Arxiv:2107.00958V1 [Math.NT] 2 Jul 2021 Step Further and Show That Tame Lattices Also Give Rise to Generic Well-Rounded Lattices

DENSE GENERIC WELL-ROUNDED LATTICES∗

CAMILLA HOLLANTI† , GUILLERMO MANTILLA-SOLER‡ , AND NIKLAS MILLER§

Abstract. It is well-known that the densest lattice sphere packings also typically have the largest kissing numbers. The sphere packing density maximization problem is known to have a n solution among well-rounded lattices, of which the integer lattice Z is the simplest example. The integer lattice is also an example of a generic well-rounded lattice, i.e., a well-rounded lattice with a minimal kissing number. However, the integer lattice has the worst density among well-rounded lattices. In this paper, the problem of constructing explicit generic well-rounded lattices with dense sphere packings is considered. To this end, so-called tame lattices recently introduced by Damir et al. are utilized. Tame lattices came to be as a generalization of the ring of integers of certain abelian number ﬁelds. The sublattices of tame lattices constructed in this paper are shown to always result n in either a generic well-rounded lattice or the lattice An, with density ranging between that of Z and An. Further, explicit deformations of some known densest lattice packings are constructed, yielding a family of generic well-rounded lattices with densities arbitrarily close to the optimum. In addition to being an interesting mathematical problem on its own, the constructions are also motivated from a more practical point of view. Namely, the constructed lattices provide good candidates for lattice codes used in secure wireless communications.

Key words. Generic well-rounded lattices, Tame lattices, Dense lattice packings, Trace forms, Kissing number.

1. Introduction. Lattices are of broad interest both in mathematics and applications. Well-rounded lattices [18, 17] are a profound family of lattices, providing a sufficient source for various optimization problems. For instance, the known densest lattice sphere packings (e.g., the Gosset lattice E8) arise from well-rounded lattices, and it has been shown that the lattice sphere packing problem can be restricted to the space of well-rounded lattices also in general. A prominent feature of well-rounded lattices is that the set of shortest vectors span the ambient space. Previously, constructions of well-rounded lattices have been considered in, e.g.,[12,1, 15], and they also relate to the famous Minkowski and Woods conjectures [22]. One way to describe a lattice is via its theta series, which can be further related to the so-called flatness factor utilized in communications theory and smoothing parameter in cryptography. In [13], it was shown that well-rounded lattices yield promising candidates for lattice codes in wireless communications [26] and physical layer security [6]. In [2] a means to approximate theta functions was given, enabling efficient com- parison of candidate lattice codes. It was also noted in [13] that the densest packings might not provide the best candidates due to their large kissing numbers, and hence a study on generic well-rounded lattices having minimal kissing numbers was proposed. Motivated by these findings, tame lattices were introduced in [14], providing a source for constructing explicit well-rounded lattices. In the present paper, we take a arXiv:2107.00958v1 [math.NT] 2 Jul 2021 step further and show that tame lattices also give rise to generic well-rounded lattices. Moreover, we show that by distorting some of the densest lattice packings, it is possible

∗Submitted to the editors July 5, 2021. N. Miller and C. Hollanti are with the Department of Mathematics and Systems Analysis, Aalto University, Finland. G. Mantilla-Soler is with the Department of Mathematics at Universidad Konrad Lorenz, and is currently a visiting professor at Aalto University. The visiting professor funding from the department and the Aalto Science Institute is gratefully acknowledged. Data availability statement: Data sharing not applicable to this article as no datasets were generated or analysed during the current study. †Email: camilla.hollanti@aalto.fi ‡Email: [email protected] §Email: niklas.miller@aalto.fi 1 to construct lattices that simultaneously have a large minimum distance and a small kissing number. Such lattices should provide ideal solutions for reliable and secure communications. More specifically, our main contributions are the following: • In Theorem 15, we state a stronger version of Theorem 4.9 in [14], namely (r,s) that a well-rounded lattice of the form Lv1 (see Definition 12) is GWR or equivalent to An. • In Proposition 18, we prove that for the center densities it holds that δ( n) ≤ δ(L(r,s)) ≤ δ(A ). Z v1 n • In Theorem 19, we give two conditions which tell when a well-rounded lattice (r,s) n Lv1 is equivalent to either Z or An. This is particularly interesting from the applications point-of-view, as discussed in Section5. • In Theorem 24, we show that the dual of a lattice of the form L(r,s) is of the T,v1 same type. α α • In Sections 4.2 and 4.3 we introduce deformed lattices Dn and E8 , where α is a real parameter, and in Theorems 32 and 38 we prove that these lattices are GWR if they are not equivalent to Dn and E8, respectively. It is also shown that the center density can be brought arbitrarily close to the optimum. Let us next provide some mathematical motivation for the concept of tame lattices. Let p be an odd prime and let K be a number field that is Galois over Q and such that Gal(K/Q) =∼ Z/pZ. In [8, IV.8] Conner and Perlis showed that if K is tame, i.e., no rational prime has wild ramification, then there exists an integral basis for OK , which they called Lagrangian basis, such that the Gram matrix of the integral trace in such basis is a p × p matrix of the form  a −h . . . −h   .. .   −h a . .    .  . .. ..   . . . −h  −h . . . −h a The values of a and h are given by d(p − 1) + 1 d − 1 a = h = p p where d is the conductor of K, i.e., the smallest positive integer such that K is inside of the cyclotomic extension Q(ζd). Since K is tame, d is also equal to the product of the primes that ramify in K. Based in this explicit description of the trace form over OK , in [15] the authors have constructed families of well-rounded lattices that are sublattices of OK . In [14], the notion of tame lattice is introduced by axiomatizing the key properties of the integral trace over degree p Galois extensions. For instance notice that that a and h satisfy a − h(p − 1) = 1 and a, h ≥ 0. In [14] these lattices are used to develop a procedure to construct well-rounded lattices, in fact strongly well-rounded, for which the previous results on Galois degree p number fields case is just a particular example. In the present paper we explore further the construction of tame lattices. For instance, one of the key elements of the construction of strongly well-rounded lattices in [14] is the construction of a family of sublattices of the form

L(r,s) , Tv1 ,v1 2 see Definition 12 for details, for a given lattice L. Such construction is inspired by the study of the trace form over real number fields but its applications go beyond such fields. For instance we show that several lattices, including An,Dn,E8 and their duals, all arise as special cases of this construction. Furthermore we study the behavior of quantities such as the center density over such families, and exhibit examples of lattices with large center density. 1.1. Organization. The rest of the paper is organized as follows. In Section 2, we introduce the mathematical preliminaries needed in later sections and discuss application potential in secure wireless communications. In Section3 we study well- rounded sublattices of tame lattices, specifically give bounds for the center densities of such lattices, and also characterize them under some conditions. In Section4 we α α define deformed lattices E8 and Dn , which are families of GWR lattices with density approaching optimal density as α → 1. In Section5, we conclude this work and give directions for further research. 2. Preliminaries. In this section we define some basic concepts about lattices and number fields that we will need later on, as well as discuss some connections to applications. Vectors will always be column vectors, and will be denoted by boldface letters. 2.1. Lattices. Definition 1. A lattice Λ is a discrete subgroup of the additive group (Rn, +). Equivalently, it is a set ( m ) X Λ = uibi : ui ∈ Z , i=1 n where {b1, b2,..., bm} is a set of linearly independent vectors in R called a basis of Λ, and m is called the rank of Λ. If m = n, then Λ is said to be full rank. All lattices considered in this article are full rank and thus, when we speak of a lattice, we implicitly mean a full rank lattice. The non-singular square matrix M whose columns consist of the basis vectors bi, 1 ≤ i ≤ n, is called a generator matrix for the lattice Λ. The generator matrix is not unique: if M is a generator matrix for Λ and U ∈ Zn×n is a unimodular matrix, then M 0 = MU is another generator matrix for Λ. We have the following two important examples of lattices frequently occuring in this article. n Example 1. The orthogonal lattice Z is a lattice with a generator matrix In, the identity matrix. n+1 Example 2. The root lattice An is a rank n lattice in R defined by

( n+1 ) T n+1 X An := (x1, . . . , xn+1) ∈ Z : xi = 0 . i=1

Basis vectors are given by vi = ei − ei+1, for 1 ≤ i ≤ n, where the ei’s are the standard basis vectors in Rn+1. Information about the inner products between basis vectors of a lattice is encoded in the Gram matrix, which we deﬁne to be the symmetric and positive deﬁnite matrix

T T G := M M = (bi bj)1≤i,j≤n. 3 Since the generator matrix of a lattice is not unique, neither is the Gram matrix. Indeed, if M is a generator matrix for Λ and M 0 = MU is another generator matrix for some unimodular U, then the corresponding Gram matrices are related by G0 = M 0T M 0 = (MU)T MU = U T GU. It is however the case that det(G0) = det(G), motivating the deﬁnition of the determinant of a lattice, which we simply denote by det(Λ) := det(G), where G is any Gram matrix of Λ. We deﬁne the volume of a lattice Λ to be vol(Λ) := pdet(Λ). For a full rank lattice this reduces to vol(Λ) = | det(M)|. If a lattice Λ has a Gram matrix G with integral entries, then the lattice is said to be integral. In this article we are interested in certain sublattices of tame lattices. What we mean by a sublattice of a lattice Λ is a lattice Λ0 ⊆ Λ. If Λ0 and Λ have the same rank, then we have the following formula for the index of Λ0 in Λ: vol(Λ0) [Λ : Λ0] = . vol(Λ)

Two lattices Λ and Λ0 with generator matrices M and M 0, respectively, are said to be equivalent, denoted Λ ∼ Λ0, if M 0 = αBMU for some α > 0, a real orthogonal matrix B and a unimodular matrix U ∈ Zn×n. If α = 1, we say that Λ and Λ0 are congruent and denote it by Λ =∼ Λ0. Remark that two lattices are equivalent if and only if one is obtained from the other by rotation, reflection and scaling. The Gram matrices of equivalent lattices Λ and Λ0, as above, are related by G0 = α2U T GU. Consequently, the volumes of the lattices are related by vol(Λ0) = αn vol(Λ). The dual lattice in some sense is the same to a lattice as what the dual space is to a vector space. The geometry of the dual lattice is related to the primal, which allows one to deduce information about the primal lattice by studying the dual lattice. The definition is as follows. Definition 2. Let Λ ⊂ Rn be a lattice. Its dual lattice Λ∗ is defined as

∗ n Λ := {v ∈ R : hv, xi ∈ Z for any x ∈ Λ} . We have an explicit generator matrix for the dual lattice given a generator matrix for the primal lattice, given by the next proposition. Proposition 3. Let Λ ⊂ Rn be a lattice with a generator matrix M. Then Λ∗ has a generator matrix (M T )−1. Remark 1. If G is a Gram matrix for the primal lattice, then G−1 is a Gram matrix for the dual lattice. To see this, let G = M T M be the Gram matrix for the primal lattice with respect to the generator matrix M. Then

G−1 = (M T M)−1 = M −1(M T )−1 = ((M T )−1)T (M T )−1

is the Gram matrix for the dual lattice with respect to the generator matrix (M T )−1. An important lattice constant is the shortest vector length, or lattice minimum, of a lattice. Given a lattice Λ ⊂ Rn, we define it to be the norm of the shortest non-zero lattice vector, i.e., λ1(Λ) := min kxk . 06=x∈Λ n Maximizing λ1(Λ) over the set of rank n lattices Λ ⊂ R of volume 1 is in general a hard problem: this problem is equivalent to finding the densest lattice packing of spheres, and the solution is known only in dimensions 1–8 and 24. A commonly used 4 measure of sphere packing density of a lattice packing is the center density, which we define to be λ (Λ)n δ(Λ) := 1 . 2n vol(Λ) The center density is invariant under lattice equivalence, meaning that maximising the center density over the set of all lattices can be reduced to the maximisation of the center density over set of volume 1 lattices. We can see from the very definition −n n that δ(Λ) = 2 λ1(Λ) for a volume 1 lattice Λ, showing that maximising the center density is equivalent to maximising the shortest vector length. We will denote by δn the largest known center density of a lattice in dimension n. The set S(Λ) = {x ∈ Λ: kxk = λ1(Λ)} is called the set of minimal vectors of Λ and its cardinality κ(Λ) := |S(Λ)| is called the kissing number of the lattice. We will be concerned with specific types of lattices called well-rounded lattices. Definition 4. A lattice Λ ⊂ Rn (i) is called well-rounded (WR) if the R-span of S(Λ) is Rn. (ii) is called strongly well-rounded (SWR) if the Z-span of S(Λ) is Λ. (iii) has a basis of minimal vectors if the Z-span of the linearly independent set {v1,..., vn} ⊆ S(Λ) is Λ. (iv) is called generic well-rounded (GWR) if it is well-rounded and has kissing number κ(Λ) = 2n. A common example of a GWR lattice is Zn. 2.2. Algebraic number theory. A popular construction of lattices come from number theory. Here we briefly recall some facts from algebraic number theory that serve as a motivation to several our definitions and results here. For details and definitions see [23] or [25].

Let n be a positive integer and let K be a degree n number field, i.e., a finite field extension of Q of degree n. The ring of integers of K, denoted by OK is the set of elements of K such that their minimal polynomial over Q has coefficients in Z. It can be shown that OK is a ring, moreover it is a free Z-module of rank n.

Definition 5. Let TrK/Q be the trace map on the free module OK . The integral trace form of K is the integral binary quadratic form associated to the pairing

OK × OK → Z (x, y) 7→ TrK/Q(xy). The integral trace pairing on K yields to a number theoretical construction of lattices as follows: Let HomQ−alg(K, C) = {σ1, . . . , σr1 , τ1, τ1, . . . , τr2 , τr2 } be the set of complex embeddings of K, where the σi are the embeddings such that σi(K) ⊆ R and the pairs {τj, τj} are complex conjugate embeddings with images outside R. Let < and = be the real part and imaginary part function on complex numbers, respectively. Definition 6. The Minkowski embedding from K into RN is the Q-linear map N jK : K → R deﬁned by mapping x to

(σ1(x), . . . , σr1 (x), <(τ1(x)), =(τ1(x)),..., <τr2 (x), =τr2 (x)) .

It is a classical fact from the geometry of numbers that jK (OK ) is a full rank N lattice in R , hence the same is true for jK (I) for any abelian sub-group I of OK of finite index. In particular, this is the case for I a non-zero ideal of OK . If the field K 5 is totally real we have that the lattice jK (OK ) can be identified with the quadratic module OK via the integral trace pairing. More explicitly, for all x, y ∈ K

TrK/Q(xy) = hjK (x), jK (y)i. Let p be a prime. Recall that the ramiﬁcation degrees of p in K are the exponents e1, ..., eg that appear in the ideal prime factorization of p in OK

e1 eg pOK = P1 · ... · Pg .

Furthermore, we say that p is ramified if at least one of the ei’s is bigger than 1. Definition 7. Let p be a prime and let K be a number field. We say that K is tamely ramified at p, or p is tame in K, if p ramifies in K and all the ramification degrees at p are co-prime to p. A prime that ramifies and is wild in K if it is not tame. The field K is tame if there no prime p that is wild in K.

Example 3. Let ` be an odd prime and let ζ` ∈ C be a primitive root of unity. Let K be Q(ζ`) the cyclotomic field. Since ` is the only ramified prime in K, and since its ramification exponent is ` − 1, the number field K is tame. Example 4. Let K be a Galois number field of prime degree p. Then K is tame if and only if p does not ramify. This follows since for Galois number fields the ramification exponents divide the degree. For Galois number fields the notion of tameness can be described in terms of the trace over the ring of integers. For details see [23, Corollary 3 to Proposition 6.2] Theorem 8. Let K be a Galois number field. Then, K is tame if and only if the trace map TrK/Q : OK → Z is surjective. Remark 2. For non-Galois number fields tameness implies surjectivity of the trace but the converse is not necesarily true. For the first part see [23, Corollary 5 to Theorem 4.24]. For the second claim we have that the cubic field defined by x3 + 15x2 + 20x + 30, example taken from [23], is not tame however the trace map is surjective over the ring of integers of such field. For abelian number fields there is another characterization of tameness that some- times is useful. See [23, Proposition 8.1] for details. Theorem 9. Let K be an abelian number field and let d be its conductor. Then, K is tame if and only if d is square free. 2.3. Secure wireless communications. In this section, we will provide a short practical motivation for the construction of generic well-rounded lattices with large packing densities. We refer the interested reader to [26,3,6,9] and references therein for more details on reliable and secure wireless communications. When communicating over a wireless channel, reliability and security of the transmission are natural concerns. The related design criteria may vary depending on the specific application. If the communication channel can be expected to be of a good quality, measured by signal-to-noise ratio (SNR), and we assume the typical model of a Rayleigh fading channel, then reliability can be maximized by maximizing the modulation diversity L and minimum product distance of the lattice [26]. The former is defined as

L := min | {i : xi 6= 0} |, 06=x∈Λ 6 the minimum number of non-zero coordinates in a non-zero lattice vector, and the latter is deﬁned as n Y dp,min := inf |xi| 06=x∈Λ i=1 for a full diversity lattice with L = n. For a low signal quality, the minimum distance of the lattice becomes crucial. The security potential of the communication channel for its part can be measured by the deviation of the lattice Gaussian PDF from the uniform distribution on the Voronoi cell V(Λ), which is characterized by the ﬂatness factor

gn(Λ + u; σ) (1) εΛ(σ) := max − 1 , n u∈R 1/ vol(Λ) where we can maximize over Rn by periodicity. The ﬂatness factor was introduced as a wiretap information theory tool in [21], and we have the following useful equalities:

vol(Λ) −1/2σ2 −2πσ2 (2) εΛ(σ) = vol(Λ)gn(Λ; σ) − 1 = √ ΘΛ(e ) − 1 = ΘΛ∗ (e ) − 1. ( 2πσ)n It can be shown that the eavesdropper’s correct decoding probability as well as the information leakage are upper bounded by the flatness factor of the eavesdropper’s lattice, which should hence be minimized. The benefit from well-rounded lattices is now multi-fold: they contain the maxi- mizers of the minimum product distance, the densest lattice sphere packings, as well as the global minimizers of the flatness factor, which is closely related to the theta series as shown above. For more details, see [13, Sec. VI],[27, 10, 16]. As a conclusion, lattice codes should be designed such that the whole signal quality range from weak to strong can be covered at once. This goal can be achieved by full diversity lattices with large minimum product distance having dense lattice packings and small flatness factors. To this end, well-rounded (sub)lattices should be utilized [20, 19]. Nevertheless, the densest possible packing does not necessarily yield the best performance, as demonstrated in [13], but the kissing number also plays a role, and hence generic well-rounded lattices are of a particular interest. 3. Tame lattices. In this section, we define the concept of a tame lattice, which where introduced in [14], and expand on the results in [14] regarding well-rounded sublattices of a tame lattice. In particular, we prove that, under specific conditions, well-rounded sublattices of a tame lattice are generic, and further, develop bounds for the center densities of said sublattices. We end up with two conditions which tell when n a sublattice is equivalent to either Z or An. We also give an explicit construction of An as a sublattice of a tame lattice, and further, show that the dual lattice of a tame lattice is tame, and that the dual lattice of a certain sublattice is actually a scaled variant of its tame superlattice. The term Lagrangian basis was used by Conner and Perlis [8, p. 193] to describe a normal integral basis of a Galois, tame number field of odd prime degree, such that the Gram matrix of the lattice obtained from this basis via the Minkowski embedding has a certain form. Later, it has been proven that such bases exist in a larger set of number fields. Particularly, when the number field is tame and has a prime conductor, such a basis has been shown to exist [5]. Here we will be interested in the tame lattices themselves, rather than the number fields they originate from. The following definition is given in [14]. 7 Definition 10. Let n ≥ 1 be an integer and L ⊂ Rn a lattice. We call the lattice ∗ L tame if there exists a basis {e1,..., en} of L and a non-zero vector v1 ∈ L ∩ L such that Pn 1. i=1 ei = v1 2. hei, v1i = 1 for all 1 ≤ i ≤ n 3. hei, eii = a for all 1 ≤ i ≤ n 4. hei, eji = −h for all 1 ≤ i 6= j ≤ n. In this case, we call {e1,..., en} a Lagrangian basis for L. Conditions 3 and 4 give an explicit expression for the Gram matrix of a tame lattice with respect to the basis {e1,..., en}:

 a −h . . . −h   .. .   −h a . .  (3) G =   .  . .. ..   . . . −h  −h . . . −h a

Moreover, the conditions imply that a − h(n − 1) = 1. Conversely, a basis with a Gram matrix of the form (3) such that a − h(n − 1) = 1 holds, is Lagrangian. We can express the volume of a tame lattice using (3). n Lemma 11. Let L ⊂ R be a tame lattice with a Lagrangian basis {e1,..., en}, a := he1, e1i and h := − he1, e2i. Then

n−1 vol(L) = (a + h) 2 .

Proof. The tame lattice L has a Gram matrix of the form (3) with respect to the basis {e1,..., en}. This matrix has the well-known determinant

det(G) = (a + h)n − nh(a + h)n−1 = (a + h)n−1(a + h(1 − n)).

We have a + h(1 − n) = 1 and thus,

p p n−1 vol(L) = det(G) = (a + h)n−1 = (a + h) 2 ,

as desired. Remark 3. If L ⊂ Rn is a tame lattice with a Gram matrix G, then since G is a positive definite matrix, we have det(G) = (a + h)n−1 > 0 and in particular, the 1 1 inequality a + h > 0 and equation a + h(1 − n) = 1 imply a > n and h > − n . 3.1. Well-rounded sublattices. In this section, we present some of the definitions and results in [14], since our goal is to expand on these results. We recall how one can produce full rank sublattices of lattices, and particularly tame lattices, via a specific linear map. We also restate the main theorem presented in [14]. Definition 12. Let L ⊂ Rn be a lattice and T: L → Z a non-trivial linear map. Let r, s be integers, v1 ∈ L \ ker T and m := r + s T(v1). Define Φ(r,s) : L → L to be the linear map x 7→ rx + s T(x)v1. Define the lattice L(r,s) to be the image of L under the map Φ : T,v1 (r,s)

L(r,s) := Φ (L). T,v1 (r,s) 8 It is clear that L(r,s) is a sublattice of L. Further, if r and m are non-zero, T,v1 then Φ(r,s) turns out to be an injection. This can be achieved for instance when 0 < |r| < | T(v1)|, whence we get the following result (Corollary 3.2 in [14]).

Lemma 13. Assume that L, T and v1 are as in Deﬁnition 12 and r, s are integers such that 0 < |r| < | T(v )|. Then L(r,s) is a full rank sublattice of L. 1 T,v1

If the conditions in the above lemma are satisﬁed, then if {e1,..., en} is a basis for L, then Φ (e ),..., Φ (e ) is a basis for L(r,s) . We will be particularly (r,s) 1 (r,s) n T,v1 interested in the following linear map:

Tv1 : L → Z, Tv1 (x) = hx, v1i ,

∗ where 0 6= v1 ∈ L ∩ L . We deﬁne

(r,s) (r,s) Lv := L . 1 Tv1 ,v1

(r,s) In [14], the authors give a condition which tells when Lv1 is well-rounded, and (r,s) in fact, possesses a minimal basis. If the condition is met, then the index [Lv1 : L] (r,s) and the lattice minimum of Lv1 are given. More speciﬁcally: Theorem 14. [14, Theorem 4.9] Let n ≥ 2 be an integer and L ⊂ Rn a tame lattice with a Lagrangian basis {e1,..., en}. Let a := he1, e1i and h := − he1, e2i. Let r, s be integers such that 0 6= |r| < n and let m = r + sn. Suppose that

na − 1 m2 (na − 1)(n + 1) (4) ≤ ≤ . n2 − 1 r n − 1

(r,s) n−1 Then Lv1 is a full rank sublattice of L of index |mr |, with minimum

m2 − r2 λ2(L(r,s)) = ar2 + 1 v1 n

and a basis of minimal vectors {re1 + sv1, . . . , ren + sv1}. (r,s) In fact a stronger result is true: if the upper bound in (4) is strict, then Lv1 is GWR. We omit the proof here since it is a straightforward analogy of the proof of Theorem 4.9 in [14]. Theorem 15. Let n ≥ 2 be an integer and let L ⊂ Rn be a tame lattice with a Lagrangian basis {e1,..., en}. Let a := he1, e1i and h := − he1, e2i. Let r, s be integers such that 0 6= |r| < n and let m = r + sn. Suppose that

na − 1 m2 (na − 1)(n + 1) ≤ < . n2 − 1 r n − 1

(r,s) Then Lv1 is GWR. An interesting observation is that, when the upper bound in (4) is achieved, then (r,s) (r,s) Lv1 has the same kissing number as An, which is n(n + 1). In particular, Lv1 is (r,s) no longer GWR. Later (in Theorem 19) we will show that Lv1 is actually equivalent to An in this case. 9 3.2. Center densities of sublattices of tame lattices. Our goal in this sec- (r,s) tion is to characterize the center densities of the lattices Lv1 in (4) and develop bounds for them. We will see that the center densities are always between δ(Zn) and δ(An). The following proposition gives an explicit formula for the center density. Proposition 16. Let n ≥ 2 be an integer and let L ⊂ Rn be a tame lattice with a Lagrangian basis {e1,..., en}. Let a := he1, e1i and h := − he1, e2i. Let r, s be integers such that 0 6= |r| < n and let m = r + sn. Suppose that na − 1 m2 (na − 1)(n + 1) ≤ ≤ . n2 − 1 r n − 1 (r,s) Then the center density of Lv1 is given by ((na − 1)r2 + m2)n/2 δ(L(r,s)) = . v1 n−1 2nnn/2(a + h) 2 |mrn−1| Proof. Note that by Lemma 11 and Theorem 14,

(r,s) (r,s) n−1 n−1 vol(L ) = vol(L)[L : L ] = (a + h) 2 |mr |. v1 v1 Using the deﬁnition of center density and the expression for the shortest vector length (r,s) of Lv1 , we get

2 (r,s) n/2 (r,s) λ1(Lv1 ) δ(Lv ) = 1 n (r,s) 2 vol(Lv1 ) n/2 2 m2−r2 ar + n = n−1 2n(a + h) 2 |mrn−1| ((na − 1)r2 + m2)n/2 = n−1 2nnn/2(a + h) 2 |mrn−1| as desired. The following lemma will turn out useful when we maximize and minimize the expression for the center density derived in the previous proposition. 1 Lemma 17. Let n ≥ 2 be an integer and let a > n . Deﬁne na − 1 l := , n2 − 1 (na − 1)(n + 1) u := . n − 1 Then the real-valued function f :[l, u] → R deﬁned by (na − 1 + x)n f(x) = x has a maximum point at x = u and the maximum value is f(u) = 2nnn(na − 1)n−1(n − 1)1−n(n + 1)−1.

na−1 Further, f has a minimum point at x = x0 := n−1 and the minimum value is n n−1 1−n f(x0) = n (na − 1) (n − 1) . 10 Proof. As a diﬀerentiable function, f achieves its extreme values at the endpoints of [l, u] or at a point where f 0 vanishes. A simple calculation shows that • f(l) = n2n(na − 1)n−1(n2 − 1)1−n • f(u) = 2nnn(na − 1)n−1(n − 1)1−n(n + 1)−1 n−1 0 (na+x−1) ((n−1)x−na+1) 0 • f (x) = x2 and thus f (x) = 0 if and only if na − 1 x = 1 − na or x = . n − 1

The ﬁrst equality is impossible since na > 1 by assumption, and negative solutions are not allowed since l > 0. This leaves us with a single zero for the derivative, na−1 x0 = n−1 ∈ [l, u]. The value of the function at this point is

n n−1 1−n f(x0) = n (na − 1) (n − 1) .

Note that f(u) 2n 1 n 22 12 = 1 + ≥ 1 + = 1, f(l) (n + 1)2 n (2 + 1)2 2 when n ≥ 2. Similarly,

f(l) 4 = nn(n + 1)1−n ≥ 22(2 + 1)1−2 = > 1, f(x0) 3

when n ≥ 2. Since f is positive, it follows that

f(x0) < f(l) ≤ f(u),

proving the lemma. (r,s) We now have the ingredients to ﬁnd bounds for δ(Lv1 ). Proposition 18. Let n ≥ 2 be an integer and let L ⊂ Rn be a tame lattice with a Lagrangian basis {e1,..., en}. Let a := he1, e1i and h := − he1, e2i. Let r, s be integers such that 0 6= |r| < n and let m = r + sn. Suppose that

na − 1 m2 (na − 1)(n + 1) ≤ ≤ . n2 − 1 r n − 1

Then 1 1 ≤ δ(L(r,s)) ≤ √ . 2n v1 2n/2 n + 1 Further, the lower bound is achieved when

m2 na − 1 = r n − 1

and the upper bound is achieved when

m2 (na − 1)(n + 1) = . r n − 1

11 (r,s) Proof. Proposition 16 gives the center density of Lv1 as

((na − 1)r2 + m2)n/2 δ(L(r,s)) = v1 n−1 2nnn/2(a + h) 2 |mrn−1| n/2 m 2 na − 1 + r = n−1 . n n/2 2 m 2 n (a + h) r The condition for q := (m/r)2 is satisﬁed when q ∈ [l, u], where l and u are deﬁned as in Lemma 17. Note that we have

pf (q) δ(L(r,s)) = , v1 n−1 2nnn/2(a + h) 2

where f :[l, u] → R is the function deﬁned in Lemma 17. It follows from the previous na−1 lemma, where x0 = n−1 , that

pf(x ) δ(L(r,s)) ≥ 0 v1 n−1 2nnn/2(a + h) 2 p f(x0) = n−1 n n/2 a−1 2 2 n (a + n−1 ) n/2 n−1 1−n n (na − 1) 2 (n − 1) 2 = n−1 n n/2 a−1 2 2 n (a + n−1 ) n/2 n−1 1−n n (na − 1) 2 (n − 1) 2 = n−1 1−n 2nnn/2(na − 1) 2 (n − 1) 2 1 = , 2n and

pf(u) δ(L(r,s)) ≤ v1 n−1 2nnn/2(a + h) 2 pf(u) = n−1 n n/2 a−1 2 2 n (a + n−1 ) n/2 n/2 n−1 1−n −1/2 2 n (na − 1) 2 (n − 1) 2 (n + 1) = n−1 1−n 2nnn/2(na − 1) 2 (n − 1) 2 1 = √ , 2n/2 n + 1 proving the claim. 4. Since δ(A ) = √1 and δ( n) = 2−n, the previous proposition Remark n 2n/2 n+1 Z says that δ( n) ≤ δ(L(r,s)) ≤ δ(A ). Z v1 n (r,s) To complete our analysis, we give a characterization of Lv1 when the bounds of the previous proposition are achieved. 12 Theorem 19. Let n ≥ 2 be an integer and let L ⊂ Rn be a tame lattice with a Lagrangian basis {e1,..., en}. Let a := he1, e1i and h := − he1, e2i. Let r, s be integers such that 0 6= |r| < n and let m = r + sn. If (a) m2 na − 1 = , r n − 1 q (r,s) ∼ na−1 n then Lv1 = |r| n−1 Z . (b) m2 (na − 1)(n + 1) = , r n − 1 (r,s) ∼ q na−1 then Lv1 = |r| n−1 An. Proof. Theorem 14 gives a basis of minimal vectors B = Φ(r,s)(ei) : 1 ≤ i ≤ n = {re1 + sv1, . . . , ren + sv1}

(r,s) for the lattice Lv1 . A direct computation shows that 2 m2−r2 1. Φ(r,s)(ei), Φ(r,s)(ei) = ar + n for all 1 ≤ i ≤ n. 2 m2−r2 2. Φ(r,s)(ei), Φ(r,s)(ej) = −r h + n for all 1 ≤ i 6= j ≤ n. 2 2 na−1 For (a), using the equations a + h(1 − n) = 1 and m = r n−1 we end up with

( (na−1)r2 n−1 if i = j Φ(r,s)(ei), Φ(r,s)(ej) = 0 if i 6= j.

(r,s) As a consequence, the Gram matrix of the lattice Lv1 with respect to the basis B is given by (na − 1)r2 G = I . B n − 1 n q (r,s) ∼ na−1 n Therefore, Lv1 = |r| n−1 Z proving (a). For (b), using the equations a + h(1 − 2 2 (na−1)(n+1) n) = 1 and m = r n−1 we end up with

( 2(na−1)r2 n−1 if i = j Φ(r,s)(ei), Φ(r,s)(ej) = (na−1)r2 n−1 if i 6= j. (r,s) As a consequence, the Gram matrix for the lattice Lv1 with respect to the basis B is given by  2 1 ... 1  . 2  .. .  2 (na − 1)r  1 2 . .  (na − 1)r GB =   =: A. n − 1  . .. ..  n − 1  . . . 1  1 ... 1 2 Deﬁne the unimodular matrix  1 0 ... 0   .. .   −1 1 . .  U :=   .  . .. ..   . . . 0  0 ... −1 1 13 Note that n − 1 UG U T = UAU T = G (na − 1)r2 B where  2 −1 ... 0   .. .   −1 2 . .  G =    . .. ..   . . . −1  0 ... −1 2

(r,s) ∼ q na−1 is a Gram matrix for An. Therefore, Lv1 = |r| n−1 An proving (b).

(r,s) 3.3. Construction of An. We know that a lattice of the form Lv1 is equivalent to An if condition (b) in Theorem 19 holds. But the question remains, for which values of the parameter a for the tame superlattice does there exist a pair of integers (r, s) such that the condition is satisfied. If we take the orthogonal lattice, with a = 1, the condition can only be satisfied when n + 1 is a square. But if we consider a tame (r,s) lattice with a = n, then we can always find pairs (r, s) such that Lv1 is equivalent to An. Propositions 20 and 21 show how one can construct a lattice equivalent to the n An lattice as a sublattice of Z in the case that n + 1 is a square, and as a sublattice of a specific tame lattice for a general n. Proposition 20. Suppose that n + 1 = d2 for some integer d > 2. Let L = Zn be the tame lattice with (a, h) = (1, 0) and a Lagrangian basis {e1,..., en}, where the n (d+1,1) ei’s are the standard basis vectors in R . Then Lv1 is a sublattice of L such that (d+1,1) ∼ Lv1 = (d + 1)An. Proof. Let (r, s) = (d + 1, 1), and note that 0 < |r| = d + 1 < (d + 1)(d − 1) = n. Moreover,

m2 r + sn2 n2 (na − 1)(n + 1) = = 1 + = d2 = n + 1 = . r r r n − 1

By Theorem 19, part (b), L(r,s) ∼ (d + 1)A . v1 = n

Proposition 21. Let n ≥ 2 be an integer and let e ,..., e be the standard basis 1 n √ n T n 0 1− n+1 vectors in R . Let v1 = (1,..., 1) ∈ R and deﬁne the vectors ei = v1 + √ n n 0 0 n + 1ei for each i = 1, . . . , n. Then the lattice L ⊂ R with the basis {e1,..., en} is tame with (a, h) = (n, 1). Further, for any integer r such that 0 < |r| < n, we have √ (r,r) (r,r) ∼ that Lv1 is a sublattice of L such that Lv1 = |r| n + 1An. Proof. First we need to check that the vectors e0 form a Lagrangian basis. √ √ i Pn 0 1− n+1 1. i=1 ei = n n v1 + n + 1v1 = v1. √ √ 0 1− n+1 2. hei, v1i = n n + n + 1 = 1 for all 1 ≤ i ≤ n. √ 2 √ √ 2 0 0 1− n+1 1− n+1 3. a = hei, eii = (n − 1) n + n + n + 1 = n for all 1 ≤ i ≤ n. √ 2 √ √ √ 0 0 1− n+1 1− n+1 1− n+1 4. −h = ei, ej = (n−2) n +2 n n + n + 1 = −1 for all 1 ≤ i 6= j ≤ n. 14 The conditions are satisﬁed, so L is tame with (a, h) = (n, 1). For the second part, suppose that r is an integer such that 0 < |r| < n. Let s := r, so that m = r(1 + n), and note that m2 (n2 − 1)(n + 1) (na − 1)(n + 1) = (1 + n)2 = = . r n − 1 n − 1 By Theorem 19, part (b),

r √ (r,r) ∼ na − 1 L = |r| An = |r| n + 1An v1 n − 1 and we are done. 3.4. Dual lattices. In this section, we show that the dual of a lattice of the (r,s) form Lv1 is a scaled version of L (Proposition 22), and that the dual of a tame lattice is also tame (Lemma 23). The most general result is Theorem 24, which says that the dual of a lattice of the form L(r,s) is of the same type. T,v1 Proposition 22. Let n ≥ 2 be an integer and let L ⊂ Rn be a tame lattice with a Lagrangian basis {e1,..., en}. Let a := he1, e1i and h := − he1, e2i. Let r, s be (r,s) integers such that 0 6= |r| < n. Suppose that s = rh. Then the dual lattice of Lv1 is 1 r(a+h) L. (r,s) Proof. A basis for Lv1 is given by the vectors rei + sv1, for i = 1, . . . , n. Let 1 1 c := r(a+h) . Then a basis for r(a+h) L is given by the vectors ce1, . . . , cen. Note that for every i, j ∈ {1, . . . , n}, ( c(ra + s), if i = j hrei + sv1, ceji = c(s − rh), if i 6= j,

where we used the fact that hv1, eji = 1 for all 1 ≤ j ≤ n, and that hei, eji = a if i = j and hei, eji = −h else. The assumption s = rh and c 6= 0 give that

hrei + sv1, ceji = c(ra + s)δij = cr(a + h)δij = δij

0 (r,s) for all 1 ≤ i 6= j ≤ n, proving that the generator matrices M and M of Lv1 and 1 T 0 r(a+h) L, respectively, are related by M M = In. The claim follows from Proposition 3. Lemma 23. Let n ≥ 2 be an integer and let L ⊂ Rn be a tame lattice with a Lagrangian basis {e1,..., en}. Let a := he1, e1i and h := − he1, e2i. Then the ∗ 0 0 dual lattice L is a tame lattice with a Lagrangian basis {e1,..., en}, such that a˜ := 0 0 1+h ˜ 0 0 −h he1, e1i = a+h and h := − he1, e2i = a+h . Proof. First note that

1 + h −h 1 + hn a + h a˜ − (n − 1)h˜ = − (n − 1) = = = 1, a + h a + h a + h a + h since a − h(n − 1) = 1, i.e., 1 + hn = a + h, by the deﬁnition of a tame lattice. Furthermore, 1 + h −h 1 a˜ + h˜ = + = > 0 a + h a + h a + h 15 since a + h > 0 by Remark3. If we can show that L∗ has a Gram matrix G˜ of the form  a˜ −h˜ . . . −h˜   .. .   −h˜ a˜ . .  G˜ =   ,  . .. ..   . . . −h˜  −h˜ . . . −h˜ a˜

0 0 ∗ with respect to some basis {e1,..., en}, then we have shown that L is a tame 0 0 0 0 1+h lattice with a Lagrangian basis {e1,..., en} such thata ˜ = he1, e1i = a+h and ˜ 0 0 −h h = − he1, e2i = a+h . Denote by {b1,..., bn} the rows of the Gram matrix G of the tame lattice L, and by {d1,..., dn} the columns of the matrix G˜. Let us compute the inner products hbi, dji for all i, j ∈ {1, . . . , n}. For i = j, we obtain

T ˜ ˜ hbi, dii = (−h, . . . , a, . . . , −h) (−h, . . . , a,˜ . . . , −h) = (n − 1)hh˜ + aa˜ = (n − 1)hh˜ + a(1 + (n − 1)h˜) = a + (n − 1)h˜(a + h) −h = a + (n − 1) (a + h) a + h = a − h(n − 1) = 1.

For i 6= j, we have

T ˜ ˜ ˜ hbi, dji = (−h, . . . , a, −h . . . , −h) (−h, . . . , −h, a,˜ . . . , −h) = (n − 2)hh˜ − ah˜ − ah˜ = h˜((n − 2)h − a) − ah˜ = −h˜(a − h(n − 1) + h) − ah˜ = −h˜(1 + h) − ah˜ h h(1 + h) = (1 + h) − a + h a + h = 0.

This shows that G and G˜ are related by GG˜ = In. But by Remark1, G˜ must be a Gram matrix for the dual lattice L∗. More generally we show in the next theorem that the family of lattices of the form L(r,s) is closed under duality. T,v1 Before doing so we recall the notion of dual map. Given two lattices Λ1 and Λ2, and a linear map between them φ :Λ1 → Λ2, there exists a unique linear map ∗ ∗ ∗ φ :Λ2 → Λ1 that is completely characterized by the following adjoint property: for ∗ ∗ every l1 ∈ Λ1 and l2 ∈ Λ2 we have that

∗ ∗ ∗ hφ (l2), l1i1 = hl2, φ(l1)i2 where h , ii denotes the inner product in the vector space Λi ⊗ R. If we pick bases for Λi, together with their respective dual basis, we can identify Hom(Λ1, Λ2) with 16 Mm,n(Z). Under this choice of coordinates the map

∗ ∗ Hom(Λ1, Λ2) → Hom(Λ2, Λ1) φ 7→ φ∗ is just transposition,

Mm,n(Z) → Mn,m(Z) A 7→ AT .

Observe that in particular we have that the map φ 7→ φ∗ is linear and that (φ ◦ ψ)∗ = ψ∗ ◦ φ∗. Theorem 24. Let L ⊂ Rn be a tame lattice, let T: L → Z be a non-trivial linear map and v1 ∈ L \ ker T. Let r, s be integers such that 0 6= |r| < | T(v1)|. Then, the dual lattice of L(r,s) is of the same type. More explicitly, there is a non-trivial linear T,v1 ∗ ∗ ∗ (r,s) ∗ (r,s) map T:e L → Z and v1 ∈ L \ ker Te such that LT,v is isomorphic to (L ) e 1 Te,ve1 and that |r| < |T(e ve1)|. Proof. By deﬁnition L(r,s) is the image of the map Φ : L → L; x 7→ rφ (x) + T,v1 r,s 1 sφ2(x), where φ1(x) = x and φ2(x) = T(x)v1. Since dualizing the maps is a linear ∗ ∗ ∗ operator we see that Φr,s = rφ1 + sφ2. Notice that the map φ2 : L → L; φ2(x) =

T(x)v1 can be written as the composition of the maps T : L → Z and µv1 : Z → L; n 7→ nv1

φ2 = µv1 ◦ T Hence, φ∗ = T∗ ◦ µ∗ . 2 v1 Let T = µ∗ and v := T∗(1) (here we are using that ∗ = ). The image of e v1 e1 Z Z ∗ ∗ ∗ T : Z → L is generated by ve1 so we can write, in a similar fashion as above, µ := T∗ and obtain that ve1 φ∗ = µ ◦ T. 2 ve1 e It follows from this that

∗ ∗ ∗ Φr,s(x) = rφ1(x) + sφ2(x) = rx + sT(e x)ve1. Finally, observe that T(v ) = µ∗ ◦ T∗(1) = (T ◦ µ )∗(1) = T(v ) which ﬁnishes the e e1 v1 v1 1 claim.

3.5. Construction of some lattices with high center density. We have seen that, under the assumptions made in Theorem 14, a well-rounded lattice of the (r,s) form Lv1 has center density less than or equal to the center density of An. However, if we replace Tv1 (x) = hx, v1i with a general Z-linear map T : L → Z, then we may get lattices with higher center density. Indeed, in [14, Example 3.6], using the trace Pn map T : L → Z, T(x) = i=1 xi, the authors construct the lattice Dn as a lattice of the form L(r,s) . The purpose of this section is to show how one can construct the T,v1 densest known lattices in dimensions 8 and 9 as lattices L(r,s) , when L is the tame T,v1 lattice Zn. This shows that there are many different types of lattices that can be obtained from a tame lattice L, using the linear map Φ(r,s). 17 3.5.1. The E8 lattice. The E8 lattice is the densest lattice packing in dimension 0 8. There are two equivalent versions of the lattice; the even (Γ8) and odd (Γ8) coordinate system version. Let us state the definition of both. Definition 25. Define

( 8 8 ) 1 X (5) Γ := x ∈ 8 ∪ + : x ∈ 2 , 8 Z Z 2 i Z i=1 ( 8 ) ( 8 8 ) X 1 X (6) Γ0 := x ∈ 8 : x ∈ 2 ∪ x ∈ + : x ∈ 2 + 1 . 8 Z i Z Z 2 i Z i=1 i=1

Proposition 26 gives the index of L(r,s) in L; this result will become useful in T,v1 Lemma 27. Proposition 26. Let L ⊂ Rn be a lattice, T: L → Z a non-trivial linear map and v1 ∈ L\ker T. Let r, s be integers such that 0 6= |r| < | T(v1)| and let m = r+s T(v1). Then [L : L(r,s) ] = |mrn−1|. T,v1 Proof. The case when L has a basis which is rigid with respect to T is proven in [14, Proposition 3.7], so let us prove that such a basis always exists. Since T: L → Z is non-trivial there is v ∈ L such that Im(T) is generated by T(v), moreover by the additivity of the rank ker T is a sublattice of L of rank n − 1. Let {w1,..., wn−1} be a basis for ker T. Notice that {w1,..., wn−1, v} is a basis for L. Thus, {w1 + v,..., wn−1 + v, v} is a basis for L and it is rigid with respect to T since T takes the value of T(v) at every element in the set.

Lemma 27. Let L = Z8 and c = (1, −1, 1, −1, 1, −1, 1, −1)T . Deﬁne the linear T T map T: L → Z, T(x) = c x. Let v1 = (−1, 1, −1, 1, 1, 1, 1, 1) ∈ L \ ker T, and (r, s) = (2, 1). Then L(r,s) = 2Γ0 . T,v1 8 Proof. First note that 0 6= |r| = 2 < 4 = | T(v )|, so Lemma 13 gives that L(r,s) 1 T,v1 is a full rank sublattice of L. Let m := r + s T(v1) = 2 + (−4) = −2. Let e1,..., e8 be the standard basis vectors in 8. A basis for L(r,s) is given by the vectors R T,v1

Φ(r,s)(ei) = rei + s T(ei)v1 = 2ei + civ1

producing the generator matrix

 1 1 −1 1 −1 1 −1 1   1 1 1 −1 1 −1 1 −1     −1 1 1 1 −1 1 −1 1     1 −1 1 1 1 −1 1 −1  M =   .  1 −1 1 −1 3 −1 1 −1     1 −1 1 −1 1 1 1 −1     1 −1 1 −1 1 −1 3 −1  1 −1 1 −1 1 −1 1 1

If we add column i to column i + 1 for every i ∈ {1,..., 7}, we get the following 18 generator matrix:  1 2 0 0 0 0 0 0   1 2 2 0 0 0 0 0     −1 0 2 2 0 0 0 0    0  1 0 0 2 2 0 0 0  M =   .  1 0 0 0 2 2 0 0     1 0 0 0 0 2 2 0     1 0 0 0 0 0 2 2  1 0 0 0 0 0 0 2

Since the columns of M 0 are contained in 2Γ0 , we can conclude that L(r,s) ⊆ 2Γ0 . 8 T,v1 8 On the other hand, since det(L) = 1 and [L : L(r,s) ] = |mrn−1| by Proposition 26, we T,v1 have

det(L(r,s) ) = [L : L(r,s) ]2 det(L) = |mrn−1|2 = |(−2) · 28−1|2 = 48 = det(2Γ0 ), T,v1 T,v1 8 proving L(r,s) = 2Γ0 . T,v1 8 3.5.2. Densest known lattice packing in dimension 9. The largest known center density of a lattice packing in dimension 9 is δ = 1√ , achieved by the 9 16 2 laminated lattice Λ [24]. Here we show how we can construct a lattice L(r,s) with 9 T,v1 this center density.

Lemma 28. Let L = Z9 and c = (1, −1, 1, −1, 1, −1, 1, −1, 1)T . Deﬁne the linear map T: L → Z, T(x) = cT x. Let

v1 = (−1, −1, −1, −1, −1, −1, −2, 1, −1) ∈ L \ ker T,

(r,s) 1 and (r, s) = (2, 1). Then δ(L ) = √ = δ9. T,v1 16 2 Proof. First note that 0 6= |r| = 2 < 4 = | T(v )|, so L(r,s) is a full rank sublattice 1 T,v1 of L. Let e ,..., e be the standard basis vectors in 9. A basis for L(r,s) is given by 1 9 R T,v1 the vectors

Φ(r,s)(ei) = rei + s T(ei)v1 = 2ei + civ1.

Thus, a generator matrix for L(r,s) is given by T,v1

 1 1 −1 1 −1 1 −1 1 −1   −1 3 −1 1 −1 1 −1 1 −1     −1 1 1 1 −1 1 −1 1 −1     −1 1 −1 3 −1 1 −1 1 −1    M =  −1 1 −1 1 1 1 −1 1 −1  .    −1 1 −1 1 −1 3 −1 1 −1     −2 2 −2 2 −2 2 0 2 −2     1 −1 1 −1 1 −1 1 1 1  −1 1 −1 1 −1 1 −1 1 1

If we add column i to column i + 1 for every i ∈ {1,..., 8}, we get the generator 19 matrix  1 2 0 0 0 0 0 0 0   −1 2 2 0 0 0 0 0 0     −1 0 2 2 0 0 0 0 0     −1 0 0 2 2 0 0 0 0  0   M =  −1 0 0 0 2 2 0 0 0  .    −1 0 0 0 0 2 2 0 0     −2 0 0 0 0 0 2 2 0     1 0 0 0 0 0 0 2 2  −1 0 0 0 0 0 0 0 2 √ √ One can verify that λ (L(r,s) ) = 22 + 22 = 8 and vol(L(r,s) ) = 29. Therefore, 1 T,v1 T,v1 √ √ λ (L(r,s) )9 9 9 9 (r,s) 1 T,v1 ( 8) 2 · ( 2) 1 δ(LT,v ) = = = = √ = δ9. 1 29 vol(L(r,s) ) 29 · 29 29 · 29 16 2 T,v1

4. Generic well-rounded lattices. In the previous section we saw how to construct GWR sublattices of tame lattices, and moreover, that the center densities of said sublattices are upper bounded by the center density of An. Thus, these lattices provide a source of GWR lattices with packing density arbitrarily close to the optimal density in dimensions 1–3. We proceed to construct GWR lattices which have a good packing density in higher dimensions. Particularly, we deform basis vectors of the densest lattice packing in dimensions 3–5, Dn, and in dimension 8, E8, to obtain lattices with good sphere packing density and minimal kissing number in α α these dimensions. We call these constructions Dn and E8 , where α is a parameter describing how much the basis vectors are distorted. We also investigate when these deformed lattices (scaled) are sublattices of Zn. One motivation for ﬁnding GWR lattices with high sphere packing density has to do with the theta series for a lattice, which is deﬁned as the function

X 2 Θ(q) := qkxk , x∈Λ where q = eiπτ , and =τ > 0. We are interested in the case τ = it, i.e. when q is real. We can express the theta function equivalently as

l2 l2 Θ(q) = 1 + k1q 1 + k2q 2 + ...,

where li, i = 1, 2,... is the length of the i:th shortest vector in the lattice and l2 ki := | {x ∈ Λ: kxk = li} |. Now if we approximate the series with Θ(q) ≈ 1 + k1q 1 = 2 1 + κ(Λ)qλ1(Λ) , we see that the kissing number and shortest vector length of a lattice play important roles on the theta series. The minimisation of the theta series for speciﬁc ranges for q is interesting in itself, but this problem has also real-life applications in contexts such as wireless communications in the presence of an eavesdropper, where it has been shown [3] that in the Rayleigh fading wiretap channel, using lattice coset codes, one can bound the eavesdropper’s correct decoding probability by the theta series of the sublattice used for the lattice coset codes. Further, the minimisation of the theta series of lattices appear when one wants to ﬁnd an optimal arrangement of particles which minimize the average energy per particle, when the interaction potential energies between particles are Gaussians [4]. 20 Densest lattice packings (with maximal λ1) have a large kissing number in all the dimensions for which the densest packings are known. Thus, a natural direction is to consider lattices with minimal kissing number but for which λ1 is close to the maximum in that dimension. This is the motivation behind the deformed lattices. As we will see, however, this does not mean that the theta function of the deformed lattice will be smaller than those for the non-deformed lattices Dn and E8. Indeed, it has been recently shown in [7] that in dimensions 8 and 24, minimizers of the theta series in these dimension are the E8 and Leech lattice, respectively. Nevertheless, the construction of GWR lattices with arbitrary good packing density is something that, as far as we are aware, has not been done before. Further, such constructions provide good candidates for secure lattice codes according to [13] (see also discussion in Section5).

4.1. The planar case. To highlight the idea behind the deformed lattices, we illustrate how the densest lattice packing in dimension 2, the hexagonal lattice Λh ∼ A2, can be deformed to produce GWR lattices with density as close as desired to the optimal density in dimension 2. In fact, we end up with a parametrization of representatives of equivalence classes of planar well-rounded lattices. √ 1 2 α Definition 29. Let 0 ≤ α ≤ 2 and α := 1 − α . Deﬁne Λh to be the planar lattice generated by the matrix

1 α MΛα := . h 0 α

1 2 0 2 It follows from the deﬁnition that Λh = Λh, the hexagonal lattice, and Λh = Z , the α orthogonal lattice. For other values of α,Λh is a planar GWR lattice. The volume α α of Λ is given by vol(Λ ) = | det(M α )| = α, and the center density is given by h h Λh α 2 α λ1(Λh ) 1 α 1 δ(Λh ) = 2 α = . In particular, α 7→ δ(Λh ) is an increasing function on 0, , 2 vol(Λh ) 4α 2 as can be seen from Figure1.

0.29

0.28

α δΛh  0.27 Out[]=

0.26

0.25 0.0 0.1 0.2 0.3 0.4 0.5 α

α Fig. 1: Center density of the deformed hexagonal lattice Λh as a function of α. The 2 bottom line shows δ(Z ) and the upper line δ(Λh).

21 4.2. Deformed Dn. The Checkerboard lattice, which is deﬁned as

( n ) T n X Dn := (x1, . . . , xn) ∈ Z : xi ≡ 0 (mod 2) , i=1 is a root lattice and the densest lattice packing in dimensions 3–5. D3 is also known as the face-centered cubic lattice. Using the same strategy as with the hexagonal α lattice, we distort√ the basis vectors of Dn to produce a family of GWR lattices {Dn }, α where α ∈ (1, 2], such that δ(D ) → δ(Dn) as α → 1. n√ √ 2 α Definition 30. Let 1 ≤ α ≤ 2 and α := 2 − α . We deﬁne Dn , n ≥ 3, to be the rank n lattice with generator matrix

 α 0 α 0 0 0 ... 0   α α 0 0 0 0 ... 0     0 α α α 0 0 ... 0     0 0 0 −α α 0 ... 0    MDα :=  ..  . n  0 0 0 0 −α α . 0     ......   ......     0 0 0 0 0 0 −α α  0 0 0 0 0 0 0 −α √ Remark that 0 ≤ α ≤ 1 when 1 ≤ α ≤ 2 and α2 + α2 = 2. Further, D1 = D √ √ √ n n 2 n (motivating the notation) and Dn = 2Z . For any α ∈ (1, 2] we get a GWR√ α α lattice Dn (Theorem 32) and further, α 7→ δ(Dn ) is strictly decreasing on [1, 2] α (Proposition 35). We start by deriving an expression for the volume of Dn . α Proposition 31. The volume of Dn is given by

α n−3 3 3 vol(Dn ) = α (α + α ).

Proof. Notice that M α has the form of an upper triangular block matrix. There- Dn fore, the determinant is equal to the product of the determinants of the diagonal blocks;

α 0 α α n−3 n−3 3 3 n−3 det(M ) = α α 0 · (−α) = (−1) (α + α )α . Dn 0 α α

α The claim follows from vol(Dn ) = | det(MDα )|. √ n α α Theorem 32. Let 1 < α ≤ 2. Then Dn has a set of minimal vectors S(Dn ) = α {±b ,..., ±b } where b is the i:th column of M α . In particular, D is GWR and 1 n i Dn n 2 α λ1(Dn ) = 2. Proof. See Appendix, Section 5.1. α Knowing the volume and shortest vector length of Dn , we are able to compute its center density. α Corollary 33. The center density of Dn is given by 1 δ(Dα) = . n 2n/2αn−3(α3 + α3) 22 2 α Proof. By Theorem 32, λ1(Dn ) = 2. Thus, using Proposition 31, λ (Dα)n 1 δ(Dα) = 1 n = . n n α n/2 n−3 3 3 2 vol(Dn ) 2 α (α + α )

α Next√ we will show that the center density δ(Dn ) is a strictly decreasing function α on [1, 2]. This implies that no two lattices in the family {Dn } are equivalent. Lemma 34 will be used to prove this fact. Lemma 34. Let n ≥ 3 be an integer. Deﬁne the real-valued function √ n−3 3 2 3/2 f : [1, 2] → R, f(x) = x (x + (2 − x ) ). √ Then f is strictly increasing on [1, 2]. Proof. A direct computation shows that

3/2 p f 0(x) = xn−4 (n − 3) x3 + 2 − x2 + 3 x − 2 − x2 x2 > 0 √ √ 3 2 3/2 2 for all x√∈ (1, 2), since (n − 3)(x + (2 − x ) ) ≥ 0 and x − 2 − x > 0 when x ∈ (1, 2). α Proposition 35. The center density of Dn satisﬁes

n α δ(Z ) ≤ δ(Dn ) ≤ δ(Dn),

and√ the upper bound is achieved when α = 1 and the√ lower bound is achieved when α α = 2. Moreover, δ(Dn ) is strictly decreasing on [1, 2]. Proof. Let f be the function deﬁned in Lemma 34. Then δ(Dα) = 1 and in n √2n/2f(α) α particular, by the previous lemma, δ(Dn ) is strictly decreasing on [1, 2]. Therefore, 1 1 δ(Dα) ≤ = = δ(D ), n 2n/2f(1) 2n/2+1 n

α 1 1 n δ(Dn ) ≥ √ = = δ(Z ). 2n/2f( 2) 2n

4.2.1. Integral deformed Dn. From a computational perspective, the lattices α Dn can be problematic, since the basis vectors might contain irrational entries, or entries close to 1 if we wanted to have a high center density. Moreover, in the context of lattice coset coding, one often looks for GWR lattices which are sublattices of the orthogonal lattice Zn. Since scaling does not change the center density of a lattice, and GWR lattices are closed under scaling, the aforementioned considerations motivate α n finding lattices cDn ⊆ Z , where c is a constant. α n One way to obtain a lattice cDn ⊆ Z is to ensure that α, α ∈ Q and then scale α Dn with the denominator of α. If we suppose that α = p/q√ for some positive and √ 2q2−p2 relatively prime p, q ∈ Z such that 1 ≤ p/q < 2, then α = q ∈ Q if and only if 2q2 −p2 is a square. Equivalently, (p, q) is a solution to the generalized Pell’s equation 2y2 −x2 = d2 for some d ∈ Z. There exists algorithms for finding a solution to such an equation (in fact, infinitely many solutions), if a solution exists. We are particularly interested in solutions for which q is small, since this gives a small determinant for 23 α qDn , and for which p/q is close to 1, since this gives a high center density. Thus, it suffices to check case by case all small pairs of integers (p, q) which have the desired properties. Table1 shows some pairs of integers ( p, q), and corresponding d, α = p/q, center density and normalized squared lattice minimum, such that 2q2 − p2 = d2 for α n some d ∈ Z and consequently, qDn ⊆ Z . We have excluded the trivial case p = q which yields Dn.

α 2 α0 p q d α δ(Dn ) λ1(Dn )

2 − n −3 − n −3 2 2 ·5n·73−n 2 2 ·5n·73−n n 7 5 1 1.4 43 4 · 43 2 − n −3 − n −3 2 2 ·13n·173−n 2 2 ·13n·173−n n 17 13 7 1.30769 657 4 · 657 2 − n −4 − n −4 2 2 ·25n·313−n 2 2 ·25n·313−n n 31 25 17 1.24 2169 4 · 2169 2 − n −4 − n −4 2 2 ·41n·493−n 2 2 ·41n·493−n n 49 41 31 1.19512 9215 4 · 9215 2 − n −3 − n −3 2 2 ·61n·713−n 2 2 ·61n·713−n n 71 61 49 1.16393 59445 4 · 59445 2 − n −3 − n −3 2 2 ·85n·973−n 2 2 ·85n·973−n n 97 85 71 1.14118 158823 4 · 158823 2 − n −5 − n −5 2 2 ·113n·1273−n 2 2 ·113n·1273−n n 127 113 97 1.12389 92533 4 · 92533 2 − n −5 − n −5 2 2 ·145n·1613−n 2 2 ·145n·1613−n n 161 145 127 1.11034 194427 4 · 194427 2 − n −3 − n −3 2 2 ·181n·1993−n 2 2 ·181n·1993−n n 199 181 161 1.09945 1506735 4 · 1506735 2 − n −4 − n −4 2 2 ·265n·2873−n 2 2 ·265n·2873−n n 287 265 241 1.08302 2352339 4 · 2352339 2 − n −3 − n −3 2 2 ·365n·3913−n 2 2 ·365n·3913−n n 391 365 337 1.07123 12256153 4 · 12256153 2 − n −6 − n −6 2 2 ·481n·5113−n 2 2 ·481n·5113−n n 511 481 449 1.06237 3499245 4 · 3499245

α n α0 Table 1: Pairs of integers (p, q) which produce a sublattice qDn ⊆ Z . Here Dn α α denotes the lattice cDn where c is chosen such that vol(cDn ) = 1.

α We provide an example which illustrates how an integral scaled Dn lattice can be obtained. p 7 Example 5. Suppose that n = 4 and (p, q) = (7, 5). Then α = q = 5 and d 1 α α = q = 5 , as seen from Table1. In this case, a generator matrix for qDn is given by  7 0 1 0   1 7 0 0  MqDα =   . n  0 1 7 1  0 0 0 −7 By Corollary 33, (or from Table1),

α 1 1 δ(Dn ) = 3 = ≈ 0.0648879. 2n/2αn−3(α3 + α ) 4/2 7 4−3 7 3 1 3 2 · 5 · 5 + 5 24 4.3. Deformed E8. Recall the definition of the E8 lattice (Definition 25), the densest lattice packing in dimension 8, with a center density of 0.0625. Our goal in 0 this section is to deform the basis vectors of the lattice Γ8 (the odd coordinate system version of E8) in the same way that we did with the Dn lattice, to obtain GWR lattices with good packing density in dimension 8. √ √ 2 α Definition 36. Let 1 ≤ α ≤ 2 and α := 2 − α . We define E8 to be the rank 8 lattice with generator matrix  1 2α 0 0 0 0 0 0   1 2α 2α 0 0 0 0 0     1 0 2α 2α 0 0 0 0    1  1 0 0 2α 2α 0 0 0  MEα :=   . 8 2  1 0 0 0 2α 2α 0 0     1 0 0 0 0 2α 2α 0     −1 0 0 0 0 0 2α 2α  1 0 0 0 0 0 0 2α

2 2 1 0 Note that α + α = 2 and that E8 = Γ8. This motivates the notation. α Lemma 37. The volume of E8 is given by 1 vol(Eα) = α2(α − α)(α4 + α2α2 + α4) + α6(α + α) . 8 2 Proof. A direct computation shows that

1 2 4 2 2 4 6 det(MEα ) = − α (α − α)(α + α α + α ) + α (α + α) . 8 2 α α The claim follows from vol(E8 ) = | det(ME )|. 8 √ α The following theorem states that E8 is GWR for all 1 < α ≤ 2. This means α that, in particular, {E8 } is a family of GWR lattices with center density approaching δ8 = 0.0625 as α → 1. √ α α Theorem 38. Let 1 < α ≤ 2. Then E8 has a set of minimal vectors S(E8 ) = α {±b ,..., ±b }, where b is the i:th column of M α . In particular, E is GWR and 1 8 i E8 8 2 α λ1(E8 ) = 2. Proof. The proof is similar to the proof of Theorem 32 (see Appendix, Section 5.1), but even longer and relies on a computer program so we simply omit it here. α We can now compute the center density of the lattice E8 . α Proposition 39. The center density of E8 is given by 1 δ(Eα) = . 8 8 α2(α − α)(α4 + α2α2 + α4) + α6(α + α) Proof. By Lemma 37 and Theorem 38, α 8 α λ1(E8 ) δ(E8 ) = 8 α 2 vol(E8 ) 24 = 8 1 2 4 2 2 4 6 2 · 2 α (α − α)(α + α α + α ) + α (α + α) 1 = . 8 α2(α − α)(α4 + α2α2 + α4) + α6(α + α)

25 4.3.1. Integral deformed E8 lattice. Our goal in this section is to ﬁnd scaled α 8 variants of the E8 lattice as a sublattice of Z . We apply the exact same strategy as with the Dα lattice. Let α = p for some relatively prime, positive p, q ∈ such that n √ q Z √ 2 2 2q −p 2 2 2 1 ≤ p/q < 2. Then α = q ∈ Q if and only if 2q − p = d for some d ∈ Z. If α 8 this is the case, then 2qE8 is a sublattice of Z . We want to ﬁnd small values for q to α p get a small determinant for 2qE8 , and a value of α = q which is close to 1 to maximize the center density. Table2 shows some pairs of integers ( p, q), and corresponding α, d, center density and squared lattice minimum, such that 2q2 −p2 = d2 for some d ∈ Z α n and thus, 2qE8 ⊆ Z . As the table indicates, if we desire a high center density, we have to accept large values for q.

α 2 α0 p q d α δ(E8 ) λ1(E8 ) 7 5 1 1.4 0.0102162 1.27169 17 13 7 1.30769 0.0124829 1.33702 31 25 17 1.24 0.0159616 1.42177 49 41 31 1.19512 0.0192763 1.49045 71 61 49 1.16393 0.0222471 1.54482 97 85 71 1.14118 0.0248757 1.58856 127 113 97 1.12389 0.0272007 1.62445 161 145 127 1.11034 0.0292647 1.65442 241 221 199 1.0905 0.0327571 1.70171 337 313 287 1.07668 0.0355924 1.7374 449 421 391 1.06651 0.0379372 1.76533 647 613 577 1.05546 0.0407789 1.7975 881 841 799 1.04756 0.0430324 1.82183 1249 1201 1151 1.03997 0.0453987 1.84638 1799 1741 1681 1.03331 0.0476548 1.8689 2591 2521 2449 1.02777 0.0496839 1.88849 4049 3961 3871 1.02222 0.0518646 1.90888 6727 6613 6497 1.01724 0.0539629 1.9279 30257 30013 29767 1.00813 0.0582025 1.9647 95047 94613 94177 1.00459 0.0600098 1.97977 301087 300313 299537 1.00258 0.0610791 1.98853

α 8 Table 2: Examples of pairs of integers (p, q) which produce a sublattice 2qE8 ⊆ Z . α0 α α Here E8 denotes the lattice cE8 where c is chosen such that vol(cE8 ) = 1.

5. Conclusions and future work. Motivated by the attempt to construct lattices suitable for lattice coset codes for the fading wiretap channel as well as by the mathematical question itself, we have in this paper explored different constructions of generic well-rounded lattices: sublattices of tame lattices and deformed dense lattices. The task of finding GWR lattices with good sphere packing density was motivated by the findings in e.g. [11], [13] where it was demonstrated, using fading wiretap channel simulations, that having a small kissing number and a high sphere packing density for the eavesdropper’s lattice can lower the correct decoding probability for the eavesdropper for some relevant channel quality ranges. Furthermore, as discussed in Sec. 2.3, when the communication channel is of a 26 good quality, the performance of a lattice code is dictated by its diversity and mini- α mum product distance, which both should be maximized. The deformed lattices Dn α and E8 defined in this paper are clearly not full diversity. However, a natural way to obtain full diversity variants of the integral deformed lattices discussed in Sections 4.2.1 and 4.3.1 is to apply orthogonal transformations which maximise modulation diversity and minimum product distance for the lattice Zn. Then, since the integral deformed lattices are sublattices of Zn, if we apply the same orthogonal transforma- tion, the obtained lattices must also have full diversity and minimum product distance lower bounded by the minimum product distance of the optimally rotated Zn lattice. For some currently best known rotations, maximizing minimum product distance of Zn, see [28]. Let us also note that in scenarios where the channel quality is constantly low (e.g., due to long communication distance, high speed, physical obstacles, or low transmission power), non-full-diversity lattices can be of great interest. As for more applied future work, it is left to evaluate how well the constructed lattices perform in actual wiretap channel simulations. As mentioned earlier, one design criterion for the fading wiretap channel is to minimize the flatness factor of the sublattice, which is equivalent to minimizing the theta series. However, at least α in dimension 8, the deformed lattices E8 cannot be minimizers of the theta function, α since E8 is the unique minimizer. Still, the theta series of E8 approaches that of the E8 lattice as the parameter α approaches 1. In this paper, we also demonstrated the fact that, at least in dimensions less than 5 and in dimension 8, there exist GWR lattices with density arbitrarily close to the optimal density. One would expect this to hold for other dimensions as well. As far as we know, the research on GWR lattices is rather scarce, even though GWR lattices are widely represented in the set of well-rounded lattices. For instance, in dimension 1, all WR lattices are GWR, and in dimension 2, all WR lattices but the hexagonal lattice are GWR.

REFERENCES

[1] C. Alves, W. L. da Silva Pinto, and A. A. de Andrade, Well-rounded lattices via polyno- mials, CoRR, abs/1904.03510 (2019), http://arxiv.org/abs/1904.03510, arXiv:1904.03510. [2] A. Barreal, M. T. Damir, R. Freij-Hollanti, and C. Hollanti, An approximation of theta functions with applications to communications, SIAM Journal on Applied Algebra and Geometry, 4 (2020), pp. 471–501. [3] J.-C. Belfiore and F. Oggier, Lattice code design for the Rayleigh fading wiretap channel, in 2011 IEEE International Conference on Communications Workshops (ICC), 2011, pp. 1–5. [4] X. Blanc and M. Lewin, The crystallization conjecture: A review, EMS Surveys in Mathe- matical Sciences, EMS, 2 (2015), pp. 255–306. [5] W. Bolanos˜ and G. Mantilla-Soler, The trace form over cyclic number fields, Canadian Journal of Mathematics, (2020), pp. 1–23. [6] A. Chorti, C. Hollanti, J. Belfiore, and H. Poor, Physical layer security: A paradigm shift in data confidentiality, vol. 358 of Springer Lecture Notes in Electrical Engineering, 2016, pp. 1–15. [7] H. Cohn, A. Kumar, S. D. Miller, D. Radchenko, and M. Viazovska, Universal optimality of the E8 and Leech lattices and interpolation formulas, abs/1902.05438 (2019), https: //arxiv.org/abs/1902.05438, arXiv:1902.05438. [8] P. E. Conner and R. Perlis, A Survey of Trace Forms of Algebraic Number Fields, World Scientific, 1984. [9] S. I. Costa, F. Oggier, A. Campello, J.-C. Belfiore, and E. Viterbo, Lattices Applied to Coding for Reliable and Secure Communications, Springer, 2017. [10] R. Coulangeon, Spherical designs and zeta functions of lattices, International Mathematics Research Notices, 2006 (2006). [11] M. T. Damir, O. Gnilke, L. Amoros,´ and C. Hollanti, Analysis of some well-rounded lattices in wiretap channels, in 2018 IEEE 19th International Workshop on Signal Processing 27 Advances in Wireless Communications (SPAWC), IEEE, 2018, pp. 1–5. [12] M. T. Damir and D. Karpuk, Well-rounded twists of ideal lattices from real quadratic fields, Journal of Number Theory, 196 (2019), pp. 168–196. [13] M. T. Damir, A. Karrila, L. Amoros,´ O. W. Gnilke, D. Karpuk, and C. Hollanti, Well- rounded lattices: Towards optimal coset codes for gaussian and fading wiretap channels, IEEE Transactions on Information Theory, 67 (2021), pp. 3645–3663. [14] M. T. Damir and G. Mantilla-Soler, Bases of minimal vectors in tame lattices, abs/2006.167944 (2020), https://arxiv.org/abs/2006.16794, arXiv:2006.16794. [15] R. R. de Araujo and S. I. Costa, Well-rounded algebraic lattices in odd prime dimension, Archiv der Mathematik, 112 (2019), pp. 139–148. [16] B. N. Delone and S. S. Ryshkov, A contribution to the theory of the extrema of a multi- dimensional ζ-function, in Doklady Akademii Nauk, vol. 173, 1967, pp. 991–994. [17] L. Fukshansky, G. Henshaw, P. Liao, M. Prince, X. Sun, and S. Whitehead, On well- rounded ideal lattices II, International Journal of Number Theory, 9 (2013), pp. 139–154. [18] L. Fukshansky and K. Petersen, On well-rounded ideal lattices, International Journal of Number Theory, 8 (2012), pp. 189–206. [19] O. W. Gnilke, A. Barreal, A. Karrila, H. T. N. Tran, D. Karpuk, and C. Hollanti, Well-rounded lattices for coset coding in MIMO wiretap channels, in 2016 26th Interna- tional Telecommunication Networks and Applications Conference (ITNAC), IEEE, 2016, pp. 289–294. [20] O. W. Gnilke, H. T. N. Tran, A. Karrila, and C. Hollanti, Well-rounded lattices for reliability and security in Rayleigh fading SISO channels, in 2016 IEEE Information Theory Workshop (ITW), IEEE, 2016, pp. 359–363. [21] C. Ling, L. Luzzi, J.-C. Belfiore, and D. Stehle´, Semantically secure lattice codes for the Gaussian wiretap channel, 60 (2014), pp. 6399–6416. [22] C. McMullen, Minkowski’s conjecture, well-rounded lattices and topological dimension, Jour- nal of the American Mathematical Society, 18 (2005), pp. 711–734. [23] W. Narkiewicz, Elementary and analytic theory of algebraic numbers, Springer, 2004. [24] G. Nebe and N. J. A. Sloane, Table of densest packings presently known. http://www.math. rwth-aachen.de/∼Gabriele.Nebe/LATTICES/density.html, 2012. [Online]. [25] J. Neukirch, Algebraic Number Theory, Springer, 2010. [26] F. Oggier and E. Viterbo, Algebraic number theory and code design for Rayleigh fading channels, Foundations and Trends in Communications and Information Theory, 1 (2004), pp. 333–415. [27] P. Sarnak and A. Strombergsson¨ , Minima of Epstein’s zeta function and heights of flat tori, Inventiones mathematicae, 165 (2006), pp. 115–151. [28] E. Viterbo and F. E. Oggier, Full diversity rotations. http://www.ecse.monash.edu.au/ staff/eviterbo/rotations/rotations.html, 2005. [Online]. Appendix. 5.1. Proof of Theorem 32. √ √ √ 2 n Proof.√ The case α = 2 is clear since Dn = 2Z . Let us therefore assume 2 2 2 2 α 1 < α < 2. Note that kbik = α +α = 2 for all 1 ≤ i ≤ n, which gives λ1(Dn ) ≤ 2. α 2 Pn Now suppose that x ∈ S(Dn ); then kxk ≤ 2. Write x = i=1 cibi, where ci ∈ Z and not all ci’s are zero. Then     x1 c1α + c3α  x2   c1α + c2α       x3   c2α + c3α + c4α       x4   −c4α + c5α  x =   =   .  x5   −c5α + c6α       .   .   .   .       xn−1   −cn−1α + cnα  xn −cnα

2 2 2 2 2 2 We have that kxk ≥ xn = cnα > cn and hence |cn| ≤ 1. If cn = 0, then kxk ≥ 2 2 2 2 xn−1 = cn−1α > cn−1 and so |cn−1| ≤ 1. Otherwise, if |cn| = 1, then also |cn−1| ≤ 1. 28 2 2 2 2 2 To see this, suppose that |cn−1| ≥ 2. Then kxk ≥ xn−1 + xn ≥ (2α − α) + α > 2 2 α + α > 2, a contradiction. Inductively, we conclude |ci| ≤ 1 for i = 4, 5, . . . , n. Consider the coeﬃcient c4. We may without loss of generality assume that c4 ∈ {0, 1}, since we can always replace x by −x. Case 1: c4 = 0. We have

2 2 2 2 2 2 2 kxk ≥ x1 + x2 + x3 = (c1α + c3α) + (c1α + c2α) + (c2α + c3α) .

Note that c1, c2, c3 cannot all be non-zero. If this was the case, then two of them, 2 say ci and cj, would have the same sign and we would get the contradiction kxk ≥ 2 2 (ciα+cjα) ≥ (α+α) = 2+2αα > 2. Now suppose that ci = 0 for some i ∈ {1, 2, 3} and denote by cj and ck the other two coeﬃcients. Then if both cj and ck are non-zero, we get 2 2 2 2 2 2 2 2 kxk ≥ cj α + ckα + (ckα + cjα) > α + α = 2, 2 since (ckα + cjα) > 0 when cj, ck ∈ {±1}. This is a contradiction. We conclude that either (i) |ci| = 1 for some i ∈ {1, 2, 3} and cj = 0 for j ∈ {1, 2, 3}\{i} or, (ii) ci = 0 for all i ∈ {1, 2, 3}. In case (i),

2 2 2 2 2 (7) kxk = α + α + x4 + ··· + xn 2 2 2 2 2 2 (8) = 2 + c5α + (−c5α + c6α) + ··· + (−cn−1α + cnα) + cnα

and thus c5 = ··· = cn = 0. In case (ii),

2 2 2 2 2 2 2 2 2 kxk = x4 + ··· + xn = c5α + (−c5α + c6α) + ··· + (−cn−1α + cnα) + cnα .

Suppose that there are at least two non-zero coeﬃcients ci where i ∈ {5, . . . , n}. Let cj and ck be two such coeﬃcients, with j minimal and k maximal. Then

2 2 2 2 2 2 2 2 2 2 2 kxk = cj α + (−cjα + cj+1α) + ··· + (−ck−1α + ckα) + ckα > cj α + ckα = 2,

a contradiction. We conclude that |ci| = 1 for one i ∈ {5, . . . , n} and ci = 0 else.

Case 2: c4 = 1. Note that

2 2 2 2 2 2 x1 + x2 + x3 = (c1α + c3α) + (c1α + c2α) + ((c2 + 1)α + c3α) 2 2 2 2 2 2 2 x4 + ··· + xn = (−α + c5α) + ··· + (−cn−1α + cnα) + cnα ≥ α .

Suppose that c1, c2, c3 are all non-zero. If c1 and c3 have the same sign or c1 and 2 2 2 2 2 c2 have the same sign, we get a contradiction kxk ≥ x1 + x2 + x3 ≥ (α + α) > 2. 2 2 On the other hand, if c2 and c3 have the same sign, then x3 ≥ α which implies 2 2 2 kxk ≥ α + α > 2, a contradiction. Therefore, ci = 0 for some i ∈ {1, 2, 3}. If c1 = 0 then 2 2 2 2 2 2 2 kxk ≥ c3α + c2α + ((c2 + 1)α + c3α) + α

which implies c2 = 0 and c3 = 0. If c2 = 0 then

2 2 2 2 2 2 kxk ≥ (c1α + c3α) + c1α + (α + c3α) + α

which implies c1 = 0 and c3 = 0. If c3 = 0 then

2 2 2 2 2 2 2 kxk ≥ c1α + (c1α + c2α) + (c2 + 1) α + α 29 which implies c1 = 0 and c2 = 0. In any case, c1 = c2 = c3 = 0. Therefore, 2 2 2 2 x1 + x2 + x3 = α and

2 2 2 2 2 2 kxk = α + (−α + c5α) + ··· + (−cn−1α + cnα) + cnα ≥ 2 with equality holding if and only if c5 = c6 = ··· = cn = 0. Cases 1 and 2 imply that x = ±bi for some i ∈ {1, . . . , n}. This shows that α α S(D ) = {±b ,..., ±b }. To see that D is WR, note that M α is non-singular and n 1 n n Dn α n α α α thus S(Dn ) spans R . Moreover, Dn is GWR since κ(Dn ) = |S(Dn )| = 2n.