Introduction
Let L L L be a lattice in R n \mathbb{R}^n R n . Its theta series is defined by
Θ L ( q ) = ∑ x ∈ L q ( x , x ) , 0 < q < 1 , \Theta_L(q) = \sum_{x\in L} q^{(x,x)}, \quad0 < q < 1, Θ L ( q ) = x ∈ L ∑ q ( x , x ) , 0 < q < 1 ,
where ( x , x ) (x,x) ( x , x ) is the Euclidean inner product. The Gaussian mass G L ( t ) G_L(t) G L ( t ) of L L L at t t t is the real number obtained upon letting q = exp ( − t ) q=\exp(-t) q = exp ( − t ) in θ L ( q ) \theta_L(q) θ L ( q ) . Applications of this invariant range from geometry of numbers [20 ] , to computer science [16 ] , to cryptography [3 ] . It was conjectured for stable lattices L L L [20 ] that
∀ t ∈ ( 0 , ∞ ) , G L ( t ) ≤ G Z n ( t ) . \forall t \in(0,\infty),\quad G_L(t) \leq G_{\mathbb{Z}^n}(t). ∀ t ∈ ( 0 , ∞ ) , G L ( t ) ≤ G Z n ( t ) .
In particular Kominers has shown in [14 ] that this bound holds for lattices L L L obtained by Construction A from a binary code. The main tool in his proof is the connection between weight enumerator of the code, Jacobi modular forms and theta series of the lattice.
This question is related to the security of the wiretap Gaussian channel, and more generally to Physical Layer security, a paradigm where the inherent randomness of communication channels is leveraged to ensure security against an eavesdropper. In particular the secrecy gain [18 ] the maximum of the ratio G Z n ( t ) / G L ( t ) G_{\mathbb{Z}^n}(t)/G_L(t) G Z n ( t ) / G L ( t ) are of critical importance in that area. The Belfiore-Solé conjecture which states that the secrecy gain is achieved in the center of the fundamental domain is proved for most unimodular lattices in dimension < 32 <32 < 32 [10 ] . The lattices where a Kominers-like bound hold have a secrecy gain > 1 >1 > 1 , a safe protection against an eavesdropper.
In this paper, we establish the following Gaussian-mass inequalities, in relation with the conjecture in [20 ] .
genus bounds for quaternary codes under conditions on their binary residues.
As usual the Type I result is harder to achieve, and more technical to derive than the Type II result and rely on shadow theory [6 , 19 ] . Extremality of the lattice determines an initial truncation of the normalized theta series. A positive Bernstein-basis representation controls this truncation, while an exact coefficient inversion of the shadow expansion controls the remaining terms. Poisson summation on a translate of the lattice then gives a uniform estimate that completes the proof.
The paper is organized as follows. § 2 collects the preliminaries. § 3 relates quaternary weight enumerators to lattice theta series, and § 4 develops the averaging arguments for Constructions A 4 A_4 A 4 and A 6 A_6 A 6 . § 5 treats extremal Type II lattices and their genus averages. § 6 proves the inequality for extremal Type I lattices. Finally, § 7 studies the higher-genus inequalities. § 8 concludes the article.
Preliminaries
We recall the basic notation concerning self-dual codes, Construction A lattices, Gaussian mass, extremal unimodular lattices, and higher-genus theta series.
Let M ≥ 2 M \ge2 M ≥ 2 and write Z M = Z / M Z \mathbb{Z}_M = \mathbb{Z}/M\mathbb{Z} Z M = Z / M Z . A linear code of length n n n over Z M \mathbb{Z}_M Z M is a Z M \mathbb{Z}_M Z M -submodule C ⊆ Z M n C \subseteq\mathbb{Z}_M^n C ⊆ Z M n . Its dual code, with respect to the standard inner product, is C ⊥ = { x ∈ Z M n : x ⋅ c = 0 ( m o d M ) for every c ∈ C } C^\perp= \{x \in\mathbb{Z}_M^n : x \cdot c = 0 \pmod{M}\text{ for every }c \in C\} C ⊥ = { x ∈ Z M n : x ⋅ c = 0 ( mod M ) for every c ∈ C } . The code C C C is called self-orthogonal if C ⊆ C ⊥ C \subseteq C^\perp C ⊆ C ⊥ , and self-dual if C = C ⊥ C = C^\perp C = C ⊥ . A code is called free of rank k k k if it is isomorphic, as a Z M \mathbb{Z}_M Z M -module, to Z M k \mathbb{Z}_M^k Z M k .
For a ∈ Z M a \in\mathbb{Z}_M a ∈ Z M , define its Euclidean weight by
w E ( a ) = min { r 2 : r ∈ Z , r ≡ a ( m o d M ) } , w_E(a) = \min\{r^2 : r \in\mathbb{Z},\ r \equiv a \pmod{M}\}, w E ( a ) = min { r 2 : r ∈ Z , r ≡ a ( mod M )} ,
and, for c = ( c 1 , … , c n ) ∈ Z M n c = (c_1,\ldots,c_n) \in\mathbb{Z}_M^n c = ( c 1 , … , c n ) ∈ Z M n , put w E ( c ) = ∑ i = 1 n w E ( c i ) w_E(c) = \sum_{i=1}^n w_E(c_i) w E ( c ) = ∑ i = 1 n w E ( c i ) . For even M M M , a self-dual code over Z M \mathbb{Z}_M Z M is called Type II if w E ( c ) ≡ 0 ( m o d 2 M ) w_E(c) \equiv0 \pmod{2M} w E ( c ) ≡ 0 ( mod 2 M ) for every c ∈ C c \in C c ∈ C ; otherwise it is called Type I. In particular, the divisibility conditions are modulo 8 over Z 4 \mathbb{Z}_4 Z 4 and modulo 12 over Z 6 \mathbb{Z}_6 Z 6 .
For a linear code C ⊆ Z M n C \subseteq\mathbb{Z}_M^n C ⊆ Z M n , its Construction A lattice is
A M ( C ) = 1 M { x ∈ Z n : x m o d M ∈ C } , A_M(C) = \frac{1}{\sqrt{M}}\{x \in\mathbb{Z}^n : x \bmod M \in C\}, A M ( C ) = M 1 { x ∈ Z n : x mod M ∈ C } ,
to simplify some formulas we let a M ( C ) = M a_M(C)=\sqrt{M} a M ( C ) = M . If C C C is self-dual, then A M ( C ) A_M(C) A M ( C ) is an integral unimodular lattice. If, in addition, C C C is Type II, then A M ( C ) A_M(C) A M ( C ) is even unimodular.
For a Z 4 \mathbb{Z}_4 Z 4 -code C C C , its binary residue code is Res ( C ) = { c m o d 2 : c ∈ C } ⊆ F 2 n \operatorname{Res}(C)=\{c\bmod2:c\in C\}\subseteq\mathbb{F}_2^n Res ( C ) = { c mod 2 : c ∈ C } ⊆ F 2 n . This binary code will be fixed in the averaging argument for free self-dual quaternary codes.
For modulus 6 6 6 , the Chinese remainder theorem gives Z 6 ≅ F 2 × F 3 \mathbb{Z}_6\cong\mathbb{F}_2\times\mathbb{F}_3 Z 6 ≅ F 2 × F 3 . Accordingly, a Z 6 \mathbb{Z}_6 Z 6 -code C C C determines the binary and ternary components C 2 = C m o d 2 ⊆ F 2 n C_2=C\bmod2\subseteq\mathbb{F}_2^n C 2 = C mod 2 ⊆ F 2 n , C 3 = C m o d 3 ⊆ F 3 n C_3=C\bmod3\subseteq\mathbb{F}_3^n C 3 = C mod 3 ⊆ F 3 n . Conversely,
C = { c ∈ Z 6 n : c m o d 2 ∈ C 2 , c m o d 3 ∈ C 3 } . C=\{c\in\mathbb{Z}_6^n:c\bmod2\in C_2,\ c\bmod3\in C_3\}. C = { c ∈ Z 6 n : c mod 2 ∈ C 2 , c mod 3 ∈ C 3 } .
Moreover, duality is compatible with the Chinese remainder decomposition, so that C C C is self-dual if and only if both C 2 C_2 C 2 and C 3 C_3 C 3 are self-dual. This decomposition will be used in the averaging argument for Construction A 6 A_6 A 6 .
For any finite ensemble E \mathcal{E} E and any function F F F on E \mathcal{E} E , we use the notation
E X ∈ E [ F ( X ) ] = 1 ∣ E ∣ ∑ X ∈ E F ( X ) . \mathbb{E}_{X\in\mathcal{E}}[F(X)]=\frac{1}{|\mathcal{E}|}\sum_{X\in\mathcal{E}}F(X). E X ∈ E [ F ( X )] = ∣ E ∣ 1 X ∈ E ∑ F ( X ) .
The two averaging procedures considered later correspond to fixing the binary component and averaging over the ternary component.
Let L ⊆ R n L\subseteq\mathbb{R}^n L ⊆ R n be a positive-definite lattice. Its theta series is
Θ L ( q ) = ∑ x ∈ L q ( x , x ) , 0 < q < 1. \Theta_L(q)=\sum_{x\in L}q^{(x,x)},\quad0<q<1. Θ L ( q ) = x ∈ L ∑ q ( x , x ) , 0 < q < 1.
We use the Jacobi theta function ϑ 3 ( q ) = ∑ m ∈ Z q m 2 \vartheta_3(q)=\sum_{m\in\mathbb{Z}}q^{m^2} ϑ 3 ( q ) = ∑ m ∈ Z q m 2 . Thus Θ Z n ( q ) = ϑ 3 ( q ) n \Theta_{\mathbb{Z}^n}(q)=\vartheta_3(q)^n Θ Z n ( q ) = ϑ 3 ( q ) n . Equivalently, putting q = e − π t q=e^{-\pi t} q = e − π t , t > 0 t>0 t > 0 , one may write
G L ( t ) = Θ L ( e − π t ) = ∑ x ∈ L e − π t ( x , x ) . G_L(t)=\Theta_L(e^{-\pi t})=\sum_{x\in L}e^{-\pi t(x,x)}. G L ( t ) = Θ L ( e − π t ) = x ∈ L ∑ e − π t ( x , x ) .
This is the Gaussian mass of L L L defined in the Introduction. The Gaussian-mass inequality considered throughout this paper is
Θ L ( q ) ≤ ϑ 3 ( q ) n , 0 < q < 1. \Theta_L(q)\leq\vartheta_3(q)^n,\quad0<q<1. Θ L ( q ) ≤ ϑ 3 ( q ) n , 0 < q < 1.
Equivalently,
∑ x ∈ L e − π t ( x , x ) ≤ ( ∑ m ∈ Z e − π t m 2 ) n , t > 0. \sum_{x\in L}e^{-\pi t(x,x)}\leq\left(\sum_{m\in\mathbb{Z}}e^{-\pi t m^2}\right)^n,\quad t>0. x ∈ L ∑ e − π t ( x , x ) ≤ ( m ∈ Z ∑ e − π t m 2 ) n , t > 0.
For lattices obtained by Construction A from binary self-dual codes, the corresponding inequality is known (see [14 ] ). The purpose of the averaging results below is to establish analogous inequalities on average for Constructions A 4 A_4 A 4 and A 6 A_6 A 6 .
For later use, we also recall the complete weight enumerator of a code C ⊆ Z M n C \subseteq\mathbb{Z}_M^n C ⊆ Z M n :
cwe C ( X a : a ∈ Z M ) = ∑ c ∈ C ∏ i = 1 n X c i . \operatorname{cwe}_C\left(X_a : a \in\mathbb{Z}_M\right)=\sum_{c\in C}\prod_{i=1}^{n}X_{c_i}. cwe C ( X a : a ∈ Z M ) = c ∈ C ∑ i = 1 ∏ n X c i .
For a ∈ Z M a \in\mathbb{Z}_M a ∈ Z M , define the local theta function
ϕ a ( M ) ( q ) = ∑ m ∈ Z q ( a + M m ) 2 / M . \phi_a^{(M)}(q)=\sum_{m\in\mathbb{Z}}q^{(a+Mm)^2/M}. ϕ a ( M ) ( q ) = m ∈ Z ∑ q ( a + M m ) 2 / M .
By decomposing A M ( C ) A_M(C) A M ( C ) into cosets of M Z n M\mathbb{Z}^n M Z n , one obtains the standard Construction A identity
Θ A M ( C ) ( q ) = cwe C ( ϕ a ( M ) ( q ) : a ∈ Z M ) . \Theta_{A_M(C)}(q)=\operatorname{cwe}_C\left(\phi_a^{(M)}(q):a\in\mathbb{Z}_M\right). Θ A M ( C ) ( q ) = cwe C ( ϕ a ( M ) ( q ) : a ∈ Z M ) .
In the quaternary case, if n j ( c ) n_j(c) n j ( c ) denotes the number of coordinates of c c c equal to j j j , the symmetrized weight enumerator is
swe C ( X , Y , Z ) = ∑ c ∈ C X n 0 ( c ) Y n 1 ( c ) + n 3 ( c ) Z n 2 ( c ) . \operatorname{swe}_C(X,Y,Z)=\sum_{c\in C}X^{n_0(c)}Y^{n_1(c)+n_3(c)}Z^{n_2(c)}. swe C ( X , Y , Z ) = c ∈ C ∑ X n 0 ( c ) Y n 1 ( c ) + n 3 ( c ) Z n 2 ( c ) .
It is the specialization cwe C ( X , Y , Z , Y ) \operatorname{cwe}_C(X,Y,Z,Y) cwe C ( X , Y , Z , Y ) of the complete weight enumerator. The corresponding relation between the symmetrized weight enumerator and the theta series of A 4 ( C ) A_4(C) A 4 ( C ) is the standard code–lattice theta-series correspondence (see [5 ] ).
We next recall the lattice terminology needed for the extremal case. For a positive-definite lattice L L L , its minimum norm is
min ( L ) = min { ( x , x ) : 0 ≠ x ∈ L } . \min(L)=\min\{(x,x):0\ne x\in L\}. min ( L ) = min {( x , x ) : 0 = x ∈ L } .
A lattice is unimodular if L = L ∗ L=L^* L = L ∗ , and is even if ( x , x ) ∈ 2 Z (x,x)\in2\mathbb{Z} ( x , x ) ∈ 2 Z for every x ∈ L x\in L x ∈ L . An integral unimodular lattice is of Type II if it is even, and of Type I if it is odd.
An even unimodular lattice can exist only in dimensions divisible by 8. If L L L is an even unimodular lattice of dimension n n n , then its minimum norm satisfies the classical bound (see [15 ] )
min ( L ) ≤ 2 ⌊ n 24 ⌋ + 2. \min(L)\le2\left\lfloor\frac{n}{24}\right\rfloor+2. min ( L ) ≤ 2 ⌊ 24 n ⌋ + 2.
Definition 1 An even unimodular lattice L L L of dimension n n n is called extremal if
min ( L ) = 2 ⌊ n 24 ⌋ + 2. \min(L) = 2\left\lfloor\frac{n}{24}\right\rfloor+ 2. min ( L ) = 2 ⌊ 24 n ⌋ + 2.
The theta series of an even unimodular lattice of dimension n n n is a modular form of weight n / 2 n/2 n /2 . Consequently, extremality imposes the maximal possible vanishing of its initial nonconstant Fourier coefficients. This modular-form description will be used later to study the Gaussian mass of extremal even unimodular lattices.
We finally introduce the higher-genus setting. For an integer g ≥ 1 g \ge1 g ≥ 1 , let
H g = { Z ∈ M g ( C ) : Z = Z T , Im ( Z ) > 0 } \mathbb{H}_g = \{Z \in M_g(\mathbb{C}) : Z = Z^T,\ \operatorname{Im}(Z) > 0\} H g = { Z ∈ M g ( C ) : Z = Z T , Im ( Z ) > 0 }
be the Siegel upper half-space of degree g g g . For x = ( x 1 , … , x g ) ∈ L g x = (x_1,\ldots,x_g) \in L^g x = ( x 1 , … , x g ) ∈ L g , let G ( x ) = ( ( x i , x j ) ) 1 ≤ i , j ≤ g G(x) = ((x_i,x_j))_{1 \le i,j \le g} G ( x ) = (( x i , x j ) ) 1 ≤ i , j ≤ g denote its Gram matrix.
Definition 2 The genus-g g g theta series of a lattice L L L is
Θ L ( g ) ( Z ) = ∑ ( x 1 , … , x g ) ∈ L g exp ( π i Tr ( G ( x ) Z ) ) , Z ∈ H g . \Theta_L^{(g)}(Z) = \sum_{(x_1,\ldots,x_g)\in L^g} \exp\left(\pi i\operatorname{Tr}(G(x)Z)\right),\quad Z \in\mathbb{H}_g. Θ L ( g ) ( Z ) = ( x 1 , … , x g ) ∈ L g ∑ exp ( π i Tr ( G ( x ) Z ) ) , Z ∈ H g .
For g = 1 g=1 g = 1 this reduces to the ordinary theta series. If L L L is even unimodular of dimension n n n , then Θ L ( g ) \Theta_L^{(g)} Θ L ( g ) is a Siegel modular form of degree g g g and weight n / 2 n/2 n /2 .
For the Gaussian formulation, let Y = Y T > 0 Y=Y^T>0 Y = Y T > 0 be a positive-definite real g × g g\times g g × g matrix and put
G L ( g ) ( Y ) = Θ L ( g ) ( i Y ) . G_L^{(g)}(Y) = \Theta_L^{(g)}(iY). G L ( g ) ( Y ) = Θ L ( g ) ( iY ) .
Equivalently,
G L ( g ) ( Y ) = ∑ x 1 , … , x g ∈ L exp ( − π ∑ a , b = 1 g Y a b ( x a , x b ) ) . G_L^{(g)}(Y) = \sum_{x_1,\ldots,x_g\in L} \exp\left(-\pi\sum_{a,b=1}^{g}Y_{ab}(x_a,x_b)\right). G L ( g ) ( Y ) = x 1 , … , x g ∈ L ∑ exp − π a , b = 1 ∑ g Y ab ( x a , x b ) .
Define
Θ g ( Y ) = ∑ m ∈ Z g e − π m T Y m . \Theta_g(Y) = \sum_{m\in\mathbb{Z}^g}e^{-\pi m^T Y m}. Θ g ( Y ) = m ∈ Z g ∑ e − π m T Y m .
Since Z n \mathbb{Z}^n Z n is an orthogonal direct product of n n n copies of Z \mathbb{Z} Z , we have
G Z n ( g ) ( Y ) = Θ g ( Y ) n . G_{\mathbb{Z}^n}^{(g)}(Y) = \Theta_g(Y)^n. G Z n ( g ) ( Y ) = Θ g ( Y ) n .
Thus the natural genus-g g g extension of the Gaussian-mass inequality is
G L ( g ) ( Y ) ≤ Θ g ( Y ) n , Y = Y T > 0. G_L^{(g)}(Y) \le\Theta_g(Y)^n,\qquad Y=Y^T>0. G L ( g ) ( Y ) ≤ Θ g ( Y ) n , Y = Y T > 0.
For g = 1 g=1 g = 1 , with Y = ( t ) Y=(t) Y = ( t ) , this is exactly the ordinary Gaussian-mass inequality. Notice also that for the scalar diagonal matrix Y = t I g Y=tI_g Y = t I g the variables separate and
G L ( g ) ( t I g ) = Θ L ( e − π t ) g . G_L^{(g)}(tI_g)=\Theta_L(e^{-\pi t})^g. G L ( g ) ( t I g ) = Θ L ( e − π t ) g .
Hence the genuinely higher-genus problem arises when Y Y Y is allowed to be an arbitrary positive-definite symmetric matrix.
To relate the higher-genus lattice theta series to codes, let C ⊆ Z M n C\subseteq\mathbb{Z}_M^n C ⊆ Z M n . Its genus-g g g complete weight enumerator is
cwe C ( g ) ( X a : a ∈ Z M g ) = ∑ ( c ( 1 ) , … , c ( g ) ) ∈ C g ∏ i = 1 n X ( c i ( 1 ) , … , c i ( g ) ) . \operatorname{cwe}_C^{(g)}\left(X_a:a\in\mathbb{Z}_M^g\right)
=\sum_{\left(c^{(1)},\ldots,c^{(g)}\right)\in C^g}\prod_{i=1}^{n}X_{\left(c_i^{(1)},\ldots,c_i^{(g)}\right)}. cwe C ( g ) ( X a : a ∈ Z M g ) = ( c ( 1 ) , … , c ( g ) ) ∈ C g ∑ i = 1 ∏ n X ( c i ( 1 ) , … , c i ( g ) ) .
For a ∈ Z M g a\in\mathbb{Z}_M^g a ∈ Z M g and Y = Y T > 0 Y=Y^T>0 Y = Y T > 0 , define
ϕ a ( M , g ) ( Y ) = ∑ m ∈ Z g exp ( − π M ( a + M m ) T Y ( a + M m ) ) . \phi_a^{(M,g)}(Y)=\sum_{m\in\mathbb{Z}^g}\exp\left(-\frac{\pi}{M}(a+Mm)^TY(a+Mm)\right). ϕ a ( M , g ) ( Y ) = m ∈ Z g ∑ exp ( − M π ( a + M m ) T Y ( a + M m ) ) .
The higher-genus Construction A identity is
G a M ( C ) ( g ) ( Y ) = cwe C ( g ) ( ϕ a ( M , g ) ( Y ) : a ∈ Z M g ) . G_{a_M(C)}^{(g)}(Y)=\operatorname{cwe}_C^{(g)}\left(\phi_a^{(M,g)}(Y):a\in\mathbb{Z}_M^g\right). G a M ( C ) ( g ) ( Y ) = cwe C ( g ) ( ϕ a ( M , g ) ( Y ) : a ∈ Z M g ) .
For M = 4 M=4 M = 4 , this gives the code-theoretic formulation used later in the genus-g g g extension of the Gaussian-mass problem (see also [2 ] ). The ordinary Gaussian-mass problem, its averaged versions for Constructions A 4 A_4 A 4 and A 6 A_6 A 6 , the extremal even-unimodular case, and the higher-genus formulation above will be considered in the following sections.
Weight enumerators and theta series
The first step toward the Z 4 \mathbb{Z}_4 Z 4 version of the Gaussian-mass problem is to make explicit the passage from the symmetrized weight enumerator of a Z 4 \mathbb{Z}_4 Z 4 -code to the theta series of its Construction A lattice. The identity below is the z = 0 z=0 z = 0 case of the complete-weight-enumerator/theta-series correspondence of Choie and Solé [5 Theorem 4.1] . We state this standard identity without proof.
Let C ⊆ Z 4 n C \subseteq\mathbb{Z}_4^n C ⊆ Z 4 n be a linear code. For a codeword c = ( c 1 , … , c n ) ∈ C c=(c_1,\ldots,c_n)\in C c = ( c 1 , … , c n ) ∈ C , let n j ( c ) n_j(c) n j ( c ) denote the number of coordinates of c c c equal to j ∈ { 0 , 1 , 2 , 3 } j\in\{0,1,2,3\} j ∈ { 0 , 1 , 2 , 3 } . Its symmetrized weight enumerator is
swe C ( X , Y , Z ) = ∑ c ∈ C X n 0 ( c ) Y n 1 ( c ) + n 3 ( c ) Z n 2 ( c ) . \operatorname{swe}_C(X,Y,Z)=\sum_{c\in C}X^{n_0(c)}Y^{n_1(c)+n_3(c)}Z^{n_2(c)}. swe C ( X , Y , Z ) = c ∈ C ∑ X n 0 ( c ) Y n 1 ( c ) + n 3 ( c ) Z n 2 ( c ) .
We use the Jacobi theta functions in the form
ϑ 3 ( q ) = ∑ m ∈ Z q m 2 , ϑ 2 ( q ) = ∑ m ∈ Z q ( m + 1 / 2 ) 2 . \vartheta_3(q)=\sum_{m\in\mathbb{Z}}q^{m^2},\qquad\vartheta_2(q)=\sum_{m\in\mathbb{Z}}q^{(m+1/2)^2}. ϑ 3 ( q ) = m ∈ Z ∑ q m 2 , ϑ 2 ( q ) = m ∈ Z ∑ q ( m + 1/2 ) 2 .
The Construction A 4 A_4 A 4 lattice associated with C C C is
A 4 ( C ) = 1 2 { x ∈ Z n : x m o d 4 ∈ C } , A_4(C)=\frac{1}{2}\{x\in\mathbb{Z}^n:x\bmod4\in C\}, A 4 ( C ) = 2 1 { x ∈ Z n : x mod 4 ∈ C } ,
and its theta series is
Θ A 4 ( C ) ( q ) = ∑ x ∈ A 4 ( C ) q ( x , x ) . \Theta_{A_4(C)}(q)=\sum_{x\in A_4(C)}q^{(x,x)}. Θ A 4 ( C ) ( q ) = x ∈ A 4 ( C ) ∑ q ( x , x ) .
Lemma 1 For every linear code C ⊆ Z 4 n C\subseteq\mathbb{Z}_4^n C ⊆ Z 4 n ,
Θ A 4 ( C ) ( q ) = swe C ( ϑ 3 ( q 4 ) , ϑ 2 ( q ) 2 , ϑ 2 ( q 4 ) ) . \Theta_{A_4(C)}(q)=\operatorname{swe}_C\left(\vartheta_3(q^4),\frac{\vartheta_2(q)}{2},\vartheta_2(q^4)\right). Θ A 4 ( C ) ( q ) = swe C ( ϑ 3 ( q 4 ) , 2 ϑ 2 ( q ) , ϑ 2 ( q 4 ) ) .
Reduction to one real variable
To express the theta-series identity in Lemma 1 using a single real parameter, recall
ϑ 4 ( q ) = ∑ m ∈ Z ( − 1 ) m q m 2 . \vartheta_4(q)=\sum_{m\in\mathbb{Z}}(-1)^m q^{m^2}. ϑ 4 ( q ) = m ∈ Z ∑ ( − 1 ) m q m 2 .
Set r = r ( q ) : = ϑ 4 ( q ) / ϑ 3 ( q ) r=r(q):=\vartheta_4(q)/\vartheta_3(q) r = r ( q ) := ϑ 4 ( q ) / ϑ 3 ( q ) . For 0 < q < 1 0<q<1 0 < q < 1 we have 0 < r < 1 0<r<1 0 < r < 1 . Separating even and odd indices in the theta series yields
ϑ 3 ( q ) = ϑ 3 ( q 4 ) + ϑ 2 ( q 4 ) , ϑ 4 ( q ) = ϑ 3 ( q 4 ) − ϑ 2 ( q 4 ) . \vartheta_3(q)=\vartheta_3(q^4)+\vartheta_2(q^4),\qquad\vartheta_4(q)=\vartheta_3(q^4)-\vartheta_2(q^4). ϑ 3 ( q ) = ϑ 3 ( q 4 ) + ϑ 2 ( q 4 ) , ϑ 4 ( q ) = ϑ 3 ( q 4 ) − ϑ 2 ( q 4 ) .
Together with Jacobi’s quartic identity ϑ 3 ( q ) 4 = ϑ 2 ( q ) 4 + ϑ 4 ( q ) 4 \vartheta_3(q)^4=\vartheta_2(q)^4+\vartheta_4(q)^4 ϑ 3 ( q ) 4 = ϑ 2 ( q ) 4 + ϑ 4 ( q ) 4 , these give
ϑ 3 ( q 4 ) ϑ 3 ( q ) = 1 + r 2 , ϑ 2 ( q 4 ) ϑ 3 ( q ) = 1 − r 2 , ϑ 2 ( q ) 2 ϑ 3 ( q ) = ( 1 − r 4 ) 1 / 4 2 . \frac{\vartheta_3(q^4)}{\vartheta_3(q)}=\frac{1+r}{2},\qquad\frac{\vartheta_2(q^4)}{\vartheta_3(q)}=\frac{1-r}{2},\qquad\frac{\vartheta_2(q)}{2\vartheta_3(q)}=\frac{(1-r^4)^{1/4}}{2}. ϑ 3 ( q ) ϑ 3 ( q 4 ) = 2 1 + r , ϑ 3 ( q ) ϑ 2 ( q 4 ) = 2 1 − r , 2 ϑ 3 ( q ) ϑ 2 ( q ) = 2 ( 1 − r 4 ) 1/4 .
These identities give the following reduction of the Gaussian-mass inequality to a single real variable.
Lemma 2. Let C ⊆ Z 4 n C \subseteq\mathbb{Z}_4^n C ⊆ Z 4 n be self-dual and define
H C ( r ) : = swe C ( 1 + r , ( 1 − r 4 ) 1 / 4 , 1 − r ) . H_C(r) := \operatorname{swe}_C\left(1+r,(1-r^4)^{1/4},1-r\right). H C ( r ) := swe C ( 1 + r , ( 1 − r 4 ) 1/4 , 1 − r ) . For r = r ( q ) = ϑ 4 ( q ) / ϑ 3 ( q ) r=r(q)=\vartheta_4(q)/\vartheta_3(q) r = r ( q ) = ϑ 4 ( q ) / ϑ 3 ( q ) ,
Θ A 4 ( C ) ( q ) Θ Z n ( q ) = H C ( r ) 2 n . (1) \frac{\Theta_{A_4(C)}(q)}{\Theta_{\mathbb{Z}^n}(q)}=\frac{H_C(r)}{2^n}. \tag*{(1)} Θ Z n ( q ) Θ A 4 ( C ) ( q ) = 2 n H C ( r ) . (1) Moreover, every c ∈ C c\in C c ∈ C satisfies
n 1 ( c ) + n 3 ( c ) ≡ 0 ( m o d 4 ) . (2) n_1(c)+n_3(c)\equiv0\pmod{4}. \tag*{(2)} n 1 ( c ) + n 3 ( c ) ≡ 0 ( mod 4 ) . (2) Hence, with j ( c ) : = n 1 ( c ) + n 3 ( c ) 4 j(c):=\frac{n_1(c)+n_3(c)}{4} j ( c ) := 4 n 1 ( c ) + n 3 ( c ) , we have j ( c ) ∈ Z ≥ 0 j(c)\in\mathbb{Z}_{\geq0} j ( c ) ∈ Z ≥ 0 and
H C ( r ) = ∑ c ∈ C ( 1 + r ) n 0 ( c ) ( 1 − r 4 ) j ( c ) ( 1 − r ) n 2 ( c ) . (3) H_C(r)=\sum_{c\in C}(1+r)^{n_0(c)}(1-r^4)^{j(c)}(1-r)^{n_2(c)}. \tag*{(3)} H C ( r ) = c ∈ C ∑ ( 1 + r ) n 0 ( c ) ( 1 − r 4 ) j ( c ) ( 1 − r ) n 2 ( c ) . (3) Thus H C ( r ) H_C(r) H C ( r ) is an ordinary polynomial, with H C ( 0 ) = H C ( 1 ) = 2 n H_C(0)=H_C(1)=2^n H C ( 0 ) = H C ( 1 ) = 2 n , and proving the Gaussian mass inequality for A 4 ( C ) A_4(C) A 4 ( C ) is equivalent to proving
H C ( r ) ≤ 2 n , 0 < r < 1. (4) H_C(r)\leq2^n,\quad0<r<1. \tag*{(4)} H C ( r ) ≤ 2 n , 0 < r < 1. (4)
Proof. By Lemma 1,
Θ A 4 ( C ) ( q ) = swe C ( ϑ 3 ( q 4 ) , ϑ 2 ( q ) 2 , ϑ 2 ( q 4 ) ) . \Theta_{A_4(C)}(q)=\operatorname{swe}_C\left(\vartheta_3(q^4),\frac{\vartheta_2(q)}{2},\vartheta_2(q^4)\right). Θ A 4 ( C ) ( q ) = swe C ( ϑ 3 ( q 4 ) , 2 ϑ 2 ( q ) , ϑ 2 ( q 4 ) ) .
while Θ Z n ( q ) = ϑ 3 ( q ) n \Theta_{\mathbb{Z}^n}(q)=\vartheta_3(q)^n Θ Z n ( q ) = ϑ 3 ( q ) n . Since swe C \operatorname{swe}_C swe C is homogeneous of degree n n n ,
Θ A 4 ( C ) ( q ) ϑ 3 ( q ) n = swe C ( ϑ 3 ( q 4 ) ϑ 3 ( q ) , ϑ 2 ( q ) 2 ϑ 3 ( q ) , ϑ 2 ( q 4 ) ϑ 3 ( q ) ) . \frac{\Theta_{A_4(C)}(q)}{\vartheta_3(q)^n}
=\operatorname{swe}_C\left(\frac{\vartheta_3(q^4)}{\vartheta_3(q)},\frac{\vartheta_2(q)}{2\vartheta_3(q)},\frac{\vartheta_2(q^4)}{\vartheta_3(q)}\right). ϑ 3 ( q ) n Θ A 4 ( C ) ( q ) = swe C ( ϑ 3 ( q ) ϑ 3 ( q 4 ) , 2 ϑ 3 ( q ) ϑ 2 ( q ) , ϑ 3 ( q ) ϑ 2 ( q 4 ) ) .
Using the preceding theta identities and homogeneity once more gives
Θ A 4 ( C ) ( q ) Θ Z n ( q ) = 1 2 n swe C ( 1 + r , ( 1 − r 4 ) 1 / 4 , 1 − r ) = H C ( r ) 2 n , \frac{\Theta_{A_4(C)}(q)}{\Theta_{\mathbb{Z}^n}(q)}
=\frac{1}{2^n}\operatorname{swe}_C\left(1+r,(1-r^4)^{1/4},1-r\right)
=\frac{H_C(r)}{2^n}, Θ Z n ( q ) Θ A 4 ( C ) ( q ) = 2 n 1 swe C ( 1 + r , ( 1 − r 4 ) 1/4 , 1 − r ) = 2 n H C ( r ) ,
which proves (3.1).
Since C C C is self-dual, it is self-orthogonal. Thus c ⋅ c ≡ 0 ( m o d 4 ) c\cdot c\equiv0\pmod{4} c ⋅ c ≡ 0 ( mod 4 ) for every c ∈ C c\in C c ∈ C . As
0 2 ≡ 0 , 1 2 ≡ 1 , 2 2 ≡ 0 , 3 2 ≡ 1 ( m o d 4 ) , (5) 0^2\equiv0,\quad1^2\equiv1,\quad2^2\equiv0,\quad3^2\equiv1\pmod{4}, \tag*{(5)} 0 2 ≡ 0 , 1 2 ≡ 1 , 2 2 ≡ 0 , 3 2 ≡ 1 ( mod 4 ) , (5)
we have c ⋅ c ≡ n 1 ( c ) + n 3 ( c ) ( m o d 4 ) c \cdot c \equiv n_1(c) + n_3(c) \pmod{4} c ⋅ c ≡ n 1 ( c ) + n 3 ( c ) ( mod 4 ) , which proves (2) . Therefore n 1 ( c ) + n 3 ( c ) = 4 j ( c ) n_1(c) + n_3(c) = 4j(c) n 1 ( c ) + n 3 ( c ) = 4 j ( c ) , and the contribution of c c c to H C ( r ) H_C(r) H C ( r ) becomes
( 1 + r ) n 0 ( c ) ( 1 − r 4 ) j ( c ) ( 1 − r ) n 2 ( c ) . (1+r)^{n_0(c)}(1-r^4)^{j(c)}(1-r)^{n_2(c)}. ( 1 + r ) n 0 ( c ) ( 1 − r 4 ) j ( c ) ( 1 − r ) n 2 ( c ) .
Summing over c ∈ C c \in C c ∈ C gives (3) , so H C ( r ) H_C(r) H C ( r ) is a polynomial. At r = 0 r=0 r = 0 ,
H C ( 0 ) = swe C ( 1 , 1 , 1 ) = ∣ C ∣ = 2 n , H_C(0)=\operatorname{swe}_C(1,1,1)=|C|=2^n, H C ( 0 ) = swe C ( 1 , 1 , 1 ) = ∣ C ∣ = 2 n ,
because a self-dual Z 4 \mathbb{Z}_4 Z 4 -code satisfies ∣ C ∣ 2 = 4 n |C|^2=4^n ∣ C ∣ 2 = 4 n . At r = 1 r=1 r = 1 ,
H C ( 1 ) = swe C ( 2 , 0 , 0 ) = 2 n , H_C(1)=\operatorname{swe}_C(2,0,0)=2^n, H C ( 1 ) = swe C ( 2 , 0 , 0 ) = 2 n ,
since only the zero codeword contributes. Finally, (1) shows that Θ A 4 ( C ) ( q ) ≤ Θ Z n ( q ) \Theta_{A_4(C)}(q) \le\Theta_{\mathbb{Z}^n}(q) Θ A 4 ( C ) ( q ) ≤ Θ Z n ( q ) is equivalent to H C ( r ) ≤ 2 n H_C(r) \le2^n H C ( r ) ≤ 2 n . Since r ( q ) r(q) r ( q ) runs through ( 0 , 1 ) (0,1) ( 0 , 1 ) as q q q runs through ( 0 , 1 ) (0,1) ( 0 , 1 ) , this gives (4) .
Averaging bound for constructions A4 and A6
Let C C C be a free self-dual Z 4 \mathbb{Z}_4 Z 4 -code of length n = 2 k n=2k n = 2 k . After a permutation of coordinates, write its generator matrix G G G in systematic form as
G = [ I k ∣ A ] , A A T ≡ − I k ( m o d 4 ) . G=[I_k \mid A], \quad AA^T \equiv-I_k \pmod{4}. G = [ I k ∣ A ] , A A T ≡ − I k ( mod 4 ) .
For 0 < r < 1 0<r<1 0 < r < 1 , define
f r ( 0 ) = 1 + r , f r ( 1 ) = f r ( 3 ) = ( 1 − r 4 ) 1 / 4 , f r ( 2 ) = 1 − r . f_r(0)=1+r,\quad f_r(1)=f_r(3)=(1-r^4)^{1/4},\quad f_r(2)=1-r. f r ( 0 ) = 1 + r , f r ( 1 ) = f r ( 3 ) = ( 1 − r 4 ) 1/4 , f r ( 2 ) = 1 − r .
and, for u = ( u 1 , … , u k ) ∈ Z 4 k u=(u_1,\ldots,u_k)\in\mathbb{Z}_4^k u = ( u 1 , … , u k ) ∈ Z 4 k ,
F r ( u ) = ∏ i = 1 k f r ( u i ) . F_r(u)=\prod_{i=1}^{k} f_r(u_i). F r ( u ) = i = 1 ∏ k f r ( u i ) .
Then
H A ( r ) = ∑ a ∈ Z 4 k F r ( a ) F r ( a A ) . H_A(r)=\sum_{a\in\mathbb{Z}_4^k} F_r(a)F_r(aA). H A ( r ) = a ∈ Z 4 k ∑ F r ( a ) F r ( a A ) .
By Lemma 2, the Gaussian-mass inequality for C C C is equivalent to H A ( r ) ≤ 4 k H_A(r) \le4^k H A ( r ) ≤ 4 k , since n = 2 k n=2k n = 2 k . Since A A A is invertible, the Kominers-type identity is
H A ( r ) = Λ ( r ) k − 1 2 ∑ a ∈ Z 4 k ( F r ( a ) − F r ( a A ) ) 2 H_A(r)=\Lambda(r)^k-\frac{1}{2}\sum_{a\in\mathbb{Z}_4^k}\left(F_r(a)-F_r(aA)\right)^2 H A ( r ) = Λ ( r ) k − 2 1 a ∈ Z 4 k ∑ ( F r ( a ) − F r ( a A ) ) 2
where
Λ ( r ) = 2 ( 1 + r 2 + 1 − r 4 ) . \Lambda(r)=2\left(1+r^2+\sqrt{1-r^4}\right). Λ ( r ) = 2 ( 1 + r 2 + 1 − r 4 ) .
Thus the desired inequality H A ( r ) ≤ 4 k H_A(r)\leq4^k H A ( r ) ≤ 4 k is equivalent to
1 2 ∑ a ∈ Z 4 k ( F r ( a ) − F r ( a A ) ) 2 ≥ Λ ( r ) k − 4 k . (6) \frac{1}{2}\sum_{a\in\mathbb{Z}_4^k}\left(F_r(a)-F_r(aA)\right)^2\geq\Lambda(r)^k-4^k. \tag*{(6)} 2 1 a ∈ Z 4 k ∑ ( F r ( a ) − F r ( a A ) ) 2 ≥ Λ ( r ) k − 4 k . (6)
The question considered here is whether the left-hand side of (6) can be controlled after averaging over the matrices A A A satisfying A A T ≡ − I k ( m o d 4 ) AA^T\equiv-I_k\pmod{4} A A T ≡ − I k ( mod 4 ) . Put B = A m o d 2 B=A\bmod2 B = A mod 2 . Then B B T = I k BB^T=I_k B B T = I k over F 2 \mathbb{F}_2 F 2 . The associated binary code
D B = { ( x , x B ) : x ∈ F 2 k } D_B=\{(x,xB):x\in\mathbb{F}_2^k\} D B = {( x , x B ) : x ∈ F 2 k }
is self-dual. Moreover it is doubly even: if ( x , x B ) (x,xB) ( x , x B ) is the residue of a codeword of C C C , self-orthogonality over Z 4 \mathbb{Z}_4 Z 4 implies that the number of odd coordinates is divisible by 4.
Fix one lift A 0 A_0 A 0 of B B B satisfying A 0 A 0 T ≡ − I k ( m o d 4 ) A_0A_0^T\equiv-I_k\pmod{4} A 0 A 0 T ≡ − I k ( mod 4 ) . Every other lift with the same residue matrix can be written uniquely as
A H = A 0 + 2 H B , (7) A_H=A_0+2HB, \tag*{(7)} A H = A 0 + 2 H B , (7)
where H H H runs through the symmetric matrices in M k ( F 2 ) M_k(\mathbb{F}_2) M k ( F 2 ) .
Indeed, write A = A 0 + 2 E A=A_0+2E A = A 0 + 2 E . Comparing the two congruences A A T ≡ A 0 A 0 T ≡ − I k ( m o d 4 ) AA^T\equiv A_0A_0^T\equiv-I_k\pmod{4} A A T ≡ A 0 A 0 T ≡ − I k ( mod 4 ) gives E B T + B E T = 0 EB^T+BE^T=0 E B T + B E T = 0 over F 2 \mathbb{F}_2 F 2 . If H = E B T H=EB^T H = E B T , then H + H T = 0 H+H^T=0 H + H T = 0 , which in characteristic 2 means H = H T H=H^T H = H T . Since B B T = I k BB^T=I_k B B T = I k , one has E = H B E=HB E = H B , giving (4.2).
Thus it is enough first to average over the symmetric binary matrix H H H , while keeping B B B fixed.
Exact average over all lifts of a fixed B B B
Set α = 1 + r \alpha=1+r α = 1 + r , γ = 1 − r \gamma=1-r γ = 1 − r , and η = ( 1 − r 4 ) 1 / 4 \eta=(1-r^4)^{1/4} η = ( 1 − r 4 ) 1/4 . Write every a ∈ Z 4 k a\in\mathbb{Z}_4^k a ∈ Z 4 k uniquely as
a = x + 2 v , x , v ∈ F 2 k . a=x+2v,\quad x,v\in\mathbb{F}_2^k. a = x + 2 v , x , v ∈ F 2 k .
Assume first that x ≠ 0 x \ne0 x = 0 . From (7) ,
a A H ≡ a A 0 + 2 x H B ( m o d 4 ) . (8) aA_H \equiv aA_0 + 2xHB \pmod{4}. \tag*{(8)} a A H ≡ a A 0 + 2 x H B ( mod 4 ) . (8)
As H H H runs uniformly over all symmetric binary matrices, the vector x H xH x H runs uniformly over F 2 k \mathbb{F}_2^k F 2 k . Hence x H B xHB x H B is also uniform, because B B B is invertible. Therefore a A H aA_H a A H is uniformly distributed among all vectors in Z 4 k \mathbb{Z}_4^k Z 4 k whose residue modulo 2 is x B xB x B .
For a coordinate with binary residue 0, the two possible Z 4 \mathbb{Z}_4 Z 4 values are 0 and 2, and their average contribution to F r F_r F r is α + γ 2 = 1 \frac{\alpha+\gamma}{2}=1 2 α + γ = 1 . For a coordinate with binary residue 1, the two possible values are 1 and 3, and both contribute η \eta η . Consequently,
E H F r ( a A H ) = η wt ( x B ) , x ≠ 0. (9) \mathbb{E}_H F_r(aA_H)=\eta^{\operatorname{wt}(xB)},\qquad x\ne0. \tag*{(9)} E H F r ( a A H ) = η wt ( x B ) , x = 0. (9)
On the other hand,
∑ v ∈ F 2 k F r ( x + 2 v ) = 2 k η wt ( x ) . (10) \sum_{v\in\mathbb{F}_2^k}F_r(x+2v)=2^k\eta^{\operatorname{wt}(x)}. \tag*{(10)} v ∈ F 2 k ∑ F r ( x + 2 v ) = 2 k η wt ( x ) . (10)
Combining (9) and (10) , the total contribution from x ≠ 0 x\ne0 x = 0 is
2 k ∑ x ≠ 0 η wt ( x ) + wt ( x B ) = 2 k ( W D B ( 1 , η ) − 1 ) , (11) 2^k\sum_{x\ne0}\eta^{\operatorname{wt}(x)+\operatorname{wt}(xB)}
=2^k\left(W_{D_B}(1,\eta)-1\right), \tag*{(11)} 2 k x = 0 ∑ η wt ( x ) + wt ( x B ) = 2 k ( W D B ( 1 , η ) − 1 ) , (11)
where W D B W_{D_B} W D B denotes the binary weight enumerator of D B D_B D B . Now consider x = 0 x=0 x = 0 . Then a = 2 v a=2v a = 2 v and 2 v A H ≡ 2 v B ( m o d 4 ) 2vA_H\equiv2vB\pmod{4} 2 v A H ≡ 2 v B ( mod 4 ) , so this part is independent of the lift. Its contribution is
∑ v ∈ F 2 k F r ( 2 v ) F r ( 2 v B ) = W D B ( α , γ ) . (12) \sum_{v\in\mathbb{F}_2^k}F_r(2v)F_r(2vB)=W_{D_B}(\alpha,\gamma). \tag*{(12)} v ∈ F 2 k ∑ F r ( 2 v ) F r ( 2 v B ) = W D B ( α , γ ) . (12)
Since D B D_B D B is binary self-dual, the MacWilliams identity gives
W D B ( α , γ ) = 2 − k W D B ( α + γ , α − γ ) = 2 k W D B ( 1 , r ) . (13) W_{D_B}(\alpha,\gamma)=2^{-k}W_{D_B}(\alpha+\gamma,\alpha-\gamma)=2^kW_{D_B}(1,r). \tag*{(13)} W D B ( α , γ ) = 2 − k W D B ( α + γ , α − γ ) = 2 k W D B ( 1 , r ) . (13)
Equations (11) –(13) yield the exact average formula
E A ∣ B H A ( r ) = 2 k ( W D B ( 1 , r ) + W D B ( 1 , η ) − 1 ) . (14) \mathbb{E}_{A\mid B}H_A(r)=2^k\left(W_{D_B}(1,r)+W_{D_B}(1,\eta)-1\right). \tag*{(14)} E A ∣ B H A ( r ) = 2 k ( W D B ( 1 , r ) + W D B ( 1 , η ) − 1 ) . (14)
Convexity argument
Because D B D_B D B is doubly even, there is a polynomial with nonnegative coefficients
P B ( t ) = ∑ j ≥ 0 A 4 j t j P_B(t) = \sum_{j\geq0} A_{4j}t^j P B ( t ) = j ≥ 0 ∑ A 4 j t j
such that W D B ( 1 , z ) = P B ( z 4 ) W_{D_B}(1,z)=P_B(z^4) W D B ( 1 , z ) = P B ( z 4 ) . Put t = r 4 t=r^4 t = r 4 . Since η 4 = 1 − r 4 = 1 − t \eta^4=1-r^4=1-t η 4 = 1 − r 4 = 1 − t , formula (4.9) becomes
E A ∣ B H A ( r ) = 2 k ( P B ( t ) + P B ( 1 − t ) − 1 ) . (15) \mathbb{E}_{A|B}H_A(r)=2^k\left(P_B(t)+P_B(1-t)-1\right). \tag*{(15)} E A ∣ B H A ( r ) = 2 k ( P B ( t ) + P B ( 1 − t ) − 1 ) . (15)
The coefficients of P B P_B P B are nonnegative, so P B P_B P B is convex on [ 0 , 1 ] [0,1] [ 0 , 1 ] . Hence Q B ( t ) = P B ( t ) + P B ( 1 − t ) Q_B(t)=P_B(t)+P_B(1-t) Q B ( t ) = P B ( t ) + P B ( 1 − t ) is also convex. It is symmetric about t = 1 / 2 t=1/2 t = 1/2 , and its endpoint values are equal. Therefore its maximum on [ 0 , 1 ] [0,1] [ 0 , 1 ] is attained at the endpoints. Since P B ( 0 ) = 1 P_B(0)=1 P B ( 0 ) = 1 and P B ( 1 ) = ∣ D B ∣ = 2 k P_B(1)=|D_B|=2^k P B ( 1 ) = ∣ D B ∣ = 2 k , we obtain
P B ( t ) + P B ( 1 − t ) ≤ 1 + 2 k . (16) P_B(t)+P_B(1-t)\leq1+2^k. \tag*{(16)} P B ( t ) + P B ( 1 − t ) ≤ 1 + 2 k . (16)
Substituting (4.11) into (4.10) gives
E A ∣ B H A ( r ) ≤ 4 k . (17) \mathbb{E}_{A|B}H_A(r)\leq4^k. \tag*{(17)} E A ∣ B H A ( r ) ≤ 4 k . (17)
Equivalently,
E A ∣ B [ 1 2 ∑ a ∈ Z 4 k ( F r ( a ) − F r ( a A ) ) 2 ] ≥ Λ ( r ) k − 4 k . (18) \mathbb{E}_{A|B}\left[\frac{1}{2}\sum_{a\in\mathbb{Z}_4^k}\left(F_r(a)-F_r(aA)\right)^2\right]\geq\Lambda(r)^k-4^k. \tag*{(18)} E A ∣ B 2 1 a ∈ Z 4 k ∑ ( F r ( a ) − F r ( a A ) ) 2 ≥ Λ ( r ) k − 4 k . (18)
Thus the desired lower bound in (4.1) is valid after averaging over all Z 4 \mathbb{Z}_4 Z 4 lifts of every fixed binary residue matrix B B B .
The amount by which the average exceeds the required lower bound can also be written explicitly. From (4.10),
4 k − E A ∣ B H A ( r ) = 2 k ( 1 + 2 k − P B ( t ) − P B ( 1 − t ) ) . 4^k-\mathbb{E}_{A|B}H_A(r)=2^k\left(1+2^k-P_B(t)-P_B(1-t)\right). 4 k − E A ∣ B H A ( r ) = 2 k ( 1 + 2 k − P B ( t ) − P B ( 1 − t ) ) .
Using P B ( t ) = ∑ j A 4 j t j P_B(t)=\sum_j A_{4j}t^j P B ( t ) = ∑ j A 4 j t j gives
4 k − E A ∣ B H A ( r ) = 2 k ∑ j ≥ 2 A 4 j ( 1 − t j − ( 1 − t ) j ) . (19) 4^k-\mathbb{E}_{A|B}H_A(r)=2^k\sum_{j\geq2}A_{4j}\left(1-t^j-(1-t)^j\right). \tag*{(19)} 4 k − E A ∣ B H A ( r ) = 2 k j ≥ 2 ∑ A 4 j ( 1 − t j − ( 1 − t ) j ) . (19)
and every term on the right-hand side is nonnegative for $0\leq t\leq 1.
Example 1 For k = 4 k = 4 k = 4 , the binary doubly-even self-dual code has weight enumerator
W ( 1 , z ) = 1 + 14 z 4 + z 8 , so P ( t ) = 1 + 14 t + t 2 . W(1,z) = 1 + 14z^{4} + z^{8}, \quad\text{so } P(t) = 1 + 14t + t^{2}. W ( 1 , z ) = 1 + 14 z 4 + z 8 , so P ( t ) = 1 + 14 t + t 2 .
Formula (4.10) gives
E A ∣ B H A ( r ) = 256 − 32 t ( 1 − t ) , t = r 4 . \mathbb{E}_{A|B}H_A(r) = 256 - 32t(1-t), \quad t = r^{4}. E A ∣ B H A ( r ) = 256 − 32 t ( 1 − t ) , t = r 4 .
Thus
E A ∣ B [ 1 2 ∑ a ( F r ( a ) − F r ( a A ) ) 2 ] = Λ ( r ) 4 − 256 + 32 r 4 ( 1 − r 4 ) , \mathbb{E}_{A|B}\left[\frac{1}{2}\sum_a\left(F_r(a)-F_r(aA)\right)^2\right] = \Lambda(r)^4 - 256 + 32r^4(1-r^4), E A ∣ B [ 2 1 a ∑ ( F r ( a ) − F r ( a A ) ) 2 ] = Λ ( r ) 4 − 256 + 32 r 4 ( 1 − r 4 ) ,
which is strictly stronger than the required averaged inequality for 0 < r < 1 0 < r < 1 0 < r < 1 .
The exact fixed-binary average for Construction A 6 A_6 A 6
We now turn to Construction A 6 A_6 A 6 . In contrast with the Z 4 \mathbb{Z}_4 Z 4 case, where one fixes the binary residue code and averages over its self-dual lifts, the Chinese remainder decomposition Z 6 ≅ F 2 × F 3 \mathbb{Z}_6 \cong\mathbb{F}_2 \times\mathbb{F}_3 Z 6 ≅ F 2 × F 3 leads to a different natural averaging ensemble.
Let n = 2 k = 4 s > 0 n = 2k = 4s > 0 n = 2 k = 4 s > 0 and fix a binary self-dual code D ⊆ F 2 n D \subseteq\mathbb{F}_2^n D ⊆ F 2 n . Let
S 3 ( n ) = { T ⊆ F 3 n : T = T ⊥ } , \mathcal{S}_3(n) = \{T \subseteq\mathbb{F}_3^n : T = T^\perp\}, S 3 ( n ) = { T ⊆ F 3 n : T = T ⊥ } ,
equipped with the uniform distribution on its actual subspaces. For T ∈ S 3 ( n ) T \in\mathcal{S}_3(n) T ∈ S 3 ( n ) , let
C ( D , T ) = { c ∈ Z 6 n : c m o d 2 ∈ D , c m o d 3 ∈ T } , C(D,T) = \{c \in\mathbb{Z}_6^n : c \bmod2 \in D,\ c \bmod3 \in T\}, C ( D , T ) = { c ∈ Z 6 n : c mod 2 ∈ D , c mod 3 ∈ T } ,
and put K = A 2 ( D ) K = A_2(D) K = A 2 ( D ) . Thus the binary component is fixed, while the ternary self-dual component is averaged.
The basic incidence probability for the ternary ensemble is the following.
Lemma 3 Let c ∈ F 3 2 k c \in\mathbb{F}_3^{2k} c ∈ F 3 2 k and let T ∈ S 3 ( 2 k ) T \in\mathcal{S}_3(2k) T ∈ S 3 ( 2 k ) be chosen uniformly. Then
P ( c ∈ T ) = { 1 , c = 0 , 1 3 k − 1 + 1 , c ≠ 0 and c ⋅ c = 0 , 0 , c ⋅ c ≠ 0. (20) \mathbb{P}(c \in T) =
\begin{cases}
1, & c = 0, \\
\frac{1}{3^{k-1}+1}, & c \ne0\ \text{and}\ c \cdot c = 0, \\
0, & c \cdot c \ne0.
\end{cases} \tag*{(20)} P ( c ∈ T ) = ⎩ ⎨ ⎧ 1 , 3 k − 1 + 1 1 , 0 , c = 0 , c = 0 and c ⋅ c = 0 , c ⋅ c = 0. (20)
Proof. We refer to chapter 1 of [12 ] for working knowledge of quadratic forms over finite fields. Since n = 2 k n=2k n = 2 k is divisible by 4, the standard quadratic space F 3 2 k \mathbb{F}_3^{2k} F 3 2 k is split. In a hyperbolic basis, its isotropic vectors may be represented by pairs ( x , y ) ∈ F 3 k × F 3 k (x,y)\in\mathbb{F}_3^k\times\mathbb{F}_3^k ( x , y ) ∈ F 3 k × F 3 k satisfying x ⋅ y = 0 x\cdot y=0 x ⋅ y = 0 . For x = 0 x=0 x = 0 there are 3 k 3^k 3 k choices for y y y , whereas for each nonzero x x x there are 3 k − 1 3^{k-1} 3 k − 1 choices. Hence the number of nonzero isotropic vectors is
3 k + ( 3 k − 1 ) 3 k − 1 − 1 = ( 3 k − 1 ) ( 3 k − 1 + 1 ) . 3^k+(3^k-1)3^{k-1}-1=(3^k-1)(3^{k-1}+1). 3 k + ( 3 k − 1 ) 3 k − 1 − 1 = ( 3 k − 1 ) ( 3 k − 1 + 1 ) .
The orthogonal group acts transitively on the nonzero isotropic vectors and permutes the members of S 3 ( 2 k ) S_3(2k) S 3 ( 2 k ) . Hence the probability that a fixed nonzero isotropic vector belongs to a uniformly chosen T T T is constant. Every T ∈ S 3 ( 2 k ) T\in S_3(2k) T ∈ S 3 ( 2 k ) contains 3 k − 1 3^k-1 3 k − 1 nonzero vectors. Double counting incidences therefore gives
P ( c ∈ T ) = 3 k − 1 ( 3 k − 1 ) ( 3 k − 1 + 1 ) = 1 3 k − 1 + 1 . \mathbb{P}(c\in T)=\frac{3^k-1}{(3^k-1)(3^{k-1}+1)}=\frac{1}{3^{k-1}+1}. P ( c ∈ T ) = ( 3 k − 1 ) ( 3 k − 1 + 1 ) 3 k − 1 = 3 k − 1 + 1 1 .
A nonisotropic vector cannot belong to a self-orthogonal subspace. ■
Define the mod-3 norm filter
P 3 ( K ; t ) = ∑ v ∈ K ∥ v ∥ 2 ≡ 0 ( m o d 3 ) e − π t ∥ v ∥ 2 / 3 . (21) P_3(K;t)=\sum_{\substack{v\in K\\ \lVert v\rVert^2\equiv0\pmod3}}e^{-\pi t\lVert v\rVert^2/3}. \tag*{(21)} P 3 ( K ; t ) = v ∈ K ∥ v ∥ 2 ≡ 0 ( mod 3 ) ∑ e − π t ∥ v ∥ 2 /3 . (21)
Since K K K is integral, the congruence condition in (21) is well-defined.
Theorem 1 For every t > 0 t>0 t > 0 ,
E T ∈ S 3 ( n ) Θ A 6 ( C ( D , T ) ) ( t ) = 3 k − 1 Θ K ( 3 t ) + P 3 ( K ; t ) 3 k − 1 + 1 . (22) \mathbb{E}_{T\in S_3(n)}\Theta_{A_6(C(D,T))}(t)=\frac{3^{k-1}\Theta_K(3t)+P_3(K;t)}{3^{k-1}+1}. \tag*{(22)} E T ∈ S 3 ( n ) Θ A 6 ( C ( D , T )) ( t ) = 3 k − 1 + 1 3 k − 1 Θ K ( 3 t ) + P 3 ( K ; t ) . (22)
Proof. Write v = u 2 ∈ K v=\frac{u}{\sqrt{2}}\in K v = 2 u ∈ K , where u ∈ Z n u\in\mathbb{Z}^n u ∈ Z n , u m o d 2 ∈ D u\bmod2\in D u mod 2 ∈ D . The corresponding potential vector in the modulus-6 Construction A lattice is u 6 = v 3 \frac{u}{\sqrt{6}}=\frac{v}{\sqrt{3}} 6 u = 3 v , and hence its Gaussian contribution is e − π t ∥ v ∥ 2 / 3 e^{-\pi t\lVert v\rVert^2/3} e − π t ∥ v ∥ 2 /3 . The ternary residue u m o d 3 u\bmod3 u mod 3 can occur in a ternary self-dual code only when it is isotropic. Since ∥ u ∥ 2 = 2 ∥ v ∥ 2 \lVert u\rVert^2=2\lVert v\rVert^2 ∥ u ∥ 2 = 2 ∥ v ∥ 2 , and 2 is invertible modulo 3, this is equivalent to ∥ v ∥ 2 ≡ 0 ( m o d 3 ) \lVert v\rVert^2\equiv0\pmod3 ∥ v ∥ 2 ≡ 0 ( mod 3 ) . Put ρ = 1 3 k − 1 + 1 \rho=\frac{1}{3^{k-1}+1} ρ = 3 k − 1 + 1 1 . By Lemma 3, every nonzero isotropic ternary residue occurs with probability ρ \rho ρ . The zero ternary residue means u = 3 w u=3w u = 3 w for some w ∈ Z n w\in\mathbb{Z}^n w ∈ Z n . Since 3 ≡ 1 ( m o d 2 ) 3\equiv1\pmod{2} 3 ≡ 1 ( mod 2 ) , the condition u m o d 2 ∈ D u\bmod2\in D u mod 2 ∈ D is equivalent to w m o d 2 ∈ D w\bmod2\in D w mod 2 ∈ D . Hence w 2 ∈ K \frac{w}{\sqrt{2}}\in K 2 w ∈ K . Thus the vectors v v v with zero ternary residue are exactly the vectors of 3 K 3K 3 K . Their contribution is
∑ x ∈ K e − π t ∥ 3 x ∥ 2 / 3 = Θ K ( 3 t ) . \sum_{x\in K}e^{-\pi t\lVert3x\rVert^2/3}=\Theta_K(3t). x ∈ K ∑ e − π t ∥ 3 x ∥ 2 /3 = Θ K ( 3 t ) .
Consequently,
E T Θ A 6 ( C ( D , T ) ) ( t ) = Θ K ( 3 t ) + ρ ( P 3 ( K ; t ) − Θ K ( 3 t ) ) = ( 1 − ρ ) Θ K ( 3 t ) + ρ P 3 ( K ; t ) . \mathbb{E}_T\Theta_{A_6(C(D,T))}(t)=\Theta_K(3t)+\rho\left(P_3(K;t)-\Theta_K(3t)\right)
=(1-\rho)\Theta_K(3t)+\rho P_3(K;t). E T Θ A 6 ( C ( D , T )) ( t ) = Θ K ( 3 t ) + ρ ( P 3 ( K ; t ) − Θ K ( 3 t ) ) = ( 1 − ρ ) Θ K ( 3 t ) + ρ P 3 ( K ; t ) .
Since 1 − ρ = 3 k − 1 3 k − 1 + 1 1-\rho=\frac{3^{k-1}}{3^{k-1}+1} 1 − ρ = 3 k − 1 + 1 3 k − 1 , equation (22) follows.
Thus, after averaging over the ternary self-dual component, all dependence on T T T disappears. The resulting expression depends only on the fixed binary Construction A lattice K = A 2 ( D ) K=A_2(D) K = A 2 ( D ) and its mod-3 norm filter P 3 ( K ; t ) P_3(K;t) P 3 ( K ; t ) . This is the modulus-6 analogue of the fixed-residue averaging formula obtained above for Construction A 4 A_4 A 4 .
The averaged Gaussian bound for Construction A 6 A_6 A 6
We next use the exact formula (4.17) to derive the Gaussian-mass bound. The main input is the Gaussian inequality for binary self-dual Construction A lattices [14 ] :
Θ K ( u ) ≤ ϑ 3 ( u ) n , u > 0. (23) \Theta_K(u)\leq\vartheta_3(u)^n,\qquad u>0. \tag*{(23)} Θ K ( u ) ≤ ϑ 3 ( u ) n , u > 0. (23)
For convenience, write w ( t ) = ϑ 3 ( t ) w(t)=\vartheta_3(t) w ( t ) = ϑ 3 ( t ) , ϕ ( t ) = − w ′ ( t ) w ( t ) \phi(t)=-\frac{w'(t)}{w(t)} ϕ ( t ) = − w ( t ) w ′ ( t ) . We first record the elementary estimates required below.
Lemma 4. One has w ( 3 ) < 1.0002 w(3)<1.0002 w ( 3 ) < 1.0002 , w ( 30 / 11 ) < 1.0004 w(30/11)<1.0004 w ( 30/11 ) < 1.0004 , w ( 11 / 10 ) > 1.062 w(11/10)>1.062 w ( 11/10 ) > 1.062 , and for t ≥ 3 2 t\geq\frac{3}{2} t ≥ 2 3 , we have ϕ ( t ) < 1 / 16 \phi(t)<1/16 ϕ ( t ) < 1/16 . In addition,
w ( 2 / 3 ) 3 < 3 4 . \frac{w(2/3)}{\sqrt{3}}<\frac{3}{4}. 3 w ( 2/3 ) < 4 3 . Moreover,
( 1.062 ) 24 > ( 1.0002 ) 24 + 3 ( 1.0004 ) 24 , (1.062)^{24}>(1.0002)^{24}+3(1.0004)^{24}, ( 1.062 ) 24 > ( 1.0002 ) 24 + 3 ( 1.0004 ) 24 , 2 ( 24 25 ) 24 < 4 5 , and 3 ⋅ 540 48 ( 3 4 ) 24 < 1 20 . 2\left(\frac{24}{25}\right)^{24}<\frac{4}{5},\qquad\text{and}\qquad\frac{3\cdot540}{48}\left(\frac{3}{4}\right)^{24}<\frac{1}{20}. 2 ( 25 24 ) 24 < 5 4 , and 48 3 ⋅ 540 ( 4 3 ) 24 < 20 1 .
Proof. Let q = e − π t q=e^{-\pi t} q = e − π t . Since j 2 ≥ 1 + 3 ( j − 1 ) j^2 \ge1+3(j-1) j 2 ≥ 1 + 3 ( j − 1 ) for every integer j ≥ 1 j \ge1 j ≥ 1 , we have
1 + 2 q ≤ w ( t ) ≤ 1 + 2 q 1 − q 3 . 1+2q \le w(t) \le1+\frac{2q}{1-q^3}. 1 + 2 q ≤ w ( t ) ≤ 1 + 1 − q 3 2 q .
Similarly,
ϕ ( t ) ≤ 2 π q 1 + q 3 ( 1 − q 3 ) 3 . \phi(t) \le2\pi q\frac{1+q^3}{(1-q^3)^3}. ϕ ( t ) ≤ 2 π q ( 1 − q 3 ) 3 1 + q 3 .
Using
e − 3 π < 1 12000 , e − 30 π / 11 < 1 5200 , e − 11 π / 10 > 31 1000 , e^{-3\pi}<\frac{1}{12000},\quad e^{-30\pi/11}<\frac{1}{5200},\quad e^{-11\pi/10}>\frac{31}{1000}, e − 3 π < 12000 1 , e − 30 π /11 < 5200 1 , e − 11 π /10 > 1000 31 ,
gives the first three estimates. For t ≥ 3 / 2 t \ge3/2 t ≥ 3/2 , one has q < 1 110 q<\frac{1}{110} q < 110 1 . Together with 2 π < 44 / 7 2\pi<44/7 2 π < 44/7 , this yields
ϕ ( t ) < 44 770 1 + 110 − 3 ( 1 − 110 − 3 ) 3 < 1 16 . \phi(t)<\frac{44}{770}\frac{1+110^{-3}}{(1-110^{-3})^3}<\frac{1}{16}. ϕ ( t ) < 770 44 ( 1 − 11 0 − 3 ) 3 1 + 11 0 − 3 < 16 1 .
The theta transformation gives
w ( 2 / 3 ) 3 = w ( 3 / 2 ) 2 . \frac{w(2/3)}{\sqrt{3}}=\frac{w(3/2)}{\sqrt{2}}. 3 w ( 2/3 ) = 2 w ( 3/2 ) .
The preceding geometric estimate gives w ( 3 / 2 ) < 1.019 w(3/2)<1.019 w ( 3/2 ) < 1.019 , and therefore
w ( 2 / 3 ) 3 < 1.019 2 < 3 4 . \frac{w(2/3)}{\sqrt{3}}<\frac{1.019}{\sqrt{2}}<\frac{3}{4}. 3 w ( 2/3 ) < 2 1.019 < 4 3 .
The remaining assertions are direct rational comparisons.
Proposition 1 Let K K K be an integral unimodular lattice of rank n = 2 k ≥ 24 n=2k\ge24 n = 2 k ≥ 24 satisfying (4.18). Then
3 k − 1 Θ K ( 3 t ) + P 3 ( K ; t ) 3 k − 1 + 1 < w ( t ) n , t ≥ 1. (24) \frac{3^{k-1}\Theta_K(3t)+P_3(K;t)}{3^{k-1}+1}<w(t)^n,\quad t\ge1. \tag*{(24)} 3 k − 1 + 1 3 k − 1 Θ K ( 3 t ) + P 3 ( K ; t ) < w ( t ) n , t ≥ 1. (24)
Proof. Put a = 3 k − 1 a=3^{k-1} a = 3 k − 1 and
B K ( t ) = a Θ K ( 3 t ) + P 3 ( K ; t ) a + 1 . B_K(t)=\frac{a\Theta_K(3t)+P_3(K;t)}{a+1}. B K ( t ) = a + 1 a Θ K ( 3 t ) + P 3 ( K ; t ) .
First suppose 1 ≤ t ≤ 2 1 \le t \le2 1 ≤ t ≤ 2 . Dropping the norm restriction in P 3 ( K ; t ) P_3(K;t) P 3 ( K ; t ) and using (4.18) give
P 3 ( K ; t ) ≤ Θ K ( t / 3 ) ≤ w ( t / 3 ) n . P_3(K;t) \le\Theta_K(t/3) \le w(t/3)^n. P 3 ( K ; t ) ≤ Θ K ( t /3 ) ≤ w ( t /3 ) n .
By the theta transformation,
w ( t / 3 ) = 3 t w ( 3 / t ) . w(t/3) = \sqrt{\frac{3}{t}}w(3/t). w ( t /3 ) = t 3 w ( 3/ t ) .
Since a = 3 k − 1 = 3 n / 2 − 1 a=3^{k-1}=3^{n/2-1} a = 3 k − 1 = 3 n /2 − 1 , it follows that
B K ( t ) ≤ a a + 1 ( w ( 3 t ) n + 3 t − n / 2 w ( 3 / t ) n ) . B_K(t) \le\frac{a}{a+1}\left(w(3t)^n+3t^{-n/2}w(3/t)^n\right). B K ( t ) ≤ a + 1 a ( w ( 3 t ) n + 3 t − n /2 w ( 3/ t ) n ) .
It is therefore sufficient to prove
w ( 3 t ) n + 3 t − n / 2 w ( 3 / t ) n < w ( t ) n , 1 ≤ t ≤ 2. (25) w(3t)^n+3t^{-n/2}w(3/t)^n < w(t)^n,\quad1 \le t \le2. \tag*{(25)} w ( 3 t ) n + 3 t − n /2 w ( 3/ t ) n < w ( t ) n , 1 ≤ t ≤ 2. (25)
Suppose first that 1 ≤ t ≤ 11 10 1 \le t \le\frac{11}{10} 1 ≤ t ≤ 10 11 . By monotonicity of w w w ,
w ( 3 t ) n + 3 t − n / 2 w ( 3 / t ) n < ( 1.0002 ) n + 3 ( 1.0004 ) n , w(3t)^n+3t^{-n/2}w(3/t)^n < (1.0002)^n+3(1.0004)^n, w ( 3 t ) n + 3 t − n /2 w ( 3/ t ) n < ( 1.0002 ) n + 3 ( 1.0004 ) n ,
whereas w ( t ) n > ( 1.062 ) n w(t)^n>(1.062)^n w ( t ) n > ( 1.062 ) n . For every n ≥ 24 n \ge24 n ≥ 24 , Lemma 4 proves (4.20).
Now suppose 11 10 ≤ t ≤ 2 \frac{11}{10} \le t \le2 10 11 ≤ t ≤ 2 and put q = e − π t q=e^{-\pi t} q = e − π t . Since w ( t ) > w ( 3 t ) ≥ 1 w(t)>w(3t)\ge1 w ( t ) > w ( 3 t ) ≥ 1 ,
w ( t ) n − w ( 3 t ) n ≥ n ( w ( t ) − w ( 3 t ) ) ≥ 2 n q ( 1 − q 2 ) . w(t)^n-w(3t)^n \ge n\left(w(t)-w(3t)\right) \ge2nq(1-q^2). w ( t ) n − w ( 3 t ) n ≥ n ( w ( t ) − w ( 3 t ) ) ≥ 2 n q ( 1 − q 2 ) .
Define
F n ( t ) = 3 2 n e π t ( w ( 3 / t ) t ) n . F_n(t)=\frac{3}{2n}e^{\pi t}\left(\frac{w(3/t)}{\sqrt{t}}\right)^n. F n ( t ) = 2 n 3 e π t ( t w ( 3/ t ) ) n .
Then
d d t log F n ( t ) = π − n 2 t + 3 n t 2 ϕ ( 3 / t ) . \frac{d}{dt}\log F_n(t)=\pi-\frac{n}{2t}+\frac{3n}{t^2}\phi(3/t). d t d log F n ( t ) = π − 2 t n + t 2 3 n ϕ ( 3/ t ) .
By Lemma 4, ϕ ( 3 / t ) < 1 16 \phi(3/t)<\frac{1}{16} ϕ ( 3/ t ) < 16 1 , and hence
d d t log F n ( t ) < π + n ( − 1 2 t + 3 16 t 2 ) ≤ π − 13 n 64 < 0. \frac{d}{dt}\log F_n(t)<\pi+n\left(-\frac{1}{2t}+\frac{3}{16t^2}\right)\le\pi-\frac{13n}{64}<0. d t d log F n ( t ) < π + n ( − 2 t 1 + 16 t 2 3 ) ≤ π − 64 13 n < 0.
Thus F n F_n F n is decreasing. Furthermore,
w ( 30 / 11 ) 11 / 10 < 24 25 , \frac{w(30/11)}{\sqrt{11/10}}<\frac{24}{25}, 11/10 w ( 30/11 ) < 25 24 ,
F n ( t ) ≤ F n ( 11 / 10 ) < 48 n ( 24 25 ) n ≤ 2 ( 24 25 ) 24 < 4 5 . F_n(t) \le F_n(11/10) < \frac{48}{n}\left(\frac{24}{25}\right)^n \le2\left(\frac{24}{25}\right)^{24} < \frac{4}{5}. F n ( t ) ≤ F n ( 11/10 ) < n 48 ( 25 24 ) n ≤ 2 ( 25 24 ) 24 < 5 4 .
On the other hand,
1 − q 2 ≥ 1 − e − 11 π / 5 > 99 100 . 1-q^2 \ge1-e^{-11\pi/5} > \frac{99}{100}. 1 − q 2 ≥ 1 − e − 11 π /5 > 100 99 .
Hence F n ( t ) < 1 − q 2 F_n(t)<1-q^2 F n ( t ) < 1 − q 2 , and (4.20) follows.
It remains to consider t ≥ 2 t\ge2 t ≥ 2 . Put G n ( t ) = w ( t ) n − 1 G_n(t)=w(t)^n-1 G n ( t ) = w ( t ) n − 1 . The scalar theta series satisfies
w ( 3 t ) − 1 ≤ e − 2 π t ( w ( t ) − 1 ) . w(3t)-1\le e^{-2\pi t}(w(t)-1). w ( 3 t ) − 1 ≤ e − 2 π t ( w ( t ) − 1 ) .
Using
( 1 + c x ) n − 1 ≤ c ( ( 1 + x ) n − 1 ) (1+cx)^n-1\le c\left((1+x)^n-1\right) ( 1 + c x ) n − 1 ≤ c ( ( 1 + x ) n − 1 )
for x ≥ 0 x\ge0 x ≥ 0 and 0 ≤ c ≤ 1 0\le c\le1 0 ≤ c ≤ 1 , we obtain
Θ K ( 3 t ) − 1 ≤ w ( 3 t ) n − 1 ≤ e − 2 π t G n ( t ) . (26) \Theta_K(3t)-1\le w(3t)^n-1\le e^{-2\pi t}G_n(t). \tag*{(26)} Θ K ( 3 t ) − 1 ≤ w ( 3 t ) n − 1 ≤ e − 2 π t G n ( t ) . (26)
Since K K K is integral, every nonzero vector counted by P 3 ( K ; t ) P_3(K;t) P 3 ( K ; t ) has squared norm at least 3 3 3 . Therefore
P 3 ( K ; t ) − 1 ≤ e − π ( t − 2 ) ( P 3 ( K ; 2 ) − 1 ) ≤ e − π ( t − 2 ) w ( 2 / 3 ) n . (27) P_3(K;t)-1\le e^{-\pi(t-2)}\left(P_3(K;2)-1\right)\le e^{-\pi(t-2)}w(2/3)^n. \tag*{(27)} P 3 ( K ; t ) − 1 ≤ e − π ( t − 2 ) ( P 3 ( K ; 2 ) − 1 ) ≤ e − π ( t − 2 ) w ( 2/3 ) n . (27)
Moreover, G n ( t ) ≥ 2 n e − π t G_n(t)\ge2ne^{-\pi t} G n ( t ) ≥ 2 n e − π t . Combining (4.21) and (4.22) gives
B K ( t ) − 1 ≤ [ e − 2 π t + 3 e 2 π 2 n ( w ( 2 / 3 ) 3 ) n ] G n ( t ) . B_K(t)-1\le\left[e^{-2\pi t}+\frac{3e^{2\pi}}{2n}\left(\frac{w(2/3)}{\sqrt{3}}\right)^n\right]G_n(t). B K ( t ) − 1 ≤ [ e − 2 π t + 2 n 3 e 2 π ( 3 w ( 2/3 ) ) n ] G n ( t ) .
Using
e − 4 π < 1 100000 , e 2 π < 540 , w ( 2 / 3 ) 3 < 3 4 , e^{-4\pi}<\frac{1}{100000},\qquad e^{2\pi}<540,\qquad\frac{w(2/3)}{\sqrt{3}}<\frac{3}{4}, e − 4 π < 100000 1 , e 2 π < 540 , 3 w ( 2/3 ) < 4 3 ,
we obtain
B K ( t ) − 1 < [ 1 100000 + 3 ⋅ 540 2 n ( 3 4 ) n ] G n ( t ) < G n ( t ) . B_K(t)-1<\left[\frac{1}{100000}+\frac{3\cdot540}{2n}\left(\frac{3}{4}\right)^n\right]G_n(t)<G_n(t). B K ( t ) − 1 < [ 100000 1 + 2 n 3 ⋅ 540 ( 4 3 ) n ] G n ( t ) < G n ( t ) .
Hence B K ( t ) < w ( t ) n B_K(t)<w(t)^n B K ( t ) < w ( t ) n for every t ≥ 2 t\ge2 t ≥ 2 , completing the proof.
Theorem 2 Let D ⊆ F 2 n D\subseteq\mathbb{F}_2^n D ⊆ F 2 n , n = 4 s > 0 n=4s>0 n = 4 s > 0 , be a binary self-dual code. Then
E T ∈ S 3 ( n ) Θ A 6 ( C ( D , T ) ) ( t ) ≤ ϑ 3 ( t ) n for every t > 0. (28) \mathop{\mathbb{E}}_{T\in\mathcal{S}_3(n)}\Theta_{A_6(C(D,T))}(t)\le\vartheta_3(t)^n\quad\text{for every }t>0. \tag*{(28)} E T ∈ S 3 ( n ) Θ A 6 ( C ( D , T )) ( t ) ≤ ϑ 3 ( t ) n for every t > 0. (28) The inequality is strict if n ≥ 24 n\ge24 n ≥ 24 . If D D D is Type II, then it is strict for every positive admissible length.
Proof. Let K = A 2 ( D ) K=A_2(D) K = A 2 ( D ) . By the Gaussian-mass inequality for binary self-dual Construction A lattices [14 ] ,
Θ K ( u ) ≤ ϑ 3 ( u ) n \Theta_K(u) \le\vartheta_3(u)^n Θ K ( u ) ≤ ϑ 3 ( u ) n
for every u > 0 u>0 u > 0 .
If n ≥ 24 n\ge24 n ≥ 24 , Theorem 1 and Proposition 1 give
E T Θ A 6 ( C ( D , T ) ) ( t ) < ϑ 3 ( t ) n \mathbb{E}_T\Theta_{A_6(C(D,T))}(t)<\vartheta_3(t)^n E T Θ A 6 ( C ( D , T )) ( t ) < ϑ 3 ( t ) n
for every t ≥ 1 t\ge1 t ≥ 1 .
Each lattice A 6 ( C ( D , T ) ) A_6(C(D,T)) A 6 ( C ( D , T )) is integral unimodular. Therefore Poisson summation gives
Θ A 6 ( C ( D , T ) ) ( t ) = t − n / 2 Θ A 6 ( C ( D , T ) ) ( 1 / t ) , \Theta_{A_6(C(D,T))}(t)=t^{-n/2}\Theta_{A_6(C(D,T))}(1/t), Θ A 6 ( C ( D , T )) ( t ) = t − n /2 Θ A 6 ( C ( D , T )) ( 1/ t ) ,
while ϑ 3 ( t ) n = t − n / 2 ϑ 3 ( 1 / t ) n \vartheta_3(t)^n=t^{-n/2}\vartheta_3(1/t)^n ϑ 3 ( t ) n = t − n /2 ϑ 3 ( 1/ t ) n . Averaging the first identity extends the strict inequality to 0 < t < 1 0<t<1 0 < t < 1 .
It remains to consider the admissible lengths n ∈ { 4 , 8 , 12 , 16 , 20 } n\in\{4,8,12,16,20\} n ∈ { 4 , 8 , 12 , 16 , 20 } below 24. In these dimensions, the Gaussian-mass upper bound for integral unimodular lattices follows from [13 Theorem 1.1] . Therefore
Θ A 6 ( C ( D , T ) ) ( t ) ≤ ϑ 3 ( t ) n \Theta_{A_6(C(D,T))}(t)\le\vartheta_3(t)^n Θ A 6 ( C ( D , T )) ( t ) ≤ ϑ 3 ( t ) n
for every T T T , and hence the same inequality holds after averaging.
If D D D is Type II, then every C ( D , T ) C(D,T) C ( D , T ) is Type II and hence every A 6 ( C ( D , T ) ) A_6(C(D,T)) A 6 ( C ( D , T )) is even unimodular. The remaining positive admissible lengths below 24 are 8 and 16. In these dimensions such a lattice cannot be isometric to Z n \mathbb{Z}^n Z n , and the strictness statement in [13 Theorem 1.1] therefore yields
Θ A 6 ( C ( D , T ) ) ( t ) < ϑ 3 ( t ) n . \Theta_{A_6(C(D,T))}(t)<\vartheta_3(t)^n. Θ A 6 ( C ( D , T )) ( t ) < ϑ 3 ( t ) n .
This proves the final assertion.
When D D D is Type II, the exact average admits a natural Hecke-theoretic interpretation. Let
f K ( τ ) = ∑ v ∈ K e π i τ ∥ v ∥ 2 , Im τ > 0. f_K(\tau)=\sum_{v\in K}e^{\pi i\tau\lVert v\rVert^2},\qquad\operatorname{Im}\tau>0. f K ( τ ) = v ∈ K ∑ e π i τ ∥ v ∥ 2 , Im τ > 0.
Since K K K is even unimodular, write ∥ v ∥ 2 = 2 m v \lVert v\rVert^2=2m_v ∥ v ∥ 2 = 2 m v , m v ∈ Z m_v\in\mathbb{Z} m v ∈ Z . The root-of-unity filter gives
P 3 ( K ; t ) = 1 3 ∑ j = 0 2 f K ( i t + j 3 ) . (29) P_3(K;t)=\frac{1}{3}\sum_{j=0}^{2}f_K\left(\frac{it+j}{3}\right). \tag*{(29)} P 3 ( K ; t ) = 3 1 j = 0 ∑ 2 f K ( 3 i t + j ) . (29)
Define
( T 3 f ) ( τ ) = 3 k − 1 f ( 3 τ ) + 1 3 ∑ j = 0 2 f ( τ + j 3 ) . (30) (T_3f)(\tau) = 3^{k-1}f(3\tau) + \frac{1}{3}\sum_{j=0}^{2} f\left(\frac{\tau+j}{3}\right). \tag*{(30)} ( T 3 f ) ( τ ) = 3 k − 1 f ( 3 τ ) + 3 1 j = 0 ∑ 2 f ( 3 τ + j ) . (30)
Corollary 1 If D D D is Type II, then
E T ∈ S 3 ( n ) Θ A 6 ( C ( D , T ) ) ( t ) = ( T 3 f K ) ( i t ) 3 k − 1 + 1 . (31) \mathbb{E}_{T\in S_3(n)}\Theta_{A_6(C(D,T))}(t) = \frac{(T_3f_K)(it)}{3^{k-1}+1}. \tag*{(31)} E T ∈ S 3 ( n ) Θ A 6 ( C ( D , T )) ( t ) = 3 k − 1 + 1 ( T 3 f K ) ( i t ) . (31) Consequently,
( T 3 f K ) ( i t ) < ( 3 k − 1 + 1 ) ϑ 3 ( t ) n (T_3f_K)(it) < (3^{k-1}+1)\vartheta_3(t)^n ( T 3 f K ) ( i t ) < ( 3 k − 1 + 1 ) ϑ 3 ( t ) n for every t > 0 t>0 t > 0 .
Proof. Equation (31) follows by substituting (29) and (30) into (22) . The strict inequality follows from Theorem 2.
Type II lattices
Extremal Type II lattices
Recall that an even unimodular lattice L L L of rank n n n is called extremal if its minimum attains the upper bound of Mallows, Odlyzko and Sloane [15 ] , namely
min L = 2 ⌊ n 24 ⌋ + 2. \min L = 2\left\lfloor\frac{n}{24}\right\rfloor+ 2. min L = 2 ⌊ 24 n ⌋ + 2.
Even unimodular lattices can exist only when n n n is a multiple of 8 [7 Chap. 7, Sec. 6, pp. 191–192] .
Theorem 3. Let L L L be an extremal even unimodular lattice of dimension n n n . Its Gaussian mass satisfies
Θ L ( q ) < ϑ 3 ( q ) n , 0 < q < 1. \Theta_L(q)<\vartheta_3(q)^n,\qquad0<q<1. Θ L ( q ) < ϑ 3 ( q ) n , 0 < q < 1. Proof. Write n = 8 m n=8m n = 8 m and put ν = ⌊ m 3 ⌋ = ⌊ n 24 ⌋ \nu=\left\lfloor\frac{m}{3}\right\rfloor=\left\lfloor\frac{n}{24}\right\rfloor ν = ⌊ 3 m ⌋ = ⌊ 24 n ⌋ . For τ \tau τ in the upper half-plane, write Q = e 2 π i τ Q=e^{2\pi i\tau} Q = e 2 π i τ and use the modular-form normalization
Θ ^ L ( τ ) = ∑ x ∈ L Q ∥ x ∥ 2 / 2 . \widehat{\Theta}_L(\tau)=\sum_{x\in L}Q^{\lVert x\rVert^2/2}. Θ L ( τ ) = x ∈ L ∑ Q ∥ x ∥ 2 /2 . Since L L L is extremal, min L = 2 ( ν + 1 ) \min L=2(\nu+1) min L = 2 ( ν + 1 ) , and therefore
Θ ^ L ( τ ) = 1 + O ( Q ν + 1 ) . \widehat{\Theta}_L(\tau)=1+O(Q^{\nu+1}). Θ L ( τ ) = 1 + O ( Q ν + 1 ) . The theta series Θ ^ L \widehat{\Theta}_L Θ L is a modular form of weight 4 m 4m 4 m for S L 2 ( Z ) SL_2(\mathbb{Z}) S L 2 ( Z ) . Since dim M 4 m ( S L 2 ( Z ) ) = ν + 1 \dim M_{4m}(SL_2(\mathbb{Z}))=\nu+1 dim M 4 m ( S L 2 ( Z )) = ν + 1 , the conditions on the first ν \nu ν nonconstant Fourier coefficients determine it uniquely. Thus Θ ^ L \widehat{\Theta}_L Θ L is the so-called gap modular form of weight 4 m 4m 4 m . It can be written in the form
Θ ^ L ( τ ) = E 4 ( τ ) m P m ( Δ ( τ ) E 4 ( τ ) 3 ) , \widehat{\Theta}_L(\tau)=E_4(\tau)^mP_m\left(\frac{\Delta(\tau)}{E_4(\tau)^3}\right), Θ L ( τ ) = E 4 ( τ ) m P m ( E 4 ( τ ) 3 Δ ( τ ) ) , where P m ( X ) P_m(X) P m ( X ) is a real polynomial of degree at most ν \nu ν and P m ( 0 ) = 1 P_m(0)=1 P m ( 0 ) = 1 . Duke and Jenkins proved that all zeros of this gap modular form in the standard fundamental domain lie on the unit-circle part of its lower boundary; see [8 Th. 1] . On this arc it is well-known that the modular invariant j ( τ ) = E 4 ( τ ) 3 / Δ ( τ ) j(\tau)=E_4(\tau)^3/\Delta(\tau) j ( τ ) = E 4 ( τ ) 3 /Δ ( τ ) is real and ranges from 0 to 1728 [1 Chapter 3] . Let α \alpha α be a finite zero of P m P_m P m . Since P m ( 0 ) = 1 P_m(0)=1 P m ( 0 ) = 1 , one has α ≠ 0 \alpha\ne0 α = 0 . Choose τ \tau τ in the standard fundamental domain with j ( τ ) = 1 α j(\tau)=\frac{1}{\alpha} j ( τ ) = α 1 . Then E 4 ( τ ) ≠ 0 E_4(\tau)\ne0 E 4 ( τ ) = 0 and
Θ ^ L ( τ ) = E 4 ( τ ) m P m ( α ) = 0. \widehat{\Theta}_L(\tau)=E_4(\tau)^mP_m(\alpha)=0. Θ L ( τ ) = E 4 ( τ ) m P m ( α ) = 0. Hence τ \tau τ lies on the lower boundary arc, so 0 < j ( τ ) ≤ 1728 0<j(\tau)\le1728 0 < j ( τ ) ≤ 1728 . It follows that every zero α \alpha α of P m P_m P m satisfies α ≥ 1 1728 \alpha\ge\frac{1}{1728} α ≥ 1728 1 . Counting the zeros with multiplicity and using P m ( 0 ) = 1 P_m(0)=1 P m ( 0 ) = 1 , we may therefore write
P m ( X ) = ∏ i ( 1 − X α i ) , α i ≥ 1 1728 . P_m(X)=\prod_i\left(1-\frac{X}{\alpha_i}\right),\qquad\alpha_i\ge\frac{1}{1728}. P m ( X ) = i ∏ ( 1 − α i X ) , α i ≥ 1728 1 .
Consequently, for 0 ≤ X ≤ 1 / 1728 0 \le X \le1/1728 0 ≤ X ≤ 1/1728 ,
0 ≤ P m ( X ) ≤ 1. 0 \le P_m(X) \le1. 0 ≤ P m ( X ) ≤ 1.
Now take τ = i y \tau= iy τ = i y with y > 0 y > 0 y > 0 and consider the auxiliary variable
u = ϑ 2 ( τ ) 4 ϑ 4 ( τ ) 4 ϑ 3 ( τ ) 8 = ϑ 2 ( τ ) 4 ϑ 3 ( τ ) 4 × ϑ 4 ( τ ) 4 ϑ 3 ( τ ) 4 . u = \frac{\vartheta_2(\tau)^4\vartheta_4(\tau)^4}{\vartheta_3(\tau)^8}
= \frac{\vartheta_2(\tau)^4}{\vartheta_3(\tau)^4} \times\frac{\vartheta_4(\tau)^4}{\vartheta_3(\tau)^4}. u = ϑ 3 ( τ ) 8 ϑ 2 ( τ ) 4 ϑ 4 ( τ ) 4 = ϑ 3 ( τ ) 4 ϑ 2 ( τ ) 4 × ϑ 3 ( τ ) 4 ϑ 4 ( τ ) 4 .
Jacobi’s identity in the form
1 = ϑ 2 ( τ ) 4 ϑ 3 ( τ ) 4 + ϑ 4 ( τ ) 4 ϑ 3 ( τ ) 4 , 1 = \frac{\vartheta_2(\tau)^4}{\vartheta_3(\tau)^4} + \frac{\vartheta_4(\tau)^4}{\vartheta_3(\tau)^4}, 1 = ϑ 3 ( τ ) 4 ϑ 2 ( τ ) 4 + ϑ 3 ( τ ) 4 ϑ 4 ( τ ) 4 ,
combined with the AGM inequality gives
0 < u ≤ 1 4 . 0 < u \le\frac{1}{4}. 0 < u ≤ 4 1 .
To get rid of E 4 E_4 E 4 we use the standard theta identities yielding the two equations
E 4 ( τ ) ϑ 3 ( τ ) 8 = 1 − u , Δ ( τ ) ϑ 3 ( τ ) 24 = u 2 256 . \frac{E_4(\tau)}{\vartheta_3(\tau)^8} = 1-u,\qquad\frac{\Delta(\tau)}{\vartheta_3(\tau)^{24}} = \frac{u^2}{256}. ϑ 3 ( τ ) 8 E 4 ( τ ) = 1 − u , ϑ 3 ( τ ) 24 Δ ( τ ) = 256 u 2 .
Therefore, taking quotients we get
Δ ( τ ) E 4 ( τ ) 3 = u 2 256 ( 1 − u ) 3 = : X ( u ) . \frac{\Delta(\tau)}{E_4(\tau)^3} = \frac{u^2}{256(1-u)^3} =: X(u). E 4 ( τ ) 3 Δ ( τ ) = 256 ( 1 − u ) 3 u 2 =: X ( u ) .
The function X ( u ) X(u) X ( u ) is increasing for 0 < u ≤ 1 / 4 0 < u \le1/4 0 < u ≤ 1/4 , and X ( 1 4 ) = 1 1728 X\left(\frac{1}{4}\right) = \frac{1}{1728} X ( 4 1 ) = 1728 1 . Hence
0 < X ( u ) ≤ 1 1728 . 0 < X(u) \le\frac{1}{1728}. 0 < X ( u ) ≤ 1728 1 .
Dividing the expression for Θ ^ L \widehat{\Theta}_L Θ L by ϑ 3 ( τ ) 8 m \vartheta_3(\tau)^{8m} ϑ 3 ( τ ) 8 m gives
Θ ^ L ( τ ) ϑ 3 ( τ ) 8 m = ( 1 − u ) m P m ( X ( u ) ) . \frac{\widehat{\Theta}_L(\tau)}{\vartheta_3(\tau)^{8m}} = (1-u)^m P_m(X(u)). ϑ 3 ( τ ) 8 m Θ L ( τ ) = ( 1 − u ) m P m ( X ( u )) .
Since 0 ≤ P m ( X ( u ) ) ≤ 1 0 \le P_m(X(u)) \le1 0 ≤ P m ( X ( u )) ≤ 1 and 0 < u ≤ 1 / 4 0 < u \le1/4 0 < u ≤ 1/4 , we obtain
Θ ^ L ( τ ) ϑ 3 ( τ ) 8 m ≤ ( 1 − u ) m < 1. \frac{\widehat{\Theta}_L(\tau)}{\vartheta_3(\tau)^{8m}} \le(1-u)^m < 1. ϑ 3 ( τ ) 8 m Θ L ( τ ) ≤ ( 1 − u ) m < 1.
Finally, for every 0 < q < 1 0 < q < 1 0 < q < 1 choose y > 0 y > 0 y > 0 such that q = e − π y q = e^{-\pi y} q = e − π y . Then Θ ^ L ( i y ) = Θ L ( q ) \widehat{\Theta}_L(iy) = \Theta_L(q) Θ L ( i y ) = Θ L ( q ) and ϑ 3 ( i y ) = ϑ 3 ( q ) \vartheta_3(iy) = \vartheta_3(q) ϑ 3 ( i y ) = ϑ 3 ( q ) , so
Θ L ( q ) < ϑ 3 ( q ) n , 0 < q < 1. \Theta_L(q) < \vartheta_3(q)^n,\qquad0 < q < 1. Θ L ( q ) < ϑ 3 ( q ) n , 0 < q < 1.
This proves the result.
Genus averages of Type II lattices
In the preceding subsection, we established a pointwise Gaussian-mass inequality for extremal even unimodular lattices. We now consider a different averaging procedure, namely averaging theta series over the genus of all positive-definite even unimodular lattices of a fixed rank. Unlike the averaging methods for Construction A 4 A_4 A 4 and Construction A 6 A_6 A 6 , this procedure does not depend on a particular code construction.
The connection between genus averages and Eisenstein series is classical and follows from Siegel’s averaging formula. It has also been used in the study of secrecy gains of unimodular lattices (see [18 ] ). Here we use this connection to obtain a uniform pointwise Gaussian-mass bound for the genus average, with an explicit comparison to the orthogonal sum of copies of the E 8 E_8 E 8 lattice.
Let n = 8 k n = 8k n = 8 k , where k ≥ 1 k \ge1 k ≥ 1 , and let G n \mathcal{G}_n G n denote the genus of positive-definite even unimodular lattices of rank n n n . The genus consists of finitely many isometry classes. Its mass is defined by
Mass ( G n ) = ∑ [ L ] ∈ G n 1 ∣ Aut ( L ) ∣ . \operatorname{Mass}(\mathcal{G}_n) = \sum_{[L]\in\mathcal{G}_n} \frac{1}{|\operatorname{Aut}(L)|}. Mass ( G n ) = [ L ] ∈ G n ∑ ∣ Aut ( L ) ∣ 1 .
The corresponding weighted genus-average theta series is
Θ G n ( q ) = 1 Mass ( G n ) ∑ [ L ] ∈ G n Θ L ( q ) ∣ Aut ( L ) ∣ . \Theta_{\mathcal{G}_n}(q) = \frac{1}{\operatorname{Mass}(\mathcal{G}_n)} \sum_{[L]\in\mathcal{G}_n} \frac{\Theta_L(q)}{|\operatorname{Aut}(L)|}. Θ G n ( q ) = Mass ( G n ) 1 [ L ] ∈ G n ∑ ∣ Aut ( L ) ∣ Θ L ( q ) .
To describe this average, let τ \tau τ belong to the complex upper half-plane and put Q = e 2 π i τ Q = e^{2\pi i\tau} Q = e 2 π i τ . For an integer r ≥ 2 r \ge2 r ≥ 2 , we use the standard normalized Eisenstein series E 2 r ( τ ) E_{2r}(\tau) E 2 r ( τ ) of weight 2 r 2r 2 r (see [1 ] ), defined by
E 2 r ( τ ) = 1 − 4 r B 2 r ∑ ℓ = 1 ∞ σ 2 r − 1 ( ℓ ) Q ℓ . E_{2r}(\tau) = 1 - \frac{4r}{B_{2r}} \sum_{\ell=1}^{\infty} \sigma_{2r-1}(\ell)Q^\ell. E 2 r ( τ ) = 1 − B 2 r 4 r ℓ = 1 ∑ ∞ σ 2 r − 1 ( ℓ ) Q ℓ .
where B 2 r B_{2r} B 2 r is the Bernoulli number and σ s ( ℓ ) = ∑ d ∣ ℓ d s \sigma_s(\ell) = \sum_{d\mid\ell} d^s σ s ( ℓ ) = ∑ d ∣ ℓ d s .
By the classical Siegel genus-average formula (see also [18 ] ), the weighted average of the theta series of even unimodular lattices satisfies
Θ ^ G 8 k ( τ ) = E 4 k ( τ ) , \widehat{\Theta}_{\mathcal{G}_{8k}}(\tau) = E_{4k}(\tau), Θ G 8 k ( τ ) = E 4 k ( τ ) ,
where
Θ ^ G 8 k ( τ ) = 1 Mass ( G 8 k ) ∑ [ L ] ∈ G 8 k Θ ^ L ( τ ) ∣ Aut ( L ) ∣ . \widehat{\Theta}_{\mathcal{G}_{8k}}(\tau) = \frac{1}{\operatorname{Mass}(\mathcal{G}_{8k})} \sum_{[L]\in\mathcal{G}_{8k}} \frac{\widehat{\Theta}_L(\tau)}{|\operatorname{Aut}(L)|}. Θ G 8 k ( τ ) = Mass ( G 8 k ) 1 [ L ] ∈ G 8 k ∑ ∣ Aut ( L ) ∣ Θ L ( τ ) .
and
Θ ^ L ( τ ) = ∑ x ∈ L e π i τ ( x , x ) . \widehat{\Theta}_L(\tau)=\sum_{x\in L}e^{\pi i\tau(x,x)}. Θ L ( τ ) = x ∈ L ∑ e π i τ ( x , x ) .
In particular, if q = e − π t q=e^{-\pi t} q = e − π t with t > 0 t>0 t > 0 , then
Θ G 8 k ( q ) = E 4 k ( i t ) . \Theta_{G_{8k}}(q)=E_{4k}(it). Θ G 8 k ( q ) = E 4 k ( i t ) .
We first establish an inequality for normalized Eisenstein series on the positive imaginary axis.
Lemma 5 For every integer m ≥ 2 m\ge2 m ≥ 2 and every t > 0 t>0 t > 0 ,
∣ E 2 m ( i t ) ∣ ≤ E 4 ( i t ) m / 2 . \lvert E_{2m}(it)\rvert\le E_4(it)^{m/2}. ∣ E 2 m ( i t )∣ ≤ E 4 ( i t ) m /2 . Equality holds for m = 2 , 4 m=2,4 m = 2 , 4 , whereas the inequality is strict for all other integers m ≥ 2 m\ge2 m ≥ 2 . In particular, for every integer k ≥ 1 k\ge1 k ≥ 1 ,
E 4 k ( i t ) ≤ E 4 ( i t ) k , E_{4k}(it)\le E_4(it)^k, E 4 k ( i t ) ≤ E 4 ( i t ) k , with equality if and only if k = 1 , 2 k=1,2 k = 1 , 2 .
Proof. For m ≥ 2 m\ge2 m ≥ 2 , consider the classical Eisenstein series (see [1 ] )
G 2 m ( τ ) = ∑ ( a , b ) ∈ Z 2 ∖ { ( 0 , 0 ) } 1 ( a + b τ ) 2 m . G_{2m}(\tau)=\sum_{(a,b)\in\mathbb{Z}^2\setminus\{(0,0)\}}\frac{1}{(a+b\tau)^{2m}}. G 2 m ( τ ) = ( a , b ) ∈ Z 2 ∖ {( 0 , 0 )} ∑ ( a + b τ ) 2 m 1 .
These series are related to the normalized Eisenstein series by
G 2 m ( τ ) = 2 ζ ( 2 m ) E 2 m ( τ ) , G_{2m}(\tau)=2\zeta(2m)E_{2m}(\tau), G 2 m ( τ ) = 2 ζ ( 2 m ) E 2 m ( τ ) ,
where ζ \zeta ζ denotes the Riemann zeta function.
We use the classical recurrence for Eisenstein series, which follows by comparing coefficients in the Laurent expansion of the Weierstrass elliptic function (see [1 ] ). For every integer m ≥ 4 m\ge4 m ≥ 4 ,
G 2 m ( τ ) = 3 ( m − 3 ) ( 2 m − 1 ) ( 2 m + 1 ) ∑ j = 2 m − 2 ( 2 j − 1 ) ( 2 m − 2 j − 1 ) G 2 j ( τ ) G 2 m − 2 j ( τ ) . G_{2m}(\tau)=\frac{3}{(m-3)(2m-1)(2m+1)}
\sum_{j=2}^{m-2}(2j-1)(2m-2j-1)G_{2j}(\tau)G_{2m-2j}(\tau). G 2 m ( τ ) = ( m − 3 ) ( 2 m − 1 ) ( 2 m + 1 ) 3 j = 2 ∑ m − 2 ( 2 j − 1 ) ( 2 m − 2 j − 1 ) G 2 j ( τ ) G 2 m − 2 j ( τ ) .
Substituting G 2 r = 2 ζ ( 2 r ) E 2 r G_{2r}=2\zeta(2r)E_{2r} G 2 r = 2 ζ ( 2 r ) E 2 r gives
E 2 m ( τ ) = ∑ j = 2 m − 2 c m , j E 2 j ( τ ) E 2 m − 2 j ( τ ) , E_{2m}(\tau)=\sum_{j=2}^{m-2}c_{m,j}E_{2j}(\tau)E_{2m-2j}(\tau), E 2 m ( τ ) = j = 2 ∑ m − 2 c m , j E 2 j ( τ ) E 2 m − 2 j ( τ ) ,
where
c m , j = 6 ( 2 j − 1 ) ( 2 m − 2 j − 1 ) ζ ( 2 j ) ζ ( 2 m − 2 j ) ( m − 3 ) ( 2 m − 1 ) ( 2 m + 1 ) ζ ( 2 m ) . c_{m,j}=\frac{6(2j-1)(2m-2j-1)\zeta(2j)\zeta(2m-2j)}{(m-3)(2m-1)(2m+1)\zeta(2m)}. c m , j = ( m − 3 ) ( 2 m − 1 ) ( 2 m + 1 ) ζ ( 2 m ) 6 ( 2 j − 1 ) ( 2 m − 2 j − 1 ) ζ ( 2 j ) ζ ( 2 m − 2 j ) .
Since ζ ( s ) > 0 \zeta(s)>0 ζ ( s ) > 0 for s > 1 s>1 s > 1 , we have c m , j > 0 c_{m,j}>0 c m , j > 0 . Moreover, every normalized Eisenstein series has constant Fourier coefficient equal to 1. Comparing constant coefficients in the preceding identity yields
∑ j = 2 m − 2 c m , j = 1. \sum_{j=2}^{m-2}c_{m,j}=1. j = 2 ∑ m − 2 c m , j = 1.
Thus E 2 m E_{2m} E 2 m is a positive convex combination of products of Eisenstein series of lower weights.
We now prove the desired inequality by strong induction on m m m .
For m = 2 m=2 m = 2 , the assertion is immediate.
For m = 3 m=3 m = 3 , recall the classical modular-form identity
E 4 ( τ ) 3 − E 6 ( τ ) 2 = 1728 Δ ( τ ) , E_4(\tau)^3-E_6(\tau)^2=1728\Delta(\tau), E 4 ( τ ) 3 − E 6 ( τ ) 2 = 1728Δ ( τ ) ,
where
Δ ( τ ) = Q ∏ ℓ = 1 ∞ ( 1 − Q ℓ ) 24 . \Delta(\tau)=Q\prod_{\ell=1}^{\infty}(1-Q^\ell)^{24}. Δ ( τ ) = Q ℓ = 1 ∏ ∞ ( 1 − Q ℓ ) 24 .
When τ = i t \tau=it τ = i t with t > 0 t>0 t > 0 , we have 0 < Q < 1 0<Q<1 0 < Q < 1 , and therefore Δ ( i t ) > 0 \Delta(it)>0 Δ ( i t ) > 0 . Since E 4 ( i t ) > 0 E_4(it)>0 E 4 ( i t ) > 0 and E 6 ( i t ) E_6(it) E 6 ( i t ) is real, it follows that E 6 ( i t ) 2 < E 4 ( i t ) 3 E_6(it)^2<E_4(it)^3 E 6 ( i t ) 2 < E 4 ( i t ) 3 . Consequently,
∣ E 6 ( i t ) ∣ < E 4 ( i t ) 3 / 2 . |E_6(it)|<E_4(it)^{3/2}. ∣ E 6 ( i t ) ∣ < E 4 ( i t ) 3/2 .
Suppose now that m ≥ 4 m\geq4 m ≥ 4 and that the assertion holds for every integer r r r with 2 ≤ r < m 2\leq r<m 2 ≤ r < m . Using the recurrence and the positivity of the coefficients c m , j c_{m,j} c m , j , we obtain
∣ E 2 m ( i t ) ∣ ≤ ∑ j = 2 m − 2 c m , j ∣ E 2 j ( i t ) ∣ ∣ E 2 m − 2 j ( i t ) ∣ ≤ ∑ j = 2 m − 2 c m , j E 4 ( i t ) j / 2 E 4 ( i t ) ( m − j ) / 2 = E 4 ( i t ) m / 2 ∑ j = 2 m − 2 c m , j = E 4 ( i t ) m / 2 . \begin{aligned}
\lvert E_{2m}(it)\rvert\le\sum_{j=2}^{m-2} c_{m,j}\lvert E_{2j}(it)\rvert\lvert E_{2m-2j}(it)\rvert\\
&\le\sum_{j=2}^{m-2} c_{m,j}E_4(it)^{j/2}E_4(it)^{(m-j)/2} \\
&= E_4(it)^{m/2}\sum_{j=2}^{m-2}c_{m,j} \\
&= E_4(it)^{m/2}.
\end{aligned} ∣ E 2 m ( i t )∣ ≤ j = 2 ∑ m − 2 c m , j ∣ E 2 j ( i t )∣ ∣ E 2 m − 2 j ( i t )∣ ≤ j = 2 ∑ m − 2 c m , j E 4 ( i t ) j /2 E 4 ( i t ) ( m − j ) /2 = E 4 ( i t ) m /2 j = 2 ∑ m − 2 c m , j = E 4 ( i t ) m /2 .
This completes the induction.
When m = 4 m=4 m = 4 , the recurrence reduces to E 8 ( τ ) = E 4 ( τ ) 2 E_8(\tau)=E_4(\tau)^2 E 8 ( τ ) = E 4 ( τ ) 2 , so equality holds. For every m ≥ 5 m\ge5 m ≥ 5 , the recurrence contains a term involving E 6 E_6 E 6 , whose bound is strict on the positive imaginary axis. Since all the coefficients c m , j c_{m,j} c m , j are positive, the resulting inequality is strict. Finally, taking m = 2 k m=2k m = 2 k gives
E 4 k ( i t ) ≤ E 4 ( i t ) k , E_{4k}(it)\le E_4(it)^k, E 4 k ( i t ) ≤ E 4 ( i t ) k ,
with equality precisely when k = 1 , 2 k=1,2 k = 1 , 2 .
We can now obtain the Gaussian-mass inequality for the genus average of Type II lattices.
Theorem 4 Let n = 8 k n=8k n = 8 k , where k ≥ 1 k\ge1 k ≥ 1 , and let G n \mathcal{G}_n G n denote the genus of positive-definite even unimodular lattices of rank n n n . Then, for every 0 < q < 1 0<q<1 0 < q < 1 ,
Θ G 8 k ( q ) ≤ Θ E 8 ⊕ k ( q ) < Θ Z 8 k ( q ) = ϑ 3 ( q ) 8 k . \Theta_{\mathcal{G}_{8k}}(q)\le\Theta_{E_8^{\oplus k}}(q)<\Theta_{\mathbb{Z}^{8k}}(q)=\vartheta_3(q)^{8k}. Θ G 8 k ( q ) ≤ Θ E 8 ⊕ k ( q ) < Θ Z 8 k ( q ) = ϑ 3 ( q ) 8 k . Equality in the first inequality holds for k = 1 , 2 k=1,2 k = 1 , 2 , whereas it is strict for every $k\ge 3.
Moreover, define
u ( q ) = ϑ 2 ( q ) 4 ϑ 4 ( q ) 4 ϑ 3 ( q ) 8 . u(q)=\frac{\vartheta_2(q)^4\vartheta_4(q)^4}{\vartheta_3(q)^8}. u ( q ) = ϑ 3 ( q ) 8 ϑ 2 ( q ) 4 ϑ 4 ( q ) 4 . Then 0 < u ( q ) ≤ 1 4 0<u(q)\le\frac{1}{4} 0 < u ( q ) ≤ 4 1 and
Θ G 8 k ( q ) ϑ 3 ( q ) 8 k ≤ ( 1 − u ( q ) ) k < 1. \frac{\Theta_{\mathcal{G}_{8k}}(q)}{\vartheta_3(q)^{8k}}\le(1-u(q))^k<1. ϑ 3 ( q ) 8 k Θ G 8 k ( q ) ≤ ( 1 − u ( q ) ) k < 1.
Proof. Fix 0 < q < 1 0 < q < 1 0 < q < 1 and choose t > 0 t > 0 t > 0 such that q = e − π t q = e^{-\pi t} q = e − π t . By the Siegel genus-average formula,
Θ G 8 k ( q ) = E 4 k ( i t ) . \Theta_{G_{8k}}(q) = E_{4k}(it). Θ G 8 k ( q ) = E 4 k ( i t ) .
Applying Lemma 5, we obtain
Θ G 8 k ( q ) ≤ E 4 ( i t ) k . \Theta_{G_{8k}}(q) \le E_4(it)^k. Θ G 8 k ( q ) ≤ E 4 ( i t ) k .
The theta series of the E 8 E_8 E 8 lattice is the normalized Eisenstein series E 4 E_4 E 4 (see [7 Chap. 4, Sec. 8.1, p. 123] ). Since theta series are multiplicative under orthogonal direct sums,
Θ E 8 ⊕ k ( q ) = Θ E 8 ( q ) k = E 4 ( i t ) k . \Theta_{E_8^{\oplus k}}(q) = \Theta_{E_8}(q)^k = E_4(it)^k. Θ E 8 ⊕ k ( q ) = Θ E 8 ( q ) k = E 4 ( i t ) k .
Therefore,
Θ G 8 k ( q ) ≤ Θ E 8 ⊕ k ( q ) . \Theta_{G_{8k}}(q) \le\Theta_{E_8^{\oplus k}}(q). Θ G 8 k ( q ) ≤ Θ E 8 ⊕ k ( q ) .
To compare this expression with the theta series of the integer lattice, recall Jacobi’s identity
ϑ 3 ( q ) 4 = ϑ 2 ( q ) 4 + ϑ 4 ( q ) 4 . \vartheta_3(q)^4 = \vartheta_2(q)^4 + \vartheta_4(q)^4. ϑ 3 ( q ) 4 = ϑ 2 ( q ) 4 + ϑ 4 ( q ) 4 .
The standard theta-function representation of E 4 E_4 E 4 (equivalently, the theta series of E 8 E_8 E 8 ; see [7 Chap. 4, Sec. 8.1, p. 123] ) gives
E 4 ( i t ) = 1 2 ( ϑ 2 ( q ) 8 + ϑ 3 ( q ) 8 + ϑ 4 ( q ) 8 ) . E_4(it) = \frac{1}{2}\left(\vartheta_2(q)^8 + \vartheta_3(q)^8 + \vartheta_4(q)^8\right). E 4 ( i t ) = 2 1 ( ϑ 2 ( q ) 8 + ϑ 3 ( q ) 8 + ϑ 4 ( q ) 8 ) .
Using Jacobi’s identity, this simplifies to
E 4 ( i t ) = ϑ 3 ( q ) 8 − ϑ 2 ( q ) 4 ϑ 4 ( q ) 4 . E_4(it) = \vartheta_3(q)^8 - \vartheta_2(q)^4\vartheta_4(q)^4. E 4 ( i t ) = ϑ 3 ( q ) 8 − ϑ 2 ( q ) 4 ϑ 4 ( q ) 4 .
Hence
E 4 ( i t ) ϑ 3 ( q ) 8 = 1 − u ( q ) . \frac{E_4(it)}{\vartheta_3(q)^8} = 1 - u(q). ϑ 3 ( q ) 8 E 4 ( i t ) = 1 − u ( q ) .
For 0 < q < 1 0 < q < 1 0 < q < 1 , the defining series show that ϑ 2 ( q ) > 0 \vartheta_2(q) > 0 ϑ 2 ( q ) > 0 and ϑ 3 ( q ) > 0 \vartheta_3(q) > 0 ϑ 3 ( q ) > 0 . For ϑ 4 \vartheta_4 ϑ 4 , the Jacobi triple-product formula (see [7 Chap. 4, Sec. 4.1, pp. 102–103] ) gives
ϑ 4 ( q ) = ∏ j = 1 ∞ ( 1 − q 2 j ) ( 1 − q 2 j − 1 ) 2 > 0. \vartheta_4(q) = \prod_{j=1}^{\infty}(1-q^{2j})(1-q^{2j-1})^2 > 0. ϑ 4 ( q ) = j = 1 ∏ ∞ ( 1 − q 2 j ) ( 1 − q 2 j − 1 ) 2 > 0.
Indeed, all factors are positive and the infinite product converges to a nonzero limit for 0 < q < 1 0 < q < 1 0 < q < 1 . Consequently, u ( q ) > 0 u(q) > 0 u ( q ) > 0 . Furthermore,
u ( q ) = ϑ 2 ( q ) 4 ϑ 4 ( q ) 4 ϑ 3 ( q ) 4 ϑ 3 ( q ) 4 ≤ 1 4 ( ϑ 2 ( q ) 4 + ϑ 4 ( q ) 4 ϑ 3 ( q ) 4 ) 2 = 1 4 . u(q) = \frac{\vartheta_2(q)^4\vartheta_4(q)^4}{\vartheta_3(q)^4\vartheta_3(q)^4} \le\frac{1}{4}\left(\frac{\vartheta_2(q)^4+\vartheta_4(q)^4}{\vartheta_3(q)^4}\right)^2 = \frac{1}{4}. u ( q ) = ϑ 3 ( q ) 4 ϑ 3 ( q ) 4 ϑ 2 ( q ) 4 ϑ 4 ( q ) 4 ≤ 4 1 ( ϑ 3 ( q ) 4 ϑ 2 ( q ) 4 + ϑ 4 ( q ) 4 ) 2 = 4 1 .
Combining the preceding identities yields
Θ G 8 k ( q ) ϑ 3 ( q ) 8 k ≤ E 4 ( i t ) k ϑ 3 ( q ) 8 k = ( 1 − u ( q ) ) k < 1. \frac{\Theta_{\mathcal{G}_{8k}}(q)}{\vartheta_3(q)^{8k}} \le\frac{E_4(it)^k}{\vartheta_3(q)^{8k}} = (1-u(q))^k < 1. ϑ 3 ( q ) 8 k Θ G 8 k ( q ) ≤ ϑ 3 ( q ) 8 k E 4 ( i t ) k = ( 1 − u ( q ) ) k < 1.
Therefore,
Θ G 8 k ( q ) ≤ Θ E 8 ⊕ k ( q ) < ϑ 3 ( q ) 8 k . \Theta_{\mathcal{G}_{8k}}(q) \le\Theta_{E_8^{\oplus k}}(q) < \vartheta_3(q)^{8k}. Θ G 8 k ( q ) ≤ Θ E 8 ⊕ k ( q ) < ϑ 3 ( q ) 8 k .
The equality cases follow from Lemma 5.
The preceding theorem also provides a quantitative lower bound for the difference between the Gaussian masses of the integer lattice and the genus average. Indeed, for every 0 < q < 1 0 < q < 1 0 < q < 1 ,
ϑ 3 ( q ) 8 k − Θ G 8 k ( q ) ≥ ϑ 3 ( q ) 8 k ( 1 − ( 1 − u ( q ) ) k ) > 0. \vartheta_3(q)^{8k} - \Theta_{\mathcal{G}_{8k}}(q) \ge\vartheta_3(q)^{8k}\left(1-(1-u(q))^k\right) > 0. ϑ 3 ( q ) 8 k − Θ G 8 k ( q ) ≥ ϑ 3 ( q ) 8 k ( 1 − ( 1 − u ( q ) ) k ) > 0.
At the symmetry point q = e − π q=e^{-\pi} q = e − π , the classical identities for Jacobi theta constants imply
ϑ 2 ( e − π ) 4 = ϑ 4 ( e − π ) 4 = 1 2 ϑ 3 ( e − π ) 4 . \vartheta_2(e^{-\pi})^4 = \vartheta_4(e^{-\pi})^4 = \frac{1}{2}\vartheta_3(e^{-\pi})^4. ϑ 2 ( e − π ) 4 = ϑ 4 ( e − π ) 4 = 2 1 ϑ 3 ( e − π ) 4 .
Consequently, u ( e − π ) = 1 4 u(e^{-\pi})=\frac{1}{4} u ( e − π ) = 4 1 , and Theorem 4 gives
Θ G 8 k ( e − π ) ϑ 3 ( e − π ) 8 k ≤ ( 3 4 ) k . \frac{\Theta_{\mathcal{G}_{8k}}(e^{-\pi})}{\vartheta_3(e^{-\pi})^{8k}} \le\left(\frac{3}{4}\right)^k. ϑ 3 ( e − π ) 8 k Θ G 8 k ( e − π ) ≤ ( 4 3 ) k .
Equivalently,
ϑ 3 ( e − π ) 8 k Θ G 8 k ( e − π ) ≥ ( 4 3 ) k . \frac{\vartheta_3(e^{-\pi})^{8k}}{\Theta_{\mathcal{G}_{8k}}(e^{-\pi})} \ge\left(\frac{4}{3}\right)^k. Θ G 8 k ( e − π ) ϑ 3 ( e − π ) 8 k ≥ ( 3 4 ) k .
We conclude by emphasizing the distinction between this result and an individual Gaussian-mass inequality. Theorem 4 concerns the weighted average of the theta series over the full genus. It does not, by itself, imply that every even unimodular lattice L L L of rank 8 k 8k 8 k satisfies Θ L ( q ) ≤ ϑ 3 ( q ) 8 k \Theta_L(q) \le\vartheta_3(q)^{8k} Θ L ( q ) ≤ ϑ 3 ( q ) 8 k for all 0 < q < 1 0 < q < 1 0 < q < 1 .
Nevertheless, the genus-average bound holds in every rank divisible by eight, without requiring a classification of the lattices in the genus or a restriction on their minimum norms. It therefore provides a complementary averaged perspective to the pointwise results for extremal Type II lattices.
Extremal Type I lattices
We now prove the strict Gaussian-mass inequality for extremal odd unimodular lattices. The argument uses the shadow and applies without an s s s -extremality assumption.
Statement and scope
For an odd unimodular lattice L L L , let
L 0 = { x ∈ L : ( x , x ) ∈ 2 Z } , S = L 0 ∗ ∖ L L_0=\{x\in L:(x,x)\in2\mathbb{Z}\},\qquad S=L_0^*\setminus L L 0 = { x ∈ L : ( x , x ) ∈ 2 Z } , S = L 0 ∗ ∖ L
be its even sublattice and its shadow, respectively, and write Θ S ( q ) = ∑ x ∈ S q ( x , x ) \Theta_S(q)=\sum_{x\in S}q^{(x,x)} Θ S ( q ) = ∑ x ∈ S q ( x , x ) . We use the theta constants defined earlier and put
Δ 8 ( q ) = ϑ 2 ( q ) 4 ϑ 4 ( q ) 4 16 . \Delta_8(q)=\frac{\vartheta_2(q)^4\vartheta_4(q)^4}{16}. Δ 8 ( q ) = 16 ϑ 2 ( q ) 4 ϑ 4 ( q ) 4 .
In this section, extremal means attaining the Rains–Sloane minimum-norm bound [19 ] : the bound is 2 ⌊ n / 24 ⌋ + 2 2\lfloor n/24\rfloor+2 2 ⌊ n /24 ⌋ + 2 when n ≠ 23 n\ne23 n = 23 , with the exceptional bound 3 3 3 when n = 23 n=23 n = 23 .
Theorem 5 Let L L L be a positive-definite integral odd unimodular lattice of rank n > 0 n>0 n > 0 . Suppose either n ≠ 23 n\ne23 n = 23 , min ( L ) = 2 ⌊ n / 24 ⌋ + 2 \min(L)=2\lfloor n/24\rfloor+2 min ( L ) = 2 ⌊ n /24 ⌋ + 2 , or n = 23 n=23 n = 23 and min ( L ) = 3 \min(L)=3 min ( L ) = 3 . Then
Θ L ( q ) < ϑ 3 ( q ) n ( 0 < q < 1 ) . \Theta_L(q)<\vartheta_3(q)^n\qquad(0<q<1). Θ L ( q ) < ϑ 3 ( q ) n ( 0 < q < 1 ) . Equivalently,
∑ x ∈ L e − s ( x , x ) < ∑ z ∈ Z n e − s ( z , z ) \sum_{x\in L}e^{-s(x,x)}<\sum_{z\in\mathbb{Z}^n}e^{-s(z,z)} x ∈ L ∑ e − s ( x , x ) < z ∈ Z n ∑ e − s ( z , z ) for every s > 0 s>0 s > 0 .
This statement covers extremal Type I lattices in every rank in which they exist.
The universal theta truncation
Set
u I = Δ 8 ϑ 3 8 , t = 64 u I , K = ⌊ n / 8 ⌋ . u_{\mathrm I}=\frac{\Delta_8}{\vartheta_3^8},\qquad t=64u_{\mathrm I},\qquad K=\lfloor n/8\rfloor. u I = ϑ 3 8 Δ 8 , t = 64 u I , K = ⌊ n /8 ⌋ .
The standard theta and shadow expansions [19 , 11 ] are
Θ L ( q ) = ∑ j = 0 K c j Δ 8 ( q ) j ϑ 3 ( q ) n − 8 j , (32) \Theta_L(q)=\sum_{j=0}^{K}c_j\Delta_8(q)^j\vartheta_3(q)^{n-8j}, \tag*{(32)} Θ L ( q ) = j = 0 ∑ K c j Δ 8 ( q ) j ϑ 3 ( q ) n − 8 j , (32)
Θ S ( q ) = ∑ j = 0 K ( − 1 ) j c j 16 j ϑ 4 ( q 2 ) 8 j ϑ 2 ( q ) n − 8 j . (33) \Theta_S(q)=\sum_{j=0}^{K}\frac{(-1)^jc_j}{16^j}\vartheta_4(q^2)^{8j}\vartheta_2(q)^{n-8j}. \tag*{(33)} Θ S ( q ) = j = 0 ∑ K 1 6 j ( − 1 ) j c j ϑ 4 ( q 2 ) 8 j ϑ 2 ( q ) n − 8 j . (33)
The theta series lies in the graded algebra generated by ϑ 3 \vartheta_3 ϑ 3 and Δ 8 \Delta_8 Δ 8 . Its weight-n / 2 n/2 n /2 component has basis Δ 8 j ϑ 3 n − 8 j \Delta_8^j\vartheta_3^{n-8j} Δ 8 j ϑ 3 n − 8 j , for 0 ≤ j ≤ K 0\leq j\leq K 0 ≤ j ≤ K . Here c j = c j ( L ) c_j=c_j(L) c j = c j ( L ) , for 0 ≤ j ≤ K 0\leq j\leq K 0 ≤ j ≤ K , denotes the uniquely determined coefficient in the theta-series expansion on this basis (6.1). In particular, c 0 = 1 c_0=1 c 0 = 1 . The same coefficients occur in the shadow expansion (6.2). Thus
P ( t ) : = Θ L ( q ) ϑ 3 ( q ) n = ∑ j = 0 K c j 64 j t j . P(t):=\frac{\Theta_L(q)}{\vartheta_3(q)^n}=\sum_{j=0}^{K}\frac{c_j}{64^j}t^j. P ( t ) := ϑ 3 ( q ) n Θ L ( q ) = j = 0 ∑ K 6 4 j c j t j .
The parameter u I u_I u I is 1 / 16 1/16 1/16 times the parameter u u u used in § 5 . Thus the theta-function identities recalled in Subsection 5.2 give
0 < t ( q ) = 4 ϑ 2 ( q ) 4 ϑ 4 ( q ) 4 ϑ 3 ( q ) 8 ≤ 1 , t ( q 0 ) = 1 , q 0 = e − π . 0<t(q)=4\frac{\vartheta_2(q)^4\vartheta_4(q)^4}{\vartheta_3(q)^8}\leq1,\quad t(q_0)=1,\quad q_0=e^{-\pi}. 0 < t ( q ) = 4 ϑ 3 ( q ) 8 ϑ 2 ( q ) 4 ϑ 4 ( q ) 4 ≤ 1 , t ( q 0 ) = 1 , q 0 = e − π .
To specify the hypergeometric identity and its branch, put
λ ( q ) = ϑ 2 ( q ) 4 ϑ 3 ( q ) 4 , 1 − λ ( q ) = ϑ 4 ( q ) 4 ϑ 3 ( q ) 4 . \lambda(q)=\frac{\vartheta_2(q)^4}{\vartheta_3(q)^4},\quad1-\lambda(q)=\frac{\vartheta_4(q)^4}{\vartheta_3(q)^4}. λ ( q ) = ϑ 3 ( q ) 4 ϑ 2 ( q ) 4 , 1 − λ ( q ) = ϑ 3 ( q ) 4 ϑ 4 ( q ) 4 .
Then t ( q ) = 4 λ ( q ) ( 1 − λ ( q ) ) t(q)=4\lambda(q)(1-\lambda(q)) t ( q ) = 4 λ ( q ) ( 1 − λ ( q )) . The classical quadratic transformation [17 Chap. 15] , with a = b = 1 / 2 a=b=1/2 a = b = 1/2 , reads
2 F 1 ( 1 2 , 1 2 ; 1 ; λ ) = 2 F 1 ( 1 4 , 1 4 ; 1 ; 4 λ ( 1 − λ ) ) , Re λ < 1 2 . {}_2F_1\left(\frac{1}{2},\frac{1}{2};1;\lambda\right)= {}_2F_1\left(\frac{1}{4},\frac{1}{4};1;4\lambda(1-\lambda)\right),\quad\operatorname{Re}\lambda<\frac{1}{2}. 2 F 1 ( 2 1 , 2 1 ; 1 ; λ ) = 2 F 1 ( 4 1 , 4 1 ; 1 ; 4 λ ( 1 − λ ) ) , Re λ < 2 1 .
Also, the elliptic-integral identities [17 Chaps. 19–20] give
2 F 1 ( 1 2 , 1 2 ; 1 ; λ ( q ) ) = 2 π K ( λ ( q ) ) = ϑ 3 ( q ) 2 , {}_2F_1\left(\frac{1}{2},\frac{1}{2};1;\lambda(q)\right)=\frac{2}{\pi}K\left(\sqrt{\lambda(q)}\right)=\vartheta_3(q)^2, 2 F 1 ( 2 1 , 2 1 ; 1 ; λ ( q ) ) = π 2 K ( λ ( q ) ) = ϑ 3 ( q ) 2 ,
where K denotes the complete elliptic integral of the first kind. Consequently, on the branch at q = 0 q=0 q = 0 ,
F ( t ) : = 2 F 1 ( 1 4 , 1 4 ; 1 ; t ) , F ( t ( q ) ) = ϑ 3 ( q ) 2 . F(t):={}_2F_1\left(\frac{1}{4},\frac{1}{4};1;t\right),\quad F(t(q))=\vartheta_3(q)^2. F ( t ) := 2 F 1 ( 4 1 , 4 1 ; 1 ; t ) , F ( t ( q )) = ϑ 3 ( q ) 2 .
Only this local, equivalently formal-power-series, identity is used to determine coefficients. It is not asserted on the same branch for all 0 < q < 1 0<q<1 0 < q < 1 . Write
b j = [ t j ] F ( t ) − n / 2 , Q n , d ( t ) = ∑ j = 0 d b j t j . b_j=[t^j]F(t)^{-n/2},\qquad Q_{n,d}(t)=\sum_{j=0}^{d}b_jt^j. b j = [ t j ] F ( t ) − n /2 , Q n , d ( t ) = j = 0 ∑ d b j t j .
Suppose min ( L ) = μ \min(L)=\mu min ( L ) = μ and d = μ − 1 ≤ K d=\mu-1\le K d = μ − 1 ≤ K . The minimum-norm condition says precisely that
Θ L ( q ) = 1 + O ( q μ ) . \Theta_L(q)=1+O(q^\mu). Θ L ( q ) = 1 + O ( q μ ) .
Since F ( t ( q ) ) − n / 2 = ϑ 3 ( q ) − n F(t(q))^{-n/2}=\vartheta_3(q)^{-n} F ( t ( q ) ) − n /2 = ϑ 3 ( q ) − n , division by the unit series ϑ 3 ( q ) n \vartheta_3(q)^n ϑ 3 ( q ) n gives
P ( t ( q ) ) − F ( t ( q ) ) − n / 2 = Θ L ( q ) − 1 ϑ 3 ( q ) n = O ( q μ ) . P(t(q))-F(t(q))^{-n/2}=\frac{\Theta_L(q)-1}{\vartheta_3(q)^n}=O(q^\mu). P ( t ( q )) − F ( t ( q ) ) − n /2 = ϑ 3 ( q ) n Θ L ( q ) − 1 = O ( q μ ) .
Now Δ 8 ( q ) = q + O ( q 2 ) \Delta_8(q)=q+O(q^2) Δ 8 ( q ) = q + O ( q 2 ) and ϑ 3 ( q ) = 1 + O ( q ) \vartheta_3(q)=1+O(q) ϑ 3 ( q ) = 1 + O ( q ) , so t ( q ) = 64 q + O ( q 2 ) t(q)=64q+O(q^2) t ( q ) = 64 q + O ( q 2 ) . Hence there is a formal inverse q ( t ) = t / 64 + O ( t 2 ) q(t)=t/64+O(t^2) q ( t ) = t /64 + O ( t 2 ) , and substitution yields
P ( t ) − F ( t ) − n / 2 = O ( t μ ) . P(t)-F(t)^{-n/2}=O(t^\mu). P ( t ) − F ( t ) − n /2 = O ( t μ ) .
Comparing coefficients of t j t^j t j , using the definitions of P P P and b j b_j b j , we obtain
c j 64 j = b j ( 0 ≤ j < μ ) . \frac{c_j}{64^j}=b_j\quad(0\le j<\mu). 6 4 j c j = b j ( 0 ≤ j < μ ) .
This proves the decomposition
P ( t ) = Q n , d ( t ) + ∑ j = μ K c j 64 j t j . (34) P(t)=Q_{n,d}(t)+\sum_{j=\mu}^{K}\frac{c_j}{64^j}t^j. \tag*{(34)} P ( t ) = Q n , d ( t ) + j = μ ∑ K 6 4 j c j t j . (34)
For clarity, the vanishing coefficients are [ q j ] Θ L ( q ) [q^j]\Theta_L(q) [ q j ] Θ L ( q ) for 1 ≤ j < μ 1\le j<\mu 1 ≤ j < μ , not the basis coefficients c j c_j c j . Although Δ 8 j ϑ 3 n − 8 j = q j + O ( q j + 1 ) \Delta_8^j\vartheta_3^{n-8j}=q^j+O(q^{j+1}) Δ 8 j ϑ 3 n − 8 j = q j + O ( q j + 1 ) , the j = 0 j=0 j = 0 term is ϑ 3 n \vartheta_3^n ϑ 3 n , not the constant series 1 1 1 . For example,
ϑ 3 ( q ) n = 1 + 2 n q + 2 n ( n − 1 ) q 2 + O ( q 3 ) , Δ 8 ( q ) = q − 8 q 2 + O ( q 3 ) . \vartheta_3(q)^n=1+2nq+2n(n-1)q^2+O(q^3),\qquad\Delta_8(q)=q-8q^2+O(q^3). ϑ 3 ( q ) n = 1 + 2 n q + 2 n ( n − 1 ) q 2 + O ( q 3 ) , Δ 8 ( q ) = q − 8 q 2 + O ( q 3 ) .
Thus μ ≥ 2 \mu\ge2 μ ≥ 2 forces 2 n + c 1 = 0 2n+c_1=0 2 n + c 1 = 0 , or c 1 = − 2 n c_1=-2n c 1 = − 2 n . If μ ≥ 3 \mu\ge3 μ ≥ 3 , the coefficient of q 2 q^2 q 2 similarly gives
2 n ( n − 1 ) + ( 2 n − 24 ) c 1 + c 2 = 0 , c 2 = 2 n ( n − 23 ) . 2n(n-1)+(2n-24)c_1+c_2=0,\qquad c_2=2n(n-23). 2 n ( n − 1 ) + ( 2 n − 24 ) c 1 + c 2 = 0 , c 2 = 2 n ( n − 23 ) .
These agree with b 1 = − n / 32 b_1=-n/32 b 1 = − n /32 and b 2 = n ( n − 23 ) / 2048 b_2=n(n-23)/2048 b 2 = n ( n − 23 ) /2048 through c j = 64 j b j c_j=64^jb_j c j = 6 4 j b j .
Lemma 6. Let n , d n,d n , d be positive integers with n ≤ 12 d + 15 n \le12d+15 n ≤ 12 d + 15 . There are strictly positive numbers g 0 , … , g d − 1 g_0,\ldots,g_{d-1} g 0 , … , g d − 1 such that
− Q n , d ′ ( t ) = ∑ i = 0 d − 1 g i t i ( 1 − t ) d − 1 − i . (35) -Q'_{n,d}(t)=\sum_{i=0}^{d-1}g_i t^i(1-t)^{d-1-i}. \tag*{(35)} − Q n , d ′ ( t ) = i = 0 ∑ d − 1 g i t i ( 1 − t ) d − 1 − i . (35) Consequently Q n , d ′ ( t ) < 0 Q'_{n,d}(t)<0 Q n , d ′ ( t ) < 0 on [ 0 , 1 ] [0,1] [ 0 , 1 ] , and
1 − Q n , d ( t ) ≥ t d ( 1 − Q n , d ( 1 ) ) , 0 ≤ t ≤ 1. (36) 1-Q_{n,d}(t)\ge t^d\left(1-Q_{n,d}(1)\right),\quad0\le t\le1. \tag*{(36)} 1 − Q n , d ( t ) ≥ t d ( 1 − Q n , d ( 1 ) ) , 0 ≤ t ≤ 1. (36)
Proof. Define
g i = [ w i ] ( 1 + w ) d − 1 [ − Q n , d ′ ( w 1 + w ) ] . g_i=[w^i](1+w)^{d-1}\left[-Q'_{n,d}\left(\frac{w}{1+w}\right)\right]. g i = [ w i ] ( 1 + w ) d − 1 [ − Q n , d ′ ( 1 + w w ) ] .
Here 0 ≤ i ≤ d − 1 0\le i\le d-1 0 ≤ i ≤ d − 1 . Set H ( t ) = F ( t ) − n / 2 H(t)=F(t)^{-n/2} H ( t ) = F ( t ) − n /2 . Since H ( t ) − Q n , d ( t ) = O ( t d + 1 ) H(t)-Q_{n,d}(t)=O(t^{d+1}) H ( t ) − Q n , d ( t ) = O ( t d + 1 ) , we have H ′ ( t ) − Q n , d ′ ( t ) = O ( t d ) H'(t)-Q'_{n,d}(t)=O(t^d) H ′ ( t ) − Q n , d ′ ( t ) = O ( t d ) . Substitution of t = w / ( 1 + w ) t=w/(1+w) t = w / ( 1 + w ) and multiplication by ( 1 + w ) d − 1 (1+w)^{d-1} ( 1 + w ) d − 1 preserve order at least d d d . Thus Q n , d ′ Q'_{n,d} Q n , d ′ may be replaced by H ′ H' H ′ when computing g i g_i g i . Writing coefficient extraction as a formal residue gives
g i = − Res w = 0 ( 1 + w ) d − 1 H ′ ( w 1 + w ) w − i − 1 d w . g_i=-\mathop{\operatorname{Res}}_{w=0}(1+w)^{d-1}H'\left(\frac{w}{1+w}\right)w^{-i-1}\,dw. g i = − Res w = 0 ( 1 + w ) d − 1 H ′ ( 1 + w w ) w − i − 1 d w .
Under t = w / ( 1 + w ) t=w/(1+w) t = w / ( 1 + w ) , we have
w = t 1 − t , 1 + w = 1 1 − t , d w = d t ( 1 − t ) 2 . w=\frac{t}{1-t},\quad1+w=\frac{1}{1-t},\quad dw=\frac{dt}{(1-t)^2}. w = 1 − t t , 1 + w = 1 − t 1 , d w = ( 1 − t ) 2 d t .
In particular,
( 1 + w ) d − 1 w − i − 1 d w = t − i − 1 ( 1 − t ) i − d d t , (1+w)^{d-1}w^{-i-1}\,dw=t^{-i-1}(1-t)^{i-d}\,dt, ( 1 + w ) d − 1 w − i − 1 d w = t − i − 1 ( 1 − t ) i − d d t ,
and therefore
g i = − Res t = 0 H ′ ( t ) t − i − 1 ( 1 − t ) i − d d t . g_i=-\mathop{\operatorname{Res}}_{t=0}H'(t)t^{-i-1}(1-t)^{i-d}\,dt. g i = − Res t = 0 H ′ ( t ) t − i − 1 ( 1 − t ) i − d d t .
Next substitute t = t ( q ) t=t(q) t = t ( q ) . Both changes of variable are valid for formal residues, because their linear coefficients at zero are nonzero. The chain rule gives the differential identity
H ′ ( t ( q ) ) d t = H ′ ( t ( q ) ) t ′ ( q ) d q = d d q ( ϑ 3 ( q ) − n ) d q = − n ϑ 3 ′ ( q ) ϑ 3 ( q ) − n − 1 d q . H'(t(q))\,dt=H'(t(q))t'(q)\,dq=\frac{d}{dq}\left(\vartheta_3(q)^{-n}\right)\,dq=-n\vartheta'_3(q)\vartheta_3(q)^{-n-1}\,dq. H ′ ( t ( q )) d t = H ′ ( t ( q )) t ′ ( q ) d q = d q d ( ϑ 3 ( q ) − n ) d q = − n ϑ 3 ′ ( q ) ϑ 3 ( q ) − n − 1 d q .
Equivalently, the factor t ′ ( q ) t'(q) t ′ ( q ) from d t = t ′ ( q ) d q dt=t'(q)\,dq d t = t ′ ( q ) d q cancels the factor 1 / t ′ ( q ) 1/t'(q) 1/ t ′ ( q ) in the chain-rule expression for H ′ ( t ( q ) ) H'(t(q)) H ′ ( t ( q )) . No derivative factor is omitted. We obtain
g i = n Res q = 0 ϑ 3 ′ ( q ) ϑ 3 ( q ) − n − 1 t ( q ) − i − 1 ( 1 − t ( q ) ) i − d d q . g_i=n\mathop{\operatorname{Res}}_{q=0}\vartheta'_3(q)\vartheta_3(q)^{-n-1}t(q)^{-i-1}(1-t(q))^{i-d}\,dq. g i = n Res q = 0 ϑ 3 ′ ( q ) ϑ 3 ( q ) − n − 1 t ( q ) − i − 1 ( 1 − t ( q ) ) i − d d q .
Finally, the Jacobi product identities give
Φ ( q ) = ∏ m ≥ 1 ( 1 + q 2 m − 1 ) , u I ( q ) = q Φ ( q ) − 24 , t ( q ) = 64 q Φ ( q ) − 24 . \Phi(q)=\prod_{m\geq1}(1+q^{2m-1}),\quad u_{\mathrm{I}}(q)=q\Phi(q)^{-24},\quad t(q)=64q\Phi(q)^{-24}. Φ ( q ) = m ≥ 1 ∏ ( 1 + q 2 m − 1 ) , u I ( q ) = q Φ ( q ) − 24 , t ( q ) = 64 q Φ ( q ) − 24 .
Using Res q = 0 q − i − 1 A ( q ) d q = [ q i ] A ( q ) \operatorname{Res}_{q=0}q^{-i-1}A(q)\,dq=[q^i]A(q) Res q = 0 q − i − 1 A ( q ) d q = [ q i ] A ( q ) yields
g i = n 64 i + 1 [ q i ] ϑ 3 ′ ( q ) ϑ 3 ( q ) − n − 1 × Φ ( q ) 24 ( i + 1 ) ( 1 − t ( q ) ) − ( d − i ) . (37) g_i=\frac{n}{64^{i+1}}[q^i]\vartheta'_3(q)\vartheta_3(q)^{-n-1}\times\Phi(q)^{24(i+1)}(1-t(q))^{-(d-i)}. \tag*{(37)} g i = 6 4 i + 1 n [ q i ] ϑ 3 ′ ( q ) ϑ 3 ( q ) − n − 1 × Φ ( q ) 24 ( i + 1 ) ( 1 − t ( q ) ) − ( d − i ) . (37)
where ϑ 3 ′ ( q ) \vartheta'_3(q) ϑ 3 ′ ( q ) denotes differentiation with respect to q q q ; H ′ H' H ′ and Q n , d ′ Q'_{n,d} Q n , d ′ denote differentiation with respect to t t t .
Put k = d − i ≥ 1 k=d-i\geq1 k = d − i ≥ 1 , R = ϑ 4 4 / ϑ 3 4 R=\vartheta_4^4/\vartheta_3^4 R = ϑ 4 4 / ϑ 3 4 , and E = 16 d + 8 i + 22 − 2 n E=16d+8i+22-2n E = 16 d + 8 i + 22 − 2 n . Since
1 − t = ϑ 4 8 ϑ 3 8 ( 1 − R ) 2 , 1-t=\frac{\vartheta_4^8}{\vartheta_3^8}(1-R)^2, 1 − t = ϑ 3 8 ϑ 4 8 ( 1 − R ) 2 ,
the series inside (6.6) is
ϑ 3 ′ ( q ) ( 1 − R ( q ) ) − 2 k ∏ m ≥ 1 ( 1 − q 2 m ) − n − 1 × ∏ m ≥ 1 ( 1 + q 2 m − 1 ) E ( 1 − q 2 m − 1 ) − 16 k . (38) \vartheta'_3(q)(1-R(q))^{-2k}\prod_{m\geq1}(1-q^{2m})^{-n-1}\times\prod_{m\geq1}(1+q^{2m-1})^E(1-q^{2m-1})^{-16k}. \tag*{(38)} ϑ 3 ′ ( q ) ( 1 − R ( q ) ) − 2 k m ≥ 1 ∏ ( 1 − q 2 m ) − n − 1 × m ≥ 1 ∏ ( 1 + q 2 m − 1 ) E ( 1 − q 2 m − 1 ) − 16 k . (38)
Both ϑ 3 ′ \vartheta'_3 ϑ 3 ′ and R R R have nonnegative coefficients, and R = 16 q + O ( q 2 ) R=16q+O(q^2) R = 16 q + O ( q 2 ) . If E ≥ 0 E\geq0 E ≥ 0 , every factor in (6.7) has nonnegative coefficients. If E < 0 E<0 E < 0 , use
( 1 + z ) E ( 1 − z ) − 16 k = ( 1 − z 2 ) E ( 1 − z ) − ( 16 k + E ) . (1+z)^E(1-z)^{-16k}=(1-z^2)^E(1-z)^{-(16k+E)}. ( 1 + z ) E ( 1 − z ) − 16 k = ( 1 − z 2 ) E ( 1 − z ) − ( 16 k + E ) .
Here
16 k + E = 32 d − 8 i + 22 − 2 n ≥ 24 d + 30 − 2 n ≥ 0. 16k+E=32d-8i+22-2n\geq24d+30-2n\geq0. 16 k + E = 32 d − 8 i + 22 − 2 n ≥ 24 d + 30 − 2 n ≥ 0.
Again all coefficients are nonnegative. The constant term ϑ 3 ′ ( 0 ) = 2 \vartheta'_3(0)=2 ϑ 3 ′ ( 0 ) = 2 and the positive coefficient of every power in ( 1 − R ) − 2 k (1-R)^{-2k} ( 1 − R ) − 2 k make g i > 0 g_i>0 g i > 0 . This proves (6.4). Since g i > 0 g_i>0 g i > 0 , the right-hand side of (6.4) is positive for 0 < t < 1 0<t<1 0 < t < 1 ; at t = 0 t=0 t = 0 and t = 1 t=1 t = 1 it equals g 0 g_0 g 0 and g d − 1 g_{d-1} g d − 1 , respectively. Hence Q n , d ′ < 0 Q'_{n,d}<0 Q n , d ′ < 0 on the whole interval [ 0 , 1 ] [0,1] [ 0 , 1 ] .
We now prove (6.5) in detail. Write Q = Q n , d Q=Q_{n,d} Q = Q n , d . Since F ( 0 ) = 1 F(0)=1 F ( 0 ) = 1 , we have Q ( 0 ) = b 0 = 1 Q(0)=b_0=1 Q ( 0 ) = b 0 = 1 . Integrating (6.4) from 0 0 0 to t t t gives
1 − Q ( t ) = ∑ i = 0 d − 1 g i I i ( t ) , 1-Q(t)=\sum_{i=0}^{d-1}g_iI_i(t), 1 − Q ( t ) = i = 0 ∑ d − 1 g i I i ( t ) ,
I i ( t ) : = ∫ 0 t x i ( 1 − x ) d − 1 − i d x . I_i(t):=\int_0^t x^i(1-x)^{d-1-i}\,dx. I i ( t ) := ∫ 0 t x i ( 1 − x ) d − 1 − i d x .
Fix 0 < t < 1 0<t<1 0 < t < 1 and 0 ≤ i ≤ d − 1 0\le i\le d-1 0 ≤ i ≤ d − 1 . The substitution x = t s x=ts x = t s yields
I i ( t ) = t i + 1 ∫ 0 1 s i ( 1 − t s ) d − 1 − i d s . I_i(t)=t^{i+1}\int_0^1s^i(1-ts)^{d-1-i}\,\mathrm{d}s. I i ( t ) = t i + 1 ∫ 0 1 s i ( 1 − t s ) d − 1 − i d s .
For 0 ≤ s ≤ 1 0\le s\le1 0 ≤ s ≤ 1 , we have 1 − t s ≥ 1 − s ≥ 0 1-ts\ge1-s\ge0 1 − t s ≥ 1 − s ≥ 0 . Because d − 1 − i d-1-i d − 1 − i is a nonnegative integer, it follows that ( 1 − t s ) d − 1 − i ≥ ( 1 − s ) d − 1 − i (1-ts)^{d-1-i}\ge(1-s)^{d-1-i} ( 1 − t s ) d − 1 − i ≥ ( 1 − s ) d − 1 − i . Moreover, i + 1 ≤ d i+1\le d i + 1 ≤ d and 0 < t < 1 0<t<1 0 < t < 1 imply t i + 1 ≥ t d t^{i+1}\ge t^d t i + 1 ≥ t d . Consequently,
I i ( t ) ≥ t i + 1 ∫ 0 1 s i ( 1 − s ) d − 1 − i d s ≥ t d ∫ 0 1 s i ( 1 − s ) d − 1 − i d s = t d I i ( 1 ) . I_i(t)\ge t^{i+1}\int_0^1s^i(1-s)^{d-1-i}\,\mathrm{d}s\ge t^d\int_0^1s^i(1-s)^{d-1-i}\,\mathrm{d}s=t^d I_i(1). I i ( t ) ≥ t i + 1 ∫ 0 1 s i ( 1 − s ) d − 1 − i d s ≥ t d ∫ 0 1 s i ( 1 − s ) d − 1 − i d s = t d I i ( 1 ) .
Multiplying these inequalities by g i > 0 g_i>0 g i > 0 and summing over i i i , we obtain
1 − Q ( t ) = ∑ i = 0 d − 1 g i I i ( t ) ≥ t d ∑ i = 0 d − 1 g i I i ( 1 ) = t d ( Q ( 0 ) − Q ( 1 ) ) = t d ( 1 − Q ( 1 ) ) . 1-Q(t)=\sum_{i=0}^{d-1}g_iI_i(t)\ge t^d\sum_{i=0}^{d-1}g_iI_i(1)=t^d(Q(0)-Q(1))=t^d(1-Q(1)). 1 − Q ( t ) = i = 0 ∑ d − 1 g i I i ( t ) ≥ t d i = 0 ∑ d − 1 g i I i ( 1 ) = t d ( Q ( 0 ) − Q ( 1 )) = t d ( 1 − Q ( 1 )) .
At t = 0 t=0 t = 0 both sides are zero, and at t = 1 t=1 t = 1 equality holds. This proves (6.5) for every 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 .
Exact inversion of the shadow coefficients
Write n = 8 K + r 8 n=8K+r_8 n = 8 K + r 8 , 0 ≤ r 8 ≤ 7 0\le r_8\le7 0 ≤ r 8 ≤ 7 , and set x = q 2 x=q^2 x = q 2 . The shadow has an expansion
Θ S ( q ) = q r 8 / 4 B ( x ) , B ( x ) = ∑ a ≥ 0 B a x a , B a ≥ 0. \Theta_S(q)=q^{r_8/4}B(x),\qquad B(x)=\sum_{a\ge0}B_ax^a,\quad B_a\ge0. Θ S ( q ) = q r 8 /4 B ( x ) , B ( x ) = a ≥ 0 ∑ B a x a , B a ≥ 0.
Define the following convergent products for ∣ x ∣ < 1 |x|<1 ∣ x ∣ < 1 :
T ( x ) = ∑ a ≥ 0 x a ( a + 1 ) / 2 = ∏ m ≥ 1 1 − x 2 m 1 − x 2 m − 1 , T(x)=\sum_{a\ge0}x^{a(a+1)/2}=\prod_{m\ge1}\frac{1-x^{2m}}{1-x^{2m-1}}, T ( x ) = a ≥ 0 ∑ x a ( a + 1 ) /2 = m ≥ 1 ∏ 1 − x 2 m − 1 1 − x 2 m ,
C ( x ) = ϑ 4 ( x ) = ∏ m ≥ 1 ( 1 − x 2 m ) ( 1 − x 2 m − 1 ) 2 , C(x)=\vartheta_4(x)=\prod_{m\ge1}(1-x^{2m})(1-x^{2m-1})^2, C ( x ) = ϑ 4 ( x ) = m ≥ 1 ∏ ( 1 − x 2 m ) ( 1 − x 2 m − 1 ) 2 ,
w ( x ) = x ( T ( x ) / C ( x ) ) 8 = x ∏ m ≥ 1 ( 1 − x 2 m − 1 ) − 24 , w(x)=x\bigl(T(x)/C(x)\bigr)^8=x\prod_{m\ge1}(1-x^{2m-1})^{-24}, w ( x ) = x ( T ( x ) / C ( x ) ) 8 = x m ≥ 1 ∏ ( 1 − x 2 m − 1 ) − 24 ,
A ( x ) = x w ′ ( x ) w ( x ) = 1 + 24 ∑ a ≥ 1 a o d d a x a 1 − x a , A(x)=\frac{xw'(x)}{w(x)}=1+24\sum_{\substack{a\ge1\\a\ \mathrm{odd}}}\frac{ax^a}{1-x^a}, A ( x ) = w ( x ) x w ′ ( x ) = 1 + 24 a ≥ 1 a odd ∑ 1 − x a a x a ,
H j ( x ) = C ( x ) − 8 j T ( x ) − ( n − 8 j ) = ∏ m ≥ 1 ( 1 − x 2 m ) − n ( 1 − x 2 m − 1 ) n − 24 j . H_j(x)=C(x)^{-8j}T(x)^{-(n-8j)}
=\prod_{m\ge1}(1-x^{2m})^{-n}(1-x^{2m-1})^{n-24j}. H j ( x ) = C ( x ) − 8 j T ( x ) − ( n − 8 j ) = m ≥ 1 ∏ ( 1 − x 2 m ) − n ( 1 − x 2 m − 1 ) n − 24 j .
Lemma 7 For 0 ≤ j ≤ K 0 \le j \le K 0 ≤ j ≤ K ,
( − 1 ) j 2 n − 12 j c j = [ x K − j ] B ( x ) H j ( x ) A ( x ) . (39) (-1)^j2^{n-12j}c_j=[x^{K-j}]B(x)H_j(x)A(x). \tag*{(39)} ( − 1 ) j 2 n − 12 j c j = [ x K − j ] B ( x ) H j ( x ) A ( x ) . (39) In particular, if 24 j ≥ n 24j\ge n 24 j ≥ n , then ( − 1 ) j c j ≥ 0 (-1)^jc_j\ge0 ( − 1 ) j c j ≥ 0 and
∣ c j ∣ 64 j ≤ A 0 2 n / 4 8 − j P ( 1 ) , A 0 = A ( e − 2 π ) < 21 20 . (40) \frac{|c_j|}{64^j}\le A_0\,2^{n/4}8^{-j}P(1),\quad A_0=A(e^{-2\pi})<\frac{21}{20}. \tag*{(40)} 6 4 j ∣ c j ∣ ≤ A 0 2 n /4 8 − j P ( 1 ) , A 0 = A ( e − 2 π ) < 20 21 . (40)
Proof. Using ϑ 2 ( q ) = 2 q 1 / 4 T ( q 2 ) \vartheta_2(q)=2q^{1/4}T(q^2) ϑ 2 ( q ) = 2 q 1/4 T ( q 2 ) , equation (33) becomes
B ( x ) = 2 n − 12 K C ( x ) 8 K T ( x ) r 8 × ∑ k = 0 K ( − 1 ) K − k c K − k 4096 k w ( x ) k . B(x)=2^{n-12K}C(x)^{8K}T(x)^{r_8}\times\sum_{k=0}^{K}(-1)^{K-k}c_{K-k}4096^kw(x)^k. B ( x ) = 2 n − 12 K C ( x ) 8 K T ( x ) r 8 × k = 0 ∑ K ( − 1 ) K − k c K − k 409 6 k w ( x ) k .
The formal coefficient change of variable, with w = x f ( x ) w=xf(x) w = x f ( x ) , is
[ u k ] G ( x ( w ) ) = [ x k ] G ( x ) f ( x ) − k x w ′ ( x ) w ( x ) . [u^k]G(x(w))=[x^k]G(x)f(x)^{-k}\frac{xw'(x)}{w(x)}. [ u k ] G ( x ( w )) = [ x k ] G ( x ) f ( x ) − k w ( x ) x w ′ ( x ) .
Apply this with k = K − j k=K-j k = K − j and f = ( T / C ) 8 f=(T/C)^8 f = ( T / C ) 8 to obtain (6.8). If 24 j ≥ n 24j\ge n 24 j ≥ n , B B B , H j H_j H j , A A A all have nonnegative coefficients. We use the following elementary coefficient estimate. If G ( x ) = ∑ ℓ ≥ 0 g ℓ x ℓ G(x)=\sum_{\ell\ge0}g_\ell x^\ell G ( x ) = ∑ ℓ ≥ 0 g ℓ x ℓ has nonnegative coefficients and converges at x 0 > 0 x_0>0 x 0 > 0 , then
[ x m ] G ( x ) = g m ≤ x 0 − m G ( x 0 ) , [x^m]G(x)=g_m\le x_0^{-m}G(x_0), [ x m ] G ( x ) = g m ≤ x 0 − m G ( x 0 ) ,
because G ( x 0 ) = ∑ ℓ ≥ 0 g ℓ x 0 ℓ ≥ g m x 0 m G(x_0)=\sum_{\ell\ge0}g_\ell x_0^\ell\ge g_mx_0^m G ( x 0 ) = ∑ ℓ ≥ 0 g ℓ x 0 ℓ ≥ g m x 0 m . Apply this estimate to G ( x ) = B ( x ) H j ( x ) A ( x ) G(x)=B(x)H_j(x)A(x) G ( x ) = B ( x ) H j ( x ) A ( x ) with m = K − j m=K-j m = K − j and x 0 = q 0 2 = e − 2 π x_0=q_0^2=e^{-2\pi} x 0 = q 0 2 = e − 2 π . Using (6.8), we get
∣ c j ∣ ≤ 2 12 j − n x 0 − ( K − j ) B ( x 0 ) H j ( x 0 ) A 0 , A 0 = A ( x 0 ) . |c_j|\le2^{12j-n}x_0^{-(K-j)}B(x_0)H_j(x_0)A_0,\quad A_0=A(x_0). ∣ c j ∣ ≤ 2 12 j − n x 0 − ( K − j ) B ( x 0 ) H j ( x 0 ) A 0 , A 0 = A ( x 0 ) .
After replacing B B B , T T T , C C C by theta functions, this is
∣ c j ∣ 64 j ≤ A 0 Θ S ( q 0 ) ϑ 2 ( q 0 ) − n ( ϑ 2 ( q 0 ) 8 4 ϑ 4 ( q 0 ) 8 ) j . \frac{|c_j|}{64^j}\le A_0\Theta_S(q_0)\vartheta_2(q_0)^{-n}\left(\frac{\vartheta_2(q_0)^8}{4\vartheta_4(q_0)^8}\right)^j. 6 4 j ∣ c j ∣ ≤ A 0 Θ S ( q 0 ) ϑ 2 ( q 0 ) − n ( 4 ϑ 4 ( q 0 ) 8 ϑ 2 ( q 0 ) 8 ) j .
At q 0 = e − π q_0=e^{-\pi} q 0 = e − π , the Jacobi transformation and duplication identities give
ϑ 2 ( q 0 ) 4 = ϑ 4 ( q 0 ) 4 = 1 2 ϑ 3 ( q 0 ) 4 , ϑ 4 ( q 0 ) 2 = ϑ 3 ( q 0 ) ϑ 4 ( q 0 ) . \vartheta_2(q_0)^4=\vartheta_4(q_0)^4=\frac{1}{2}\vartheta_3(q_0)^4,\quad\vartheta_4(q_0)^2=\vartheta_3(q_0)\vartheta_4(q_0). ϑ 2 ( q 0 ) 4 = ϑ 4 ( q 0 ) 4 = 2 1 ϑ 3 ( q 0 ) 4 , ϑ 4 ( q 0 ) 2 = ϑ 3 ( q 0 ) ϑ 4 ( q 0 ) .
Thus the parenthesis is 1 / 8 1/8 1/8 and ϑ 2 ( q 0 ) − n = 2 n / 4 ϑ 3 ( q 0 ) − n \vartheta_2(q_0)^{-n}=2^{n/4}\vartheta_3(q_0)^{-n} ϑ 2 ( q 0 ) − n = 2 n /4 ϑ 3 ( q 0 ) − n . The shadow is one translate v + L v+L v + L of L L L , since [ L 0 ∗ : L ] = 2 [L_0^*:L]=2 [ L 0 ∗ : L ] = 2 and S = L 0 ∗ ∖ L S=L_0^*\setminus L S = L 0 ∗ ∖ L . We recall the Gaussian translate inequality of Micciancio and Regev [16 Lemma 2.9] , including its derivation from their Poisson summation formula [16 Lemma 2.8] . Use the Fourier-transform convention
f ^ ( z ) = ∫ R n f ( x ) e − 2 π i ( x , z ) d x . \widehat{f}(z)=\int_{\mathbb{R}^{n}}f(x)e^{-2\pi i(x,z)}\,dx. f ( z ) = ∫ R n f ( x ) e − 2 π i ( x , z ) d x .
For a Schwartz function f f f and g ( x ) = f ( x + v ) g(x)=f(x+v) g ( x ) = f ( x + v ) , the substitution r = x + v r=x+v r = x + v gives g ^ ( z ) = e 2 π i ( z , v ) f ^ ( z ) \widehat{g}(z)=e^{2\pi i(z,v)}\widehat{f}(z) g ( z ) = e 2 π i ( z , v ) f ( z ) . Applying Poisson summation to g g g on a full-rank lattice M M M yields
∑ x ∈ M + v f ( x ) = 1 covol ( M ) ∑ z ∈ M ∗ e 2 π i ( z , v ) f ^ ( z ) . \sum_{x\in M+v}f(x)=\frac{1}{\operatorname{covol}(M)}\sum_{z\in M^*}e^{2\pi i(z,v)}\widehat{f}(z). x ∈ M + v ∑ f ( x ) = covol ( M ) 1 z ∈ M ∗ ∑ e 2 π i ( z , v ) f ( z ) .
For f y ( x ) = e − π y ∥ x ∥ 2 f_y(x)=e^{-\pi y\lVert x\rVert^2} f y ( x ) = e − π y ∥ x ∥ 2 , y > 0 y>0 y > 0 , its Fourier transform is f y ^ ( z ) = y − n / 2 e − π ∥ z ∥ 2 / y \widehat{f_y}(z)=y^{-n/2}e^{-\pi\lVert z\rVert^2/y} f y ( z ) = y − n /2 e − π ∥ z ∥ 2 / y . Pairing the terms indexed by z z z and − z -z − z , and then using cos ( 2 π ( z , v ) ) ≤ 1 \cos(2\pi(z,v))\leq1 cos ( 2 π ( z , v )) ≤ 1 , gives
∑ x ∈ M + v e − π y ∥ x ∥ 2 = y − n / 2 covol ( M ) ∑ z ∈ M ∗ e − π ∥ z ∥ 2 / y cos ( 2 π ( z , v ) ) ≤ y − n / 2 covol ( M ) ∑ z ∈ M ∗ e − π ∥ z ∥ 2 / y = ∑ x ∈ M e − π y ∥ x ∥ 2 . \begin{aligned}
\sum_{x\in M+v}e^{-\pi y\lVert x\rVert^2}
&=\frac{y^{-n/2}}{\operatorname{covol}(M)}
\sum_{z\in M^*}e^{-\pi\lVert z\rVert^2/y}\cos(2\pi(z,v))\\
&\leq\frac{y^{-n/2}}{\operatorname{covol}(M)}
\sum_{z\in M^*}e^{-\pi\lVert z\rVert^2/y}\\
&=\sum_{x\in M}e^{-\pi y\lVert x\rVert^2}.
\end{aligned} x ∈ M + v ∑ e − π y ∥ x ∥ 2 = covol ( M ) y − n /2 z ∈ M ∗ ∑ e − π ∥ z ∥ 2 / y cos ( 2 π ( z , v )) ≤ covol ( M ) y − n /2 z ∈ M ∗ ∑ e − π ∥ z ∥ 2 / y = x ∈ M ∑ e − π y ∥ x ∥ 2 .
The final equality is Poisson summation with v = 0 v=0 v = 0 . All Gaussian sums here converge absolutely. In particular Θ S ( q 0 ) ≤ Θ L ( q 0 ) \Theta_S(q_0)\leq\Theta_L(q_0) Θ S ( q 0 ) ≤ Θ L ( q 0 ) , proving the main estimate. Finally x 0 < 1 / 500 x_0<1/500 x 0 < 1/500 and
A 0 ≤ 1 + 24 x 0 ( 1 − x 0 ) 3 < 1 + 24 / 500 ( 499 / 500 ) 3 < 21 20 . A_0\leq1+\frac{24x_0}{(1-x_0)^3}<1+\frac{24/500}{(499/500)^3}<\frac{21}{20}. A 0 ≤ 1 + ( 1 − x 0 ) 3 24 x 0 < 1 + ( 499/500 ) 3 24/500 < 20 21 .
The elementary bound x 0 < 1 / 500 x_0<1/500 x 0 < 1/500 follows, for example, from π > 25 / 8 \pi>25/8 π > 25/8 and ∑ a = 0 16 ( 25 / 4 ) a / a ! > 500 \sum_{a=0}^{16}(25/4)^a/a!>500 ∑ a = 0 16 ( 25/4 ) a / a ! > 500 .
A uniform bound for the additional terms
From now on let
n = 24 m + r , m ≥ 1 , 0 ≤ r ≤ 23 , μ = 2 m + 2 , d = 2 m + 1. n=24m+r,\quad m\geq1,\quad0\leq r\leq23,\quad\mu=2m+2,\quad d=2m+1. n = 24 m + r , m ≥ 1 , 0 ≤ r ≤ 23 , μ = 2 m + 2 , d = 2 m + 1.
Then n = 12 d + c n = 12d + c n = 12 d + c , where c = r − 12 ∈ { − 12 , … , 11 } c = r - 12 \in\{-12,\ldots,11\} c = r − 12 ∈ { − 12 , … , 11 } , and n ≤ 12 d + 15 n \le12d + 15 n ≤ 12 d + 15 . Moreover 24 j ≥ n 24j \ge n 24 j ≥ n for every j ≥ μ j \ge\mu j ≥ μ . Since μ \mu μ is even, Lemma 7 implies that the positive terms in the tail of (6.3) have even indices. Discard the nonpositive odd-index terms and use t j ≤ t μ t^j \le t^\mu t j ≤ t μ for 0 ≤ t ≤ 1 0 \le t \le1 0 ≤ t ≤ 1 to obtain
P ( t ) ≤ Q n , d ( t ) + κ r t μ P ( 1 ) , κ r = A 0 1 − 1 / 64 2 r / 4 − 6 . (41) P(t) \le Q_{n,d}(t) + \kappa_r t^\mu P(1), \qquad\kappa_r = \frac{A_0}{1 - 1/64}2^{r/4-6}. \tag*{(41)} P ( t ) ≤ Q n , d ( t ) + κ r t μ P ( 1 ) , κ r = 1 − 1/64 A 0 2 r /4 − 6 . (41)
This also holds when the tail is empty. In particular,
0 < κ r < 16 15 2 r / 4 − 6 < 9 10 . (42) 0 < \kappa_r < \frac{16}{15}2^{r/4-6} < \frac{9}{10}. \tag*{(42)} 0 < κ r < 15 16 2 r /4 − 6 < 10 9 . (42)
For the last inequality, use r ≤ 23 r \le23 r ≤ 23 and 2 1 / 4 > 32 / 27 2^{1/4} > 32/27 2 1/4 > 32/27 , verified by fourth powers.
Proposition 2 With the above parameters,
Q n , d ( 1 ) + κ r < 1. Q_{n,d}(1) + \kappa_r < 1. Q n , d ( 1 ) + κ r < 1.
Proof. Write Q = Q n , d Q = Q_{n,d} Q = Q n , d . By Lemma 6,
1 − Q ( 1 ) = ∑ i = 0 d − 1 g i d ( d − 1 i ) . 1 - Q(1) = \sum_{i=0}^{d-1} \frac{g_i}{d\binom{d-1}{i}}. 1 − Q ( 1 ) = i = 0 ∑ d − 1 d ( i d − 1 ) g i .
All summands are positive. We use only the first one, two, or five. If 0 ≤ r ≤ 12 0 \le r \le12 0 ≤ r ≤ 12 , then κ r < 2 / 15 \kappa_r < 2/15 κ r < 2/15 , while
1 − Q ( 1 ) ≥ g 0 d = n 32 d ≥ 12 d − 12 32 d ≥ 1 4 . 1 - Q(1) \ge\frac{g_0}{d} = \frac{n}{32d} \ge\frac{12d-12}{32d} \ge\frac{1}{4}. 1 − Q ( 1 ) ≥ d g 0 = 32 d n ≥ 32 d 12 d − 12 ≥ 4 1 .
If 13 ≤ r ≤ 19 13 \le r \le19 13 ≤ r ≤ 19 , so 1 ≤ c ≤ 7 1 \le c \le7 1 ≤ c ≤ 7 , then κ r < 1 / 2 \kappa_r < 1/2 κ r < 1/2 , and
g 0 d + g 1 d ( d − 1 ) − 1 2 = 112 d 2 + 40 d c + 20 d − c 2 − 41 c 1024 d ( d − 1 ) > 0. \frac{g_0}{d} + \frac{g_1}{d(d-1)} - \frac{1}{2} = \frac{112d^2 + 40dc + 20d - c^2 - 41c}{1024d(d-1)} > 0. d g 0 + d ( d − 1 ) g 1 − 2 1 = 1024 d ( d − 1 ) 112 d 2 + 40 d c + 20 d − c 2 − 41 c > 0.
The numerator is positive because d ≥ 3 d \ge3 d ≥ 3 and 40 d − c − 41 > 0 40d - c - 41 > 0 40 d − c − 41 > 0 . It remains to consider 20 ≤ r ≤ 23 20 \le r \le23 20 ≤ r ≤ 23 , or 8 ≤ c ≤ 11 8 \le c \le11 8 ≤ c ≤ 11 . For d = 3 d = 3 d = 3 , direct substitution gives
Table 1.
and each value is less than 1 / 10 1/10 1/10 . For odd d ≥ 5 d \ge5 d ≥ 5 , put
I 4 = ∑ i = 0 4 g i d ( d − 1 i ) . I_4=\sum_{i=0}^{4}\frac{g_i}{d\binom{d-1}{i}}. I 4 = i = 0 ∑ 4 d ( i d − 1 ) g i .
The exact polynomial calculation in Subsection 6.5 proves
I 4 > 1 − ( 5 / 8 ) 5 > 9 / 10. I_4>1-(5/8)^5>9/10. I 4 > 1 − ( 5/8 ) 5 > 9/10.
Hence Q ( 1 ) < 1 / 10 Q(1)<1/10 Q ( 1 ) < 1/10 in both cases, whereas κ r < 9 / 10 \kappa_r<9/10 κ r < 9/10 .
Exact certificate for the five-term estimate
The first coefficients of F ( t ) − n / 2 F(t)^{-n/2} F ( t ) − n /2 are
b 0 = 1 , b 1 = − n 32 , b 2 = n ( n − 23 ) 2048 , b_0=1,\quad b_1=-\frac{n}{32},\quad b_2=\frac{n(n-23)}{2048}, b 0 = 1 , b 1 = − 32 n , b 2 = 2048 n ( n − 23 ) ,
b 3 = − n ( n 2 − 69 n + 1208 ) 196608 , b_3=-\frac{n(n^2-69n+1208)}{196608}, b 3 = − 196608 n ( n 2 − 69 n + 1208 ) ,
b 4 = n ( n 3 − 138 n 2 + 6419 n − 100677 ) 25165824 , b_4=\frac{n(n^3-138n^2+6419n-100677)}{25165824}, b 4 = 25165824 n ( n 3 − 138 n 2 + 6419 n − 100677 ) ,
b 5 = − n ( n 4 − 230 n 3 + 20015 n 2 − 781225 n + 11542704 ) 4026531840 . b_5=-\frac{n(n^4-230n^3+20015n^2-781225n+11542704)}{4026531840}. b 5 = − 4026531840 n ( n 4 − 230 n 3 + 20015 n 2 − 781225 n + 11542704 ) .
For 0 ≤ i ≤ 4 0\le i\le4 0 ≤ i ≤ 4 ,
g i = − ∑ j = 1 i + 1 j b j ( d − j i + 1 − j ) . g_i=-\sum_{j=1}^{i+1}jb_j\binom{d-j}{i+1-j}. g i = − j = 1 ∑ i + 1 j b j ( i + 1 − j d − j ) .
Set x = d − 5 ≥ 0 x=d-5\ge0 x = d − 5 ≥ 0 . With n = 12 d + c n=12d+c n = 12 d + c and c = 8 , 9 , 10 , 11 c=8,9,10,11 c = 8 , 9 , 10 , 11 , exact simplification gives
I 4 − ( 1 − ( 5 / 8 ) 5 ) = R c ( x ) 33554432 d ( d − 1 ) ( d − 2 ) ( d − 3 ) ( d − 4 ) . I_4-\left(1-(5/8)^5\right)=\frac{R_c(x)}{33554432d(d-1)(d-2)(d-3)(d-4)}. I 4 − ( 1 − ( 5/8 ) 5 ) = 33554432 d ( d − 1 ) ( d − 2 ) ( d − 3 ) ( d − 4 ) R c ( x ) .
where
R 8 ( x ) = 16960000 x 4 + 162644800 x 3 + 547478960 x 2 + 756889760 x + 355317360 , R 9 ( x ) = 17760000 x 4 + 167620800 x 3 + 557975520 x 2 + 765530880 x + 357599985 , R 10 ( x ) = 18560000 x 4 + 172436800 x 3 + 567910480 x 2 + 773570720 x + 359702340 , R 11 ( x ) = 19360000 x 4 + 177092800 x 3 + 577307840 x 2 + 781046960 x + 361639185. \begin{aligned}
R_{8}(x) &= 16960000x^{4} + 162644800x^{3} + 547478960x^{2} \\
&\quad+ 756889760x + 355317360,\\
R_{9}(x) &= 17760000x^{4} + 167620800x^{3} + 557975520x^{2} \\
&\quad+ 765530880x + 357599985,\\
R_{10}(x) &= 18560000x^{4} + 172436800x^{3} + 567910480x^{2} \\
&\quad+ 773570720x + 359702340,\\
R_{11}(x) &= 19360000x^{4} + 177092800x^{3} + 577307840x^{2} \\
&\quad+ 781046960x + 361639185.
\end{aligned} R 8 ( x ) R 9 ( x ) R 10 ( x ) R 11 ( x ) = 16960000 x 4 + 162644800 x 3 + 547478960 x 2 + 756889760 x + 355317360 , = 17760000 x 4 + 167620800 x 3 + 557975520 x 2 + 765530880 x + 357599985 , = 18560000 x 4 + 172436800 x 3 + 567910480 x 2 + 773570720 x + 359702340 , = 19360000 x 4 + 177092800 x 3 + 577307840 x 2 + 781046960 x + 361639185.
Every coefficient is positive. This is an exact proof for all d ≥ 5 d \ge5 d ≥ 5 , not a check up to a finite dimension. The separate d = 3 d = 3 d = 3 values in the proof are the only small-degree exceptions needed.
Completion of the proof
For n ≥ 24 n \ge24 n ≥ 24 , equation (41) at t = 1 t = 1 t = 1 gives
( 1 − κ r ) P ( 1 ) ≤ Q ( 1 ) < 1 − κ r , (1-\kappa_r)P(1) \le Q(1) < 1-\kappa_r, ( 1 − κ r ) P ( 1 ) ≤ Q ( 1 ) < 1 − κ r ,
so P ( 1 ) < 1 P(1) < 1 P ( 1 ) < 1 . Notice that P ( 1 ) > 0 P(1) > 0 P ( 1 ) > 0 also forces Q ( 1 ) > 0 Q(1) > 0 Q ( 1 ) > 0 whenever the assumed lattice exists; no independent existence or positivity assertion about the formal truncation is needed. For every 0 < t ≤ 1 0 < t \le1 0 < t ≤ 1 , Lemma 6 and Proposition 2 give
P ( t ) ≤ Q ( t ) + κ r t d + 1 P ( 1 ) < Q ( t ) + κ r t d < Q ( t ) + t d ( 1 − Q ( 1 ) ) ≤ 1. P(t) \le Q(t) + \kappa_r t^{d+1}P(1) < Q(t) + \kappa_r t^d < Q(t) + t^d(1-Q(1)) \le1. P ( t ) ≤ Q ( t ) + κ r t d + 1 P ( 1 ) < Q ( t ) + κ r t d < Q ( t ) + t d ( 1 − Q ( 1 )) ≤ 1.
This is the desired dimension-independent estimate.
For completeness the small ranks can be treated without enumerating lattices. If n < 8 n < 8 n < 8 , equation (32) forces Θ L = ϑ 3 n \Theta_L = \vartheta_3^n Θ L = ϑ 3 n , so no lattice with minimum 2 2 2 exists. If 8 ≤ n ≤ 15 8 \le n \le15 8 ≤ n ≤ 15 and min ( L ) = 2 \min(L) = 2 min ( L ) = 2 , then
P ( t ) = 1 − n t / 32 < 1. P(t) = 1-nt/32 < 1. P ( t ) = 1 − n t /32 < 1.
For 16 ≤ n ≤ 22 16 \le n \le22 16 ≤ n ≤ 22 and min ( L ) = 2 \min(L) = 2 min ( L ) = 2 , write N 2 N_2 N 2 for the number of roots. The ADE root-system formulas imply N 2 ≤ 2 n ( n − 1 ) N_2 \le2n(n-1) N 2 ≤ 2 n ( n − 1 ) : every irreducible component has root-to-rank ratio at most 2 ( n − 1 ) 2(n-1) 2 ( n − 1 ) , using 30 ≤ 2 ( n − 1 ) 30 \le2(n-1) 30 ≤ 2 ( n − 1 ) for the exceptional components. Comparing the coefficient of q 2 q^2 q 2 gives
c 1 = − 2 n , c 2 = N 2 − 2 n ( 23 − n ) ≤ 4 n ( n − 12 ) . c_1 = -2n,\qquad c_2 = N_2 - 2n(23-n) \le4n(n-12). c 1 = − 2 n , c 2 = N 2 − 2 n ( 23 − n ) ≤ 4 n ( n − 12 ) .
Consequently
P ( t ) ≤ 1 − n 32 t + n ( n − 12 ) 1024 t 2 ≤ 1 − n ( 44 − n ) 1024 t < 1. P(t) \le1-\frac{n}{32}t+\frac{n(n-12)}{1024}t^2 \le1-\frac{n(44-n)}{1024}t<1. P ( t ) ≤ 1 − 32 n t + 1024 n ( n − 12 ) t 2 ≤ 1 − 1024 n ( 44 − n ) t < 1.
For n = 23 n=23 n = 23 and minimum 3 3 3 , the same coefficient comparison with N 2 = 0 N_2=0 N 2 = 0 gives c 2 = 0 c_2=0 c 2 = 0 , hence
P ( t ) = 1 − 23 t / 32 < 1. P(t)=1-23t/32<1. P ( t ) = 1 − 23 t /32 < 1.
This completes Theorem 5.
Remark 2 The estimate (6.10) uses the extremal minimum norm. Consequently, the argument establishes the stated inequality for extremal Type I lattices; it does not establish it for arbitrary nonextremal odd unimodular lattices.
Higher genus
Binary case
In this section we generalize the Kominers inequality to higher genus theta series. Let C ⊆ F 2 2 k C\subseteq\mathbb{F}_2^{2k} C ⊆ F 2 2 k be a binary self-dual code with generator matrix
G = [ I k ∣ A ] , A A T = I . G=[I_k\mid A],\quad AA^T=I. G = [ I k ∣ A ] , A A T = I .
For ε ∈ F 2 g \varepsilon\in\mathbb{F}_2^g ε ∈ F 2 g , define
ψ ε ( Y ) = ∑ m ∈ Z g exp [ − π 2 ( ε + 2 m ) T Y ( ε + 2 m ) ] . \psi_{\varepsilon}(Y)=\sum_{m\in\mathbb{Z}^g}\exp\left[-\frac{\pi}{2}(\varepsilon+2m)^T Y(\varepsilon+2m)\right]. ψ ε ( Y ) = m ∈ Z g ∑ exp [ − 2 π ( ε + 2 m ) T Y ( ε + 2 m ) ] .
For U ∈ F 2 g × k U\in\mathbb{F}_2^{g\times k} U ∈ F 2 g × k , let F ( U ) = ∏ j = 1 k ψ U j ( Y ) F(U)=\prod_{j=1}^k\psi_{U_j}(Y) F ( U ) = ∏ j = 1 k ψ U j ( Y ) . Then
G A 2 ( C ) ( g ) ( Y ) = ∑ U F ( U ) F ( U A ) . G_{A_2(C)}^{(g)}(Y)=\sum_U F(U)F(UA). G A 2 ( C ) ( g ) ( Y ) = U ∑ F ( U ) F ( U A ) .
By the Cauchy–Schwarz inequality,
G A 2 ( C ) ( g ) ( Y ) ≤ ∑ U F ( U ) 2 . G_{A_2(C)}^{(g)}(Y)\leq\sum_U F(U)^2. G A 2 ( C ) ( g ) ( Y ) ≤ U ∑ F ( U ) 2 .
On the other hand, the change of variables p = u + v 2 p=\frac{u+v}{2} p = 2 u + v , q = u − v 2 q=\frac{u-v}{2} q = 2 u − v gives
∑ ε ∈ F 2 g ψ ε ( Y ) 2 = Θ g ( Y ) 2 . \sum_{\varepsilon\in\mathbb{F}_2^g}\psi_{\varepsilon}(Y)^2=\Theta_g(Y)^2. ε ∈ F 2 g ∑ ψ ε ( Y ) 2 = Θ g ( Y ) 2 .
Therefore
G A 2 ( C ) ( g ) ( Y ) ≤ Θ g ( Y ) 2 k = G Z 2 k ( g ) ( Y ) (43) G_{A_2(C)}^{(g)}(Y) \le\Theta_g(Y)^{2k} = G_{\mathbb{Z}_{2k}}^{(g)}(Y) \tag*{(43)} G A 2 ( C ) ( g ) ( Y ) ≤ Θ g ( Y ) 2 k = G Z 2 k ( g ) ( Y ) (43)
for every genus g g g .
Quaternary case
For a lattice L ⊆ R n L \subseteq\mathbb{R}^n L ⊆ R n and a real symmetric positive-definite matrix Y ∈ Sym g + ( R ) Y \in\operatorname{Sym}^{+}_g(\mathbb{R}) Y ∈ Sym g + ( R ) , write
G L ( g ) ( Y ) = ∑ x 1 , … , x g ∈ L exp ( − π ∑ a , b = 1 g Y a b ⟨ x a , x b ⟩ ) . (44) G_L^{(g)}(Y) = \sum_{x_1,\ldots,x_g \in L} \exp\left(-\pi\sum_{a,b=1}^{g} Y_{ab}\langle x_a,x_b\rangle\right). \tag*{(44)} G L ( g ) ( Y ) = x 1 , … , x g ∈ L ∑ exp − π a , b = 1 ∑ g Y ab ⟨ x a , x b ⟩ . (44)
This is the genus-g g g Siegel theta series of L L L evaluated at i Y iY iY . Set
Θ g ( Y ) = ∑ z ∈ Z g exp ( − π z T Y z ) . (45) \Theta_g(Y) = \sum_{z \in\mathbb{Z}^g} \exp\left(-\pi z^{T}Yz\right). \tag*{(45)} Θ g ( Y ) = z ∈ Z g ∑ exp ( − π z T Y z ) . (45)
Then G Z n ( g ) ( Y ) = Θ g ( Y ) n G_{\mathbb{Z}_n}^{(g)}(Y) = \Theta_g(Y)^n G Z n ( g ) ( Y ) = Θ g ( Y ) n . For a self-dual code C ⊆ Z 4 n C \subseteq\mathbb{Z}_4^n C ⊆ Z 4 n , its Construction A lattice is
A 4 ( C ) = 1 2 { z ∈ Z n : z m o d 4 ∈ C } . A_4(C) = \frac{1}{2}\{z \in\mathbb{Z}^n : z \bmod4 \in C\}. A 4 ( C ) = 2 1 { z ∈ Z n : z mod 4 ∈ C } .
We use R = Res ( C ) R = \operatorname{Res}(C) R = Res ( C ) , k = dim F 2 R k = \dim_{\mathbb{F}_2} R k = dim F 2 R only as notation; throughout, R R R is doubly even and Tor ( C ) = R ⊥ \operatorname{Tor}(C) = R^\perp Tor ( C ) = R ⊥ . For 1 ≤ s ≤ k 1 \le s \le k 1 ≤ s ≤ k , the s s s th generalized Hamming weight of R R R is
d s ( R ) = min { # supp ( U ) : U ≤ R , dim U = s } , (46) d_s(R) = \min\{\#\operatorname{supp}(U) : U \le R,\ \dim U = s\}, \tag*{(46)} d s ( R ) = min { # supp ( U ) : U ≤ R , dim U = s } , (46)
where supp ( U ) = ⋃ u ∈ U supp ( u ) \operatorname{supp}(U) = \bigcup_{u \in U}\operatorname{supp}(u) supp ( U ) = ⋃ u ∈ U supp ( u ) .
We first record the theta identities needed below. Fix g ≥ 1 g \ge1 g ≥ 1 and Y > 0 Y > 0 Y > 0 , and put d = 2 g d = 2^g d = 2 g and Θ = Θ g ( Y ) \Theta= \Theta_g(Y) Θ = Θ g ( Y ) . For ε , η ∈ F 2 g \varepsilon,\eta\in\mathbb{F}_2^g ε , η ∈ F 2 g , define
B ε , η ( Y ) = ∑ z ∈ Z g ( − 1 ) η ⋅ z exp [ − π ( z + ε / 2 ) T Y ( z + ε / 2 ) ] . (47) B_{\varepsilon,\eta}(Y) = \sum_{z \in\mathbb{Z}^g} (-1)^{\eta\cdot z}\exp\left[-\pi(z+\varepsilon/2)^T Y(z+\varepsilon/2)\right]. \tag*{(47)} B ε , η ( Y ) = z ∈ Z g ∑ ( − 1 ) η ⋅ z exp [ − π ( z + ε /2 ) T Y ( z + ε /2 ) ] . (47)
Let H = F 2 g × F 2 g \mathcal{H} = \mathbb{F}_2^g \times\mathbb{F}_2^g H = F 2 g × F 2 g and write b ε , η = B ε , η / Θ b_{\varepsilon,\eta} = B_{\varepsilon,\eta}/\Theta b ε , η = B ε , η /Θ .
Choose a section σ : R → C \sigma:R\to C σ : R → C of reduction modulo 2, and write σ ( r ) = r + 2 h ( r ) \sigma(r)=r+2h(r) σ ( r ) = r + 2 h ( r ) with h ( r ) ∈ F 2 n h(r)\in\mathbb{F}_2^n h ( r ) ∈ F 2 n . Every codeword of C C C is uniquely expressible as σ ( r ) + 2 t \sigma(r)+2t σ ( r ) + 2 t , where r ∈ R r\in R r ∈ R and t ∈ R ⊥ t\in R^\perp t ∈ R ⊥ . For r = ( r 1 , … , r g ) ∈ R g r=(r_1,\ldots,r_g)\in R^g r = ( r 1 , … , r g ) ∈ R g and s = ( s 1 , … , s g ) ∈ R g s=(s_1,\ldots,s_g)\in R^g s = ( s 1 , … , s g ) ∈ R g , set
ε i = ( r 1 i , … , r g i ) , η i = ( s 1 i , … , s g i ) . \varepsilon_i=(r_{1i},\ldots,r_{gi}),\qquad\eta_i=(s_{1i},\ldots,s_{gi}). ε i = ( r 1 i , … , r g i ) , η i = ( s 1 i , … , s g i ) .
Character orthogonality applied to each torsion row gives
G A 4 ( C ) ( g ) ( Y ) = d − k ∑ r , s ∈ R g ( − 1 ) ∑ j = 1 g s j ⋅ h ( r j ) ∏ i = 1 n B ε i , η i ( Y ) . (48) G_{A_4(C)}^{(g)}(Y)=d^{-k}\sum_{\mathbf{r},\mathbf{s}\in R^g}(-1)^{\sum_{j=1}^{g}s_j\cdot h(r_j)}\prod_{i=1}^{n}B_{\varepsilon_i,\eta_i}(Y). \tag*{(48)} G A 4 ( C ) ( g ) ( Y ) = d − k r , s ∈ R g ∑ ( − 1 ) ∑ j = 1 g s j ⋅ h ( r j ) i = 1 ∏ n B ε i , η i ( Y ) . (48)
Indeed, the projector onto R ⊥ R^\perp R ⊥ is 2 − k ∑ s ∈ R ( − 1 ) s ⋅ t 2^{-k}\sum_{s\in R}(-1)^{s\cdot t} 2 − k ∑ s ∈ R ( − 1 ) s ⋅ t . After translating the torsion coordinate by h ( r ) h(r) h ( r ) , its local Fourier transform is exactly (7.5).
Lemma 8 For every g ≥ 1 g\geq1 g ≥ 1 and Y ∈ Sym g + ( R ) Y\in\operatorname{Sym}_{g}^{+}(\mathbb{R}) Y ∈ Sym g + ( R ) ,
∣ b x ∣ ≤ 1 , ∑ x ∈ H ∣ b x ∣ 4 = d . (49) |b_x|\leq1,\qquad\sum_{x\in\mathcal{H}}|b_x|^4=d. \tag*{(49)} ∣ b x ∣ ≤ 1 , x ∈ H ∑ ∣ b x ∣ 4 = d . (49)
Proof. Removing the signs in (7.5) gives
∣ B ε , η ∣ ≤ ∑ z ∈ Z g exp [ − π ( z + ε / 2 ) T Y ( z + ε / 2 ) ] ≤ Θ . |B_{\varepsilon,\eta}|\leq\sum_{z\in\mathbb{Z}^g}\exp\left[-\pi(z+\varepsilon/2)^{T}Y(z+\varepsilon/2)\right]\leq\Theta. ∣ B ε , η ∣ ≤ z ∈ Z g ∑ exp [ − π ( z + ε /2 ) T Y ( z + ε /2 ) ] ≤ Θ.
The last inequality follows from Poisson summation: the periodized Gaussian has positive Fourier coefficients and attains its maximum at an integer shift. For the quartic identity, let E = { t ∈ F 2 4 : t 1 + t 2 + t 3 + t 4 = 0 } E=\{t\in\mathbb{F}_2^4:t_1+t_2+t_3+t_4=0\} E = { t ∈ F 2 4 : t 1 + t 2 + t 3 + t 4 = 0 } and consider the self-dual code
C 4 = ⟨ ( 1 , 1 , 1 , 1 ) , 2 E ⟩ Z 4 . C_4=\langle(1,1,1,1),2E\rangle_{\mathbb{Z}_4}. C 4 = ⟨( 1 , 1 , 1 , 1 ) , 2 E ⟩ Z 4 .
Its lattice is
A 4 ( C 4 ) = D 4 ∪ ( D 4 + 1 2 ( 1 , 1 , 1 , 1 ) ) , A_4(C_4)=D_4\cup\left(D_4+\tfrac{1}{2}(1,1,1,1)\right), A 4 ( C 4 ) = D 4 ∪ ( D 4 + 2 1 ( 1 , 1 , 1 , 1 ) ) ,
where D 4 = { z ∈ Z 4 : ∑ i z i ≡ 0 ( m o d 2 ) } D_4=\{z\in\mathbb{Z}^4:\sum_i z_i\equiv0\pmod{2}\} D 4 = { z ∈ Z 4 : ∑ i z i ≡ 0 ( mod 2 )} . The rows of H / 2 H/2 H /2 , with
H = ( 1 1 1 1 1 1 − 1 − 1 1 − 1 1 − 1 1 − 1 − 1 1 ) , H H T = 4 I 4 , (50) H=\begin{pmatrix}
1&1&1&1\\
1&1&-1&-1\\
1&-1&1&-1\\
1&-1&-1&1
\end{pmatrix},\qquad HH^{T}=4I_4, \tag*{(50)} H = 1 1 1 1 1 1 − 1 − 1 1 − 1 1 − 1 1 − 1 − 1 1 , H H T = 4 I 4 , (50)
belong to this lattice and form an orthonormal basis. Since the lattice is unimodular, they generate it. Thus A 4 ( C 4 ) ≅ Z 4 A_4(C_4)\cong\mathbb{Z}^4 A 4 ( C 4 ) ≅ Z 4 . Taking the section σ ( 1 , 1 , 1 , 1 ) = ( 1 , 1 , 1 , 1 ) \sigma(1,1,1,1)=(1,1,1,1) σ ( 1 , 1 , 1 , 1 ) = ( 1 , 1 , 1 , 1 ) in (7.6) yields
d − 1 ∑ ε , η ∈ F 2 g B ε , η 4 = Θ 4 . d^{-1}\sum_{\varepsilon,\eta\in\mathbb{F}_2^g}B_{\varepsilon,\eta}^4=\Theta^4. d − 1 ε , η ∈ F 2 g ∑ B ε , η 4 = Θ 4 .
The quantities B ε , η B_{\varepsilon,\eta} B ε , η are real, so division by Θ 4 \Theta^4 Θ 4 proves the second assertion. □ \square □
Theorem 6. Let C ⊆ Z 4 n C \subseteq\mathbb{Z}_4^n C ⊆ Z 4 n be self-dual, and let R = Res ( C ) R = \operatorname{Res}(C) R = Res ( C ) have dimension k k k .
If
d s ( R ) ≥ 4 s for every 1 ≤ s ≤ k , (51) d_s(R) \geq4s \text{ for every } 1 \leq s \leq k, \tag*{(51)} d s ( R ) ≥ 4 s for every 1 ≤ s ≤ k , (51) then, for every g ≥ 1 g \geq1 g ≥ 1 and every Y ∈ Sym g + ( R ) Y \in\operatorname{Sym}_g^+(\mathbb{R}) Y ∈ Sym g + ( R ) ,
G A 4 ( C ) ( g ) ( Y ) ≤ Θ g ( Y ) n . (52) G_{A_4(C)}^{(g)}(Y) \leq\Theta_g(Y)^n. \tag*{(52)} G A 4 ( C ) ( g ) ( Y ) ≤ Θ g ( Y ) n . (52)
Proof. The assertion is immediate when k = 0 k=0 k = 0 , since then C = 2 Z 4 n C=2\mathbb{Z}_4^n C = 2 Z 4 n and A 4 ( C ) = Z n A_4(C)=\mathbb{Z}^n A 4 ( C ) = Z n . Assume k > 0 k>0 k > 0 and choose a generator matrix G R = ( v 1 , … , v n ) G_R=(v_1,\ldots,v_n) G R = ( v 1 , … , v n ) with columns v i ∈ F 2 k v_i\in\mathbb{F}_2^k v i ∈ F 2 k .
We claim that its column matroid contains four pairwise disjoint bases. For a set A A A of column indices, let V = ⟨ v i : i ∈ A ⟩ V=\langle v_i:i\in A\rangle V = ⟨ v i : i ∈ A ⟩ and put s = k − dim V s=k-\dim V s = k − dim V . If s > 0 s>0 s > 0 , the subcode { a G R : a ∈ V ⊥ } \{aG_R:a\in V^\perp\} { a G R : a ∈ V ⊥ } has dimension s s s , and its support is { i : v i ∉ V } \{i:v_i\notin V\} { i : v i ∈ / V } . Therefore
# ( { 1 , … , n } ∖ A ) ≥ # { i : v i ∉ V } ≥ d s ( R ) ≥ 4 ( k − dim V ) . \#(\{1,\ldots,n\}\setminus A)\geq\#\{i:v_i\notin V\}\geq d_s(R)\geq4(k-\dim V). # ({ 1 , … , n } ∖ A ) ≥ # { i : v i ∈ / V } ≥ d s ( R ) ≥ 4 ( k − dim V ) .
For s = 0 s=0 s = 0 the same inequality is automatic. The matroid base-packing theorem [9 ] now supplies four disjoint index sets I 1 , … , I 4 I_1,\ldots,I_4 I 1 , … , I 4 , each consisting of k k k columns that form a basis of F 2 k \mathbb{F}_2^k F 2 k .
Expressing the residue and Fourier variables in this basis rewrites (7.6) as
G A 4 ( C ) ( g ) ( Y ) Θ n = d − k ∑ x = ( x 1 , … , x k ) ∈ H k χ C ( x ) ∏ i = 1 n b v i ⋅ x , (53) \frac{G_{A_4(C)}^{(g)}(Y)}{\Theta^n}
= d^{-k}\sum_{x=(x_1,\ldots,x_k)\in\mathcal{H}^k}\chi_C(x)\prod_{i=1}^{n} b_{v_i\cdot x}, \tag*{(53)} Θ n G A 4 ( C ) ( g ) ( Y ) = d − k x = ( x 1 , … , x k ) ∈ H k ∑ χ C ( x ) i = 1 ∏ n b v i ⋅ x , (53)
where χ C ( x ) ∈ { 1 , − 1 } \chi_C(x)\in\{1,-1\} χ C ( x ) ∈ { 1 , − 1 } and v i ⋅ x = ∑ a = 1 k ( v i ) a x a ∈ H v_i\cdot x=\sum_{a=1}^{k}(v_i)_a x_a\in\mathcal{H} v i ⋅ x = ∑ a = 1 k ( v i ) a x a ∈ H . For 1 ≤ j ≤ 4 1\leq j\leq4 1 ≤ j ≤ 4 , put F j ( x ) = ∏ i ∈ I j ∣ b v i ⋅ x ∣ F_j(x)=\prod_{i\in I_j}|b_{v_i\cdot x}| F j ( x ) = ∏ i ∈ I j ∣ b v i ⋅ x ∣ . By Lemma 8 and the disjointness of these index sets,
∣ ∏ i = 1 n b v i ⋅ x ∣ ≤ ∏ j = 1 4 F j ( x ) . \left|\prod_{i=1}^{n}b_{v_i\cdot x}\right|\leq\prod_{j=1}^{4}F_j(x). i = 1 ∏ n b v i ⋅ x ≤ j = 1 ∏ 4 F j ( x ) .
Hölder’s inequality gives
∑ x ∈ H k ∏ j = 1 4 F j ( x ) ≤ ∏ j = 1 4 ( ∑ x ∈ H k F j ( x ) 4 ) 1 / 4 . \sum_{x\in\mathcal{H}^k}\prod_{j=1}^{4}F_j(x)
\leq\prod_{j=1}^{4}\left(\sum_{x\in\mathcal{H}^k}F_j(x)^4\right)^{1/4}. x ∈ H k ∑ j = 1 ∏ 4 F j ( x ) ≤ j = 1 ∏ 4 ( x ∈ H k ∑ F j ( x ) 4 ) 1/4 .
For each j j j , the map x ↦ ( v i ⋅ x ) i ∈ I j x\mapsto(v_i\cdot x)_{i\in I_j} x ↦ ( v i ⋅ x ) i ∈ I j is a bijection of H k \mathcal{H}^k H k . Consequently,
∑ x ∈ H k F j ( x ) 4 = ( ∑ y ∈ H ∣ b y ∣ 4 ) k = d k . \sum_{x\in\mathcal{H}^k}F_j(x)^4
=\left(\sum_{y\in\mathcal{H}}|b_y|^4\right)^k=d^k. x ∈ H k ∑ F j ( x ) 4 = y ∈ H ∑ ∣ b y ∣ 4 k = d k .
Thus the absolute value of the sum in (53) is at most d k d^{k} d k , which proves (52) . □
Example 2 Let m ≥ 5 m \ge5 m ≥ 5 and consider the binary first-order Reed–Muller code
R m = RM ( 1 , m ) = { ( a 0 + ∑ i = 1 m a i x i ) x ∈ F 2 m : a 0 , a 1 , … , a m ∈ F 2 } ⊆ F 2 2 m . R_m=\operatorname{RM}(1,m)=\left\{\left(a_0+\sum_{i=1}^{m}a_i x_i\right)_{x\in\mathbb{F}_2^m}:a_0,a_1,\ldots,a_m\in\mathbb{F}_2\right\}\subseteq\mathbb{F}_2^{2^m}. R m = RM ( 1 , m ) = ⎩ ⎨ ⎧ ( a 0 + i = 1 ∑ m a i x i ) x ∈ F 2 m : a 0 , a 1 , … , a m ∈ F 2 ⎭ ⎬ ⎫ ⊆ F 2 2 m .
It has dimension k = m + 1 k=m+1 k = m + 1 , and its generalized Hamming weights are
d s ( R m ) = { 2 m − 2 m − s , 1 ≤ s ≤ m , 2 m , s = m + 1. d_s(R_m)=
\begin{cases}
2^m-2^{m-s}, & 1\le s\le m,\\
2^m, & s=m+1.
\end{cases} d s ( R m ) = { 2 m − 2 m − s , 2 m , 1 ≤ s ≤ m , s = m + 1.
Indeed, the common zero set of s s s independent affine functions is either empty or has 2 m − s 2^{m-s} 2 m − s elements. For m ≥ 5 m\ge5 m ≥ 5 , these values satisfy d s ( R m ) ≥ 4 s d_s(R_m)\ge4s d s ( R m ) ≥ 4 s for all s s s . For m = 5 m=5 m = 5 , the weight hierarchy is
( d 1 , … , d 6 ) = ( 16 , 24 , 28 , 30 , 31 , 32 ) . (d_1,\ldots,d_6)=(16,24,28,30,31,32). ( d 1 , … , d 6 ) = ( 16 , 24 , 28 , 30 , 31 , 32 ) .
The condition need not hold for smaller m m m ; for instance, d 4 ( RM ( 1 , 4 ) ) = 15 < 16 d_4(\operatorname{RM}(1,4))=15<16 d 4 ( RM ( 1 , 4 )) = 15 < 16 . Consequently, by Theorem 6, every self-dual quaternary code C C C with Res ( C ) = R m \operatorname{Res}(C)=R_m Res ( C ) = R m satisfies, for all g ≥ 1 g\ge1 g ≥ 1 and Y ∈ Sym g + ( R ) Y\in\operatorname{Sym}_{g}^{+}(\mathbb{R}) Y ∈ Sym g + ( R ) ,
G A 4 ( C ) ( g ) ( Y ) ≤ Θ g ( Y ) 2 m . G_{A_4(C)}^{(g)}(Y)\le\Theta_g(Y)^{2^m}. G A 4 ( C ) ( g ) ( Y ) ≤ Θ g ( Y ) 2 m .
The quaternary code family C m , 1 = RM ( 1 , m ) + 2 RM ( m − 2 , m ) C_{m,1}=\operatorname{RM}(1,m)+2\operatorname{RM}(m-2,m) C m , 1 = RM ( 1 , m ) + 2 RM ( m − 2 , m ) was constructed by Bonnecaze et al. [4 Sec. IV-B, Proposition 7] . These codes are Type II self-dual and have residue code RM ( 1 , m ) \operatorname{RM}(1,m) RM ( 1 , m ) . Thus, for m ≥ 5 m\ge5 m ≥ 5 , they provide an explicit family to which Theorem 6 applies.
The next lemma changes the code representation while preserving the isometry class of its lattice. All vectors in its proof are row vectors.
Lemma 9 Let C ⊆ Z 4 n C\subseteq\mathbb{Z}_4^n C ⊆ Z 4 n be self-dual and k = dim Res ( C ) k=\dim\operatorname{Res}(C) k = dim Res ( C ) . If its residue contains a word of weight 4 4 4 , then there is a self-dual code C ′ ⊆ Z 4 n C'\subseteq\mathbb{Z}_4^n C ′ ⊆ Z 4 n such that
A 4 ( C ′ ) ≅ A 4 ( C ) , dim Res ( C ′ ) ≤ k − 1. A_4(C')\cong A_4(C),\quad\dim\operatorname{Res}(C')\le k-1. A 4 ( C ′ ) ≅ A 4 ( C ) , dim Res ( C ′ ) ≤ k − 1.
Proof. After a coordinate permutation, take r = ( 1 , 1 , 1 , 1 , 0 , … , 0 ) r=(1,1,1,1,0,\ldots,0) r = ( 1 , 1 , 1 , 1 , 0 , … , 0 ) in R = Res ( C ) R=\operatorname{Res}(C) R = Res ( C ) . Choose a lift c c c of r r r . Changing coordinate signs, an isometry of the lattice, allows an integer representative c = ( 1 , 1 , 1 , 1 , 2 h 5 , … , 2 h n ) c=(1,1,1,1,2h_5,\ldots,2h_n) c = ( 1 , 1 , 1 , 1 , 2 h 5 , … , 2 h n ) . Let L = A 4 ( C ) L=A_4(C) L = A 4 ( C ) and Q = diag ( H / 2 , I n − 4 ) Q=\operatorname{diag}(H/2,I_{n-4}) Q = diag ( H /2 , I n − 4 ) , with H H H as in (50) .
If z / 2 ∈ L z/2\in L z /2 ∈ L , then z m o d 2 ∈ R z\bmod2\in R z mod 2 ∈ R and R ⊆ R ⊥ R\subseteq R^\perp R ⊆ R ⊥ gives z 1 + z 2 + z 3 + z 4 ≡ 0 ( m o d 2 ) z_1+z_2+z_3+z_4\equiv0\pmod2 z 1 + z 2 + z 3 + z 4 ≡ 0 ( mod 2 ) . Since all entries of H H H are odd, z Q ∈ Z n zQ\in\mathbb{Z}^n z Q ∈ Z n . Thus L Q ⊆ 1 2 Z n LQ\subseteq\frac{1}{2}\mathbb{Z}^n L Q ⊆ 2 1 Z n . For 1 ≤ i ≤ 4 1\leq i\leq4 1 ≤ i ≤ 4 , the vector 2 e i Q T 2e_iQ^T 2 e i Q T is a row of H H H , extended by zeros. Its reduction modulo 2 is r r r , so 2 ( 2 e i Q T ) m o d 4 = 2 r ∈ C 2(2e_iQ^T)\bmod4=2r\in C 2 ( 2 e i Q T ) mod 4 = 2 r ∈ C and 2 e i Q T ∈ L 2e_iQ^T\in L 2 e i Q T ∈ L . For i > 4 i>4 i > 4 , 2 e i Q T = 2 e i ∈ L 2e_iQ^T=2e_i\in L 2 e i Q T = 2 e i ∈ L . Hence
2 Z n ⊆ L Q ⊆ 1 2 Z n . 2\mathbb{Z}^n\subseteq LQ\subseteq\frac{1}{2}\mathbb{Z}^n. 2 Z n ⊆ L Q ⊆ 2 1 Z n .
Define C ′ = ( 2 L Q ) m o d 4 C'=(2LQ)\bmod4 C ′ = ( 2 L Q ) mod 4 . These inclusions imply L Q = A 4 ( C ′ ) LQ=A_4(C') L Q = A 4 ( C ′ ) . Since L Q LQ L Q is integral and unimodular, C ′ C' C ′ is self-dual: integrality gives self-orthogonality modulo 4, and covolume 1 gives # C ′ = 2 n \#C'=2^n # C ′ = 2 n .
Extend r r r to a basis of R R R and choose corresponding lifts in C C C . The integer lattice 2 L 2L 2 L is generated by these lifts, by 2 t 2t 2 t for t ∈ R ⊥ t\in R^\perp t ∈ R ⊥ , and by 4 Z n 4\mathbb{Z}^n 4 Z n . The selected lift satisfies ( 1 , 1 , 1 , 1 ) H / 2 = ( 2 , 0 , 0 , 0 ) (1,1,1,1)H/2=(2,0,0,0) ( 1 , 1 , 1 , 1 ) H /2 = ( 2 , 0 , 0 , 0 ) , so c Q cQ c Q is even. For t ∈ R ⊥ t\in R^\perp t ∈ R ⊥ , the first four coordinates of ( 2 t ) Q (2t)Q ( 2 t ) Q are ( t 1 , t 2 , t 3 , t 4 ) H (t_1,t_2,t_3,t_4)H ( t 1 , t 2 , t 3 , t 4 ) H , which are even because t ⋅ r = 0 t\cdot r=0 t ⋅ r = 0 . The other coordinates are also even, as are all coordinates of 4 Z n Q 4\mathbb{Z}^nQ 4 Z n Q . Only the remaining k − 1 k-1 k − 1 lifts can therefore contribute to the residue of C ′ C' C ′ . This proves the dimension bound. □
Theorem 7. Let C ⊆ Z 4 n C\subseteq\mathbb{Z}_4^n C ⊆ Z 4 n be a self-dual code satisfying dim F 2 Res ( C ) ≤ 4 \dim_{\mathbb{F}_2}\operatorname{Res}(C)\leq4 dim F 2 Res ( C ) ≤ 4 . Then, for every g ≥ 1 g\geq1 g ≥ 1 and every Y ∈ Sym g + ( R ) Y\in\operatorname{Sym}_g^+(\mathbb{R}) Y ∈ Sym g + ( R ) ,
G A 4 ( C ) ( g ) ( Y ) ≤ Θ g ( Y ) n . (54) G_{A_4(C)}^{(g)}(Y)\leq\Theta_g(Y)^n. \tag*{(54)} G A 4 ( C ) ( g ) ( Y ) ≤ Θ g ( Y ) n . (54)
Proof. We argue by induction on k = dim R k=\dim R k = dim R , where R = Res ( C ) R=\operatorname{Res}(C) R = Res ( C ) . For k = 0 k=0 k = 0 , the lattice is Z n \mathbb{Z}^n Z n . If R R R contains a word of weight 4, apply Lemma 9 and the induction hypothesis. Gaussian mass is invariant under lattice isometries. We may therefore assume that every nonzero word of R R R has weight at least 8.
For an s s s -dimensional subcode U ≤ R U\leq R U ≤ R , each coordinate in its support is nonzero in exactly 2 s − 1 2^{s-1} 2 s − 1 words of U U U . Hence
# supp ( U ) = 2 1 − s ∑ 0 ≠ u ∈ U wt ( u ) ≥ 8 ( 2 − 2 1 − s ) . (55) \#\operatorname{supp}(U)=2^{1-s}\sum_{0\neq u\in U}\operatorname{wt}(u)\geq8(2-2^{1-s}). \tag*{(55)} # supp ( U ) = 2 1 − s 0 = u ∈ U ∑ wt ( u ) ≥ 8 ( 2 − 2 1 − s ) . (55)
In particular, d 1 ( R ) ≥ 8 d_1(R)\geq8 d 1 ( R ) ≥ 8 , d 2 ( R ) ≥ 12 d_2(R)\geq12 d 2 ( R ) ≥ 12 , and d 3 ( R ) ≥ 14 d_3(R)\geq14 d 3 ( R ) ≥ 14 . For k ≤ 3 k\leq3 k ≤ 3 , Theorem 6 applies directly. For k = 4 k=4 k = 4 , it also applies if $N=#\operatorname{supp}(R)\geq 16.
It remains to consider k = 4 k = 4 k = 4 and N ≤ 15 N \le15 N ≤ 15 . Equation (55) gives N ≥ 15 N \ge15 N ≥ 15 , so N = 15 N = 15 N = 15 and every nonzero residue word has weight 8. Coordinates outside the support of R R R split off copies of Z \mathbb{Z} Z : if the i i i th residue coordinate is zero, then e i ∈ R ⊥ e_i \in R^\perp e i ∈ R ⊥ , 2 e i ∈ C 2e_i \in C 2 e i ∈ C , and e i ∈ A 4 ( C ) e_i \in A_4(C) e i ∈ A 4 ( C ) ; all lattice vectors have integral i i i th coordinate. We can therefore restrict to these fifteen coordinates and multiply the final inequality by Θ n − 15 \Theta^{n-15} Θ n − 15 .
Choose a binary generator matrix for the restricted residue and let m v m_v m v denote the multiplicity of its column v ∈ F 2 4 v \in\mathbb{F}_2^4 v ∈ F 2 4 . For v ≠ 0 v \ne0 v = 0 , Walsh inversion gives
m v = − 1 8 ∑ a ∈ F 2 4 ( − 1 ) a ⋅ v wt ( a G R ) = 1. m_v = -\frac{1}{8}\sum_{a\in\mathbb{F}_2^4}(-1)^{a\cdot v}\operatorname{wt}(aG_R)=1. m v = − 8 1 a ∈ F 2 4 ∑ ( − 1 ) a ⋅ v wt ( a G R ) = 1.
Thus the residue is the binary simplex code [ 15 , 4 , 8 ] [15,4,8] [ 15 , 4 , 8 ] .
We next show that all self-dual lifts of this residue are equivalent by coordinate sign changes. Choose residue generators r 1 , … , r 4 r_1,\ldots,r_4 r 1 , … , r 4 and lifts c a = r a + 2 h a c_a = r_a + 2h_a c a = r a + 2 h a , and put M a b = r b ⋅ h a M_{ab}=r_b\cdot h_a M ab = r b ⋅ h a in F 2 \mathbb{F}_2 F 2 . Distinct generators have integer inner product 4, so self-orthogonality gives M a b = M b a M_{ab}=M_{ba} M ab = M ba . The matrix M M M determines the code: it determines each h a h_a h a modulo R ⊥ R^\perp R ⊥ , and the torsion is 2 R ⊥ 2R^\perp 2 R ⊥ . Changing the sign of a coordinate of column type v v v adds v v T vv^T v v T to M M M . All nonzero types occur, and the matrices v v T vv^T v v T span the space of symmetric 4 × 4 4 \times4 4 × 4 binary matrices. Indeed, v = e i v=e_i v = e i gives the diagonal matrices, while v = e i + e j v=e_i+e_j v = e i + e j , together with the diagonals, gives each symmetric off-diagonal matrix. Consequently every lift is sign-equivalent to
C 0 = ⟨ r 1 , r 2 , r 3 , r 4 , 2 R ⊥ ⟩ Z 4 . C_0=\langle r_1,r_2,r_3,r_4,2R^\perp\rangle_{\mathbb{Z}^4}. C 0 = ⟨ r 1 , r 2 , r 3 , r 4 , 2 R ⊥ ⟩ Z 4 .
This code is self-dual: the integer residue generators have weights 8 and pairwise inner products 4, the torsion is orthogonal to them, and # C 0 = 2 15 \#C_0=2^{15} # C 0 = 2 15 .
Order the fifteen coordinates by the nonzero column types v = 1 , … , 15 v=1,\ldots,15 v = 1 , … , 15 , where the four entries of v v v are its binary digits, starting with the least significant digit. Retain the first three coordinates and set
B 1 = { 4 , 5 , 6 , 7 } , B 2 = { 8 , 9 , 10 , 11 } , B 3 = { 12 , 13 , 14 , 15 } . B_1=\{4,5,6,7\},\quad B_2=\{8,9,10,11\},\quad B_3=\{12,13,14,15\}. B 1 = { 4 , 5 , 6 , 7 } , B 2 = { 8 , 9 , 10 , 11 } , B 3 = { 12 , 13 , 14 , 15 } .
Apply the orthogonal transformation
Q 15 = diag ( I 3 , H / 2 , H / 2 , H / 2 ) . Q_{15}=\operatorname{diag}(I_3,H/2,H/2,H/2). Q 15 = diag ( I 3 , H /2 , H /2 , H /2 ) .
Every residue word has even weight on each block B j B_j B j , so 2 A 4 ( C 0 ) Q 15 ⊆ Z 15 2A_4(C_0)Q_{15}\subseteq\mathbb{Z}^{15} 2 A 4 ( C 0 ) Q 15 ⊆ Z 15 . Also, the indicator of each B j B_j B j belongs to R ⊥ R^\perp R ⊥ . The inverse images of the new frame vectors 2 e i 2e_i 2 e i consequently belong to A 4 ( C 0 ) A_4(C_0) A 4 ( C 0 ) , exactly as in Lemma 9 . Thus A 4 ( C 0 ) Q 15 = A 4 ( C 0 ′ ) A_4(C_0)Q_{15}=A_4(C'_0) A 4 ( C 0 ) Q 15 = A 4 ( C 0 ′ ) for a self-dual code C 0 ′ C'_0 C 0 ′ .
The transformed lifts of r 3 r_3 r 3 and r 4 r_4 r 4 are even, because
supp ( r 3 ) = B 1 ∪ B 3 , supp ( r 4 ) = B 2 ∪ B 3 , \operatorname{supp}(r_3)=B_1\cup B_3,\qquad\operatorname{supp}(r_4)=B_2\cup B_3, supp ( r 3 ) = B 1 ∪ B 3 , supp ( r 4 ) = B 2 ∪ B 3 ,
and ( 1 , 1 , 1 , 1 ) H / 2 = ( 2 , 0 , 0 , 0 ) (1,1,1,1)H/2=(2,0,0,0) ( 1 , 1 , 1 , 1 ) H /2 = ( 2 , 0 , 0 , 0 ) . For t ∈ R ⊥ t\in R^\perp t ∈ R ⊥ , write τ j = ∑ i ∈ B j t i \tau_j=\sum_{i\in B_j}t_i τ j = ∑ i ∈ B j t i in F 2 \mathbb{F}_2 F 2 . Orthogonality to r 3 r_3 r 3 and r 4 r_4 r 4 gives τ 1 = τ 2 = τ 3 \tau_1=\tau_2=\tau_3 τ 1 = τ 2 = τ 3 . On B j B_j B j , the residue of ( 2 t ) Q 15 (2t)Q_{15} ( 2 t ) Q 15 is τ j ( 1 , 1 , 1 , 1 ) \tau_j(1,1,1,1) τ j ( 1 , 1 , 1 , 1 ) ; on the first three coordinates it is zero. All torsion generators therefore contribute at most one residue direction, while the transformed lifts of r 1 r_1 r 1 and r 2 r_2 r 2 contribute at most two. The images of 4 Z 15 4\mathbb{Z}^{15} 4 Z 15 are even. It follows that dim Res ( C 0 ′ ) ≤ 3 \dim\operatorname{Res}(C'_0)\leq3 dim Res ( C 0 ′ ) ≤ 3 .
The already established case k ≤ 3 k\leq3 k ≤ 3 , together with lattice isometry and the split-off integer coordinates, proves (54) . □
Conclusion
In this paper we have explored inequalities on the Gaussian mass of unimodular lattices of the type
∀ t ∈ ( 0 , ∞ ) , G L ( t ) ≤ G Z n ( t ) . \forall t\in(0,\infty),\quad G_L(t)\leq G_{\mathbb{Z}^n}(t). ∀ t ∈ ( 0 , ∞ ) , G L ( t ) ≤ G Z n ( t ) .
We have shown that they hold exactly for extremal unimodular lattices of both types I and II and on average for
all Type II lattices in a given dimension,
lattices obtained by Construction A from free self-dual Z 4 \mathbb{Z}_4 Z 4 -codes,
lattices obtained by Construction A from self-dual Z 6 \mathbb{Z}_6 Z 6 -codes.
Similar results have been derived for higher genus theta series, in particular the direct analogue of [14 ] for the binary Construction A. It seems very hard to prove the Z 4 \mathbb{Z}_4 Z 4 analogue of [14 ] (even for g = 1 g=1 g = 1 !) with the current technology. This is the main problem left open by this research.
Acknowledgement: The last author is grateful to Professor Kominers for suggesting to look at higher genus theta series.
Declaration of AI use: The authors used chatGPT (Open AI) for several technical proofs. All the AI-assisted content was reviewed, verified, and revised by the authors, who take full responsibility for the results, the arguments, and the final form of the article.