ON SPARSE GEOMETRY OF NUMBERS

LENNY FUKSHANSKY, PAVEL GUERZHOY, AND STEFAN KÜHNLEIN

Abstract. Let L be a lattice of full rank in n-dimensional real space. A vec-

tor in L is called i-sparse if it has no more than i nonzero coordinates. We
define the i-th successive sparsity level of L, si (L), to be the minimal s so
that L has s linearly independent i-sparse vectors, then si (L) ≤ n for each
1 ≤ i ≤ n. We investigate sufficient conditions for si (L) to be smaller than n
and obtain explicit bounds on the sup-norms of the corresponding linearly in-
dependent sparse vectors in L. These results can be viewed as a partial sparse
analogue of Minkowski’s successive minima theorem. We then use this result
to study virtually rectangular lattices, establishing conditions for the lattice to
be virtually rectangular and determining the index of a rectangular sublattice.
We further investigate the 2-dimensional situation, showing that virtually rect-
angular lattices in the plane correspond to elliptic curves isogenous to those
with real j-invariant. We also identify planar virtually rectangular lattices in
terms of a natural rationality condition of the geodesics on the modular curve
carrying the corresponding points.

1. Introduction
Let n ≥ 2 be an integer. For each x ∈ Rn , we write
n
!1/2
X
kxk = x2i , |x| = max |xi |
1≤i≤n
i=1

for the usual Euclidean norm and sup-norm on Rn , respectively. We also define the
0-norm on Rn :
n
X
kxk0 := x0i ,
i=1

where we use the convention that 00 = 0. The 0-norm counts the number of nonzero
coordinates of a vector, which we refer to as the sparsity level of this vector; if
sparsity level of some vector is no larger than m, we say that this vector is m-sparse.
Sparsity has been actively investigated in the context of compressed sensing, which
is a signal recovery paradigm based on the idea that most signals are sparse and can
therefore be reconstructed from a small number of linear measurements [6]. More
recently, the sparsity phenomenon has also been studied in discrete mathematics
and discrete geometry, in particular in the context of lattices [7], [2], [1]. In this
paper, we want to take a first stab at a systematic approach to what we see as a
“sparse analogue” of the classical geometry of numbers.

2010 Mathematics Subject Classification. Primary: 11H06, 52C07, 11G05.

Key words and phrases. lattices, sparse vectors, virtually rectangular lattices, Siegel’s lemma,
elliptic curve, j-invariant, isogeny, modular curve, geodesics.
Fukshansky was partially supported by the Simons Foundation grant #519058.
1
2 LENNY FUKSHANSKY, PAVEL GUERZHOY, AND STEFAN KÜHNLEIN

Let A = (aij ) ∈ GLn (R), and define

|A| := max |aij |.
1≤i,j≤n
n n
Let L = AZ ⊂ R , then L is a lattice of full rank with basis matrix A. The
minimal norm of L is defined as
|L| := min {kxk : x ∈ L \ {0}} .
Previous research has focused on sparsity of integer representations of lattice vec-
tors, i.e. on representing a vector x ∈ L as x = Ay with y ∈ Zn being as sparse
as possible. In this paper, we will focus on the sparsity of the lattice vectors them-
selves. Specifically, we define the successive sparsity levels s1 , . . . , sn of the lattice
L to be

si (L) := min s : ∃ i linearly independent vectors x1 , . . . , xi ∈ L

with kx1 k0 , . . . , kxi k0 ≤ s .
Then 1 ≤ s1 ≤ · · · ≤ sn ≤ n. Given a lattice L, what can be said about its
successive sparsity levels? Further, assuming we know that some s` ≤ k, can we
find ` such k-sparse vectors in L? To answer these questions, we need some more
notation.
For every nonzero vector x ∈ L define
d(x) := dimQ spanQ {x1 , . . . , xn }
to be its rational dimension. If A is an n × n real matrix with row vectors P ai for
1 ≤ i ≤ n, then for each subset I ⊆ [n] := {1, . . . , n} we define dI (A) := i∈I d(ai ).
We write d(A) for d[n] (A), and define d(L) = d(A), where A is any basis matrix for
L. Indeed, this definition does not depend on the choice of a basis matrix: if A and
B are two basis matrices for L, then B = AU for some U ∈ GLn (Z), and so each
row vector bi of B is of the form bi = ai U for the corresponding row vector ai of
A, which implies that d(bi ) = d(ai ). More generally, d(A) = d(AU ) holds for any
U ∈ GLn (Q). Notice that d(L) ≥ n. We refer to d(L) as the rational dimension
of L: the smaller d(L) is the “closer” L is to being rational, meaning L ⊂ Qn .
Recall that L is integral if kxk2 ∈ Z for any x ∈ L, and L is arithmetic if it is a
scalar multiple of an integral lattice, so rational lattices are arithmetic. Certainly
d(L) = n for all rational lattices, but there also exist non-rational arithmetic lattices
for which d(L) = n, for instance
√ √ 
2 2 √2 Z2
L= √
2 3 2
is one such example. On the other hand, there exist arithmetic lattices with rational
dimension > n, for instance
√ 
 
 1 √1 1
√1 2
Z2 , √1 2 0 Z3 ,
2 −1
2 −1 0
and similar examples can be constructed in every dimension. Non-arithmetic lat-
tices can also have rational dimension > n, for instance the planar lattice
 √ 
1 3
(1) Λ1 = Z2
0 1
 ON SPARSE GEOMETRY OF NUMBERS 3

with d(Λ1 ) = 3, as well as = n, for instance the planar lattice

 
π 2π
Λ2 = Z2
2 1
with d(Λ2 ) = 2.
We will also define two “measures of irrationality” of vectors in L and of L itself.
First, if x ∈ L has d(x) = k, then it can be written as
k
X
(2) x= αi f i ,
i=1

where f 1 , . . . , f k are integer vectors with relatively prime coordinates. Notice that
this decomposition is unique only if k = 1, for example
 √ √  √ √
1, 2, −2 3 = 1 · (1, 0, 0) + 2 · (0, 1, 0) − 2 3 · (0, 0, 1)
√ ! √ !
2 √ 2 √
= 1 · (1, 0, 0) + − 3 · (0, 1, 1) + + 3 · (0, 1, −1) .
2 2
Then define 
|α1 | if k = 1,
ν(x) :=
0 if k > 1.
For our basis matrix A, we define
n
Y
ν(A) := ν(ai ),
i=1

and for the lattice L = AZn , we let ν(L) = ν(A). This definition does not depend
on the choice of the basis matrix A. Clearly, there are many lattices for which
ν(L) = 0, for example ν(Λ1 ) = 0, where Λ1 is as in (1). In fact, it is not difficult
to show that ν(L) > 0 if and only if d(L) = n (see Lemma 3.1 below).
Second, let hLi be the additive abelian group generated by the entries of vectors
of L, and suppose that hLi has rank k ≥ 1. Fix a basis α = (α1 , . . . , αk ) for
hLi over Z, then every x ∈ L has a representation of the form (2) with vectors
f 1 , . . . , f k ∈ Zn . Define a map Φα : L → Rnk by
 
k
! f1
X  .. 
(3) Φα αi f i =  .  .
i=1 fk
The map Φα is additive, and hence extends to a map R ⊗ L = Rn → Rnk . We
can then pull back the sup-norm | · | on Rnk to Rn under Φα by defining |x|Φα :=
|Φα (x)|, this way obtaining a norm | · |Φα on Rn , which can then be compared to
the sup-norm on Rn . Specifically, we can define
 
|x|Φα n
(4) µ(α) := sup : x ∈ R \ {0} .
|x|
We can now state our first result.
Theorem 1.1. Let A ∈ GLn (R) and let L = AZn . Fix a basis α for hLi as above
and let µ(α) be as given in (4). Let 1 ≤ k < n and suppose that there exists a
subset I ⊂ [n] of n − k distinct indices such that dI (A) < n. Let ` = n − dI (A).
4 LENNY FUKSHANSKY, PAVEL GUERZHOY, AND STEFAN KÜHNLEIN

Then s` (L) ≤ k, and there exist ` linearly independent vectors x1 , . . . , x` ∈ L with

kxi k0 ≤ k and
`
Y
(5) |xi | ≤ nn−dI (A)/2 |A|n µ(α)dI (A) .
i=1

This theorem can be viewed as a “sparse” partial analogue of Minkowski’s successive

minima theorem. Indeed, if we know that s` (L) ≤ k, we can define the k-sparse
successive minima λ1 (L, k) ≤ · · · ≤ λ` (L, k) with respect to sup-norm to be

λi (L, k) := min {t ∈ R>0 : ∃ lin. ind. x1 , . . . , xi ∈ L with kxj k0 ≤ k, |xj | ≤ t} ,

so the usual successive minima are λi (L) := λi (L, n). Then (5) is an upper bound
on the product of these k-sparse successive minima. We prove Theorem 1.1 in
Section 2. Our main tool here is the celebrated Siegel’s lemma, which is known
to be sharp with respect to the exponent (we state it below as Theorem 2.3). We
comment on the quality of the bound (5) at the end of Section 2. Unfortunately,
the upper bound of (5) depends on the choice of the basis for L and for hLi. We
can alleviate this dependence for lattices L with rational dimension d(L) = n. We
need some more notation.
Let us say that a lattice is rectangular if it has an orthogonal basis. Follow-
ing [11], we will say that a lattice is virtually rectangular if it contains a rectangular
sublattice of finite index. Two lattices L, L0 ⊂ Rn are called isometric if there
exists a real orthogonal matrix U such that L0 = U L; on the other hand, L, L0 are
called similar if there exists a real orthogonal matrix U and a positive real number
β such that L0 = βU L. In other words, similarity as a linear map is a composition
of an isometry and a dilation. It is easy to notice that the virtually rectangular
property is preserved under isometry (and under similarity), however the sparsity
levels, rational dimension and the irrationality measure ν are not necessarily pre-
served under isometry (they are preserved under dilation). In Section 3 we give the
following characterization of virtually rectangular lattices, using the invariants we
have just introduced.

Theorem 1.2. Let L ⊂ Rn be a lattice of full rank. The following three statements
are equivalent:
(1) d(L) = n,
(2) ν(L) > 0,
(3) s1 (L) = · · · = sn (L) = 1.
Further, a full-rank lattice L0 ⊂ Rn is virtually rectangular if and only if it is
isometric to some lattice L satisfying the three equivalent conditions above.

In Section 3 we also prove the following effective result and show it to be optimal
(Example 3.1).

Theorem 1.3. Let A ∈ GLn (R) be such that the lattice L = AZn satisfies the
equivalent conditions of Theorem 1.2. Then L contains a rectangular sublattice M
with a basis of 1-sparse vectors so that
 n−1
det(L)
(6) [L : M ] = .
ν(L)
 ON SPARSE GEOMETRY OF NUMBERS 5

More generally, if L0 ⊂ Rn be a virtually rectangular lattice, then there exists a

 n−1
rectangular sublattice M 0 of L0 such that [L0 : M 0 ] = det(L)
ν(L) , where L is a
lattice isometric to L0 which satisfies the equivalent conditions of Theorem 1.2.

In the 2-dimensional case our results imply a certain property of elliptic curves
over C. An elliptic curve E can be realized as a complex torus C/Λ for a planar
lattice Λ, called the period lattice of this curve. Given two elliptic curves, E and E 0
a morphism φ : E → E 0 between them such that φ(0) = 0 and φ(E) 6= {0} is called
an isogeny. It is a remarkable fact that an isogeny is always surjective and has a
finite kernel, the order of which is called the degree of the isogeny, denoted δ(E/E 0 ).
If an isogeny E → E 0 exists, then there also exists the dual isogeny E 0 → E of the
same degree such that their composition is simply the multiplication-by-δ(E/E 0 )
map, and hence the curves are called isogenous: this is an equivalence relation. An
injective isogeny is called an isomorphism, and the set of isomorphism classes of
elliptic curves over C is parameterized by
(7) D := {τ = a+bi ∈ C : −1/2 < a ≤ 1/2, b ≥ 0, |τ | ≥ 1}\{eiθ : π/2 < θ < 2π/3}
in the following way. For each τ = a + bi ∈ D we can define a lattice
(8) Γτ = Z + Zτ,
 
1 a
which can be thought of as Z2 in R2 . Then every elliptic curve is isomor-
0 b
phic to an elliptic curve Eτ with period lattice Γτ for some τ ∈ D. In fact, it is
more natural to identify the set of isomorphic classes of elliptic curves with the
quotient space of the upper half-plane under the action of SL2 (Z) by linear frac-
tional transformations, where D is a fundamental domain for this action. There is
a unique bijective holomorphic map j : D → C taking e2πi/3 to 0 and i to 1728,
called the Klein j-function, which is modular and gives the j-invariant j(τ ) for
each isomorphism class Eτ of elliptic curves. In [8] the relevant properties of the
j-invariant are outlined, and in particular it is noted that for τ ∈ D, j(τ ) ∈ R if
and only if τ belongs to the set
n √ o 
1/2 + it : t ∈ R, t ≥ 3/2 ∪ eiθ : θ ∈ [π/3, π/2] ∪ {it : t ∈ R, t ≥ 1} ,

(9)

and j maps the first of these three subsets bijectively onto the interval (−∞, 0], the
second onto [0, 1], and the third onto [1, ∞) (see also Proposition on p. 160 of [10]
for an earlier appearance of this observation). With this notation, we can state the
following result which we prove in Section 4.
Theorem 1.4. Let τ = a + bi ∈ D and let Eτ be the corresponding elliptic curve
with the period lattice Γτ as above. The following statements are equivalent:
(1) Either a ∈ Q or there exists some t ∈ R such that a − bt, a + b/t ∈ Q,
(2) Γτ is virtually rectangular,
(3) Eτ is isogenous to an elliptic curve E 0 with real j-invariant ≥ 1,
(4) Eτ is isogenous to an elliptic curve E 0 with real j-invariant in [0, 1].
If these equivalent conditions hold with a ∈ Q, then there exists such an isogeny
E 0 → Eτ with δ(E 0 /Eτ ) = the denominator of a. If the conditions hold with a ∈
/Q
and t is any real number satisfying (1), then there exists such an isogeny E 0 → Eτ
6 LENNY FUKSHANSKY, PAVEL GUERZHOY, AND STEFAN KÜHNLEIN

with
|b|vw(t2 + 1)
(10) δ(E 0 /Eτ ) = ,
|t|
where v, w > 0 are denominators of the rational numbers a − bt and a + b/t, respec-
tively.
Our proof of this theorem uses Theorem 1.2. In particular, condition (1) of Theo-
rem 1.4 is equivalent to condition (1) of Theorem 1.2 in this 2-dimensional situation.
Further, (10) is just a reformulation of (6) in this case, since δ(E 0 /Eτ ) is precisely
the index of the rectangular period lattice of E 0 as a sublattice in the virtually
rectangular period lattice Γτ of Eτ . In the case when a = p/q ∈ Q, d(Γτ ) = 2.
We will refer to elliptic curves satisfying the conditions of Theorem 1.4 as virtu-
ally rectangular. The class of their period lattices includes all Γτ so that j(τ ) ∈ R
(see (9)). Further, this class includes all arithmetic planar lattices (see Lemma 2.5
of [11]), which are Γτ for τ ∈ D being a quadratic irrationality (see [8]): these are
τ = a+bi with a, b2 ∈ Q, which correspond precisely to elliptic curves with complex
multiplication (CM). We will discuss this situation in more details in Section 4, in
particular proving that CM elliptic curves are the only ones whose period lattice
contains non-parallel rectangular sublattices (Proposition 4.2 and Corollary 4.3):
in the CM case, there are infinitely many t satisfying condition (1) of Theorem 1.4
(each corresponding to a different rectangular sublattice), whereas for all other vir-
tually rectangular elliptic curves such t is essentially unique. Finally, in Section 5
we will show that virtually rectangular lattices in the plane have intrinsic geometric
meaning in terms of the corresponding points on the modular curve: they corre-
spond precisely to the points that lie on geodesics closed at infinity (Theorem 5.1).

2. Successive sparsity levels

In this section we prove Theorem 1.1. We start with a lemma on successive
sparsity levels.
Lemma 2.1. Let A ∈ GLn (R) and let L = AZn . Let 1 ≤ k < n and suppose that
there exists a subset I ⊂ {1, . . . , n} of n − k distinct indices such that dI (A) < n.
Let ` = n − dI (A). Then s` (L) ≤ k.
Proof. Let I = {i1 , . . . , in−k } for some 1 ≤ i1 < i2 < · · · < in−k ≤ n, and let us
write dj := d(aij ) for each 1 ≤ j ≤ n − k. Let AI be the (n − k) × n submatrix of A
consisting of the rows indexed by I. We want to show that there exists a nonzero
vector x ∈ Zn such that AI x = 0. For each 1 ≤ j ≤ n − k, let
Vj = x ∈ Qn : aij · x = 0 ,


then dimQ Vj = n − dj . Further, let us prove that

 
n−k
\
dimQ  Vj  ≥ ` = n − dI (A).
j=1

We argue by induction on n − k ≥ 1. If n − k = 1, then I = {i1 } and so dI (A) = d1 ,

in which case
dimQ V1 = n − d1 = `.
 ON SPARSE GEOMETRY OF NUMBERS 7

Now assume the result for all 1 ≤ n − k < m ≤ n − 1, and let us prove it for
n − k = m. Let
I 0 = {i1 , . . . , im−1 }, so I = I 0 ∪ {im },
Tm−1
then dI (A) = d(I 0 ) + dm . Let V 0 = j=1 Vj and V = V 0 ∩ Vm . By induction
hypothesis,
dimQ V 0 ≥ n − d(I 0 ).
Since d(I 0 ) < dI (A) < n, this implies that dimQ V 0 > 0, and so V 0 6= {0}. Now, by
a well-known identity in linear algebra,
dimQ V = dimQ V 0 + dimQ Vm − dimQ spanQ {V 0 , Vm }
≥ (n − d(I 0 )) + (n − dm ) − dimQ spanQ {V 0 , Vm }
≥ n − dI (A) = `,
since spanQ {V 0 , Vm } ⊆ Qn , and so dimQ spanQ {V 0 , Vm } ≤ n.
T 
n−k
This implies that dimQ j=1 Vj = ` > 0, and so there exist ` nonzero linearly
Tn−k
independent vectors y 1 , . . . , y ` ∈ j=1 Vj . These vectors are in Qn and satisfy
the equation AI y i = 0. Multiplying y 1 , . . . , y ` by the least common denominator
of their coordinates, we obtain linearly independent vectors x1 , . . . , x` ∈ Zn such
that AI xi = 0. This means that the vectors Ax1 , . . . , Ax` ∈ L have at least
n − k coordinates equal to 0. Since x1 , . . . , x` are linearly independent and A is
a nonsingular matrix, we must have Ax1 , . . . , Ax` linearly independent, and so
s` (L) ≤ k. 
Remark 2.1. Notice that
dI (A) ≥ dimQ AI := dimQ {ai1 1 , . . . , aik n }.
The converse of Lemma 2.1 is not true: if s1 (L) = k, there may not exist any
I ⊂ {1, . . . , n} of cardinality n − k so that dI (A) < n. Indeed, consider the example
 √ √ 
√1 √3 2√3
(11) A = √5 √3 2√ 3
2 3 5
and let L = AZ3 . Then s1 (L) = 1, since
   
0 0
A  2  =  √ 0 √  ∈ L.
−1 2 3− 5
On the other hand, d(a1 ) = d(a2 ) = 2 and d(a3 ) = 3. Thus for any I of cardinality
3 − 1 = 2, dI (A) ≥ 4 > n = 3. Further, even dimQ AI here is at least 3. On the
other hand, notice for comparison purposes that if a lattice L = AZn is virtually
rectangular, then dimQ (A> A) ≤ n (see [11]).
For each row vector ai of A let di = d(ai ). Then there exist Q-linearly indepen-
dent real numbers αi1 , . . . , αidi such that
di
X
(12) ai = αij f ij ,
j=1

where f ij are integer vectors with relatively prime coefficients for all 1 ≤ i ≤ n,
Pn
1 ≤ j ≤ di . Let d = i=1 di and let F (A) be the d × n matrix with rows f ij .
8 LENNY FUKSHANSKY, PAVEL GUERZHOY, AND STEFAN KÜHNLEIN

Example 2.1. For instance, in case of the matrix A in (11), we have d1 = 2, d2 = 2,

d3 = 3, and
a1 = α11 f 11 + α12 f 12 , a2 = α21 f 21 + α22 f 22 , a3 = α31 f 31 + α32 f 32 + α33 f 33 ,
where √ √ √
α11 = 1, α12 = α22 = α32 = 3, α21 = α33 = 5, α31 = 2,
and
f 11 = f 21 = f 31 = (1, 0, 0), f 12 = f 21 = (0, 1, 2), f 32 = (0, 1, 0), f 33 = (0, 0, 1).
Therefore d = 2 + 2 + 3 = 7 in this example, and the 7 × 3 matrix F (A) is
 
1 0 0
0 1 2
 
1 0 0
 
F (A) = 0 1 2 .

1 0 0
 
0 1 0
0 0 1
Define
|F (A)| := max max |f ij |,
1≤i≤n 1≤j≤di

where |f ij | is the sup-norm of the vector f ij .

Lemma 2.2. For a fixed choice of α = (αij : 1 ≤ i ≤ n, 1 ≤ j ≤ di ) as above, we
have
|F (A)| ≤ µ(α)|A|,
where µ(α) is as in (4).
Proof. Notice that the rank of the additive group hLi is equal to d, and α is a
basis for it. Using the notation of Section 1, specifically (3), notice that for each
1 ≤ i ≤ n,
|ai |Φα
≤ µ(α).
|ai |
On the other hand, |ai |Φα = |Φα (ai )| = max1≤j≤di |f ij |, and hence
|F (A)| = max max |f ij | ≤ µ(α) max |ai | = µ(α)|A|.
1≤i≤n 1≤j≤di 1≤i≤n


Our next result relies heavily on the use of Siegel’s lemma (Theorem 2 of [4])
and its adaptation to sup-norm using Fisher’s inequality (equation (1.8) of [3]). We
state it here for the reader’s convenience.
Theorem 2.3 (Siegel’s lemma with matrix sup-norm). Let B be an m × n inte-
ger matrix of rank m < n. Then there exist n − m linearly independent vectors
z 1 , . . . , z n−m ∈ Zn such that Bz i = 0 for every 1 ≤ i ≤ n − m and
n−m
Y √
m
|z i | ≤ n|B| .
i=1

The exponent m in this upper bound cannot in general be improved.

 ON SPARSE GEOMETRY OF NUMBERS 9

Lemma 2.4. Let A ∈ GLn (R), L = AZn and let α be a fixed basis for hLi. Let
1 ≤ k < n and suppose that there exists a subset I ⊂ {1, . . . , n} of n − k distinct
indices
1 ≤ i1 < i2 < · · · < in−k
Pn−k
such that dI (A) := j=1 dij < n. Let ` = n − dI (A). Then there exist ` linearly
independent vectors x1 , . . . , x` ∈ L with kxi k0 ≤ k and
`
Y
|xi | ≤ nn−dI (A)/2 |A|n µ(α)dI (A) .
i=1

Proof. Let AI be the (n − k) × n submatrix of A consisting of the rows indexed

by I. Let F (A)I be the dI (A) × n submatrix of F (A) consisting of rows with f il j
for il ∈ I and 1 ≤ j ≤ dil . Notice that AI y = 0 for some y ∈ Zn if and only if
F (A)I y = 0. Using the notation in the proof of Lemma 2.1, we have
 
n−k
\
V = Vj  = {y ∈ Qn : AI y = 0} = {y ∈ Qn : F (A)I y = 0} ,
j=1

which is an `-dimensional subspace of Qn . By Theorem 2.3, there exist ` linearly

independent vectors y 1 , . . . , y ` ∈ V ∩ Zn such that
`
Y √
dI (A)
(13) |y j | ≤ n|F (A)| .
j=1

For each 1 ≤ j ≤ `, define xj = Ay j . Since A is a nonsingular matrix, x1 , . . . , x`

are nonzero linearly independent vectors in L which are at least k-sparse. Now
notice that
(14) |xj | = |Ay j | ≤ n|A||y j |,
and so
`
Y `
Y
|xi | ≤ n` |A|` |y j |.
i=1 j=1
Combining this observation with (13) and Lemma 2.2 completes the proof. 
Theorem 1.1 now follows by combining Lemmas 2.1 and 2.4. Notice that the
exponent dI (A) in the upper bound of (13) is sharp due to the optimality of the
bound in Theorem 2.3. On the other hand, dependence of the bound of (14) on
|A| also cannot be improved in general. This suggests that the dependence of the
bound of Theorem 1.1 on |A| and µ(α) has correct order of magnitude.

3. Virtually rectangular lattices

In this section we focus on virtually rectangular lattices. We start by presenting
the proof of Theorem 1.2, split into two parts.
Lemma 3.1. Let L ⊂ Rn be a lattice of full rank. The following three statements
are equivalent:
(1) d(L) = n,
(2) ν(L) > 0,
(3) s1 (L) = · · · = sn (L) = 1.
10 LENNY FUKSHANSKY, PAVEL GUERZHOY, AND STEFAN KÜHNLEIN

Proof. Let L = AZn , where A is a basis matrix with rows a1 , . . . , an . We will prove
that (1) is equivalent to (2) and that (1) is equivalent to (3). First assume that
d(L) = n, then d(ai ) = 1 for each row ai of the basis matrix A. This means that
each ai = αi z i , where αi ∈ R \ {0} and z i ∈ Zn is a vector with relatively prime
coordinates, and so ν(ai ) = |αi |. Then
n
Y
ν(L) = ν(A) = |αi | > 0,
i=1

and so (1) implies (2). Further, this means that for any subset I ⊂ [n] of cardinality
n − 1, the linear system AI y = 0 has a nontrivial integer solution. Since such sets I
are of the form I = [n]\{i}, 1 ≤ i ≤ n, we can index corresponding integer solutions
by y i . Then each vector Ay i has only the i-th coordinate nonzero, and hence all
of such vectors are linearly independent (they are multiples of the standard basis
vectors). Therefore s1 (L) = · · · = sn (L) = 1, and so (1) implies (3).
Next assume ν(L) > 0, and let A be a basis matrix for L. Then
n
Y
ν(L) = ν(A) = ν(ai ) > 0,
i=1

which means that each ai is of the form ai = αi z i for some αi ∈ R \ {0} and
z i ∈ Zn a primitive vector. Hence d(L) = n, and so (2) implies (1).
Finally, suppose s1 (L) = · · · = sn (L) = 1. Then there exist linearly independent
vectors a1 e1 , . . . , an en ∈ L, where e1 , . . . , en are the standard basis vectors. Hence
there exists U ∈ GLn (Q) such that AU is a nonsingular diagonal matrix, which
implies d(A) = d(AU ) = n. Thus (3) implies (1). 
Lemma 3.2. A lattice L is virtually rectangular if and only if it is isometric to
some lattice L0 with s1 (L0 ) = · · · = sn (L0 ) = 1.
Proof. Suppose that L contains a rectangular sublattice M , and let B be an or-
thogonal basis matrix for M . Then there exists a real orthogonal matrix U such
that U B is a diagonal matrix. Let M 0 = U M = U BZn be a sublattice of the lattice
L0 = U L. Since U B is diagonal, M 0 has a basis consisting of scalar multiples of the
standard basis vectors, and thus all successive sparsity levels of L0 are equal to 1.
Conversely, assume L is isometric to some L0 with successive sparsity levels
equal to 1, say L = U L0 for some orthogonal matrix U . Then L0 contains n linearly
independent vectors x1 , . . . , xn with kxi k0 = 1. These vectors must therefore be
constant multiples of standard basis vectors. Let M 0 = spanZ {x1 , . . . , xn }, then
M = U M 0 is a rectangular sublattice of L. 
Then Theorem 1.2 follows by combining Lemmas 3.1 and 3.2. Next we prove
Theorem 1.3.
Proof of Theorem 1.3. Since d(L) = n, we must have d(ai ) = 1 for each row vector
ai of A. Then there must exist nonzero real numbers α1 , . . . , αn and primitive
integer row vectors f 1 , . . . , f n so that ai = αi f i for each 1 ≤ i ≤ n. Hence the
matrix F (A) with row vectors f i is n × n and A = AF (A), where A is the diagonal
matrix with diagonal entries α1 , . . . , αn . Let adj(F (A)) be the adjugate of F (A),
then adj(F (A)) is an integer matrix in GLn (Q), det(adj(F (A))) = det(F (A))n−1
and
F (A) adj(F (A)) = det(F (A))In ,
 ON SPARSE GEOMETRY OF NUMBERS 11

where In is the n × n identity matrix. Then

B := A adj(F (A)) = AF (A) adj(F (A)) = det(F (A))A,
which is a diagonal matrix with diagonal entries α1 det(F (A)), . . . , αn det(F (A)).
This implies that
Yn
det(B) = det(F (A))n det(A) = det(F (A))n αi .
i=1
On the other hand, since adj(F (A)) is an integer matrix in GLn (Q), the lattice
M := BZn is a full-rank sublattice of L, which is rectangular with a basis of
1-sparse vectors. Further,
det(M ) | det(B)|
[L : M ] = = = | det(adj(F (A))|
det(L) | det(A)|
 det(A) n−1
   n−1
det(L)
= | det(F (A))|n−1 =   = .
det(A)  ν(L)
Now, if L0 ⊂ Rn is any virtually rectangular lattice, then it is isometric to some lat-
tice L satisfying the equivalent conditions of Theorem 1.2. This L has a rectangular
sublattice M as we just constructed satisfying (6). Then L0 contains a rectangular
sublattice M 0 isometric to M with [L0 : M 0 ] = [L : M ]. 
Example 3.1. Theorem 1.3 is optimal, i.e. the lattice L = AZn as in the statement
of the theorem may not contain a rectangular sublattice M with a basis of 1-sparse
vectors and smaller index than given in (6). Indeed, this is easily seen to be the
case when  
1 0 ... 0 0
0 1 . . . 0 0
 
A =  ... ... . . . ... ... 
 
 
0 0 . . . 1 0
1 1 ... 1 d
for some integer d > 1. Here ν(L) = 1, det(L) = d, and the smallest-index
rectangular sublattice with a basis of 1-sparse vectors is
 
d 0 ... 0 0
0 d . . . 0 0
 
M =  ... ... . . . ... ...  Zn ,
 
 
0 0 . . . d 0
0 0 ... 0 d
which has index dn−1 in L.
Remark 3.1. It may be instructive to separately consider the 2-dimensional case of
Theorem 1.3, where the computation becomes completely elementary. Let L0 ⊂ R2
be a virtually rectangular lattice and let L be a lattice isometric to L0 which satisfies
the equivalent conditions of Theorem 1.2. Then L = AZ2 , where
 
α1 u1 α1 v1
A=
α2 u2 α2 v2
with u1 , u2 , v1 , v2 relatively prime integers and α1 , α2 real numbers. Then
det(L) = α1 α2 |u1 v2 − u2 v1 |, ν(L) = |α1 α2 |.
12 LENNY FUKSHANSKY, PAVEL GUERZHOY, AND STEFAN KÜHNLEIN

Further, it is easy to see that the orthogonal vectors

   
0 α1 (u2 v1 − u1 v2 )
z1 = , z2 =
α2 (u1 v2 − u2 v1 ) 0
are in L, and so M = spanZ {z 1 , z 2 } is a rectangular sublattice of L. Then
det(L)2
det(M ) = |α1 α2 (u1 v2 − u2 v1 )(u2 v1 − u1 v2 )| = .
|α1 α2 |
Let M 0 be a sublattice of L0 isometric to M in L, then
det(M ) det(L)
[L0 : M 0 ] = [L : M ] = = .
det(L) ν(L)

4. Isogenies of elliptic curves

In this section we prove Theorem 1.4 and discuss some of its consequences. To
start with, we state a technical lemma that will be of use to us: it is a combination
of Lemma 5.3 and Proposition 5.4 of [8] (see also Proposition on p. 160 of [10], as
mentioned in Section 1).
Lemma 4.1. Let D be as in (7). For τ ∈ D, the value j(τ ) is real if and only if
τ belongs to the set described in (9). Further, Γτ is WR if and only if j(τ ) is real
and belongs to the interval [0, 1].
Proof of Theorem 1.4. First suppose that a = pq ∈ Q, then Γτ contains orthogonal
vectors        
1 0 a 1
, =q −p .
0 qb b 0
These two vectors span a rectangular sublattice of Γτ of determinant qb, i.e. of
index q, which in particular implies that Γτ is virtually rectangular. If a ∈ / Q,
assume that there exists some t ∈ R such that a − bt, a + b/t ∈ Q. Define the lattice
 
1 1 a − bt
Lt := √ Z2 ,
1 + t2 t at + b
then it is easy to see that d(Lt ) = 2, and so it is virtually rectangular by Theo-
1 t
rem 1.2. Let θ = arctan t, then cos θ = √1+t 2
and sin θ = √1+t 2
, meaning that
 
1 1 −t
Ut = √
1 + t2 t 1
is an orthogonal matrix. Notice that Ut Γτ = Lt , meaning that Γτ is isometric to
Lt , hence it is also virtually rectangular. This shows that condition (1) implies (2).
Suppose now Γτ is virtually rectangular, then it contains a rectangular sublat-
tice Γ0 . Let E 0 be the elliptic curve (up to isomorphism) with period lattice Γ0 (up
to similarity). We can then assume that
 
1 0
(15) Γ0 = Z2 ,
0 q
which means that E 0 = Eτ 0 for τ 0 = iq. Hence τ 0 is in the third component set
of (9), and so j(τ 0 ) ≥ 1 (see Lemma 4.1 above). Now, since the period lattice of
E 0 is a sublattice of the period lattice of Eτ , there must exist an isogeny E 0 → Eτ
induced by the projection C/Γ0 → C/Γτ . This shows that condition (2) implies (3).
 ON SPARSE GEOMETRY OF NUMBERS 13

An isogeny E 0 → E exists if and only if the period lattice for E 0 is (up to

similarity) a sublattice of the period lattice for E. A planar lattice is called well-
rounded (WR) if it has two linearly independent shortest vectors with respect to
Euclidean norm. Now, the period lattice is rectangular if and only if the corre-
sponding j-invariant is real and ≥ 1 while the period lattice is WR if and only if
the corresponding j-invariant is real and in the interval [0, 1] (Lemma 4.1). Hence
to prove that conditions (3) and (4) are equivalent it is sufficient to show that Γτ
contains a rectangular sublattice if and only if it contains a WR sublattice. This is
guaranteed by Lemma 2.1 of [11].
Next assume that Eτ is isogenous to some elliptic curve E 0 = Eτ 0 with real
nonnegative j-invariant. Let Γ0 = Γτ 0 be the period lattice of E 0 , so j(τ 0 ) ∈ R≥0 and
Γ0 is (up to similarity) a sublattice of Γ. If j(τ 0 ) ≥ 1, then Lemma 4.1 guarantees
that τ 0 = iq for some q ≥ 1, and so Γ0 is of the form (15), which is rectangular.
If, on the other hand, 0 ≤ j(τ 0 ) < 1, then Lemma 4.1 implies that the lattice
Γ0 is WR. Now, Lemma 2.1 of [11] asserts that a lattice has WR sublattices if
and only if it is virtually rectangular. Hence we conclude in any case that Γτ is
virtually rectangular. Therefore Γτ is isometric to some lattice L with d(L) = 2,
by Theorem 1.2. Let  
cos θ − sin θ
U (θ) =
sin θ cos θ
for some angle θ be the corresponding isometry matrix, and let τ = a + bi so that
Γτ is of the form (8). Then
   
cos θ a cos θ − b sin θ 1 1 a − bt
L = U (θ)Γτ = Z2 = √ Z2 ,
sin θ a sin θ + b cos θ 1 + t2 t at + b
at+b
where t = tan θ. Since d(L) = 2, we must have a − bt ∈ Q and t = a + b/t ∈ Q.
This shows that condition (3) implies (1).
Finally, assume that the equivalent conditions of Theorem 1.4 hold. If a = pq ∈ Q,
then Γτ contains a rectangular sublattice Γ0 of index q. Let E 0 be the elliptic curve
(up to isomorphism) corresponding to Γ0 , then the degree of the isogeny E 0 → Eτ
is precisely this index q. If a ∈
/ Q, then equivalent conditions of Theorem 1.4 hold
with some t ∈ R. Then the period lattice Γτ of the curve Eτ is virtually rectangular
and isometric to the lattice Lt = At Z2 with
!
√ 1 √ 1 (a − bt)
 
1 1 a − bt 2 2
At = √ = √1+t t
1+t
√ t

,
1 + t2 t at + b 1+t2 1+t2
a + bt

and d(L0t ) = 2. Since a − bt, a + b/t ∈ Q, we can write

u b q
a − bt = , a+ =
v t w
with u, v, q, w ∈ Z and v, w > 0. Then, repeating the argument of Remark 3.1 for
this specific situation,
 
b
u − (a − bt)v = 0, q − a + w = 0,
t
and so the vectors
   
t 0
1 q − (a − bt)w
√ , √
1 + t2 u − a + bt v 1 + t2 0
14 LENNY FUKSHANSKY, PAVEL GUERZHOY, AND STEFAN KÜHNLEIN

are in Lt . These two vectors span a rectangular sublattice Rt of Lt , whose deter-

minant is
 b2 vw(t2 + 1)
     
 t b
det Rt =  2
u − a + v (q − (a − bt)w)= .
1+t t  |t|
Let Γ0 be a rectangular sublattice of Γt isometric to Rt , then
det Rt |b|vw(t2 + 1)
[Γt : Γ0 ] = [Lt : Rt ] = = .
det Lt |t|
Now, let E 0 be the elliptic curve (up to isomorphism) corresponding to Γ0 (up to
similarity), then the degree of the isogeny E 0 → Eτ is
|b|vw(t2 + 1)
δ(E 0 /Eτ ) = [Γτ : Γ0 ] = .
|t|
This completes the proof. 

Notice that if a lattice Γτ for τ = a + bi ∈ D contains a rectangular sublattice,

then it contains infinitely many non-similar rectangular sublattices: these can be
obtained for instance
 by multiplying the original rectangular sublattice by matri-
l 0
ces of the form for relatively prime integers l, m. However all of these
0 m
sublattices are parallel to each other, meaning that they are spanned by parallel
pairs of orthogonal basis vectors. Can Γτ have non-parallel rectangular sublattices?
This condition is equivalent to saying that there are multiple ways to rotate Γτ so
that some sublattice will have an orthogonal basis along the coordinate axes. Since
the parameter t of Theorem 1.4 is the tangent of the angle of rotation, this can
be possible if and only if there exist distinct t1 , t2 ∈ R satisfying condition (1) of
Theorem 1.4 so that t1 6= −1/t2 (t and −1/t correspond to rotations resulting in
the same lattice). This turns out to be possible if and only if τ is a quadratic
irrationality, in which case the corresponding elliptic curve Eτ is said to be a curve
with complex multiplication (CM): this is precisely the situation when the endomor-
phism ring of Eτ is larger than Z (specifically, an order in the imaginary quadratic
field Q(τ ); see Corollary III.9.4 of [14]). We now prove that for such τ there are
infinitely many different real numbers t satisfying condition (1) of Theorem 1.4.
Proposition 4.2. With notation as in Theorem 1.4, suppose that there exist t1 , t2 ∈
R satisfying condition (1). Define
(16) α1 := a − bt1 ∈ Q
(17) β1 := a + b/t1 ∈ Q
(18) α2 := a − bt2 ∈ Q
(19) β2 := a + b/t2 ∈ Q
Assume also that t1 6= t2 , −1/t2 . Then t21 , t22 ∈ Q and b2 ∈ Q, meaning that
τ = a + bi is a quadratic irrationality, and hence Eτ is a CM elliptic curve.
Proof. Subtract (16) from (18) and (17) from (19) to obtain
(20) b(t1 − t2 ) = α1 − α2
t1 − t2
(21) b = β1 − β2 .
t1 t2
 ON SPARSE GEOMETRY OF NUMBERS 15

Divide (20) by (21) (note that t1 6= t2 implies that β1 6= β2 ) to conclude that

α1 − α2
t1 t2 = ∈ Q.
β1 − β2
Multiply (19) by t1 t2 and take into the account bt1 = a − α1 (from (16)) to obtain
a(t1 t2 + 1) = α1 + β2 t1 t2
and conclude that a ∈ Q since t1 t2 6= −1 by assumption, and the quantities α1 , β2 ∈
Q, and t1 t2 ∈ Q are already known to be rational. Since a ∈ Q, we conclude from
(16) and (18) that bt1 , bt2 ∈ Q, and therefore their ratio (note that t1 t2 6= 0 because
τ∈ / R) is rational: t1 /t2 ∈ Q. Multiplication (and division) by the rational quantity
t1 t2 now allows us to conclude that t21 , t22 ∈ Q. Finally, since bt1 = a − α1 ∈ Q,
we square it to conclude that b2 t21 ∈ Q, and divide by t21 to obtain that b2 ∈ Q as
claimed. 
Proposition 4.2 asserts essential uniqueness of the real number t satisfying condition
(1) of Theorem 1.4 in the generic (non-CM) situation. In the case when CM occurs,
there is an infinite family of such t.
Corollary 4.3. With notation as in Proposition 4.2, assume τ = a + bi with
a, b2 ∈ Q. Then every t satisfying condition (1) of Theorem 1.4 is of the form
(22) qb or q/b for some q ∈ Q.
Conversely, for every rational q, t = qb satisfies condition (1) of Theorem 1.4.
Proof. Assume t1 , t2 are two different values of t satisfying condition (1) of Theo-
rem 1.4. From the proof of Proposition 4.2, we know that t1 /t2 ∈ Q, so t1 = qt2
for some q ∈ Q. Then t1 satisfies (22) if and only if t2 does, so it is enough to
show that there exists a t of the form (22) satisfying condition (1) of Theorem 1.4.
Indeed, for any q ∈ Q, take t = qb, then
a − bt = a − b2 q ∈ Q, a + b/t = a + 1/q ∈ Q.
On the other hand, take t = q/b, then
a − bt = a − q ∈ Q, a + b/t = a + b2 /q ∈ Q.
This completes the proof. 
Let us now provide an interpretation of this result for elliptic curves. For every
N > 1 there is a symmetric polynomial FN (X, Y ) = FN (Y, X) in two variables with
integer coefficients, which has the following property: two elliptic curves E1 and E2
with corresponding j-invariants j1 and j2 are isogenous with an isogeny of degree
N if and only if FN (j1 , j2 ) = 0. The polynomial FN (X, Y ) is commonly referred to
as N -th modular polynomial. Our Theorem 1.4 now implies that an elliptic curve E
with a non-real j-invariant j(E) is virtually rectangular if and only if for some N
the polynomial FN (X, j(E)), now monic with complex coefficients, has a real root.
Another observation that is not difficult to prove is that the curve Eτ is virtually
rectangular if and only if E−τ̄ is virtually rectangular, and if this is the case then
the curves Eτ and E−τ̄ are isogenous (i.e., any one of the corresponding lattices is
similar to a sublattice of the other one).
Further, consider an elliptic curve E over C with a non-real j-invariant j(E)
which is not isomorphic to an elliptic curve over R. Assume that E is nevertheless
isogenous to an elliptic curve E 0 over R with j-invariant j(E 0 ). This implies that
16 LENNY FUKSHANSKY, PAVEL GUERZHOY, AND STEFAN KÜHNLEIN

j(E) and j(E 0 ) are algebraically dependent over Q. Then the upper bound on the
inequality (10) on Theorem 1.4 gives a bound on the degree of the field extension
[Q(j(E), j(E 0 )) : Q(j(E))].

5. Planar virtually rectangular lattices on the modular curve

Our goal here is to present a geometric interpretation which justifies the consid-
eration of virtually rectangular lattices as very natural objects. Specifically, we look
at how the points corresponding to the virtually rectangular lattices are positioned
in the moduli space of all lattices.
WR lattices can be clearly seen in the fundamental domain (see Lemma 4.1
above), as can be the rectangular ones: they correspond to the set {it : t ∈ R, t ≥ 1}.
Virtually WR lattices (those that contain a finite-index WR sublattice) are the
same as virtually rectangular in R2 (see Lemma 2.1 of [11]), however they are not
as easily identified in the fundamental domain. Indeed, for a rational a ∈ Q with
−1/2 < a ≤ 1/2, the lattice corresponding to every point τ = a + ib ∈ D in the
fundamental domain is virtually rectangular. While this set is already too large to
be contained on any continuous path in the upper half-plane, there are many more
points in the fundamental domain corresponding to virtually rectangular lattices.
Surprisingly, the picture becomes much clearer if we look at these points as points
on the modular curve instead.
Let
H = {τ = x + iy ∈ C | y = =(τ ) > 0}
be the complex upper half-plane. It comes with the Poincaré metric
ds2 = y −2 (dx2 + dy 2 ),
which is invariant under the action of PGL2 (R)+ (and is uniquely defined by that
property up to a constant multiplication). Let Y = PSL2 (Z)\H. The points of
Y classify elliptic curves up to isomorphism over C and correspond to (orientation
preserving) similarity classes of lattices Γτ = h1, τ iZ in the plane.
Since PSL2 (Z) ⊂ PGL2 (R), the space Y inherits the metric from H, in particu-
lar, geodesics on Y are precisely the images of the geodesics on H under the natural
projection π : H → Y (recall that geodesics for a given metric are the paths of short-
est length). The modular curve is the compact Riemann surface X = PSL2 (Z)\H,
where H = H ∪ Q ∪ {∞}. We then have X = Y ∪ ∞, and the point at infinity ∞
does not correspond to any lattice.
Geodesics on H (for the metric ds2 ) are the vertical lines together with the
semicircles orthogonal to the real axis (see e.g. [5, Proposition 4.5.5], [12, Lemma
1.4.1]). While all geodesics on H are of infinite length, under the map π, some
geodesics become closed, while others are still of infinite length. We say that a
geodesic on H passes through ∞ if it is either a vertical line or a semicircle which
meets the real line at a rational point, and we say that a geodesic is closed at ∞
if it is either vertical with a rational x-coordinate or both ends of the semicircle
meet the real line at rational points. We apply the same terminology to the images
of these geodesics under the projection π. This terminology is not standard. For
example, while for any two points in Y there exists exactly one geodesic which
passes through these two points, there are infinitely many geodesics which pass
through ∞ and any given point on Y ; furthermore, the geodesics which are closed
 ON SPARSE GEOMETRY OF NUMBERS 17

at ∞ are never closed in Y . However, one may possibly argue that, for example,
a semicircle which meets the real line at two rational points is really closed at ∞
because both of its ends map to ∞ under π.
Theorem 5.1. A point on Y corresponds to a virtually rectangular lattice if and
only if this point belongs to a geodesic that is closed at ∞.
Proof. Let p ∈ Y , and assume that the corresponding lattice is virtually rectangu-
lar. Thus p = π(τ ) with τ = a + bi ∈ H, and the lattice Γτ is virtually rectangular.
Then the lattice Γτ has two orthogonal vectors, say, Aτ + B and Cτ + D where
A, B, C, D are integers and AD − BC 6= 0. The orthogonality condition can be
written as
(23) (a2 + b2 )AC + a(AD + BC) + BD = 0.
If A = 0, then BC 6= 0, and τ belongs to the vertical line x = −D/C. If C = 0,
then AD 6= 0, and τ belongs to the vertical line x = −B/A. Finally, if AC 6= 0
then (23) becomes an equation of the semicircle
2
(AD − BC)2

AD + BC
a+ + b2 =
2AC 4A2 C 2
satisfied by (a, b) with a rational radius of |(AD − BC)/2AC| and the center on
real line at x = −(AD + BC)/2AC ∈ Q. Thus in either case π(τ ) belongs to a
geodesic that is closed at ∞.
Conversely, assume that π(τ ) belongs to a geodesic which is closed at ∞. If this
geodesic is the image under π of a vertical line in H with a rational x-coordinate,
then τ = a + bi with a ∈ Q, and the lattice Γτ is virtually rectangular. Otherwise,
τ belongs to a semicircle which meets the real line at rational points, say α and β.
Pick σ ∈ SL2 (Z) such that σ(α) = ∞. Then σ(β) ∈ Q, and σ takes the semicircle
to the vertical line with a rational x-coordinate of σ(β) (Möbius transformations
preserve the Poincaré metric, therefore take geodesics to geodesics). Thus the
lattice h1, σ(τ )iZ is virtually rectangular, and this lattice is similar to Γτ . We thus
conclude that Γτ is virtually rectangular. 

Acknowledgement: We wish to thank the anonymous referees for their helpful

suggestions, which improved the quality of presentation.

References
[1] I. Aliev, G. Averkov, J. De Loera and T. Oertel, Optimizing sparsity over lattices and semi-
groups. Lecture Notes in Computer Science, 2020.
[2] I. Aliev, J. De Loera, T. Oertel and C. O’Neill, Sparse solutions of linear diophantine equa-
tions. SIAM Journal on Applied Algebra and Geometry, 1(1), 239–253, 2017.
[3] R. Baker and D. Masser, Siegel’s lemma is sharp for almost all linear systems. Bull. Lond.
Math. Soc., 51(5), 853–867, 2019.
[4] E. Bombieri and J. D. Vaaler, On Siegel’s lemma. Invent. Math., 73(1), 11–32, 1983.
[5] H. Cohen and F. Strömberg, Modular Forms. A Classical Approach. Graduate Studies in
Mathematics, 179. American Mathematical Society, Providence, RI, 2017.
[6] Y. C. Eldar and G. Kutyniok, Compressed sensing: theory and applications. Cambridge
University Press, 2012.
[7] L. Fukshansky, D. Needell and B. Sudakov. An algebraic perspective on integer sparse recov-
ery. Appl. Math. Comput., 340, 31–42, 2019.
18 LENNY FUKSHANSKY, PAVEL GUERZHOY, AND STEFAN KÜHNLEIN

[8] L. Fukshansky, P. Guerzhoy and F. Luca. On arithmetic lattices in the plane. Proc. Amer.
Math. Soc., 145(4), 1453–1465, 2017.
[9] P. M. Gruber and C. G. Lekkerkerker, Geometry of Numbers. North-Holland Publishing Co.,
1987.
[10] M. Koecher and A. Krieg, Elliptische Funktionen und Modulformen. Springer-Verlag, Berlin,
1998.
[11] S. Kühnlein, Well-rounded sublattices. Int. J. Number Theory, 8(5), 1133–1144, 2012.
[12] T. Miyake, Modular Forms Modular forms, Translated from the 1976 Japanese original by
Yoshitaka Maeda. Springer Monographs in Mathematics. Springer-Verlag, Berlin, 2006.
[13] W. M. Schmidt, Diophantine approximations and Diophantine equations. Lecture Notes in
Mathematics, 1467. Springer-Verlag, Berlin, 1991.
[14] J. H. Silverman, The arithmetic of elliptic curves. Graduate Texts in Mathematics, 106.
Springer-Verlag, New York, 1986.

Department of Mathematics, 850 Columbia Avenue, Claremont McKenna College,

Claremont, CA 91711
Email address: [email protected]

Department of Mathematics, University of Hawaii, 2565 McCarthy Mall, Honolulu,

HI, 96822-2273
Email address: [email protected]

Institut für Algebra und Geometrie, Fakultät für Mathematik, KIT, FRG-76128
Karlsruhe
Email address: [email protected]