Introduction
We prove that BCC uniquely maximizes the first positive Neumann eigenvalue at fixed covolume by Selling reduction, local eigenvalue splitting and global Rayleigh bounds.
Let Λ ⊂ R 3 \Lambda\subset\mathbb{R}^{3} Λ ⊂ R 3 be a full-rank lattice, and let
V Λ = { x ∈ R 3 : ∣ x ∣ ≤ ∣ x − v ∣ for all v ∈ Λ } (1) V_{\Lambda}=\{x\in\mathbb{R}^{3}: |x|\leq|x-v|\text{ for all }v\in\Lambda\} \tag*{(1)} V Λ = { x ∈ R 3 : ∣ x ∣ ≤ ∣ x − v ∣ for all v ∈ Λ } (1)
be its Voronoi cell. For a bounded connected Lipschitz domain Ω \Omega Ω , write μ 1 ( Ω ) \mu_{1}(\Omega) μ 1 ( Ω ) for the first positive Neumann eigenvalue:
μ 1 ( Ω ) = inf u ∈ H 1 ( Ω ) ∖ { 0 } ∫ Ω u = 0 ∫ Ω ∣ ∇ u ∣ 2 ∫ Ω u 2 . (2) \mu_{1}(\Omega)=\inf_{\substack{u\in H^{1}(\Omega)\setminus\{0\}\\ \int_{\Omega}u=0}}\frac{\int_{\Omega}|\nabla u|^{2}}{\int_{\Omega}u^{2}}. \tag*{(2)} μ 1 ( Ω ) = u ∈ H 1 ( Ω ) ∖ { 0 } ∫ Ω u = 0 inf ∫ Ω u 2 ∫ Ω ∣∇ u ∣ 2 . (2)
Since
∣ V Λ ∣ = covol Λ , μ 1 ( a Ω ) = a − 2 μ 1 ( Ω ) , (3) |V_{\Lambda}|=\operatorname{covol}\Lambda,\qquad\mu_{1}(a\Omega)=a^{-2}\mu_{1}(\Omega), \tag*{(3)} ∣ V Λ ∣ = covol Λ , μ 1 ( a Ω ) = a − 2 μ 1 ( Ω ) , (3)
we normalize covol Λ = 1 \operatorname{covol}\Lambda=1 covol Λ = 1 .
The Szegő–Weinberger inequality [17 , 19 ] gives
∣ Ω ∣ 2 / 3 μ 1 ( Ω ) ≤ ∣ B ∣ 2 / 3 μ 1 ( B ) |\Omega|^{2/3}\mu_{1}(\Omega)\leq|B|^{2/3}\mu_{1}(B) ∣Ω ∣ 2/3 μ 1 ( Ω ) ≤ ∣ B ∣ 2/3 μ 1 ( B )
where B ⊂ R 3 B\subset\mathbb{R}^{3} B ⊂ R 3 is a ball; equality holds precisely when Ω \Omega Ω is a ball. Brasco and Pratelli [3 ] established sharp quantitative stability. We determine the maximizer among lattice Voronoi cells. Set the body-centered cubic lattice as follows:
B C C = 2 1 / 3 [ Z 3 ∪ ( Z 3 + 1 2 ( 1 , 1 , 1 ) ) ] , covol B C C = 1. (4) \mathrm{BCC}=2^{1/3}\left[\mathbb{Z}^{3}\cup\left(\mathbb{Z}^{3}+\frac{1}{2}(1,1,1)\right)\right],\qquad\operatorname{covol}\mathrm{BCC}=1. \tag*{(4)} BCC = 2 1/3 [ Z 3 ∪ ( Z 3 + 2 1 ( 1 , 1 , 1 ) ) ] , covol BCC = 1. (4)
Theorem 1.1. For every full-rank lattice Λ ⊂ R 3 \Lambda\subset\mathbb{R}^{3} Λ ⊂ R 3 of covolume one,
μ 1 ( V Λ ) ≤ μ 1 ( V B C C ) , (5) \mu_{1}(V_{\Lambda})\leq\mu_{1}(V_{\mathrm{BCC}}), \tag*{(5)} μ 1 ( V Λ ) ≤ μ 1 ( V BCC ) , (5) with equality if and only if Λ = R B C C \Lambda=R\mathrm{BCC} Λ = R BCC for some R ∈ O ( 3 ) R\in O(3) R ∈ O ( 3 ) . The maximizing cell is the regular truncated octahedron.
2020 Mathematics Subject Classification. Primary 35P15; Secondary 11H06, 52C07.
Keywords and phrases. Neumann eigenvalue, lattice Voronoi cell, body-centered cubic lattice, spectral optimization, Selling reduction.
Selling reduction and geometric bounds
Six nonnegative Selling conorms parametrize lattice shapes. Dual sine trials and projected moment comparisons give the short-vector bounds.
Put I = { 0 , 1 , 2 , 3 } I=\{0,1,2,3\} I = { 0 , 1 , 2 , 3 } and list the six edges as
E = { 01 , 02 , 03 , 12 , 13 , 23 } , i j = { i , j } ( 0 ≤ i < j ≤ 3 ) . E=\{01,02,03,12,13,23\},\qquad ij=\{i,j\}\quad(0\le i<j\le3). E = { 01 , 02 , 03 , 12 , 13 , 23 } , ij = { i , j } ( 0 ≤ i < j ≤ 3 ) .
We write p = ( p 01 , p 02 , p 03 , p 12 , p 13 , p 23 ) p=(p_{01},p_{02},p_{03},p_{12},p_{13},p_{23}) p = ( p 01 , p 02 , p 03 , p 12 , p 13 , p 23 ) and set p j i = p i j p_{ji}=p_{ij} p j i = p ij for i ≠ j i\ne j i = j . A superbase ( b i ) i = 0 3 (b_i)_{i=0}^3 ( b i ) i = 0 3 satisfies ∑ i = 0 3 b i = 0 \sum_{i=0}^3b_i=0 ∑ i = 0 3 b i = 0 , with any three vectors forming a lattice basis. Every three-dimensional lattice has an obtuse superbase [6 ] . Indeed, if b i ⋅ b j > 0 b_i\cdot b_j>0 b i ⋅ b j > 0 , the Selling move
( b i , b j , b k , b l ) ⟼ ( − b i , b j , b k + b i , b l + b i ) (b_i,b_j,b_k,b_l)\longmapsto(-b_i,b_j,b_k+b_i,b_l+b_i) ( b i , b j , b k , b l ) ⟼ ( − b i , b j , b k + b i , b l + b i )
preserves the superbase property and satisfies
∑ r = 0 3 ∣ b r ′ ∣ 2 − ∑ r = 0 3 ∣ b r ∣ 2 = 2 b i ⋅ ( b k + b l ) + 2 ∣ b i ∣ 2 = − 2 b i ⋅ b j < 0. \sum_{r=0}^{3}|b'_r|^2-\sum_{r=0}^{3}|b_r|^2=2b_i\cdot(b_k+b_l)+2|b_i|^2=-2b_i\cdot b_j<0. r = 0 ∑ 3 ∣ b r ′ ∣ 2 − r = 0 ∑ 3 ∣ b r ∣ 2 = 2 b i ⋅ ( b k + b l ) + 2∣ b i ∣ 2 = − 2 b i ⋅ b j < 0.
Only finitely many superbases have bounded squared length, so the reduction terminates.
Normalize the conorms by
p i j = − b i ⋅ b j ∑ 0 ≤ a < b ≤ 3 ( − b a ⋅ b b ) , p ∈ Δ : = { p ≥ 0 : p 01 + p 02 + p 03 + p 12 + p 13 + p 23 = 1 } . (8) p_{ij}=\frac{-b_i\cdot b_j}{\sum_{0\le a<b\le3}(-b_a\cdot b_b)},\qquad
p\in\Delta:=\{p\ge0:p_{01}+p_{02}+p_{03}+p_{12}+p_{13}+p_{23}=1\}. \tag*{(8)} p ij = ∑ 0 ≤ a < b ≤ 3 ( − b a ⋅ b b ) − b i ⋅ b j , p ∈ Δ := { p ≥ 0 : p 01 + p 02 + p 03 + p 12 + p 13 + p 23 = 1 } . (8)
Put e 0 = 0 e_0=0 e 0 = 0 , let e 1 , e 2 , e 3 e_1,e_2,e_3 e 1 , e 2 , e 3 be the coordinate vectors, and define
d i j = e i − e j , G ( p ) = ∑ 0 ≤ i < j ≤ 3 p i j d i j d i j T , (9) d_{ij}=e_i-e_j,\qquad G(p)=\sum_{0\le i<j\le3}p_{ij}d_{ij}d_{ij}^{T}, \tag*{(9)} d ij = e i − e j , G ( p ) = 0 ≤ i < j ≤ 3 ∑ p ij d ij d ij T , (9)
D ( p ) = det G ( p ) , C i j ( p ) = d i j T adj ( G ( p ) ) d i j ( 0 ≤ i < j ≤ 3 ) . D(p)=\det G(p),\qquad C_{ij}(p)=d_{ij}^{T}\operatorname{adj}(G(p))d_{ij}\quad(0\le i<j\le3). D ( p ) = det G ( p ) , C ij ( p ) = d ij T adj ( G ( p )) d ij ( 0 ≤ i < j ≤ 3 ) .
For e = i j ∈ E e=ij\in E e = ij ∈ E , the edge notation means p e = p i j p_e=p_{ij} p e = p ij , d e = d i j d_e=d_{ij} d e = d ij , and C e = C i j C_e=C_{ij} C e = C ij . In particular,
G ( p ) = ( p 01 + p 12 + p 13 − p 12 − p 13 − p 12 p 02 + p 12 + p 23 − p 23 − p 13 − p 23 p 03 + p 13 + p 23 ) . G(p)=\begin{pmatrix}
p_{01}+p_{12}+p_{13} & -p_{12} & -p_{13}\\
-p_{12} & p_{02}+p_{12}+p_{23} & -p_{23}\\
-p_{13} & -p_{23} & p_{03}+p_{13}+p_{23}
\end{pmatrix}. G ( p ) = p 01 + p 12 + p 13 − p 12 − p 13 − p 12 p 02 + p 12 + p 23 − p 23 − p 13 − p 23 p 03 + p 13 + p 23 .
The identity
∣ ∑ i = 1 3 n i b i ∣ 2 = ∑ 0 ≤ i < j ≤ 3 ( − b i ⋅ b j ) ( n i − n j ) 2 , n 0 = 0 , \left|\sum_{i=1}^{3}n_i b_i\right|^2=\sum_{0\le i<j\le3}(-b_i\cdot b_j)(n_i-n_j)^2,\qquad n_0=0, i = 1 ∑ 3 n i b i 2 = 0 ≤ i < j ≤ 3 ∑ ( − b i ⋅ b j ) ( n i − n j ) 2 , n 0 = 0 ,
gives
Gram ( b 1 , b 2 , b 3 ) = ( ∑ 0 ≤ a < b ≤ 3 ( − b a ⋅ b b ) ) G ( p ) , det ( D − 1 / 6 G 1 / 2 ) = 1. \operatorname{Gram}(b_1,b_2,b_3)=\left(\sum_{0\le a<b\le3}(-b_a\cdot b_b)\right)G(p),\qquad\det\left(D^{-1/6}G^{1/2}\right)=1. Gram ( b 1 , b 2 , b 3 ) = ( 0 ≤ a < b ≤ 3 ∑ ( − b a ⋅ b b ) ) G ( p ) , det ( D − 1/6 G 1/2 ) = 1.
Thus
Λ ( p ) = D ( p ) − 1 / 6 G ( p ) 1 / 2 Z 3 , p ∈ Δ , D ( p ) > 0 , (10) \Lambda(p)=D(p)^{-1/6}G(p)^{1/2}\mathbb{Z}^3,\qquad p\in\Delta,\quad D(p)>0, \tag*{(10)} Λ ( p ) = D ( p ) − 1/6 G ( p ) 1/2 Z 3 , p ∈ Δ , D ( p ) > 0 , (10)
parametrizes unit-covolume lattices up to orthogonal equivalence. For n 0 = 0 n_0=0 n 0 = 0 ,
n T G n = 0 ⟺ n i = n j whenever p i j > 0 , n^{T}Gn=0\Longleftrightarrow n_i=n_j\quad\text{whenever }p_{ij}>0, n T G n = 0 ⟺ n i = n j whenever p ij > 0 ,
and therefore
D ( p ) > 0 ⟺ ( I , { i j ∈ E : p i j > 0 } ) is connected . D(p)>0\Longleftrightarrow(I,\{ij\in E:p_{ij}>0\})\text{ is connected}. D ( p ) > 0 ⟺ ( I , { ij ∈ E : p ij > 0 }) is connected .
For ∅ ≠ S ⊊ I \varnothing\ne S\subsetneq I ∅ = S ⊊ I , set
W S ( p ) = ∑ i ∈ S ∑ j ∈ I ∖ S p i j , Y ( p ) = ∑ 0 ≤ i < j ≤ 3 [ − p i j d i j / 2 , p i j d i j / 2 ] . W_S(p)=\sum_{i\in S}\sum_{j\in I\setminus S}p_{ij},\qquad Y(p)=\sum_{0\le i<j\le3}[-p_{ij}d_{ij}/2,p_{ij}d_{ij}/2]. W S ( p ) = i ∈ S ∑ j ∈ I ∖ S ∑ p ij , Y ( p ) = 0 ≤ i < j ≤ 3 ∑ [ − p ij d ij /2 , p ij d ij /2 ] .
With y 0 = − y 1 − y 2 − y 3 y_0=-y_1-y_2-y_3 y 0 = − y 1 − y 2 − y 3 , we have
Y ( p ) = { y ∈ R 3 : ∑ i ∈ S y i ≤ 1 2 W S ( p ) ( ∅ ≠ S ⊊ I ) } , (11) Y(p)=\left\{y\in\mathbb{R}^3:\sum_{i\in S}y_i\le\frac{1}{2}W_S(p)\quad(\varnothing\ne S\subsetneq I)\right\}, \tag*{(11)} Y ( p ) = { y ∈ R 3 : i ∈ S ∑ y i ≤ 2 1 W S ( p ) ( ∅ = S ⊊ I ) } , (11)
V Λ ( p ) = D − 1 / 6 G − 1 / 2 Y ( p ) . (12) V_{\Lambda(p)}=D^{-1/6}G^{-1/2}Y(p). \tag*{(12)} V Λ ( p ) = D − 1/6 G − 1/2 Y ( p ) . (12)
In y y y coordinates, the Voronoi inequalities are
2 n ⋅ y ≤ ∑ 0 ≤ i < j ≤ 3 p i j ( n i − n j ) 2 , n ∈ Z 3 , n 0 = 0. 2n\cdot y\le\sum_{0\le i<j\le3}p_{ij}(n_i-n_j)^2,\qquad n\in\mathbb{Z}^3,\quad n_0=0. 2 n ⋅ y ≤ 0 ≤ i < j ≤ 3 ∑ p ij ( n i − n j ) 2 , n ∈ Z 3 , n 0 = 0.
For integral n n n ,
∣ n i − n j ∣ ≤ ( n i − n j ) 2 , n i = 1 S ( i ) − 1 S ( 0 ) ⟹ n T G n = W S . |n_i-n_j|\le(n_i-n_j)^2,\qquad n_i=\mathbf{1}_S(i)-\mathbf{1}_S(0)\Longrightarrow n^{T}Gn=W_S. ∣ n i − n j ∣ ≤ ( n i − n j ) 2 , n i = 1 S ( i ) − 1 S ( 0 ) ⟹ n T G n = W S .
The segment sum satisfies the Voronoi inequalities; cut indicators recover the constraints in (2.4). Conversely, for u 0 = 0 u_0=0 u 0 = 0 ,
2 u ⋅ y = 2 ∫ min 0 ≤ i ≤ 3 u i max 0 ≤ i ≤ 3 u i ∑ 0 ≤ i ≤ 3 u i > s y i d s ≤ ∫ min 0 ≤ i ≤ 3 u i max 0 ≤ i ≤ 3 u i W { i ∈ I : u i > s } d s = ∑ 0 ≤ i < j ≤ 3 p i j ∣ u i − u j ∣ . 2u\cdot y=2\int_{\min_{0\le i\le3}u_i}^{\max_{0\le i\le3}u_i}\sum_{\substack{0\le i\le3\\u_i>s}}y_i\,\mathrm{d}s\le\int_{\min_{0\le i\le3}u_i}^{\max_{0\le i\le3}u_i}W_{\{i\in I:u_i>s\}}\,\mathrm{d}s=\sum_{0\le i<j\le3}p_{ij}|u_i-u_j|. 2 u ⋅ y = 2 ∫ m i n 0 ≤ i ≤ 3 u i m a x 0 ≤ i ≤ 3 u i 0 ≤ i ≤ 3 u i > s ∑ y i d s ≤ ∫ m i n 0 ≤ i ≤ 3 u i m a x 0 ≤ i ≤ 3 u i W { i ∈ I : u i > s } d s = 0 ≤ i < j ≤ 3 ∑ p ij ∣ u i − u j ∣.
The last expression is twice the support function of the segment sum, proving (2.4)–(2.5).
By (12) and (2) ,
F ( p ) : = μ 1 ( V Λ ( p ) ) = D 1 / 3 inf f ∈ H 1 ( Y ( p ) ) ∖ { 0 } ∫ Y ( p ) f = 0 ∫ Y ( p ) ∇ f T G ∇ f ∫ Y ( p ) f 2 . (13) F(p) := \mu_1(V_{\Lambda(p)}) = D^{1/3}\inf_{\substack{f\in H^1(Y^{(p)})\setminus\{0\}\\ \int_{Y^{(p)}} f=0}} \frac{\int_{Y^{(p)}} \nabla f^T G \nabla f}{\int_{Y^{(p)}} f^2}. \tag*{(13)} F ( p ) := μ 1 ( V Λ ( p ) ) = D 1/3 f ∈ H 1 ( Y ( p ) ) ∖ { 0 } ∫ Y ( p ) f = 0 inf ∫ Y ( p ) f 2 ∫ Y ( p ) ∇ f T G ∇ f . (13)
In particular,
Y ( p ) = − Y ( p ) , f ( − y ) = − f ( y ) ⟹ ∫ Y ( p ) f = 0 , Y^{(p)}=-Y^{(p)},\qquad f(-y)=-f(y)\quad\Longrightarrow\quad\int_{Y^{(p)}}f=0, Y ( p ) = − Y ( p ) , f ( − y ) = − f ( y ) ⟹ ∫ Y ( p ) f = 0 ,
and, for s > 0 s>0 s > 0 ,
G ( s p ) = s G ( p ) , D ( s p ) = s 3 D ( p ) , Λ ( s p ) = Λ ( p ) , F ( s p ) = F ( p ) . G(sp)=sG(p),\qquad D(sp)=s^3D(p),\qquad\Lambda(sp)=\Lambda(p),\qquad F(sp)=F(p). G ( s p ) = s G ( p ) , D ( s p ) = s 3 D ( p ) , Λ ( s p ) = Λ ( p ) , F ( s p ) = F ( p ) .
At the equal-conorm point
p ∗ = ( 1 / 6 , … , 1 / 6 ) , G ( p ∗ ) = 1 6 ( 3 − 1 − 1 − 1 3 − 1 − 1 − 1 3 ) , D ( p ∗ ) = 2 27 . (14) p_*=(1/6,\ldots,1/6),\qquad G(p_*)=\frac{1}{6}\begin{pmatrix}3&-1&-1\\-1&3&-1\\-1&-1&3\end{pmatrix},\qquad D(p_*)=\frac{2}{27}. \tag*{(14)} p ∗ = ( 1/6 , … , 1/6 ) , G ( p ∗ ) = 6 1 3 − 1 − 1 − 1 3 − 1 − 1 − 1 3 , D ( p ∗ ) = 27 2 . (14)
The three permutations of ( 1 , − 1 , − 1 ) / 2 (1,-1,-1)/2 ( 1 , − 1 , − 1 ) /2 form a BCC basis with Gram matrix 3 2 G ( p ∗ ) \frac{3}{2}G(p_*) 2 3 G ( p ∗ ) ; hence Λ ( p ∗ ) = R B C C \Lambda(p_*)=R\mathrm{BCC} Λ ( p ∗ ) = R BCC for some R ∈ O ( 3 ) R\in O(3) R ∈ O ( 3 ) . Fix
T = 10837 1000 . (15) T=\frac{10837}{1000}. \tag*{(15)} T = 1000 10837 . (15)
The proof of Theorem 1.1 reduces to
F ( p ∗ ) > T , Section 3 , (16) F(p_*)>T,\qquad\text{Section 3}, \tag*{(16)} F ( p ∗ ) > T , Section 3 , (16)
0 < ∥ p − p ∗ ∥ 2 ≤ 73 6000 ⟹ F ( p ) < F ( p ∗ ) , Section 4 , (17) 0<\lVert p-p_*\rVert_2\leq\frac{73}{6000}\quad\Longrightarrow\quad F(p)<F(p_*),\qquad\text{Section 4}, \tag*{(17)} 0 < ∥ p − p ∗ ∥ 2 ≤ 6000 73 ⟹ F ( p ) < F ( p ∗ ) , Section 4 , (17)
∥ p − p ∗ ∥ 2 > 73 6000 ⟹ F ( p ) < T , Section 5 . (18) \lVert p-p_*\rVert_2>\frac{73}{6000}\quad\Longrightarrow\quad F(p)<T,\qquad\text{Section 5}. \tag*{(18)} ∥ p − p ∗ ∥ 2 > 6000 73 ⟹ F ( p ) < T , Section 5 . (18)
Write λ 1 ( Λ ) \lambda_1(\Lambda) λ 1 ( Λ ) for the shortest nonzero vector length, and let Λ ∗ = { w : w ⋅ Λ ⊂ Z } \Lambda^*=\{w:w\cdot\Lambda\subset\mathbb{Z}\} Λ ∗ = { w : w ⋅ Λ ⊂ Z } .
Lemma 2.1. For p ∈ Δ p\in\Delta p ∈ Δ with D ( p ) > 0 D(p)>0 D ( p ) > 0 ,
λ 1 ( Λ ( p ) ) 2 = D − 1 / 3 min ∅ ≠ S ⊊ I W S , λ 1 ( Λ ( p ) ∗ ) 2 = D − 2 / 3 min 0 ≤ i < j ≤ 3 C i j . (19) \lambda_1(\Lambda(p))^2=D^{-1/3}\min_{\varnothing\neq S\subsetneq I}W_S,\qquad\lambda_1(\Lambda(p)^*)^2=D^{-2/3}\min_{0\leq i<j\leq3}C_{ij}. \tag*{(19)} λ 1 ( Λ ( p ) ) 2 = D − 1/3 ∅ = S ⊊ I min W S , λ 1 ( Λ ( p ) ∗ ) 2 = D − 2/3 0 ≤ i < j ≤ 3 min C ij . (19)
Proof. For n ∈ Z 3 ∖ { 0 } n\in\mathbb{Z}^3\setminus\{0\} n ∈ Z 3 ∖ { 0 } , put
n 0 = 0 , S = { i ∈ I : n i > min 0 ≤ j ≤ 3 n j } , u i = 1 S ( i ) − 1 S ( 0 ) ( 1 ≤ i ≤ 3 ) . n_0=0,\qquad S=\left\{i\in I:n_i>\min_{0\leq j\leq3}n_j\right\},\qquad u_i=1_S(i)-1_S(0)\quad(1\leq i\leq3). n 0 = 0 , S = { i ∈ I : n i > 0 ≤ j ≤ 3 min n j } , u i = 1 S ( i ) − 1 S ( 0 ) ( 1 ≤ i ≤ 3 ) .
Then ∅ ≠ S ⊊ I \varnothing\neq S\subsetneq I ∅ = S ⊊ I and
n T G n ≥ W S ≥ min ∅ ≠ T ⊊ I W T , u T G u = W S . n^T G n\geq W_S\geq\min_{\varnothing\neq T\subsetneq I}W_T,\qquad u^T G u=W_S. n T G n ≥ W S ≥ ∅ = T ⊊ I min W T , u T G u = W S .
For the dual lattice, put
m ~ = ( − ∑ i = 1 3 m i , m 1 , m 2 , m 3 ) , S = { i ∈ I : m ~ i > 0 } , u i = 1 S ( i ) − 1 S ( 0 ) . \widetilde m=\left(-\sum_{i=1}^{3}m_i,m_1,m_2,m_3\right),\qquad S=\{i\in I:\widetilde m_i>0\},\qquad u_i=1_S(i)-1_S(0). m = ( − i = 1 ∑ 3 m i , m 1 , m 2 , m 3 ) , S = { i ∈ I : m i > 0 } , u i = 1 S ( i ) − 1 S ( 0 ) .
If m ∉ { 0 , ± d e : e ∈ E } m\notin\{0,\pm d_e:e\in E\} m ∈ / { 0 , ± d e : e ∈ E } , then
m ⋅ u = ∑ i ∈ S m ~ i ≥ 2 , 4 ≤ ( m ⋅ u ) 2 ≤ ( m T G − 1 m ) W S ≤ m T G − 1 m . m\cdot u=\sum_{i\in S}\widetilde m_i\geq2,\qquad4\leq(m\cdot u)^2\leq(m^T G^{-1}m)W_S\leq m^T G^{-1}m. m ⋅ u = i ∈ S ∑ m i ≥ 2 , 4 ≤ ( m ⋅ u ) 2 ≤ ( m T G − 1 m ) W S ≤ m T G − 1 m .
On the other hand,
∑ 0 ≤ i < j ≤ 3 p i j d i j T G − 1 d i j = tr ( G − 1 G ) = 3 , \sum_{0\leq i<j\leq3}p_{ij}d_{ij}^TG^{-1}d_{ij}=\operatorname{tr}(G^{-1}G)=3, 0 ≤ i < j ≤ 3 ∑ p ij d ij T G − 1 d ij = tr ( G − 1 G ) = 3 ,
so
min m ∈ Z 3 ∖ { 0 } m T G − 1 m = min 0 ≤ i < j ≤ 3 d i j T G − 1 d i j ≤ 3. \min_{m\in\mathbb{Z}^3\setminus\{0\}}m^TG^{-1}m=\min_{0\leq i<j\leq3}d_{ij}^TG^{-1}d_{ij}\leq3. m ∈ Z 3 ∖ { 0 } min m T G − 1 m = 0 ≤ i < j ≤ 3 min d ij T G − 1 d ij ≤ 3.
Since
Λ ( p ) ∗ = D 1 / 6 G − 1 / 2 Z 3 , G − 1 = D − 1 adj G , \Lambda(p)^* = D^{1/6}G^{-1/2}\mathbb{Z}^3,\qquad G^{-1}=D^{-1}\operatorname{adj}G, Λ ( p ) ∗ = D 1/6 G − 1/2 Z 3 , G − 1 = D − 1 adj G ,
the scaling factors give (19) . □
Fundamental-cell orthogonality and central symmetry give, for 0 ≠ w ∈ Λ ∗ 0\ne w\in\Lambda^* 0 = w ∈ Λ ∗ ,
∫ V Λ e 2 π i w ⋅ x = 0 , ∫ V Λ sin ( π w ⋅ x ) = 0 , \int_{V_\Lambda}e^{2\pi i w\cdot x}=0,\qquad\int_{V_\Lambda}\sin(\pi w\cdot x)=0, ∫ V Λ e 2 π i w ⋅ x = 0 , ∫ V Λ sin ( π w ⋅ x ) = 0 ,
∫ V Λ sin 2 ( π w ⋅ x ) = ∫ V Λ cos 2 ( π w ⋅ x ) = ∣ V Λ ∣ 2 . \int_{V_\Lambda}\sin^2(\pi w\cdot x)=\int_{V_\Lambda}\cos^2(\pi w\cdot x)=\frac{|V_\Lambda|}{2}. ∫ V Λ sin 2 ( π w ⋅ x ) = ∫ V Λ cos 2 ( π w ⋅ x ) = 2 ∣ V Λ ∣ .
Consequently,
∫ V Λ ∣ ∇ sin ( π w ⋅ x ) ∣ 2 ∫ V Λ sin 2 ( π w ⋅ x ) = π 2 ∣ w ∣ 2 , \frac{\int_{V_\Lambda}|\nabla\sin(\pi w\cdot x)|^2}{\int_{V_\Lambda}\sin^2(\pi w\cdot x)}=\pi^2|w|^2, ∫ V Λ sin 2 ( π w ⋅ x ) ∫ V Λ ∣∇ sin ( π w ⋅ x ) ∣ 2 = π 2 ∣ w ∣ 2 ,
and
μ 1 ( V Λ ) ≤ π 2 λ 1 ( Λ ∗ ) 2 , F ( p ) ≤ π 2 D − 2 / 3 min 0 ≤ i < j ≤ 3 C i j ≤ 3 π 2 D 1 / 3 . (20) \mu_1(V_\Lambda)\leq\pi^2\lambda_1(\Lambda^*)^2,\qquad F(p)\leq\pi^2D^{-2/3}\min_{0\leq i<j\leq3}C_{ij}\leq3\pi^2D^{1/3}. \tag*{(20)} μ 1 ( V Λ ) ≤ π 2 λ 1 ( Λ ∗ ) 2 , F ( p ) ≤ π 2 D − 2/3 0 ≤ i < j ≤ 3 min C ij ≤ 3 π 2 D 1/3 . (20)
In particular,
p ∈ Δ , D ( p ) ⟶ 0 ⟹ F ( p ) ⟶ 0. p\in\Delta,\qquad D(p)\longrightarrow0\quad\Longrightarrow\quad F(p)\longrightarrow0. p ∈ Δ , D ( p ) ⟶ 0 ⟹ F ( p ) ⟶ 0.
For a primitive v ∈ Λ v\in\Lambda v ∈ Λ , put
a = ∣ v ∣ , Γ = proj v ⊥ Λ , covol Γ = a − 1 . a=|v|,\qquad\Gamma=\operatorname{proj}_{v^\perp}\Lambda,\qquad\operatorname{covol}\Gamma=a^{-1}. a = ∣ v ∣ , Γ = proj v ⊥ Λ , covol Γ = a − 1 .
Let w ( y ) w(y) w ( y ) be the fiber length of V Λ V_\Lambda V Λ above y ∈ v ⊥ y\in v^\perp y ∈ v ⊥ . The bisectors of ± v \pm v ± v and the lattice tiling give
0 ≤ w ( y ) ≤ a , ∑ γ ∈ Γ w ( y + γ ) = a a.e. (21) 0\leq w(y)\leq a,\qquad\sum_{\gamma\in\Gamma}w(y+\gamma)=a\quad\text{a.e.} \tag*{(21)} 0 ≤ w ( y ) ≤ a , γ ∈ Γ ∑ w ( y + γ ) = a a.e. (21)
For a locally Lipschitz profile g g g with g ( 0 ) = 0 g(0)=0 g ( 0 ) = 0 , let
u i ( y ) = g ( ∣ y ∣ ) y i ∣ y ∣ ( i = 1 , 2 ) . u_i(y)=g(|y|)\frac{y_i}{|y|}\qquad(i=1,2). u i ( y ) = g ( ∣ y ∣ ) ∣ y ∣ y i ( i = 1 , 2 ) .
Then
∑ i = 1 2 u i ( y ) 2 = g ( ∣ y ∣ ) 2 , ∑ i = 1 2 ∣ ∇ u i ( y ) ∣ 2 = g ′ ( ∣ y ∣ ) 2 + g ( ∣ y ∣ ) 2 ∣ y ∣ 2 , \sum_{i=1}^{2}u_i(y)^2=g(|y|)^2,\qquad\sum_{i=1}^{2}|\nabla u_i(y)|^2=g'(|y|)^2+\frac{g(|y|)^2}{|y|^2}, i = 1 ∑ 2 u i ( y ) 2 = g ( ∣ y ∣ ) 2 , i = 1 ∑ 2 ∣∇ u i ( y ) ∣ 2 = g ′ ( ∣ y ∣ ) 2 + ∣ y ∣ 2 g ( ∣ y ∣ ) 2 ,
∫ R 2 u i ( y ) w ( y ) d y = 0 ( i = 1 , 2 ) , \int_{\mathbb{R}^2}u_i(y)w(y)\,dy=0\qquad(i=1,2), ∫ R 2 u i ( y ) w ( y ) d y = 0 ( i = 1 , 2 ) ,
since w ( − y ) = w ( y ) w(-y)=w(y) w ( − y ) = w ( y ) . Suppose r ↦ g ( r ) 2 r\mapsto g(r)^2 r ↦ g ( r ) 2 is nondecreasing, r ↦ g ′ ( r ) 2 + g ( r ) 2 / r 2 r\mapsto g'(r)^2+g(r)^2/r^2 r ↦ g ′ ( r ) 2 + g ( r ) 2 / r 2 is nonincreasing, and ∫ V Γ g ( ∣ y ∣ ) 2 d y > 0 \int_{V_\Gamma}g(|y|)^2\,dy>0 ∫ V Γ g ( ∣ y ∣ ) 2 d y > 0 . For y ∈ V Γ y\in V_\Gamma y ∈ V Γ , ∣ y + γ ∣ ≥ ∣ y ∣ |y+\gamma|\geq|y| ∣ y + γ ∣ ≥ ∣ y ∣ for every γ ∈ Γ \gamma\in\Gamma γ ∈ Γ ; hence
∑ γ ∈ Γ ( g ′ ( ∣ y + γ ∣ ) 2 + g ( ∣ y + γ ∣ ) 2 ∣ y + γ ∣ 2 ) w ( y + γ ) ≤ a ( g ′ ( ∣ y ∣ ) 2 + g ( ∣ y ∣ ) 2 ∣ y ∣ 2 ) , ∑ γ ∈ Γ g ( ∣ y + γ ∣ ) 2 w ( y + γ ) ≥ a g ( ∣ y ∣ ) 2 . \begin{aligned}
\sum_{\gamma\in\Gamma}\left(g'(|y+\gamma|)^2+\frac{g(|y+\gamma|)^2}{|y+\gamma|^2}\right)w(y+\gamma)
&\leq a\left(g'(|y|)^2+\frac{g(|y|)^2}{|y|^2}\right),\\
\sum_{\gamma\in\Gamma}g(|y+\gamma|)^2w(y+\gamma)&\geq ag(|y|)^2.
\end{aligned} γ ∈ Γ ∑ ( g ′ ( ∣ y + γ ∣ ) 2 + ∣ y + γ ∣ 2 g ( ∣ y + γ ∣ ) 2 ) w ( y + γ ) γ ∈ Γ ∑ g ( ∣ y + γ ∣ ) 2 w ( y + γ ) ≤ a ( g ′ ( ∣ y ∣ ) 2 + ∣ y ∣ 2 g ( ∣ y ∣ ) 2 ) , ≥ a g ( ∣ y ∣ ) 2 .
Rayleigh’s principle and (21) give
μ 1 ( V Λ ) ≤ min i = 1 , 2 ∫ R 2 u i 2 w > 0 ∫ R 2 ∣ ∇ u i ∣ 2 w ∫ R 2 u i 2 w ≤ ∑ i = 1 2 ∫ R 2 ∣ ∇ u i ∣ 2 w ∑ i = 1 2 ∫ R 2 u i 2 w = ∫ R 2 ( g ′ ( ∣ y ∣ ) 2 + g ( ∣ y ∣ ) 2 ∣ y ∣ 2 ) w ( y ) d y ∫ R 2 g ( ∣ y ∣ ) 2 w ( y ) d y ≤ ∫ V Γ ( g ′ ( ∣ y ∣ ) 2 + g ( ∣ y ∣ ) 2 ∣ y ∣ 2 ) d y ∫ V Γ g ( ∣ y ∣ ) 2 d y . (22) \mu_1(V_\Lambda) \le\min_{\substack{i=1,2\\ \int_{\mathbb{R}^2}u_i^2w>0}} \frac{\int_{\mathbb{R}^2}|\nabla u_i|^2w}{\int_{\mathbb{R}^2}u_i^2w}
\le\frac{\sum_{i=1}^{2}\int_{\mathbb{R}^2}|\nabla u_i|^2w}{\sum_{i=1}^{2}\int_{\mathbb{R}^2}u_i^2w}
= \frac{\int_{\mathbb{R}^2}\left(g'(|y|)^2+\frac{g(|y|)^2}{|y|^2}\right)w(y)\,dy}{\int_{\mathbb{R}^2}g(|y|)^2w(y)\,dy}
\le\frac{\int_{V_\Gamma}\left(g'(|y|)^2+\frac{g(|y|)^2}{|y|^2}\right)\,dy}{\int_{V_\Gamma}g(|y|)^2\,dy}. \tag*{(22)} μ 1 ( V Λ ) ≤ i = 1 , 2 ∫ R 2 u i 2 w > 0 min ∫ R 2 u i 2 w ∫ R 2 ∣∇ u i ∣ 2 w ≤ ∑ i = 1 2 ∫ R 2 u i 2 w ∑ i = 1 2 ∫ R 2 ∣∇ u i ∣ 2 w = ∫ R 2 g ( ∣ y ∣ ) 2 w ( y ) d y ∫ R 2 ( g ′ ( ∣ y ∣ ) 2 + ∣ y ∣ 2 g ( ∣ y ∣ ) 2 ) w ( y ) d y ≤ ∫ V Γ g ( ∣ y ∣ ) 2 d y ∫ V Γ ( g ′ ( ∣ y ∣ ) 2 + ∣ y ∣ 2 g ( ∣ y ∣ ) 2 ) d y . (22)
Lemma 2.2 (Planar disk overlap). For a planar lattice Voronoi polygon P P P and the centered regular hexagon H H H with ∣ H ∣ = ∣ P ∣ |H|=|P| ∣ H ∣ = ∣ P ∣ ,
∣ P ∩ B r ∣ ≤ ∣ H ∩ B r ∣ ( r > 0 ) . (23) |P\cap B_r|\le|H\cap B_r| \qquad(r>0). \tag*{(23)} ∣ P ∩ B r ∣ ≤ ∣ H ∩ B r ∣ ( r > 0 ) . (23)
For integrable radial f f f ,
d f ≥ 0 ⟹ ∫ P f ≥ ∫ H f , d f ≤ 0 ⟹ ∫ P f ≤ ∫ H f . df\ge0 \Longrightarrow\int_P f\ge\int_H f,\qquad df\le0 \Longrightarrow\int_P f\le\int_H f. df ≥ 0 ⟹ ∫ P f ≥ ∫ H f , df ≤ 0 ⟹ ∫ P f ≤ ∫ H f .
Proof. We use the planar moment comparison of [10 , 2 ] . A Voronoi facet has relative-interior midpoint v / 2 v/2 v /2 . If w ≡ v ( m o d 2 Γ ) w\equiv v\pmod{2\Gamma} w ≡ v ( mod 2Γ ) , w ≠ ± v w\ne\pm v w = ± v , and ∣ w ∣ ≤ ∣ v ∣ |w|\le|v| ∣ w ∣ ≤ ∣ v ∣ , then
v + w 2 ∈ Γ ∖ { 0 , v } , ∣ v + w 2 − v 2 ∣ = ∣ w ∣ 2 ≤ ∣ v ∣ 2 , \frac{v+w}{2}\in\Gamma\setminus\{0,v\},\qquad\left|\frac{v+w}{2}-\frac v2\right|=\frac{|w|}{2}\le\frac{|v|}{2}, 2 v + w ∈ Γ ∖ { 0 , v } , 2 v + w − 2 v = 2 ∣ w ∣ ≤ 2 ∣ v ∣ ,
contradicting the relative-interior Voronoi inequalities. In particular, w = 0 w=0 w = 0 excludes v ∈ 2 Γ v\in2\Gamma v ∈ 2Γ , and
# { facets of P } ≤ 2 ( # ( Γ / 2 Γ ) − 1 ) = 6. \#\{\text{facets of }P\}\le2\bigl(\#(\Gamma/2\Gamma)-1\bigr)=6. # { facets of P } ≤ 2 ( # ( Γ/2Γ ) − 1 ) = 6.
By scaling, take r = 1 r=1 r = 1 . A facet triangle of area A A A and half-angle θ \theta θ has altitude A cot θ \sqrt{A}\cot\theta A cot θ and disk overlap
H ( A , θ ) = ∫ 0 θ min { 1 , A cot θ sec 2 φ } d φ . H(A,\theta)=\int_0^\theta\min\{1,A\cot\theta\sec^2\varphi\}\,d\varphi. H ( A , θ ) = ∫ 0 θ min { 1 , A cot θ sec 2 φ } d φ .
Equivalently,
H ( A , θ ) = { A , A cot θ ≤ cos 2 θ , θ + A cot θ ( 1 − A cot θ ) − arccos A cot θ , cos 2 θ < A cot θ < 1 , θ , A cot θ ≥ 1. H(A,\theta)=
\begin{cases}
A, & A\cot\theta\le\cos^2\theta,\\
\theta+\sqrt{A\cot\theta(1-A\cot\theta)}-\arccos\sqrt{A\cot\theta}, & \cos^2\theta<A\cot\theta<1,\\
\theta, & A\cot\theta\ge1.
\end{cases} H ( A , θ ) = ⎩ ⎨ ⎧ A , θ + A cot θ ( 1 − A cot θ ) − arccos A cot θ , θ , A cot θ ≤ cos 2 θ , cos 2 θ < A cot θ < 1 , A cot θ ≥ 1.
In the intermediate region,
H A A = − cot 2 θ 2 ( A cot θ ) 3 / 2 1 − A cot θ < 0 , H_{AA}=-\frac{\cot^2\theta}{2(A\cot\theta)^{3/2}\sqrt{1-A\cot\theta}}<0, H AA = − 2 ( A cot θ ) 3/2 1 − A cot θ cot 2 θ < 0 ,
det D 2 H = 1 + cot 2 θ A cot θ ( ( 1 + cot 2 θ ) A cot θ − cot 2 θ ) > 0. \det D^2H=\frac{1+\cot^2\theta}{A\cot\theta}\left((1+\cot^2\theta)A\cot\theta-\cot^2\theta\right)>0. det D 2 H = A cot θ 1 + cot 2 θ ( ( 1 + cot 2 θ ) A cot θ − cot 2 θ ) > 0.
The interface gradients are ( 1 , 0 ) (1,0) ( 1 , 0 ) and ( 0 , 1 ) (0,1) ( 0 , 1 ) , so H H H is jointly concave. Include degenerate triangles as needed, using continuity at ( 0 , 0 ) (0,0) ( 0 , 0 ) . Since
∑ ν = 1 6 A ν = ∣ P ∣ , ∑ ν = 1 6 θ ν = π , \sum_{\nu=1}^{6}A_\nu=|P|,\qquad\sum_{\nu=1}^{6}\theta_\nu=\pi, ν = 1 ∑ 6 A ν = ∣ P ∣ , ν = 1 ∑ 6 θ ν = π ,
Jensen’s inequality gives
∣ P ∩ B 1 ∣ = ∑ ν = 1 6 H ( A ν , θ ν ) ≤ 6 H ( ∣ P ∣ / 6 , π / 6 ) = ∣ H ∩ B 1 ∣ . |P \cap B_1| = \sum_{\nu=1}^{6} H(A_\nu,\theta_\nu) \le6H(|P|/6,\pi/6) = |H \cap B_1|. ∣ P ∩ B 1 ∣ = ν = 1 ∑ 6 H ( A ν , θ ν ) ≤ 6 H ( ∣ P ∣/6 , π /6 ) = ∣ H ∩ B 1 ∣.
For a nondecreasing radial profile f f f , Stieltjes integration gives
∫ P f − ∫ H f = ∫ 0 ∞ ( ∣ H ∩ B r ∣ − ∣ P ∩ B r ∣ ) d f ( r ) ≥ 0. \int_P f-\int_H f=\int_0^\infty\left(|H\cap B_r|-|P\cap B_r|\right)\,\mathrm{d}f(r)\ge0. ∫ P f − ∫ H f = ∫ 0 ∞ ( ∣ H ∩ B r ∣ − ∣ P ∩ B r ∣ ) d f ( r ) ≥ 0.
For nonincreasing f f f , apply this identity to − f -f − f . □ \square □
Use the radial profile y ↦ g ( π a ∣ y ∣ ) y\mapsto g(\sqrt{\pi a}|y|) y ↦ g ( π a ∣ y ∣ ) , where
g ( s ) = s − 5 12 s 3 + 1 20 s 5 ( 0 ≤ s ≤ 1 ) , g ( s ) = 19 30 ( s ≥ 1 ) . (24) g(s)=s-\frac{5}{12}s^3+\frac{1}{20}s^5 \quad(0\le s\le1),\qquad g(s)=\frac{19}{30}\quad(s\ge1). \tag*{(24)} g ( s ) = s − 12 5 s 3 + 20 1 s 5 ( 0 ≤ s ≤ 1 ) , g ( s ) = 30 19 ( s ≥ 1 ) . (24)
On [ 0 , 1 ] [0,1] [ 0 , 1 ] ,
g ′ = ( 1 − s 2 ) ( 1 − s 2 / 4 ) ≥ 0 , g ′ ′ = s ( s 2 − 5 / 2 ) ≤ 0 , ( g / s ) ′ = s ( s 2 / 5 − 5 / 6 ) ≤ 0. g'=(1-s^2)(1-s^2/4)\ge0,\qquad g''=s(s^2-5/2)\le0,\qquad(g/s)'=s(s^2/5-5/6)\le0. g ′ = ( 1 − s 2 ) ( 1 − s 2 /4 ) ≥ 0 , g ′′ = s ( s 2 − 5/2 ) ≤ 0 , ( g / s ) ′ = s ( s 2 /5 − 5/6 ) ≤ 0.
Thus
( g 2 ) ′ ≥ 0 , d d s ( ( g ′ ( s ) ) 2 + g ( s ) 2 s 2 ) ≤ 0 , ( g ′ ( s ) ) 2 + g ( s ) 2 s 2 = 361 900 s 2 ( s ≥ 1 ) . (g^2)'\ge0,\qquad\frac{\mathrm{d}}{\mathrm{d}s}\left((g'(s))^2+\frac{g(s)^2}{s^2}\right)\le0,\qquad(g'(s))^2+\frac{g(s)^2}{s^2}=\frac{361}{900s^2}\quad(s\ge1). ( g 2 ) ′ ≥ 0 , d s d ( ( g ′ ( s ) ) 2 + s 2 g ( s ) 2 ) ≤ 0 , ( g ′ ( s ) ) 2 + s 2 g ( s ) 2 = 900 s 2 361 ( s ≥ 1 ) .
By (22) and Lemma 2.2,
μ 1 ( V Λ ) ≤ π a ∫ H ( g ′ ( ∣ y ∣ ) 2 + g ( ∣ y ∣ ) 2 / ∣ y ∣ 2 ) d y ∫ H g ( ∣ y ∣ ) 2 d y , \mu_1(V_\Lambda)\le\pi a\frac{\int_H\left(g'(|y|)^2+g(|y|)^2/|y|^2\right)\,\mathrm{d}y}{\int_H g(|y|)^2\,\mathrm{d}y}, μ 1 ( V Λ ) ≤ π a ∫ H g ( ∣ y ∣ ) 2 d y ∫ H ( g ′ ( ∣ y ∣ ) 2 + g ( ∣ y ∣ ) 2 /∣ y ∣ 2 ) d y ,
where H H H now has area π \pi π .
For 0 ≤ s ≤ 1 0\le s\le1 0 ≤ s ≤ 1 ,
g s = 19 60 ( 3 − s 2 ) + ( 1 − s 2 ) 2 20 , \frac{g}{s}=\frac{19}{60}(3-s^2)+\frac{(1-s^2)^2}{20}, s g = 60 19 ( 3 − s 2 ) + 20 ( 1 − s 2 ) 2 ,
g ′ ( s ) 2 + g ( s ) 2 s 2 − 361 900 ≥ 361 900 ( ( 1 − s 2 ) + ( 1 − s 2 ) 2 ) . g'(s)^2+\frac{g(s)^2}{s^2}-\frac{361}{900}\ge\frac{361}{900}\left((1-s^2)+(1-s^2)^2\right). g ′ ( s ) 2 + s 2 g ( s ) 2 − 900 361 ≥ 900 361 ( ( 1 − s 2 ) + ( 1 − s 2 ) 2 ) .
For s ≥ 1 s\ge1 s ≥ 1 ,
361 900 − g ′ ( s ) 2 − g ( s ) 2 s 2 = 361 900 s 2 − 1 s 2 ≥ 361 900 ( ( s 2 − 1 ) − ( s 2 − 1 ) 2 ) . \frac{361}{900}-g'(s)^2-\frac{g(s)^2}{s^2}=\frac{361}{900}\frac{s^2-1}{s^2}\ge\frac{361}{900}\left((s^2-1)-(s^2-1)^2\right). 900 361 − g ′ ( s ) 2 − s 2 g ( s ) 2 = 900 361 s 2 s 2 − 1 ≥ 900 361 ( ( s 2 − 1 ) − ( s 2 − 1 ) 2 ) .
Since ∣ B 1 ∣ = ∣ H ∣ |B_1|=|H| ∣ B 1 ∣ = ∣ H ∣ ,
∫ B 1 ( g ′ ( ∣ y ∣ ) 2 + g ( ∣ y ∣ ) 2 ∣ y ∣ 2 ) d y − ∫ H ( g ′ ( ∣ y ∣ ) 2 + g ( ∣ y ∣ ) 2 ∣ y ∣ 2 ) d y ≥ 361 900 ( ∫ H ( 3 ∣ y ∣ 2 − ∣ y ∣ 4 ) − ∫ B 1 ( 3 ∣ y ∣ 2 − ∣ y ∣ 4 ) ) = 361 π 900 ( 5 π 6 3 − 14 π 2 405 − 7 6 ) . \begin{aligned}
&\int_{B_1}\left(g'(|y|)^2+\frac{g(|y|)^2}{|y|^2}\right)\,\mathrm{d}y-\int_H\left(g'(|y|)^2+\frac{g(|y|)^2}{|y|^2}\right)\,\mathrm{d}y \\
&\ge\frac{361}{900}\left(\int_H(3|y|^2-|y|^4)-\int_{B_1}(3|y|^2-|y|^4)\right) \\
&=\frac{361\pi}{900}\left(\frac{5\pi}{6\sqrt{3}}-\frac{14\pi^2}{405}-\frac{7}{6}\right).
\end{aligned} ∫ B 1 ( g ′ ( ∣ y ∣ ) 2 + ∣ y ∣ 2 g ( ∣ y ∣ ) 2 ) d y − ∫ H ( g ′ ( ∣ y ∣ ) 2 + ∣ y ∣ 2 g ( ∣ y ∣ ) 2 ) d y ≥ 900 361 ( ∫ H ( 3∣ y ∣ 2 − ∣ y ∣ 4 ) − ∫ B 1 ( 3∣ y ∣ 2 − ∣ y ∣ 4 ) ) = 900 361 π ( 6 3 5 π − 405 14 π 2 − 6 7 ) .
With N 0 = { 0 , 1 , 2 , … } \mathbb{N}_0=\{0,1,2,\ldots\} N 0 = { 0 , 1 , 2 , … } , the barycentric coordinates on a facet triangle △ \triangle △ give
∣ y ∣ 2 = 2 π 3 3 ( λ 2 + λ μ + μ 2 ) , 1 ∣ △ ∣ ∫ △ λ i μ j = 2 i ! j ! ( i + j + 2 ) ! ( i , j ∈ N 0 ) . |y|^2=\frac{2\pi}{3\sqrt{3}}(\lambda^2+\lambda\mu+\mu^2),\qquad\frac{1}{|\triangle|}\int_\triangle\lambda^i\mu^j=\frac{2i!j!}{(i+j+2)!}\quad(i,j\in\mathbb{N}_0). ∣ y ∣ 2 = 3 3 2 π ( λ 2 + λ μ + μ 2 ) , ∣△∣ 1 ∫ △ λ i μ j = ( i + j + 2 )! 2 i ! j ! ( i , j ∈ N 0 ) .
Consequently,
1 π ∫ H ∣ y ∣ 2 = 5 π 18 3 , 1 π ∫ H ∣ y ∣ 4 = 14 π 2 405 , ∫ B 1 ∣ y ∣ 2 k = π k + 1 ( k ∈ N 0 ) . \frac{1}{\pi}\int_H |y|^2=\frac{5\pi}{18\sqrt{3}},\qquad\frac{1}{\pi}\int_H |y|^4=\frac{14\pi^2}{405},\qquad\int_{B_1}|y|^{2k}=\frac{\pi}{k+1}\quad(k\in\mathbb{N}_0). π 1 ∫ H ∣ y ∣ 2 = 18 3 5 π , π 1 ∫ H ∣ y ∣ 4 = 405 14 π 2 , ∫ B 1 ∣ y ∣ 2 k = k + 1 π ( k ∈ N 0 ) .
The disk integrals are
∫ 0 1 ( s g ′ ( s ) 2 + g ( s ) 2 s ) d s = 6469 13500 , ∫ 0 1 s g ( s ) 2 d s = 1357 9600 . \int_0^1\left(sg'(s)^2+\frac{g(s)^2}{s}\right)\,\mathrm{d}s=\frac{6469}{13500},\qquad\int_0^1 sg(s)^2\,\mathrm{d}s=\frac{1357}{9600}. ∫ 0 1 ( s g ′ ( s ) 2 + s g ( s ) 2 ) d s = 13500 6469 , ∫ 0 1 s g ( s ) 2 d s = 9600 1357 .
Lemma 2.2 gives ∫ H g 2 ≥ 1357 π / 4800 \int_H g^2\ge1357\pi/4800 ∫ H g 2 ≥ 1357 π /4800 , and hence
μ 1 ( V Λ ) ≤ a ( 34232 π 6785 + 80864 π 3 1648755 − 14440 π 2 12213 3 ) < 5317 500 a . (25) \mu_1(V_\Lambda)\le a\left(\frac{34232\pi}{6785}+\frac{80864\pi^3}{1648755}-\frac{14440\pi^2}{12213\sqrt{3}}\right)<\frac{5317}{500}a. \tag*{(25)} μ 1 ( V Λ ) ≤ a ( 6785 34232 π + 1648755 80864 π 3 − 12213 3 14440 π 2 ) < 500 5317 a . (25)
For a = λ 1 ( Λ ( p ) ) a=\lambda_1(\Lambda(p)) a = λ 1 ( Λ ( p )) , Lemma 2.1 yields
F ( p ) ≤ 5317 500 D − 1 / 6 ( min ∅ ≠ S ⊊ I W S ) 1 / 2 . (26) F(p)\le\frac{5317}{500}D^{-1/6}\left(\min_{\varnothing\ne S\subsetneq I}W_S\right)^{1/2}. \tag*{(26)} F ( p ) ≤ 500 5317 D − 1/6 ( ∅ = S ⊊ I min W S ) 1/2 . (26)
The rational constants also satisfy
132379 100000 π 6 < T 3 , 28 25 ( 5317 500 ) 6 < T 6 . (27) \frac{132379}{100000}\pi^6<T^3,\qquad\frac{28}{25}\left(\frac{5317}{500}\right)^6<T^6. \tag*{(27)} 100000 132379 π 6 < T 3 , 25 28 ( 500 5317 ) 6 < T 6 . (27)
The spectrum of the BCC cell
Parity decomposition identifies the three axial first modes of the BCC cell; a mixed-problem residual estimate gives (3.9).
Let
W = [ − 1 , 1 ] 3 ∩ { ∣ x 1 ∣ + ∣ x 2 ∣ + ∣ x 3 ∣ ≤ 3 / 2 } , Ω + = W ∩ [ 0 , 1 ] 3 . (28) W=[-1,1]^3\cap\{|x_1|+|x_2|+|x_3|\le3/2\},\qquad\Omega_+=W\cap[0,1]^3. \tag*{(28)} W = [ − 1 , 1 ] 3 ∩ { ∣ x 1 ∣ + ∣ x 2 ∣ + ∣ x 3 ∣ ≤ 3/2 } , Ω + = W ∩ [ 0 , 1 ] 3 . (28)
The bisectors of ± 2 1 / 2 e i \pm2^{1/2}\mathbf{e}_i ± 2 1/2 e i and 2 − 2 / 3 ( ± 1 , ± 1 , ± 1 ) 2^{-2/3}(\pm1,\pm1,\pm1) 2 − 2/3 ( ± 1 , ± 1 , ± 1 ) give
V B C C ⊂ 2 − 2 / 3 W , ∣ Ω + ∣ = 1 2 , ∣ V B C C ∣ = ∣ 2 − 2 / 3 W ∣ = 1. V_{\mathrm{BCC}}\subset2^{-2/3}W,\qquad|\Omega_+|=\frac{1}{2},\qquad|V_{\mathrm{BCC}}|=|2^{-2/3}W|=1. V BCC ⊂ 2 − 2/3 W , ∣ Ω + ∣ = 2 1 , ∣ V BCC ∣ = ∣ 2 − 2/3 W ∣ = 1.
Hence V B C C = 2 − 2 / 3 W V_{\mathrm{BCC}}=2^{-2/3}W V BCC = 2 − 2/3 W . The vertices of W W W are the signed permutations of ( 1 , 1 / 2 , 0 ) (1,1/2,0) ( 1 , 1/2 , 0 ) ; its square and regular hexagonal facets have common edge length 1 / 2 1/\sqrt{2} 1/ 2 .
Opposite parities in any coordinate give
∫ W u v = 0 , ∫ W ∇ u ⋅ ∇ v = 0. \int_W uv=0,\qquad\int_W\nabla u\cdot\nabla v=0. ∫ W uv = 0 , ∫ W ∇ u ⋅ ∇ v = 0.
Coordinate reflections therefore split the Neumann spectrum into eight orthogonal sectors. On Ω + \Omega_+ Ω + , odd and even parity in x i x_i x i give Dirichlet and Neumann conditions on x i = 0 x_i=0 x i = 0 , respectively. Let λ , λ 2 \lambda,\lambda_2 λ , λ 2 be the first two eigenvalues with Dirichlet conditions on x 1 = 0 x_1=0 x 1 = 0 and Neumann conditions elsewhere.
Lemma 3.1.
λ ≥ π 2 4 , λ 2 > π 2 . (29) \lambda\ge\frac{\pi^2}{4},\qquad\lambda_2>\pi^2. \tag*{(29)} λ ≥ 4 π 2 , λ 2 > π 2 . (29)
Proof. For f ∈ H 1 ( 0 , 1 ) f\in H^1(0,1) f ∈ H 1 ( 0 , 1 ) with f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 , the Dirichlet–Neumann inequality follows from
∫ 0 1 ( ∣ f ′ ∣ 2 − π 2 4 ∣ f ∣ 2 ) d x = ∫ 0 1 sin 2 π x 2 ∣ ( f sin ( π x / 2 ) ) ′ ∣ 2 d x ≥ 0. \int_0^1\left(|f'|^2-\frac{\pi^2}{4}|f|^2\right)\,\mathrm{d}x
=\int_0^1\sin^2\frac{\pi x}{2}\left|\left(\frac{f}{\sin(\pi x/2)}\right)'\right|^2\,\mathrm{d}x\ge0. ∫ 0 1 ( ∣ f ′ ∣ 2 − 4 π 2 ∣ f ∣ 2 ) d x = ∫ 0 1 sin 2 2 π x ( sin ( π x /2 ) f ) ′ 2 d x ≥ 0.
Rescaling on each x 1 x_1 x 1 -fibre gives
∫ Ω + ∣ ∂ x 1 u ∣ 2 ≥ ∫ Ω + π 2 ∣ u ∣ 2 4 min ( 1 , 3 / 2 − x 2 − x 3 ) 2 ≥ π 2 4 ∫ Ω + ∣ u ∣ 2 . \int_{\Omega_+}|\partial_{x_1}u|^2\ge\int_{\Omega_+}\frac{\pi^2|u|^2}{4\min(1,3/2-x_2-x_3)^2}\ge\frac{\pi^2}{4}\int_{\Omega_+}|u|^2. ∫ Ω + ∣ ∂ x 1 u ∣ 2 ≥ ∫ Ω + 4 min ( 1 , 3/2 − x 2 − x 3 ) 2 π 2 ∣ u ∣ 2 ≥ 4 π 2 ∫ Ω + ∣ u ∣ 2 .
To bound the second eigenvalue, decompose under x 2 ↔ x 3 x_2\leftrightarrow x_3 x 2 ↔ x 3 and restrict to x 2 ≥ x 3 x_2\ge x_3 x 2 ≥ x 3 . In the odd sector, u ∣ x 2 = x 3 = 0 u|_{x_2=x_3}=0 u ∣ x 2 = x 3 = 0 . Slicing in x 1 x_1 x 1 and ( x 2 − x 3 ) / 2 (x_2-x_3)/\sqrt{2} ( x 2 − x 3 ) / 2 , with s = x 2 + x 3 s=x_2+x_3 s = x 2 + x 3 , gives
∫ ∣ ∇ u ∣ 2 ∫ ∣ u ∣ 2 ≥ inf 0 < s < 3 / 2 π 2 { 1 4 min ( 1 , 3 / 2 − s ) 2 + 1 2 min ( s , 2 − s ) 2 } ≥ 5 4 π 2 . \frac{\int|\nabla u|^2}{\int|u|^2}\ge\inf_{0<s<3/2}\pi^2\left\{\frac{1}{4\min(1,3/2-s)^2}+\frac{1}{2\min(s,2-s)^2}\right\}\ge\frac{5}{4}\pi^2. ∫ ∣ u ∣ 2 ∫ ∣∇ u ∣ 2 ≥ 0 < s < 3/2 inf π 2 { 4 min ( 1 , 3/2 − s ) 2 1 + 2 min ( s , 2 − s ) 2 1 } ≥ 4 5 π 2 .
Indeed, Hölder’s inequality yields
1 4 min ( 1 , 3 / 2 − s ) 2 + 1 2 min ( s , 2 − s ) 2 ≥ { 9 / 4 , 0 < s ≤ 1 / 2 , ( 1 + 2 1 / 3 ) 3 / 9 > 5 / 4 , 1 / 2 < s < 1 , 3 / 2 , 1 ≤ s < 3 / 2. \frac{1}{4\min(1,3/2-s)^2}+\frac{1}{2\min(s,2-s)^2}\ge
\begin{cases}
9/4, & 0<s\le1/2,\\
(1+2^{1/3})^3/9>5/4, & 1/2<s<1,\\
3/2, & 1\le s<3/2.
\end{cases} 4 min ( 1 , 3/2 − s ) 2 1 + 2 min ( s , 2 − s ) 2 1 ≥ ⎩ ⎨ ⎧ 9/4 , ( 1 + 2 1/3 ) 3 /9 > 5/4 , 3/2 , 0 < s ≤ 1/2 , 1/2 < s < 1 , 1 ≤ s < 3/2.
In the even sector, insert a Neumann cut at x 1 = 1 / 2 x_1=1/2 x 1 = 1/2 and put
Ω L = Ω + ∩ { x 2 ≥ x 3 , x 1 < 1 / 2 } , Ω R = Ω + ∩ { x 2 ≥ x 3 , x 1 > 1 / 2 } . \Omega_L=\Omega_+\cap\{x_2\ge x_3,\ x_1<1/2\},\qquad
\Omega_R=\Omega_+\cap\{x_2\ge x_3,\ x_1>1/2\}. Ω L = Ω + ∩ { x 2 ≥ x 3 , x 1 < 1/2 } , Ω R = Ω + ∩ { x 2 ≥ x 3 , x 1 > 1/2 } .
On Ω L \Omega_L Ω L , Dirichlet–Neumann fibres have length at most 1 / 2 1/2 1/2 , with strict inequality on a set of positive transverse measure. Positivity of the ground state gives λ 1 ( Ω L ) > π 2 \lambda_1(\Omega_L)>\pi^2 λ 1 ( Ω L ) > π 2 . On Ω R \Omega_R Ω R , set
( x 2 , x 3 ) = ( 3 / 2 − x 1 ) ( Y , Z ) , T 0 = { 0 ≤ Z ≤ Y , Y + Z ≤ 1 } , ∣ T 0 ∣ = 1 4 . (x_2,x_3)=(3/2-x_1)(Y,Z),\qquad
T_0=\{0\le Z\le Y,\ Y+Z\le1\},\qquad|T_0|=\frac14. ( x 2 , x 3 ) = ( 3/2 − x 1 ) ( Y , Z ) , T 0 = { 0 ≤ Z ≤ Y , Y + Z ≤ 1 } , ∣ T 0 ∣ = 4 1 .
The triangle and its reflection across Z = 0 Z=0 Z = 0 form a square of side 1 / 2 1/\sqrt{2} 1/ 2 . Thus
d V = ( 3 / 2 − x 1 ) 2 d x 1 d Y d Z , ∫ T 0 ∣ f − 4 ∣ T 0 ∣ ∫ T 0 f ∣ 2 ≤ 1 2 π 2 ∫ T 0 ∣ ∇ f ∣ 2 . dV=(3/2-x_1)^2\,dx_1\,dY\,dZ,\qquad
\int_{T_0}\left|f-\frac{4}{|T_0|}\int_{T_0}f\right|^2\le\frac{1}{2\pi^2}\int_{T_0}|\nabla f|^2. d V = ( 3/2 − x 1 ) 2 d x 1 d Y d Z , ∫ T 0 f − ∣ T 0 ∣ 4 ∫ T 0 f 2 ≤ 2 π 2 1 ∫ T 0 ∣∇ f ∣ 2 .
For the first nonconstant weighted Neumann mode,
− ( ( 3 / 2 − x 1 ) 2 f ′ ) ′ = ν ( 3 / 2 − x 1 ) 2 f , f ′ ( 1 / 2 ) = f ′ ( 1 ) = 0 , -((3/2-x_1)^2f')'=\nu(3/2-x_1)^2f,\qquad f'(1/2)=f'(1)=0, − (( 3/2 − x 1 ) 2 f ′ ) ′ = ν ( 3/2 − x 1 ) 2 f , f ′ ( 1/2 ) = f ′ ( 1 ) = 0 ,
and hence
ν ∫ 1 / 2 1 ( 3 / 2 − x 1 ) 2 ( f ′ ) 2 = ∫ 1 / 2 1 ( 3 / 2 − x 1 ) 2 ( f ′ ′ ) 2 + 2 ∫ 1 / 2 1 ( f ′ ) 2 ≥ ∫ 1 / 2 1 ( ( ( 3 / 2 − x 1 ) f ′ ) ′ ) 2 ≥ 4 π 2 ∫ 1 / 2 1 ( 3 / 2 − x 1 ) 2 ( f ′ ) 2 . \begin{aligned}
\nu\int_{1/2}^{1}(3/2-x_1)^2(f')^2
=\int_{1/2}^{1}(3/2-x_1)^2(f'')^2+2\int_{1/2}^{1}(f')^2 \\
&\ge\int_{1/2}^{1}\left(((3/2-x_1)f')'\right)^2
\ge4\pi^2\int_{1/2}^{1}(3/2-x_1)^2(f')^2.
\end{aligned} ν ∫ 1/2 1 ( 3/2 − x 1 ) 2 ( f ′ ) 2 = ∫ 1/2 1 ( 3/2 − x 1 ) 2 ( f ′′ ) 2 + 2 ∫ 1/2 1 ( f ′ ) 2 ≥ ∫ 1/2 1 ( (( 3/2 − x 1 ) f ′ ) ′ ) 2 ≥ 4 π 2 ∫ 1/2 1 ( 3/2 − x 1 ) 2 ( f ′ ) 2 .
Integration by parts gives
∫ 1 / 2 1 ( ( ( 3 / 2 − x 1 ) f ′ ) ′ ) 2 = ∫ 1 / 2 1 ( 3 / 2 − x 1 ) 2 ( f ′ ′ ) 2 . \int_{1/2}^{1}\left(((3/2-x_1)f')'\right)^2
=\int_{1/2}^{1}(3/2-x_1)^2(f'')^2. ∫ 1/2 1 ( (( 3/2 − x 1 ) f ′ ) ′ ) 2 = ∫ 1/2 1 ( 3/2 − x 1 ) 2 ( f ′′ ) 2 .
Set
m ( x 1 ) = 4 ∫ 0 1 ∫ 0 min ( Y , 1 − Y ) u ( x 1 , ( 3 / 2 − x 1 ) Y , ( 3 / 2 − x 1 ) Z ) d Z d Y . m(x_1)=4\int_0^1\int_0^{\min(Y,1-Y)}
u(x_1,(3/2-x_1)Y,(3/2-x_1)Z)\,dZ\,dY. m ( x 1 ) = 4 ∫ 0 1 ∫ 0 m i n ( Y , 1 − Y ) u ( x 1 , ( 3/2 − x 1 ) Y , ( 3/2 − x 1 ) Z ) d Z d Y .
Then
4 ∫ T 0 ( Y 2 + Z 2 ) d Y d Z = 1 3 , 4\int_{T_0}(Y^2+Z^2)\,dY\,dZ=\frac13, 4 ∫ T 0 ( Y 2 + Z 2 ) d Y d Z = 3 1 ,
m ′ ( x 1 ) = 4 ∫ T 0 ( u x 1 − Y u x 2 − Z u x 3 ) d Y d Z , m'(x_1)=4\int_{T_0}(u_{x_1}-Yu_{x_2}-Zu_{x_3})\,dY\,dZ, m ′ ( x 1 ) = 4 ∫ T 0 ( u x 1 − Y u x 2 − Z u x 3 ) d Y d Z ,
∣ m ′ ( x 1 ) ∣ 2 ≤ 8 ∫ T 0 u x 1 2 d Y d Z + 8 3 ∫ T 0 ( u x 2 2 + u x 3 2 ) d Y d Z , |m'(x_1)|^2\le8\int_{T_0}u_{x_1}^2\,dY\,dZ+\frac83\int_{T_0}(u_{x_2}^2+u_{x_3}^2)\,dY\,dZ, ∣ m ′ ( x 1 ) ∣ 2 ≤ 8 ∫ T 0 u x 1 2 d Y d Z + 3 8 ∫ T 0 ( u x 2 2 + u x 3 2 ) d Y d Z ,
where the derivatives are evaluated at ( x 1 , ( 3 / 2 − x 1 ) Y , ( 3 / 2 − x 1 ) Z ) (x_1,(3/2-x_1)Y,(3/2-x_1)Z) ( x 1 , ( 3/2 − x 1 ) Y , ( 3/2 − x 1 ) Z ) . Since
∣ Ω R ∣ = 7 96 , 96 7 ∫ Ω R u = 24 7 ∫ 1 / 2 1 ( 3 / 2 − x 1 ) 2 m ( x 1 ) d x 1 , |\Omega_R|=\frac{7}{96},\qquad
\frac{96}{7}\int_{\Omega_R}u=\frac{24}{7}\int_{1/2}^{1}(3/2-x_1)^2m(x_1)\,dx_1, ∣ Ω R ∣ = 96 7 , 7 96 ∫ Ω R u = 7 24 ∫ 1/2 1 ( 3/2 − x 1 ) 2 m ( x 1 ) d x 1 ,
1 2 ≤ 3 / 2 − x 1 ≤ 1 , \frac12\le3/2-x_1\le1, 2 1 ≤ 3/2 − x 1 ≤ 1 ,
the transverse and weighted inequalities give
∫ Ω R ∣ u − m ∣ 2 ≤ 1 2 π 2 ∫ Ω R ( 3 / 2 − x 1 ) 2 ( u x 2 2 + u x 3 2 ) , \int_{\Omega_R} |u-m|^2 \le\frac{1}{2\pi^2}\int_{\Omega_R} (3/2-x_1)^2(u_{x_2}^2+u_{x_3}^2), ∫ Ω R ∣ u − m ∣ 2 ≤ 2 π 2 1 ∫ Ω R ( 3/2 − x 1 ) 2 ( u x 2 2 + u x 3 2 ) ,
∣ ∫ Ω R ( m − 96 7 ∫ Ω R u ) ∣ 2 ≤ 1 16 π 2 ∫ 1 / 2 1 ( 3 / 2 − x 1 ) 2 ∣ m ′ ∣ 2 d x 1 ≤ 1 2 π 2 ∫ Ω R u x 1 2 + 1 6 π 2 ∫ Ω R ( u x 2 2 + u x 3 2 ) . \begin{aligned}
\left|\int_{\Omega_R}\left(m-\frac{96}{7}\int_{\Omega_R}u\right)\right|^2 \le\frac{1}{16\pi^2}\int_{1/2}^{1}(3/2-x_1)^2|m'|^2\,dx_1 \\
&\le\frac{1}{2\pi^2}\int_{\Omega_R}u_{x_1}^2+\frac{1}{6\pi^2}\int_{\Omega_R}(u_{x_2}^2+u_{x_3}^2).
\end{aligned} ∫ Ω R ( m − 7 96 ∫ Ω R u ) 2 ≤ 16 π 2 1 ∫ 1/2 1 ( 3/2 − x 1 ) 2 ∣ m ′ ∣ 2 d x 1 ≤ 2 π 2 1 ∫ Ω R u x 1 2 + 6 π 2 1 ∫ Ω R ( u x 2 2 + u x 3 2 ) .
Thus
∫ Ω R ∣ u − 96 7 ∫ Ω R u ∣ 2 = ∫ Ω R ∣ u − m ∣ 2 + ∫ Ω R ∣ m − 96 7 ∫ Ω R u ∣ 2 ≤ 2 3 π 2 ∫ Ω R ∣ ∇ u ∣ 2 . \begin{aligned}
\int_{\Omega_R}\left|u-\frac{96}{7}\int_{\Omega_R}u\right|^2
=\int_{\Omega_R}|u-m|^2+\int_{\Omega_R}\left|m-\frac{96}{7}\int_{\Omega_R}u\right|^2 \\
&\le\frac{2}{3\pi^2}\int_{\Omega_R}|\nabla u|^2.
\end{aligned} ∫ Ω R u − 7 96 ∫ Ω R u 2 = ∫ Ω R ∣ u − m ∣ 2 + ∫ Ω R m − 7 96 ∫ Ω R u 2 ≤ 3 π 2 2 ∫ Ω R ∣∇ u ∣ 2 .
Neumann bracketing therefore gives
λ 1 ( Ω R ) = 0 , λ 2 ( Ω R ) ≥ 3 2 π 2 , \lambda_1(\Omega_R)=0,\qquad\lambda_2(\Omega_R)\ge\frac{3}{2}\pi^2, λ 1 ( Ω R ) = 0 , λ 2 ( Ω R ) ≥ 2 3 π 2 ,
λ 2 e v e n ≥ min { λ 1 ( Ω L ) , λ 2 ( Ω R ) } > π 2 . \lambda_2^{\mathrm{even}}\ge\min\{\lambda_1(\Omega_L),\lambda_2(\Omega_R)\}>\pi^2. λ 2 even ≥ min { λ 1 ( Ω L ) , λ 2 ( Ω R )} > π 2 .
The odd-sector bound completes (29) .
For complementary bounds using H ( div ) H(\operatorname{div}) H ( div ) fluxes, see Šebestová and Vejchodský [16 ] and Vejchodský [18 ] . Assume
v ∈ H 1 ( Ω + ) , ∥ v ∥ L 2 ( Ω + ) = 1 , v ∣ x 1 = 0 = 0 , v\in H^1(\Omega_+),\qquad\|v\|_{L^2(\Omega_+)}=1,\qquad v|_{x_1=0}=0, v ∈ H 1 ( Ω + ) , ∥ v ∥ L 2 ( Ω + ) = 1 , v ∣ x 1 = 0 = 0 ,
q ∈ H ( div ; Ω + ) , ( q ⋅ n ) ∣ ∂ Ω + ∖ { x 1 = 0 } = 0. q\in H(\operatorname{div};\Omega_+),\qquad(q\cdot n)|_{\partial\Omega_+\setminus\{x_1=0\}}=0. q ∈ H ( div ; Ω + ) , ( q ⋅ n ) ∣ ∂ Ω + ∖ { x 1 = 0 } = 0.
Lemma 3.2. If ∥ ∇ v ∥ 2 2 < γ ≤ λ 2 \|\nabla v\|_2^2<\gamma\le\lambda_2 ∥∇ v ∥ 2 2 < γ ≤ λ 2 , then
0 ≤ ∥ ∇ v ∥ 2 2 − λ ≤ γ γ − ∥ ∇ v ∥ 2 2 ( 4001 4000 ∥ ∇ v − q ∥ 2 2 + 16004 9 ∥ div q + ∥ ∇ v ∥ 2 2 v ∥ 2 2 ) . (30) 0\le\|\nabla v\|_2^2-\lambda\le\frac{\gamma}{\gamma-\|\nabla v\|_2^2}\left(\frac{4001}{4000}\|\nabla v-q\|_2^2+\frac{16004}{9}\|\operatorname{div}q+\|\nabla v\|_2^2v\|_2^2\right). \tag*{(30)} 0 ≤ ∥∇ v ∥ 2 2 − λ ≤ γ − ∥∇ v ∥ 2 2 γ ( 4000 4001 ∥∇ v − q ∥ 2 2 + 9 16004 ∥ div q + ∥∇ v ∥ 2 2 v ∥ 2 2 ) . (30)
Proof. For every mixed test function w w w , integration by parts and Lemma 3.1 give
∣ ∫ Ω + ( ∇ v ⋅ ∇ w − ∥ ∇ v ∥ 2 2 v w ) ∣ = ∣ ∫ Ω + ( ( ∇ v − q ) ⋅ ∇ w − ( div q + ∥ ∇ v ∥ 2 2 v ) w ) ∣ ≤ ( ∥ ∇ v − q ∥ 2 + 2 3 ∥ div q + ∥ ∇ v ∥ 2 2 v ∥ 2 ) ∥ ∇ w ∥ 2 . \begin{aligned}
\left|\int_{\Omega_+}(\nabla v\cdot\nabla w-\|\nabla v\|_2^2vw)\right|
=\left|\int_{\Omega_+}((\nabla v-q)\cdot\nabla w-(\operatorname{div}q+\|\nabla v\|_2^2v)w)\right| \\
&\le\left(\|\nabla v-q\|_2+\frac{2}{3}\|\operatorname{div}q+\|\nabla v\|_2^2v\|_2\right)\|\nabla w\|_2.
\end{aligned} ∫ Ω + ( ∇ v ⋅ ∇ w − ∥∇ v ∥ 2 2 v w ) = ∫ Ω + (( ∇ v − q ) ⋅ ∇ w − ( div q + ∥∇ v ∥ 2 2 v ) w ) ≤ ( ∥∇ v − q ∥ 2 + 3 2 ∥ div q + ∥∇ v ∥ 2 2 v ∥ 2 ) ∥∇ w ∥ 2 .
In an orthonormal mixed eigenbasis,
v = ∑ j = 1 ∞ v j ϕ j , ∑ j = 1 ∞ ∣ v j ∣ 2 = 1 , ∥ ∇ v ∥ 2 2 = ∑ j = 1 ∞ λ j ∣ v j ∣ 2 , v=\sum_{j=1}^{\infty}v_j\phi_j,\qquad\sum_{j=1}^{\infty}|v_j|^2=1,\qquad\|\nabla v\|_2^2=\sum_{j=1}^{\infty}\lambda_j|v_j|^2, v = j = 1 ∑ ∞ v j ϕ j , j = 1 ∑ ∞ ∣ v j ∣ 2 = 1 , ∥∇ v ∥ 2 2 = j = 1 ∑ ∞ λ j ∣ v j ∣ 2 ,
( γ − ∥ ∇ v ∥ 2 2 ) ( 1 − ∣ v 1 ∣ 2 ) ≤ ( ∥ ∇ v ∥ 2 2 − λ ) ∣ v 1 ∣ 2 , ∣ v 1 ∣ 2 ≥ γ − ∥ ∇ v ∥ 2 2 γ − λ . (\gamma-\|\nabla v\|_2^2)(1-|v_1|^2)\le(\|\nabla v\|_2^2-\lambda)|v_1|^2,\qquad|v_1|^2\ge\frac{\gamma-\|\nabla v\|_2^2}{\gamma-\lambda}. ( γ − ∥∇ v ∥ 2 2 ) ( 1 − ∣ v 1 ∣ 2 ) ≤ ( ∥∇ v ∥ 2 2 − λ ) ∣ v 1 ∣ 2 , ∣ v 1 ∣ 2 ≥ γ − λ γ − ∥∇ v ∥ 2 2 .
Duality and Cauchy–Schwarz yield
4001 4000 ∥ ∇ v − q ∥ 2 2 + 16004 9 ∥ div q + ∥ ∇ v ∥ 2 2 v ∥ 2 2 ≥ ∑ j = 1 ∞ ( λ j − ∥ ∇ v ∥ 2 2 ) 2 λ j ∣ v j ∣ 2 ≥ ( ∥ ∇ v ∥ 2 2 − λ ) ∣ v 1 ∣ 2 ( ∥ ∇ v ∥ 2 2 − λ λ + 1 − ∥ ∇ v ∥ 2 2 γ ) ≥ ∥ ∇ v ∥ 2 2 − λ λ ∥ ∇ v ∥ 2 2 ( γ − ∥ ∇ v ∥ 2 2 ) γ ≥ ( ∥ ∇ v ∥ 2 2 − λ ) γ − ∥ ∇ v ∥ 2 2 γ . \begin{aligned}
\frac{4001}{4000}\lVert\nabla v-q\rVert_{2}^{2}+\frac{16004}{9}\lVert\operatorname{div}q+\lVert\nabla v\rVert_{2}^{2}v\rVert_{2}^{2} \\
&\ge\sum_{j=1}^{\infty}\frac{(\lambda_j-\lVert\nabla v\rVert_{2}^{2})^{2}}{\lambda_j}|v_j|^{2} \\
&\ge(\lVert\nabla v\rVert_{2}^{2}-\lambda)|v_1|^{2}\left(\frac{\lVert\nabla v\rVert_{2}^{2}-\lambda}{\lambda}+1-\frac{\lVert\nabla v\rVert_{2}^{2}}{\gamma}\right) \\
&\ge\frac{\lVert\nabla v\rVert_{2}^{2}-\lambda}{\lambda}\frac{\lVert\nabla v\rVert_{2}^{2}(\gamma-\lVert\nabla v\rVert_{2}^{2})}{\gamma} \\
&\ge(\lVert\nabla v\rVert_{2}^{2}-\lambda)\frac{\gamma-\lVert\nabla v\rVert_{2}^{2}}{\gamma}.
\end{aligned} 4000 4001 ∥ ∇ v − q ∥ 2 2 + 9 16004 ∥ div q + ∥ ∇ v ∥ 2 2 v ∥ 2 2 ≥ j = 1 ∑ ∞ λ j ( λ j − ∥ ∇ v ∥ 2 2 ) 2 ∣ v j ∣ 2 ≥ (∥ ∇ v ∥ 2 2 − λ ) ∣ v 1 ∣ 2 ( λ ∥ ∇ v ∥ 2 2 − λ + 1 − γ ∥ ∇ v ∥ 2 2 ) ≥ λ ∥ ∇ v ∥ 2 2 − λ γ ∥ ∇ v ∥ 2 2 ( γ − ∥ ∇ v ∥ 2 2 ) ≥ (∥ ∇ v ∥ 2 2 − λ ) γ γ − ∥ ∇ v ∥ 2 2 .
This proves (3.3). □
Set b = ( 1 / 2 , 1 / 2 , 1 / 2 ) b=(1/2,1/2,1/2) b = ( 1/2 , 1/2 , 1/2 ) and v i j = e i + e j / 2 v_{ij}=e_i+e_j/2 v ij = e i + e j /2 . Coning the non-coordinate faces of Ω + \Omega_+ Ω + to 0 0 0 gives the nine tetrahedra of T + \mathcal{T}_+ T + :
( 0 , e i , v i j , v i k ) , ( 0 , b , v i j , v i k ) ( i = 1 , 2 , 3 , j < k , { i , j , k } = { 1 , 2 , 3 } ) , (0,e_i,v_{ij},v_{ik}),\quad(0,b,v_{ij},v_{ik})\quad(i=1,2,3,\ j<k,\ \{i,j,k\}=\{1,2,3\}), ( 0 , e i , v ij , v ik ) , ( 0 , b , v ij , v ik ) ( i = 1 , 2 , 3 , j < k , { i , j , k } = { 1 , 2 , 3 }) ,
( 0 , b , v i j , v j i ) ( 1 ≤ i < j ≤ 3 ) . (31) (0,b,v_{ij},v_{ji})\quad(1\le i<j\le3). \tag*{(31)} ( 0 , b , v ij , v j i ) ( 1 ≤ i < j ≤ 3 ) . (31)
Order the first six by increasing i i i , with the square cone first for each i i i , and the last three lexicographically in ( i , j ) (i,j) ( i , j ) . For barycentric coordinates ζ \zeta ζ on P P P , use
α ∈ N 0 4 , ∣ α ∣ = ∑ i = 0 3 α i , α ! = ∏ i = 0 3 α i ! . \alpha\in\mathbb{N}_{0}^{4},\qquad|\alpha|=\sum_{i=0}^{3}\alpha_i,\qquad\alpha!=\prod_{i=0}^{3}\alpha_i!. α ∈ N 0 4 , ∣ α ∣ = i = 0 ∑ 3 α i , α ! = i = 0 ∏ 3 α i ! .
Then
∫ P ζ α = 6 ∣ P ∣ ∏ i = 0 3 α i ! ( 3 + ∣ α ∣ ) ! , α ∈ N 0 4 . (32) \int_P\zeta^\alpha=6|P|\frac{\prod_{i=0}^{3}\alpha_i!}{(3+|\alpha|)!},\qquad\alpha\in\mathbb{N}_{0}^{4}. \tag*{(32)} ∫ P ζ α = 6∣ P ∣ ( 3 + ∣ α ∣ )! ∏ i = 0 3 α i ! , α ∈ N 0 4 . (32)
For the trial and flux construction, use Bernstein coefficients:
f ∣ P = ∑ α ∈ N 0 4 ∣ α ∣ = m f P , α m ! α ! ζ α . f|_P=\sum_{\substack{\alpha\in\mathbb{N}_{0}^{4}\\|\alpha|=m}}f_{P,\alpha}\frac{m!}{\alpha!}\zeta^\alpha. f ∣ P = α ∈ N 0 4 ∣ α ∣ = m ∑ f P , α α ! m ! ζ α .
For f , h f,h f , h of degrees m , n m,n m , n , (3.5) gives
∫ P f h = 6 ∣ P ∣ m ! n ! ( m + n + 3 ) ! ∑ α ∈ N 0 4 ∣ α ∣ = m ∑ β ∈ N 0 4 ∣ β ∣ = n f α h β ∏ i = 0 3 ( α i + β i α i ) . (33) \int_P fh=\frac{6|P|m!n!}{(m+n+3)!}\sum_{\substack{\alpha\in\mathbb{N}_{0}^{4}\\|\alpha|=m}}\sum_{\substack{\beta\in\mathbb{N}_{0}^{4}\\|\beta|=n}}f_\alpha h_\beta\prod_{i=0}^{3}\binom{\alpha_i+\beta_i}{\alpha_i}. \tag*{(33)} ∫ P f h = ( m + n + 3 )! 6∣ P ∣ m ! n ! α ∈ N 0 4 ∣ α ∣ = m ∑ β ∈ N 0 4 ∣ β ∣ = n ∑ f α h β i = 0 ∏ 3 ( α i α i + β i ) . (33)
In particular,
( f , h ) Ω + = ∑ P ∈ T + ∫ P f h . (f,h)_{\Omega_+}=\sum_{P\in\mathcal{T}_+}\int_P fh. ( f , h ) Ω + = P ∈ T + ∑ ∫ P f h .
Numerical Lemma 1.
∃ v ∈ H 1 ( Ω + ) ∩ C ( Ω + ‾ ) , ∃ q ∈ H ( div ; Ω + ) , \exists v\in H^{1}(\Omega_+)\cap C(\overline{\Omega_+}),\qquad\exists q\in H(\operatorname{div};\Omega_+), ∃ v ∈ H 1 ( Ω + ) ∩ C ( Ω + ) , ∃ q ∈ H ( div ; Ω + ) , ∥ v ∥ 2 = 1 , v ∣ x 1 = 0 = 0 , ( q ⋅ n ) ∣ ∂ Ω + ∖ { x 1 = 0 } = 0 , \lVert v\rVert_2=1,\qquad v|_{x_1=0}=0,\qquad(q\cdot n)|_{\partial\Omega_+\setminus\{x_1=0\}}=0, ∥ v ∥ 2 = 1 , v ∣ x 1 = 0 = 0 , ( q ⋅ n ) ∣ ∂ Ω + ∖ { x 1 = 0 } = 0 , v ∣ P , q i ∣ P ∈ R [ x 1 , x 2 , x 3 ] , deg ( v ∣ P ) ≤ 9 , deg ( q i ∣ P ) ≤ 10 ( P ∈ T + , 1 ≤ i ≤ 3 ) , v|_P,\ q_i|_P\in\mathbb{R}[x_1,x_2,x_3],\qquad\deg(v|_P)\le9,\qquad\deg(q_i|_P)\le10\quad(P\in\mathcal{T}_+,\ 1\le i\le3), v ∣ P , q i ∣ P ∈ R [ x 1 , x 2 , x 3 ] , deg ( v ∣ P ) ≤ 9 , deg ( q i ∣ P ) ≤ 10 ( P ∈ T + , 1 ≤ i ≤ 3 ) ,
43008469 10 7 < ∥ ∇ v ∥ 2 2 < 43008470 10 7 , ∥ ∇ v − q ∥ 2 2 < 9002 10 9 , ∥ div q + ∥ ∇ v ∥ 2 v ∥ 2 2 < 23 10 13 . (34) \frac{43008469}{10^{7}} < \lVert\nabla v\rVert_{2}^{2} < \frac{43008470}{10^{7}},\\
\lVert\nabla v-q\rVert_{2}^{2} < \frac{9002}{10^{9}},\\
\left\lVert\operatorname{div}q+\lVert\nabla v\rVert_{2}v\right\rVert_{2}^{2} < \frac{23}{10^{13}}. \tag*{(34)} 1 0 7 43008469 < ∥ ∇ v ∥ 2 2 < 1 0 7 43008470 , ∥ ∇ v − q ∥ 2 2 < 1 0 9 9002 , ∥ div q + ∥ ∇ v ∥ 2 v ∥ 2 2 < 1 0 13 23 . (34)
The Appendix gives the construction and finite rational evaluations. Taking γ = 49 / 5 \gamma=49/5 γ = 49/5 in Lemma 3.2 and using (3.2) and (34) gives
43008308 10 7 < λ < 43008470 10 7 . (35) \frac{43008308}{10^{7}} < \lambda< \frac{43008470}{10^{7}}. \tag*{(35)} 1 0 7 43008308 < λ < 1 0 7 43008470 . (35)
Proposition 3.3.
μ 1 ( W ) = λ , ker ( − Δ W N − λ ) = span { u 1 , u 2 , u 3 } , \mu_{1}(W)=\lambda,\qquad\ker(-\Delta_{W}^{N}-\lambda)=\operatorname{span}\{u_{1},u_{2},u_{3}\}, μ 1 ( W ) = λ , ker ( − Δ W N − λ ) = span { u 1 , u 2 , u 3 } , ∫ W u i u j = δ i j , u i ( x − 2 x j e j ) = ( − 1 ) δ i j u i ( x ) , \int_{W}u_{i}u_{j}=\delta_{ij},\qquad u_{i}(x-2x_{j}e_{j})=(-1)^{\delta_{ij}}u_{i}(x), ∫ W u i u j = δ ij , u i ( x − 2 x j e j ) = ( − 1 ) δ ij u i ( x ) , u i ( x ) > 0 ( x ∈ int W , x i > 0 ) , 1 ≤ i , j ≤ 3. u_{i}(x)>0\quad(x\in\operatorname{int}W,\ x_{i}>0),\qquad1\leq i,j\leq3. u i ( x ) > 0 ( x ∈ int W , x i > 0 ) , 1 ≤ i , j ≤ 3. μ 1 ( V B C C ) = 2 4 / 3 λ , 10.837414 < μ 1 ( V B C C ) < 10.837456. (36) \mu_{1}(V_{\mathrm{BCC}})=2^{4/3}\lambda,\qquad10.837414<\mu_{1}(V_{\mathrm{BCC}})<10.837456. \tag*{(36)} μ 1 ( V BCC ) = 2 4/3 λ , 10.837414 < μ 1 ( V BCC ) < 10.837456. (36)
Proof. By (35) , λ < π 2 / 2 \lambda<\pi^{2}/2 λ < π 2 /2 . For u u u odd in at least two coordinates, slicing gives
∫ W ∣ ∇ u ∣ 2 ≥ π 2 2 ∫ W ∣ u ∣ 2 . \int_{W}|\nabla u|^{2}\geq\frac{\pi^{2}}{2}\int_{W}|u|^{2}. ∫ W ∣∇ u ∣ 2 ≥ 2 π 2 ∫ W ∣ u ∣ 2 .
Let u u u be a mean-zero all-even eigenfunction with ν < π 2 / 2 \nu<\pi^{2}/2 ν < π 2 /2 . For a coordinate transposition τ \tau τ ,
Δ ( u ∘ τ ) = ( Δ u ) ∘ τ , − Δ ( u − u ∘ τ ) = ν ( u − u ∘ τ ) . \Delta(u\circ\tau)=(\Delta u)\circ\tau,\qquad-\Delta(u-u\circ\tau)=\nu(u-u\circ\tau). Δ ( u ∘ τ ) = ( Δ u ) ∘ τ , − Δ ( u − u ∘ τ ) = ν ( u − u ∘ τ ) .
The difference is odd in the corresponding difference coordinate, whose half-fibres have length at most 1 / 2 1/\sqrt{2} 1/ 2 . Hence
u − u ∘ τ ≠ 0 ⟹ ν = ∫ Ω + ∣ ∇ ( u − u ∘ τ ) ∣ 2 ∫ Ω + ∣ u − u ∘ τ ∣ 2 ≥ π 2 2 . u-u\circ\tau\ne0\quad\Longrightarrow\quad
\nu=\frac{\int_{\Omega_{+}}|\nabla(u-u\circ\tau)|^{2}}{\int_{\Omega_{+}}|u-u\circ\tau|^{2}}\geq\frac{\pi^{2}}{2}. u − u ∘ τ = 0 ⟹ ν = ∫ Ω + ∣ u − u ∘ τ ∣ 2 ∫ Ω + ∣∇ ( u − u ∘ τ ) ∣ 2 ≥ 2 π 2 .
Thus u = u ∘ τ u=u\circ\tau u = u ∘ τ for every coordinate transposition, and
Vert ( Ω + ∩ { x 1 ≥ x 2 ≥ x 3 } ) = { 0 , ( 1 , 0 , 0 ) , ( 3 / 4 , 3 / 4 , 0 ) , ( 1 / 2 , 1 / 2 , 1 / 2 ) , ( 1 , 1 / 2 , 0 ) , ( 1 , 1 / 4 , 1 / 4 ) } . \operatorname{Vert}(\Omega_{+}\cap\{x_{1}\geq x_{2}\geq x_{3}\})
=\{0,\ (1,0,0),\ (3/4,3/4,0),\ (1/2,1/2,1/2),\ (1,1/2,0),\ (1,1/4,1/4)\}. Vert ( Ω + ∩ { x 1 ≥ x 2 ≥ x 3 }) = { 0 , ( 1 , 0 , 0 ) , ( 3/4 , 3/4 , 0 ) , ( 1/2 , 1/2 , 1/2 ) , ( 1 , 1/2 , 0 ) , ( 1 , 1/4 , 1/4 )} .
The Payne–Weinberger inequality [15 ] gives
diam ( Ω + ∩ { x 1 ≥ x 2 ≥ x 3 } ) = 5 2 , \operatorname{diam}(\Omega_{+}\cap\{x_{1}\geq x_{2}\geq x_{3}\})=\frac{\sqrt{5}}{2}, diam ( Ω + ∩ { x 1 ≥ x 2 ≥ x 3 }) = 2 5 ,
∫ Ω + ∩ { x 1 ≥ x 2 ≥ x 3 } u = 1 6 ∫ Ω + u = 0 , \int_{\Omega_{+}\cap\{x_{1}\geq x_{2}\geq x_{3}\}}u=\frac{1}{6}\int_{\Omega_{+}}u=0, ∫ Ω + ∩ { x 1 ≥ x 2 ≥ x 3 } u = 6 1 ∫ Ω + u = 0 ,
ν = ∫ Ω + ∩ { x 1 ≥ x 2 ≥ x 3 } ∣ ∇ u ∣ 2 ∫ Ω + ∩ { x 1 ≥ x 2 ≥ x 3 } ∣ u ∣ 2 ≥ 4 π 2 5 > π 2 2 . \nu=\frac{\int_{\Omega_{+}\cap\{x_{1}\geq x_{2}\geq x_{3}\}}|\nabla u|^{2}}{\int_{\Omega_{+}\cap\{x_{1}\geq x_{2}\geq x_{3}\}}|u|^{2}}
\geq\frac{4\pi^{2}}{5}>\frac{\pi^{2}}{2}. ν = ∫ Ω + ∩ { x 1 ≥ x 2 ≥ x 3 } ∣ u ∣ 2 ∫ Ω + ∩ { x 1 ≥ x 2 ≥ x 3 } ∣∇ u ∣ 2 ≥ 5 4 π 2 > 2 π 2 .
Simplicity and positivity of the mixed ground state give
μ 1 ( W ) = λ , ker ( − Δ W N − λ ) = span { u 1 , u 2 , u 3 } , μ 1 ( V B C C ) = 2 4 / 3 λ . \mu_{1}(W)=\lambda,\qquad\ker(-\Delta_{W}^{N}-\lambda)=\operatorname{span}\{u_{1},u_{2},u_{3}\},\qquad\mu_{1}(V_{\mathrm{BCC}})=2^{4/3}\lambda. μ 1 ( W ) = λ , ker ( − Δ W N − λ ) = span { u 1 , u 2 , u 3 } , μ 1 ( V BCC ) = 2 4/3 λ .
Scaling (35) gives (36) and μ 1 ( V B C C ) > T \mu_{1}(V_{\mathrm{BCC}})>T μ 1 ( V BCC ) > T . □
The ten-place numerical value in (1.7) was computed using conforming high-order approximations on geometrically graded meshes and checked under polynomial enrichment, mesh refinement, and increased quadrature order. The proof uses the rigorous enclosure (3.8).
Strict comparison near BCC
We prove strict local maximality of BCC by transporting the axial trial space and controlling its linear splitting and higher-order terms.
Proposition 4.1.
0 < ∥ p − p ∗ ∥ 2 ≤ 73 6000 ⟹ F ( p ) < F ( p ∗ ) . 0 < \lVert p-p_* \rVert_2 \le\frac{73}{6000} \quad\Longrightarrow\quad F(p)<F(p_*). 0 < ∥ p − p ∗ ∥ 2 ≤ 6000 73 ⟹ F ( p ) < F ( p ∗ ) .
For the cell W W W of volume 4 in § 3 , fix
( b 0 , b 1 , b 2 , b 3 ) = ( ( 1 , 1 , 1 ) , ( 1 , − 1 , − 1 ) , ( − 1 , 1 , − 1 ) , ( − 1 , − 1 , 1 ) ) . (b_0,b_1,b_2,b_3)=((1,1,1),(1,-1,-1),(-1,1,-1),(-1,-1,1)). ( b 0 , b 1 , b 2 , b 3 ) = (( 1 , 1 , 1 ) , ( 1 , − 1 , − 1 ) , ( − 1 , 1 , − 1 ) , ( − 1 , − 1 , 1 )) .
The superbase metric is I + K I+K I + K , where
K = 1 16 ∑ 0 ≤ i < j ≤ 3 ( 6 p i j − 1 ) ( b i − b j ) ( b i − b j ) T , tr K = 0 , K=\frac{1}{16}\sum_{0\le i<j\le3}(6p_{ij}-1)(b_i-b_j)(b_i-b_j)^T,\qquad\operatorname{tr}K=0, K = 16 1 0 ≤ i < j ≤ 3 ∑ ( 6 p ij − 1 ) ( b i − b j ) ( b i − b j ) T , tr K = 0 ,
36 ∥ p − p ∗ ∥ 2 2 = 8 ∥ diag K ∥ F 2 + 4 ∥ off K ∥ F 2 . (37) 36\lVert p-p_* \rVert_2^2=8\lVert\operatorname{diag}K\rVert_F^2+4\lVert\operatorname{off}K\rVert_F^2. \tag*{(37)} 36 ∥ p − p ∗ ∥ 2 2 = 8 ∥ diag K ∥ F 2 + 4 ∥ off K ∥ F 2 . (37)
∑ i = 0 3 b i b i T = 4 I , b i T K b j = 1 − 6 p i j ( i ≠ j ) , \sum_{i=0}^{3}b_i b_i^T=4I,\qquad b_i^T Kb_j=1-6p_{ij}\quad(i\ne j), i = 0 ∑ 3 b i b i T = 4 I , b i T K b j = 1 − 6 p ij ( i = j ) ,
and hence
16 ∥ K ∥ F 2 = ∑ i = 0 3 ∑ j = 0 3 ( b i T K b j ) 2 = 72 ∥ p − p ∗ ∥ 2 2 + 8 ∥ off K ∥ F 2 , 16\lVert K\rVert_F^2=\sum_{i=0}^{3}\sum_{j=0}^{3}(b_i^T Kb_j)^2=72\lVert p-p_*\rVert_2^2+8\lVert\operatorname{off}K\rVert_F^2, 16 ∥ K ∥ F 2 = i = 0 ∑ 3 j = 0 ∑ 3 ( b i T K b j ) 2 = 72 ∥ p − p ∗ ∥ 2 2 + 8 ∥ off K ∥ F 2 ,
3 2 ∥ p − p ∗ ∥ 2 ≤ ∥ K ∥ F ≤ 3 ∥ p − p ∗ ∥ 2 . \frac{3}{\sqrt{2}}\lVert p-p_*\rVert_2\le\lVert K\rVert_F\le3\lVert p-p_*\rVert_2. 2 3 ∥ p − p ∗ ∥ 2 ≤ ∥ K ∥ F ≤ 3 ∥ p − p ∗ ∥ 2 .
In the hyperplane ∑ i = 0 3 y i = 0 \sum_{i=0}^{3}y_i=0 ∑ i = 0 3 y i = 0 , use x = 1 4 ∑ i = 0 3 y i b i x=\frac{1}{4}\sum_{i=0}^{3}y_i b_i x = 4 1 ∑ i = 0 3 y i b i . A vertex indexed by a strict order σ \sigma σ is
x σ ( p ) = 3 4 ∑ 0 ≤ i < j ≤ 3 p i j sgn ( σ i − σ j ) ( b i − b j ) . (38) x_\sigma(p)=\frac{3}{4}\sum_{0\le i<j\le3}p_{ij}\operatorname{sgn}(\sigma_i-\sigma_j)(b_i-b_j). \tag*{(38)} x σ ( p ) = 4 3 0 ≤ i < j ≤ 3 ∑ p ij sgn ( σ i − σ j ) ( b i − b j ) . (38)
Since 6 p i j ≥ 1 − 6 ∥ p − p ∗ ∥ 2 > 0 6p_{ij}\ge1-6\lVert p-p_*\rVert_2>0 6 p ij ≥ 1 − 6 ∥ p − p ∗ ∥ 2 > 0 , the face incidences persist. Triangulate each face from its vertex average and cone the resulting triangles to 0; matching the vertices and face averages of the resulting 72 tetrahedra defines an odd piecewise linear map with matrix I + E I+E I + E on each tetrahedron.
For the fan (31) of § 3 , the vertex displacements satisfy
δ v i j = K i i e i + ( K i j − K i i ) e j + ( K i k + K j k ) e k , δ e i = K e i , δ b = K b . (39) \delta v_{ij}=K_{ii}e_i+(K_{ij}-K_{ii})e_j+(K_{ik}+K_{jk})e_k,\qquad\delta e_i=Ke_i,\qquad\delta b=Kb. \tag*{(39)} δ v ij = K ii e i + ( K ij − K ii ) e j + ( K ik + K j k ) e k , δ e i = K e i , δ b = K b . (39)
Thus, on T = ( 0 , c , v , w ) T=(0,c,v,w) T = ( 0 , c , v , w ) ,
E T ( K ) = [ δ c , δ v , δ w ] [ c , v , w ] − 1 . (40) E_T(K)=[\delta c,\delta v,\delta w][c,v,w]^{-1}. \tag*{(40)} E T ( K ) = [ δ c , δ v , δ w ] [ c , v , w ] − 1 . (40)
For a coordinate reflection S S S , the matrix on S T ST S T is S E T ( S K S ) S SE_T(SKS)S S E T ( S K S ) S . We claim
∥ E ∥ F ≤ 6 ∥ p − p ∗ ∥ 2 , \lVert E\rVert_F\le6\lVert p-p_*\rVert_2, ∥ E ∥ F ≤ 6 ∥ p − p ∗ ∥ 2 ,
∣ tr E s q ∣ 2 ≤ 27 ∥ p − p ∗ ∥ 2 2 , ∣ tr E h e x ∣ 2 ≤ 12 ∥ p − p ∗ ∥ 2 2 . (41) \lvert\operatorname{tr}E_{\mathrm{sq}}\rvert^2\le27\lVert p-p_*\rVert_2^2,\qquad\lvert\operatorname{tr}E_{\mathrm{hex}}\rvert^2\le12\lVert p-p_*\rVert_2^2. \tag*{(41)} ∣ tr E sq ∣ 2 ≤ 27 ∥ p − p ∗ ∥ 2 2 , ∣ tr E hex ∣ 2 ≤ 12 ∥ p − p ∗ ∥ 2 2 . (41)
The active vectors at v i j v_{ij} v ij are 2 e i 2e_i 2 e i and e i + e j ± e k e_i+e_j\pm e_k e i + e j ± e k . Differentiating their bisector equations v ⋅ x = v T ( I + K ) v / 2 v\cdot x=v^T(I+K)v/2 v ⋅ x = v T ( I + K ) v /2 gives
2 e i ⋅ δ v i j = 2 K i i , ( e i + e j ± e k ) ⋅ δ v i j = K i j ± ( K i k + K j k ) . 2e_i\cdot\delta v_{ij}=2K_{ii},\qquad(e_i+e_j\pm e_k)\cdot\delta v_{ij}=K_{ij}\pm(K_{ik}+K_{jk}). 2 e i ⋅ δ v ij = 2 K ii , ( e i + e j ± e k ) ⋅ δ v ij = K ij ± ( K ik + K j k ) .
These equations give (4.3); face averaging gives δ e i = K e i \delta e_i=Ke_i δ e i = K e i , δ b = K b \delta b=Kb δ b = K b . Since $\operatorname{tr}K=0,
36 ∥ p − p ∗ ∥ 2 2 = 12 K i i 2 + 4 ( K j j − K k k ) 2 + 8 ( K i j 2 + K i k 2 + K j k 2 ) . 36\lVert p-p_*\rVert_2^2=12K_{ii}^2+4(K_{jj}-K_{kk})^2+8(K_{ij}^2+K_{ik}^2+K_{jk}^2). 36 ∥ p − p ∗ ∥ 2 2 = 12 K ii 2 + 4 ( K j j − K k k ) 2 + 8 ( K ij 2 + K ik 2 + K j k 2 ) .
The square cones give
E s q = ( K i i 0 0 K i j − 2 K i i 2 K j k K i k 2 K j k − 2 K i i ) . E_{\mathrm{sq}}=\begin{pmatrix} K_{ii} & 0 & 0 \\ K_{ij} & -2K_{ii} & 2K_{jk} \\ K_{ik} & 2K_{jk} & -2K_{ii} \end{pmatrix}. E sq = K ii K ij K ik 0 − 2 K ii 2 K j k 0 2 K j k − 2 K ii .
In the basis e i , ( e j + e k ) / 2 , ( e j − e k ) / 2 e_i,(e_j+e_k)/\sqrt{2},(e_j-e_k)/\sqrt{2} e i , ( e j + e k ) / 2 , ( e j − e k ) / 2 , the two hexagon edge pairs ( v i j , v i k ) (v_{ij},v_{ik}) ( v ij , v ik ) and ( v j k , v k j ) (v_{jk},v_{kj}) ( v j k , v k j ) give
E h e x = ( K i i − ( K i j + K i k ) / 3 4 ( K i j + K i k ) / ( 3 2 ) 0 ( − K i i + K i j + K i k + 2 K j k / 3 ) / 2 2 K j k / 3 0 ( K i j − K i k − ( K j j − K k k ) / 3 ) / 2 2 ( K j j − K k k ) / 3 − 2 ( K i i + K j k ) ) . E_{\mathrm{hex}}=\begin{pmatrix} K_{ii}-(K_{ij}+K_{ik})/3 & 4(K_{ij}+K_{ik})/(3\sqrt{2}) & 0 \\ (-K_{ii}+K_{ij}+K_{ik}+2K_{jk}/3)/\sqrt{2} & 2K_{jk}/3 & 0 \\ (K_{ij}-K_{ik}-(K_{jj}-K_{kk})/3)/\sqrt{2} & 2(K_{jj}-K_{kk})/3 & -2(K_{ii}+K_{jk}) \end{pmatrix}. E hex = K ii − ( K ij + K ik ) /3 ( − K ii + K ij + K ik + 2 K j k /3 ) / 2 ( K ij − K ik − ( K j j − K k k ) /3 ) / 2 4 ( K ij + K ik ) / ( 3 2 ) 2 K j k /3 2 ( K j j − K k k ) /3 0 0 − 2 ( K ii + K j k ) .
Consequently,
36 ∥ p − p ∗ ∥ 2 2 − ∥ E s q ∥ F 2 = 3 K i i 2 + 4 ( K j j − K k k ) 2 + 7 K i j 2 + 7 K i k 2 ≥ 0 , 36\lVert p-p_*\rVert_2^2-\lVert E_{\mathrm{sq}}\rVert_F^2=3K_{ii}^2+4(K_{jj}-K_{kk})^2+7K_{ij}^2+7K_{ik}^2\geq0, 36 ∥ p − p ∗ ∥ 2 2 − ∥ E sq ∥ F 2 = 3 K ii 2 + 4 ( K j j − K k k ) 2 + 7 K ij 2 + 7 K ik 2 ≥ 0 ,
36 ∥ p − p ∗ ∥ 2 2 − ∥ E h e x ∥ F 2 = ( K i i K i j + K i k K j k ) ( 13 / 2 5 / 6 − 11 / 3 5 / 6 5 / 2 − 1 / 3 − 11 / 3 − 1 / 3 10 / 3 ) ( K i i K i j + K i k K j k ) + ( K j j − K k k K i j − K i k ) ( 7 / 2 1 / 6 1 / 6 7 / 2 ) ( K j j − K k k K i j − K i k ) ≥ 0. \begin{aligned}
36\lVert p-p_*\rVert_2^2-\lVert E_{\mathrm{hex}}\rVert_F^2
={}&(K_{ii}\quad K_{ij}+K_{ik}\quad K_{jk})
\begin{pmatrix} 13/2 & 5/6 & -11/3 \\ 5/6 & 5/2 & -1/3 \\ -11/3 & -1/3 & 10/3 \end{pmatrix}
\begin{pmatrix} K_{ii} \\ K_{ij}+K_{ik} \\ K_{jk} \end{pmatrix}\\
&+(K_{jj}-K_{kk}\quad K_{ij}-K_{ik})
\begin{pmatrix} 7/2 & 1/6 \\ 1/6 & 7/2 \end{pmatrix}
\begin{pmatrix} K_{jj}-K_{kk} \\ K_{ij}-K_{ik} \end{pmatrix}\geq0.
\end{aligned} 36 ∥ p − p ∗ ∥ 2 2 − ∥ E hex ∥ F 2 = ( K ii K ij + K ik K j k ) 13/2 5/6 − 11/3 5/6 5/2 − 1/3 − 11/3 − 1/3 10/3 K ii K ij + K ik K j k + ( K j j − K k k K ij − K ik ) ( 7/2 1/6 1/6 7/2 ) ( K j j − K k k K ij − K ik ) ≥ 0.
The two coefficient blocks are positive definite. Also,
∣ tr E s q ∣ 2 = 9 K i i 2 ≤ 27 ∥ p − p ∗ ∥ 2 2 , \lvert\operatorname{tr}E_{\mathrm{sq}}\rvert^2=9K_{ii}^2\leq27\lVert p-p_*\rVert_2^2, ∣ tr E sq ∣ 2 = 9 K ii 2 ≤ 27 ∥ p − p ∗ ∥ 2 2 ,
∣ tr E h e x ∣ 2 = ( K i i + K i j + K i k + 4 K j k 3 ) 2 ≤ 12 ∥ p − p ∗ ∥ 2 2 . \lvert\operatorname{tr}E_{\mathrm{hex}}\rvert^2=\left(K_{ii}+\frac{K_{ij}+K_{ik}+4K_{jk}}{3}\right)^2\leq12\lVert p-p_*\rVert_2^2. ∣ tr E hex ∣ 2 = ( K ii + 3 K ij + K ik + 4 K j k ) 2 ≤ 12 ∥ p − p ∗ ∥ 2 2 .
This proves (41) .
The map det ( I + K ) − 1 / 6 ( I + K ) − 1 / 2 \det(I+K)^{-1/6}(I+K)^{-1/2} det ( I + K ) − 1/6 ( I + K ) − 1/2 gives the physical volume-four cell. Its constant Jacobian cancels from the Rayleigh quotient, and its inverse metric is det ( I + K ) 1 / 3 ( I + K ) \det(I+K)^{1/3}(I+K) det ( I + K ) 1/3 ( I + K ) , with
det ( I + K ) 1 / 3 ≤ tr ( I + K ) 3 = 1 , ∥ E ∥ o p ≤ 6 ∥ p − p ∗ ∥ 2 < 1. \det(I+K)^{1/3}\leq\frac{\operatorname{tr}(I+K)}{3}=1,\qquad\lVert E\rVert_{\mathrm{op}}\leq6\lVert p-p_*\rVert_2<1. det ( I + K ) 1/3 ≤ 3 tr ( I + K ) = 1 , ∥ E ∥ op ≤ 6 ∥ p − p ∗ ∥ 2 < 1.
Put
J = det ( I + E ) > 0 , C = J ( I + E ) − 1 ( I + K ) ( I + E ) − T . (42) J=\det(I+E)>0,\qquad C=J(I+E)^{-1}(I+K)(I+E)^{-T}. \tag*{(42)} J = det ( I + E ) > 0 , C = J ( I + E ) − 1 ( I + K ) ( I + E ) − T . (42)
Define
B θ ( f , g ; A , b ) = ∫ W ( ∇ f T A ∇ g − θ b f g ) . B_\theta(f,g;A,b)=\int_W(\nabla f^T A\nabla g-\theta bfg). B θ ( f , g ; A , b ) = ∫ W ( ∇ f T A ∇ g − θ b f g ) .
For 0 ≠ f ∈ H 1 ( W ) 0\ne f\in H^1(W) 0 = f ∈ H 1 ( W ) with f ( − x ) = − f ( x ) f(-x)=-f(x) f ( − x ) = − f ( x ) ,
J ( − x ) = J ( x ) , ∫ W J f = 0 , J(-x)=J(x),\qquad\int_W Jf=0, J ( − x ) = J ( x ) , ∫ W J f = 0 ,
so the Rayleigh principle gives
2 − 4 / 3 F ( p ) − θ ≤ B θ ( f , f ; C , J ) ∫ W J f 2 , ∫ W J f 2 > 0. (43) 2^{-4/3}F(p)-\theta\leq\frac{B_\theta(f,f;C,J)}{\int_W Jf^2},\qquad\int_W Jf^2>0. \tag*{(43)} 2 − 4/3 F ( p ) − θ ≤ ∫ W J f 2 B θ ( f , f ; C , J ) , ∫ W J f 2 > 0. (43)
On Ω + \Omega_+ Ω + define
p 1 = x 1 ∑ i , j , k ∈ N 0 i + j + k ≤ 7 , j ≤ k c i j k ∑ a = 0 i ∑ b = 0 j ∑ c = 0 k ( − 1 ) i + j + k − a − b − c ( i + a a ) ( j + b b ) ( k + c c ) x 1 a ( x 2 b x 3 c + 1 j < k x 2 c x 3 b ) . (44) p_1=x_1\sum_{\substack{i,j,k\in\mathbb{N}_0\\i+j+k\leq7,\ j\leq k}}c_{ijk}\sum_{a=0}^{i}\sum_{b=0}^{j}\sum_{c=0}^{k}(-1)^{i+j+k-a-b-c}
\binom{i+a}{a}\binom{j+b}{b}\binom{k+c}{c}x_1^a\left(x_2^b x_3^c+1_{j<k}x_2^c x_3^b\right). \tag*{(44)} p 1 = x 1 i , j , k ∈ N 0 i + j + k ≤ 7 , j ≤ k ∑ c ij k a = 0 ∑ i b = 0 ∑ j c = 0 ∑ k ( − 1 ) i + j + k − a − b − c ( a i + a ) ( b j + b ) ( c k + c ) x 1 a ( x 2 b x 3 c + 1 j < k x 2 c x 3 b ) . (44)
Read the 70 coefficients row by row, from left to right and top to bottom, in the nested order
i = 0 , … , 7 , j = 0 , … , ⌊ 7 − i 2 ⌋ , k = j , … , 7 − i − j . i=0,\ldots,7,\qquad j=0,\ldots,\left\lfloor\frac{7-i}{2}\right\rfloor,\qquad k=j,\ldots,7-i-j. i = 0 , … , 7 , j = 0 , … , ⌊ 2 7 − i ⌋ , k = j , … , 7 − i − j .
229900428 − 139871184 − 52930218 − 14629008 − 13261341 − 5686815 − 889274 73926 − 83267960 − 41119660 − 26184507 − 21182213 − 8427964 − 1388016 − 11380873 − 6083225 − 7172094 − 2162237 282052 − 402975 − 775093 78837850 21184188 − 10453204 − 10599016 − 5677196 − 1112077 116720818 67372472 14741791 − 3555916 − 3386892 47315301 14147076 2266381 2353273 − 3445608 11601610 − 2439555 − 3913125 − 1623965 − 757041 − 10857375 − 10285059 − 624273 1512055 − 7195789 − 1214356 9282897 10229380 4231107 1067499 337892 854262 − 3570128 − 1636479 − 674516 8206128 11338133 3833423 487474 13246110 3072687 3185632 4391020 1201058 4365604 340142 95828 − 177693 \begin{array}{rrrrr}
229900428 & -139871184 & -52930218 & -14629008 & -13261341 \\
-5686815 & -889274 & 73926 & -83267960 & -41119660 \\
-26184507 & -21182213 & -8427964 & -1388016 & -11380873 \\
-6083225 & -7172094 & -2162237 & 282052 & -402975 \\
-775093 & 78837850 & 21184188 & -10453204 & -10599016 \\
-5677196 & -1112077 & 116720818 & 67372472 & 14741791 \\
-3555916 & -3386892 & 47315301 & 14147076 & 2266381 \\
2353273 & -3445608 & 11601610 & -2439555 & -3913125 \\
-1623965 & -757041 & -10857375 & -10285059 & -624273 \\
1512055 & -7195789 & -1214356 & 9282897 & 10229380 \\
4231107 & 1067499 & 337892 & 854262 & -3570128 \\
-1636479 & -674516 & 8206128 & 11338133 & 3833423 \\
487474 & 13246110 & 3072687 & 3185632 & 4391020 \\
1201058 & 4365604 & 340142 & 95828 & -177693
\end{array} 229900428 − 5686815 − 26184507 − 6083225 − 775093 − 5677196 − 3555916 2353273 − 1623965 1512055 4231107 − 1636479 487474 1201058 − 139871184 − 889274 − 21182213 − 7172094 78837850 − 1112077 − 3386892 − 3445608 − 757041 − 7195789 1067499 − 674516 13246110 4365604 − 52930218 73926 − 8427964 − 2162237 21184188 116720818 47315301 11601610 − 10857375 − 1214356 337892 8206128 3072687 340142 − 14629008 − 83267960 − 1388016 282052 − 10453204 67372472 14147076 − 2439555 − 10285059 9282897 854262 11338133 3185632 95828 − 13261341 − 41119660 − 11380873 − 402975 − 10599016 14741791 2266381 − 3913125 − 624273 10229380 − 3570128 3833423 4391020 − 177693
Thus c 000 = 229900428 c_{000}=229900428 c 000 = 229900428 , c 005 = − 5686815 c_{005}=-5686815 c 005 = − 5686815 , and c 700 = − 177693 c_{700}=-177693 c 700 = − 177693 . Obtain p i p_i p i by interchanging coordinates 1 and i i i . Set
∫ Ω + p 1 2 = 6755696638648162400908255727 675569664000 , \int_{\Omega_+}p_1^2=\frac{6755696638648162400908255727}{675569664000}, ∫ Ω + p 1 2 = 675569664000 6755696638648162400908255727 ,
P i ( x ) = sgn ( x i ) 8 ∫ Ω + p 1 2 p i ( ∣ x 1 ∣ , ∣ x 2 ∣ , ∣ x 3 ∣ ) , x ∈ W . (45) P_i(x)=\frac{\operatorname{sgn}(x_i)}{\sqrt{8\int_{\Omega_+}p_1^2}}p_i(|x_1|,|x_2|,|x_3|),\qquad x\in W. \tag*{(45)} P i ( x ) = 8 ∫ Ω + p 1 2 sgn ( x i ) p i ( ∣ x 1 ∣ , ∣ x 2 ∣ , ∣ x 3 ∣ ) , x ∈ W . (45)
For α ∈ N 3 \alpha\in\mathbb{N}^3 α ∈ N 3 , write ∣ α ∣ = α 1 + α 2 + α 3 |\alpha|=\alpha_1+\alpha_2+\alpha_3 ∣ α ∣ = α 1 + α 2 + α 3 and α ! = α 1 ! α 2 ! α 3 ! \alpha!=\alpha_1!\alpha_2!\alpha_3! α ! = α 1 ! α 2 ! α 3 ! . Barycentric integration on T = conv ( w 0 , … , w 3 ) ∈ T + T=\operatorname{conv}(w_0,\ldots,w_3)\in T_+ T = conv ( w 0 , … , w 3 ) ∈ T + gives
∫ T x α d x = 6 ∣ T ∣ α ! ( ∣ α ∣ + 3 ) ! [ z α ] ∏ i = 0 3 ( 1 − w i ⋅ z ) − 1 . (46) \int_Tx^\alpha\,dx=\frac{6|T|\alpha!}{(|\alpha|+3)!}[z^\alpha]\prod_{i=0}^{3}(1-w_i\cdot z)^{-1}. \tag*{(46)} ∫ T x α d x = ( ∣ α ∣ + 3 )! 6∣ T ∣ α ! [ z α ] i = 0 ∏ 3 ( 1 − w i ⋅ z ) − 1 . (46)
For the doubled vertices, the identities
∑ α ∈ N 0 3 g α ( 0 ) z α = 1 , ( 1 − 2 w i ⋅ z ) ∑ α ∈ N 0 3 g α ( i + 1 ) z α = ∑ α ∈ N 0 3 g α ( i ) z α \sum_{\alpha\in\mathbb{N}_0^3}g_\alpha^{(0)}z^\alpha=1,\qquad
(1-2w_i\cdot z)\sum_{\alpha\in\mathbb{N}_0^3}g_\alpha^{(i+1)}z^\alpha=\sum_{\alpha\in\mathbb{N}_0^3}g_\alpha^{(i)}z^\alpha α ∈ N 0 3 ∑ g α ( 0 ) z α = 1 , ( 1 − 2 w i ⋅ z ) α ∈ N 0 3 ∑ g α ( i + 1 ) z α = α ∈ N 0 3 ∑ g α ( i ) z α
give g α ( 0 ) = 1 α = 0 g_\alpha^{(0)}=1_{\alpha=0} g α ( 0 ) = 1 α = 0 and, for i = 0 , 1 , 2 , 3 i=0,1,2,3 i = 0 , 1 , 2 , 3 ,
g α ( i + 1 ) = g α ( i ) + ∑ j = 1 3 ( 2 w i ) j g α − e j ( i + 1 ) , g β ( i + 1 ) = 0 if min 1 ≤ j ≤ 3 β j < 0. (47) g_\alpha^{(i+1)}=g_\alpha^{(i)}+\sum_{j=1}^{3}(2w_i)_jg_{\alpha-e_j}^{(i+1)},\qquad
g_\beta^{(i+1)}=0\ \text{if}\ \min_{1\le j\le3}\beta_j<0. \tag*{(47)} g α ( i + 1 ) = g α ( i ) + j = 1 ∑ 3 ( 2 w i ) j g α − e j ( i + 1 ) , g β ( i + 1 ) = 0 if 1 ≤ j ≤ 3 min β j < 0. (47)
Evaluation in increasing ∣ α ∣ |\alpha| ∣ α ∣ through 16 gives
[ z α ] ∏ i = 0 3 ( 1 − w i ⋅ z ) − 1 = 2 − ∣ α ∣ g α ( 4 ) , [z^\alpha]\prod_{i=0}^{3}(1-w_i\cdot z)^{-1}=2^{-|\alpha|}g_\alpha^{(4)}, [ z α ] i = 0 ∏ 3 ( 1 − w i ⋅ z ) − 1 = 2 − ∣ α ∣ g α ( 4 ) ,
8 ( 6 ∣ T ∣ ) ∈ Z , 2 19 19 ! ∫ T x α d x ∈ Z ( ∣ α ∣ ≤ 16 ) . 8(6|T|)\in\mathbb{Z},\qquad2^{19}19!\int_Tx^\alpha\,dx\in\mathbb{Z}\quad(|\alpha|\le16). 8 ( 6∣ T ∣ ) ∈ Z , 2 19 19 ! ∫ T x α d x ∈ Z ( ∣ α ∣ ≤ 16 ) .
Summing these moments against the coefficients of p 1 2 p_1^2 p 1 2 and ∣ ∇ p 1 ∣ 2 |\nabla p_1|^2 ∣∇ p 1 ∣ 2 yields (4.9) and (4.12). Parity and coordinate permutation give
∫ W P i P j = δ i j , ∫ W ∇ P i ⋅ ∇ P j = r δ i j , \int_WP_iP_j=\delta_{ij},\qquad\int_W\nabla P_i\cdot\nabla P_j=r\delta_{ij}, ∫ W P i P j = δ ij , ∫ W ∇ P i ⋅ ∇ P j = r δ ij ,
r = 377741884455339525044322247762 87824056302426111211807324451 < 430113 100000 . (48) r=\frac{377741884455339525044322247762}{87824056302426111211807324451}<\frac{430113}{100000}. \tag*{(48)} r = 87824056302426111211807324451 377741884455339525044322247762 < 100000 430113 . (48)
By (35) , 4300827 / 10 6 < λ 4300827/10^{6}<\lambda 4300827/1 0 6 < λ . With unit octant normalization and a common sign choice for the eigenfunctions, write
P = α u + w , ( w , u ) = 0 , α ≥ 0 , α 2 + ∥ w ∥ 2 2 = 1. P=\alpha u+w,\qquad(w,u)=0,\qquad\alpha\geq0,\qquad\alpha^{2}+\lVert w\rVert_{2}^{2}=1. P = α u + w , ( w , u ) = 0 , α ≥ 0 , α 2 + ∥ w ∥ 2 2 = 1.
Since λ 2 > π 2 > 49 / 5 \lambda_{2}>\pi^{2}>49/5 λ 2 > π 2 > 49/5 , the mixed spectral gap yields
r − λ = ∥ ∇ w ∥ 2 2 − λ ∥ w ∥ 2 2 ≥ 549887 980000 ∥ ∇ w ∥ 2 2 , r-\lambda=\lVert\nabla w\rVert_{2}^{2}-\lambda\lVert w\rVert_{2}^{2}\geq\frac{549887}{980000}\lVert\nabla w\rVert_{2}^{2}, r − λ = ∥ ∇ w ∥ 2 2 − λ ∥ w ∥ 2 2 ≥ 980000 549887 ∥ ∇ w ∥ 2 2 ,
∥ ∇ w ∥ 2 2 < 1 1800 , 1 − α ≤ ∥ w ∥ 2 2 < 1 17640 , \lVert\nabla w\rVert_{2}^{2}<\frac{1}{1800},\qquad1-\alpha\leq\lVert w\rVert_{2}^{2}<\frac{1}{17640}, ∥ ∇ w ∥ 2 2 < 1800 1 , 1 − α ≤ ∥ w ∥ 2 2 < 17640 1 ,
∥ ∇ ( P − u ) ∥ 2 2 = ∥ ∇ w ∥ 2 2 + λ ( 1 − α ) 2 < ( 3 125 ) 2 , \lVert\nabla(P-u)\rVert_{2}^{2}=\lVert\nabla w\rVert_{2}^{2}+\lambda(1-\alpha)^{2}<\left(\frac{3}{125}\right)^{2}, ∥ ∇ ( P − u ) ∥ 2 2 = ∥ ∇ w ∥ 2 2 + λ ( 1 − α ) 2 < ( 125 3 ) 2 ,
∥ P − u ∥ 2 2 = ∥ w ∥ 2 2 + ( 1 − α ) 2 < ( 1 125 ) 2 . \lVert P-u\rVert_{2}^{2}=\lVert w\rVert_{2}^{2}+(1-\alpha)^{2}<\left(\frac{1}{125}\right)^{2}. ∥ P − u ∥ 2 2 = ∥ w ∥ 2 2 + ( 1 − α ) 2 < ( 125 1 ) 2 .
For u ξ = ∑ i = 1 3 ξ i u i u_{\xi}=\sum_{i=1}^{3}\xi_{i}u_{i} u ξ = ∑ i = 1 3 ξ i u i , P ξ = ∑ i = 1 3 ξ i P i P_{\xi}=\sum_{i=1}^{3}\xi_{i}P_{i} P ξ = ∑ i = 1 3 ξ i P i , ∣ ξ ∣ = 1 |\xi|=1 ∣ ξ ∣ = 1 , orthogonality of the parity sectors gives
∥ ∇ ( u ξ − P ξ ) ∥ 2 < 3 125 , ∥ u ξ − P ξ ∥ 2 < 1 125 . (49) \lVert\nabla(u_{\xi}-P_{\xi})\rVert_{2}<\frac{3}{125},\qquad\lVert u_{\xi}-P_{\xi}\rVert_{2}<\frac{1}{125}. \tag*{(49)} ∥ ∇ ( u ξ − P ξ ) ∥ 2 < 125 3 , ∥ u ξ − P ξ ∥ 2 < 125 1 . (49)
In dimension three,
J = 1 + tr E + ( tr E ) 2 − tr ( E 2 ) 2 + det E . J=1+\operatorname{tr}E+\frac{(\operatorname{tr}E)^{2}-\operatorname{tr}(E^{2})}{2}+\det E. J = 1 + tr E + 2 ( tr E ) 2 − tr ( E 2 ) + det E .
For M 0 = I M_{0}=I M 0 = I and M m = 0 M_{m}=0 M m = 0 when m < 0 m<0 m < 0 , define
M m = ∑ a = 0 m ( − E ) a ( − E T ) m − a + ∑ a = 0 m − 1 ( − E ) a K ( − E T ) m − 1 − a ( m ≥ 1 ) , (50) M_{m}=\sum_{a=0}^{m}(-E)^{a}(-E^{T})^{m-a}+\sum_{a=0}^{m-1}(-E)^{a}K(-E^{T})^{m-1-a}\qquad(m\geq1), \tag*{(50)} M m = a = 0 ∑ m ( − E ) a ( − E T ) m − a + a = 0 ∑ m − 1 ( − E ) a K ( − E T ) m − 1 − a ( m ≥ 1 ) , (50)
C m = M m + ( tr E ) M m − 1 + ( tr E ) 2 − tr ( E 2 ) 2 M m − 2 + ( det E ) M m − 3 . C_{m}=M_{m}+(\operatorname{tr}E)M_{m-1}+\frac{(\operatorname{tr}E)^{2}-\operatorname{tr}(E^{2})}{2}M_{m-2}+(\det E)M_{m-3}. C m = M m + ( tr E ) M m − 1 + 2 ( tr E ) 2 − tr ( E 2 ) M m − 2 + ( det E ) M m − 3 .
Then C = ∑ m = 0 ∞ C m C=\sum_{m=0}^{\infty}C_{m} C = ∑ m = 0 ∞ C m . For p ≠ p ∗ p\ne p_{*} p = p ∗ , (41) and (37) give
∣ tr E ∣ < 26 5 ∥ p − p ∗ ∥ 2 , (51) |\operatorname{tr}E|<\frac{26}{5}\lVert p-p_{*}\rVert_{2}, \tag*{(51)} ∣ tr E ∣ < 5 26 ∥ p − p ∗ ∥ 2 , (51)
1 2 ∣ ( tr E ) 2 − tr ( E 2 ) ∣ ≤ 1 2 ( ( tr E ) 2 + ∥ E ∥ F 2 ) ≤ 63 2 ∥ p − p ∗ ∥ 2 2 , \frac{1}{2}|(\operatorname{tr}E)^{2}-\operatorname{tr}(E^{2})|\leq\frac{1}{2}\left((\operatorname{tr}E)^{2}+\lVert E\rVert_{F}^{2}\right)\leq\frac{63}{2}\lVert p-p_{*}\rVert_{2}^{2}, 2 1 ∣ ( tr E ) 2 − tr ( E 2 ) ∣ ≤ 2 1 ( ( tr E ) 2 + ∥ E ∥ F 2 ) ≤ 2 63 ∥ p − p ∗ ∥ 2 2 ,
∣ det E ∣ ≤ ( ∥ E ∥ F 3 ) 3 < 216 5 ∥ p − p ∗ ∥ 2 3 , |\det E|\leq\left(\frac{\lVert E\rVert_{F}}{\sqrt{3}}\right)^{3}<\frac{216}{5}\lVert p-p_{*}\rVert_{2}^{3}, ∣ det E ∣ ≤ ( 3 ∥ E ∥ F ) 3 < 5 216 ∥ p − p ∗ ∥ 2 3 ,
∥ M m ∥ o p ≤ ( m + 1 ) ( 6 ∥ p − p ∗ ∥ 2 ) m + m ∥ K ∥ o p ( 6 ∥ p − p ∗ ∥ 2 ) m − 1 ≤ ( 1 + 3 m / 2 ) ( 6 ∥ p − p ∗ ∥ 2 ) m ( m ≥ 1 ) , \begin{aligned}
\lVert M_{m}\rVert_{\mathrm{op}}\leq(m+1)(6\lVert p-p_{*}\rVert_{2})^{m}+m\lVert K\rVert_{\mathrm{op}}(6\lVert p-p_{*}\rVert_{2})^{m-1} \\
&\leq(1+3m/2)(6\lVert p-p_{*}\rVert_{2})^{m}\qquad(m\geq1),
\end{aligned} ∥ M m ∥ op ≤ ( m + 1 ) ( 6 ∥ p − p ∗ ∥ 2 ) m + m ∥ K ∥ op ( 6 ∥ p − p ∗ ∥ 2 ) m − 1 ≤ ( 1 + 3 m /2 ) ( 6 ∥ p − p ∗ ∥ 2 ) m ( m ≥ 1 ) ,
∥ C 2 ∥ o p < 261 ∥ p − p ∗ ∥ 2 2 , \lVert C_{2}\rVert_{\mathrm{op}}<261\lVert p-p_{*}\rVert_{2}^{2}, ∥ C 2 ∥ op < 261 ∥ p − p ∗ ∥ 2 2 ,
∥ C 3 ∥ o p < 2484 ∥ p − p ∗ ∥ 2 3 . \lVert C_{3}\rVert_{\mathrm{op}}<2484\lVert p-p_{*}\rVert_{2}^{3}. ∥ C 3 ∥ op < 2484 ∥ p − p ∗ ∥ 2 3 .
The first homogeneous term is
C 1 = K − E − E T + ( tr E ) I . C_{1}=K-E-E^{T}+(\operatorname{tr}E)I. C 1 = K − E − E T + ( tr E ) I .
Let
( L θ ( K / 2 ) ) i j = B θ ( f i , f j ; C 1 , tr E ) , ( θ , f i ) = { ( λ , u i ) , ( r , P i ) . (L_{\theta}(K/2))_{ij}=B_{\theta}(f_{i},f_{j};C_{1},\operatorname{tr}E),\qquad
(\theta,f_{i})=
\begin{cases}
(\lambda,u_{i}),\\
(r,P_{i}).
\end{cases} ( L θ ( K /2 ) ) ij = B θ ( f i , f j ; C 1 , tr E ) , ( θ , f i ) = { ( λ , u i ) , ( r , P i ) .
For the eigenfunctions, this is the first-order compression to their eigenspace [11 ] . For every signed permutation S S S ,
L θ ( S K S T / 2 ) = S L θ ( K / 2 ) S T . L_{\theta}(SKS^{T}/2)=SL_{\theta}(K/2)S^{T}. L θ ( S K S T /2 ) = S L θ ( K /2 ) S T .
Coordinate reflections and the exchange 1 ↔ 2 1 \leftrightarrow2 1 ↔ 2 give
L θ ( diag ( 1 , − 1 , 0 ) ) = ( L θ ( diag ( 1 , − 1 , 0 ) ) ) 11 diag ( 1 , − 1 , 0 ) , L_{\theta}(\operatorname{diag}(1,-1,0)) = \bigl(L_{\theta}(\operatorname{diag}(1,-1,0))\bigr)_{11}\operatorname{diag}(1,-1,0), L θ ( diag ( 1 , − 1 , 0 )) = ( L θ ( diag ( 1 , − 1 , 0 )) ) 11 diag ( 1 , − 1 , 0 ) ,
L θ ( e 1 e 2 T + e 2 e 1 T ) = ( L θ ( e 1 e 2 T + e 2 e 1 T ) ) 12 ( e 1 e 2 T + e 2 e 1 T ) . L_{\theta}(e_{1}e_{2}^{T}+e_{2}e_{1}^{T}) = \bigl(L_{\theta}(e_{1}e_{2}^{T}+e_{2}e_{1}^{T})\bigr)_{12}(e_{1}e_{2}^{T}+e_{2}e_{1}^{T}). L θ ( e 1 e 2 T + e 2 e 1 T ) = ( L θ ( e 1 e 2 T + e 2 e 1 T ) ) 12 ( e 1 e 2 T + e 2 e 1 T ) .
Since L θ L_{\theta} L θ is linear and tr K = 0 \operatorname{tr}K=0 tr K = 0 ,
K = K 11 diag ( 1 , 0 , − 1 ) + K 22 diag ( 0 , 1 , − 1 ) + ∑ 1 ≤ i < j ≤ 3 K i j ( e i e j T + e j e i T ) . K=K_{11}\operatorname{diag}(1,0,-1)+K_{22}\operatorname{diag}(0,1,-1)+\sum_{1\leq i<j\leq3}K_{ij}(e_{i}e_{j}^{T}+e_{j}e_{i}^{T}). K = K 11 diag ( 1 , 0 , − 1 ) + K 22 diag ( 0 , 1 , − 1 ) + 1 ≤ i < j ≤ 3 ∑ K ij ( e i e j T + e j e i T ) .
Permutation covariance therefore yields
L θ ( K / 2 ) = 1 2 ( L θ ( diag ( 1 , − 1 , 0 ) ) ) 11 diag K + 1 2 ( L θ ( e 1 e 2 T + e 2 e 1 T ) ) 12 off K . (52) L_{\theta}(K/2)=\frac{1}{2}\bigl(L_{\theta}(\operatorname{diag}(1,-1,0))\bigr)_{11}\operatorname{diag}K
+\frac{1}{2}\bigl(L_{\theta}(e_{1}e_{2}^{T}+e_{2}e_{1}^{T})\bigr)_{12}\operatorname{off}K. \tag*{(52)} L θ ( K /2 ) = 2 1 ( L θ ( diag ( 1 , − 1 , 0 )) ) 11 diag K + 2 1 ( L θ ( e 1 e 2 T + e 2 e 1 T ) ) 12 off K . (52)
For m ∈ { 2 , 3 } m\in\{2,3\} m ∈ { 2 , 3 } , put
( B m ( K ) ) i j = B r ( P i , P j ; C m , [ z m ] det ( I + z E ) ) . (B_{m}(K))_{ij}=\mathcal{B}_{r}(P_{i},P_{j};C_{m},[z^{m}]\det(I+zE)). ( B m ( K ) ) ij = B r ( P i , P j ; C m , [ z m ] det ( I + z E )) .
For A = A T A=A^{T} A = A T and pairs ( P , u ) (P,u) ( P , u ) , ( Q , v ) (Q,v) ( Q , v ) satisfying (49) , expansion gives
B λ ( u , v ; A , b ) − B r ( P , Q ; A , b ) = B r ( u − P , Q ; A , b ) + B r ( P , v − Q ; A , b ) + B r ( u − P , v − Q ; A , b ) + ( r − λ ) ∫ W b u . \mathcal{B}_{\lambda}(u,v;A,b)-\mathcal{B}_{r}(P,Q;A,b)
=\mathcal{B}_{r}(u-P,Q;A,b)+\mathcal{B}_{r}(P,v-Q;A,b)
+\mathcal{B}_{r}(u-P,v-Q;A,b)+(r-\lambda)\int_{W}bu\,. B λ ( u , v ; A , b ) − B r ( P , Q ; A , b ) = B r ( u − P , Q ; A , b ) + B r ( P , v − Q ; A , b ) + B r ( u − P , v − Q ; A , b ) + ( r − λ ) ∫ W b u .
and hence, by Cauchy–Schwarz,
∣ B λ ( u , v ; A , b ) − B r ( P , Q ; A , b ) ∣ ≤ 6 125 max { ∥ A ∇ P ∥ 2 , ∥ A ∇ Q ∥ 2 } + 9 15625 ∥ A ∥ ∞ , op + 430113 100000 ( 2 125 max { ∥ b P ∥ 2 , ∥ b Q ∥ 2 } + ∥ b ∥ ∞ 15625 ) + 303 10 6 ∣ ∫ W b u ∣ . (53) \left|\mathcal{B}_{\lambda}(u,v;A,b)-\mathcal{B}_{r}(P,Q;A,b)\right|
\leq\frac{6}{125}\max\{\lVert A\nabla P\rVert_{2},\lVert A\nabla Q\rVert_{2}\}
+\frac{9}{15625}\lVert A\rVert_{\infty,\operatorname{op}}
+\frac{430113}{100000}\left(\frac{2}{125}\max\{\lVert bP\rVert_{2},\lVert bQ\rVert_{2}\}+\frac{\lVert b\rVert_{\infty}}{15625}\right)
+\frac{303}{10^{6}}\left|\int_{W}bu\right|. \tag*{(53)} ∣ B λ ( u , v ; A , b ) − B r ( P , Q ; A , b ) ∣ ≤ 125 6 max {∥ A ∇ P ∥ 2 , ∥ A ∇ Q ∥ 2 } + 15625 9 ∥ A ∥ ∞ , op + 100000 430113 ( 125 2 max {∥ b P ∥ 2 , ∥ b Q ∥ 2 } + 15625 ∥ b ∥ ∞ ) + 1 0 6 303 ∫ W b u . (53)
Numerical Lemma 2 (BCC trial moments). For (44) –(45) , componentwise,
\begin{pmatrix}
\bigl(L_{r}(\operatorname{diag}(1,-1,0))\bigr)_{11} & \bigl(L_{r}(e_{1}e_{2}^{T}+e_{2}e_{1}^{T})\bigr)_{12}\\
\bigl(L_{\lambda}(\operatorname{diag}(1,-1,0))\bigr)_{11} & \bigl(L_{\lambda}(e_{1}e_{2}^{T}+e_{2}e_{1}^{T})\bigr)_{12}
\end{pmatrix}
<
\begin{pmatrix}
-97/50 & -263/50\\
-7/5 & -24/5
\end{pmatrix}. \tag*{(54)}\end{pmatrix}
For p ≠ p ∗ p\ne p_{\ast} p = p ∗ ,
λ max ( B 2 ( K ) ) < 45 ∥ p − p ∗ ∥ 2 2 , ∥ B 3 ( K ) ∥ F < 2376 5 ∥ p − p ∗ ∥ 2 2 . (55) \lambda_{\max}(B_{2}(K))<45\lVert p-p_{\ast}\rVert_{2}^{2},\qquad\lVert B_{3}(K)\rVert_{F}<\frac{2376}{5}\lVert p-p_{\ast}\rVert_{2}^{2}. \tag*{(55)} λ m a x ( B 2 ( K )) < 45 ∥ p − p ∗ ∥ 2 2 , ∥ B 3 ( K ) ∥ F < 5 2376 ∥ p − p ∗ ∥ 2 2 . (55)
Proof of Numerical Lemma 2. Use
( K , A , b ) = { ( 2 diag ( 1 , − 1 , 0 ) , C 1 + 2 I , tr E + 2 ) , a = 1 , ( 2 ( e 1 e 2 T + e 2 e 1 T ) , C 1 , tr E ) , a = 2. (K,A,b)=
\begin{cases}
(2\operatorname{diag}(1,-1,0),C_{1}+2I,\operatorname{tr}E+2), & a=1,\\
(2(e_{1}e_{2}^{T}+e_{2}e_{1}^{T}),C_{1},\operatorname{tr}E), & a=2.
\end{cases} ( K , A , b ) = { ( 2 diag ( 1 , − 1 , 0 ) , C 1 + 2 I , tr E + 2 ) , ( 2 ( e 1 e 2 T + e 2 e 1 T ) , C 1 , tr E ) , a = 1 , a = 2.
The diagonal shift is valid because B θ ( f 1 , f 1 ; I , 1 ) = 0 \mathcal{B}_{\theta}(f_{1},f_{1};I,1)=0 B θ ( f 1 , f 1 ; I , 1 ) = 0 . Under coordinate reflections, the diagonal pairing is invariant and the two x 1 x 2 x_{1}x_{2} x 1 x 2 characters in the shear pairing cancel. Equations (40) and (46) give, componentwise,
( − 1941429 43700852 8980830 307200 ) < 10 6 ∫ Ω + p 1 2 ∑ T ∈ T + ∫ T ( ∇ p 1 T A ∇ p a − r b p 1 p a ∇ p 1 T A 2 ∇ p 1 b 2 p 1 2 b p 1 p a ) < ( − 1941428 43700853 8980831 307201 ) , a = 1 , ( − 5269574 31266387 1733164 − 685749 ) < 10 6 ∫ Ω + p 1 2 ∑ T ∈ T + ∫ T ( ∇ p 1 T A ∇ p a − r b p 1 p a ∇ p 1 T A 2 ∇ p 1 b 2 p 1 2 b p 1 p a ) < ( − 5269573 31266388 1733165 − 685748 ) , a = 2. (56) \begin{aligned}
\begin{pmatrix}
-1941429\\
43700852\\
8980830\\
307200
\end{pmatrix}
&<
\frac{10^{6}}{\int_{\Omega_+}p_1^2}
\sum_{T\in\mathcal{T}_+}\int_T
\begin{pmatrix}
\nabla p_1^T A\nabla p_a-rbp_1p_a\\
\nabla p_1^T A^2\nabla p_1\\
b^2p_1^2\\
bp_1p_a
\end{pmatrix}
<
\begin{pmatrix}
-1941428\\
43700853\\
8980831\\
307201
\end{pmatrix},
\qquad a=1,\\
\begin{pmatrix}
-5269574\\
31266387\\
1733164\\
-685749
\end{pmatrix}
&<
\frac{10^{6}}{\int_{\Omega_+}p_1^2}
\sum_{T\in\mathcal{T}_+}\int_T
\begin{pmatrix}
\nabla p_1^T A\nabla p_a-rbp_1p_a\\
\nabla p_1^T A^2\nabla p_1\\
b^2p_1^2\\
bp_1p_a
\end{pmatrix}
<
\begin{pmatrix}
-5269573\\
31266388\\
1733165\\
-685748
\end{pmatrix},
\qquad a=2. \tag*{(56)}
\end{aligned} − 1941429 43700852 8980830 307200 − 5269574 31266387 1733164 − 685749 < ∫ Ω + p 1 2 1 0 6 T ∈ T + ∑ ∫ T ∇ p 1 T A ∇ p a − r b p 1 p a ∇ p 1 T A 2 ∇ p 1 b 2 p 1 2 b p 1 p a < − 1941428 43700853 8980831 307201 , a = 1 , < ∫ Ω + p 1 2 1 0 6 T ∈ T + ∑ ∫ T ∇ p 1 T A ∇ p a − r b p 1 p a ∇ p 1 T A 2 ∇ p 1 b 2 p 1 2 b p 1 p a < − 5269573 31266388 1733165 − 685748 , a = 2. (56)
The vector inequalities are componentwise. On each tetrahedron,
( max 1 ≤ i ≤ 3 ∑ j = 1 3 ∣ A i j ∣ , ∣ b ∣ ) ≤ { ( 10 , 8 ) , a = 1 , ( 6 , 8 / 3 ) , a = 2. \begin{pmatrix}
\max_{1\leq i\leq3}\sum_{j=1}^{3}|A_{ij}|,|b|
\end{pmatrix}
\leq
\begin{cases}
(10,8), & a=1,\\
(6,8/3), & a=2.
\end{cases} ( max 1 ≤ i ≤ 3 ∑ j = 1 3 ∣ A ij ∣ , ∣ b ∣ ) ≤ { ( 10 , 8 ) , ( 6 , 8/3 ) , a = 1 , a = 2.
The shear test is invariant under 1 ↔ 2 1\leftrightarrow2 1 ↔ 2 . For (53) , the integral bounds give
( ∥ A ∇ P 1 ∥ 2 ∥ b P 1 ∥ 2 ) < { ( 331 / 50 3 ) , a = 1 , ( 6 7 / 5 ) , a = 2. \begin{pmatrix}
\lVert A\nabla P_1\rVert_2\\
\lVert bP_1\rVert_2
\end{pmatrix}
<
\begin{cases}
\begin{pmatrix}331/50\\3\end{pmatrix}, & a=1,\\
\begin{pmatrix}6\\7/5\end{pmatrix}, & a=2.
\end{cases} ( ∥ A ∇ P 1 ∥ 2 ∥ b P 1 ∥ 2 ) < ⎩ ⎨ ⎧ ( 331/50 3 ) , ( 6 7/5 ) , a = 1 , a = 2.
For a = 2 a=2 a = 2 the same bounds hold with P 2 P_2 P 2 . The first components yield
( L r ( diag ( 1 , − 1 , 0 ) ) ) 11 < − 97 / 50 , ( L r ( e 1 e 2 T + e 2 e 1 T ) ) 12 < − 263 / 50. \left(L_r(\operatorname{diag}(1,-1,0))\right)_{11}<-97/50,
\qquad
\left(L_r(e_1e_2^T+e_2e_1^T)\right)_{12}<-263/50. ( L r ( diag ( 1 , − 1 , 0 )) ) 11 < − 97/50 , ( L r ( e 1 e 2 T + e 2 e 1 T ) ) 12 < − 263/50.
For the diagonal and shear pairings, respectively,
∣ ∫ W b u 1 2 ∣ ≤ ∫ W b P 1 2 + 2 125 ∥ b P 1 ∥ 2 + 8 15625 < 307201 10 6 + 6 125 + 8 15625 ( a = 1 ) , ∣ ∫ W b u 1 u 2 ∣ ≤ ∥ b ∥ ∞ ∥ u 1 ∥ 2 ∥ u 2 ∥ 2 ≤ 8 / 3 ( a = 2 ) . \begin{aligned}
\left|\int_W bu_1^2\right|
&\leq\int_W bP_1^2+\frac{2}{125}\lVert bP_1\rVert_2+\frac{8}{15625}\\
&<\frac{307201}{10^6}+\frac{6}{125}+\frac{8}{15625}
\qquad(a=1),\\
\left|\int_W bu_1u_2\right|
&\leq\lVert b\rVert_\infty\lVert u_1\rVert_2\lVert u_2\rVert_2\leq8/3
\qquad(a=2).
\end{aligned} ∫ W b u 1 2 ∫ W b u 1 u 2 ≤ ∫ W b P 1 2 + 125 2 ∥ b P 1 ∥ 2 + 15625 8 < 1 0 6 307201 + 125 6 + 15625 8 ( a = 1 ) , ≤ ∥ b ∥ ∞ ∥ u 1 ∥ 2 ∥ u 2 ∥ 2 ≤ 8/3 ( a = 2 ) .
Equation (53) gives
( L λ ( diag ( 1 , − 1 , 0 ) ) ) 11 < − 7 / 5 , ( L λ ( e 1 e 2 T + e 2 e 1 T ) ) 12 < − 24 / 5. \left(L_\lambda(\operatorname{diag}(1,-1,0))\right)_{11}<-7/5,
\qquad
\left(L_\lambda(e_1e_2^T+e_2e_1^T)\right)_{12}<-24/5. ( L λ ( diag ( 1 , − 1 , 0 )) ) 11 < − 7/5 , ( L λ ( e 1 e 2 T + e 2 e 1 T ) ) 12 < − 24/5.
Coordinate reflection gives
( B m ( K ) ) i j = 1 8 ∫ Ω + p 1 2 ∑ s ∈ { − 1 , 1 } 3 s i s j ∑ T ∈ T + × ∫ T ( ∇ p i T C m ( ( s a s b K a b ) a , b = 1 3 ) ∇ p j ) − r [ z m ] det ( I + z E T ( ( s a s b K a b ) a , b = 1 3 ) p i p j ) , (57) \begin{aligned}
(B_m(K))_{ij}
&=\frac{1}{8\int_{\Omega_+}p_1^2}
\sum_{s\in\{-1,1\}^3}s_i s_j
\sum_{T\in\mathcal{T}_+}\\
&\quad\times\int_T
\left(\nabla p_i^T C_m\left((s_as_bK_{ab})_{a,b=1}^{3}\right)\nabla p_j\right)\\
&\quad-r\left[z^m\right]\det\left(I+zE_T\left((s_as_bK_{ab})_{a,b=1}^{3}\right)p_ip_j\right), \tag*{(57)}
\end{aligned} ( B m ( K ) ) ij = 8 ∫ Ω + p 1 2 1 s ∈ { − 1 , 1 } 3 ∑ s i s j T ∈ T + ∑ × ∫ T ( ∇ p i T C m ( ( s a s b K ab ) a , b = 1 3 ) ∇ p j ) − r [ z m ] det ( I + z E T ( ( s a s b K ab ) a , b = 1 3 ) p i p j ) , (57)
where the coefficient fields use E T ( ( s a s b K a b ) a , b = 1 3 ) E_T((s_as_bK_{ab})_{a,b=1}^3) E T (( s a s b K ab ) a , b = 1 3 ) from (40) . For a distinguished index d 0 d_0 d 0 and a < b a < b a < b with { a , b , d 0 } = { 1 , 2 , 3 } \{a,b,d_0\}=\{1,2,3\} { a , b , d 0 } = { 1 , 2 , 3 } , use
X = ( K d 0 d 0 , K a a − K b b , K 23 , K 13 , K 12 ) , (58) X=(K_{d_0d_0},K_{aa}-K_{bb},K_{23},K_{13},K_{12}), \tag*{(58)} X = ( K d 0 d 0 , K aa − K bb , K 23 , K 13 , K 12 ) , (58)
K a a = ( − X 1 + X 2 ) / 2 , K b b = ( − X 1 − X 2 ) / 2 , K_{aa}=(-X_1+X_2)/2,\qquad K_{bb}=(-X_1-X_2)/2, K aa = ( − X 1 + X 2 ) /2 , K bb = ( − X 1 − X 2 ) /2 ,
36 ∥ p − p ∗ ∥ 2 2 = 12 X 1 2 + 4 X 2 2 + 8 ∑ i = 3 5 X i 2 . 36\lVert p-p_*\rVert_2^2=12X_1^2+4X_2^2+8\sum_{i=3}^{5}X_i^2. 36 ∥ p − p ∗ ∥ 2 2 = 12 X 1 2 + 4 X 2 2 + 8 i = 3 ∑ 5 X i 2 .
Reflections K i j ↦ s i s j K i j K_{ij}\mapsto s_is_jK_{ij} K ij ↦ s i s j K ij and permutations imply, for { i , j , k } = { 1 , 2 , 3 } \{i,j,k\}=\{1,2,3\} { i , j , k } = { 1 , 2 , 3 } ,
( B 2 ) i i = α K i i 2 + β ( K j j − K k k ) 2 + γ 2 ( K i j 2 + K i k 2 ) + δ 2 K j k 2 , (B_2)_{ii}=\alpha K_{ii}^2+\beta(K_{jj}-K_{kk})^2+\gamma_2(K_{ij}^2+K_{ik}^2)+\delta_2K_{jk}^2, ( B 2 ) ii = α K ii 2 + β ( K j j − K k k ) 2 + γ 2 ( K ij 2 + K ik 2 ) + δ 2 K j k 2 ,
( B 2 ) i j = η K k k K i j + ζ K i k K j k ( i ≠ j ) . (59) (B_2)_{ij}=\eta K_{kk}K_{ij}+\zeta K_{ik}K_{jk}\qquad(i\ne j). \tag*{(59)} ( B 2 ) ij = η K k k K ij + ζ K ik K j k ( i = j ) . (59)
In (57) , extract ( X 1 2 , X 2 2 , X 4 2 , X 3 2 ) (X_1^2,X_2^2,X_4^2,X_3^2) ( X 1 2 , X 2 2 , X 4 2 , X 3 2 ) from entry 11 with d 0 = 1 d_0=1 d 0 = 1 , and ( X 1 X 5 , X 3 X 4 ) (X_1X_5,X_3X_4) ( X 1 X 5 , X 3 X 4 ) from entry 12 with d 0 = 3 d_0=3 d 0 = 3 . The resulting enclosures are
( 4853 , 2868 , 5418 , 2931 , − 9533 , 2691 ) < 1000 ( α , β , γ 2 , δ 2 , η , ζ ) < ( 4854 , 2869 , 5419 , 2932 , − 9532 , 2692 ) . (60) \begin{aligned}
(4853,2868,5418,2931,-9533,2691)<1000(\alpha,\beta,\gamma_2,\delta_2,\eta,\zeta) \\
&< (4854,2869,5419,2932,-9532,2692). \tag*{(60)}
\end{aligned} ( 4853 , 2868 , 5418 , 2931 , − 9533 , 2691 ) < 1000 ( α , β , γ 2 , δ 2 , η , ζ ) < ( 4854 , 2869 , 5419 , 2932 , − 9532 , 2692 ) . (60)
For ∣ ξ ∣ = 1 |\xi|=1 ∣ ξ ∣ = 1 and { i , j , k } = { 1 , 2 , 3 } \{i,j,k\}=\{1,2,3\} { i , j , k } = { 1 , 2 , 3 } ,
∑ i = 1 3 ξ i 2 ( α K i i 2 + β ( K j j − K k k ) 2 ) = 2 β ∑ i = 1 3 K i i 2 + ( α − 3 β ) ∑ i = 1 3 K i i 2 ξ i 2 ≤ 2 β ∑ i = 1 3 K i i 2 , \begin{aligned}
\sum_{i=1}^{3}\xi_i^2\left(\alpha K_{ii}^2+\beta(K_{jj}-K_{kk})^2\right) \\
&=2\beta\sum_{i=1}^{3}K_{ii}^2+(\alpha-3\beta)\sum_{i=1}^{3}K_{ii}^2\xi_i^2\le2\beta\sum_{i=1}^{3}K_{ii}^2,
\end{aligned} i = 1 ∑ 3 ξ i 2 ( α K ii 2 + β ( K j j − K k k ) 2 ) = 2 β i = 1 ∑ 3 K ii 2 + ( α − 3 β ) i = 1 ∑ 3 K ii 2 ξ i 2 ≤ 2 β i = 1 ∑ 3 K ii 2 ,
( K 23 ξ 1 + K 13 ξ 2 + K 12 ξ 3 ) 2 (K_{23}\xi_1+K_{13}\xi_2+K_{12}\xi_3)^2 ( K 23 ξ 1 + K 13 ξ 2 + K 12 ξ 3 ) 2
≤ 3 ( K 23 2 ξ 1 2 + K 13 2 ξ 2 2 + K 12 2 ξ 3 2 ) , \le3(K_{23}^2\xi_1^2+K_{13}^2\xi_2^2+K_{12}^2\xi_3^2), ≤ 3 ( K 23 2 ξ 1 2 + K 13 2 ξ 2 2 + K 12 2 ξ 3 2 ) ,
( K 23 ξ 1 + K 13 ξ 2 + K 12 ξ 3 ) 2 ≤ ∑ 1 ≤ i < j ≤ 3 K i j 2 , (K_{23}\xi_1+K_{13}\xi_2+K_{12}\xi_3)^2\le\sum_{1\le i<j\le3}K_{ij}^2, ( K 23 ξ 1 + K 13 ξ 2 + K 12 ξ 3 ) 2 ≤ 1 ≤ i < j ≤ 3 ∑ K ij 2 ,
δ 2 − γ 2 − ζ < 0 , δ 2 − γ 2 + 2 ζ > 0. \delta_2-\gamma_2-\zeta<0,\qquad\delta_2-\gamma_2+2\zeta>0. δ 2 − γ 2 − ζ < 0 , δ 2 − γ 2 + 2 ζ > 0.
Hence
γ 2 ∑ 1 ≤ i < j ≤ 3 K i j 2 + ζ ( K 23 ξ 1 + K 13 ξ 2 + K 12 ξ 3 ) 2 + ( δ 2 − γ 2 − ζ ) ( K 23 2 ξ 1 2 + K 13 2 ξ 2 2 + K 12 2 ξ 3 2 ) ≤ γ 2 ∑ 1 ≤ i < j ≤ 3 K i j 2 + δ 2 − γ 2 + 2 ζ 3 ( K 23 ξ 1 + K 13 ξ 2 + K 12 ξ 3 ) 2 ≤ 2 γ 2 + δ 2 + 2 ζ 3 ∑ 1 ≤ i < j ≤ 3 K i j 2 . \begin{aligned}
\gamma_2\sum_{1\le i<j\le3}K_{ij}^2+\zeta(K_{23}\xi_1+K_{13}\xi_2+K_{12}\xi_3)^2 \\
&+(\delta_2-\gamma_2-\zeta)(K_{23}^2\xi_1^2+K_{13}^2\xi_2^2+K_{12}^2\xi_3^2) \\
&\le\gamma_2\sum_{1\le i<j\le3}K_{ij}^2+\frac{\delta_2-\gamma_2+2\zeta}{3}(K_{23}\xi_1+K_{13}\xi_2+K_{12}\xi_3)^2 \\
&\le\frac{2\gamma_2+\delta_2+2\zeta}{3}\sum_{1\le i<j\le3}K_{ij}^2.
\end{aligned} γ 2 1 ≤ i < j ≤ 3 ∑ K ij 2 + ζ ( K 23 ξ 1 + K 13 ξ 2 + K 12 ξ 3 ) 2 + ( δ 2 − γ 2 − ζ ) ( K 23 2 ξ 1 2 + K 13 2 ξ 2 2 + K 12 2 ξ 3 2 ) ≤ γ 2 1 ≤ i < j ≤ 3 ∑ K ij 2 + 3 δ 2 − γ 2 + 2 ζ ( K 23 ξ 1 + K 13 ξ 2 + K 12 ξ 3 ) 2 ≤ 3 2 γ 2 + δ 2 + 2 ζ 1 ≤ i < j ≤ 3 ∑ K ij 2 .
If ∑ a = 1 3 K a a 2 > 0 \sum_{a=1}^{3}K_{aa}^2>0 ∑ a = 1 3 K aa 2 > 0 , then
K i i 2 = ( K j j + K k k ) 2 ≤ 2 ( K j j 2 + K k k 2 ) , K i i 2 ∑ a = 1 3 K a a 2 ≤ 2 3 . K_{ii}^2=(K_{jj}+K_{kk})^2\le2(K_{jj}^2+K_{kk}^2),\qquad
\frac{K_{ii}^2}{\sum_{a=1}^{3}K_{aa}^2}\le\frac{2}{3}. K ii 2 = ( K j j + K k k ) 2 ≤ 2 ( K j j 2 + K k k 2 ) , ∑ a = 1 3 K aa 2 K ii 2 ≤ 3 2 .
Ordering the coordinate squares of ξ \xi ξ as r 1 ≥ r 2 ≥ r 3 r_1\ge r_2\ge r_3 r 1 ≥ r 2 ≥ r 3 gives
K 11 2 ξ 2 2 ξ 3 2 + K 22 2 ξ 1 2 ξ 3 2 + K 33 2 ξ 1 2 ξ 2 2 ∑ a = 1 3 K a a 2 ≤ 2 r 1 r 2 + r 1 r 3 3 ≤ 2 3 r 1 ( 1 − r 1 ) ≤ 1 6 . \frac{K_{11}^2\xi_2^2\xi_3^2+K_{22}^2\xi_1^2\xi_3^2+K_{33}^2\xi_1^2\xi_2^2}{\sum_{a=1}^{3}K_{aa}^2}
\le\frac{2r_1r_2+r_1r_3}{3}\le\frac{2}{3}r_1(1-r_1)\le\frac{1}{6}. ∑ a = 1 3 K aa 2 K 11 2 ξ 2 2 ξ 3 2 + K 22 2 ξ 1 2 ξ 3 2 + K 33 2 ξ 1 2 ξ 2 2 ≤ 3 2 r 1 r 2 + r 1 r 3 ≤ 3 2 r 1 ( 1 − r 1 ) ≤ 6 1 .
The mixed term satisfies
2 ∣ η ∣ ∣ K 11 K 23 ξ 3 ξ 3 + K 22 K 13 ξ 1 ξ 3 + K 33 K 12 ξ 1 ξ 2 ∣ 2|\eta|\left|K_{11}K_{23}\xi_3\xi_3+K_{22}K_{13}\xi_1\xi_3+K_{33}K_{12}\xi_1\xi_2\right| 2∣ η ∣ ∣ K 11 K 23 ξ 3 ξ 3 + K 22 K 13 ξ 1 ξ 3 + K 33 K 12 ξ 1 ξ 2 ∣
≤ 2 ∣ η ∣ ( ∑ 1 ≤ i < j ≤ 3 K i j 2 ) 1 / 2 ( K 11 2 ξ 2 2 ξ 3 2 + K 22 2 ξ 1 2 ξ 3 2 + K 33 2 ξ 1 2 ξ 2 2 ) 1 / 2 ≤ ∣ η ∣ ( 2 3 ( ∑ i = 1 3 K i i 2 ) ( ∑ 1 ≤ i < j ≤ 3 K i j 2 ) ) 1 / 2 . \begin{aligned}
\le2|\eta|\left(\sum_{1\le i<j\le3}K_{ij}^{2}\right)^{1/2}\left(K_{11}^{2}\xi_2^{2}\xi_3^{2}+K_{22}^{2}\xi_1^{2}\xi_3^{2}+K_{33}^{2}\xi_1^{2}\xi_2^{2}\right)^{1/2} \\
&\le|\eta|\left(\frac{2}{3}\left(\sum_{i=1}^{3}K_{ii}^{2}\right)\left(\sum_{1\le i<j\le3}K_{ij}^{2}\right)\right)^{1/2}.
\end{aligned} ≤ 2∣ η ∣ ( 1 ≤ i < j ≤ 3 ∑ K ij 2 ) 1/2 ( K 11 2 ξ 2 2 ξ 3 2 + K 22 2 ξ 1 2 ξ 3 2 + K 33 2 ξ 1 2 ξ 2 2 ) 1/2 ≤ ∣ η ∣ ( 3 2 ( i = 1 ∑ 3 K ii 2 ) ( 1 ≤ i < j ≤ 3 ∑ K ij 2 ) ) 1/2 .
Consequently,
ξ T B 2 ξ ≤ 2 β ∑ i = 1 3 K i i 2 + 2 γ 2 + δ 2 + 2 ζ 3 ∑ 1 ≤ i < j ≤ 3 K i j 2 + ∣ η ∣ ( 2 3 ( ∑ i = 1 3 K i i 2 ) ( ∑ 1 ≤ i < j ≤ 3 K i j 2 ) ) 1 / 2 . \begin{aligned}
\xi^{T}B_{2}\xi\le2\beta\sum_{i=1}^{3}K_{ii}^{2}+\frac{2\gamma_{2}+\delta_{2}+2\zeta}{3}\sum_{1\le i<j\le3}K_{ij}^{2} \\
&+|\eta|\left(\frac{2}{3}\left(\sum_{i=1}^{3}K_{ii}^{2}\right)\left(\sum_{1\le i<j\le3}K_{ij}^{2}\right)\right)^{1/2}.
\end{aligned} ξ T B 2 ξ ≤ 2 β i = 1 ∑ 3 K ii 2 + 3 2 γ 2 + δ 2 + 2 ζ 1 ≤ i < j ≤ 3 ∑ K ij 2 + ∣ η ∣ ( 3 2 ( i = 1 ∑ 3 K ii 2 ) ( 1 ≤ i < j ≤ 3 ∑ K ij 2 ) ) 1/2 .
By (60) ,
10 − 2 β > 0 , det ( 10 I − ( 2 β ∣ η ∣ / 6 ∣ η ∣ / 6 ( 2 γ 2 + δ 2 + 2 ζ ) / 3 ) ) > 0. 10-2\beta>0,\qquad\det\left(10I-\begin{pmatrix}2\beta&|\eta|/\sqrt{6}\\|\eta|/\sqrt{6}&(2\gamma_{2}+\delta_{2}+2\zeta)/3\end{pmatrix}\right)>0. 10 − 2 β > 0 , det ( 10 I − ( 2 β ∣ η ∣/ 6 ∣ η ∣/ 6 ( 2 γ 2 + δ 2 + 2 ζ ) /3 ) ) > 0.
Thus
ξ T B 2 ξ < 10 ( ∑ i = 1 3 K i i 2 + ∑ 1 ≤ i < j ≤ 3 K i j 2 ) = 45 ∥ p − p ∗ ∥ 2 2 . \xi^{T}B_{2}\xi<10\left(\sum_{i=1}^{3}K_{ii}^{2}+\sum_{1\le i<j\le3}K_{ij}^{2}\right)=45\|p-p_{*}\|_{2}^{2}. ξ T B 2 ξ < 10 ( i = 1 ∑ 3 K ii 2 + 1 ≤ i < j ≤ 3 ∑ K ij 2 ) = 45∥ p − p ∗ ∥ 2 2 .
The same symmetries give
( B 3 ) 11 = a 1 K 11 3 + a 2 K 11 ( K 22 − K 33 ) 2 (B_{3})_{11}=a_{1}K_{11}^{3}+a_{2}K_{11}(K_{22}-K_{33})^{2} ( B 3 ) 11 = a 1 K 11 3 + a 2 K 11 ( K 22 − K 33 ) 2
+ a 3 K 11 ( K 12 2 + K 13 2 ) + a 4 K 11 K 23 2 \qquad+a_{3}K_{11}(K_{12}^{2}+K_{13}^{2})+a_{4}K_{11}K_{23}^{2} + a 3 K 11 ( K 12 2 + K 13 2 ) + a 4 K 11 K 23 2
+ a 5 ( K 22 − K 33 ) ( K 12 2 − K 13 2 ) + a 6 K 12 K 13 K 23 , (61) \qquad+a_{5}(K_{22}-K_{33})(K_{12}^{2}-K_{13}^{2})+a_{6}K_{12}K_{13}K_{23}, \tag*{(61)} + a 5 ( K 22 − K 33 ) ( K 12 2 − K 13 2 ) + a 6 K 12 K 13 K 23 , (61)
( B 3 ) 12 = b 1 K 33 2 K 12 + b 2 ( K 11 − K 22 ) 2 K 12 (B_{3})_{12}=b_{1}K_{33}^{2}K_{12}+b_{2}(K_{11}-K_{22})^{2}K_{12} ( B 3 ) 12 = b 1 K 33 2 K 12 + b 2 ( K 11 − K 22 ) 2 K 12
+ b 3 K 33 K 13 K 23 + b 4 K 12 3 + b 5 K 12 ( K 13 2 + K 23 2 ) . \qquad+b_{3}K_{33}K_{13}K_{23}+b_{4}K_{12}^{3}+b_{5}K_{12}(K_{13}^{2}+K_{23}^{2}). + b 3 K 33 K 13 K 23 + b 4 K 12 3 + b 5 K 12 ( K 13 2 + K 23 2 ) .
Use d 0 = 1 d_{0}=1 d 0 = 1 in entry 11, d 0 = 3 d_{0}=3 d 0 = 3 in entry 12, and extract the first displayed monomial of each factor (X 2 X 5 2 X_{2}X_{5}^{2} X 2 X 5 2 for a 5 a_{5} a 5 ). Summation gives
1000 ( ∣ a 1 ∣ , … , ∣ a 6 ∣ ) < ( 4430 , 6415 , 15209 , 7801 , 15233 , 3184 ) , 1000(|a_{1}|,\ldots,|a_{6}|)<(4430,6415,15209,7801,15233,3184), 1000 ( ∣ a 1 ∣ , … , ∣ a 6 ∣ ) < ( 4430 , 6415 , 15209 , 7801 , 15233 , 3184 ) ,
1000 ( ∣ b 1 ∣ , … , ∣ b 5 ∣ ) < ( 26526 , 223 , 4934 , 8108 , 951 ) . (62) 1000(|b_{1}|,\ldots,|b_{5}|)<(26526,223,4934,8108,951). \tag*{(62)} 1000 ( ∣ b 1 ∣ , … , ∣ b 5 ∣ ) < ( 26526 , 223 , 4934 , 8108 , 951 ) . (62)
For each cubic entry, the multinomial formula and Cauchy–Schwarz give
46656 ∥ p − p ∗ ∥ 2 6 = ∑ ν ∈ N 0 5 ∣ ν ∣ = 3 ( 3 ν ) 12 ν 1 4 ν 2 8 ν 3 + ν 4 + ν 5 X 2 ν , 46656\|p-p_{*}\|_{2}^{6}=\sum_{\substack{\nu\in\mathbb{N}_{0}^{5}\\|\nu|=3}}\binom{3}{\nu}12^{\nu_{1}}4^{\nu_{2}}8^{\nu_{3}+\nu_{4}+\nu_{5}}X^{2\nu}, 46656∥ p − p ∗ ∥ 2 6 = ν ∈ N 0 5 ∣ ν ∣ = 3 ∑ ( ν 3 ) 1 2 ν 1 4 ν 2 8 ν 3 + ν 4 + ν 5 X 2 ν ,
∣ ∑ ν ∈ N 0 5 ∣ ν ∣ = 3 a ν X ν ∣ 2 ≤ 46656 ∥ p − p ∗ ∥ 2 6 ∑ ν ∈ N 0 5 ∣ ν ∣ = 3 a ν 2 ( 3 ν ) 12 ν 1 4 ν 2 8 ν 3 + ν 4 + ν 5 . \left|\sum_{\substack{\nu\in\mathbb{N}_{0}^{5}\\|\nu|=3}}a_{\nu}X^{\nu}\right|^{2}\le46656\|p-p_{*}\|_{2}^{6}\sum_{\substack{\nu\in\mathbb{N}_{0}^{5}\\|\nu|=3}}\frac{a_{\nu}^{2}}{\binom{3}{\nu}12^{\nu_{1}}4^{\nu_{2}}8^{\nu_{3}+\nu_{4}+\nu_{5}}}. ν ∈ N 0 5 ∣ ν ∣ = 3 ∑ a ν X ν 2 ≤ 46656∥ p − p ∗ ∥ 2 6 ν ∈ N 0 5 ∣ ν ∣ = 3 ∑ ( ν 3 ) 1 2 ν 1 4 ν 2 8 ν 3 + ν 4 + ν 5 a ν 2 .
The sums for entries 11 and 12 satisfy, respectively,
1 512 ( 8 27 a 1 2 + 8 9 a 2 2 + 4 9 a 3 2 + 2 9 a 4 2 + 4 3 a 5 2 + 1 6 a 6 2 ) < 459 500 . \frac{1}{512}\left(\frac{8}{27}a_{1}^{2}+\frac{8}{9}a_{2}^{2}+\frac{4}{9}a_{3}^{2}+\frac{2}{9}a_{4}^{2}+\frac{4}{3}a_{5}^{2}+\frac{1}{6}a_{6}^{2}\right)<\frac{459}{500}. 512 1 ( 27 8 a 1 2 + 9 8 a 2 2 + 9 4 a 3 2 + 9 2 a 4 2 + 3 4 a 5 2 + 6 1 a 6 2 ) < 500 459 .
1 512 ( 4 27 b 1 2 + 4 3 b 2 2 + 1 9 b 3 2 + b 4 2 + 2 3 b 5 2 ) < 339 1000 . \frac{1}{512}\left(\frac{4}{27}b_1^2+\frac{4}{3}b_2^2+\frac{1}{9}b_3^2+b_4^2+\frac{2}{3}b_5^2\right)<\frac{339}{1000}. 512 1 ( 27 4 b 1 2 + 3 4 b 2 2 + 9 1 b 3 2 + b 4 2 + 3 2 b 5 2 ) < 1000 339 .
Permutation symmetry gives
∥ B 3 ∥ F 2 = ∑ i = 1 3 ( B 3 ) i i 2 + 2 ∑ 1 ≤ i < j ≤ 3 ( B 3 ) i j 2 < ( 2376 5 ) 2 ∥ p − p ∗ ∥ 2 6 . \begin{aligned}
\lVert B_3\rVert_F^2=\sum_{i=1}^{3}(B_3)_{ii}^2+2\sum_{1\le i<j\le3}(B_3)_{ij}^2 \\
&<\left(\frac{2376}{5}\right)^2\lVert p-p_*\rVert_2^6.
\end{aligned} ∥ B 3 ∥ F 2 = i = 1 ∑ 3 ( B 3 ) ii 2 + 2 1 ≤ i < j ≤ 3 ∑ ( B 3 ) ij 2 < ( 5 2376 ) 2 ∥ p − p ∗ ∥ 2 6 .
Proof of Proposition 4.1. Since λ min ( Q ) ≤ − ∥ Q ∥ F / 6 \lambda_{\min}(Q)\le-\lVert Q\rVert_F/\sqrt{6} λ m i n ( Q ) ≤ − ∥ Q ∥ F / 6 for Q = Q T Q=Q^T Q = Q T with tr Q = 0 \operatorname{tr}Q=0 tr Q = 0 , equations (37) and (52) give
λ min ( L θ ( K / 2 ) ) ≤ − ∥ L θ ( K / 2 ) ∥ F 6 ≤ − 3 4 ∥ p − p ∗ ∥ 2 min { ∣ ( L θ ( diag ( 1 , − 1 , 0 ) ) ) 11 ∣ , 2 ∣ ( L θ ( e 1 e 2 T + e 2 e 1 T ) ) 12 ∣ } . \begin{aligned}
\lambda_{\min}(L_\theta(K/2))\le-\frac{\lVert L_\theta(K/2)\rVert_F}{\sqrt{6}} \\
&\le-\frac{\sqrt{3}}{4}\lVert p-p_*\rVert_2\min\left\{\left|(L_\theta(\operatorname{diag}(1,-1,0)))_{11}\right|,\sqrt{2}\left|(L_\theta(e_1e_2^T+e_2e_1^T))_{12}\right|\right\}.
\end{aligned} λ m i n ( L θ ( K /2 )) ≤ − 6 ∥ L θ ( K /2 ) ∥ F ≤ − 4 3 ∥ p − p ∗ ∥ 2 min { ∣ ( L θ ( diag ( 1 , − 1 , 0 )) ) 11 ∣ , 2 ( L θ ( e 1 e 2 T + e 2 e 1 T ) ) 12 } .
Numerical Lemma 2 gives
λ min ( L λ ( K / 2 ) ) < − 291 500 ∥ p − p ∗ ∥ 2 , λ min ( L r ( K / 2 ) ) < − 21 25 ∥ p − p ∗ ∥ 2 . (63) \lambda_{\min}(L_\lambda(K/2))<-\frac{291}{500}\lVert p-p_*\rVert_2,\qquad\lambda_{\min}(L_r(K/2))<-\frac{21}{25}\lVert p-p_*\rVert_2. \tag*{(63)} λ m i n ( L λ ( K /2 )) < − 500 291 ∥ p − p ∗ ∥ 2 , λ m i n ( L r ( K /2 )) < − 25 21 ∥ p − p ∗ ∥ 2 . (63)
Using (53) , ∥ ∇ P ξ ∥ 2 < 21 / 10 \lVert\nabla P_\xi\rVert_2<21/10 ∥ ∇ P ξ ∥ 2 < 21/10 , and (51) , the bounds (55) give
B λ ( u ξ , u ξ ; C 2 , ( tr E ) 2 − tr ( E 2 ) 2 ) < 153 2 ∥ p − p ∗ ∥ 2 2 , \mathcal{B}_\lambda\left(u_\xi,u_\xi;C_2,\frac{(\operatorname{tr}E)^2-\operatorname{tr}(E^2)}{2}\right)<\frac{153}{2}\lVert p-p_*\rVert_2^2, B λ ( u ξ , u ξ ; C 2 , 2 ( tr E ) 2 − tr ( E 2 ) ) < 2 153 ∥ p − p ∗ ∥ 2 2 ,
B λ ( u ξ , u ξ ; C 3 , det E ) < 756 ∥ p − p ∗ ∥ 2 3 . \mathcal{B}_\lambda(u_\xi,u_\xi;C_3,\det E)<756\lVert p-p_*\rVert_2^3. B λ ( u ξ , u ξ ; C 3 , det E ) < 756 ∥ p − p ∗ ∥ 2 3 .
C − ∑ m = 0 3 C m = J ∑ m = 4 ∞ M m + ( J − 1 ) M 3 C-\sum_{m=0}^{3}C_m=J\sum_{m=4}^{\infty}M_m+(J-1)M_3 C − m = 0 ∑ 3 C m = J m = 4 ∑ ∞ M m + ( J − 1 ) M 3
+ ( ( tr E ) 2 − tr ( E 2 ) 2 + det E ) M 2 + ( det E ) M 1 , \qquad+\left(\frac{(\operatorname{tr}E)^2-\operatorname{tr}(E^2)}{2}+\det E\right)M_2+(\det E)M_1, + ( 2 ( tr E ) 2 − tr ( E 2 ) + det E ) M 2 + ( det E ) M 1 ,
∥ ∑ m = 4 ∞ M m ∥ o p ≤ ∑ m = 4 ∞ ( 1 + 3 m / 2 ) ( 6 ∥ p − p ∗ ∥ 2 ) m = 1296 ∥ p − p ∗ ∥ 2 4 ( 7 − 33 ∥ p − p ∗ ∥ 2 ) ( 1 − 6 ∥ p − p ∗ ∥ 2 ) 2 ≤ 9072 ∥ p − p ∗ ∥ 2 4 ( 1 − 6 ∥ p − p ∗ ∥ 2 ) 2 . \begin{aligned}
\left\lVert\sum_{m=4}^{\infty}M_m\right\rVert_{\mathrm{op}}\le\sum_{m=4}^{\infty}(1+3m/2)(6\lVert p-p_*\rVert_2)^m \\
&=\frac{1296\lVert p-p_*\rVert_2^4(7-33\lVert p-p_*\rVert_2)}{(1-6\lVert p-p_*\rVert_2)^2} \\
&\le\frac{9072\lVert p-p_*\rVert_2^4}{(1-6\lVert p-p_*\rVert_2)^2}.
\end{aligned} m = 4 ∑ ∞ M m op ≤ m = 4 ∑ ∞ ( 1 + 3 m /2 ) ( 6 ∥ p − p ∗ ∥ 2 ) m = ( 1 − 6 ∥ p − p ∗ ∥ 2 ) 2 1296 ∥ p − p ∗ ∥ 2 4 ( 7 − 33 ∥ p − p ∗ ∥ 2 ) ≤ ( 1 − 6 ∥ p − p ∗ ∥ 2 ) 2 9072 ∥ p − p ∗ ∥ 2 4 .
Since
∣ J − 1 ∣ ≤ 26 5 ∥ p − p ∗ ∥ 2 2 + 63 2 ∥ p − p ∗ ∥ 2 2 + 216 5 ∥ p − p ∗ ∥ 2 3 , |J-1|\le\frac{26}{5}\lVert p-p_*\rVert_2^2+\frac{63}{2}\lVert p-p_*\rVert_2^2+\frac{216}{5}\lVert p-p_*\rVert_2^3, ∣ J − 1∣ ≤ 5 26 ∥ p − p ∗ ∥ 2 2 + 2 63 ∥ p − p ∗ ∥ 2 2 + 5 216 ∥ p − p ∗ ∥ 2 3 ,
∥ C − ∑ m = 0 3 C m ∥ o p 1296 ∥ p − p ∗ ∥ 2 4 ≤ 7 ( 1 + ∣ J − 1 ∣ ) ( 1 − 6 ∥ p − p ∗ ∥ 2 2 ) 2 + 11 ∣ J − 1 ∣ 12 ∥ p − p ∗ ∥ 2 + 4 ( 7 8 + 6 5 ∥ p − p ∗ ∥ 2 ) + 1 2 < { 35 / 2 , ∥ p − p ∗ ∥ 2 ≤ 1 / 120 , 37 / 2 , ∥ p − p ∗ ∥ 2 ≤ 1 / 75. \frac{\left\|C-\sum_{m=0}^{3}C_m\right\|_{\mathrm{op}}}{1296\left\|p-p_*\right\|_2^4}
\leq\frac{7(1+\lvert J-1\rvert)}{(1-6\left\|p-p_*\right\|_2^2)^2}+\frac{11\lvert J-1\rvert}{12\left\|p-p_*\right\|_2}+4\left(\frac{7}{8}+\frac{6}{5}\left\|p-p_*\right\|_2\right)+\frac{1}{2}
<\begin{cases}
35/2, & \left\|p-p_*\right\|_2\leq1/120,\\
37/2, & \left\|p-p_*\right\|_2\leq1/75.
\end{cases} 1296 ∥ p − p ∗ ∥ 2 4 C − ∑ m = 0 3 C m op ≤ ( 1 − 6 ∥ p − p ∗ ∥ 2 2 ) 2 7 ( 1 + ∣ J − 1 ∣) + 12 ∥ p − p ∗ ∥ 2 11 ∣ J − 1 ∣ + 4 ( 8 7 + 5 6 ∥ p − p ∗ ∥ 2 ) + 2 1 < { 35/2 , 37/2 , ∥ p − p ∗ ∥ 2 ≤ 1/120 , ∥ p − p ∗ ∥ 2 ≤ 1/75.
For each unit mode f f f ,
∣ ∫ W ∇ f T ( C − ∑ m = 0 3 C m ) ∇ f ∣ ≤ 430113 100000 ∥ C − ∑ m = 0 3 C m ∥ o p < { 98496 ∥ p − p ∗ ∥ 2 4 , ∥ p − p ∗ ∥ 2 ≤ 1 / 120 , 103680 ∥ p − p ∗ ∥ 2 4 , ∥ p − p ∗ ∥ 2 ≤ 1 / 75. \left|\int_W \nabla f^T\left(C-\sum_{m=0}^{3}C_m\right)\nabla f\right|
\leq\frac{430113}{100000}\left\|C-\sum_{m=0}^{3}C_m\right\|_{\mathrm{op}}
<\begin{cases}
98496\left\|p-p_*\right\|_2^4, & \left\|p-p_*\right\|_2\leq1/120,\\
103680\left\|p-p_*\right\|_2^4, & \left\|p-p_*\right\|_2\leq1/75.
\end{cases} ∫ W ∇ f T ( C − m = 0 ∑ 3 C m ) ∇ f ≤ 100000 430113 C − m = 0 ∑ 3 C m op < { 98496 ∥ p − p ∗ ∥ 2 4 , 103680 ∥ p − p ∗ ∥ 2 4 , ∥ p − p ∗ ∥ 2 ≤ 1/120 , ∥ p − p ∗ ∥ 2 ≤ 1/75.
For 0 < ∥ p − p ∗ ∥ 2 ≤ 27 / 4000 0<\left\|p-p_*\right\|_2\leq27/4000 0 < ∥ p − p ∗ ∥ 2 ≤ 27/4000 , choose u ξ u_\xi u ξ to minimize the linear term. Then
B λ ( u ξ , u ξ ; C , J ) < − 291 500 ∥ p − p ∗ ∥ 2 + 153 2 ∥ p − p ∗ ∥ 2 2 + 756 ∥ p − p ∗ ∥ 2 3 + 98496 ∥ p − p ∗ ∥ 2 4 < 0. (64) B_\lambda(u_\xi,u_\xi;C,J)<-\frac{291}{500}\left\|p-p_*\right\|_2+\frac{153}{2}\left\|p-p_*\right\|_2^2+756\left\|p-p_*\right\|_2^3+98496\left\|p-p_*\right\|_2^4<0. \tag*{(64)} B λ ( u ξ , u ξ ; C , J ) < − 500 291 ∥ p − p ∗ ∥ 2 + 2 153 ∥ p − p ∗ ∥ 2 2 + 756 ∥ p − p ∗ ∥ 2 3 + 98496 ∥ p − p ∗ ∥ 2 4 < 0. (64)
The quotient by ∥ p − p ∗ ∥ 2 \left\|p-p_*\right\|_2 ∥ p − p ∗ ∥ 2 is increasing and negative at ∥ p − p ∗ ∥ 2 = 27 / 4000 \left\|p-p_*\right\|_2=27/4000 ∥ p − p ∗ ∥ 2 = 27/4000 . For 27 / 4000 ≤ ∥ p − p ∗ ∥ 2 ≤ 73 / 6000 27/4000\leq\left\|p-p_*\right\|_2\leq73/6000 27/4000 ≤ ∥ p − p ∗ ∥ 2 ≤ 73/6000 , choose P ξ P_\xi P ξ to minimize its linear term. Since
B 4300827 10 6 ( P ξ , P ξ ; C , J ) = B r ( P ξ , P ξ ; C , J ) + ( r − 4300827 10 6 ) ∫ W J P ξ 2 , B_{\frac{4300827}{10^6}}(P_\xi,P_\xi;C,J)=B_r(P_\xi,P_\xi;C,J)+\left(r-\frac{4300827}{10^6}\right)\int_W J P_\xi^2, B 1 0 6 4300827 ( P ξ , P ξ ; C , J ) = B r ( P ξ , P ξ ; C , J ) + ( r − 1 0 6 4300827 ) ∫ W J P ξ 2 ,
∫ W J P ξ 2 ≤ 1 + 57 10 ∥ p − p ∗ ∥ 2 , \int_W JP_\xi^2\leq1+\frac{57}{10}\left\|p-p_*\right\|_2, ∫ W J P ξ 2 ≤ 1 + 10 57 ∥ p − p ∗ ∥ 2 ,
we obtain
B 4300827 10 6 ( P ξ , P ξ ; C , J ) < − 21 25 ∥ p − p ∗ ∥ 2 + 45 ∥ p − p ∗ ∥ 2 2 + 2376 5 ∥ p − p ∗ ∥ 2 3 + 103680 ∥ p − p ∗ ∥ 2 4 + 303 10 6 ( 1 + 57 10 ∥ p − p ∗ ∥ 2 ) < 0. (65) B_{\frac{4300827}{10^6}}(P_\xi,P_\xi;C,J)<-\frac{21}{25}\left\|p-p_*\right\|_2+45\left\|p-p_*\right\|_2^2+\frac{2376}{5}\left\|p-p_*\right\|_2^3+103680\left\|p-p_*\right\|_2^4+\frac{303}{10^6}\left(1+\frac{57}{10}\left\|p-p_*\right\|_2\right)<0. \tag*{(65)} B 1 0 6 4300827 ( P ξ , P ξ ; C , J ) < − 25 21 ∥ p − p ∗ ∥ 2 + 45 ∥ p − p ∗ ∥ 2 2 + 5 2376 ∥ p − p ∗ ∥ 2 3 + 103680 ∥ p − p ∗ ∥ 2 4 + 1 0 6 303 ( 1 + 10 57 ∥ p − p ∗ ∥ 2 ) < 0. (65)
The last polynomial is convex and negative at both endpoints. Thus (4.7) gives
2 − 4 / 3 F ( p ) < { λ , 0 < ∥ p − p ∗ ∥ 2 ≤ 27 / 4000 , 4300827 10 6 , 27 / 4000 < ∥ p − p ∗ ∥ 2 ≤ 73 / 6000 , 2^{-4/3}F(p)<\begin{cases}
\lambda, & 0<\left\|p-p_*\right\|_2\leq27/4000,\\
\frac{4300827}{10^6}, & 27/4000<\left\|p-p_*\right\|_2\leq73/6000,
\end{cases} 2 − 4/3 F ( p ) < { λ , 1 0 6 4300827 , 0 < ∥ p − p ∗ ∥ 2 ≤ 27/4000 , 27/4000 < ∥ p − p ∗ ∥ 2 ≤ 73/6000 ,
4300827 10 6 < λ = 2 − 4 / 3 F ( p ∗ ) . (66) \frac{4300827}{10^6}<\lambda=2^{-4/3}F(p_*). \tag*{(66)} 1 0 6 4300827 < λ = 2 − 4/3 F ( p ∗ ) . (66)
The global comparison
Geometric exclusions and polynomial Rayleigh tests give the remaining comparison:
D ( p ) > 0 , ∥ p − p ∗ ∥ 2 > 73 6000 ⟹ F ( p ) < 10837 1000 = T < F ( p ∗ ) . (67) D(p)>0,\qquad\left\|p-p_*\right\|_2>\frac{73}{6000}
\quad\Longrightarrow\quad
F(p)<\frac{10837}{1000}=T<F(p_*). \tag*{(67)} D ( p ) > 0 , ∥ p − p ∗ ∥ 2 > 6000 73 ⟹ F ( p ) < 1000 10837 = T < F ( p ∗ ) . (67)
Together with Proposition 4.1, this proves Theorem 1.1. Relabel an edge of largest weight as 01. Its stabilizer acts transitively on 02, 03, 12, 13; permutation and rescaling give
F ( s p ) = F ( p ) ( s > 0 ) , q e = p e p 01 , q 02 = max ( q 02 , q 03 , q 12 , q 13 ) . F(sp)=F(p)\qquad(s>0),\qquad q_e=\frac{p_e}{p_{01}},\qquad q_{02}=\max(q_{02},q_{03},q_{12},q_{13}). F ( s p ) = F ( p ) ( s > 0 ) , q e = p 01 p e , q 02 = max ( q 02 , q 03 , q 12 , q 13 ) .
It therefore suffices to consider
C = { q : q 01 = 1 , 0 ≤ q e ≤ 1 ( e ∈ E ) , q 02 ≥ max ( q 03 , q 12 , q 13 ) } . (68) \mathcal{C}=\{q:q_{01}=1,\ 0\leq q_e\leq1\ (e\in E),\ q_{02}\geq\max(q_{03},q_{12},q_{13})\}. \tag*{(68)} C = { q : q 01 = 1 , 0 ≤ q e ≤ 1 ( e ∈ E ) , q 02 ≥ max ( q 03 , q 12 , q 13 )} . (68)
We evaluate F , D , C e , W S F,D,C_e,W_S F , D , C e , W S at these unnormalized conorms. For w = w 1 ⋯ w h ∈ { 0 , 1 } h w=w_1\cdots w_h\in\{0,1\}^h w = w 1 ⋯ w h ∈ { 0 , 1 } h , bisect 02, 03, 12, 13, 23 cyclically, choosing the lower half for 0 and the upper half for 1:
B w = { 1 } × ∏ i = 1 5 [ j i , j i + 2 − k i ] , B_w=\{1\}\times\prod_{i=1}^{5}[j_i,j_i+2^{-k_i}], B w = { 1 } × i = 1 ∏ 5 [ j i , j i + 2 − k i ] ,
k i = # ( { i , i + 5 , … } ∩ { 1 , … , h } ) , j i = ∑ r = 0 k i − 1 2 k i − 1 − r w i + 5 r . k_i=\#(\{i,i+5,\ldots\}\cap\{1,\ldots,h\}),\qquad j_i=\sum_{r=0}^{k_i-1}2^{k_i-1-r}w_{i+5r}. k i = # ({ i , i + 5 , … } ∩ { 1 , … , h }) , j i = r = 0 ∑ k i − 1 2 k i − 1 − r w i + 5 r .
Here k i = j i = 0 k_i=j_i=0 k i = j i = 0 for an empty subsequence; q − q^- q − , q + q^+ q + denote the box endpoints. We have
q 02 + < max ( q 03 − , q 12 − , q 13 − ) ⟹ B w ∩ C = ∅ . (69) q_{02}^{+}<\max(q_{03}^{-},q_{12}^{-},q_{13}^{-})\ \Longrightarrow\ B_w\cap\mathcal{C}=\varnothing. \tag*{(69)} q 02 + < max ( q 03 − , q 12 − , q 13 − ) ⟹ B w ∩ C = ∅ . (69)
max v ∈ Vert ( B w ) ( ∑ 0 ≤ i < j ≤ 3 v i j 2 ( ∑ 0 ≤ i < j ≤ 3 v i j ) 2 − 1 6 ) ≤ ( 73 6000 ) 2 , ⟹ ∥ q ∑ 0 ≤ i < j ≤ 3 q i j − p ∗ ∥ 2 ≤ 73 6000 ( q ∈ B w ) . (70) \begin{aligned}
\max_{v\in\operatorname{Vert}(B_w)}
\left(\frac{\sum_{0\leq i<j\leq3}v_{ij}^{2}}{\left(\sum_{0\leq i<j\leq3}v_{ij}\right)^2}-\frac{1}{6}\right)
&\leq\left(\frac{73}{6000}\right)^2,\\
\Longrightarrow\quad
\left\|\frac{q}{\sum_{0\leq i<j\leq3}q_{ij}}-p_*\right\|_2
&\leq\frac{73}{6000}\qquad(q\in B_w). \tag*{(70)}
\end{aligned} v ∈ Vert ( B w ) max ( ∑ 0 ≤ i < j ≤ 3 v ij ) 2 ∑ 0 ≤ i < j ≤ 3 v ij 2 − 6 1 ⟹ ∑ 0 ≤ i < j ≤ 3 q ij q − p ∗ 2 ≤ ( 6000 73 ) 2 , ≤ 6000 73 ( q ∈ B w ) . (70)
Indeed, for q = ∑ v ∈ Vert ( B w ) a v v q=\sum_{v\in\operatorname{Vert}(B_w)}a_vv q = ∑ v ∈ Vert ( B w ) a v v , a v ≥ 0 a_v\geq0 a v ≥ 0 , ∑ v ∈ Vert ( B w ) a v = 1 \sum_{v\in\operatorname{Vert}(B_w)}a_v=1 ∑ v ∈ Vert ( B w ) a v = 1 ,
q ∑ 0 ≤ i < j ≤ 3 q i j = ∑ v ∈ Vert ( B w ) a v ∑ 0 ≤ i < j ≤ 3 v i j ∑ 0 ≤ i < j ≤ 3 q i j v ∑ 0 ≤ i < j ≤ 3 v i j . \frac{q}{\sum_{0\leq i<j\leq3}q_{ij}}
=\sum_{v\in\operatorname{Vert}(B_w)}
\frac{a_v\sum_{0\leq i<j\leq3}v_{ij}}{\sum_{0\leq i<j\leq3}q_{ij}}
\frac{v}{\sum_{0\leq i<j\leq3}v_{ij}}. ∑ 0 ≤ i < j ≤ 3 q ij q = v ∈ Vert ( B w ) ∑ ∑ 0 ≤ i < j ≤ 3 q ij a v ∑ 0 ≤ i < j ≤ 3 v ij ∑ 0 ≤ i < j ≤ 3 v ij v .
Convexity of the squared norm therefore gives
∥ q ∑ 0 ≤ i < j ≤ 3 q i j − p ∗ ∥ 2 2 = ∑ 0 ≤ i < j ≤ 3 q i j 2 ( ∑ 0 ≤ i < j ≤ 3 q i j ) 2 − 1 6 ≤ max v ∈ Vert ( B w ) ∥ v ∑ 0 ≤ i < j ≤ 3 v i j − p ∗ ∥ 2 2 . \left\|\frac{q}{\sum_{0\leq i<j\leq3}q_{ij}}-p_*\right\|_2^2
=\frac{\sum_{0\leq i<j\leq3}q_{ij}^2}{\left(\sum_{0\leq i<j\leq3}q_{ij}\right)^2}-\frac{1}{6}
\leq\max_{v\in\operatorname{Vert}(B_w)}
\left\|\frac{v}{\sum_{0\leq i<j\leq3}v_{ij}}-p_*\right\|_2^2. ∑ 0 ≤ i < j ≤ 3 q ij q − p ∗ 2 2 = ( ∑ 0 ≤ i < j ≤ 3 q ij ) 2 ∑ 0 ≤ i < j ≤ 3 q ij 2 − 6 1 ≤ v ∈ Vert ( B w ) max ∑ 0 ≤ i < j ≤ 3 v ij v − p ∗ 2 2 .
The sixteen spanning trees of K 4 K_4 K 4 , in lexicographic order, are J 1 = { 01 , 02 , 03 } J_1=\{01,02,03\} J 1 = { 01 , 02 , 03 } , J 2 = { 01 , 02 , 13 } J_2=\{01,02,13\} J 2 = { 01 , 02 , 13 } , J 3 = { 01 , 02 , 23 } J_3=\{01,02,23\} J 3 = { 01 , 02 , 23 } , J 4 = { 01 , 03 , 12 } J_4=\{01,03,12\} J 4 = { 01 , 03 , 12 } , J 5 = { 01 , 03 , 23 } J_5=\{01,03,23\} J 5 = { 01 , 03 , 23 } , J 6 = { 01 , 12 , 13 } J_6=\{01,12,13\} J 6 = { 01 , 12 , 13 } , J 7 = { 01 , 12 , 23 } J_7=\{01,12,23\} J 7 = { 01 , 12 , 23 } , J 8 = { 01 , 13 , 23 } J_8=\{01,13,23\} J 8 = { 01 , 13 , 23 } , J 9 = { 02 , 03 , 12 } J_9=\{02,03,12\} J 9 = { 02 , 03 , 12 } , J 10 = { 02 , 03 , 13 } J_{10}=\{02,03,13\} J 10 = { 02 , 03 , 13 } , J 11 = { 02 , 12 , 13 } J_{11}=\{02,12,13\} J 11 = { 02 , 12 , 13 } , J 12 = { 02 , 12 , 23 } J_{12}=\{02,12,23\} J 12 = { 02 , 12 , 23 } , J 13 = { 02 , 13 , 23 } J_{13}=\{02,13,23\} J 13 = { 02 , 13 , 23 } , J 14 = { 03 , 12 , 13 } J_{14}=\{03,12,13\} J 14 = { 03 , 12 , 13 } , J 15 = { 03 , 12 , 23 } J_{15}=\{03,12,23\} J 15 = { 03 , 12 , 23 } , J 16 = { 03 , 13 , 23 } J_{16}=\{03,13,23\} J 16 = { 03 , 13 , 23 } .
Cauchy–Binet and the matrix–tree identity [5 ] give
D ( q ) = ∑ J ⊆ E ∣ J ∣ = 3 ( det [ d e ] e ∈ J ) 2 ∏ e ∈ J q e = ∑ r = 1 16 ∏ e ∈ J r q e . (71) D(q)=\sum_{\substack{J\subseteq E\\\lvert J\rvert=3}}\left(\det[d_e]_{e\in J}\right)^2\prod_{e\in J}q_e=\sum_{r=1}^{16}\prod_{e\in J_r}q_e. \tag*{(71)} D ( q ) = J ⊆ E ∣ J ∣ = 3 ∑ ( det [ d e ] e ∈ J ) 2 e ∈ J ∏ q e = r = 1 ∑ 16 e ∈ J r ∏ q e . (71)
C i j ( q ) = ∂ q i j D ( q ) = d i j T adj ( G ( q ) ) d i j ( 0 ≤ i < j ≤ 3 ) . C_{ij}(q)=\partial_{q_{ij}}D(q)=d_{ij}^{T}\operatorname{adj}(G(q))d_{ij}\qquad(0\leq i<j\leq3). C ij ( q ) = ∂ q ij D ( q ) = d ij T adj ( G ( q )) d ij ( 0 ≤ i < j ≤ 3 ) .
By (2.13), (2.19), and (2.20),
C e ( q ) 3 D ( q ) 2 < 132379 100000 or W S ( q ) 3 D ( q ) < 28 25 , \frac{C_e(q)^3}{D(q)^2}<\frac{132379}{100000}\quad\text{or}\quad\frac{W_S(q)^3}{D(q)}<\frac{28}{25}, D ( q ) 2 C e ( q ) 3 < 100000 132379 or D ( q ) W S ( q ) 3 < 25 28 ,
D ( q ) > 0 ⟹ F ( q ) < T . (72) D(q)>0\quad\Longrightarrow\quad F(q)<T. \tag*{(72)} D ( q ) > 0 ⟹ F ( q ) < T . (72)
For a , b , c , d , z ≥ 0 a,b,c,d,z \ge0 a , b , c , d , z ≥ 0 and c + d z > 0 c+dz>0 c + d z > 0 ,
d 2 d z 2 ( a + b z ) 3 ( c + d z ) m ≥ 0 , m ∈ { 1 , 2 } . \frac{\mathrm{d}^{2}}{\mathrm{d}z^{2}}\frac{(a+bz)^{3}}{(c+dz)^{m}}\ge0,\qquad m\in\{1,2\}. d z 2 d 2 ( c + d z ) m ( a + b z ) 3 ≥ 0 , m ∈ { 1 , 2 } .
Their maxima are therefore attained at corners:
max q ∈ B w C e ( q ) 3 D ( q ) 2 = max v ∈ Vert ( B w ) C e ( v ) 3 D ( v ) 2 , max q ∈ B w W S ( q ) 3 D ( q ) 2 = max v ∈ Vert ( B w ) W S ( v ) 3 D ( v ) 2 . \max_{q\in B_w}\frac{C_e(q)^{3}}{D(q)^{2}}=\max_{v\in\operatorname{Vert}(B_w)}\frac{C_e(v)^{3}}{D(v)^{2}},\qquad
\max_{q\in B_w}\frac{W S(q)^{3}}{D(q)^{2}}=\max_{v\in\operatorname{Vert}(B_w)}\frac{W S(v)^{3}}{D(v)^{2}}. q ∈ B w max D ( q ) 2 C e ( q ) 3 = v ∈ Vert ( B w ) max D ( v ) 2 C e ( v ) 3 , q ∈ B w max D ( q ) 2 W S ( q ) 3 = v ∈ Vert ( B w ) max D ( v ) 2 W S ( v ) 3 .
For cut tests require min v ∈ Vert ( B w ) D ( v ) > 0 \min_{v\in\operatorname{Vert}(B_w)}D(v)>0 min v ∈ Vert ( B w ) D ( v ) > 0 ; for edge tests require q e > 0 q_e>0 q e > 0 and use the continuous extension by zero at D = 0 D=0 D = 0 :
D ( q ) = D ( q ) ∣ q e = 0 + q e C e ( q ) ≥ q e C e ( q ) , 0 ≤ C e 3 D 2 ≤ D ( q e − ) 3 ⟶ 0. D(q)=D(q)|_{q_e=0}+q_eC_e(q)\ge q_eC_e(q),\qquad0\le\frac{C_e^{3}}{D^{2}}\le\frac{D}{(q_e^{-})^{3}}\longrightarrow0. D ( q ) = D ( q ) ∣ q e = 0 + q e C e ( q ) ≥ q e C e ( q ) , 0 ≤ D 2 C e 3 ≤ ( q e − ) 3 D ⟶ 0.
For an odd trial in (2.6), let M , K M,K M , K be its mass and energy after canceling their common positive Jacobian. For the box center c c c ,
F ( q ) ≤ D ( q ) 1 / 3 K ( q ) M ( q ) ( M > 0 ) , F(q)\le D(q)^{1/3}\frac{K(q)}{M(q)}\qquad(M>0), F ( q ) ≤ D ( q ) 1/3 M ( q ) K ( q ) ( M > 0 ) ,
∑ 0 ≤ i < j ≤ 3 C i j ( c ) q i j = tr { adj ( G ( c ) ) G ( q ) } , ∑ 0 ≤ i < j ≤ 3 C i j ( c ) c i j = 3 D ( c ) . \sum_{0\le i<j\le3}C_{ij}(c)q_{ij}=\operatorname{tr}\{\operatorname{adj}(G(c))G(q)\},\qquad
\sum_{0\le i<j\le3}C_{ij}(c)c_{ij}=3D(c). 0 ≤ i < j ≤ 3 ∑ C ij ( c ) q ij = tr { adj ( G ( c )) G ( q )} , 0 ≤ i < j ≤ 3 ∑ C ij ( c ) c ij = 3 D ( c ) .
The arithmetic–geometric mean inequality now gives
( D ( q ) D ( c ) ) 1 / 3 ≤ 1 3 tr { G ( c ) − 1 / 2 G ( q ) G ( c ) − 1 / 2 } = ∑ 0 ≤ i < j ≤ 3 C i j ( c ) q i j 3 D ( c ) , \left(\frac{D(q)}{D(c)}\right)^{1/3}\le\frac{1}{3}\operatorname{tr}\{G(c)^{-1/2}G(q)G(c)^{-1/2}\}=\frac{\sum_{0\le i<j\le3}C_{ij}(c)q_{ij}}{3D(c)}, ( D ( c ) D ( q ) ) 1/3 ≤ 3 1 tr { G ( c ) − 1/2 G ( q ) G ( c ) − 1/2 } = 3 D ( c ) ∑ 0 ≤ i < j ≤ 3 C ij ( c ) q ij ,
so either test
0 < τ 3 < 27 T 3 D ( c ) 2 , τ M − ( ∑ 0 ≤ i < j ≤ 3 C i j ( c ) q i j ) K > 0 (73) 0<\tau^{3}<27T^{3}D(c)^{2},\qquad\tau M-\left(\sum_{0\le i<j\le3}C_{ij}(c)q_{ij}\right)K>0 \tag*{(73)} 0 < τ 3 < 27 T 3 D ( c ) 2 , τ M − ( 0 ≤ i < j ≤ 3 ∑ C ij ( c ) q ij ) K > 0 (73)
M > 0 , T 3 M 3 − D K 3 > 0 (74) M>0,\qquad T^{3}M^{3}-DK^{3}>0 \tag*{(74)} M > 0 , T 3 M 3 − D K 3 > 0 (74)
gives F ( q ) < T F(q)<T F ( q ) < T . For (5.8),
K ≥ 0 ⟹ M > 0 , F ( q ) ≤ ( ∑ 0 ≤ i < j ≤ 3 C i j ( c ) q i j ) K 3 D ( c ) 2 / 3 M < τ 3 D ( c ) 2 / 3 < T . K\ge0\quad\Longrightarrow\quad M>0,\qquad F(q)\le\frac{\left(\sum_{0\le i<j\le3}C_{ij}(c)q_{ij}\right)K}{3D(c)^{2/3}M}<\frac{\tau}{3D(c)^{2/3}}<T. K ≥ 0 ⟹ M > 0 , F ( q ) ≤ 3 D ( c ) 2/3 M ( ∑ 0 ≤ i < j ≤ 3 C ij ( c ) q ij ) K < 3 D ( c ) 2/3 τ < T .
Order the edges as ( 01 , 02 , 03 , 12 , 13 , 23 ) (01,02,03,12,13,23) ( 01 , 02 , 03 , 12 , 13 , 23 ) and the trees as in (5.5), with heights
( h 01 , h 02 , h 03 , h 12 , h 13 , h 23 ) = ( 1 , 8 , 64 , 512 , 4096 , 32768 ) . (h_{01},h_{02},h_{03},h_{12},h_{13},h_{23})=(1,8,64,512,4096,32768). ( h 01 , h 02 , h 03 , h 12 , h 13 , h 23 ) = ( 1 , 8 , 64 , 512 , 4096 , 32768 ) .
For each J ∈ { J 1 , … , J 16 } J\in\{J_1,\ldots,J_{16}\} J ∈ { J 1 , … , J 16 } , solve
d e ⋅ ℓ J = h e ( e ∈ J ) . d_e\cdot\ell_J=h_e\qquad(e\in J). d e ⋅ ℓ J = h e ( e ∈ J ) .
The signed sums of distinct tree heights satisfy h e − d e ⋅ ℓ J ≠ 0 h_e-d_e\cdot\ell_J\ne0 h e − d e ⋅ ℓ J = 0 for e ∉ J e\notin J e ∈ / J . For q > 0 q>0 q > 0 , the parallelotopes
1 2 ∑ e ∈ E ∖ J sgn ( h e − d e ⋅ ℓ J ) q e d e + ∑ e ∈ J [ − q e d e / 2 , q e d e / 2 ] (75) \frac{1}{2}\sum_{e\in E\setminus J}\operatorname{sgn}(h_e-d_e\cdot\ell_J)q_ed_e+\sum_{e\in J}[-q_ed_e/2,q_ed_e/2] \tag*{(75)} 2 1 e ∈ E ∖ J ∑ sgn ( h e − d e ⋅ ℓ J ) q e d e + e ∈ J ∑ [ − q e d e /2 , q e d e /2 ] (75)
tile Y ( q ) Y(q) Y ( q ) with disjoint interiors. Indeed, under ∑ 0 ≤ i < j ≤ 3 z i j d i j = y \sum_{0\le i<j\le3}z_{ij}d_{ij}=y ∑ 0 ≤ i < j ≤ 3 z ij d ij = y , ∣ z e ∣ ≤ q e / 2 |z_e|\le q_e/2 ∣ z e ∣ ≤ q e /2 ,
∑ 0 ≤ i < j ≤ 3 h i j z i j = ℓ J ⋅ y + ∑ e ∈ E ∖ J ( h e − d e ⋅ ℓ J ) z e ≤ ℓ J ⋅ y + 1 2 ∑ e ∈ E ∖ J ∣ h e − d e ⋅ ℓ J ∣ q e , \sum_{0\le i<j\le3}h_{ij}z_{ij}=\ell_J\cdot y+\sum_{e\in E\setminus J}(h_e-d_e\cdot\ell_J)z_e
\le\ell_J\cdot y+\frac{1}{2}\sum_{e\in E\setminus J}|h_e-d_e\cdot\ell_J|q_e, 0 ≤ i < j ≤ 3 ∑ h ij z ij = ℓ J ⋅ y + e ∈ E ∖ J ∑ ( h e − d e ⋅ ℓ J ) z e ≤ ℓ J ⋅ y + 2 1 e ∈ E ∖ J ∑ ∣ h e − d e ⋅ ℓ J ∣ q e ,
with equality precisely for z e = sgn ( h e − d e ⋅ ℓ J ) q e / 2 z_e=\operatorname{sgn}(h_e-d_e\cdot\ell_J)q_e/2 z e = sgn ( h e − d e ⋅ ℓ J ) q e /2 off J J J . At a vertex maximizer, the free columns are independent and, for almost every y y y , form a tree J J J . A wrong off-tree sign permits, for small t > 0 t>0 t > 0 ,
δ z e = t sgn ( h e − d e ⋅ ℓ J ) , δ z J = − [ d a ] a ∈ J − 1 d e δ z e , δ z a = 0 ( a ∈ E ∖ ( J ∪ { e } ) ) , \delta z_e=t\operatorname{sgn}(h_e-d_e\cdot\ell_J),\qquad\delta z_J=-[d_a]_{a\in J}^{-1}d_e\delta z_e,\qquad\delta z_a=0\quad(a\in E\setminus(J\cup\{e\})), δ z e = t sgn ( h e − d e ⋅ ℓ J ) , δ z J = − [ d a ] a ∈ J − 1 d e δ z e , δ z a = 0 ( a ∈ E ∖ ( J ∪ { e })) ,
so that
∑ 0 ≤ i < j ≤ 3 d i j δ z i j = 0 , ∑ 0 ≤ i < j ≤ 3 h i j δ z i j = t ∣ h e − d e ⋅ ℓ J ∣ > 0 , \sum_{0\le i<j\le3}d_{ij}\delta z_{ij}=0,\qquad\sum_{0\le i<j\le3}h_{ij}\delta z_{ij}=t\lvert h_e-d_e\cdot\ell_J\rvert>0, 0 ≤ i < j ≤ 3 ∑ d ij δ z ij = 0 , 0 ≤ i < j ≤ 3 ∑ h ij δ z ij = t ∣ h e − d e ⋅ ℓ J ∣ > 0 ,
contradicting maximality. The off-tree signs and the tree coordinates are unique; hence the interiors are disjoint, and closure gives (75) .
In the fixed coordinates of § 4 ,
Γ ( q ) = 1 4 ( q 02 + q 03 + q 12 + q 13 q 03 − q 12 q 02 − q 13 q 03 − q 12 q 01 + q 03 + q 12 + q 23 q 01 − q 23 q 02 − q 13 q 01 − q 23 q 01 + q 02 + q 13 + q 23 ) , x J ( z ) = 1 4 ∑ 0 ≤ i < j ≤ 3 q i j ( b i − b j ) z i j , z e = sgn ( h e − d e ⋅ ℓ J ) / 2 ( e ∉ J ) . (76) \begin{aligned}
\Gamma(q)&=\frac{1}{4}\begin{pmatrix}
q_{02}+q_{03}+q_{12}+q_{13} & q_{03}-q_{12} & q_{02}-q_{13}\\
q_{03}-q_{12} & q_{01}+q_{03}+q_{12}+q_{23} & q_{01}-q_{23}\\
q_{02}-q_{13} & q_{01}-q_{23} & q_{01}+q_{02}+q_{13}+q_{23}
\end{pmatrix},\\
x_J(z)&=\frac{1}{4}\sum_{0\le i<j\le3}q_{ij}(b_i-b_j)z_{ij},\qquad z_e=\operatorname{sgn}(h_e-d_e\cdot\ell_J)/2\quad(e\notin J). \tag*{(76)}
\end{aligned} Γ ( q ) x J ( z ) = 4 1 q 02 + q 03 + q 12 + q 13 q 03 − q 12 q 02 − q 13 q 03 − q 12 q 01 + q 03 + q 12 + q 23 q 01 − q 23 q 02 − q 13 q 01 − q 23 q 01 + q 02 + q 13 + q 23 , = 4 1 0 ≤ i < j ≤ 3 ∑ q ij ( b i − b j ) z ij , z e = sgn ( h e − d e ⋅ ℓ J ) /2 ( e ∈ / J ) . (76)
Since ∣ det [ d e ] e ∈ J r ∣ = 1 \lvert\det[d_e]_{e\in J_r}\rvert=1 ∣ det [ d e ] e ∈ J r ∣ = 1 , 1 ≤ r ≤ 16 1\le r\le16 1 ≤ r ≤ 16 , canceling the common frame Jacobian gives
M = ∑ r = 1 16 ( ∏ e ∈ J r q e ) ∫ [ − 1 / 2 , 1 / 2 ] 3 f ( x J r ( z ) ) 2 d z J r , K = ∑ r = 1 16 ( ∏ e ∈ J r q e ) ∫ [ − 1 / 2 , 1 / 2 ] 3 ( ∇ f T Γ ∇ f ) ( x J r ( z ) ) d z J r . (77) \begin{aligned}
M&=\sum_{r=1}^{16}\left(\prod_{e\in J_r}q_e\right)\int_{[-1/2,1/2]^3}f(x_{J_r}(z))^2\,dz_{J_r},\\
K&=\sum_{r=1}^{16}\left(\prod_{e\in J_r}q_e\right)\int_{[-1/2,1/2]^3}(\nabla f^T\Gamma\nabla f)(x_{J_r}(z))\,dz_{J_r}. \tag*{(77)}
\end{aligned} M K = r = 1 ∑ 16 ( e ∈ J r ∏ q e ) ∫ [ − 1/2 , 1/2 ] 3 f ( x J r ( z ) ) 2 d z J r , = r = 1 ∑ 16 ( e ∈ J r ∏ q e ) ∫ [ − 1/2 , 1/2 ] 3 ( ∇ f T Γ∇ f ) ( x J r ( z )) d z J r . (77)
These polynomial expressions extend to q e = 0 q_e=0 q e = 0 , using
∫ − 1 / 2 1 / 2 z 2 j d z = 2 − 2 j 2 j + 1 , ∫ − 1 / 2 1 / 2 z 2 j + 1 d z = 0 ( j ∈ N 0 ) . \int_{-1/2}^{1/2}z^{2j}\,dz=\frac{2^{-2j}}{2j+1},\qquad\int_{-1/2}^{1/2}z^{2j+1}\,dz=0\quad(j\in\mathbb{N}_0). ∫ − 1/2 1/2 z 2 j d z = 2 j + 1 2 − 2 j , ∫ − 1/2 1/2 z 2 j + 1 d z = 0 ( j ∈ N 0 ) .
The one-dimensional rules are
Q − 1 f = 0 , Q 0 f = f ( 0 ) , Q_{-1}f=0,\qquad Q_0f=f(0), Q − 1 f = 0 , Q 0 f = f ( 0 ) ,
Q j f = ∑ k = − j j f ( k 2 j ) ∫ − 1 / 2 1 / 2 ∏ ℓ ∈ Z − j ≤ ℓ ≤ j , ℓ ≠ k 2 j x − ℓ k − ℓ d x ( j ≥ 1 ) . Q_jf=\sum_{k=-j}^{j}f\left(\frac{k}{2j}\right)\int_{-1/2}^{1/2}\prod_{\substack{\ell\in\mathbb{Z}\\-j\le\ell\le j,\ \ell\ne k}}\frac{2jx-\ell}{k-\ell}\,dx\quad(j\ge1). Q j f = k = − j ∑ j f ( 2 j k ) ∫ − 1/2 1/2 ℓ ∈ Z − j ≤ ℓ ≤ j , ℓ = k ∏ k − ℓ 2 j x − ℓ d x ( j ≥ 1 ) .
For each s ∈ { 3 , 5 } s\in\{3,5\} s ∈ { 3 , 5 } , add coincident rational weights before rounding and define w y w_y w y , zero at all other points, by
∑ y ∈ [ − 1 / 2 , 1 / 2 ] 3 w y ≠ 0 w y f ( y ) = ∑ i = 0 s ∑ j = 0 s − i ∑ k = 0 s − i − j [ ( Q i − Q i − 1 ) ⊗ ( Q j − Q j − 1 ) ⊗ ( Q k − Q k − 1 ) ] f . (78) \sum_{\substack{y\in[-1/2,1/2]^3\\w_y\ne0}}w_yf(y)=\sum_{i=0}^{s}\sum_{j=0}^{s-i}\sum_{k=0}^{s-i-j}[(Q_i-Q_{i-1})\otimes(Q_j-Q_{j-1})\otimes(Q_k-Q_{k-1})]f. \tag*{(78)} y ∈ [ − 1/2 , 1/2 ] 3 w y = 0 ∑ w y f ( y ) = i = 0 ∑ s j = 0 ∑ s − i k = 0 ∑ s − i − j [( Q i − Q i − 1 ) ⊗ ( Q j − Q j − 1 ) ⊗ ( Q k − Q k − 1 )] f . (78)
Telescoping gives, for a , b , c ∈ N 0 a,b,c\in\mathbb{N}_0 a , b , c ∈ N 0 with a + b + c ≤ s a+b+c\le s a + b + c ≤ s ,
∑ y ∈ [ − 1 / 2 , 1 / 2 ] 3 w y ≠ 0 w y y 1 2 a y 2 2 b y 3 2 c = 2 − 2 ( a + b + c ) ( 2 a + 1 ) ( 2 b + 1 ) ( 2 c + 1 ) . \sum_{\substack{y\in[-1/2,1/2]^3\\w_y\ne0}}w_yy_1^{2a}y_2^{2b}y_3^{2c}=\frac{2^{-2(a+b+c)}}{(2a+1)(2b+1)(2c+1)}. y ∈ [ − 1/2 , 1/2 ] 3 w y = 0 ∑ w y y 1 2 a y 2 2 b y 3 2 c = ( 2 a + 1 ) ( 2 b + 1 ) ( 2 c + 1 ) 2 − 2 ( a + b + c ) .
# { y ∈ [ − 1 / 2 , 1 / 2 ] 3 : w y ≠ 0 } = { 69 , s = 3 , 357 , s = 5. \#\{y\in[-1/2,1/2]^3:w_y\ne0\}=
\begin{cases}
69,&s=3,\\
357,&s=5.
\end{cases} # { y ∈ [ − 1/2 , 1/2 ] 3 : w y = 0 } = { 69 , 357 , s = 3 , s = 5.
Monomials with an odd exponent vanish by symmetry.
The cubic monomial order is
x 1 , x 2 , x 3 , x 1 3 , x 2 3 , x 3 3 , x 1 2 x 2 , x 1 2 x 3 , x 1 x 2 2 , x 2 2 x 3 , x 1 x 3 2 , x 2 x 3 2 , x 1 x 2 x 3 . x_1,x_2,x_3,x_1^3,x_2^3,x_3^3,x_1^2x_2,x_1^2x_3,x_1x_2^2,x_2^2x_3,x_1x_3^2,x_2x_3^2,x_1x_2x_3. x 1 , x 2 , x 3 , x 1 3 , x 2 3 , x 3 3 , x 1 2 x 2 , x 1 2 x 3 , x 1 x 2 2 , x 2 2 x 3 , x 1 x 3 2 , x 2 x 3 2 , x 1 x 2 x 3 .
For quintics, order x 1 a x 2 b x 3 c x_1^a x_2^b x_3^c x 1 a x 2 b x 3 c , a + b + c ∈ { 1 , 3 , 5 } a+b+c \in\{1,3,5\} a + b + c ∈ { 1 , 3 , 5 } , by increasing ( a + b + c , a , b ) (a+b+c,a,b) ( a + b + c , a , b ) .
Apply coefficientwise truncation
T h ( a ) = h sgn ( a ) ⌊ ∣ a ∣ h ⌋ . \mathcal{T}_h(a)=h\operatorname{sgn}(a)\left\lfloor\frac{\lvert a\rvert}{h}\right\rfloor. T h ( a ) = h sgn ( a ) ⌊ h ∣ a ∣ ⌋ .
For FCC, the box center c c c determines the coordinates
( z ~ 1 , z ~ 2 , z ~ 3 ) = 2 − 24 [ 12 ⋅ 2 24 3 ( c 01 + c 02 + c 13 + c 23 ) + 4 ( c 03 + c 12 ) ] ( x 3 , x 1 + x 2 , x 1 − x 2 ) . (79) (\widetilde{z}_1,\widetilde{z}_2,\widetilde{z}_3)=2^{-24}\left[\frac{12\cdot2^{24}}{3(c_{01}+c_{02}+c_{13}+c_{23})+4(c_{03}+c_{12})}\right](x_3,x_1+x_2,x_1-x_2). \tag*{(79)} ( z 1 , z 2 , z 3 ) = 2 − 24 [ 3 ( c 01 + c 02 + c 13 + c 23 ) + 4 ( c 03 + c 12 ) 12 ⋅ 2 24 ] ( x 3 , x 1 + x 2 , x 1 − x 2 ) . (79)
Define
10 6 g ( z ) = z 1 [ 10 6 − 860262 ( z 2 2 + z 3 2 ) − 109471 ( z 2 4 + z 3 4 ) + 3253797 z 2 2 z 3 2 − z 1 2 { 460101 + 291610 ( z 2 2 + z 3 2 ) } + 99272 z 1 4 ] . (80) 10^6g(z)=z_1\left[10^6-860262(z_2^2+z_3^2)-109471(z_2^4+z_3^4)+3253797z_2^2z_3^2-z_1^2\{460101+291610(z_2^2+z_3^2)\}+99272z_1^4\right]. \tag*{(80)} 1 0 6 g ( z ) = z 1 [ 1 0 6 − 860262 ( z 2 2 + z 3 2 ) − 109471 ( z 2 4 + z 3 4 ) + 3253797 z 2 2 z 3 2 − z 1 2 { 460101 + 291610 ( z 2 2 + z 3 2 )} + 99272 z 1 4 ] . (80)
Interchanging z 1 , z i z_1,z_i z 1 , z i gives g i g_i g i and the trial
T 2 − 40 ( ∑ i = 1 3 v i T 2 − 40 g i ) . \mathcal{T}_{2^{-40}}\left(\sum_{i=1}^{3}v_i\mathcal{T}_{2^{-40}}g_i\right). T 2 − 40 ( i = 1 ∑ 3 v i T 2 − 40 g i ) .
Its energy matrix is
2 − 50 [ 12 ⋅ 2 24 3 ( c 01 + c 02 + c 13 + c 23 ) + 4 ( c 03 + c 12 ) ] 2 ( q 01 + q 02 + q 13 + q 23 q 01 + q 02 − q 13 − q 23 − q 01 + q 02 − q 13 + q 23 q 01 + q 02 − q 13 − q 23 q 01 + q 02 + 4 q 03 + q 13 + q 23 − q 01 + q 02 + q 13 − q 23 − q 01 + q 02 − q 13 + q 23 − q 01 + q 02 + q 13 − q 23 q 01 + q 02 + 4 q 12 + q 13 + q 23 ) . (81) 2^{-50}\left[\frac{12\cdot2^{24}}{3(c_{01}+c_{02}+c_{13}+c_{23})+4(c_{03}+c_{12})}\right]^2
\begin{pmatrix}
q_{01}+q_{02}+q_{13}+q_{23} & q_{01}+q_{02}-q_{13}-q_{23} & -q_{01}+q_{02}-q_{13}+q_{23}\\
q_{01}+q_{02}-q_{13}-q_{23} & q_{01}+q_{02}+4q_{03}+q_{13}+q_{23} & -q_{01}+q_{02}+q_{13}-q_{23}\\
-q_{01}+q_{02}-q_{13}+q_{23} & -q_{01}+q_{02}+q_{13}-q_{23} & q_{01}+q_{02}+4q_{12}+q_{13}+q_{23}
\end{pmatrix}. \tag*{(81)} 2 − 50 [ 3 ( c 01 + c 02 + c 13 + c 23 ) + 4 ( c 03 + c 12 ) 12 ⋅ 2 24 ] 2 q 01 + q 02 + q 13 + q 23 q 01 + q 02 − q 13 − q 23 − q 01 + q 02 − q 13 + q 23 q 01 + q 02 − q 13 − q 23 q 01 + q 02 + 4 q 03 + q 13 + q 23 − q 01 + q 02 + q 13 − q 23 − q 01 + q 02 − q 13 + q 23 − q 01 + q 02 + q 13 − q 23 q 01 + q 02 + 4 q 12 + q 13 + q 23 . (81)
For transported trials, use the common BCC fan (31) , with its fixed tetrahedron order and coordinate reflections. Let F A ( q ) F_A(q) F A ( q ) send each reference vertex and face average to its corresponding point in the zonotope with weights q e q_e q e . The vertex of strict order σ \sigma σ is
V q ( σ ) = 1 8 ∑ 0 ≤ i < j ≤ 3 q i j sgn ( σ i − σ j ) ( b i − b j ) . (82) V_q(\sigma)=\frac{1}{8}\sum_{0\le i<j\le3}q_{ij}\operatorname{sgn}(\sigma_i-\sigma_j)(b_i-b_j). \tag*{(82)} V q ( σ ) = 8 1 0 ≤ i < j ≤ 3 ∑ q ij sgn ( σ i − σ j ) ( b i − b j ) . (82)
Positive segment lengths preserve the normal fan, and face averages remain in relative interiors. Hence
q e > 0 ( e ∈ E ) ⟹ the face incidences persist , det F A ( q ) > 0. q_e>0\ (e\in E)\quad\Longrightarrow\quad\text{the face incidences persist},\quad\det F_A(q)>0. q e > 0 ( e ∈ E ) ⟹ the face incidences persist , det F A ( q ) > 0.
The coned triangulations are therefore nondegenerate.
Transporting the normalized modes of Numerical Lemma 2 preserves oddness and gives
M = ∑ T ∈ T + ∑ s ∈ { − 1 , 1 } 3 det ( F diag ( s 1 , s 2 , s 3 ) T ) ∫ diag ( s 1 , s 2 , s 3 ) T ( ∑ i = 1 3 v i P i ) 2 , K = ∑ T ∈ T + ∑ s ∈ { − 1 , 1 } 3 1 det F diag ( s 1 , s 2 , s 3 ) T ∫ diag ( s 1 , s 2 , s 3 ) T ( ∑ i = 1 3 v i ∇ P i ) T × adj ( F diag ( s 1 , s 2 , s 3 ) T ) Γ adj ( F diag ( s 1 , s 2 , s 3 ) T ) T × ( ∑ j = 1 3 v j ∇ P j ) . (83) \begin{aligned}
M={}&\sum_{T\in\mathcal{T}_+}\sum_{s\in\{-1,1\}^3}\det(F_{\operatorname{diag}(s_1,s_2,s_3)}T)\int_{\operatorname{diag}(s_1,s_2,s_3)T}\left(\sum_{i=1}^{3}v_iP_i\right)^2,\\
K={}&\sum_{T\in\mathcal{T}_+}\sum_{s\in\{-1,1\}^3}\frac{1}{\det F_{\operatorname{diag}(s_1,s_2,s_3)}T}\int_{\operatorname{diag}(s_1,s_2,s_3)T}\left(\sum_{i=1}^{3}v_i\nabla P_i\right)^{T}\\
&\qquad\times\operatorname{adj}(F_{\operatorname{diag}(s_1,s_2,s_3)}T)\Gamma\operatorname{adj}(F_{\operatorname{diag}(s_1,s_2,s_3)}T)^{T}\\
&\qquad\times\left(\sum_{j=1}^{3}v_j\nabla P_j\right). \tag*{(83)}
\end{aligned} M = K = T ∈ T + ∑ s ∈ { − 1 , 1 } 3 ∑ det ( F diag ( s 1 , s 2 , s 3 ) T ) ∫ diag ( s 1 , s 2 , s 3 ) T ( i = 1 ∑ 3 v i P i ) 2 , T ∈ T + ∑ s ∈ { − 1 , 1 } 3 ∑ det F diag ( s 1 , s 2 , s 3 ) T 1 ∫ diag ( s 1 , s 2 , s 3 ) T ( i = 1 ∑ 3 v i ∇ P i ) T × adj ( F diag ( s 1 , s 2 , s 3 ) T ) Γ adj ( F diag ( s 1 , s 2 , s 3 ) T ) T × ( j = 1 ∑ 3 v j ∇ P j ) . (83)
All reference integrals are rational by (46) .
The Appendix specifies the center coefficients, enclosure arithmetic, and test order. Accept a box at the first successful test, and set
W 0 = { ∅ } , W h + 1 = { w b : w ∈ W h , B w is not accepted , b ∈ { 0 , 1 } } . (84) \mathcal{W}_0=\{\varnothing\},\qquad\mathcal{W}_{h+1}=\{wb:w\in\mathcal{W}_h,\ B_w\text{ is not accepted},\ b\in\{0,1\}\}. \tag*{(84)} W 0 = { ∅ } , W h + 1 = { w b : w ∈ W h , B w is not accepted , b ∈ { 0 , 1 }} . (84)
Lemma 5.1. If W H + 1 = ∅ \mathcal{W}_{H+1}=\varnothing W H + 1 = ∅ for some H H H , then every q ∈ C q\in\mathcal{C} q ∈ C with D ( q ) > 0 D(q)>0 D ( q ) > 0 satisfies
F ( q ) < T or ∥ q ∑ 0 ≤ i < j ≤ 3 q i j − p ∗ ∥ 2 ≤ 73 6000 . F(q)<T\qquad\text{or}\qquad\left\lVert\frac{q}{\sum_{0\leq i<j\leq3}q_{ij}}-p_*\right\rVert_2\leq\frac{73}{6000}. F ( q ) < T or ∑ 0 ≤ i < j ≤ 3 q ij q − p ∗ 2 ≤ 6000 73 .
Proof. Since B w = B w 0 ∪ B w 1 B_w=B_{w0}\cup B_{w1} B w = B w 0 ∪ B w 1 , (5.19) gives, by induction, for every $h\geq 0,
{ 1 } × [ 0 , 1 ] 5 = ⋃ j = 0 h − 1 ⋃ w ∈ W j B w is accepted B w ∪ ⋃ w ∈ W h B w . \{1\}\times[0,1]^5=\bigcup_{j=0}^{h-1}\ \bigcup_{\substack{w\in\mathcal{W}_j\\B_w\text{ is accepted}}}B_w\ \cup\ \bigcup_{w\in\mathcal{W}_h}B_w. { 1 } × [ 0 , 1 ] 5 = j = 0 ⋃ h − 1 w ∈ W j B w is accepted ⋃ B w ∪ w ∈ W h ⋃ B w .
If W H + 1 = ∅ \mathcal{W}_{H+1}=\varnothing W H + 1 = ∅ , every q ∈ C q\in\mathcal{C} q ∈ C lies in an accepted box. Since the chamber test excludes no such q q q ,
(5.4) ⟹ ∥ q ∑ 0 ≤ i < j ≤ 3 q i j − p ∗ ∥ 2 ≤ 73 6000 , (5.7), (5.8), (5.9) ⟹ F ( q ) < T . \text{(5.4)}\quad\Longrightarrow\quad\left\lVert\frac{q}{\sum_{0\leq i<j\leq3}q_{ij}}-p_*\right\rVert_2\leq\frac{73}{6000},\qquad\text{(5.7), (5.8), (5.9)}\quad\Longrightarrow\quad F(q)<T. (5.4) ⟹ ∑ 0 ≤ i < j ≤ 3 q ij q − p ∗ 2 ≤ 6000 73 , (5.7), (5.8), (5.9) ⟹ F ( q ) < T .
□
Numerical Lemma 3. For (5.19) and the Appendix constructions,
# W 37 = 4 , W 38 = ∅ . \#\mathcal{W}_{37}=4,\qquad\mathcal{W}_{38}=\varnothing. # W 37 = 4 , W 38 = ∅ .
The Appendix gives the acceptance counts and strict sign bounds.
Proof of Theorem 1.1. By Lemma 5.1, Numerical Lemma 3, and Proposition 4.1, a reduced representative satisfies
q ∈ C , D ( q ) > 0 , F ( q ) ≥ F ( p ∗ ) > T q\in\mathcal{C},\qquad D(q)>0,\qquad F(q)\geq F(p_*)>T q ∈ C , D ( q ) > 0 , F ( q ) ≥ F ( p ∗ ) > T
⟹ ∥ q ∑ 0 ≤ i < j ≤ 3 q i j − p ∗ ∥ 2 ≤ 73 6000 ⟹ q ∑ 0 ≤ i < j ≤ 3 q i j = p ∗ . \Longrightarrow\quad\left\lVert\frac{q}{\sum_{0\leq i<j\leq3}q_{ij}}-p_*\right\rVert_2\leq\frac{73}{6000}\quad\Longrightarrow\quad\frac{q}{\sum_{0\leq i<j\leq3}q_{ij}}=p_*. ⟹ ∑ 0 ≤ i < j ≤ 3 q ij q − p ∗ 2 ≤ 6000 73 ⟹ ∑ 0 ≤ i < j ≤ 3 q ij q = p ∗ .
Equal conorms represent BCC; for R ∈ O ( 3 ) R\in O(3) R ∈ O ( 3 ) , μ 1 ( R V B C C ) = μ 1 ( V B C C ) \mu_1(RV_{\mathrm{BCC}})=\mu_1(V_{\mathrm{BCC}}) μ 1 ( R V BCC ) = μ 1 ( V BCC ) . Finally, the mixed eigenvalue enclosure and
V B C C = 2 − 2 / 3 W , μ 1 ( V B C C ) = 2 4 / 3 μ 1 ( W ) V_{\mathrm{BCC}}=2^{-2/3}W,\qquad\mu_1(V_{\mathrm{BCC}})=2^{4/3}\mu_1(W) V BCC = 2 − 2/3 W , μ 1 ( V BCC ) = 2 4/3 μ 1 ( W )
give the stated enclosure.
□
Appendix
We give the coefficient construction for Numerical Lemma 1 and the rational enclosures for Numerical Lemma 3.
Proof of Numerical Lemma 1. Use (3.4) and the Bernstein convention of Section 3. For ordered vertices w 0 , … , w 3 w_0,\ldots,w_3 w 0 , … , w 3 , with lexicographic sorting,
k ( P , α ) = sort { ( 2 w i , α i ) : 0 ≤ i ≤ 3 , α i > 0 } , k(P,\alpha)=\operatorname{sort}\{(2w_i,\alpha_i):0\leq i\leq3,\ \alpha_i>0\}, k ( P , α ) = sort {( 2 w i , α i ) : 0 ≤ i ≤ 3 , α i > 0 } ,
k ∗ = sort { ( ( u 1 , u 3 , u 2 ) , a ) : ( ( u 1 , u 2 , u 3 ) , a ) ∈ k } . k^*=\operatorname{sort}\{((u_1,u_3,u_2),a):((u_1,u_2,u_3),a)\in k\}. k ∗ = sort {(( u 1 , u 3 , u 2 ) , a ) : (( u 1 , u 2 , u 3 ) , a ) ∈ k } .
The scalar space. Lexicographically index the distinct min ( k , k ∗ ) \min(k,k^*) min ( k , k ∗ ) with ∣ α ∣ = 9 |\alpha|=9 ∣ α ∣ = 9 and support not contained in { x 1 = 0 } \{x_1=0\} { x 1 = 0 } . Then
( s i ) P , α = { 1 , min ( k ( P , α ) , k ( P , α ) ∗ ) has index i , 0 , otherwise . s i ∈ C ( Ω ‾ + ) , s i ∣ x 1 = 0 = 0 ( 1 ≤ i ≤ 710 ) . (s_i)_{P,\alpha}=
\begin{cases}
1,&\min(k(P,\alpha),k(P,\alpha)^*)\text{ has index }i,\\
0,&\text{otherwise}.
\end{cases}
\qquad s_i\in C(\overline{\Omega}_+),\qquad s_i\big|_{x_1=0}=0\quad(1\leq i\leq710). ( s i ) P , α = { 1 , 0 , min ( k ( P , α ) , k ( P , α ) ∗ ) has index i , otherwise . s i ∈ C ( Ω + ) , s i x 1 = 0 = 0 ( 1 ≤ i ≤ 710 ) .
Set
M i j = ( s i , s j ) , F i j = ( ∇ s i , ∇ s j ) − 107 25 ( s i , s j ) , (85) M_{ij}=(s_i,s_j),\qquad F_{ij}=(\nabla s_i,\nabla s_j)-\frac{107}{25}(s_i,s_j), \tag*{(85)} M ij = ( s i , s j ) , F ij = ( ∇ s i , ∇ s j ) − 25 107 ( s i , s j ) , (85)
with inner products over Ω + \Omega_+ Ω + and 1 ≤ i , j ≤ 710 1\le i,j\le710 1 ≤ i , j ≤ 710 .
The flux space. For ∣ α ∣ = 10 |\alpha|=10 ∣ α ∣ = 10 , group coefficients by k k k , ordered by first occurrence in ( P , α ) (P,\alpha) ( P , α ) : P P P follows (3.4), and α \alpha α is lexicographic. Within each group, order by ( P , component ) (P,\text{component}) ( P , component ) . With supp k = { u / 2 : ( u , a ) ∈ k } \operatorname{supp}k=\{u/2:(u,a)\in k\} supp k = { u /2 : ( u , a ) ∈ k } , impose
( ξ P , k − ξ P ′ , k ) ⋅ ( ( u b − u a ) × ( u c − u a ) ) = 0 ( F = P ∩ P ′ , supp k ⊂ F ) , (\xi_{P,k}-\xi_{P',k})\cdot\left((u_b-u_a)\times(u_c-u_a)\right)=0
\qquad(F=P\cap P',\ \operatorname{supp}k\subset F), ( ξ P , k − ξ P ′ , k ) ⋅ ( ( u b − u a ) × ( u c − u a ) ) = 0 ( F = P ∩ P ′ , supp k ⊂ F ) ,
ξ P , k ⋅ ( ( u b − u a ) × ( u c − u a ) ) = 0 ( F ⊂ ∂ Ω + Neumann , supp k ⊂ F ) . \xi_{P,k}\cdot\left((u_b-u_a)\times(u_c-u_a)\right)=0
\qquad(F\subset\partial\Omega_+\ \text{Neumann},\ \operatorname{supp}k\subset F). ξ P , k ⋅ ( ( u b − u a ) × ( u c − u a ) ) = 0 ( F ⊂ ∂ Ω + Neumann , supp k ⊂ F ) .
where u a , u b , u c u_a,u_b,u_c u a , u b , u c are the lexicographically ordered doubled face vertices. For each free index i i i , solve the reduced-row-echelon constraints with free coordinates ( z i ) j = δ i j (z_i)_j=\delta_{ij} ( z i ) j = δ ij , then successively
z i ← ( lcm j o v e r e n t r i e s den ( ( z i ) j ) ) z i , z i ← z i gcd j o v e r e n t r i e s ∣ ( z i ) j ∣ . z_i\leftarrow\left(\operatorname*{lcm}_{j\ \mathrm{over\ entries}}\operatorname{den}((z_i)_j)\right)z_i,\qquad
z_i\leftarrow\frac{z_i}{\operatorname*{gcd}_{j\ \mathrm{over\ entries}}|(z_i)_j|}. z i ← ( j over entries lcm den (( z i ) j ) ) z i , z i ← gcd j over entries ∣ ( z i ) j ∣ z i .
Denominators are positive. Relabel consecutively by ( group , free index ) (\text{group},\text{free index}) ( group , free index ) :
{ z i } i = 1 6138 ⊂ H ( div ; Ω + ) , z i ⋅ n = 0 on the Neumann boundary . \{z_i\}_{i=1}^{6138}\subset H(\operatorname{div};\Omega_+),\qquad z_i\cdot n=0\quad\text{on the Neumann boundary}. { z i } i = 1 6138 ⊂ H ( div ; Ω + ) , z i ⋅ n = 0 on the Neumann boundary .
Rational coefficient generation. For scalar a a a , square brackets mean rounding to the nearest finite IEEE 754 binary64 value:
[ a ] ∈ argmin b = m 2 e , m , e ∈ Z ∣ m ∣ < 2 53 , − 1074 ≤ e ≤ 971 ∣ a − b ∣ , [ G ] i j = [ G i j ] . [a]\in\operatorname*{argmin}_{\substack{b=m2^e,\ m,e\in\mathbb{Z}\\ |m|<2^{53},\ -1074\le e\le971}}|a-b|,\qquad[G]_{ij}=[G_{ij}]. [ a ] ∈ b = m 2 e , m , e ∈ Z ∣ m ∣ < 2 53 , − 1074 ≤ e ≤ 971 argmin ∣ a − b ∣ , [ G ] ij = [ G ij ] .
At a midpoint choose the even canonical significand; a zero result retains the input sign. Vector and matrix inputs are rounded entrywise, and every nested bracket is a separate rounding. Define nearest-integer rounding by
nint ( a ) = { ⌊ a + 1 2 ⌋ , a ∉ Z + 1 2 , 2 ⌊ a 2 + 1 4 ⌋ , a ∈ Z + 1 2 . \operatorname{nint}(a)=
\begin{cases}
\lfloor a+\frac12\rfloor,&a\notin\mathbb{Z}+\frac12,\\
2\left\lfloor\frac{a}{2}+\frac14\right\rfloor,&a\in\mathbb{Z}+\frac12.
\end{cases} nint ( a ) = { ⌊ a + 2 1 ⌋ , 2 ⌊ 2 a + 4 1 ⌋ , a ∈ / Z + 2 1 , a ∈ Z + 2 1 .
Thus midpoint ties select the even integer. For symmetric G ∈ R N × N G\in\mathbb{R}^{N\times N} G ∈ R N × N ,
G i j ( 1 ) = [ G i j ] ( 1 ≤ j ≤ i ≤ N ) . G_{ij}^{(1)}=[G_{ij}]\qquad(1\le j\le i\le N). G ij ( 1 ) = [ G ij ] ( 1 ≤ j ≤ i ≤ N ) .
For k = 1 , … , N k=1,\ldots,N k = 1 , … , N ,
l i k = [ G i k ( k ) / G k k ( k ) ] ( i > k ) , l_{ik}=[G_{ik}^{(k)}/G_{kk}^{(k)}]\qquad(i>k), l ik = [ G ik ( k ) / G k k ( k ) ] ( i > k ) ,
G i j ( k + 1 ) = [ G i j ( k ) − [ G i k ( k ) l j k ] ] ( i ≥ j > k ) . (86) G_{ij}^{(k+1)}=[G_{ij}^{(k)}-[G_{ik}^{(k)}l_{jk}]]\qquad(i\ge j>k). \tag*{(86)} G ij ( k + 1 ) = [ G ij ( k ) − [ G ik ( k ) l j k ]] ( i ≥ j > k ) . (86)
For rounded y y y , compute S G ( y ) S_G(y) S G ( y ) by the ordered substitutions
k = 1 , … , N : y i ← [ y i − [ l i k y k ] ] , i = k + 1 , … , N , k=1,\ldots,N:\quad y_i\leftarrow[y_i-[l_{ik}y_k]],\qquad i=k+1,\ldots,N, k = 1 , … , N : y i ← [ y i − [ l ik y k ]] , i = k + 1 , … , N ,
k = 1 , … , N : y k ← [ y k / G k k ( k ) ] , k=1,\ldots,N:\quad y_k\leftarrow[y_k/G_{kk}^{(k)}], k = 1 , … , N : y k ← [ y k / G k k ( k ) ] ,
k = N , … , 1 : y k ← [ y k − [ l i k y i ] ] , i = k + 1 , … , N . k=N,\ldots,1:\quad y_k\leftarrow[y_k-[l_{ik}y_i]],\qquad i=k+1,\ldots,N. k = N , … , 1 : y k ← [ y k − [ l ik y i ]] , i = k + 1 , … , N .
All pivots below are positive.
For doubled vertices u i = 2 w i u_i=2w_i u i = 2 w i , initialize a ( 0 ) a^{(0)} a ( 0 ) by rounding
( x 1 ) P , α = 1 18 ∑ i = 0 3 ( u i ) 1 α i , α ∈ N 0 4 , ∣ α ∣ = 9. (x_1)_{P,\alpha}=\frac{1}{18}\sum_{i=0}^{3}(u_i)_{1}\alpha_i,\qquad\alpha\in\mathbb{N}_0^4,\quad|\alpha|=9. ( x 1 ) P , α = 18 1 i = 0 ∑ 3 ( u i ) 1 α i , α ∈ N 0 4 , ∣ α ∣ = 9.
For j = 0 , … , 7 j=0,\ldots,7 j = 0 , … , 7 and i = 1 , … , 710 i=1,\ldots,710 i = 1 , … , 710 , set
a i ( j + 1 ) = [ ( S F ( [ M ] a ( j ) ) ) i max 1 ≤ k ≤ 710 ∣ ( S F ( [ M ] a ( j ) ) ) k ∣ ] . a_i^{(j+1)}=
\left[
\frac{(S_F([M]a^{(j)}))_i}
{\max_{1\le k\le710}|(S_F([M]a^{(j)}))_k|}
\right]. a i ( j + 1 ) = [ max 1 ≤ k ≤ 710 ∣ ( S F ([ M ] a ( j ) ) ) k ∣ ( S F ([ M ] a ( j ) ) ) i ] .
Evaluate the solve and its fixed maximum once per step. In [ M ] a ( j ) [M]a^{(j)} [ M ] a ( j ) , round products and partial sums in increasing column order. Set
v = 10 − 10 ∑ i = 1 710 nint ( 10 10 a i ( 8 ) ) s i . (87) v = 10^{-10}\sum_{i=1}^{710}\operatorname{nint}\left(10^{10}a_i^{(8)}\right)s_i. \tag*{(87)} v = 1 0 − 10 i = 1 ∑ 710 nint ( 1 0 10 a i ( 8 ) ) s i . (87)
Set
H i j = ( z i , z j ) + 20 ( div z i , div z j ) , H_{ij}=(z_i,z_j)+20(\operatorname{div}z_i,\operatorname{div}z_j), H ij = ( z i , z j ) + 20 ( div z i , div z j ) ,
q = 10 − 10 ∑ i = 1 6138 nint ( 10 10 ( S H ( [ ( z j , ∇ v ) − 20 ∥ ∇ v ∥ 2 2 ∥ v ∥ 2 2 ( div z j , v ) ] j = 1 6138 ) ) i ) z i . (88) q=10^{-10}\sum_{i=1}^{6138}\operatorname{nint}\left(10^{10}\left(S_H\left(\left[(z_j,\nabla v)-20\frac{\lVert\nabla v\rVert_2^2}{\lVert v\rVert_2^2}(\operatorname{div}z_j,v)\right]_{j=1}^{6138}\right)\right)_i\right)z_i. \tag*{(88)} q = 1 0 − 10 i = 1 ∑ 6138 nint ( 1 0 10 ( S H ( [ ( z j , ∇ v ) − 20 ∥ v ∥ 2 2 ∥ ∇ v ∥ 2 2 ( div z j , v ) ] j = 1 6138 ) ) i ) z i . (88)
Use rational arithmetic for the following integrations.
Integration. On P P P ,
∇ ζ i = [ w 1 − w 0 , w 2 − w 0 , w 3 − w 0 ] − T e i ( 1 ≤ i ≤ 3 ) , ∇ ζ 0 = − ∑ i = 1 3 ∇ ζ i . \nabla\zeta_i=[w_1-w_0,w_2-w_0,w_3-w_0]^{-T}e_i\quad(1\le i\le3),\qquad\nabla\zeta_0=-\sum_{i=1}^{3}\nabla\zeta_i. ∇ ζ i = [ w 1 − w 0 , w 2 − w 0 , w 3 − w 0 ] − T e i ( 1 ≤ i ≤ 3 ) , ∇ ζ 0 = − i = 1 ∑ 3 ∇ ζ i .
For degree m m m ,
( ∂ j f ) P , α = m ∑ i = 0 3 ( ∂ j ζ i ) f P , α + e i , α ∈ N 0 4 , ∣ α ∣ = m − 1 , 1 ≤ j ≤ 3. (89) (\partial_j f)_{P,\alpha}=m\sum_{i=0}^{3}(\partial_j\zeta_i)f_{P,\alpha+e_i},\qquad\alpha\in\mathbb{N}_0^4,\quad|\alpha|=m-1,\quad1\le j\le3. \tag*{(89)} ( ∂ j f ) P , α = m i = 0 ∑ 3 ( ∂ j ζ i ) f P , α + e i , α ∈ N 0 4 , ∣ α ∣ = m − 1 , 1 ≤ j ≤ 3. (89)
Equation (3.6) evaluates (A.1), (A.4), and
( v , v ) , ( ∇ v , ∇ v ) , ( q , q ) , ( ∇ v , q ) , ( div q , div q ) . (v,v),\qquad(\nabla v,\nabla v),\qquad(q,q),\qquad(\nabla v,q),\qquad(\operatorname{div}q,\operatorname{div}q). ( v , v ) , ( ∇ v , ∇ v ) , ( q , q ) , ( ∇ v , q ) , ( div q , div q ) .
The boundary conditions give
( v , div q ) + ( ∇ v , q ) = ∫ ∂ Ω + v q ⋅ n = 0. (v,\operatorname{div}q)+(\nabla v,q)=\int_{\partial\Omega_+}v\,q\cdot n=0. ( v , div q ) + ( ∇ v , q ) = ∫ ∂ Ω + v q ⋅ n = 0.
Evaluation gives
⌈ 10 10 min 1 ≤ k ≤ 710 G k ( k ) ⌉ = 198287 ( G = F ) , \left\lceil10^{10}\min_{1\le k\le710}G_k^{(k)}\right\rceil=198287\qquad(G=F), ⌈ 1 0 10 1 ≤ k ≤ 710 min G k ( k ) ⌉ = 198287 ( G = F ) ,
⌈ 10 12 min 1 ≤ k ≤ 6138 G k ( k ) ⌉ = 3063 ( G = H ) , \left\lceil10^{12}\min_{1\le k\le6138}G_k^{(k)}\right\rceil=3063\qquad(G=H), ⌈ 1 0 12 1 ≤ k ≤ 6138 min G k ( k ) ⌉ = 3063 ( G = H ) ,
⌈ 10 12 ∥ ∇ v ∥ 2 2 ∥ v ∥ 2 2 ⌉ = 4300846940707 , \left\lceil10^{12}\frac{\lVert\nabla v\rVert_2^2}{\lVert v\rVert_2^2}\right\rceil=4300846940707, ⌈ 1 0 12 ∥ v ∥ 2 2 ∥ ∇ v ∥ 2 2 ⌉ = 4300846940707 ,
⌈ 10 12 ∥ ∇ v − q ∥ 2 2 ∥ v ∥ 2 2 ⌉ = 9001492 , \left\lceil10^{12}\frac{\lVert\nabla v-q\rVert_2^2}{\lVert v\rVert_2^2}\right\rceil=9001492, ⌈ 1 0 12 ∥ v ∥ 2 2 ∥ ∇ v − q ∥ 2 2 ⌉ = 9001492 ,
⌈ 10 17 ∥ v ∥ 2 2 ∥ div q + ∥ ∇ v ∥ 2 2 ∥ v ∥ 2 2 v ∥ 2 2 ⌉ = 226155 , \left\lceil\frac{10^{17}}{\lVert v\rVert_2^2}\left\lVert\operatorname{div}q+\frac{\lVert\nabla v\rVert_2^2}{\lVert v\rVert_2^2}v\right\rVert_2^2\right\rceil=226155, ⌈ ∥ v ∥ 2 2 1 0 17 div q + ∥ v ∥ 2 2 ∥ ∇ v ∥ 2 2 v 2 2 ⌉ = 226155 ,
min 0 ≤ j ≤ 7 max 1 ≤ i ≤ 710 ∣ ( S F ( [ M ] a ( j ) ) i ) ∣ > 43. (90) \min_{0\le j\le7}\max_{1\le i\le710}\left|(S_F([M]a^{(j)})_i)\right|>43. \tag*{(90)} 0 ≤ j ≤ 7 min 1 ≤ i ≤ 710 max ( S F ([ M ] a ( j ) ) i ) > 43. (90)
The last three integer parts follow from (3.6) by cross-multiplication. Thus (3.7) holds. □
Proof of Numerical Lemma 3. In the five free coordinates,
q e = c e + r e u e , ∣ u e ∣ ≤ 1. q_e=c_e+r_eu_e,\qquad|u_e|\le1. q e = c e + r e u e , ∣ u e ∣ ≤ 1.
The enclosure ( P , r ) (P,r) ( P , r ) satisfies
∣ h ( u ) − P ( u ) ∣ ≤ r ( u ∈ [ − 1 , 1 ] 5 ) , deg P ≤ d , P ∈ 2 − 40 Z [ u ] , r ∈ 2 − 40 Z ≥ 0 . |h(u)-P(u)|\le r\quad(u\in[-1,1]^5),\qquad\deg P\le d,\qquad P\in2^{-40}\mathbb{Z}[u],\qquad r\in2^{-40}\mathbb{Z}_{\ge0}. ∣ h ( u ) − P ( u ) ∣ ≤ r ( u ∈ [ − 1 , 1 ] 5 ) , deg P ≤ d , P ∈ 2 − 40 Z [ u ] , r ∈ 2 − 40 Z ≥ 0 .
( P , r ) + ( Q , s ) = ( P + Q , r + s ) , − ( P , r ) = ( − P , r ) , ⌈ a ⌉ 2 − 40 = 2 − 40 ⌈ 2 40 a ⌉ . (P,r)+(Q,s)=(P+Q,r+s),\qquad-(P,r)=(-P,r),\qquad\lceil a\rceil_{2^{-40}}=2^{-40}\lceil2^{40}a\rceil. ( P , r ) + ( Q , s ) = ( P + Q , r + s ) , − ( P , r ) = ( − P , r ) , ⌈ a ⌉ 2 − 40 = 2 − 40 ⌈ 2 40 a ⌉ .
For d ∈ { 1 , 2 } d \in\{1,2\} d ∈ { 1 , 2 } ,
P = ∑ α ∈ N 0 5 ∣ α ∣ ≤ d p α u α , Q = ∑ β ∈ N 0 5 ∣ β ∣ ≤ d q β u β . P=\sum_{\substack{\alpha\in\mathbb{N}_0^5\\|\alpha|\le d}}p_{\alpha}u^{\alpha},\qquad Q=\sum_{\substack{\beta\in\mathbb{N}_0^5\\|\beta|\le d}}q_{\beta}u^{\beta}. P = α ∈ N 0 5 ∣ α ∣ ≤ d ∑ p α u α , Q = β ∈ N 0 5 ∣ β ∣ ≤ d ∑ q β u β .
The product enclosure is
P ′ = ∑ γ ∈ N 0 5 ∣ γ ∣ ≤ d T 2 − 40 ( ∑ α , β ∈ N 0 5 α + β = γ p α q β ) u γ , P'=\sum_{\substack{\gamma\in\mathbb{N}_0^5\\|\gamma|\le d}}\mathcal{T}_{2^{-40}}\left(\sum_{\substack{\alpha,\beta\in\mathbb{N}_0^5\\\alpha+\beta=\gamma}}p_{\alpha}q_{\beta}\right)u^{\gamma}, P ′ = γ ∈ N 0 5 ∣ γ ∣ ≤ d ∑ T 2 − 40 α , β ∈ N 0 5 α + β = γ ∑ p α q β u γ ,
r ′ = ⌈ s ∑ α ∈ N 0 5 ∣ α ∣ ≤ d ∣ p α ∣ + r ∑ β ∈ N 0 5 ∣ β ∣ ≤ d ∣ q β ∣ + r s + ∑ α , β ∈ N 0 5 ∣ α ∣ , ∣ β ∣ ≤ d , ∣ α + β ∣ > d ∣ p α q β ∣ ⌉ 2 − 40 + 2 − 40 ⌈ { # { γ ∈ N 0 5 : ∣ γ ∣ ≤ 1 , ∑ α , β ∈ N 0 5 α + β = γ p α q β ∉ 2 − 40 Z } , d = 1 , 21 , d = 2. (91) r'=\left\lceil s\sum_{\substack{\alpha\in\mathbb{N}_0^5\\|\alpha|\le d}}|p_{\alpha}|+r\sum_{\substack{\beta\in\mathbb{N}_0^5\\|\beta|\le d}}|q_{\beta}|+rs
+\sum_{\substack{\alpha,\beta\in\mathbb{N}_0^5\\|\alpha|,|\beta|\le d,\ |\alpha+\beta|>d}}|p_{\alpha}q_{\beta}|\right\rceil_{2^{-40}}
+2^{-40}\left\lceil
\begin{cases}
\#\left\{\gamma\in\mathbb{N}_0^5:|\gamma|\le1,\ \sum_{\substack{\alpha,\beta\in\mathbb{N}_0^5\\\alpha+\beta=\gamma}}p_{\alpha}q_{\beta}\notin2^{-40}\mathbb{Z}\right\}, & d=1,\\
21, & d=2.
\end{cases}
\right. \tag*{(91)} r ′ = s α ∈ N 0 5 ∣ α ∣ ≤ d ∑ ∣ p α ∣ + r β ∈ N 0 5 ∣ β ∣ ≤ d ∑ ∣ q β ∣ + r s + α , β ∈ N 0 5 ∣ α ∣ , ∣ β ∣ ≤ d , ∣ α + β ∣ > d ∑ ∣ p α q β ∣ 2 − 40 + 2 − 40 ⎩ ⎨ ⎧ # { γ ∈ N 0 5 : ∣ γ ∣ ≤ 1 , ∑ α , β ∈ N 0 5 α + β = γ p α q β ∈ / 2 − 40 Z } , 21 , d = 1 , d = 2. (91)
Since ∣ u α ∣ ≤ 1 |u^{\alpha}|\le1 ∣ u α ∣ ≤ 1 , the discarded coefficients and input errors give (A.7). Use rational arithmetic for each coefficient sum before truncation and for the radius before its single ceiling. For dyadic a a a , the degree-two rule is
S a ( P , r ) = ( T 2 − 40 ( a P ) , ∥ a ∥ r ) 2 − 40 + 21 ⋅ 2 − 40 , (92) S_a(P,r)=\left(\mathcal{T}_{2^{-40}}(aP),\left\|a\right\|r\right)_{2^{-40}}+21\cdot2^{-40}, \tag*{(92)} S a ( P , r ) = ( T 2 − 40 ( a P ) , ∥ a ∥ r ) 2 − 40 + 21 ⋅ 2 − 40 , (92)
a ∈ { 0 , 1 , − 1 } a\in\{0,1,-1\} a ∈ { 0 , 1 , − 1 } acts without rounding.
For the degree-one reciprocal,
a = P ( 0 ) , A = r + ∑ α ∈ N 0 5 ∣ α ∣ = 1 ∣ p α ∣ , a=P(0),\qquad A=r+\sum_{\substack{\alpha\in\mathbb{N}_0^5\\|\alpha|=1}}|p_{\alpha}|, a = P ( 0 ) , A = r + α ∈ N 0 5 ∣ α ∣ = 1 ∑ ∣ p α ∣ ,
a > A , β = ⌊ A / a ⌋ 2 − 40 < 1. a>A,\qquad\beta=\left\lfloor A/a\right\rfloor_{2^{-40}}<1. a > A , β = ⌊ A / a ⌋ 2 − 40 < 1.
Set b = T 2 − 40 ( 1 / a ) b=\mathcal{T}_{2^{-40}}(1/a) b = T 2 − 40 ( 1/ a ) . Evaluate b − ( b b ) ( P − a , r ) b-(bb)(P-a,r) b − ( bb ) ( P − a , r ) by (A.7), then add to its radius
2 − 40 + ⌈ 2 − 40 ( 2 b + 2 − 40 ) A ⌉ 2 − 40 + ⌈ ( b + 2 − 40 ) [ β 2 ] 2 − 40 1 − β ⌉ 2 − 40 . (93) 2^{-40}+\left\lceil2^{-40}(2b+2^{-40})A\right\rceil_{2^{-40}}+\left\lceil\frac{(b+2^{-40})[\beta^2]_{2^{-40}}}{1-\beta}\right\rceil_{2^{-40}}. \tag*{(93)} 2 − 40 + ⌈ 2 − 40 ( 2 b + 2 − 40 ) A ⌉ 2 − 40 + ⌈ 1 − β ( b + 2 − 40 ) [ β 2 ] 2 − 40 ⌉ 2 − 40 . (93)
Indeed,
1 a + x = 1 a − x a 2 + x 2 a 2 ( a + x ) , 0 ≤ a − 1 − b < 2 − 40 , x 2 a 2 ( a + x ) ≤ ( b + 2 − 40 ) β 2 1 − β ( ∣ x ∣ ≤ A ) . \frac{1}{a+x}=\frac{1}{a}-\frac{x}{a^2}+\frac{x^2}{a^2(a+x)},\qquad0\le a^{-1}-b<2^{-40},\qquad\frac{x^2}{a^2(a+x)}\le\frac{(b+2^{-40})\beta^2}{1-\beta}\quad(|x|\le A). a + x 1 = a 1 − a 2 x + a 2 ( a + x ) x 2 , 0 ≤ a − 1 − b < 2 − 40 , a 2 ( a + x ) x 2 ≤ 1 − β ( b + 2 − 40 ) β 2 ( ∣ x ∣ ≤ A ) .
A lower endpoint is
L ( P , r ) = p 0 − r − ∑ α ∈ N 0 5 , 0 < ∣ α ∣ ≤ d p α < 0 or α i odd for some 1 ≤ i ≤ 5 ∣ p α ∣ (94) \mathcal{L}(P,r)=p_0-r-\sum_{\substack{\alpha\in\mathbb{N}_0^5,\ 0<|\alpha|\le d\\p_{\alpha}<0\ \text{or}\ \alpha_i\ \text{odd for some }1\le i\le5}}|p_{\alpha}| \tag*{(94)} L ( P , r ) = p 0 − r − α ∈ N 0 5 , 0 < ∣ α ∣ ≤ d p α < 0 or α i odd for some 1 ≤ i ≤ 5 ∑ ∣ p α ∣ (94)
with
∣ u α ∣ ≤ 1 , p α ≥ 0 , α ∈ ( 2 N 0 ) 5 ⟹ p α u α ≥ 0 , h ( u ) ≥ L ( P , r ) . |u^{\alpha}|\le1,\qquad p_{\alpha}\ge0,\ \alpha\in(2\mathbb{N}_0)^5\Longrightarrow p_{\alpha}u^{\alpha}\ge0,\qquad h(u)\ge\mathcal{L}(P,r). ∣ u α ∣ ≤ 1 , p α ≥ 0 , α ∈ ( 2 N 0 ) 5 ⟹ p α u α ≥ 0 , h ( u ) ≥ L ( P , r ) .
For polynomial trials, use degree two in (73) and
τ = 2 − 18 max { j ∈ Z : 0 ≤ j < 2 26 , j 3 < 2 54 27 T 3 D ( c ) 2 } . (95) \tau= 2^{-18}\max\{j \in\mathbb{Z}: 0 \le j < 2^{26},\ j^3 < 2^{54}27T^3D(c)^2\}. \tag*{(95)} τ = 2 − 18 max { j ∈ Z : 0 ≤ j < 2 26 , j 3 < 2 54 27 T 3 D ( c ) 2 } . (95)
For h ≤ 40 h \le40 h ≤ 40 ,
c 01 = 1 , c e ∈ 2 − 9 Z ∩ [ 2 − 9 , 1 ] ( e ≠ 01 ) , c_{01}=1,\qquad c_e \in2^{-9}\mathbb{Z}\cap[2^{-9},1]\quad(e\ne01), c 01 = 1 , c e ∈ 2 − 9 Z ∩ [ 2 − 9 , 1 ] ( e = 01 ) ,
so
D ( c ) ≥ c 01 c 02 c 03 ≥ 2 − 18 , 1 < 2 54 27 T 3 D ( c ) 2 , τ > 0. D(c)\ge c_{01}c_{02}c_{03}\ge2^{-18},\qquad1<2^{54}27T^3D(c)^2,\qquad\tau>0. D ( c ) ≥ c 01 c 02 c 03 ≥ 2 − 18 , 1 < 2 54 27 T 3 D ( c ) 2 , τ > 0.
Use
q e ⟼ ( q e , 0 ) , q e ⟼ ( q e , 2 − 39 ) ( F C C ) , a ∈ Q ⟼ ( T 2 − 40 ( a ) , 2 − 40 ) . q_e\longmapsto(q_e,0),\qquad q_e\longmapsto(q_e,2^{-39})\quad(\mathrm{FCC}),\qquad a\in\mathbb{Q}\longmapsto(\mathcal{T}_{2^{-40}}(a),2^{-40}). q e ⟼ ( q e , 0 ) , q e ⟼ ( q e , 2 − 39 ) ( FCC ) , a ∈ Q ⟼ ( T 2 − 40 ( a ) , 2 − 40 ) .
Compute in order
∑ 0 ≤ i < j ≤ 3 S C i j ( c ) q i j , ∏ e ∈ J q e = ( q i q j ) q k ( J = { i , j , k } , i < j < k ) , \sum_{0\le i<j\le3}S_{C_{ij}}(c)q_{ij},\qquad\prod_{e\in J}q_e=(q_iq_j)q_k\quad(J=\{i,j,k\},\ i<j<k), 0 ≤ i < j ≤ 3 ∑ S C ij ( c ) q ij , e ∈ J ∏ q e = ( q i q j ) q k ( J = { i , j , k } , i < j < k ) ,
S τ M − ( ∑ 0 ≤ i < j ≤ 3 S C i j ( c ) q i j ) K . S_{\tau}M-\left(\sum_{0\le i<j\le3}S_{C_{ij}}(c)q_{ij}\right)K. S τ M − ( 0 ≤ i < j ≤ 3 ∑ S C ij ( c ) q ij ) K .
Products use (A.7); dyadic scalars use S a S_a S a .
At each cubature node,
x J , i ⟼ ( T 2 − 40 x J , i , 6 ⋅ 2 − 40 ) ( 1 ≤ i ≤ 3 ) , x_{J,i}\longmapsto(\mathcal{T}_{2^{-40}}x_{J,i},6\cdot2^{-40})\qquad(1\le i\le3), x J , i ⟼ ( T 2 − 40 x J , i , 6 ⋅ 2 − 40 ) ( 1 ≤ i ≤ 3 ) ,
Γ i j = ∑ 0 ≤ a < b ≤ 3 S ( b a − b b ) i ( b a − b b ) j / 16 q a b ( 1 ≤ i , j ≤ 3 ) . \Gamma_{ij}=\sum_{0\le a<b\le3}S_{(b_a-b_b)i(b_a-b_b)j/16}q_{ab}\qquad(1\le i,j\le3). Γ ij = 0 ≤ a < b ≤ 3 ∑ S ( b a − b b ) i ( b a − b b ) j /16 q ab ( 1 ≤ i , j ≤ 3 ) .
For FCC, apply in order
( z 1 , z 2 , z 3 ) ← ( x 3 , x 1 + x 2 , x 1 − x 2 ) , (z_1,z_2,z_3)\leftarrow(x_3,x_1+x_2,x_1-x_2), ( z 1 , z 2 , z 3 ) ← ( x 3 , x 1 + x 2 , x 1 − x 2 ) ,
( Γ 1 j , Γ 2 j , Γ 3 j ) ← ( Γ 3 j , Γ 1 j + Γ 2 j , Γ 1 j − Γ 2 j ) ( 1 ≤ j ≤ 3 ) , (\Gamma_{1j},\Gamma_{2j},\Gamma_{3j})\leftarrow(\Gamma_{3j},\Gamma_{1j}+\Gamma_{2j},\Gamma_{1j}-\Gamma_{2j})\qquad(1\le j\le3), ( Γ 1 j , Γ 2 j , Γ 3 j ) ← ( Γ 3 j , Γ 1 j + Γ 2 j , Γ 1 j − Γ 2 j ) ( 1 ≤ j ≤ 3 ) ,
( Γ i 1 , Γ i 2 , Γ i 3 ) ← ( Γ i 3 , Γ i 1 + Γ i 2 , Γ i 1 − Γ i 2 ) ( 1 ≤ i ≤ 3 ) . (\Gamma_{i1},\Gamma_{i2},\Gamma_{i3})\leftarrow(\Gamma_{i3},\Gamma_{i1}+\Gamma_{i2},\Gamma_{i1}-\Gamma_{i2})\qquad(1\le i\le3). ( Γ i 1 , Γ i 2 , Γ i 3 ) ← ( Γ i 3 , Γ i 1 + Γ i 2 , Γ i 1 − Γ i 2 ) ( 1 ≤ i ≤ 3 ) .
Each assignment is simultaneous. Apply (A.8) with factor
2 − 24 [ 12 ⋅ 2 24 3 ( c 01 + c 02 + c 13 + c 23 ) + 4 ( c 03 + c 12 ) ] 2^{-24}\left[\frac{12\cdot2^{24}}{3(c_{01}+c_{02}+c_{13}+c_{23})+4(c_{03}+c_{12})}\right] 2 − 24 [ 3 ( c 01 + c 02 + c 13 + c 23 ) + 4 ( c 03 + c 12 ) 12 ⋅ 2 24 ]
once to each z i z_i z i and twice successively to each Γ i j \Gamma_{ij} Γ ij ; evaluate the scale once and reuse it. For cubic monomials and their derivatives, start with the enclosure ( 1 , 0 ) (1,0) ( 1 , 0 ) and use successive powers, grouped as ( x 1 a x 2 b ) x 3 c (x_1^a x_2^b)x_3^c ( x 1 a x 2 b ) x 3 c . For quintics, in increasing degree,
x 0 = ( 1 , 0 ) , x e i = x i , x α = x α − e i x i , i = min { j ∈ { 1 , 2 , 3 } : α j > 0 } , ∣ α ∣ ≥ 2 , x^0=(1,0),\qquad x^{e_i}=x_i,\qquad x^\alpha=x^{\alpha-e_i}x_i,\qquad i=\min\{j\in\{1,2,3\}:\alpha_j>0\},\qquad|\alpha|\ge2, x 0 = ( 1 , 0 ) , x e i = x i , x α = x α − e i x i , i = min { j ∈ { 1 , 2 , 3 } : α j > 0 } , ∣ α ∣ ≥ 2 ,
For FCC, use coordinate powers grouped as
v a l u e , ∂ 1 : z 1 a ( z 2 b z 3 c ) , ∂ 2 , ∂ 3 : ( z 1 a z 2 b ) z 3 c . \mathrm{value},\ \partial_1:z_1^a(z_2^bz_3^c),\qquad\partial_2,\ \partial_3:(z_1^az_2^b)z_3^c. value , ∂ 1 : z 1 a ( z 2 b z 3 c ) , ∂ 2 , ∂ 3 : ( z 1 a z 2 b ) z 3 c .
Lower the differentiated exponent; omit zero coefficients. For independently evaluated enclosures of f f f and ∂ i f \partial_i f ∂ i f , each tree gives
m a s s : ∑ y ∈ [ − 1 / 2 , 1 / 2 ] 3 w y ≠ 0 w y ( f f ) , \mathrm{mass}:\qquad\sum_{\substack{y\in[-1/2,1/2]^3\\w_y\ne0}}w_y(f f), mass : y ∈ [ − 1/2 , 1/2 ] 3 w y = 0 ∑ w y ( f f ) ,
e n e r g y : ∑ y ∈ [ − 1 / 2 , 1 / 2 ] 3 w y ≠ 0 w y ∑ i = 1 3 ∑ j = 1 i S 2 − 1 i = j ( Γ i j ( ( ∂ i f ) ( ∂ j f ) ) ) . \mathrm{energy}:\qquad\sum_{\substack{y\in[-1/2,1/2]^3\\w_y\ne0}}w_y\sum_{i=1}^{3}\sum_{j=1}^{i}S_{2^{-1}i=j}\bigl(\Gamma_{ij}((\partial_i f)(\partial_j f))\bigr). energy : y ∈ [ − 1/2 , 1/2 ] 3 w y = 0 ∑ w y i = 1 ∑ 3 j = 1 ∑ i S 2 − 1 i = j ( Γ ij (( ∂ i f ) ( ∂ j f )) ) .
Omit zero factors; multiply by ( q i q j ) q k (q_iq_j)q_k ( q i q j ) q k , then sum over trees.
For A = ( 0 , a 1 , a 2 , a 3 ) ∈ T + A=(0,a_1,a_2,a_3)\in\mathcal{T}_+ A = ( 0 , a 1 , a 2 , a 3 ) ∈ T + and s ∈ { − 1 , 1 } 3 s\in\{-1,1\}^3 s ∈ { − 1 , 1 } 3 , let V q ( w ) V_q(w) V q ( w ) denote the corresponding vertex or face average in (82) :
H i j = ∑ k = 1 3 s i ( V q ( diag ( s 1 , s 2 , s 3 ) a k ) ) i ( [ a 1 , a 2 , a 3 ] − 1 ) k j ( 1 ≤ i , j ≤ 3 ) . H_{ij}=\sum_{k=1}^{3}s_i\bigl(V_q(\operatorname{diag}(s_1,s_2,s_3)a_k)\bigr)_i\bigl([a_1,a_2,a_3]^{-1}\bigr)_{kj}\qquad(1\leq i,j\leq3). H ij = k = 1 ∑ 3 s i ( V q ( diag ( s 1 , s 2 , s 3 ) a k ) ) i ( [ a 1 , a 2 , a 3 ] − 1 ) k j ( 1 ≤ i , j ≤ 3 ) .
Use affine coefficient multiplication and addition with all radii, and
a ∈ 2 − 40 Z : ( a , 0 ) , a ∉ 2 − 40 Z : ( T 2 − 40 ( a ) , 2 − 40 ) . a\in2^{-40}\mathbb{Z}:(a,0),\qquad a\notin2^{-40}\mathbb{Z}:(T_{2^{-40}}(a),2^{-40}). a ∈ 2 − 40 Z : ( a , 0 ) , a ∈ / 2 − 40 Z : ( T 2 − 40 ( a ) , 2 − 40 ) .
With indices modulo three,
( adj H ) i j = H j + 1 , i + 1 H j + 2 , i + 2 − H j + 1 , i + 2 H j + 2 , i + 1 , (\operatorname{adj}H)_{ij}=H_{j+1,i+1}H_{j+2,i+2}-H_{j+1,i+2}H_{j+2,i+1}, ( adj H ) ij = H j + 1 , i + 1 H j + 2 , i + 2 − H j + 1 , i + 2 H j + 2 , i + 1 ,
det H = ∑ i = 1 3 H 1 i ( adj H ) i 1 . \det H=\sum_{i=1}^{3}H_{1i}(\operatorname{adj}H)_{i1}. det H = i = 1 ∑ 3 H 1 i ( adj H ) i 1 .
Form ( s i s j Γ i j ) i , j = 1 3 (s_is_j\Gamma_{ij})_{i,j=1}^{3} ( s i s j Γ ij ) i , j = 1 3 by sign changes without rounding. Multiply first on the left by adj H \operatorname{adj}H adj H , then on the right by ( adj H ) T (\operatorname{adj}H)^T ( adj H ) T ; reuse the resulting entries below. Each of the 378 reference integrals below, evaluated by (46) , is enclosed once by its nearest 2 − 40 2^{-40} 2 − 40 -multiple (ties even), with radius 2 − 40 2^{-40} 2 − 40 ; the formulas below reuse these enclosures:
M = ∑ 1 ≤ i ≤ j ≤ 3 ( 2 − 1 i = j ) v i v j ∑ A ∈ T + ∑ s ∈ { − 1 , 1 } 3 s i s j ( ( det H ) ∫ A P i P j ) , M=\sum_{1\leq i\leq j\leq3}(2-1_{i=j})v_iv_j\sum_{A\in\mathcal{T}_+}\sum_{s\in\{-1,1\}^3}s_is_j\left((\det H)\int_A P_iP_j\right), M = 1 ≤ i ≤ j ≤ 3 ∑ ( 2 − 1 i = j ) v i v j A ∈ T + ∑ s ∈ { − 1 , 1 } 3 ∑ s i s j ( ( det H ) ∫ A P i P j ) ,
K = ∑ 1 ≤ i ≤ j ≤ 3 ( 2 − 1 i = j ) v i v j ∑ A ∈ T + ∑ s ∈ { − 1 , 1 } 3 s i s j × ( ( ∑ 1 ≤ k ≤ l ≤ 3 ( ( adj H ) ( s a s b Γ a b ) a , b = 1 3 ( adj H ) T ) k l ) × ∫ A ( ∂ k P i ∂ l P j + 1 k < l ∂ l P i ∂ k P j ) ( det H ) − 1 ) . (96) \begin{aligned}
K={}&\sum_{1\leq i\leq j\leq3}(2-1_{i=j})v_iv_j\sum_{A\in\mathcal{T}_+}\sum_{s\in\{-1,1\}^3}s_is_j\\
&\times\left(\left(\sum_{1\leq k\leq l\leq3}\bigl((\operatorname{adj}H)(s_as_b\Gamma_{ab})_{a,b=1}^{3}(\operatorname{adj}H)^T\bigr)_{kl}\right)\right.\\
&\qquad\left.\times\int_A(\partial_kP_i\partial_lP_j+1_{k<l}\partial_lP_i\partial_kP_j)(\det H)^{-1}\right). \tag*{(96)}
\end{aligned} K = 1 ≤ i ≤ j ≤ 3 ∑ ( 2 − 1 i = j ) v i v j A ∈ T + ∑ s ∈ { − 1 , 1 } 3 ∑ s i s j × ( ( 1 ≤ k ≤ l ≤ 3 ∑ ( ( adj H ) ( s a s b Γ ab ) a , b = 1 3 ( adj H ) T ) k l ) × ∫ A ( ∂ k P i ∂ l P j + 1 k < l ∂ l P i ∂ k P j ) ( det H ) − 1 ) . (96)
Keep s i s j ∈ { − 1 , 1 } s_is_j\in\{-1,1\} s i s j ∈ { − 1 , 1 } unrounded and enclose ( 2 − 1 i = j ) v i v j (2-1_{i=j})v_iv_j ( 2 − 1 i = j ) v i v j as one rational. Use the affine reciprocal for det H \det H det H and
D = ∑ { a , b , c } ⊂ { J 1 , … , J 16 } a < b < c ( q a q b ) q c . D=\sum_{\substack{\{a,b,c\}\subset\{J_1,\ldots,J_{16}\}\\a<b<c}}(q_aq_b)q_c. D = { a , b , c } ⊂ { J 1 , … , J 16 } a < b < c ∑ ( q a q b ) q c .
Enclose T T T as one rational constant; evaluate T M TM T M once and reuse it. Test the lower endpoints of M M M and
( ( T M ) ( T M ) ) ( T M ) − ( ( D K ) K ) K . (97) ((TM)(TM))(TM)-((DK)K)K. \tag*{(97)} (( T M ) ( T M )) ( T M ) − (( D K ) K ) K . (97)
For center selection, round rational inputs once and every operation and square root to binary64 (ties even, signed zero). Sums start at 0 in displayed order, otherwise increasing-index order; square roots use squared dyadic midpoints.
For cubic and quintic trials, s = 3 , 5 s=3,5 s = 3 , 5 , let m α m_\alpha m α be the rounded cubature moment of x α x^\alpha x α at c c c , α ∈ N 0 3 \alpha\in\mathbb{N}_0^3 α ∈ N 0 3 , ∣ α ∣ ≤ 2 s |\alpha|\leq2s ∣ α ∣ ≤ 2 s even. Use lexicographic tree and node-coordinate order, rounded coordinates x J ( y ; c ) x_J(y;c) x J ( y ; c ) , successive powers, and weights
[ ( ∏ e ∈ J c e ) T 2 − 40 ( w y ) ] , \left[\left(\prod_{e\in J}c_e\right)T_{2^{-40}}(w_y)\right], [ ( e ∈ J ∏ c e ) T 2 − 40 ( w y ) ] ,
with rational w y w_y w y . Accumulate α = ( a , b , d ) \alpha=(a,b,d) α = ( a , b , d ) terms as ( ( w x 1 a ) x 2 b ) x 3 d ((wx_1^a)x_2^b)x_3^d (( w x 1 a ) x 2 b ) x 3 d . The lower triangles are
M α β = m α + β , K α β = ∑ 1 ≤ i , j ≤ 3 α i β j > 0 ( α i β j Γ i j ( c ) ) m α + β − e i − e j . (98) M_{\alpha\beta}=m_{\alpha+\beta},\qquad K_{\alpha\beta}=\sum_{\substack{1\leq i,j\leq3\\\alpha_i\beta_j>0}}(\alpha_i\beta_j\Gamma_{ij}(c))m_{\alpha+\beta-e_i-e_j}. \tag*{(98)} M α β = m α + β , K α β = 1 ≤ i , j ≤ 3 α i β j > 0 ∑ ( α i β j Γ ij ( c )) m α + β − e i − e j . (98)
i i i is the outer summation index; complete by symmetry. For FCC at q = c q=c q = c ,
( z 1 , z 2 , z 3 ) ← ( [ x 3 ] , [ x 1 + x 2 ] , [ x 1 − x 2 ] ) . (z_1,z_2,z_3)\gets([x_3],[x_1+x_2],[x_1-x_2]). ( z 1 , z 2 , z 3 ) ← ([ x 3 ] , [ x 1 + x 2 ] , [ x 1 − x 2 ]) .
Use the displayed row/column updates and one coordinate or two successive metric multiplications by the factor in (79) . Evaluate it once; round every multiplication to binary64. Use g i g_i g i without coefficient truncation, in the monomial order
z 1 , z 1 z 2 2 , z 1 z 3 2 , z 1 z 2 4 , z 1 z 3 4 , z 1 z 2 2 z 3 2 , z 1 3 , z 1 3 z 2 2 , z 1 3 z 3 2 , z 1 5 . z_1,z_1z_2^2,z_1z_3^2,z_1z_2^4,z_1z_3^4,z_1z_2^2z_3^2,z_1^3,z_1^3z_2^2,z_1^3z_3^2,z_1^5. z 1 , z 1 z 2 2 , z 1 z 3 2 , z 1 z 2 4 , z 1 z 3 4 , z 1 z 2 2 z 3 2 , z 1 3 , z 1 3 z 2 2 , z 1 3 z 3 2 , z 1 5 .
Preserve this order under coordinate interchange. In the lower triangle of (A.14), multiply by both monomial coefficients and sum pairs lexicographically. For transport, use the constant coefficients of the six affine enclosures of each of M ( P i , P j ) M(P_i,P_j) M ( P i , P j ) and K ( P i , P j ) K(P_i,P_j) K ( P i , P j ) .
For each ( M , K ) (M,K) ( M , K ) , the serial Cholesky recurrence is
r 0 = M i j , r k = [ r k − 1 − [ L i k L j k ] ] ( 1 ≤ k < j ≤ i ) , r_0=M_{ij},\qquad r_k=[r_{k-1}-[L_{ik}L_{jk}]]\quad(1\le k<j\le i), r 0 = M ij , r k = [ r k − 1 − [ L ik L j k ]] ( 1 ≤ k < j ≤ i ) ,
L i j = { r j − 1 , i = j , [ r j − 1 / L j j ] , i > j . L_{ij}=
\begin{cases}
\sqrt{r_{j-1}}, & i=j,\\
[r_{j-1}/L_{jj}], & i>j.
\end{cases} L ij = { r j − 1 , [ r j − 1 / L j j ] , i = j , i > j .
A nonpositive diagonal pivot rejects the test. By forward substitution, W = L − 1 W=L^{-1} W = L − 1 ; compute W K WK W K , then the lower triangle of E = ( W K ) W T E=(WK)W^{T} E = ( W K ) W T , and complete by symmetry.
Start V = I V=I V = I . For twelve lexicographic sweeps of i < j i<j i < j , update whenever ∣ E i j ∣ ≥ [ 10 − 13 ] |E_{ij}|\ge[10^{-13}] ∣ E ij ∣ ≥ [ 1 0 − 13 ] :
u = E j j − E i i 2 E i j , t = { 1 ∣ u ∣ + 1 + u 2 , u > 0 or u is + 0 , − 1 ∣ u ∣ + 1 + u 2 , u < 0 or u is − 0 , u=\frac{E_{jj}-E_{ii}}{2E_{ij}},\qquad
t=
\begin{cases}
\dfrac{1}{|u|+\sqrt{1+u^2}}, & u>0\text{ or }u\text{ is }+0,\\
\dfrac{-1}{|u|+\sqrt{1+u^2}}, & u<0\text{ or }u\text{ is }-0,
\end{cases} u = 2 E ij E j j − E ii , t = ⎩ ⎨ ⎧ ∣ u ∣ + 1 + u 2 1 , ∣ u ∣ + 1 + u 2 − 1 , u > 0 or u is + 0 , u < 0 or u is − 0 ,
a = 1 1 + t 2 , b = a t . (99) a=\frac{1}{\sqrt{1+t^2}},\qquad b=at. \tag*{(99)} a = 1 + t 2 1 , b = a t . (99)
Here + 0 +0 + 0 and − 0 -0 − 0 are the signed binary64 zeros. Apply
( E i i , E j j ) ← ( E i i − t E i j , E j j + t E i j ) , E i j = E j i ← 0 , (E_{ii},E_{jj})\gets(E_{ii}-tE_{ij},E_{jj}+tE_{ij}),\qquad E_{ij}=E_{ji}\gets0, ( E ii , E j j ) ← ( E ii − t E ij , E j j + t E ij ) , E ij = E j i ← 0 ,
( E k i , E k j ) = ( x , y ) ← ( a x − b y , b x + a y ) ( k ≠ i , j ) , (E_{ki},E_{kj})=(x,y)\gets(ax-by,bx+ay)\quad(k\ne i,j), ( E k i , E k j ) = ( x , y ) ← ( a x − b y , b x + a y ) ( k = i , j ) ,
E i k ← E k i , E j k ← E k j , E_{ik}\gets E_{ki},\qquad E_{jk}\gets E_{kj}, E ik ← E k i , E j k ← E k j ,
and the same two-column update rowwise to V V V . For the first minimizing diagonal index i i i ,
V ⋅ i ← W T V ⋅ i , v j = 2 − 24 nint ( 2 24 [ V j i max k ∣ V k i ∣ ] ) . V_{\cdot i}\gets W^{T}V_{\cdot i},\qquad
v_j=2^{-24}\operatorname{nint}\left(2^{24}\left[\frac{V_{ji}}{\max_k|V_{ki}|}\right]\right). V ⋅ i ← W T V ⋅ i , v j = 2 − 24 nint ( 2 24 [ max k ∣ V k i ∣ V j i ] ) .
Fix the maximum before normalization; reject the test if v = 0 v=0 v = 0 .
For ∣ w ∣ = h |w|=h ∣ w ∣ = h , stop at the first successful test in this order:
Chamber → \to → corner → \to → local (Section 5).
h ≥ 12 , q e − > 0 ( e ∈ E ) , min 0 ≤ i < j ≤ 3 c i j ≥ 1 2 ⟹ transported test . h\ge12,\qquad q_e^{-}>0\ (e\in E),\qquad\min_{0\le i<j\le3}c_{ij}\ge\frac{1}{2}
\quad\Longrightarrow\quad\text{transported test}. h ≥ 12 , q e − > 0 ( e ∈ E ) , 0 ≤ i < j ≤ 3 min c ij ≥ 2 1 ⟹ transported test .
h ≥ 11 , w 1 = ⋯ = w 11 = 1 ⟹ omit subsequent tests . h\ge11,\qquad w_1=\cdots=w_{11}=1
\quad\Longrightarrow\quad\text{omit subsequent tests}. h ≥ 11 , w 1 = ⋯ = w 11 = 1 ⟹ omit subsequent tests .
h ≥ 12 , ( h even or h ≥ 35 ) ⟹ cubic test . h\ge12,\qquad(h\text{ even or }h\ge35)
\quad\Longrightarrow\quad\text{cubic test}. h ≥ 12 , ( h even or h ≥ 35 ) ⟹ cubic test .
For h ≥ 20 h\ge20 h ≥ 20 with h h h even or h ≥ 35 h\ge35 h ≥ 35 ,
min ( c 03 , c 12 ) ≤ 1 3 , min ( c 02 , c 13 , c 23 ) ≥ 2 3 ⟹ FCC test , \min(c_{03},c_{12})\le\frac{1}{3},\qquad
\min(c_{02},c_{13},c_{23})\ge\frac{2}{3}
\quad\Longrightarrow\quad\text{FCC test}, min ( c 03 , c 12 ) ≤ 3 1 , min ( c 02 , c 13 , c 23 ) ≥ 3 2 ⟹ FCC test ,
followed by the quintic test. A failed reciprocal condition rejects that trial. Use the preceding center trials and (84) through depth 41;
∣ w ∣ ≤ 40 ⟹ q e ± ∈ 2 − 8 Z . |w| \le40 \quad\Longrightarrow\quad q_e^{\pm} \in2^{-8}\mathbb{Z}. ∣ w ∣ ≤ 40 ⟹ q e ± ∈ 2 − 8 Z .
The first-success counts are
( c h a m b e r , c o r n e r , l o c a l ) = ( 8655 , 1813 , 507 ) , (\mathrm{chamber},\mathrm{corner},\mathrm{local})=(8655,1813,507), ( chamber , corner , local ) = ( 8655 , 1813 , 507 ) ,
( c u b i c , q u i n t i c , F C C , t r a n s p o r t ) = ( 14556 , 3373 , 1653 , 61342 ) . (\mathrm{cubic},\mathrm{quintic},\mathrm{FCC},\mathrm{transport})=(14556,3373,1653,61342). ( cubic , quintic , FCC , transport ) = ( 14556 , 3373 , 1653 , 61342 ) .
The minimum lower endpoints are
min L ( τ M − ( ∑ 0 ≤ i < j ≤ 3 S C i j ( c ) q i j ) K ) c u b i c = 12339663 ⋅ 2 − 40 > 0 , \min\mathcal{L}\left(\tau_M-\left(\sum_{0\le i<j\le3}S_{C_{ij}(c)}q_{ij}\right)K\right)_{\mathrm{cubic}}=12339663\cdot2^{-40}>0, min L ( τ M − ( 0 ≤ i < j ≤ 3 ∑ S C ij ( c ) q ij ) K ) cubic = 12339663 ⋅ 2 − 40 > 0 ,
min L ( τ M − ( ∑ 0 ≤ i < j ≤ 3 S C i j ( c ) q i j ) K ) q u i n t i c = 2997515 ⋅ 2 − 40 > 0 , \min\mathcal{L}\left(\tau_M-\left(\sum_{0\le i<j\le3}S_{C_{ij}(c)}q_{ij}\right)K\right)_{\mathrm{quintic}}=2997515\cdot2^{-40}>0, min L ( τ M − ( 0 ≤ i < j ≤ 3 ∑ S C ij ( c ) q ij ) K ) quintic = 2997515 ⋅ 2 − 40 > 0 ,
min L ( τ M − ( ∑ 0 ≤ i < j ≤ 3 S C i j ( c ) q i j ) K ) F C C = 3874396 ⋅ 2 − 40 > 0 , \min\mathcal{L}\left(\tau_M-\left(\sum_{0\le i<j\le3}S_{C_{ij}(c)}q_{ij}\right)K\right)_{\mathrm{FCC}}=3874396\cdot2^{-40}>0, min L ( τ M − ( 0 ≤ i < j ≤ 3 ∑ S C ij ( c ) q ij ) K ) FCC = 3874396 ⋅ 2 − 40 > 0 ,
min L ( T 3 M 3 − D K 3 ) t r a n s p o r t = 799999930 ⋅ 2 − 40 > 0 , \min\mathcal{L}\left(T^3M^3-DK^3\right)_{\mathrm{transport}}=799999930\cdot2^{-40}>0, min L ( T 3 M 3 − D K 3 ) transport = 799999930 ⋅ 2 − 40 > 0 ,
min L ( M ) t r a n s p o r t = 684637240747 ⋅ 2 − 40 > 0. \min\mathcal{L}(M)_{\mathrm{transport}}=684637240747\cdot2^{-40}>0. min L ( M ) transport = 684637240747 ⋅ 2 − 40 > 0.
Finally,
# W 0 = 1 , # W h + 1 = 2 # { w ∈ W h : B w is not accepted } , # W 37 = 4. \#\mathcal{W}_0=1,\qquad\#\mathcal{W}_{h+1}=2\#\{w\in\mathcal{W}_h:B_w\text{ is not accepted}\},\qquad\#\mathcal{W}_{37}=4. # W 0 = 1 , # W h + 1 = 2# { w ∈ W h : B w is not accepted } , # W 37 = 4.
∑ h = 0 37 # { w ∈ W h : B w is accepted } = 91899 , \sum_{h=0}^{37}\#\{w\in\mathcal{W}_h:B_w\text{ is accepted}\}=91899, h = 0 ∑ 37 # { w ∈ W h : B w is accepted } = 91899 ,
∑ h = 0 36 # { w ∈ W h : B w is not accepted } = 91898 , \sum_{h=0}^{36}\#\{w\in\mathcal{W}_h:B_w\text{ is not accepted}\}=91898, h = 0 ∑ 36 # { w ∈ W h : B w is not accepted } = 91898 ,
{ w ∈ W 37 : B w is not accepted } = ∅ ⟹ W 38 = ∅ . \{w\in\mathcal{W}_{37}:B_w\text{ is not accepted}\}=\varnothing\quad\Longrightarrow\quad\mathcal{W}_{38}=\varnothing. { w ∈ W 37 : B w is not accepted } = ∅ ⟹ W 38 = ∅ .
Acknowledgements. The authors used AI to assist the long but finite computations.
(K. Deng) School of Mathematics and statistics, Jiangxi Normal University, Nanchang, 330022, China
Email address , K. Deng: Dengkaixin1999@163.com
(S. Luo) School of Mathematics and statistics, Jiangxi Normal University, Nanchang, 330022, China
Email address , S. Luo: luosp1989@163.com