_i|\ge1. ξ = i = 0 ∑ d λ i ϕ d ( i ) ( tθ ) , 0 ≤ i ≤ m + 1 max r i / d ∣ λ i ∣ ≥ 1.
Truncation can only decrease the number of active parents. As in Appendix § 16 .§ 1 , layer cake reduces the proof to levels t ≥ α 2 / 2 t\ge\alpha^2/2 t ≥ α 2 /2 , and Chebyshev’s inequality together with (2.10) bounds the contribution of the discarded wave packets by the second sum in the lemma.
For the remaining wave packets, apply (8.8) at order m + 1 m+1 m + 1 , scale r r r , and amplitude t / 4 t/4 t /4 . At m = d − 3 m=d-3 m = d − 3 , use (8.5) and Proposition 8.6 instead. Fix a box B B B selected by this estimate, with parent ρ \rho ρ of width u u u . Its weighted average is ≳ ( t / ( r ϵ N ( u ) ) ) p d − m − 1 / 2 \gtrsim(t/(r^\epsilon N(u)))^{p_{d-m-1}/2} ≳ ( t / ( r ϵ N ( u )) ) p d − m − 1 /2 . To compare it with the average in the first sum of the lemma, observe that
∣ h θ , t ∣ 2 ≲ min { ∣ h θ ∣ , t } 2 ≤ t 2 − p d − m / p d − m − 1 ∣ h θ ∣ p d − m / p d − m − 1 . |h_{\theta,t}|^2\lesssim\min\{|h_\theta|,t\}^2\le t^{2-p_{d-m}/p_{d-m-1}}|h_\theta|^{p_{d-m}/p_{d-m-1}}. ∣ h θ , t ∣ 2 ≲ min { ∣ h θ ∣ , t } 2 ≤ t 2 − p d − m / p d − m − 1 ∣ h θ ∣ p d − m / p d − m − 1 .
All sums and convolutions are positive, and the powers in the average telescope to p d − m − 1 / 4 p_{d-m-1}/4 p d − m − 1 /4 . Thus the average with ∣ h θ ∣ p d − m / p d − m − 1 |h_\theta|^{p_{d-m}/p_{d-m-1}} ∣ h θ ∣ p d − m / p d − m − 1 in place of ∣ h θ , t ∣ 2 |h_{\theta,t}|^2 ∣ h θ , t ∣ 2 is bounded below, up to a constant, by
t p d − m / 4 − p d − m − 1 / 2 ( t r ϵ N ( u ) ) p d − m − 1 / 2 = t p d − m / 4 ( r ϵ N ( u ) ) p d − m − 1 / 2 ≳ ( α r ϵ N ( u ) ) p d − m / 2 . (187) t^{p_{d-m}/4-p_{d-m-1}/2}\left(\frac{t}{r^\epsilon N(u)}\right)^{p_{d-m-1}/2}
=\frac{t^{p_{d-m}/4}}{(r^\epsilon N(u))^{p_{d-m-1}/2}}
\gtrsim\left(\frac{\alpha}{r^\epsilon N(u)}\right)^{p_{d-m}/2}. \tag*{(187)} t p d − m /4 − p d − m − 1 /2 ( r ϵ N ( u ) t ) p d − m − 1 /2 = ( r ϵ N ( u ) ) p d − m − 1 /2 t p d − m /4 ≳ ( r ϵ N ( u ) α ) p d − m /2 . (187)
Here t ≥ α 2 / 2 t \ge\alpha^2/2 t ≥ α 2 /2 and p d − m − 1 < p d − m p_{d-m-1}<p_{d-m} p d − m − 1 < p d − m . Taking a smaller ϵ \epsilon ϵ in the input estimate absorbs the fixed constant in the selection threshold. We may therefore sum over the boxes selected at amplitude α / r ϵ \alpha/r^\epsilon α / r ϵ before integrating in t t t .
Multiply the level set bound for t p d − m − 1 ∣ { ∣ ∑ θ h θ , t ∣ > t / 4 } ∣ t^{p_{d-m-1}}\bigl|\{\lvert\sum_\theta h_{\theta,t}\rvert>t/4\}\bigr| t p d − m − 1 {∣ ∑ θ h θ , t ∣ > t /4 } by d μ ( t ) = t p d − m / 2 − p d − m − 1 − 1 d t d\mu(t)=t^{p_{d-m}/2-p_{d-m-1}-1}\,dt d μ ( t ) = t p d − m /2 − p d − m − 1 − 1 d t . For each fixed box and chain of scales, take the L t 2 ( d μ ) L_t^2(d\mu) L t 2 ( d μ ) norm of the average inside the outer square. Minkowski moves this norm inside F B \mathcal{F}_B F B and the sum over subcaps. At each successive convolution, write a = p d − m − ℓ / p d − m − ℓ − 1 a=p_{d-m-\ell}/p_{d-m-\ell-1} a = p d − m − ℓ / p d − m − ℓ − 1 for the power at that level. Pointwise in the spatial variable, Minkowski gives
∥ ( F ∗ w B τ ) a ∥ L t p d − m − ℓ − 1 / 2 ( d μ ) ≲ ( ∥ F ∥ L t p d − m − ℓ / 2 ( d μ ) ∗ w B τ ) a ( F ≥ 0 ) . \left\|(F*w_{B_{\tau}})^a\right\|_{L_t^{p_{d-m-\ell-1}/2}(d\mu)}
\lesssim\left(\|F\|_{L_t^{p_{d-m-\ell}/2}(d\mu)}*w_{B_{\tau}}\right)^a \qquad(F\ge0). ∥ ( F ∗ w B τ ) a ∥ L t p d − m − ℓ − 1 /2 ( d μ ) ≲ ( ∥ F ∥ L t p d − m − ℓ /2 ( d μ ) ∗ w B τ ) a ( F ≥ 0 ) .
Starting at ℓ = d − m − 3 \ell=d-m-3 ℓ = d − m − 3 and ending at ℓ = 1 \ell=1 ℓ = 1 , the time exponent therefore increases from 2 = p 2 / 2 2=p_2/2 2 = p 2 /2 to p d − m − 1 / 2 p_{d-m-1}/2 p d − m − 1 /2 . At the innermost sum, the scalar integral used in Appendix A.1 gives
∥ ∣ h θ , t ∣ 2 ∥ L t p d − m − 1 / 2 ( d μ ) ≲ ∣ h θ ∣ p d − m / p d − m − 1 , p d − m / 2 < p d − m − 1 . \left\||h_{\theta,t}|^2\right\|_{L_t^{p_{d-m-1}/2}(d\mu)}\lesssim|h_\theta|^{p_{d-m}/p_{d-m-1}},\qquad p_{d-m}/2<p_{d-m-1}. ∣ h θ , t ∣ 2 L t p d − m − 1 /2 ( d μ ) ≲ ∣ h θ ∣ p d − m / p d − m − 1 , p d − m /2 < p d − m − 1 .
When m = d − 3 m=d-3 m = d − 3 , the average has no intermediate convolutions and the time norm is already L t 2 ( t − 3 / 2 d t ) L_t^2(t^{-3/2}\,dt) L t 2 ( t − 3/2 d t ) , giving F B ∑ θ ⊂ ρ ∣ h θ ∣ 7 / 4 \mathcal{F}_B\sum_{\theta\subset\rho}|h_\theta|^{7/4} F B ∑ θ ⊂ ρ ∣ h θ ∣ 7/4 . Squaring this bound for each average and summing over the boxes in (A.3) gives the first sum in the lemma. □
Appendix B. From weak type to strong type
Proof of Lemma 6.1. Fix r r r , g g g , and α \alpha α as in the lemma. By the layer cake formula, ∫ ∣ g ∣ > α ∣ g ∣ p n ≲ ∫ α / 2 ∞ t p n − 1 ∣ { ∣ g ∣ > t } ∣ d t \int_{|g|>\alpha}|g|^{p_n}\lesssim\int_{\alpha/2}^{\infty}t^{p_n-1}|\{|g|>t\}|\,dt ∫ ∣ g ∣ > α ∣ g ∣ p n ≲ ∫ α /2 ∞ t p n − 1 ∣ { ∣ g ∣ > t } ∣ d t . For t > 0 t>0 t > 0 , set
m τ , t = ∑ T ∥ τ ∗ : t / ( C R ) < ∥ ψ τ , T ∥ ∞ ≤ t ψ τ , T . m_{\tau,t}=\sum_{T\parallel\tau_*:t/(CR)<\|\psi_{\tau,T}\|_{\infty}\le t}\psi_{\tau,T}. m τ , t = T ∥ τ ∗ : t / ( C R ) < ∥ ψ τ , T ∥ ∞ ≤ t ∑ ψ τ , T .
The wave packet partition of Section 2.2 gives
g = ∑ τ , T ∥ τ ∗ ∥ ψ τ , T g τ ∥ ∞ ≤ t / ( C R ) ψ τ , T g τ + ∑ τ g τ m τ , t + ∑ τ , T ∥ τ ∗ ∥ ψ τ , T g τ ∥ ∞ > t ψ τ , T g τ . (188) g=\sum_{\substack{\tau,\ T\parallel\tau_*\\ \|\psi_{\tau,T}g_\tau\|_\infty\le t/(CR)}}\psi_{\tau,T}g_\tau+\sum_\tau g_\tau m_{\tau,t}+\sum_{\substack{\tau,\ T\parallel\tau_*\\ \|\psi_{\tau,T}g_\tau\|_\infty>t}}\psi_{\tau,T}g_\tau. \tag*{(188)} g = τ , T ∥ τ ∗ ∥ ψ τ , T g τ ∥ ∞ ≤ t / ( C R ) ∑ ψ τ , T g τ + τ ∑ g τ m τ , t + τ , T ∥ τ ∗ ∥ ψ τ , T g τ ∥ ∞ > t ∑ ψ τ , T g τ . (188)
The first sum on the right of (B.1) has absolute value at most t C R ∑ τ ∑ T ψ τ , T 1 / 2 ≤ t / 4 \frac{t}{CR}\sum_\tau\sum_T\psi_{\tau,T}^{1/2}\le t/4 C R t ∑ τ ∑ T ψ τ , T 1/2 ≤ t /4 , since there are O ( r 1 / n ) ≤ O ( R 1 / n ) O(r^{1/n})\le O(R^{1/n}) O ( r 1/ n ) ≤ O ( R 1/ n ) caps at scale r − 1 / n r^{-1/n} r − 1/ n . Thus the layer cake formula and (B.1) give
∫ ∣ g ∣ > α ∣ g ∣ p n ≲ ∫ α / 2 ∞ t p n − 1 ( ∣ { ∣ ∑ τ g τ m τ , t ∣ > t / 4 } ∣ + ∣ { ∣ ∑ τ , T ∥ τ ∗ ψ τ , T g τ ∣ > t / 4 } ∣ ) d t . (189) \int_{|g|>\alpha}|g|^{p_n}\lesssim\int_{\alpha/2}^{\infty}t^{p_n-1}\left(\left|\left\{\left|\sum_\tau g_\tau m_{\tau,t}\right|>t/4\right\}\right|+\left|\left\{\left|\sum_{\substack{\tau,\ T\parallel\tau_*}}\psi_{\tau,T}g_\tau\right|>t/4\right\}\right|\right)\,dt. \tag*{(189)} ∫ ∣ g ∣ > α ∣ g ∣ p n ≲ ∫ α /2 ∞ t p n − 1 { τ ∑ g τ m τ , t > t /4 } + ⎩ ⎨ ⎧ τ , T ∥ τ ∗ ∑ ψ τ , T g τ > t /4 ⎭ ⎬ ⎫ d t . (189)
We first estimate the first term in parentheses in (B.2) using the hypothesis of Lemma 6.1 with g g g replaced by ∑ τ g τ m τ , t \sum_\tau g_\tau m_{\tau,t} ∑ τ g τ m τ , t and α \alpha α replaced by t / 4 t/4 t /4 . Each g τ m τ , t g_\tau m_{\tau,t} g τ m τ , t has Fourier support in a fixed enlargement of τ \tau τ . Since g τ m τ , t = 0 g_\tau m_{\tau,t}=0 g τ m τ , t = 0 whenever g τ = 0 g_\tau=0 g τ = 0 , the active counts in the weak estimate for ∑ τ g τ m τ , t \sum_\tau g_\tau m_{\tau,t} ∑ τ g τ m τ , t are at most N ( s ) N(s) N ( s ) . For each V ∥ V ρ , r V\parallel V_{\rho,r} V ∥ V ρ , r on its RHS, the selection condition, 0 ≤ m τ , t ≤ 1 0\le m_{\tau,t}\le1 0 ≤ m τ , t ≤ 1 , and t ≥ α / 2 t\ge\alpha/2 t ≥ α /2 give
∫ V ( ∑ τ ⊂ ρ ∣ g τ ∣ 2 ∗ w τ ) p n / 4 ≥ ∫ V ( ∑ τ ⊂ ρ ( ∣ g τ ∣ 2 m τ , t 2 ) ∗ w τ ) p n / 4 ≳ ( t Λ ( R ) N ( s ) ) p n / 2 ≳ ( α Λ ( R ) N ( s ) ) p n / 2 . (190) \int_V\left(\sum_{\tau\subset\rho}|g_\tau|^2*w_\tau\right)^{p_n/4}\ge\int_V\left(\sum_{\tau\subset\rho}(|g_\tau|^2m_{\tau,t}^2)*w_\tau\right)^{p_n/4}
\gtrsim\left(\frac{t}{\Lambda(R)N(s)}\right)^{p_n/2}\gtrsim\left(\frac{\alpha}{\Lambda(R)N(s)}\right)^{p_n/2}. \tag*{(190)} ∫ V ( τ ⊂ ρ ∑ ∣ g τ ∣ 2 ∗ w τ ) p n /4 ≥ ∫ V ( τ ⊂ ρ ∑ ( ∣ g τ ∣ 2 m τ , t 2 ) ∗ w τ ) p n /4 ≳ ( Λ ( R ) N ( s ) t ) p n /2 ≳ ( Λ ( R ) N ( s ) α ) p n /2 . (190)
Hence V ∈ G ρ ( α R ϵ Λ ( R ) ) V\in\mathcal{G}_{\rho}\left(\frac{\alpha}{R^\epsilon\Lambda(R)}\right) V ∈ G ρ ( R ϵ Λ ( R ) α ) . Multiplying the weak estimate by t − 1 t^{-1} t − 1 , integrating over t ≥ α / 2 t\ge\alpha/2 t ≥ α /2 , and using (B.3), we obtain
∫ α / 2 ∞ t p n − 1 ∣ { ∣ ∑ τ g τ m τ , t ∣ > t / 4 } ∣ d t ≲ Λ ( R ) p n ∑ r − 1 / n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∈ G ρ ( α R ϵ Λ ( R ) ) ∣ V ∣ ∫ 0 ∞ [ ∫ V ( ∑ τ ⊂ ρ ( ∣ g τ ∣ 2 m τ , t 2 ) ∗ w τ ) p n / 4 ] 2 d t t . (191) \int_{\alpha/2}^{\infty}t^{p_n-1}\left|\left\{\left|\sum_\tau g_\tau m_{\tau,t}\right|>t/4\right\}\right|\,dt
\lesssim\Lambda(R)^{p_n}\sum_{r^{-1/n}\le s\le1}\sum_{\rho\in\Theta(s)}\sum_{V\in\mathcal{G}_{\rho}\left(\frac{\alpha}{R^\epsilon\Lambda(R)}\right)}|V|\int_0^\infty\left[\int_V\left(\sum_{\tau\subset\rho}(|g_\tau|^2m_{\tau,t}^2)*w_\tau\right)^{p_n/4}\right]^2\frac{dt}{t}. \tag*{(191)} ∫ α /2 ∞ t p n − 1 { τ ∑ g τ m τ , t > t /4 } d t ≲ Λ ( R ) p n r − 1/ n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∈ G ρ ( R ϵ Λ ( R ) α ) ∑ ∣ V ∣ ∫ 0 ∞ ∫ V ( τ ⊂ ρ ∑ ( ∣ g τ ∣ 2 m τ , t 2 ) ∗ w τ ) p n /4 2 t d t . (191)
We estimate the integral in t t t in each summand on the right of (B.4) by the square of the corresponding weighted average over V V V with m τ , t m_{\tau,t} m τ , t removed. By the definition of m τ , t m_{\tau,t} m τ , t and 0 ≤ m τ , t ≤ 1 0 \le m_{\tau,t} \le1 0 ≤ m τ , t ≤ 1 ,
∫ 0 ∞ m τ , t ( x ) p n d t t ≤ ∑ τ ∈ T ∗ ψ τ , T ( x ) ∥ ψ τ , T g τ ∥ ∞ C R ∥ ψ τ , T 1 / 2 g τ ∥ ∞ C R d t t ≲ 1. (192) \int_{0}^{\infty}m_{\tau,t}(x)^{p_n}\frac{dt}{t}\leq\sum_{\tau\in\mathcal{T}^{*}}\psi_{\tau,T}(x)\frac{\|\psi_{\tau,T}g_{\tau}\|_{\infty}^{CR}}{\|\psi_{\tau,T}^{1/2}g_{\tau}\|_{\infty}^{CR}}\frac{dt}{t}\lesssim1. \tag*{(192)} ∫ 0 ∞ m τ , t ( x ) p n t d t ≤ τ ∈ T ∗ ∑ ψ τ , T ( x ) ∥ ψ τ , T 1/2 g τ ∥ ∞ C R ∥ ψ τ , T g τ ∥ ∞ C R t d t ≲ 1. (192)
Apply Minkowski first between L t 2 ( d t / t ) L_t^2(dt/t) L t 2 ( d t / t ) and L x 1 ( W V d x / ∣ V ∣ ) L_x^1(W_V\,dx/|V|) L x 1 ( W V d x /∣ V ∣ ) , and then between L t p n / 2 ( d t / t ) L_t^{p_n/2}(dt/t) L t p n /2 ( d t / t ) and ℓ τ 1 L y 1 ( w τ ( y ) d y ) \ell_\tau^1L_y^1(w_\tau(y)\,dy) ℓ τ 1 L y 1 ( w τ ( y ) d y ) . By (B.5), for each V V V on the right of (B.4),
∥ ∫ V ( ∑ τ ⊂ ρ ∣ g τ ∣ 2 m τ , t 2 ) ∗ w τ ∥ L t 2 ( d t / t ) p n / 4 ≤ ∫ V ∥ ∑ τ ⊂ ρ ∫ ∣ g τ ( x − y ) ∣ 2 m τ , t ( x − y ) 2 w τ ( y ) d y ∥ L t p n / 2 ( d t / t ) p n / 4 ≤ ∫ V [ ∑ τ ⊂ ρ ∫ ∣ g τ ( x − y ) ∣ 2 ∥ m τ , t ( x − y ) ∥ L t p n ( d t / t ) 2 w τ ( y ) d y ] p n / 4 (193) \begin{aligned}
\left\| \int_V\left(\sum_{\tau\subset\rho}|g_\tau|^2m_{\tau,t}^2\right)*w_\tau\right\|_{L_t^2(dt/t)}^{p_n/4}
\leq\int_V\left\|\sum_{\tau\subset\rho}\int|g_\tau(x-y)|^2m_{\tau,t}(x-y)^2w_\tau(y)\,dy\right\|_{L_t^{p_n/2}(dt/t)}^{p_n/4} \\
&\leq\int_V\left[\sum_{\tau\subset\rho}\int|g_\tau(x-y)|^2\|m_{\tau,t}(x-y)\|_{L_t^{p_n}(dt/t)}^2w_\tau(y)\,dy\right]^{p_n/4} \tag*{(193)}
\end{aligned} ∫ V ( τ ⊂ ρ ∑ ∣ g τ ∣ 2 m τ , t 2 ) ∗ w τ L t 2 ( d t / t ) p n /4 ≤ ∫ V τ ⊂ ρ ∑ ∫ ∣ g τ ( x − y ) ∣ 2 m τ , t ( x − y ) 2 w τ ( y ) d y L t p n /2 ( d t / t ) p n /4 ≤ ∫ V [ τ ⊂ ρ ∑ ∫ ∣ g τ ( x − y ) ∣ 2 ∥ m τ , t ( x − y ) ∥ L t p n ( d t / t ) 2 w τ ( y ) d y ] p n /4 (193)
≲ ∫ V ( ∑ τ ⊂ ρ ∣ g τ ∣ 2 ∗ w τ ) p n / 4 . \lesssim\int_V\left(\sum_{\tau\subset\rho}|g_\tau|^2*w_\tau\right)^{p_n/4}. ≲ ∫ V ( τ ⊂ ρ ∑ ∣ g τ ∣ 2 ∗ w τ ) p n /4 .
Squaring (B.6) and substituting into (B.4) therefore gives
∫ α / 2 ∞ t p n − 1 ∣ { ∣ ∑ τ g τ m τ , t ∣ > t / 4 } ∣ d t ≲ R ϵ Λ ( R ) p n ∑ r − 1 / n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∈ G ρ ( R ϵ Λ ( R ) R ) ∣ V ∣ ∥ ( ∑ τ ⊂ ρ ∣ g τ ∣ 2 ∗ w τ ) 1 / 2 ∥ L # p n / 2 ( V ) p n . (194) \int_{\alpha/2}^{\infty}t^{p_n-1}\left|\left\{ \left|\sum_\tau g_\tau m_{\tau,t}\right|>t/4\right\}\right|\,dt\lesssim R^{\epsilon}\Lambda(R)^{p_n}\sum_{r^{-1/n}\leq s\leq1}\sum_{\rho\in\Theta(s)}\sum_{V\in\mathcal{G}_\rho\left(\frac{R^\epsilon\Lambda(R)}{R}\right)}|V|\left\|\left(\sum_{\tau\subset\rho}|g_\tau|^2*w_\tau\right)^{1/2}\right\|_{L_\#^{p_n/2}(V)}^{p_n}. \tag*{(194)} ∫ α /2 ∞ t p n − 1 { τ ∑ g τ m τ , t > t /4 } d t ≲ R ϵ Λ ( R ) p n r − 1/ n ≤ s ≤ 1 ∑ ρ ∈ Θ ( s ) ∑ V ∈ G ρ ( R R ϵ Λ ( R ) ) ∑ ∣ V ∣ ( τ ⊂ ρ ∑ ∣ g τ ∣ 2 ∗ w τ ) 1/2 L # p n /2 ( V ) p n . (194)
We now estimate the second term in parentheses in (B.2) after integration in t t t . Chebyshev’s inequality and ∥ ψ τ , T g τ ∥ 1 ≲ ∣ T ∣ ∥ ψ τ , T 1 / 2 g τ ∥ ∞ \|\psi_{\tau,T}g_\tau\|_1\lesssim|T|\|\psi_{\tau,T}^{1/2}g_\tau\|_\infty ∥ ψ τ , T g τ ∥ 1 ≲ ∣ T ∣∥ ψ τ , T 1/2 g τ ∥ ∞ give
∫ α / 2 ∞ t p n − 1 ∣ { ∣ ∑ τ , T ∣ τ ∗ ψ τ , T g τ ∣ > t / 4 } ∣ d t ≲ ∑ τ ∑ T ∣ τ ∗ ∥ ψ τ , T 1 / 2 g τ ∥ ∞ > t ∣ T ∣ ∥ ψ τ , T 1 / 2 g τ ∥ ∞ p n (195) \int_{\alpha/2}^{\infty}t^{p_n-1}\left|\left\{\left|\sum_{\tau,T\mid\tau^*}\psi_{\tau,T}g_\tau\right|>t/4\right\}\right|\,dt\lesssim\sum_\tau\sum_{\substack{T\mid\tau^*\\\|\psi_{\tau,T}^{1/2}g_\tau\|_\infty>t}}|T|\|\psi_{\tau,T}^{1/2}g_\tau\|_\infty^{p_n} \tag*{(195)} ∫ α /2 ∞ t p n − 1 ⎩ ⎨ ⎧ τ , T ∣ τ ∗ ∑ ψ τ , T g τ > t /4 ⎭ ⎬ ⎫ d t ≲ τ ∑ T ∣ τ ∗ ∥ ψ τ , T 1/2 g τ ∥ ∞ > t ∑ ∣ T ∣∥ ψ τ , T 1/2 g τ ∥ ∞ p n (195)
≲ ∑ τ ∑ T ∣ τ ∗ ∫ T ∣ g τ ∣ p n / 2 ≳ α p n / 2 ∣ T ∣ [ ∫ T ∣ g τ ∣ p n / 2 ] 2 . \lesssim\sum_\tau\sum_{\substack{T\mid\tau^*\\\int_T|g_\tau|^{p_n/2}\gtrsim\alpha^{p_n/2}}}|T|\left[\int_T|g_\tau|^{p_n/2}\right]^2. ≲ τ ∑ T ∣ τ ∗ ∫ T ∣ g τ ∣ p n /2 ≳ α p n /2 ∑ ∣ T ∣ [ ∫ T ∣ g τ ∣ p n /2 ] 2 .
The last inequality is (21) with exponent p n / 2 p_n/2 p n /2 . Since V τ , r = τ ∗ V_{\tau,r}=\tau^* V τ , r = τ ∗ and ∣ g τ ∣ 2 ≲ ∣ g τ ∣ 2 ∗ w τ |g_\tau|^2\lesssim|g_\tau|^2*w_\tau ∣ g τ ∣ 2 ≲ ∣ g τ ∣ 2 ∗ w τ , the last sum in (B.8) is bounded by the s = r − 1 / n s=r^{-1/n} s = r − 1/ n terms on the right of (B.7). Integrating the two terms in parentheses in (B.2), we obtain in each case a bound by the right hand side of (B.7), proving the lemma. □ Sangwon Jeong, Department of Mathematical Sciences, Seoul National University, Republic of Korea Email address: hawl1209@snu.ac.kr
Changkeun Oh, Department of Mathematical Sciences and RIM, Seoul National University, Republic of Korea Email address: changkeun.math@gmail.com