Introduction
Let n ≥ 2 n\ge2 n ≥ 2 and α ≥ 0 \alpha\ge0 α ≥ 0 . For Schwartz function f f f , the Bochner–Riesz operator of order α \alpha α is defined by
B α f ( x ) = ∫ R n ( 1 − ∣ ξ ∣ 2 ) + α f ^ ( ξ ) e 2 π i x ⋅ ξ d ξ B^\alpha f(x)=\int_{\mathbb{R}^n}(1-|\xi|^2)_{+}^{\alpha}\widehat{f}(\xi)e^{2\pi i x\cdot\xi}\,d\xi B α f ( x ) = ∫ R n ( 1 − ∣ ξ ∣ 2 ) + α f ( ξ ) e 2 π i x ⋅ ξ d ξ
where r + = max { r , 0 } r_+=\max\{r,0\} r + = max { r , 0 } and f ^ ( ξ ) = ∫ R n f ( x ) e − 2 π i x ⋅ ξ d x \widehat{f}(\xi)=\int_{\mathbb{R}^n}f(x)e^{-2\pi i x\cdot\xi}\,dx f ( ξ ) = ∫ R n f ( x ) e − 2 π i x ⋅ ξ d x . The study of the L p L^p L p boundedness of B α B^\alpha B α is a fundamental problem in harmonic analysis.
We can rewrite the Bochner–Riesz operator B α B^\alpha B α in kernel form
B α f ( x ) = ∫ R n K α ( x − y ) f ( y ) d y B^\alpha f(x)=\int_{\mathbb{R}^n}K^\alpha(x-y)f(y)\,dy B α f ( x ) = ∫ R n K α ( x − y ) f ( y ) d y
where
K α ( y ) = Γ ( α + 1 ) π α J n 2 + α ( 2 π ∣ y ∣ ) ∣ y ∣ n 2 + α K^\alpha(y)=\frac{\Gamma(\alpha+1)}{\pi^\alpha}\frac{J_{\frac{n}{2}+\alpha}(2\pi|y|)}{|y|^{\frac{n}{2}+\alpha}} K α ( y ) = π α Γ ( α + 1 ) ∣ y ∣ 2 n + α J 2 n + α ( 2 π ∣ y ∣ )
and J ν J_\nu J ν is the Bessel function of order ν \nu ν . By the asymptotic expansion of the Bessel function, we have
∣ K α ( y ) ∣ ≲ ( 1 + ∣ y ∣ ) − n + 1 2 − α . |K^\alpha(y)|\lesssim(1+|y|)^{-\frac{n+1}{2}-\alpha}. ∣ K α ( y ) ∣ ≲ ( 1 + ∣ y ∣ ) − 2 n + 1 − α .
Thus, K α ∈ L 1 ( R n ) K^\alpha\in L^1(\mathbb{R}^n) K α ∈ L 1 ( R n ) when α > n − 1 2 \alpha> \frac{n-1}{2} α > 2 n − 1 . Consequently, B α B^\alpha B α is bounded on L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) for every 1 ≤ p ≤ ∞ 1 \le p \le\infty 1 ≤ p ≤ ∞ in this range. Therefore, the index α = n − 1 2 \alpha= \frac{n-1}{2} α = 2 n − 1 is referred to as the critical index. At this index, the kernel is no longer integrable, and the strong-type estimates do not hold on L 1 L^1 L 1 and L ∞ L^\infty L ∞ , see [31 ] .
Below this index, the boundedness problem depends on p p p . For n ≥ 2 n \ge2 n ≥ 2 and 1 ≤ p ≤ ∞ 1 \le p \le\infty 1 ≤ p ≤ ∞ , p ≠ 2 p \ne2 p = 2 , it is conjectured that B α B^\alpha B α is bounded on L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) if and only if
α > α ( p ) : = max { n ∣ 1 p − 1 2 ∣ − 1 2 , 0 } . \alpha> \alpha(p) := \max\left\{ n\left|\frac{1}{p}-\frac{1}{2}\right|-\frac{1}{2},0\right\}. α > α ( p ) := max { n p 1 − 2 1 − 2 1 , 0 } .
When α = 0 \alpha= 0 α = 0 , B 0 B^0 B 0 is commonly known as the ball multiplier operator. Fefferman [14 ] proved that the ball multiplier is unbounded on L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) for n ≥ 2 n \ge2 n ≥ 2 except p = 2 p = 2 p = 2 . Carleson and Sjölin [12 ] established the conjecture for the Bochner–Riesz operator B α B^\alpha B α in dimension two. In higher dimensions, there has been partial progress made over the years. We refer to [33 , 25 , 2 , 16 ] and the references therein in this direction.
At the index α = α ( p ) \alpha= \alpha(p) α = α ( p ) , the strong-type L p L^p L p estimate fails for p ≠ 2 p \ne2 p = 2 when n ≥ 2 n \ge2 n ≥ 2 . However, weak-type estimates at this index have been studied extensively, see for example [6 , 28 , 32 ] . For the critical index α = n − 1 2 \alpha= \frac{n-1}{2} α = 2 n − 1 , Christ [7 ] proved that
∥ B n − 1 2 f ∥ L 1 , ∞ ( R n ) ≲ ∥ f ∥ L 1 ( R n ) . \left\|B^{\frac{n-1}{2}}f\right\|_{L^{1,\infty}(\mathbb{R}^n)} \lesssim\|f\|_{L^1(\mathbb{R}^n)}. B 2 n − 1 f L 1 , ∞ ( R n ) ≲ ∥ f ∥ L 1 ( R n ) .
Vargas [Var96] extended this result to weighted weak-type ( 1 , 1 ) (1,1) ( 1 , 1 ) estimates for A 1 A_1 A 1 weights. The proof of Christ’s estimate is closely related to similar estimates for rough operators. At the critical index, the Bochner–Riesz kernel has decay of order ∣ x ∣ − n |x|^{-n} ∣ x ∣ − n at infinity, but its derivatives do not satisfy the usual Calderón–Zygmund estimates due to oscillation. To overcome this difficulty, Christ used a Calderón–Zygmund decomposition together with L 2 L^2 L 2 estimates for dyadic pieces of the kernel.
In the same paper, Christ [7 ] proved weak-type ( 1 , 1 ) (1,1) ( 1 , 1 ) estimates for maximal functions associated with rough kernels in dimension two. Christ and Rubio de Francia [CRdF88] extended this result to all dimensions, assuming that the angular function belongs to L log L L\log L L log L . They also established weak-type ( 1 , 1 ) (1,1) ( 1 , 1 ) estimates for rough singular integrals in dimension two. Later, Seeger [29 ] extended the result to all dimensions n ≥ 2 n \ge2 n ≥ 2 .
The weak-type estimates for these operators have been strengthened further through sparse domination principle. Conde-Alonso et al. [CACDPO17] proved sparse bounds for both the Bochner–Riesz operator at the critical index and rough singular integral operators. Their proof uses a common sparse domination method, together with oscillatory estimates from Christ’s argument for the Bochner–Riesz operator and estimates from Seeger’s work for rough singular integrals. More recently, Beltran, Roos and Seeger [4 ] obtained further sparse bounds for Bochner–Riesz operators.
Bilinear Bochner–Riesz operator
We consider the bilinear Bochner–Riesz operator of order α ≥ 0 \alpha\ge0 α ≥ 0
B α ( f , g ) ( x ) = ∬ R n × R n ( 1 − ∣ ξ ∣ 2 − ∣ η ∣ 2 ) + α f ^ ( ξ ) g ^ ( η ) e 2 π i x ⋅ ( ξ + η ) d ξ d η . B^\alpha(f,g)(x)=\iint_{\mathbb{R}^n\times\mathbb{R}^n}(1-|\xi|^2-|\eta|^2)^\alpha_+\widehat{f}(\xi)\widehat{g}(\eta)e^{2\pi ix\cdot(\xi+\eta)}\,d\xi\,d\eta. B α ( f , g ) ( x ) = ∬ R n × R n ( 1 − ∣ ξ ∣ 2 − ∣ η ∣ 2 ) + α f ( ξ ) g ( η ) e 2 π i x ⋅ ( ξ + η ) d ξ d η .
where f , g ∈ S ( R n ) f,g \in\mathcal{S}(\mathbb{R}^n) f , g ∈ S ( R n ) . We can express B α B^\alpha B α in kernel form
B α ( f , g ) ( x ) = ∬ R n × R n K α ( y , z ) f ( x − y ) g ( x − z ) d y d z . B^\alpha(f,g)(x)=\iint_{\mathbb{R}^n\times\mathbb{R}^n}K^\alpha(y,z)f(x-y)g(x-z)\,dy\,dz. B α ( f , g ) ( x ) = ∬ R n × R n K α ( y , z ) f ( x − y ) g ( x − z ) d y d z .
where
K α ( y , z ) = Γ ( α + 1 ) J n + α ( 2 π ∣ ( y , z ) ∣ ) π α ∣ ( y , z ) ∣ n + α . K^\alpha(y,z)=\frac{\Gamma(\alpha+1)J_{n+\alpha}(2\pi|(y,z)|)}{\pi^\alpha|(y,z)|^{n+\alpha}}. K α ( y , z ) = π α ∣ ( y , z ) ∣ n + α Γ ( α + 1 ) J n + α ( 2 π ∣ ( y , z ) ∣ ) .
As in the linear case, the asymptotic expansion of Bessel function gives
∣ K α ( y , z ) ∣ ≲ ( 1 + ∣ ( y , z ) ∣ ) − n − α − 1 2 . |K^\alpha(y,z)|\lesssim(1+|(y,z)|)^{-n-\alpha-\frac{1}{2}}. ∣ K α ( y , z ) ∣ ≲ ( 1 + ∣ ( y , z ) ∣ ) − n − α − 2 1 .
When α > n − 1 2 \alpha> n-\frac{1}{2} α > n − 2 1 , this kernel is dominated by a product of integrable functions on R n \mathbb{R}^n R n . Hence
∥ B α ( f , g ) ∥ L p ( R n ) ≲ ∥ f ∥ L 1 ( R n ) ∥ g ∥ L 2 ( R n ) \|B^\alpha(f,g)\|_{L^p(\mathbb{R}^n)}\lesssim\|f\|_{L^1(\mathbb{R}^n)}\|g\|_{L^2(\mathbb{R}^n)} ∥ B α ( f , g ) ∥ L p ( R n ) ≲ ∥ f ∥ L 1 ( R n ) ∥ g ∥ L 2 ( R n )
for all 1 ≤ p 1 , p 2 ≤ ∞ 1 \le p_1,p_2 \le\infty 1 ≤ p 1 , p 2 ≤ ∞ satisfying 1 p = 1 p 1 + 1 p 2 \frac{1}{p}=\frac{1}{p_1}+\frac{1}{p_2} p 1 = p 1 1 + p 2 1 . Thus α c = n − 1 2 \alpha_c=n-\frac{1}{2} α c = n − 2 1 is the critical index for the bilinear Bochner–Riesz operator.
Likewise the linear case, the operator B 0 \mathcal{B}^0 B 0 is commonly known as bilinear ball multiplier operator. Grafakos and Li [15 ] proved boundedness of the bilinear ball multiplier in dimension one in the local L 2 L^2 L 2 range (2 ≤ p 1 , p 2 , p ′ < ∞ 2\le p_1,p_2,p'<\infty 2 ≤ p 1 , p 2 , p ′ < ∞ where p ′ p' p ′ denotes the conjugate exponent of p p p given by 1 = 1 p + 1 p ′ 1=\frac{1}{p}+\frac{1}{p'} 1 = p 1 + p ′ 1 ). In dimensions n ≥ 2 n\ge2 n ≥ 2 , Diestel and Grafakos [13 ] showed that B 0 \mathcal{B}^0 B 0 is unbounded in the Banach range when exactly one of p 1 , p 2 , p ′ p_1,p_2,p' p 1 , p 2 , p ′ is less than 2. In [3 ] , Bernicot et al. first studied the boundedness of B α \mathcal{B}^\alpha B α for positive α \alpha α . Later, the boundedness range for a fixed α \alpha α was improved by Jeong, Lee and Vargas [22 ] in Banach range and Liu and Wang [27 ] in the non-Banach range p < 1 p<1 p < 1 . In dimension one, Jotsaroop and Shrivastava [23 ] extended the results to the non-Banach range. Maximal and square-function estimates for bilinear Bochner–Riesz operators have been studied in [21 , 23 , 8 , 11 ] . These results have also been extended to multilinear Bochner–Riesz operators in [30 , 19 ] .
At the critical index α = n − 1 2 \alpha=n-\frac{1}{2} α = n − 2 1 , B n − 1 2 \mathcal{B}^{n-\frac{1}{2}} B n − 2 1 is known to be bounded from L p 1 × L p 2 L^{p_1}\times L^{p_2} L p 1 × L p 2 to L p L^p L p for all 1 ≤ p 1 , p 2 ≤ ∞ 1\le p_1,p_2\le\infty 1 ≤ p 1 , p 2 ≤ ∞ , with 1 p = 1 p 1 + 1 p 2 \frac{1}{p}=\frac{1}{p_1}+\frac{1}{p_2} p 1 = p 1 1 + p 2 1 , except at ( p 1 , p 2 ) = ( 1 , 1 ) (p_1,p_2)=(1,1) ( p 1 , p 2 ) = ( 1 , 1 ) and ( ∞ , ∞ ) (\infty,\infty) ( ∞ , ∞ ) . However, Bernicot et al. [3 Proposition 4.2] showed that the strong-type estimate fails at ( p 1 , p 2 ; p ) = ( 1 , 1 ; 1 2 ) (p_1,p_2;p)=(1,1;\frac{1}{2}) ( p 1 , p 2 ; p ) = ( 1 , 1 ; 2 1 ) . Jotsaroop, Shrivastava and Shuin [24 ] proved weighted estimates at the critical index for bilinear Muckenhoupt weights when 1 < p 1 , p 2 ≤ ∞ 1<p_1,p_2\le\infty 1 < p 1 , p 2 ≤ ∞ . More recently, Carro, Luque and Sánchez-Pascuala [9 ] obtained weighted endpoint estimates using bilinear extrapolation methods, including estimates in Orlicz-type spaces near L 1 L^1 L 1 .
Unlike the linear Bochner–Riesz operator at the critical index, the weak-type ( 1 , 1 ; 1 2 ) (1,1;\frac{1}{2}) ( 1 , 1 ; 2 1 ) boundedness of B n − 1 2 \mathcal{B}^{n-\frac{1}{2}} B n − 2 1 remained open. For bilinear Calderón–Zygmund operators with sufficiently smooth kernels, Grafakos and Torres [18 ] established weak-type ( 1 , 1 ; 1 2 ) (1,1;\frac{1}{2}) ( 1 , 1 ; 2 1 ) estimates. However, such estimates need not hold for rough kernels. Recently, Honzík [20 ] showed that, for every n ≥ 1 n\ge1 n ≥ 1 , there exist bounded mean-zero angular functions on S 2 n − 1 \mathbb{S}^{2n-1} S 2 n − 1 for which the associated bilinear rough singular integral operators fail to map L 1 × L 1 L^1\times L^1 L 1 × L 1 into L 1 2 , ∞ L^{\frac{1}{2},\infty} L 2 1 , ∞ .
Another such contrast between linear and bilinear operators appears in the study of lacunary spherical maximal functions. In the linear case, the lacunary spherical maximal function is bounded on L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) for all p > 1 p>1 p > 1 and n ≥ 2 n\ge2 n ≥ 2 . Moreover, Christ [7 ] proved that it maps H 1 ( R n ) H^1(\mathbb{R}^n) H 1 ( R n ) into L 1 , ∞ ( R n ) L^{1,\infty}(\mathbb{R}^n) L 1 , ∞ ( R n ) for every n ≥ 2 n\ge2 n ≥ 2 . However, Lee, Li, Wen and Wu [26 ] recently showed that the bilinear lacunary spherical maximal function fails to map H 1 ( R n ) × H 1 ( R n ) H^1(\mathbb{R}^n)\times H^1(\mathbb{R}^n) H 1 ( R n ) × H 1 ( R n ) into L 1 2 , ∞ ( R n ) L^{\frac{1}{2},\infty}(\mathbb{R}^n) L 2 1 , ∞ ( R n ) for every n ≥ 1 n\ge1 n ≥ 1 . Yet another instance is the recent work by Alves and Grafakos [1 ] showing that the bilinear Hilbert transform fails to satisfy the restricted weak-type ( 1 , 1 ; 1 2 ) (1,1;\frac{1}{2}) ( 1 , 1 ; 2 1 ) estimate.
These results show that weak-type endpoint estimates for linear operators need not have their bilinear counterparts. In particular, the counterexample of Honzík [20 ] is closely related to our problem, since both kernels have the same decay of O ( ∣ ( y , z ) ∣ − 2 n ) O(|(y,z)|^{-2n}) O ( ∣ ( y , z ) ∣ − 2 n ) . The main idea of his counterexample is to choose the angular function so that contributions from different spatial scales have the same sign. On the other hand, the bilinear Bochner–Riesz kernel has a fixed radial oscillation. Indeed, at the critical index, the asymptotic expansion of the Bessel function gives
K n − 1 2 ( y , z ) = c + e 2 π i ∣ ( y , z ) ∣ + c − e − 2 π i ∣ ( y , z ) ∣ ∣ ( y , z ) ∣ 2 n + O ( ∣ ( y , z ) ∣ − 2 n − 1 ) (1) K^{n-\frac{1}{2}}(y,z)=\frac{c_+e^{2\pi i|(y,z)|}+c_-e^{-2\pi i|(y,z)|}}{|(y,z)|^{2n}}+O(|(y,z)|^{-2n-1}) \tag*{(1)} K n − 2 1 ( y , z ) = ∣ ( y , z ) ∣ 2 n c + e 2 π i ∣ ( y , z ) ∣ + c − e − 2 π i ∣ ( y , z ) ∣ + O ( ∣ ( y , z ) ∣ − 2 n − 1 ) (1)
as ∣ ( y , z ) ∣ → ∞ |(y,z)|\to\infty ∣ ( y , z ) ∣ → ∞ , where c + c_+ c + and c − c_- c − are constants. Although the kernel has the same decay as a bilinear rough singular integral kernel, its radial oscillation provides additional cancellation. By exploiting this oscillation, we prove that the bilinear Bochner–Riesz operator at the critical index satisfies the weak-type ( 1 , 1 ; 1 2 ) (1,1;\frac{1}{2}) ( 1 , 1 ; 2 1 ) estimate.
Theorem 1.1. Let n ≥ 1 n\ge1 n ≥ 1 . Then the bilinear Bochner–Riesz operator B n − 1 2 \mathcal{B}^{n-\frac{1}{2}} B n − 2 1 is bounded from L 1 ( R n ) × L 1 ( R n ) L^1(\mathbb{R}^n)\times L^1(\mathbb{R}^n) L 1 ( R n ) × L 1 ( R n ) to L 1 2 , ∞ ( R n ) L^{\frac{1}{2},\infty}(\mathbb{R}^n) L 2 1 , ∞ ( R n ) . In particular,
∥ B n − 1 2 ( f , g ) ∥ L 1 2 , ∞ ( R n ) ≤ C n ∥ f ∥ L 1 ( R n ) ∥ g ∥ L 1 ( R n ) . \left\|\mathcal{B}^{n-\frac{1}{2}}(f,g)\right\|_{L^{\frac{1}{2},\infty}(\mathbb{R}^n)}\le C_n\|f\|_{L^1(\mathbb{R}^n)}\|g\|_{L^1(\mathbb{R}^n)}. B n − 2 1 ( f , g ) L 2 1 , ∞ ( R n ) ≤ C n ∥ f ∥ L 1 ( R n ) ∥ g ∥ L 1 ( R n ) .
Methodology
Our proof is motivated by Christ’s method for the linear Bochner–Riesz operator at the critical index [7 ] . The main idea is to combine the Calderón–Zygmund decomposition with estimates that use the oscillation of the kernel. We now briefly explain the main steps of the proof.
We first apply the Calderón–Zygmund decomposition to both f f f and g g g . The terms containing at least one good function are controlled using the mixed L 1 × L ∞ → L 1 L^1 \times L^\infty\to L^1 L 1 × L ∞ → L 1 estimates. Thus, the main difficulty is to estimate the term containing two bad functions.
To estimate this term, we decompose the Bochner–Riesz kernel into dyadic annuli. We then group the bad functions according to the sizes of their supporting cubes relative to the kernel scale.
We next divide the bad–bad term into two parts, depending on whether the variables y y y and z z z are close to each other or well separated. For the part close to the diagonal y = z y=z y = z , we obtain an L 1 L^1 L 1 estimate using the support of the kernel and the disjointness of the bad cubes.
For the part away from the diagonal, we use the cancellation of the bad functions together with a T T ∗ TT^* T T ∗ argument. The main difficulty is that the bilinear kernel is defined on R 2 n \mathbb{R}^{2n} R 2 n , whereas the correlation kernels arising from the T T ∗ TT^* T T ∗ argument are obtained by integrating over R n \mathbb{R}^n R n . Thus, we cannot directly apply the corresponding estimates from the linear case.
To overcome this difficulty, we prove oscillatory estimates for these correlation kernels. When n ≥ 2 n \ge2 n ≥ 2 , we use polar coordinates and angular stationary phase, following the ideas of Christ [7 ] . When n = 1 n=1 n = 1 , we instead use the first and second derivative forms of van der Corput’s lemma. These estimates give an L 2 L^2 L 2 bound with decay in the relative scale.
Finally, we combine the L 1 L^1 L 1 estimate for the part close to the diagonal with the L 2 L^2 L 2 estimate for the remaining part. The decay obtained in the L 2 L^2 L 2 estimate allows us to sum over the relative scales.
Organization of the paper. In § 2 , we apply the Calderón–Zygmund decomposition and reduce the proof of Theorem 1.1 to estimating the bad–bad term. In § 3 , we decompose this term into two parts and reduce the desired estimate to an L 1 L^1 L 1 estimate (Proposition 3.1) and an L 2 L^2 L 2 estimate (Proposition 3.2). In § 4 and § 5 , we prove these two estimates, respectively. Finally, in § 6 , we prove the oscillatory kernel estimates needed for the L 2 L^2 L 2 argument.
Calderón–Zygmund decomposition
By homogeneity, it is enough to consider ∥ f ∥ 1 = ∥ g ∥ 1 = 1 \|f\|_1=\|g\|_1=1 ∥ f ∥ 1 = ∥ g ∥ 1 = 1 . Let λ > 0 \lambda>0 λ > 0 . We apply the dyadic Calderón–Zygmund decomposition to f f f and g g g at height λ 1 2 \lambda^{\frac12} λ 2 1 to get
f = f 0 + b , b = ∑ Q b Q , and g = g 0 + c , c = ∑ P c P . f=f_0+b,\qquad b=\sum_Q b_Q,\qquad\text{and}\qquad g=g_0+c,\qquad c=\sum_P c_P. f = f 0 + b , b = Q ∑ b Q , and g = g 0 + c , c = P ∑ c P .
where the cubes Q Q Q (respectively P P P ) are pairwise disjoint and satisfy the following properties.
∑ Q ∣ Q ∣ + ∑ P ∣ P ∣ ≲ λ − 1 2 , \sum_Q |Q|+\sum_P |P|\lesssim\lambda^{-\frac12}, Q ∑ ∣ Q ∣ + P ∑ ∣ P ∣ ≲ λ − 2 1 ,
supp b Q ⊂ Q , ∫ b Q = 0 , ∥ b Q ∥ 1 ≲ λ 1 2 ∣ Q ∣ , \operatorname{supp} b_Q\subset Q,\qquad\int b_Q=0,\qquad\|b_Q\|_1\lesssim\lambda^{\frac12}|Q|, supp b Q ⊂ Q , ∫ b Q = 0 , ∥ b Q ∥ 1 ≲ λ 2 1 ∣ Q ∣ ,
supp c P ⊂ P , ∫ c P = 0 , ∥ c P ∥ 1 ≲ λ 1 2 ∣ P ∣ , \operatorname{supp} c_P\subset P,\qquad\int c_P=0,\qquad\|c_P\|_1\lesssim\lambda^{\frac12}|P|, supp c P ⊂ P , ∫ c P = 0 , ∥ c P ∥ 1 ≲ λ 2 1 ∣ P ∣ ,
∥ f 0 ∥ ∞ + ∥ g 0 ∥ ∞ ≲ λ 1 2 , \|f_0\|_\infty+\|g_0\|_\infty\lesssim\lambda^{\frac12}, ∥ f 0 ∥ ∞ + ∥ g 0 ∥ ∞ ≲ λ 2 1 ,
∥ f 0 ∥ 2 2 + ∥ g 0 ∥ 2 2 ≲ λ 1 2 . \|f_0\|_2^2+\|g_0\|_2^2\lesssim\lambda^{\frac12}. ∥ f 0 ∥ 2 2 + ∥ g 0 ∥ 2 2 ≲ λ 2 1 .
We can organize the bad functions based on the length of the cubes. For k ∈ Z k \in\mathbb{Z} k ∈ Z , define
b k = ∑ ℓ ( Q ) = 2 k b Q , c k = ∑ ℓ ( P ) = 2 k c P . b_k=\sum_{\ell(Q)=2^k} b_Q,\qquad c_k=\sum_{\ell(P)=2^k} c_P. b k = ℓ ( Q ) = 2 k ∑ b Q , c k = ℓ ( P ) = 2 k ∑ c P .
where ℓ ( Q ) \ell(Q) ℓ ( Q ) is the side length of Q Q Q . Consider the set
E = ⋃ Q 5 Q ∪ ⋃ P 5 P . E=\bigcup_Q 5Q\cup\bigcup_P 5P. E = Q ⋃ 5 Q ∪ P ⋃ 5 P .
Then, from the above conditions, we get
∣ E ∣ ≲ λ − 1 2 . (2) |E|\lesssim\lambda^{-\frac{1}{2}}. \tag*{(2)} ∣ E ∣ ≲ λ − 2 1 . (2)
By Calderón–Zygmund decomposition, we have
∣ { x : ∣ B n − 1 2 ( f , g ) ( x ) ∣ > λ } ∣ ≤ ∣ E ∣ + ∣ { x ∉ E : ∣ B n − 1 2 ( f 0 , g 0 ) ( x ) ∣ > λ 4 } ∣ + ∣ { x ∉ E : ∣ B n − 1 2 ( b , g 0 ) ( x ) ∣ > λ 4 } ∣ + ∣ { x ∉ E : ∣ B n − 1 2 ( f 0 , c ) ( x ) ∣ > λ 4 } ∣ + ∣ { x ∉ E : ∣ B n − 1 2 ( b , c ) ( x ) ∣ > λ 4 } ∣ . \begin{aligned}
\left|\left\{x: \left|\mathcal{B}^{n-\frac{1}{2}}(f,g)(x)\right|>\lambda\right\}\right|
\leq{}& |E|+\left|\left\{x\notin E:\left|\mathcal{B}^{n-\frac{1}{2}}(f_0,g_0)(x)\right|>\frac{\lambda}{4}\right\}\right|\\
&+\left|\left\{x\notin E:\left|\mathcal{B}^{n-\frac{1}{2}}(b,g_0)(x)\right|>\frac{\lambda}{4}\right\}\right|\\
&+\left|\left\{x\notin E:\left|\mathcal{B}^{n-\frac{1}{2}}(f_0,c)(x)\right|>\frac{\lambda}{4}\right\}\right|\\
&+\left|\left\{x\notin E:\left|\mathcal{B}^{n-\frac{1}{2}}(b,c)(x)\right|>\frac{\lambda}{4}\right\}\right|.
\end{aligned} { x : B n − 2 1 ( f , g ) ( x ) > λ } ≤ ∣ E ∣ + { x ∈ / E : B n − 2 1 ( f 0 , g 0 ) ( x ) > 4 λ } + { x ∈ / E : B n − 2 1 ( b , g 0 ) ( x ) > 4 λ } + { x ∈ / E : B n − 2 1 ( f 0 , c ) ( x ) > 4 λ } + { x ∈ / E : B n − 2 1 ( b , c ) ( x ) > 4 λ } .
We first note that B n − 1 2 \mathcal{B}^{n-\frac{1}{2}} B n − 2 1 is bounded from L 1 ( R n ) × L ∞ ( R n ) L^1(\mathbb{R}^n)\times L^\infty(\mathbb{R}^n) L 1 ( R n ) × L ∞ ( R n ) into L 1 ( R n ) L^1(\mathbb{R}^n) L 1 ( R n ) for every n ≥ 1 n\geq1 n ≥ 1 . This follows from [3 Theorem 4.1] when n = 1 n=1 n = 1 and [3 Theorem 4.9] when n ≥ 2 n\geq2 n ≥ 2 . Using this estimate and Chebyshev’s inequality, we control all the terms containing at least one good function in the Calderón–Zygmund decomposition.
Good–good term. By L 1 × L ∞ → L 1 L^1\times L^\infty\to L^1 L 1 × L ∞ → L 1 -bound for B n − 1 2 \mathcal{B}^{n-\frac{1}{2}} B n − 2 1 and Chebyshev’s inequality, we get
∣ { x : ∣ B n − 1 2 ( f 0 , g 0 ) ( x ) ∣ > λ 4 } ∣ ≲ λ − 1 ∥ f 0 ∥ 1 ∥ g 0 ∥ ∞ ≲ λ − 1 2 . (3) \left|\left\{x:\left|\mathcal{B}^{n-\frac{1}{2}}(f_0,g_0)(x)\right|>\frac{\lambda}{4}\right\}\right|\lesssim\lambda^{-1}\|f_0\|_1\|g_0\|_\infty\lesssim\lambda^{-\frac{1}{2}}. \tag*{(3)} { x : B n − 2 1 ( f 0 , g 0 ) ( x ) > 4 λ } ≲ λ − 1 ∥ f 0 ∥ 1 ∥ g 0 ∥ ∞ ≲ λ − 2 1 . (3)
Good–bad terms. Again by L 1 × L ∞ → L 1 L^1\times L^\infty\to L^1 L 1 × L ∞ → L 1 -bound for B n − 1 2 \mathcal{B}^{n-\frac{1}{2}} B n − 2 1 and Chebyshev’s inequality, we get
∣ { x : ∣ B n − 1 2 ( b , g 0 ) ( x ) ∣ > λ 4 } ∣ ≲ λ − 1 ∥ b ∥ 1 ∥ g 0 ∥ ∞ ≲ λ − 1 2 . (4) \left|\left\{x:\left|\mathcal{B}^{n-\frac{1}{2}}(b,g_0)(x)\right|>\frac{\lambda}{4}\right\}\right|\lesssim\lambda^{-1}\|b\|_1\|g_0\|_\infty\lesssim\lambda^{-\frac{1}{2}}. \tag*{(4)} { x : B n − 2 1 ( b , g 0 ) ( x ) > 4 λ } ≲ λ − 1 ∥ b ∥ 1 ∥ g 0 ∥ ∞ ≲ λ − 2 1 . (4)
Similarly, the L ∞ × L 1 → L 1 L^\infty\times L^1\to L^1 L ∞ × L 1 → L 1 -bound for B n − 1 2 \mathcal{B}^{n-\frac{1}{2}} B n − 2 1 together with Chebyshev’s inequality gives
∣ { x : ∣ B n − 1 2 ( f 0 , c ) ( x ) ∣ > λ 4 } ∣ ≲ λ − 1 ∥ f 0 ∥ ∞ ∥ c ∥ 1 ≲ λ − 1 2 . (5) \left|\left\{x:\left|\mathcal{B}^{n-\frac{1}{2}}(f_0,c)(x)\right|>\frac{\lambda}{4}\right\}\right|\lesssim\lambda^{-1}\|f_0\|_\infty\|c\|_1\lesssim\lambda^{-\frac{1}{2}}. \tag*{(5)} { x : B n − 2 1 ( f 0 , c ) ( x ) > 4 λ } ≲ λ − 1 ∥ f 0 ∥ ∞ ∥ c ∥ 1 ≲ λ − 2 1 . (5)
Bad–bad term. Recall that the kernel of B n − 1 2 \mathcal{B}^{n-\frac{1}{2}} B n − 2 1 is given by
K n − 1 2 ( y , z ) = Γ ( n + 1 2 ) J 2 n − 1 2 ( 2 π ∣ ( y , z ) ∣ ) π n − 1 2 ∣ ( y , z ) ∣ 2 n − 1 2 . K^{n-\frac{1}{2}}(y,z)=\frac{\Gamma\left(n+\frac{1}{2}\right)J_{2n-\frac{1}{2}}\left(2\pi|(y,z)|\right)}{\pi^{n-\frac{1}{2}}|(y,z)|^{2n-\frac{1}{2}}}. K n − 2 1 ( y , z ) = π n − 2 1 ∣ ( y , z ) ∣ 2 n − 2 1 Γ ( n + 2 1 ) J 2 n − 2 1 ( 2 π ∣ ( y , z ) ∣ ) .
From the derivative properties of Bessel functions, we have for every positive integer N N N
∣ ∂ γ K ( y , z ) ∣ ≲ ∣ ( y , z ) ∣ − 2 n , ∣ ( y , z ) ∣ ≥ 1 , ∣ γ ∣ ≤ N . (6) |\partial^\gamma K(y,z)|\lesssim|(y,z)|^{-2n},\qquad|(y,z)|\geq1,\quad|\gamma|\leq N. \tag*{(6)} ∣ ∂ γ K ( y , z ) ∣ ≲ ∣ ( y , z ) ∣ − 2 n , ∣ ( y , z ) ∣ ≥ 1 , ∣ γ ∣ ≤ N . (6)
Let ψ 0 ∈ C ∞ ( R 2 n ) \psi_0\in C^\infty(\mathbb{R}^{2n}) ψ 0 ∈ C ∞ ( R 2 n ) be a radial function such that 0 ≤ ψ 0 ≤ 1 0\leq\psi_0\leq1 0 ≤ ψ 0 ≤ 1 , ψ 0 ( y , z ) = 1 \psi_0(y,z)=1 ψ 0 ( y , z ) = 1 when ∣ ( y , z ) ∣ ≤ 1 |(y,z)|\leq1 ∣ ( y , z ) ∣ ≤ 1 and supp ψ 0 ⊂ { ∣ ( y , z ) ∣ ≤ 2 } \operatorname{supp}\psi_0\subset\{|(y,z)|\leq2\} supp ψ 0 ⊂ { ∣ ( y , z ) ∣ ≤ 2 } . Define ψ ( y , z ) = ψ 0 ( y , z ) − ψ 0 ( 2 ( y , z ) ) \psi(y,z)=\psi_0(y,z)-\psi_0(2(y,z)) ψ ( y , z ) = ψ 0 ( y , z ) − ψ 0 ( 2 ( y , z )) . Then
ψ 0 ( y , z ) + ∑ j ≥ 1 ψ ( 2 − j ( y , z ) ) = 1. \psi_0(y,z)+\sum_{j\geq1}\psi(2^{-j}(y,z))=1. ψ 0 ( y , z ) + j ≥ 1 ∑ ψ ( 2 − j ( y , z )) = 1.
Write K 0 = ψ 0 K K_0=\psi_0K K 0 = ψ 0 K , K j ( y , z ) = ψ ( 2 − j ( y , z ) ) K ( y , z ) K_j(y,z)=\psi(2^{-j}(y,z))K(y,z) K j ( y , z ) = ψ ( 2 − j ( y , z )) K ( y , z ) for j ≥ 1 j\geq1 j ≥ 1 . Let T j T_j T j denote the bilinear operator associated with the kernel K j K_j K j ,
T j ( f , g ) ( x ) = ∫ R n ∫ R n K j ( x − y , x − z ) f ( y ) g ( z ) d y d z . T_j(f,g)(x)=\int_{\mathbb{R}^n}\int_{\mathbb{R}^n}K_j(x-y,x-z)f(y)g(z)\,dy\,dz. T j ( f , g ) ( x ) = ∫ R n ∫ R n K j ( x − y , x − z ) f ( y ) g ( z ) d y d z .
Using the support of ψ \psi ψ , (1) and (6) , we get for every j ≥ 1 j \ge1 j ≥ 1
supp K j ⊂ { 2 j − 1 ≤ ∣ X ∣ ≤ 2 j + 1 } , \operatorname{supp} K_j \subset\{2^{j-1} \le|X| \le2^{j+1}\}, supp K j ⊂ { 2 j − 1 ≤ ∣ X ∣ ≤ 2 j + 1 } ,
∥ ∂ γ K j ∥ ∞ ≲ γ 2 − 2 n j , ∣ γ ∣ ≤ N , \|\partial^\gamma K_j\|_\infty\lesssim_\gamma2^{-2nj}, \qquad|\gamma| \le N, ∥ ∂ γ K j ∥ ∞ ≲ γ 2 − 2 nj , ∣ γ ∣ ≤ N ,
sup y , z ∈ R n ∫ R n ∣ ∂ γ K j ( x − y , x − z ) ∣ d x ≲ γ 2 − n j , ∣ γ ∣ ≤ N . (7) \sup_{y,z\in\mathbb{R}^n}\int_{\mathbb{R}^n}|\partial^\gamma K_j(x-y,x-z)|\,dx \lesssim_\gamma2^{-nj}, \qquad|\gamma| \le N. \tag*{(7)} y , z ∈ R n sup ∫ R n ∣ ∂ γ K j ( x − y , x − z ) ∣ d x ≲ γ 2 − nj , ∣ γ ∣ ≤ N . (7)
Since K 0 K_0 K 0 is bounded, it is easy to see that the operator T 0 T_0 T 0 maps L p 1 × L p 2 L^{p_1}\times L^{p_2} L p 1 × L p 2 to L p L^p L p for all 1 ≤ p 1 , p 2 ≤ ∞ 1\le p_1,p_2\le\infty 1 ≤ p 1 , p 2 ≤ ∞ satisfying 1 p = 1 p 1 + 1 p 2 \frac{1}{p}=\frac{1}{p_1}+\frac{1}{p_2} p 1 = p 1 1 + p 2 1 . We claim the following estimate holds for ∑ j ≥ 1 T j \sum_{j\ge1}T_j ∑ j ≥ 1 T j .
∣ { x ∉ E : ∣ ∑ j ≥ 1 T j ( b , c ) ( x ) ∣ > λ 8 } ∣ ≲ λ − 1 2 . (8) \left|\left\{x\notin E:\left|\sum_{j\ge1}T_j(b,c)(x)\right|>\frac{\lambda}{8}\right\}\right|\lesssim\lambda^{-\frac12}. \tag*{(8)} { x ∈ / E : j ≥ 1 ∑ T j ( b , c ) ( x ) > 8 λ } ≲ λ − 2 1 . (8)
If we assume (8) , then together with the estimates of T 0 T_0 T 0 , we have
∣ { x ∉ E : ∣ B n − 1 2 ( b , c ) ( x ) ∣ > λ 4 } ∣ ≲ λ − 1 2 . (9) \left|\left\{x\notin E:\left|\mathcal{B}^{n-\frac12}(b,c)(x)\right|>\frac{\lambda}{4}\right\}\right|\lesssim\lambda^{-\frac12}. \tag*{(9)} { x ∈ / E : B n − 2 1 ( b , c ) ( x ) > 4 λ } ≲ λ − 2 1 . (9)
Combining the estimates in (2) , (3) , (4) , (5) and (9) proves the Theorem 1.1. Therefore, we need to prove the claim (8) .
Decomposition of the bad–bad term
We now prove (8) for the bad–bad term. We first organize the bad functions according to the sizes of their supporting cubes relative to the scale of K j K_j K j . We then divide the interactions into two parts, depending on whether the variables y y y and z z z are close to each other or well separated.
Fix j ≥ 1 j\ge1 j ≥ 1 and x ∉ E x\notin E x ∈ / E . Suppose that y ∈ Q y\in Q y ∈ Q , z ∈ P z\in P z ∈ P and K j ( x − y , x − z ) ≠ 0 K_j(x-y,x-z)\ne0 K j ( x − y , x − z ) = 0 . Since x ∉ 5 Q ∪ 5 P x\notin5Q\cup5P x ∈ / 5 Q ∪ 5 P , we have
∣ x − y ∣ > 2 ℓ ( Q ) , ∣ x − z ∣ > 2 ℓ ( P ) , |x-y|>2\ell(Q), \qquad|x-z|>2\ell(P), ∣ x − y ∣ > 2 ℓ ( Q ) , ∣ x − z ∣ > 2 ℓ ( P ) ,
whereas ∣ x − y ∣ , ∣ x − z ∣ ≤ 2 j + 1 |x-y|,|x-z|\le2^{j+1} ∣ x − y ∣ , ∣ x − z ∣ ≤ 2 j + 1 . Consequently, we obtain ℓ ( Q ) < 2 j \ell(Q)<2^j ℓ ( Q ) < 2 j and ℓ ( P ) < 2 j \ell(P)<2^j ℓ ( P ) < 2 j . Thus we only need to consider cubes of the length ℓ ( Q ) = 2 j − s \ell(Q)=2^{j-s} ℓ ( Q ) = 2 j − s and ℓ ( P ) = 2 j − t \ell(P)=2^{j-t} ℓ ( P ) = 2 j − t for s , t ≥ 0 s,t\ge0 s , t ≥ 0 . Using this observation, we can rewrite the operator ∑ j ≥ 1 T j \sum_{j\ge1}T_j ∑ j ≥ 1 T j as
∑ j ≥ 1 T j ( b , c ) ( x ) = ∑ s , t ≥ 0 H s , t ( b , c ) ( x ) , (10) \sum_{j\ge1}T_j(b,c)(x)=\sum_{s,t\ge0}H_{s,t}(b,c)(x), \tag*{(10)} j ≥ 1 ∑ T j ( b , c ) ( x ) = s , t ≥ 0 ∑ H s , t ( b , c ) ( x ) , (10)
where
H s , t ( b , c ) ( x ) = ∑ j ≥ 1 T j ( b j − s , c j − t ) ( x ) . H_{s,t}(b,c)(x)=\sum_{j\ge1}T_j(b_{j-s},c_{j-t})(x). H s , t ( b , c ) ( x ) = j ≥ 1 ∑ T j ( b j − s , c j − t ) ( x ) .
Note that the function H s , t H_{s,t} H s , t is symmetric in the sense that H s , t ( b , c ) = H t , s ( c , b ) H_{s,t}(b,c)=H_{t,s}(c,b) H s , t ( b , c ) = H t , s ( c , b ) . Thus, it is enough to estimate ∑ s ≥ t ≥ 0 H s , t ( b , c ) \sum_{s\ge t\ge0}H_{s,t}(b,c) ∑ s ≥ t ≥ 0 H s , t ( b , c ) .
Further, observe that K j ( x − y , x − z ) ≠ 0 K_j(x-y,x-z)\ne0 K j ( x − y , x − z ) = 0 implies ∣ y − z ∣ ≤ 2 j + 2 |y-z|\le2^{j+2} ∣ y − z ∣ ≤ 2 j + 2 . Therefore, for fixed s , t , j s,t,j s , t , j , it is enough to consider pairs of bad cubes satisfying dist ( Q , P ) ≤ 2 j + 2 \operatorname{dist}(Q,P)\le2^{j+2} dist ( Q , P ) ≤ 2 j + 2 . Define
F j ( y , z ) = ∑ ℓ ( Q ) = 2 j − s , ℓ ( P ) = 2 j − t dist ( Q , P ) ≤ 2 j + 2 b Q ( y ) c P ( z ) . (11) F_j(y,z)=\sum_{\substack{\ell(Q)=2^{j-s},\ \ell(P)=2^{j-t}\\ \operatorname{dist}(Q,P)\le2^{j+2}}}b_Q(y)c_P(z). \tag*{(11)} F j ( y , z ) = ℓ ( Q ) = 2 j − s , ℓ ( P ) = 2 j − t dist ( Q , P ) ≤ 2 j + 2 ∑ b Q ( y ) c P ( z ) . (11)
Then, for x ∉ E x\notin E x ∈ / E , we have
T j ( b j − s , c j − t ) ( x ) = ∬ K j ( x − y , x − z ) F j ( y , z ) d y d z . (12) T_j(b_{j-s},c_{j-t})(x)=\iint K_j(x-y,x-z)F_j(y,z)\,dy\,dz. \tag*{(12)} T j ( b j − s , c j − t ) ( x ) = ∬ K j ( x − y , x − z ) F j ( y , z ) d y d z . (12)
Since each pair ( Q , P ) (Q,P) ( Q , P ) in F j ( y , z ) F_j(y,z) F j ( y , z ) occurs once and the rectangles Q × P Q\times P Q × P are pairwise disjoint in R 2 n \mathbb{R}^{2n} R 2 n . Thus for y ∈ Q y\in Q y ∈ Q and z ∈ P z\in P z ∈ P , we have
∣ y − z ∣ ≤ dist ( Q , P ) + n ℓ ( Q ) + n ℓ ( P ) ≤ 2 j + 2 + n 2 j − s + n 2 j − t ≤ ( 2 n + 4 ) 2 j . |y-z|\le\operatorname{dist}(Q,P)+\sqrt{n}\ell(Q)+\sqrt{n}\ell(P)\le2^{j+2}+\sqrt{n}2^{j-s}+\sqrt{n}2^{j-t}\le(2n+4)2^j. ∣ y − z ∣ ≤ dist ( Q , P ) + n ℓ ( Q ) + n ℓ ( P ) ≤ 2 j + 2 + n 2 j − s + n 2 j − t ≤ ( 2 n + 4 ) 2 j .
Consequently, we obtain
∣ F j ( y , z ) ∣ ≤ ∣ b j − s ( y ) ∣ ∣ c j − t ( z ) ∣ 1 { ∣ y − z ∣ ≤ ( 2 n + 4 ) 2 j } . (13) |F_j(y,z)| \leq|b_{j-s}(y)||c_{j-t}(z)|\mathbf{1}_{\{|y-z|\leq(2n+4)2^j\}}. \tag*{(13)} ∣ F j ( y , z ) ∣ ≤ ∣ b j − s ( y ) ∣∣ c j − t ( z ) ∣ 1 { ∣ y − z ∣ ≤ ( 2 n + 4 ) 2 j } . (13)
Let φ ∈ C ∞ ( R n ) \varphi\in C^\infty(\mathbb{R}^n) φ ∈ C ∞ ( R n ) such that 0 ≤ φ ≤ 1 0 \leq\varphi\leq1 0 ≤ φ ≤ 1 , φ = 1 \varphi= 1 φ = 1 on B ( 0 , 1 ) B(0,1) B ( 0 , 1 ) and supp φ ⊂ B ( 0 , 2 ) \operatorname{supp}\varphi\subset B(0,2) supp φ ⊂ B ( 0 , 2 ) . For every Q Q Q with ℓ ( Q ) = 2 j − s \ell(Q)=2^{j-s} ℓ ( Q ) = 2 j − s , let h j = 2 j − s 4 h_j=2^{j-\frac{s}{4}} h j = 2 j − 4 s and define
φ Q , j ( z ) = φ ( z − ρ Q ( n + 1 ) h j ) , \varphi_{Q,j}(z)=\varphi\left(\frac{z-\rho_Q}{(n+1)h_j}\right), φ Q , j ( z ) = φ ( ( n + 1 ) h j z − ρ Q ) ,
where ρ Q \rho_Q ρ Q denotes the center of Q Q Q . Note that φ Q , j ( z ) ≠ 0 \varphi_{Q,j}(z)\ne0 φ Q , j ( z ) = 0 implies ∣ z − ρ Q ∣ ≤ 2 ( n + 1 ) h j |z-\rho_Q|\leq2(n+1)h_j ∣ z − ρ Q ∣ ≤ 2 ( n + 1 ) h j whereas φ Q , j ( z ) ≠ 1 \varphi_{Q,j}(z)\ne1 φ Q , j ( z ) = 1 implies ∣ z − ρ Q ∣ ≥ ( n + 1 ) h j |z-\rho_Q|\geq(n+1)h_j ∣ z − ρ Q ∣ ≥ ( n + 1 ) h j .
We decompose F j = F j 1 + F j 2 F_j=F_j^1+F_j^2 F j = F j 1 + F j 2 , where
F j 1 ( y , z ) = ∑ ( Q , P ) b Q ( y ) c P ( z ) φ Q , j ( z ) , F_j^1(y,z)=\sum_{(Q,P)}b_Q(y)c_P(z)\varphi_{Q,j}(z), F j 1 ( y , z ) = ( Q , P ) ∑ b Q ( y ) c P ( z ) φ Q , j ( z ) ,
F j 2 ( y , z ) = ∑ ( Q , P ) b Q ( y ) c P ( z ) ( 1 − φ Q , j ( z ) ) . F_j^2(y,z)=\sum_{(Q,P)}b_Q(y)c_P(z)(1-\varphi_{Q,j}(z)). F j 2 ( y , z ) = ( Q , P ) ∑ b Q ( y ) c P ( z ) ( 1 − φ Q , j ( z )) .
Consequently, we get that
H s , t ( b , c ) = H s , t 1 ( b , c ) + H s , t 2 ( b , c ) . H_{s,t}(b,c)=H_{s,t}^1(b,c)+H_{s,t}^2(b,c). H s , t ( b , c ) = H s , t 1 ( b , c ) + H s , t 2 ( b , c ) .
Since 0 ≤ φ ≤ 1 0\leq\varphi\leq1 0 ≤ φ ≤ 1 , the estimate in (13) continues to hold for both the pieces F j 1 F_j^1 F j 1 and F j 2 F_j^2 F j 2 .
We prove the following estimates for H s , t 1 ( b , c ) H_{s,t}^1(b,c) H s , t 1 ( b , c ) and H s , t 2 ( b , c ) H_{s,t}^2(b,c) H s , t 2 ( b , c ) .
Proposition 3.1. For s ≥ t ≥ 0 s\geq t\geq0 s ≥ t ≥ 0 , we have
∥ H s , t 1 ( b , c ) ∥ L 1 ( E c ) ≲ 2 − s 4 λ 1 2 . \|H_{s,t}^1(b,c)\|_{L^1(E^c)}\lesssim2^{-\frac{s}{4}}\lambda^{\frac{1}{2}}. ∥ H s , t 1 ( b , c ) ∥ L 1 ( E c ) ≲ 2 − 4 s λ 2 1 .
Proposition 3.2. For s ≥ t ≥ 0 s\geq t\geq0 s ≥ t ≥ 0 , we have
∥ H s , t 2 ( b , c ) ∥ 2 ≲ 2 − s 8 λ 3 4 . \|H_{s,t}^2(b,c)\|_2\lesssim2^{-\frac{s}{8}}\lambda^{\frac{3}{4}}. ∥ H s , t 2 ( b , c ) ∥ 2 ≲ 2 − 8 s λ 4 3 .
We assume both the results above for the moment. Then
∣ { x ∉ E : ∣ ∑ j ≥ 1 T j ( b , c ) ( x ) ∣ > λ 8 } ∣ = ∣ { x ∉ E : ∣ ∑ s , t ≥ 0 H s , t ( b , c ) ( x ) ∣ > λ 8 } ∣ ≲ ∣ { x ∉ E : ∣ ∑ s ≥ t ≥ 0 H s , t ( b , c ) ( x ) ∣ > λ 16 } ∣ + ∣ { x ∉ E : ∣ ∑ s ≥ t ≥ 0 H s , t ( c , b ) ( x ) ∣ > λ 16 } ∣ ≲ ∣ { x ∉ E : ∣ ∑ s ≥ t ≥ 0 H s , t 1 ( b , c ) ( x ) ∣ > λ 32 } ∣ + ∣ { x : ∣ ∑ s ≥ t ≥ 0 H s , t 2 ( b , c ) ( x ) ∣ > λ 32 } ∣ + ∣ { x ∉ E : ∣ ∑ s ≥ t ≥ 0 H s , t 1 ( c , b ) ( x ) ∣ > λ 32 } ∣ + ∣ { x : ∣ ∑ s ≥ t ≥ 0 H s , t 2 ( c , b ) ( x ) ∣ > λ 32 } ∣ ≲ ∑ s ≥ t ≥ 0 ( ∣ { x ∉ E : ∣ H s , t 1 ( b , c ) ( x ) ∣ > λ s , t } ∣ + ∣ { x : ∣ H s , t 2 ( b , c ) ( x ) ∣ > λ s , t } ∣ ) + ∑ s ≥ t ≥ 0 ( ∣ { x ∉ E : ∣ H s , t 1 ( c , b ) ( x ) ∣ > λ s , t } ∣ + ∣ { x : ∣ H s , t 2 ( c , b ) ( x ) ∣ > λ s , t } ∣ ) , \begin{aligned}
\left|\left\{x\notin E:\left|\sum_{j\geq1}T_j(b,c)(x)\right|>\frac{\lambda}{8}\right\}\right|
&=\left|\left\{x\notin E:\left|\sum_{s,t\geq0}H_{s,t}(b,c)(x)\right|>\frac{\lambda}{8}\right\}\right|\\
&\lesssim\left|\left\{x\notin E:\left|\sum_{s\geq t\geq0}H_{s,t}(b,c)(x)\right|>\frac{\lambda}{16}\right\}\right|\\
&\quad+\left|\left\{x\notin E:\left|\sum_{s\geq t\geq0}H_{s,t}(c,b)(x)\right|>\frac{\lambda}{16}\right\}\right|\\
&\lesssim\left|\left\{x\notin E:\left|\sum_{s\geq t\geq0}H_{s,t}^1(b,c)(x)\right|>\frac{\lambda}{32}\right\}\right|\\
&\quad+\left|\left\{x:\left|\sum_{s\geq t\geq0}H_{s,t}^2(b,c)(x)\right|>\frac{\lambda}{32}\right\}\right|\\
&\quad+\left|\left\{x\notin E:\left|\sum_{s\geq t\geq0}H_{s,t}^1(c,b)(x)\right|>\frac{\lambda}{32}\right\}\right|\\
&\quad+\left|\left\{x:\left|\sum_{s\geq t\geq0}H_{s,t}^2(c,b)(x)\right|>\frac{\lambda}{32}\right\}\right|\\
&\lesssim\sum_{s\geq t\geq0}\left(\left|\{x\notin E:|H_{s,t}^1(b,c)(x)|>\lambda_{s,t}\}\right|+\left|\{x:|H_{s,t}^2(b,c)(x)|>\lambda_{s,t}\}\right|\right)\\
&\quad+\sum_{s\geq t\geq0}\left(\left|\{x\notin E:|H_{s,t}^1(c,b)(x)|>\lambda_{s,t}\}\right|+\left|\{x:|H_{s,t}^2(c,b)(x)|>\lambda_{s,t}\}\right|\right),
\end{aligned} { x ∈ / E : j ≥ 1 ∑ T j ( b , c ) ( x ) > 8 λ } = { x ∈ / E : s , t ≥ 0 ∑ H s , t ( b , c ) ( x ) > 8 λ } ≲ { x ∈ / E : s ≥ t ≥ 0 ∑ H s , t ( b , c ) ( x ) > 16 λ } + { x ∈ / E : s ≥ t ≥ 0 ∑ H s , t ( c , b ) ( x ) > 16 λ } ≲ { x ∈ / E : s ≥ t ≥ 0 ∑ H s , t 1 ( b , c ) ( x ) > 32 λ } + { x : s ≥ t ≥ 0 ∑ H s , t 2 ( b , c ) ( x ) > 32 λ } + { x ∈ / E : s ≥ t ≥ 0 ∑ H s , t 1 ( c , b ) ( x ) > 32 λ } + { x : s ≥ t ≥ 0 ∑ H s , t 2 ( c , b ) ( x ) > 32 λ } ≲ s ≥ t ≥ 0 ∑ ( { x ∈ / E : ∣ H s , t 1 ( b , c ) ( x ) ∣ > λ s , t } + { x : ∣ H s , t 2 ( b , c ) ( x ) ∣ > λ s , t } ) + s ≥ t ≥ 0 ∑ ( { x ∈ / E : ∣ H s , t 1 ( c , b ) ( x ) ∣ > λ s , t } + { x : ∣ H s , t 2 ( c , b ) ( x ) ∣ > λ s , t } ) ,
where λ s , t = c 0 λ 2 − s 16 \lambda_{s,t}=c_0\lambda2^{-\frac{s}{16}} λ s , t = c 0 λ 2 − 16 s with c 0 > 0 c_0>0 c 0 > 0 a fixed constant such that ∑ s ≥ t ≥ 0 λ s ≤ λ 16 \sum_{s\ge t\ge0}\lambda_s\le\frac{\lambda}{16} ∑ s ≥ t ≥ 0 λ s ≤ 16 λ . Applying Chebyshev's inequality along with estimates from Proposition 3.1 and Proposition 3.2, we get that
∣ { x : ∣ ∑ j ≥ 1 T j ( b , c ) ( x ) ∣ > λ 8 } ∣ ≲ ∑ s ≥ t ≥ 0 ( ∥ H s , t 1 ( b , c ) ∥ L 1 ( E c ) λ s , t + ∥ H s , t 2 ( b , c ) ∥ 2 2 λ s , t 2 ) + ∑ s ≥ t ≥ 0 ( ∥ H s , t 1 ( c , b ) ∥ L 1 ( E c ) λ s , t + ∥ H s , t 2 ( c , b ) ∥ 2 2 λ s , t 2 ) \begin{aligned}
\left|\left\{x:\left|\sum_{j\ge1}T_j(b,c)(x)\right|>\frac{\lambda}{8}\right\}\right|
\lesssim\sum_{s\ge t\ge0}\left(\frac{\left\|H^1_{s,t}(b,c)\right\|_{L^1(E^c)} }{\lambda_{s,t}}+\frac{\left\|H^2_{s,t}(b,c)\right\|_2^2}{\lambda_{s,t}^2}\right) \\
&+\sum_{s\ge t\ge0}\left(\frac{\left\|H^1_{s,t}(c,b)\right\|_{L^1(E^c)}}{\lambda_{s,t}}+\frac{\left\|H^2_{s,t}(c,b)\right\|_2^2}{\lambda_{s,t}^2}\right)
\end{aligned} { x : j ≥ 1 ∑ T j ( b , c ) ( x ) > 8 λ } ≲ s ≥ t ≥ 0 ∑ ( λ s , t H s , t 1 ( b , c ) L 1 ( E c ) + λ s , t 2 H s , t 2 ( b , c ) 2 2 ) + s ≥ t ≥ 0 ∑ ( λ s , t H s , t 1 ( c , b ) L 1 ( E c ) + λ s , t 2 H s , t 2 ( c , b ) 2 2 )
≲ ∑ s ≥ t ≥ 0 ( 2 − n s 4 λ 1 2 2 − s 16 λ + 2 − s 4 λ 3 2 2 − s 8 λ 2 ) \lesssim\sum_{s\ge t\ge0}\left(\frac{2^{-\frac{ns}{4}}\lambda^{\frac12}}{2^{-\frac{s}{16}}\lambda}+\frac{2^{-\frac{s}{4}}\lambda^{\frac32}}{2^{-\frac{s}{8}}\lambda^2}\right) ≲ s ≥ t ≥ 0 ∑ ( 2 − 16 s λ 2 − 4 n s λ 2 1 + 2 − 8 s λ 2 2 − 4 s λ 2 3 )
≲ λ − 1 2 . \lesssim\lambda^{-\frac12}. ≲ λ − 2 1 .
Therefore, the estimate in (2.7) is reduced to proving Propositions 3.1 and 3.2. We first prove Proposition 3.1.
The near-diagonal L 1 L^1 L 1 estimate
Proof of Proposition 3.1. Note that from (2.6), we have
∥ ∬ K j ( ⋅ − y , ⋅ − z ) F j 1 ( y , z ) d y d z ∥ 1 ≤ 2 − n j ∥ F j 1 ∥ 1 . (14) \left\|\iint K_j(\cdot-y,\cdot-z)F_j^1(y,z)\,dy\,dz\right\|_1\le2^{-nj}\|F_j^1\|_1. \tag*{(14)} ∬ K j ( ⋅ − y , ⋅ − z ) F j 1 ( y , z ) d y d z 1 ≤ 2 − nj ∥ F j 1 ∥ 1 . (14)
Therefore, we need to estimate the L 1 L^1 L 1 norm of F j 1 F_j^1 F j 1 . Using (3.4) and the definition of F j 1 F_j^1 F j 1 , we have
∥ F j 1 ∥ 1 ≤ ∑ ( Q , P ) ∬ Q × P ∣ b Q ( y ) ∣ ∣ c P ( z ) ∣ φ Q , j ( z ) d y d z ≤ ∑ ( Q , P ) ∥ b Q ∥ 1 ∫ P ∣ c P ( z ) ∣ φ Q , j ( z ) d z ≤ ∫ ∣ c j − t ( z ) ∣ ( ∑ Q ∥ b Q ∥ 1 φ Q , j ( z ) ) d z . \begin{aligned}
\|F_j^1\|_1\le\sum_{(Q,P)}\iint_{Q\times P}|b_Q(y)||c_P(z)|\varphi_{Q,j}(z)\,dy\,dz \\
&\le\sum_{(Q,P)}\|b_Q\|_1\int_P|c_P(z)|\varphi_{Q,j}(z)\,dz \\
&\le\int|c_{j-t}(z)|\left(\sum_Q\|b_Q\|_1\varphi_{Q,j}(z)\right)\,dz.
\end{aligned} ∥ F j 1 ∥ 1 ≤ ( Q , P ) ∑ ∬ Q × P ∣ b Q ( y ) ∣∣ c P ( z ) ∣ φ Q , j ( z ) d y d z ≤ ( Q , P ) ∑ ∥ b Q ∥ 1 ∫ P ∣ c P ( z ) ∣ φ Q , j ( z ) d z ≤ ∫ ∣ c j − t ( z ) ∣ Q ∑ ∥ b Q ∥ 1 φ Q , j ( z ) d z .
For a fixed z ∈ R n z\in\mathbb{R}^n z ∈ R n , consider the collection of intervals Q Q Q given by
Q ( z ) = { Q : ℓ ( Q ) = 2 j − s , φ Q , j ( z ) ≠ 0 } . \mathcal{Q}(z)=\{Q:\ell(Q)=2^{j-s},\ \varphi_{Q,j}(z)\ne0\}. Q ( z ) = { Q : ℓ ( Q ) = 2 j − s , φ Q , j ( z ) = 0 } .
We can see that φ Q , j ( z ) ≠ 0 \varphi_{Q,j}(z)\ne0 φ Q , j ( z ) = 0 implies ∣ z − ρ Q ∣ ≤ 2 ( n + 1 ) h j |z-\rho_Q|\le2(n+1)h_j ∣ z − ρ Q ∣ ≤ 2 ( n + 1 ) h j , therefore every Q ∈ Q ( z ) Q\in\mathcal{Q}(z) Q ∈ Q ( z ) has its center in B ( z , C h j ) B(z,Ch_j) B ( z , C h j ) . Since cubes in Q ( z ) \mathcal{Q}(z) Q ( z ) are pairwise disjoint and ℓ ( Q ) = 2 j − s \ell(Q)=2^{j-s} ℓ ( Q ) = 2 j − s , we have
∑ Q ∈ Q ( z ) ∣ Q ∣ ≲ h j n + 2 n ( j − s ) ≲ h j n . \sum_{Q\in\mathcal{Q}(z)}|Q|\lesssim h_j^n+2^{n(j-s)}\lesssim h_j^n. Q ∈ Q ( z ) ∑ ∣ Q ∣ ≲ h j n + 2 n ( j − s ) ≲ h j n .
Therefore, we get that
∥ F j 1 ∥ 1 ≤ ∫ ∣ c j − t ( z ) ∣ ( ∑ Q ∥ b Q ∥ 1 φ Q , j ( z ) ) d z \|F_j^1\|_1\le\int|c_{j-t}(z)|\left(\sum_{Q}\|b_Q\|_1\varphi_{Q,j}(z)\right)\,dz ∥ F j 1 ∥ 1 ≤ ∫ ∣ c j − t ( z ) ∣ Q ∑ ∥ b Q ∥ 1 φ Q , j ( z ) d z
≲ ( ∑ Q ∈ Q ( z ) ∥ b Q ∥ 1 ) ∫ ∣ c j − t ( z ) ∣ d z \lesssim\left(\sum_{Q\in\mathcal{Q}(z)}\|b_Q\|_1\right)\int|c_{j-t}(z)|\,dz ≲ Q ∈ Q ( z ) ∑ ∥ b Q ∥ 1 ∫ ∣ c j − t ( z ) ∣ d z
≲ λ 1 2 ( ∑ Q ∈ Q ( z ) ∣ Q ∣ ) ∫ ∣ c j − t ( z ) ∣ d z \lesssim\lambda^{\frac12}\left(\sum_{Q\in\mathcal{Q}(z)}|Q|\right)\int|c_{j-t}(z)|\,dz ≲ λ 2 1 Q ∈ Q ( z ) ∑ ∣ Q ∣ ∫ ∣ c j − t ( z ) ∣ d z
≲ λ 1 2 h j n ∥ c j − t ∥ 1 \lesssim\lambda^{\frac{1}{2}}h_j^n\|c_{j-t}\|_1 ≲ λ 2 1 h j n ∥ c j − t ∥ 1 .
Substituting the above estimate in (4.1), we get
∥ ∬ K j ( ⋅ − y , ⋅ − z ) F j 1 ( y , z ) d y d z ∥ 1 ≲ λ 1 2 h j n 2 n j ∥ c j − t ∥ 1 ≲ 2 − n s 4 λ 1 2 ∥ c j − t ∥ 1 . \left\|\iint K_j(\cdot-y,\cdot-z)F_j^1(y,z)\,dy\,dz\right\|_1\lesssim\lambda^{\frac{1}{2}}\frac{h_j^n}{2^{nj}}\|c_{j-t}\|_1\lesssim2^{-\frac{ns}{4}}\lambda^{\frac{1}{2}}\|c_{j-t}\|_1. ∬ K j ( ⋅ − y , ⋅ − z ) F j 1 ( y , z ) d y d z 1 ≲ λ 2 1 2 nj h j n ∥ c j − t ∥ 1 ≲ 2 − 4 n s λ 2 1 ∥ c j − t ∥ 1 .
Now, summing in j j j , we obtain
∥ H s , t 1 ( b , c ) ∥ L 1 ( E c ) ≲ 2 − n s 4 λ 1 2 . \|H_{s,t}^1(b,c)\|_{L^1(E^c)}\lesssim2^{-\frac{ns}{4}}\lambda^{\frac{1}{2}}. ∥ H s , t 1 ( b , c ) ∥ L 1 ( E c ) ≲ 2 − 4 n s λ 2 1 .
This completes the proof of Proposition 3.1. □
The L 2 L^2 L 2 estimate for the separated term
To estimate H s , t 2 ( b , c ) H_{s,t}^2(b,c) H s , t 2 ( b , c ) , we follow ideas from Christ [7 ] and use a T T ∗ TT^* T T ∗ argument that combines the cancellation of the bad functions with the oscillation of the kernel. We first obtain some local L 1 L^1 L 1 estimates for the bad functions.
Lemma 5.1. Let B ⊂ R n B\subset\mathbb{R}^n B ⊂ R n be a ball of radius r > 0 r>0 r > 0 . Then for every k k k ,
∫ B ∣ b k ( y ) ∣ d y ≲ λ 1 2 ( r + 2 k ) n , ∫ B ∣ c k ( z ) ∣ d z ≲ λ 1 2 ( r + 2 k ) n . (15) \int_B |b_k(y)|\,dy\lesssim\lambda^{\frac{1}{2}}(r+2^k)^n,\qquad\int_B |c_k(z)|\,dz\lesssim\lambda^{\frac{1}{2}}(r+2^k)^n. \tag*{(15)} ∫ B ∣ b k ( y ) ∣ d y ≲ λ 2 1 ( r + 2 k ) n , ∫ B ∣ c k ( z ) ∣ d z ≲ λ 2 1 ( r + 2 k ) n . (15) Moreover,
∑ j ≥ 1 2 − n j ∥ F j ∥ 1 ≲ λ 1 2 . (16) \sum_{j\geq1}2^{-nj}\|F_j\|_1\lesssim\lambda^{\frac{1}{2}}. \tag*{(16)} j ≥ 1 ∑ 2 − nj ∥ F j ∥ 1 ≲ λ 2 1 . (16) Proof. If a cube Q Q Q of side length 2 k 2^k 2 k intersects B B B , then it is contained in the C 2 k C2^k C 2 k -neighborhood of B B B . Since the cubes Q Q Q are disjoint and are of sidelength 2 k 2^k 2 k , we have
∫ B ∣ b k ( y ) ∣ d y ≤ ∑ Q ∩ B ≠ ∅ ∫ Q ∣ b Q ( y ) ∣ d y ≲ λ 1 2 ∑ Q ∩ B ≠ ∅ ∣ Q ∣ ≲ λ 1 2 ( r + 2 k ) n . \begin{aligned}
\int_B |b_k(y)|\,dy
&\leq\sum_{Q\cap B\ne\varnothing}\int_Q |b_Q(y)|\,dy\\
&\lesssim\lambda^{\frac{1}{2}}\sum_{Q\cap B\ne\varnothing}|Q|\\
&\lesssim\lambda^{\frac{1}{2}}(r+2^k)^n.
\end{aligned} ∫ B ∣ b k ( y ) ∣ d y ≤ Q ∩ B = ∅ ∑ ∫ Q ∣ b Q ( y ) ∣ d y ≲ λ 2 1 Q ∩ B = ∅ ∑ ∣ Q ∣ ≲ λ 2 1 ( r + 2 k ) n . Similarly, we can obtain the estimate for c k c_k c k . By (3.4) and (5.1) with a ball of radius ( 2 n + 4 ) 2 j (2n+4)2^j ( 2 n + 4 ) 2 j , we get
∥ F j ∥ 1 ≤ ∫ ∣ y − z ∣ ≤ ( 2 n + 4 ) 2 j ∣ b j − s ( y ) ∣ ∣ c j − t ( z ) ∣ d z d y ≲ λ 1 2 2 n j ∥ b j − s ∥ 1 . \|F_j\|_1\leq\int_{|y-z|\leq(2n+4)2^j}|b_{j-s}(y)|\,|c_{j-t}(z)|\,dz\,dy
\lesssim\lambda^{\frac{1}{2}}2^{nj}\|b_{j-s}\|_1. ∥ F j ∥ 1 ≤ ∫ ∣ y − z ∣ ≤ ( 2 n + 4 ) 2 j ∣ b j − s ( y ) ∣ ∣ c j − t ( z ) ∣ d z d y ≲ λ 2 1 2 nj ∥ b j − s ∥ 1 . Thus
∑ j ≥ 1 2 − n j ∥ F j ∥ 1 ≲ ∑ j ≥ 1 λ 1 2 ∥ b j − s ∥ 1 ≲ λ 1 2 . \sum_{j\geq1}2^{-nj}\|F_j\|_1\lesssim\sum_{j\geq1}\lambda^{\frac{1}{2}}\|b_{j-s}\|_1\lesssim\lambda^{\frac{1}{2}}. j ≥ 1 ∑ 2 − nj ∥ F j ∥ 1 ≲ j ≥ 1 ∑ λ 2 1 ∥ b j − s ∥ 1 ≲ λ 2 1 .
Proof of Proposition 3.2. We exploit the cancellation of the bad functions to obtain a suitable representation of H s , t 2 ( b , c ) H_{s,t}^2(b,c) H s , t 2 ( b , c ) . Cancellation. On the support of F j 2 F_j^2 F j 2 , we have 1 − φ Q , j ( z ) ≠ 0 1-\varphi_{Q,j}(z)\ne0 1 − φ Q , j ( z ) = 0 , and hence ∣ z − ρ Q ∣ ≥ ( n + 1 ) h j |z-\rho_Q|\ge(n+1)h_j ∣ z − ρ Q ∣ ≥ ( n + 1 ) h j . Since ℓ ( Q ) = 2 j − s \ell(Q)=2^{j-s} ℓ ( Q ) = 2 j − s , every y ∈ Q y\in Q y ∈ Q satisfies
∣ y − ρ Q ∣ ≤ n 2 2 j − s ≤ n 2 h j . |y-\rho_Q|\le\frac{\sqrt{n}}{2}2^{j-s}\le\frac{\sqrt{n}}{2}h_j. ∣ y − ρ Q ∣ ≤ 2 n 2 j − s ≤ 2 n h j .
Consequently, we get
∣ y − z ∣ ≥ ∣ z − ρ Q ∣ − ∣ y − ρ Q ∣ ≥ 3 2 h j . (17) |y-z|\ge|z-\rho_Q|-|y-\rho_Q|\ge\frac{3}{2}h_j. \tag*{(17)} ∣ y − z ∣ ≥ ∣ z − ρ Q ∣ − ∣ y − ρ Q ∣ ≥ 2 3 h j . (17)
We now consider the case j < s j<s j < s . Fix an admissible pair ( Q , P ) (Q,P) ( Q , P ) and z z z such that 1 − φ Q , j ( z ) ≠ 0 1-\varphi_{Q,j}(z)\ne0 1 − φ Q , j ( z ) = 0 . Since ∫ b Q = 0 \int b_Q=0 ∫ b Q = 0 , the fundamental theorem of calculus gives
∫ b Q ( y ) K j ( x − y , x − z ) d y = ∫ b Q ( y ) ( K j ( x − y , x − z ) − K j ( x − ρ Q , x − z ) ) d y = − ∫ 0 1 ∫ b Q ( y ) ( y − ρ Q ) ⋅ ∇ 1 K j ( x − ρ Q − ϑ ( y − ρ Q ) , x − z ) d y d ϑ . \begin{aligned}
\int b_Q(y)K_j(x-y,x-z)\,dy
&=\int b_Q(y)\bigl(K_j(x-y,x-z)-K_j(x-\rho_Q,x-z)\bigr)\,dy\\
&=-\int_0^1\int b_Q(y)(y-\rho_Q)\cdot\nabla_1K_j\bigl(x-\rho_Q-\vartheta(y-\rho_Q),x-z\bigr)\,dy\,d\vartheta.
\end{aligned} ∫ b Q ( y ) K j ( x − y , x − z ) d y = ∫ b Q ( y ) ( K j ( x − y , x − z ) − K j ( x − ρ Q , x − z ) ) d y = − ∫ 0 1 ∫ b Q ( y ) ( y − ρ Q ) ⋅ ∇ 1 K j ( x − ρ Q − ϑ ( y − ρ Q ) , x − z ) d y d ϑ .
where ∇ 1 K j \nabla_1K_j ∇ 1 K j denotes the gradient in the first n n n variables. For fixed 0 < ϑ ≤ 1 0<\vartheta\le1 0 < ϑ ≤ 1 , substituting y = ρ Q + w − ρ Q ϑ y=\rho_Q+\frac{w-\rho_Q}{\vartheta} y = ρ Q + ϑ w − ρ Q , we obtain
∫ b Q ( y ) K j ( x − y , x − z ) d y = 2 j − s ∫ 0 1 ∫ b Q ϑ ( w ) ⋅ ∇ 1 K j ( x − w , x − z ) d w d ϑ . (18) \int b_Q(y)K_j(x-y,x-z)\,dy
=2^{j-s}\int_0^1\int b_Q^\vartheta(w)\cdot\nabla_1K_j(x-w,x-z)\,dw\,d\vartheta. \tag*{(18)} ∫ b Q ( y ) K j ( x − y , x − z ) d y = 2 j − s ∫ 0 1 ∫ b Q ϑ ( w ) ⋅ ∇ 1 K j ( x − w , x − z ) d w d ϑ . (18)
where the vector-valued function b Q ϑ b_Q^\vartheta b Q ϑ is defined by
b Q ϑ ( w ) = − w − ρ Q 2 j − s ϑ n + 1 b Q ( ρ Q + w − ρ Q ϑ ) 1 ρ Q + ϑ ( Q − ρ Q ) ( w ) . b_Q^\vartheta(w)=-\frac{w-\rho_Q}{2^{j-s}\vartheta^{n+1}}b_Q\left(\rho_Q+\frac{w-\rho_Q}{\vartheta}\right)\mathbf{1}_{\rho_Q+\vartheta(Q-\rho_Q)}(w). b Q ϑ ( w ) = − 2 j − s ϑ n + 1 w − ρ Q b Q ( ρ Q + ϑ w − ρ Q ) 1 ρ Q + ϑ ( Q − ρ Q ) ( w ) .
From the definition of b Q ϑ b_Q^\vartheta b Q ϑ , we have
supp b Q ϑ ⊆ ρ Q + ϑ ( Q − ρ Q ) ⊆ Q . \operatorname{supp}b_Q^\vartheta\subseteq\rho_Q+\vartheta(Q-\rho_Q)\subseteq Q. supp b Q ϑ ⊆ ρ Q + ϑ ( Q − ρ Q ) ⊆ Q .
Changing variables from w w w back to y y y , we also have
∥ b Q ϑ ∥ 1 = 2 − ( j − s ) ∫ Q ∣ y − ρ Q ∣ ∣ b Q ( y ) ∣ d y ≲ ∥ b Q ∥ 1 ≲ λ 1 2 ∣ Q ∣ , \|b_Q^\vartheta\|_1=2^{-(j-s)}\int_Q|y-\rho_Q||b_Q(y)|\,dy\lesssim\|b_Q\|_1\lesssim\lambda^{\frac12}|Q|, ∥ b Q ϑ ∥ 1 = 2 − ( j − s ) ∫ Q ∣ y − ρ Q ∣∣ b Q ( y ) ∣ d y ≲ ∥ b Q ∥ 1 ≲ λ 2 1 ∣ Q ∣ ,
uniformly in ϑ ∈ ( 0 , 1 ] \vartheta\in(0,1] ϑ ∈ ( 0 , 1 ] . Here, ∣ ⋅ ∣ |\cdot| ∣ ⋅ ∣ of a vector-valued function denotes its Euclidean norm.
We define the analogous notation for the modified functions. Let b j − s ϑ = ∑ Q b Q ϑ b_{j-s}^\vartheta=\sum_Qb_Q^\vartheta b j − s ϑ = ∑ Q b Q ϑ , where the sum is over cubes Q Q Q of side length 2 j − s 2^{j-s} 2 j − s . Note that b j − s ϑ b_{j-s}^\vartheta b j − s ϑ is vector-valued. Define
F j 2 , ϑ ( y , z ) = { ∑ ( Q , P ) b Q ϑ ( y ) c P ( z ) ( 1 − φ Q , j ( z ) ) , j < s , F j 2 ( y , z ) , j ≥ s . F_j^{2,\vartheta}(y,z)=
\begin{cases}
\sum_{(Q,P)}b_Q^\vartheta(y)c_P(z)(1-\varphi_{Q,j}(z)),&j<s,\\
F_j^2(y,z),&j\ge s.
\end{cases} F j 2 , ϑ ( y , z ) = { ∑ ( Q , P ) b Q ϑ ( y ) c P ( z ) ( 1 − φ Q , j ( z )) , F j 2 ( y , z ) , j < s , j ≥ s .
Thus, F j 2 , ϑ F_j^{2,\vartheta} F j 2 , ϑ is vector-valued when j < s j<s j < s and scalar-valued when j ≥ s j\ge s j ≥ s . Since supp b Q ϑ ⊆ Q \operatorname{supp}b_Q^\vartheta\subseteq Q supp b Q ϑ ⊆ Q and ∥ b Q ϑ ∥ 1 ≲ ∥ b Q ∥ 1 \|b_Q^\vartheta\|_1\lesssim\|b_Q\|_1 ∥ b Q ϑ ∥ 1 ≲ ∥ b Q ∥ 1 uniformly in ϑ \vartheta ϑ . Thus the estimates (3.4), (5.1), and (5.2) hold for the modified functions, with constants independent of ϑ \vartheta ϑ . In particular, summing over the admissible pairs, we obtain
∥ F j 2 , ϑ ∥ 1 ≲ ∥ F j 2 ∥ 1 ≤ ∥ F j ∥ 1 . \|F_j^{2,\vartheta}\|_1\lesssim\|F_j^2\|_1\le\|F_j\|_1. ∥ F j 2 , ϑ ∥ 1 ≲ ∥ F j 2 ∥ 1 ≤ ∥ F j ∥ 1 .
We now combine the cases j < s j<s j < s and j ≥ s j\ge s j ≥ s . Recall that
H s , t 2 ( b , c ) ( x ) = ∑ j ≥ 1 ∬ K j ( x − y , x − z ) F j 2 ( y , z ) d y d z . H_{s,t}^2(b,c)(x)=\sum_{j\ge1}\iint K_j(x-y,x-z)F_j^2(y,z)\,dy\,dz. H s , t 2 ( b , c ) ( x ) = j ≥ 1 ∑ ∬ K j ( x − y , x − z ) F j 2 ( y , z ) d y d z .
When j < s j<s j < s , we use the cancellation identity (18) to replace K j K_j K j by its gradient in the first variable. When j ≥ s j\ge s j ≥ s , we keep the original kernel K j K_j K j . To write both cases in a single expression, let γ j = min { 1 , 2 j − s } \gamma_j=\min\{1,2^{j-s}\} γ j = min { 1 , 2 j − s } . For each 0 < ϑ ≤ 1 0<\vartheta\le1 0 < ϑ ≤ 1 , define
V j ϑ ( x ) = { γ j ∬ F j 2 , ϑ ( y , z ) ⋅ ∇ 1 K j ( x − y , x − z ) d y d z , j < s , γ j ∬ K j ( x − y , x − z ) F j 2 , ϑ ( y , z ) d y d z , j ≥ s . V_j^\vartheta(x)=
\begin{cases}
\gamma_j\iint F_j^{2,\vartheta}(y,z)\cdot\nabla_1K_j(x-y,x-z)\,dy\,dz, & j<s,\\
\gamma_j\iint K_j(x-y,x-z)F_j^{2,\vartheta}(y,z)\,dy\,dz, & j\ge s.
\end{cases} V j ϑ ( x ) = { γ j ∬ F j 2 , ϑ ( y , z ) ⋅ ∇ 1 K j ( x − y , x − z ) d y d z , γ j ∬ K j ( x − y , x − z ) F j 2 , ϑ ( y , z ) d y d z , j < s , j ≥ s .
Then, we can write
H s , t 2 ( b , c ) ( x ) = ∑ j ≥ 1 ∫ 0 1 V j ϑ ( x ) d ϑ . H_{s,t}^2(b,c)(x)=\sum_{j\ge1}\int_0^1V_j^\vartheta(x)\,d\vartheta. H s , t 2 ( b , c ) ( x ) = j ≥ 1 ∑ ∫ 0 1 V j ϑ ( x ) d ϑ .
From (2.6) and the estimates for F j 2 , ϑ F_j^{2,\vartheta} F j 2 , ϑ , we have
∫ R n ∫ 0 1 ∑ j ≥ 1 ∣ V j ϑ ( x ) ∣ d ϑ d x ≲ ∑ j ≥ 1 γ j 2 − n j ∫ 0 1 ∥ F j 2 , ϑ ∥ 1 d ϑ \int_{\mathbb{R}^n}\int_0^1\sum_{j\ge1}|V_j^\vartheta(x)|\,d\vartheta\,dx
\lesssim\sum_{j\ge1}\gamma_j2^{-nj}\int_0^1\|F_j^{2,\vartheta}\|_1\,d\vartheta ∫ R n ∫ 0 1 j ≥ 1 ∑ ∣ V j ϑ ( x ) ∣ d ϑ d x ≲ j ≥ 1 ∑ γ j 2 − nj ∫ 0 1 ∥ F j 2 , ϑ ∥ 1 d ϑ
≲ ∑ j ≥ 1 2 − n j ∥ F j ∥ 1 ≲ λ 1 2 , \lesssim\sum_{j\ge1}2^{-nj}\|F_j\|_1\lesssim\lambda^{\frac12}, ≲ j ≥ 1 ∑ 2 − nj ∥ F j ∥ 1 ≲ λ 2 1 ,
where the last inequality follows from (5.2). Therefore, Fubini’s theorem gives
H s , t 2 ( b , c ) ( x ) = ∫ 0 1 ∑ j ≥ 1 V j ϑ ( x ) d ϑ H_{s,t}^2(b,c)(x)=\int_0^1\sum_{j\ge1}V_j^\vartheta(x)\,d\vartheta H s , t 2 ( b , c ) ( x ) = ∫ 0 1 j ≥ 1 ∑ V j ϑ ( x ) d ϑ
for almost every x ∈ R n x\in\mathbb{R}^n x ∈ R n .
By Minkowski’s inequality, we have
∥ H s , t 2 ( b , c ) ∥ 2 ≤ ∫ 0 1 ∥ ∑ j ≥ 1 V j ϑ ∥ 2 d ϑ ≤ sup 0 < ϑ ≤ 1 ∥ ∑ j ≥ 1 V j ϑ ∥ 2 . \begin{aligned}
\|H_{s,t}^2(b,c)\|_2\le\int_0^1\left\|\sum_{j\ge1}V_j^\vartheta\right\|_2\,d\vartheta\\
&\le\sup_{0<\vartheta\le1}\left\|\sum_{j\ge1}V_j^\vartheta\right\|_2.
\end{aligned} ∥ H s , t 2 ( b , c ) ∥ 2 ≤ ∫ 0 1 j ≥ 1 ∑ V j ϑ 2 d ϑ ≤ 0 < ϑ ≤ 1 sup j ≥ 1 ∑ V j ϑ 2 .
Thus, it remains to obtain an L 2 L^2 L 2 estimate for ∑ j ≥ 1 V j ϑ \sum_{j\ge1}V_j^\vartheta ∑ j ≥ 1 V j ϑ , uniformly in ϑ \vartheta ϑ .
Proposition 5.2. The following estimate holds for V j ϑ V_j^\vartheta V j ϑ uniformly in ϑ \vartheta ϑ .
∥ ∑ j ≥ 1 V j ϑ ∥ 2 ≲ 2 − s 8 λ 3 4 . (19) \left\|\sum_{j\ge1}V_j^\vartheta\right\|_2\lesssim2^{-\frac{s}{8}}\lambda^{\frac34}. \tag*{(19)} j ≥ 1 ∑ V j ϑ 2 ≲ 2 − 8 s λ 4 3 . (19)
If we assume that Proposition 5.2 holds true, then the proof of Proposition 3.2 is complete. □
Before giving the proof of Proposition 5.2, we apply a change of variables to V j ϑ V_j^\vartheta V j ϑ .
Change of variables. Let Ψ : R 2 n → R 2 n \Psi:\mathbb{R}^{2n}\to\mathbb{R}^{2n} Ψ : R 2 n → R 2 n be given by
Ψ ( y , z ) = ( y + z 2 , y − z 2 ) = ( u , v ) , ( y , z ) ∈ R n × R n . (20) \Psi(y,z)=\left(\frac{y+z}{\sqrt{2}},\frac{y-z}{\sqrt{2}}\right)=(u,v),\qquad(y,z)\in\mathbb{R}^n\times\mathbb{R}^n. \tag*{(20)} Ψ ( y , z ) = ( 2 y + z , 2 y − z ) = ( u , v ) , ( y , z ) ∈ R n × R n . (20)
Then
Ψ ( x − y , x − z ) = ( 2 x − y − z 2 , z − y 2 ) = ( 2 x , 0 ) − Ψ ( y , z ) . \Psi(x-y,x-z)=\left(\frac{2x-y-z}{\sqrt{2}},\frac{z-y}{\sqrt{2}}\right)=(\sqrt{2}x,0)-\Psi(y,z). Ψ ( x − y , x − z ) = ( 2 2 x − y − z , 2 z − y ) = ( 2 x , 0 ) − Ψ ( y , z ) .
Since K j K_j K j is radial and Ψ \Psi Ψ is orthogonal, we have
K j ( x − y , x − z ) = K j ( Ψ ( x − y , x − z ) ) = K j ( ( 2 x , 0 ) − Ψ ( y , z ) ) . K_j(x-y,x-z)=K_j\bigl(\Psi(x-y,x-z)\bigr)=K_j\bigl((\sqrt{2}x,0)-\Psi(y,z)\bigr). K j ( x − y , x − z ) = K j ( Ψ ( x − y , x − z ) ) = K j ( ( 2 x , 0 ) − Ψ ( y , z ) ) .
Since the Jacobian of Ψ \Psi Ψ has absolute value one, applying a change of variable Ψ ( y , z ) = ( u , v ) \Psi(y,z)=(u,v) Ψ ( y , z ) = ( u , v ) , we get
∬ K j ( x − y , x − z ) F j ( y , z ) d y d z = ∬ K j ( ( 2 x , 0 ) − ( u , v ) ) F ~ j ( u , v ) d u d v , \iint K_j(x-y,x-z)F_j(y,z)\,dy\,dz=\iint K_j\left((\sqrt{2}x,0)-(u,v)\right)\widetilde{F}_j(u,v)\,du\,dv, ∬ K j ( x − y , x − z ) F j ( y , z ) d y d z = ∬ K j ( ( 2 x , 0 ) − ( u , v ) ) F j ( u , v ) d u d v ,
where F ~ j = F j ∘ Ψ − 1 \widetilde{F}_j=F_j\circ\Psi^{-1} F j = F j ∘ Ψ − 1 , i.e. F ~ j ( u , v ) = F j ( u + v 2 , u − v 2 ) \widetilde{F}_j(u,v)=F_j\left(\frac{u+v}{\sqrt{2}},\frac{u-v}{\sqrt{2}}\right) F j ( u , v ) = F j ( 2 u + v , 2 u − v ) . Moreover, from (5.3), we have
∣ v ∣ = ∣ y − z ∣ 2 ≥ 3 2 2 h j > h j . (21) |v|=\frac{|y-z|}{\sqrt{2}}\geq\frac{3}{2\sqrt{2}}h_j>h_j. \tag*{(21)} ∣ v ∣ = 2 ∣ y − z ∣ ≥ 2 2 3 h j > h j . (21)
With this change of variables, we first obtain the following estimate for F ~ j \widetilde{F}_j F j .
Lemma 5.3. For every u 0 ∈ R n u_0\in\mathbb{R}^n u 0 ∈ R n , we have
∫ ∣ u − u 0 ∣ ≤ r ∫ R n ∣ F ~ j ( u , v ) ∣ d v d u ≲ λ 2 n j ( r + 2 j − s ) n , 0 < r ≲ 2 j , (22) \int_{|u-u_0|\leq r}\int_{\mathbb{R}^n}|\widetilde{F}_j(u,v)|\,dv\,du\lesssim\lambda2^{nj}(r+2^{j-s})^n,\qquad0<r\lesssim2^j, \tag*{(22)} ∫ ∣ u − u 0 ∣ ≤ r ∫ R n ∣ F j ( u , v ) ∣ d v d u ≲ λ 2 nj ( r + 2 j − s ) n , 0 < r ≲ 2 j , (22) ∫ ∣ u − u 0 ∣ ≤ r ∫ R n ∣ F ~ j ( u , v ) ∣ d u d v ≲ λ 2 n j r n , r ≳ 2 j . (23) \int_{|u-u_0|\leq r}\int_{\mathbb{R}^n}|\widetilde{F}_j(u,v)|\,du\,dv\lesssim\lambda2^{nj}r^n,\qquad r\gtrsim2^j. \tag*{(23)} ∫ ∣ u − u 0 ∣ ≤ r ∫ R n ∣ F j ( u , v ) ∣ d u d v ≲ λ 2 nj r n , r ≳ 2 j . (23) Moreover, F ~ j ( u , v ) = 0 \widetilde{F}_j(u,v)=0 F j ( u , v ) = 0 unless ∣ v ∣ ≲ 2 j |v|\lesssim2^j ∣ v ∣ ≲ 2 j .
Proof. The support condition of F j F_j F j in (3.4) implies that ∣ y − z ∣ ≤ ( 2 n + 4 ) 2 j |y-z|\leq(2n+4)2^j ∣ y − z ∣ ≤ ( 2 n + 4 ) 2 j whenever F j ( y , z ) ≠ 0 F_j(y,z)\neq0 F j ( y , z ) = 0 . Since v = y − z 2 v=\frac{y-z}{\sqrt{2}} v = 2 y − z , we obtain ∣ v ∣ ≲ 2 j |v|\lesssim2^j ∣ v ∣ ≲ 2 j on the support of F ~ j \widetilde{F}_j F j .
By change of variable y = u + v 2 y=\frac{u+v}{\sqrt{2}} y = 2 u + v and z = u − v 2 z=\frac{u-v}{\sqrt{2}} z = 2 u − v , we have
∫ ∣ u − u 0 ∣ ≤ r ∫ R n ∣ F ~ j ( u , v ) ∣ d v d u = ∬ S ∣ F j ( y , z ) ∣ d y d z , \int_{|u-u_0|\leq r}\int_{\mathbb{R}^n}|\widetilde{F}_j(u,v)|\,dv\,du=\iint_S|F_j(y,z)|\,dy\,dz, ∫ ∣ u − u 0 ∣ ≤ r ∫ R n ∣ F j ( u , v ) ∣ d v d u = ∬ S ∣ F j ( y , z ) ∣ d y d z ,
where
S = { ( y , z ) : ∣ y + z − 2 u 0 ∣ ≤ 2 r } . S=\{(y,z):|y+z-\sqrt{2}u_0|\leq\sqrt{2}r\}. S = {( y , z ) : ∣ y + z − 2 u 0 ∣ ≤ 2 r } .
For ( y , z ) ∈ S (y,z)\in S ( y , z ) ∈ S with F j ( y , z ) ≠ 0 F_j(y,z)\neq0 F j ( y , z ) = 0 , we have
∣ 2 z − 2 u 0 ∣ = ∣ ( y + z − 2 u 0 ) − ( y − z ) ∣ ≤ 2 r + ( 2 n + 4 ) 2 j . |2z-\sqrt{2}u_0|=|(y+z-\sqrt{2}u_0)-(y-z)|\leq\sqrt{2}r+(2n+4)2^j. ∣2 z − 2 u 0 ∣ = ∣ ( y + z − 2 u 0 ) − ( y − z ) ∣ ≤ 2 r + ( 2 n + 4 ) 2 j .
First consider r ≲ 2 j r\lesssim2^j r ≲ 2 j . Then ∣ 2 z − 2 u 0 ∣ ≲ 2 j |2z-\sqrt{2}u_0|\lesssim2^j ∣2 z − 2 u 0 ∣ ≲ 2 j . Let
B 0 = { z : ∣ 2 z − 2 u 0 ∣ ≲ 2 j } , B z = { y : ∣ y + z − 2 u 0 ∣ ≤ 2 r } . B_0=\{z:|2z-\sqrt{2}u_0|\lesssim2^j\},\qquad B_z=\{y:|y+z-\sqrt{2}u_0|\leq\sqrt{2}r\}. B 0 = { z : ∣2 z − 2 u 0 ∣ ≲ 2 j } , B z = { y : ∣ y + z − 2 u 0 ∣ ≤ 2 r } .
For each z ∈ B 0 z\in B_0 z ∈ B 0 , the corresponding variable y y y in S S S lies in the ball B z B_z B z with radius 2 r \sqrt{2}r 2 r . Using (5.1), we obtain
∬ S ∣ F j ( y , z ) ∣ d y d z ≤ ∫ B 0 ∣ c j − t ( z ) ∣ ( ∫ B z ∣ b j − s ( y ) ∣ d y ) d z \iint_S|F_j(y,z)|\,dy\,dz\leq\int_{B_0}|c_{j-t}(z)|\left(\int_{B_z}|b_{j-s}(y)|\,dy\right)\,dz ∬ S ∣ F j ( y , z ) ∣ d y d z ≤ ∫ B 0 ∣ c j − t ( z ) ∣ ( ∫ B z ∣ b j − s ( y ) ∣ d y ) d z
≲ λ 1 2 ( r + 2 j − s ) n ∫ B 0 ∣ c j − t ( z ) ∣ d z \lesssim\lambda^{\frac12}(r+2^{j-s})^n\int_{B_0}|c_{j-t}(z)|\,dz ≲ λ 2 1 ( r + 2 j − s ) n ∫ B 0 ∣ c j − t ( z ) ∣ d z
≲ λ 1 2 ( r + 2 j − s ) n λ 1 2 ( 2 j + 2 j − t ) n \lesssim\lambda^{\frac12}(r+2^{j-s})^n\lambda^{\frac12}(2^j+2^{j-t})^n ≲ λ 2 1 ( r + 2 j − s ) n λ 2 1 ( 2 j + 2 j − t ) n
≲ λ 2 n j ( r + 2 j − s ) n . \lesssim\lambda2^{nj}(r+2^{j-s})^n. ≲ λ 2 nj ( r + 2 j − s ) n .
This proves (5.8).
Next, when r ≳ 2 j r\gtrsim2^j r ≳ 2 j , for ( y , z ) ∈ S (y,z)\in S ( y , z ) ∈ S with F j ( y , z ) ≠ 0 F_j(y,z)\neq0 F j ( y , z ) = 0 , we have
∣ 2 y − 2 u 0 ∣ = ∣ ( y + z − 2 u 0 ) + ( y − z ) ∣ ≤ ∣ y + z − 2 u 0 ∣ + ∣ y − z ∣ \begin{aligned}
|2y-\sqrt{2}u_0|=|(y+z-\sqrt{2}u_0)+(y-z)| \\
&\leq|y+z-\sqrt{2}u_0|+|y-z|
\end{aligned} ∣2 y − 2 u 0 ∣ = ∣ ( y + z − 2 u 0 ) + ( y − z ) ∣ ≤ ∣ y + z − 2 u 0 ∣ + ∣ y − z ∣
≲ r + 2 j ≲ r . \lesssim r+2^j\lesssim r. ≲ r + 2 j ≲ r .
Let
B 0 = { y : ∣ 2 y − 2 u 0 ∣ ≲ r } , B y = { z : ∣ z − y ∣ ≤ ( 2 n + 4 ) 2 j } . B_0=\{y:|2y-\sqrt{2}u_0|\lesssim r\},\qquad B_y=\{z:|z-y|\leq(2n+4)2^j\}. B 0 = { y : ∣2 y − 2 u 0 ∣ ≲ r } , B y = { z : ∣ z − y ∣ ≤ ( 2 n + 4 ) 2 j } .
Then
S ∩ supp F j ⊆ { ( y , z ) : y ∈ B 0 , z ∈ B y } . S \cap\operatorname{supp} F_j \subseteq\{(y,z): y \in B_0,\ z \in B_y\}. S ∩ supp F j ⊆ {( y , z ) : y ∈ B 0 , z ∈ B y } .
Using (5.1), we obtain
∬ S ∣ F j ( y , z ) ∣ d y d z ≤ ∫ B 0 ∣ b j − s ( y ) ∣ ( ∫ B y ∣ c j − t ( z ) ∣ d z ) d y \iint_S |F_j(y,z)|\,dy\,dz \le\int_{B_0} |b_{j-s}(y)|\left(\int_{B_y}|c_{j-t}(z)|\,dz\right)dy ∬ S ∣ F j ( y , z ) ∣ d y d z ≤ ∫ B 0 ∣ b j − s ( y ) ∣ ( ∫ B y ∣ c j − t ( z ) ∣ d z ) d y
≲ λ 1 2 ( 2 j + 2 j − t ) n ∫ B 0 ∣ b j − s ( y ) ∣ d y \lesssim\lambda^{\frac{1}{2}}(2^j+2^{j-t})^n \int_{B_0}|b_{j-s}(y)|\,dy ≲ λ 2 1 ( 2 j + 2 j − t ) n ∫ B 0 ∣ b j − s ( y ) ∣ d y
≲ λ 1 2 2 n j λ 1 2 ( r + 2 j − s ) n \lesssim\lambda^{\frac{1}{2}}2^{nj}\lambda^{\frac{1}{2}}(r+2^{j-s})^n ≲ λ 2 1 2 nj λ 2 1 ( r + 2 j − s ) n
≲ λ 2 n j r n . \lesssim\lambda2^{nj}r^n. ≲ λ 2 nj r n .
This proves (5.9). □
The estimates in the above lemma also hold for F ~ j 2 , ϑ = F j 2 , ϑ ∘ Ψ − 1 \widetilde{F}_j^{2,\vartheta}=F_j^{2,\vartheta}\circ\Psi^{-1} F j 2 , ϑ = F j 2 , ϑ ∘ Ψ − 1 , uniformly in ϑ \vartheta ϑ , as they have the same support properties and satisfy the same local L 1 L^1 L 1 estimates. In particular, F ~ j 2 , ϑ \widetilde{F}_j^{2,\vartheta} F j 2 , ϑ satisfies (5.8) and (5.9).
Finally, we rewrite V j ϑ V_j^\vartheta V j ϑ after this change of variables. Since
∇ 1 K j ( y , z ) = 1 2 ( ∇ u + ∇ v ) K j ( Ψ ( y , z ) ) , \nabla_1K_j(y,z)=\frac{1}{\sqrt{2}}(\nabla_u+\nabla_v)K_j(\Psi(y,z)), ∇ 1 K j ( y , z ) = 2 1 ( ∇ u + ∇ v ) K j ( Ψ ( y , z )) ,
we define
K j ( u , v ) = { 1 2 ( ∇ u + ∇ v ) K j ( u , v ) , j < s , K j ( u , v ) , j ≥ s . \mathcal{K}_j(u,v)=
\begin{cases}
\frac{1}{\sqrt{2}}(\nabla_u+\nabla_v)K_j(u,v), & j<s,\\
K_j(u,v), & j\ge s.
\end{cases} K j ( u , v ) = { 2 1 ( ∇ u + ∇ v ) K j ( u , v ) , K j ( u , v ) , j < s , j ≥ s .
Thus K j \mathcal{K}_j K j and F ~ j 2 , ϑ \widetilde{F}_j^{2,\vartheta} F j 2 , ϑ are both vector-valued when j < s j<s j < s and both scalar-valued when j ≥ s j\ge s j ≥ s . Therefore
V j ϑ ( x ) = γ j ∫ R 2 n K j ( ( 2 x , 0 ) − ( u , v ) ) ⋅ F ~ j 2 , ϑ ( u , v ) d u d v . (24) V_j^\vartheta(x)=\gamma_j\int_{\mathbb{R}^{2n}}\mathcal{K}_j((\sqrt{2}x,0)-(u,v))\cdot\widetilde{F}_j^{2,\vartheta}(u,v)\,du\,dv. \tag*{(24)} V j ϑ ( x ) = γ j ∫ R 2 n K j (( 2 x , 0 ) − ( u , v )) ⋅ F j 2 , ϑ ( u , v ) d u d v . (24)
where the ‘⋅ \cdot ⋅ ’ denotes the Euclidean scalar product in the vector-valued case and ordinary multiplication in the scalar-valued case.
For every j ≥ 1 j\ge1 j ≥ 1 , let
d j = { n , j < s , 1 , j ≥ s . d_j=
\begin{cases}
n, & j<s,\\
1, & j\ge s.
\end{cases} d j = { n , 1 , j < s , j ≥ s .
We can write K j = ( K j , m ) m = 1 d j \mathcal{K}_j=(\mathcal{K}_{j,m})_{m=1}^{d_j} K j = ( K j , m ) m = 1 d j , where K j , m \mathcal{K}_{j,m} K j , m is the m m m -th component given by
K j , m ( u , v ) = { 1 2 ( ∂ u m + ∂ v m ) K j ( u , v ) , j < s , 1 ≤ m ≤ n , K j ( u , v ) , j ≥ s , m = 1. \mathcal{K}_{j,m}(u,v)=
\begin{cases}
\frac{1}{\sqrt{2}}(\partial_{u_m}+\partial_{v_m})K_j(u,v), & j<s,\quad1\le m\le n,\\
K_j(u,v), & j\ge s,\quad m=1.
\end{cases} K j , m ( u , v ) = { 2 1 ( ∂ u m + ∂ v m ) K j ( u , v ) , K j ( u , v ) , j < s , 1 ≤ m ≤ n , j ≥ s , m = 1.
Similarly, we can write F ~ j 2 , ϑ = ( F ~ j , m 2 , ϑ ) m = 1 d j \widetilde{F}_j^{2,\vartheta}=(\widetilde{F}_{j,m}^{2,\vartheta})_{m=1}^{d_j} F j 2 , ϑ = ( F j , m 2 , ϑ ) m = 1 d j . Then (5.10) can be written as
V j ϑ ( x ) = γ j ∑ m = 1 d j ∫ R 2 n K j , m ( 2 x − u , − v ) F ~ j , m 2 , ϑ ( u , v ) d u d v . (25) V_j^\vartheta(x)=\gamma_j\sum_{m=1}^{d_j}\int_{\mathbb{R}^{2n}}\mathcal{K}_{j,m}(\sqrt{2}x-u,-v)\widetilde{F}_{j,m}^{2,\vartheta}(u,v)\,du\,dv. \tag*{(25)} V j ϑ ( x ) = γ j m = 1 ∑ d j ∫ R 2 n K j , m ( 2 x − u , − v ) F j , m 2 , ϑ ( u , v ) d u d v . (25)
Thus, although the number of components depends on j j j , the function V j ϑ ( x ) V_j^\vartheta(x) V j ϑ ( x ) is always scalar-valued. We now state the kernel estimates that will be used to prove Proposition 5.2.
For j , k ≥ 1 j,k\ge1 j , k ≥ 1 , 1 ≤ m ≤ d j 1\le m\le d_j 1 ≤ m ≤ d j and 1 ≤ ℓ ≤ d k 1\le\ell\le d_k 1 ≤ ℓ ≤ d k , define
A j k m , ℓ ( u , v , u ′ , v ′ ) = ∫ R n K j , m ( x − u , − v ) K k , ℓ ( x − u ′ , − v ′ ) d x . (26) A_{jk}^{m,\ell}(u,v,u',v')=\int_{\mathbb{R}^n}\mathcal{K}_{j,m}(x-u,-v)\mathcal{K}_{k,\ell}(x-u',-v')\,dx. \tag*{(26)} A j k m , ℓ ( u , v , u ′ , v ′ ) = ∫ R n K j , m ( x − u , − v ) K k , ℓ ( x − u ′ , − v ′ ) d x . (26)
Lemma 5.4. Let j , k ≥ 1 j,k \ge1 j , k ≥ 1 , 1 ≤ m ≤ d j 1 \le m \le d_j 1 ≤ m ≤ d j and 1 ≤ ℓ ≤ d k 1 \le\ell\le d_k 1 ≤ ℓ ≤ d k . Then the following estimates hold uniformly in m m m and ℓ \ell ℓ .
(1) Suppose n ≥ 2 n \ge2 n ≥ 2 . Then
∣ A j j m , ℓ ( u , v , u ′ , v ′ ) ∣ ≲ 2 − 3 n j ( 1 + ∣ u − u ′ ∣ ) − n − 2 2 . (27) \lvert A_{jj}^{m,\ell}(u,v,u',v')\rvert\lesssim2^{-3nj}\left(1+\lvert u-u'\rvert\right)^{-\frac{n-2}{2}}. \tag*{(27)} ∣ A j j m , ℓ ( u , v , u ′ , v ′ )∣ ≲ 2 − 3 nj ( 1 + ∣ u − u ′ ∣ ) − 2 n − 2 . (27) and if k ≥ j + 4 k \ge j+4 k ≥ j + 4 ,
∣ A j k m , ℓ ( u , v , u ′ , v ′ ) ∣ ≲ 2 − n j 2 − 2 n k 2 − j ( n − 1 ) 2 . (28) \lvert A_{jk}^{m,\ell}(u,v,u',v')\rvert\lesssim2^{-nj}2^{-2nk}2^{-\frac{j(n-1)}{2}}. \tag*{(28)} ∣ A j k m , ℓ ( u , v , u ′ , v ′ )∣ ≲ 2 − nj 2 − 2 nk 2 − 2 j ( n − 1 ) . (28) (2) Suppose n = 1 n=1 n = 1 . If ∣ v ∣ , ∣ v ′ ∣ ≥ h \lvert v\rvert,\lvert v'\rvert\ge h ∣ v ∣ , ∣ v ′ ∣ ≥ h and 0 < h ≤ 2 j 0<h\le2^j 0 < h ≤ 2 j , then
∣ A j j m , ℓ ( u , v , u ′ , v ′ ) ∣ ≲ 2 − 3 j ( 1 + ∣ u − u ′ ∣ h 2 2 2 j ) − 1 2 . (29) \lvert A_{jj}^{m,\ell}(u,v,u',v')\rvert\lesssim2^{-3j}\left(1+\frac{\lvert u-u'\rvert h^2}{2^{2j}}\right)^{-\frac{1}{2}}. \tag*{(29)} ∣ A j j m , ℓ ( u , v , u ′ , v ′ )∣ ≲ 2 − 3 j ( 1 + 2 2 j ∣ u − u ′ ∣ h 2 ) − 2 1 . (29) If k ≥ j + 4 k \ge j+4 k ≥ j + 4 , ∣ v ∣ ≥ h \lvert v\rvert\ge h ∣ v ∣ ≥ h , and 0 < h ≤ 2 j 0<h\le2^j 0 < h ≤ 2 j , then
∣ A j k m , ℓ ( u , v , u ′ , v ′ ) ∣ ≲ 2 − 2 j − 2 k ( 1 + h 2 2 2 j ) − 1 2 . (30) \lvert A_{jk}^{m,\ell}(u,v,u',v')\rvert\lesssim2^{-2j-2k}\left(1+\frac{h^2}{2^{2j}}\right)^{-\frac{1}{2}}. \tag*{(30)} ∣ A j k m , ℓ ( u , v , u ′ , v ′ )∣ ≲ 2 − 2 j − 2 k ( 1 + 2 2 j h 2 ) − 2 1 . (30) In particular, for every n ≥ 1 n\ge1 n ≥ 1 , if ∣ v ∣ , ∣ v ′ ∣ ≥ h \lvert v\rvert,\lvert v'\rvert\ge h ∣ v ∣ , ∣ v ′ ∣ ≥ h and 0 < h ≤ 2 j 0<h\le2^j 0 < h ≤ 2 j , then
∣ A j j m , ℓ ( u , v , u ′ , v ′ ) ∣ ≲ 2 − 3 n j ( 1 + ∣ u − u ′ ∣ h 2 2 2 j ) − 1 2 . (31) \lvert A_{jj}^{m,\ell}(u,v,u',v')\rvert\lesssim2^{-3nj}\left(1+\frac{\lvert u-u'\rvert h^2}{2^{2j}}\right)^{-\frac{1}{2}}. \tag*{(31)} ∣ A j j m , ℓ ( u , v , u ′ , v ′ )∣ ≲ 2 − 3 nj ( 1 + 2 2 j ∣ u − u ′ ∣ h 2 ) − 2 1 . (31) Similarly, if k ≥ j + 4 k\ge j+4 k ≥ j + 4 , ∣ v ∣ ≥ h \lvert v\rvert\ge h ∣ v ∣ ≥ h and 0 < h ≤ 2 j 0<h\le2^j 0 < h ≤ 2 j , then
∣ A j k m , ℓ ( u , v , u ′ , v ′ ) ∣ ≲ 2 − n j 2 − 2 n k ( 1 + h 2 2 2 j ) − 1 2 . (32) \lvert A_{jk}^{m,\ell}(u,v,u',v')\rvert\lesssim2^{-nj}2^{-2nk}\left(1+\frac{h^2}{2^{2j}}\right)^{-\frac{1}{2}}. \tag*{(32)} ∣ A j k m , ℓ ( u , v , u ′ , v ′ )∣ ≲ 2 − nj 2 − 2 nk ( 1 + 2 2 j h 2 ) − 2 1 . (32)
We will prove the Lemma 5.4 in Section 6. We first show how it implies the desired L 2 L^2 L 2 bound in Proposition 5.2.
Proof of Proposition 5.2. We fix ϑ \vartheta ϑ and omit it from the notation. In particular, we write F ~ j , m = F ~ j , m 2 , ϑ \widetilde{F}_{j,m}= \widetilde{F}_{j,m}^{2,\vartheta} F j , m = F j , m 2 , ϑ and V j = V j ϑ V_j=V_j^\vartheta V j = V j ϑ . Expanding the L 2 L^2 L 2 norm, we have
∥ ∑ j ≥ 1 V j ∥ 2 2 = ∑ j , k ≥ 1 ⟨ V j , V k ⟩ . \left\lVert\sum_{j\ge1}V_j\right\rVert_2^2=\sum_{j,k\ge1}\langle V_j,V_k\rangle. j ≥ 1 ∑ V j 2 2 = j , k ≥ 1 ∑ ⟨ V j , V k ⟩ .
The T T ∗ TT^* T T ∗ identity. By (5.11), we can write the inner product as
\begin{aligned}
\langle V_j,V_k\rangle
&=\int_{\mathbb{R}^n}V_j(x)\overline{V_k(x)}\,dx\\
&=\gamma_j\gamma_k\sum_{m=1}^{d_j}\sum_{\ell=1}^{d_k}\iiiint_{\mathbb{R}^{4n}}\widetilde{F}_{j,m}(u,v)\overline{\widetilde{F}_{k,\ell}(u',v')}\\
&\qquad\times\left(\int_{\mathbb{R}^n}K_{j,m}((\sqrt{2}x,0)-(u,v))\overline{K_{k,\ell}((\sqrt{2}x,0)-(u',v'))}\,dx\right)\,du\,dv\,du'\,dv'.
\end{aligned}
By the change of variables x ↦ x 2 x\mapsto\frac{x}{\sqrt{2}} x ↦ 2 x in the inner integral, we obtain
⟨ V j , V k ⟩ = 2 − n 2 γ j γ k ∑ m = 1 d j ∑ ℓ = 1 d k ∫ R 4 n A j k m , ℓ ( u , v , u ′ , v ′ ) F ~ j , m ( u , v ) F ~ k , ℓ ( u ′ , v ′ ) ‾ d u d v d u ′ d v ′ . (33) \langle V_j,V_k\rangle
=2^{-\frac n2}\gamma_j\gamma_k\sum_{m=1}^{d_j}\sum_{\ell=1}^{d_k}\int_{\mathbb{R}^{4n}}A_{jk}^{m,\ell}(u,v,u',v')\widetilde{F}_{j,m}(u,v)\overline{\widetilde{F}_{k,\ell}(u',v')}\,du\,dv\,du'\,dv'. \tag*{(33)} ⟨ V j , V k ⟩ = 2 − 2 n γ j γ k m = 1 ∑ d j ℓ = 1 ∑ d k ∫ R 4 n A j k m , ℓ ( u , v , u ′ , v ′ ) F j , m ( u , v ) F k , ℓ ( u ′ , v ′ ) d u d v d u ′ d v ′ . (33)
Since the estimates for ∣ A j k m , ℓ ∣ |A_{jk}^{m,\ell}| ∣ A j k m , ℓ ∣ in Lemma 5.4 are uniform in m , ℓ m,\ell m , ℓ , and d j ≤ n d_j \le n d j ≤ n , we can apply the Cauchy–Schwarz inequality for sums over m , ℓ m,\ell m , ℓ in (5.19) to obtain
∑ m = 1 d j ∣ F ~ j , m 2 , ϑ ( u , v ) ∣ ≤ d j ( ∑ m = 1 d j ∣ F ~ j , m 2 , ϑ ( u , v ) ∣ 2 ) 1 2 ≤ n ∣ F ~ j 2 , ϑ ( u , v ) ∣ . (34) \sum_{m=1}^{d_j} |\widetilde{F}_{j,m}^{2,\vartheta}(u,v)| \le\sqrt{d_j}\left(\sum_{m=1}^{d_j}|\widetilde{F}_{j,m}^{2,\vartheta}(u,v)|^2\right)^{\frac{1}{2}} \le\sqrt{n}|\widetilde{F}_{j}^{2,\vartheta}(u,v)|. \tag*{(34)} m = 1 ∑ d j ∣ F j , m 2 , ϑ ( u , v ) ∣ ≤ d j m = 1 ∑ d j ∣ F j , m 2 , ϑ ( u , v ) ∣ 2 2 1 ≤ n ∣ F j 2 , ϑ ( u , v ) ∣. (34)
Thus, after summation over m , ℓ m,\ell m , ℓ in (5.19), the resulting estimate has the form of a scalar T T ∗ TT^* T T ∗ estimate, with an implicit constant depending only on n n n .
We estimate the near-diagonal terms with ∣ j − k ∣ < 4 |j-k|<4 ∣ j − k ∣ < 4 and off-diagonal terms with ∣ j − k ∣ ≥ 4 |j-k|\ge4 ∣ j − k ∣ ≥ 4 separately.
Near-diagonal terms. Applying Cauchy–Schwarz in x x x , we have
∑ j , k ≥ 1 ∣ j − k ∣ < 4 ∣ ⟨ V j , V k ⟩ ∣ ≤ ∑ j , k ≥ 1 ∣ j − k ∣ < 4 ∥ V j ∥ 2 ∥ V k ∥ 2 ≲ ∑ j ≥ 1 ∥ V j ∥ 2 2 . (35) \sum_{\substack{j,k\ge1\\ |j-k|<4}}|\langle V_j,V_k\rangle| \le\sum_{\substack{j,k\ge1\\ |j-k|<4}}\|V_j\|_2\|V_k\|_2 \lesssim\sum_{j\ge1}\|V_j\|_2^2. \tag*{(35)} j , k ≥ 1 ∣ j − k ∣ < 4 ∑ ∣ ⟨ V j , V k ⟩ ∣ ≤ j , k ≥ 1 ∣ j − k ∣ < 4 ∑ ∥ V j ∥ 2 ∥ V k ∥ 2 ≲ j ≥ 1 ∑ ∥ V j ∥ 2 2 . (35)
For a fixed j j j , if the integral in (5.12) is nonzero for k = j k=j k = j , then the two kernels have supports of size O ( 2 j ) O(2^j) O ( 2 j ) . Therefore ∣ u − u ′ ∣ ≤ ∣ x − u ∣ + ∣ x − u ′ ∣ ≲ 2 j |u-u'|\le|x-u|+|x-u'|\lesssim2^j ∣ u − u ′ ∣ ≤ ∣ x − u ∣ + ∣ x − u ′ ∣ ≲ 2 j . Also, by (5.7), both ∣ v ∣ |v| ∣ v ∣ and ∣ v ′ ∣ |v'| ∣ v ′ ∣ are at least h j = 2 j − s 4 h_j=2^{j-\frac{s}{4}} h j = 2 j − 4 s on the supports of the modified functions F ~ j 2 , ϑ \widetilde{F}_j^{2,\vartheta} F j 2 , ϑ . Since h j 2 2 2 j = 2 − s 2 \frac{h_j^2}{2^{2j}}=2^{-\frac{s}{2}} 2 2 j h j 2 = 2 − 2 s , the first part of Lemma 5.4 gives, for every m , ℓ m,\ell m , ℓ ,
∣ A j j m , ℓ ( u , v , u ′ , v ′ ) ∣ ≲ 2 − 3 n j ( 1 + 2 − s 2 ∣ u − u ′ ∣ ) − 1 2 1 { ∣ u − u ′ ∣ ≲ 2 j } . (36) |A_{jj}^{m,\ell}(u,v,u',v')|\lesssim2^{-3nj}\left(1+2^{-\frac{s}{2}}|u-u'|\right)^{-\frac{1}{2}}\mathbf{1}_{\{|u-u'|\lesssim2^j\}}. \tag*{(36)} ∣ A j j m , ℓ ( u , v , u ′ , v ′ ) ∣ ≲ 2 − 3 nj ( 1 + 2 − 2 s ∣ u − u ′ ∣ ) − 2 1 1 { ∣ u − u ′ ∣ ≲ 2 j } . (36)
Combining (5.19), (5.20) and (5.22), we obtain
∥ V j ∥ 2 2 ≲ γ j 2 2 − 3 n j ∬ ∣ F ~ j ( u ′ , v ′ ) ∣ ( ∫ ∣ u − u ′ ∣ ≲ 2 j ∫ R n ∣ F ~ j ( u , v ) ∣ ( 1 + 2 − s 2 ∣ u − u ′ ∣ ) 1 2 d v d u ) d u ′ d v ′ . (37) \|V_j\|_2^2\lesssim\gamma_j^2 2^{-3nj}\iint|\widetilde{F}_j(u',v')|\left(\int_{|u-u'|\lesssim2^j}\int_{\mathbb{R}^n}\frac{|\widetilde{F}_j(u,v)|}{(1+2^{-\frac{s}{2}}|u-u'|)^{\frac{1}{2}}}\,dv\,du\right)du'\,dv'. \tag*{(37)} ∥ V j ∥ 2 2 ≲ γ j 2 2 − 3 nj ∬ ∣ F j ( u ′ , v ′ ) ∣ ( ∫ ∣ u − u ′ ∣ ≲ 2 j ∫ R n ( 1 + 2 − 2 s ∣ u − u ′ ∣ ) 2 1 ∣ F j ( u , v ) ∣ d v d u ) d u ′ d v ′ . (37)
We claim that, for every u 0 ∈ R n u_0\in\mathbb{R}^n u 0 ∈ R n ,
∫ ∣ u − u 0 ∣ ≲ 2 j ∫ R n ∣ F ~ j ( u , v ) ∣ ( 1 + 2 − s 2 ∣ u − u 0 ∣ ) 1 2 d v d u ≲ λ 2 2 n j [ 2 − n s + min { 1 , 2 s 4 − j 2 } ] . (38) \int_{|u-u_0|\lesssim2^j}\int_{\mathbb{R}^n}\frac{|\widetilde{F}_j(u,v)|}{(1+2^{-\frac{s}{2}}|u-u_0|)^{\frac{1}{2}}}\,dv\,du\lesssim\lambda2^{2nj}\left[2^{-ns}+\min\left\{1,2^{\frac{s}{4}-\frac{j}{2}}\right\}\right]. \tag*{(38)} ∫ ∣ u − u 0 ∣ ≲ 2 j ∫ R n ( 1 + 2 − 2 s ∣ u − u 0 ∣ ) 2 1 ∣ F j ( u , v ) ∣ d v d u ≲ λ 2 2 nj [ 2 − n s + min { 1 , 2 4 s − 2 j } ] . (38)
If we assume (5.24) for the moment and substitute it into (5.23) along with ∥ F ~ j ∥ 1 ≲ ∥ F j ∥ 1 \|\widetilde{F}_j\|_1\lesssim\|F_j\|_1 ∥ F j ∥ 1 ≲ ∥ F j ∥ 1 , we get
∥ V j ∥ 2 2 ≲ λ 2 − n j ∥ F j ∥ 1 γ j 2 [ 2 − n s + min { 1 , 2 s 4 − j 2 } ] . \|V_j\|_2^2\lesssim\lambda2^{-nj}\|F_j\|_1\gamma_j^2\left[2^{-ns}+\min\left\{1,2^{\frac{s}{4}-\frac{j}{2}}\right\}\right]. ∥ V j ∥ 2 2 ≲ λ 2 − nj ∥ F j ∥ 1 γ j 2 [ 2 − n s + min { 1 , 2 4 s − 2 j } ] .
Note that the term γ j 2 2 − n s \gamma_j^2 2^{-ns} γ j 2 2 − n s is at most 2 − n s 2^{-ns} 2 − n s . For the other term, we consider three ranges. If j < s 2 j<\frac{s}{2} j < 2 s , then γ j = 2 j − s \gamma_j=2^{j-s} γ j = 2 j − s and
γ j 2 min { 1 , 2 s 4 − j 2 } ≤ 2 2 ( j − s ) ≤ 2 − s . \gamma_j^2\min\left\{1,2^{\frac{s}{4}-\frac{j}{2}}\right\}\le2^{2(j-s)}\le2^{-s}. γ j 2 min { 1 , 2 4 s − 2 j } ≤ 2 2 ( j − s ) ≤ 2 − s .
If s 2 ≤ j < s \frac{s}{2}\le j<s 2 s ≤ j < s , then
γ j 2 2 s 4 − j 2 = 2 3 j 2 − 7 s 4 ≤ 2 − s 4 . \gamma_j^2 2^{\frac{s}{4}-\frac{j}{2}}=2^{\frac{3j}{2}-\frac{7s}{4}}\le2^{-\frac{s}{4}}. γ j 2 2 4 s − 2 j = 2 2 3 j − 4 7 s ≤ 2 − 4 s .
Finally, if j ≥ s j\ge s j ≥ s , then γ j = 1 \gamma_j=1 γ j = 1 and
2 s 4 − j 2 ≤ 2 − s 4 . 2^{\frac{s}{4}-\frac{j}{2}}\le2^{-\frac{s}{4}}. 2 4 s − 2 j ≤ 2 − 4 s .
Hence
∑ j ≥ 1 ∥ V j ∥ 2 2 ≲ 2 − s 4 λ ∑ j ≥ 1 2 − n j ∥ F j ∥ 1 ≲ 2 − s 4 λ 3 2 , (39) \sum_{j\ge1}\|V_j\|_2^2\lesssim2^{-\frac{s}{4}}\lambda\sum_{j\ge1}2^{-nj}\|F_j\|_1\lesssim2^{-\frac{s}{4}}\lambda^{\frac{3}{2}}, \tag*{(39)} j ≥ 1 ∑ ∥ V j ∥ 2 2 ≲ 2 − 4 s λ j ≥ 1 ∑ 2 − nj ∥ F j ∥ 1 ≲ 2 − 4 s λ 2 3 , (39)
where the last inequality follows from (5.2). Thus, it remains to prove the claim (5.24).
Proof of claim (38) : If 2 j ≲ 2 s 2 2^j \lesssim2^{\frac{s}{2}} 2 j ≲ 2 2 s , then ( 1 + 2 − s 2 ∣ u − u 0 ∣ ) − 1 2 ≤ 1 (1+2^{-\frac{s}{2}}|u-u_0|)^{-\frac{1}{2}}\le1 ( 1 + 2 − 2 s ∣ u − u 0 ∣ ) − 2 1 ≤ 1 . Thus, applying (22) with r ∼ 2 j r\sim2^j r ∼ 2 j , we obtain
∫ ∣ u − u 0 ∣ ≲ 2 j ∫ R n ∣ F ~ j ( u , v ) ∣ ( 1 + 2 − s 2 ∣ u − u 0 ∣ ) 1 2 d v d u ≲ λ 2 n j . \int_{|u-u_0|\lesssim2^j}\int_{\mathbb{R}^n}\frac{|\widetilde{F}_j(u,v)|}{(1+2^{-\frac{s}{2}}|u-u_0|)^{\frac{1}{2}}}\,dv\,du\lesssim\lambda2^{nj}. ∫ ∣ u − u 0 ∣ ≲ 2 j ∫ R n ( 1 + 2 − 2 s ∣ u − u 0 ∣ ) 2 1 ∣ F j ( u , v ) ∣ d v d u ≲ λ 2 nj .
This proves the bound corresponding to the term 1 1 1 in the minimum in (38) .
Now suppose 2 s 2 ≲ 2 j 2^{\frac{s}{2}}\lesssim2^j 2 2 s ≲ 2 j . We split the integration region into the ball ∣ u − u 0 ∣ ≤ 2 s 2 |u-u_0|\le2^{\frac{s}{2}} ∣ u − u 0 ∣ ≤ 2 2 s and the annuli
2 s 2 + ℓ < ∣ u − u 0 ∣ ≤ 2 s 2 + ℓ + 1 , ℓ = 0 , 1 , … , 2 s 2 + ℓ ≲ 2 j . 2^{\frac{s}{2}+\ell}<|u-u_0|\le2^{\frac{s}{2}+\ell+1},\qquad\ell=0,1,\ldots,\quad2^{\frac{s}{2}+\ell}\lesssim2^j. 2 2 s + ℓ < ∣ u − u 0 ∣ ≤ 2 2 s + ℓ + 1 , ℓ = 0 , 1 , … , 2 2 s + ℓ ≲ 2 j .
Then, ( 1 + 2 − s 2 ∣ u − u 0 ∣ ) − 1 2 = O ( 2 − ℓ 2 ) (1+2^{-\frac{s}{2}}|u-u_0|)^{-\frac{1}{2}}=O(2^{-\frac{\ell}{2}}) ( 1 + 2 − 2 s ∣ u − u 0 ∣ ) − 2 1 = O ( 2 − 2 ℓ ) on the ℓ \ell ℓ -th annulus. Using (22) on the balls ∣ u − u 0 ∣ ≤ 2 s 2 + ℓ + 1 |u-u_0|\le2^{\frac{s}{2}+\ell+1} ∣ u − u 0 ∣ ≤ 2 2 s + ℓ + 1 containing these annuli, we obtain
∫ ∣ u − u 0 ∣ ≲ 2 j ∫ R n ∣ F ~ j ( u , v ) ∣ ( 1 + 2 − s 2 ∣ u − u 0 ∣ ) 1 2 d v d u ≲ λ 2 n j [ ( 2 s 2 + 2 j − s ) n + ∑ 2 s 2 + ℓ ≲ 2 j 2 − ℓ 2 ( 2 s 2 + ℓ + 2 j − s ) n ] \int_{|u-u_0|\lesssim2^j}\int_{\mathbb{R}^n}\frac{|\widetilde{F}_j(u,v)|}{(1+2^{-\frac{s}{2}}|u-u_0|)^{\frac{1}{2}}}\,dv\,du\lesssim\lambda2^{nj}\left[\left(2^{\frac{s}{2}}+2^{j-s}\right)^n+\sum_{2^{\frac{s}{2}+\ell}\lesssim2^j}2^{-\frac{\ell}{2}}\left(2^{\frac{s}{2}+\ell}+2^{j-s}\right)^n\right] ∫ ∣ u − u 0 ∣ ≲ 2 j ∫ R n ( 1 + 2 − 2 s ∣ u − u 0 ∣ ) 2 1 ∣ F j ( u , v ) ∣ d v d u ≲ λ 2 nj ( 2 2 s + 2 j − s ) n + 2 2 s + ℓ ≲ 2 j ∑ 2 − 2 ℓ ( 2 2 s + ℓ + 2 j − s ) n
≲ λ 2 n j [ 2 n ( j − s ) + 2 n s 2 ∑ 2 s 2 + ℓ ≲ 2 j 2 ℓ ( n − 1 2 ) ] \lesssim\lambda2^{nj}\left[2^{n(j-s)}+2^{\frac{ns}{2}}\sum_{2^{\frac{s}{2}+\ell}\lesssim2^j}2^{\ell(n-\frac{1}{2})}\right] ≲ λ 2 nj 2 n ( j − s ) + 2 2 n s 2 2 s + ℓ ≲ 2 j ∑ 2 ℓ ( n − 2 1 )
≲ λ 2 n j [ 2 n ( j − s ) + 2 n j 2 s 2 − j ] . \lesssim\lambda2^{nj}\left[2^{n(j-s)}+2^{nj}2^{\frac{s}{2}-j}\right]. ≲ λ 2 nj [ 2 n ( j − s ) + 2 nj 2 2 s − j ] .
This proves (38) in the second case.
Off-diagonal terms. By symmetry of the inner product, it is enough to consider k ≥ j + 4 k\ge j+4 k ≥ j + 4 . For fixed ( u ′ , v ′ ) (u',v') ( u ′ , v ′ ) , the integral defining A j k m , ℓ ( u , v , u ′ , v ′ ) A_{jk}^{m,\ell}(u,v,u',v') A j k m , ℓ ( u , v , u ′ , v ′ ) can be nonzero only if ∣ x − u ∣ ≲ 2 j |x-u|\lesssim2^j ∣ x − u ∣ ≲ 2 j and ∣ x − u ′ ∣ ≲ 2 k |x-u'|\lesssim2^k ∣ x − u ′ ∣ ≲ 2 k for some x ∈ R n x\in\mathbb{R}^n x ∈ R n . Thus
∣ u − u ′ ∣ ≲ 2 k . |u-u'|\lesssim2^k. ∣ u − u ′ ∣ ≲ 2 k .
Also, on the support of F ~ j \widetilde{F}_j F j , we have ∣ v ∣ ≲ 2 j |v|\lesssim2^j ∣ v ∣ ≲ 2 j . From (23) for F ~ j 2 , ϑ \widetilde{F}_j^{2,\vartheta} F j 2 , ϑ with u 0 = u ′ u_0=u' u 0 = u ′ and r ∼ 2 k r\sim2^k r ∼ 2 k , we obtain
∫ ∣ u − u ′ ∣ ≲ 2 k ∫ R n ∣ F ~ j 2 , ϑ ( u , v ) ∣ d v d u ≲ λ 2 n j 2 n k , \int_{|u-u'|\lesssim2^k}\int_{\mathbb{R}^n}|\widetilde{F}_j^{2,\vartheta}(u,v)|\,dv\,du\lesssim\lambda2^{nj}2^{nk}, ∫ ∣ u − u ′ ∣ ≲ 2 k ∫ R n ∣ F j 2 , ϑ ( u , v ) ∣ d v d u ≲ λ 2 nj 2 nk ,
uniformly in ϑ \vartheta ϑ .
Next, from the second part of Lemma 5.4 with h = h j = 2 j − s 4 h=h_j=2^{j-\frac{s}{4}} h = h j = 2 j − 4 s , we have
∣ A j k m , ℓ ( u , v , u ′ , v ′ ) ∣ ≲ 2 − n j 2 − 2 n k ( 1 + 2 j − s 2 ) − 1 2 . |A_{jk}^{m,\ell}(u,v,u',v')|\lesssim2^{-nj}2^{-2nk}(1+2^{j-\frac{s}{2}})^{-\frac{1}{2}}. ∣ A j k m , ℓ ( u , v , u ′ , v ′ ) ∣ ≲ 2 − nj 2 − 2 nk ( 1 + 2 j − 2 s ) − 2 1 .
Combining this bound with (34) , (23) , and ∥ F ~ k ∥ 1 ≲ ∥ F k ∥ 1 \|\widetilde{F}_k\|_1\lesssim\|F_k\|_1 ∥ F k ∥ 1 ≲ ∥ F k ∥ 1 , we obtain
∣ ⟨ V j , V k ⟩ ∣ ≲ γ j γ k 2 − n j 2 − 2 n k ( 1 + 2 j − s 2 ) − 1 2 λ 2 n j 2 n k ∥ F k ∥ 1 |\langle V_j,V_k\rangle|\lesssim\gamma_j\gamma_k2^{-nj}2^{-2nk}(1+2^{j-\frac{s}{2}})^{-\frac{1}{2}}\lambda2^{nj}2^{nk}\|F_k\|_1 ∣ ⟨ V j , V k ⟩ ∣ ≲ γ j γ k 2 − nj 2 − 2 nk ( 1 + 2 j − 2 s ) − 2 1 λ 2 nj 2 nk ∥ F k ∥ 1
≲ λ γ j ( 1 + 2 j − s 2 ) − 1 2 2 − n k ∥ F k ∥ 1 , \lesssim\lambda\gamma_j(1+2^{j-\frac{s}{2}})^{-\frac{1}{2}}2^{-nk}\|F_k\|_1, ≲ λ γ j ( 1 + 2 j − 2 s ) − 2 1 2 − nk ∥ F k ∥ 1 ,
where we used the fact that γ k ≤ 1 \gamma_k\le1 γ k ≤ 1 .
It remains to sum the factors depending on j j j . Since γ j = 2 j − s \gamma_j=2^{j-s} γ j = 2 j − s for j < s j<s j < s and γ j = 1 \gamma_j=1 γ j = 1 for j ≥ s j\ge s j ≥ s , we obtain
∑ j ≥ 1 γ j ( 1 + 2 j − s 2 ) − 1 2 ≲ ∑ 1 ≤ j < s 2 j 2 − 3 s 4 + ∑ j ≥ max { 1 , s } 2 − j 2 + s 4 \sum_{j\ge1}\gamma_j(1+2^{j-\frac{s}{2}})^{-\frac{1}{2}}\lesssim\sum_{1\le j<s}2^{\frac{j}{2}-\frac{3s}{4}}+\sum_{j\ge\max\{1,s\}}2^{-\frac{j}{2}+\frac{s}{4}} j ≥ 1 ∑ γ j ( 1 + 2 j − 2 s ) − 2 1 ≲ 1 ≤ j < s ∑ 2 2 j − 4 3 s + j ≥ m a x { 1 , s } ∑ 2 − 2 j + 4 s
≲ 2 − s 4 . \lesssim2^{-\frac{s}{4}}. ≲ 2 − 4 s .
Hence
∑ k ≥ 1 ∑ j ≤ k − 4 ∣ ⟨ V j , V k ⟩ ∣ ≲ 2 − s 4 λ ∑ k ≥ 1 2 − n k ∥ F k ∥ 1 ≲ 2 − s 4 λ 3 2 . (40) \sum_{k\ge1}\sum_{j\le k-4}|\langle V_j,V_k\rangle|\lesssim2^{-\frac{s}{4}}\lambda\sum_{k\ge1}2^{-nk}\|F_k\|_1\lesssim2^{-\frac{s}{4}}\lambda^{\frac{3}{2}}. \tag*{(40)} k ≥ 1 ∑ j ≤ k − 4 ∑ ∣ ⟨ V j , V k ⟩ ∣ ≲ 2 − 4 s λ k ≥ 1 ∑ 2 − nk ∥ F k ∥ 1 ≲ 2 − 4 s λ 2 3 . (40)
The terms with j ≥ k + 4 j\ge k+4 j ≥ k + 4 have the same bound by symmetry.
Combining (35) , (39) , and (40) , we obtain
∑ j , k ≥ 1 ∣ ⟨ V j , V k ⟩ ∣ ≲ 2 − 4 5 λ 3 2 . \sum_{j,k\geq1} \lvert\langle V_j,V_k\rangle\rvert\lesssim2^{-\frac{4}{5}}\lambda^{\frac{3}{2}}. j , k ≥ 1 ∑ ∣⟨ V j , V k ⟩∣ ≲ 2 − 5 4 λ 2 3 .
This proves (19) and completes the proof of Proposition 5.2. □
Oscillatory kernel estimates
We now prove Lemma 5.4 by estimating the correlations of the dyadic bilinear Bochner–Riesz kernels. Using the Bessel function expansion, we write these correlations as oscillatory integrals. The main difference from the linear case is that the bilinear kernels are defined on R 2 n \mathbb{R}^{2n} R 2 n , while their correlations are integrated over R n \mathbb{R}^{n} R n . For n ≥ 2 n \geq2 n ≥ 2 , following the ideas of Christ [7 ] for the linear case, we obtain the required decay using angular stationary phase. The case n = 1 n=1 n = 1 , which does not arise in the linear case, requires a different approach based on van der Corput’s lemma.
At the critical index, the order of the Bessel function is 2 n − 1 2 2n-\frac{1}{2} 2 n − 2 1 . Using the recurrence formulas for Bessel functions (see [17 Appendix B, Section B.2] ), we obtain the finite expansion
K ( u , v ) = ∑ τ = ± 1 e 2 π i τ ∣ u ∣ 2 + ∣ v ∣ 2 ∑ m = 0 2 n − 1 c τ , m ( ∣ u ∣ 2 + ∣ v ∣ 2 ) − n − m 2 . K(u,v)=\sum_{\tau=\pm1}e^{2\pi i\tau\sqrt{|u|^2+|v|^2}}\sum_{m=0}^{2n-1}c_{\tau,m}(|u|^2+|v|^2)^{-n-\frac{m}{2}}. K ( u , v ) = τ = ± 1 ∑ e 2 π i τ ∣ u ∣ 2 + ∣ v ∣ 2 m = 0 ∑ 2 n − 1 c τ , m ( ∣ u ∣ 2 + ∣ v ∣ 2 ) − n − 2 m .
where c τ , m c_{\tau,m} c τ , m are constants depending only on n n n , τ \tau τ and m m m .
Consequently, after multiplying by the dyadic cutoff ψ ( 2 − j ⋅ ) \psi(2^{-j}\cdot) ψ ( 2 − j ⋅ ) , we can write
K j ( u , v ) = 2 − 2 n j ∑ τ = ± 1 e 2 π i τ ∣ u ∣ 2 + ∣ v ∣ 2 a j , τ ( 2 − j u , 2 − j v ) . (41) K_j(u,v)=2^{-2nj}\sum_{\tau=\pm1}e^{2\pi i\tau\sqrt{|u|^2+|v|^2}}a_{j,\tau}(2^{-j}u,2^{-j}v). \tag*{(41)} K j ( u , v ) = 2 − 2 nj τ = ± 1 ∑ e 2 π i τ ∣ u ∣ 2 + ∣ v ∣ 2 a j , τ ( 2 − j u , 2 − j v ) . (41)
where
a j , τ ( u , v ) = ψ ( u , v ) ∑ m = 0 2 n − 1 c τ , m 2 − j m ( ∣ u ∣ 2 + ∣ v ∣ 2 ) − n − m 2 . a_{j,\tau}(u,v)=\psi(u,v)\sum_{m=0}^{2n-1}c_{\tau,m}2^{-jm}(|u|^2+|v|^2)^{-n-\frac{m}{2}}. a j , τ ( u , v ) = ψ ( u , v ) m = 0 ∑ 2 n − 1 c τ , m 2 − j m ( ∣ u ∣ 2 + ∣ v ∣ 2 ) − n − 2 m .
Since ψ \psi ψ is smooth and supported in B ( 0 , 2 ) ∖ B ( 0 , 1 2 ) B(0,2)\setminus B(0,\frac{1}{2}) B ( 0 , 2 ) ∖ B ( 0 , 2 1 ) , the functions a j , τ a_{j,\tau} a j , τ are smooth, supported in the same annulus, and satisfy
sup j ≥ 1 ∥ a j , τ ∥ C N ≲ 1 (42) \sup_{j\geq1}\|a_{j,\tau}\|_{C^N}\lesssim1 \tag*{(42)} j ≥ 1 sup ∥ a j , τ ∥ C N ≲ 1 (42)
for every nonnegative integer N N N .
Recall that K j , m = K j K_{j,m}=K_j K j , m = K j when j ≥ s j\geq s j ≥ s and m = 1 m=1 m = 1 , whereas for j < s j<s j < s , K j , m = 1 2 ( ∂ u m + ∂ v m ) K j K_{j,m}=\frac{1}{\sqrt{2}}(\partial_{u_m}+\partial_{v_m})K_j K j , m = 2 1 ( ∂ u m + ∂ v m ) K j . Note that the differentiation of (41) changes only the amplitude as
( ∂ u m + ∂ v m ) e 2 π i τ ∣ u ∣ 2 + ∣ v ∣ 2 = 2 π i τ u m + v m ∣ u ∣ 2 + ∣ v ∣ 2 e 2 π i τ ∣ u ∣ 2 + ∣ v ∣ 2 . (\partial_{u_m}+\partial_{v_m})e^{2\pi i\tau\sqrt{|u|^2+|v|^2}}
=2\pi i\tau\frac{u_m+v_m}{\sqrt{|u|^2+|v|^2}}e^{2\pi i\tau\sqrt{|u|^2+|v|^2}}. ( ∂ u m + ∂ v m ) e 2 π i τ ∣ u ∣ 2 + ∣ v ∣ 2 = 2 π i τ ∣ u ∣ 2 + ∣ v ∣ 2 u m + v m e 2 π i τ ∣ u ∣ 2 + ∣ v ∣ 2 .
More precisely, define
b j , τ , m ( u , v ) = { 2 π i τ ( u m + v m ) 2 ∣ u ∣ 2 + ∣ v ∣ 2 a j , τ ( u , v ) + 2 − j 2 ( ∂ u m + ∂ v m ) a j , τ ( u , v ) , j < s , a j , τ ( u , v ) , j ≥ s , m = 1. b_{j,\tau,m}(u,v)=
\begin{cases}
\frac{2\pi i\tau(u_m+v_m)}{\sqrt{2}\sqrt{|u|^2+|v|^2}}a_{j,\tau}(u,v)+\frac{2^{-j}}{\sqrt{2}}(\partial_{u_m}+\partial_{v_m})a_{j,\tau}(u,v), & j<s,\\
a_{j,\tau}(u,v), & j\geq s,\ m=1.
\end{cases} b j , τ , m ( u , v ) = { 2 ∣ u ∣ 2 + ∣ v ∣ 2 2 π i τ ( u m + v m ) a j , τ ( u , v ) + 2 2 − j ( ∂ u m + ∂ v m ) a j , τ ( u , v ) , a j , τ ( u , v ) , j < s , j ≥ s , m = 1.
Then
K j , m ( u , v ) = 2 − 2 n j ∑ τ = ± 1 e 2 π i τ ∣ u ∣ 2 + ∣ v ∣ 2 b j , τ , m ( 2 − j u , 2 − j v ) . K_{j,m}(u,v)=2^{-2nj}\sum_{\tau=\pm1}e^{2\pi i\tau\sqrt{|u|^2+|v|^2}}b_{j,\tau,m}(2^{-j}u,2^{-j}v). K j , m ( u , v ) = 2 − 2 nj τ = ± 1 ∑ e 2 π i τ ∣ u ∣ 2 + ∣ v ∣ 2 b j , τ , m ( 2 − j u , 2 − j v ) .
Since u m + v m ∣ u ∣ 2 + ∣ v ∣ 2 \frac{u_m+v_m}{\sqrt{|u|^2+|v|^2}} ∣ u ∣ 2 + ∣ v ∣ 2 u m + v m is smooth on the support of a j , τ a_{j,\tau} a j , τ , b j , τ , m b_{j,\tau,m} b j , τ , m satisfies the same bounds (42) , uniformly in m m m .
For the product of two kernels, write
a j k , τ , τ ′ m , ℓ ( x , u , v , u ′ , v ′ ) = b j , τ , m ( 2 − j ( x − u ) , − 2 − j v ) b k , τ ′ , ℓ ( 2 − k ( x − u ′ ) , − 2 − k v ′ ) . (43) a^{m,\ell}_{jk,\tau,\tau'}(x,u,v,u',v')=b_{j,\tau,m}(2^{-j}(x-u),-2^{-j}v)b_{k,\tau',\ell}(2^{-k}(x-u'),-2^{-k}v'). \tag*{(43)} a j k , τ , τ ′ m , ℓ ( x , u , v , u ′ , v ′ ) = b j , τ , m ( 2 − j ( x − u ) , − 2 − j v ) b k , τ ′ , ℓ ( 2 − k ( x − u ′ ) , − 2 − k v ′ ) . (43)
Thus, we can rewrite A j k m , ℓ A_{jk}^{m,\ell} A j k m , ℓ in (5.12) as
A j k m , ℓ ( u , v , u ′ , v ′ ) = 2 − 2 n j 2 − 2 n k ∑ τ , τ ′ = ± 1 ∫ R n e 2 π i [ τ ∣ x − u ∣ 2 + ∣ v ∣ 2 − τ ′ ∣ x − u ′ ∣ 2 + ∣ v ′ ∣ 2 ] a j k , τ , τ ′ m , ℓ ( x , u , v , u ′ , v ′ ) d x . (44) A_{jk}^{m,\ell}(u,v,u',v')=2^{-2nj}2^{-2nk}\sum_{\tau,\tau'=\pm1}\int_{\mathbb{R}^n}e^{2\pi i\left[\tau\sqrt{|x-u|^2+|v|^2}-\tau'\sqrt{|x-u'|^2+|v'|^2}\right]}a_{jk,\tau,\tau'}^{m,\ell}(x,u,v,u',v')\,dx. \tag*{(44)} A j k m , ℓ ( u , v , u ′ , v ′ ) = 2 − 2 nj 2 − 2 nk τ , τ ′ = ± 1 ∑ ∫ R n e 2 π i [ τ ∣ x − u ∣ 2 + ∣ v ∣ 2 − τ ′ ∣ x − u ′ ∣ 2 + ∣ v ′ ∣ 2 ] a j k , τ , τ ′ m , ℓ ( x , u , v , u ′ , v ′ ) d x . (44)
The phase in the above oscillatory integral is given by
Φ τ , τ ′ ( x ) = τ ∣ x − u ∣ 2 + ∣ v ∣ 2 − τ ′ ∣ x − u ′ ∣ 2 + ∣ v ′ ∣ 2 , (45) \Phi_{\tau,\tau'}(x)=\tau\sqrt{|x-u|^2+|v|^2}-\tau'\sqrt{|x-u'|^2+|v'|^2}, \tag*{(45)} Φ τ , τ ′ ( x ) = τ ∣ x − u ∣ 2 + ∣ v ∣ 2 − τ ′ ∣ x − u ′ ∣ 2 + ∣ v ′ ∣ 2 , (45)
and a j k , τ , τ ′ m , ℓ a_{jk,\tau,\tau'}^{m,\ell} a j k , τ , τ ′ m , ℓ vanishes unless
$2^{j-1}\leq\sqrt{|x-u|^2+|v|^2}\leq 2\cdot2^j,\qquad 2^{k-1}\leq\sqrt{|x-u'|^2+|v'|^2}\leq 2\cdot2^k.$$
By the chain rule and the bounds for b j , τ , m b_{j,\tau,m} b j , τ , m , we get
∣ ∂ x β a j k , τ , τ ′ m , ℓ ( x , u , v , u ′ , v ′ ) ∣ ≲ 2 − j ∣ β ∣ , k ≥ j . (46) \left|\partial_x^\beta a_{jk,\tau,\tau'}^{m,\ell}(x,u,v,u',v')\right|\lesssim2^{-j|\beta|},\qquad k\geq j. \tag*{(46)} ∂ x β a j k , τ , τ ′ m , ℓ ( x , u , v , u ′ , v ′ ) ≲ 2 − j ∣ β ∣ , k ≥ j . (46)
Proof of Lemma 5.4. We prove the estimates separately for n ≥ 2 n\geq2 n ≥ 2 and n = 1 n=1 n = 1 .
n ≥ 2 n\geq2 n ≥ 2
We first prove the diagonal estimate, followed by the off-diagonal estimate.
Diagonal estimate k = j k=j k = j . The support conditions of a j k , τ , τ ′ m , ℓ a_{jk,\tau,\tau'}^{m,\ell} a j k , τ , τ ′ m , ℓ gives ∣ u − u ′ ∣ ≲ 2 j |u-u'|\lesssim2^j ∣ u − u ′ ∣ ≲ 2 j . Moreover, it restricts x x x to a ball of radius C 2 j C2^j C 2 j . When ∣ u − u ′ ∣ ≤ 1 |u-u'|\leq1 ∣ u − u ′ ∣ ≤ 1 , from (6.3), we get
∣ A j j m , ℓ ∣ ≲ 2 − 3 n j . |A_{jj}^{m,\ell}|\lesssim2^{-3nj}. ∣ A j j m , ℓ ∣ ≲ 2 − 3 nj .
Therefore (5.13) holds in this case. We now assume ∣ u − u ′ ∣ > 1 |u-u'|>1 ∣ u − u ′ ∣ > 1 . By a change of variables into polar coordinates centered at u ′ u' u ′ , we write
x = u ′ + r ω , r > 0 , ω ∈ S n − 1 . x=u'+r\omega,\qquad r>0,\quad\omega\in S^{n-1}. x = u ′ + r ω , r > 0 , ω ∈ S n − 1 .
Then d x = r n − 1 d r d σ ( ω ) dx=r^{n-1}\,dr\,d\sigma(\omega) d x = r n − 1 d r d σ ( ω ) and
Φ τ , τ ′ ( u ′ + r ω ) = τ r 2 + ∣ u − u ′ ∣ 2 − 2 r ω ⋅ ( u − u ′ ) + ∣ v ∣ 2 − τ ′ r 2 + ∣ v ′ ∣ 2 . \Phi_{\tau,\tau'}(u'+r\omega)=\tau\sqrt{r^2+|u-u'|^2-2r\omega\cdot(u-u')+|v|^2}-\tau'\sqrt{r^2+|v'|^2}. Φ τ , τ ′ ( u ′ + r ω ) = τ r 2 + ∣ u − u ′ ∣ 2 − 2 r ω ⋅ ( u − u ′ ) + ∣ v ∣ 2 − τ ′ r 2 + ∣ v ′ ∣ 2 .
Consequently, (6.3) can be written as
A j j m , ℓ ( u , v , u ′ , v ′ ) = 2 − 4 n j ∑ τ , τ ′ = ± 1 ∫ 0 ∞ ∫ S n − 1 e 2 π i [ τ r 2 + ∣ u − u ′ ∣ 2 − 2 r ω ⋅ ( u − u ′ ) + ∣ v ∣ 2 − τ ′ r 2 + ∣ v ′ ∣ 2 ] × a j j , τ , τ ′ m , ℓ ( u ′ + r ω , u , v , u ′ , v ′ ) d σ ( ω ) r n − 1 d r . (47) \begin{aligned}
A_{jj}^{m,\ell}(u,v,u',v')={}&2^{-4nj}\sum_{\tau,\tau'=\pm1}\int_0^\infty\int_{S^{n-1}}e^{2\pi i\left[\tau\sqrt{r^2+|u-u'|^2-2r\omega\cdot(u-u')+|v|^2}-\tau'\sqrt{r^2+|v'|^2}\right]}\\
&\qquad\times a_{jj,\tau,\tau'}^{m,\ell}(u'+r\omega,u,v,u',v')\,d\sigma(\omega)\,r^{n-1}\,dr. \tag*{(47)}
\end{aligned} A j j m , ℓ ( u , v , u ′ , v ′ ) = 2 − 4 nj τ , τ ′ = ± 1 ∑ ∫ 0 ∞ ∫ S n − 1 e 2 π i [ τ r 2 + ∣ u − u ′ ∣ 2 − 2 r ω ⋅ ( u − u ′ ) + ∣ v ∣ 2 − τ ′ r 2 + ∣ v ′ ∣ 2 ] × a j j , τ , τ ′ m , ℓ ( u ′ + r ω , u , v , u ′ , v ′ ) d σ ( ω ) r n − 1 d r . (47)
Since the amplitude is supported where r ≲ 2 j r\lesssim2^j r ≲ 2 j , the radial integral is taken over 0 < r ≲ 2 j 0<r\lesssim2^j 0 < r ≲ 2 j . Let
e = u − u ′ ∣ u − u ′ ∣ . e=\frac{u-u'}{|u-u'|}. e = ∣ u − u ′ ∣ u − u ′ .
Differentiating the phase with respect to ω \omega ω , we obtain
∇ ω Φ τ , τ ′ ( u ′ + r ω ) = − τ r ∣ u − u ′ ∣ r 2 + ∣ u − u ′ ∣ 2 − 2 r ω ⋅ ( u − u ′ ) + ∣ v ∣ 2 ( e − ( e ⋅ ω ) ω ) . \nabla_\omega\Phi_{\tau,\tau'}(u'+r\omega)=-\tau\frac{r|u-u'|}{\sqrt{r^2+|u-u'|^2-2r\omega\cdot(u-u')+|v|^2}}(e-(e\cdot\omega)\omega). ∇ ω Φ τ , τ ′ ( u ′ + r ω ) = − τ r 2 + ∣ u − u ′ ∣ 2 − 2 r ω ⋅ ( u − u ′ ) + ∣ v ∣ 2 r ∣ u − u ′ ∣ ( e − ( e ⋅ ω ) ω ) .
Thus the critical points are ω = ± e \omega=\pm e ω = ± e . Let U U U be a small neighborhood of these critical points. For θ = ( θ 1 , … , θ n − 1 ) \theta=(\theta_1,\ldots,\theta_{n-1}) θ = ( θ 1 , … , θ n − 1 ) and ∣ θ ∣ ≪ 1 |\theta|\ll1 ∣ θ ∣ ≪ 1 , we write ω ∈ U \omega\in U ω ∈ U as
ω = ± e 1 − ∣ θ ∣ 2 + θ 1 e 1 + ⋯ + θ n − 1 e n − 1 , \omega=\pm e\sqrt{1-|\theta|^2}+\theta_1e_1+\cdots+\theta_{n-1}e_{n-1}, ω = ± e 1 − ∣ θ ∣ 2 + θ 1 e 1 + ⋯ + θ n − 1 e n − 1 ,
where e 1 , … , e n − 1 e_1,\ldots,e_{n-1} e 1 , … , e n − 1 form an orthonormal basis of e ⊥ e^\perp e ⊥ . Since ω ⋅ ( u − u ′ ) = ± ∣ u − u ′ ∣ 1 − ∣ θ ∣ 2 \omega\cdot(u-u')=\pm|u-u'|\sqrt{1-|\theta|^2} ω ⋅ ( u − u ′ ) = ± ∣ u − u ′ ∣ 1 − ∣ θ ∣ 2 , differentiating the phase function twice, we get
D θ 2 Φ τ , τ ′ = ± τ r ∣ u − u ′ ∣ R ± ( θ ) ( I n − 1 + O ( ∣ θ ∣ 2 ) ) , D_\theta^2\Phi_{\tau,\tau'}=\pm\tau\frac{r|u-u'|}{R_\pm(\theta)}\left(I_{n-1}+O(|\theta|^2)\right), D θ 2 Φ τ , τ ′ = ± τ R ± ( θ ) r ∣ u − u ′ ∣ ( I n − 1 + O ( ∣ θ ∣ 2 ) ) ,
where I n − 1 I_{n-1} I n − 1 is the identity matrix of order ( n − 1 ) × ( n − 1 ) (n-1)\times(n-1) ( n − 1 ) × ( n − 1 ) and
R ± ( θ ) = r 2 + ∣ u − u ′ ∣ 2 ∓ 2 r ∣ u − u ′ ∣ 1 − ∣ θ ∣ 2 + ∣ v ∣ 2 . R_{\pm}(\theta)=\sqrt{r^2+\lvert u-u'\rvert^2\mp2r\lvert u-u'\rvert\sqrt{1-\lvert\theta\rvert^2}+\lvert v\rvert^2}. R ± ( θ ) = r 2 + ∣ u − u ′ ∣ 2 ∓ 2 r ∣ u − u ′ ∣ 1 − ∣ θ ∣ 2 + ∣ v ∣ 2 .
When θ = 0 \theta=0 θ = 0 , R ± ( 0 ) ∼ 2 j R_{\pm}(0)\sim2^j R ± ( 0 ) ∼ 2 j on the support of a j k , τ , τ ′ m , ℓ a_{jk,\tau,\tau'}^{m,\ell} a j k , τ , τ ′ m , ℓ . Therefore
∣ det D θ 2 Φ τ , τ ′ ( u ′ + r ω ) ∣ θ = 0 ∣ ∼ ( r ∣ u − u ′ ∣ 2 j ) n − 1 . \left\lvert\det D_\theta^2\Phi_{\tau,\tau'}(u'+r\omega)\big\vert_{\theta=0}\right\rvert\sim\left(\frac{r\lvert u-u'\rvert}{2^j}\right)^{n-1}. det D θ 2 Φ τ , τ ′ ( u ′ + r ω ) θ = 0 ∼ ( 2 j r ∣ u − u ′ ∣ ) n − 1 .
When ω ∈ U \omega\in U ω ∈ U , θ \theta θ is close to 0 0 0 . Consequently, R ± ( θ ) ∼ 2 j R_{\pm}(\theta)\sim2^j R ± ( θ ) ∼ 2 j and r ∣ u − u ′ ∣ ≲ 2 2 j r\lvert u-u'\rvert\lesssim2^{2j} r ∣ u − u ′ ∣ ≲ 2 2 j on the support of a j k , τ , τ ′ m , ℓ a_{jk,\tau,\tau'}^{m,\ell} a j k , τ , τ ′ m , ℓ . Thus
∣ det D θ 2 Φ τ , τ ′ ( u ′ + r ω ) ∣ ≳ ( r ∣ u − u ′ ∣ 2 j ) n − 1 . \left\lvert\det D_\theta^2\Phi_{\tau,\tau'}(u'+r\omega)\right\rvert\gtrsim\left(\frac{r\lvert u-u'\rvert}{2^j}\right)^{n-1}. det D θ 2 Φ τ , τ ′ ( u ′ + r ω ) ≳ ( 2 j r ∣ u − u ′ ∣ ) n − 1 .
Observe that from (6.5), we have
∣ ∂ ω β a j j , τ , τ ′ m , ℓ ( u ′ + r ω , u , v , u ′ , v ′ ) ∣ ≲ 1 , r ≲ 2 j . \left\lvert\partial_\omega^\beta a_{jj,\tau,\tau'}^{m,\ell}(u'+r\omega,u,v,u',v')\right\rvert\lesssim1,\qquad r\lesssim2^j. ∂ ω β a j j , τ , τ ′ m , ℓ ( u ′ + r ω , u , v , u ′ , v ′ ) ≲ 1 , r ≲ 2 j .
Also, for every ∣ β ∣ ≥ 1 \lvert\beta\rvert\geq1 ∣ β ∣ ≥ 1 , we have
∣ ∂ ω β Φ τ , τ ′ ( u ′ + r ω ) ∣ ≲ r ∣ u − u ′ ∣ 2 j . \left\lvert\partial_\omega^\beta\Phi_{\tau,\tau'}(u'+r\omega)\right\rvert\lesssim\frac{r\lvert u-u'\rvert}{2^j}. ∂ ω β Φ τ , τ ′ ( u ′ + r ω ) ≲ 2 j r ∣ u − u ′ ∣ .
Thus, when r ∣ u − u ′ ∣ 2 j ≥ 1 \frac{r\lvert u-u'\rvert}{2^j}\geq1 2 j r ∣ u − u ′ ∣ ≥ 1 , by stationary-phase method, we obtain
\left\lvert\int_U e^{2\pi i[\tau\sqrt{r^2+\lvert u-u'\rvert^2-2r\omega\cdot(u-u')+\lvert v\rvert^2-\tau'\sqrt{r^2+\lvert v'\rvert^2}]}a_{jj,\tau,\tau'}^{m,\ell}(u'+r\omega,u,v,u',v')\,d\sigma(\omega)\right\rvert
\lesssim\left(1+\frac{r\lvert u-u'\rvert}{2^j}\right)^{-\frac{n-1}{2}}.}
Away from the critical points on U c U^c U c , we have
∣ ∇ ω Φ τ , τ ′ ( u ′ + r ω ) ∣ ≳ r ∣ u − u ′ ∣ 2 j . \left\lvert\nabla_\omega\Phi_{\tau,\tau'}(u'+r\omega)\right\rvert\gtrsim\frac{r\lvert u-u'\rvert}{2^j}. ∣ ∇ ω Φ τ , τ ′ ( u ′ + r ω ) ∣ ≳ 2 j r ∣ u − u ′ ∣ .
Therefore, when r ∣ u − u ′ ∣ 2 j ≥ 1 \frac{r\lvert u-u'\rvert}{2^j}\geq1 2 j r ∣ u − u ′ ∣ ≥ 1 , by an integration by parts argument, we obtain
\left\lvert\int_{U^c} e^{2\pi i[\tau\sqrt{r^2+\lvert u-u'\rvert^2-2r\omega\cdot(u-u')+\lvert v\rvert^2-\tau'\sqrt{r^2+\lvert v'\rvert^2}]}a_{jj,\tau,\tau'}^{m,\ell}(u'+r\omega,u,v,u',v')\,d\sigma(\omega)\right\rvert
\lesssim\left(1+\frac{r\lvert u-u'\rvert}{2^j}\right)^{-\frac{n-1}{2}}.}
When r ∣ u − u ′ ∣ 2 j < 1 \frac{r\lvert u-u'\rvert}{2^j}<1 2 j r ∣ u − u ′ ∣ < 1 , the above estimate follows directly by integrating (6.3). Combining the estimates for r ∣ u − u ′ ∣ 2 j < 1 \frac{r\lvert u-u'\rvert}{2^j}<1 2 j r ∣ u − u ′ ∣ < 1 and r ∣ u − u ′ ∣ 2 j ≥ 1 \frac{r\lvert u-u'\rvert}{2^j}\geq1 2 j r ∣ u − u ′ ∣ ≥ 1 including both near and away from critical points, we obtain
\left\lvert\int_{S^{n-1}} e^{2\pi i[\tau\sqrt{r^2+\lvert u-u'\rvert^2-2r\omega\cdot(u-u')+\lvert v\rvert^2-\tau'\sqrt{r^2+\lvert v'\rvert^2}]}a_{jj,\tau,\tau'}^{m,\ell}(u'+r\omega,u,v,u',v')\,d\sigma(\omega)\right\rvert
\lesssim\left(1+\frac{r\lvert u-u'\rvert}{2^j}\right)^{-\frac{n-1}{2}}.} \tag*{(48)}
Substituting (6.7) into (6.6) and then splitting the integral in r r r at r = 2 j ∣ u − u ′ ∣ r=\frac{2^j}{\lvert u-u'\rvert} r = ∣ u − u ′ ∣ 2 j , we obtain
∣ A j j m , ℓ ( u , v , u ′ , v ′ ) ∣ ≲ 2 − 4 n j ∫ 0 C 2 j r n − 1 ( 1 + r ∣ u − u ′ ∣ 2 j ) − n − 1 2 d r \left\lvert A_{jj}^{m,\ell}(u,v,u',v')\right\rvert\lesssim2^{-4nj}\int_0^{C2^j}r^{n-1}\left(1+\frac{r\lvert u-u'\rvert}{2^j}\right)^{-\frac{n-1}{2}}\,dr A j j m , ℓ ( u , v , u ′ , v ′ ) ≲ 2 − 4 nj ∫ 0 C 2 j r n − 1 ( 1 + 2 j r ∣ u − u ′ ∣ ) − 2 n − 1 d r
≲ 2 − 4 n j [ ( 2 j ∣ u − u ′ ∣ ) n + ( 2 j ∣ u − u ′ ∣ ) n − 1 2 ∫ 0 C 2 j r n − 1 2 d r ] . \lesssim2^{-4nj}\left[\left(\frac{2^j}{\lvert u-u'\rvert}\right)^n+\left(\frac{2^j}{\lvert u-u'\rvert}\right)^{\frac{n-1}{2}}\int_0^{C2^j}r^{\frac{n-1}{2}}\,dr\right]. ≲ 2 − 4 nj [ ( ∣ u − u ′ ∣ 2 j ) n + ( ∣ u − u ′ ∣ 2 j ) 2 n − 1 ∫ 0 C 2 j r 2 n − 1 d r ] .
≲ 2 − 3 n j ∣ u − u ′ ∣ − n − 1 2 , \lesssim2^{-3nj}|u-u'|^{-\frac{n-1}{2}}, ≲ 2 − 3 nj ∣ u − u ′ ∣ − 2 n − 1 ,
where we used the fact that ∣ u − u ′ ∣ > 1 |u-u'|>1 ∣ u − u ′ ∣ > 1 . This completes the proof of (5.13). Since n ≥ 2 n\ge2 n ≥ 2 and 0 < h ≤ 2 j 0<h\le2^j 0 < h ≤ 2 j , we have
( 1 + ∣ u − u ′ ∣ ) − n − 1 2 ≤ ( 1 + ∣ u − u ′ ∣ h 2 2 2 j ) − 1 2 . (1+|u-u'|)^{-\frac{n-1}{2}}\le\left(1+\frac{|u-u'|h^2}{2^{2j}}\right)^{-\frac{1}{2}}. ( 1 + ∣ u − u ′ ∣ ) − 2 n − 1 ≤ ( 1 + 2 2 j ∣ u − u ′ ∣ h 2 ) − 2 1 .
which also proves (5.17) in this range.
Off-diagonal estimate k ≥ j + 4 k\ge j+4 k ≥ j + 4 . As in the diagonal case, we use polar coordinates centered at u ′ u' u ′ and rewrite (6.3) as
A j k m , ℓ ( u , v , u ′ , v ′ ) = 2 − 2 n j 2 − 2 n k ∑ τ , τ ′ = ± 1 ∫ 0 ∞ ∫ S n − 1 e 2 π i Φ τ , τ ′ ( u ′ + r ω ) a j k , τ , τ ′ m , ℓ ( u ′ + r ω , u , v , u ′ , v ′ ) d σ ( ω ) r n − 1 d r . (49) A_{jk}^{m,\ell}(u,v,u',v')
=2^{-2nj}2^{-2nk}\sum_{\tau,\tau'=\pm1}\int_0^\infty\int_{S^{n-1}}e^{2\pi i\Phi_{\tau,\tau'}(u'+r\omega)}a_{jk,\tau,\tau'}^{m,\ell}(u'+r\omega,u,v,u',v')\,d\sigma(\omega)\,r^{n-1}\,dr. \tag*{(49)} A j k m , ℓ ( u , v , u ′ , v ′ ) = 2 − 2 nj 2 − 2 nk τ , τ ′ = ± 1 ∑ ∫ 0 ∞ ∫ S n − 1 e 2 π i Φ τ , τ ′ ( u ′ + r ω ) a j k , τ , τ ′ m , ℓ ( u ′ + r ω , u , v , u ′ , v ′ ) d σ ( ω ) r n − 1 d r . (49)
From the support of a j k , τ , τ ′ m , ℓ a_{jk,\tau,\tau'}^{m,\ell} a j k , τ , τ ′ m , ℓ , we have
∣ r − ∣ u − u ′ ∣ ∣ ≤ 2 j + 1 , r 2 + ∣ v ′ ∣ 2 ≥ 2 k − 1 . |r-|u-u'||\le2^{j+1},\qquad\sqrt{r^2+|v'|^2}\ge2^{k-1}. ∣ r − ∣ u − u ′ ∣∣ ≤ 2 j + 1 , r 2 + ∣ v ′ ∣ 2 ≥ 2 k − 1 .
We consider the integrals in two cases.
Case 1: ∣ u − u ′ ∣ ≤ 2 j + 2 |u-u'|\le2^{j+2} ∣ u − u ′ ∣ ≤ 2 j + 2 . In this case, we have r ≤ ∣ u − u ′ ∣ + 2 j + 1 ≤ 6 ⋅ 2 j r\le|u-u'|+2^{j+1}\le6\cdot2^j r ≤ ∣ u − u ′ ∣ + 2 j + 1 ≤ 6 ⋅ 2 j . Hence, we can obtain the following estimate similar to (6.7).
∣ ∫ S n − 1 e 2 π i Φ τ , τ ′ ( u ′ + r ω ) a j k , τ , τ ′ m , ℓ ( u ′ + r ω , u , v , u ′ , v ′ ) d σ ( ω ) ∣ ≲ ( 1 + r ∣ u − u ′ ∣ 2 j ) − n − 1 2 . \left|\int_{S^{n-1}}e^{2\pi i\Phi_{\tau,\tau'}(u'+r\omega)}a_{jk,\tau,\tau'}^{m,\ell}(u'+r\omega,u,v,u',v')\,d\sigma(\omega)\right|
\lesssim\left(1+\frac{r|u-u'|}{2^j}\right)^{-\frac{n-1}{2}}. ∫ S n − 1 e 2 π i Φ τ , τ ′ ( u ′ + r ω ) a j k , τ , τ ′ m , ℓ ( u ′ + r ω , u , v , u ′ , v ′ ) d σ ( ω ) ≲ ( 1 + 2 j r ∣ u − u ′ ∣ ) − 2 n − 1 .
With this estimate, it remains to compute the integral in r r r . We first consider
r ≤ 4 max { ∣ u − u ′ ∣ , 2 j 2 } . r\le4\max\{|u-u'|,2^{\frac{j}{2}}\}. r ≤ 4 max { ∣ u − u ′ ∣ , 2 2 j } .
If ∣ u − u ′ ∣ ≤ 2 j 2 |u-u'|\le2^{\frac{j}{2}} ∣ u − u ′ ∣ ≤ 2 2 j , then we have from the trivial bound,
∫ 0 C 2 j 2 r n − 1 d r ≲ 2 j n 2 . \int_0^{C2^{\frac{j}{2}}}r^{n-1}\,dr\lesssim2^{\frac{jn}{2}}. ∫ 0 C 2 2 j r n − 1 d r ≲ 2 2 j n .
If 2 j 2 < ∣ u − u ′ ∣ ≤ 2 j + 2 2^{\frac{j}{2}}<|u-u'|\le2^{j+2} 2 2 j < ∣ u − u ′ ∣ ≤ 2 j + 2 , splitting the integral at r = 2 j ∣ u − u ′ ∣ r=\frac{2^j}{|u-u'|} r = ∣ u − u ′ ∣ 2 j , we obtain
∫ 0 C ∣ u − u ′ ∣ r n − 1 ( 1 + r ∣ u − u ′ ∣ 2 j ) − n − 2 2 d r ≲ ( 2 j ∣ u − u ′ ∣ ) n + ( 2 j ∣ u − u ′ ∣ ) n − 2 2 ∫ 0 C ∣ u − u ′ ∣ r n 2 − 1 d r ≲ 2 n j ∣ u − u ′ ∣ n + 2 j ( n − 2 ) 2 ∣ u − u ′ ∣ ≲ 2 j ( n + 1 ) 2 . \begin{aligned}
\int_0^{C|u-u'|}r^{n-1}\left(1+\frac{r|u-u'|}{2^j}\right)^{-\frac{n-2}{2}}\,dr
&\lesssim\left(\frac{2^j}{|u-u'|}\right)^n
+\left(\frac{2^j}{|u-u'|}\right)^{\frac{n-2}{2}}\int_0^{C|u-u'|}r^{\frac{n}{2}-1}\,dr\\
&\lesssim\frac{2^{nj}}{|u-u'|^n}+2^{\frac{j(n-2)}{2}}|u-u'|\\
&\lesssim2^{\frac{j(n+1)}{2}}.
\end{aligned} ∫ 0 C ∣ u − u ′ ∣ r n − 1 ( 1 + 2 j r ∣ u − u ′ ∣ ) − 2 n − 2 d r ≲ ( ∣ u − u ′ ∣ 2 j ) n + ( ∣ u − u ′ ∣ 2 j ) 2 n − 2 ∫ 0 C ∣ u − u ′ ∣ r 2 n − 1 d r ≲ ∣ u − u ′ ∣ n 2 nj + 2 2 j ( n − 2 ) ∣ u − u ′ ∣ ≲ 2 2 j ( n + 1 ) .
Next, we consider
r ≥ 4 max { ∣ u − u ′ ∣ , 2 j 2 } . r\ge4\max\{|u-u'|,2^{\frac{j}{2}}\}. r ≥ 4 max { ∣ u − u ′ ∣ , 2 2 j } .
From the support of a j k , τ , τ ′ m , ℓ a_{jk,\tau,\tau'}^{m,\ell} a j k , τ , τ ′ m , ℓ , we obtain
τ ∂ r Φ τ , τ ′ ( u ′ + r ω ) = r − ω ⋅ ( u − u ′ ) ∣ r ω − ( u − u ′ ) ∣ 2 + ∣ v ∣ 2 − τ τ ′ r r 2 + ∣ v ′ ∣ 2 ≥ r − ∣ u − u ′ ∣ 2 j + 1 − r 2 k − 1 ≥ r 4 ⋅ 2 j , \begin{aligned}
\tau\partial_r\Phi_{\tau,\tau'}(u'+r\omega)
=\frac{r-\omega\cdot(u-u')}{\sqrt{|r\omega-(u-u')|^2+|v|^2}}
-\tau\tau'\frac{r}{\sqrt{r^2+|v'|^2}} \\
&\ge\frac{r-|u-u'|}{2^{j+1}}-\frac{r}{2^{k-1}}
\ge\frac{r}{4\cdot2^j},
\end{aligned} τ ∂ r Φ τ , τ ′ ( u ′ + r ω ) = ∣ r ω − ( u − u ′ ) ∣ 2 + ∣ v ∣ 2 r − ω ⋅ ( u − u ′ ) − τ τ ′ r 2 + ∣ v ′ ∣ 2 r ≥ 2 j + 1 r − ∣ u − u ′ ∣ − 2 k − 1 r ≥ 4 ⋅ 2 j r ,
where we used the fact that k ≥ j + 4 k \ge j+4 k ≥ j + 4 . Moreover, we have the upper bounds
∣ ∂ r Φ τ , τ ′ ( u ′ + r ω ) ∣ ≲ r 2 2 j , \left|\partial_r\Phi_{\tau,\tau'}(u'+r\omega)\right|\lesssim\frac{r}{2^{2j}}, ∣ ∂ r Φ τ , τ ′ ( u ′ + r ω ) ∣ ≲ 2 2 j r ,
and for every integer M ≥ 2 M\ge2 M ≥ 2 ,
∣ ∂ r M Φ τ , τ ′ ( u ′ + r ω ) ∣ ≲ 2 j ( 1 − M ) . \left|\partial_r^M\Phi_{\tau,\tau'}(u'+r\omega)\right|\lesssim2^{j(1-M)}. ∂ r M Φ τ , τ ′ ( u ′ + r ω ) ≲ 2 j ( 1 − M ) .
Since r ≲ 2 j r\lesssim2^j r ≲ 2 j , differentiating 1 ∂ r Φ τ , τ ′ ( u ′ + r ω ) \frac{1}{\partial_r\Phi_{\tau,\tau'}(u'+r\omega)} ∂ r Φ τ , τ ′ ( u ′ + r ω ) 1 , we obtain
∣ ∂ r M ( 1 ∂ r Φ τ , τ ′ ( u ′ + r ω ) ) ∣ ≲ 2 j r M + 1 , M ≥ 0. \left|\partial_r^M\left(\frac{1}{\partial_r\Phi_{\tau,\tau'}(u'+r\omega)}\right)\right|\lesssim\frac{2^j}{r^{M+1}},\qquad M\ge0. ∂ r M ( ∂ r Φ τ , τ ′ ( u ′ + r ω ) 1 ) ≲ r M + 1 2 j , M ≥ 0.
Also, by (6.5), we have
∣ ∂ r M a j k , τ , τ ′ m , ℓ ( u ′ + r ω , u , v , u ′ , v ′ ) ∣ ≲ 2 − j M ≲ r − M . \left|\partial_r^M a_{jk,\tau,\tau'}^{m,\ell}(u'+r\omega,u,v,u',v')\right|\lesssim2^{-jM}\lesssim r^{-M}. ∂ r M a j k , τ , τ ′ m , ℓ ( u ′ + r ω , u , v , u ′ , v ′ ) ≲ 2 − j M ≲ r − M .
Since e 2 π i Φ τ , τ ′ = 1 2 π i ∂ r Φ τ , τ ′ ∂ r e 2 π i Φ τ , τ ′ e^{2\pi i\Phi_{\tau,\tau'}}=\frac{1}{2\pi i\partial_r\Phi_{\tau,\tau'}}\partial_r e^{2\pi i\Phi_{\tau,\tau'}} e 2 π i Φ τ , τ ′ = 2 π i ∂ r Φ τ , τ ′ 1 ∂ r e 2 π i Φ τ , τ ′ , integrating by parts in r r r along with the above derivative estimates, we obtain
∣ ∫ r ≳ max { ∣ u − u ′ ∣ , 2 j / 2 } e 2 π i Φ τ , τ ′ ( u ′ + r ω ) a j k , τ , τ ′ m , ℓ ( u ′ + r ω , u , v , u ′ , v ′ ) r n − 1 d r ∣ ≲ 2 j N ∫ r ≳ max { ∣ u − u ′ ∣ , 2 j / 2 } r n − 1 − 2 N d r . \left|\int_{r\gtrsim\max\{|u-u'|,2^{j/2}\}} e^{2\pi i\Phi_{\tau,\tau'}(u'+r\omega)}a_{jk,\tau,\tau'}^{m,\ell}(u'+r\omega,u,v,u',v')r^{n-1}\,dr\right|
\lesssim2^{jN}\int_{r\gtrsim\max\{|u-u'|,2^{j/2}\}}r^{n-1-2N}\,dr. ∫ r ≳ m a x { ∣ u − u ′ ∣ , 2 j /2 } e 2 π i Φ τ , τ ′ ( u ′ + r ω ) a j k , τ , τ ′ m , ℓ ( u ′ + r ω , u , v , u ′ , v ′ ) r n − 1 d r ≲ 2 j N ∫ r ≳ m a x { ∣ u − u ′ ∣ , 2 j /2 } r n − 1 − 2 N d r .
We choose N > n 2 N>\frac{n}{2} N > 2 n and obtain
2 j N ∫ r ≳ max { ∣ u − u ′ ∣ , 2 j / 2 } r n − 1 − 2 N d r ≲ 2 j N ( max { ∣ u − u ′ ∣ , 2 j / 2 } ) n − 2 N ≲ 2 j n 2 . 2^{jN}\int_{r\gtrsim\max\{|u-u'|,2^{j/2}\}}r^{n-1-2N}\,dr
\lesssim2^{jN}\left(\max\{|u-u'|,2^{j/2}\}\right)^{n-2N}
\lesssim2^{\frac{jn}{2}}. 2 j N ∫ r ≳ m a x { ∣ u − u ′ ∣ , 2 j /2 } r n − 1 − 2 N d r ≲ 2 j N ( max { ∣ u − u ′ ∣ , 2 j /2 } ) n − 2 N ≲ 2 2 j n .
Since the above bound is uniform in ω \omega ω , integrating over S n − 1 S^{n-1} S n − 1 we obtain the same bound.
Case 2: ∣ u − u ′ ∣ > 2 j + 2 |u-u'|>2^{j+2} ∣ u − u ′ ∣ > 2 j + 2 . Let e = u − u ′ ∣ u − u ′ ∣ e=\frac{u-u'}{|u-u'|} e = ∣ u − u ′ ∣ u − u ′ . From the support of a j k , τ , τ ′ m , ℓ a_{jk,\tau,\tau'}^{m,\ell} a j k , τ , τ ′ m , ℓ , we have
∣ r ω − ( u − u ′ ) ∣ 2 = ( r − ∣ u − u ′ ∣ ) 2 + 2 r ∣ u − u ′ ∣ ( 1 − ω ⋅ e ) ≲ 2 2 j . |r\omega-(u-u')|^2=(r-|u-u'|)^2+2r|u-u'|(1-\omega\cdot e)\lesssim2^{2j}. ∣ r ω − ( u − u ′ ) ∣ 2 = ( r − ∣ u − u ′ ∣ ) 2 + 2 r ∣ u − u ′ ∣ ( 1 − ω ⋅ e ) ≲ 2 2 j .
Thus, r ∼ ∣ u − u ′ ∣ r\sim|u-u'| r ∼ ∣ u − u ′ ∣ and
1 − ω ⋅ e ≲ 2 2 j r ∣ u − u ′ ∣ . 1-\omega\cdot e\lesssim\frac{2^{2j}}{r|u-u'|}. 1 − ω ⋅ e ≲ r ∣ u − u ′ ∣ 2 2 j .
Since ∣ ω ∣ = ∣ e ∣ = 1 |\omega|=|e|=1 ∣ ω ∣ = ∣ e ∣ = 1 , this implies
∣ ω − ( ω ⋅ e ) e ∣ ≲ 2 j r ∣ u − u ′ ∣ . |\omega-(\omega\cdot e)e|\lesssim\frac{2^j}{\sqrt{r|u-u'|}}. ∣ ω − ( ω ⋅ e ) e ∣ ≲ r ∣ u − u ′ ∣ 2 j .
Hence ω \omega ω lies in a small neighborhood of e e e . As in the diagonal case, we choose an orthonormal basis e 1 , … , e n − 1 e_1,\ldots,e_{n-1} e 1 , … , e n − 1 of e ⊥ e^\perp e ⊥ . Rescaling the angular coordinates according to the size of this neighborhood, we write
ω = e 1 − 2 2 j r ∣ u − u ′ ∣ ∣ θ ∣ 2 + 2 j r ∣ u − u ′ ∣ ( θ 1 e 1 + ⋯ + θ n − 1 e n − 1 ) , ∣ θ ∣ ≲ 1. \omega=e\sqrt{1-\frac{2^{2j}}{r|u-u'|}|\theta|^2}+\frac{2^j}{\sqrt{r|u-u'|}}(\theta_1e_1+\cdots+\theta_{n-1}e_{n-1}),\qquad|\theta|\lesssim1. ω = e 1 − r ∣ u − u ′ ∣ 2 2 j ∣ θ ∣ 2 + r ∣ u − u ′ ∣ 2 j ( θ 1 e 1 + ⋯ + θ n − 1 e n − 1 ) , ∣ θ ∣ ≲ 1.
By change of variables in above coordinates, we get
d σ ( ω ) = ( 2 2 j r ∣ u − u ′ ∣ ) n − 1 2 ( 1 − 2 2 j r ∣ u − u ′ ∣ ∣ θ ∣ 2 ) − 1 2 d θ . d\sigma(\omega)=\left(\frac{2^{2j}}{r|u-u'|}\right)^{\frac{n-1}{2}}\left(1-\frac{2^{2j}}{r|u-u'|}|\theta|^2\right)^{-\frac{1}{2}}d\theta. d σ ( ω ) = ( r ∣ u − u ′ ∣ 2 2 j ) 2 n − 1 ( 1 − r ∣ u − u ′ ∣ 2 2 j ∣ θ ∣ 2 ) − 2 1 d θ .
Moreover, in these coordinates, the phase function becomes
Φ τ , τ ′ ( u ′ + r ω ) = τ [ r − ∣ u − u ′ ∣ 2 + ∣ v ∣ 2 + 2 r ∣ u − u ′ ∣ ( 1 − 1 − 2 2 j r ∣ u − u ′ ∣ ∣ θ ∣ 2 ) ] 1 2 − τ ′ r 2 + ∣ v ′ ∣ 2 . \Phi_{\tau,\tau'}(u' + r\omega) = \tau\left[ r - |u-u'|^2 + |v|^2 + 2r|u-u'|\left(1-\sqrt{1-\frac{2^{2j}}{r|u-u'|}|\theta|^2}\right)\right]^{\frac{1}{2}} - \tau'\sqrt{r^2+|v'|^2}. Φ τ , τ ′ ( u ′ + r ω ) = τ [ r − ∣ u − u ′ ∣ 2 + ∣ v ∣ 2 + 2 r ∣ u − u ′ ∣ ( 1 − 1 − r ∣ u − u ′ ∣ 2 2 j ∣ θ ∣ 2 ) ] 2 1 − τ ′ r 2 + ∣ v ′ ∣ 2 .
Differentiating with respect to θ \theta θ , we obtain
∇ θ ( 2 − j Φ τ , τ ′ ( u ′ + r ω ) ) = τ θ 2 − j ∣ r ω − ( u − u ′ ) ∣ 2 + ∣ v ∣ 2 1 − 2 2 j r ∣ u − u ′ ∣ ∣ θ ∣ 2 . \nabla_\theta(2^{-j}\Phi_{\tau,\tau'}(u' + r\omega)) = \frac{\tau\theta}{2^{-j}\sqrt{|r\omega-(u-u')|^2+|v|^2}\sqrt{1-\frac{2^{2j}}{r|u-u'|}|\theta|^2}}. ∇ θ ( 2 − j Φ τ , τ ′ ( u ′ + r ω )) = 2 − j ∣ r ω − ( u − u ′ ) ∣ 2 + ∣ v ∣ 2 1 − r ∣ u − u ′ ∣ 2 2 j ∣ θ ∣ 2 τ θ .
Since the first square root is comparable to 2 j 2^j 2 j on the support of the amplitude, the gradient vanishes only when θ = 0 \theta=0 θ = 0 .
Let U U U be a fixed small neighborhood of θ = 0 \theta=0 θ = 0 . Differentiating once more on the set U U U , we get
D θ 2 ( 2 − j Φ τ , τ ′ ( u ′ + r ω ) ) = τ 2 − j ∣ r ω − ( u − u ′ ) ∣ 2 + ∣ v ∣ 2 ( I n − 1 + O ( ∣ θ ∣ 2 ) ) . D_\theta^2(2^{-j}\Phi_{\tau,\tau'}(u' + r\omega)) = \frac{\tau}{2^{-j}\sqrt{|r\omega-(u-u')|^2+|v|^2}}(I_{n-1}+O(|\theta|^2)). D θ 2 ( 2 − j Φ τ , τ ′ ( u ′ + r ω )) = 2 − j ∣ r ω − ( u − u ′ ) ∣ 2 + ∣ v ∣ 2 τ ( I n − 1 + O ( ∣ θ ∣ 2 )) .
Thus, the Hessian is uniformly nondegenerate in U U U . On U c U^c U c , we have
∣ ∇ θ ( 2 − j Φ τ , τ ′ ( u ′ + r ω ) ) ∣ ≳ ∣ θ ∣ ≥ δ . \left|\nabla_\theta(2^{-j}\Phi_{\tau,\tau'}(u' + r\omega))\right| \gtrsim|\theta| \geq\delta. ∇ θ ( 2 − j Φ τ , τ ′ ( u ′ + r ω )) ≳ ∣ θ ∣ ≥ δ .
Since r ∼ ∣ u − u ′ ∣ r\sim|u-u'| r ∼ ∣ u − u ′ ∣ and 2 j ∣ u − u ′ ∣ ≤ 1 4 \frac{2^j}{|u-u'|}\leq\frac{1}{4} ∣ u − u ′ ∣ 2 j ≤ 4 1 , for every ∣ β ∣ ≥ 1 |\beta|\geq1 ∣ β ∣ ≥ 1 , we have
∣ ∂ θ β ( r ω − ( u − u ′ ) 2 j ) ∣ ≲ r ∣ u − u ′ ∣ ≲ 1. \left|\partial_\theta^\beta\left(\frac{r\omega-(u-u')}{2^j}\right)\right| \lesssim\sqrt{\frac{r}{|u-u'|}}\lesssim1. ∂ θ β ( 2 j r ω − ( u − u ′ ) ) ≲ ∣ u − u ′ ∣ r ≲ 1.
Similarly, using r ≲ 2 k r\lesssim2^k r ≲ 2 k , we have
∣ ∂ θ β ( 2 − k r ω ) ∣ ≲ 1. \left|\partial_\theta^\beta(2^{-k}r\omega)\right|\lesssim1. ∂ θ β ( 2 − k r ω ) ≲ 1.
Consequently, by the chain rule, for every multi-index β \beta β , we have
∣ ∂ θ β a j k , τ , τ ′ m , ℓ ( u ′ + r ω , u , u ′ , v ′ ) ∣ + ∣ ∂ θ β ( 1 − 2 2 j r ∣ u − u ′ ∣ ∣ θ ∣ 2 ) − 1 2 ∣ ≲ 1. \left|\partial_\theta^\beta a_{jk,\tau,\tau'}^{m,\ell}(u'+r\omega,u,u',v')\right|+\left|\partial_\theta^\beta\left(1-\frac{2^{2j}}{r|u-u'|}|\theta|^2\right)^{-\frac{1}{2}}\right|\lesssim1. ∂ θ β a j k , τ , τ ′ m , ℓ ( u ′ + r ω , u , u ′ , v ′ ) + ∂ θ β ( 1 − r ∣ u − u ′ ∣ 2 2 j ∣ θ ∣ 2 ) − 2 1 ≲ 1.
Therefore, applying stationary phase on U U U and integration by parts on U c U^c U c , as in the diagonal case, we obtain
∣ ∫ S n − 1 e 2 π i Φ τ , τ ′ ( u ′ + r ω ) a j k , τ , τ ′ m , ℓ ( u ′ + r ω , u , u ′ , v ′ ) d σ ( ω ) ∣ ≲ ( 2 2 j r ∣ u − u ′ ∣ ) n − 1 2 2 − j ( n − 1 ) 2 . \left|\int_{S^{n-1}}e^{2\pi i\Phi_{\tau,\tau'}(u'+r\omega)}a_{jk,\tau,\tau'}^{m,\ell}(u'+r\omega,u,u',v')\,d\sigma(\omega)\right|\lesssim\left(\frac{2^{2j}}{r|u-u'|}\right)^{\frac{n-1}{2}}2^{-\frac{j(n-1)}{2}}. ∫ S n − 1 e 2 π i Φ τ , τ ′ ( u ′ + r ω ) a j k , τ , τ ′ m , ℓ ( u ′ + r ω , u , u ′ , v ′ ) d σ ( ω ) ≲ ( r ∣ u − u ′ ∣ 2 2 j ) 2 n − 1 2 − 2 j ( n − 1 ) .
Since ∣ r − ∣ u − u ′ ∣ ∣ ≤ 2 j + 1 |r-|u-u'||\leq2^{j+1} ∣ r − ∣ u − u ′ ∣∣ ≤ 2 j + 1 and r ∼ ∣ u − u ′ ∣ r\sim|u-u'| r ∼ ∣ u − u ′ ∣ , integrating in r r r , we get
∫ ∣ r − ∣ u − u ′ ∣ ∣ ≤ 2 j + 1 r n − 1 ( 2 2 j r ∣ u − u ′ ∣ ) n − 1 2 2 − j ( n − 1 ) 2 d r = 2 j ( n − 1 ) 2 ∫ ∣ r − ∣ u − u ′ ∣ ∣ ≤ 2 j + 1 ( r ∣ u − u ′ ∣ ) n − 1 2 d r \int_{|r-|u-u'||\leq2^{j+1}}r^{n-1}\left(\frac{2^{2j}}{r|u-u'|}\right)^{\frac{n-1}{2}}2^{-\frac{j(n-1)}{2}}\,dr
=2^{\frac{j(n-1)}{2}}\int_{|r-|u-u'||\leq2^{j+1}}\left(\frac{r}{|u-u'|}\right)^{\frac{n-1}{2}}\,dr ∫ ∣ r − ∣ u − u ′ ∣∣ ≤ 2 j + 1 r n − 1 ( r ∣ u − u ′ ∣ 2 2 j ) 2 n − 1 2 − 2 j ( n − 1 ) d r = 2 2 j ( n − 1 ) ∫ ∣ r − ∣ u − u ′ ∣∣ ≤ 2 j + 1 ( ∣ u − u ′ ∣ r ) 2 n − 1 d r
≲ 2 j ( n + 1 ) 2 . \lesssim2^{\frac{j(n+1)}{2}}. ≲ 2 2 j ( n + 1 ) .
Combining both cases, each oscillatory integral in (6.3) is bounded by C 2 j ( n + 1 ) 2 C2^{\frac{j(n+1)}{2}} C 2 2 j ( n + 1 ) . Consequently, we obtain
∣ A j k m , ℓ ( u , u ′ , v ′ ) ∣ ≲ 2 − 2 n j 2 − 2 n k 2 j ( n + 1 ) 2 = 2 − n j 2 − 2 n k 2 − j ( n − 1 ) 2 . \begin{aligned}
|A_{jk}^{m,\ell}(u,u',v')|\lesssim2^{-2nj}2^{-2nk}2^{\frac{j(n+1)}{2}} \\
&=2^{-nj}2^{-2nk}2^{-\frac{j(n-1)}{2}}.
\end{aligned} ∣ A j k m , ℓ ( u , u ′ , v ′ ) ∣ ≲ 2 − 2 nj 2 − 2 nk 2 2 j ( n + 1 ) = 2 − nj 2 − 2 nk 2 − 2 j ( n − 1 ) .
This completes the proof of (28) . Now, for n ≥ 2 n\geq2 n ≥ 2 and 0 < h ≤ 2 j 0<h\leq2^j 0 < h ≤ 2 j , we have
2 − j ( n − 1 ) 2 ≲ ( 1 + h 2 2 2 j ) − 1 2 , 2^{-\frac{j(n-1)}{2}}\lesssim\left(1+\frac{h^2}{2^{2j}}\right)^{-\frac{1}{2}}, 2 − 2 j ( n − 1 ) ≲ ( 1 + 2 2 j h 2 ) − 2 1 ,
which proves (32) .
6.2. n = 1 n=1 n = 1 . In this case, (6.3) is
A j k m , ℓ ( u , v , u ′ , v ′ ) = 2 − 2 j 2 − 2 k ∑ τ , τ ′ = ± 1 ∫ R e 2 π i Φ τ , τ ′ ( x ) a j k , τ , τ ′ m , ℓ ( x , u , v , u ′ , v ′ ) d x , (50) A_{jk}^{m,\ell}(u,v,u',v')=2^{-2j}2^{-2k}\sum_{\tau,\tau'=\pm1}\int_{\mathbb{R}}e^{2\pi i\Phi_{\tau,\tau'}(x)}a_{jk,\tau,\tau'}^{m,\ell}(x,u,v,u',v')\,dx, \tag*{(50)} A j k m , ℓ ( u , v , u ′ , v ′ ) = 2 − 2 j 2 − 2 k τ , τ ′ = ± 1 ∑ ∫ R e 2 π i Φ τ , τ ′ ( x ) a j k , τ , τ ′ m , ℓ ( x , u , v , u ′ , v ′ ) d x , (50)
where
Φ τ , τ ′ ( x ) = τ ( x − u ) 2 + v 2 − τ ′ ( x − u ′ ) 2 + ( v ′ ) 2 . \Phi_{\tau,\tau'}(x)=\tau\sqrt{(x-u)^2+v^2}-\tau'\sqrt{(x-u')^2+(v')^2}. Φ τ , τ ′ ( x ) = τ ( x − u ) 2 + v 2 − τ ′ ( x − u ′ ) 2 + ( v ′ ) 2 .
Differentiating the phase function, we obtain
Φ τ , τ ′ ′ ( x ) = τ x − u ( x − u ) 2 + v 2 − τ ′ x − u ′ ( x − u ′ ) 2 + ( v ′ ) 2 , \Phi'_{\tau,\tau'}(x)=\tau\frac{x-u}{\sqrt{(x-u)^2+v^2}}-\tau'\frac{x-u'}{\sqrt{(x-u')^2+(v')^2}}, Φ τ , τ ′ ′ ( x ) = τ ( x − u ) 2 + v 2 x − u − τ ′ ( x − u ′ ) 2 + ( v ′ ) 2 x − u ′ ,
Φ τ , τ ′ ′ ′ ( x ) = τ v 2 ( ( x − u ) 2 + v 2 ) 3 2 − τ ′ ( v ′ ) 2 ( ( x − u ′ ) 2 + ( v ′ ) 2 ) 3 2 . \Phi''_{\tau,\tau'}(x)=\tau\frac{v^2}{((x-u)^2+v^2)^{\frac{3}{2}}}-\tau'\frac{(v')^2}{((x-u')^2+(v')^2)^{\frac{3}{2}}}. Φ τ , τ ′ ′′ ( x ) = τ (( x − u ) 2 + v 2 ) 2 3 v 2 − τ ′ (( x − u ′ ) 2 + ( v ′ ) 2 ) 2 3 ( v ′ ) 2 .
When τ = τ ′ \tau=\tau' τ = τ ′ , we write
Φ ( x ) = ( x − u ) 2 + v 2 − ( x − u ′ ) 2 + ( v ′ ) 2 , \Phi(x)=\sqrt{(x-u)^2+v^2}-\sqrt{(x-u')^2+(v')^2}, Φ ( x ) = ( x − u ) 2 + v 2 − ( x − u ′ ) 2 + ( v ′ ) 2 ,
so that Φ τ , τ = τ Φ \Phi_{\tau,\tau}=\tau\Phi Φ τ , τ = τ Φ .
For fixed u , v ∈ R u,v\in\mathbb{R} u , v ∈ R and j , k ≥ 1 j,k\geq1 j , k ≥ 1 , a j k , τ , τ ′ m , ℓ a_{jk,\tau,\tau'}^{m,\ell} a j k , τ , τ ′ m , ℓ is supported in x x x on the region
I j k = { x : 2 j − 1 ≤ ( x − u ) 2 + v 2 ≤ 2 j + 1 , 2 k − 1 ≤ ( x − u ′ ) 2 + ( v ′ ) 2 ≤ 2 k + 1 } . I_{jk}=\{x:2^{j-1}\leq\sqrt{(x-u)^2+v^2}\leq2^{j+1},\ 2^{k-1}\leq\sqrt{(x-u')^2+(v')^2}\leq2^{k+1}\}. I j k = { x : 2 j − 1 ≤ ( x − u ) 2 + v 2 ≤ 2 j + 1 , 2 k − 1 ≤ ( x − u ′ ) 2 + ( v ′ ) 2 ≤ 2 k + 1 } .
Each shell condition gives at most two intervals. Thus, I j k I_{jk} I j k is a union of at most four intervals, with total length O ( 2 j ) O(2^j) O ( 2 j ) . Note that the support of a j k , τ , τ ′ m , ℓ a_{jk,\tau,\tau'}^{m,\ell} a j k , τ , τ ′ m , ℓ itself need not have finitely many connected components.
On I j k I_{jk} I j k , the two square roots are comparable to 2 j 2^j 2 j and 2 k 2^k 2 k , respectively. Moreover, by (6.5), we have
∥ a j k , τ , τ ′ m , ℓ ∥ ∞ + 2 j ∥ ∂ x a j k , τ , τ ′ m , ℓ ∥ ∞ ≲ 1 , k ≥ j . \lVert a_{jk,\tau,\tau'}^{m,\ell}\rVert_\infty+2^j\lVert\partial_xa_{jk,\tau,\tau'}^{m,\ell}\rVert_\infty\lesssim1,\qquad k\geq j. ∥ a j k , τ , τ ′ m , ℓ ∥ ∞ + 2 j ∥ ∂ x a j k , τ , τ ′ m , ℓ ∥ ∞ ≲ 1 , k ≥ j .
We apply the first and second derivative tests of van der Corput’s lemma.
Diagonal estimate k = j k=j k = j . Suppose ∣ v ∣ , ∣ v ′ ∣ ≥ h |v|,|v'|\geq h ∣ v ∣ , ∣ v ′ ∣ ≥ h . The support conditions of a j k , τ , τ ′ m , ℓ a_{jk,\tau,\tau'}^{m,\ell} a j k , τ , τ ′ m , ℓ give
( x − u ) 2 + v 2 ∼ 2 j , ( x − u ′ ) 2 + ( v ′ ) 2 ∼ 2 j , \sqrt{(x-u)^2+v^2}\sim2^j,\qquad\sqrt{(x-u')^2+(v')^2}\sim2^j, ( x − u ) 2 + v 2 ∼ 2 j , ( x − u ′ ) 2 + ( v ′ ) 2 ∼ 2 j ,
and ∣ u − u ′ ∣ ≲ 2 j |u-u'|\lesssim2^j ∣ u − u ′ ∣ ≲ 2 j . When u = u ′ u=u' u = u ′ , the trivial estimate ∣ A j j m , ℓ ∣ ≲ 2 − 3 j |A_{jj}^{m,\ell}|\lesssim2^{-3j} ∣ A j j m , ℓ ∣ ≲ 2 − 3 j proves (5.15). Hence we assume u ≠ u ′ u\ne u' u = u ′ .
First, suppose τ = − τ ′ \tau=-\tau' τ = − τ ′ . Then
∣ Φ τ , τ ′ ′ ′ ( x ) ∣ = v 2 ( ( x − u ) 2 + v 2 ) 3 2 + ( v ′ ) 2 ( ( x − u ′ ) 2 + ( v ′ ) 2 ) 3 2 ≳ h 2 2 3 j . |\Phi''_{\tau,\tau'}(x)|=\frac{v^2}{((x-u)^2+v^2)^{\frac{3}{2}}}+\frac{(v')^2}{((x-u')^2+(v')^2)^{\frac{3}{2}}}
\gtrsim\frac{h^2}{2^{3j}}. ∣ Φ τ , τ ′ ′′ ( x ) ∣ = (( x − u ) 2 + v 2 ) 2 3 v 2 + (( x − u ′ ) 2 + ( v ′ ) 2 ) 2 3 ( v ′ ) 2 ≳ 2 3 j h 2 .
Since the second derivative has a fixed sign, by van der Corput’s lemma, we get
∣ ∫ R e 2 π i Φ τ , τ ′ ( x ) a j j , τ , τ ′ m , ℓ ( x , u , v , u ′ , v ′ ) d x ∣ ≲ min { 2 j , ( 2 3 j h 2 ) 1 2 } \left|\int_{\mathbb{R}}e^{2\pi i\Phi_{\tau,\tau'}(x)}a_{jj,\tau,\tau'}^{m,\ell}(x,u,v,u',v')\,dx\right|
\lesssim\min\left\{2^j,\left(\frac{2^{3j}}{h^2}\right)^{\frac{1}{2}}\right\} ∫ R e 2 π i Φ τ , τ ′ ( x ) a j j , τ , τ ′ m , ℓ ( x , u , v , u ′ , v ′ ) d x ≲ min { 2 j , ( h 2 2 3 j ) 2 1 }
≲ 2 j ( 1 + h 2 2 j ) − 1 2 \lesssim2^j\left(1+\frac{h^2}{2^j}\right)^{-\frac{1}{2}} ≲ 2 j ( 1 + 2 j h 2 ) − 2 1
≲ 2 j ( 1 + ∣ u − u ′ ∣ h 2 2 2 j ) − 1 2 , \lesssim2^j\left(1+\frac{|u-u'|h^2}{2^{2j}}\right)^{-\frac{1}{2}}, ≲ 2 j ( 1 + 2 2 j ∣ u − u ′ ∣ h 2 ) − 2 1 ,
where we used ∣ u − u ′ ∣ ≲ 2 j |u-u'|\lesssim2^j ∣ u − u ′ ∣ ≲ 2 j in the last inequality.
Next, suppose τ = τ ′ \tau=\tau' τ = τ ′ . Observe that
x − u ( x − u ) 2 + v 2 1 − ( x − u ) 2 ( x − u ) 2 + v 2 v 2 ( ( x − u ) 2 + v 2 ) 3 / 2 = x − u . (51) \frac{x-u}{\sqrt{(x-u)^2+v^2}}\frac{1-\frac{(x-u)^2}{(x-u)^2+v^2}}{\frac{v^2}{((x-u)^2+v^2)^{3/2}}}=x-u. \tag*{(51)} ( x − u ) 2 + v 2 x − u (( x − u ) 2 + v 2 ) 3/2 v 2 1 − ( x − u ) 2 + v 2 ( x − u ) 2 = x − u . (51)
The same identity holds with ( u , v ) (u,v) ( u , v ) replaced by ( u ′ , v ′ ) (u',v') ( u ′ , v ′ ) . Since r ( 1 − r 2 ) r(1-r^2) r ( 1 − r 2 ) is Lipschitz on [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] and
v 2 ( ( x − u ) 2 + v 2 ) 3 / 2 , ( v ′ ) 2 ( ( x − u ′ ) 2 + ( v ′ ) 2 ) 3 / 2 ≳ h 2 2 3 j , \frac{v^2}{((x-u)^2+v^2)^{3/2}},\ \frac{(v')^2}{((x-u')^2+(v')^2)^{3/2}}\gtrsim\frac{h^2}{2^{3j}}, (( x − u ) 2 + v 2 ) 3/2 v 2 , (( x − u ′ ) 2 + ( v ′ ) 2 ) 3/2 ( v ′ ) 2 ≳ 2 3 j h 2 ,
from (51) , we get
∣ u − u ′ ∣ ≤ ∣ x − u − ( x − u ′ ) ∣ |u-u'|\leq|x-u-(x-u')| ∣ u − u ′ ∣ ≤ ∣ x − u − ( x − u ′ ) ∣
≲ 2 3 j h 2 ( ∣ Φ ′ ( x ) ∣ + ∣ x − u ′ ∣ ∣ Φ ′ ′ ( x ) ∣ ) \lesssim\frac{2^{3j}}{h^2}\left(|\Phi'(x)|+|x-u'|\lvert\Phi''(x)\rvert\right) ≲ h 2 2 3 j ( ∣ Φ ′ ( x ) ∣ + ∣ x − u ′ ∣ ∣ Φ ′′ ( x )∣ )
≲ 2 3 j h 2 ( ∣ Φ ′ ( x ) ∣ + 2 j ∣ Φ ′ ′ ( x ) ∣ ) . \lesssim\frac{2^{3j}}{h^2}\left(|\Phi'(x)|+2^j|\Phi''(x)|\right). ≲ h 2 2 3 j ( ∣ Φ ′ ( x ) ∣ + 2 j ∣ Φ ′′ ( x ) ∣ ) .
Thus
∣ Φ ′ ( x ) ∣ + 2 j ∣ Φ ′ ′ ( x ) ∣ ≳ ∣ u − u ′ ∣ h 2 2 3 j . (52) |\Phi'(x)|+2^j|\Phi''(x)|\gtrsim\frac{|u-u'|h^2}{2^{3j}}. \tag*{(52)} ∣ Φ ′ ( x ) ∣ + 2 j ∣ Φ ′′ ( x ) ∣ ≳ 2 3 j ∣ u − u ′ ∣ h 2 . (52)
Next, the zeros of Φ ′ ′ \Phi'' Φ ′′ satisfy the equation
∣ v ∣ 4 3 ( ( x − u ′ ) 2 + ( v ′ ) 2 ) = ∣ v ′ ∣ 4 3 ( ( x − u ) 2 + v 2 ) . |v|^{\frac{4}{3}}\left((x-u')^2+(v')^2\right)=|v'|^{\frac{4}{3}}\left((x-u)^2+v^2\right). ∣ v ∣ 3 4 ( ( x − u ′ ) 2 + ( v ′ ) 2 ) = ∣ v ′ ∣ 3 4 ( ( x − u ) 2 + v 2 ) .
This is a quadratic equation in x x x . Since u ≠ u ′ u\ne u' u = u ′ , it cannot hold identically. Thus, Φ ′ ′ \Phi'' Φ ′′ has at most two zeros. Dividing the intervals of I j j I_{jj} I j j at these zeros, we obtain a bounded number of intervals on which Φ ′ \Phi' Φ ′ is monotone. Moreover, the support conditions and (52) hold on each of these intervals.
If ∣ u − u ′ ∣ h 2 ≤ 2 2 j |u-u'|h^2\leq2^{2j} ∣ u − u ′ ∣ h 2 ≤ 2 2 j , the trivial bound ∣ A j j m , ℓ ∣ ≲ 2 − 3 j |A_{jj}^{m,\ell}|\lesssim2^{-3j} ∣ A j j m , ℓ ∣ ≲ 2 − 3 j is sufficient. Otherwise, we divide each of these intervals into the regions where
∣ Φ ′ ( x ) ∣ ≥ c ∣ u − u ′ ∣ h 2 2 3 j and ∣ Φ ′ ( x ) ∣ < c ∣ u − u ′ ∣ h 2 2 3 j , |\Phi'(x)|\geq c\frac{|u-u'|h^2}{2^{3j}}
\qquad\text{and}\qquad
|\Phi'(x)|<c\frac{|u-u'|h^2}{2^{3j}}, ∣ Φ ′ ( x ) ∣ ≥ c 2 3 j ∣ u − u ′ ∣ h 2 and ∣ Φ ′ ( x ) ∣ < c 2 3 j ∣ u − u ′ ∣ h 2 ,
where c > 0 c>0 c > 0 is sufficiently small.
On the first set, by the first derivative test we get
∣ ∫ e 2 π i τ Φ ( x ) a j j , τ , τ m , ℓ ( x , u , v , u ′ , v ′ ) d x ∣ ≲ 2 3 j ∣ u − u ′ ∣ h 2 . \left|\int e^{2\pi i\tau\Phi(x)}a_{jj,\tau,\tau}^{m,\ell}(x,u,v,u',v')\,dx\right|\lesssim\frac{2^{3j}}{|u-u'|h^2}. ∫ e 2 π i τ Φ ( x ) a j j , τ , τ m , ℓ ( x , u , v , u ′ , v ′ ) d x ≲ ∣ u − u ′ ∣ h 2 2 3 j .
On the second set, from (52) , we have
∣ Φ ′ ′ ( x ) ∣ ≳ ∣ u − u ′ ∣ h 2 2 4 j . |\Phi''(x)|\gtrsim\frac{|u-u'|h^2}{2^{4j}}. ∣ Φ ′′ ( x ) ∣ ≳ 2 4 j ∣ u − u ′ ∣ h 2 .
Thus, by the second derivative test, we get
∣ ∫ e 2 π i τ Φ ( x ) a j j , τ , τ m , ℓ ( x , u , v , u ′ , v ′ ) d x ∣ ≲ ( 2 4 j ∣ u − u ′ ∣ h 2 ) 1 2 . \left|\int e^{2\pi i\tau\Phi(x)}a_{jj,\tau,\tau}^{m,\ell}(x,u,v,u',v')\,dx\right|\lesssim\left(\frac{2^{4j}}{|u-u'|h^2}\right)^{\frac{1}{2}}. ∫ e 2 π i τ Φ ( x ) a j j , τ , τ m , ℓ ( x , u , v , u ′ , v ′ ) d x ≲ ( ∣ u − u ′ ∣ h 2 2 4 j ) 2 1 .
Since ∣ u − u ′ ∣ h 2 > 2 2 j |u-u'|h^2>2^{2j} ∣ u − u ′ ∣ h 2 > 2 2 j , the second bound is larger than the first. Combining these estimates, we obtain
∣ ∫ R e 2 π i τ Φ ( x ) a j j , τ , τ m , ℓ ( x , u , v , u ′ , v ′ ) d x ∣ ≲ 2 j ( 1 + ∣ u − u ′ ∣ h 2 2 2 j ) − 1 2 . \left|\int_{\mathbb{R}}e^{2\pi i\tau\Phi(x)}a_{jj,\tau,\tau}^{m,\ell}(x,u,v,u',v')\,dx\right|\lesssim2^j\left(1+\frac{|u-u'|h^2}{2^{2j}}\right)^{-\frac{1}{2}}. ∫ R e 2 π i τ Φ ( x ) a j j , τ , τ m , ℓ ( x , u , v , u ′ , v ′ ) d x ≲ 2 j ( 1 + 2 2 j ∣ u − u ′ ∣ h 2 ) − 2 1 .
Substituting the estimates for both sign choices into (50) proves (29) .
Off-diagonal estimate k ≥ j + 4 k\geq j+4 k ≥ j + 4 . Suppose ∣ v ∣ ≥ h |v|\geq h ∣ v ∣ ≥ h . We follow the same argument as in the diagonal case.
First, suppose τ = − τ ′ \tau=-\tau' τ = − τ ′ . Then
∣ Φ τ , τ ′ ′ ′ ( x ) ∣ = v 2 ( ( x − u ) 2 + v 2 ) 3 2 + ( v ′ ) 2 ( ( x − u ′ ) 2 + ( v ′ ) 2 ) 3 2 ≳ h 2 2 3 j . \left|\Phi_{\tau,\tau'}^{\prime\prime}(x)\right|=\frac{v^2}{\left((x-u)^2+v^2\right)^{\frac{3}{2}}}+\frac{(v')^2}{\left((x-u')^2+(v')^2\right)^{\frac{3}{2}}}
\gtrsim\frac{h^2}{2^{3j}}. Φ τ , τ ′ ′′ ( x ) = ( ( x − u ) 2 + v 2 ) 2 3 v 2 + ( ( x − u ′ ) 2 + ( v ′ ) 2 ) 2 3 ( v ′ ) 2 ≳ 2 3 j h 2 .
Thus, by the second derivative test and the trivial bound, we get
∣ ∫ R e 2 π i Φ τ , τ ′ ( x ) a j k , τ , τ ′ m , ℓ ( x , u , v , u ′ , v ′ ) d x ∣ ≲ 2 j ( 1 + h 2 2 j ) − 1 2 . \left|\int_{\mathbb{R}}e^{2\pi i\Phi_{\tau,\tau'}(x)}a_{jk,\tau,\tau'}^{m,\ell}(x,u,v,u',v')\,dx\right|\lesssim2^j\left(1+\frac{h^2}{2^j}\right)^{-\frac{1}{2}}. ∫ R e 2 π i Φ τ , τ ′ ( x ) a j k , τ , τ ′ m , ℓ ( x , u , v , u ′ , v ′ ) d x ≲ 2 j ( 1 + 2 j h 2 ) − 2 1 .
Now suppose τ = τ ′ \tau=\tau' τ = τ ′ . Observe that
( v ′ ) 2 ( ( x − u ′ ) 2 + ( v ′ ) 2 ) 3 2 = 1 ( x − u ′ ) 2 + ( v ′ ) 2 ( 1 − ( x − u ′ ) 2 ( x − u ′ ) 2 + ( v ′ ) 2 ) . \frac{(v')^2}{\left((x-u')^2+(v')^2\right)^{\frac{3}{2}}}=\frac{1}{\sqrt{(x-u')^2+(v')^2}}\left(1-\frac{(x-u')^2}{(x-u')^2+(v')^2}\right). ( ( x − u ′ ) 2 + ( v ′ ) 2 ) 2 3 ( v ′ ) 2 = ( x − u ′ ) 2 + ( v ′ ) 2 1 ( 1 − ( x − u ′ ) 2 + ( v ′ ) 2 ( x − u ′ ) 2 ) .
Using ∣ r 2 − ( r ′ ) 2 ∣ ≤ 2 ∣ r − r ′ ∣ |r^2-(r')^2|\leq2|r-r'| ∣ r 2 − ( r ′ ) 2 ∣ ≤ 2∣ r − r ′ ∣ for ∣ r ∣ , ∣ r ′ ∣ ≤ 1 |r|,|r'|\leq1 ∣ r ∣ , ∣ r ′ ∣ ≤ 1 and the support conditions of a j k , τ , τ ′ m , ℓ a_{jk,\tau,\tau'}^{m,\ell} a j k , τ , τ ′ m , ℓ , we obtain
( v ′ ) 2 ( ( x − u ′ ) 2 + ( v ′ ) 2 ) 3 2 ≤ 2 2 k ( 1 − ( x − u ) 2 ( x − u ) 2 + v 2 + 2 ∣ Φ ′ ( x ) ∣ ) \frac{(v')^2}{\left((x-u')^2+(v')^2\right)^{\frac{3}{2}}}\leq\frac{2}{2^k}\left(1-\frac{(x-u)^2}{(x-u)^2+v^2}+2|\Phi'(x)|\right) ( ( x − u ′ ) 2 + ( v ′ ) 2 ) 2 3 ( v ′ ) 2 ≤ 2 k 2 ( 1 − ( x − u ) 2 + v 2 ( x − u ) 2 + 2∣ Φ ′ ( x ) ∣ )
≤ 4 ⋅ 2 j − k v 2 ( ( x − u ) 2 + v 2 ) 3 2 + 4 2 k ∣ Φ ′ ( x ) ∣ ≤ 1 4 v 2 ( ( x − u ) 2 + v 2 ) 3 2 + 1 4 ⋅ 2 j ∣ Φ ′ ( x ) ∣ , \begin{aligned}
\leq4\cdot2^{j-k}\frac{v^2}{\left((x-u)^2+v^2\right)^{\frac{3}{2}}}+\frac{4}{2^k}|\Phi'(x)| \\
&\leq\frac{1}{4}\frac{v^2}{\left((x-u)^2+v^2\right)^{\frac{3}{2}}}+\frac{1}{4\cdot2^j}|\Phi'(x)|,
\end{aligned} ≤ 4 ⋅ 2 j − k ( ( x − u ) 2 + v 2 ) 2 3 v 2 + 2 k 4 ∣ Φ ′ ( x ) ∣ ≤ 4 1 ( ( x − u ) 2 + v 2 ) 2 3 v 2 + 4 ⋅ 2 j 1 ∣ Φ ′ ( x ) ∣ ,
where we used k ≥ j + 4 k\geq j+4 k ≥ j + 4 in the last inequality. Consequently, we get
Φ ′ ′ ( x ) ≥ 3 4 v 2 ( ( x − u ) 2 + v 2 ) 3 2 − 1 4 ⋅ 2 j ∣ Φ ′ ( x ) ∣ . \Phi^{\prime\prime}(x)\geq\frac{3}{4}\frac{v^2}{\left((x-u)^2+v^2\right)^{\frac{3}{2}}}-\frac{1}{4\cdot2^j}|\Phi'(x)|. Φ ′′ ( x ) ≥ 4 3 ( ( x − u ) 2 + v 2 ) 2 3 v 2 − 4 ⋅ 2 j 1 ∣ Φ ′ ( x ) ∣.
Since v 2 ( ( x − u ) 2 + v 2 ) 3 2 ≳ h 2 2 3 j \frac{v^2}{\left((x-u)^2+v^2\right)^{\frac{3}{2}}}\gtrsim\frac{h^2}{2^{3j}} ( ( x − u ) 2 + v 2 ) 2 3 v 2 ≳ 2 3 j h 2 , we get
∣ Φ ′ ( x ) ∣ + 2 j ∣ Φ ′ ′ ( x ) ∣ ≳ h 2 2 2 j . (53) |\Phi'(x)|+2^j|\Phi^{\prime\prime}(x)|\gtrsim\frac{h^2}{2^{2j}}. \tag*{(53)} ∣ Φ ′ ( x ) ∣ + 2 j ∣ Φ ′′ ( x ) ∣ ≳ 2 2 j h 2 . (53)
As in the diagonal case, the equation Φ ′ ′ ( x ) = 0 \Phi^{\prime\prime}(x)=0 Φ ′′ ( x ) = 0 is quadratic in x x x when v ′ ≠ 0 v'\neq0 v ′ = 0 . It cannot hold identically, since this would imply u = u ′ u=u' u = u ′ and ∣ v ∣ = ∣ v ′ ∣ |v|=|v'| ∣ v ∣ = ∣ v ′ ∣ , which is not possible for k ≥ j + 4 k\geq j+4 k ≥ j + 4 . If v ′ = 0 v'=0 v ′ = 0 , then Φ ′ ′ ( x ) > 0 \Phi^{\prime\prime}(x)>0 Φ ′′ ( x ) > 0 . As in the diagonal case, we divide the intervals of I j k I_{jk} I j k at the zeros of Φ ′ ′ \Phi^{\prime\prime} Φ ′′ . This gives a bounded number of intervals on which Φ ′ \Phi' Φ ′ is monotone. Moreover, the support conditions and (6.12) hold on each of these intervals.
If h 2 ≤ 2 j h^2\leq2^j h 2 ≤ 2 j , the trivial bound ∣ A j k m , ℓ ∣ ≲ 2 − j 2 − 2 k |A_{jk}^{m,\ell}|\lesssim2^{-j}2^{-2k} ∣ A j k m , ℓ ∣ ≲ 2 − j 2 − 2 k is enough. Otherwise, on each monotonic interval, we consider the sets where
∣ Φ ′ ( x ) ∣ ≥ c h 2 2 2 j and ∣ Φ ′ ( x ) ∣ < c h 2 2 2 j . |\Phi'(x)|\geq c\frac{h^2}{2^{2j}}
\qquad\text{and}\qquad
|\Phi'(x)|<c\frac{h^2}{2^{2j}}. ∣ Φ ′ ( x ) ∣ ≥ c 2 2 j h 2 and ∣ Φ ′ ( x ) ∣ < c 2 2 j h 2 .
On the first set, by the first derivative test we get
∣ ∫ e 2 π i τ Φ ( x ) a j k , τ , τ m , ℓ ( x , u , v , u ′ , v ′ ) d x ∣ ≲ 2 2 j h 2 . \left|\int e^{2\pi i\tau\Phi(x)}a_{jk,\tau,\tau}^{m,\ell}(x,u,v,u',v')\,dx\right|\lesssim\frac{2^{2j}}{h^2}. ∫ e 2 π i τ Φ ( x ) a j k , τ , τ m , ℓ ( x , u , v , u ′ , v ′ ) d x ≲ h 2 2 2 j .
On the second set, (6.12) implies
∣ Φ ′ ′ ( x ) ∣ ≳ h 2 2 3 j , |\Phi^{\prime\prime}(x)|\gtrsim\frac{h^2}{2^{3j}}, ∣ Φ ′′ ( x ) ∣ ≳ 2 3 j h 2 ,
and by the second derivative test we get
∣ ∫ e 2 π i τ Φ ( x ) a j k , τ , τ m , ℓ ( x , u , v , u ′ , v ′ ) d x ∣ ≲ ( 2 3 j h 2 ) 1 2 . \left|\int e^{2\pi i\tau\Phi(x)}a_{jk,\tau,\tau}^{m,\ell}(x,u,v,u',v')\,dx\right|\lesssim\left(\frac{2^{3j}}{h^2}\right)^{\frac{1}{2}}. ∫ e 2 π i τ Φ ( x ) a j k , τ , τ m , ℓ ( x , u , v , u ′ , v ′ ) d x ≲ ( h 2 2 3 j ) 2 1 .
Since h 2 > 2 j h^{2} > 2^{j} h 2 > 2 j , the second estimate is larger than the first. Combining both estimates, we obtain
∣ ∫ R e 2 π i τ Φ ( x ) m j k , τ , τ ′ ℓ ( x , u , v , u ′ , v ′ ) d x ∣ ≲ 2 j ( 1 + h 2 2 j ) − 1 2 . \left| \int_{\mathbb{R}} e^{2\pi i \tau\Phi(x)} m_{jk,\tau,\tau'}^{\ell}(x,u,v,u',v')\,dx \right| \lesssim2^{j}\left(1+\frac{h^{2}}{2^{j}}\right)^{-\frac{1}{2}}. ∫ R e 2 π i τ Φ ( x ) m j k , τ , τ ′ ℓ ( x , u , v , u ′ , v ′ ) d x ≲ 2 j ( 1 + 2 j h 2 ) − 2 1 .
Substituting the estimates for both sign choices into (50) proves (30) .
All the estimates are uniform in m m m , ℓ \ell ℓ and τ , τ ′ \tau,\tau' τ , τ ′ . This completes the proof of Lemma 5.4. □
Acknowledgements. Surjeet Singh Choudhary is supported by the National Center for Theoretical Sciences through National Science and Technology Council of Taiwan under research grant number 115-2124-M-002-009-. Saurabh Shrivastava acknowledges the support from Anusandhan National Research Foundation (ANRF), India under the projects ANRF/ARG/2025/000940/MS.
Surjeet Singh Choudhary, National Center for Theoretical Sciences, National Taiwan University, Taipei-106, Taiwan. Email address: surjeet19@ncts.ntu.edu.tw
Saurabh Shrivastava, Department of Mathematics, Indian Institute of Science Education and Research Bhopal, Bhopal-462066, India. Email address: saurabhk@iiserb.ac.in