268

H. WEYL

[Dec.

15

ON GEOMETRY OF NUMBERS

By HERMANN WEYL.

[Received 27 November, 1939.Read 15 December, 1939.]

1. Minkowski stated his fundamental theorem concerning convex solids in relationship to a lattice as follows. Let the lattice in the w-dimensional space of vectors jr = {xv ..., xn) consist of all vectors o with integral components, and let/0?) be any "gauge function", i.e. any continuous function in vector space enjoying the following properties: f(t) > 0 except for the origin jc = o = (0, ..., 0); /(ft) = |$|./0f) for any real factor t;

/(*+*') <f(t)+f(f>Then/(]r) < 1 defines an (open) convex solid $ with o as its centre. V being the volume of $ and 2M the minimum of /(a) for all lattice vectors a ^ o, Minkowski's universal inequality is (1.1) Jf7<l.

If is any set in vector space, we can place it round a given point a of the w-dimensional affine point space; the ensuing point set a consists of all points x such that the vector jr = ax is in . Minkowski's argument is based on the simple remark that the convey solid

homothetic with $, when placed round each of the lattice points a, causes no overlapping so long as q < M, and hence up to that limit its volume q71 V stays less than or equal to 1. With q growing beyond M, the solids of the lattice begin to overlap. Studying the portion of the space covered by

1939.]

ON GEOMETRY OF NUMBERS.

269

them, Minkowski arrives at the sharper inequality f (i.2) if1...irllF<i,

where the numbers Mk may be described thus: A bubble of shape $ while being blown up describes the series of figures &{q) with increasing q\ q = 2M = 2MX is the first moment when a lattice point b j ^ o enters into the bubble, 2Mk that stage at which for the first time a lattice point bk

enters off the (k l)-dimensional linear manifold [bl5 ..., &&_J spanned by
i> > &&-1- Clearly M1^.M2^i... <Jf. In other words, a = bk is that one among the lattice vectors ct outside the manifold [bv ..., i>k-i\ f r which f(a) assumes the least possible value 2Mk. So long as q < 2Mk the body &(q) contains less than k linearly independent lattice vectors, while for q > 2Mk their number is at least k. The vectors
&

t>JWXt

t)2/2M2)

.... bn/2Mn>

and hence their convex closure, are contained in Si. If P = \b1...pn\ is the volume of the parallelotope spanned by the lattice vectors bl3 ..., &n, the volume of that closure is p Pi n\ 2M1...2M2~n\ Mx...Mn' p -^ < M,... Mn V < 1,
2n

Therefore

and thus the positive integer P ^.nl. This note consists of two parts, thefirststating Minkowski's problem in terms of characteristic functions <j> which are capable of the two " truth values " 0 and 1 only, and then generalizing it to probability functions that move in the range 0 ^ <f> ^ 1. Overlapping is not excluded, and is taken care of by using Boolean rather than ordinary sums. The substance of the argument remains unaltered, yet the essential points come out more clearly. If P were equal to 1, the inequality (1.2) would tell an important fact about reduction under arithmetic equivalence, namely about normalization of the basis of our lattice in terms of a given convex solid $. As it stands, it points in a slightly different direction. However, by an almost trivial twist, one can make it tally with the problem of reduction. This is done in the second part, which concludes with some remarks about the special case of the reduction of quadratic forms.
f Geometric der Zahlen (Leipzig, 1896, 1910), 211-218.

270

H.

WEYL

[Dec.

15,

[Added January, 1940. Prof. L. J. Mordell, to whom I sent a copy of the MS., kindly pointed out to me two papers which I had overlooked; one by H. Davenport, Quart. J. of Math., 10 (1939), 119-121, bears closely on Part I by containing another proof for the inequality (1.2), while the other by K. Mahler, Quart. J. of Math., 9 (1938), 259-262, actually anticipates the main result of Part II, namely Theorem V. I am still surprised that the discovery of this fact in 1938 by Dr. Mahler, and of the simple remark on which it is based, had to wait for forty-two years after Minkowski established his inequality (1.2).]

I. 2. A set in the affine space of points x = (xv ..., xn) may be described by its characteristic function <f>(x) which is equal to 1 inside the set and to 0 outside the set. The characteristic function <f> for the union of two sets with the characteristic functions (f>v <j>2 is their "Boolean" sum

which reduces to the ordinary sum only if the sets do not overlap. Any

number a in the interval 0 ^ a ^ 1 is said to be a probability. If a and b are

probabilities of two statistically independent events A and B, then their Boolean sum c = av6 = a+6 ab 1(1a)(l6), which again satisfies the inequality 0 < c < 1, is the probability of the event "A or B". The Boolean sum is commutative and associative, a' < a implies (a' v b) ^ (a v b), as the expression avb = a{\b)-\-b at once shows; and consequently (2.1) a'<a, fc'<6 imply (a'v6') < (avb).

The Boolean sum of a finite number of probabilities av ..., ah is

h

= 1 II (1 a{)
(0 (i, ft) (i.*.O

= ^-y La, yr L a,afc+^T S Summation extends to all throws of 1 or 2 or 3 or ... different indices from the set 1, ..., h, in the second line without regard to their order, in the

1939.]

ON GEOMETRY OF NUMBERS.

271

third line with regard to their order. Brackets as in (*, k, I) indicate the first kind of summation. A function <f>(x) in our space which satisfies the condition 0 ^ <f> ^ 1 is called a probability function. Let <f>(z) now be an (integrable) probability function vanishing outside a finite region # of the aj-space which may be fixed as a parallelotope: Af < x t < B { (i = 1, ..., n).

The lattice translation a = (als ..., an) with integral components a,-,
x

carries <f>(x) into a function

<f>(zt a) = <f>{z1-a1, ..., xn-an).

In these circumstances, the Boolean sum ifi(z) = S<f>(z, a)

a

extending to all lattice translations a has a meaning, since for each x only a limited number of its terms do not disappear. tfj(z) is a probability function with period 1 in each of the variables x1} ..., xn. We form the integral =
}E

ih(x)dx [dx = dxx dz2... dxn]

extending over the fundamental mesh of the lattice E: 0 < a ; 1 < l , ..., 0 < a ; n < l .

Obviously 0 < tp(z) < 1 entails (2.2) Moreover, for two probabih'ty functions <f>2 and <f>x ^ <f>2 we have 0X < tfi2, because of the law of monotony (2.1) for Boolean sums, and hence (2.3) The explicit expression for \ji is

S *(*, a)<f>(z, h)<f>(z, c ) - . . .

I a, b, c

[a, &, c, ... lattice veotors].

272

H. WEYL

pec.

16,

In substituting a+h, a+c, ... for b, c, ... we see that ift(x) may be written as an ordinary sum where

[b, c, ..., lattice vectors # o]. This trick works because the lattice translations form a group. When we return to " combinations " without regard to order, the second factor on the right in the last formula is

(2.4)

<f>**(x) = 1 - ^ S ftx, ) + J S <f>{x, a) <f>(x, b ) - . . . .

(a) (a, h)

[a, b, ... lattice vectors ^ o Take a real parameter t ranging from 0 to 1, and introduce the product
a 9*0

(a)

(a, b)

Its integration from t = 0 to 1 leads to (2.4). Hence our final definitions are:

(2.5) (2.6)

<*(*) = <(*).<**(*),

<f>**(x)=\1 n
JO a _t

^ M = f <l>*{x)dx.

J oo indicates that the integral extends over the whole re-space. Notice that 0 < ^ * * ( a ; ) < l and ^ ? * ^ 2 * * if & < ^ a . (2.6) proves J[<j>] to be invariant under translation, namely

(2.7)

J[<I>'] = J[<I>] for <f>f(x) = <f>(x-x).

If 0 < <f>x ^ ^ and ^^i < e, we find that

and thus, from (2.6), derive the result

(2.8)

^[^]-^M < f

{iW-faix^

where H is the volume of Jp. We summarize: I. The integral (2.6) is less than or equal to 1 for any probability function ^ which vanishes outside a finite region. J[<f>] is
THEOREM

1939.]

ON GEOMETRY OF NUMBERS.

273

invariant under translation, and

if *i(*K*a().
Specilization of the theorem for the characteristic function (f> of an (open) set of definite Jordan volume V yields the simple MinkowskiBlichfeldt-Scherrer theorem: If 0 has no point in common with the sets a placed round the other lattice points a # 0, or, what is the same thing: if 0 never contains both end points x and + a of a lattice vector a ^ 0, then its volume is less than or equal to 1. I consider Theorem I to be the natural general form of this principle, and our proof is nothing but a neat analytical form of the simplest argument by which that principle has been provedf. [Our affine #-space could be replaced by any differentiate manifold U of points x which carries an infinitesimal volume measure, and the group of lattice translations by any group of differentiable volume-preserving one-to-one transformations x-s-x8 of U which (1) is discrete in the strictest sense, so that the set of points Xs, equivalent to any given point x, never has a condensation point, (2) has a compact fundamental domain; i.e., the manifold, arising from U by identification of equivalent points, is supposed to be compact.] 3. With SiegelJ we introduce the Fourier coefficients of the periodic function *fi(x): J{ = \ tfi(x). e(lx xx+...+ln
JE

xn) dx

corresponding to the various lattice-vectors 1= (lv ...,ln) in the dual space, with e(x) as an abbreviation for e2frte, and formulate the Parseval equation
l"l

12

The zero coefficient Jo is our J = J[<f>].

Making use of the equation

ip{x).a>(x)dx=\ <f>*(x).a)(x)dx
}E Joo f G. Haj6s, Act. Litt. Set., Szeged, 6 (1934), 224-225. H. F. Bliohfeldt, Trans. American Math. Soc, 16 (1914), 227-235. W. Scherrer, Math. Annalen, 86 (1922), 99-107. X Acta Math., 65 (1935), 307-323. T 8BR. 2. TOL. 47. NO. 2306.

274

H. WEYL

[Bee. 15,

which holds for any periodic function co(x), we get

(3.1)
and We find that

J I = [ cf>*(x).e(llx1+...+lnxn)dx
[ [lp(x))2dx^[ <f>*(x). ijj(x)dx.

and thus ** (3.2)

witn

<M*) = n {l-<(z, a)}.

THEOBEM II. / / 0* a?wZ <f>** are defined by (3.2), (2.4) and J[ denotes the integral (3.1), then the following equation holds for J = J[<f>]:

It shows very clearly how 0 < <f> ^ 1 leads to 0 < J < 1. [This argument works on any diflferentiable manifold U on which we have a transitive group F of differentiate one-to-one transformations, the lattice consisting of a discrete subgroup of F whose fundamental domain is
compact. The harmonics replacing e(l1x1-{-...-\-lnxn) are obtained by

a slight modification of the process described by the author, Annals of

Math., 35 (1934), 486-499.]

4. We now introduce an assumption of convexity by supposing {<f>{x) ^ r}x, i.e. the locus r of the points x where <f>(x) ^ r, to be a convex set for any r > 0. The relations <f>(x) ^ r and <f>{x') ^ r then imply (4.1) <f>(tx+(lt)x')^r

for any number t between 0 and 1. A function <f> having this property in our n-dimensional space preserves it on any linear subspace. The "terrace" {r < <f>(x) < r'}x between the levels r and r' >r has a Jordan volume because

1939.]

ON GEOMETRY OF NUMBERS.

275

it is the difference between the two convex bodies $r> and %r. Let

(4.2)

O=

ro<r1<r2<...<rh=l

be any division of the interval 0 < r < 1 into subintervals of lengths and let <f>A (x), <f>*(x) be the corresponding step functions ri+ir^e, which approximate <f> from below and above with an error less than or equal to e. They vanish outside and are defined inside J? by fa{x) = r{ a n d <f>A{x) = ri+l if The Biemann integrals of <f>A and ^A clearly differ by not more than He; this proves the Riemann integrability of <f>. Moreover, by (2.3) and (2.8), JM<J[<f>]<J[<f>A] and

and these inequalities could also serve to define J[(j>] by mean of the Biemann integrals J[<j>^\ and /[<*] of the approximate terraced functions. We repeat our assumptions (A) about <f>: (i) 0 <<(*)< 1; (ii) tf>{x) = 0 outside a finite region ; (iii) the sets r : {<f>{x)^r)x are convex. Minkowski's second step deals with the process of dilation as denned by

(4-3)

W"fc.-..^) =

and, properly generalized, leads to

THEOREM

III. Under the assumptions (A),

(4.4) Proof. Suppose that we have a point x where (f> (x) takes on its maximum value. If a; is any point and ^(a;) = r, we then have <f>(x) > r, and, by (4.1),

(4.5)

<f>(jx+(l-j}z)^r

= <l>(x) for any j f > l .

For the moment denote the function on the left-hand side by <j>Q'(z). Formula (4.5) entails and thus, since <f>a' arises from <f>a by a translation,
T2

276

H. WEYL

[Dec.

15,

Since we have neither assumed <f>(x) to be continuous nor the sets f to be closed, the existence of a maximum is doubtful. But the following slight modification will do. Let r{ in the finite sequence (4.2) be the highest level for which %r, is not empty. We cut off the top of the mountain at the altitude r(, i.e. we form

and choose a;0 as a point in r.. Then we get

instead of (4.5), and hence

However, according to (2.8)

and by driving towards 0 the inequality (4.4) is re-established. (Instead of the decapitated mountain, we might have used the terraced mountain <A.) Since {<f>Q.)Q = ^ a ,

we may claim Theorem III to assert that

J[<f>Q] = J(q) is a monotonically increasing function of the parameter q of dilation.

5. Our next move is to modify the procedure by taking into consideration only the lattice vectors (5.1) in the linear subspace = . . . = * = 0. (f>* = <(*) is now defined by the last equation (2.5) with the product ranging over the vectors o ^ o of this A;-dimensional lattice, and (av ..., ak, 0, ..., 0) (a, integers, k ^ 1)

1939.]

ON GEOMETRY OF NUMBERS.

277

The integration is then conveniently carried out in two steps, first with respect toxv ..., xk for any constant xk+1, ..., xn, and secondly with respect to xk+1, ..., xn. Thus we find that implies while the inequality

holding under the condition 0 < <f>fa < e, follows by the same argument as (2.8), without a break in integration. Like <f> itself, the function <D(*i **) = Hxi **. 4 + i *) of k variables satisfies the assumptions (A), and therefore (5.2) Notice that with J[OJ>J[O] Q^...xk) for g>l. = ^ q ( x x . . . x k , a#+1...xQn)

$q(x) = <f>(j, ., Y > **+!, ..-, xn)

Thus, by integrating (5. 2) with respect to xk+1, ..., xn, we get

The substitution

(5.3)

performed on the last nk variables results in the equation

Hence

J(k)[<f>q] > <t~k J(k\$\

for

q^l.

For our Riemann integrals the performance of the integration in two steps xx... xk\xk+1... xn and of the substitution (5.3) offers no difficulty. But maybe it is preferable first to argue for the step function <f>, thus avoiding the difficulty about the maximum of <f> mentioned above, and then to combine the ensuing inequality

with

J(k)[fa] > J(k)[<l>]-H.

278

H. WBYL

[Dec.

15,

For k = 0 we define <(0) = (f> and therefore

Any ^-dimensional discrete lattice Zk may serve here instead of (5.1), since the first k coordinate vectors tx, ..., efc may be so chosen as to span the pre-assigned lattice (that is, so that the lattice consists of the vectors a1t1-\-...-\-aktk with integral components ty).
THEOBEM

IV. By means of a k-dimensional lattice %k we define

4f(z) = <f>(x). f n{l-#(a>, a)} A,

Jo a the product extending to all vectors d ^ O of the lattice %k, and jwM
=

f ^ (a .) dx>

jk{q)=

jiW.

Then Jk(<l)lqn~k w an increasing function of g. In particular,

The application to the situation considered by Minkowski is immediate. <f> is how the characteristic function of the convex solid $: f(x) < 1. Let Zk consist of the lattice vectors in [&ls ..., bk]. So long as we have Indeed, the solids R(q) and Sta(q) do not overlap if a is a lattice point outside [g^ ..., gfc] and q^-Mk+1. We thus find that
n

.V

for

in particular Then

^(-^i) = ^ i n V-

{^J~\
and hence

= M1<TKV for
7(lf2) > Jf j.

Continuing in the same manner, we get

M1...Mk_xcr^.V (k=\,...,n)

for

1939.] and ultimately

ON GEOMETRY OF NUMBERS.

279

..Mn.V

for q>Mn.

Since J(q) always remains less than or equal to 1, Mmkowski's inequality (1.2) is proved II. 6. Contrary to the procedure in I, yielding the vectors b& and ratios of dilation 2Mk, the problem of reduction requires the determination of vectors tk and ratios Nk according to the following inductive rule: Let a range over all lattice vectors outside the (ft1)-dimensional manifold [ex ... efc_J such that every lattioe vector in the ^-dimensional manifold [ex ... eft_xa], h-i i n Oi has an integral a-component x. a = c& is a vector within this range 9t& for which /(a) takes on the least possible value Nk. As the first step, we may take cr = blt so that Nt = 2MX. By induction we readily conclude that e^ ..., efc span {i.e. form a basis of) the whole lattice in the manifold [tx ... cfc]; hence cls ..., en span the whole n-dimensional lattice. Relative to this coordinate system, the lattice vectors e B =(a; 1 , .... a>n) are again described byintegral components xi} and our reduction is equivalent to the following conditions: (6.1) /(a^afc ...,*) > / ( 8 A ...,8 n *) = ^ which hold provided that xlf xz, ..., xn are integers and xk, ..., xn are without common divisor. Thefy*are the Kronecker 8's and thus c & = (8r&, ..., Snfc), the k-th unit vector. This characterization (6.1) obviously entails the inequalities

We have normalized the basis of our ^-dimensional lattice relative to the convex solid by means of the n sets (6.1) of infinitely many inequalities, and this is what the problem of reduction demands. The question arises whether one can establish a universal upper bound for Nx... Nn V. I now prove
THEOREM V. If a reduced lattice basis efc is so chosen as to satisfy the conditions (6.1), then

(6.2)

Nx...NnV^n,

280

H. WBYL

[Deo. 15,

where (6.3) /a ft = 2 .
(

Proof. One of the vectors t>v ..., bk, say 6, lies outside the linear manifold
E

k-i=

ih " efc-i]-

In

^fc*=[e1...eA:_1)]

we can find a lattice vector e* which, together with ev ..., tk_v spans the whole lattice in Ek*. Out of the finite number of lattice vectors of the form

with

- * < * * < * (=1, .... fc-1), 0 < i / < l ,

select one, say e*, with the lowest y. Since b itself must be an integral combination of ex, ..., ik_x, e*, this y is the reciprocal 1/e of a positive integer e. If e = 1, e* coincides with t>. Our e* is in 3ftfc and thus in the competition for the lowest value Nk of/(a). But, since/(>), as one of the numbers
2Mlt ..., 2Mk>

is certainly less than or equal to 2Mk, we infer from these remarks and (6.4) that if e = l ,
w

) if e > 2 .

Hence Nk cannot exceed the larger of the two numbers (6.5) (6.6) Mk+t(Ni+...+N*-i) Ni^26iMi. Once it holds good for and 2Mk.

This allows us to establish universal inequalities of the kind

Indeed (6. 6) is true for i = 1 with 81=1.

t=l, ...,*-1,

the bound (6. 5) for Nk yields a similar inequality for i = k with 6k as the greater of the two numbers 1 and $ ( 1 + ^ + . . . + ^ ) . This determines the factors 6k by recursion, and we readily verify that 0 i = A, 0* = (f)&-2 for fc>2.

1939.]

ON GEOMETRY OP NUMBERS.

281

Therefore (1.2) implies

It would be an interesting problem to ascertain the lowest constant fx,n for which the relation (6.2) holds. Fuller exploitation of the method yields the following
GENERALIZED THEOREM V. Let qv ..., qn be given positive numbers greater than or equal to 1. If the inequality

f(xvx2,...,xn)^.f(81*,...}8n)
holds whenever xv ..., xn are integers and xk, ..., xn have no common factor (k= 1, ..., n), then the values
satisfy the relations and qkNk+X ^ N k (k=l, ...,n1)

The proof is practically the same as before. for the 8k are now

The recursive equations (k>2)

, = <?, and ^ = 1 + 9 , + ..+*-,

i.e.
with the solution

f*-l+fc. ^+0* = ^
9z Ik V

(*>*).

7. Here is another similar proposition:

THEOREM VI. Put

*r = v - * v - l and

(v=l, 2, ...

for
We can determine a lattice basis tx*,..., cn* such that the distances /(c&*) = Mk* satisfy the inequality (7.1) Jf JL *...Jf n *

282

H. WBYL

[Dec.

15,

This constant pn is much lower than the one in Theorem V. with n is best judged by the recurrence formulae

Its increase

and
e

Pn

2+.'v)+\

Pn-i for

odd v and n

Proo/. Out of the vector basis bv ..., &, we construct a lattice basis i*"- en* according to the recurrent equations e&* = t>k+aktk^t>k_1+...
e

(7 .2)

+ak>

bx

with the conditions

the ek being positive integers. If ek = 1, all the coefficients akth_lt vanish, and /(e&*) = 2Mk. However, if ek^2,

...,aktl

Mk* =/(efc*) < A

e

*l

with

^ == i.

To cover both cases we put r\k = 2 for ek = 1. Prom (1.2) we now get

and it remains to discuss the product ^ i . . . i?n. We put 8^ = 2 or hk = 1 according as e f c = l or e f c ^ 2 ; ^ fc '==2if8 fc = 2, and Then ^fc < r^fc'
Vk'

= 8 fc +i(8*_ 1 +...+8 l ) if 8*= 1. ... -qn'.

and ^ ... r?n ^ ^

Among the 2n values of the product (7.3)

n

%'...*?'

resulting from the 2 possible sequences hlt ..., 8n of l's and 2's, we nave to ascertain the largest. An interchange of two consecutive figures 1 and 2 in this sequence whereby 8& = 1, 8fc+1 = 2 is ohanged into 8& = 2, 8 ^ ! = 1

1939.]

ON GEOMETRY OF NUMBERS.

283

(all other figures remaining unaltered) affects merely the two factors ikVk+i m (7-3), turning their values [l+*(8*-i+.+8i)]-2 into 2.[2+J(8 fr _ 1 +...+8 1 )] f

and so increasing (7.3). For a given number v of 2's in the sequence, the maximum value is thus attained when they come first (8ls ..., ), and then (7. 3) is equal to

2H-2 2v+3
i.e. 2a"-*.(

Let v pass through the values v = 0, 1, ..., w. Since

is on the upgrade (wv > wv_x or wy ^ wv_x) so long as or v( and then goes down. Hence the maximum is attained for the last v for which v{v\) does not exceed n-\-\. Here is a table for the lowest values of pn: n=l, 2, 3, 4, 5, 6, 7 ...,

pn = 2, 4,

12, 42,

168, 756, 3960,

Higher bounds, for instance

can even more easily be obtained. Again we have the question of the best value of pn in the inequality (7.1). 8. A positive-definite quadratic form

(8.1)
(8.2) /(*) ^gkk

/(*) = S0
for every x in 3tfc and k= 1, ..., n,

is said to be reduced if it satisfies the inequalities

where 0tfc is the set of all lattice points x = (xv ..., xn) for which xk, ..., a;n are without common divisor. While for every positive form of discriminant D = det (gik),

284

H. WEYL

[Dec. 16,

we have found that for a reduced form (8-3) K9n-9n<D with An=KK)*.

Here u)n denotes the volume of the w-dimensional unit sphere, most handily described by the recurrent relation a>n = .<on_2 together with a> 0 =l, a>1 = 2. Indeed, with V / a s g a u g e function, the convex body $ : / < 1, becomes an ellipsoid of volume wJ-y/D, as is readily proved by linear transformation x^-\-...-\-x^. of f(x) into the normal form The linear substitutions 8 of the variables x{ with integral coefficients s{k of determinant \aik | = 1- (or only those of determinant 1) form a group {8}. Two quadratic forms / and 8f which arise from each other by a substitution of the group are said to be equivalent. Every positive form is equivalent to a reduced form. All quadratic forms / = {gik} with real coefficients gik = gki form an N = r&(i-f-l)-dimensional linear space R, the positive ones an open convex region & in R, the reduced forms a convex subset Z of O which is closed relative to 0. Both 0 and Z are cones and may, therefore, be considered as sets in the (N 1)-dimensional sphere of rays in R issuing from the origin. A convex cone H is a set having these three properties: (i) / = 0 is not in H; (ii) if / is in H, so is tf, where t is any positive factor; (iii) with / and / ' in H, the sum / + / ' is in H. We operate throughout in 0; all topological words referring to subcones of O are meant relative to O (not R). Each of the inequalities (8.2) for the variables git is of the form (8.4) a t ,<to>0
2TT

with integral coefficients a^. We expressly exclude from the set $ of inequalities (8.2) those for which all coefficients ai} vanish, namely (8.2) with x ={hxk, ...,8,*). Each individual inequality (8.4) of $ defines a half-space bounded by a plane (8.5)
i

1939.]

ON GEOMETRY OP NUMBERS.

285

A substitution S carries the central cell Z into an equivalent cell Zs whioh is also convex. The equivalent cells jointly cover the whole spaco 0. Zj coincides with Z for any of the 2 n substitutions

J: x1-*x1, ..., xn->xn,

and 8 and SJ have the same effect upon Z* From (8. 3) Minkowski deduced two important theorems of finiteness: I. A finite number of the inequalities 3 suffices to define the cell Z. II. There is only a finite number of substitutions 8 capable of carrying a point of Z into a point of Z (worded more precisely, the demand on the element 8 of {8} which limits it to a finite set is to the effect that there should exist two points /, / ' in Z such that / ' = 8f). Minkowski has a straightforward algebraic proof of II. in 7 of his paper " Diskontinuitatsbereich fur arithmetische Aquivalenz "f. His proof of I is involved because he establishes I simultaneously with an inequality of the nature of (8. 3) by a complicated induction. This entanglement was dissolved in a joint paper by L. Bieberbach and I. SchurJ (oh tempi passati!). They improve on I by breaking it up into the two theorems:
I*. There is only a finite number of lattice points x in 9tfc for which there exists a reduced form f0 = {g$} satisfying the equation (8.6) /<>(*) = <$*.

I**. If we substitute for x in succession each of the lattice points assigned by Theorem I*, then the corresponding inequalities (8-7) for k= 1, ..., n completely close in the cell Z in G.

While the proof of I* again is a straightforward algebraic affair based on (8.3), I** depends on the topological argument that a straight line joining a point inside to a point outside a convex region must cross the border. I explain the Bieberbach-Schur argument for I** in this simplified form.
t Journalfiir Math., 129 (1906), 220^274; Oesammelte Abhandlungen 2 (Leipzig, 1911),t 63-100. J Sitzungsber. Berlin Akad., (1928), 519-523; (1929), 508.

286

H. WEYL

pec.

15,

I say that those points / (in 0) for which all the inequalities $ hold with the sign > rather than > form the core of Z. Any diagonal form

...+gnxn*

with 0 <g1 <ig2<... <gn

evidently belongs to the core of Z. So do- the inner points of Z. I maintain that conversely every point of the core is an inner point of Z. Indeed if/ is a positive form, Ue a sufficiently small neighbourhood of/ and A any positive number, then the points x of the #-space such that ft(x) ^ A holds for at least one /in Vt belong to a limited region. Therefore all lattice points x with a finite number of exceptions satisfy the inequality /,(#) > A for every/, in Z7e. This follows readily from Jacobi's step transformation of/e into a square sum. Hence we have a finite subset of $ such that every form /, in Ut satisfies all inequalities of -3. Provided that the inequalities 3 hold for/with the > sign, we may shrink Ue so as to have the inequalities 3 satisfied throughout Ut. This establishes I** by the topological argument mentioned above. Let/ 0 again be a point on the boundary of Z and let x range over the finite number of lattice points x in $tk which satisfy (8.6). Studying the convex wedge defined by the corresponding simultaneous inequalities (8.7), we findf that those inequalities 3 suffice which are satisfied with the = sign for Nl linearly independent points / in Z. They are at the same time indispensable, and the corresponding equations describe the (Nl)dimensional planes which contain the faces of the pyramidal cell Z in. 0. Another way of constructing these faces would be by determining the common boundaries of Z with the adjacent equivalent cells Zs. But Z and Zs have a point / 0 in common only if both / 0 and S'1^ are in Z. Hence thefinitenessof the number of plane faces must also be a consequence of II, so that I ought to be deducible from II. Let us examine the situation more carefully. Only the substitutions J are capable of transforming a point in the core of Z into a point of Z. Hence the two convex cells Z and Zs have no inner point in common provided that S is not a J. The points which they have in common, if any, form a convex cone W in a linear subxnanifold of N 1 or ... or 1 dimensions. If Nl is the correct number of dimensions we speak of a wall separating Z from Zs. Choose / 0 as an inner point of W in its (Nl)-dimensional plane. It satisfies one of the inequalities , say (8.4), with the sign = while every point of W satisfies it with the ^ sign. Therefore the whole W clearly lies in the plane (8.5), and thus I is again established, provided that we can show that every point f on the boundary of Z
t For an elementary treatment see H. Weyl, Comm. Math. Helvet., 7 (1934-5), 290-306.

1939.]

ON GEOMETRY OF NUMBERS.

287

lies on a wall separating Z from an adjacent cell Zs. We deduce this from the fact, important in itself, that the equivalent cells cluster only towards the boundary of 0. On closer examination Minkowski's proof of II reveals that there is not more than a finite number of 8 capable of carrying a point of Z into a point of O(p, fi). Here p is any number greater than 1 and /x is any positive number. If Dk denotes the determinant of the form f(xx... xk 0 ... 0), the subcone Q(p, fi) of 0 is defined by the following simultaneous inequalities:

9kk > Qu-i, fc-i faoo = P

f{xlt ..., x^v

1, 0, ..., 0)

for any integers xlt ..., xk-v All points of Z are inner points of O(p, /x). This is more than sufficient to settle our question. Choose a definite p > 1 and /J, > 0. Let / be a point on the boundary of Z and let flt ...,/*> ... be a sequence of points outside Z approaching /. ,We may assume that they lie in G(p, fi). A certain substitution 8'1 of {8} will carry / into a point /,' = /Sf71/H of Z. In consequence of our statement about G(p, p), a certain 8 recurs infinitely often among the 8V. This 8 is no J. The points fv of the corresponding selected sequence all lie in Zs, and so does their limit/. Consequently any point/ on the boundary of Z belongs to the common boundary of Z and an equivalent Zs. The domain O(p, JJ.) increases when p and //, increase, and with p->oo, fx-^-oo exhausts the whole O. At every stage G(p, yC) has points in common with only a finite number of cells Zs. An elementary argument like the one applied before shows that among the common boundaries Z\ZS (which exist in finite number only) none but the "walls", the (N l)-dimensional ones, matter. The substitutions 8 which carry Z into cells Zs bordering on Z along a wall, together with the J, generate the entire group {8}. The Generalized Theorem V proves that all our statements about Q(p, fi) remain true if we define it by the following linear inequalities: f{xv ..., xn)>jgkk

whenever the xi are integers and xk, ..., xn have no common factor,

288

H. WEYL

[Dec.

15,

whenever the xt are integers and (xk, xk+1, ..., xn)=(l, 0, ..., 0). Incidentally the estimates may be improved considerably if we carry out the comparison of the coefficients gkk, g'kk of two equivalent forms /, / ' in Z and O(p, fi) respectively for any convex bodies rather than ellipsoids by the procedure of 6. Throughout we have studied Z relative to 0; this is really the more natural standpoint. However, the results I and I** are true even relative to the space R of all quadratic forms. Z is not closed in R; there may be boundary points f of Z relative to R which are not positive forms. I maintain that for such a form the equation gn = 0 holds. Since / is nonnegative, it can be brought into the form

by a transformation with real coefficients. We can determine a lattice point x ^ 0 for which the m linear forms tjv ..., m are in absolute value less than any preassigned positive number e, either by Minkowski's theorem (1.1) for parallelotopes, or by a simple application of Dirichlet's principle of the distribution of v-f 1 objects over v boxes. Hence gn > 0 would contradict the inequality f(x) ^gn which holds for all lattice points x^O, and Z relative to R is defined by the inequalities described under I* and I** together with gu ^ 0. But even the addition of glx ^ 0 becomes superfluous if n ^ 2. The substitution
Xx = 1, X% = * J
3 = . . . = Xn =
0

yields the inequality

for any reduced form, or 2|gr12| <<7 n . The same holds for g19, ..., gln. Hence the boundary form / even satisfies the n equations
(8.8) ^ = ^ = . . . = ^ = 0.

The fact that Z borders on the plane gxl = 0 only along this linear variety (8.8) of n dimensions less, proves our point. Minkowski's investigations set out from the reduction of quadratic forms. Owing to some "unexpected difficulties" the second half of his book Geometrie der Zahlen never appeared, and fourteen years after the discovery of the basic idea of the geometry of numbers he published his theory of reduction in a strictly arithmetical form, without a hint of geometric approach. What was the snag he struck ? After having taken the main hurdle with the general inequality (1.2), did he fail to notice the simple remark that led us on to (6.2), or was he unable to unravel the

1939.J

ON GEOMETRY OF NUMBERS.

289

inductive kink by which in his final publication the establishment of the inequality (6. 2) is tied to the selection of a finite number of inequalities $ ? It is now idle to speculate. But of him it might be said as of Saul that he went out to look after bis father's asses and found a kingdom. For quadratic forms there is a simple algebraic method of deriving (1.2) from (1.1). Minkowski describes it in Geometrie der Zahlen, 51; and very likely this suggested to him the general inequality (1.2). Hence every improvement of the constant yn in the inequality for positive forms in which M is the minimum of f(x) for all lattice points x ^ 0 entails a proportional improvement in (1.2) and for the constants fxn and pn in the estimates (6.2), (7.1) for ellipsoids. Our value was yn (2n/con)2. Blichfeldtf succeeded in improving it by the factor

This leads to the following Xn: (8. 9) 1/An = Remak recently obtained a better estimate^ which differs from (8.9) essentially by having f instead of f. However, the chief interest of our constant jxn in (6.2) lies in its validity for all convex solids whatsoever. Theorem VI together with Blichfeldt's improved yn shows that every positive form has an equivalent one / satisfying the relation

where

1 / ^ * = ( 1 + W 2-">-*pn*.

Such an estimate is best suited to prove that there is only a finite number of classes of equivalent positive forms with integral coefficients. The Institute for Advanced Study, Princeton, N.J., U.S.A.

t Trans. American Math. Soc, 15 (1914), 227-235; cf. R. Remak, Math. Zeitechrift, 26 (1927), 694-699, and Blichfeldt, Math. Annalen, 101 (1929), 605-608. % Oompoaitio Math., 5 (1938), 368-391. B B . 2. VOL. 47. MO. 2307. U