Introduction
Following Bourne, Cagnetti and Roper [4 Eqs. (1.1)–(1.4)] , let A = ( t 1 ∣ t 2 ∣ t 3 ) ∈ G L ( 3 , R ) A=(t_1\mid t_2\mid t_3)\in GL(3,\mathbb{R}) A = ( t 1 ∣ t 2 ∣ t 3 ) ∈ G L ( 3 , R ) and write
Λ ( A ) = span Z { t 1 , t 2 , t 3 } = A Z 3 , V ( A ) = { x ∈ R 3 : ∣ x ∣ ≤ ∣ x − w ∣ for every w ∈ Λ ( A ) } . \Lambda(A)=\operatorname{span}_{\mathbb{Z}}\{t_1,t_2,t_3\}=A\mathbb{Z}^3,\qquad V(A)=\{x\in\mathbb{R}^3:|x|\leq|x-w|\text{ for every }w\in\Lambda(A)\}. Λ ( A ) = span Z { t 1 , t 2 , t 3 } = A Z 3 , V ( A ) = { x ∈ R 3 : ∣ x ∣ ≤ ∣ x − w ∣ for every w ∈ Λ ( A )} .
The volume ∣ V ( A ) ∣ = ∣ det A ∣ |V(A)|=|\det A| ∣ V ( A ) ∣ = ∣ det A ∣ is the covolume of Λ ( A ) \Lambda(A) Λ ( A ) . For p ≥ 1 p\geq1 p ≥ 1 define the normalized Voronoi moment by
E p ( A ) = ∣ V ( A ) ∣ − 1 − p / 3 ∫ V ( A ) ∣ x ∣ p d x . E_p(A)=|V(A)|^{-1-p/3}\int_{V(A)}|x|^p\,dx. E p ( A ) = ∣ V ( A ) ∣ − 1 − p /3 ∫ V ( A ) ∣ x ∣ p d x .
For c > 0 c>0 c > 0 , O ∈ O ( 3 ) O\in O(3) O ∈ O ( 3 ) and U ∈ G L ( 3 , Z ) U\in GL(3,\mathbb{Z}) U ∈ G L ( 3 , Z ) , V ( c O A U ) = c O V ( A ) V(cOAU)=cOV(A) V ( c O A U ) = c O V ( A ) , so
E p ( c O A U ) = ( c 3 ∣ V ( A ) ∣ ) − 1 − p / 3 c p + 3 ∫ V ( A ) ∣ x ∣ p d x = E p ( A ) . (1) E_p(cOAU)=(c^3|V(A)|)^{-1-p/3}c^{p+3}\int_{V(A)}|x|^p\,dx=E_p(A). \tag*{(1)} E p ( c O A U ) = ( c 3 ∣ V ( A ) ∣ ) − 1 − p /3 c p + 3 ∫ V ( A ) ∣ x ∣ p d x = E p ( A ) . (1)
Taking c = ∣ V ( A ) ∣ − 1 / 3 c=|V(A)|^{-1/3} c = ∣ V ( A ) ∣ − 1/3 reduces the minimization to ∣ V ( A ) ∣ = 1 |V(A)|=1 ∣ V ( A ) ∣ = 1 .
Use the BCC basis of [4 Eq. (4.13)] ,
A b c c = ( 1 0 1 / 2 0 1 1 / 2 0 0 1 / 2 ) , D 3 ∗ = Λ ( A b c c ) . A_{\mathrm{bcc}}=
\begin{pmatrix}
1 & 0 & 1/2\\
0 & 1 & 1/2\\
0 & 0 & 1/2
\end{pmatrix},\qquad D_3^*=\Lambda(A_{\mathrm{bcc}}). A bcc = 1 0 0 0 1 0 1/2 1/2 1/2 , D 3 ∗ = Λ ( A bcc ) .
Here ∣ V ( A b c c ) ∣ = 1 / 2 |V(A_{\mathrm{bcc}})|=1/2 ∣ V ( A bcc ) ∣ = 1/2 . For a lattice L = Λ ( A ) L=\Lambda(A) L = Λ ( A ) , write ∣ L ∣ = ∣ V ( A ) ∣ |L|=|V(A)| ∣ L ∣ = ∣ V ( A ) ∣ and E p ( L ) = E p ( A ) E_p(L)=E_p(A) E p ( L ) = E p ( A ) . This value is independent of the choice of basis by (1) . Set
L B = 2 1 / 3 D 3 ∗ = Λ ( 2 1 / 3 A b c c ) , ∣ L B ∣ = 1. L_{\mathrm{B}}=2^{1/3}D_3^*=\Lambda(2^{1/3}A_{\mathrm{bcc}}),\qquad|L_{\mathrm{B}}|=1. L B = 2 1/3 D 3 ∗ = Λ ( 2 1/3 A bcc ) , ∣ L B ∣ = 1.
Theorem 1.1. Let 1 ≤ p < ∞ 1\leq p<\infty 1 ≤ p < ∞ . Among full-rank lattices L ⊂ R 3 L\subset\mathbb{R}^3 L ⊂ R 3 of covolume one,
min ∣ L ∣ = 1 E p ( L ) = E p ( L B ) = E p ( A b c c ) , argmin ∣ L ∣ = 1 E p ( L ) = { O L B : O ∈ O ( 3 ) } . (2) \min_{|L|=1}E_p(L)=E_p(L_{\mathrm{B}})=E_p(A_{\mathrm{bcc}}),\qquad\underset{|L|=1}{\operatorname{argmin}}E_p(L)=\{OL_{\mathrm{B}}:O\in O(3)\}. \tag*{(2)} ∣ L ∣ = 1 min E p ( L ) = E p ( L B ) = E p ( A bcc ) , ∣ L ∣ = 1 argmin E p ( L ) = { O L B : O ∈ O ( 3 )} . (2)
This resolves the Graf–Luschgy Conjecture [13 Conjecture 8.17] . The normalization of the sharp constant is
E p ( L B ) = ∫ V ( 2 1 / 3 A b c c ) ∣ x ∣ p d x = 2 1 + p / 3 ∫ V ( A b c c ) ∣ x ∣ p d x . E_p(L_{\mathrm{B}})=\int_{V(2^{1/3}A_{\mathrm{bcc}})}|x|^p\,dx=2^{1+p/3}\int_{V(A_{\mathrm{bcc}})}|x|^p\,dx. E p ( L B ) = ∫ V ( 2 1/3 A bcc ) ∣ x ∣ p d x = 2 1 + p /3 ∫ V ( A bcc ) ∣ x ∣ p d x .
The fixed-square expression for this constant is given in (11) .
Let X X X be a nonempty finite set in a domain Ω \Omega Ω of volume one. With the masses on X X X free, the transport cost is
min f : Ω → X f m e a s u r a b l e ∫ Ω ∣ x − f ( x ) ∣ p d x = ∫ Ω min w ∈ X ∣ x − w ∣ p d x = ∫ Ω dist ( x , X ) p d x . \min_{\substack{f:\Omega\to X\\ f\ \mathrm{measurable}}}\int_{\Omega}|x-f(x)|^{p}\,dx=\int_{\Omega}\min_{w\in X}|x-w|^{p}\,dx=\int_{\Omega}\operatorname{dist}(x,X)^{p}\,dx. f : Ω → X f measurable min ∫ Ω ∣ x − f ( x ) ∣ p d x = ∫ Ω w ∈ X min ∣ x − w ∣ p d x = ∫ Ω dist ( x , X ) p d x .
A nearest-site map attains the minimum [4 Lemma 2.2] . On the unit cube in R d \mathbb{R}^{d} R d , the asymptotic quantization theorem gives [22 , 5 ]
lim N → ∞ N p / d inf X ⊂ [ 0 , 1 ] d # X = N ∫ [ 0 , 1 ] d min w ∈ X ∣ x − w ∣ p d x ∈ ( 0 , ∞ ) . \lim_{N\to\infty}N^{p/d}\inf_{\substack{X\subset[0,1]^{d}\\ \#X=N}}\int_{[0,1]^{d}}\min_{w\in X}|x-w|^{p}\,dx\in(0,\infty). N → ∞ lim N p / d X ⊂ [ 0 , 1 ] d # X = N inf ∫ [ 0 , 1 ] d w ∈ X min ∣ x − w ∣ p d x ∈ ( 0 , ∞ ) .
Gruber extended this asymptotic theory to Riemannian manifolds and normed spaces [15 ] . For a bounded open Ω ⊂ R 3 \Omega\subset\mathbb{R}^{3} Ω ⊂ R 3 of volume one with ∂ Ω = φ ( K ) \partial\Omega=\varphi(K) ∂ Ω = φ ( K ) , where K ⊂ R 2 K\subset\mathbb{R}^{2} K ⊂ R 2 is compact and φ : R 2 → R 3 \varphi:\mathbb{R}^{2}\to\mathbb{R}^{3} φ : R 2 → R 3 is Lipschitz, Bourne, Cagnetti and Roper [4 Theorem A.1] prove the pointwise limit
lim ε ↓ 0 ε − p ∣ V ( A ) ∣ − p / 3 ∫ Ω dist ( x , Λ ( ε A ) ∩ Ω ) p d x = E p ( A ) . \lim_{\varepsilon\downarrow0}\varepsilon^{-p}|V(A)|^{-p/3}\int_{\Omega}\operatorname{dist}(x,\Lambda(\varepsilon A)\cap\Omega)^{p}\,dx=E_{p}(A). ε ↓ 0 lim ε − p ∣ V ( A ) ∣ − p /3 ∫ Ω dist ( x , Λ ( ε A ) ∩ Ω ) p d x = E p ( A ) .
The same theorem gives Γ \Gamma Γ -convergence of the rescaled distortion on G L ( 3 , R ) GL(3,\mathbb{R}) G L ( 3 , R ) with the Frobenius metric.
In dimension two the moment theorem of Fejes Tóth yields triangular-lattice optimality for every finite p ≥ 1 p\ge1 p ≥ 1 . The regular hexagon also gives the unrestricted planar asymptotic constant [11 , 14 ] . See [4, Sect. 2]. In dimension three the unrestricted problem is the setting of Gersho’s conjecture. It concerns the asymptotic cell geometry of unconstrained optimal point sets [12 ] . For a precise quadratic formulation see Choksi and Lu [7 Conjecture 1.1] . Du and Wang obtained numerical evidence for the BCC cell in the three-dimensional quadratic problem [10 ] . For the high-rate variable-rate problem on periodic tilings, Sloane and Vaishampayan computed the asymptotic constant in terms of the cell volumes and second moments. In their comparison, the BCC constant is smaller than that of the A15 quantizer [21 ] . For quadratic quantization, Aydın and Iacobelli establish asymptotic local optimality of BCC under small periodic perturbations [1 ] , while Luo and Xie obtain nonlinear stability estimates uniform over finite-index periods [18 ] .
Among three-dimensional lattices, Barnes and Sloane proved BCC optimality for p = 2 p=2 p = 2 [3 ] , and Bambah proved it for the covering radius at fixed covolume [2 ] . The latter is the limiting problem
lim p → ∞ E p ( A ) 1 / p = ∣ V ( A ) ∣ − 1 / 3 max x ∈ V ( A ) ∣ x ∣ . \lim_{p\to\infty}E_{p}(A)^{1/p}=|V(A)|^{-1/3}\max_{x\in V(A)}|x|. p → ∞ lim E p ( A ) 1/ p = ∣ V ( A ) ∣ − 1/3 x ∈ V ( A ) max ∣ x ∣.
Graf and Luschgy asked whether the same lattice minimizes every finite p p p -th moment [13 Conjecture 8.17] . Bourne, Cagnetti and Roper proved strict local optimality in lattice shape for every finite p ≥ 1 p\ge1 p ≥ 1 and global optimality near p = 2 p=2 p = 2 [4 Theorems 1.2–1.3] . They also found a local branch of body-centered tetragonal critical points issuing from the face-centered cubic lattice at p = 2 p=2 p = 2 [4 Theorem 1.5] . Theorem 1.1 gives the global minimizer in the Bravais-lattice class for the entire finite exponent range.
The regular truncated octahedron, the Voronoi cell of BCC, also minimizes surface area among three-dimensional parallelohedra of fixed volume [6 , 16 ] . The proof by Hales and Song uses an exact Bernstein certificate for the final polynomial inequality [16 Appendix A] . Outline of the proof.
Voronoi moments on a fixed square. Section 2 writes the Voronoi cell in Selling coordinates and subdivides one facet from each opposite pair. Radial integration over the fifteen tile pyramids and their opposites gives
E p ( A ) = 1 4 ( p + 3 ) ( det Q ) 1 + p / 6 ∑ T ∈ T W T ∫ [ − 1 , 1 ] 2 q T p / 2 d x d y . E_{p}(A)=\frac{1}{4(p+3)(\det Q)^{1+p/6}}\sum_{T\in\mathcal{T}}W_{T}\int_{[-1,1]^{2}}q_{T}^{p/2}\,dx\,dy. E p ( A ) = 4 ( p + 3 ) ( det Q ) 1 + p /6 1 T ∈ T ∑ W T ∫ [ − 1 , 1 ] 2 q T p /2 d x d y .
Here Q Q Q is the Gram matrix of a Selling basis, q T q_T q T is the squared radius on tile T T T , and W T W_T W T is its weight. At ρ i j = 1 \rho_{ij}=1 ρ ij = 1 the fifteen quadratics take only two forms. The BCC equality case is determined by
0 < det Q ≤ 16 ( ∑ 0 ≤ i < j ≤ 3 ρ i j 6 ) 3 0 < \det Q \le16\left(\frac{\sum_{0\le i<j\le3}\rho_{ij}}{6}\right)^3 0 < det Q ≤ 16 ( 6 ∑ 0 ≤ i < j ≤ 3 ρ ij ) 3
equality ⟺ \Longleftrightarrow ⟺ ρ 01 = ρ 02 = ρ 03 = ρ 12 = ρ 13 = ρ 23 \rho_{01}=\rho_{02}=\rho_{03}=\rho_{12}=\rho_{13}=\rho_{23} ρ 01 = ρ 02 = ρ 03 = ρ 12 = ρ 13 = ρ 23 .
(2) Polynomial comparison for 1 ≤ p ≤ 35 1\le p\le35 1 ≤ p ≤ 35 . Section 3 bounds the p / 2 p/2 p /2 power of the scaled ratio of q T q_T q T to its BCC value. The lower polynomials are Taylor polynomials of degrees four and three for p ∈ [ 1 , 2 ] p\in[1,2] p ∈ [ 1 , 2 ] and p ∈ [ 2 , 4 ] p\in[2,4] p ∈ [ 2 , 4 ] , and weighted tangents of degrees two and three for p ∈ [ 4 , 6 ] p\in[4,6] p ∈ [ 4 , 6 ] and p ∈ [ 6 , 35 ] p\in[6,35] p ∈ [ 6 , 35 ] , respectively. Bounds on the determinant factors then give polynomials in the conorms. For each of the thirty edge orders up to relabelling, the substitution
ρ σ ( i ) = X i + ⋯ + X 5 , X 0 , … , X 5 ≥ 0 , 0 ≤ i ≤ 5 \rho_{\sigma(i)}=X_i+\cdots+X_5,\qquad X_0,\ldots,X_5\ge0,\qquad0\le i\le5 ρ σ ( i ) = X i + ⋯ + X 5 , X 0 , … , X 5 ≥ 0 , 0 ≤ i ≤ 5
reduces the comparison to coefficientwise nonnegativity. Numerical Lemmas 5.1–5.4 give these inequalities by exact interval arithmetic.
(3) Linear comparison for p ≥ 35 p\ge35 p ≥ 35 . Section 4 uses the tangent s p / 2 ≥ 1 + p 2 ( s − 1 ) s^{p/2}\ge1+\frac{p}{2}(s-1) s p /2 ≥ 1 + 2 p ( s − 1 ) and reduces each coefficient to a linear combination of six weighted reference integrals. Numerical Lemma 5.6 gives nonnegativity on [ 35 , 1028 ] [35,1028] [ 35 , 1028 ] . For p ≥ 1026 p\ge1026 p ≥ 1026 , the signs follow from estimates at the maximizing corners of the two reference quadratics. In both comparisons, the determinant bounds give E p ( A ) > E p ( A b c c ) E_p(A)>E_p(A_{\mathrm{bcc}}) E p ( A ) > E p ( A bcc ) when the conorms are unequal. Equal conorms and ∣ V ( A ) ∣ = 1 |V(A)|=1 ∣ V ( A ) ∣ = 1 give the orbit in (1.2).
(4) Exact numerical bounds. Section 5 gives the rational computations for both comparisons. For bounded exponents, Numerical Lemmas 5.1–5.4 give coefficient bounds and counts. Numerical Lemma 5.5 computes the degree-ten coefficient pairs by finite differences. Numerical Lemma 5.6 bounds the corresponding weighted integrals by Bernstein coefficients.
Selling coordinates and the Voronoi moment
We obtain the fifteen fixed-square integrals in (9) by subdividing the facets of the zonotope (4) and integrating along rays. The BCC quadratics are given by (10) . The determinant bound (14) identifies the equality case, and (18) gives the nonnegative variables for the two comparisons.
We use Selling’s reduction [20 ] in the obtuse-superbase form of Conway and Sloane [9 Theorem 8] .
Lemma 2.1 (Selling reduction). Let A ∈ G L ( 3 , R ) A\in GL(3,\mathbb{R}) A ∈ G L ( 3 , R ) . There exist t 0 , t 1 , t 2 , t 3 ∈ Λ ( A ) t_0,t_1,t_2,t_3\in\Lambda(A) t 0 , t 1 , t 2 , t 3 ∈ Λ ( A ) such that
∑ i = 0 3 t i = 0 , t i ⋅ t j ≤ 0 ( i ≠ j ) , \sum_{i=0}^{3}t_i=0,\qquad t_i\cdot t_j\le0\quad(i\ne j), i = 0 ∑ 3 t i = 0 , t i ⋅ t j ≤ 0 ( i = j ) , and any three of these vectors form a basis of Λ ( A ) \Lambda(A) Λ ( A ) .
Proof. Take a basis and the negative of its sum. If t i ⋅ t j > 0 t_i\cdot t_j>0 t i ⋅ t j > 0 , set
t i ′ = − t i , t j ′ = t j , t k ′ = t k + t i ( k ∉ { i , j } ) , ∑ k = 0 3 t k ′ = 0. t'_i=-t_i,\qquad t'_j=t_j,\qquad t'_k=t_k+t_i\quad(k\notin\{i,j\}),\qquad\sum_{k=0}^{3}t'_k=0. t i ′ = − t i , t j ′ = t j , t k ′ = t k + t i ( k ∈ / { i , j }) , k = 0 ∑ 3 t k ′ = 0. For { i , j , k , ℓ } = { 0 , 1 , 2 , 3 } \{i,j,k,\ell\}=\{0,1,2,3\} { i , j , k , ℓ } = { 0 , 1 , 2 , 3 } ,
det ( t i ′ , t k ′ , t ℓ ′ ) = − det ( t i , t k , t ℓ ) . \det(t'_i,t'_k,t'_\ell)=-\det(t_i,t_k,t_\ell). det ( t i ′ , t k ′ , t ℓ ′ ) = − det ( t i , t k , t ℓ ) .
The relation ∑ k = 0 3 t k ′ = 0 \sum_{k=0}^{3}t'_k=0 ∑ k = 0 3 t k ′ = 0 lets us replace any column of this basis by t j ′ t'_j t j ′ , preserving the absolute determinant. Moreover,
∑ k = 0 3 ( ∣ t k ′ ∣ 2 − ∣ t k ∣ 2 ) = 2 ∣ t i ∣ 2 + 2 t i ⋅ ∑ k ∉ { i , j } t k = − 2 t i ⋅ t j < 0. \sum_{k=0}^{3}\left(\lvert t'_k\rvert^2-\lvert t_k\rvert^2\right)=2\lvert t_i\rvert^2+2t_i\cdot\sum_{k\notin\{i,j\}}t_k=-2t_i\cdot t_j<0. k = 0 ∑ 3 ( ∣ t k ′ ∣ 2 − ∣ t k ∣ 2 ) = 2 ∣ t i ∣ 2 + 2 t i ⋅ k ∈ / { i , j } ∑ t k = − 2 t i ⋅ t j < 0.
The iteration stays in a bounded subset of Λ ( A ) \Lambda(A) Λ ( A ) and therefore terminates. □
By Lemma 2.1 and (1.1), take A = ( t 1 ∣ t 2 ∣ t 3 ) A=(t_1\mid t_2\mid t_3) A = ( t 1 ∣ t 2 ∣ t 3 ) with t 0 = − t 1 − t 2 − t 3 t_0=-t_1-t_2-t_3 t 0 = − t 1 − t 2 − t 3 and t i ⋅ t j ≤ 0 t_i\cdot t_j\leq0 t i ⋅ t j ≤ 0 for i ≠ j i\ne j i = j . Set ρ i j = − t i ⋅ t j \rho_{ij}=-t_i\cdot t_j ρ ij = − t i ⋅ t j for i ≠ j i\ne j i = j . With e 0 = 0 e_0=0 e 0 = 0 and e 1 , e 2 , e 3 e_1,e_2,e_3 e 1 , e 2 , e 3 the standard coordinate vectors, write g i j = e i − e j g_{ij}=e_i-e_j g ij = e i − e j for the edges i j ij ij in E = { 01 , 02 , 03 , 12 , 13 , 23 } \mathcal{E}=\{01,02,03,12,13,23\} E = { 01 , 02 , 03 , 12 , 13 , 23 } . Then
Q = A T A = ∑ e ∈ E ρ e g e g e T , det Q = ∣ V ( A ) ∣ 2 . Q=A^{T}A=\sum_{e\in\mathcal{E}}\rho_e g_e g_e^{T},\qquad\det Q=\lvert V(A)\rvert^2. Q = A T A = e ∈ E ∑ ρ e g e g e T , det Q = ∣ V ( A ) ∣ 2 .
In coordinates,
Q = ( ρ 01 + ρ 12 + ρ 13 − ρ 12 − ρ 13 − ρ 12 ρ 02 + ρ 12 + ρ 23 − ρ 23 − ρ 13 − ρ 23 ρ 03 + ρ 13 + ρ 23 ) . Q=\begin{pmatrix}
\rho_{01}+\rho_{12}+\rho_{13} & -\rho_{12} & -\rho_{13}\\
-\rho_{12} & \rho_{02}+\rho_{12}+\rho_{23} & -\rho_{23}\\
-\rho_{13} & -\rho_{23} & \rho_{03}+\rho_{13}+\rho_{23}
\end{pmatrix}. Q = ρ 01 + ρ 12 + ρ 13 − ρ 12 − ρ 13 − ρ 12 ρ 02 + ρ 12 + ρ 23 − ρ 23 − ρ 13 − ρ 23 ρ 03 + ρ 13 + ρ 23 .
The inverse relations are
ρ i j = − Q i j ( 1 ≤ i < j ≤ 3 ) , ρ 0 i = ∑ j = 1 3 Q i j ( 1 ≤ i ≤ 3 ) . \rho_{ij}=-Q_{ij}\quad(1\leq i<j\leq3),\qquad\rho_{0i}=\sum_{j=1}^{3}Q_{ij}\quad(1\leq i\leq3). ρ ij = − Q ij ( 1 ≤ i < j ≤ 3 ) , ρ 0 i = j = 1 ∑ 3 Q ij ( 1 ≤ i ≤ 3 ) .
Expanding the determinant gives
det Q = ρ 01 ρ 02 ρ 03 + ρ 01 ρ 02 ( ρ 13 + ρ 23 ) + ρ 01 ρ 03 ( ρ 12 + ρ 23 ) + ρ 02 ρ 03 ( ρ 12 + ρ 13 ) + ( ρ 01 + ρ 02 + ρ 03 ) ( ρ 12 ρ 13 + ρ 12 ρ 23 + ρ 13 ρ 23 ) . (3) \begin{aligned}
\det Q={}&\rho_{01}\rho_{02}\rho_{03}\\
&+\rho_{01}\rho_{02}(\rho_{13}+\rho_{23})+\rho_{01}\rho_{03}(\rho_{12}+\rho_{23})+\rho_{02}\rho_{03}(\rho_{12}+\rho_{13})\\
&+(\rho_{01}+\rho_{02}+\rho_{03})(\rho_{12}\rho_{13}+\rho_{12}\rho_{23}+\rho_{13}\rho_{23}). \tag*{(3)}
\end{aligned} det Q = ρ 01 ρ 02 ρ 03 + ρ 01 ρ 02 ( ρ 13 + ρ 23 ) + ρ 01 ρ 03 ( ρ 12 + ρ 23 ) + ρ 02 ρ 03 ( ρ 12 + ρ 13 ) + ( ρ 01 + ρ 02 + ρ 03 ) ( ρ 12 ρ 13 + ρ 12 ρ 23 + ρ 13 ρ 23 ) . (3)
In dimension three, Voronoi cells of lattices are zonotopes [19 Sect. 3.2] . In the coordinates z = A T x z=A^{T}x z = A T x , the six generating segments give the Minkowski sum
A T V ( A ) = ∑ e ∈ E [ − ρ e g e / 2 , ρ e g e / 2 ] , ∣ x ∣ 2 = z T Q − 1 z . (4) A^{T}V(A)=\sum_{e\in\mathcal{E}}[-\rho_e g_e/2,\rho_e g_e/2],\qquad\lvert x\rvert^2=z^{T}Q^{-1}z. \tag*{(4)} A T V ( A ) = e ∈ E ∑ [ − ρ e g e /2 , ρ e g e /2 ] , ∣ x ∣ 2 = z T Q − 1 z . (4)
For z z z in the sum and k ∈ Z 3 k\in\mathbb{Z}^3 k ∈ Z 3 ,
2 z ⋅ k ≤ ∑ e ∈ E ρ e ∣ g e ⋅ k ∣ ≤ ∑ e ∈ E ρ e ( g e ⋅ k ) 2 = k T Q k , 2z\cdot k\leq\sum_{e\in\mathcal{E}}\rho_e\lvert g_e\cdot k\rvert\leq\sum_{e\in\mathcal{E}}\rho_e(g_e\cdot k)^2=k^{T}Qk, 2 z ⋅ k ≤ e ∈ E ∑ ρ e ∣ g e ⋅ k ∣ ≤ e ∈ E ∑ ρ e ( g e ⋅ k ) 2 = k T Q k ,
since g e ⋅ k g_e\cdot k g e ⋅ k is integral. Thus
∣ x − A k ∣ 2 − ∣ x ∣ 2 = k T Q k − 2 z ⋅ k ≥ 0 ( k ∈ Z 3 ) , \lvert x-Ak\rvert^2-\lvert x\rvert^2=k^{T}Qk-2z\cdot k\geq0\qquad(k\in\mathbb{Z}^3), ∣ x − A k ∣ 2 − ∣ x ∣ 2 = k T Q k − 2 z ⋅ k ≥ 0 ( k ∈ Z 3 ) ,
and the sum is contained in A T V ( A ) A^{T}V(A) A T V ( A ) . For three distinct edges, det ( g e , g f , g h ) = 0 \det(g_e,g_f,g_h)=0 det ( g e , g f , g h ) = 0 on a cycle and det ( g e , g f , g h ) = ± 1 \det(g_e,g_f,g_h)=\pm1 det ( g e , g f , g h ) = ± 1 on a tree, by expansion at the leaves. Thus the zonotope volume formula and Cauchy–Binet give
∣ ∑ e ∈ E [ − ρ e g e / 2 , ρ e g e / 2 ] ∣ = ∑ e < f < h ρ e ρ f ρ h ∣ det ( g e , g f , g h ) ∣ = ∑ e < f < h ρ e ρ f ρ h det ( g e , g f , g h ) 2 = det Q = ∣ det A ∣ ∣ V ( A ) ∣ = ∣ A T V ( A ) ∣ . \begin{aligned}
\left\lvert\sum_{e\in\mathcal{E}}[-\rho_e g_e/2,\rho_e g_e/2]\right\rvert
&=\sum_{e<f<h}\rho_e\rho_f\rho_h\lvert\det(g_e,g_f,g_h)\rvert\\
&=\sum_{e<f<h}\rho_e\rho_f\rho_h\det(g_e,g_f,g_h)^2\\
&=\det Q=\lvert\det A\rvert\lvert V(A)\rvert=\lvert A^{T}V(A)\rvert.
\end{aligned} e ∈ E ∑ [ − ρ e g e /2 , ρ e g e /2 ] = e < f < h ∑ ρ e ρ f ρ h ∣ det ( g e , g f , g h )∣ = e < f < h ∑ ρ e ρ f ρ h det ( g e , g f , g h ) 2 = det Q = ∣ det A ∣ ∣ V ( A )∣ = ∣ A T V ( A )∣ .
The two full-dimensional compact convex sets are equal by inclusion and equality of volume. Suppose first that ρ e > 0 \rho_e>0 ρ e > 0 for every e ∈ E e\in\mathcal{E} e ∈ E . For a normal η ≠ 0 \eta\ne0 η = 0 , the supporting face has dimension
dim span { g i j : η ⋅ e i = η ⋅ e j } = 4 − # { η ⋅ e 0 , η ⋅ e 1 , η ⋅ e 2 , η ⋅ e 3 } . \dim\operatorname{span}\{g_{ij}:\eta\cdot e_i=\eta\cdot e_j\}=4-\#\{\eta\cdot e_0,\eta\cdot e_1,\eta\cdot e_2,\eta\cdot e_3\}. dim span { g ij : η ⋅ e i = η ⋅ e j } = 4 − # { η ⋅ e 0 , η ⋅ e 1 , η ⋅ e 2 , η ⋅ e 3 } .
Indeed, within each block of equal values, the differences to one chosen vertex form a basis for the differences within that block. The union of these bases is independent by affine independence of e 0 , e 1 , e 2 , e 3 e_0,e_1,e_2,e_3 e 0 , e 1 , e 2 , e 3 . A facet therefore has exactly two blocks. In each opposite pair take the normal for which the block containing vertex 3 has the smaller value, scaled so that the two values differ by one.
For an edge with η ⋅ g e = 0 \eta\cdot g_e=0 η ⋅ g e = 0 the full segment contributes to the face. For any other edge, η ⋅ g e = ± 1 \eta\cdot g_e=\pm1 η ⋅ g e = ± 1 and the contribution is its maximizing endpoint ρ e ( η ⋅ g e ) g e / 2 \rho_e(\eta\cdot g_e)g_e/2 ρ e ( η ⋅ g e ) g e /2 . The facet center z η z_\eta z η and the supporting plane thus satisfy
2 z η = ∑ e ∈ E ρ e ( η ⋅ g e ) g e = Q η , 2 η ⋅ z = η T Q η on the facet . (5) 2z_\eta=\sum_{e\in E}\rho_e(\eta\cdot g_e)g_e=Q\eta,\qquad2\eta\cdot z=\eta^TQ\eta\quad\text{on the facet}. \tag*{(5)} 2 z η = e ∈ E ∑ ρ e ( η ⋅ g e ) g e = Q η , 2 η ⋅ z = η T Q η on the facet . (5)
Lemma 2.2 (Tile parametrizations). Assume ρ e > 0 \rho_e>0 ρ e > 0 for every e ∈ E e\in\mathcal{E} e ∈ E . Each parallelogram facet consists of the single tile
z T ( x , y ) = z η + 1 2 ( x ρ u g u + y ρ v g v ) , z_T(x,y)=z_\eta+\frac{1}{2}(x\rho_u g_u+y\rho_v g_v), z T ( x , y ) = z η + 2 1 ( x ρ u g u + y ρ v g v ) ,
where the free-edge pair and normal are
( ( u , v ) , η ) ∈ { ( ( 01 , 23 ) , ( 0 , − 1 , − 1 ) ) , ( ( 02 , 13 ) , ( − 1 , 0 , − 1 ) ) , ( ( 12 , 03 ) , ( 1 , 1 , 0 ) ) } . ((u,v),\eta)\in\{((01,23),(0,-1,-1)),((02,13),(-1,0,-1)),((12,03),(1,1,0))\}. (( u , v ) , η ) ∈ {(( 01 , 23 ) , ( 0 , − 1 , − 1 )) , (( 02 , 13 ) , ( − 1 , 0 , − 1 )) , (( 12 , 03 ) , ( 1 , 1 , 0 ))} .
The free edges of each hexagonal facet are i j ij ij , i k ik ik , j k jk j k , with
( ( i , j , k ) , η ) ∈ { ( ( 1 , 2 , 3 ) , ( − 1 , − 1 , − 1 ) ) , ( ( 0 , 2 , 3 ) , ( 1 , 0 , 0 ) ) , ( ( 0 , 1 , 3 ) , ( 0 , 1 , 0 ) ) , ( ( 0 , 1 , 2 ) , ( 0 , 0 , − 1 ) ) } . ((i,j,k),\eta)\in\{((1,2,3),(-1,-1,-1)),((0,2,3),(1,0,0)),((0,1,3),(0,1,0)),((0,1,2),(0,0,-1))\}. (( i , j , k ) , η ) ∈ {(( 1 , 2 , 3 ) , ( − 1 , − 1 , − 1 )) , (( 0 , 2 , 3 ) , ( 1 , 0 , 0 )) , (( 0 , 1 , 3 ) , ( 0 , 1 , 0 )) , (( 0 , 1 , 2 ) , ( 0 , 0 , − 1 ))} .
Each hexagonal facet consists of the three tiles
z η + 1 2 ( x ρ i j g i j + y ρ j k g j k − ρ i k g i k ) , z η + 1 2 ( x ρ i j g i j + y ρ i k g i k + ρ j k g j k ) , z η + 1 2 ( x ρ j k g j k + y ρ i k g i k + ρ i j g i j ) . (6) \begin{aligned}
&z_\eta+\frac{1}{2}(x\rho_{ij}g_{ij}+y\rho_{jk}g_{jk}-\rho_{ik}g_{ik}),\\
&z_\eta+\frac{1}{2}(x\rho_{ij}g_{ij}+y\rho_{ik}g_{ik}+\rho_{jk}g_{jk}),\\
&z_\eta+\frac{1}{2}(x\rho_{jk}g_{jk}+y\rho_{ik}g_{ik}+\rho_{ij}g_{ij}). \tag*{(6)}
\end{aligned} z η + 2 1 ( x ρ ij g ij + y ρ j k g j k − ρ ik g ik ) , z η + 2 1 ( x ρ ij g ij + y ρ ik g ik + ρ j k g j k ) , z η + 2 1 ( x ρ j k g j k + y ρ ik g ik + ρ ij g ij ) . (6)
Here ( x , y ) ∈ [ − 1 , 1 ] 2 (x,y)\in[-1,1]^2 ( x , y ) ∈ [ − 1 , 1 ] 2 and the edges multiplying x , y x,y x , y are the free edges of the tile.
Proof. The two free segments give the parallelogram. On a hexagonal facet, write 2 ( z − z η ) = X g i j + Y g j k 2(z-z_\eta)=Xg_{ij}+Yg_{jk} 2 ( z − z η ) = X g ij + Y g j k . Since g i k = g i j + g j k g_{ik}=g_{ij}+g_{jk} g ik = g ij + g j k , the face consists of the points for which
max { − ρ i k , X − ρ i j , Y − ρ j k } ≤ t ≤ min { ρ i k , X + ρ i j , Y + ρ j k } \max\{-\rho_{ik},X-\rho_{ij},Y-\rho_{jk}\}\le t\le\min\{\rho_{ik},X+\rho_{ij},Y+\rho_{jk}\} max { − ρ ik , X − ρ ij , Y − ρ j k } ≤ t ≤ min { ρ ik , X + ρ ij , Y + ρ j k }
for some t t t , the coefficient of g i k g_{ik} g ik in the segment sum. Take the left endpoint. The three active bounds give the three tiles in (2.4):
t = − ρ i k ⟹ ( X , Y ) = ( x ρ i j − ρ i k , y ρ j k − ρ i k ) , t = Y − ρ j k ⟹ ( X , Y ) = ( x ρ i j + y ρ i k , ρ j k + y ρ i k ) , t = X − ρ i j ⟹ ( X , Y ) = ( ρ i j + y ρ i k , x ρ j k + y ρ i k ) . \begin{aligned}
t=-\rho_{ik}&\Longrightarrow(X,Y)=(x\rho_{ij}-\rho_{ik},y\rho_{jk}-\rho_{ik}),\\
t=Y-\rho_{jk}&\Longrightarrow(X,Y)=(x\rho_{ij}+y\rho_{ik},\rho_{jk}+y\rho_{ik}),\\
t=X-\rho_{ij}&\Longrightarrow(X,Y)=(\rho_{ij}+y\rho_{ik},x\rho_{jk}+y\rho_{ik}).
\end{aligned} t = − ρ ik t = Y − ρ j k t = X − ρ ij ⟹ ( X , Y ) = ( x ρ ij − ρ ik , y ρ j k − ρ ik ) , ⟹ ( X , Y ) = ( x ρ ij + y ρ ik , ρ j k + y ρ ik ) , ⟹ ( X , Y ) = ( ρ ij + y ρ ik , x ρ j k + y ρ ik ) .
In each case the other bounds are exactly ∣ x ∣ , ∣ y ∣ ≤ 1 |x|,|y|\le1 ∣ x ∣ , ∣ y ∣ ≤ 1 , and ties have ∣ x ∣ = 1 |x|=1 ∣ x ∣ = 1 or ∣ y ∣ = 1 |y|=1 ∣ y ∣ = 1 . □
Let T \mathcal{T} T be the set of 3 + 4 ⋅ 3 = 15 3+4\cdot3=15 3 + 4 ⋅ 3 = 15 tiles. The ordered free-edge pair ( u , v ) (u,v) ( u , v ) identifies each tile uniquely. On the facet with normal η \eta η , write
q u , v ( x , y ) = q T ( x , y ) = z T ( x , y ) T Q − 1 z T ( x , y ) , W u , v = W T = ρ u ρ v η T Q η . (7) q_{u,v}(x,y)=q_T(x,y)=z_T(x,y)^TQ^{-1}z_T(x,y),\qquad W_{u,v}=W_T=\rho_u\rho_v\eta^TQ\eta. \tag*{(7)} q u , v ( x , y ) = q T ( x , y ) = z T ( x , y ) T Q − 1 z T ( x , y ) , W u , v = W T = ρ u ρ v η T Q η . (7)
Since η T Q η = ∑ e ∈ E ρ e ( η ⋅ g e ) 2 \eta^TQ\eta=\sum_{e\in E}\rho_e(\eta\cdot g_e)^2 η T Q η = ∑ e ∈ E ρ e ( η ⋅ g e ) 2 , the three parallelogram weights are
W 01 , 23 = ρ 01 ρ 23 ( ρ 02 + ρ 03 + ρ 12 + ρ 13 ) , W_{01,23}=\rho_{01}\rho_{23}(\rho_{02}+\rho_{03}+\rho_{12}+\rho_{13}), W 01 , 23 = ρ 01 ρ 23 ( ρ 02 + ρ 03 + ρ 12 + ρ 13 ) ,
W 02 , 13 = ρ 02 ρ 13 ( ρ 01 + ρ 03 + ρ 12 + ρ 23 ) , W_{02,13}=\rho_{02}\rho_{13}(\rho_{01}+\rho_{03}+\rho_{12}+\rho_{23}), W 02 , 13 = ρ 02 ρ 13 ( ρ 01 + ρ 03 + ρ 12 + ρ 23 ) ,
W 12 , 03 = ρ 12 ρ 03 ( ρ 01 + ρ 02 + ρ 13 + ρ 23 ) . W_{12,03}=\rho_{12}\rho_{03}(\rho_{01}+\rho_{02}+\rho_{13}+\rho_{23}). W 12 , 03 = ρ 12 ρ 03 ( ρ 01 + ρ 02 + ρ 13 + ρ 23 ) .
The hexagonal weights in the order of (2.4) are
( W 12 , 23 , W 12 , 13 , W 23 , 13 ) = ( ρ 01 + ρ 02 + ρ 03 ) ( ρ 12 ρ 23 , ρ 12 ρ 13 , ρ 23 ρ 13 ) , (W_{12,23},W_{12,13},W_{23,13})=(\rho_{01}+\rho_{02}+\rho_{03})(\rho_{12}\rho_{23},\rho_{12}\rho_{13},\rho_{23}\rho_{13}), ( W 12 , 23 , W 12 , 13 , W 23 , 13 ) = ( ρ 01 + ρ 02 + ρ 03 ) ( ρ 12 ρ 23 , ρ 12 ρ 13 , ρ 23 ρ 13 ) ,
( W 02 , 23 , W 02 , 03 , W 23 , 03 ) = ( ρ 01 + ρ 12 + ρ 13 ) ( ρ 02 ρ 23 , ρ 02 ρ 03 , ρ 23 ρ 03 ) , (W_{02,23},W_{02,03},W_{23,03})=(\rho_{01}+\rho_{12}+\rho_{13})(\rho_{02}\rho_{23},\rho_{02}\rho_{03},\rho_{23}\rho_{03}), ( W 02 , 23 , W 02 , 03 , W 23 , 03 ) = ( ρ 01 + ρ 12 + ρ 13 ) ( ρ 02 ρ 23 , ρ 02 ρ 03 , ρ 23 ρ 03 ) ,
( W 01 , 13 , W 01 , 03 , W 13 , 03 ) = ( ρ 02 + ρ 12 + ρ 23 ) ( ρ 01 ρ 13 , ρ 01 ρ 03 , ρ 13 ρ 03 ) , (W_{01,13},W_{01,03},W_{13,03})=(\rho_{02}+\rho_{12}+\rho_{23})(\rho_{01}\rho_{13},\rho_{01}\rho_{03},\rho_{13}\rho_{03}), ( W 01 , 13 , W 01 , 03 , W 13 , 03 ) = ( ρ 02 + ρ 12 + ρ 23 ) ( ρ 01 ρ 13 , ρ 01 ρ 03 , ρ 13 ρ 03 ) ,
( W 01 , 12 , W 01 , 02 , W 12 , 02 ) = ( ρ 03 + ρ 13 + ρ 23 ) ( ρ 01 ρ 12 , ρ 01 ρ 02 , ρ 12 ρ 02 ) . (W_{01,12},W_{01,02},W_{12,02})=(\rho_{03}+\rho_{13}+\rho_{23})(\rho_{01}\rho_{12},\rho_{01}\rho_{02},\rho_{12}\rho_{02}). ( W 01 , 12 , W 01 , 02 , W 12 , 02 ) = ( ρ 03 + ρ 13 + ρ 23 ) ( ρ 01 ρ 12 , ρ 01 ρ 02 , ρ 12 ρ 02 ) .
The coordinates in Lemma 2.2 also give a direct polynomial evaluation of the squared radius. By (2.3),
4 q u , v det Q = ( η T Q η ) det Q + ( 2 z T − 2 z η ) T adj ( Q ) ( 2 z T − 2 z η ) . (8) 4q_{u,v}\det Q=(\eta^{T}Q\eta)\det Q+(2z_{T}-2z_{\eta})^{T}\operatorname{adj}(Q)(2z_{T}-2z_{\eta}). \tag*{(8)} 4 q u , v det Q = ( η T Q η ) det Q + ( 2 z T − 2 z η ) T adj ( Q ) ( 2 z T − 2 z η ) . (8)
The cross term vanishes because z η T Q − 1 ( z T − z η ) = η T ( z T − z η ) / 2 = 0 z_{\eta}^{T}Q^{-1}(z_{T}-z_{\eta})=\eta^{T}(z_{T}-z_{\eta})/2=0 z η T Q − 1 ( z T − z η ) = η T ( z T − z η ) /2 = 0 . In particular,
q T = 1 4 η T Q η + ( z T − z η ) T Q − 1 ( z T − z η ) ≥ 1 4 η T Q η > 0. q_{T}=\frac{1}{4}\eta^{T}Q\eta+(z_{T}-z_{\eta})^{T}Q^{-1}(z_{T}-z_{\eta})\geq\frac{1}{4}\eta^{T}Q\eta>0. q T = 4 1 η T Q η + ( z T − z η ) T Q − 1 ( z T − z η ) ≥ 4 1 η T Q η > 0.
Proposition 2.3 (Fixed-domain moment identity). For a reduced basis A A A with conorms ρ e ≥ 0 \rho_e\geq0 ρ e ≥ 0 , one has
E p ( A ) = 1 4 ( p + 3 ) ∣ V ( A ) ∣ 2 + p / 3 ∑ T ∈ T W T ∫ [ − 1 , 1 ] 2 q T ( x , y ) p / 2 d x d y . (9) E_p(A)=\frac{1}{4(p+3)|V(A)|^{2+p/3}}\sum_{T\in\mathcal{T}}W_T\int_{[-1,1]^2}q_T(x,y)^{p/2}\,dx\,dy. \tag*{(9)} E p ( A ) = 4 ( p + 3 ) ∣ V ( A ) ∣ 2 + p /3 1 T ∈ T ∑ W T ∫ [ − 1 , 1 ] 2 q T ( x , y ) p /2 d x d y . (9) Moreover, ∑ T ∈ T W T = 3 det Q \sum_{T\in\mathcal{T}}W_T=3\det Q ∑ T ∈ T W T = 3 det Q .
Proof. For the general facet formula for homogeneous integrands, see Lasserre [17 ] . For positive conorms the pyramid over T T T is parametrized by z = r z T ( x , y ) z=rz_T(x,y) z = r z T ( x , y ) , 0 ≤ r ≤ 1 0\leq r\leq1 0 ≤ r ≤ 1 . The free-edge vectors in Lemma 2.2 satisfy g u × g v = ± η g_u\times g_v=\pm\eta g u × g v = ± η . By (2.3),
∣ det ( z T , ∂ x z T , ∂ y z T ) ∣ = ρ u ρ v 4 ∣ η ⋅ z T ∣ = W T 8 . \left|\det(z_T,\partial_xz_T,\partial_yz_T)\right|=\frac{\rho_u\rho_v}{4}|\eta\cdot z_T|=\frac{W_T}{8}. ∣ det ( z T , ∂ x z T , ∂ y z T ) ∣ = 4 ρ u ρ v ∣ η ⋅ z T ∣ = 8 W T .
The radial factor is given by
d z = r 2 W T 8 d r d x d y , z T Q − 1 z = r 2 q T ( x , y ) , ∫ 0 1 r p + 2 d r = 1 p + 3 . dz=\frac{r^2W_T}{8}\,dr\,dx\,dy,\qquad z^TQ^{-1}z=r^2q_T(x,y),\qquad\int_0^1r^{p+2}\,dr=\frac{1}{p+3}. d z = 8 r 2 W T d r d x d y , z T Q − 1 z = r 2 q T ( x , y ) , ∫ 0 1 r p + 2 d r = p + 3 1 .
Including the opposite pyramids gives
∫ A T V ( A ) ( z T Q − 1 z ) p / 2 d z = 1 4 ( p + 3 ) ∑ T ∈ T W T ∫ [ − 1 , 1 ] 2 q T ( x , y ) p / 2 d x d y . \int_{A^TV(A)}(z^TQ^{-1}z)^{p/2}\,dz=\frac{1}{4(p+3)}\sum_{T\in\mathcal{T}}W_T\int_{[-1,1]^2}q_T(x,y)^{p/2}\,dx\,dy. ∫ A T V ( A ) ( z T Q − 1 z ) p /2 d z = 4 ( p + 3 ) 1 T ∈ T ∑ W T ∫ [ − 1 , 1 ] 2 q T ( x , y ) p /2 d x d y .
Since ∣ det A ∣ = ∣ V ( A ) ∣ |\det A|=|V(A)| ∣ det A ∣ = ∣ V ( A ) ∣ , multiplication by ∣ V ( A ) ∣ − 2 − p / 3 |V(A)|^{-2-p/3} ∣ V ( A ) ∣ − 2 − p /3 gives (2.7). At p = 0 p=0 p = 0 the unnormalized integral gives
det Q = ∣ A T V ( A ) ∣ = 1 3 ∑ T ∈ T W T . \det Q=|A^TV(A)|=\frac{1}{3}\sum_{T\in\mathcal{T}}W_T. det Q = ∣ A T V ( A ) ∣ = 3 1 T ∈ T ∑ W T .
For vanishing conorms use ρ e + ε \rho_e+\varepsilon ρ e + ε , with basis A ε = ( Q + ε ∑ e ∈ E g e g e T ) 1 / 2 A_\varepsilon=(Q+\varepsilon\sum_{e\in\mathcal{E}}g_eg_e^T)^{1/2} A ε = ( Q + ε ∑ e ∈ E g e g e T ) 1/2 . By (2.2),
A T V ( A ) ⊆ A ε T V ( A ε ) ⊆ A T V ( A ) + { z : ∣ z ∣ ≤ ε 2 ∑ e ∈ E ∣ g e ∣ } . A^TV(A)\subseteq A_\varepsilon^TV(A_\varepsilon)\subseteq A^TV(A)+\left\{z:|z|\leq\frac{\varepsilon}{2}\sum_{e\in\mathcal{E}}|g_e|\right\}. A T V ( A ) ⊆ A ε T V ( A ε ) ⊆ A T V ( A ) + { z : ∣ z ∣ ≤ 2 ε e ∈ E ∑ ∣ g e ∣ } .
Every point outside the compact set A T V ( A ) A^TV(A) A T V ( A ) has positive distance from it. The inclusions therefore give, pointwise,
1 A ε T V ( A ε ) ( z ) ⟶ 1 A T V ( A ) ( z ) ( ε ↓ 0 ) . 1_{A_\varepsilon^TV(A_\varepsilon)}(z)\longrightarrow1_{A^TV(A)}(z)\qquad(\varepsilon\downarrow0). 1 A ε T V ( A ε ) ( z ) ⟶ 1 A T V ( A ) ( z ) ( ε ↓ 0 ) .
Also ( A ε T A ε ) − 1 → Q − 1 (A_\varepsilon^TA_\varepsilon)^{-1}\to Q^{-1} ( A ε T A ε ) − 1 → Q − 1 and z T ( A ε T A ε ) − 1 z ≤ z T Q − 1 z z^T(A_\varepsilon^TA_\varepsilon)^{-1}z\leq z^TQ^{-1}z z T ( A ε T A ε ) − 1 z ≤ z T Q − 1 z . For 0 < ε ≤ 1 0<\varepsilon\leq1 0 < ε ≤ 1 the domains are contained in a fixed bounded set, on which ( z T Q − 1 z ) p / 2 (z^TQ^{-1}z)^{p/2} ( z T Q − 1 z ) p /2 is bounded. Dominated convergence for the integrals and continuity of the tile expressions give both identities at ε = 0 \varepsilon=0 ε = 0 . □
Lemma 2.4 (BCC reference quadratics). At ρ = 1 \rho= 1 ρ = 1 , write Q 0 = Q ∣ ρ = 1 Q_0 = Q|_{\rho=1} Q 0 = Q ∣ ρ = 1 and q 0 , T = q T ∣ ρ = 1 q_{0,T} = q_T|_{\rho=1} q 0 , T = q T ∣ ρ = 1 . The three square tiles have weight 4 4 4 and reference quadratic q 0 , s q q_{0,\mathrm{sq}} q 0 , sq . The twelve rhombus tiles have weight 3 3 3 . On the hexagonal facet with free edges i j ij ij , i k ik ik , j k jk j k , the reference quadratics are
q 0 , T ( x , y ) = { q 0 , r h ( x , y ) , ( u , v ) = ( i j , j k ) , q 0 , r h ( x , − y ) , ( u , v ) ∈ { ( i j , i k ) , ( j k , i k ) } . q_{0,T}(x,y) =
\begin{cases}
q_{0,\mathrm{rh}}(x,y), & (u,v)=(ij,jk),\\
q_{0,\mathrm{rh}}(x,-y), & (u,v)\in\{(ij,ik),(jk,ik)\}.
\end{cases} q 0 , T ( x , y ) = { q 0 , rh ( x , y ) , q 0 , rh ( x , − y ) , ( u , v ) = ( ij , j k ) , ( u , v ) ∈ {( ij , ik ) , ( j k , ik )} . Here
q 0 , s q ( x , y ) = 1 + x 2 + y 2 8 , q 0 , r h ( x , y ) = 3 4 + ( x − 1 ) 2 − ( x − 1 ) ( y − 1 ) + ( y − 1 ) 2 8 . (10) q_{0,\mathrm{sq}}(x,y)=1+\frac{x^2+y^2}{8},
\qquad
q_{0,\mathrm{rh}}(x,y)=\frac{3}{4}+\frac{(x-1)^2-(x-1)(y-1)+(y-1)^2}{8}. \tag*{(10)} q 0 , sq ( x , y ) = 1 + 8 x 2 + y 2 , q 0 , rh ( x , y ) = 4 3 + 8 ( x − 1 ) 2 − ( x − 1 ) ( y − 1 ) + ( y − 1 ) 2 . (10) All these quadratics take values in [ 3 / 4 , 5 / 4 ] [3/4,5/4] [ 3/4 , 5/4 ] on [ − 1 , 1 ] 2 [-1,1]^2 [ − 1 , 1 ] 2 .
Proof. At ρ = 1 \rho=1 ρ = 1 the Gram matrix and its inverse are
Q 0 = ( 3 − 1 − 1 − 1 3 − 1 − 1 − 1 3 ) , 4 Q 0 − 1 = ( 2 1 1 1 2 1 1 1 2 ) , det Q 0 = 16. Q_0=
\begin{pmatrix}
3&-1&-1\\
-1&3&-1\\
-1&-1&3
\end{pmatrix},
\qquad
4Q_0^{-1}=
\begin{pmatrix}
2&1&1\\
1&2&1\\
1&1&2
\end{pmatrix},
\qquad
\det Q_0=16. Q 0 = 3 − 1 − 1 − 1 3 − 1 − 1 − 1 3 , 4 Q 0 − 1 = 2 1 1 1 2 1 1 1 2 , det Q 0 = 16. The columns ( − 1 , 1 , 1 ) T (-1,1,1)^T ( − 1 , 1 , 1 ) T , ( 1 , − 1 , 1 ) T (1,-1,1)^T ( 1 , − 1 , 1 ) T , ( 1 , 1 , − 1 ) T (1,1,-1)^T ( 1 , 1 , − 1 ) T have Gram matrix Q 0 Q_0 Q 0 . They belong to 2 D 3 ∗ 2D_3^* 2 D 3 ∗ and generate it, since both lattices have covolume 4 4 4 . By (1) , the corresponding moment is E p ( A b c c ) E_p(A_{\mathrm{bcc}}) E p ( A bcc ) .
At ρ = 1 \rho=1 ρ = 1 the center contributes z η T Q 0 − 1 z η = η T Q 0 η / 4 = W T / 4 z_\eta^TQ_0^{-1}z_\eta=\eta^TQ_0\eta/4=W_T/4 z η T Q 0 − 1 z η = η T Q 0 η /4 = W T /4 and is orthogonal to the free directions in the Q 0 − 1 Q_0^{-1} Q 0 − 1 inner product. The inverse matrix gives
g e T Q 0 − 1 g e = 1 2 , g e T Q 0 − 1 g f = 0 if e , f have disjoint vertices , g_e^TQ_0^{-1}g_e=\frac{1}{2},
\qquad
g_e^TQ_0^{-1}g_f=0\quad\text{if }e,f\text{ have disjoint vertices}, g e T Q 0 − 1 g e = 2 1 , g e T Q 0 − 1 g f = 0 if e , f have disjoint vertices , g i j T Q 0 − 1 g j k = − 1 4 , g i j T Q 0 − 1 g i k = g j k T Q 0 − 1 g i k = 1 4 ( i < j < k ) . g_{ij}^TQ_0^{-1}g_{jk}=-\frac{1}{4},
\qquad
g_{ij}^TQ_0^{-1}g_{ik}=g_{jk}^TQ_0^{-1}g_{ik}=\frac{1}{4}
\quad(i<j<k). g ij T Q 0 − 1 g j k = − 4 1 , g ij T Q 0 − 1 g ik = g j k T Q 0 − 1 g ik = 4 1 ( i < j < k ) . The hexagonal displacements 2 ( z T − z η ) 2(z_T-z_\eta) 2 ( z T − z η ) in order are
( x − 1 ) g i j + ( y − 1 ) g j k , ( x − 1 ) g i j + ( y + 1 ) g i k , ( x − 1 ) g j k + ( y + 1 ) g i k . (x-1)g_{ij}+(y-1)g_{jk},
\qquad
(x-1)g_{ij}+(y+1)g_{ik},
\qquad
(x-1)g_{jk}+(y+1)g_{ik}. ( x − 1 ) g ij + ( y − 1 ) g j k , ( x − 1 ) g ij + ( y + 1 ) g ik , ( x − 1 ) g j k + ( y + 1 ) g ik . The listed conorm weights give W T = 4 W_T=4 W T = 4 or 3 3 3 , so the inner products yield (10) and its reflections. The lower bounds follow from (10) , and convexity gives the upper bounds from the vertex values
q 0 , s q ( ± 1 , ± 1 ) = 5 / 4 , q_{0,\mathrm{sq}}(\pm1,\pm1)=5/4, q 0 , sq ( ± 1 , ± 1 ) = 5/4 , q 0 , r h ( − 1 , − 1 ) = q 0 , r h ( − 1 , 1 ) = q 0 , r h ( 1 , − 1 ) = 5 / 4 , q 0 , r h ( 1 , 1 ) = 3 / 4. q_{0,\mathrm{rh}}(-1,-1)=q_{0,\mathrm{rh}}(-1,1)=q_{0,\mathrm{rh}}(1,-1)=5/4,
\qquad
q_{0,\mathrm{rh}}(1,1)=3/4. q 0 , rh ( − 1 , − 1 ) = q 0 , rh ( − 1 , 1 ) = q 0 , rh ( 1 , − 1 ) = 5/4 , q 0 , rh ( 1 , 1 ) = 3/4.
Applying (9) at ρ = 1 \rho=1 ρ = 1 gives
E p ( A b c c ) = 3 16 ( p + 3 ) 2 2 p / 3 ∫ [ − 1 , 1 ] 2 ( q 0 , s q p / 2 + 3 q 0 , r h p / 2 ) d x d y . (11) E_p(A_{\mathrm{bcc}})=\frac{3}{16(p+3)2^{2p/3}}\int_{[-1,1]^2}\left(q_{0,\mathrm{sq}}^{p/2}+3q_{0,\mathrm{rh}}^{p/2}\right)\,dx\,dy. \tag*{(11)} E p ( A bcc ) = 16 ( p + 3 ) 2 2 p /3 3 ∫ [ − 1 , 1 ] 2 ( q 0 , sq p /2 + 3 q 0 , rh p /2 ) d x d y . (11)
For nonnegative integers a , b a,b a , b , direct integration gives
1 4 ∫ [ − 1 , 1 ] 2 x a y b d x d y = { 1 ( a + 1 ) ( b + 1 ) , a , b even , 0 , otherwise . (12) \frac{1}{4}\int_{[-1,1]^2}x^ay^b\,dx\,dy=
\begin{cases}
\dfrac{1}{(a+1)(b+1)}, & a,b\text{ even},\\
0, & \text{otherwise}.
\end{cases} \tag*{(12)} 4 1 ∫ [ − 1 , 1 ] 2 x a y b d x d y = ⎩ ⎨ ⎧ ( a + 1 ) ( b + 1 ) 1 , 0 , a , b even , otherwise . (12)
For a quadratic q ( x , y ) = ∑ r + s ≤ 2 d r s x r y s q(x,y)=\sum_{r+s\le2}d_{rs}x^ry^s q ( x , y ) = ∑ r + s ≤ 2 d r s x r y s , the multinomial theorem and (12) give
1 4 ∫ [ − 1 , 1 ] 2 x a y b q ( x , y ) n d x d y = ∑ ∑ r + s ≤ 2 m r s = n a + ∑ r + s ≤ 2 r m r s e v e n b + ∑ r + s ≤ 2 s m r s e v e n n ! ∏ r + s ≤ 2 ( d r s m r s / m r s ! ) ( a + 1 + ∑ r + s ≤ 2 r m r s ) ( b + 1 + ∑ r + s ≤ 2 s m r s ) . (13) \frac{1}{4}\int_{[-1,1]^2}x^ay^bq(x,y)^n\,dx\,dy
=
\sum_{\substack{\sum_{r+s\le2}m_{rs}=n\\a+\sum_{r+s\le2}rm_{rs}\ \mathrm{even}\\b+\sum_{r+s\le2}sm_{rs}\ \mathrm{even}}}
\frac{n!\prod_{r+s\le2}\left(d_{rs}^{m_{rs}}/m_{rs}!\right)}
{\left(a+1+\sum_{r+s\le2}rm_{rs}\right)\left(b+1+\sum_{r+s\le2}sm_{rs}\right)}. \tag*{(13)} 4 1 ∫ [ − 1 , 1 ] 2 x a y b q ( x , y ) n d x d y = ∑ r + s ≤ 2 m r s = n a + ∑ r + s ≤ 2 r m r s even b + ∑ r + s ≤ 2 s m r s even ∑ ( a + 1 + ∑ r + s ≤ 2 r m r s ) ( b + 1 + ∑ r + s ≤ 2 s m r s ) n ! ∏ r + s ≤ 2 ( d r s m r s / m r s ! ) . (13)
All indices r , s , m r s r,s,m_{rs} r , s , m r s and exponents a , b , n a,b,n a , b , n are nonnegative integers. A factor with m r s = 0 m_{rs}=0 m r s = 0 is one.
Numerical Lemma 2.1 (First and second moments at BCC). One has
0.46908851641917105 < E 1 ( A b c c ) < 0.46908851641917107 , E 2 ( A b c c ) = 19 64 2 1 / 3 . 0.46908851641917105<E_1(A_{\mathrm{bcc}})<0.46908851641917107,\qquad E_2(A_{\mathrm{bcc}})=\frac{19}{64\,2^{1/3}}. 0.46908851641917105 < E 1 ( A bcc ) < 0.46908851641917107 , E 2 ( A bcc ) = 64 2 1/3 19 . Proof. Integration of the quadratics in (10) gives
∫ [ − 1 , 1 ] 2 q 0 , s q d x d y = 13 3 , ∫ [ − 1 , 1 ] 2 q 0 , r h d x d y = 23 6 . \int_{[-1,1]^2}q_{0,\mathrm{sq}}\,dx\,dy=\frac{13}{3},\qquad
\int_{[-1,1]^2}q_{0,\mathrm{rh}}\,dx\,dy=\frac{23}{6}. ∫ [ − 1 , 1 ] 2 q 0 , sq d x d y = 3 13 , ∫ [ − 1 , 1 ] 2 q 0 , rh d x d y = 6 23 . Thus (11) gives
E 2 ( A b c c ) = 3 80 ⋅ 2 4 / 3 ( 13 3 + 3 ⋅ 23 6 ) = 19 64 ⋅ 2 1 / 3 . E_2(A_{\mathrm{bcc}})=\frac{3}{80\cdot2^{4/3}}\left(\frac{13}{3}+3\cdot\frac{23}{6}\right)=\frac{19}{64\cdot2^{1/3}}. E 2 ( A bcc ) = 80 ⋅ 2 4/3 3 ( 3 13 + 3 ⋅ 6 23 ) = 64 ⋅ 2 1/3 19 . Conway and Sloane [8 Eq. (17)] give the same moment with the per-coordinate normalization E 2 / 3 E_2/3 E 2 /3 .
For q = q 0 , s q q=q_{0,\mathrm{sq}} q = q 0 , sq or q = q 0 , r h q=q_{0,\mathrm{rh}} q = q 0 , rh , the bounds ∣ q − 1 ∣ ≤ 1 / 4 \lvert q-1\rvert\le1/4 ∣ q − 1 ∣ ≤ 1/4 and ∣ ( 1 / 2 ) j ∣ ≤ 1 \lvert(1/2)_j\rvert\le1 ∣( 1/2 ) j ∣ ≤ 1 give
∣ 1 4 ∫ [ − 1 , 1 ] 2 q 1 / 2 d x d y − 1 4 ∑ j = 0 40 ( 1 / 2 j ) ∫ [ − 1 , 1 ] 2 ( q − 1 ) j d x d y ∣ ≤ ∑ j = 41 ∞ 4 − j = 1 3 ⋅ 4 40 . \left\lvert\frac{1}{4}\int_{[-1,1]^2}q^{1/2}\,dx\,dy-\frac{1}{4}\sum_{j=0}^{40}\binom{1/2}{j}\int_{[-1,1]^2}(q-1)^j\,dx\,dy\right\rvert
\le\sum_{j=41}^{\infty}4^{-j}=\frac{1}{3\cdot4^{40}}. 4 1 ∫ [ − 1 , 1 ] 2 q 1/2 d x d y − 4 1 j = 0 ∑ 40 ( j 1/2 ) ∫ [ − 1 , 1 ] 2 ( q − 1 ) j d x d y ≤ j = 41 ∑ ∞ 4 − j = 3 ⋅ 4 40 1 . Apply (13) to the quadratic q − 1 q-1 q − 1 with a = b = 0 a=b=0 a = b = 0 and n = j n=j n = j . The resulting rational sums and the error bound give
10 405 278 818 962 246 491 10 19 < 1 4 ∫ [ − 1 , 1 ] 2 q 0 , s q 1 / 2 d x d y < 10 405 278 818 962 246 492 10 19 , \frac{10\,405\,278\,818\,962\,246\,491}{10^{19}}
<
\frac{1}{4}\int_{[-1,1]^2}q_{0,\mathrm{sq}}^{1/2}\,dx\,dy
<
\frac{10\,405\,278\,818\,962\,246\,492}{10^{19}}, 1 0 19 10 405 278 818 962 246 491 < 4 1 ∫ [ − 1 , 1 ] 2 q 0 , sq 1/2 d x d y < 1 0 19 10 405 278 818 962 246 492 , 9 769 468 916 879 147 255 10 19 < 1 4 ∫ [ − 1 , 1 ] 2 q 0 , r h 1 / 2 d x d y < 9 769 468 916 879 147 256 10 19 . \frac{9\,769\,468\,916\,879\,147\,255}{10^{19}}
<
\frac{1}{4}\int_{[-1,1]^2}q_{0,\mathrm{rh}}^{1/2}\,dx\,dy
<
\frac{9\,769\,468\,916\,879\,147\,256}{10^{19}}. 1 0 19 9 769 468 916 879 147 255 < 4 1 ∫ [ − 1 , 1 ] 2 q 0 , rh 1/2 d x d y < 1 0 19 9 769 468 916 879 147 256 . Cubing rational endpoints gives 1.587401051968199474 < 4 3 < 1.587401051968199475 1.587401051968199474<\sqrt[3]{4}<1.587401051968199475 1.587401051968199474 < 3 4 < 1.587401051968199475 .
Now (11) at p = 1 p=1 p = 1 reads
E 1 ( A b c c ) = 3 64 4 3 ∫ [ − 1 , 1 ] 2 ( q 0 , s q 1 / 2 + 3 q 0 , r h 1 / 2 ) d x d y . E_1(A_{\mathrm{bcc}})=\frac{3}{64\sqrt[3]{4}}\int_{[-1,1]^2}\left(q_{0,\mathrm{sq}}^{1/2}+3q_{0,\mathrm{rh}}^{1/2}\right)\,dx\,dy. E 1 ( A bcc ) = 64 3 4 3 ∫ [ − 1 , 1 ] 2 ( q 0 , sq 1/2 + 3 q 0 , rh 1/2 ) d x d y . The displayed rational bounds give the stated first moment. □
The arithmetic–geometric mean inequality for the eigenvalues of Q 0 − 1 / 2 Q Q 0 − 1 / 2 Q_0^{-1/2}QQ_0^{-1/2} Q 0 − 1/2 Q Q 0 − 1/2 gives
0 < det Q 16 ≤ ( tr ( Q 0 − 1 / 2 Q Q 0 − 1 / 2 ) 3 ) 3 = ( ∑ e ∈ E ρ e 6 ) 3 . (14) 0<\frac{\det Q}{16}\le\left(\frac{\operatorname{tr}(Q_0^{-1/2}QQ_0^{-1/2})}{3}\right)^3=\left(\frac{\sum_{e\in\mathcal{E}}\rho_e}{6}\right)^3. \tag*{(14)} 0 < 16 det Q ≤ ( 3 tr ( Q 0 − 1/2 Q Q 0 − 1/2 ) ) 3 = ( 6 ∑ e ∈ E ρ e ) 3 . (14)
The trace is tr ( Q 0 − 1 Q ) = ∑ e ∈ E ρ e g e T Q 0 − 1 g e = 1 2 ∑ e ∈ E ρ e \operatorname{tr}(Q_0^{-1}Q)=\sum_{e\in\mathcal{E}}\rho_e g_e^TQ_0^{-1}g_e=\frac{1}{2}\sum_{e\in\mathcal{E}}\rho_e tr ( Q 0 − 1 Q ) = ∑ e ∈ E ρ e g e T Q 0 − 1 g e = 2 1 ∑ e ∈ E ρ e . Equality holds exactly when
Q = ∑ e ∈ E ρ e 6 Q 0 ⟺ ρ 01 = ρ 02 = ρ 03 = ρ 12 = ρ 13 = ρ 23 . Q=\frac{\sum_{e\in\mathcal{E}}\rho_e}{6}Q_0\quad\Longleftrightarrow\quad\rho_{01}=\rho_{02}=\rho_{03}=\rho_{12}=\rho_{13}=\rho_{23}. Q = 6 ∑ e ∈ E ρ e Q 0 ⟺ ρ 01 = ρ 02 = ρ 03 = ρ 12 = ρ 13 = ρ 23 .
In this case Λ ( A ) = 2 ( ∑ e ∈ E ρ e / 6 ) 1 / 2 O D 3 ∗ \Lambda(A)=2\left(\sum_{e\in\mathcal{E}}\rho_e/6\right)^{1/2}OD_3^* Λ ( A ) = 2 ( ∑ e ∈ E ρ e /6 ) 1/2 O D 3 ∗ for some O ∈ O ( 3 ) O\in O(3) O ∈ O ( 3 ) . At unit covolume,
1 = 4 ( ∑ e ∈ E ρ e 6 ) 3 / 2 , Λ ( A ) = 2 1 / 3 O D 3 ∗ = O L B for some O ∈ O ( 3 ) . 1=4\left(\frac{\sum_{e\in\mathcal{E}}\rho_e}{6}\right)^{3/2},\qquad\Lambda(A)=2^{1/3}OD_3^*=OL_B\quad\text{for some }O\in O(3). 1 = 4 ( 6 ∑ e ∈ E ρ e ) 3/2 , Λ ( A ) = 2 1/3 O D 3 ∗ = O L B for some O ∈ O ( 3 ) .
Conversely, every lattice in this orbit has moment E p ( A b c c ) E_p(A_{\mathrm{bcc}}) E p ( A bcc ) by (1) . With α = p / 2 \alpha=p/2 α = p /2 , dividing (9) by (11) gives
E p ( A ) E p ( A b c c ) = ( 16 det Q ) 1 + α / 3 ∑ T ∈ T W T ∫ [ − 1 , 1 ] 2 q T α d x d y ∫ [ − 1 , 1 ] 2 ( 12 q 0 , s q α + 36 q 0 , r h α ) d x d y (15) \frac{E_p(A)}{E_p(A_{\mathrm{bcc}})}
=\left(\frac{16}{\det Q}\right)^{1+\alpha/3}
\frac{\sum_{T\in\mathcal{T}}W_T\int_{[-1,1]^2}q_T^\alpha\,dx\,dy}
{\int_{[-1,1]^2}(12q_{0,\mathrm{sq}}^\alpha+36q_{0,\mathrm{rh}}^\alpha)\,dx\,dy} \tag*{(15)} E p ( A bcc ) E p ( A ) = ( det Q 16 ) 1 + α /3 ∫ [ − 1 , 1 ] 2 ( 12 q 0 , sq α + 36 q 0 , rh α ) d x d y ∑ T ∈ T W T ∫ [ − 1 , 1 ] 2 q T α d x d y (15)
Lemma 2.5 (Tile polynomials). For a tile with coordinates z T = ( z 1 , z 2 , z 3 ) T z_T=(z_1,z_2,z_3)^T z T = ( z 1 , z 2 , z 3 ) T ,
q T det Q = ( ρ 02 ρ 03 + ρ 02 ρ 23 + ρ 03 ρ 23 ) z 1 2 + ( ρ 01 ρ 03 + ρ 01 ρ 13 + ρ 03 ρ 13 ) z 2 2 + ( ρ 01 ρ 02 + ρ 01 ρ 12 + ρ 02 ρ 12 ) z 3 2 + ρ 03 ρ 12 ( z 1 + z 2 ) 2 + ρ 02 ρ 13 ( z 1 + z 3 ) 2 + ρ 01 ρ 23 ( z 2 + z 3 ) 2 + ( ρ 12 ρ 13 + ρ 12 ρ 23 + ρ 13 ρ 23 ) ( z 1 + z 2 + z 3 ) 2 . (16) \begin{aligned}
q_T\det Q={}&(\rho_{02}\rho_{03}+\rho_{02}\rho_{23}+\rho_{03}\rho_{23})z_1^2\\
&+(\rho_{01}\rho_{03}+\rho_{01}\rho_{13}+\rho_{03}\rho_{13})z_2^2\\
&+(\rho_{01}\rho_{02}+\rho_{01}\rho_{12}+\rho_{02}\rho_{12})z_3^2\\
&+\rho_{03}\rho_{12}(z_1+z_2)^2+\rho_{02}\rho_{13}(z_1+z_3)^2\\
&+\rho_{01}\rho_{23}(z_2+z_3)^2\\
&+(\rho_{12}\rho_{13}+\rho_{12}\rho_{23}+\rho_{13}\rho_{23})(z_1+z_2+z_3)^2. \tag*{(16)}
\end{aligned} q T det Q = ( ρ 02 ρ 03 + ρ 02 ρ 23 + ρ 03 ρ 23 ) z 1 2 + ( ρ 01 ρ 03 + ρ 01 ρ 13 + ρ 03 ρ 13 ) z 2 2 + ( ρ 01 ρ 02 + ρ 01 ρ 12 + ρ 02 ρ 12 ) z 3 2 + ρ 03 ρ 12 ( z 1 + z 2 ) 2 + ρ 02 ρ 13 ( z 1 + z 3 ) 2 + ρ 01 ρ 23 ( z 2 + z 3 ) 2 + ( ρ 12 ρ 13 + ρ 12 ρ 23 + ρ 13 ρ 23 ) ( z 1 + z 2 + z 3 ) 2 . (16) Equivalently,
4 q T det Q = ∑ e , f ∈ E e < f ρ e ρ f det ( g e , g f , 2 z T ) 2 . (17) 4q_T\det Q=\sum_{\substack{e,f\in\mathcal{E}\\e<f}}\rho_e\rho_f\det(g_e,g_f,2z_T)^2. \tag*{(17)} 4 q T det Q = e , f ∈ E e < f ∑ ρ e ρ f det ( g e , g f , 2 z T ) 2 . (17) For k ∈ Z ≥ 0 k\in\mathbb{Z}_{\geq0} k ∈ Z ≥ 0 , the polynomial W T ( 4 q T det Q ) k W_T(4q_T\det Q)^k W T ( 4 q T det Q ) k has integer coefficients, is homogeneous of degree 4 k + 3 4k+3 4 k + 3 in the conorms, and has degree at most 2 k 2k 2 k in ( x , y ) (x,y) ( x , y ) .
Proof. The rank-one determinant identity and Cauchy–Binet give
det ( Q + t ( 2 z T ) ( 2 z T ) T ) − det Q = t ( 2 z T ) T adj ( Q ) ( 2 z T ) \det\left(Q+t(2z_T)(2z_T)^T\right)-\det Q
=t(2z_T)^T\operatorname{adj}(Q)(2z_T) det ( Q + t ( 2 z T ) ( 2 z T ) T ) − det Q = t ( 2 z T ) T adj ( Q ) ( 2 z T ) = t ∑ e , f ∈ E e < f ρ e ρ f det ( g e , g f , 2 z T ) 2 . =t\sum_{\substack{e,f\in\mathcal{E}\\e<f}}\rho_e\rho_f\det(g_e,g_f,2z_T)^2. = t e , f ∈ E e < f ∑ ρ e ρ f det ( g e , g f , 2 z T ) 2 . Since ( 2 z T ) T adj ( Q ) ( 2 z T ) = 4 q T det Q (2z_T)^T\operatorname{adj}(Q)(2z_T)=4q_T\det Q ( 2 z T ) T adj ( Q ) ( 2 z T ) = 4 q T det Q , this gives (17) . For each pair of edges,
det ( g e , g f , 2 z T ) 2 = 4 ( ( g e × g f ) ⋅ z T ) 2 . \det(g_e,g_f,2z_T)^2=4\bigl((g_e\times g_f)\cdot z_T\bigr)^2. det ( g e , g f , 2 z T ) 2 = 4 ( ( g e × g f ) ⋅ z T ) 2 . The seven possible cross products, up to sign, are
e 1 , e 2 , e 3 , e 1 + e 2 , e 1 + e 3 , e 2 + e 3 , e 1 + e 2 + e 3 . e_1,\ e_2,\ e_3,\ e_1+e_2,\ e_1+e_3,\ e_2+e_3,\ e_1+e_2+e_3. e 1 , e 2 , e 3 , e 1 + e 2 , e 1 + e 3 , e 2 + e 3 , e 1 + e 2 + e 3 . Grouping the fifteen pairs by their cross products gives (16) . The tile parametrizations give 2 z T 2z_T 2 z T with integer coefficients, linear homogeneous in the conorms and affine in ( x , y ) (x,y) ( x , y ) . The weight in (7) is cubic with integer coefficients, giving the claimed degrees. □ \square □
Order the edges as 01, 02, 03, 12, 13, 23. If σ : { 0 , … , 5 } → E \sigma:\{0,\ldots,5\}\to\mathcal{E} σ : { 0 , … , 5 } → E is a bijection with ρ σ ( 0 ) ≥ ⋯ ≥ ρ σ ( 5 ) \rho_{\sigma(0)}\geq\cdots\geq\rho_{\sigma(5)} ρ σ ( 0 ) ≥ ⋯ ≥ ρ σ ( 5 ) , write
ρ σ ( i ) = ∑ j = i 5 X j , X i ≥ 0 , 0 ≤ i ≤ 5. (18) \rho_{\sigma(i)}=\sum_{j=i}^{5}X_j,\qquad X_i\geq0,\qquad0\leq i\leq5. \tag*{(18)} ρ σ ( i ) = j = i ∑ 5 X j , X i ≥ 0 , 0 ≤ i ≤ 5. (18)
Thus X i = ρ σ ( i ) − ρ σ ( i + 1 ) X_i=\rho_{\sigma(i)}-\rho_{\sigma(i+1)} X i = ρ σ ( i ) − ρ σ ( i + 1 ) for 0 ≤ i < 5 0\leq i<5 0 ≤ i < 5 and X 5 = ρ σ ( 5 ) X_5=\rho_{\sigma(5)} X 5 = ρ σ ( 5 ) . There are 6 ! / 4 ! = 30 6!/4!=30 6 ! /4 ! = 30 orderings modulo vertex relabellings, since a vertex permutation fixing each edge is the identity. In each orbit take the lexicographically least order. Relabelling the superbase preserves E p ( A ) E_p(A) E p ( A ) by (1) . The closures of the thirty sectors include the cases with ties.
For μ ∈ Z ≥ 0 6 \boldsymbol{\mu} \in\mathbb{Z}_{\geq0}^{6} μ ∈ Z ≥ 0 6 , expanding the powers of X 0 , … , X 5 X_0,\ldots,X_5 X 0 , … , X 5 in (18) gives
∏ i = 0 5 ( ∑ j = i 5 X j ) μ i = ∑ ν ∈ Z ≥ 0 6 ∑ i = 0 5 ν i = ∑ i = 0 5 μ i ( ∏ j = 0 5 ( ∑ i = 0 j μ i − ∑ i = 0 j − 1 ν i ν j ) ) ∏ i = 0 5 X i ν i . (19) \prod_{i=0}^{5}\left(\sum_{j=i}^{5}X_j\right)^{\mu_i}
=
\sum_{\substack{\boldsymbol{\nu}\in\mathbb{Z}_{\geq0}^{6}\\ \sum_{i=0}^{5}\nu_i=\sum_{i=0}^{5}\mu_i}}
\left(\prod_{j=0}^{5}
\binom{\sum_{i=0}^{j}\mu_i-\sum_{i=0}^{j-1}\nu_i}{\nu_j}\right)
\prod_{i=0}^{5}X_i^{\nu_i}. \tag*{(19)} i = 0 ∏ 5 ( j = i ∑ 5 X j ) μ i = ν ∈ Z ≥ 0 6 ∑ i = 0 5 ν i = ∑ i = 0 5 μ i ∑ ( j = 0 ∏ 5 ( ν j ∑ i = 0 j μ i − ∑ i = 0 j − 1 ν i ) ) i = 0 ∏ 5 X i ν i . (19)
The second sum in the upper entry is empty at j = 0 j=0 j = 0 . A factor vanishes if its upper entry is smaller than its lower entry. Indeed, at step j j j there are ∑ i = 0 j μ i − ∑ i = 0 j − 1 ν i \sum_{i=0}^{j}\mu_i-\sum_{i=0}^{j-1}\nu_i ∑ i = 0 j μ i − ∑ i = 0 j − 1 ν i powers of ∑ i = j 5 X i \sum_{i=j}^{5}X_i ∑ i = j 5 X i from which to select X j ν j X_j^{\nu_j} X j ν j .
For homogeneous P P P of degree n ≥ 1 n\geq1 n ≥ 1 , the terms of degree at most one in X 0 , … , X 4 X_0,\ldots,X_4 X 0 , … , X 4 in P ( ρ ( X ) ) P(\rho(X)) P ( ρ ( X )) are
P ( 1 ) X 5 n + X 5 n − 1 ∑ i = 0 4 X i ∑ j = 0 i ∂ ρ σ ( j ) P ( 1 ) . P(1)X_5^n+X_5^{n-1}\sum_{i=0}^{4}X_i\sum_{j=0}^{i}\partial_{\rho_{\sigma(j)}}P(1). P ( 1 ) X 5 n + X 5 n − 1 i = 0 ∑ 4 X i j = 0 ∑ i ∂ ρ σ ( j ) P ( 1 ) .
Indeed, with X 0 = ⋯ = X 4 = 0 X_0=\cdots=X_4=0 X 0 = ⋯ = X 4 = 0 and X 5 = 1 X_5=1 X 5 = 1 , (18) gives ρ = 1 \rho=1 ρ = 1 and the derivative in X i X_i X i is the sum of the derivatives in ρ σ ( 0 ) , … , ρ σ ( i ) \rho_{\sigma(0)},\ldots,\rho_{\sigma(i)} ρ σ ( 0 ) , … , ρ σ ( i ) . We call these six monomials the first-order positions.
Comparison for bounded exponents
We prove BCC optimality and uniqueness for 1 ≤ p ≤ 35 1\leq p\leq35 1 ≤ p ≤ 35 . A lower bound for the power of the scaled squared radius is combined with upper bounds for the fractional determinant powers. The resulting coefficient inequalities are quartic Taylor for 1 ≤ p ≤ 2 1\leq p\leq2 1 ≤ p ≤ 2 , cubic Taylor for 2 ≤ p ≤ 4 2\leq p\leq4 2 ≤ p ≤ 4 , weighted quadratic for 4 ≤ p ≤ 6 4\leq p\leq6 4 ≤ p ≤ 6 , and weighted cubic for 6 ≤ p ≤ 35 6\leq p\leq35 6 ≤ p ≤ 35 .
Proposition 3.1. Theorem 1.1 holds for 1 ≤ p ≤ 35 1\leq p\leq35 1 ≤ p ≤ 35 .
Lemma 3.2. Let 0 < β < 1 0<\beta<1 0 < β < 1 and 0 < t ≤ 1 0<t\leq1 0 < t ≤ 1 . Then
t β ≤ ( 1 − β ) + ( 1 + β ) t ( 1 + β ) + ( 1 − β ) t (20) t^\beta\leq\frac{(1-\beta)+(1+\beta)t}{(1+\beta)+(1-\beta)t} \tag*{(20)} t β ≤ ( 1 + β ) + ( 1 − β ) t ( 1 − β ) + ( 1 + β ) t (20) t j / 3 ≤ 1 − j 3 ( 1 − t ) , j = 1 , 2 , 3. (21) t^{j/3}\leq1-\frac{j}{3}(1-t),\qquad j=1,2,3. \tag*{(21)} t j /3 ≤ 1 − 3 j ( 1 − t ) , j = 1 , 2 , 3. (21) For t < 1 t<1 t < 1 the inequality in (20) is strict, as is (21) for j = 1 , 2 j=1,2 j = 1 , 2 .
Proof. Write t = e − 2 u t=e^{-2u} t = e − 2 u . Strict concavity of tanh \tanh tanh on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) gives
1 − t β 1 + t β = tanh ( β u ) ≥ β tanh u = β 1 − t 1 + t . \frac{1-t^\beta}{1+t^\beta}=\tanh(\beta u)\geq\beta\tanh u=\beta\frac{1-t}{1+t}. 1 + t β 1 − t β = tanh ( β u ) ≥ β tanh u = β 1 + t 1 − t .
The inequality is strict for u > 0 u>0 u > 0 and rearranges to (20) . For (21) , use concavity of t j / 3 t^{j/3} t j /3 , which is strict for j = 1 , 2 j=1,2 j = 1 , 2 . □
Lemma 3.3. For r ≥ 0 r\geq0 r ≥ 0 and 1 / 2 ≤ α ≤ 2 1/2\leq\alpha\leq2 1/2 ≤ α ≤ 2 , one has
r α ≥ ∑ j = 0 3 ( α j ) ( r − 1 ) j + { ( α 3 ) ( r − 1 ) 4 , 1 / 2 ≤ α ≤ 1 , 0 , 1 ≤ α ≤ 2. (22) r^\alpha\geq\sum_{j=0}^{3}\binom{\alpha}{j}(r-1)^j+
\begin{cases}
\binom{\alpha}{3}(r-1)^4, & 1/2\leq\alpha\leq1,\\
0, & 1\leq\alpha\leq2.
\end{cases} \tag*{(22)} r α ≥ j = 0 ∑ 3 ( j α ) ( r − 1 ) j + { ( 3 α ) ( r − 1 ) 4 , 0 , 1/2 ≤ α ≤ 1 , 1 ≤ α ≤ 2. (22)
Proof. For r > 0 r>0 r > 0 , r ≠ 1 r\ne1 r = 1 , the integral remainder gives
r α − ∑ j = 0 3 ( α j ) ( r − 1 ) j ( r − 1 ) 4 = α ( α − 1 ) ( α − 2 ) ( α − 3 ) 3 ! ∫ 0 1 ( 1 − τ ) 3 ( 1 + τ ( r − 1 ) ) α − 4 d τ . \frac{r^\alpha-\sum_{j=0}^{3}\binom{\alpha}{j}(r-1)^j}{(r-1)^4}
=
\frac{\alpha(\alpha-1)(\alpha-2)(\alpha-3)}{3!}
\int_{0}^{1}(1-\tau)^3(1+\tau(r-1))^{\alpha-4}\,\mathrm{d}\tau. ( r − 1 ) 4 r α − ∑ j = 0 3 ( j α ) ( r − 1 ) j = 3 ! α ( α − 1 ) ( α − 2 ) ( α − 3 ) ∫ 0 1 ( 1 − τ ) 3 ( 1 + τ ( r − 1 ) ) α − 4 d τ .
For 1 ≤ α ≤ 2 1\leq\alpha\leq2 1 ≤ α ≤ 2 the right-hand side is nonnegative. For 1 / 2 ≤ α < 1 1/2\leq\alpha<1 1/2 ≤ α < 1 its prefactor is negative, and
( 1 + τ ( r − 1 ) ) α − 4 ≤ ( 1 − τ ) α − 4 , 0 ≤ τ < 1. (1+\tau(r-1))^{\alpha-4}\leq(1-\tau)^{\alpha-4},\qquad0\leq\tau<1. ( 1 + τ ( r − 1 ) ) α − 4 ≤ ( 1 − τ ) α − 4 , 0 ≤ τ < 1.
Thus the quotient is bounded below by
α ( α − 1 ) ( α − 2 ) ( α − 3 ) 3 ! ∫ 0 1 ( 1 − τ ) α − 1 d τ = ( α − 1 3 ) . \frac{\alpha(\alpha-1)(\alpha-2)(\alpha-3)}{3!}\int_0^1(1-\tau)^{\alpha-1}\,\mathrm{d}\tau=\binom{\alpha-1}{3}. 3 ! α ( α − 1 ) ( α − 2 ) ( α − 3 ) ∫ 0 1 ( 1 − τ ) α − 1 d τ = ( 3 α − 1 ) .
Continuity gives (22) also at r = 0 , 1 r=0,1 r = 0 , 1 . □
For α = p / 2 \alpha=p/2 α = p /2 and k ∈ Z ≥ 0 k\in\mathbb{Z}_{\geq0} k ∈ Z ≥ 0 , let
I k = 1 4 ∑ T ∈ T W T ∫ [ − 1 , 1 ] 2 ( 4 q T det Q ) k q 0 , T α − k d x d y . I_k=\frac{1}{4}\sum_{T\in\mathcal{T}}W_T\int_{[-1,1]^2}(4q_T\det Q)^kq_{0,T}^{\alpha-k}\,\mathrm{d}x\,\mathrm{d}y. I k = 4 1 T ∈ T ∑ W T ∫ [ − 1 , 1 ] 2 ( 4 q T det Q ) k q 0 , T α − k d x d y .
By (10) , the parallelogram terms use q 0 , s q q_{0,\mathrm{sq}} q 0 , sq . The hexagon terms use q 0 , r h q_{0,\mathrm{rh}} q 0 , rh after y ↦ − y y\mapsto-y y ↦ − y on the tiles with pairs ( i j , i ℓ ) (ij,i\ell) ( ij , i ℓ ) and ( j ℓ , i ℓ ) (j\ell,i\ell) ( j ℓ , i ℓ ) . Thus
4 I k ( 4 det Q ) k = ∫ [ − 1 , 1 ] 2 q 0 , s q α − k [ W 01 , 23 q 01 , 23 ( x , y ) k + W 02 , 13 q 02 , 13 ( x , y ) k + W 12 , 03 q 12 , 03 ( x , y ) k ] d x d y + ∑ ( i , j , ℓ ) ∈ { ( 1 , 2 , 3 ) , ( 0 , 2 , 3 ) , ( 0 , 1 , 3 ) , ( 0 , 1 , 2 ) } ∫ [ − 1 , 1 ] 2 q 0 , r h α − k [ W i j , j ℓ q i j , j ℓ ( x , y ) k + W i j , i ℓ q i j , i ℓ ( x , − y ) k + W j ℓ , i ℓ q j ℓ , i ℓ ( x , − y ) k ] d x d y . (23) \begin{aligned}
\frac{4I_k}{(4\det Q)^k}=\int_{[-1,1]^2}q_{0,\mathrm{sq}}^{\alpha-k}\left[W_{01,23}q_{01,23}(x,y)^k+W_{02,13}q_{02,13}(x,y)^k+W_{12,03}q_{12,03}(x,y)^k\right]\,\mathrm{d}x\,\mathrm{d}y \\
&+\sum_{(i,j,\ell)\in\{(1,2,3),(0,2,3),(0,1,3),(0,1,2)\}}\int_{[-1,1]^2}q_{0,\mathrm{rh}}^{\alpha-k}\left[W_{ij,j\ell}q_{ij,j\ell}(x,y)^k+W_{ij,i\ell}q_{ij,i\ell}(x,-y)^k+W_{j\ell,i\ell}q_{j\ell,i\ell}(x,-y)^k\right]\,\mathrm{d}x\,\mathrm{d}y. \tag*{(23)}
\end{aligned} ( 4 det Q ) k 4 I k = ∫ [ − 1 , 1 ] 2 q 0 , sq α − k [ W 01 , 23 q 01 , 23 ( x , y ) k + W 02 , 13 q 02 , 13 ( x , y ) k + W 12 , 03 q 12 , 03 ( x , y ) k ] d x d y + ( i , j , ℓ ) ∈ {( 1 , 2 , 3 ) , ( 0 , 2 , 3 ) , ( 0 , 1 , 3 ) , ( 0 , 1 , 2 )} ∑ ∫ [ − 1 , 1 ] 2 q 0 , rh α − k [ W ij , j ℓ q ij , j ℓ ( x , y ) k + W ij , i ℓ q ij , i ℓ ( x , − y ) k + W j ℓ , i ℓ q j ℓ , i ℓ ( x , − y ) k ] d x d y . (23)
Taking k = 0 k=0 k = 0 and using ∑ T ∈ T W T = 3 det Q \sum_{T\in\mathcal{T}}W_T=3\det Q ∑ T ∈ T W T = 3 det Q gives
4 I 0 = ∫ [ − 1 , 1 ] 2 [ ( W 01 , 23 + W 02 , 13 + W 12 , 03 ) ( q 0 , s q α − q 0 , r h α ) + 3 ( det Q ) q 0 , r h α ] d x d y . (24) 4I_0=\int_{[-1,1]^2}\left[(W_{01,23}+W_{02,13}+W_{12,03})(q_{0,\mathrm{sq}}^\alpha-q_{0,\mathrm{rh}}^\alpha)+3(\det Q)q_{0,\mathrm{rh}}^\alpha\right]\,\mathrm{d}x\,\mathrm{d}y. \tag*{(24)} 4 I 0 = ∫ [ − 1 , 1 ] 2 [ ( W 01 , 23 + W 02 , 13 + W 12 , 03 ) ( q 0 , sq α − q 0 , rh α ) + 3 ( det Q ) q 0 , rh α ] d x d y . (24)
With 4 q T det Q 4q_T\det Q 4 q T det Q given by (16) , expand
W T ( 4 q T det Q ) k = ∑ ν ∈ Z ≥ 0 a , b ∈ Z ≥ 0 , a + b ≤ 2 k ∑ e ∈ E ν e = 4 k + 3 d ν , a , b ( T , k ) ( ∏ e ∈ E ρ e ν e ) x a y b . W_T(4q_T\det Q)^k=\sum_{\substack{\nu\in\mathbb{Z}_{\geq0}\\ a,b\in\mathbb{Z}_{\geq0},\ a+b\leq2k}}\sum_{\substack{e\in\mathcal{E}\\ \nu_e=4k+3}}d_{\nu,a,b}(T,k)\left(\prod_{e\in\mathcal{E}}\rho_e^{\nu_e}\right)x^ay^b. W T ( 4 q T det Q ) k = ν ∈ Z ≥ 0 a , b ∈ Z ≥ 0 , a + b ≤ 2 k ∑ e ∈ E ν e = 4 k + 3 ∑ d ν , a , b ( T , k ) ( e ∈ E ∏ ρ e ν e ) x a y b .
Termwise integration yields the conorm polynomial
I k = 1 4 ∑ T ∈ T ∑ ν ∈ Z ≥ 0 a , b ∈ Z ≥ 0 , a + b ≤ 2 k ∑ e ∈ E ν e = 4 k + 3 d ν , a , b ( T , k ) ( ∏ e ∈ E ρ e ν e ) ∫ [ − 1 , 1 ] 2 x a y b q 0 , T α − k d x d y . (25) I_k=\frac{1}{4}\sum_{T\in\mathcal{T}}\sum_{\substack{\nu\in\mathbb{Z}_{\geq0}\\ a,b\in\mathbb{Z}_{\geq0},\ a+b\leq2k}}\sum_{\substack{e\in\mathcal{E}\\ \nu_e=4k+3}}d_{\nu,a,b}(T,k)\left(\prod_{e\in\mathcal{E}}\rho_e^{\nu_e}\right)\int_{[-1,1]^2}x^ay^bq_{0,T}^{\alpha-k}\,\mathrm{d}x\,\mathrm{d}y. \tag*{(25)} I k = 4 1 T ∈ T ∑ ν ∈ Z ≥ 0 a , b ∈ Z ≥ 0 , a + b ≤ 2 k ∑ e ∈ E ν e = 4 k + 3 ∑ d ν , a , b ( T , k ) ( e ∈ E ∏ ρ e ν e ) ∫ [ − 1 , 1 ] 2 x a y b q 0 , T α − k d x d y . (25)
Write the right-hand side of (22) as ∑ k = 0 K c k ( α ) r k \sum_{k=0}^{K}c_k(\alpha)r^k ∑ k = 0 K c k ( α ) r k , with K = 4 K=4 K = 4 on [ 1 / 2 , 1 ] [1/2,1] [ 1/2 , 1 ] and K = 3 K=3 K = 3 on [ 1 , 2 ] [1,2] [ 1 , 2 ] . The two polynomials agree at α = 1 \alpha=1 α = 1 . Expanding the powers of r − 1 r-1 r − 1 and using
∑ j = 0 m ( − 1 ) j ( j m ) = ( − 1 ) m ( x − 1 m ) \sum_{j=0}^{m}(-1)^j\binom{j}{m} = (-1)^m\binom{x-1}{m} j = 0 ∑ m ( − 1 ) j ( m j ) = ( − 1 ) m ( m x − 1 )
gives
c k ( α ) = { ( − 1 ) 3 − k ( α k ) ( α − k − 1 3 − k ) , K = 3 , 0 ≤ k ≤ 3 , ( − 1 ) k [ ( 4 k ) ( α − 1 3 ) − ( α k ) ( α − k − 1 3 − k ) ] , K = 4 , 0 ≤ k ≤ 4. c_k(\alpha)=
\begin{cases}
(-1)^{3-k}\binom{\alpha}{k}\binom{\alpha-k-1}{3-k}, & K=3,\quad0\leq k\leq3,\\
(-1)^k\left[\binom{4}{k}\binom{\alpha-1}{3}-\binom{\alpha}{k}\binom{\alpha-k-1}{3-k}\right], & K=4,\quad0\leq k\leq4.
\end{cases} c k ( α ) = { ( − 1 ) 3 − k ( k α ) ( 3 − k α − k − 1 ) , ( − 1 ) k [ ( k 4 ) ( 3 α − 1 ) − ( k α ) ( 3 − k α − k − 1 ) ] , K = 3 , 0 ≤ k ≤ 3 , K = 4 , 0 ≤ k ≤ 4.
where ( x − 1 ) = 0 \binom{x}{-1}=0 ( − 1 x ) = 0 . In both cases $\sum_{k=0}^{K}c_k(\alpha)=1.`
The Taylor comparison uses the polynomial
P α T = [ ( 1 + α 3 ) ( ∑ e ∈ E ρ e ) 3 + 9 2 ( 3 − α ) det Q ] P_{\alpha}^{\mathrm{T}}=\left[\left(1+\frac{\alpha}{3}\right)\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3+\frac{9}{2}(3-\alpha)\det Q\right] P α T = ( 1 + 3 α ) ( e ∈ E ∑ ρ e ) 3 + 2 9 ( 3 − α ) det Q
× ∑ k = 0 K 3 k c k ( α ) ( 2 det Q ∑ e ∈ E ρ e ) K − k I k {}\times\sum_{k=0}^{K}3^k c_k(\alpha)\left(2\det Q\sum_{e\in\mathcal{E}}\rho_e\right)^{K-k}I_k × k = 0 ∑ K 3 k c k ( α ) ( 2 det Q e ∈ E ∑ ρ e ) K − k I k
− det Q 16 ( 2 det Q ∑ e ∈ E ρ e ) K [ ( 1 − α 3 ) ( ∑ e ∈ E ρ e ) 3 + 9 2 ( 3 + α ) det Q ] {}-\frac{\det Q}{16}\left(2\det Q\sum_{e\in\mathcal{E}}\rho_e\right)^K\left[\left(1-\frac{\alpha}{3}\right)\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3+\frac{9}{2}(3+\alpha)\det Q\right] − 16 det Q ( 2 det Q e ∈ E ∑ ρ e ) K ( 1 − 3 α ) ( e ∈ E ∑ ρ e ) 3 + 2 9 ( 3 + α ) det Q
× ∫ [ − 1 , 1 ] 2 ( 3 q 0 , s q α + 9 q 0 , r h α ) d x d y . (26) {}\times\int_{[-1,1]^2}\left(3q_{0,\mathrm{sq}}^\alpha+9q_{0,\mathrm{rh}}^\alpha\right)\,dx\,dy. \tag*{(26)} × ∫ [ − 1 , 1 ] 2 ( 3 q 0 , sq α + 9 q 0 , rh α ) d x d y . (26)
The quadratic and cubic comparisons use
P α ( 2 ) = 864 ( α − 1 ) ( ∑ e ∈ E ρ e ) 3 I 2 P_{\alpha}^{(2)}=864(\alpha-1)\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3I_2 P α ( 2 ) = 864 ( α − 1 ) ( e ∈ E ∑ ρ e ) 3 I 2
− 96 ( α − 2 ) det Q ( ∑ e ∈ E ρ e ) [ 4 ( ∑ e ∈ E ρ e ) 3 + 27 det Q ] I 1 {}-96(\alpha-2)\det Q\left(\sum_{e\in\mathcal{E}}\rho_e\right)\left[4\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3+27\det Q\right]I_1 − 96 ( α − 2 ) det Q ( e ∈ E ∑ ρ e ) 4 ( e ∈ E ∑ ρ e ) 3 + 27 det Q I 1
− det Q ( 2 det Q ∑ e ∈ E ρ e ) 2 [ 2 ( ∑ e ∈ E ρ e ) 3 + 54 det Q ] {}-\det Q\left(2\det Q\sum_{e\in\mathcal{E}}\rho_e\right)^2\left[2\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3+54\det Q\right] − det Q ( 2 det Q e ∈ E ∑ ρ e ) 2 2 ( e ∈ E ∑ ρ e ) 3 + 54 det Q
× ∫ [ − 1 , 1 ] 2 ( 3 q 0 , s q α + 9 q 0 , r h α ) d x d y , (27) {}\times\int_{[-1,1]^2}\left(3q_{0,\mathrm{sq}}^\alpha+9q_{0,\mathrm{rh}}^\alpha\right)\,dx\,dy, \tag*{(27)} × ∫ [ − 1 , 1 ] 2 ( 3 q 0 , sq α + 9 q 0 , rh α ) d x d y , (27)
P α ( 3 ) = 288 ( α − 2 ) ( ∑ e ∈ E ρ e ) 3 I 3 P_{\alpha}^{(3)}=288(\alpha-2)\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3I_3 P α ( 3 ) = 288 ( α − 2 ) ( e ∈ E ∑ ρ e ) 3 I 3
− 32 ( α − 3 ) det Q ( ∑ e ∈ E ρ e ) [ 4 ( ∑ e ∈ E ρ e ) 3 + 27 det Q ] I 2 {}-32(\alpha-3)\det Q\left(\sum_{e\in\mathcal{E}}\rho_e\right)\left[4\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3+27\det Q\right]I_2 − 32 ( α − 3 ) det Q ( e ∈ E ∑ ρ e ) 4 ( e ∈ E ∑ ρ e ) 3 + 27 det Q I 2
− 9 ( det Q ) 2 ( 2 det Q ∑ e ∈ E ρ e ) 3 ∫ [ − 1 , 1 ] 2 ( 3 q 0 , s q α + 9 q 0 , r h α ) d x d y . (28) {}-9(\det Q)^2\left(2\det Q\sum_{e\in\mathcal{E}}\rho_e\right)^3\int_{[-1,1]^2}\left(3q_{0,\mathrm{sq}}^\alpha+9q_{0,\mathrm{rh}}^\alpha\right)\,dx\,dy. \tag*{(28)} − 9 ( det Q ) 2 ( 2 det Q e ∈ E ∑ ρ e ) 3 ∫ [ − 1 , 1 ] 2 ( 3 q 0 , sq α + 9 q 0 , rh α ) d x d y . (28)
The moments I k I_k I k are invariant under vertex relabelling by (34) , as are det Q \det Q det Q and ∑ e ∈ E ρ e \sum_{e\in\mathcal{E}}\rho_e ∑ e ∈ E ρ e . Each comparison polynomial is therefore invariant. Every cornn tuple can be relabelled into one of the thirty sectors in (18) , where X i ≥ 0 X_i\geq0 X i ≥ 0 for all i i i . Nonnegative coefficients in these sectors imply nonnegativity for every cornn tuple.
Proof of Proposition 3.1. By (14) ,
0 < t = 27 det Q 2 ( ∑ e ∈ E ρ e ) 3 ≤ 1. 0<t=\frac{27\det Q}{2\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3}\leq1. 0 < t = 2 ( ∑ e ∈ E ρ e ) 3 27 det Q ≤ 1.
For each tile,
6 q T q 0 , T ∑ e ∈ E ρ e = t 1 / 3 ( 16 det Q ) 1 / 3 q T q 0 , T = 3 ( 4 q T det Q ) 2 q 0 , T ( det Q ) ∑ e ∈ E ρ e . (29) \frac{6q_T}{q_{0,T}\sum_{e\in\mathcal{E}}\rho_e}=t^{1/3}\left(\frac{16}{\det Q}\right)^{1/3}\frac{q_T}{q_{0,T}}=\frac{3(4q_T\det Q)}{2q_{0,T}(\det Q)\sum_{e\in\mathcal{E}}\rho_e}. \tag*{(29)} q 0 , T ∑ e ∈ E ρ e 6 q T = t 1/3 ( det Q 16 ) 1/3 q 0 , T q T = 2 q 0 , T ( det Q ) ∑ e ∈ E ρ e 3 ( 4 q T det Q ) . (29)
In particular,
1 4 ∑ T ∈ T W T ∫ [ − 1 , 1 ] 2 q 0 , T α ( 6 q T q 0 , T ∑ e ∈ E ρ e ) k d x d y = ( 3 2 det Q ∑ e ∈ E ρ e ) k I k . (30) \frac{1}{4}\sum_{T\in\mathcal{T}}W_T\int_{[-1,1]^2}q_{0,T}^\alpha\left(\frac{6q_T}{q_{0,T}\sum_{e\in\mathcal{E}}\rho_e}\right)^k\,dx\,dy=\left(\frac{3}{2\det Q\sum_{e\in\mathcal{E}}\rho_e}\right)^kI_k. \tag*{(30)} 4 1 T ∈ T ∑ W T ∫ [ − 1 , 1 ] 2 q 0 , T α ( q 0 , T ∑ e ∈ E ρ e 6 q T ) k d x d y = ( 2 det Q ∑ e ∈ E ρ e 3 ) k I k . (30)
Case 1. 1 ≤ p ≤ 4 1 \le p \le4 1 ≤ p ≤ 4 . By (22) , (30) and (15) ,
det Q 16 t α / 3 E p ( A ) E p ( A b c c ) ∫ [ − 1 , 1 ] 2 ( 3 q 0 , s q α + 9 q 0 , r h α ) d x d y ≥ ∑ k = 0 K c k ( α ) ( 3 2 det Q ∑ e ∈ E ρ e ) k I k ≥ det Q 16 ( 1 − α / 3 ) + ( 1 + α / 3 ) t ( 1 + α / 3 ) + ( 1 − α / 3 ) t ∫ [ − 1 , 1 ] 2 ( 3 q 0 , s q α + 9 q 0 , r h α ) d x d y . \frac{\det Q}{16}t^{\alpha/3}\frac{E_p(A)}{E_p(A_{\mathrm{bcc}})}\int_{[-1,1]^2}(3q_{0,\mathrm{sq}}^\alpha+9q_{0,\mathrm{rh}}^\alpha)\,dx\,dy
\ge\sum_{k=0}^{K}c_k(\alpha)\left(\frac{3}{2\det Q\sum_{e\in\mathcal{E}}\rho_e}\right)^k I_k
\ge\frac{\det Q}{16}\frac{(1-\alpha/3)+(1+\alpha/3)t}{(1+\alpha/3)+(1-\alpha/3)t}\int_{[-1,1]^2}(3q_{0,\mathrm{sq}}^\alpha+9q_{0,\mathrm{rh}}^\alpha)\,dx\,dy. 16 det Q t α /3 E p ( A bcc ) E p ( A ) ∫ [ − 1 , 1 ] 2 ( 3 q 0 , sq α + 9 q 0 , rh α ) d x d y ≥ k = 0 ∑ K c k ( α ) ( 2 det Q ∑ e ∈ E ρ e 3 ) k I k ≥ 16 det Q ( 1 + α /3 ) + ( 1 − α /3 ) t ( 1 − α /3 ) + ( 1 + α /3 ) t ∫ [ − 1 , 1 ] 2 ( 3 q 0 , sq α + 9 q 0 , rh α ) d x d y .
The last inequality is P α T ≥ 0 P_\alpha^{\mathsf{T}}\ge0 P α T ≥ 0 from Numerical Lemmas 5.1 and 5.2, divided by
( 2 det Q ∑ e ∈ E ρ e ) K ( ∑ e ∈ E ρ e ) 3 [ ( 1 + α / 3 ) + ( 1 − α / 3 ) t ] > 0. \left(2\det Q\sum_{e\in\mathcal{E}}\rho_e\right)^K\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3[(1+\alpha/3)+(1-\alpha/3)t]>0. ( 2 det Q e ∈ E ∑ ρ e ) K ( e ∈ E ∑ ρ e ) 3 [( 1 + α /3 ) + ( 1 − α /3 ) t ] > 0.
By (20) ,
E p ( A ) E p ( A b c c ) ≥ ( 1 − α / 3 ) + ( 1 + α / 3 ) t t α / 3 [ ( 1 + α / 3 ) + ( 1 − α / 3 ) t ] ≥ 1 , \frac{E_p(A)}{E_p(A_{\mathrm{bcc}})}\ge\frac{(1-\alpha/3)+(1+\alpha/3)t}{t^{\alpha/3}[(1+\alpha/3)+(1-\alpha/3)t]}\ge1, E p ( A bcc ) E p ( A ) ≥ t α /3 [( 1 + α /3 ) + ( 1 − α /3 ) t ] ( 1 − α /3 ) + ( 1 + α /3 ) t ≥ 1 ,
with strict inequality for t < 1 t<1 t < 1 .
Case 2. 4 ≤ p ≤ 35 4\le p\le35 4 ≤ p ≤ 35 . Take k = 2 k=2 k = 2 for 4 ≤ p ≤ 6 4\le p\le6 4 ≤ p ≤ 6 and k = 3 k=3 k = 3 for 6 < p ≤ 35 6<p\le35 6 < p ≤ 35 . The tangent to s α − k + 1 s^{\alpha-k+1} s α − k + 1 at one gives
s α ≥ ( α − k + 1 ) s k − ( α − k ) s k − 1 , s ≥ 0. (31) s^\alpha\ge(\alpha-k+1)s^k-(\alpha-k)s^{k-1},\qquad s\ge0. \tag*{(31)} s α ≥ ( α − k + 1 ) s k − ( α − k ) s k − 1 , s ≥ 0. (31)
Use (31) with s = ( 16 / det Q ) 1 / 3 q T / q 0 , T s=(16/\det Q)^{1/3}q_T/q_{0,T} s = ( 16/ det Q ) 1/3 q T / q 0 , T in the tile integrals. Equations (15) , (29) and (30) give
det Q 16 ( 2 det Q ∑ e ∈ E ρ e ) k / 3 ( E p ( A ) E p ( A b c c ) − 1 ) ∫ [ − 1 , 1 ] 2 ( 3 q 0 , s q α + 9 q 0 , r h α ) d x d y ≥ 3 k ( α − k + 1 ) I k − 3 k − 1 ( α − k ) ( 2 det Q ∑ e ∈ E ρ e ) t 1 / 3 I k − 1 − det Q 16 ( 2 det Q ∑ e ∈ E ρ e ) k t k / 3 ∫ [ − 1 , 1 ] 2 ( 3 q 0 , s q α + 9 q 0 , r h α ) d x d y ≥ 1 96 ( ∑ e ∈ E ρ e ) 3 { P α ( 2 ) , k = 2 , 9 P α ( 3 ) , k = 3 ≥ 0. \begin{aligned}
\frac{\det Q}{16}\left(2\det Q\sum_{e\in\mathcal{E}}\rho_e\right)^{k/3}\left(\frac{E_p(A)}{E_p(A_{\mathrm{bcc}})}-1\right)\int_{[-1,1]^2}(3q_{0,\mathrm{sq}}^\alpha+9q_{0,\mathrm{rh}}^\alpha)\,dx\,dy
&\ge3^k(\alpha-k+1)I_k-3^{k-1}(\alpha-k)\left(2\det Q\sum_{e\in\mathcal{E}}\rho_e\right)t^{1/3}I_{k-1}\\
&\quad-\frac{\det Q}{16}\left(2\det Q\sum_{e\in\mathcal{E}}\rho_e\right)^{k}t^{k/3}\int_{[-1,1]^2}(3q_{0,\mathrm{sq}}^\alpha+9q_{0,\mathrm{rh}}^\alpha)\,dx\,dy\\
&\ge\frac{1}{96\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3}
\begin{cases}
P_\alpha^{(2)},& k=2,\\
9P_\alpha^{(3)},& k=3
\end{cases}
\ge0.
\end{aligned} 16 det Q ( 2 det Q e ∈ E ∑ ρ e ) k /3 ( E p ( A bcc ) E p ( A ) − 1 ) ∫ [ − 1 , 1 ] 2 ( 3 q 0 , sq α + 9 q 0 , rh α ) d x d y ≥ 3 k ( α − k + 1 ) I k − 3 k − 1 ( α − k ) ( 2 det Q e ∈ E ∑ ρ e ) t 1/3 I k − 1 − 16 det Q ( 2 det Q e ∈ E ∑ ρ e ) k t k /3 ∫ [ − 1 , 1 ] 2 ( 3 q 0 , sq α + 9 q 0 , rh α ) d x d y ≥ 96 ( ∑ e ∈ E ρ e ) 3 1 { P α ( 2 ) , 9 P α ( 3 ) , k = 2 , k = 3 ≥ 0.
The second inequality uses (21) and the identities
6 ( ∑ e ∈ E ρ e ) 3 ( 2 / 3 + t / 3 ) = 4 ( ∑ e ∈ E ρ e ) 3 + 27 det Q , 6\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3(2/3+t/3)=4\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3+27\det Q, 6 ( e ∈ E ∑ ρ e ) 3 ( 2/3 + t /3 ) = 4 ( e ∈ E ∑ ρ e ) 3 + 27 det Q ,
6 ( ∑ e ∈ E ρ e ) 3 ( 1 / 3 + 2 t / 3 ) = 2 ( ∑ e ∈ E ρ e ) 3 + 54 det Q , 6\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3(1/3+2t/3)=2\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3+54\det Q, 6 ( e ∈ E ∑ ρ e ) 3 ( 1/3 + 2 t /3 ) = 2 ( e ∈ E ∑ ρ e ) 3 + 54 det Q ,
6 ( ∑ e ∈ E ρ e ) 3 t = 81 det Q . 6\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3t=81\det Q. 6 ( e ∈ E ∑ ρ e ) 3 t = 81 det Q .
The last inequality is Numerical Lemma 5.3 for k = 2 k=2 k = 2 and Numerical Lemma 5.4 for k = 3 k=3 k = 3 . For t < 1 t<1 t < 1 , the error from the tangent bounds in (21) includes a positive multiple of
{ ( 1 3 + 2 3 t − t 2 / 3 ) ∫ [ − 1 , 1 ] 2 ( 3 q 0 , s q α + 9 q 0 , r h α ) d x d y , k = 2 , ( α − 3 ) ( 2 3 + 1 3 t − t 1 / 3 ) I 2 , k = 3. \begin{cases}
\left(\frac{1}{3}+\frac{2}{3}t-t^{2/3}\right)\int_{[-1,1]^2}(3q_{0,\mathrm{sq}}^\alpha+9q_{0,\mathrm{rh}}^\alpha)\,dx\,dy,& k=2,\\
(\alpha-3)\left(\frac{2}{3}+\frac{1}{3}t-t^{1/3}\right)I_2,& k=3.
\end{cases} { ( 3 1 + 3 2 t − t 2/3 ) ∫ [ − 1 , 1 ] 2 ( 3 q 0 , sq α + 9 q 0 , rh α ) d x d y , ( α − 3 ) ( 3 2 + 3 1 t − t 1/3 ) I 2 , k = 2 , k = 3.
The integral is positive. Also α > 3 \alpha>3 α > 3 when k = 3 k=3 k = 3 , and I 2 > 0 I_2>0 I 2 > 0 since ∑ T ∈ T W T = 3 det Q > 0 \sum_{T\in\mathcal{T}}W_T=3\det Q>0 ∑ T ∈ T W T = 3 det Q > 0 and q T ≥ η T Q η / 4 > 0 q_T\ge\eta^{\mathsf{T}}Q\eta/4>0 q T ≥ η T Q η /4 > 0 by (8) . Thus the comparison is strict for t < 1 t<1 t < 1 . By (14) , t = 1 t=1 t = 1 gives precisely the unit-covolume orbit in (2) . □
At ρ = 1 \rho=1 ρ = 1 , the values in the comparison polynomials are
det Q = 16 , ∑ e ∈ E ρ e = 6 , I k ( 1 ) = 64 k ∫ [ − 1 , 1 ] 2 ( 3 q 0 , s q k + 9 q 0 , r h k ) d x d y . \det Q=16,\qquad\sum_{e\in\mathcal{E}}\rho_e=6,\qquad I_k(1)=64^k\int_{[-1,1]^2}(3q_{0,\mathrm{sq}}^k+9q_{0,\mathrm{rh}}^k)\,dx\,dy. det Q = 16 , e ∈ E ∑ ρ e = 6 , I k ( 1 ) = 6 4 k ∫ [ − 1 , 1 ] 2 ( 3 q 0 , sq k + 9 q 0 , rh k ) d x d y .
In (26) , both bracketed factors equal 432, and 2 det Q ∑ e ∈ E ρ e = 192 = 3 ⋅ 64 2\det Q\sum_{e\in\mathcal{E}}\rho_e=192=3\cdot64 2 det Q ∑ e ∈ E ρ e = 192 = 3 ⋅ 64 . Consequently,
P α T ( 1 ) = 432 ⋅ 192 K I 0 ( 1 ) ( ∑ k = 0 K c k ( α ) − 1 ) = 0 , P_{\alpha}^{\mathrm{T}}(1)=432\cdot192^K I_0(1)\left(\sum_{k=0}^{K}c_k(\alpha)-1\right)=0, P α T ( 1 ) = 432 ⋅ 19 2 K I 0 ( 1 ) ( k = 0 ∑ K c k ( α ) − 1 ) = 0 ,
P α ( 2 ) ( 1 ) = 864 ⋅ 6 3 I 2 ( 1 ) ( ( α − 1 ) − ( α − 2 ) − 1 ) = 0 , P_{\alpha}^{(2)}(1)=864\cdot6^3I_2(1)((\alpha-1)-(\alpha-2)-1)=0, P α ( 2 ) ( 1 ) = 864 ⋅ 6 3 I 2 ( 1 ) (( α − 1 ) − ( α − 2 ) − 1 ) = 0 ,
P α ( 3 ) ( 1 ) = 288 ⋅ 6 3 I 3 ( 1 ) ( ( α − 2 ) − ( α − 3 ) − 1 ) = 0. P_{\alpha}^{(3)}(1)=288\cdot6^3I_3(1)((\alpha-2)-(\alpha-3)-1)=0. P α ( 3 ) ( 1 ) = 288 ⋅ 6 3 I 3 ( 1 ) (( α − 2 ) − ( α − 3 ) − 1 ) = 0.
Each polynomial is homogeneous and, by (34) , invariant under relabelling the superbase. Since the vertex permutations act transitively on the six edges, Euler’s identity gives
∂ ρ e P ( 1 ) = 1 6 ∑ f ∈ E ∂ ρ f P ( 1 ) = deg P 6 P ( 1 ) = 0. \partial_{\rho_e}P(1)=\frac{1}{6}\sum_{f\in\mathcal{E}}\partial_{\rho_f}P(1)=\frac{\deg P}{6}P(1)=0. ∂ ρ e P ( 1 ) = 6 1 f ∈ E ∑ ∂ ρ f P ( 1 ) = 6 deg P P ( 1 ) = 0.
Hence the six first-order coefficients in each sector vanish.
Comparison for large exponents
For p ≥ 35 p\ge35 p ≥ 35 , the tangent inequality for the normalized squared radius reduces the comparison to coefficient inequalities for a degree-ten polynomial. The six BCC integrals in its coefficients are controlled by Bernstein bounds for 35 ≤ p ≤ 1028 35\le p\le1028 35 ≤ p ≤ 1028 and by an estimate at the maximizing corners for p ≥ 1026 p\ge1026 p ≥ 1026 .
Proposition 4.1. For 35 ≤ p < ∞ 35\le p<\infty 35 ≤ p < ∞ and A ∈ G L ( 3 , R ) A\in GL(3,\mathbb{R}) A ∈ G L ( 3 , R ) ,
E p ( A ) ≥ E p ( A b c c ) . E_p(A)\ge E_p(A_{\mathrm{bcc}}). E p ( A ) ≥ E p ( A bcc ) . Among lattices of unit covolume, equality holds exactly on the orbit in (2) .
Let α = p / 2 \alpha=p/2 α = p /2 . By (10) , 3 / 5 ≤ 4 q 0 , T / 5 ≤ 1 3/5\le4q_{0,T}/5\le1 3/5 ≤ 4 q 0 , T /5 ≤ 1 on [ − 1 , 1 ] 2 [-1,1]^2 [ − 1 , 1 ] 2 .
Lemma 4.2 (Polynomial comparison). Let α ≥ 35 / 2 \alpha\ge35/2 α ≥ 35/2 . For a reduced basis matrix A A A with conorms ρ \rho ρ , define
P α ( 1 ) = 384 α [ 4 ( ∑ e ∈ E ρ e ) 3 + 27 det Q ] × ∑ T ∈ T W T ∫ [ − 1 , 1 ] 2 q T ( det Q ) ( 4 5 q 0 , T ) α − 1 d x d y − 5 det Q ( ∑ e ∈ E ρ e ) [ 2 ( ∑ e ∈ E ρ e ) 3 + 54 det Q ] × [ 16 ( α − 1 ) ∑ T ∈ T W T ∫ [ − 1 , 1 ] 2 ( 4 5 q 0 , T ) α d x d y + ( det Q ) ( 4 / 5 ) α ∫ [ − 1 , 1 ] 2 ( 12 q 0 , s q + 36 q 0 , r h ) d x d y ] . (32) \begin{aligned}
P_{\alpha}^{(1)}={}&384\alpha\left[4\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3+27\det Q\right]\\
&\times\sum_{T\in\mathcal{T}}W_T\int_{[-1,1]^2}q_T(\det Q)\left(\frac{4}{5}q_{0,T}\right)^{\alpha-1}\,dx\,dy\\
&-5\det Q\left(\sum_{e\in\mathcal{E}}\rho_e\right)\left[2\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3+54\det Q\right]\\
&\times\left[16(\alpha-1)\sum_{T\in\mathcal{T}}W_T\int_{[-1,1]^2}\left(\frac{4}{5}q_{0,T}\right)^\alpha\,dx\,dy\right.\\
&\qquad\left.+(\det Q)(4/5)^\alpha\int_{[-1,1]^2}(12q_{0,\mathrm{sq}}+36q_{0,\mathrm{rh}})\,dx\,dy\right]. \tag*{(32)}
\end{aligned} P α ( 1 ) = 384 α 4 ( e ∈ E ∑ ρ e ) 3 + 27 det Q × T ∈ T ∑ W T ∫ [ − 1 , 1 ] 2 q T ( det Q ) ( 5 4 q 0 , T ) α − 1 d x d y − 5 det Q ( e ∈ E ∑ ρ e ) 2 ( e ∈ E ∑ ρ e ) 3 + 54 det Q × [ 16 ( α − 1 ) T ∈ T ∑ W T ∫ [ − 1 , 1 ] 2 ( 5 4 q 0 , T ) α d x d y + ( det Q ) ( 4/5 ) α ∫ [ − 1 , 1 ] 2 ( 12 q 0 , sq + 36 q 0 , rh ) d x d y ] . (32) If P α ( 1 ) ≥ 0 P_{\alpha}^{(1)}\ge0 P α ( 1 ) ≥ 0 , then E p ( A ) ≥ E p ( A b c c ) E_p(A)\ge E_p(A_{\mathrm{bcc}}) E p ( A ) ≥ E p ( A bcc ) . The inequality is strict when the conorms are unequal.
Proof. Apply s α ≥ α s + 1 − α s^\alpha\ge\alpha s+1-\alpha s α ≥ α s + 1 − α with s = ( 16 / det Q ) 1 / 3 q T / q 0 , T s=(16/\det Q)^{1/3}q_T/q_{0,T} s = ( 16/ det Q ) 1/3 q T / q 0 , T and multiply by q 0 , T α q_{0,T}^\alpha q 0 , T α . This gives
( 16 det Q ) α / 3 q T α ≥ α ( 16 det Q ) 1 / 3 q T q 0 , T α − 1 − ( α − 1 ) q 0 , T α . \left(\frac{16}{\det Q}\right)^{\alpha/3}q_T^\alpha\ge\alpha\left(\frac{16}{\det Q}\right)^{1/3}q_Tq_{0,T}^{\alpha-1}-(\alpha-1)q_{0,T}^\alpha. ( det Q 16 ) α /3 q T α ≥ α ( det Q 16 ) 1/3 q T q 0 , T α − 1 − ( α − 1 ) q 0 , T α .
For t = 27 det Q / ( 2 ( ∑ e ∈ E ρ e ) 3 ) ∈ ( 0 , 1 ] t=27\det Q/(2(\sum_{e\in\mathcal{E}}\rho_e)^3)\in(0,1] t = 27 det Q / ( 2 ( ∑ e ∈ E ρ e ) 3 ) ∈ ( 0 , 1 ] , (20) with β = 1 / 3 \beta=1/3 β = 1/3 gives
( 16 det Q ) 1 / 3 = 6 t − 1 / 3 ∑ e ∈ E ρ e ≥ 6 ∑ e ∈ E ρ e 4 ( ∑ e ∈ E ρ e ) 3 + 27 det Q 2 ( ∑ e ∈ E ρ e ) 3 + 54 det Q . \left(\frac{16}{\det Q}\right)^{1/3}=\frac{6t^{-1/3}}{\sum_{e\in\mathcal{E}}\rho_e}\geq\frac{6}{\sum_{e\in\mathcal{E}}\rho_e}\frac{4(\sum_{e\in\mathcal{E}}\rho_e)^3+27\det Q}{2(\sum_{e\in\mathcal{E}}\rho_e)^3+54\det Q}. ( det Q 16 ) 1/3 = ∑ e ∈ E ρ e 6 t − 1/3 ≥ ∑ e ∈ E ρ e 6 2 ( ∑ e ∈ E ρ e ) 3 + 54 det Q 4 ( ∑ e ∈ E ρ e ) 3 + 27 det Q .
Integrate the tangent bound against W T W_T W T and sum over the tiles. Using (15) and then substituting the Padé bound gives
( E p ( A ) E p ( A b c c ) − 1 ) ( det Q ) ∫ [ − 1 , 1 ] 2 ( 3 q 0 , s q α + 9 q 0 , r h α ) d x d y ≥ 4 α det Q ( 16 det Q ) 1 / 3 I 1 − 16 ( α − 1 ) I 0 − ( det Q ) ∫ [ − 1 , 1 ] 2 ( 3 q 0 , s q α + 9 q 0 , r h α ) d x d y ≥ P α ( 1 ) 20 ( 4 / 5 ) α ( det Q ) ( ∑ e ∈ E ρ e ) [ 2 ( ∑ e ∈ E ρ e ) 3 + 54 det Q ] . \begin{aligned}
\left(\frac{E_p(A)}{E_p(A_{\mathrm{bcc}})}-1\right)(\det Q)\int_{[-1,1]^2}(3q_{0,\mathrm{sq}}^\alpha+9q_{0,\mathrm{rh}}^\alpha)\,dx\,dy \\
&\geq\frac{4\alpha}{\det Q}\left(\frac{16}{\det Q}\right)^{1/3}I_1-16(\alpha-1)I_0-(\det Q)\int_{[-1,1]^2}(3q_{0,\mathrm{sq}}^\alpha+9q_{0,\mathrm{rh}}^\alpha)\,dx\,dy \\
&\geq\frac{P_\alpha^{(1)}}{20(4/5)^\alpha(\det Q)(\sum_{e\in\mathcal{E}}\rho_e)\left[2(\sum_{e\in\mathcal{E}}\rho_e)^3+54\det Q\right]}.
\end{aligned} ( E p ( A bcc ) E p ( A ) − 1 ) ( det Q ) ∫ [ − 1 , 1 ] 2 ( 3 q 0 , sq α + 9 q 0 , rh α ) d x d y ≥ det Q 4 α ( det Q 16 ) 1/3 I 1 − 16 ( α − 1 ) I 0 − ( det Q ) ∫ [ − 1 , 1 ] 2 ( 3 q 0 , sq α + 9 q 0 , rh α ) d x d y ≥ 20 ( 4/5 ) α ( det Q ) ( ∑ e ∈ E ρ e ) [ 2 ( ∑ e ∈ E ρ e ) 3 + 54 det Q ] P α ( 1 ) .
The numerator in the last line is exactly (32) . For unequal conorms the bound from (20) is strict, and I 1 > 0 I_1>0 I 1 > 0 since ∑ T ∈ T W T = 3 det Q > 0 \sum_{T\in\mathcal{T}}W_T=3\det Q>0 ∑ T ∈ T W T = 3 det Q > 0 and q T ≥ η T Q η / 4 > 0 q_T\geq\eta^TQ\eta/4>0 q T ≥ η T Q η /4 > 0 by (8) . □
Write the six reference integrals as
F α = ( 4 / 5 ) α − 1 4 ( ∫ [ − 1 , 1 ] 2 ( 1 , x 2 ) q 0 , s q α − 1 d x d y , ∫ [ − 1 , 1 ] 2 ( 1 , x , x 2 , x y ) q 0 , r h α − 1 d x d y ) . F_\alpha=\frac{(4/5)^{\alpha-1}}{4}\left(\int_{[-1,1]^2}(1,x^2)q_{0,\mathrm{sq}}^{\alpha-1}\,dx\,dy,\ \int_{[-1,1]^2}(1,x,x^2,xy)q_{0,\mathrm{rh}}^{\alpha-1}\,dx\,dy\right). F α = 4 ( 4/5 ) α − 1 ( ∫ [ − 1 , 1 ] 2 ( 1 , x 2 ) q 0 , sq α − 1 d x d y , ∫ [ − 1 , 1 ] 2 ( 1 , x , x 2 , x y ) q 0 , rh α − 1 d x d y ) .
For R ( x , y ) = ∑ i + j ≤ 2 r i j x i y j R(x,y)=\sum_{i+j\leq2}r_{ij}x^iy^j R ( x , y ) = ∑ i + j ≤ 2 r ij x i y j , define the vector Φ T ( R ) \Phi_T(R) Φ T ( R ) by
Φ T ( R ) = { ( r 00 , r 20 + r 02 , 0 , 0 , 0 , 0 ) , q 0 , T = q 0 , s q , ( 0 , 0 , r 00 , r 10 + r 01 , r 20 + r 02 , r 11 ) , q 0 , T = q 0 , r h ( x , y ) , ( 0 , 0 , r 00 , r 10 − r 01 , r 20 + r 02 , − r 11 ) , q 0 , T = q 0 , r h ( x , − y ) . (33) \Phi_T(R)=
\begin{cases}
(r_{00},r_{20}+r_{02},0,0,0,0),&q_{0,T}=q_{0,\mathrm{sq}},\\
(0,0,r_{00},r_{10}+r_{01},r_{20}+r_{02},r_{11}),&q_{0,T}=q_{0,\mathrm{rh}}(x,y),\\
(0,0,r_{00},r_{10}-r_{01},r_{20}+r_{02},-r_{11}),&q_{0,T}=q_{0,\mathrm{rh}}(x,-y).
\end{cases} \tag*{(33)} Φ T ( R ) = ⎩ ⎨ ⎧ ( r 00 , r 20 + r 02 , 0 , 0 , 0 , 0 ) , ( 0 , 0 , r 00 , r 10 + r 01 , r 20 + r 02 , r 11 ) , ( 0 , 0 , r 00 , r 10 − r 01 , r 20 + r 02 , − r 11 ) , q 0 , T = q 0 , sq , q 0 , T = q 0 , rh ( x , y ) , q 0 , T = q 0 , rh ( x , − y ) . (33)
The symmetry in x , y x,y x , y , evenness on the square, and the change y ↦ − y y\mapsto-y y ↦ − y on the reflected rhombus give
Φ T ( R ) ⋅ F α = 1 4 ∫ [ − 1 , 1 ] 2 R ( 4 q 0 , T / 5 ) α − 1 d x d y . \Phi_T(R)\cdot F_\alpha=\frac{1}{4}\int_{[-1,1]^2}R(4q_{0,T}/5)^{\alpha-1}\,dx\,dy. Φ T ( R ) ⋅ F α = 4 1 ∫ [ − 1 , 1 ] 2 R ( 4 q 0 , T /5 ) α − 1 d x d y .
A vertex permutation π \pi π acts on the dual coordinates by U π e i = e π ( i ) − e π ( 0 ) U_\pi e_i=e_{\pi(i)}-e_{\pi(0)} U π e i = e π ( i ) − e π ( 0 ) for i = 1 , 2 , 3 i=1,2,3 i = 1 , 2 , 3 . With primes denoting the relabelled data,
U π g i j = e π ( i ) − e π ( j ) , Q ′ = U π Q U π T , det Q ′ = det Q , ∑ e ∈ E ρ e ′ = ∑ e ∈ E ρ e . U_\pi g_{ij}=e_{\pi(i)}-e_{\pi(j)},\qquad Q'=U_\pi Q U_\pi^T,\qquad\det Q'=\det Q,\qquad\sum_{e\in\mathcal{E}}\rho'_e=\sum_{e\in\mathcal{E}}\rho_e. U π g ij = e π ( i ) − e π ( j ) , Q ′ = U π Q U π T , det Q ′ = det Q , e ∈ E ∑ ρ e ′ = e ∈ E ∑ ρ e .
Here U π − 1 U_{\pi^{-1}} U π − 1 is the map associated to π − 1 \pi^{-1} π − 1 , so det U π = ± 1 \det U_\pi=\pm1 det U π = ± 1 . The relabelled facet representative has normal η ′ = ± U π − T η \eta'=\pm U_\pi^{-T}\eta η ′ = ± U π − T η , hence η ′ T Q ′ η ′ = η T Q η \eta'^TQ'\eta'=\eta^TQ\eta η ′ T Q ′ η ′ = η T Q η . Match the free-edge pair, and on each hexagon also the remaining edge that determines the tile’s shift in (6) . For each matched pair of tiles there is a signed permutation ( x , y ) ↦ ( x ′ , y ′ ) (x,y)\mapsto(x',y') ( x , y ) ↦ ( x ′ , y ′ ) for which
z T ′ ′ ( x ′ , y ′ ) − z η ′ ′ = ± U π ( z T ( x , y ) − z η ) . z'_{T'}(x',y')-z'_{\eta'}=\pm U_\pi(z_T(x,y)-z_\eta). z T ′ ′ ( x ′ , y ′ ) − z η ′ ′ = ± U π ( z T ( x , y ) − z η ) .
On a hexagon choose the overall sign to match the fixed-edge shift, and the two variable signs to match the free directions. By (7) and (8) ,
W T ′ ′ = W T , q T ′ ′ ( x ′ , y ′ ) = q T ( x , y ) , q 0 , T ′ ′ ( x ′ , y ′ ) = q 0 , T ( x , y ) . W'_{T'}=W_T,\qquad q'_{T'}(x',y')=q_T(x,y),\qquad q'_{0,T'}(x',y')=q_{0,T}(x,y). W T ′ ′ = W T , q T ′ ′ ( x ′ , y ′ ) = q T ( x , y ) , q 0 , T ′ ′ ( x ′ , y ′ ) = q 0 , T ( x , y ) .
The last identity is the preceding one at ρ = 1 \rho=1 ρ = 1 . The changes of variables preserve [ − 1 , 1 ] 2 [-1,1]^2 [ − 1 , 1 ] 2 and its area, so
I k ( ρ ′ ) = I k ( ρ ) ( k ≥ 0 ) . (34) I_k(\rho')=I_k(\rho)\qquad(k\geq0). \tag*{(34)} I k ( ρ ′ ) = I k ( ρ ) ( k ≥ 0 ) . (34)
Consequently the sum of P α ( 1 ) P_\alpha^{(1)} P α ( 1 ) over the twenty-four relabellings is P α s y m = 24 P α ( 1 ) P_\alpha^{\mathrm{sym}}=24P_\alpha^{(1)} P α sym = 24 P α ( 1 ) .
For the weights 16 q 0 , T 16q_{0,T} 16 q 0 , T in (33) , the three square tiles give ( 16 , 4 , 0 , 0 , 0 , 0 ) (16,4,0,0,0,0) ( 16 , 4 , 0 , 0 , 0 , 0 ) and the twelve rhombus tiles give ( 0 , 0 , 14 , − 4 , 4 , − 2 ) (0,0,14,-4,4,-2) ( 0 , 0 , 14 , − 4 , 4 , − 2 ) . Thus
U ( ρ ) = ∑ T ∈ T W T Φ T ( 16 q 0 , T ) = ( W 01 , 23 + W 02 , 13 + W 12 , 03 ) ( 16 , 4 , − 14 , − 4 , 4 , 2 ) + 3 ( det Q ) ( 0 , 0 , 14 , − 4 , 4 , − 2 ) , \begin{aligned}
U(\rho)=\sum_{T\in\mathcal{T}}W_T\Phi_T(16q_{0,T}) \\
&=(W_{01,23}+W_{02,13}+W_{12,03})(16,4,-14,-4,4,2) \\
&+3(\det Q)(0,0,14,-4,4,-2),
\end{aligned} U ( ρ ) = T ∈ T ∑ W T Φ T ( 16 q 0 , T ) = ( W 01 , 23 + W 02 , 13 + W 12 , 03 ) ( 16 , 4 , − 14 , − 4 , 4 , 2 ) + 3 ( det Q ) ( 0 , 0 , 14 , − 4 , 4 , − 2 ) ,
U ( 1 ) = ( 192 , 48 , 504 , − 144 , 144 , − 72 ) . U(1)=(192,48,504,-144,144,-72). U ( 1 ) = ( 192 , 48 , 504 , − 144 , 144 , − 72 ) .
The first tile sum in (32) is ∑ T ∈ T W T ( Φ T ( 4 q T det Q ) ⋅ F α ) \sum_{T\in\mathcal{T}}W_T\left(\Phi_T(4q_T\det Q)\cdot F_\alpha\right) ∑ T ∈ T W T ( Φ T ( 4 q T det Q ) ⋅ F α ) . Also
U ( ρ ) ⋅ F α = 5 ∑ T ∈ T W T ∫ [ − 1 , 1 ] 2 ( 4 q 0 , T / 5 ) α d x d y . U(\rho)\cdot F_\alpha=5\sum_{T\in\mathcal{T}}W_T\int_{[-1,1]^2}(4q_{0,T}/5)^\alpha\,dx\,dy. U ( ρ ) ⋅ F α = 5 T ∈ T ∑ W T ∫ [ − 1 , 1 ] 2 ( 4 q 0 , T /5 ) α d x d y .
The last bracket in (32) is therefore [ 16 ( α − 1 ) U ( ρ ) + ( det Q ) U ( 1 ) ] ⋅ F α / 5 [16(\alpha-1)U(\rho)+(\det Q)U(1)]\cdot F_\alpha/5 [ 16 ( α − 1 ) U ( ρ ) + ( det Q ) U ( 1 )] ⋅ F α /5 , so
P α s y m ( ρ ) = 9216 α [ 4 ( ∑ e ∈ E ρ e ) 3 + 27 det Q ] × ∑ T ∈ T W T ( Φ T ( 4 q T det Q ) ⋅ F α ) − 24 ( det Q ) ( ∑ e ∈ E ρ e ) [ 2 ( ∑ e ∈ E ρ e ) 3 + 54 det Q ] × ( [ 16 ( α − 1 ) U ( ρ ) + ( det Q ) U ( 1 ) ] ⋅ F α ) . (35) \begin{aligned}
P_\alpha^{\mathrm{sym}}(\rho)
={}&9216\alpha\left[4\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3+27\det Q\right]\\
&\times\sum_{T\in\mathcal{T}}W_T\left(\Phi_T(4q_T\det Q)\cdot F_\alpha\right)\\
&-24(\det Q)\left(\sum_{e\in\mathcal{E}}\rho_e\right)
\left[2\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3+54\det Q\right]\\
&\times\left([16(\alpha-1)U(\rho)+(\det Q)U(1)]\cdot F_\alpha\right). \tag*{(35)}
\end{aligned} P α sym ( ρ ) = 9216 α 4 ( e ∈ E ∑ ρ e ) 3 + 27 det Q × T ∈ T ∑ W T ( Φ T ( 4 q T det Q ) ⋅ F α ) − 24 ( det Q ) ( e ∈ E ∑ ρ e ) 2 ( e ∈ E ∑ ρ e ) 3 + 54 det Q × ( [ 16 ( α − 1 ) U ( ρ ) + ( det Q ) U ( 1 )] ⋅ F α ) . (35)
Numerical Lemma 5.5 gives the vectors C σ , ν C_{\sigma,\nu} C σ , ν and B σ , ν B_{\sigma,\nu} B σ , ν by the direct sum (41) . The coefficient of X 0 ν 0 ⋯ X 5 ν 5 X_0^{\nu_0}\cdots X_5^{\nu_5} X 0 ν 0 ⋯ X 5 ν 5 in (35) after ρ σ ( i ) = X i + ⋯ + X 5 \rho_{\sigma(i)}=X_i+\cdots+X_5 ρ σ ( i ) = X i + ⋯ + X 5 is ( C σ , ν + α B σ , ν ) ⋅ F α (C_{\sigma,\nu}+\alpha B_{\sigma,\nu})\cdot F_\alpha ( C σ , ν + α B σ , ν ) ⋅ F α . The set C C C consists of the distinct pairs obtained by dividing each nonzero pair in (41) by the positive gcd of its twelve entries.
At ρ = 1 \rho=1 ρ = 1 ,
I 0 ( 1 ) = ∫ [ − 1 , 1 ] 2 ( 3 q 0 , s q α + 9 q 0 , r h α ) d x d y , I 1 ( 1 ) = 64 I 0 ( 1 ) , P α s y m ( 1 ) = 0. I_0(1)=\int_{[-1,1]^2}(3q_{0,\mathrm{sq}}^\alpha+9q_{0,\mathrm{rh}}^\alpha)\,dx\,dy,\qquad I_1(1)=64I_0(1),\qquad P_\alpha^{\mathrm{sym}}(1)=0. I 0 ( 1 ) = ∫ [ − 1 , 1 ] 2 ( 3 q 0 , sq α + 9 q 0 , rh α ) d x d y , I 1 ( 1 ) = 64 I 0 ( 1 ) , P α sym ( 1 ) = 0.
Vertex invariance and homogeneity give the first-order vanishing in each sector, since for every e ∈ E e\in\mathcal{E} e ∈ E ,
∂ ρ e P α s y m ( 1 ) = 1 6 ∑ f ∈ E ∂ ρ f P α s y m ( 1 ) = 10 6 P α s y m ( 1 ) = 0. \partial_{\rho_e}P_\alpha^{\mathrm{sym}}(1)=\frac{1}{6}\sum_{f\in\mathcal{E}}\partial_{\rho_f}P_\alpha^{\mathrm{sym}}(1)=\frac{10}{6}P_\alpha^{\mathrm{sym}}(1)=0. ∂ ρ e P α sym ( 1 ) = 6 1 f ∈ E ∑ ∂ ρ f P α sym ( 1 ) = 6 10 P α sym ( 1 ) = 0.
Put v = ( 1 , 1 , 1 / 2 , − 1 / 4 , 1 / 2 , 0 ) v=(1,1,1/2,-1/4,1/2,0) v = ( 1 , 1 , 1/2 , − 1/4 , 1/2 , 0 ) .
Lemma 4.3 (Estimate of the reference moments). For α ≥ 513 \alpha\geq513 α ≥ 513 ,
∥ α ( α + 1 ) 25 F α − v ∥ ∞ < 1 10 . (36) \left\lVert\frac{\alpha(\alpha+1)}{25}F_\alpha-v\right\rVert_\infty<\frac{1}{10}. \tag*{(36)} 25 α ( α + 1 ) F α − v ∞ < 10 1 . (36)
Proof. Set s = 1 − 4 5 q 0 s=1-\frac{4}{5}q_0 s = 1 − 5 4 q 0 for either reference quadratic q 0 q_0 q 0 . By convexity and (10) ,
arg max [ − 1 , 1 ] 2 q 0 , s q = { ( − 1 , − 1 ) , ( − 1 , 1 ) , ( 1 , − 1 ) , ( 1 , 1 ) } , \underset{[-1,1]^2}{\operatorname{arg\,max}}\ q_{0,\mathrm{sq}}=\{(-1,-1),(-1,1),(1,-1),(1,1)\}, [ − 1 , 1 ] 2 arg max q 0 , sq = {( − 1 , − 1 ) , ( − 1 , 1 ) , ( 1 , − 1 ) , ( 1 , 1 )} ,
arg max [ − 1 , 1 ] 2 q 0 , r h = { ( − 1 , − 1 ) , ( − 1 , 1 ) , ( 1 , − 1 ) } , \underset{[-1,1]^2}{\operatorname{arg\,max}}\ q_{0,\mathrm{rh}}=\{(-1,-1),(-1,1),(1,-1)\}, [ − 1 , 1 ] 2 arg max q 0 , rh = {( − 1 , − 1 ) , ( − 1 , 1 ) , ( 1 , − 1 )} ,
max [ 0 , 1 ] 2 4 5 q 0 , r h = 7 10 , max ( x , y ) ∈ [ − 1 , 1 ] 2 , x y = 0 4 5 q 0 = 9 10 . \underset{[0,1]^2}{\max}\ \frac{4}{5}q_{0,\mathrm{rh}}=\frac{7}{10},\qquad
\underset{(x,y)\in[-1,1]^2,\ xy=0}{\max}\ \frac{4}{5}q_0=\frac{9}{10}. [ 0 , 1 ] 2 max 5 4 q 0 , rh = 10 7 , ( x , y ) ∈ [ − 1 , 1 ] 2 , x y = 0 max 5 4 q 0 = 10 9 .
Thus s ≤ 1 / 20 s\leq1/20 s ≤ 1/20 lies in the quadrants with a maximizing corner. In such a quadrant write the inward distances from its corner c c c as r w rw r w and r ( 1 − w ) r(1-w) r ( 1 − w ) , where 0 ≤ w ≤ 1 0\leq w\leq1 0 ≤ w ≤ 1 and $0\leq r\leq 1 / max ( w , 1 − w ) 1/\max(w,1-w) 1/ max ( w , 1 − w ) . Then
s = ℓ r − Q ∗ r 2 , ℓ = a w + b ( 1 − w ) , s=\ell r-Q_{\ast}r^2,\qquad\ell=aw+b(1-w), s = ℓ r − Q ∗ r 2 , ℓ = a w + b ( 1 − w ) ,
( a , b , 10 Q ∗ − w 2 − ( 1 − w ) 2 ) = { ( 1 / 5 , 1 / 5 , 0 ) , q 0 = q 0 , s q , ( 1 / 5 , 1 / 5 , − w ( 1 − w ) ) , q 0 = q 0 , r h , c = ( − 1 , − 1 ) , ( 2 / 5 , 1 / 5 , w ( 1 − w ) ) , q 0 = q 0 , r h , c = ( − 1 , 1 ) , ( 1 / 5 , 2 / 5 , w ( 1 − w ) ) , q 0 = q 0 , r h , c = ( 1 , − 1 ) . (a,b,10Q_{\ast}-w^2-(1-w)^2)=
\begin{cases}
(1/5,1/5,0), & q_0=q_{0,\mathrm{sq}},\\
(1/5,1/5,-w(1-w)), & q_0=q_{0,\mathrm{rh}},\quad c=(-1,-1),\\
(2/5,1/5,w(1-w)), & q_0=q_{0,\mathrm{rh}},\quad c=(-1,1),\\
(1/5,2/5,w(1-w)), & q_0=q_{0,\mathrm{rh}},\quad c=(1,-1).
\end{cases} ( a , b , 10 Q ∗ − w 2 − ( 1 − w ) 2 ) = ⎩ ⎨ ⎧ ( 1/5 , 1/5 , 0 ) , ( 1/5 , 1/5 , − w ( 1 − w )) , ( 2/5 , 1/5 , w ( 1 − w )) , ( 1/5 , 2/5 , w ( 1 − w )) , q 0 = q 0 , sq , q 0 = q 0 , rh , c = ( − 1 , − 1 ) , q 0 = q 0 , rh , c = ( − 1 , 1 ) , q 0 = q 0 , rh , c = ( 1 , − 1 ) .
Here ℓ ≥ 1 / 5 \ell\ge1/5 ℓ ≥ 1/5 , 0 < Q ∗ ≤ 1 / 10 0<Q_{\ast}\le1/10 0 < Q ∗ ≤ 1/10 and
s ( 0 ) = 0 , s ( 1 max ( w , 1 − w ) ) ≥ 1 10 , d s d r = ℓ − 2 Q ∗ r . s(0)=0,\qquad s\left(\frac{1}{\max(w,1-w)}\right)\ge\frac{1}{10},\qquad\frac{ds}{dr}=\ell-2Q_{\ast}r. s ( 0 ) = 0 , s ( max ( w , 1 − w ) 1 ) ≥ 10 1 , d r d s = ℓ − 2 Q ∗ r .
Thus 0 < s ≤ 1 / 20 0<s\le1/20 0 < s ≤ 1/20 occurs only on the increasing branch, where d x d y = r d r d w dx\,dy=r\,dr\,dw d x d y = r d r d w and
u = 1 − 4 Q ∗ s / ℓ 2 ≥ 1 / 2 , r = 2 s ℓ ( 1 + u ) ≤ 6 s , d r d s = 1 ℓ u , u=\sqrt{1-4Q_{\ast}s/\ell^2}\ge1/\sqrt{2},\qquad r=\frac{2s}{\ell(1+u)}\le6s,\qquad\frac{dr}{ds}=\frac{1}{\ell u}, u = 1 − 4 Q ∗ s / ℓ 2 ≥ 1/ 2 , r = ℓ ( 1 + u ) 2 s ≤ 6 s , d s d r = ℓ u 1 ,
0 ≤ ℓ 2 s d r d s − 1 = 4 Q ∗ s ℓ 2 1 + 2 / u ( 1 + u ) 2 ≤ 10 ( 8 2 − 10 ) ≤ 40 3 s . 0\le\frac{\ell^2}{s}\frac{dr}{ds}-1=\frac{4Q_{\ast}s}{\ell^2}\frac{1+2/u}{(1+u)^2}\le10(8\sqrt{2}-10)\le\frac{40}{3}s. 0 ≤ s ℓ 2 d s d r − 1 = ℓ 2 4 Q ∗ s ( 1 + u ) 2 1 + 2/ u ≤ 10 ( 8 2 − 10 ) ≤ 3 40 s .
For each monomial R R R in F α F_{\alpha} F α one has ∣ R ∣ ≤ 1 |R|\le1 ∣ R ∣ ≤ 1 and ∣ R ( x , y ) − R ( c ) ∣ ≤ 2 r |R(x,y)-R(c)|\le2r ∣ R ( x , y ) − R ( c ) ∣ ≤ 2 r . Hence
∣ ℓ 2 s d r d s R ( x , y ) − R ( c ) ∣ ≤ 40 3 s + 2 r ≤ 76 3 s . \left|\frac{\ell^2}{s}\frac{dr}{ds}R(x,y)-R(c)\right|\le\frac{40}{3}s+2r\le\frac{76}{3}s. s ℓ 2 d s d r R ( x , y ) − R ( c ) ≤ 3 40 s + 2 r ≤ 3 76 s .
The leading integral at each corner, extended to 0 ≤ s ≤ 1 0\le s\le1 0 ≤ s ≤ 1 , is
R ( c ) 4 ∫ 0 1 d w ( a w + b ( 1 − w ) ) 2 ∫ 0 1 s ( 1 − s ) α − 1 d s = R ( c ) 4 a b α ( α + 1 ) . \frac{R(c)}{4}\int_0^1\frac{dw}{(aw+b(1-w))^2}\int_0^1s(1-s)^{\alpha-1}\,ds=\frac{R(c)}{4ab\alpha(\alpha+1)}. 4 R ( c ) ∫ 0 1 ( a w + b ( 1 − w ) ) 2 d w ∫ 0 1 s ( 1 − s ) α − 1 d s = 4 ab α ( α + 1 ) R ( c ) .
The sums of R ( c ) / ( 4 a b ) R(c)/(4ab) R ( c ) / ( 4 ab ) give the square and rhombus entries of 25 v 25v 25 v , respectively,
25 4 ∑ c ∈ { ± 1 } 2 ( 1 , c 1 2 ) = ( 25 , 25 ) , 25 4 ( 1 , − 1 , 1 , 1 ) + 25 8 ( 2 , 0 , 2 , − 2 ) = 25 ( 1 2 , 1 4 , 1 2 , 0 ) . \frac{25}{4}\sum_{c\in\{\pm1\}^2}(1,c_1^2)=(25,25),\qquad
\frac{25}{4}(1,-1,1,1)+\frac{25}{8}(2,0,2,-2)=25\left(\frac12,\frac14,\frac12,0\right). 4 25 c ∈ { ± 1 } 2 ∑ ( 1 , c 1 2 ) = ( 25 , 25 ) , 4 25 ( 1 , − 1 , 1 , 1 ) + 8 25 ( 2 , 0 , 2 , − 2 ) = 25 ( 2 1 , 4 1 , 2 1 , 0 ) .
For each quadratic ∑ c 1 / ( 4 a b ) ≤ 25 \sum_c1/(4ab)\le25 ∑ c 1/ ( 4 ab ) ≤ 25 . The extended leading integral has tail
α ( α + 1 ) ∫ 1 / 20 1 s ( 1 − s ) α − 1 d s = ( 19 / 20 ) α ( 1 + α / 20 ) . \alpha(\alpha+1)\int_{1/20}^1s(1-s)^{\alpha-1}\,ds=(19/20)^\alpha(1+\alpha/20). α ( α + 1 ) ∫ 1/20 1 s ( 1 − s ) α − 1 d s = ( 19/20 ) α ( 1 + α /20 ) .
The error on s ≤ 1 / 20 s\le1/20 s ≤ 1/20 and the two tails therefore satisfy
∥ α ( α + 1 ) 25 F α − v ∥ ∞ ≤ 76 3 α ( α + 1 ) ∫ 0 1 s 2 ( 1 − s ) α − 1 d s + α ( α + 1 ) ∫ 1 / 20 1 s ( 1 − s ) α − 1 d s + α ( α + 1 ) 25 ( 19 / 20 ) α − 1 ≤ 152 3 ( α + 2 ) + ( α + 1 ) ( α + 25 ) 25 ( 19 / 20 ) α − 1 . \begin{aligned}
\left\|\frac{\alpha(\alpha+1)}{25}F_\alpha-v\right\|_\infty
\le\frac{76}{3}\alpha(\alpha+1)\int_0^1s^2(1-s)^{\alpha-1}\,ds
+\alpha(\alpha+1)\int_{1/20}^1s(1-s)^{\alpha-1}\,ds
+\frac{\alpha(\alpha+1)}{25}(19/20)^{\alpha-1} \\
&\le\frac{152}{3(\alpha+2)}+\frac{(\alpha+1)(\alpha+25)}{25}(19/20)^{\alpha-1}.
\end{aligned} 25 α ( α + 1 ) F α − v ∞ ≤ 3 76 α ( α + 1 ) ∫ 0 1 s 2 ( 1 − s ) α − 1 d s + α ( α + 1 ) ∫ 1/20 1 s ( 1 − s ) α − 1 d s + 25 α ( α + 1 ) ( 19/20 ) α − 1 ≤ 3 ( α + 2 ) 152 + 25 ( α + 1 ) ( α + 25 ) ( 19/20 ) α − 1 .
Both terms decrease for α ≥ 513 \alpha\ge513 α ≥ 513 , since
d d α log ( ( α + 1 ) ( α + 25 ) ( 19 / 20 ) α − 1 ) = 1 α + 1 + 1 α + 25 − log ( 20 / 19 ) < 2 513 − 1 20 < 0. \frac{d}{d\alpha}\log\left((\alpha+1)(\alpha+25)(19/20)^{\alpha-1}\right)
=\frac{1}{\alpha+1}+\frac{1}{\alpha+25}-\log(20/19)<\frac{2}{513}-\frac{1}{20}<0. d α d log ( ( α + 1 ) ( α + 25 ) ( 19/20 ) α − 1 ) = α + 1 1 + α + 25 1 − log ( 20/19 ) < 513 2 − 20 1 < 0.
As ( 19 / 20 ) 4 < 5 / 6 (19/20)^4<5/6 ( 19/20 ) 4 < 5/6 and ( 5 / 6 ) 8 < 1 / 4 (5/6)^8<1/4 ( 5/6 ) 8 < 1/4 , at α = 513 \alpha=513 α = 513 the bound is less than
152 1545 + 11062 2 − 32 < 1 10 . \frac{152}{1545}+11062\,2^{-32}<\frac{1}{10}. 1545 152 + 11062 2 − 32 < 10 1 .
Proof of Proposition 4.1. For α ≥ 513 \alpha\ge513 α ≥ 513 and ( C , B ) ∈ C (C,B) \in\mathcal{C} ( C , B ) ∈ C , Numerical Lemma 5.5 gives
4 B ⋅ v = 4 B 1 + 4 B 2 + 2 B 3 − B 4 + 2 B 5 ≥ 13 25 ∥ B ∥ 1 , ∥ C ∥ 1 ≤ ∥ B ∥ 1 . 4B \cdot v = 4B_1 + 4B_2 + 2B_3 - B_4 + 2B_5 \ge\frac{13}{25}\lVert B\rVert_1,\qquad\lVert C\rVert_1 \le\lVert B\rVert_1. 4 B ⋅ v = 4 B 1 + 4 B 2 + 2 B 3 − B 4 + 2 B 5 ≥ 25 13 ∥ B ∥ 1 , ∥ C ∥ 1 ≤ ∥ B ∥ 1 .
Together with Lemma 4.3, these inequalities give
α ( α + 1 ) 25 ( C + α B ) ⋅ F α ≥ ( C + α B ) ⋅ v − 1 10 ( ∥ C ∥ 1 + α ∥ B ∥ 1 ) ≥ ( 3 α 100 − 11 10 ) ∥ B ∥ 1 > 0. \begin{aligned}
\frac{\alpha(\alpha+1)}{25}(C+\alpha B)\cdot F_\alpha\ge(C+\alpha B)\cdot v-\frac{1}{10}(\lVert C\rVert_1+\alpha\lVert B\rVert_1) \\
&\ge\left(\frac{3\alpha}{100}-\frac{11}{10}\right)\lVert B\rVert_1>0.
\end{aligned} 25 α ( α + 1 ) ( C + α B ) ⋅ F α ≥ ( C + α B ) ⋅ v − 10 1 (∥ C ∥ 1 + α ∥ B ∥ 1 ) ≥ ( 100 3 α − 10 11 ) ∥ B ∥ 1 > 0.
Numerical Lemma 5.6 gives the same positivity for 35 / 2 ≤ α ≤ 514 35/2\le\alpha\le514 35/2 ≤ α ≤ 514 . Thus every coefficient in P α s y m P_\alpha^{\mathrm{sym}} P α sym after (18) is nonnegative for α ∈ [ 35 / 2 , 514 ] ∪ [ 513 , ∞ ) \alpha\in[35/2,514]\cup[513,\infty) α ∈ [ 35/2 , 514 ] ∪ [ 513 , ∞ ) . Since all sector variables are nonnegative, P α ( 1 ) = P α s y m / 24 ≥ 0 P_\alpha^{(1)}=P_\alpha^{\mathrm{sym}}/24\ge0 P α ( 1 ) = P α sym /24 ≥ 0 . Lemma 4.2 gives the desired moment inequality with strictness for unequal conorms. The equality case is (14) . □
Completion of the proof of Theorem 1.1. Propositions 3.1 and 4.1 give (2) for 1 ≤ p ≤ 35 1\le p\le35 1 ≤ p ≤ 35 and p ≥ 35 p\ge35 p ≥ 35 , respectively. □
Appendix: Numerical Computations
The following numerical lemmas use finite rational calculations. For intervals [ l , u ] [l,u] [ l , u ] , [ l ′ , u ′ ] [l',u'] [ l ′ , u ′ ] and a rational number r r r , use
[ l , u ] + [ l ′ , u ′ ] = [ l + l ′ , u + u ′ ] , [l,u]+[l',u']=[l+l',u+u'], [ l , u ] + [ l ′ , u ′ ] = [ l + l ′ , u + u ′ ] ,
r [ l , u ] = [ min ( r l , r u ) , max ( r l , r u ) ] , r[l,u]=[\min(rl,ru),\max(rl,ru)], r [ l , u ] = [ min ( r l , r u ) , max ( r l , r u )] ,
R b [ l , u ] = 2 − b [ ⌊ 2 b l ⌋ , ⌈ 2 b u ⌉ ] . R_b[l,u]=2^{-b}[\lfloor2^b l\rfloor,\lceil2^b u\rceil]. R b [ l , u ] = 2 − b [⌊ 2 b l ⌋ , ⌈ 2 b u ⌉] .
For the product of an exact polynomial and an interval polynomial, the coefficient at multi-index ν \nu ν is
∑ 0 ≤ μ ≤ ν a μ [ l ν − μ , u ν − μ ] , \sum_{0\le\mu\le\nu}a_\mu[l_{\nu-\mu},u_{\nu-\mu}], 0 ≤ μ ≤ ν ∑ a μ [ l ν − μ , u ν − μ ] ,
where a μ a_\mu a μ are the exact coefficients. This rule includes the exponent variable when present. Missing coefficients are zero. In (19) one can apply X i ↦ X i + X i + 1 X_i\mapsto X_i+X_{i+1} X i ↦ X i + X i + 1 for i = 0 , 1 , 2 , 3 , 4 i=0,1,2,3,4 i = 0 , 1 , 2 , 3 , 4 in that order, starting from ρ σ ( i ) = X i \rho_{\sigma(i)}=X_i ρ σ ( i ) = X i .
The two Bernstein calculations below use the identity
t i = ∑ j = i m ( j i ) ( m i ) ( m j ) t j ( 1 − t ) m − j , 0 ≤ i ≤ m . t^i=\sum_{j=i}^{m}\frac{\binom{j}{i}}{\binom{m}{i}}\binom{m}{j}t^j(1-t)^{m-j},\qquad0\le i\le m. t i = j = i ∑ m ( i m ) ( i j ) ( j m ) t j ( 1 − t ) m − j , 0 ≤ i ≤ m .
It follows by expanding t i ( t + ( 1 − t ) ) m − i t^i(t+(1-t))^{m-i} t i ( t + ( 1 − t ) ) m − i . On [ 0 , 1 ] [0,1] [ 0 , 1 ] the degree-m m m Bernstein basis is nonnegative and sums to one.
Bounded exponents
For a comparison polynomial of degree d d d , count the coefficients of X 0 ν 0 ⋯ X 5 ν 5 X_0^{\nu_0}\cdots X_5^{\nu_5} X 0 ν 0 ⋯ X 5 ν 5 separately in each of the thirty orders ρ σ ( i ) = X i + ⋯ + X 5 \rho_{\sigma(i)}=X_i+\cdots+X_5 ρ σ ( i ) = X i + ⋯ + X 5 . The total is 30 ( d + 5 5 ) 30\binom{d+5}{5} 30 ( 5 d + 5 ) , of which 180 have ν 5 ≥ d − 1 \nu_5\ge d-1 ν 5 ≥ d − 1 and vanish by § 3 .
Numerical Lemma 5.1 (Quartic Taylor comparison). In P α T P_\alpha^{\mathrm{T}} P α T with K = 4 K=4 K = 4 and d = 22 d=22 d = 22 , exactly 2 301 600 2\,301\,600 2 301 600 coefficients exceed 5 / 2 5/2 5/2 for every α ∈ [ 1 / 2 , 1 ] \alpha\in[1/2,1] α ∈ [ 1/2 , 1 ] . The other 120 300 120\,300 120 300 coefficients vanish identically on this interval.
Numerical Lemma 5.2 (Cubic Taylor comparison). In P α T P_\alpha^{\mathrm{T}} P α T with K = 3 K=3 K = 3 and d = 18 d=18 d = 18 , exactly 960 780 960\,780 960 780 coefficients exceed 7 7 7 for every α ∈ [ 1 , 2 ] \alpha\in[1,2] α ∈ [ 1 , 2 ] . The other 48 690 48\,690 48 690 coefficients vanish identically on this interval.
Numerical Lemma 5.3 (Weighted quadratic comparison). For each I ∈ { [ 2 , 5 / 2 ] , [ 5 / 2 , 3 ] } I\in\{[2,5/2],[5/2,3]\} I ∈ {[ 2 , 5/2 ] , [ 5/2 , 3 ]} , exactly 332 730 332\,730 332 730 coefficients of P α ( 2 ) P_\alpha^{(2)} P α ( 2 ) of degree d = 14 d=14 d = 14 exceed 3 / 4 3/4 3/4 for every α ∈ I \alpha\in I α ∈ I . The other 16 110 16\,110 16 110 coefficients vanish identically on I I I .
Numerical Lemma 5.4 (Weighted cubic comparison). For each I ∈ { [ j , j + 1 ] : j = 3 , … , 15 } ∪ { [ 16 , 35 / 2 ] } I \in\{[j,j+1]:j=3,\ldots,15\}\cup\{[16,35/2]\} I ∈ {[ j , j + 1 ] : j = 3 , … , 15 } ∪ {[ 16 , 35/2 ]} , exactly 960 780 960\,780 960 780 coefficients of P α ( 3 ) P_{\alpha}^{(3)} P α ( 3 ) of degree d = 18 d=18 d = 18 exceed 260 260 260 for every α ∈ I \alpha\in I α ∈ I . The other 48 690 48\,690 48 690 coefficients vanish identically on I I I .
Proof of Numerical Lemmas 5.1–5.4. For I = [ α − , α + ] I=[\alpha_{-},\alpha_{+}] I = [ α − , α + ] , write
α = c + r w , c = α − + α + 2 , r = α + − α − 2 , ∣ w ∣ ≤ 1. \alpha=c+rw,\qquad c=\frac{\alpha_{-}+\alpha_{+}}{2},\qquad r=\frac{\alpha_{+}-\alpha_{-}}{2},\qquad|w|\le1. α = c + r w , c = 2 α − + α + , r = 2 α + − α − , ∣ w ∣ ≤ 1.
For q ∈ { q 0 , s q , q 0 , r h } q\in\{q_{0,\mathrm{sq}},q_{0,\mathrm{rh}}\} q ∈ { q 0 , sq , q 0 , rh } , the moment is
1 4 ∫ [ − 1 , 1 ] 2 x a y b q c + r w − k d x d y , k , a , b ∈ Z ≥ 0 , a + b ≤ 2 k . \frac{1}{4}\int_{[-1,1]^{2}}x^{a}y^{b}q^{c+rw-k}\,dx\,dy,\qquad k,a,b\in\mathbb{Z}_{\ge0},\qquad a+b\le2k. 4 1 ∫ [ − 1 , 1 ] 2 x a y b q c + r w − k d x d y , k , a , b ∈ Z ≥ 0 , a + b ≤ 2 k .
The coefficients in
( c − k + r w ℓ ) = ∑ j = 0 ℓ b ℓ , j w j \left(\frac{c-k+rw}{\ell}\right)=\sum_{j=0}^{\ell}b_{\ell,j}w^{j} ( ℓ c − k + r w ) = j = 0 ∑ ℓ b ℓ , j w j
satisfy
b 0 , 0 = 1 , b ℓ + 1 , j = ( c − k − ℓ ) b ℓ , j + r b ℓ , j − 1 ℓ + 1 , b ℓ , j = 0 ( j < 0 or j > ℓ ) . b_{0,0}=1,\qquad b_{\ell+1,j}=\frac{(c-k-\ell)b_{\ell,j}+rb_{\ell,j-1}}{\ell+1},\qquad b_{\ell,j}=0\quad(j<0\text{ or }j>\ell). b 0 , 0 = 1 , b ℓ + 1 , j = ℓ + 1 ( c − k − ℓ ) b ℓ , j + r b ℓ , j − 1 , b ℓ , j = 0 ( j < 0 or j > ℓ ) .
The coefficient of w j w^{j} w j in the moment truncated at degree J J J in q − 1 q-1 q − 1 is
1 4 ∑ ℓ = j J b ℓ , j ∫ [ − 1 , 1 ] 2 x a y b ( q − 1 ) ℓ d x d y . (37) \frac{1}{4}\sum_{\ell=j}^{J}b_{\ell,j}\int_{[-1,1]^{2}}x^{a}y^{b}(q-1)^{\ell}\,dx\,dy. \tag*{(37)} 4 1 ℓ = j ∑ J b ℓ , j ∫ [ − 1 , 1 ] 2 x a y b ( q − 1 ) ℓ d x d y . (37)
All integrals here are given by (13) .
For the Taylor comparisons, retain 0 ≤ j ≤ h 0\le j\le h 0 ≤ j ≤ h and use
( J , h ) = { ( 40 , 10 ) , K = 4 , ( 50 , 12 ) , K = 3. (J,h)=
\begin{cases}
(40,10),&K=4,\\
(50,12),&K=3.
\end{cases} ( J , h ) = { ( 40 , 10 ) , ( 50 , 12 ) , K = 4 , K = 3.
The coefficient ℓ 1 \ell^{1} ℓ 1 norm of the spatial term of order ℓ \ell ℓ is bounded by
4 − ℓ ∏ i = 0 ℓ − 1 ∣ c − k − i ∣ + r i + 1 ≤ 4 − ℓ ( ℓ + 3 3 ) , 0 ≤ k ≤ 4. 4^{-\ell}\prod_{i=0}^{\ell-1}\frac{|c-k-i|+r}{i+1}\le4^{-\ell}\binom{\ell+3}{3},\qquad0\le k\le4. 4 − ℓ i = 0 ∏ ℓ − 1 i + 1 ∣ c − k − i ∣ + r ≤ 4 − ℓ ( 3 ℓ + 3 ) , 0 ≤ k ≤ 4.
The empty product is one. The spatial tail in the retained powers of w w w satisfies
∑ ℓ = J + 1 ∞ 4 − ℓ ( ℓ + 3 3 ) ≤ 4 − ( J + 1 ) ( J + 4 3 ) 1 − ( J + 5 ) / ( 4 ( J + 2 ) ) . \sum_{\ell=J+1}^{\infty}4^{-\ell}\binom{\ell+3}{3}\le\frac{4^{-(J+1)}\binom{J+4}{3}}{1-(J+5)/(4(J+2))}. ℓ = J + 1 ∑ ∞ 4 − ℓ ( 3 ℓ + 3 ) ≤ 1 − ( J + 5 ) / ( 4 ( J + 2 )) 4 − ( J + 1 ) ( 3 J + 4 ) .
For the omitted powers of w w w use the exponential series for q c − k exp ( r w log q ) q^{c-k}\exp(rw\log q) q c − k exp ( r w log q ) . Since q c − k ≤ 4 q^{c-k}\le4 q c − k ≤ 4 and r ∣ log q ∣ < r / 3 r|\log q|<r/3 r ∣ log q ∣ < r /3 , the total error is bounded by
4 − ( J + 1 ) ( J + 4 3 ) 1 − ( J + 5 ) / ( 4 ( J + 2 ) ) + 4 ( r / 3 ) h + 1 ( h + 1 ) ! ( 1 − r / ( 3 ( h + 2 ) ) ) < 2 − 60 . \frac{4^{-(J+1)}\binom{J+4}{3}}{1-(J+5)/(4(J+2))}+\frac{4(r/3)^{h+1}}{(h+1)!(1-r/(3(h+2)))}<2^{-60}. 1 − ( J + 5 ) / ( 4 ( J + 2 )) 4 − ( J + 1 ) ( 3 J + 4 ) + ( h + 1 )! ( 1 − r / ( 3 ( h + 2 ))) 4 ( r /3 ) h + 1 < 2 − 60 .
For Numerical Lemma 5.4, retain 0 ≤ j ≤ 19 0\le j\le19 0 ≤ j ≤ 19 on each stated interval. Scale the moments by 2 − s 2^{-s} 2 − s , with
s = min { j ∈ Z ≥ 1 : 2 j ≥ ( 5 / 4 ) ⌈ c ⌉ } . s=\min\{j\in\mathbb{Z}_{\ge1}:2^{j}\ge(5/4)^{\lceil c\rceil}\}. s = min { j ∈ Z ≥ 1 : 2 j ≥ ( 5/4 ) ⌈ c ⌉ } .
Indeed, for 0 ≤ k ≤ 4 0\le k\le4 0 ≤ k ≤ 4 ,
q c − k ≤ { ( 5 / 4 ) ⌈ c ⌉ , c ≥ k , 4 / 3 , c < k , 4 / 3 < ( 5 / 4 ) ⌈ c ⌉ ( c ≥ 7 / 2 ) . q^{c-k}\le
\begin{cases}
(5/4)^{\lceil c\rceil},&c\ge k,\\
4/3,&c<k,
\end{cases}
\qquad4/3<(5/4)^{\lceil c\rceil}\quad(c\ge7/2). q c − k ≤ { ( 5/4 ) ⌈ c ⌉ , 4/3 , c ≥ k , c < k , 4/3 < ( 5/4 ) ⌈ c ⌉ ( c ≥ 7/2 ) .
From 2 − s q c − k ≤ 1 2^{-s}q^{c-k}\le1 2 − s q c − k ≤ 1 and ∣ log q ∣ < 3 / 10 |\log q|<3/10 ∣ log q ∣ < 3/10 , the normalized exponential tail is at most
( 3 r / 10 ) 20 20 ! ( 1 − ( 3 r / 10 ) / 21 ) < 2 − 61 . \frac{(3r/10)^{20}}{20!(1-(3r/10)/21)}<2^{-61}. 20 ! ( 1 − ( 3 r /10 ) /21 ) ( 3 r /10 ) 20 < 2 − 61 .
Take the least J ≥ ⌈ c ⌉ J\ge\lceil c\rceil J ≥ ⌈ c ⌉ for which
4 3 2 − s max 0 ≤ k ≤ 4 k ∈ Z ( 4 − ( J + 1 ) ∏ i = 0 J ∣ c − k − i ∣ + r i + 1 ) < 2 − 61 . \frac{4}{3}2^{-s}\max_{\substack{0\le k\le4\\ k\in\mathbb{Z}}}\left(4^{-(J+1)}\prod_{i=0}^{J}\frac{|c-k-i|+r}{i+1}\right)<2^{-61}. 3 4 2 − s 0 ≤ k ≤ 4 k ∈ Z max ( 4 − ( J + 1 ) i = 0 ∏ J i + 1 ∣ c − k − i ∣ + r ) < 2 − 61 .
Since c − r ≥ 3 c-r \ge3 c − r ≥ 3 , k ≤ 4 k \le4 k ≤ 4 , and ℓ ≥ ⌈ c ⌉ \ell\ge\lceil c\rceil ℓ ≥ ⌈ c ⌉ ,
∣ c − k − ℓ ∣ + r 4 ( ℓ + 1 ) = ℓ + k − ( c − r ) 4 ( ℓ + 1 ) ≤ 1 4 . \frac{|c-k-\ell|+r}{4(\ell+1)}=\frac{\ell+k-(c-r)}{4(\ell+1)}\le\frac{1}{4}. 4 ( ℓ + 1 ) ∣ c − k − ℓ ∣ + r = 4 ( ℓ + 1 ) ℓ + k − ( c − r ) ≤ 4 1 .
The factor 4 / 3 4/3 4/3 sums this geometric tail. The total error is less than 2 − 60 2^{-60} 2 − 60 .
For Numerical Lemma 5.3, use α = n + w / 2 \alpha=n+w/2 α = n + w /2 with 0 ≤ k ≤ 2 0\le k\le2 0 ≤ k ≤ 2 and
( n , w ) ∈ { { 2 } × [ 0 , 1 ] , I = [ 2 , 5 / 2 ] , { 3 } × [ − 1 , 0 ] , I = [ 5 / 2 , 3 ] . (n,w)\in
\begin{cases}
\{2\}\times[0,1], & I=[2,5/2],\\
\{3\}\times[-1,0], & I=[5/2,3].
\end{cases} ( n , w ) ∈ { { 2 } × [ 0 , 1 ] , { 3 } × [ − 1 , 0 ] , I = [ 2 , 5/2 ] , I = [ 5/2 , 3 ] .
Use b ℓ , j b_{\ell,j} b ℓ , j from the same recurrence with c − k = 0 c-k=0 c − k = 0 and r = 1 / 2 r=1/2 r = 1/2 . For the moments scaled by ( 4 / 5 ) n (4/5)^n ( 4/5 ) n , the retained coefficients are
( 4 / 5 ) n 4 ∑ ℓ = j 36 b ℓ , j ∑ i = 0 n − k ( n − k i ) ∫ [ − 1 , 1 ] 2 x a y b ( q − 1 ) ℓ + i d x d y , 0 ≤ j ≤ 12. \frac{(4/5)^n}{4}\sum_{\ell=j}^{36}b_{\ell,j}\sum_{i=0}^{n-k}\binom{n-k}{i}\int_{[-1,1]^2}x^a y^b(q-1)^{\ell+i}\,dx\,dy,\qquad0\le j\le12. 4 ( 4/5 ) n ℓ = j ∑ 36 b ℓ , j i = 0 ∑ n − k ( i n − k ) ∫ [ − 1 , 1 ] 2 x a y b ( q − 1 ) ℓ + i d x d y , 0 ≤ j ≤ 12.
The coefficient norm of ( w ℓ / 2 ) (w^\ell/2) ( w ℓ /2 ) is at most one. Using ( 4 / 5 ) n q n − k ≤ ( 4 / 3 ) k (4/5)^nq^{n-k}\le(4/3)^k ( 4/5 ) n q n − k ≤ ( 4/3 ) k and ∣ log q ∣ / 2 < 1 / 6 |\log q|/2<1/6 ∣ log q ∣/2 < 1/6 gives the following bound for the two tails,
16 9 ( 4 3 ⋅ 4 37 + 2 6 13 13 ! ) < 2 − 60 . \frac{16}{9}\left(\frac{4}{3\cdot4^{37}}+\frac{2}{6^{13}13!}\right)<2^{-60}. 9 16 ( 3 ⋅ 4 37 4 + 6 13 13 ! 2 ) < 2 − 60 .
On I = [ 5 / 2 , 3 ] I=[5/2,3] I = [ 5/2 , 3 ] , replace w w w by − w -w − w in the power expansion before testing the coefficients on [ 0 , 1 ] [0,1] [ 0 , 1 ] .
Parity and the substitution y ↦ − y y\mapsto-y y ↦ − y give
∫ [ − 1 , 1 ] 2 x a y b q 0 , s q α − k d x d y = 0 if a or b is odd \int_{[-1,1]^2}x^a y^bq_{0,\mathrm{sq}}^{\alpha-k}\,dx\,dy=0\quad\text{if }a\text{ or }b\text{ is odd} ∫ [ − 1 , 1 ] 2 x a y b q 0 , sq α − k d x d y = 0 if a or b is odd
∫ [ − 1 , 1 ] 2 x a y b q 0 , r h ( x , − y ) α − k d x d y = ( − 1 ) b ∫ [ − 1 , 1 ] 2 x a y b q 0 , r h ( x , y ) α − k d x d y . \int_{[-1,1]^2}x^a y^bq_{0,\mathrm{rh}}(x,-y)^{\alpha-k}\,dx\,dy=(-1)^b\int_{[-1,1]^2}x^a y^bq_{0,\mathrm{rh}}(x,y)^{\alpha-k}\,dx\,dy. ∫ [ − 1 , 1 ] 2 x a y b q 0 , rh ( x , − y ) α − k d x d y = ( − 1 ) b ∫ [ − 1 , 1 ] 2 x a y b q 0 , rh ( x , y ) α − k d x d y .
If the retained coefficients of the scaled moment are a j a_j a j , its interval coefficient at degree j j j is
2 − 60 [ ⌊ 2 60 a j ⌋ − 1 { j = 0 } , ⌈ 2 60 a j ⌉ + 1 { j = 0 } ] . 2^{-60}\left[\left\lfloor2^{60}a_j\right\rfloor-\mathbf{1}_{\{j=0\}},\ \left\lceil2^{60}a_j\right\rceil+\mathbf{1}_{\{j=0\}}\right]. 2 − 60 [ ⌊ 2 60 a j ⌋ − 1 { j = 0 } , ⌈ 2 60 a j ⌉ + 1 { j = 0 } ] .
Use [ 0 , 0 ] [0,0] [ 0 , 0 ] at every degree for an odd square moment. On a reflected rhombus apply ( − 1 ) b (-1)^b ( − 1 ) b to the geometric weight of x a y b x^a y^b x a y b .
For the Taylor comparisons compute the geometric coefficients d ν , a , b ( T , k ) d_{\nu,a,b}(T,k) d ν , a , b ( T , k ) in (3.6) by multiplication of W T ( 4 q T det Q ) k W_T(4q_T\det Q)^k W T ( 4 q T det Q ) k . Insert the moment intervals separately for every tile and sum the contributions to I k I_k I k . For the reference term below use the constant-weight moments at k = 0 k=0 k = 0 , one per tile, with factor 4 on the three square tiles and 3 on the twelve rhombus tiles. With the 1 / 4 1/4 1/4 in the moments, this gives the reference integral in (3.7). Evaluate the following product in the displayed order and apply R 32 R_{32} R 32 to its coefficients before (2.17).
9216 P α T = [ 24 ( 1 + α 3 ) ( ∑ e ∈ E ρ e ) 3 + 108 ( 3 − α ) det Q ] × ∑ k = 0 K [ 384 ⋅ 3 k c k ( α ) ( 2 det Q ∑ e ∈ E ρ e ) K − k ] I k 9216P_\alpha^T=
\left[24\left(1+\frac{\alpha}{3}\right)\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3+108(3-\alpha)\det Q\right]
\times\sum_{k=0}^{K}\left[384\cdot3^k c_k(\alpha)\left(2\det Q\sum_{e\in\mathcal{E}}\rho_e\right)^{K-k}\right]I_k 9216 P α T = 24 ( 1 + 3 α ) ( e ∈ E ∑ ρ e ) 3 + 108 ( 3 − α ) det Q × k = 0 ∑ K 384 ⋅ 3 k c k ( α ) ( 2 det Q e ∈ E ∑ ρ e ) K − k I k
− [ 24 ( 1 − α 3 ) ( ∑ e ∈ E ρ e ) 3 + 108 ( 3 + α ) det Q ] × [ 24 det Q ( 2 det Q ∑ e ∈ E ρ e ) K ∫ [ − 1 , 1 ] 2 ( 3 q 0 , s q α + 9 q 0 , r h α ) d x d y ] . -\left[24\left(1-\frac{\alpha}{3}\right)\left(\sum_{e\in\mathcal{E}}\rho_e\right)^3+108(3+\alpha)\det Q\right]
\times\left[24\det Q\left(2\det Q\sum_{e\in\mathcal{E}}\rho_e\right)^K\int_{[-1,1]^2}\left(3q_{0,\mathrm{sq}}^\alpha+9q_{0,\mathrm{rh}}^\alpha\right)\,dx\,dy\right]. − 24 ( 1 − 3 α ) ( e ∈ E ∑ ρ e ) 3 + 108 ( 3 + α ) det Q × 24 det Q ( 2 det Q e ∈ E ∑ ρ e ) K ∫ [ − 1 , 1 ] 2 ( 3 q 0 , sq α + 9 q 0 , rh α ) d x d y .
For the weighted comparisons, let ϵ T = − 1 \epsilon_T=-1 ϵ T = − 1 on the reflected rhombus tiles and ϵ T = 1 \epsilon_T=1 ϵ T = 1 on the other tiles. In (25) the exact weight of 1 4 ∫ [ − 1 , 1 ] 2 x a y b q α − k d x d y \frac{1}{4}\int_{[-1,1]^2}x^a y^b q^{\alpha-k}\,\mathrm{d}x\,\mathrm{d}y 4 1 ∫ [ − 1 , 1 ] 2 x a y b q α − k d x d y is
∑ T ∈ T q 0 , T ( x , ϵ T y ) = q ( x , y ) ϵ T b ∑ ν ∈ Z ≥ 0 6 ∑ e ∈ E ν e = 4 k + 3 d ν , a , b ( T , k ) ∏ e ∈ E ρ e ν e , q ∈ { q 0 , s q , q 0 , r h } . \sum_{\substack{T\in\mathcal{T}\\q_{0,T}(x,\epsilon_Ty)=q(x,y)}}\epsilon_T^b\sum_{\substack{\nu\in\mathbb{Z}_{\geq0}^6\\\sum_{e\in\mathcal{E}}\nu_e=4k+3}}d_{\nu,a,b}(T,k)\prod_{e\in\mathcal{E}}\rho_e^{\nu_e},\qquad q\in\{q_{0,\mathrm{sq}},q_{0,\mathrm{rh}}\}. T ∈ T q 0 , T ( x , ϵ T y ) = q ( x , y ) ∑ ϵ T b ν ∈ Z ≥ 0 6 ∑ e ∈ E ν e = 4 k + 3 ∑ d ν , a , b ( T , k ) e ∈ E ∏ ρ e ν e , q ∈ { q 0 , sq , q 0 , rh } .
The equality in the outer index is an identity in x , y x,y x , y . Omit the zero square moments. The reference integral uses the two ( k , a , b ) = ( 0 , 0 , 0 ) (k,a,b)=(0,0,0) ( k , a , b ) = ( 0 , 0 , 0 ) moments with weights 12 and 36. Multiply these weights by the exact conorm factors of (27) or (28) before inserting the moment intervals. The three exponent factors are
2 ( α − 1 ) , 2 ( α − 2 ) , 2 in the quadratic comparison, 2(\alpha-1),\qquad2(\alpha-2),\qquad2\quad\text{in the quadratic comparison,} 2 ( α − 1 ) , 2 ( α − 2 ) , 2 in the quadratic comparison,
4 ( α − 2 ) , 4 ( α − 3 ) , 4 in the cubic comparison. 4(\alpha-2),\qquad4(\alpha-3),\qquad4\quad\text{in the cubic comparison.} 4 ( α − 2 ) , 4 ( α − 3 ) , 4 in the cubic comparison.
For the quadratic comparison substitute (19) in the exact conorm weights and apply R 52 R_{52} R 52 to the moment coefficients before taking their products.
For the cubic comparison take the products and sum in the conorms. At each conorm coefficient write the resulting exponent coefficients as [ l j , u j ] [l_j,u_j] [ l j , u j ] . Retain degrees up to
h = { 8 , I = [ m , m + 1 ] , m = 3 , … , 15 , 11 , I = [ 16 , 35 / 2 ] . h=\begin{cases}
8,&I=[m,m+1],\quad m=3,\ldots,15,\\
11,&I=[16,35/2].
\end{cases} h = { 8 , 11 , I = [ m , m + 1 ] , m = 3 , … , 15 , I = [ 16 , 35/2 ] .
Add the interval [ − ∑ j > h max { ∣ l j ∣ , ∣ u j ∣ } , ∑ j > h max { ∣ l j ∣ , ∣ u j ∣ } ] [-\sum_{j>h}\max\{|l_j|,|u_j|\},\sum_{j>h}\max\{|l_j|,|u_j|\}] [ − ∑ j > h max { ∣ l j ∣ , ∣ u j ∣ } , ∑ j > h max { ∣ l j ∣ , ∣ u j ∣ }] to [ l 0 , u 0 ] [l_0,u_0] [ l 0 , u 0 ] . Apply R 32 R_{32} R 32 to the retained coefficients before (19) . The total positive factors, including the moment scales, are
{ 9216 , P α T , K = 3 or 4 , 2 ( 4 / 5 ) n , P α ( 2 ) , 4 ⋅ 2 − s , P α ( 3 ) . \left\{\begin{aligned}
9216,&\qquad P_\alpha^{\mathrm{T}},\quad K=3\text{ or }4,\\
2(4/5)^n,&\qquad P_\alpha^{(2)},\\
4\cdot2^{-s},&\qquad P_\alpha^{(3)}.
\end{aligned}\right. ⎩ ⎨ ⎧ 9216 , 2 ( 4/5 ) n , 4 ⋅ 2 − s , P α T , K = 3 or 4 , P α ( 2 ) , P α ( 3 ) .
Recall that ν \nu ν gives the powers of X 0 , … , X 5 X_0,\ldots,X_5 X 0 , … , X 5 after ρ σ ( i ) = X i + ⋯ + X 5 \rho_{\sigma(i)}=X_i+\cdots+X_5 ρ σ ( i ) = X i + ⋯ + X 5 . For ν 5 ≤ d − 2 \nu_5\leq d-2 ν 5 ≤ d − 2 , the scaled coefficient lies in ∑ j ≥ 0 w j [ l j , u j ] \sum_{j\geq0}w^j[l_j,u_j] ∑ j ≥ 0 w j [ l j , u j ] . The Taylor comparison has degree at most 14 for K = 4 K=4 K = 4 and at most 16 for K = 3 K=3 K = 3 . Thus for ∣ w ∣ ≤ 1 |w|\leq1 ∣ w ∣ ≤ 1 the scaled coefficient is bounded below by
∑ j ≥ 0 min { w j l j , w j u j } ≥ l 0 − ∑ j ≥ 1 max { ∣ l j ∣ , ∣ u j ∣ } . \sum_{j\geq0}\min\{w^jl_j,w^ju_j\}\geq l_0-\sum_{j\geq1}\max\{|l_j|,|u_j|\}. j ≥ 0 ∑ min { w j l j , w j u j } ≥ l 0 − j ≥ 1 ∑ max { ∣ l j ∣ , ∣ u j ∣ } .
In each Taylor or weighted cubic position the rational arithmetic gives either l j = u j = 0 l_j=u_j=0 l j = u j = 0 for every j j j or
l 0 − ∑ j ≥ 1 max { ∣ l j ∣ , ∣ u j ∣ } > { 23040 , P α T , K = 4 , 64512 , P α T , K = 3 , 260 , P α ( 3 ) . (38) l_0-\sum_{j\geq1}\max\{|l_j|,|u_j|\}>
\begin{cases}
23040,&P_\alpha^{\mathrm{T}},\quad K=4,\\
64512,&P_\alpha^{\mathrm{T}},\quad K=3,\\
260,&P_\alpha^{(3)}.
\end{cases} \tag*{(38)} l 0 − j ≥ 1 ∑ max { ∣ l j ∣ , ∣ u j ∣ } > ⎩ ⎨ ⎧ 23040 , 64512 , 260 , P α T , K = 4 , P α T , K = 3 , P α ( 3 ) . (38)
In the quadratic comparison, if all exact moment weights are zero, the coefficient vanishes. In every other position, on the parameter interval 0 ≤ w ≤ 1 0\leq w\leq1 0 ≤ w ≤ 1 , its lower bound is
max { l 0 + ∑ j = 1 13 min ( 0 , l j ) , min 0 ≤ i ≤ 13 ∑ j = 0 i ( i j ) ( 13 j ) 2 − 28 [ 2 28 l j ] } > 1. \max\left\{l_0+\sum_{j=1}^{13}\min(0,l_j),\ \min_{0\leq i\leq13}\sum_{j=0}^{i}\frac{\binom{i}{j}}{\binom{13}{j}}2^{-28}\left[2^{28}l_j\right]\right\}>1. max { l 0 + j = 1 ∑ 13 min ( 0 , l j ) , 0 ≤ i ≤ 13 min j = 0 ∑ i ( j 13 ) ( j i ) 2 − 28 [ 2 28 l j ] } > 1.
The first bound uses 0 ≤ w j ≤ 1 0\leq w^j\leq1 0 ≤ w j ≤ 1 . The second uses the degree-13 Bernstein coefficients of the polynomial with power coefficients 2 − 28 [ 2 28 l j ] 2^{-28}\left[2^{28}l_j\right] 2 − 28 [ 2 28 l j ] . The exact scale conversions are
23040 9216 = 5 2 , 64512 9216 = 7 , 1 2 ( 4 / 5 ) n ≥ 25 32 > 3 4 , 260 4 ⋅ 2 − s ≥ 260 , \frac{23040}{9216}=\frac{5}{2},\qquad
\frac{64512}{9216}=7,\qquad
\frac{1}{2(4/5)^n}\geq\frac{25}{32}>\frac{3}{4},\qquad
\frac{260}{4\cdot2^{-s}}\geq260, 9216 23040 = 2 5 , 9216 64512 = 7 , 2 ( 4/5 ) n 1 ≥ 32 25 > 4 3 , 4 ⋅ 2 − s 260 ≥ 260 ,
where s ≥ 2 s\geq2 s ≥ 2 by ( 5 / 4 ) ⌈ c ⌉ ≥ ( 5 / 4 ) 4 > 2 (5/4)^{\lceil c\rceil}\geq(5/4)^4>2 ( 5/4 ) ⌈ c ⌉ ≥ ( 5/4 ) 4 > 2 .
For each interval, enumerate the thirty orders from (2.16) and
0 ≤ ν i ≤ d − ∑ j = 0 i − 1 ν j ( 0 ≤ i < 5 ) , ν 5 = d − ∑ j = 0 4 ν j . 0 \le\nu_i \le d-\sum_{j=0}^{i-1}\nu_j \quad(0 \le i<5), \qquad\nu_5=d-\sum_{j=0}^{4}\nu_j. 0 ≤ ν i ≤ d − j = 0 ∑ i − 1 ν j ( 0 ≤ i < 5 ) , ν 5 = d − j = 0 ∑ 4 ν j .
Among the positions with ν 5 ≤ d − 2 \nu_5 \le d-2 ν 5 ≤ d − 2 , the zero tests give 120 120, 48 510, 15 930, and 48 510 zeros in the four comparisons, respectively, on each indicated interval. Adding the 180 positions with ν 5 ≥ d − 1 \nu_5 \ge d-1 ν 5 ≥ d − 1 gives the stated zero counts. The other positions pass the strict lower-bound tests. □
An explicit calculation of one coefficient. Consider P α ( 2 ) P_{\alpha}^{(2)} P α ( 2 ) in the sector ρ 01 ≥ ρ 02 ≥ ρ 03 ≥ ρ 12 ≥ ρ 13 ≥ ρ 23 \rho_{01}\ge\rho_{02}\ge\rho_{03}\ge\rho_{12}\ge\rho_{13}\ge\rho_{23} ρ 01 ≥ ρ 02 ≥ ρ 03 ≥ ρ 12 ≥ ρ 13 ≥ ρ 23 . Since the polynomial is homogeneous of degree 14, setting X 2 = 1 X_2=1 X 2 = 1 and X i = 0 X_i=0 X i = 0 for i ≠ 2 i\ne2 i = 2 makes X 2 14 = 1 X_2^{14}=1 X 2 14 = 1 and every other monomial zero. At this point, (2.16) gives
( ρ 01 , ρ 02 , ρ 03 , ρ 12 , ρ 13 , ρ 23 ) = ( 1 , 1 , 1 , 0 , 0 , 0 ) , Q = Id 3 . (\rho_{01},\rho_{02},\rho_{03},\rho_{12},\rho_{13},\rho_{23})=(1,1,1,0,0,0), \qquad Q=\operatorname{Id}_3. ( ρ 01 , ρ 02 , ρ 03 , ρ 12 , ρ 13 , ρ 23 ) = ( 1 , 1 , 1 , 0 , 0 , 0 ) , Q = Id 3 .
The three tiles with free-edge pairs ( 01 , 02 ) (01,02) ( 01 , 02 ) , ( 01 , 03 ) (01,03) ( 01 , 03 ) and ( 02 , 03 ) (02,03) ( 02 , 03 ) have W T = 1 W_T=1 W T = 1 and 4 q T = 1 + x 2 + y 2 4q_T=1+x^2+y^2 4 q T = 1 + x 2 + y 2 . All other weights vanish. The reflected hexagon parametrizations in (3.4) give
I k = 3 4 ∫ [ − 1 , 1 ] 2 ( 1 + x 2 + y 2 ) k q 0 , r h α − k d x d y . I_k=\frac{3}{4}\int_{[-1,1]^2}(1+x^2+y^2)^kq_{0,\mathrm{rh}}^{\alpha-k}\,dx\,dy. I k = 4 3 ∫ [ − 1 , 1 ] 2 ( 1 + x 2 + y 2 ) k q 0 , rh α − k d x d y .
Substitution in (3.8) gives the coefficient as
1 4 ∫ [ − 1 , 1 ] 2 [ 69984 ( α − 1 ) ( 1 + x 2 + y 2 ) 2 q 0 , r h α − 2 − 116640 ( α − 2 ) ( 1 + x 2 + y 2 ) 2 q 0 , r h α − 1 − 46656 q 0 , s q α − 139968 q 0 , r h α ] d x d y . (39) \frac{1}{4}\int_{[-1,1]^2}\left[69984(\alpha-1)(1+x^2+y^2)^2q_{0,\mathrm{rh}}^{\alpha-2}
-116640(\alpha-2)(1+x^2+y^2)^2q_{0,\mathrm{rh}}^{\alpha-1}
-46656q_{0,\mathrm{sq}}^\alpha-139968q_{0,\mathrm{rh}}^\alpha\right]\,dx\,dy. \tag*{(39)} 4 1 ∫ [ − 1 , 1 ] 2 [ 69984 ( α − 1 ) ( 1 + x 2 + y 2 ) 2 q 0 , rh α − 2 − 116640 ( α − 2 ) ( 1 + x 2 + y 2 ) 2 q 0 , rh α − 1 − 46656 q 0 , sq α − 139968 q 0 , rh α ] d x d y . (39)
For α = 2 + w / 2 \alpha=2+w/2 α = 2 + w /2 , 0 ≤ w ≤ 1 0\le w\le1 0 ≤ w ≤ 1 , use the degree-36 binomial and degree-12 power truncations for the moments scaled by ( 4 / 5 ) 2 (4/5)^2 ( 4/5 ) 2 , as in Numerical Lemma 5.3. Round the moment intervals outward by R 52 R_{52} R 52 and multiply by twice the corresponding factors in (5.3). This encloses 32 / 25 32/25 32/25 times the coefficient in (5.3). The lower endpoints l j l_j l j satisfy
2 52 l 0 = 122896820705070320064 , 2 52 ∑ j = 1 13 min ( 0 , l j ) = − 962510680643051232. 2^{52}l_0=122896820705070320064,\qquad2^{52}\sum_{j=1}^{13}\min(0,l_j)=-962510680643051232. 2 52 l 0 = 122896820705070320064 , 2 52 j = 1 ∑ 13 min ( 0 , l j ) = − 962510680643051232.
Thus on 2 ≤ α ≤ 5 / 2 2\le\alpha\le5/2 2 ≤ α ≤ 5/2 the exact coefficient is bounded below by
25 32 ( l 0 + ∑ j = 1 13 min ( 0 , l j ) ) = 25 ⋅ 121934310024427268832 2 57 > 21152. \frac{25}{32}\left(l_0+\sum_{j=1}^{13}\min(0,l_j)\right)=\frac{25\cdot121934310024427268832}{2^{57}}>21152. 32 25 ( l 0 + j = 1 ∑ 13 min ( 0 , l j ) ) = 2 57 25 ⋅ 121934310024427268832 > 21152.
At α = 2 \alpha=2 α = 2 , direct integration in (5.3) gives 106596 / 5 106596/5 106596/5 .
Large exponents
Numerical Lemma 5.5 (Degree-ten coefficient pairs). For a conorm tuple ρ \rho ρ , set
C ( ρ ) = 24 ( det Q ) ( ∑ ϵ ∈ E ρ ϵ ) [ 2 ( ∑ ϵ ∈ E ρ ϵ ) 3 + 54 det Q ] × [ 16 ( W 01 , 23 + W 02 , 13 + W 12 , 03 ) − 12 det Q ] ( 16 , 4 , − 14 , 4 , − 4 , 2 ) , B ( ρ ) = 9216 [ 4 ( ∑ ϵ ∈ E ρ ϵ ) 3 + 27 det Q ] ∑ T ∈ T W T Φ T ( 4 q T det Q ) − 384 ( det Q ) ( ∑ ϵ ∈ E ρ ϵ ) [ 2 ( ∑ ϵ ∈ E ρ ϵ ) 3 + 54 det Q ] × [ ( W 01 , 23 + W 02 , 13 + W 12 , 03 ) ( 16 , 4 , − 14 , 4 , − 4 , 2 ) + 3 ( det Q ) ( 0 , 0 , 14 , − 4 , 4 , − 2 ) ] . (40) \begin{aligned}
C(\rho)={}&24(\det Q)\left(\sum_{\epsilon\in\mathcal{E}}\rho_\epsilon\right)\left[2\left(\sum_{\epsilon\in\mathcal{E}}\rho_\epsilon\right)^3+54\det Q\right]\\
&\times[16(W_{01,23}+W_{02,13}+W_{12,03})-12\det Q](16,4,-14,4,-4,2),\\
B(\rho)={}&9216\left[4\left(\sum_{\epsilon\in\mathcal{E}}\rho_\epsilon\right)^3+27\det Q\right]\sum_{T\in\mathcal{T}}W_T\Phi_T(4q_T\det Q)\\
&-384(\det Q)\left(\sum_{\epsilon\in\mathcal{E}}\rho_\epsilon\right)\left[2\left(\sum_{\epsilon\in\mathcal{E}}\rho_\epsilon\right)^3+54\det Q\right]\\
&\times[(W_{01,23}+W_{02,13}+W_{12,03})(16,4,-14,4,-4,2)\\
&\qquad+3(\det Q)(0,0,14,-4,4,-2)]. \tag*{(40)}
\end{aligned} C ( ρ ) = B ( ρ ) = 24 ( det Q ) ( ϵ ∈ E ∑ ρ ϵ ) 2 ( ϵ ∈ E ∑ ρ ϵ ) 3 + 54 det Q × [ 16 ( W 01 , 23 + W 02 , 13 + W 12 , 03 ) − 12 det Q ] ( 16 , 4 , − 14 , 4 , − 4 , 2 ) , 9216 4 ( ϵ ∈ E ∑ ρ ϵ ) 3 + 27 det Q T ∈ T ∑ W T Φ T ( 4 q T det Q ) − 384 ( det Q ) ( ϵ ∈ E ∑ ρ ϵ ) 2 ( ϵ ∈ E ∑ ρ ϵ ) 3 + 54 det Q × [( W 01 , 23 + W 02 , 13 + W 12 , 03 ) ( 16 , 4 , − 14 , 4 , − 4 , 2 ) + 3 ( det Q ) ( 0 , 0 , 14 , − 4 , 4 , − 2 )] . (40)
For the tile with free edges u u u , v v v and normal η \eta η , W T = ρ u ρ v η T Q η W_T=\rho_u\rho_v\eta^TQ\eta W T = ρ u ρ v η T Q η and Φ T \Phi_T Φ T is given by (33) . At every evaluation use
4 q T det Q = ∑ e , f ∈ E e < f ρ e ρ f det ( g e , g f , 2 z T ) 2 . 4q_T\det Q=\sum_{\substack{e,f\in E\\e<f}}\rho_e\rho_f\det(g_e,g_f,2z_T)^2. 4 q T det Q = e , f ∈ E e < f ∑ ρ e ρ f det ( g e , g f , 2 z T ) 2 .
This polynomial expression also applies when det Q = 0 \det Q=0 det Q = 0 . For each of the thirty orders σ \sigma σ and ν ∈ Z ≥ 0 6 \nu\in\mathbb{Z}_{\geq0}^{6} ν ∈ Z ≥ 0 6 with ν 0 + ⋯ + ν 5 = 10 \nu_0+\cdots+\nu_5=10 ν 0 + ⋯ + ν 5 = 10 , set ρ σ ( i ) ( m ) = m i + ⋯ + m 5 \rho_{\sigma(i)}(m)=m_i+\cdots+m_5 ρ σ ( i ) ( m ) = m i + ⋯ + m 5 in each summand below. Then
( C σ , ν , B σ , ν ) = 1 ν 0 ! ⋯ ν 5 ! ∑ m 0 = 0 ν 0 ⋯ ∑ m 5 = 0 ν 5 ( − 1 ) 10 − m 0 − ⋯ − m 5 ( ∏ i = 0 5 ( ν i m i ) ) ( C ( ρ ( m ) ) , B ( ρ ( m ) ) ) . (41) (C_{\sigma,\nu},B_{\sigma,\nu})=\frac{1}{\nu_0!\cdots\nu_5!}\sum_{m_0=0}^{\nu_0}\cdots\sum_{m_5=0}^{\nu_5}(-1)^{10-m_0-\cdots-m_5}\left(\prod_{i=0}^{5}\binom{\nu_i}{m_i}\right)(C(\rho(m)),B(\rho(m))). \tag*{(41)} ( C σ , ν , B σ , ν ) = ν 0 ! ⋯ ν 5 ! 1 m 0 = 0 ∑ ν 0 ⋯ m 5 = 0 ∑ ν 5 ( − 1 ) 10 − m 0 − ⋯ − m 5 ( i = 0 ∏ 5 ( m i ν i ) ) ( C ( ρ ( m )) , B ( ρ ( m ))) . (41)
There are 86 172 nonzero pairs, 3 738 zero pairs with ν 5 ≤ 8 \nu_5\leq8 ν 5 ≤ 8 , and 180 zero pairs with ν 5 ≥ 9 \nu_5\geq9 ν 5 ≥ 9 . Divide each nonzero pair by the positive gcd of its twelve entries and retain each resulting pair ( C , B ) (C,B) ( C , B ) once. There are 30 812 such pairs, and each satisfies
0 < ∑ ℓ = 1 6 ∣ B ℓ ∣ , ∑ ℓ = 1 6 ∣ C ℓ ∣ ≤ ∑ ℓ = 1 6 ∣ B ℓ ∣ , 0<\sum_{\ell=1}^{6}|B_\ell|,\qquad\sum_{\ell=1}^{6}|C_\ell|\leq\sum_{\ell=1}^{6}|B_\ell|, 0 < ℓ = 1 ∑ 6 ∣ B ℓ ∣ , ℓ = 1 ∑ 6 ∣ C ℓ ∣ ≤ ℓ = 1 ∑ 6 ∣ B ℓ ∣ ,
4 B 1 + 4 B 2 + 2 B 3 − B 4 + 2 B 5 ≥ 13 25 ∑ ℓ = 1 6 ∣ B ℓ ∣ . 4B_1+4B_2+2B_3-B_4+2B_5\geq\frac{13}{25}\sum_{\ell=1}^{6}|B_\ell|. 4 B 1 + 4 B 2 + 2 B 3 − B 4 + 2 B 5 ≥ 25 13 ℓ = 1 ∑ 6 ∣ B ℓ ∣.
Numerical Lemma 5.6 (Finite large-exponent comparison). For every pair ( C , B ) (C,B) ( C , B ) obtained by dividing a nonzero pair in (41) by its positive gcd, and every 35 / 2 ≤ α ≤ 514 35/2\leq\alpha\leq514 35/2 ≤ α ≤ 514 ,
( 4 / 5 ) α − 1 4 { ∫ [ − 1 , 1 ] 2 [ C 1 + α B 1 + ( C 2 + α B 2 ) x 2 ] q 0 , s q α − 1 d x d y + ∫ [ − 1 , 1 ] 2 [ C 3 + α B 3 + ( C 4 + α B 4 ) x + ( C 5 + α B 5 ) x 2 + ( C 6 + α B 6 ) x y ] q 0 , r h α − 1 d x d y } > 47 100 . \frac{(4/5)^{\alpha-1}}{4}\left\{\int_{[-1,1]^2}\left[C_1+\alpha B_1+(C_2+\alpha B_2)x^2\right]q_{0,\mathrm{sq}}^{\alpha-1}\,dx\,dy+\int_{[-1,1]^2}\left[C_3+\alpha B_3+(C_4+\alpha B_4)x+(C_5+\alpha B_5)x^2+(C_6+\alpha B_6)xy\right]q_{0,\mathrm{rh}}^{\alpha-1}\,dx\,dy\right\}>\frac{47}{100}. 4 ( 4/5 ) α − 1 { ∫ [ − 1 , 1 ] 2 [ C 1 + α B 1 + ( C 2 + α B 2 ) x 2 ] q 0 , sq α − 1 d x d y + ∫ [ − 1 , 1 ] 2 [ C 3 + α B 3 + ( C 4 + α B 4 ) x + ( C 5 + α B 5 ) x 2 + ( C 6 + α B 6 ) x y ] q 0 , rh α − 1 d x d y } > 100 47 .
Proof of Numerical Lemma 5.5. The identity
16 U ( ρ ) − ( det Q ) U ( 1 ) = [ 16 ( W 01 , 23 + W 02 , 13 + W 12 , 03 ) − 12 det Q ] ( 16 , 4 , − 14 , 4 , − 4 , 2 ) 16U(\rho)-(\det Q)U(1)=[16(W_{01,23}+W_{02,13}+W_{12,03})-12\det Q](16,4,-14,4,-4,2) 16 U ( ρ ) − ( det Q ) U ( 1 ) = [ 16 ( W 01 , 23 + W 02 , 13 + W 12 , 03 ) − 12 det Q ] ( 16 , 4 , − 14 , 4 , − 4 , 2 )
gives P α s y m = ( C ( ρ ) + α B ( ρ ) ) ⋅ F α P_\alpha^{\mathrm{sym}}=(C(\rho)+\alpha B(\rho))\cdot F_\alpha P α sym = ( C ( ρ ) + α B ( ρ )) ⋅ F α in (35) . Each polynomial in (40) is homogeneous of degree ten. For nonnegative integers n , k n,k n , k ,
∑ m = 0 n ( − 1 ) n − m ( n m ) m k = { 0 , k < n , n ! , k = n . ( k ≤ n , 0 0 = 1 ) . \sum_{m=0}^{n}(-1)^{n-m}\binom{n}{m}m^k=
\begin{cases}
0,&k<n,\\
n!,&k=n.
\end{cases}
\qquad(k\leq n,\ 0^0=1). m = 0 ∑ n ( − 1 ) n − m ( m n ) m k = { 0 , n ! , k < n , k = n . ( k ≤ n , 0 0 = 1 ) .
Indeed, the left side is the n n n th forward difference of x k x^k x k at zero, and each difference lowers the degree by one and multiplies the leading coefficient by the previous degree. A monomial X 0 k 0 ⋯ X 5 k 5 X_0^{k_0}\cdots X_5^{k_5} X 0 k 0 ⋯ X 5 k 5 of degree ten with ( k 0 , … , k 5 ) ≠ ν (k_0,\ldots,k_5)\ne\nu ( k 0 , … , k 5 ) = ν has k i < ν i k_i<\nu_i k i < ν i for some i i i , so it contributes zero to (41) . The monomial with exponents ν \nu ν contributes its coefficient times ν 0 ! ⋯ ν 5 ! \nu_0!\cdots\nu_5! ν 0 ! ⋯ ν 5 ! before division. This proves the formula for the two coefficient vectors.
For the numerical counts, in each order evaluate (40) at ρ σ ( i ) = m i + ⋯ + m 5 \rho_{\sigma(i)}=m_i+\cdots+m_5 ρ σ ( i ) = m i + ⋯ + m 5 for the ( 16 6 ) = 8 008 \binom{16}{6}=8\,008 ( 6 16 ) = 8 008 tuples of nonnegative integers with m 0 + ⋯ + m 5 ≤ 10 m_0+\cdots+m_5\leq10 m 0 + ⋯ + m 5 ≤ 10 . Apply (41) to the ( 30 5 ) = 90 090 \binom{30}{5}=90\,090 ( 5 30 ) = 90 090 positions. The exact zero tests give
90 090 = 86 172 + 3 738 + 180. 90\,090=86\,172+3\,738+180. 90 090 = 86 172 + 3 738 + 180.
The gcd division gives 30 812 distinct pairs. For each of them the stated inequalities follow by summing the six absolute values in each vector and testing the integer 100 B 1 + 100 B 2 + 50 B 3 − 25 B 4 + 50 B 5 − 13 ∑ ℓ = 1 6 ∣ B ℓ ∣ 100B_1+100B_2+50B_3-25B_4+50B_5-13\sum_{\ell=1}^{6}|B_\ell| 100 B 1 + 100 B 2 + 50 B 3 − 25 B 4 + 50 B 5 − 13 ∑ ℓ = 1 6 ∣ B ℓ ∣ , which is nonnegative.
An explicit calculation of the coefficient pair for X 2 10 X_2^{10} X 2 10 . Use σ = ( 01 , 02 , 03 , 12 , 13 , 23 ) \sigma=(01,02,03,12,13,23) σ = ( 01 , 02 , 03 , 12 , 13 , 23 ) and ν = ( 0 , 0 , 10 , 0 , 0 , 0 ) \nu=(0,0,10,0,0,0) ν = ( 0 , 0 , 10 , 0 , 0 , 0 ) . The coefficient vectors of X 2 10 X_2^{10} X 2 10 are the values of the homogeneous degree-ten polynomials C ( ρ ) C(\rho) C ( ρ ) and B ( ρ ) B(\rho) B ( ρ ) at X 2 = 1 X_2=1 X 2 = 1 and X i = 0 X_i=0 X i = 0 for i ≠ 2 i\ne2 i = 2 . By (18) , this gives
( ρ 01 , ρ 02 , ρ 03 , ρ 12 , ρ 13 , ρ 23 ) = ( 1 , 1 , 1 , 0 , 0 , 0 ) , Q = Id 3 . (\rho_{01},\rho_{02},\rho_{03},\rho_{12},\rho_{13},\rho_{23})=(1,1,1,0,0,0),\qquad Q=\operatorname{Id}_3. ( ρ 01 , ρ 02 , ρ 03 , ρ 12 , ρ 13 , ρ 23 ) = ( 1 , 1 , 1 , 0 , 0 , 0 ) , Q = Id 3 .
Only the three rhombus tiles of weight one remain, each with 4 q τ = 1 + x 2 + y 2 4q\tau=1+x^2+y^2 4 q τ = 1 + x 2 + y 2 . Hence
U ( ρ ) = ( 0 , 0 , 42 , − 12 , 12 , − 6 ) , ∑ T ∈ T W T Φ T ( 4 q τ det Q ) = ( 0 , 0 , 3 , 0 , 6 , 0 ) . U(\rho)=(0,0,42,-12,12,-6),\qquad\sum_{T\in\mathcal{T}}W_T\Phi_T(4q\tau\det Q)=(0,0,3,0,6,0). U ( ρ ) = ( 0 , 0 , 42 , − 12 , 12 , − 6 ) , T ∈ T ∑ W T Φ T ( 4 q τ det Q ) = ( 0 , 0 , 3 , 0 , 6 , 0 ) .
Inserting these values and ∑ ρ ∈ E ρ ℓ = 3 \sum_{\rho\in\mathcal{E}}\rho_\ell=3 ∑ ρ ∈ E ρ ℓ = 3 in (40) gives
( C σ , ν , B σ , ν ) = 186 624 ( ( − 8 , − 2 , 7 , − 2 , 2 , − 1 ) , ( 0 , 0 , − 8 , 8 , 32 , 4 ) ) . (42) (C_{\sigma,\nu},B_{\sigma,\nu})=186\,624\bigl((-8,-2,7,-2,2,-1),(0,0,-8,8,32,4)\bigr). \tag*{(42)} ( C σ , ν , B σ , ν ) = 186 624 ( ( − 8 , − 2 , 7 , − 2 , 2 , − 1 ) , ( 0 , 0 , − 8 , 8 , 32 , 4 ) ) . (42)
The positive gcd of the twelve entries is 186 624. Dividing both vectors by this number gives the pair in C \mathcal{C} C .
Proof of Numerical Lemma 5.6. Use the intervals
I = { [ 35 / 2 , 18 ] } ∪ { [ j , j + 1 ] : j = 18 , … , 57 } \mathcal{I}=\{[35/2,18]\}\cup\{[j,j+1]:j=18,\ldots,57\} I = {[ 35/2 , 18 ]} ∪ {[ j , j + 1 ] : j = 18 , … , 57 }
I ∪ { [ 58 + 8 j , 66 + 8 j ] : j = 0 , … , 56 } , \phantom{\mathcal{I}}{}\cup\{[58+8j,66+8j]:j=0,\ldots,56\}, I ∪ {[ 58 + 8 j , 66 + 8 j ] : j = 0 , … , 56 } ,
whose union is [ 35 / 2 , 514 ] [35/2,514] [ 35/2 , 514 ] . Fix I = [ a , b ] ∈ I I=[a,b]\in\mathcal{I} I = [ a , b ] ∈ I and put n = ⌊ a ⌋ − 1 n=\lfloor a\rfloor-1 n = ⌊ a ⌋ − 1 . Then
α − n − 1 ∈ { [ 1 / 2 , 1 ] , I = [ 35 / 2 , 18 ] , [ 0 , 1 ] , I = [ j , j + 1 ] , j = 18 , … , 57 , [ 0 , 8 ] , I = [ 58 + 8 j , 66 + 8 j ] , j = 0 , … , 56. \alpha-n-1\in
\begin{cases}
[1/2,1],& I=[35/2,18],\\
[0,1],& I=[j,j+1],\quad j=18,\ldots,57,\\
[0,8],& I=[58+8j,66+8j],\quad j=0,\ldots,56.
\end{cases} α − n − 1 ∈ ⎩ ⎨ ⎧ [ 1/2 , 1 ] , [ 0 , 1 ] , [ 0 , 8 ] , I = [ 35/2 , 18 ] , I = [ j , j + 1 ] , j = 18 , … , 57 , I = [ 58 + 8 j , 66 + 8 j ] , j = 0 , … , 56.
Recall that ( F α ) ℓ = 1 4 ∫ [ − 1 , 1 ] 2 R u α − 1 d x d y (F_\alpha)_\ell=\frac{1}{4}\int_{[-1,1]^2}Ru^{\alpha-1}\,dx\,dy ( F α ) ℓ = 4 1 ∫ [ − 1 , 1 ] 2 R u α − 1 d x d y , with
u = { ( 4 / 5 ) q 0 , s q , ℓ = 1 , 2 , ( 4 / 5 ) q 0 , r h , ℓ = 3 , 4 , 5 , 6 , R = 1 , x 2 , 1 , x , x 2 , x y in order of ℓ . u=
\begin{cases}
(4/5)q_{0,\mathrm{sq}},& \ell=1,2,\\
(4/5)q_{0,\mathrm{rh}},& \ell=3,4,5,6,
\end{cases}
\qquad R=1,x^2,1,x,x^2,xy\quad\text{in order of }\ell. u = { ( 4/5 ) q 0 , sq , ( 4/5 ) q 0 , rh , ℓ = 1 , 2 , ℓ = 3 , 4 , 5 , 6 , R = 1 , x 2 , 1 , x , x 2 , x y in order of ℓ .
Since ∣ R ∣ ≤ 1 |R|\le1 ∣ R ∣ ≤ 1 and 0 ≤ 1 − u ≤ 2 / 5 0\le1-u\le2/5 0 ≤ 1 − u ≤ 2/5 , the binomial series gives
∣ ( F α ) ℓ − 1 4 ∑ j = 0 30 ( − 1 ) j ( α − n − 1 j ) ∫ [ − 1 , 1 ] 2 R u n ( 1 − u ) j d x d y ∣ ≤ ε ℓ , \left|(F_\alpha)_\ell-\frac{1}{4}\sum_{j=0}^{30}(-1)^j\binom{\alpha-n-1}{j}\int_{[-1,1]^2}Ru^n(1-u)^j\,dx\,dy\right|\le\varepsilon_\ell, ( F α ) ℓ − 4 1 j = 0 ∑ 30 ( − 1 ) j ( j α − n − 1 ) ∫ [ − 1 , 1 ] 2 R u n ( 1 − u ) j d x d y ≤ ε ℓ ,
where
ε ℓ = ( 2 / 5 ) 31 1 − 2 / 5 { 1 , I ⊆ [ 35 / 2 , 58 ] , 8 8 4 ⋅ 8 ! ∫ [ − 1 , 1 ] 2 u n d x d y , I ⊆ [ 58 , 514 ] . \varepsilon_\ell=\frac{(2/5)^{31}}{1-2/5}
\begin{cases}
1,& I\subseteq[35/2,58],\\
\dfrac{8^8}{4\cdot8!}\int_{[-1,1]^2}u^n\,dx\,dy,& I\subseteq[58,514].
\end{cases} ε ℓ = 1 − 2/5 ( 2/5 ) 31 ⎩ ⎨ ⎧ 1 , 4 ⋅ 8 ! 8 8 ∫ [ − 1 , 1 ] 2 u n d x d y , I ⊆ [ 35/2 , 58 ] , I ⊆ [ 58 , 514 ] .
For the first case, the product for ( α − n − 1 j ) \binom{\alpha-n-1}{j} ( j α − n − 1 ) has absolute value at most one. In the second case,
∣ ( α − n − 1 8 ) ∣ ≤ 8 8 8 ! , \left|\binom{\alpha-n-1}{8}\right|\le\frac{8^8}{8!}, ( 8 α − n − 1 ) ≤ 8 ! 8 8 ,
∣ ( α − n − 1 j + 1 ) ∣ = ∣ j − ( α − n − 1 ) j + 1 ∣ ∣ ( α − n − 1 j ) ∣ ≤ ∣ ( α − n − 1 j ) ∣ ( j ≥ 8 ) . \left|\binom{\alpha-n-1}{j+1}\right|
=\left|\frac{j-(\alpha-n-1)}{j+1}\right|\left|\binom{\alpha-n-1}{j}\right|
\le\left|\binom{\alpha-n-1}{j}\right|\qquad(j\ge8). ( j + 1 α − n − 1 ) = j + 1 j − ( α − n − 1 ) ( j α − n − 1 ) ≤ ( j α − n − 1 ) ( j ≥ 8 ) .
Summing the geometric tail gives the displayed bound.
For α = a + ( b − a ) t \alpha=a+(b-a)t α = a + ( b − a ) t , 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 , and r = 0 , 1 r=0,1 r = 0 , 1 , the power coefficients a ℓ , i ( r ) a_{\ell,i}^{(r)} a ℓ , i ( r ) are given by
∑ i = 0 31 a ℓ , i ( r ) t i = ( a + ( b − a ) t ) r 4 ∑ j = 0 30 ( − 1 ) j ( a + ( b − a ) t − n − 1 j ) ∫ [ − 1 , 1 ] 2 R u n ( 1 − u ) j d x d y . \sum_{i=0}^{31}a_{\ell,i}^{(r)}t^i
=\frac{(a+(b-a)t)^r}{4}\sum_{j=0}^{30}(-1)^j\binom{a+(b-a)t-n-1}{j}\int_{[-1,1]^2}Ru^n(1-u)^j\,dx\,dy. i = 0 ∑ 31 a ℓ , i ( r ) t i = 4 ( a + ( b − a ) t ) r j = 0 ∑ 30 ( − 1 ) j ( j a + ( b − a ) t − n − 1 ) ∫ [ − 1 , 1 ] 2 R u n ( 1 − u ) j d x d y .
This polynomial differs from α r ( F α ) ℓ \alpha^{r}(F_{\alpha})_{\ell} α r ( F α ) ℓ by at most b r ε ℓ b^{r}\varepsilon_{\ell} b r ε ℓ . The spatial integrals are
∫ [ − 1 , 1 ] 2 R u n ( 1 − u ) j d x d y = ∑ i = 0 j ( − 1 ) i ( j i ) ∫ [ − 1 , 1 ] 2 R u n + i d x d y , 0 ≤ j ≤ 30. \int_{[-1,1]^{2}} R u^{n}(1-u)^{j}\,dx\,dy=\sum_{i=0}^{j}(-1)^{i}\binom{j}{i}\int_{[-1,1]^{2}}R u^{n+i}\,dx\,dy,\qquad0\leq j\leq30. ∫ [ − 1 , 1 ] 2 R u n ( 1 − u ) j d x d y = i = 0 ∑ j ( − 1 ) i ( i j ) ∫ [ − 1 , 1 ] 2 R u n + i d x d y , 0 ≤ j ≤ 30.
For I ⊆ [ 35 / 2 , 58 ] I\subseteq[35/2,58] I ⊆ [ 35/2 , 58 ] and R = x a y b R=x^{a}y^{b} R = x a y b , (2.11) gives 1 4 ∫ [ − 1 , 1 ] 2 R u n + i d x d y \frac{1}{4}\int_{[-1,1]^{2}}R u^{n+i}\,dx\,dy 4 1 ∫ [ − 1 , 1 ] 2 R u n + i d x d y by using u u u as the quadratic and n + i n+i n + i as its power. For I ⊆ [ 58 , 514 ] I\subseteq[58,514] I ⊆ [ 58 , 514 ] , use the coordinates Y Y Y and domains Ω \Omega Ω given by
u = 4 5 q 0 , s q : Y = ( x , y ) , Ω = [ − 1 , 1 ] 2 , u = 4 5 + Y 1 2 + Y 2 2 10 , u=\frac{4}{5}q_{0,\mathrm{sq}}:\quad Y=(x,y),\quad\Omega=[-1,1]^{2},\quad u=\frac{4}{5}+\frac{Y_{1}^{2}+Y_{2}^{2}}{10}, u = 5 4 q 0 , sq : Y = ( x , y ) , Ω = [ − 1 , 1 ] 2 , u = 5 4 + 10 Y 1 2 + Y 2 2 ,
u = 4 5 q 0 , r h : Y = ( x − 1 , y − 1 ) , Ω = [ − 2 , 0 ] 2 , u = 3 5 + Y 1 2 − Y 1 Y 2 + Y 2 2 10 . u=\frac{4}{5}q_{0,\mathrm{rh}}:\quad Y=(x-1,y-1),\quad\Omega=[-2,0]^{2},\quad u=\frac{3}{5}+\frac{Y_{1}^{2}-Y_{1}Y_{2}+Y_{2}^{2}}{10}. u = 5 4 q 0 , rh : Y = ( x − 1 , y − 1 ) , Ω = [ − 2 , 0 ] 2 , u = 5 3 + 10 Y 1 2 − Y 1 Y 2 + Y 2 2 .
The translated weights are 1 1 1 , Y 1 + 1 Y_{1}+1 Y 1 + 1 , ( Y 1 + 1 ) 2 (Y_{1}+1)^{2} ( Y 1 + 1 ) 2 and ( Y 1 + 1 ) ( Y 2 + 1 ) (Y_{1}+1)(Y_{2}+1) ( Y 1 + 1 ) ( Y 2 + 1 ) . Since Y ⋅ ∇ u = 2 ( u − u ( 0 ) ) Y\cdot\nabla u=2(u-u(0)) Y ⋅ ∇ u = 2 ( u − u ( 0 )) , the divergence theorem gives, for m ≥ 1 m\geq1 m ≥ 1 ,
( h + k + 2 + 2 m ) ∫ Ω Y 1 h Y 2 k u m d Y = ∫ ∂ Ω ( Y ⋅ ν ) Y 1 h Y 2 k u m d S + 2 m u ( 0 ) ∫ Ω Y 1 h Y 2 k u m − 1 d Y . (h+k+2+2m)\int_{\Omega}Y_{1}^{h}Y_{2}^{k}u^{m}\,dY=\int_{\partial\Omega}(Y\cdot\nu)Y_{1}^{h}Y_{2}^{k}u^{m}\,dS+2mu(0)\int_{\Omega}Y_{1}^{h}Y_{2}^{k}u^{m-1}\,dY. ( h + k + 2 + 2 m ) ∫ Ω Y 1 h Y 2 k u m d Y = ∫ ∂ Ω ( Y ⋅ ν ) Y 1 h Y 2 k u m d S + 2 m u ( 0 ) ∫ Ω Y 1 h Y 2 k u m − 1 d Y .
∫ [ L , H ] 2 Y 1 h Y 2 k d Y = ( H h + 1 − L h + 1 ) ( H k + 1 − L k + 1 ) ( h + 1 ) ( k + 1 ) . \int_{[L,H]^{2}}Y_{1}^{h}Y_{2}^{k}\,dY=\frac{(H^{h+1}-L^{h+1})(H^{k+1}-L^{k+1})}{(h+1)(k+1)}. ∫ [ L , H ] 2 Y 1 h Y 2 k d Y = ( h + 1 ) ( k + 1 ) ( H h + 1 − L h + 1 ) ( H k + 1 − L k + 1 ) .
Here ν \nu ν is the outward unit normal, and the last line gives the initial values at m = 0 m=0 m = 0 . Every edge with Y ⋅ ν ≠ 0 Y\cdot\nu\ne0 Y ⋅ ν = 0 has u = ( 9 + s 2 ) / 10 u=(9+s^{2})/10 u = ( 9 + s 2 ) /10 for − 1 ≤ s ≤ 1 -1\leq s\leq1 − 1 ≤ s ≤ 1 . Define
K j , m = ∫ − 1 1 s j ( 9 + s 2 10 ) m d s . K_{j,m}=\int_{-1}^{1}s^{j}\left(\frac{9+s^{2}}{10}\right)^{m}\,ds. K j , m = ∫ − 1 1 s j ( 10 9 + s 2 ) m d s .
Integrating the derivative of s j + 1 ( ( 9 + s 2 ) / 10 ) m s^{j+1}((9+s^{2})/10)^{m} s j + 1 (( 9 + s 2 ) /10 ) m gives
K j , 0 = 1 − ( − 1 ) j + 1 j + 1 , K_{j,0}=\frac{1-(-1)^{j+1}}{j+1}, K j , 0 = j + 1 1 − ( − 1 ) j + 1 ,
( j + 1 + 2 m ) K j , m = 1 − ( − 1 ) j + 1 + 9 m 5 K j , m − 1 ( m ≥ 1 ) . (j+1+2m)K_{j,m}=1-(-1)^{j+1}+\frac{9m}{5}K_{j,m-1}\qquad(m\geq1). ( j + 1 + 2 m ) K j , m = 1 − ( − 1 ) j + 1 + 5 9 m K j , m − 1 ( m ≥ 1 ) .
On the square take s = Y j s=Y_{j} s = Y j on the edge with fixed coordinate Y i = ± 1 Y_{i}=\pm1 Y i = ± 1 . On the translated rhombus the edges Y i = 0 Y_{i}=0 Y i = 0 have Y ⋅ ν = 0 Y\cdot\nu=0 Y ⋅ ν = 0 , and on Y i = − 2 Y_{i}=-2 Y i = − 2 take s = Y j + 1 s=Y_{j}+1 s = Y j + 1 . The boundary term for Y 1 h Y 2 k Y_{1}^{h}Y_{2}^{k} Y 1 h Y 2 k is therefore
{ ( 1 + ( − 1 ) h ) K k , m + ( 1 + ( − 1 ) k ) K h , m , Ω = [ − 1 , 1 ] 2 , 2 ( − 2 ) h ∑ j = 0 k ( k j ) ( − 1 ) k − j K j , m + 2 ( − 2 ) k ∑ j = 0 h ( h j ) ( − 1 ) h − j K j , m , Ω = [ − 2 , 0 ] 2 . \begin{cases}
(1+(-1)^{h})K_{k,m}+(1+(-1)^{k})K_{h,m}, & \Omega=[-1,1]^{2},\\
2(-2)^{h}\sum_{j=0}^{k}\binom{k}{j}(-1)^{k-j}K_{j,m}+2(-2)^{k}\sum_{j=0}^{h}\binom{h}{j}(-1)^{h-j}K_{j,m}, & \Omega=[-2,0]^{2}.
\end{cases} { ( 1 + ( − 1 ) h ) K k , m + ( 1 + ( − 1 ) k ) K h , m , 2 ( − 2 ) h ∑ j = 0 k ( j k ) ( − 1 ) k − j K j , m + 2 ( − 2 ) k ∑ j = 0 h ( j h ) ( − 1 ) h − j K j , m , Ω = [ − 1 , 1 ] 2 , Ω = [ − 2 , 0 ] 2 .
For j = 0 , … , 31 j=0,\ldots,31 j = 0 , … , 31 , add the residual interval [ − b r ε ℓ , b r ε ℓ ] [-b^{r}\varepsilon_{\ell},b^{r}\varepsilon_{\ell}] [ − b r ε ℓ , b r ε ℓ ] to each Bernstein coefficient of the approximating polynomial. The resulting integer endpoints at scale 2 60 2^{60} 2 60 are
l ℓ , j ( r ) = ⌊ 2 60 ( ∑ i = 0 j ( j i ) ( 31 i ) a ℓ , i ( r ) − b r ε ℓ ) ⌋ , l_{\ell,j}^{(r)}=\left\lfloor2^{60}\left(\sum_{i=0}^{j}\frac{\binom{j}{i}}{\binom{31}{i}}a_{\ell,i}^{(r)}-b^{r}\varepsilon_{\ell}\right)\right\rfloor, l ℓ , j ( r ) = ⌊ 2 60 ( i = 0 ∑ j ( i 31 ) ( i j ) a ℓ , i ( r ) − b r ε ℓ ) ⌋ ,
u ℓ , j ( r ) = ⌈ 2 60 ( ∑ i = 0 j ( j i ) ( 31 i ) a ℓ , i ( r ) + b r ε ℓ ) ⌉ . u_{\ell,j}^{(r)}=\left\lceil2^{60}\left(\sum_{i=0}^{j}\frac{\binom{j}{i}}{\binom{31}{i}}a_{\ell,i}^{(r)}+b^{r}\varepsilon_{\ell}\right)\right\rceil. u ℓ , j ( r ) = ⌈ 2 60 ( i = 0 ∑ j ( i 31 ) ( i j ) a ℓ , i ( r ) + b r ε ℓ ) ⌉ .
The Bernstein basis is nonnegative and sums to one on [ 0 , 1 ] [0,1] [ 0 , 1 ] , so
α r ( F α ) ℓ ∈ 2 − 60 ∑ j = 0 31 ( 31 j ) t j ( 1 − t ) 31 − j [ l ℓ , j ( r ) , u ℓ , j ( r ) ] . \alpha^{r}(F_{\alpha})_{\ell}\in2^{-60}\sum_{j=0}^{31}\binom{31}{j}t^{j}(1-t)^{31-j}[l_{\ell,j}^{(r)},u_{\ell,j}^{(r)}]. α r ( F α ) ℓ ∈ 2 − 60 j = 0 ∑ 31 ( j 31 ) t j ( 1 − t ) 31 − j [ l ℓ , j ( r ) , u ℓ , j ( r ) ] .
Using the same basis for both terms of ( C + α B ) ⋅ F α (C+\alpha B)\cdot F_\alpha ( C + α B ) ⋅ F α gives the lower bound
1 2 60 min 0 ≤ j ≤ 31 ∑ ℓ = 1 6 [ min { C ℓ ℓ , j ( 0 ) , C ℓ u , j ( 0 ) } + min { B ℓ ℓ , j ( 1 ) , B ℓ u , j ( 1 ) } ] . \frac{1}{2^{60}}\min_{0\le j\le31}\sum_{\ell=1}^{6}\left[\min\{C_{\ell\ell,j}^{(0)},C_{\ell u,j}^{(0)}\}+\min\{B_{\ell\ell,j}^{(1)},B_{\ell u,j}^{(1)}\}\right]. 2 60 1 0 ≤ j ≤ 31 min ℓ = 1 ∑ 6 [ min { C ℓℓ , j ( 0 ) , C ℓ u , j ( 0 ) } + min { B ℓℓ , j ( 1 ) , B ℓ u , j ( 1 ) } ] .
Taking the minimum of this expression over ( C , B ) ∈ C (C,B)\in\mathcal{C} ( C , B ) ∈ C and I ∈ I I\in\mathcal{I} I ∈ I with I ⊆ J I\subseteq J I ⊆ J gives
{ 4 288 754 204 510 563 386 2 60 , J = [ 35 / 2 , 58 ] , 552 536 073 955 686 419 2 60 , J = [ 58 , 514 ] . \begin{cases}
\dfrac{4\,288\,754\,204\,510\,563\,386}{2^{60}}, & J=[35/2,58],\\
\dfrac{552\,536\,073\,955\,686\,419}{2^{60}}, & J=[58,514].
\end{cases} ⎩ ⎨ ⎧ 2 60 4 288 754 204 510 563 386 , 2 60 552 536 073 955 686 419 , J = [ 35/2 , 58 ] , J = [ 58 , 514 ] .
Both minima exceed 47 / 100 47/100 47/100 . They occur at j = 31 j=31 j = 31 on [ 57 , 58 ] [57,58] [ 57 , 58 ] and [ 506 , 514 ] [506,514] [ 506 , 514 ] , respectively, for
( C , B ) = ( ( − 8 , − 2 , 7 , − 2 , 2 , − 1 ) , ( 0 , 0 , − 8 , 8 , 32 , 4 ) ) . (C,B)=((-8,-2,7,-2,2,-1),(0,0,-8,8,32,4)). ( C , B ) = (( − 8 , − 2 , 7 , − 2 , 2 , − 1 ) , ( 0 , 0 , − 8 , 8 , 32 , 4 )) .
Acknowledgements. The research of S. Luo is partially supported by the National Natural Science Foundation of China (NSFC) under Grant Nos. 12261045 and 12001253, and by the Jiangxi Jieqing Fund under Grant No. 20242BAB23001. The author used AI to assist the long but finite computations.
Statements and Declarations: there is no conflict of interest.
Data availability: the manuscript has no associated data.
References