Integer Points in Polyhedra Gennady Shmonin

Minkowskis theorem and its applications

March 3, 2009
1 Characterization of lattices
In this section, we prove that there is another, equivalent denition of lattices: a lattice is a
discrete subgroup of R
n
, i.e., a set of vectors R
n
such that
(a) z z
t
whenever z, z
t
, and
(b) there is a ball B

:
_
x R
n
: |x| <
_
such that B

{0}.
Since all norms in a nite-dimensional space are equivalent up to a constant factor, condition
(b) is invariant over different choices of the normin R
n
. For simplicity, for the rest of this section
we shall assume that || denotes the
2
-norm. Yet, condition (b) actually implies that (z +B

)
{z} for all vectors z (since z
t
(z +B

)implies z
t
z B

). The latter can be read

as follows: if z and z
t
are two vectors in and |z z
t
| <, then z z
t
. This property turns out to
be helpful when, for example, we consider convergent sequences of lattice vectors: if z
1
, z
2
, . . . is
a sequence of vectors in that converges to a vector z, then there is an index i
0
such that z
i
z
for all i >i
0
, and therefore, z is also a lattice vector.
1.1 Lattices are discrete groups
It is clear that every lattice (B) satises condition (a) in the denition of a discrete group.
There are many ways to check that condition (b) is also true; we present one which will be es-
pecially useful for us later. For a basis B
_
b
1
, b
2
, . . . , b
n
_
we denote B

_
b

1
, b

2
, . . . , b

n
_
the
corresponding GramSchmidt orthogonal basis (B

spans the same linear subspace as B, but

not necessarily generates the same lattice). Then the shortest vector in B

gives a lower bound

on the required value of .
Lemma 1. Let (B) be a lattice. Then
inf
_
|z|
2
: z \{0}
_
>min
i
|b

i
|
2
.
Proof. Any vector z in the span of B may be expressed as a linear combination of vectors in B

:
z
n

i 1
z,b

i

b

i
,b

i

b

i
.
Furthermore, since B

is an orthogonal basis, we have

|z|
2

i 1
_
_
_
z,b

i

b

i
,b

i

b

i
_
_
_
2

i 1
z,b

i

2
b

i
,b

i

2
|b

i
|
2
.
1
Suppose that z \ {0}, i.e., z

n
i 1
x
i
b
i
for some integers x
1
, x
2
, . . . , x
n
, not all equal to 0, and
let k be the biggest index such that x
k
/0. Then
|z|
2

i 1
z,b

i

2
b

i
,b

i

2
|b
i
|
2
>
z,b

2
b

k
,b

2
|b
k
|
2

z,b

2
b

k
,b

.
Now, it sufces to showthat [z, b

k
[ >b

k
, b

k
; this will imply that |z| >|b

k
|, and consequently,
|z| > min
i
|b

i
| for all z \ {0}. By the choice of k, x
i
0 for all i > k. Since b

k
is orthogonal
to the span of vectors b

1
, b

2
, . . . , b

k1
, which is also the span of vectors b
1
, b
2
, . . . , b
k1
, we have
b
i
, b

k
0 for all i < k. Finally, b
k
, b

k
b

k
, b

k
follows immediately from the description of
the GramSchmidt orthogonalization procedure. All together, this yields
[z, b

k
[

i 1
x
i
b
i
, b

[x
i
[ [b
k
, b

k
[ [x
i
[ b

k
, b

k
>b

k
, b

k
,
because [x
k
[ >1. This completes the proof.
Thus, lattices are indeed discrete subgroups of R
n
. Moreover, we obtained a reasonable lower
bound on the norm of the shortest vector in a lattice. However, this bound depends on the
choice of a basis B, and even on the order of vectors in B.
1.2 Discrete subgroups of R
n
are lattices
It remains to showthat every discrete subgroup of R
n
is a lattice. Let be a discrete subgroup of
R
n
. We need to show that there is a set of linearly independent vectors B
_
b
1
, b
2
, . . . , b
k
_
such
that (B). First, we show that there is a lattice vector attaining the inmum in Lemma 1.
Lemma 2. Let R
n
be a discrete group. Then there is a vector b
1
such that
|b
1
|
1
:inf
_
|z| : z \{0}
_
.
Proof. Let z
1
, z
2
, . . . be a sequence of vectors from such that the sequence of their norms |z
i
|
converges to
1
. We may assume that all z
i
s belong to the closed ball
_
x : |x| < 2
1
_
, which
is a compact set. Then there is a subsequence z
i
1
, z
i
2
, . . . that converges to some vector, say z.
Furthermore, there is an index j
0
such that |z
i
j
z| <

2
for all j > j
0
(where comes from the
denition of a discrete group). Thus, for any j , k > j
0
, we have
|z
i
j
z
i
k
| <|z
i
j
z|+|z z
i
k
| <

2
+

2
,
and since both z
i
j
and z
i
k
belong to , we must have z
i
j
z
i
k
. Consequently, z z
i
j
is also a
vector from, and we may assign b
1
z.
It is easy to see that every vector z , belonging to the span of b
1
, is an integer multiple of b
1
.
Indeed, if b
1
and is not an integer, then the vector b
t
1
(]])b
1
also belongs to but
|b
t
1
| <|b
1
|, a contradiction.
Now, we proceed by induction, assuming that || is the
2
-norm. Suppose that we have
constructed vectors b
1
, b
2
, . . . , b
k
such that every vector z in the span of b
1
, b
2
, . . . , b
k
, is an
2
integral linear combination of these vectors. Let L denote the span of b
1
, b
2
, . . . , b
k
. If L,
there is nothing to do. Otherwise, we dene b
k+1
as a vector from\ L, whose distance to L is
as small as possible. Existence of such a vector can be shown in a similar way to the proof of
Lemma 2. Indeed, let denotes the inmum of the distances d(z, L) from z \ L to L. Every
vector z \ L can be expressed as z x +z

, where x L and z

. Then d(z, L) |z

|.
Particularly, if z x +z

and x
1
b
1
+
2
b
2
+. . . +
k
b
k
, then the vector
z

+
k

i 1
(
i
]
i
])b
i
is at the same distance |z

| from L as vector z. Therefore, we may restrict our attention onto

vectors z x +z

, where 0 <
i
< 1 for all i (i.e., x is taken from the closed parallelepiped).
Let z
1
, z
2
, . . . be a sequence of vectors from\L, satisfying the above assumption, such that the
sequence |z

i
| converges to . We may further assume that |z

i
| < 2 for all i . Then all z
i
s
are taken from a compact set, and therefore, there is a convergent subsequence z
i
1
, z
i
2
, . . .. As
above, we may argue that starting from some index j
0
, all vectors z
i
j
are the same, say z
i
j
z.
Then |z

| >0 and we assign b

k+1
z.
It remains to showthat all vectors frombelonging to the spanof L{b
k+1
} canbe expressed
as an integral linear combination of vectors b
1
, b
2
, . . . , b
k+1
. Indeed, suppose that z

k+1
i 1

i
b
i
is a vector from. If
k+1
is not an integer, then the vector
z
t

i 1

i
b
i
+(
k+1
]
k+1
])b
k+1
is closer to L than b
k+1
:
d(z
t
, L) (
k+1
]
k+1
])|b

k+1
| <|b

k+1
| d(b
k+1
, L),
a contradiction to our choice of b
k+1
. Therefore,
k+1
is an integer. Then the vector z
k+1
b
k+1
belongs to L, and hence, can be expressed as an integral linear combination of vectors
b
1
, b
2
, . . . , b
k
.
After at most n steps, the process terminates. Thus, we proved the following theorem.
Theorem3. A set R
n
is a lattice if and only if it is a discrete group.
2 Fundamental parallelepiped
Let (B) be a full-dimensional lattice in R
n
. The set
(B) :
_
Bx : x [0, 1)
n
_
is called the fundamental parallelepiped associated to B. The volume of the fundamental par-
allelepiped vol(B) is exactly the absolute value of the determinant of B, [det(B)[, which, as we
know from the previous lectures, is the same for all bases of lattice . This suggests the way to
generalize the notion of the lattice determinant to lower-dimensional lattices, too.
3
Let B R
nk
be a matrix of full column rank and let (B) be the lattice generated by B.
We dene the determinant of as
det() :vol((B))
_
det(B
T
B),
where by vol((B)) we mean the relative volume of (B), i.e., its volume in the linear space
spanned by B. It is straight-forward to show that this quantity is invariant over unimodular
transformations of B, and therefore, det() is well-dened.
We shall need the following property of fundamental parallelepipeds.
Lemma 4. Let (B) be a full-dimensional lattice in R
n
. Then the sets
z +(B) :
_
z +x : x (B)
_
, z ,
form a partition of R
n
.
Proof. We showthat every vector x R
n
can be uniquely expressed as x z +y, where z and
y (B). Since B spans R
n
, we have
x
n

i 1

i
b
i

n

i 1
]
i
]b
i
+
n

i 1
{
i
}b
i
for some numbers
1
,
2
, . . . ,
n
; here ]] denotes the largest integer not exceeding and let {}
denotes ]]. Now, the vector z

n
i 1
]
i
]b
i
belongs to the lattice , while y

n
i 1
{
i
}b
i
belongs to the fundamental parallelepiped (B), and the claim follows.
Suppose that x z + y z
t
+ y
t
with z, z
t
and y, y
t
(B). Then the vector z z
t
also
belongs to the lattice, and therefore, can be expressed as
z z
t

i 1

i
b
i
,
i
Z, i 1, 2, . . . , n.
On the other hand,
z z
t
y
t
y
n

i 1

i
b
i
, 1 <
i
<1, i 1, 2, . . . , n,
since both y and y
t
belong to the fundamental parallelepiped (B). The vectors b
1
, b
2
, . . . , b
n
are
linearly independent, and therefore, both expressions must be the same, i.e.,
i

i
for each i .
But then the only possible choice is
i
0 for each i , which means z z
t
and y y
t
.
3 Minkowskis theorem
In this section, we consider the classical result of Minkowski, which, in some sense, originated
the whole geometry of numbers. Intuitively, it established the connexion between the geomet-
ric properties of the sets (convexity, symmetry, volume) and its arithmetic properties (existence
of an integral vector in a set).
As it is quite common nowadays, we shall not follow the original proof of Minkowski, but
exploit a later result of Blichfeldt.
4
Theorem5 (Blichfeldt). Let be a full-dimensional lattice in R
n
and let C R
n
be a measurable
set. Suppose that either
(a) vol(C) >m det(), or
(b) vol(C) >m det() and C is compact.
Then there are m+1 vectors x
0
, x
1
, . . . , x
m
C such that x
i
x
j
for each i , j .
Proof. For simplicity, we give a proof for the case m1. Let B be a basis of and suppose that
vol(C) >det(). We consider the sets
C
z
:C (z +(B)), z .
By Lemma 4, these sets partition C, and therefore,
vol(C)

z
vol(C
z
).
Now, consider the translates C
z
z
_
x z : x C
z
_
. Clearly, C
z
z (C z)(B) (B). Now,
we argue that these translates C
z
z cannot be pairwise disjoint. Indeed, since vol(C
z
z)
vol(C
z
), the total volume of all these sets is

z
vol(C
z
z)

z
vol(C
z
) vol(C) >det() vol((B)).
As all the sets C
z
z lie in (B), there must be z
0
, z
1
such that (C
z
0
z
0
) (C
z
1
z
1
) is non-
empty. Let y be a vector in (C
z
0
z
0
) (C
z
1
z
1
). If we dene x
0
y +z
0
C
z
0
C and x
1

y +z
1
C
z
1
C, then x
1
x
0
z
1
z
0
. This completes the proof of Part (a). Part (b) follows by
a compactness argument, similarly to what we did in the proof of Theorem 3. Considering the
sets (1+)C with 0, we nd a sequence of vector pairs (x
0
, x
1
) such that x
1
x
0
. Since C
is a compact set, there is a subsequence of this sequence that converges to some pair of vectors
in C. Then the sequence of the differences x
1
x
0
also converges, and since all these vectors
belong to , the limit also belongs to .
The proof for m > 1 is very similar, except that we use inequality vol(C) > m vol((B)) to
argue that there is a vector y contained in m+1 of the sets C
z
z.
Theorem 6. Let be a full-dimensional lattice in R
n
and let C R
n
be a convex set symmetric
about the origin (i.e., x C implies x C). Suppose that either
(a) vol(C) >m 2
n
det(), or
(b) vol(C) >m 2
n
det() and C is compact.
Then there are m different pairs of vectors z
1
, z
2
, . . . , z
m
C \{0}.
Proof. It is easy to see that the volume of the set
1
2
C
_
1
2
x : x C
_
is 2
n
vol(C), and therefore,
we can apply Theorem 5 to nd vectors
1
2
x
0
,
1
2
x
1
, . . . ,
1
2
x
m

1
2
C such that
1
2
x
i

1
2
x
j
for all i
and j . Without loss of generality, we assume that vector x
0
is lexicographically smaller than any
other vector x
1
, x
2
, . . . , x
m
(we write x
0
x
i
for all i >0). Dene
z
i

1
2
x
i

1
2
x
0
, i 1, 2, . . . , m.
Clearly, all z
i
s are different; moreover, since 0 z
i
for all i , we have z
i
/ z
j
for all i and j . It
remains to show that z
i
C. Indeed, x
0
C implies x
0
C, as C is symmetric about the origin,
and
1
2
x
i

1
2
x
0

1
2
x
i
+
1
2
(x
0
) C, as C is convex.
5
4 Applications of Minkowskis theorem
Now, we consider some applications of Minkowskis theorem. We shall not use them on our
course, the only purpose of these applications is to demonstrate how wide is the range of these
applications.
4.1 Dirichlets theoremon Diophantine approximation
Let be a real number and suppose that we wish to approximate with rational numbers.
Clearly, we can do it with any precision, but we put an extra requirement on the denominator of
a rational number: it should not exceed a givenQ. Obviously, we cannd a rational number p/Q
such that

p
Q

<
1
Q
. But, in fact, we can do much better. The following theoremwas derived by
Dirichlet using rather elementary methods. However, it can also be viewed as a direct corollary
of Minkowskis theorem.
Theorem7 (Dirichlet). Let be a real number and Q a positive integer. Then there are integers p
and q with 0 <q <Q and

p
Q

<
1
qQ
.
Proof. We assume that Q >1. Consider the following set of points C in R
d
dened by the follow-
ing inequalities:
y <x +
1
Q
, y >x
1
Q
, x <Q, x >Q.
Thus, C is a closed parallelepiped of volume vol(C) 4Q
1
Q
4, symmetric about the origin. By
Minkowskis theorem, it contains a point (q, p) Z
2
(the determinant of lattice Z
2
is 1). We may
assume that q >0. Moreover, q 0 is impossible, since then we would have
1
Q
<p <
1
Q
, whence
p 0. Therefore, 0 <q <Q holds. Yet, the rst two inequalities imply
[p q[ <
1
Q
.
We remark that the theorem can be generalized to the case of simultaneous Diophantine ap-
proximation, where we are given n real numbers
1
,
2
, . . . ,
n
and a positive integer Q, and
the task is to nd rational numbers
p
1
q
,
p
2
q
, . . . ,
p
n
q
with 0 < q < Q, approximating
1
,
2
, . . . ,
n
,
respectively.
4.2 Lagranges four-square theorem
The following theoremwas rst proved by Lagrange and states that every positive integer can be
expressedas the sumof four squares of integers. It is also a consequence of Minkowski theorem,
although not as straight-forward as our previous example.
Theorem8. For every positive integer x, there are integers x
1
, x
2
, x
3
, and x
4
such that
x x
2
1
+x
2
2
+x
2
3
+x
2
4
.
6
Proof. First, we remark that we only need to consider integers x that are prime: if
x x
2
1
+x
2
2
+x
2
3
+x
2
4
and y y
2
1
+y
2
2
+y
2
3
+y
2
4
,
then
xy (x
1
y
1
+x
2
y
2
+x
3
y
3
+x
4
y
4
)
2
+(x
1
y
2
x
2
y
1
+x
3
y
4
x
4
y
3
)
2
+(x
1
y
3
x
2
y
4
x
3
y
1
+x
4
y
2
)
2
+(x
1
y
4
+x
2
y
3
x
3
y
2
x
4
y
1
)
2
.
Thus, let x be a prime number. First, we show that there are integers y and z such that y
2
+z
2
+
1 0 (mod x). For x 2 this is trivially true; otherwise, x is odd. Consider the set
S
1
:
_
y
2
mod x : 0 < y <
x1
2
_
.
It is easy to see that (y
1
mod x) /(y
2
mod x) for different 0 < y
1
, y
2
<
x1
2
: otherwise
(y
1
y
2
)(y
1
+y
2
) (y
2
1
y
2
2
) 0 (mod x),
but both terms in the product (y
1
y
2
)(y
1
+y
2
) are not divisible by x, a contradiction. Conse-
quently, [S
1
[
x+1
2
. Similarly, we may show that
S
2
:
_
z
2
1 mod x : 0 <z <
x1
2
_
.
has cardinality
x+1
2
. Therefore, these two sets must intersect, and this intersection yields y and
z as required.
Now, we choose a suitable lattice and a suitable set to apply Minkowskis theorem. We dene
(B), where
B
_
_
_
_
_
x 0 y z
0 x z y
0 0 1 0
0 0 0 1
_
_
_
_
_
For every vector
_
x
1
, x
2
, x
3
, x
4
_
B
_

1
,
2
,
3
,
4
_
T
, we have
x
2
1
+x
2
2
+x
2
3
+x
2
4
(x
1
+y
3
+z
4
)
2
+(x
2
+z
3
y
4
)
2
+
2
3
+
2
4
(1+y
2
+z
2
)(
2
3
+
2
4
) (mod x)
0 (mod x).
Yet, we have det() x
2
. Consider the ball
B
__
x
1
, x
2
, x
3
, x
4
_
: x
2
1
+x
2
2
+x
2
3
+x
2
4
<2x
_
.
The volume of this ball is
vol(B)
1
2

2
(
_
2x)
4
2
2
x
2
>2
4
x
2
.
Therefore, by Minkowskis theorem, there is a non-zero vector
_
x
1
, x
2
, x
3
, x
4
_
B, i.e.,
0 <x
2
1
+x
2
2
+x
2
3
+x
2
4
<2x and x
2
1
+x
2
2
+x
2
3
+x
2
4
0 mod x,
and the latter is only possible when x x
2
1
+x
2
2
+x
2
3
+x
2
4
.
7