You are on page 1of 36

Annals of Mathematics

An Analogue to Minkowski's Geometry of Numbers in a Field of Series


Author(s): Kurt Mahler
Reviewed work(s):
Source: The Annals of Mathematics, Second Series, Vol. 42, No. 2 (Apr., 1941), pp. 488-522
Published by: Annals of Mathematics
Stable URL: http://www.jstor.org/stable/1968914 .
Accessed: 29/07/2012 15:20
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

Annals of Mathematics is collaborating with JSTOR to digitize, preserve and extend access to The Annals of
Mathematics.

http://www.jstor.org

ANNALS OF MATHEMATICS
Vol. 42, No. 2, April, 1941

AN ANALOGUE TO MINKOWSKI'S GEOMETRY


FIELD OF SERIES

OF NUMBERS IN A

BY KURT MAHLER

(Received November 16, 1939)

i'Iinkowski,in his "Geometrie der Zahlen" (Leipzig 1910), studied properties


of a convex body in a space RI? of n dimensionswith respect to the set of all
lattice points. Let F(X) = F(x1, ***, Xn) be a distance function,i.e. a function satisfyingthe conditions
F(O) = 0, F(X) > 0 if X $ 0;
F(tX) = I t I F(X) forall real t;
F(X - Y) _ F(X) + F(Y).
The inequality F(X) ? 1 definesa convex body in Rn which has its centre at
the originX = 0. Suppose that this body has the volume V. The well known
2 , thenthebodycontains at least one
result of Minkowski asserts that if V
from0. This theorem is contained
(and so at least two) latticepoints different
in the followingdeeper result of Minkowski (G.d.Z. ??50-53): "There are n
in Rnwiththefollowingproperties:
An)
independentlatticepointsX"), X(2) ... , X
(1) F(X(1') = arl) is theminimumof F(X) in all latticepoints X $ 0, and for
k _ 2, F(X(k)) = a(k) is theminimumof F(X) in all latticepoints X whichare
(k-1)
(2) The determinantD of the points X,
independentof V),
the
X(n) satisfies inequalities
.._

1 _ ID

I ? n!.

(3) The numbersa(k) depend only on F(X) and not on the special choice of the
latticepointsX(k), and theysatisfytheinequalities
< __ or(2)

<

<

...

...

(A new simpleproofforthe last part of this theoremwas given by H. Davenport,


Quart. Journ.Math. (OxfordSer.), Vol. 10 (1939), 119-121).
From Minkowski's theorem, properties of general classes of convex bodies
can be obtained. For instance,thereis a convexbodyG(Y) _ 1 polar to F(X) <
T(n)
1, and to this body correspondby the theorem n minima T(1), T(2) ...
I have proved (Casopis 68 (1939), 93-102), that theseminima are relatedto the
a's bytheinequalities
1

T(h)T(n-h+)
488

< (n!)2

(h=

1, 2,...

ANALOGUE

S GEOMETRY

TO MINKOWSKI

OF NUMBERS

489

From this result,applicationsto inhomogeneousDiophantine inequalities can


be made, and in particular,generalizationsof Kronecker'stheorem can be
obtained.
The presentpaper does notdeal with ordinaryconvexbodies in a real space.
The n-dimensional
space Pnwithwhichwe shall be concernedhas its coordinates
valuation I x l; a distance functionis any
in a fieldS witha non-Archimedean
functionsatisfying
F(O) = 0, F(X)

> 0 if X # 0,

F(tX) = I t I F(X) forall t in 9,

F(X - Y) < max (F(X), F(Y)).


The inequalityF(X) < r then'definesthe convex body C(-r),if r > 0. We
show that everyconvexbody is bounded,and that it has propertiessimilarto
a parallelepipedin real space.
In particular,let ftbe the fieldof all Laurent series
x = afzf + af-,zf-Z +

af2Zf-2

..

in an arbitraryfieldf; the valuation I x I is definedas 0 I = 0,


withcoefficients
and I x == ef if af 0 0. Furtherlet Anbe the modul of all pointsin P. , the
in f; these pointswe
coordinatesof whichare polynomialsin z withcoefficients
call latticepoints. We consider only distance functionsF(X) which for all
X 0 0 in P, are always as integralpowerof e. We shall definea certainpositive constantV as the volumeof C(1); thisconstantis invariantunderall linear
of P, with determinant1, and the volume of C(1) and that of
transformations
its polar reciprocalbody C'(1) have the product1. In analogy to Minkowski's
theorem,the followingtheoremholds: "There are n independentlatticepoints
X(I) ... * , (n) in P, withthefollowingproperties:1) F(X(')) is theminimumof
F(X) in all latticepointsX $ 0, and fork > 2, F(X(k)) is theminimumofF(X)
in all latticepointsX whichare independent
of Xl, ..., X(k-1). 2) The deter,
is 1. 3) The numbersF(X(k)) = _(k)) which
minantof thepointsX(') ... X*n)
dependonlyon F(X) and noton thespecial choiceof thelatticepointsX, satisfy
theformulae
<

(1) <

'T(2) <

.<

(n)

(1)

(2) .

(n)

1A

, r(n) forthe distance functionG(Y)


Further,we have similarminima r(1)...
whichdefinesthepolarbody C'(1); theseare relatedwiththe cr'sby theequations
(h) (n-h+)=
(h = 1, 2, ... , n).
1

These two resultscan be used to study special Diophantineproblemsin Pn;


a few of them are consideredas examples. All the proofsin this paper are
based on the methodsof Minkowski,and in one finalparagraphI make use of
ideas of C. L. Siegel.

490

KURT MAHLER

I. CONVEX DQMAINS IN NON-ARCHIMEDEAN


1. Notation.

SPACES

In this chapter, we denote by

9?

an arbitrary field,

lnon-Archimedean
valuationof theelementsx of Ts,'
Qt theperfectextensionof R withrespectto thisvaluation,
Pn then-dimensionalspace of all pointsor vectors

Ix Ia

X =

(X1i

*..*

Xn)y

where the coordinates xi , * , xn lie in 9,


X I thelengthof thevectorX, viz.
(I xi

max

{X]

.n Ixn)

We applyr the usual notation forvectors in Pn; thus if


X = (xi,.**, x,)

and

Y =(y,

**,yn),

and a belongs to .-, then we write

X +T Y = (Xi +- Yl **
aX
XY

IX

(2)

aXI = IaIIX

(3)
(4)

IXFYI

(ax,,

n
E

h-1

For instance, the length I X


(1)

Xn+T Yn),

,
aXn)

XhYh

I of X has the properties:

? 0, with equality if and only if X = (0.*.

Iifaisanyelementof

, 0) = 0;

9;

wmax(I XlIY);

iXY I ?IXI

Y.

If Z is any sub-ring of R, and X"), ... , Xtr) are vectors in Pn , then these
are called TZ-dependent,or Z-indepcndent, according as there exist, or do not
exist elements al , '

ar of Z not all zero, such that


aiX

")

*** +

arX(r) =

0.

A set of vectors of P,, is called a '-modul, if with X and Y it also contains


aX + bY, wherea and b are arbitraryelementsof ); the modul has the dimen'1 This means that the function I x I satisfies the conditions:

oI = 0, but I x

I xy

X T Y !I

Ix

> 0 for x $ 0,

11y 1,

rMax(I X 1, YI)

ANALOGUE

S GEOMETRY

TO MINKOWSKI

OF

NUMBERS

491

sion m, if there are an, but not m + 1, Z-independent elements in it. The dimensionof a Tr-modulis at most n, while that of any other class of moduli need
not be finite.
2. The distance functionF(X). A functionF(X) of the variable point X
in P. is called a generaldistance function,if it has the properties:
F(X) _ 0;

(A):

F(aX) = I a I F(X) forall a in 9, hence F(O) = 0;

(B):

F(X + Y) < max (F(X), F(Y));

(C):

it is called a special distance functionor simply a distance function,if instead


of (A) it satisfies,the strongercondition
F(X) > O forX # 0.

(A'):
If

is a positive number,then the set C(T) of all points X with


F(X) <

is called a convex set;2 if F(X) is a special distance function,then it is called a


convex body. It is clear from the definitionof F(X) that a convex set C(r)
contains the origin0, and that with Y and Y also aX + bY belong to it, if a
and b are elements of f such that I a I < 1, I b << 1. Further,if
E(1)

(1, O 1..

I]

0),

(O)i..1

E(2

...

0))

En)=(.O*-,1

are the n unit vectors of the coordinate system,then

X = xiE(l) +

i.e. F(X) ?

+ xnE(n),

max

h-1,2,- **,n

(IxhiF(E(h))),

and therefore

F(X) < riXI,

(5)

where r is the positive constant


r=

max (F(E

h-1,2.

^))).

,* n

C(T) contains thereforeall points of the cube


IXI

<.

We prove now that for special distancefunctionsthereis a second positiveconstantay,such thatfor all points in Pn
F(X)

(6)

yI X

1.

We consider only convex sets and bodies as defined; they are obviously symmetrical
with respect to the origin.
2

492

KURT MAHLER

PROOF:
We assume that (6) is not true and show that this leads to a contradiction.
By hypothesis,there is an infinitesequence S of points
X(h)

(Xh)

(h

#0

,X(h))

..

1, 2, 3,

*..),

such that
Fj

F(Xh)

Since
F(X)
IX I

F(aX)

I aX I

for all a # 0 in ft, we may assume that for the elements of S


X()

lim F(X(h)) = 0

h%-o

1,

so that in particular the n real sequences


|X(

|, |X

,(i)

X
,X(2)

(k = 1, 2, *.. ,n)

are bounded.
Hence we can replace S by an infinitesub-sequence which we again call S:
, sueb that the n real limits
X(1),
X(2),
X(3),
(7)

ak =

lim

(k =1,

I Xk |h

h-0X

2, ..*,

n)

exist and satisfythe equation


max

k- 1,2,.* * , n

ak=1.

We call S a sequence of rank m, if exactly m of the limitsa,, a2, ..* , an do not


vanish; without loss of generality,these are the m firstlimits al, a2, .**, a".
Obviously 1 ? m _ n.
If the rank m = 1, then forlarge h
iX()I=l,

and

(h) =

(h'

(1,h)

*,h)

U.,

Ekl)+ X*(h)

say, where
lim

|X*(h)

0.

h-0X

Hence by (5)

F(E(1)) = F

(_X*()

_<max

( (h))

II

F(X*(h))

_ ma

=<max(F(x(h)), r

X*(h)j)

and thereforefor h -*

493

OF NUMBERS

GEOMETRY

! 0, i.e. F(E())

0 < F(E(')
which is not true.
Hence the rank m

'S

TO MINKOWSKI

ANALOGUE

=0,

Put

2.

xi(gh)

X(g)

X(h)

Xm

Xm

Then from (7) for large g, h


A(g
h)

< max

F (X

F (X

2 max (F(X(g)), F(X(h))),

<

and therefore

limF(X(Q'))

0.

h-0X
g-4m0

Two cases are now possible:


a: The limit

lim IX(g.)

limmax (I x(gA) |,

g-4

D--#00

hx?)

l)

h-oo

h-Xo

exists and is zero. Hence the n limitsin St


(h)

(8)

(k=

lrn

h-co

Xmn

1,2, ***, n)

all exist, and in particular


x*

lim 1 = 1,

so that
if*

(X1,*--

Xn

i4

0.

By the continuityof F(X),3


F(X*) = lim F
h-co

\ 1h)
~~X~~h)
limF(X(h))0
) am

\Xm

amnh--*o

which is not true.


b: The limit
lim X gh

g-4m0

h-oo

3 If e > O is given, then there is a a > O, such that| F(X)


as follows easily from the properties (B), (C), and (5).

-F(Y)

I < e for X-

i < 5,

KURT MAHLER

494

fromzero. That implies that at


eitherdoes not exist, or exists and is different
least one of the limits (8) does not exist. Now obviously
lim

(k

I=

l X(g'h)

g-900

= m, m + 1***,n),

h-oo

since for large g, h


XMh)

0;

kx(Qh)

_
Xmh

Xi

Xm(g

(i xk

ikmax m
am

i,

j)

Xk

(k = m + 1,

,n).

Hence the index IA of this non-existinglimit (8) is < m - 1. For this index,
lim x,,X

g-Xoo
h-oo

either does not exist or exists and is differentfrom zero. Hence there is an
infiniteone-dimensionalsub-sequence

(i = 1, 2, 3, *.)

X(h)

(9)
of the double sequence X(

h),

such that forall i


I >

I X(g"hi

wherec is a positive constant. Further obviously


lim F(X(g'

h)

O0

i-00

limi x (2$hi)I = 0

(k =m, m + 1,

,n),

i-00

and all m - 1 firstcoordinates


(k =1, 2, ... ,m -1)

X(g,,hi)
are bounded fori m*.
Let i,, foreveryi, be the coordinate

(k

X(gihi)
of maximumvalue Jx(2;ih;)

1, 2,

**

,m - 1)

J;hence

| it | >-C,

I I

since

i.

Then there is an infinitesubsequence


X(gi i Ihii)

(j

1, 2, 3,

..

of the sequence (9), such that, if


X -)

Xoi

aii

(xj

Xn

(j= 1, 2, 3, ... ),

TO

ANALOGUE

495

OF NUMBERS

S GEOMETRY

MINKOWSKI'

then all n limits


lim ix(|

(k = 1, 2, ...,

=a

n)

exist and satisfythe equations


max (al,...

+
am= a =+l

an)=1,

=an=O.

and
0 ? lim F(X'(j))

<

i-*oo

lim F(X(Q

,hii))

0,

i.e.

lim F(X'(j))
0

j-400,

0,

Thereforethe new sequence S'


X/(1)

X/(2))XIW3

has the same propertiesas S, but is of lower rank. Hence by induction with
respect to the rank, a contradictionfollowsalso in this case.By the inequality (6), all points of the convex body C(r) lie in the finitecube
lX

a convex body is thereforebounded. Conversely, if a convex set is bounded,


then it is a convex body. For if its distance functionF(X) is not special, then
thereis at least one point X(0) 5 0, such that F(X(?)) = 0; hence all points of
the straightline passing throughX(?) and the origin 0 belong to the set.
3. The character of a convex body. Let C(r) be a convex body, F(X) its
distance function. If X' - 0 is an arbitraryvector, then the point X = aX',
where a is an element of S, lies in C(r) provided that Ja I is either sufficiently
small and positive, or 0. Hence for every index h = 1, 2, *., n, the set Sh
of all points
X = (xi,

with

,xn)

x=

Xh-1= O

Xh# O

of elements. By (6),
of C(r) is not empty and contains an infinity,
| Xh |<

forthe points of Sh. Therefore x,


I Jhas a positive upper bound
and to every e > 0 there is a point
Xh

*=

(x (h)

th

in this set,

(h),

for which
F(Xt

) <

7r

Xh-le =0

?_

1+

<

496

KURT MAHLER

whereasthereis no point X for which


F(X)?r -,

**=Xh=O,

X=*

Xh

>

The systemof the n points


(1)

x,,2)

x,(n)

correspondingto e is obviously a-independent, and any point X of Pn can be


writtenas
X = ul.X?() +

+ UneX ')

where the u's belong to S and are given explicitlyby


n

k
Uhe

k-1

(h =

ahke Xk

1, 2, *,n)

with a matrix
(ahke)

h o1 92, *,n

of non-vanishingdeterminantand elements depending on e, but not on X.


We distinguishnow whetherthe valuation i x i of S, is discreteor not.
If i x i is discrete,then thereis a constant b > 1, such that forall x F 0 in
I xj = bg
4 If I x I is discreet,then u'(X) has a similar property: The set a of its values for X in
Pn has no pointof accumulationexcept0. This is clear forn = 1, forthen all vectors are

multipla of the unit vector (1). Suppose that the statementhas already been proved for
all spaces of n - 1 dimensions,but that it is not true in Pn . There is thereforean infinite
sequence 2 of points
X(k) =

(X(k),

(k = 1,2,3,*

X(*))

..

in P, , such that all numbers


F(X(1)),

F(X(2)),

F(X(3)),

are differentfrom each other, and that the limit


lim F(X(k)

= X

k-_o

exists and is positive.

Write
X(k)

Xjk) E()

X(k)*

(k =

1, 2, 3,

1, 2, 3,

*..)

where
X()*=

lies in the (n - l)-dimensional


2:; hence we may assume that

(0,

X(k),

subspace Pnp1: xi = 0, of Pn.


lim I x(*) J =

k-ao

(k

x(k))

By (6),

x(k) I is bounded in

ANALOGUE

TO MINKOWSKI'S

GEOMETRY

with a rational integerg depending oil x.


satisfiesthe equations
he

OF

497

NUMBERS

In this case the set of values I x

(h = 1,2 ... * n)

forall sufficiently
small e. We assume that e is sufficiently
small and omit the
index e. Put
4P(X)

(Ilu 1,*

r max

I, ua

1) = r

max

h-1,2,- * *,n

(E

k-1

ahkXk)

Theii obviously
F(X) ? X,

if

+,(X) <

T.

Conversely let X be any point in C(T). Then

I X1I - 41
and therefore
lul

lxii?(1) 1.

=
-

we can replace ^ by an infinitesubsequence. If ,


since, if necessary,

sufficientlylarge k

F(X(k))

so that the sequence


hypothesis on P,1.
Hence if

X(1)*, X(2)*,

Xjk+1)

X(3)*,

4~k)

q(k),

0, then for all

F(X(k*))

has the same properties as 2, contrary to the

...

lim I q(k) 1

then

1,

so that for all sufficientlylarge k


q(k)

1.

Obviously
X*(k)
lies in P,j1,

q(k)X(k) =X(k+l)*

X(k+-)

q(k)X(k)*

and for all large k


F(X(k))

F(q(k)X(k))

$ F(X(k+l)).

Hence
F(X*(k))

max (F(X(h)), F(X(k+l))).

Therefore the sequence of positive numbers


F(X*(l)),

F(X*(2)), F(X*(2)),

contains an infinityof differentelements and has the limit X, so that again a contradiction
is obtained.

498

KURT

MAHLER

Hence, if
,

U1X

(O,

En),

then
F(X')

< max (F(X),

I ul F(X('"))

<

T.

and so X' also belongs to C(T). Therefore


l X2

so that
1u2!

___

1.

?2l

Ix2)

Continuingin this way, we obtain all inequalities


Iu I _

1,*,unI<1,

i.e. we have proved

t,

4)T(X) ?

if

F(X)

< r.

The domain definedby


T

1~~~~~~~~
- mT(X)
=

CEhkXk

h-1,2, - -, n

jk-e

?1

is called a parallelepiped; our result may thereforebe expressed in the form:


If thevaluationI x I is discrete,thaneveryconvexbodyC(r) is a parallelepiped.
As we have proved, the two domains

F(X) < r

and

dT (X)

<

are identical. In general, this does not imply the identity'


F(X)

b(X)

for all X, and the function47(X) depends on r. Suppose, however, that the
set of values of F(X) is the same as that of the values of I x 1,and that r is also
an element of this set.6 Then
bT(X) = 4(X)
becomes independentof r, and forall X in Pnidentically
F(X) =

(10)

(X))

as followseasily fromthe property(B) of the distance functions.-E.g., if 91= R is the p-adic field(p > 3), n = 2, and
F(X)

= max (I x1 |P, 2 | x2 Ip)

6
It sufficesto assume that F(X) does not assume every positive value, and that the
equation F(X) = T has no solution.

ANALOGUE

TO M.INKOWSKI

S GEOMETRY

OF

499

NUMBERS

Next assume that thevaluation j x I is not discrete,so that its values lie everywheredense on thepositivereal axis. Now the n vectors
x

X (2)

(1)

(n)

will depend on e, and so does the function


4TE(X)

max

(I Uh, I)

h-1,2,---,n

max

(
\

h-1,2,-,n

k-1

ahkeXk

Evidently
F (X)

Conversely,suppose F(X) <

r.

(11)

if

bTf(X)

Then

xI <?
and therefore
I Ulfl

-(l) I < 1+

Ui I

Hence, if
X,

UlfXI)

xi

(0,

then

< max (F(X), IUlf IF(Xl)))

F(X')

< (1 + ()r.

There is a numbera, in ftsuch that


F(XYe) ? j

(1 + O)r, i.e.

|?;
Ijr

F(a1 X') _

T-

Hence
4

?
-2,

jalX2j

?-

(1 +

E)02,

and therefore

UeI

< (1+

jUE

I2

so that, if

Xe

'U2eXf

(u u2eXel)
X? +-

(0, 0, X

***

4)

then

F(X') ? max (F(X'), I U2e I F(XE2)) < (1 +

E)T

Continuingin the same way, we obtain the n inequalities


I Uhe

I <

(1

(h =

E)

hence
(12)

bT((X)

<

(1 +

E)

if F(X)

From (11) and (12), since e > 0 is arbitrarilysmall:

<

1, 2,

n),

500

KURT

MAHLER

If thevaluationi x I is everywhere
dense on thepositiveaxis, thentheconvex
bodyC(r) can be approximatedarbitrarily
near bothfromtheinside and outside
bymeansofparallelepipeds.
Take now, say r = 1 and put
'ib (X)

(Di(X)

max

h=1,2,-***,n

(|

ahkaXkcx)

k=1

To everypoint X, thereare two elementsa and j3 of 9, such that


4,(X)

< I a I < (1 + E)cibe(X)

F(X) < |

and

< (1 + e)F(X).

Hence from(11)
F

(a)1

.(,,)

F (X) _ 1a _ ( + e) 4,(X),

and from(12)
F (X)<_ 1,

(e

(1 +

E)n,

be,(X)

(1

E)I

n<

(1

.) n+1 'F(X)

and thereforeuniformlyin X
(1 +

(13)

F(X) ? (1 + e)16(X).

?(X)

E)-(n+l).

In general,these inequalitiescannot be improvedto an equation analogous to


(10), e.g. ifF(X) = r has no solution.
4. The characterof a convex set. If F(X) is not special, then the set M of
all solutionsof F(X) = 0 containselementsotherthan X = 0. From (B) and
(C), with X and Y also aX + bY belongsto M, if a and b are elementsof R.
Hence M is a t-modul,say of dimensionn - m. Obviouslym < n; it is possible that m = 0, but then F(X) vanishes identicallyand C(r) is the whole
space. Suppose therefore,
that 1 < m < n1, and let
p(m+l)

be n

p(m+2)

p(n)

mg-independentelementsof M,
p(l)

p(m)

p(2)I...

m otherpointsof P , so that the systemof n vectors


p(l)

p(n)

p(2))..

is still h-independent. Then everypointX in Pn can be writtenas


X

with elements vi,

.-*,

v1P(

) +

v. of f, viz.
Vh =

* +

n
Z fhkXk

k=1

VnP(n)

(h = 1,2, ..., n),

ANALOGUE

OF NUMBERS

GEOMETRY

TO MINKOWSKIS

501

wherethe constantmatrixin ?
(flhk)h,k=1,2,...,n

has non-vanishingdeterminant. Since


n

h=m+l

=0 O

VkP)

we have
F(X)

= F(

vP(h)

=*(V)

where
n

n
*(Vt)

Vm) =

@*

(VI,

fkl

kXA;

#nkXk)

space Pmof
is now obviouslya special distance functionin the m-dimensional
all points V = (vI, **, vm). Every convex set with m > 0 can therefore
be consideredas a cylinder,the basis of whichis a convexbody of m < n dimensions.
5. The polar body of C(X). Let F(X) be the generaldistancefunctionof ?4,
Y an arbitraryvectorin P,. Then we definea functionG(Y) by
(14) G(O) = 0; G(Y) = lim sup (I XY j) forall X withF(X)

1, if Y F 0.

In orderto determinethis function,let


Q(2), ...

Q(l)

Q(n)

be the n pointsin P,,, whichsatisfythe equations


p(h)

Q(W)

{1 forh =
forh

k,
,
ky

and write
Y = w1Q() +?

WnQ

then
n

WA =

(h

hkYk

Be
k=1

wherethe determinantof the matrixin St


(7/hk)h,k=1,2,...,n

does not vanish. Then


XY

V1W1 +

Vn~n

1, 2, *.*.* n),

502

KURT

MAHLER

Hence obviously

G(Y) = x,

unless

= Wn = 0.

wm+1=

Suppose thereforethat
(15)

W+l

Wm+2 -

Wn

and put
G(Y) X=(W)
where W
(16)

, Wi) is a vector in Pm. Then from(14),

(wi,

X(0) = 0;

X(W) = lim sup (I VWVI)forall Vwith'I'(V) _ 1, if W'

0,

so that the relation of X(W) to 4(V) is the same as that of G(Y) to 17(X).
By ?4, I(V) is a special distance function,and so is X(W), as follows easily
from(16) and the properties(A'), (B), and (C) of [(V).
We call G(Y) the polar functionto F(X); for ?n < n it is not itselfa distance
function,but becomes one in the m-dimensionalspace (15), where it coincides
with X(W). The set C'(1/r): G(Y) < 1/r, is furthercalled the polar set to
C(r); it lies entirely in (15) and here is identical with the convex body
X(W) = I/r.
Suppose now that rn = n, i.e. both F(X) and G(Y) are special distance functions; then the polar set C'(1 1-) becomes a convex body. We shall prove that
in this case therelationbetweenF(X) and G(Y) is reciprocal,i.e. F(X) is the polar
functionto G(Y) and C(r) the polar body to C'(1/r).
This assertion is evident, if F(X) = X 1,for then obviously G(Y) = IYI.
Further let
=

(ahk)h,k-1,2,...,n;

(ahk)hk-1,2...,n

be an arbitrarymatrix in Pt with nonvanishingdeterminant,and its complementarymatrix,so that forall X and Y the scalar product7
UX4qKY

XY.

Then the transformeddistance functionsG'(Y) = G( Y) and F'(X) = F(UX)


have still the propertythat the firstone is polar to the second, since
G'(Y)

= G(O KY)

= IiM

sup (I X.

F (X) S1

YI)

uiMsup (IOX.KYI)
F (fX)?1

= limsup (IXYI).
F ' (X)?1

Further,if F1(X) and F2(X) are two distance functionssuch that for all X
F1(X) <. 12 (X),
= (x,
7The vectorX'

, x )=QXis

definedbyx,

aEd

X for
1,2,*
h=

, n.

ANALOGUE

'S

MINKOWSKI

TO

OF NUMBERS

GEOMETRY

503

then the polar distance functions G1(Y) and G2(Y) satisfy the inverted inequality

G1(Y) > G2(Y).

We distinguishI1ow the same two cases as in ?3.


discrete,then we showed the existenceof a matrix
A

If the valuation I x I is

(ahk)hL,k=1,2,...,,

in f with determinantdifferentfromzero, such that


identically in X.

F(X) = D(X) = I AX I
The polar functionto F(X) is therefore

G(Y) = IAKYI,
and since (AK)K = A, the statementfollowsat once.-I-n this case, the definition
of G(Y) can obviously be replaced by the simpler one:

~~G(Y) =max IXYI

(17)

x&o F(X)
( )
(17)
Secondly, let I x I be everywheredense on the positive real axis. Then to
every 5 > 0, thereare two matrices

A1=

and

(c))h,A=1,2,...,n

A2 = (ahk )h,k-1,2,...,n

in 9 with non-vanishingdeterminants,such that if


F1(X)

| A1X I,

F2(X)

| A2X

then forall X
F1(X) < F(X) ? F2(X)
as follows easily from (13). Hence if

G1(Y) =

AKYI,

(1 + 5)F1(X),

G2(Y)

A YI

=I

are the polar functionsto F(X) and F(X), then also


G2(Y) < G(Y) < G,(Y);
and8
G2(Y)'
8

(1 + 26)G1(Y)

There is a number a in f sUhelthat


1 +

TlieIi by hypothesis

6 ?

polar
IFIA)

<

1 +

26.

(1 + 6)1F2(X) _ F22(aX)

Hence
1+26

ys tlle polar
since

fulletioll

to P'.(a.X)

Y) <
is G,.a

G2

(Y),

504

KURT MAHLER

forall Y. Since 6 can be taken arbitrarilysmall,the assertionfollowsagain for


the same reason.-In this case, the definitionof G(Y) is easily replaced by
G(Y) = lim sup IXY
X31o F(X)'

(17')

By the proved reciprocityof F(X) and G(Y), the formulae(17) and (17') remain trueif G(Y) is replacedby F(X) and vice versa.
II.

"GEOMETRY

OF NUMBERS"

IN A DOMAIN OF POWER SERIES

6. Notation. We specialize now the fields9S and 9 of ?1, and denote by


field,
f an arbitrary
z an indeterminate,
in C,
4k[z] theringofall polynomialsin z withcoefficients
? = k(z) thequotient
fieldof Z, i.e. thefieldof all rationalfunctionsin z with
in a,
coefficients
the special valuationof ? definedby
IxI
0=

if x =0,

1ef, if x $ O is oforderf,9
t theperfect
extensionof 9Twithrespectto thisvaluation,i.e. thefieldof
Laurent
series
all formal
X =

afZf + afiZf-'

afr2-2Z

...

withhighest
in f; if af is thenon-vanishing
coefficient
withcoefficients
index 0, thenIx = e
An thesetofall "latticepoints"in Pn, i.e. thatofall pointswithcoordinates
in Z.
The valuation I x I is by definitiona powerof e withintegralexponent. We
assumethe same forall distancefunctionswhichwe considerfromnow onwards,
and we shall consideronlyconvexsets or bodies C(r), wherer is an exact power
ofe, say r- = e.
7. The volume V of a convex body C(1). Let F(X) be a special distance
function,C(et) the convexbody F(X) < et,wheret is an arbitraryinteger. It is
obvious that the set m(t) of all lattice pointsin C(e') formsa f-modul. In the
special case F(X) = X 1,this set has exactly
Mo](t)= n(t + 1)
elements. Hence, by theinequalities(5) and (6), m(t)has always
C-independent
a finitedimensionAM(t),and thisdimensionis certainlypositiveforlarge t.
9 The order of a rational function is the degree of its numerator minus the degree of its
denominator.

ANALOGUE

TO MINKOWSKI

'S

GEOMETRY

OF

NUMBERS

505

Obviously
Mo(t + 1) = Mo(t) + n.

(18)

Suppose that t is already so large that


>

et

Then a lattice point in C(et+l) can be writtenas


X = Xo + zX1,
whereX0 and Xi are again lattice points, and the coordinates of X0 lie in t, i.e.
X0j ? 1,

F(Xo) ? r ? et+'.

Hence

F(zX1) ? max (F(X), F(Xo)) <

F(Xi) ? et,

et'+

so that X1 lies in m(t). Conversely,if Xi belongs to m(t), then


F(X) ? max (F(zX1), F(Xo)) ? e".
Now the two vectorsXo and zX1, whereXo and X1 are lattice points and I xo0 <
1, are f-independent,and the X0 forma f-modulof dimension n. Hence
M(t + 1) = M(t) + n.

(19)

The two equations (18) and (19) show that for large t, the functionM(t)
Mo(t) of t is independent of t. Hence the limit
V

(20)

lim eM(t)-Ho(t)

t-00

exists; it is called the volumeof theconvexbodyC(1).10 In particular,if F(X)


X 1,then obviously V = 1.
8. The invariance of V.
Q

Let

(ahk)h,k=1,2,...,

and Q' =

(ahkJhk==1,2...,n

be a matrixwith elementsin S and determinantD # 0, and its inversematrix.


The linear transformation

Y = OX or X = Q

changes F(X) into the new distance function


F(Y)

= F(X) =F(QY);

let C'(et) be the correspondingconvex body F'(Y)


of C'(1). Then
(21)

<

et,

and V' the volume

V' = ID IV.

10Thisdefinition
is analogousto thatofthevolumeofa bodybymeansoflatticepointB
in an ordinary
reaispace.

506

KURT MAHLER

PROOF: We denote by m'(t) the r-modulof all lattice points in C'(e'), by


M'(t) the dimensionof m'(t), and prove the statementin a numberof steps.
1: The elementsof S lie in ?, and D belongs to f.
The formulaeY = QX, X = W'Y establish a (1, 1)-correspondencebetween
the elementsX of m(t) and Y of m'(t). Obviously,this correspondencechanges
every linear relation

a1Xl

+ *

+ arX( )

withcoefficients
in f into the identical relationin the Y's, and vice versa; therefore f-independentelements of m(t) or m'(t) are transformedinto f-independent members of the other modul. Hence both moduls have the same
dimension:M(t) = M'(t), q.e.d.
2: Q is a trianglematrix
all
0

a2l a22
an2

an,

...

ann

an

with elements in S and determinant

D = alla22...

0.

The equation Y = OX denotes that


y,

a11x1,

Y2

a21X1

Yn =

an1X1 +

a2X2,

an2X2 +

annuX;

henceeverylattice point Y can be writtenas11


Y = OX* + Y*1
whereX* and Y* are again lattice points and Y*
inequalities
Iy* I <

Therefore

I all

I y*
2 I|<

I a22

|,

(yr, *

..
** Y*l2|o

<

y) satisfiesthe

|jann |Ia

Y* I < cl, i.e. F'(Y*) < c1r',

wherecl is a positive constantdependingonly on Q, and r' is the constantin (5)


belongingto F'(Y ). The set of all vectors Y* formsa f-modulm* of dimension d, where
ed =

IaiiI1a221 ...

Iann

IDI.

11We use the trivial lemma: "To a and b = 0 in Z thereis a q and an r in X, such that
a = bq + r and I r i < i b 1.

ANALOGUE

TO MINKOWSKI

S GEOMETRY

OF

507

NUMBERS

Let t be so large that


e' _ cC'.
Then for X* in m(t)
F'(Y)

= F(Q'Y)

= F(X* + QiY*) _ max (F(X*), F'(Y*))

< e',

and converselyfor Y in mn'(t)


F(X* + Qi2Y*)

et,

i.e. F(X*) ? max (F(X* + QIY*), F'(Y*))

et.

There is thereforea (1, 1)-correspondencebetween the elements Y of m'(t) and


the pairs (X*, Y*) of one elementX* of m(t) and one element Y* of m*. Hence
M'(t) = M(t) + d, q.e.d.
3: The elementsof Q belong to Z.
The result follows immediately from the two previous steps, since Q, as is
well known,12can be writtenas Q = Q1Q22, wherethe two factorsare of the
classes 1 and 2.
4: The elementsof Q lie in S.
Now Q = QQ2 whereboth Qa and Qb are of the class 3, so that the statement
followsat once.
5: Q has elements in A, such that
ID I = 1,

I ahk l_ 1

(h,

k = 1, 2, ... , n).

Then the same inequalities hold for the inverse matrix Q?, so that for every
point X
IQXI _ AXIS

I XI =i2IQQX

_
I

QX

and therefore
X=

I.

IQXI = IQ

Now to every lattice point X thereis a second lattice point Y such that with
a suitable point Y*

QX= Y+

Y*

IY*

<

1;

then conversely
Q?Y = X + X*,

X*I < 1

and
X* = -QIY*

QX*

Y*.

The relation between X and Y is thereforea (1, 1)-correspondencewhich obviously leaves invariant the propertyof f-independence. Suppose that

e' > r.
12 This can be proved, e.g.
by a method analogous to Minkowki's "adaptation"
tice; Geometrie der Zahlen ?46.

of a lat-

508

MAHLER

KURT

Then for X in m(t)

F(X*) < r < e',


and therefore
F'(Y)

= F(X + X*) < max (F(X), F(X*))

= F(i'Y)

< et,

so that Y lies in m'(t); conversely,if Y belongs to m'(t). then X is an element


of m(t). Hence M(t) = M'(t), q.e.d.
6: Finally, let Q have elementsin R. Then it can be split into

Q = Q4 + Q*
whereU24is of the class 4, while the elements of Q* lie in 9 and have so small
values that
is of the class 5. Then the resultfollowsat once, since Q = Q4Q5
Two conclusions are immediate from (21). The convex body C(et), i.e.
F(z-'X') ? 1, is obtained fromC(1) by the transformationX' = ztX; hence
it has the volume V(e') = entV. Secondly, let G(Y) be thepolar distancefunctionto F(X), and V' thevolumeof theconvexbodyC'(1), i.e. G(Y) ? 1. Then V
and V' are relatedby theequation
VV'=

(22)

1.

For by ?5, thereis a matrixA with non-vanishingdeterminant,such that

F(X) =

and

AX|

G(Y)

AKY 1,

hence
V=

and

QAI)

V'a=

(IAkIj)-

IA

the statement is thereforeobvious.


To the distance function F(X), there exist n R-

9. The minima of F(X).


independentlattice points

Xik

(Xk)

x(k)

(k = 1, 2, *

, n),

such that

F(X(1)) = a(1) = e"1is the rntinimm of F(X) in all latticepointsX $ 0,


F(X 2))

o(2)

F(X(n))= (n)

e"2 is theminimumof F(X) in all latticepointsX whichare

egn

of X(l), etc., and finally


f2-independent

is theminimumofF(X) in all latticepointsX whichare


X(2) ... , x(n-1)
2-independent of Xylem

The numbers -(1) (2), * * , a- are called then successiveminima ofF(X).


this construction,the determinant
D

|~

Ihk==1,2,..n

By

ANALOGUE

TO

'S

MINKOWSKI

GEOMETRY

OF NUMBERS

509

lies in : and does not vanish; furtherobviously


(23)

<

(1) < ?(2)

<

(n)

...?<.

and

<1-<2<*..-gn

'We shall prove the two equations


IDI = 1,

(24)
(25)

a()

(n) =

(2)

in the second one, V is again the volume of C(1). Thus, in particular,D is an


elementof F, and may obviouslybe takenas equal to 1.
A: PROOF OF (24). Every point X in Pn can be writtenas
X = y1X( ) +

* ** +

wherethe y's are elementsof R. Then the coordinatesXhof X are linear functions with determinantD of the coordinates Yh of Y = (y1, * , yn). We
definea new distance functionI1(X) by
I1(X) = I Y
By (21), the convex body 11(X) < 1 has the volume I D 1; we determineit in
the followingway:
If X is a lattice point, their Y also has its coordinates yhin Z. For since
with Y also X is obviously a lattice point, we may-assume without loss of
generalitythat
(26)

IY < 1,

II(X) =

and have to show that no lattice point X F 0 satisfiesthis inequality. Let m,


where 1 < m < n, be the greatest index for which ymid 0. Thei
X

jE

h-1

XhX,

X(l

e.

X(m-1)

are T-independentlattice points, and by (26)


F(X) ? max (I yl F(X(l)),

y* I|m I F(X(m))) <

('

in contradictionto the minimumpropertyof jm)*


Hence there are exactly Mi(t) = n(t + 1) f-independentlattice points such
that 11(X) < et,viz. all points correspondingto a basis of f-independentpoints
Y with I Y I < Ct. Therefore
DI =
B: PROOF OF (25).
writtenas

ne

(t)-

_ 1

t-00

q.e.d.

Now we use the fact that every point X in Pn caii be


X =

YJZ-gX~l)

Z-tnX(n7

510

KURT

MAHLER

wherethe y's belongto fP. Let .(X) be the distancefunctiongivenby


z(X)

jYj.

Since
(h = 1, 2, ** , n),

F(z-g?^X(^) = 1

obviously
F(X) < 1, if 2:(X) _ 1.
But the converseis also true:If
F(X) ? 1, then 2(X) ? 1,
and therefore
evidently

IY I,

F(X) = 2(X) =

identically
in X.
For supposethaton the contrary
fora certainpointX in Pn
F(X) ? 1, but 2(X) > 1.
Thenlet m with1 < m ? n be thegreatestindexforwhichI YmI > 1; hence
if m < n

Ym+l|1

@*X|Yn |_1

Write
Yh =

zyh +

(h = 1, 2, ..

Yh

n),

wherethe yZ are elementsof X, the yh**elementsof 9, and


Ym506 01

Ym+l

Yn = ?0,

Yn |S

,*

and put
Y = (y1,

yn)

= (Y1

Yn*)

so that
Y = zY* + Y**.
Obviously,Y* is a latticepoint,Y** a pointsuch that Y** < 1. Alsowrite
n
m
n
X*

h-1

Z yghX

zgOAX(h)

h-1

so that

X = zX* + X**.
Then fromZ:(X**) =

Y**

1,
F(X**) ? 1.

X** =

h-1

-h

(h)

ANAIDOGUE

'B GEOMETRY

TO MINKOWSKI

OF

NUMBERS

511

Hence
F(zX*) < max (F(X), F(X**)) < 1,

F(X*) < 1,

and
F(X0) <

whereX0 = z"X*.
points

(i)

sincethe m lattice
is impossible,
This inequality,however,

XO=

m
E

h-1

X)X

yAm-ghX(h)

X(-)

...

are S-independent,so that by the minimumpropertyof a(m)

F(X(?') -> (r(m).


(27) is true,so thatby the invariancetheoremof ?8
Therefore
V

(1) a(2)

=(~)a2

(n)

.1

~)-

of X into Y has the determinant


sincethe transformation
* +g3)

Dz-(g1+g2+

proved.
The equation(25) is therefore
Fromthisequationand from(23) in particular
(1)

<

i.e. to everydistancefunctionF(X) thereis a latticepoint X # 0 such that

F (X) <

V1

Here equalityholdsif and onlyif all minima


a

(1)

(2)

*.

(n)

a,

thuscertainlynot,if V is not an integralpowerofen.


10. The relationsbetweenthe minimaof F(X) and G(Y). To the n lattice
n points
pointsX), X2 *... ,IX-n) definedin thelast paragraph,we construct
y(l)

y(2)

(27)

since I D
numbers
(28)

y(n)

satisfying

X(h) y(n-k+l)

=1

for h = k

t0 for h sPk;
= 1, thesepointsare latticepoints. XWe
definen positive
further

~(n-h+l)

(h

= 1 2 ...

n),

512

MAHLER

KURT

so that
?

(29)

<

(1) <

r(2) <

n)

<

and

j1

-< j2

-< * ..<

Jn

Then F(X) and the polar functionG(Y) can be writtenas


(30)

F(X) =

(31)

G(Y) =

max

(a(h)

| Xy(n-h+l)

I)

max

(r(h)

i YX(n-h+l)

I)

h-1,2, ** n
h=1,2,- * *,n

thus in an entirelysymmetricalway. For we proved in the precedingparagraph


that if X is writtenas
n

Z yhz

X =

(32)

X(h)

h=1

then

F(X) = |Y,

Y=

y(l) we get by (27)

But by multiplying(32) scalar with y(n)2 ...,


Z * (XY(

Yh =

* yn)

(Y1,Y2X
*Y

(h =

1, 2, *

, n)

and therefore(30). The formula (31) is a consequence of (30) by the results


in ?5.13
From (27) and (31)
(33)

(h) = ejh.

G(y(h))

We prove now that these numbers-(h) in theirnatural orderare the n successive


minimaof G(Y) in An. Obviously it sufficesto show that if
g(1)

(2)

Z(n)

are any n R-independentlatticepoints, such that


G (Z 1)) <
13

G(Z(2))

<

G (Z(n)),

..<

We can prove (31) directly in the following way: Obviously


n

X =

E3(XY(n-h+l))X(h),

h-1

where the brackets are again the scalar products.


G(Y)

= max (I XY l) = max(

Hence from (14)


(XY(nh+1))(X(h)Y)

where the maximum extends over all points X of C(1), i.e. for which
|XY(n-hil)

|'

1_

= 7(n-h+l)

(h = 1, 2, .,

n).

By choosing X such that there is equality in one of these conditions, but that all other
vanish, the assertionfollowsafterreplacing h by n - h + 1.

scalar productsXY(n-h+1)

ANALOGUE

TO MINKOWSKI'

5 GEOMETRY

513

OF NUMBERS

14
then

G(Zh

h)

(yh)

Considerthe n + 1 vectors
(1)

X(2)Y.

.I

(n-h+l)

z(h)

Z(1)IZ(2)

At most n of these are k-independent;hence the scalar products


i

XWiZ(j)

-11

2) ..

n -h + 1)

do not all vanish simultaneously,and at least one of them, say X(')Z(j), is


fromzero. Since it is an elementof Z, therefore
different
X(i)Z(j) I > 1.
Now by (17)
l XY

forall points X and Y.

I _ F(X)G(Y),

Therefore

1 ? |X" Zi'j ? F(X('))G(Z(j)) ? F(X(nh+l))G(Z(h))

as was to be proved.
From (28) and (29) in particular
(34)

(1)\~
/n1v-1n1~

(1 <

and xr(1? (<-'l)"l',

so that if the minimumof F(X) in ? is small,thenthe same is trueforthat of


G(Y), and vice versa.
11. The relationbetweenthe homogeneousand the inhomogeneousproblem.
The reciprocityformulaeof the precedingparagraph can be applied to inhomogeneousproblems. Let P be an arbitrarypoint in Pn whichis not necessarilya latticepoint; it can be writtenas
P = p1X"'

...

(8)

wherethe p's lie in Mt. Put


Ph =

Xh+

rh

(h

= 1, 2,

n),

whereXh is an elementof Z and


|rh

14

<-

The minimav(h) of F(X) have the analogous property.

(h = 1, 2, ... ,n).

514

KURT

Then the latticepoint X = (xi,

MAHLER

,
*,n)
satisfiesthe inequality

*.

F(P + X) =F(ZrhXc))

(n)

<

or by (28)

F(P + X)

(35)

This inequalitycannotin generalbe improved,


since

F(1 ~~X(n)
+ X)>eTl

(36)
(36)F

of the
forall lattice pointsX, as followsimmediatelyfromthe Se-independence
n vectors
XM I X(2

* n**,

n
X)

ZX.

These twoinequalities(35) and (36) relatethe inhomogeneous


F-problemto the
homogeneous
G-problem,in analogy with similar relations in many parts of
mathematics.
As an application,considerthe two polar distancefunctions
F(X)

max (l axn

G(Y) = max ( y 1,*

xi 1, **

Ian-lXn - Xn- 1,e

t IX

, I Yn- 1 et I aly1 + *** + an-lyn-1+ Yn |)

where t is a positive integer. Assume that the numbers1, al, ...


T-independent, so that forall lattice points Y = (Yi, **, yn) $ 0
aiyi +

* - - + an-lyn-1

+ Yn 5$

I,an- are

0.

Then, as t -- c, the firstminimum r(1)of G(Y)


T(1)

)*

0)C

Hence by (35), foreverye > 0 and foreverypointP = (pi, *


latticepoint X = (X1, * , Xn) satisfyingtheinequalities

I alXn

X1 + pi I < e, - - * ,I an-lXn-Xn-

+ Pn-1

, pn)

thereis a

|<E

Thus we have establisheda resultanalogous to Kronecker'stheorem.


12. A propertyof matrices. Let
Q =

(ahk)hk=1,2, .*,n

be a matrixin S withdeterminant
1; thenthereis a matrix
UT =

(uhk )hk 12e

S GEOMETRY

TO MINKOWSKI'

ANALOGUE

OF NUMBERS51

1, such thattheproductmatrix
in Z and determinant
withelements
Qi* =(a~h)h,k=1,2,

U=

..

satisfiestheequation

max (a~I)=1

h=1 k=1,2,- ,

PROOF

:15

To the convex body C(1) belongingto the distance function


F(X)

thereare n latticepoints Xl)


the n minima

max2.,n

XI)

Z:ahk

,X~n

...

F(X (h))

=-

Xk

of determinant D

(h

(h)

=1,

such that

=1,

2, ..

n)

satisfy

Let~XX=(xk,

) and X be the matrix

n.

(h)~~,,-,

with elements in Z and determinant 1. We introduce new coordinates


,i ..1Yn by putting

X = yiX('1+..

i.e., Xh

YnXn

thenF(X) changesinto
F(X)

F'(Y)

max

(h =1,2,.n)

x X~k

aikk)

where
=a

The n points X

=X(h)

oix.

kh~==,

into Y
are transformed
FP(E (h))

E
E(h) (h

(h)

,n);

hence

1, 2, .

n,

1, 2, ...
(h

that is

(37)

ahkI
max~~~~~~(

()(k

=1, 2, ...

,n).

15Ananalogous
in the real fieldwas provedsome timeago by C. L. Siegel in
theorem
a letterto L. J. Mordell. The presentproofand theorem,thoughnot stated in Siegel's
paper,are obtainedfromit withonlyslightchangesby makinguse of the resultsin ?9.

516

KURT MAHLER

Hence everyminorAm of orderm formedfromthe m firstcolumnsand m arbitraryrows of Q' satisfiesthe inequality


|
jAd <(1)f

(38)

(2) ...

a(i)

On the otherhand, any determinantA of orderm can be writtenas


A

n
=

xahth,
h=1

where the ah are the elementsof its last column,and the ah their cofactors;
therefore
max

({6h|)

h-1,2t* * -,n

Al

>

max

h=1,2, * n

(iah DV}

We apply thisinequalityrepeatedlyto the determinant


(1)

An

(n)

(2)

o' **.o'
1oof ?2'and use (37) and (38); thenit followsthat thereexists
an (n - 1)th orderminorAn-, ofAnformedfromthen - 1 firstcolumnsof R'
and satisfying
(n-1)

| Ao.1j =o (1), (2)

fromthen
an (n - 2)th orderminorAn-2ofAn-1formed
and satisfying
(2)

|I A~...2j~(1)

2 firstcolumnsofU'

(n-2);

etc.; a secondorderminorA2 of A3 formedfromthetwofirstcolumnsof U' and


satisfying
A2

(1)

(2)

A1of A2 lyingin thefirstcolumnof i' and satisfying


and finallyan element

I A1I =

LT1.

Without loss of generality,we may assume that the determinantsso constructedare exactlythe principledeterminants

Iahk Ihk=1,2,-..

Ar =

(r = 1, 2,

*.

We shall now constructa set of matricesof ordern


01 0 .-. O
0
...*g**

..*~

Um

g(f)

o2
.

n',i

0
1

O ...0 .
?

...
*

m rows
?

1)

in-m
1
1
t~~~~~i

rows

(m

*
1) 22

n).

ANALOGUE TO MINKOWSKI'S

517

GEOMETRY OF NUMBERS

where the g's lie in Z, and Ui is the unit matrix. If


Ur =

Qm = Q'U1U2 *

and forh, k=

then Qi-=',
ahk

ah%

(a(k

1, 2,

)hk1,2

...

n),

1, 2, *,n

if k # m, and

qfma

...

t +

g(m2a)aa-.

The n principaldeterminantsof Om
ahk

Ar=

(r =

hki,2,...,r

1, 2,

n)

. .,

are thereforeequal to the corresponding ones of Om-,and so of !2'.


By construction, the elements of Q1 satisfy the inequalities
l a (k) |

(h, k - 1, 2, ...*

<

n),

and therefore also the inequalities

Assume now that Ui,

1, 2,

*..,n).

II1 2,

*.**,n)

***, Um-i were determined such that

1< a k)

Ia(m-')

(3)

(h

a ()

|a(')

(h, k

-<(

for h = 1, 2,

, n; k

*.

1, 2,

, m -1.

Then Ur, as we shall prove now, can be constructed such that Qr satisfies the
stronger inequalities

(40)

(
)

(hk

(k)

?ahk
|

cr

|<

foi h

., n; k=

12,

1,2,

**,

1, 2, *-,r.

To this purpose put


a('

1y

...

+ ajm)

+ a rm-1n

th(y1,

...

7m-)

th

tl =

t2 =

**

tm-1 =

Am-iYr =
I)th

row

Hence

ILet the clement

Ami

Am-,r
q(m)

If

(r

+:m-lr

On solving,
=

order minor of Amobtained

whereeAm-,r,is the (m -

IYr!r

...n)

0.

This system of linear equations has the determinant Am-,.

the rth colujmn,.

1,2,

Yrn-i of 2 such that

and determine elements Y1,I72

and

(h

fi'om

of Urn, now

be

1),

by omitting the mth


(m)

(r)
the

I ain)

(37),
(mi)

O'(1)

<

- ,.m

1, 2,

number
Yrl

<I 1

>
in Z

satisfying

(r

1.2

the
*..

inequality

1m

),

518

KURT MAHLER

so that
(?n)

Then fromthe firstsystem of inequalities (39) forh = 1, 2, ... , n


1
)
g~n-)I)= gim a~l + ...+ g9m-1ahm-1+ahm
=
|(qi
t
,~

(i)

/ (m)

_?max

and fromthe second systemforh = 1, 2,


| (m) |=

|th(9()
-=t

g(m)

(qf9f

* *-

* v(1)
,

*..

\
(m-i)

(m)

(mi)

9 (M

y1)aM

(gqMZ -

ym-Dahm1i"-

<

Since the remaininginequalities (40) are contained in (39), the matrix Um has
the required property. Hence if
U1 = XUI U2 * *

then this matrixsatisfiesthe statement of our theorem.


13. A propertyof the product of n inhomogeneous linear polynomials in n

variables. Let 52 =

(ahk)hk-1,2,.. ,n

be again a matrixwithelementsin ? of

determinant1. We formthe distance function

ce
wherefi, f2

F(X If)

max

h91,2,. .n

I ahlX1

(efh

ah2x2

ahnXn

I),

, fn
are n integerssuch that fi + * + fn= 0. By the theoremoflast paragraph,thereis a matrix U withelementsin Z and determinant1,
such that the product matrix
=
1U = (ahk)
.

satisfiesthe equation
n

IImax

h-1 k-1,2,

,n

(I ah*k)=1

Let us choose the integersfh such that


(41)

e-h=

max

k-1,2,-**,n

(IahkZ)

and put
ahk = zh ahk
Then by the transformationX = U Y, F(X)
function

F(X If0)= F'(Y)

(h,k = 2,*,
2.(hk1,

If0)changes into a

max (IaZ*y*+ *

h-1,2, * *,n

(h =1, 2, .. ,n)

n).

new distance

+ aZ,"yn1),

TO

ANALOGUE

519

GEOMNIETRY OF NUMBERS

'S

MINKOWSKI

where now all coefficientsah** satisfy the inequalities I ak. I<


1, and their
determinantis still 1. Obviously, for all n,fa-independentvectors Y1() =E(1)
(2
,(1)
= E(n) the value of this function
y(2) = E
Y
1

FP(Yr(h)) <

(h = 1, 2,

n).

Thereforeby the equation (25), necessarily


=

Ff()r(l))

F'(Y''9)

F'(Y-")

1,

and so all minima of F(X If0), where the fO's are given by (41), have the same
value 1, and in particular, the firstminimumof F(X f0)has the exact value
1
where V = 1 is the volume of F(X f0) ? 1.
As an application, let a1, a2, ... , an be any n elements of f, and q1l
tn
7 n elementsoff satisfyingthe equations

'12 X **

ah~l-0 +

If y1, Y2,

*X+

ahn

7h

ah

(h

=1, 2, ...

n).

n are the elementsof : forwhich


yn
--Oh I

IYh

(h

1) 2,@n),

then obviously

Hence thelattice point X


|

|ahlyl

ahlxl

(Xl

ahnyn +

X n) =

X,X2

ahnXn

ah

ah

(h

=1, 2, ...,

n).

U IY satisfies theinequalities
I

(h =

1,

2, ... , n),

and therefore the inequality


n

TI

h=1

Here the constant e-

Iahlxl

On

ahn Xn +

ah I?<

the right-handside is the best possible, as is clear if,

e.g. U is the unit matrix and all

ah,

lz.

14. Distance functionsin Mp. The field 9 of all rational functions with
coefficientsin f has valuations differentfrom the "infinite" valuation I x
.
whichexpressesthe behavior of x at the point z =
Let r be any elementof f, and p the "finite" point z = t. Then we definea
valuation I x Ioby puttingforx $ 0
x I

=e

wherefois that integer,for which neitherthe numeratornor the denominator


are divisible by z - t; we denote by Ro
of the simplifiedfraction (z - t)'fx

520

KURT

MAHLER

the perfectextensionof 9t with respectto this valuation; it consistsof all formal


Laurent series
X =

af(Z

CXf+1(Z -

;)f +

t)f+1

Caf+2(2 _

t)f+2

with coefficients
in r,and if af # 0, then I x = e-f.
Let now F(X) be any special distance functionof 9; we use it as the measure
for the size of X. Further let F(X I p) be a general distance functionof Rf.
Since
F((z

)f

e-fF(X

this distance functionmay assume arbitrarilysmall values, ifX lies in the modul
Anof all lattice points. By (5), there is a constant rp > 0 such that

) < rFIx

F(X
here forX = (xi,@

, x)
XIX

max (I xi

x,,

Hence
F(X

IF, forall lattice points X.

Let t be an integersuch that


e-I

and C(e-t

<

rp, i.e. t >logr)

) the convex set of all points X in Pn forwhich


F(X I p)

_ e-t.

Then the set m(- t | ,) of all lattice points ill C(e-t | ) contains with X and Y
also aX + bY, when a and b lie in Z; it is thereforean 5E-modul. By the
general theoryof polynomial ideals,'6 this modul has a basis of n lattice points
p

(k - 1, 2, *

(p(k)

, n),

such that everypointX in Anbelongsto m(- t I p), ifand onlyit can be writtenas
X = YiP(1) +

+ ynP(n) with y, *

,yn in H.

The determinant

D(-t) =

|(k)

h,k=1,2,-.-,n

0,

and thereforethe number


A(-t)=

I D(-t)

is positive.
16 Compare the basis theoremin ?80 of van der Waerden's "Moderne Algebra", Vol. II,
1st ed.

521

ANALOGUE TO MINKOWSKI 'S GEOMETRY OF NUMBERS

The functionF(X) changes into a new distance function


F'(Y)

= F(X)

= F(QY),

52 = (p (k)

The convex body F'(Y)

by the transformation(42).

V' =

_ 1 has the volume

A\(-t)-1Vj

where V denotes the volume of F(X) < 1. By the results in ?9, there are n
, y(n) with determinant1, such that
lattice points y(l)
...

F(Y())

F(Y(n))

The transformedlattice points X

..,

Q y(k)

X(k) =

_(Vt).

V.

X'n) given by
(k = 1, 2, ... , n)

X(k)

(X(k)

have the determinant


D(-t)

|IXh

|h,k=1,2,..-.n

and satisfy the relations


F(X('))

*..

F(X(n)) =

t)

-(

(k = 1, 2, *.,n).

_ e-'

F(X?k)

It is not difficultto prove that for large t

IP= O(e-t).

I D(-t)

A(- t) = O(ent),

In the followingcase, sharperresultsare obtained. Let


F(X | p) =

max

h=1,2,*

(I

ahlxl

ahnmXn-

Xn-m+h

Ip),

where the a's are elements in go such that

I ahk

lP <?1

(k

n)

1 2

Then to every positive integert thereare elements Ahk in Z satisfying


-

akAhk

Hence, if Y1, *

yn belong to Z, and x1!


Xi =

(42)

l < e-t

Xn-m+h =

(Z -

Yi, .*

b)tYn-m+h

Xn-m =
-

(k

Xn

1, 2,

..

In)

are defined by

Yn-m;

(Ahlyl +

Ahn-mYn-m),

(h = I1,2, .. * *,m),

then F(X Ip) < e-'. Let F'(Y) = F(X) be the special distance functionin 9
< 1 has the
derived from F(X) by the transformation(42). Then F'(Y)

522

KURT

volume I (z points y(l), ...,

y()

MAHLER

IV

= e-mV. Hence there are n M-independentlattice


of determinant1 such that
F'(Y~l))...

F'Y))

mt

= V

The n lattice points X(1,*..., Xn) derived fromthese by (42) have the determinant (z -)mt
and satisfythe conditions
mt

F(X(1)) ... F(X(n)) = e


MANCHESTER,

ENGLAND.

F(X(k) Ip) < et

(k = 1, 2, * *

n).