(t~t~-l)u between W and itself. Proposition 6 of wI I shows that T~s is biregular at any pair of points on its graph. Therefore the varieties S~ (with empty " f r o n t i e r s " ) and the T~s may be used to define an abstract variety S. Call ~ the generic point of S over K with the representatives us and write ( t ~ ( u ) , u ~ , I , ( ~ ) ; ~ is a birational correspondence between W and S, and 9 is its inverse; both are defined over K. Let a be any point of W; by Prop. 4 of the Appendix, there is an a such that ts is generic over k'(a) ; as W is a chunk, taa is then defined; this means that 9 is defined at a, ~ ( a ) being the point of S with the representative tsa on Ss. As ts-~(t~a) is then defined and has the value a,~I, is defined at the point r with the value a. This shows that ~ is a biregular mapping of W onto its set-theoretic image r on S; as the latter is the set of points of S where ~I, is defined, it is K-open on S by Prop. 8 of the Appendix. Once and for all, we will agree to denote by d the image r of a ~ W by 9 in O(W). All this can be applied to the case when W is taken to be the same as V, V acting upon itself by left-translations. Let G be the abstract variety thus obtained from V; call ~o the birational correspondence between V and (7 which takes the place of the mapping (t~x-~)ts is defined
[1955a]
211 ON ALGEBRAIC GROUPS OF TRA~qSFOR)/[ATIONS.
369
at s~; this means that 2-1 is defined and has a representative on G~. Similarly, if we write t ~ t ~ t ~ -1, the representative of 29 on G-t is t ~ x y ~ ((tx~)t~-l)y~; let ~, ~ be two points of G with representatives r~, s~ on G~, G~ respectively; by Prop. 4 of the Appendix, we can choose 7 so that t.~ is generic on V over l~(r~,s~,t~,t~); the same will then be true of t, and also of tr~ and of (tr~)t~ -1 since V is a group-chunk; for a similar reason, this implies that ( x , y ) - - > ((tx)t~-l)y is defined at (r~,s~), and this completes the proof that G is a group. Now, going back to the space S constructed before, we transfer to G, S, by means of the birationa] correspondences 4)o, r the normal law given for V, W; in other words, for x, u generic and independent over K on V, W, and for 2 ~ o ( X ) , ~t~(u), we define ~ ( x u ) , and prove that this makes S into a transformation-space with respect to G. I n fact, the representative of 2~ on S~ is ((tx~)t~-~)u~, where t ~ t.~t~-~ as before; the rest of the proof is then quite similar to the proof given above. Naturally, if W is non-singular, S is non-singular; if W is everywhere normal, S is everywhere normal. Finally, if W is a homogeneous chunk, S is a homogeneous space. I n fact, in that case, let ~, ~ be any two points of S, with representatives as, b~ in S~, S~ respectively. Take x generic over K ( ~ , ~ ) on V; put 2'~2Ct, ~ " ~ 2 ~ . For u generic over K(x) on W, we have ~ I , ( 2 ~ ) ~ (xt~-!)u~; as W is a homogeneous chunk, x ' ~ (xt~-Z)a~ is defined and generic over K(~t, ~) on W, and therefore we have x ' ~ ( 2 ' ) ; similarly we have x" ~ ~I,(2") with x" ~ (xt~ -~) b~ generic over K(~,, 5) on W. That being so, there is an isomorphism of K(~t, b,x p) onto K(d, D, x") over K ( g , 5) which maps x' onto x " ; this can be extended to an isomorphism ~ of K((t, b, ~) onto some extension of K(~, b,x"). Then we have ~ ( ~ 2 " ~ b , and so 5 ~ ~-~2~. 7. From now on, it will be assumed that W and consequently S are everywhere normal. With this assumption, we shall construct an abstract variety S t, defined over k, and a birational correspondence F between S' and W, also defined over ~, so that the birational correspondence r o F, defined over K. between S' and S is an everywhere biregular mapping of S' onto S. This construction can then be applied to V itself, giving a variety G' and a birational correspondence Fo between G' and V, both defined over /~, such that r o Fo is biregular between G' and G. Transferring the normal laws for V, W to G', S" by means of F, Fo, we see that we have thus constructed a group G ' and a transformation-space S', birationally equivalent to V, W over k ; if W is pre-homogeneous and we have constructed S as a homogeneous space, S' will be a homogeneous space.
212
[1955a] 370
ANDRE W E I L .
I n constructing S', we may assmne that V operates faithfully on W; in fact, if this were not so, one could replace V by another pre-group ? satisfying this condition, according to Prop. 2 of w I, no. 3. Notations will now be the same as in no. 6, with the additional assumptions that W and consequently S are everywhere normal, and that V acts faithfully on W, so that G acts faithfully on S. Let /C' be any field containing /C. Let ~2 (s~) be a cycle of dimension 0 4=1
on V, rational over /C', and assume that s~ =2~s~ whenever i J : j . Then, if we put /C"=/c'(sl,. ' ' , s~), /C'" is a Galois extension of/C', i.e. separably algebraic and normal over/C'. Call K " the compositum of K and/C"; let u be a generic point of W over K " , and put w~ ~ s#. I f m is the dimension of the ambient affine space to W, we write w ~ = (W~l, 9 ',w~m). P u t now
~=i
/~=I
where T, U1,'" ', U m are indeterminates; let y be the point, in an afflne space of suitable dimension, whose coordinates are all the coefficients of the homogeneous polynomial y(T,U) except that of Tr; this is the so-called " C h o w point" of the cycle ~ (w~), and y(T, U) is its " C h o w form." As V acts faithfully on W, and the s~ have been assumed to be distinct, the eorollary of Prop. 5, w II, no. 5, shows that the w~ are all distinct. W e can therefore apply to them the following~ general result: LEM~A. I f in (1) we ta/ce the w~ to be any set of distinct points, and ko is the prime field, then the w~ are separably algebraic over ~co(y). By F-I.~, Th. 1, we need only show that a derivation D of the field /co(W~,-" ",wr) over /co(y) must be trivial. I n fact, applying D to (1), we get : r
as the w~ are all distinct, this cannot be an identity in T, U~,- 9 ", U,, unless all the Dwi, are 0. PROPOSITION 7. Notations being as defined above, we have/c'(y) ~ / c ' ( u ) provided the s~ are all distinct and satisfy the following condition: (S) The set of points ~ = ~ o ( s ~ ) any right-translation.
on G is not mapped onto itself by
The cycle ~ (wi) is the image of the cycle ~ (s~) by the mapping
[1955a]
213 ON A L G E B R A I C GROUPS OF T R A N S F O R ) s
371
x-,xu
of V into W ; it is therefore rational over U(u). By the main theorem on symmetric functions (VA, no. 7, Th. 1), this implies that y is rational over U ( u ) , i.e. that U ( y ) C U(u). On the other hand, the lemma shows that the w, are separably algebraic over U'(y); as we have u=s,-lw, by Prop. 5 of w I I , no. 5, u is therefore separably algebraic over U'(y), hence also over U ( y ) . Let (r be any automorphism over U ( y ) of the algebraic closure of U ( y ) ; as it induces an isomorphism of U(u) onto k ' ( u ~) over U, u ~ is generic on W over k', so that s~u~ is defined by Prop. 5. This gives (s,u)r ~, i.e. w~r162 ~. But the decomposition of the homogeneous polynomial y(T, U) into linear factors is uniquely determined; applying to (1), we see thus that the w,~ must be the same as the w, except for a permutation, i.e. that there is a permutation i--->r such that w ( = w o ( o . This can be written as s,r162 sr as the s.f are the same as the s, except for a permutation, we can write them as sinks,(o, where i--->r(i) is a permutation. Then we have r ~) =r which can be written as ,~(~)q)(u~) =2~(,)~, i.e. (~(u ~) ~ ( ~ ) - ~ g ~ ( o ~ . As G acts faithfully on S, the corollary of Prop. 5 shows that all the elements ~(~)-~2~(~) of G, for 1 ~ i ~ _ r, must coincide; if ~ is their common value, we have 2~(o ~ ( ~ ) T , which shows that the right-translation [ maps the set 2~ onto itself. By (S), this implies that [ is the neutral element of G, so that (~ (u ~) ~ ~, and therefore u ~ u . As u is separably algebraic over U(y), this shows that
~'(u) c k'(y). 8. Proposition 7 shows that we may write y ~ f ( u ) , where f is a birational correspondence, defined over ~,, between W and the locus Y of y over /cp in affine space. I f k'[y] is the ring generated over U by the coordinates of y, it is well-known that the integral closure of k'[y] in ~'(y) is a finitely generated ring over U, i.e. that it can be written as U [ y * ] , where y* is a point in a suitable a:ifine space; call Y* the locus of y* over /d in that Mfine space. As we have /c'(y*) ~ U ( y ) ~ U ( u ) , we may write y * ~ f * ( u ) , f* being a birational correspondence between W and Y*, defined over U. I t is usual to say that Y* is derived from Y by " n o r m a l i z a t i o n " over ~'. By Prop. 14 of the Appendix, since U ' is separably algebraic over k', U'[y*] is integrally closed in Jc"(y*). PROPOSITION 8. With the notations explained above, y* and ~t are: corresponding generic points over K" on Y* and S in a birational correspondence between Y* and S which maps Y* biregularly onto the {("-open set i
214
[1955a] 372
ANDR]~ WEIL.
In the first place, we prove that the coordinates wiz of the w, are all in U ' [ y * ] ; as they are in U ' ( y ) because Qf the relations w,=s~u and k'(u) = U ( y ) , it will be enough to show that they are integral over the ring k"[y], or in other words (e.g. by F-App. II, Prop. 6) that they are everywhere finite on W. I n fact, let ~r be any place of k"(y) such that y0r) is finite. Take r independent variables A1," " ",% over ]d'(y), and extend ~r to a place ~r' of U'(y, hl," 9 ",s at which every one of the r points (Xi, M w , , ' ' ",Mw,,~) is finite and =/=(0,.- . , 0 ) . The relation (1), by which y was defined, can be written
~,~ 9 ~ v (T, U) = II (~,T-- X (~,~,~) U~). i:1
#
Taking the values of both sides at ~', we see that the right-hand side does not become identically 0 at that place; as y(~) is finite, this implies that no ~ can become 0 at v ' ; but then wi~(~r) can be written as (~w~)0r')/X~(v') and is finite. This proves the assertion about the w~. We have thus shown that the mappings y*--)w~ of Y* into W are everywhere defined on Y*; as we have ~ t ~ A - ~ ( w ~ ) , this implies that y*--->~ is everywhere defined and maps Y* into the set ~ defined in Prop. 8. Conversely, the definition of y can be written
y(T, U) = H ( T - - X,I,,,(~,a) U~,) {=I
/z
if we call $~(~) the coordinates of ~(~). As 9 is everywhere defined on 9 (W), this shows that the mapping ~-->y is defined at every point of the set ~. As U[y*] is the integral closure of lc'[y] in U(y), it is therefore contained in the integral closure of the specialization-ring of every point of ~ on 8. But we have assumed that W and consequently S are normal, i. e. that the specialization-ring of every point of S (over any field of definition for 8) is integrally closed. This proves that ~--->y* is everywhere defined on the set ~. In view of what we have proved above, ~t is therefore the set of points of S where this mapping is defined, and is K"-open by Prop. 8 of the Appendix; more precisely, it is K'-open if K ' is the compositum of K and It'. This completes the proof. 9.
Denote now by S any cycle ~ (s~) on V, rational over the ground-
field to, consisting of distinct points st and satisfying condition (S). From such a set S, and taking U ~ k, we can derive as above a point y, which we now write as ys, and furthermore a point ys* such that k [ys*] is the integral
[1955a]
215 ON A L G E B R A I C GROUPS OF T R A N S F O R ~ s
373
closure of /c[ys] in /c(ys) ; as above, we call Ys* the locus of ys* over /C; we write gts for the open subset of S denoted by ~ in Prop. 8. I f we allow S to run through any finite set of cycles with the properties stated above, then all the varieties Ys* will be birational]y equivalent to W and to each other, and we can take the points ys* to be corresponding generic points of these varieties over /C. It is then an immediate consequence of Prop. 8 that the affine varieties Ys* (with empty " f r o n t i e r s " ) , and the birational correspondences between them for which the ys* are corresponding generic points of the Ys* over/C, determine an abstract variety S', and that this is biregularly equivalent over a suitable field (as a matter of fact, over K itself) with the union of the open sets ~s on 8. In order to prove that S' will be biregularly equivalent to S itself for a suitable choice of the cycles S, it is therefore enough, in view of the well-known " c o m p a c t o i d " property of open sets in the Zariski topology, to show that the family of all open sets ~s is a covering of 8. I n other words, we have to prove the following: PRoPosITIo~r 9.
Given any point ~t on S, there is a cycle S ~ ~, (s~)
on V, rational over/C, consisting of distinct points s~ and satisfying condition (S), and such that 2 i a e ~ ( W ) for all i. Assume that ~ has a representative as on S~; take x generic over K ( ~ ) ~ K ( a ~ ) on V, and put u ~ (xt~-~)a~, this being defined because W is a chunk. If we put, as usual, ~ o ( x ) and ( ~ a ~ ( u ) , we have then ~2g, so that u ~ , ~ ( 2 ~ ) . As the mapping x--->2a is everywhere defined on 11, this shows that the mapping x--->u of V into W is defined at the points s of V such that 2aE(P(W), and at those points only. Let F be the closed subset of V where the mapping x--->u is not defined; by Prop. 12 of the Appendix, there is a maximal ~-c]osed subset Fo of V contained in F ; then an algebraic point of V over /C is in F if and only if it is in Fo. Call F~ the union of the conjugates over /C of all the components of Fo; this is a /c-closed set on V, and its definition shows that the cycle S on V will satisfy the last one of the conditions stated in Prop. 9 if and only if it lies in V - - F , Now assume first that the field /C is infinite. Applying Prop. 13 of the Appendix to the variety V ~ ~ V - - F , and to the empty subset of V' X V', we obtain a separably algebraic point Sl over /C on V'; call s ~ , . . . , s~ all the distinct conjugates of s~ over /C; if this set satisfies condition (S), which will be the case in particular if d ~ 1, then it solves our problem. Suppose that this is not so, and therefore that d ~ l . For any r ~ d , let sa+~," " ",Sr be any set of r - - d points on V - - F ~ , distinct from one another and from s~," 9 .,sa; put S ' ~ { g ~ , ' 9 ",,~a} and S"~{2a+~," 9 ",~}. I f the set 12
216
[1955a] 374
ANDRI~ WEIL.
S ' U S " is mapped into itself by a right-translation r other than the identity, one of the following circumstances must occur: (i) r maps each one of the sets S', S " onto itself; then r is of the form s':lt ", with s', t" in S', and there must be two elements s", t" of S" such that t " ~ s " r ; (it) r maps S' into S " ; as d ~ 1, we can choose two distinct elements s', t' in S', and then s " ~ s%, t " ~ t ' r are in S", so that we have t " ~ (t's'-~)s"; (iii) r maps some s'~S' onto some # ' ~ S" and some t" e S' ofito some tlP~ S'; then s " ~ s't'-Itl ". Thus, in order to satisfy the requirements of Prop. 9, it is enough to take as sa+l," 9 ", sr the conjugates over k of a point s ~ sa+l of V - - F ~ , separably algebraic over/c, satisfying the following conditions : (a) no 3z, for d -~ 1 <-- l <-- r, coincides with any of the points .~ or 3jf13~ for 1 ~ i,j, h ~ d; (b) no pair of distinct conjugates of s over k lies on the graph of any of the birational correspondences x--->~(22~-~j), x-->~(&gj-~2) for l<=i,j<__d. As to (a), it will be satisfied provided we take s on V - - F ~ , where F2 is the union of F~, of the set Sl," 9 9 sa, and of the set of all conjugates over k of those algebraic points on V whose image on G coincides with one of the points .~jj-~a. Then our result follows at once by applying Prop. 13 of the Appendix to the variety V - - F 2 and to the union of the graphs of the birational correspondences in (b). I f ~ is finite, we have to proceed differently. Take any algebraic point s~ over /~ on V - - F ~ ; call s~,. 9 .,s~ its distinct conjugates over /~; if this set satisfies condition (S), it solves our problem. I f not, we use a result of Lang-Weil (this Jovl~N~tT,, vol. 76 (1954), p. 819) which says that, if 1 is sufficiently large, there must be a point s on V - - F 1 which is rational over the (unique) extension of k of degree 1. We take 1 prime and ~ d. I f s is rational over k, the cycle (s) solves our problem; if not, it is of degree 1 over ~; call sa+~,. --,sa+z its distinct conjugates over k; they are distinct from sl,. 9 sa, since the latter are of degree d over k. The set sa+~," 9 ", sa+z may solve our problem. I f it does not, the group g of right-translations mapping the set (2~+1,' 9 ',2d+~} onto itself is of order v ~ 1; as that set must be the union of cosets with respect to ~, v must divide l, and so g is cyclic of order l; call T a generator of g. Let r' be a right-translation mapping onto itself the set (g~,- 9 .,2~+~}. I f r' is not the identity and maps some element of the set {2a+1," 9 ',2a+~) into an element of the same set, it must be of the form r i, and therefore of order l; but this cannot be, since d ~ 1 is not a multiple of I. Therefore r' must map the set (2a+~,- 9 ",3a+;} into the set ( g , ' ' ' , 2 a } . As d ~ l , this is also impossible. Therefore the set s~,...,sa+~ solves our problem. This completes the proof of the results announced at the beginning of
[1955a]
217 ON ALGEBRAIC GROUPS OF T R A N S F O R : ~ A T I O N S .
375
no. 7. Writing now G, S instead of G', S', we may restate them in a somewhat more complete form as follows: THEOrEm. (i) To every pre-group V, defined over a field k, there is a birationally equivalent group G, also defined over lc ; this is uniquely determined up to an isomorphism. (ii) To every pre-homogeneous space W with respect to V, defined over k, there is a birationally equivalent homogeneous space with respect to G, also defined over k; this is uniquely determined up to an isomorphism. (Hi) Let W be a pre-transformation space with respect to V, defined over ]c; let a be a point of W such that W is normal at a and that, if x is generic over ~(a) on V, xa and x-l(xa) are defined. Then there is a transformation-space S with respect to G, birationally equivalent to W over tr in such a way that the birational correspondence between them is biregular at a; S may be taken everywhere normal, and it may be taken to be non-singular if a is simple on W. Moreover, S is uniquely determined up to a birational correspondence which is biregular at every point of the form ,~, where ~ is the point corresponding to a on S and ~ is any point of G. Except for the statements about unicity, all this has been proved above. As to unicity, the statements in (i) and (ii) are special cases of the statement in ( i i i ) ; and the latter is an immediate consequence of the fact that the operations o~ G are everywhere biregular mappings of S onto S.
Appendix. I f X is any cycle, we denote by I X I the support of X , i.e. the closed set which is the set-theoretic union of the components of X. PROPOSITION 1. Let ~(x) be a regular extension of a field k, and/c(x,y) a regular extension of k ( x ) . Then ~(x,y) is a regular extension of k. This is an immediate consequence of F-IT, Th. 5. PROPOSITmN 2. Let ~(x) be a regular extension of a field k; let K be an overfield of ~, linearly disjoint from ~(x) over Ic; let 7c" be the algebraic closure of k in K. Then ~'(x) is the algebraic closure of k(x) in K ( x ) . Let y be an element of K ( x ) , algebraic over /c(x) ; we may take x to be
218
[1955a] 376
ANDRE WEIL.
a generic point over K of a variety V, defined over k, in an affine space; and then we m a y write y ~ F ( x ) , where F is a function on V, defined over K ; call F the g r a p h of F. As y is algebraic over k ( x ) , there is a polynomial P ~ k [ X , Y ] such that P ( x , Y ) ~=0 and P ( x , y ) - - 0 ; then P induces on the product V X D of V and of the afflne space D of dimension 1 a function which is not 0 on Y X D and is 0 on F. As I ~ has the same dimension as V, it m u s t be a component of the divisor ( P ) of P, and is therefore algebraic over /c. The smallest field of definition of 1~ containing k must ~hen be contained in U, so t h a t F is defined over U ; this implies t h a t y is in k ' ( x ) . C0~OL~A~Y.
I f K is primary over k, K ( x ) is primary over k ( x ) .
I n fact, the assumption means that U is pure]y inseparable over ~; this implies t h a t U ( x ) is purely inseparable over ~ ( x ) . PROPOSITION 3. Let ~(x) be a finitely generated extension of a field ~; then every field K such that ]c C K C l~(x) is finitely generated over k. Let t ~ ( t l , ' ' ' , t , ) be a m a x i m a l set of algebraically independent elements of K over /~; then K is algebraic over k ( t ) . Replacing k by k ( t ) , we see t h a t it is enough to prove our proposition in the case when K is algebraic over k. This being assumed, call /c' the smallest :field of definition containing k for the locus of x over the algebraic closure ~, of /~; then /c' is a finite algebraic extension of k and is algebraically closed in lc'(x) since k' (z) is regular over U. But then /d is the algebraic closure of k in /c'(x) and therefore contains the algebraic closure of k in /c(x), so t h a t K is contained in /c'. C01~0LLAI~Y. I f k ( x ) is regular over lc, so is K. PI~OI'OSlTIOX 4. Let t be a point, k a field, and let t l , " " ", tN be N independent generic specializations of t over lc. Let x be a point of dimension d < N over lc and such that to(x), k ( t ) are linearly disjoint over lc. Then there is an ~ such that t= is a generic specialization of t over l~(x). Call n the dimension of /~(t) over /~. By F-I6, Th. 3, every t= is a specialization of t over k ( x ) ; if none is generic, every t= m u s t have over /~(,) a dimension ~ n - - 1 ; but then ( x , t ~ , ' . . , t . ~ ) has over /~ a dimension <=d+N(n--1) < N n , which is impossibl% since (t~," 9 ",tN) has the dimension N n over lc. PROPOSITIONr 5. Let V be a variety, defined over a field ~; let K be an overfield of ]c and x a point of V. Let A and A' be the prime rational cycles,
[1955a]
219 ON A L G E B R A I C GROUPS OF
TRANSFOR]~ATIONS.
377
over ~ and over K respectively, with the generic point x. Then A is the same as A' if and only if K and ~(x) are linearly disjoint over lc. We m a y replace V by any representative of V on which x has a representative, so that it is enough to prove our result for cycles in the affme n-space. F o r A to be the same as A', it is at any rate necessary t h a t they should have the same dimension, so t h a t K and k ( x ) m u s t be independent over /~; assume from now on t h a t this is so. A m o n g the coordinates of x, let (xl," 9 ", x~) be a m a x i m a l set of independent variables over /~ and therefore also over K ; write y for the point (xl,. 9 -,x~) and z for (X~+x,- 9 . , x ~ ) . By F - V I I 6 , Th. 12, A ~ A " if and only if A . ( y X S ~-~) is the same as A ' - ( y X S~-r) ; by F-VI3, Th. 12, this is so if and only if z has the same complete set of conjugates over K ( y ) as over /c(y), and therefore, by F-I4, Prop. 12 and F-I~, Prop. 6, if and only if K ( y ) and k ( y , z ) ~ T c ( x ) are linearly disjoint over k ( y ) . The l a t t e r condition means t h a t there is no relation ~ u ( ~ ( y ) ~ 0 in which the u~ are linearly independent elements of ]~(x) over /c(y) and the 4p~(y) are in K [ y ] and not all 0. Assume t h a t there is such a r e l a t i o n ; we may write O ~ ( y ) ~ ~,P~J(Y)~s, where the ~j are J
linearly independent elements of K over ~ and the P~s(Y) are in /c[y] and not all 0. Then we have F . v ~ 0 with v ~ F . u ~ P . ~ i ( y ) ; as the vj are in J
k ( x ) and not all 0 because of the assumptions on the u~ and P~s(Y), this shows that, when t h a t is so, K and /c(x) are not linearly disjoint over k. Conversely, assume t h a t there is a relation F. v~j ~ 0 in which the ~j are J
linearly independent elements of K over /~ and the v3. are in k ( x ) and not all 0; as the ~i are then also linearly independent elements of K ( y ) over k ( y ) , this implies t h a t K ( y ) and /c(x) are not linearly disjoint over 7c(y). C0nOLLA~Y. Let V be a variety, deft~ned over a field lc. Let A be a prime rational cycle on V over an overfield K of It. Then, if K' is any field such that tc C K' C K over which A is rational, A i~s prime rational over K' ; of all such fields K', there ~ one smallest one Ko; and an automorphism z of K over ~ transforms A into itself if and only if it induces the identity 0~
K o.
As in the proof of Prop. 5, it is enough to consider cycles in an affme space. Assume t h a t A is prime rational over K and rational over K ' C K, and write it as A ~ ~ n~A~, where the Ai are distinct prime rational cycles over K ' . L e t Z be a component of A1; it is algebraic over K', and so every conjugate of Z over K is a fortiori such over K% so t h a t every component of
220
[1955a] 378
ANDRl~ WEIL.
A is a component of A1 ; therefore we must have A ~ nlA1,. By F-Is, Prop. 26, the coefficient of Z in A is at most equal to its coefficient in A1 ; therefore we have A ~ A 1 . That being so, it follows from Prop. 5 and from F-I6, Th. 3 and F-IT, Lemma 2, that there is a smallest field Ko with the properties stated in our corollary; in fact, if x is a generic point of A over K, and if is the prime ideal in K [ X ] consisting of all polynomials in K [ X ] which are 0 at x, Ko is the smallest subfield of K such that ~ has a set of generators in Ko[X]. As ~ is also the ideal in K [ X ] whose set of zeros is the support I A ] of A, the last assertion follows from F-IT, Lemma 2. PROPOSITION 6. Let V be a variety, defined over a field k, and A a cycle on V ; assume either that A is a divisor on V or that the coefficients in A of all the components of A are ~ 0 modp, p being the characteristic. Then, of all the overfields of lc over which A is rational, there is one smallest one ko, k o is finitely generated over k; and an isomorphism a of leo over k onto some extension of lc leaves A invariant if and only if it leaves every element of ko invariant. Except for the last statement, this result is due to Chow. Let A be any cycle on Vi for every representative V, of V, call A, the sum of the terms in the reduced expression for A which pertain to components with representatives in V. ; then A is rational over an overfield K of k if and only if every A. is rational over K ; and an isomorphism of K which leaves A invariant must leave all the As invariant. Therefore it is enough to deal with cycles on an affine variety V. For such a cycle A, put A ~ ~ nA,~, where A, is the sum of the terms with the coefficient n in the reduced expression for A; then A is rational if and only if every cycle n A , is rational; and an isomorphism which leaves A invariant must leave all the A, invariant. Finally, if n = p~n' with n' prime to p, nA is rational if and only if p~A is rational. Therefore it will be enough to deal with the following two cases: (i) A is a cycl~ in affine space, consisting of a sum of distinct components; (it) A is a divisor on an affine variety V and of the form A ~ qAo, where q is a power of p and Ao is a sum of distinct components. (i) Let ~ be the ideal of all polynomials (with coefficients in the universal domain) which are 0 on the support ] A I of A ; this is the intersection of the prime ideals determined similarly by the components of A. The first assertion in our proposition will then be a consequence of F-IT, Lemma 2, if we prove that A is rational over a field K if and only if ~ has a set of generators in K [ X ] , i.e. if it is the extension to the universal domain
[1955a]
221 ON A L G E B R A I C GROUPS OF T R A N S F O R I ~ A T I O N S .
379
of the ideal 0i N K [ X ] ; the second assertion in our proposition also follows from the same lemma, provided one observes that, if /Co is the smallest field such that 9~ has a set of generators in /Co[X], an isomorphism which leaves A invariant must map /Co onto /Co, i.e. it must induce an automorphism in /%, so that the lemma in question is applicable. I f 9~ has a set of generators (P~) in K[X], the support I AI of A is the set of zeros of the P~ and is therefore K-closed. On the other hand, I A ] must also be K-closed if A is rational over K. In order to prove the equivalence of those two properties, one may then begin by assuming that [A I is K-closed. Consider first the case in which all the components of A are the conjugates of one of them, say Z, over K ; let x be a generic point of Z over ~ ; then A is rational over K if and only if K(x) is separable over K. Put K ' ~ K p-~, this being the smallest "perfect" field containing K. P u t : ~=~nK[X],
~'=~
n K'[X],
and call ~ ' the extension of ~3 to K'[X]. By F-IV2, Th. 4, and F-II1, Prop. 3, !13 and ~ ' consist of the polynomials, in K[X] and in KP[X] respectively, which are 0 at x; they are prime ideals; moreover, if P ' ~ r , some power prn of P ' is in ~3' and hence in ~ ' ; as ~ ' c ~ ' , this implies that ~ ' is primary and belongs to the prime ideal !13'. By F-Is, Th. 3, and F-I~, Prop. 19, we see that ~ 3 ' = ~ ' if and only if K(x) is separable over K, and therefore, as we have shown, if and only if A is rational over K. But, if 9~ is the extension of ~ to the universal domain, ~r must a f o r t i o r i be the extension of !13 to K'[X]. Conversely, if ~ ' = ~ ' , the extension of ~ to the universal domain is the same as that of ~ ' ; but it is well-known and easily verified that the latter must be a "radical" ideal, i.e. one consisting of all the polynomials which are 0 on a closed set; then one sees at once that it must be the same as 9~. This completes the proof in the special case we were considering. Now assume that I A 1 is any K-closed set; then we can write A as the sum of cycles A~ such that the components of each A~ are mutually conjugate over K, and ~ is the intersection of the ideals 9~ similarly determined b~y the A~. P u t : !13,= 9i, fq K [X],
~; = 9~,A g'[x],
and call ~ ' the extension of ~ to K'[X]. I f A is rational over K, all the A~ must be so, so that, as shown above, the 9I~ must be the extensions of the !13~ to the universal domain. I t is then easily seen that 9/ is the extension of the intersection of the ~ , i.e. of 9~ fq K[X]. Assume, on the other hand, that A is not rational over K ; then we have ~ ' - - ~ { for at least one i; from
222
[1955a] 380
AND~ WEIL.
the unicity of the decomposition of an ideal into an intersection of primary ideals, it follows then that the intersection of the ~,', which is the extension of ~f n K [ X ] to K ' [ X ] , cannot be the same as the intersection of the ~ ( , which is 91 N K' I X ] . A f o r t i o r i , 9d cannot then be the extension of 91 N K [X] to the universal domain. This completes the proof for ease (i). (it) Let V be a variety, defined over It, in an aMne space; let Ao be a divisor on V and the sum of distinct components; let q be a power of the characteristic p 5/= 0; put A ~ qAo. I f P is any polynomial which is not 0 on V, denote by ( P ) v the divisor of the function induced by P on V. Call 91 the ideal of all the polynomials P, with coefficients in the universal domain, such that either P = 0 on V or ( P ) v F A . I f A is rational over an overfield K of /c, 91 is then the extension of 91 N K IX] to the universal domain, as follows at once from F - V I I I , , Th. 10. Conversely, assume that 91 is the extension of 91 N K [ X ] to the universal domain; we will prove that A is then rational over K ; our proposition will then follow from this as in ease (i). As a polynomial P is 0 on ]A ] if and only if some power P~ of P is in 91, our assumption on A implies that A is K-closed, and therefore that Ao is rational over K' = Kp -~. Let Z be a component of A. As well-known, there is a polynomial P such that ( P ) v = A o + B, where B has no component in common with Ao ; write P as P = Y. ~,P~, where the ~, are linearly indeper~dent over K' and the P, are in K ' [ X ] ; by F-VIII~, Th. 10, we have (Pi)v } A o for all i; and Z must have the coefficient 1 in at least one of the P,, since otherwise it would occur in B ; if we call that polynomial P', P' is then in K ' [ X ] , Z has the coefficient 1 in (P')v, and we have (P')v }Ao. But then P'q is in 91, and therefore, by hypothesis, may be written as Y~,j.Q~, where ihe y Oj are in 91 N K [ X ] . The latter fact implies that Z has at least the coefficient q in all the ( @ ) v ; as it has the coefficient q in P'q, it must have the eoeMeient q in one at least of the divisors ( Q j ) v ; as these divisors are rational over K, this implies that, if A~ is the sum of Z and its conjugates over K, qA~ is rational over K. As this is so for every component Z of A, A is therefore rational over K. PROPOSITION 7. Let U, V be two varieties, defined over a field k; let F be a l~-closed subset of U X V. Then the set A of the points a on U such that a )< V c F is k-closed. Let W~,. 9 -, Wm be those components of F which have the "projection" V on V (in the sense of F-IV3, F - V I I 3 ) ; if v is a generic point of V over /~, W~ has a generic point over ~ of the form (u~,v) ; and a e A if and only if,
[1955a]
223 ON ALGEBR~_IC GROUPS OF T R A N S F O R ~ i A T I O N S .
381
for v' generic over k(a) on V, (a, v') is a specialization of some (u~, v) over k. Let V~ be any representative of the abstract variety V; let vl be the representative of v on V1; the ambient affine space for F~ being embedded in a projective space, let Vo be the locus of v~ over k in that projective space. Let Fo be the union of the loci of the points (u~,vl) over ~ in U X Vo; Fo is k-closed on U X V o . Then A is the set of the points a on U such that Fo A (a X Vo) has a component of dimension ~ dim (Vo). As Vo is complete, our conclusion is now contained in Lemma 7 of my paper in Math. Ann., vol. 128 (1954), p. 104. PRoPosITIO~ 8. Let ~ be a mapping of a variety U into a variety V; let k be a field of definition for U, V and ~. Then the set of points of U where ~ is defined is k-open. (i) Assume first that U is an affine variety and V is the affine space of dimension 1. Let x be a generic point of U over k; put y ~ ( x ) . Let be the set of all polynomials P in /c[X] such that P ( x ) y is in k[x] ; this is an ideal in k [ X ] , containing the ideal ~ of those polynomials which are 0 at x and therefore on ~/. Since y may be written as Q ( x ) / P ( x ) , with P, Q in k [ X ] and P ( x ) ~ = 0 , we have ~[:/=~. As the points where ~ is not defined are the zeros of ~, the set of such points is k-closed. (ii) Take F as in (i), and assume that U is an abstract variety, with the representatives U~, on each of which a " f r o n t i e r " F~ (i. e. a k-closed set) is given, according to the definitions in F-VIIi. Call F~' the k-closed subset of U~ where ~ is not defined; the set F of the points of U where ~ is not defined is then the union of the images of the sets F~" N (U~--F~) by the canonical birational mappings of the U~ into U. I t is easily seen that F must be k-closed provided the following assertion is true: if x is a point of U with a representative x~ on some U~ which is a generic point over ~ of a component of F~', then every specialization zJ of x over k is in F. I n fact, let fl be such that x' has a representative x f on U~ ; then x must also have a representative x~ on U~, and, from the biregularity of the correspondence between U~, U~ at (x~,x~), it follows that x~ must be in F f ; as x f is a specialization of x~ over k, and as F f is ~-closed, x~' must then be in F~', so that x' is in F. This proves our result for this case. It follows trivially from this that our result remains true when U is an abstract variety and V is an affine space or more generally an affine variety. (iii) Let U be an abstract variety and let V be a k-open subset of an affine variety V~ ; let Vo be the projective variety whose part " a t finite dis-
224
[1955a] 382
ANDR]~ WEIL.
tance" is V1; then V o - - V is a L-closed subset Fo of Vo. Call r the graph of ~ on U X V o ; the set F of the points of U where ~, considered as a mapping of U into V, is not defined, is then the union of the set F1 of the points of U where r is not defined as a mapping of U into V1 and of the set-theoretic projection of r N (U X Fo) on U. As Vo is complete, the latter set coincides with the "projection" in the sense of F-IV3 and F-VII~ and is k-closed (e.g. by F-VIIi, Prop. 10 and 11); and F1 is to-closed, as shown in (it). Therefore F is k-closed. (iv) Let U, V be arbitrary abstract varieties; let x be a generic point of U over k; let the Vs be those representatives of V on which ~b(x) has a representative r and let Fs be the "frontiers" on the Vs. Then 4) is defined at a point of U if there is an ~ such that ~bs, considered as a mapping of U into V s - - F s , is defined there. Therefore the set where r is not defined is the intersection of the sets where the Cs are not defined; as the latter sets are k-closed by (iii), this completes the proof. COROLLARY 1. Let V be an abstract variety, defined over k, with the representatives Vs. Then, for each ~, the set ~s of the points of V which have a representative on Vs is k-open; and the canonical correspondence between V and Vs is an everywhere biregular mapping of ~ts onto V s - - F a if Fa is the frontier for V~. Let x be a generic point of V over k, and let xs be its representative on Vs; if we put xs ~ Ca(x), r is the "canonical correspondence" between V and Vs. Then ~s is the set of points where Ca, considered as a mapping of V into V s - - F ~ , is defined; it is k-open by Prop. 8. The rest is obvious. COROLLARY 2. Let V be an abstract variety, defined over a field k. Then there is a finite covering of V by k-open subsets of V, each of which is biregularly equivalent to an a~ne variety. Corollary 1 says that V has a covering by the k-open sets ~ , each of which is biregularly equivalent to the k-open subset V a - - F s of the aitine variety Vs. I t is therefore enough to prove our assertion for a k-open subset V - - F of an affine variety V defined over k. Let ~[ be the set of all polynomials in k[X] which are 0 on F ; it is an ideal in k[X], and, as F is /~-closed, it is the set of zeros of ~[. Let P~," 9 ",P,~ be a set of generators for ~ ; as F :/: V, they are not all 0 on V, and we may assume that P~,. 9 P~ are not 0 on V while P ~ + ~ , ' ' ' , P ~ are 0 on V, with l ~ r ~ m . For 1 "< p ~ r, call ~p the k-open subset of V consisting of the points where Pp is not 0 ; t h e ~ a r e a c o v e r i n g o f V---F. L e t x ~ ( x l , ' - .,x~) b e a g e n e r i c
[1955a]
225 ON ALGEBRAIC GROUPS OF TRANSFORIVIATIONS.
383
point of V over k; let Vp be the locus of the point (z,.
9
z,, 1/Pp ( x , . 9 -, xn) )
in the affine space of dimension n ~-1.
Then Vp is biregularly equivalent
to tip. COROLLARY 3. Let V be a variety, defined over a field k. The set of points of V where V is normal (resp. relatively normal with respect to k) is a k-open subset of V. Let V* be the variety derived from V by normalization with reference to the smallest perfect field k ' ~ kp-~ containing k (resp. with reference to k ) ; let x be a generic point of V over k ; let x* be the corresponding point of V*, which is generic over k' (resp. over k) on V*. We may then write x * ~ @(x), where @ is a mapping of V into V*, defined over k' (resp. over k), Then the points where V is normal (resp. relatively normal) are those where is defined. As any k'-open set is also k-open, this proves the corollary. PROPOSITIO~ 9. Let V be a variety, defined over a field k; let F be a closed subset of V. For F to be k-closed, it is necessary that it should contain all the specializations over k of all its points; it is sufficient that it should contain all the generic specializations over k of all its points, or also that it should be invariant under all isomorphisms over k of a common field of definition K D k for its components. The necessity of the first condition follows from F-IVz, Th. 4; we first prove that this condition is sufficient. I n fact, it implies that, if z is a generic point over K of a component Z of F, the locus Z" of z over ~ is contained in F ; as Z is the locus of z over/~, Z' contains Z ; as z cannot be in any other component of F than Z, we get Z ' ~ Z ; thus all components of F are algebraic over k, and then F-IVz, Th. 4, shows that all the conjugates of Z over k must be contained in F. Now we show that the second condition implies the first one. Let x be any point of F and let x' be a specialization of x over k. Then if Y is the locus of x over ~, F-IV2, Th. 4, shows that x' must be on a conjugate V' of V over k. Let x 'p be a generic point of V' over /~; then x~' is a generic specialization of x over k by F-IVz, Th. 4, and is therefore in F by hypothesis, and x' is a specialization of x" over /7" and a f o r t i o r i over K and so is in F since F is K-closed. Finally the last condition implies the second one; for let x' be a generic specialization over k of a point x in a component Z of F ; then the isomorphism of k(x) onto k ( x ' ) over k which maps x onto x' can be extended to an isomorphism a of K ( x )
226
[1955a] 384~
ANDa~ WEIL.
onto a field Kr and then x' is on Z r by F-IV2, Th. 3, Coroll. 2, and is therefore in F if F is invariant under ~.
Let W be a subvariety of a product U X V, with the "'projection'" U on U; let k be a field of definition for U, V, W. Then the set-theoretic projection of W on U contains an open subset of U; and the union of all such sets is k-open. PROPOSITION 10.
The assumption means that, if (u,v) is a generic point of W over k, u is generic over k on U. Let V1 be a representative of V on which v has a representative v l ; F~ being the corresponding frontier, p u t V I ' = V~--F1, so that vl is in V~'; let W1 be the locus of (u, vl) over k on U X VI'. Let Vo be the projective variety whose p a r t " a t finite d i s t a n c e " is V~; p u t F o = V 0 - - V I ' ; fhis is a k-closed set on Vo. The set WI (1 (u X VI') can be writteIl as u X X, where X is either V~r (in the trivial case W = U X V) or else a k ( u ) - c l o s e d subset of V~'; as v~ is in X, X is not empty, so t h a t we can choose in it a point w which is algebraic over k ( u ) . Let W r be the locus of (u,w) over ~ on U X Vo, which has the same dimension as U ; call n t h a t dimension. Then the set C = W ' A (U X Fo) is a ~-closed subset of W', so t h a t all its components are of dimension ~ n - - 1 . As Vo is complete, the set-theoretic projection C' of C on U is then a ~-elosed subset of U. Let a be any point in U - - C ' ; as Vo is complete, there is a point ( a , b ) on W' with the projection a on U ; as ~ is not in C', b cannot be iI1 Fo and is therefore in V~', so that (a,b) is in WI. Therefore the ~-open set U - - C ' on U is contained in the set-theoretic pr6jection of W~ and a f o r t i o r i in that of W. The last assertion in our proposition is then an immediate consequence of the sufficiency of the last condition in Prop. 9. PROPOSITION 11. Let U, V, W be three varieties and U X V into W, all defined over a field k. Assume that, f is defined at (a,x) for x generic on V over k(a). Let those a e U such that, for x generic over k(a) orv V, f ( a , x ) k(a) on W. Then ~ is either empty or k-open on U.
f a m g p i n g of for every acU, ~ be the set of is generic over
Let r be the dimension of W ; for z generic over k on W, let z~,. 9 .,z~ be r algebraically independent elements of k ( z ) over k ; p u t z , ~ $ , ( z ) , where ~b* is a ~unction on W, defined over k. I t is clear that a point z' of W is generic over an overfield K of /r if and only if the ~, are all defined at z' and their values r are independent over K. L e t u , x be independent generic points of U, V over k; we m a y assume t h a t f ( u , x ) is generic over k on W, since otherwise ~ is empty. P u t f , = $ ~ o f ; ~ is then the set of
[1955a]
227 ON A L G E B R A I C GROUPS OF TRANSFOR~][ATIONS.
385
those points a on U such that, for x generic over ]c(a) on V, the f i ( a , x ) are all defined and are algebraically independent over ]~(a). Take u, x as above; assume t h a t u is not in e ; we prove t h a t e m u s t then be empty. I n fact, the assumption on u means t h a t there is a polynomial 2 with coefficients in ~ ( u ) such that p(fl(~,x),.
- ",fr(~,x))
=o.
Write P ~ Y. t,M~(Z), where the My(Z) are monomials (with coefficient 1) y
in the indeterminates ZI," 9 ",Zr, and the t, are in k(u) and not all 0. Let a be any point of U, and take J generic over K ~ k (a) on V. Take a variable quantity X over ]c(u, z) ; extend the specialization u---~ a over k to a /~-valued place ~- of the field ]c(u, ,~) such t h a t the elements ;~t~ of ]~(u, X) are all finite and not all 0 at ~r; call t / t i l e value of ~t~ at ~r. As k ( u , ; Q and ]c(x) are independent regular extensions of k, the place ~ of k(u,X) and the isom o r p h i s m of k ( x ) onto k ( x ' ) over k which maps x onto x' make up a specialization of the set of quantities k(u,;~) U k ( x ) , which can be extended to a place ~r' of k ( u , A , x ) at which u, x and the ;~t, have respectively the values a, x" and t / . I f the f i ( a , x ' ) are not all defined, a is not in fl; if they are all defined, they are the values at ~' of the e l e m e n t s / ~ ( u , x ) of /~(u, ;~, x). I n the l a t t e r case, the relation
~ P ( f ~ ( ~ , ~ ) , . 9 . , f , ( ~ , ~) ) = o, taken at ~', gives an algebraic relation between the f~(a,~) whose coefficients t / are in /~ and are not all 0; this implies t h a t the f~(a,J) are not independent over K ~ ( a ) , so t h a t a is again not in ~t. This shows that, for ur a m u s t be empty. F r o m now on, therefore, we may assume t h a t n is in ft. We prove now t h a t a m u s t contain a /~-open set. Since the assumptions and the conclusion of our proposition are not affected if V is replaced by any birationally equivalent variety to V over ~ (the m a p p i n g f being transferred to the l a t t e r in an obvious m a n n e r ) , we may take for V an afflne v a r i e t y ; p u t x = (Xl,' ' ",x,~). Then we can write the f~ as f~(~, ~) = e ~ ( ~ ) / f o ( ~ ) , where Po, P~,'" ",P~ are polynomials in the indeterminates X ~ , ' ' - , X , ~ with coefficients in ~ ( u ) , and P , (x) ~ 0 for 0 G i _< r. Call M , ( X ) , with 0 ~ v G N , all the monomials in Xz,- 9 .,X~ which either are of degree 0 or 1 (i.e. equal to one of the monomials 1, X~," . .,X~) or occur in one at least of the P~; call 2 the point in the projective space p N with the homo-
228
[1955a] 386
~ND~ W~IL.
geneous coordinates (My(x)), and call Y- the locus of ~ over k. We maY' replace V by the birationally equivalent F ; then, writing V, x instead of ~, 5, and calling (xo,' 9 ",xiv) the homogeneous coordinates for x, we see that the f~(u,x) are expressed as z~/zo, with N
z~~ F~t~x~ ~:o
(0 ~-- i ~ r ) ,
where the t~ are elements of k (u). I f V is contained in any linear subvariety of pN, then the smallest linear subvariety of pY which contains V is defined over k; if this is of dimension N', we can express N - - N ' of the coordinates x~ linearly in terms of the others, with coefficients in k; thus we may assume that V is not contained in any linear subvariety of pN. We may write t ~ , = r where the r are functions on U, defined over k; as zo is not 0, we may assume that too ~ 1. By Prop. 8, the subset U" of U where all the ~ are defined and finite is k-open. Call n the dimension of V; as n is then the dimension of k ( u , x ) over k ( u ) , and the ft(u,x) are independent over k(u), we have r ~ n . P u t z j ~ , w ~ x ~ for r + 1 ~-- j ~ n, where the %.~, for r + 1 ~ j ~ n, 0 --~ v --~ N~ are (n - - r ) ( N + 1) independent variables over k ( u , x ) ; call S the affine space of dimension ( n - - r ) ( N - I - 1 ) . By F-II~, Prop. 24, the n - - r quantities z~+~/Zo,'' ",z,/Zo are algebraically independent over the field K ~ k ( U , W, Z1/ZO," " ", Zr/Zo).
Now take any a~ U'; take 2, v~ generic and independent over k(a) on V, S ; put ~ ~ r (a), 5~~ Z i ~ p for 0 ~ i _~ r, and ~ ~ Z w-j~ for r + 1 --__j ~ n. Y
As too ~ 1, and V is not contained in any linear variety, 2o is not 0; therefore, if we put f j ~ Z J Z o for r + l ~ ] _ ~ n , the functions f~,' 9 ' , f ~ on U" X S are defined at (a, ~) and have the values 2~/~o," 9", 2~/~o respectively. I f one assumes that ~/5o," 9 ", ~/2o are algebraically independent over k(a, ~) this implies a f o r t i o r i that ~/2o," 9 ",~/2o are so over k(a), i.e. that a ~ . Therefore, if we prove that there is a k-closed subset C of U' }( S such that, with the notations just introduced, the quantities ~/2o," 9",~n/2o are algebraically independent over k ( a , ~ ) whenever ( a , ~ ) is not in C, it will follow that ~ contains the set of those points a on U' such that a ~ S is not contained in C; and this set wilt be k-open by Prop. 7. I n other words, as long as we merely wish to prove that ~ contains a k-open set, it is enough to prove it for U ' / S and the functions f ~ z~/zo for 1 <: i ~ n instead of for U and for f , . 9 .,f,.. This means that, writing U instead of U ' X S,
[1955a ]
229
387
ON ALGEBRAIC GROUPS OF T R A N S F O R M A T I O N S .
it is enough to prove our assertion under the additional assumption r ~ n, the ~ being now everywhere defined on U, with boo ~ 1. This being now assumed, put Z~+l~ ~ w~x~, where the w, are N ~ 1 independent variables over ~ ( u , x ) . As l r is of dimension n over ~ ( u , w ) , there is a homogeneous polynomial P, with coefficients in k ( u , w ) and not all 0, such t h a t P(zo," 9 ", z~, zn+l) = O, and P is uniquely determined up to a factor in ~:(u,w). As zl/zo," " ",zn/zo are algebraically independent, there is at least one t e r m in P where z~§ occurs with a non-zero exponent; after m u l t i p l y i n g P with a suitable element of ~(u, w), we may assume t h a t the coefficient of this term is 1. Write all the other coefficients in P as C p ( u , w ) , where the Cp are functions on U X S n§ defined over It. We now prove our assertion about ~ by showing t h a t ~ contains the set of all points a on U such t h a t all the Cp are defined at (a, ~ ) for ~ generic on S n§ over ( a ) ; this is a k-open subset of U by Prop. 8 and 7. I n fact, let a be a point with t h a t p r o p e r t y ; take ~ generic over ~(a, ~ ) on V. P u t t~v ~ ~b~(a), these being all defined, according to our present assumptions ; p u t 2t = ~ ~ for 0 --< i <: n, and ~§ ~ ~ ~p2~. If we specialize the relation P(zo,"
9
",z.+~)
=0
over the specialization (~,~,~2) of ( u , w , x ) with respect to k, since the Cp are all defined at ( a , ~ ) , a homogeneous relation 5o," 9 ",5~+~ with coefficients in ] c ( a , ~ ) , containing 5~+~ with a exponent in a term of coefficient 1. This shows that 5~+~/5o is then over the field L ( ~ ) , where we have p u t L =
~ ( ~ , ~1/~o," 9 ,
we get, between non-zero algebraic
~,/~o).
Now take n ( N + l ) independent variables w~ over /c(a), for l _ < i < : n , 0 --< v --< N ; p u t y~ = ~ w~2~ for 1 --< i --< n ; what we have proved above shows that, for each i, YJZo is algebraic over L ( w ~ o , ' ' ' , w i z r fortiori over the field
and therefore a
L ' = L ( w~o, " 9 ", w~N ) = ~ ( a, W~o, " 9 ", w,~r ~/5o, " 9 ", 5~/5o ).
On the other hand, one sees j u s t as before, using F-II~, Prop. 24, t h a t the y~/5o, for l < - - i < _ n , are algebraically independent over the field ~(a,w~o," 9 " , w , ~ ) ; as they are algebraic over L', this implies that L ' has at ]east the dimension n over the l a t t e r field, so that the 2J2o, for 1 <--i <--n, m u s t be algebraically independent over it. But then they m u s t a f o r t i o r i be so over ~ ( a ) , which means t h a t a is in ft.
230
[1955a] 388
ANDR]~ WEIL.
This completes the proof of the following statement: the assumptions being again those of Prop. 11, ~ must either be empty or contain a 7~-open subset of U. Now we prove Prop. 11 by induction on the dimension of U, the conclusion being trivially true if that dimension is 0. Assume that is not empty; put X = U - - a ; we have proved that X is contained in a k-closed subset C of U. Call U~ the components of C; they are algebraic over k, and of dimension < dim(U). By the induction assumption, ~ A U~ is either empty or a ~-open subset of ~f~; in both cases its complement C~ on U,~ is a ~-closed subset of U. As X is the union of the C~, this shows that X is ~-closed. As it is obviously invariant by all automorphisms of ~ over ~, it must then be k-closed. PR0rOSlTIO~ 12. Let U be a variety defined over a field ~ ; let F be a closed subset of U. Then, among all the ~-closed subsets of U contained in F, there is one maximal set Fo. Let K be the smallest common field of definition containing k for all the components of F ; let ~ run through all the isonmrphisms of K over into the universal domain. As such an isomorphism ~ ]eaves all k-closed sets invariant, every k-closed subset of U which is contained in F is contained in all the sets F r and therefore in their intersection Fo; Fo is closed, since it is the intersection of closed sets; and it is k-closed, by Prop. 9. This proves the proposition. PROPOSITION 13. Let U be a variety defined over an infinite field k; let F be a closed subset of U X U. Then there is a point a on U, separably algebraic over k and such that no pair (a',a") of distinct conjugates of a over k is in F. Applying Prop. 12 to U X U, F and the algebraic closure ~ of k, we see that there is a ~-closed subset Fo of U }( U such that a subvariety of U }( U which is algebraic over k is contained in F if and only if it is contained in Fo ; this applies in particular to algebraic points over k oi1 U X U. By replacing F by the union of all conjugates over ]~ of all the components of Fo, we see that it is enough to prove our result in the case in which F is k-closed. We may assume that no component o[ F is contained in the diagonal oi U }( U, since the omission of such components does not affect the content of our proposition. Furthermore, we may, in order to prove our proposition, replace U by any k-open subset of U; in view of this, we first replace U by the set of its simple points, and then use Corollary 2 of Prop. 8 to replace U by an affine variety. Thus we may assume that U is a non-singular affin~
[1955a]
231 ON A L G E B R A I C GROUPS 0 E
389
TRANSPOR]VIATIO1WS.
variety, defined over ]C, and that F is a k-closed subset of U X U, no component of which is contained in the diagonal of U X U. Let n and N be the dimensions of U and of t h e ambient affine space, respectively. The case n ~ N is trivial, since in that case any rational point of U over k, e.g. 0, would solve our problem; therefore we assume n < N. Consider all sets of n linear equations: N
(1)
(1_<: i_<,~),
Xt,~x~=t,o V=I
and identify the set (1) with the point t = (t~o,t~) in the afl]ne space T of dimension n ( N + 1). In the space T, we consider the following sets: (a) Call A the set of those points t for which the left-hand sides of (1) are not linearly independent; as A can be described as the set of zeros of certain determinants, it is ko-closed, ]co being the prime field (one could easily see that A is in fact a variety, defined over ko). Put T ' = T - - A ; for
t~T', (1) defines a linear variety L(t) of dimension N
n.
(b) Take t generic over k on T ; by F-V1, Th. 1, U A L ( t ) is not empty, and, if u is a point in it, u is algebraic over k ( t ) and is generic on U over the field K=]C(t11," 9 .,t,,~). As the t~o are then in K(u), we have k(u,t) = K ( u ) , so that k(u,t) is a regular extension of k. Let W be the locus of (u,t) over ]C on U }< T ; by F-u Prop, 4, if t'eT', a point u' is in U A L ( t ' ) if and only if (u',t') is in W. By Prop. 10, there is a It-closed subset B of T such that T - - B is contained in the set-theoretic projection of W o n T ; then, if t ' ~ T - - (A U B), UN L ( t ' ) is not empty. ( c ) Let P p ( X ) ~ 0 , for l ~ p ~ r , be a set of equations for U with coefficients in k ; put Pp~ = OPp/OX~. Let D be the subset of U X T consisting of the points where the matrix
t~ P ~ , ( u )
,
(l<-i~n,l
=
l
--
is of rank < N ; since this can he expressed by the vanishing of determinants, D is a ]c-closed subset of U X T (as U is non-singular, it could be shown that D is actually a variety, defined over k). As W is not contained in D, D N W is a k-closed subset of W (also, in ~act, a variety), so that its cornportents have a dimension < n ( N + l ) . Let D' be the "projection" of D N W on T (i.e. the closure of the set-theoretic projection); this is a k-closed subset of T. Let u' be a point in U N L(t'), for t ' c T ' ; then, if L(t') is not transversal to U at u', (u', t') must be in D and therefore in
13
232
[1955a] 390
DNW,
ANDRE WEIL.
and t' must be in D'. Therefore, if t ' r transversal to U at every point of U A L(t').
L(t') is
(d) Let X be any component of F ; let (u,v) be a generic point of X over ~. As X is not in the diagonal of U X U, we have u :f: v and may assume for instance that Ul-7(=Vl. Take the t,~ independent over k(u,v) for 1--
[1955a]
233 ON ALGEBRAIC GROUPS OF T R A N S F O R M A T I O N S .
391
P u t n ~ [U :/el ; as U is separably algebraic over lc, there are n distinct isomorphisms a of U into the algebraic closure of k. Each a can then be extended uniquely to an isomorphism, which we also denote by a, of U(x) onto U~(x) over ~ ( x ) . Let z be an element of U ( x ) , integral over U [ x ] and therefore also over k [ x ] ; then all the z~ are also integral over /~[x]. I f ~1," ' " , ~ are n linearly independent elements of to' over ~, it is wellknown that d e t ( ~ ~) is not 0; therefore all the z% and z among them, can be expressed as linear combinations of the n elements w ~ = ~ ( z ~ ; but these are integral over ~[xJ and are traces o v e r / c ( x ) of elements o f / d ( x ) , so that they are in ]~(x) ; as ~ [ x ] is integrally closed, the w~ are therefore in k [ x ] , so that z is in U [ x ] . T H E UNIVERSITY OF CHICAGO.
[ 1955b] On algebraic groups and homogeneous spaces I n a recent paper in the same JOURNAL ( [4] 0J~ the bibliography ; quoted hereafter as A G ) , I gave some results oil algebraic groups and transformationspaces, which supplement those in my Varidtds abdliennes ( [ 3 ] ; quoted as V A ) . Applications will now be made of that theory to somewhat more specific questions. I n no. 1, a rather general procedure is described for obtaining, from a given transformation-space S with respect to a group G and from a suitable cycle on S, another transformation-space with respect to the same group. As shown in no. 2, this includes as a special case the construction of coset-spaces and of factor-groups; thanks to the main theorem in AG, these can now be defined without enlarging the groundfield, whereas such an enlargement was required in their construction as previously given by S. Nakano ( [ 2 ] ) ; except for this, we have substantially followed his method. The rest of this paper is chiefly devoted to "principal homogeneous spaces," i.e. to those homogeneous spaces on which the group operates in a simply transitive manner. The pair consisting of such a space and of one point on it does not differ materially from a group ; thus there is little incentive for studying those spaces as long as one is not paying any attention to the groundfield or if the groundfield is algebraically closed. But it can happen that a principal homogeneous space contains no rational point over the groundfield over which it is defined; an example of this is given by the plane curve Xa-] - p Y Z ~ p 2 over the rational number-field, where p is a rational prime; this may be considered as a principal homogeneous space with respect to its jacobian variety, which is the plane curve Y Y ~ 4 X ~ - - 2 7 . More generally, Chow's work (cf. [1]) has shown that it is not always possible to map a curve " c a n o n i c a l l y " into its jacobian variety by a mapping defined over the groundfield, but that the curve can always be so mapped into a suitable principal homogeneous space with respect to its jacobian variety This, among other results, will be proved again here by a different method, which can be extended at once to a variety V of higher dimension and to its Albanese variety, provided the groundfield is one over which the latter is defined. I t will also be shown that the classes of principal homogeneous Received December 27, 1954. 493 Reprinted from the Am. J. of Math. 77, 1955, pp. 493-512, by permission of the editors, 9 1955The Johns Hopkins University Press. 235
236
[1955b] 494
ANDRI~ WEIL.
spaces with respect to a commutative group can be arranged into a torsiongroup, i.e. a group whose elements are all of finite order; and it follows at once from the results of no. 5 t h a t this group must be countable if the groundfield is finitely generated over the prime field. There seems to be no reason why it should be finite, even if the groundfield is the field of rational n u m b e r s ; a more detailed investigation of its structure, e.g. for the case of an elliptic curve over the field of ratiofial numbers, would be of considerable interest from the point of view of the theory of diophantine equations. 1. Let G be a group and S a transformation-space with respect to G, both defined over a field /r Let Z be either a divisor on S or a cycle on S whose components have coefficients which are prime to the characteristic of the universal domain. We denote by sZ, for any s e G, the transform of Z by the m a p p i n g u-->su of S onto itself. Let K be an overfield of k over which Z is rational. Let x be generic over K on G; by Prop. 6 of the Appendix of AG, there is a finitely generated extension k ( t ) of k which is the smallest overfield of k over which xZ is rational ; as xZ is rational over K(x), we have k(t) C K(x). I f x' is also a generic point of G over K , and ~ is the isomorphism of K(x) onto K(x') over K which maps x onto x', a will transform xZ into x'Z; if t' is the image of t under a, k ( t ' ) will then be the smallest overfield of k over which x'Z is rational, and, by Prop. 6 of the A p p e n d i x of AG, the cycle x'Z depends only upon t ' ; in other words, if a~ is an isomorphism of K(x) onto a field K(xl'), m a p p i n g x onto x l ' and t onto tl', we have x'Z~x~'Z if and only i f = t ' ~ t l ' . I n particular, take x ' = y x , with y generic over K(x) on G; then k ( t ' ) is the smallest overfield of k over which yzZ is r a t i o n a l ; as yxZ can be written as y(xZ), it is rational over /c(y, t ) , so that /c(t') C k(y, t) ; similarly, xZ can be written as y~(yxZ) and is therefore rational over k(y, t ' ) , so that k ( t ) C k ( y , t ' ) . This shows that lc(y, t ) = k ( y , t ' ) . P u t now z==yx, so that we have k ( t ' ) C K ( z ) ; take y' generic over K(x,y) on G, and call r the isomorphism of K(z) onto K(y'z) over K which maps z onto y'z=y'yx; let t" be the image of t' under r. Then t " is the image of t under the isomorphism r O a of K(x) onto K(y'yx) over K which maps x onto y'yx. I f Z is such that k ( t ) is a regular extension of k, then we may call T the locus of t over/c and, with the above notations, we may write t'-~-g(y, t), where g is a m a p p i n g of G X T into T, defined over /~. Then the results we have just proved mean that g satisfies ( T G 1, 2) of AG, no. 2, i.e. t h a t it is a normal law between G and T. A p p l y i n g now the main theorem of AG, we .e'er the following result:
[1955b]
237 ON A L G E B R A I C GROUPS A N D H 0 ~ I O G E N E O U S
SPACES.
495
PROPOSITION 1. Let G be a group and S a transformation-space with respect to G, both defined over a field It. Let Z be either a divisor on S or a cycle on S whose components have coeficients which are prime to the characteristic. Let K be an overfield of k over which Z ks rational; and assume that, if x is generic over K on G, the smallest extension U of k over which xZ is rational ks a regular extension of k. Then there is a transformationspace T with respect to G, defined over ~, and an everywhere defined mapping F of G into T, defined over K, such that the point t ~ F ( x ) is generic over k on T, that l c ' ~ ( t ) , and that F ( s s ' ) ~ s F ( s ' ) for all s, s' on G. For any s, s" in G, we have F ( s ) ~ F ( s ' ) if and only if sZ ~ gZ. If Z ks algebraic over ]c, one can talce for T a homogeneous space with respect to G. The existence of a transformation-space T with a generic point t over such that ] c ' ~ / c ( t ) has been proved above; moreover, with the same notations as above, we have It(t) C K ( x ) , k ( t ' ) C K ( y x ) , t ' ~ y t ; as K ( x ) , K ( y x ) are independent extensions of K , this shows, if K is algebraic over k, t h a t k ( t ) , k ( t ' ) are then independent extensions of k, i.e. t h a t T is pre-homogeneous, so that, by the m a i n theorem of AG, we m a y replace it by a birationally equivalent homogeneous space. As ]c(t) C K ( x ) , we m a y write t ~ F(x), with F defined over K ; then we have yt ~ F(yx), i. e. F(yx) ~ yF(x), for y generic over K ( x ) on G. This m a y be written as F ( x ) = y - ~ F ( y x ) , which shows, if s is any point of G and y is taken generic over K ( s ) on G, that F is defined at s. As F is everywhere defined, the relation F(yx) ~ y F ( x ) implies F(ss') = s F ( s ' ) for all s, s' on G. Now, for any s, s" on G, take x generic over K ( s , s p) on G; then xs, xs' are both generic over K on G, and therefore, if a is the isomorphism of K ( x s ) onto K ( x s ' ) over K which maps xs onto xs', ~ will map F ( x s ) onto F ( x s ' ) and the cycle xsZ onto xs'Z; then, as we have seen above, we have x s Z ~ x / Z if and only if F ( x s ) ~ F ( x s ' ) ; as the latter relation can be written x F ( s ) ~ x F ( s ' ) , so that these two relations are respectively equivalent to s Z = s ' Z and to F ( s ) ~ F ( s P ) , this completes our proof 2. We first apply Prop. 1 to the construction of the homogeneous space defined by a group G and a subgroup of G. PROPOSITION 2. Let G be a group, defined over a field It; let Z be a rational cycle over k on G, consisting of components with coe~cient 1, and such that its support I Z I is a subgroup of G. Then there ks a homogeneous space H with respect to G, defined over t~, and a rational point a over lc on H, with the following properties: (i) if we put, for a generic x over t~ on G,
238
[1955b] 496
n~DR~ WEIL.
F ( x ) ~ xa, the mapping F of G onto H determines a one-to-one mapping of the cosets of I Z] in G onto the points of H ; (it) l~(x) is separable over k ( F ( x ) ) ; (iii) if ~ is a mapping of G into a variety V, deflated over an overfield K of te, and such that ~ ( x s ) ~ r whenever s ~ l Z [ and x is generic over K ( s ) on G, there is a mapping ~ of ~t into V, defined over K, such that ~ t p o F . I f ] Z J is a normal subgroup of G, then one can define on H a group-law, defined over ~, such that F is a homo~norphism of G onto H. The " s u p p o r t " of a cycle was defined at the beginning of the Appendix of AG. The assumption on ]Z I implies that Z has one and only one component Zo containing e, that this is a subgroup of G, and that all the components of Z are cosets of Zo in G. We apply Prop. 1 to the cycle Z on S ~ G, G acting on itself by the left-translations, and to the field le; if x is generic over 1~ on G, the smallest extension of /c over which xZ is rational is contained in k ( x ) , and hence, by AG-App., Prop. 3, Coroll., it is regular over /e. Therefore, by Prop. 1, there is a homogeneous space with respect to G, which we now call H, defined over k, and a mapping F of G into H, defined over le, with the properties stated in Prop. 1; in particular, F is everywhere defined, t ~ F ( x ) is generic over k on H, and k ( t ) is the smallest extension of k over which xZ is rational. I f we put a ~ F ( e ) , a is rational over k, and we have, for all s~G, F ( s ) ~ s a . I f s, s' are any two points on G, we have sa ~ s'a if and only if sZ ~ s'Z, i.e. if and only if s-~s'Z ~ Z; by the assumption on Z, the latter relation is equivalent to s-~s'~ J Z 1. Thus the points of H are in a one-to-one correspondence with the cosets of t Z] in G. Call F the graph of F on G X H . For any b r F N ( G X b ) is the set of those points (s, b) which are such that s a = b; in particular, if x is generic over k on G and if we put t = F ( x ) = x a , FN (GXt) is the set of the points (s, t) such that sa ~ xa, i.e. x - ~ s ~ l Z I ; this set can be written as I x Z l X t . As F . ( G X t ) is the prime rational cycle overTc(t) on G X H with the generic point (x, t) over k ( t ) , this shows that the prime rational cycle Z' over /e(t) on G with the generic point x has the same components as the cycle x Z ; as the latter is rational over k ( t ) and its components have the coefficient 1, this implies that Z ' ~ x Z . As the components of the prime rational cycle with the generic point x over k(t) have the coefficient 1, te(x) must be separable (i. e. "separably generated") over /~(t). As to (iii), let x be generic over K on G; put t = F ( x ) and w = r ; let w' be any generic specialization of w over K ( t ) ; this can be extended to a generic specialization x" of x over K ( t ) ; we have then w ' ~ r
[1955b]
239 ON ALGEBRAIC GROUPS AND :[tOMOGENEOUS SPACES.
497
and t ~ F ( x ' ) , and the latter relation implies that x' is on I xZ I, i.e. that it is of the form xs with s e l Z I . Let ~ be generic on G over K ( s ) ; we have (~(~s) ~ r specializing ~ to z over K ( s ) , we get d p ( x s ) ~ r since both sides are defined, i.e. w ' ~ w. This shows that w is purely inseparable over K ( t ) ; as it is at the same time rational over K ( x ) which is separable over K ( t ) , it is therefore rational over K ( t ) , and we may write w ~ r where r is a mapping of H into V, defined over K. This proves (iii). Finally, assume that I z t is a normal subgroup of G; let x, y be indeWe have pendent generic points of G over /~; put t = F ( x ) , u ~ F ( y ) . F ( x y ) ~ xF (y) ~ xu; this is a function of x, defined over I Z I, we have x s y = x y s ' with s ' = y - l s y e f Z l , and therefore F(xsy) ~ F ( x y ) ; by (iii) applied to the mapping x--->xu of G into H and to K = k ( u ) , this implies that F ( x y ) is rational over k ( u , t ) . Therefore the mapping u-->xu of H into H is defined over k ( t ) ; on the other hand, if U is any field of definition for that mapping, containing k, the image x a ~ t of a by it is rational over U, so that k ( t ) C U ; thus k ( t ) is the smallest field of definition for the mapping u ~ xu. This shows that G is not operating faithfully on H ; applying Prop. 2 of AG, no. 3, to G and H, we see that we can define on H a normal law of composition f such that F ( x y ) ~ f ( F ( x ) , F ( y ) ) . By the main theorem of AG, we can then replace H by a birationally equivalent group H', defined over k, with a mapping F' of G into H', also defined over k, such that F ' ( x y ) ~ F ' ( x ) F ' ( y ) , F ( x ) and F'(x) being corresponding generic points of H and H' over ]c when x is generic over k on G. As usual, from the relation F ' ( x ) = F ' ( y ) - l F ' ( y x ) which holds for x, y generic and independent over k, we deduce that F' is everywhere defined 1; therefore, if G is made to operate on [I" by the law (x,w) --->F'(x)w for x, w generic and independent over k on G and H', H ' is a homogeneous space with respect to G. But then the unicity assertion in the main theorem of AG can be applied to H and H ' and shows that they are biregularly equivalent; in other words, H itself, with the law f, is a group. This completes the proof. I t is easily seen that the pair (H,a) is uniquely determined, up to an isonlorphism, by the conditions (i), (it), (iii) in Prop. 2 ; in other words, if H ' and a' have similar properties, there is an everywhere biregular birational correspondence between H and H ' which maps a onto a' and transforms the law of composition between G and H into the law between G and H'. The space H may be called the coset-space determined by G and Z, and may be denoted by G/Z; if ]Z 1 is a normal subgroup of G, the space H, with the group-law 1 This is Theorem 1 of Nakano ( [2] ).
240
[195561 498
ANDRE WEIL.
determined by Prop. 2, is called the quotient-group (or factor-group) of G by Z, and is denoted by G/Z. 3. Before making another application of Prop. 1, we will introduce a new condition which a law of composition may satisfy. Let V, W be two varieties, g a mapping of V X W into W, and ~ a field of definition for V, W and g; consider the following condition: (TGI')
v~g(x,u),
If x, u are independent generic points of V, W over ~, and then k ( x , u ) ~ k ( x , v ) ~ k ( u , v ) .
The condition k ( x , u ) ~ ( x , v ) is equivalent to ( T G 1 ) of AG, no. 2. The condition k ( x , u ) - - k ( u , v ) implies that the dimension of V, which is the dimension of x over k ( u ) , is the same as that of v over ~(u), and therefore at most that of W; if the dimensions of V and of W are the same, this implies that v is generic over ~(u) on W, which is condition ( I t ) of AG, no. 2. Let k' be any field of definition containing ~ for the birational correspondence u--> v ~ g(x, u) between W and itself; if u is taken generic over lc'(x) on W, we have M(u) ~ / d ( v ) ; since ~(x) C ~'(u,v) by ( T G I ' ) , we have k(x) C ~'(u). Taking u' generic on W over k ' ( x , u ) , we get in the same manner k(x) C ~ ' ( u ' ) . As k ' ( u ) , k'(u') are independent regular extensions of k', their intersection is ~', so that k(x) C k'. This shows that ( T G I ' ) implies ( T G 3 ) . In view of the results of AG, end of no. 3, this shows that, if g satisfies ( T G I ' ) and ( T G 2 ' ) , or if two mappings f, g of V X V into V and of V X W into W are given and satisfy ( T G 1 r,2), then V is a pre-group and W a pre-transformation space, and V operates faithfully o n W.
I f a pre-group V and a pre-transformation space W satisfy ( T G Y ) , we say that W is a pre-principal space with respect to V; if at the same time V and W have the same dimension, so that, as we have shown, W is prehomogeneous with respect to V, we also say that V is simply pre-transitive o n W.
Let IV be a pre-principal space with respect to a pre-group V; by the main theorem of AG, we can construct a group G and a transformation-space S, birationally equivalent to V, W and defined over the same field k; then S is also pre-prineipal with respect to G. Let T be the locus of (u, xu) over ~ on S X S, x and u being independent generic points of G, S over k; put t ~ (u, xu); then ( T G I ' ) implies that /r C ~ ( t ) , i.e. that we may write x ~ ~b(t), where ~b is a mapping of T into G, defined over/~; conversely, if this is so for a transformation-space S with respect to G, S is pre-prineipal.
[1955b]
241 ON A L G E B R A I C GROUPS A N D ]-IOMOGENEOUS SPACES.
499
The space S will be called a principal space with respect to G if, for x, u generic and independent over/r on G, S and for t ~ (u, xu), we have x ~ ( t ) where ~b is an everywhere defined mapping, defined over k, of the locus T of t over k into the group G. I f at the same time S is homogeneous, it will be called a principal homogeneous space with respect to G. PROrOSITION 3. Let S be a pre-principal transformation-space with respect to a group G, both being defined over a field k. Then there is a l~-open subset P of S which is a princ@al transformation-space with respect to G; if G and S have the same dimension, P is uniquely determined and is homogeneous. Let T and ~ be defined as above; call F the It-closed subset of T where q~ is not defined. We first show that, if (a, b) is in F, (sa, s'b) is in F for all s, s' in G. I n fact, take x, u generic and independent over k(s,s') on G, S ; p u t v ~ xu, ul ~ su, Vl ~ s'v, x~ ~ s'xs-1; then we have vl ~ x~ul, and xl, ul are generic and independent over k(s,s') on G, S, so that (u,v) and ( u , Vl) are generic points of T over k(s, s'), and that x ~ ~(u, v), xl ~ ~(u~, v~) by the definition of ~ ; this gives
~(u,v) --s'-l~(su, s'v)s. I f (a,b) is in T, it is a specialization of (u,v) over lc(s,s'), and therefore (sa, s'b) is also in T ; then the above relation shows that r is defined at (a, b) if it is defined at (sa, s'b), i.e. that (a, b ) c F implies (sa, s'b) 9 F. As (e, u) is a specialization of (x, u) over k, e being the neutral element of G, T contains the diagonal A of S X S. As the projection of A on either factor of S X S is everywhere biregular, the projection of the /~-closed subset F A A of A onto S is a /~-closed subset F ' of S, consisting of the points a 9S such that ~ is not defined at (a,a). From what we have proved above, it follows that, if a~ F', sa ~F" for all s ~ G. For the same reason, if a is in S - - F ' , then ~b(a, sa) is defined for all s e G; as (a, sa, s) is then a specialization of (u, xu, x) over k, and x ~ ( u , xu), this shows that ~(a, s a ) - - s for all a e S - - F ' and all s e G, and therefore ~ ( a ~ s ) ~ l~(a, sa); in particular, if x is generic over/c(a) on G, the locus of xa over k ( a ) has a dimension equal to that of G. I f G is complete, every specialization (a, b) of (u, xu) over k can be extended to a specialization (a,b,s) of (u, xu, x) over k, so that b ~ s a ; in other words, every point of T must be of the form (a, sa), with a~S and s c G; then it follows from what we have proved above that such a point cannot be in F unless a and sa are in F p. Without attempting to decide whether this is still so in the general case, we shall merely show that, if u
242
[1955b] 500
ANDRE WEIL.
is generic over k on S and (u,u') is in F, then u' must be in F'. In fact, suppose that this is not so; take x generic over Ir on G; call X, X' the loci of xu, xu' over ~(u,u') on S; by what we have proved above, they have the same dimension, which is that of G. By F-VI3, Th. 11, we have T N (u X S) ~ u X X ; at the same time, since we have shown that (u, xu') is in T, u X X ' is contained in T N (u X S ) ; as X and X ' have the same dimension, this implies that X ~ X : . But, as we have shown, since (u, u') is in F, (u, xu') must be in F, and therefore, since F is k-closed, u X X' must be contained in F ; as X ~ X ' , this implies that F contains (u, xu), which is generic on T over ~, and contradicts the definition of F. One may observe that, if S is pre-homogeneous, this again Shows that (a, b) cannot be in F unless a, b are in F ' ; for, if a~F" and x is generic on G over tc(a,b), xa has then over k(a) a dimension equal to that of G, and therefore is generic on S over/c(a) since in the present case the dimensions of S and G are equal; then if (a, b) is in F, so is (xa, b), and so b must be in F'. Now replace first S by S - - F ' ; as F ' is mapped onto itself by all operations of G, S - - F ' is again a transformation-space with respect to G, defined over k, and satisfies our other assumptions. Writing again S instead of S - - F ' , we see that it is enough to prove our result under the additional assumption that / 7 ' ~ . If G is complete or S is prchomogeneous, this already implies that S is principal. Otherwise we observe that, since T is also the locus of (z-lu, u) over /r and since we have ~(x-lu, u) --x, T and F are mapped onto themselves by the mapping (u,v)---. (v,u) of S X S onto itself. Call now F " the "projection" of F on either factor of S X S (in the sense of F-IV3 and F-VIIi, i.e. the closure of the set-theoretic projection); this will be the same, whether we project F onto the first or the second factor, and it is not S by what we have proved above, since F ' is empty; it is therefore a ~-closed subset of S. By what we have proved, F " is mapped onto itself by all operations of G. Then S - - F " is the principal space whose existence was to be proved. Finally, assume that the space S from which we first started was prehomogeneous; this means that T ~ S }(S. Let a, b be any two points in S - - F ' ; then, if x is generic on G over to(a, b), xa, xb are generic on S over k(a, b), and so there is an isomorphism r of to(a, b, xa) onto k(a, b, xb) over ]c(a,b), mapping xa onto xb; then we have x~a--xb, i.e. b=x-~x~a. This shows that S - - F " is homogeneous, and also that an open subset of S which is a transformation-space for G cannot contain a point of S - - F " without containing S - - F ' . Therefore S - - F ' is the only open subset of S which is a principal space with respect to G.
[1955b]
243 ON A L G E B R A I C GROUPS A N D H O M O G E N E O U S SPACES.
501
I f N is a principal homogeneous space, the mapping $ of ~ r , S X S into G which has been defined above will be called the canonical mapping of S X S into G. For any a, b on S and s on G, the relations b=sa, s=ck(a, b) are equivalent; in particular, for any a on S and for x generic over /c(a) on G, the mapping x--->xa~v of G into S has the inverse v--->x= $(a,v); as both are everywhere defined, this is therefore an everywhere biregular mapping of G onto S, defined over /c(a). I n particular, if there is at least one rational point a over k on S, S is biregularly equivalent to G over k. 4. Let G be a group, V and W two varieties, and F a mapping of V ; K W into G, all defined over a field k. We may consider W ? < G as a transformation-space with respect to G, the law of composition between them being ( x , ( N , y ) ) ~ ( N , xy) for any N ill W and any x, y in G. We now apply Prop. 1 of no. 1 to the ease when we take for S this transformationspace W X G and for Z the graph of the mapping N-.-->F(M,N) of W into G, where M, N are independent generic points of V, W over /c. We must then consider the smallest field of definition /e' containing ~ for the mapping N-->xF(M,N) of W into G, where x is generic over lc(M,N) on G. As this mapping is defined over /c(x,M), U is a regular extension of /c, contained in/c(x, M). Then Prop. 1 shows that we may write / c ' = It(u), where u is a generic point over /c of a transformation-space U with respect to G; as /~(u) c / c ( x , M ) , we may write u = f ( x , M ) , where f is a mapping of G X V into U, defined over /~; moreover, as the mapping x.--->f(x,M) of G into U is no other than the mapping F defined in Prop. 1, we see that f is defined at every point (s, M) of G ;4 M, and that f(ss', M) = sf(s', 3f) ; taking s ' = e , and writing f(M) instead of f(e,M), this gives f ( s , M ) = s f ( M ) , and in particular u ~ xf(M). As the mapping N--~ xF(3i, N) is defined over lc(u), xF(M,N) is rational over / c ( u , N ) ; similarly, if y is generic over lc(x,M,N) on G, the mapping N..->yxF(M,N) is defined over l~(yu), and so yxF(M,N) is rational over lc(yu, N). As we can write
y = (yxF(M, N) ) (xF(M, N) )-1, this shows that y is rational over k(u, yu, N). If N' is generic on W over lc(x,y,M,N), y must then also be rational over /c(u, yu, N'); thus /c(y) is contained in lc(u, yu, N) and in lc(u, yu, N'); as these are independent regular extensions of tc(u, yu), their intersection is ]c(u, yu), and so we have ]c(y) C lc(u, yu). This means that U is a pre-principal space and may therefore, by Prop. 3, be replaced by a principal space, birationally equivalent to it.
244
[1955b] 502
ANDRI~ WELL.
The most interesting case is that in which there are two mappings F1, F2 of V, W into G, defined over some overfield K of k, such t h a t F ( M , N ) ~ F I ( M ) F 2 ( N ) for M, N generic and independent over K on V, W ; by the corollary of Th. 7, VA-18, this is always so whenever G is an abelian variety. Take x generic over K(M, N) on G, and put z ~ xF~(M) ; then the m a p p i n g N - - > x F ( M , N . ) ~ z F 2 ( N ) is defined over K(z), so that k(u) C K(z). As x, M are generic and independent over K ( N ) on' G, V, u is then generic over K ( N ) on U, and the dimension of U is t h a t of u over K ; the relation K(u) C K(z) shows t h a t this is at most the dimension of G. Therefore U is pre-homogeneous and may be taken to be a principal homogeneous space with respect to G. Moreover, we may write u ~ O ( z ) , where is a m a p p i n g of G into U, defined over K. I f we substitute yx for x, with y generic over K ( M , N , x ) on G, z is replaced by yz, and u by yu; this gives r ~ y ~ ( z ) , which may be written as r and thus shows t h a t ~ is everywhere defined. P u t t i n g now a ~ , ( e ) , we see t h a t a is rational over K and that u ~ z a , i.e. f(M) ~ F I ( M ) a . P u t now g(N) ~ F 2 ( N ) - l a ; g is a m a p p i n g of W into U, defined over K. As we have also g(N) ~ F ( M , N ) - I f ( M ) , g is also defined over the field k(M), and therefore also over k(M') if M ' is another generic point of V over K ; if we take M, M ' generic and independent over K on V, k(M) and k(M') are independent r e g u l a r extensions of k, so t h a t their intersection is k ; hence g is defined over k. Thus we have proved the following r e s u l t : PROPOSlTIO~ 4. Let G be a group, V and W two varieties and F a mapping of V X W into G, all defined over k. Assume that there are two mappings F~, F2 of V, W into G, defined over some overfield K of k, such that F ( M , N ) ~ F ~ ( M ) F 2 ( N ) for M, N generic and independent over K on V, W. Then there is a princ@al homogeneous space U with respect to G, and two mappings f, g of V, W into U, all defined over ~, such that f ( M ) = F ( M , N ) g ( N ) , i.e. F ( M , N ) ~ r where ~ is the canonical mapping of U X U into G. COROLLARY. Notations being as in Prop. 4, U, f and g are uniquely determined by G, V, W and F, up to an isomorphism. I n fact, assume t h a t Up, f', g" have similar properties ; then we have xf'(M)), where r is the canonical m a p p i n g for U'. This shows that the m a p p i n g N---->xF(M,N) is defined over k ( u t) with u ' ~ x f ' ( M ) ; thus, if we put u ~ x f ( M ) as before, we have k ( u ) C k ( u ' ) and m a y write u ~ r where r is a m a p p i n g of U" into U, defined over ~.
xF(M,N) ~ r
[195561
245 ON ALGEBRAIC GROUPS A]qD H O M O G E N E O U S SPACES.
503
Replacing x by yx, with y generic on G over k (M, N, x), we get yu ~ ~ (yu p) ; from this we conclude, in the usual manner, t h a t ~ is everywhere defined. Take any point a' on U', and put a ~ ( a ' ) ; as we have x a ~ r and as the mappings x--->xa, x - o x a ' are everywhere biregular mappings of G onto U and U', defined over k(a'), we see t h a t ~ is an everywhere biregular m a p p i n g of U' onto U. Moreover, we have ~ ( u ' ) ~ xtp(f'(M)), and therefore f ~ ~ o f'; from this one easily concludes t h a t g ~ ~ o g'. This proves our assertion. 5. Let G be a group, defined over a field k. We will now prove t h a t the classes of principal homogeneous spaces with respect to G, for birational equivalence over k, form a set. I n fact, let x, y be independent generic points of G over k; let (~ be the isomorphism of ~ ( x ) onto ~(yx) over ~ which maps x onto yx. Let H be any principal homogeneous space with respect to G, defined over k ; let a be an algebraic point over k on H , and p u t u ~ xa, so that u is generic over 7~ on H. Then ]c(u) is a regular extension of 7c contained in ~ ( x ) and such t h a t ~ ( u ) = ~ ( x ) ; moreover, we have
~(y,u) =k(y,u~)
= ~ ( u , u~)
since u ~ y u . Conversely, let ~ ( u ) be any such extension of k, and call U the locus of u over k; then we may write u r g (y, u), where g is a m a p p i n g of G X U into U, defined over ~; and one verifies at once t h a t this makes U into a pre-principal pre-homogeneous space with respect to G, and thus determines uniquely a class of birationally equivalent principal homogeneous spaces with respect to G. As every such class is determined by at least one such extension, this shows that these classes form a set. I f G is commutative, one can define canonically a commutative groupstructure on the set of classes of principal homogeneous spaces with respect to G. I n order to do this, we first observe that, if H is any transformationspace over a commutative group G, then the law (x,u)-->x-lu, for x~G, u e H , defines on H a structure of transformation-space with respect to G; this will be called the opposite transformation-space to H and will be denoted by H - ; it is a principal homogeneous space if H is such. PROPOSITION 5. Let G be a commutative group, defined over a field k. Let H~, for 1 ~-- i X n, be principal homogeneous spaces with respect to G, defined over k. Then there is a principal homogeneous space H with respect to G, defined over tr and an everywhere defined mapping f of HI }( H2 X " 99X H,~ into H, defined over ~, such that f(s~a,.
9 .,sna~)
=
s~"
. . s,,f(a,.
9
a~)
246
[1955b] 504
ANDRe] ~,VEIL.
for all s~e G and a~eH~. Moreover, H and f are uniquely determined up to an isomorphism of H. Put V ~ W ~ HI X H2 X " " " X 1t,,; call ~ the canonical mapping of H, X H~ into G, so that b~~ sa, is equivalent to s ~ q~,(04, b~) for a~, b~ in H~ and s in G. Let u = ( u ~ , . - .,u~) and v ~ ( v ~ , . 9 .,v,,) be two points of V; put
F(u, v) = fI +,(u,, ~,), i=I
where the right-hand side has a meaning since G is commutative. H~, choose a point 04, and put a ~ ( a , . 9 .,a~). We have
r
vd ~ r
vdr
On each
ud -~
for all i, as one verifies at once, and therefore, again because of the commutativity of G :
F ( u , v ) = F ( a , v ) F ( a , u ) -1. Thus the assumptions of Prop. 4 are satisfied, so that there is a principal homogeneous space U and two mappings f, g of V into U, all defined over /c, such that (1)
f(u) =F(u,v)g(v),
F ( u , v ) = , ~ ( g ( v ) , f ( u ) ),
where ~ is the canonical mapping of U )< U into G. Take any point b on V, and take v generic over k(b) on V; as F is defined at (b,v), the relation (1) shows that f is defined at b. Thus f is everywhere defined. As F ( u , u ) ~ e, the relation (1) gives g ( u ) = f ( u ) , i.e. f ~ g . If s l , ' ' ' , s ~ are any elements of G, and we put s ~ s l " 9 .s.~ and u ' ~ (slu~,. 9 ",s,u,,), we have F ( u ' , v ) ~ s - I F ( u , v ) and therefore, by (1), f ( u ' ) ~ s - l f ( u ) . If we now put H ~ U-, i.e. if we take for H the opposite space to U, H and f will have the properties stated in Prop. 5. Let us now assume that ~ and ~ have similar properties; put F ~ / ~ - . Put 2 = F ( u , v ) - ~ ] ' ( u ) , the multiplication in the right-hand side being that of ~7. I f the s~, s and u" have the same meaning as above, we have ]'(u')=s-~]'(u), so that ~ does not change if one replaces u,v by u',v. Therefore /c(2) is contained both in k ( u , v ) and in k(u',v). If the s~ have been taken generic and independent over lc(u,v) on G, t~(u,v) and lc(u',v) will be independent regular extensions of k ( v ) ; this gives k ( 5 ) C k(v), so that we may write 2 = t T ( v ) , with j defined over k. Then we have ]'(u) = F ( u , v ) g ( v ) ; by the corollary of Prop. 4, ~7, ~ and j must then be
[1955b]
247 ON
ALGEBRAIC GROUPS AND I~O~IOGENEOUS SPACES.
505
the same as U, f and f, respectively, except for an isomorphism of U onto U. This proves the assertion about unicity in Prop. 5. In Prop. 5, take n ~ 2; call ~1, ~ 2 the classes of H1, H2, and denote by t~l ~ ~ 2 the class of H. This defines on the set of classes of principal homogeneous spaces with respect to G a commutative group-structure. In fact, commutativity is obvious. Call ~ o the class of G, and therefore of all principal homogeneous spaces with respect to G which have a rational point over k. For any principal homogeneous space H with respect to G, the mapping f ( x , u ) ~ x u of G X H into H satisfies the condition of Prop. 5; there-fore we have ~ o ~-t~ ~ 9/ for all classes ~ . If ~ is the canonical mapping of H X H into G, then ~, considered as a mapping of H X H- into G, satisfies the condition of Prop. 5 ; therefore, if ~ - is the class of H-, we have 5~ -~ &t- ~ 9/0. Finally, let H1, H2, H~ be three principal homogeneous spaces with respect to G ; apply Prop. 5 successively to the following spaces: (a) to H~, H2, obtaining a space H12 and a mapping f12; (b) to H12, H3, obtaining H', f'; (c) to H2, H3, obtaining H2~, f2a; (d) to H~, H~3, obtaining H", f"; (e) to H1, H2, H3, obtaining H, f. Then the two mappings
f'(f~,~(u,u~),u,),
f"(u,f~(u~,m))
of H1 }( H2 X Ha into H', H'" satisfy the same condition as the mapping f. By the unicity assertion of Prop. 5, this shows that H', H" are isomorphic to H. This means that the addition &t~ + ~2 is associative. One proves quite similarly, by induction on n, that if :~ and ~ , are the classes of the spaces H, H, in Prop. 5, then ~ ~ ~ . ~ . In fact, let H', f' be the space and the mapping obtained by applying Prop. 5 to H~,. 9 -, H,~, so that ~ ' ~ & q + . . . + 5 ~ _ ~ by the induction assumption; and let H", f" be the space and the mapping obtained by applying Prop. 5 to H', H,, so that ~ " ~ ~ ' - ~ ~ by definition. Then the mapping
(ui,
.,u.) ~f"(f'(~.
,~_~),u,,)
of H~ X" 9 ' } ( H ~ into H " has the properties stated for f in Prop. 5~ so that, by the unicity assertion in Prop. 5, H" is isomorphic to H. From this one deduces that every element ~ of the group we have just described is of finite order. In fact, take on a space H of class ~t any positive cycle ~ a~ of dimension 0, rational over k.
Call H , any space of
class nt~; then there is a mapping f(u~," . ",u,) of the product of n factor~ equal to H into H , with the properties stated in Prop. 5. From the unicity assertion in Prop. 5, it follows that any permutation of the u, will change f
248
[1955b] 506
ANDRI~ W E I L .
into sf, with s~ G; as f is everywhere defined, we see that s ~ e by t a k i n g ul . . . . . u~; therefore f is a symmetric function, so t h a t f ( a l , - - -,a~) is rationM by the m a i n theorem on symmetric functions (VA-7, Th. 1). So H~ has a rational point over k, and is therefore isomorphic to G. Now, ~ being as before, put Ho ~ G and take, for each integer n ~ 0, a space H~ of class n ~ so that all the H . are disjoint. On the set | ~ U H~ yt
(which is of course not an algebraic v a r i e t y ) , we will define a commutative group-law f (in the sense of group-theory, not of algebraic geometry) such that Ho ~ G will be a subgroup of @ and that f induces on H ~ X H~, for all m, n, a m a p p i n g fm.~ of H,~ X H~ into H ~ satisfying the conditions in Prop. 5. As there is such a m a p p i n g f~.~ for each m, n, and as it is uniquely determined up to an automorphism of H ~ (i. e. up to left-multiplication by a rational point of G), we merely have to choose the f,~.~ so that the m a p p i n g f of (~ X ~ into @ which coincides with f~,~ on Hm X H , for all m, n satisfies the axioms for groups ; we do this as follows. F o r any n, we take fo,~(x, u) ~ xu for x c G , u~H~. We choose f-l.~ and, for all n ~ 0, f~,~ and f_~_~ so as to satisfy the conditions in Prop. 5. Now, for elements Ul," " ",u~+l of H1 in any number, we define ul" 9 .u~+~ inductively as being equal to ul for n ~ 0 and to f~,~(u~" 9 - u~,u.+~) for n ~ 1, similarly, for elements v~," 9 ',v,+~ of H_~, we define v~. 9 9v~+l as equal to vl for n ~ 0 and to f-~,-l(vl" " "v., v~+~) for n ~ 1. I t is then easily seen that, whenever m, n are both ~ 0, there is one and only one way of choosing fro,, so t h a t it satisfies the condition f ~ , , ( u ~ ' 9 -u~, u~+l" - 9u~+,) ~ ul" - 9u ~ when the u~ are in I-I1, we determine f_~_~ similarly, using H_~ instead of H , Finally, for m => n ~ 0, we choose f,~_~ and f_~,, so as to satisfy the conditions n
f~,~(ul"
9 .u~, vl"
9 .v,)
~
I I f - l . ~ ( v , , ~ , ) 9 u~+~. 9 . u ~
I-~,-(~,"
9 '~m,~"
" .u.)
=fIf-~,l(~,,~)
'v.+l"
9 .~
4=1
respectively, the u~ being any elements of H1 and the v~ any elements of H _ , I t is then a trivial m a t t e r to verify t h a t these choices of the f~., satisfy all the requirements for a commutative group-law on | The points on the H~ which are rational over k form a subgroup g of | As we have shown t h a t there are such points for some n ~ 0 , there is a smallest n ~ 0 for which there is such a point a ~ H ~ , this n is the order of ~ in the group of classes of principal homogeneous spaces with respect to G. Then ~ is the direct product of the group g o n g A G of rational
[1955b]
249 507
ON ALGEBRAIC GROUPS AND ~IO~MCOGENEOUS SPACES.
points over k on G and of the infinite cyclic group ), generated by a. The quotient-group | may be described as an algebraic group consisting of n components respectively isomorphic to Ho ~ G, H ~ , - - - , H , _ , 6. PROPOSITION 6. Let A be an abelian variety and H a principal homogeneous space with respect to A, both being defined over a field k. Let V1," 9 ", V~ be varieties, and F a mapping of VI X 9 " " X V~ into H, all these being defined over k. Then there is for each i a principal homogeneous space H~ with respect to A and a mapping F~ of V~ into H~, Ht and F~ being defined over k, and there is a mapping f of H1 X " 9 " X H , into H with the properties stated in Prop. 5, such that, for (M1," 9 ",M~) generic over k on V 1 X " 9 " X Vn, we have F ( M I , . 9 .,Mn) ~ f ( F l ( M ~ ) , "
9 ",F,(M~) ).
Moreover, all these are uniquely determined up to isomorph@ms. For n ~ 1, there is nothing to prove. I f the assertion is proved for a product of two factors, then this can be applied to the product V I X ( V 2 X " ' " X V ~ ) of V1 and V2X" 9 " X V ~ , so that the general case follows by induction on n. Thus it is enough to treat the case of two factors V, W and of a mapping F of V X W into H. Call ~ the canonical mapping of H X H into A; let ( M , N ) and ( M ' , N ' ) be two independent generic points of V X W over k; and put x~(F(M,N'),F(M',N')
),
y~(F(M,N),F(M,N')
).
so that we have xy = ~,(F(M, N ) , F ( M ' , N') ). As the mapping ( (M, M'), N p) --> x of (V X V) X W into A has the constant value e on the variety (M, M) X W, Th. 7 of VA-18 shows that x is rational over k ( M , M ' ) ; for a similar reason, y must be rational over k ( N , N ' ) ; in other words, there are mappings (I), $ of V X V, W X W into A, both defined over k, such that x ~ r and y ~ r By the corollary of Th. 7' of VA-18, 9 and q, satisfy the assumptions of Prop. 4, so that there are two. principal homogeneous spaces U~, U.~ with respect to A, mappings F~, G~ of V into U~ and mappings F2, G2 of W into U2, all defined over k, such that F~(M) ~ x G I ( M ' ) , F 2 ( N ) ~ y G 2 ( N P ) ; moreover, as r ~ ( N , N ) are defined and equal to e, we have G1 ~ F 1 , G~ ~ F 2 . Now call H~, H2 the spaces respectively opposite to U1, U2, and apply Prop. 5 to H~, H2: let / t be the principal homogeneous space and ~ the mapping of H~ X H~ into /~ with
250
[1955b] 508
ANDRI~ WEIL.
the properties stated in that proposition. As H1, H2 are opposite to U~, U2, w e have
F~(M') ~ x F I ( M ) ,
Put P(M,N)=](FI(M),F2(N)).
F2(N') ~ y F 2 ( N ) ,
multiplication in the right-hand sides being understood in the sense of H1, H2. By the definition of ~, we have then:
P(M',N') = (xy)#(M,N), while, as we have seen above, the same relation holds if F is substituted for iV. But then, as the corollary of Th. 7, VA-18, shows that the mapping
( (M',N'), (M,N) ) ---->xy of (V X W) X (V X W) into A satifies the condition of Prop. 4, the corollary of Prop. 4 shows that H, F must be the same as H, F except for an isomorphism of H onto H. Then, replacing ]~ by a mapping f of:H1 X H2 into H by means of that isomorphism, we have the spaces HI, H2 and the mappings F1, F2, f whose existence was asserted in our proposition. As to unicity, assume that there are spaces Hi*, H~* and mappings F~*, F2*, f* with the same properties. Then, x being defined as above, or equivalently by F(M', N') ~ xF(M, N'), we have
f*(F~*(M'), F2*(N')) ~ xf*(F~*(M), F~*(N')) = f*(xF~*(M), F2*(N')) and therefore F~*(M')= xFI*(M) since the mapping u---->f*(u,v) of H~* into H is, as easily seen. an everywhere biregular mapping of H~* onto H. But then the corollary of Prop. 4, applied to the mapping (M',M)-->x of V X V into A, shows that H~*, FI* are the same as H~, F~ except for an isomorphism. The same argument applied to y instead of x shows that H2, F~ are uniquely determined up to an isomorphism. Then Prop. 5 shows that f is uniquely determined. This completes the proof. 7. The foregoing results will now be applied to the theory of Jacobian varieties. As in VA-35, we consider a complete non-singular curve F of genus g > 0, defined over a field ~. I f a is any divisor on F, Prop. 6 of the Appendix of AG shows that there is a smallest field containing k over which a is rational; this field will be denoted by/r (a). In particular, if M1," 9", M~ are independent generic points of F over k and if we put m ~ ~ M~, then, i
by VA-4, Lemma 1, to(m) is the field k(M1,- 9 .,Mg)~ of symmetric functions of M~,.. ",M~, defined over k, i.e. the subfield of to(M1,. 9 .,Mg) consisting of those elements which are invariant under all permutations of
[1955b1
251 ON
ALGEBRAIC
GROUPS
AND
ts
SPACES.
509
M1," 9 ",Mg; such a divisor m will be called generic over k. As lz(m) is a regular extension of k, we may write it as k(u), where u is a generic point of a variety W over k, and we may write u ~ F ( M 1 , ' 9 .,Mg), with F defined over k ; as F is symmetric in the M~, this may also be written as u ~ F ( m ) . Now let the N~, P~, for 1 ~ i <_ g, be 2g independent generic points of Y over k ( m ) ; and p u t : x = ( N , . - . , K s , P1,. 9 ",Pg), this being a generic point over k of the product V ~ F X ' " X r of 2g factors equal to F. By VA-35, Lemma 11, there is a positive divisor m' g
g
on r linearly equivMent to m + ~. N ~ - - ~ P~, and it is uniquely determined i=1
~=I
and such that k (x, m ) ~ ~(x, m ' ) ; this implies that it is generic over k (x). Then, if we write u ' = F ( m ' ) , we have k ( x , u ) ~ I c ( x , u ' ) ; we may thus write u ' = g ( x , u ) , where g is a mapping of V X W into W which satisfies (TG 1). We now show that this mapping satisfies the condition ( T G 2 ' ) of AG, no. 3, Prop. 2, so that this proposition may be applied to it. I n fact, let y be a generic point of V over k(x,u); we may write
y = (Q_I," " ", Qg, RI," 9 ",Rg). Then the point u " ~ g ( y , u ' ) will be determined by u " ~ F ( m ' ) , where m" is the positive divisor linearly equivalent to m ' - f - ~ . Q ~ - - ~ . R~. Applying again VA-35, Lemma 11, we see that there is a positive divisor ~ S~ linearly equivalent to ~..N~--~P~+Y.Q~, and that it is generic over k(y,u) and rational over k(x,y).
But then In" is linearly equivalent to m ~ - ~ S ~ - - ~ , R ~ ,
which shows that, if we put z~
($1,. 9 .,Sg, R , .
9 .,Rg)
z is generic on V over k ( u ) and that we have u " ~ g ( z , u ) . This shows that g satisfies (TG 2'). Applying Prop. 2 of AG, no. 3 and the main theorem of AG, we see that there is a group J, a normal law g between J and W, and a mapping r of V into J such that g ( x , u ) ~ j ( ~ ( x ) , u ) for x, u generic and independent over k on V, W. P u t now K ~ k ' ( P ~ , . . . , P ~ ) , n ~ N ~ , and i
w =
(s.-
9 -, s ~ , 0 . "
9 ", Q ~ ) .
Since the P~, O~, N~ are generic and independent over k, Lemma 11 of VA-35
252
[1955b] 510
ANDR~ WELL.
shows that w is generic over K on V. At the same time, the linear equivalence by which the S~ were defined shows at once that g ( x , u ) ~ g ( w , u ) for u generic over k(x,w), and therefore g(c~(x),u)~g(~(w),u). Since J, by definition, operates faithfully on W, this implies r 1 6 2 as .w is generic over K on V, this shows that ~ ( x ) is generic over K on J. I t will now be shown that K(~(x)) ~ K 0 r ). I n fact, if u and u ' ~ g ( x , u ) are as before, u" is rational over K ( m , n ) ~ K ( u , n ) since the divisor m' is so by Lemma 11 of VA-35; therefore the mapping u-->u' is defined over K ( n ) , so that K(~(x)) C K ( n ) . P u t now K ' ~ K ( ~ ( x ) ) , so that u' is rational over K'(u) ; then m and m' are both rational over Kt(u). But Lemma 11 of VA-35 shows that n is rational over K ( m , m') and therefore over K'(u). I f ul is a generic point of W over K'(u), n will also be rational over K'(ul) ; as K'(u), K'(ux) are independent regular extensions of K', this implies that n is rational over K'. But now a comparison with the construction of the jacobian variety given in VA-36 shows that the latter coincides with our J over a suitably extended groundfield; more precisely, substituting K for k, ~ P~ for a, n for m and ~(x) for z in the treatment given in VA-36, we get the same law of composition for the field K(~(x)) as has been defined above. Alternatively, one may also reason as follows. Let Jx be the jacobian variety as defined in V A ; let ~bl be the "canonical m a p p i n g " of F into J1, also according to the definition of VA-37 (which will soon be replaced by a more appropriate one) ; let K1 be an overfield of the field K defined above, over which J~ and r are defined; take n generic over K1; put t ~ ( x ) , x being as above, and z ~S[~l(n)]. As we have then K~(x) ~K~(z) ~ K l ( n ) , the mapping x---->z defines a birational correspondence between J and J~, defined over KI. I f we write it as z ~ f ( x ) , f is everywhere de:fined by VA-15, Th. 6; and this, by the results at the beginning of VA-19, must then be of the form f(x) ~fo(X) + a, where a ~ f ( e ) and where fo is a homomorphism, so that (using the additive notation on J~ and the multiplicative notation on J ) we have f o ( x x ' ) ~ f o ( x ) + fo(x'). But then fo is again a birational correspondence, and, if g is the inverse mapping to fo, we have g(z + z') ~g(z)g(z') for z, z' generic and independent over K~ on J , This can be written as g (z) ~ g(z + z')g(z')-~; if then z~ is any point of Jx, and we take z" generic on J t over K~(zt), this shows that g is defined at zl. As fo, g are everywhere defined, they determine an isomorphism between J and J~. One could also, without making use of the results of VA, verify directly (for instance by making use of the criterion for the completeness of a group
[1955b ]
253 ON A L G E B R A I C GROUPS A N D HO2YIOGENEOUS SPACES.
511
given by VA-33, Th. 16) that the group J we have constructed here is complete and is therefore an abelian variety. Then the results we have proved above, combined with the corollary of Th. 7, VA-18, show at once that J has the properties stated in VA-36, Th. 18; since the whole theory of the jacobian variety depends upon nothing else, and these properties (as proved in VA-37) are characteristic of the jacobian variety, this would suffice for a complete treatment. From this discussion, we conclude that J is an abelian variety. Now apply Prop. 6 to the mapping ~ of V into J ; this defines 2g mappings of r into principal homogeneous spaces with respect to J, all defined over ~. As ~b is symmetric in the N~ and also in the P~, the unicity assertion in Prop. 6 shows at once that the first g mappings must coincide, and that the last g mappings must coincide; call F, F ' these mappings, and H, H ' the spaces into which they map r. Now, notations being the same as above in no. 6, put z' ~
(PI,"
" ", P g , N1," " ', N g ) .
Then we have, always with the same notations as before, u ~ g ( x ' , u ' ) , and therefore ~ b ( x ' ) ~ r -1. This, combined with the unicity assertion in Prop. 6, shows at once that H ' is the opposite space to H while F ' must be the same as F. We now embed J and H, in the manner explained at the end of no. 5, into a commutative group | consisting of principal homogeneous spaces H , with respect to J, all defined over ~, with Ho ~ J, H1 ~ H, in such a way that | is an infinite cyclic group, that the H . are the cosets of J in (~ and that the group-law in (~ induces on Hm X H., for all m, n, a mapping of Hm X H , into Hm+, defined over k and satisfying the conditions in Prop. 5. At the same time, we change from the multiplicative to the additive notation, not only in J but also in @. With this notation, we have, if x, 4~(x) and F have the same meaning as before,
4)(x) ~ ~ F(N~) - - ~ F(P,). Let us now extend the mapping F into a homomorphism of the group of divisors on F into (~, by putting F(a) ~ ~, n.~F(A~) for any divisor a ~ ~ n~A~, so that F ( a ) e l l , if n is the degree of a; in particular, F ( a ) is in J if and only if a is of degree 0. I f a is any point of H and M is a generic point of F over /c(a), the mapping M . - - > F ( M ) - - a of r into J, which is defined over ~(a), is a "canonical m a p p i n g " in the sense of VA-37; naturally it is only defined up to an additive constant; and, by the unicity assertion in Prop. 6,
254
[1955b1
512
ANDR~ WEIL.
no such m a p p i n g can be defined over ~ unless H is isomorphic to G, i. e. unless H has a rational point over It. F r o m Th. 19 of VA-38, one deduces immediately that a divisor a on r is linearly equivalent to 0 if and only if
F(a)
=o.
In other words, the homomorphism a--) F(a) determines an isomorphism of the g r o u p of all divisor-classes (of any degree) on I' onto the g r o u p (~.
From the foregoing results, one concludes easily that these properties are characteristic for (~ and F, up to isomorphisms on J and its cosets H~ in (~. One may call (~ the J a c o b i a n g r o u p of F, and F the canonical m a p p i n g of F, and of the group of divisors on F, into the Jacobian group. I n substance, the construction of the varieties H~ has already been given by Chow (in [ 1 ] ) by a method belonging to projective geometry. TI{E UNIVIgIr
OF CHICAGO.
BIBLIOGRAPHY.
[1] W. L. Chow, " The Jacobian variety of an algebraic curve," American Journal of Mathematics, vol. 76 (1954), pp. 453-476. [2] S. Nakano, " Note on group varieties," Memoirs of the College of Science, University of Kyoto, Series A, vol. 27 (1952), Math. no. 1, pp. 55-66. [3] A. ~Veil, Vari~t~s abe;liennes et courbes algdbriques, Paris, ttermann et Cie, ]948. [4] - - , " On algebraic groups of transformations," American Jour~*al of Mathematics, vol. 77 (1955), pp. 355-391.
[1955c] On a certain type of characters of the id6le-class group of an algebraic number-field
Notations will be the same as in my previous work on class-field theory (Sur la th~orie du corps de classes, J. Math. Soc. Japan, 3 (1951), pp. 1-35 ; cf. also Sur les " f o r m u l e s explicites" de la th~orie des hombres premiers, Comm. Lund (M. Riesz jubilee volume), 1952, pp. 252-265). If K is any field, K* denotes the multiplicative group of non-zero elements of K. We consider an algebraic number-field k; k, means its completion with respect to a valuation v; in particular, k~, k~(1~ p ~ r~), k~(r, + 1 <~~<_rl + rf) denote the completions of k with respect to the prime ideal ~, to the real archimedian valuation vp and to the imaginary archimedian valuation v~, respectively; k~ may be identified (canonically) with the real number-field R, and k, may be identified (non-canonically) with the complex number-field C; put ~ = [k~: RJ. The id~le group Ik is the subgroup of ]-[ k* consisting of the a = ( a , ) s u c h that almost all a, (i.e., all except a finite number) are units. We denote by Pk the group of principal id~les, and by Ck=Ik/Pk the group of id~le-classes. Each id~le a = ( a , ) determines in an obvious manner an ideal a - ( a ) of k; we put: [] a ]]=N(a)-' ~ [a~ I'z. Then a ~ ] ] a [] is a representation of Ik into R* (in fact, into R*), taking the value 1 on Pk. Group characters will be understood in the extended sense, i.e. as continuous representations into C* (not necessarily of absolute value 1). The groups Ik, C~ will be topologized in the usual manner. A character )~ of Ck may also be regarded as a character of Ik, taking the value 1 on Pk; because of the known structure of Ck, such a character can always be written as )~,(a) ]] a ][~, where a e R and )~1 is a character of absolute value 1. The Hecke L-series attached to a character X of Ck can be constructed as follows. Let f be the conductor of )~; ff a = ( a , ) is an id~le such that a z = l for l ~ 2 ~ r ~ + r ~ and a ~ = l for every prime divisor ~ of f, )/(a) depends only upon the ideal a=(a); under those Reprinted from Proc. Intern. Symp. on Algebraic Number Theory, Tokyo-Nikko, 1955, pp. 1-7.
255
256
[1955cj 2
A. WEIL
circumstances, we p u t ~ ( a ) = x ( a ) . Then the L-series attached to x is ~ , ~ ( a ) N a .3, the sum being extended to all integral ideals r p r i m e to f. We shall denote by G(f) the group of the fractional non-zero ideals in k whose expression in t e r m s of p r i m e ideals does not involve any p r i m e divisor of f. We have t h u s attached, to e v e r y character )~ of Cx with the conductor f, a c h a r a c t e r % of G(f). Clearly ~ is completely d e t e r m i n e d by its values at the p r i m e ideals which do not divide f. A t the same time, X induces on the subgroup ]~k~ of Ix a character X
of t h a t g r o u p ; if we make use of the fact t h a t X m u s t be the product of a character of absolute value 1 and of a character I[ a I]~ we see t h a t )/, on t h a t group, can be w r i t t e n as:
X((a~))=F[(az/] az [)-s~, I a~ [~x<~+~~
(1)
w h e r e the f~, are integers and a and the ~o~ are real n u m b e r s . Now denote by k*(f) the subgroup of k* consisting of all elements a/a', w h e r e ~, ~' are integers in k such t h a t a------d----1 mod. f. Then X((a)) is a character of k*(f), which coincides on k*(f) w i t h the character • of k* given by the formula
=
I
I
(2 )
in which a~ denotes the image of ~ in kz (the l a t t e r being identified w i t h R or with C, as the case m a y be). Conversely, assume that for some integral ideal m of k we have a character ~ of the group G(m), and that t h e r e are integers f~ and real n u m b e r s a, ~o~ such t h a t ~ ( ( a ) ) = X ( a ) for a e k*(m), X being defined by (2). L e t a be an id~le; t h e r e is a S e k * such that, if we p u t b--~a, then, for e v e r y p r i m e divisor p of m, b~ is a u n i t in k~ and is ------1modulo the highest power of p dividing m; and $ is u n i q u e l y d e t e r m i n e d in k* modulo the subgroup k*(m) of k*. Now put: ~(a) =~((b)) 1-[ (b~/l bz I)-~l bz I~x(~+~x~. X
Our assumption on ~ implies t h a t t h e right-hand side does not depend upon the choice of ~ when a is given; and one sees at once that X is a c h a r a c t e r of Ix, taking the value 1 on Px and satisfying (1), t h a t its conductor f divides m, and t h a t the c h a r a c t e r ~ of G (f) associated w i t h X coincides w i t h @ on G(m). I t is clear t h a t a-->(a) defines a homomorphism of k*(m) into G(m) whose kernel is the group E(m) of all units e in k such t h a t ,.~-1 mod. m; E(m) is of finite index in the g r o u p E of all units in k. Notations being as above, we see that • takes the value 1 on
[1955c]
257 On a Certain Type of Characters of the Id~le-Class Group
E(f), so that, if m is the index of E(f) in E, • takes the value 1 on E . Conversely, let the f~, a, cpz be given; let X be defined by (2); and assume that t h e r e is an i n t e g e r m > 0 such t h a t • is 1 on E . Then • is 1 on a subgroup E ' of E of finite index. By a t h e o r e m of Chevalley, this implies t h a t t h e r e is an ideal m such t h a t E ' ~ E ( m ) ; t h e n • is I on E(m) and t h e r e f o r e d e t e r m i n e s a character of the image of k*(m) in G(m), which can "then be extended to a c h a r a c t e r of G(m), hence, for a suitable divisor f of m, to a c h a r a c t e r ~ of G(f) associated with a c h a r a c t e r 1~ of Ck with the conductor f. A character x of C~ is of finite order (in the group of all characters of CD if and only if it is 1 on the connected component of 1 in [k, i.e. if and only if f , = 0 for all ~, ~ = 0 for all ~, and a = 0 ; by class-field theory, such characters are those associated with the cyclic extensions of k; for such a )r all values of ~, are roots of u n i t y . Our purpose is now to show t h a t all the values of ~ m a y be algebraic for certain characters )r which are not of finite order. In fact, assume t h a t all the ~ , are 0 and that ~ is rational; t h e n ~((a)) has algebraic values on k*(f), i.e. ~ has algebraic values on t h e imuge of k*(~) in G(f); as t h a t image is of finite index in G(f), all the values of ~ m u s t be algebraic. The f~ and a being given, a necessary and sufficient condition for the existence of such a c h a r a c t e r % is, as we have seen, t h a t t h e r e should be an i n t e g e r m such t h a t ~ ( ~ / I r I)~z~ :=1 for all ~ e E ; replacing m b y 2 m , ~his can also b y w r i t t e n as
I I (~,/~,Y, = 1.
(3)
We shall say that )r is of type (A) if all the q~ are 0 and a is rational; for such a character, the integers f, will be such t h a t (3) holds, for a suitable m, for all ~ e E . Conversely, if the f , are given integers, and if t h e r e is an m such that (3) holds for all e e E , t h e n t h e r e will be a character )r of t y p e (A) belonging to the f~; and all such characters will be of the f o r m )~(a):~o(a)lta]I ~, w h e r e X0 is a c h a r a c t e r of finite order and p is rational. In partiaular, if k is a totally i m a g i n a r y quadratic extension of a totally real number-field k0, then, by Dirichlet's theorem, the group E0 of the units in k0 is of finite index in E ; if m is that index, e ~ m u s t then be totally real for e v e r y ~ e E , so t h a t (3) holds on E , for t h a t value of m and for a r b i t r a r y values of the f , . More generally (as A r t i n pointed out to me d u r i n g the symposium), Minkowski's t h e o r e m on units in absolutely normal number-fields makes it possible to reduce the problem of finding all characters of t y p e (A) of a field k to an exercise in Galois theory, and it will be enough
258
[1955c] 4
A. WEIL
here to state the result. Let /Co be the maximal totally real subfield of k; then k contains at most one totally imaginary quadratic extension of /Co; for two such extensions could be written as k o ( l / - a ) , ko(V'-/~), with a, fl totally positive in ko; then # contains the totally real field ko (V'a~), which must be the same as ko, so that the two extensions must be the same. Now let us call trivial those characters of type (A) which are of the form Xo(a)I[ all ~, with Xo of finite order and p rational. In order for a field k to have non-trivial characters of type (A), it is necessary and sdfficient that it should contain a totally imaginary quadratic extension kl of its maximal totally real subfield k0; and then all such characters are of the form
x(a) = x~(Nk/k,(a))Xo(a)
(4)
where Xo is of finite order, X~ is a character of type (A) of #~, and Nk/,, denotes the relative norm from I, to I,,, which extends the relative norm of elements of Iv over #~ in the obvious manner. Thus, in a certain sense, all non-trivial characters of type (A) come from totally imaginary quadratic extensions of totally real fields. We shall say that a character X is of type (A0) if the character • of k* associated with it according to (2) is of the form where the rz, sx are integers, and the sign may depend on a; such a character is called trivial if it is of the form Xo(a) Ilall ~, with Xo of finite order and m an integer. Non-trivial characters of type (Ao) are those non-trivial characters of type (A) for which 2o is an integer andf,~--2a mod. 2 for all b. It is easily seen that the character X of C, defined by (4) is of type (Ao) if and only if the character )/1 of C,, which appears in (4) is of type (Ao). I f x is of type (Ao), the values taken by ~ on the image of k*(f) in G(f), which are the values taken by • on k*(f), are all contained in the compositum of k and of its conjugates over Q (the rational number-field). As that image is of finite index in G(f), the values of on G(f) must all lie in a finite extension of this field. Thus: I f a character X of the id~le-class group C, of the field k is of type (A), the coe~cients of the Hec#e L-series associated with X are algebraic numbers; i f X is of type (Ao), these coefficients all lie in a finite algebraic extension K of Q. It is tempting to conjecture that the converse statements are also true; but I have not examined this question. In the second statement, it would be of interest to determine the smallest field K containing all the coefficients of the L-series, i.e. containing all the
[1955c]
259 On a Certain Type of Characters of the Id~te-Class Group
values taken by ~ on G(f). If N is the index in G(f) of the image of k*(f) in G(f), then it is clear at any rate that all the values taken by x ~ on G(f) lie in the field Ko generated by the values taken by • on #*. The determination os K0 amounts to an exercise in Galois theory; one should observe that Ko need not contain k. We now come back to the construction given above for X when the values of X are given on G(m), m being a multiple of f. It obviously depends upon the following fact (which is equivalent to the theorem on the independence os valuations on k): Is I(m) is the group of the id~les a=(a~) such that a ~ = l for all 2, and a,~-I for every prime divisor P of m, then the group Pk/(m) is everywhere dense in Ik. It amounts to the same to say that the image of I(m) in C~ is everywhere dense in Ck. This implies that a character os Ck is completely determined by its values on I(m). We shall denote by I'(m) the compact subgroup of I(m) consisting of the id~les a e I(m) such that ( a ) : l ; then I(m)/I'(m) is discrete and may be identified with G(m). Let ~o be any representation os Ck into a complete topological group /'; as usual, we make no distinction between ~o and the corresponding representation of Ik into F. Assume that there is an m such that ~ : 1 on I'(m). Then ~ determines a representation ~ of G(m) into F, and ~ is uniquely determined by ~ since the image of I(m) in Ck is everywhere dense. We may now ask, conversely, whether, if a representation ~ of G(m) into F is given, it determines a representation ~ os Ck into F. This will be so if and only if the representation into F of the image of I0n) in Ck which is determined by ~5 is continuous for the topology induced on that image by the topology os Ck; for then i~ will be uniformly continuous, and can be extended by continuity. This is easily seen to amount to the following condition. To every neighborhood V of the neutral element in F, there must be an integer N and an ~ > 0 such that we have ~((a)) e V for every a e k*(m ~) which satisfies the conditions I a ~ - i I ~ for all 2. Now let z be a character of Ck os type (A0); then ~ takes its values in a subfield K of C, of finite deawee over Q. If m is any multiple of the conductor of X, we have, for ~ e k*(m)and a~>0 for all p:
~((,~))= x(,~) = T[,~I~,~;,':,. 4,.
Let w be any valuation of K, and let K~ be the completion os K with respect to w; the above criterion shows that ~ determines a representation ~ of C, into K~*, satisfying )~w(a)=~((a)) for a e I(m),
260
[1955c] 6
A, WElL
provided e i t h e r w is a valuation at infinity o r w is attached to an ideal ~ and we take m =pf, where p is the rational prime which is a multiple of ~. As K is embedded in C, we may of course take for w the valuation Wo induced by the ordinary absolute value on C; then ~,Oo=)~. Other valuations of K at infinity determine characters of Ck in the usual sense, i.e. representations of Ck into C; the corresponding L-series are the conjugates over Q of the series attached to the given X. On the other hand, for each prime ideal ~ in K, we get a representation ~ of Ck into K~, invariantly associated with )~. As the connected component of 1 in the group K~ is {1}, )r takes the value 1 on the connected component of 1 in Ck. As C~ is the direct product of its maximal compact subgroup and of a group isomorphic to R, and as ) ~ takes the value 1 on the latter group, ) ~ must map C, onto a compact subgroup of K~ and therefore onto a subgroup of the group U~ of units in K~. Now let co be any character of the compact group U~; as U~ is the projective limit of finite groups, ~o must be of finite order; therefore o2oX~ is a character of finite order of C,, which, by class-field theory, determines a cyclic extension k' of k. If, for a given )~ and ~, we make all possible choices of o,, these cyclic extensions will generate a certain abelian extension k()~, ~) of k; the compositum of these for all ~ will be an abelian extension k(x) of k which is thus invariantly attached to x. If X is of finite order n, its values on Ik (not merely those on some I(m)) are n-th roots of unity; then, for every w, ~ . is the transform of ~ by an isomorphism into K,* of the multiplicative group of the n-th roots of unity; in that case, k(%) is the cyclic extension attached to )~ by class-field theory. In all other cases k()~) will be an infinite extension of k. If X is the trivial character :g(a)=Ha]], k()~) is the maximal cyclotomic extension of k; more generally, if )~ is any trivial character of type (A0), k()~)will be contained in the maximal cyclotomic extension of a cyclic extension of /c of finite degree. As to the non-trivial characters of type (Ao), some of them arise in connection with the theory of abelian varieties with complex multiplication; in fact, all the characters of type (Ao) can be expressed in terms of those which arise in that manner and of the trivial ones. Taniyama has'proved that the L-series attached to the characters of type (A0) belonging to abelian varieties with complex multiplication are precisely those which occur in the zeta-functions of such varieties; and his recent work (done since the symposium) has shown that the fields generated by the points of finite order on these varieties are
[1955c]
261 On a Certain Type of Characters of the Id~le-Class Group
closely related to the fields k(z) defined above. For more general results, including these as rather special cases, the reader must be referred to his forthcoming publications; all that can be said here is that they tend to emphasize the imDortance of the characters which we have discussed and of their remarkable properties. UNIVERSITY OF CHICAGO
[ 1955d] On the theory of complex multiplication I shall concentrate chiefly on those aspects of m y work which have not been duplicated by the parallel and independent investigations of Shimura and of Taniyama. A preliminary account of their results, which are more complete than m y own in several important respects, appears in this same volume; it is understood that t hey will later give a full exposition of the whole theory. We need the concept of polarized variety; the word "polarization" is chosen so as to suggest an analogy with the concept of "oriented manifold " in topology. Let V be a complete non-singular variety; X being a divisor on V, denote by C(X) the class of all the divisors X ' such that there are two integers m, m', both >0, for which m'X' is algebraically equivalent to reX. We say that the class C(X) determines a polarization of V if it contains at least one ample complete linear system, or in other words if there exists a projective embedding of V for which the hyperplane sections belong to C(X). Thus a polarized variety may be regarded as a variety with a distinguished class of projective embeddings. The class C(X) is uniquely determined by any divisor in it;'every divisor in C(X) will be called a polar divisor of V for the polarization determined by that class. It is clearly the same to say that a variety is polarizable or that it is projectively embeddable. Let V be a variety, defined over a field k. Let X he a divisor on V, defining a polarization of V. If the smallest field containing k, over which X is rational, is not algebraic over k, then X belongs to an algebraic family, defined over an algebraic extension of k, and may be replaced by a member of that family, algebraically equivalent to X and algebraic over k. Thus we may assume that X itself is algebraic over k. Call Y the sum of all conjugates of X over k; "if p is the characteristic, then, for a suitable m, p~"Y will be rational over k; and one sees immediately that it determines a polarization of V, although not necessarily the same as the original one. We say that a polarized variety V is defined over k if V is defined over k and if there is on V a polar divisor which is rational over k; this Reprinted from Proc. Intern. Symp. on Algebraic Number Theory, Tokyo-Nikko, 1955, pp. 9-22.
263
264
[1955d] lO
A. WEIL
amounts to saying that V has a projective embedding which is defined over k. As an important example, we mention the case of the jacobian variety J of a curve F ; the canonical divisor 0 on J (canonical, that is to say, up to a translation) determines a polarization of J which will be called its canonical polarization. A classical result, due to Torelli, and for which it would be .worth while to give a modernized proof covering the abstract case, asserts that two curves are isomorphic if and only if their canonically polarized jacobians are isomorphic. Let A be an abelian variety; we denote by A* its dual, and by C1 the canonical homomorphism of ~(A) onto A*, with the kernel Q~(A). Every divisor X on A determines a homomorphism ~Ox of A into A*, defined by cpxu=C1 ( X ~ - X ) . If p = 0 , the degree v(~ox) of ~ox is always the square of an integer. If X>-0, ,(~x) is >0, i.e. 9x is surjective, if and only if there is an m > 0 such that m X determines an ample complete linear system on A, i.e. if and only if X determines a polarization of A. Conversely, let A be polarized; then every polar divisor X on A determines a homomorphism ~Ox of A onto A*; in the extension ~ ( A , A * ) ~ ) Q by Q of the group of homomorphisms of A into A*, q~x is uniquely determined by the polarization of A up to a positive rational factor. If ~ is a homomorphism of A* onto A such that ~q~x is of the form m ~ , then ~-I(X) determines a polarization of A* which is canonically associated with that of A. In the case p = 0 , there will be a polar divisor X on A such that every polar divisor is algebraically equivalent to a multiple m X of X; such a divisor will be called basic; if, for such a divisor X, we put ,(~o~)=r ~, r is called the rank of the polarized variety A. As usual, if A, B are abelian varieties, J((A, B) will denote the additive group of homomorphisms of A iuto B, ~r B) its extension by Q (i.e. the vector-space ~ ( A , B ) ~ Q over Q), ~ ( A ) the ring of endomorphisms of A, ~0(A) its extension by Q. If i e J/0(A, B ) a n d ,(i)-~0 (which implies that A, B have the same dimension, since the " d e g r e e " ,(~) of i is not defined otherwise), then i-~ is defined and is in ~o(B, A). If ~ is a homomorphism of A into B, its transpose ~2 is the homomorphism of B* into A* defined by ti(C1Z)=C1 (~-~(Z)) for every Z e Q~,(B); this extends to an isomorphism of J(o(A, B) onto
J(o(B*, A *). If A is a polarized abelian variety, and X is a polar divisor of A, put, for every a ~ ~o(A), d=~o~t'%'~ox; then ~ - ~ a ' is an involutory antiautomorphism of the algebra (Ao(A), canonically attached to the polarization of A. The trace a being defined on ~o(A) as
[1955d]
265 On the Theory of Complex Multiplication
11
usual, we have a ( ~ a ' ) > 0 for every a-7~=0 in ~go(A). If A is the (canonically polarized) jacobian of a curve, then ~ - a ' is no other than the so-called "Rosati antiautomorphism" Let a be a homomorphism of an abelian variety A onto an abelian variety B of the same dimension; if Y is a divisor on B, and if we put X=a-I(Y), we have ~ox=~a.~o~.a; in particular, if Y determines a polarization on B, so does X on A. If A is polarized and X is a polar divisor of A, and if ~ is an automorphism of the non-polarized A, then it will be an automorphism of the polarized A if and only if there are integers m, m', both >0, such that m~x=m'%.~Ox.Ot; taking degrees on both sides, we get m = m ' . But this may be written as d~=SA and implies a ( ~' ) : a ~ A ) . As a(~'~) is a positive nondegenerate quadratic form on ~Zo(A), and the additive group of ~ ( A ) is finitely generated, this shows that the group of automorphisms of a polarized abelian variety is finite (a result originally due to Matsusaka, whose proof, based on a different idea, is to appear shortly). From now on, A will be a polarized abelian variety of dimension n; we usually write ~ , ~0 instead of J ( A ) , ~o(A); on ~o, we have the trace a and the antiautomorphism a ~ d . For every prime l, not equal to the characteristic, ~ has a faithful representation R, of trace a by endomorphisms of a free module of rank 2n over the l-adic integers; this can be extended to a representation R, of (~o by endomorphisms of a vector-space of dimension 2n over l-adic numbers. If the characteristic is 0, J / has a faithful representation R of trace a by endomorphisms of a free abelian group of rank 2n (viz., the fundamental group of the complex torus defined by A under any embedding of its field of definition into C); this can be extended to a representation R of ~Zo in a vector-space of dimension 2n over Q; and the representations R~ can be derived from R by extending the group (resp. the vector-space) on which R operates by the ring of /-adic integers (resp. by the /-adic number-field). Moreover, ~//may also be considered as operating on the Lie algebra of A, i.e. on the tangent vector-space to A at 0; if p = 0 , this can be extended to a representation Ro of ~o by endomorphisms of a vector-space of dimension n over any common field of definition for A and its endomorphisms. By embedding such a field into C, one finds that R decomposes over C into Ro and the imaginary conjugate representation Ro; if we call 60 the trace of R0, we have a=~o+~o. Let ~ , . . . , e~ be orthogonal idempotents in ~o, i.e. elements such that ~ = ~ for all i and ~ j = 0 for i=/==j; put ~ o = ~ A - ~ , ~ ; we can write -z~=o~,/m, where m is an integer and d0,'" ", al, are in ~ . Call
266
[1955d] 12
A. WEIL
A~ the image of A by a~; then it is easy to see that A is isogenous to Ao • "" 9• A~, and that ~(~)--2 dim (A~). Let 0 be a semi-simple commutative subalgebra of ~o; in terms of suitable orthogonal idempotents r ~ , it can be written as O=~,K~r where the K, are fields. As ~o has faithful representations with the rational-valued trace o, C has representations of the same type; this implies that, if ~---~$,~, is in C, with $, e K~ for l _ ~ i ~ h , we have ~($)=~,,~ Tr(~,), where Tr is the ordinary trace (taken in K, over Q for each i) and the ,, are integers > 0. If the A~ are defined as above, we have 2dim(A~)=a(r QJ, hence ~, ,,EK~ : Q~ ~ 2n. Assume now that ~. ~K, : Q~ ~ 2n ; then the latter inequality must be an equality, and we must have v , = l for all i. That being so, a representation of ~o of trace a is equivalent (over an algebraically closed field) to one in which all elements of C appear as diagonal matrices and in which the diagonal elements corresponding to some element of C are all distinct; then the commuter C' of 0 in ~0 i~ also represented by diagonal matrices, which implies that it is commutative and semi-simple; what we have said about 0 can now also be applied to 0', and it easily follows from this that C'=C. In particular (as Shimura also proved), if U/o contains a field K of degree ~ 2 n , K must be of degree 2n, must contain ~A and the center of ~-~o, and is a maximal commutative subalgebra of ~o. When that is so, A must be isogenous to a product B • . - . • B, where B is simple; in fact, if this were not so, ~ o would be the direct sum of algebras ~o(A,), the A, being proper subvarieties of A, at least one of which would have to contain a field isomorphic to K, while we have just shown that ~o(A,) cannot contain a field of degree > 2 d i m (A~). Assume now that A is isogenous to a product /~• . . . • of r factors B of dimension m, so that n=rm; then ~o is the ring of matrices of order r over the division-algebra ~o=~_~o(B). Call /c the center of ~0, which we identify with the center of ~o, so tha~ K ~ k; call , the degree of k, p~ the dimension of -~o as a vectorspace over k. As K is of degree 2n/, over k and is maximally commutative in ~4o, it is known that ~o, as a vector-space over k, must be of dimension (2n/~)~; this gives rg=2n/,, hence 2 r e = g , ; therefore a maximal subfield of -~0, containing k, is of degree 2m. If now we assume that p = 0 , ~o must have a faithful representation by rational matrices of order 2m; as it is known that the order of such a representation must be a multiple of ,p~, this gives p = l , ~ o = k . Moreover, any polarization of B determines an automorphism ~-->U of k, of order 1 or 2, such that Tr ($$')~ 0; if ko consists of the elements
[1955d]
267 On the Theory of Complex Multiplication
13
of k invariant under that automorphism, this implies that ko must be a totally real field, and that k is either ko or a totally imaginary quadratic extension of ko. As before, call R0 the representation of ~o determined by the Lie algebra of A; call So the representation of -~0 which is similarly defined; then the representation of k of trace Tr~/Q decomposes into So and 20; this implies that k=/=ko and that So is the direct sum of m one-dimensional representations of ~o, i.e. of m isomorpbisms q~, of k into the universal domain, inducing on k0 all its distinct isomorphisms into the algebraic closure Q of Q. Moreover, Ro induces on k the representation (n/m)So; this implies that Ro induces on K the sum of the one-dimensional representations ~ ( 1 ~ ~ ~ m, 1 ~ i ~ n/m), where, for each i, the e ~ are all the isomorphisms of K into Q which induce q~ on k. Still assuming p = 0 , consider now any field K of degree 2n containing a totally imaginary qua~lratic extension k of a totally real field ko, the latter being of degree m. Let the ~ , be all the isomorphisms of ko into Q; for each ~, let ~ be an isomorphism of k into Q, inducing -#~, on ko, and let the ~oz~,for 1 ~ i ~ n/m, be all the isomorphisms of ~" into Q which induce e~ on k. We ask for the abelian varieties A of dimension n such that ~o(A) contains a field isomorphic to K and that Ro induces on K a representation which is the sum of the ~o~. Taking C as universal domain, it is easily seen that A is uniquely defined by these conditions up to an isogeny over C and that it can be constructed as follows. Consider the mapping $-4(~0~(~)) of K into C~; let M be the image under that mapping of a " m o d u l e " m in K, i.e. of a free abelian subgroup of rank 2n of the additive group of K; then the complex torus C~/M defines an abelian variety A with the required properties. If ( $ , , . . - , $./~) is a basis for K considered as a vector-space over k, we may in particular take m = ~ , ~ n , where n is a module in k; then one finds that A is the product of n/m varieties B of dimension m. This shows that A cannot be simple unless n==m. By a CM-extension of a totally, real field Ko of degre n over Q, we shall understand a system (K; {~}) consisting of a totally imaginary quadratic extension K of Ko and of n isomorphisms ~oz of K into Q, inducing on K0 all the isomorphisms of Ko into Q.
If we
consider Q as embedded in C, K can then be written as Ko(Q, where is such that -{2 is a totally positive element of _~ and that all the ~o~({) have a positive imaginary part; $ is uniquely determined by that condition up to a totally positive factor in Ko; conversely,
268
[1955d] 14
A. WEIL
the CM-extension (K; [~ox}) is uniquely determined by I(o and $ and will also be written as/(o((5)). The CM-extension Ko((4)) will be called primitive if it cannot be written as Ko((~)) with 4~ lying in a proper subfield of K0; Ko((4)) is primitive if and only if there is no conjugate ~' of ~ over Q, other than ~, such that C/~ is a totally positive algebraic number. The proof given above shows that every CMextension of a totally real field of degree n determines a " c a t e g o r y " of mutually isogenous abelian varieties of dimension n, and that the latter are simple if and only if the former is primitive. In consequence, it seems reasonable to deal first with the simple abelian varieties belonging to primitive CM-extensions, even though some important results have already been obtained by Taniyama for more general cases. From now on, let (k; [~o~}) be a primitive CMextension, given once for all, of a totally real field ko of degree n; we consider the abelian varieties A of dimension n which belong to it in the sense described above. This means that there is an isomorphism ~o of k onto JT0(A) such than Roo~O decomposes into the sum of the ~o~. As (k; [~o~}) is primitive, it is easily seen that ~ is uniquely determined by this condition, so that it may be used to identify k with ~0(A); this identification will be made from now on. Then the ring ~4(A) is identified with a subring r of the ring o of all integers in k. If K is a field of definition for A and for all the endomorphisms of A, k will have a representation of trace ~_] ~o~ by matrices of order n over the field K. One finds that, for k to have such a representation, it is necessary and sufficient that K should contain the field k, generated over Q by the values taken by that trace on k. Conversely, if K is a field of definition for A, containing kt, it must be a field of definition for all the endomorphisms of A. One should observe that k~ need not contain k. We now consider polarized abelian varieties belonging to a given CM-extension. The rank of such a variety, for p = 0 , has been defined above as the integer r=,(~Ox) ~/2 if X is a basic polar divisor. By using the representation of our varieties as complex toruses when C is taken as universal domain, one finds that, f o r a given CM-extension (k; [~o~}), a given ring of endomorphisms r, and a given value of the rank r, there is at most a finite number of distinct types of polarized abelian varieties with respect to isomorphism over the universal domain. If A is such a variety, its group of automorphisms is the multiplicative group of the roots of unity in r. Call o~ a generator of that group, and N its order. Let K be a field of definition for the polar-
[1955d]
269 15
On the Theory of Complex Multiplication
ized variety A and for its automorphism co; let X be a positive polar divisor on A, of which we may assume that it is rational over K and that it determines an ample complete linear system; after replacing X, if necessary, by the sum of its transforms by the N automorphisms of A, we may also assume that it is invariant by (o. Now identify A with its image under the projective embedding of A defined by the complete linear system determined by X; then ~ is induced on A by an automorphism ;2 of the ambient projective space which leaves invariant the hyperplane Ho such that Ho.A=X. If we take the homogeneous coordinates (X0,. 9-, X~) in that space so that H0 is defined by 2(o=0, $2 will appear as a linear substitution: m
(Xo, xl,
. . .,
(Xo,
clix
, . . .,
cmiX
).
For any r ~ m, let the Pr~ be a base for the space of homogeneous polynomials of degree r N in the X~ which are invariant under that substitution; let U~ be the locus of the point r with the homogeneous coordinates Pr~(x), in a projective space of suitable dimension, when x is a generic point of the ambient space of A. By adjoining the N-th roots of unity, if necessary, to the groundfield, and writing the substitution ~9 in diagonal form, one shows that all the varieties Ur are isomorphic to one another. Call U any one of them; call V the image of A in U b y r and call F the mapping of A onto V induced by ~ ; we say that 11, together with the mapping F, is the quotient of A by the group generated by ~o. Now, for each one of the finitely many types of varieties belonging to given data (k; [~o~}), r, r, we can construct a representative A by means of a complex torus. A variety A obtained by this method need of course not be defined over an algebraic number-field. However, I have given (in a paper j u s t published in the Amer. J. of Math".) a criterion for a variety, defined over a field/(1, to be isomorphic to a variety defined over a subfield /Co of /(1; by using this criterion, it is easily seen that, for each type of varieties belonging to the given data, there is a representative which is defined over an algebraic number-field. As this is only a special case of some important unpublished results of T. Matsusaka on the field of moduli of a polarized abelian variety, I need not give more details here; however, it will be worthwhile to consider more closely the case in which A is defined over an algebraic number-field, even though Matsusaka's results could also be applied to that case. Let therefore A be a 1) A. Weil, The field of definition of a variety, Amer. J. of Math., 78 (1956), pp. 509-524.
270
[1955d1 16
A. WEIL
polarized abelian variety belonging to the given data and defined over an algebraic number-field which we may assume to be a finite Galois extension K of kt. Let Ko be the field of the elements of K which are invariant under all those automorphisms of K over ks which transform A into a variety isomorphic to A; the degree of Ko over kt is at most equal to the number of possible types of varieties belonging to the given data. Let a be an automorphism of K over /(o; there is an isomorphism ao of A onto A ~ uniquely determined up to an automorphism of A and algebraic over K; therefore every conjugate of a~ over K i s of the form a.o~~. Call V t h e quotient of A by its group of automorphisms, and F the canonical mapping of A onto V; then there is an isomorphism /~ of V onto V ~ uniquely determined by the condition / ~ o F = F % a ~ ; it must be the same as its conjugates over K, and is therefore defined over K; and we have p~o=fl~o/~, for any two automorphisms r, a of K over K0. Applying the results of the paper quoted above, we conclude from this that there is a variety V0 defined over I(o and an isomorphism ~o of V0 onto V, defined over K, such that ~o=q~%~o-1. Let now A1 be any variety, isomorphic to A, defined over an algebraic number-field /(1 containing kt. If K~ does not contain K0, there must be an automorphism r of the field of all algebraic numbers over /s which does not leave invariant all elements of Ko; then, if a~ is an isomorphism of A onto A1, its transform by ~ is an isomorphism of A ~ onto A , so that A and A ~ must be isomorphic; but this contradicts the definition of K0. Therefore we have /(1 ~ Ko. If now V~ is the quotient of A~ by its group of automorphisms, F~ the canonical mapping of A1 onto V , P~ the isomorphism of V onto V~ such that / ~ o F = F ~ o ~ , and a any automorphism of KK~ over /s we have/~1-/~o/~o , hence ( ~ o ) ~ which shows that /3~q~ is an isomorphism of V0 onto V , defined over K , Call z a generic point of A over K, and w the corresponding point on V0, i.e. w=~o-~(F(z)). To each primitive N-th root of unity ~, we can associate the set of those functions 0 on A, defined over Q, which satisfy 0(o~z)=r for each such function, there is a funct i o n f on Vo such that f(w)=O(z)~; call ~ the set consisting of those functions f on Vo. If f ~ ~ , and ~ is another primitive N-th root of unity, , being an integer prime to N, then ~ consists of the functions f~h N, where h runs through the set of all functions on V0, defined over Q; also, if an automorphism of Q over /40 maps ~ onto r it will transform the functions in ~ into the functions in ~ . We say that V0, together with the sets of functions !~, is the Kummer wriety attached to the given type of abelian varieties (for a more
[1955d]
271 On the Theory of Complex Multiplication
17
general definition, valid for arbitrary polarized abelian varieties, we refer the reader to a forthcoming publication by Matsusaka); and we say that this Kummer variety is defined over K0. It is clear that a type of abelian varieties is completely determined by its Kummer variety. We can now formulate the basic problems of complex multiplication for simple abelian varieties: I. Characterize the fields Ko for the types of abelian varieties belonging to given data (k; [qozt), l" and r. II. For each such type, characterize the fields generated over Ko by the images on Vo of the points of finite order on a variety A of that type. III. Determine the zeta-function of any abelian variety of the given type, over a field of definition of that variety containing kt and therefore Ko. For n - - l , the complete solution of problems (I) and (II) is given by the classical theory of complex multiplication, and problem (III) was solved recently by Deuring. For arbitrary n, Taniyama has now solved a problem which includes the general case of (III) as a special case; the independent and overlapping investigations of Shimura, Taniyama and myself give the solution of (I) and (II) in the case r=L~; and one may hope that the general case will not offer insurmountable difficulties any more. The basic tool here is Shimura's theory of reduction modulo a prime ideal, by means of which our problems can be reduced to problems on abelian varieties over finite fields. I shall sketch briefly the main ideas involved here. As above, let A be a variety of one of the given types, defined over a field K containing kt. Shimura's theory shows that, for almost all prime ideals ~ in /( (i.e., for all except a finite number), one can reduce A and its endomorphisms modulo ~, obtaining an abelian variety A(~3) of dimension n defined over the finite field with N(~) elements and an isomorphism of r = , d ( A ) into ~J4(A(~)). Then the Frobenius endomorphism of A(~) (induced by the automorphism of the universal domain which raises every element to its N(~)-th power) commutes with every element of the image of r in ~?/(A(~3)), since such an element is an endomorphism of A(~) which is defined over the field with N(~3) elements. By the results proved above, this implies that the Frobenius endomorphism can be identified with an element 7r of the field k:=~ ~o(A), and more precisely with an integer in k (not necessarily in r). The mapping ~-~Tr determines the zeta-function of A over K; and Taniyama has shown that a more detailed study of
2 72
[1955d] 18
A. WEIL
the properties of this mapping leads directly to an expression of the zeta-function in terms of Hecke L-functions attached to characters of type (Ao) of the field K (cf. p. 4 of this volume); as could be expected, these are characters which come from the quadratic extension k of the totally real /co (in the sense of formula (4), p. 4). In fact, the connection between characters " o f type (Ao)" and abelian varieties with complex multiplication appears to be so close that it can hardly be accidental; and any f u t u r e arithmetical interpretation of the characters of type (Ao), corresponding to the interpretation given by class-field theory for the characters of finite order of the id~le-class group, ought to take complex multiplication into account. As to problems (I) and (II), I will consider only the case ~=-o. The method sketched below could perhaps be applied without substantial changes to a ring r such that r=~ and that the classes of ideals in r which belong properly to r (i.e. which consist of ideals m such that, in Dedekind's notation, m : m = r ) form a group. If n = l , all the subrings of o have these properties; for n > l , it does not seem to be known whether any proper subring of o has them; in order to treat the general case of problems (I) and (II), one will presumably have to rely more heavily upon the /-adic representations R~ than is done here. We first have to look more closely into the relation between k and k,. Taking C as universal domain, and taking k to be embedded in it, call k' the compositum of k and all its conjugates over Q; call G the Galois group of k' over Q; call H, H, the subgroups of G corresponding respectively to the subfields k, kt of M; call ~ the automorphism $-->~ of k'. Call S the set of those automorphisms of k' over Q which induce on k one of the isomorphisms ~p~. Thus S is the union of cosets with respect to H, i.e. we have HS=H: we have G=S""S,:,; more generally, if o' is any transform of 6 by an inner automorphism of G, we have G - S = S a ' = a ' S . The assumption that (k; [~o~}) is primitive amounts to saying that H consists of all the elements 7 of G such that 7 S = S . On the other hand, H, consists of Che elements 7' of G such that S,y'=S. The subgroup of G corresponding to ko is H"-"Ha; and one finds that Ht'-%Ht is a group, corresponding to a totally real subfield of k' of which k~ is a totally imaginary quadratic extension. Write S as the union of distinct cosets t~-lH, with respect to H~; for each t~, let ~ be the isomorphism of ks into k' induced by ~ on k~. Then (k,; [~,}) is a primitive CM-extension, and the relation between (k; [~o~}) and (kt; [~,}) is symmetric. This suggests that one should look for a relation between the categories
[1955d]
2 73 On the Theory of Complex Multiplication
19
of abelian varieties belonging to these CM-extensions; as to what this may be, I have no conjecture to offer. Before coming back to our problems, we must also observe that, for any abelian varieties A and B, J((A, B) is a right J(A)-module and a left ~(B)-module. If cp is an isomorphism of a commutative subring C of ~yT(A) onto a subring of ~ ( B ) , and if one considers only those e cJ((A, B) for which s for all 7 e C, the distinction between right and left is not necessary. In particular, consider two abelian varieties A, A' of dimension n, belonging as above to the primitive CM-extension (k; {q~}). Then they are isogenous; and, by considering the operation of J((A, A') on the Lie algebra of A, one sees that a S = Sa for all ~ e J/o(A, A') and all $ e k. Thus J(o(A, A') is a vectorspace over k; as such, it is clearly of dimension 1; and J/[A, A') is a module over the ring generated in k by .Y/(A) and ~ ( A ' ) . If now we assume that J ~ ( A ) = ~ ( A ' ) = o , then J((A, A') is an o-module, isomorphic to an o-ideal whose class is uniquely determined; if a is a nonzero element of ~o(A, A'), and if ~t is the set of the $ e k such that $~ e J((A, A'), ~ is an ideal in that class. If we take a e SEt(A, A'), we have 1 e a, so that a-~ is an ideal in o. In particular, the dual A* of A is isogenous to A; and, if A is polarized and Y is a basic divisor on A, ~o~ is in ~a((A, A*). If we assume A to have o as its ring of endomorphisms, the same will be true of A*, and the set of the S e k such that Sq~ye ~r(A, A*) will be an o-ideal in k. One finds, in fact, that it can be written as f;~0, where fo is an ideal in the ring of integers of /co, and that the rank r of A is r =N(fo). When that is so, we say that A belongs to fo; it is clear that all the conjugates of A over kt will belong to fo. Thus, in discussing our problems (I) and (II) for r=o, we may confine our attention to those types which belong to (k; {~o~}), r = o and a fixed fo. Let A, A' be two such varieties; let Y, Y' be basic divisors on A, A'; if ~ e J((A, A'), and if we put Z = ~ - ' ( Y ' ) , q~;~o~ will be in k; one finds that in fact it must be a totally positive integer; call it f(~). Take an a-J=0 in J/o(A, A'), so that we can write J((A, A')=a,~, where a is an o-ideal in k; then one finds that there is a totally positive peko such that pr
and that f(Sa)=p$~ for all Sea.
One may
call this a positive hermitian form on t~. The form psi, defined on ~, and the form p~$~, defined on an idea] a~, will be called equivalent if there is a ~ e k such that a~=~-~a and g~=p~; the class determined for this equivalence relation by the form pS~ on (~ will be denoted by (a ; p). That being so, the class of the form f($a) on the ideal a deter-
274
[1955d] 20
A. WEIL
mined by ~ ( A , A')=ae~ is independent of the choice of ~ and will be denoted by {A': A}; A and A' are isomorphic if and only if this class is (o ; 1). On the classes of forms, we define a group law by putting
(a ; p). (a'; p') = (~';
pp').
Then, if A, A', A" all belong to the same type, we have: {A": A} = {A": A'} 9[A': A}; and, if r is any automorphism of f~ over kt, we have {A'~: A *} = {A': A}. It immediately follows from this that every field Ko occurring in problem (I) for r = o is abelian over ks, with a Galois group which is isomorphic to a subgroup of the group of classes of forms (a; g). Take a field of definition K for A, A' and their endomorphisms and homomorphisms; again by Shimura's theory, we can reduce all of these modulo almost all prime ideals in K. For sueh a prime ~, d((A, A') is mapped isomorphieally onto its image in d f ( ~ ) = ~ ' ( A ( ~ ) , A'(~)) and may be identified with that image; similarly we can identify d/o(A, A') with its image in the extension of d((~) by Q. One sees at once that an element of the latter set is in d(o(A, A') if and only if it commutes with all elements of k. Now put ~ " = J t ( ~ J ) ~ o ( A , A'); this is clearly an o-module containing d((A, A'); we show that JC'=~C((A, A'). In fact, assume that this is not so; as both are o-modules, there will be a $ in k and not in such that ~r A')cd('. But (e.g. by using a representation of A, A' as complex toruses over C) one can see that there are elements a~ of ~'(A, A ' ) a n d elements a'~ of d((A', A) such that 3 ~ = ~ , a ~ (this is a speeial ease of the faet that, if A, A', A" are three varieties of the given type, d((A, A") is no other than the tensor-produet, taken over o, of the o-modules d((A, A') and df(A', A")). This gives $=~,a~.(&r~), so that ~ must be an endomorphism of A(~), which is absurd. Now let Pc be a prime ideal in k,; we assume that it is not ramified in K0 and also that it has in a suitable field K a non-exceptional prime divisor ~, i.e. one modulo which one can reduce A, all its conjugates over k~, and the endomorphisms and homomorphisms of these varieties. Put q=N0~t). Take for A' the transform of A by an automorphism T of Q over k~ which induces on K0 the Frobenius substitution for p,. By what we have seen above, the Frobenius homomorphism of A(~) onto A'(~), induced by the automorphism x ~ x q of the universal domain, will be the image of an element ~ of J((A, A'); and we have f(w)--q. Then, if we put 3'(A, A ' ) = q - ~ , q is an ideal in o, such that q~=q% and we have {A': A}--(q-~; q).
[1955d]
275 On the Theory of Complex Multiplication
21
By class-field theory, an abelian extension is completely determined by the knowledge of the Frobenius substitution for almost all prime ideals; therefore (I) will be solved if we determine the correspondence p~-> c~. Let m be a multiple of the order of the Frobenius substitution for ~ in Ko; then r~ transforms A into a variety A1 isomorphic to A. Call al an isomorphism of A onto A~; this is uniquely determined up to an automorphism, i.e. up to a root of unity. Then the automorphism x - ~ x ~'~ of the universal domain induces a homomorphism of A(~) onto AI(~), which, as above, may be identified with an element of ~ ( A , A1); this can be written as Tral with 7r e o; ~r is uniquely determined up to a root of unity. Proceeding as above, we find that 7r~=q ~ and that 7r~=q% One should observe that, if N ( ~ ) = q ~ and m is a multiple of h, then A I ( ~ ) - - A ( ~ ) , so that in that case a~ can be determined uniquely by prescribing that it should reduce to the identity mapping on A(~); then ~r also is completely determined. Now, in order to find q, it is enough to determine the prime ideal decomposition of v in a suitable field for some suitable choice of m, e.g. for m = h. B u t this has been done by T a n i y a m a (cf. w3 of his contribution to this volume). The conclusion is that, for almost all t~, we have q=l~I ~ ( ~ ) provided of course ideals in subfields of U (the smallest Galois extension of Q containing k) are identified in the customary way with the ideals they generate in U. This formula contains the solution of problem (I) for r=~. One should observe that, while the prime ideal decomposition of ~r, together with the relation ~r~=q *'~, determines 7r up to a root of unity, this is not enough for the calculation of the zeta-function, where a more vrecise result (also contained in Taniyama's work) is required. For r=~, problem (II) can be treated by an entirely similar method. We consider the pairs (A, a), where A is an abelian variety of one of the types discussed above, and a is a point of finite order on A. If ~ is the ideal in % consisting of those ~ for which Sa=0, we say that a belongs to ~. The type of the pair (A, a) will be considered as given by the type of A, i.e. by the data (k; [~o~}), ~--o and f0, and by the ideal zT in o. Consider two such pairs (A, a) and (A', a'). If we write, as above, J((A, A ' ) = a a and f ( $ ~ ) - p ~ , an element ~o of a such that ~0~ maps a onto a'; it uniquely modulo a~ and is such that $00+~m=a. That define an equivalence relation between triplets a, p,
there will be is determined being so, we $0, where a,
276
[1955d] 22
A. WEIL
are as before and $o is such that $oO+an=a, by defining two triplets a, g, $o and a', p', $~ to be equivalent if there is a ~.ek such that a'=~-'% g ' = p ~ and ~ - ' ~ o mod. a'n; and we denote by (a; p; ~o) the class of such a triplet. Then the class of the triplet a, p, $o attached as above to the two pairs (A, a) and (A', a') is independent of the choice of ~; it will be denoted by {(A', a'} : (A, a)} ; the two pairs are isomorphic if and only if the class is (o; 1 ; 1). A group law between equivalence classes is defined by putting (a ; p; -:o)"(a'; p'; $'o)= (aa'; pp'; ~o~). Proceeding as above, one finds that the Frobenius substitution for Pt, in the field generated over Ko by the image of the point a on Vo, is (q_t; q; 1). This solves problem (II).
UNIVERSITY OF CHICAGO
[1955e] Science Franqaise?
L ' a n dernier, un h o m m e politique assez c o n n u d6jA, et qui l'est encore plus A pr6sent, s'6tonnait que, depuis fort longtemps, a u c u n s a v a n t fran?ais n'efit re?u de prix Nobel. L'occasion 6tait solennelle ; il exposait son p r o g r a m m e de g o u v e r n e m e n t . S'il faisait cette constaration, ce n'6tait pas seulement, sans doute, p o u r s'attfister d ' u n e situation si humiliante p o u r notre a m o u r propre national. C'est qu'il e n t e n d a i t que la prise du pouvoir lui d o n n e r a i t la ,facult6 d ' y p o r t e r remade. Or: sont-ils, ces rem6des ? Sont-ils fort cach6s ? E t ce qui est p o u r nos h o m m e s politiques u n sujet d'6tonn e m e n t en est-il un pour les initi6s ? Ici, je d e m a n d e la permission de raconter m o n histoire, ou p l u t 6 t de copier quelques passages d ' u n article que j'6crivis, fort jeune encore, A m o n retour d ' u n v o y a g e en Am6rique, il y a pros de vingt ans. Ce v o y a g e faisait suite A b e a u c o u p d'autres, en Allemagne, en Angleterre, en Italie, en Russie mSme (est-il p r u d e n t de l'avouer ?), et j u s q u ' e n Asie. Mon article fur soumis ~ quelques revues, qui le jug~rent impubliable ; il y a des v6rit6s qui ne sont pas bonnes A dire ; on ne se priva pas de me le faire savoir ; je ne profitai gu~re de la lemon... ~ J ' e n ai assez, 6cfivais-je. J ' a i m e v o y a g e r A l'6trang e r ; mes amis s a v e n t que m o n a m o u r - p r o p r e national n'est pas chatouilleux A l'exc~s, et j'ai pfis d6s longt e m p s l'habitude d'entendre, sans trop m'6mouvoir, qu'on discute, parfois sans bienveillance, de m o n pays, de ses h6tels, de ses femmes, de ses politiciens. Q u ' y Reprinted from La Nouvelle N.R.F., Paris, Imp. Crrtr, 3e annre, n~ 25, pp. 97-109.
277
278
[1955e]
4
LA N O U V E L L E
N.R.F.
ferais-je, si t o u t cela est vrai ? que m ' i m p o r t e , si t o u t cela est faux ? Mais j'en ai assez, q u a n d je rencontre un ehimiste, qu'il me d e m a n d e i n v a r i a b l e m e n t : ~ Pourquoi la chimie fran~aise est-elle tombfie si bus ? ~; si c'est u n biologiste : ~cPourquoi la biologic fran~aise va-t-elle si mal ? ~; si c'est un physicien : ~ Pourquoi la physique fran~aise... ~; mais je n'ach~ve pas, c'est toujours la mfime question d o n t on me rebat les oreilles, et j'en suis encore ~ chercher la r@onse. Bien stir, q u a n d je d e m a n d e des pr~eisions, il arrive qu'on reconnaisse qu'il existe encore chez nous, duns tel ou tel domaine, des s a v a n t s fort distingu~s. Q u a n t ~ moi, m a t h ~ m a t i c i e n t o u t ~ fait ignorant de t o u t e science sinon de la mienne, je ne puis discuter ; souvent je me hasarde, en r@onse l'~ternelle question, ~ suggfirer : ccMais un tel... ? ~ et je cite un nom, illustre chez nous ; mais j'ai fini par y renoncer, car pour une lois qu'on m ' a v o u e : c~E n effet, il y a t o u t de mfime un tel ~, trop souvent, l'illustre coll~gne est assomm~ aussit6t d ' u n m o t dfidaigneux, d ' n n sourire, on s i m p l e m e n t d ' u n h a n s s e m e n t d'@aules... (( E n t e n d o n s - n o u s . Les mceurs de la gent universitaire, depuis qnelque douze ans que je la frfiquente, me sont un peu connues ; qu'on ne vienne pas me parler ici de jalousie, d'ignorance ou de pr~jugfi : on n'expliquera pas ainsi que tous ces coll~gues fitrangers, et s u r t o u t les jeunes, me posent toujours, ~ pen pros dans les m~mes termes, la m~me question. Ils reconnaissent sans se gfiner, de quelque pays qu'ils soient, l'importance des centres scientifiques anglais, am~ricains, russes, allemands ; ils s a v e n t apprficier aussi, parfois avec beaucoup de chaleur, les mfirites de tel s a v a n t fran~ais. Ce ne p e u t fitre la jalousie qui les fair tous parler ; il y a a u t r e chose ; il y a, faut-il le dire, un fait : ils doivent avoir raison. Cela est fficheux; expliquons-le c o m m e nous pouvons ; mais m i e u x v a u t le reconnaitre ; m i e u x v a u t mfime, c o m m e je le fais ici ~ dessein, s'exagfirer pent-~tre l%tendue du mal que de s o t t e m e n t fermer les y e u x . Assez parlfi (avec des majuscules) de Science F r a n ~ a i s e ; assez invoqufis les m~nes de Pasteur, de Poincar~, de Lavoisier : qu'ils se reposent en paix, car ils l'ont bien m~rit~, ce repos qu'on ne v e u t pas accorder leurs ombres ; la Science Fran~aise, apr~s tout, c'est nous, c'est les vivants, et leurs noms ne sont pas une
[1955e]
279
SCIENCE
FRAN~AISE
?
m i n e d o n t on nous ait octroy6 la concession ~ perp6tuit6 ; si nous ne savons pas nous e x a m i n e r avee s~v6rit6, sans complaisance facile, d ' a u t r e s le font p o u r nous. Quelques-uns diront : c
280
[1955e]
6
LA N O U V E L L E
N.R.F.
esdme, parfois avec respect. Je sais qu'on les c o m p t e sur les doigts, et que leur 6minence ne rend que plus sensible la platitude de la contr6e environnante. Pourquoi n'y en a-t-il pas plus ? D~s 1937, j'avais c m en apercevoir les raisons : ~cSupposons que, dans tel ou tel domaine, disons la th6orie des hombres (il ne me co~te rien d'en parler, elle n'est pas enseign6e dans les universit6s fran~aises), les maitres v~ritables soient venus ~ faire d6faut ; que les chaires les plus en vue et les positions dominantes se trouvent occup6es par des hommes, non pas ignorants ou sans comp6tence, mais sans 6clat, ou, chose peut-~tre plus grave encore, par de ces savants (ilssont nombreux, et, pour des raisons qu'il faudrait bien examiner, ils le sont tout particuli6rement dans les universit6s fran~aises) ~ qui quelques travaux brillantsont valu au d6but de leur carri6re une r6putation qu'ils n'ont pu on ne se sont pas souci6s de soutenir. Que va-t-il se passer, si de tels h o m m e s (charg6s d'honneurs, sans doute, et de titres) sont install6s an pouvoir ? Car, reconnaissons-le, c'est un pouvoir v6ritable qu'ils d6tiennent ; pouvoir de distribuer les places; pouvoir, plus important encore lorsqu'il s'agit de science exp6rimentale (et c'est pourquoi chaque jour en m e levant je remercie Dieu de m'avoir fair math6maticien), d'allouer les cr6dits de laboratoire et les moyens de recherche; pouvoir, de par les positions qu'ils occupent, d'attirer~ soi les jeunes, et de conserver pour soi les collaborateurs qui ~ d'autres sont refus6s. De ces jeunes, que va-t-il arriver ? quel est l'avenir d'une science dont l'enseignement est une lois tomb6 entre les mains de pontifes de cette esp~ce ? Maints exemples, que j'ai pu 6tudier (et non pas seulement en France, qu'on le croie bien; je ne crois pas tout parfait ailleurs,et j'ai observ6 en d'autres pays des ph6nom~nes tout semblables), permettent de donner de ce qui doit se passer une description assez pr6cise : le tableau clinique de la maladie (comme disent, je crois, les m6decins) est bien connu. De tels h o m m e s ne tardent pas ~ tomber en dehors des grands courants de la science : non pas de la Science Fran~aise, mais de la science (sans majuscule) qui est universelle; ils travaillent, souvent honn~tement, de tr6s bonne foi et non sans talent, ou, d'autres fois,ils font semblant de
[1955e]
281
SCIENCE FRAN~AISE ?
7
travailler, mais en t o u t cas ils sont 6trangers aux grands probl~mes, aux id6es vivantes de la science de leur 6poque ; et A leur suite, c'est t o u t e lear 6cole qui se t r o u v e 6gar6e dans des eaux stagnantes (parfois bourbeuses, mais, cela, e'est une autre histoire) : des jeunes gens bien dou6s passent les ann~es les plus importantes de lear carri~re scientifique, les premieres, ~ travailler des probl~mes sans port@ et dans des voies sans issue. I1 faudrait les envoyer ~ l'6tranger, ces jeunes gens, les initier ~ toutes les m6thodes, ~ toutes les id6es ; car, q u a n d bien m~me il s'agirait du maitre le plus 6minent, qu'est-ce qae l'61~ve d ' u n seal maitre ? Mais quoi ? L'on a trop peur de perdre des collaborateurs et des disciples, et, ~ leur place, de voir revenir des juges, des juges s6v~res. Qu'il est pr6f6rable de les garder aupr~s de soi, de s'en faire aider, de les maintenir, a u t a n t qu'il se peut, dans des voies trac6es! Qu'ils aient du talent, c'est b i e n ; qu'ils soient sages de plus, et (sans nuire la hi6rarchie ni ~ l'ordre d'aneiennet6) routes les voies leur sont ouvertes ; et, s'ils sont sages, le talent m~me apr~s t o u t n'est pas indispensable, une bonne petite chaire les r6compensera. ~> J'ai copi6 m o t pour mot, et m o n manuscrit de 1937 est 1~ pour le prouver. J'avais bien 6crit <; je n'avais pour cela d ' a u t r e raison que celle que j'en donnais, c'est-~-dire qu'il n'existait alors aueune chaire de ee titre. J'6tais assez naif pour esp6rer ainsi ne blesser personne, dans eet article or1 je blessais t o u t le monde. Je ne pr6voyais pas q u ' u n jour serait cr66e ~ la Sorbonne une chaire de th6orie des nombres, ni que ce serait au profit d ' u n administrateur chevronn6, ancien reeteur et directeur de ministate ; encore moins pouvais-je pr6voir que la n o m i n a t i o n de son successeur serait l'occasion d ' u n 6pisode qu'il v a u t la peine de raeonter iei, car il complete, fort h e u r e u s e m e n t du point de r u e du clinieien, fort f~cheusement de t o u t autre point de rue, le <
282
[1955e]
8
LA NOUVELLE N.R.F.
mand, l'autre fran~ais. Nous les appellerons A e t ]3. Tous deux se sont partieuli+rement distingu6s en th6orie des hombres. A 6tait rest6 eu Allemagne jusqu'au d6but de 194o. I1 n'6tait pas juif. Ses sentiments d'hostilit6 au r6gime 6talent bien connus de ses coll~gues et n'6taient pas ignor6s des autorit6s universitaires; mais il n'avait jamais eu aucune activit6 politique et n'avait pas 6t6 inqui6t6. Peut&tre aurait-il quitt6 son pays plus t6t s'il n'avait pas pens6, par sa pr6sence, renforcer ce qui restait alors en Allemagne de pens6e libre ; il est vrai aussi qu'un voyage en Am6rique l'avait convaincu que le elimat intellectuel de ce pays lui convenait real. S'il se d6cida ~ 6migrer en 194o, ce qui n'alla pas sans difficult6s ni risques, ce fut sans doute qu'alors il d6sesp6ra de l'avenir de l'Europe. P e n d a n t la guerre, il fur trait6 par les Am6ricains en r6fugi6, c'est-~-dire assez real. Du moins y trouva-t-il de quoi vivre et poursuivre ses travaux, tandis que la plupart et les meilleurs des savants allemands qui avaient cherch6 un asile en France ~ la suite des premieres pers6cutions hitl6riennes avaient dfi en repartir faute de possibilit6s de travail. Vers la fin de la guerre, l'une des chaires si envi6es de l'Institute for Advanced Study, de Princeton, lui rut offerte ; il l'aecepta, et se fit citoyen am6ricain. Mais l'Allemagne se relevait de ses ruines mat6rielles et intellectuelles plus vite qu'on n'avait pu s'y attendre ; le travail seientifique y redevenait possible. Malgr6 la longueur de son s6jour en Am6rique, A n'avait pu s'aecoutumer ~ bien des aspects de la vie am6ricaine qui lui avaient d6plu d~s l'abord ; quelques-uns de ses amis rest6s en Allemagne s'en aper~urent. I1 n'en fallut pas plus. Bient6t, la plus c61~bre des universit6s d'Allemagne occidentale lui offrit une chaire. I1 ne pouvait ~tre question 1~ d'une situation mat6rielle comparable celle qu'il avait ~ P r i n c e t o n ; mais on salt que les universit6s allemandes peuvent, dans une certaine mesure, proportionner le traitement ~ la valeur et ~ la r6putation seientifiques ; on lui en offrit un fort sup6rieur ~ celui de la plupart de ses coll~gues allemands ; pour l'attirer, on lui offrit le remboursement de tous ses frais de d6m6nagement. Suivant la loi allemande, la nomination d'un 6tranger ~ une chaire universitaire
[1955e]
283
SCIENCE FRAN~AISE ?
9
lui conf6re de plein droit la naturalisation ; on offrit au professeur A, ~ son choix, de reprendre la nationalit~ allemande ou de conserver la nationalitfi am~ricaine. E t t o u t cela fur fair sans qu'il efit rien demandS, sans qu'il efit eu ~ rien demander. I1 accepta ; et sa pr~sence et son enseignement n ' o n t pas peu contribu~ ~ rendre l'Universit~ dont il s'agit une pattie du lustre qu'elle avait eu autrefois. Quant au Fran~ais, la d6claration de g u e r r e l ' a v a i t surpris 5 P r i n c e t o n ; mobilisable, il w i t l'avis de l'ambassade de France, qui lui r e c o m m a n d a de rester o~ il ~tait ; un professeur fran~ais ~ l%tranger, pensait-on alors assez raisonnablement, ~tait plus utile 5 la France q u ' u n soldat de plus sons l'uniforme. I1 passa done la guerre aux l~tats-Unis. Celle-ci finie, les postes les plus brillants lui furent bient6t offerts ; Princeton, Harvard, Columbia se le disputbrent. C'est dans cette derni6re universit~ qu'il se fixa, dans l'une des meilleures chaires qu'elle efit ~ offrir ; quelque t e m p s auparavant, il s%tait fait naturaliser am~rieain. MMs lui aussi se lassa des ~;tats-Unis, et dfisira rentrer en France. C'est ici que son histoire cesse de ressembler 5 la prficfidente. T o u t d'abord, en France, on n'offre pas une chaire nn savant, si distingufi soit-il ; il faut qu'il fasse acte de candidature ; il faut le plus souvent qu'il fasse ses visites de candidature, formalitfi destin~e principalement p e r m e t t r e ~ ceux d o n t il postule les suffrages de juger de la souplesse de son ~chine. Les amis du professeur B, mis au courant de ses intentions, a t t e n d i r e n t longtemps une occasion favorable. Plusienrs chaires devinrent vaeantes, mais chaque fois les jeux ~taient faits. Enfin le titulaire de la chaire de th~orie des nombres prit sa retraite ; les amis de B pens6rent qu'il ne convenait pas de difffirer plus longtemps. Pour eette chaire, B ~tait si fiminemment qualififi qu'il ne semblait pas que quieonque, en France ou ailleurs, pfit la lui disputer sans ridicule. Mais il fallait d'abord que cette candidature ffit recevable. La loi fran~aise n ' a d m e t plus, depuis longtemps d~j~, que nos universitfis puissent s'enrichir par des nominations de savants ~trangers, comme c'est l'usage dans presque t o u s l e s autres pays ; il e n e s t ainsi, q u a n d bien m~me l%tranger ne serait tel que par naturali-
284
[1955e] 10
LA NOUVELLE N.R.F.
sation. Mais, par hasard, B n'avait pas perdu sa nationalit6 frangaise en acqu6rant la nationalit6 am6ricaine. Force fur ~ la Sorbonne d'enregistrer sa candidature. Alors se d6clencha une campagne d'une violence extraordinaire. On vit un membre de l'Institut monter en personne sur la br~che pour d6fendre la citadelle menac6e. On feignit de mettre en doute la valeur math6matique du candidat. Parmi les 6loges que la critique avait d6cern6s ~ ses ouvrages, on rechercha les r6serves et les objections de d6tail ; par un montage habile de citations tronqu6es, on composa un texte qui pflt impressionner d6favorablement les incomp6tents; or, comme c'est l'ensemble d'une Facult6 qui vote sur chaque nomination, routes sp6cialit6s r6unies (depuis les math6matiques jusqu'~ la botanique), c'est n6cessairement, en chaque cas, une majorit6 d'incomp6tents qui d6cide. On reprocha ~ B de n'~tre pas rentr6 endosser un uniforme en 1939; on mobilisa contre lui les (~anciens c o m b a t t a n t s ~ et (( anciens r6sistants ~ professionnels ; il n'en manque pas dans l'Universit6, dont toute la carri~re ne se fonde que l~-dessus ; et je ne parle pas de ces patriotes tardifs, toujours cherchant ~ faire oublier qu'ils se sont d6shonor6s, et obligeant par 1~ m~me, quoi qu'on en air, ~ s'en souvenir toujours. On gonfla les m6rites du suppl6ant du pr6c6dent titulaire de la chaire. On fit si bien que ce suppl6ant l'emporta. La seule consolation de B fur que son 61oignement l'avait pr6serv6 de participer ~ cette m~16e sordide. Ses amis s%taient charg6s pour lui des visites de c a n d i d a t u r e ; c'6taient des savants fort distingu6s, eux aussi; le r6sultat a prouv6 qu'ils auraient pu mieux employer leur temps. Mais l'histoire ne finit pas 1A. A tort ou ~ raison, le bruit se r6pandit que la direction de l'enseignement sup~rieur d~sirait r~cup~rer pour la France un math6maticien si 6minent et ne se refuserait pas ~ une cr6ation de chaire en sa faveur pour peu que la Sorbonne la demand~t. N'6tait-ce pas l'occasion pour ces Messieurs de r6parer leur erreur sans chagriner personne ? Si le minist~re n'avait pas les intentions qu'on lui pr~tait, du moins l'honneur de la Sorbonne serait sauf, ou presque. Mais non : c'6tait bien le talent trop distingu6 du candidat qui l'excluait. Nouveau vote, nouvel 6chec.
[1955e]
285 SCIENCE FRAN~AISE ?
II
E t le professeur B est toujours ~ l'Universit6 Columbia, qui s'eu f61icite et efit 6t6 bien en peine de le remplacer. Ainsi joue la loi de la cooptation des m6diocres, que je m'imaginais d6couvrir en 1937; loi d ' a u t a n t plus fatale qu'il faut ~ un h o m m e des qualit6s de premier ordre pour qu'il d6sire attirer aupr~s de lui ses 6gaux, au risque qu'ils lui soient sup6rieurs. U n h o m m e m6diocre, au contraire, cherchera toujours ~ s'entourer non pas seulement de m6diocres, rnais de plus m6diocres que lui ; il le faut bien, pour faire briller ses minces m6rites. Ce n'est pas d'hier que la plupart de IIos institutions scientifiques sont prises dans les rouages de ce m6canisme inexorable. La situation est-elle sans remade ? Peut-on imagiiier des r6formes qui en am~neraient le redressement ? Ce n'est pas douteux. I1 reste assez d'616ments sains dans le m o n d e scientifique fran~ais, il y a assez de talent parmi les jeunes pour permettre les plus s6rieux espoirs si on se d6cidait ~ faire le n6cessaire. J'eii ai assez dit pour faire comprendre qu'une telle r6forme ne p e u t partir que d'en haut. I1 y faudrait un acte d'autorit6 ; et elle se heurterait A la plus violente r6sistance de la part de la majorit6 des universitaires fran~ais, de l'Institut, du Coll~ge de France, de corps constitu6s et de personIIalit6s d o n t il est d'usage de ne parler en public que sur un t o n de profond respect. Peut-~tre, apr~s tout, n'y faudrait-il qu'encore un peu plus de courage que pour s'attaquer aux int6r~ts des viticulteurs ou au privilege des bouilleurs de cru. La politique, dit-on, est l'art du possible; oh, en pareille mati~re, est le possible ? Je ne suis pas politicien ; ce n'est pas m o n m6tier de le savoir. Rien de ce que je vais dire n'est impossible en soi, puisque t o u t cela se pratique sous nos yeux dans les pays qui sont ~ la t~te du m o u v e m e n t scientifique moderne. Sommes-iious encore capables de nous instruire leur exemple ? Je n'eii sais rien ; si nous ne le sommes pas, t a n t pis pour nous. Quelles sont donc ces r6formes qui pourraient nous tirer de la profonde orni~re oh nous sornmes ? I1 n'y a pas l~ grand myst~re ; tous ceux qui y ont quelque peu r6fl6chi sans pr6jug6 et de bonne foi saveiit bien ~ quoi s'en tenir l~-dessus. I1 suffira ici d'indiquer bri~vement quelques points essentiels.
286
[1955e1 I2
LA N O U V E L L E N . R . F .
D'abord, il faut s'attaquer ~ une organisation vicieuse, qui fait de l'Universit6 de France un monstre hydroc6phale, dont la Sorbonne est la t~te difforme et les universit6s de province sont les membres exsangues. Lors de la r6forme de Liard, il est notoire que celui-ci c6da ~ des pressions 61ectorales en acceptant beaucoup plus d'universit6s qu'il ne le jugeait souhaitable. I1 disait, paralt-il, que cela n'avait pas d'importance, parce que la p l u p a r t mourraient d'elles-m6mes. I1 n'avait pas pr6vu que les autorit6s locales, municipalit6s, chambres de commerce, fi~res du prestige qui en rejaillissait sur elles, leur accorderaient t o u t juste le soutien n6cessaire pour les faire subsister et en tirer quelques menus services, sans bien e n t e n d u leur donner les ressources qui en auraient fait de vrais centres intellectuels. LA oil par hasard se forme en province un noyau scientifique int6ressant, il v6g6te faute d'6tudiants ; les bons 6tudiants se dirigent sur Paris ot~ ils se t r o u v e n t noy6s dans la foule et ne p e u v e n t que rarement tirer profit de l'enseignement de maitres d6bord6s de tous c6t6s. M~me dans les quelques domaines off la France tient encore son rang, il ne saurait ~tre question de trouver assez de maitres et de chercheurs pour m o n t e r plus de quatre ou cinq grands centres. Done, la premiere r6forme dolt consister ~ rabaisser la p l u p a r t de nos universit6s au rang de centres prop6deutiques interm6diaires entre le secondaire et le sup6fieur; ~ cr6er, en province, environ quatre grands centres scientifiques bien dot6s en h o m m e s et en moyens, dans des localit6s bien choisies qui ne seraient pas n6cessairement des grandes villes ; ~ d6charget Paris de son trop-plein sur ces centres par des mesures appropri6es, dont le d6tail ne serait pas difficile formuler. E n second lieu, il faut changer radicalement le mode de nomination des professeurs. Le mieux serait de s'inspirer du syst~me anglais et de m e t t r e toutes les nominations importantes entre les mains de comit6s restreints offrant un m i n i m u m de garanties d'impartialit6 et de comp6tence, comit6s qm devraient obligatoirement (comme il se fait en Angleterre, avee d ' a u t a n t plus de soin qu'il s'agit d'une chaire plus importante) consulter largement l'opinion scientifique internationale et en tenir le plus grand compte. Dans ces comit6s devraient
[1955e]
287
SCIENCE FRAN~AISE ?
13
entrer pour une large p a r t des savants d6sign6s par le ministre et choisis eux-m~mes en t e n a n t compte de l'opinion internationale. Notons en passant qu'en Angleterre une visite de candidature serait suffisante pour disqualifier aussit6t un candidat. E n troisi~me lieu, mais en troisi6me lieu seulement, il faudrait donner, non aux universit6s actuelles, mais aux quatre ou cinq grands centres qu'il s'agit de constituer, non seumement des ressources, mais aussi une autonomie financi~re qui les m l t sur le m6me plan que les grandes universit6s anglaises, allemandes, am6ricaines, et que notre H a u t Commissariat de l'~nergie Atomique. Je ne veux pas me donner le ridicule ici de r6p6ter, apr~s t a n t d'autres qui le crient bien fort depuis vingt ans, que la science cofite chef. C'est vrai, encore qu'on ait pu assez souvent autrefois, et qu'on puisse peut-~tre encore (mais exceptionnellement) aujourd'hui faire avec des moyens modestes d ' i m p o r t a n t e s d6couvertes. Je ne veux pas rappeler les statistiques humiliantes qu'on a publi6es maintes reprises sur le b u d g e t de la recherche scientifique en France compar6 ~ celui qu'on y consacre ailleurs. ~cJe vous ferai de bonne chore, disait Maitre Jacques, si vous me donnez bien de l'argent. ~ I1 avait raison; et, q u a n d m~me il aurait 6t6 un fripon, cela n'aurait pas suffi ~ lui donner tort sur ce point. I1 faudra donc de l'argent, bieu de l'argent, pour les laboratoires, les biblioth~ques, le personnel subalterne. I1 faudra bien aussi se d6cider ~ payer d 6 c e m m e n t le personnel scientifique p r o p r e m e n t dit. I1 faudra permettre ~ nos grands centres scientifiques de recruter celui-ci par contrats individuels, comme le fait le H a u t Commissariat de l']~nergie Atomique et comme le font les grandes Universit6s 6trang~res. I1 faudra qu'on puisse, le cas 6ch6ant, n o m m e r ~ nos grandes chaires et ~ la direction de nos grands laboratoires des 6trangers qualifi6s. Les Anglais, d o n t les traditions scientifiques valent bien les n6tres, le font parfois et s'en t r o u v e n t bien ; la France l'a fait autrefois; pourquoi faut-il que notre amour-propre national en soit arriv6 ~ l'emporter sur notre int6r~t bien compris ? I1 faudra que les contrats que nos grands centres seront en mesure d'offrir leur p e r m e t t e n t d'entrer en concurrence, avec quelque chance de succ~s, avec les institutions similaires ~ l'6tranger.
288
[1955e ]
14
LA NOUVELLE N.R.F.
Bien entendu, lorsqu'on se sera r6solu ~ traiter conven a b l e m e n t nos savants, on sera en droit d ' a t t e n d r e d'eux qu'ils se consacrent honn~tement ~ leur enseig n e m e n t et ~ leurs recherches. Mais il ne sera pas besoin pour cela de r~glements draconiens. Ce n'est pas de gaiet6 de cceur que t a n t d'universitaires, chez nous, cherchent un suppl6ment h de maigres ressources dans des pratiques vari6es otl se consume leur t e m p s et leur 6nergie : cumul d'enseignements de bas 6tage, trust des examens et concours, et trop souvent raise ~ la disposition de l'ilidustrie priv6e de laboratoires officiellement consacr6s la science pure. E n Angleterre, en Am6riqlie, en Allemagne, l'industrie priv6e a ses propres laboratoires de recherche, souvent si largement con~us qu'il s'y fait de n o m b r e u x t r a v a u x scientifiques de grande valeur. E n France, les industriels, quand ils lie travaillent pas sur licences 6trang~res, t r o u v e n t trop souvent plus 6conomique de faire travailler ~ leur compte uli laboratoire universitaire en 6change d ' u n suppl6melit de t r a i t e m e n t d6risoire accord6 au professeur qui le dirige. Bien entendu, la liaison entre science pure et science appliqu6e est chose h a u t e m e n t souhaitable ou plut6t indispensable, mais qui ne s'obtient pas en 6touffant celle-l~ au profit de celle-ci par des arrangements qui constituent de v6ritables escroqueries aux d6pens de l']~tat. Assur6ment, bien d'autres questions se posent : recrut e m e n t des jeulies, r61e des c~grandes 6coles ,, liaison entre l'enseignement et la recherche. Je ne crois pas utile d'en discuter ici. Je ne pense pas qu'aucune d'elles puisse offrir de difficult6 s6rieuse dans un climat redeveliu favorable. Le redeviendra-t-il ? Se trouvera-t-il un chirurgieli pour m e t t r e sur la table d'op6ration un malade qui pr6t e n d qu'il ne s'est jamais mieux port6 ? S'il lie s'en rencontre pas, faut-il d6sesp6rer ? ou attendre le salut de l'ilitervention miraculeuse du g6nie ? ~Bien stir, disait m o n article de 193 7, le g6liie perce q u a n d m~me ; le g6nie se fait toujours sa place, ~ travers t o u s l e s obstacles; bien stir... (je n'en suis pas si stir que ~a). Oui, mais pour le g6nie m~me, que d'alin6es perdues ; quels retards, quelles sordides difficult6s ; et tous les autres, ceux qui auraient pu faire oeuvre utile, maintenir,
[1955e ]
289 SCIENCE
FRAN~AISE
15
?
en a t t e n d a n t la venue d u g6nie, une tradition honorable et parfois glorieuse, tous ces autres, quoi d ' e u x ? Souvent, ils s'aperqoivent des ann6es perdues ; un peu trop tard, ils se r e m e t t e n t ~ l'6cole ; ils t e n t e n t de se refaire une place dans la eolonne en marehe, q u a n d leur esprit a perdu sa souplesse et sa plastieit6 ; fls se hissent avec difficult6 ~ u n 6chelon off d ' a u t r e s a v a n t eux parvinrent, puis, l'effort fourni, ils y restent, ils sont d6pass6s. Ils y restent, et l'histoire recommence. Car voil~ Oil l'on se sent d6sesp6rer, le tragique cerele vicieux : l'histoire recommence... U n e fois provincialis6, une fois tomb6 darts l'orni6re, on y reste. Sauf miracle, bien stir ; ear l'esprit, c'est le m i r a c l e ; mais n ' y eomptons pas trop, ou plut6t, le miracle arrive ~ qui a u r a su le m6riter. ~ J e n'6tais gu6re optimiste en I937. J e ne le suis pas plus ~ pr6sent. ANDR]~ W E I L ,
Pro]esseur ~ l' Universitd de Chicago.
6159-x-I955. - - Imp. CR]~Tt~, C or b eil- Essonnes (S.-et-O.}.
[ 1956] The field of definition of a variety
Let V be a variety, defined over an overfield K of a groundfield k. Consider the following problems: ( P ) Among the varieties, birationally equivalent to V over K, find one which is defined over t~. ( P ' ) Among the varieties, birationally and biregularly equivalent to V over K, find one which is defined over ~. Froblems o~ these types arise, for instance, in Chow's recent work on abelian varieties over function-fields ( [ 1 ] ) , in my work on algebraic groups ( [ 4 ] ) , and also in the unpublished work of Shimura and of Taniyama on complex multiplication. Criteria for those problems to have a solution are implicitly contained in Chow's paper ( [ 1 ] ) and in Lang's subsequent note on a related subject ( [ 2 ] ) ; the purpose of the present paper is to develop them more explicitly. Without restricting the generality of the problem, we may assume that K is finitely generated over /c; we shall make the restrictive assumption that it is separable over /c. Then it is a regular extension of the algebraic closure /c' of k in K ; and ]c' is a separably algebraic extension of /~ of finite degree. Thus it will be enough for our purpose to discuss ( P ) and ( P ' ) , firstly when K is separably algebraic over k, and secondly when it is regular over k. As usual, we do not distinguish between mappings and their graphs. I n particular, we do not distinguish between a birational correspondence T between two varieties V and W (this being defined as a subvariety of V X W which satisfies certain conditions) and the m a p p i n g of V into W determined by T. The inverse mapping, T -~, is then a m a p p i n g of W into V or also a birational correspondence between W and V. To prevent misunderstandings, I take this opportunity for pointing out that (by abuse of language) I call T everywhere bireguMr only when T is biregular at every point of V and T -~ is so st every point of W; when that is so, T might more suitably be ~*Received January 16, 1956. 509 Reprinted from the Am. J. of Math. 78, 1956, pp. 509-524, by permission of the editors, 9 1956 The Johns Hopkins University Press.
291
292
[1956] 510
A N DR 1:~' W E L L .
called an i.somorphi~m between V and IV, and a ~C-isomorphism if /C is a ticld of definition for V, IV and T.
Section I. Separably Algebraic Extensions of the Groundfield. 1. Let k be a separably algebraic extension of a groundfield he, of tlnite degree n. Call ~ the set of all distinct isomorphisms of /C (over /Co, i.e., leaving all elements of tee invariant) into the algebraic closure go of Ice; ~t consists of n distinct isomorphisms, including the identity automorphisnl of /c. I f a c ~ , and if oJ is any isomorphism over leo of an overfield of /C~, we denote by ~o~ the isomorphism of k defined by putting ~ ( ~ ) ~ for every ~ c k. Let V be a variety, defined over /C; assume that there is a variety Vo, defined over he, and a birational correspondence f, defined over k, between Vo and V. Then, for a e ~ t , r c ~ , the mapping f ~ . ~ f f o ( f r -1 is a birational correspondence between V r and V ~. We now modify our problem ( P ) as follows : (A) Let k be a separably algebraic extension of /co of finite degree; let ~ be the set of all distinct isomorphisms of k into ~o. Let V be a variety,
defined over k; for each pair (a,'r) of elements of ,~, let f~.~ be a birational correspondence between V ~ and V ~. Find a variety Vo, defined over he, and a birational correspondence f, defined over k, between Vo and V, such that f ~ . a ~ f~o (f~)-~ for all ac ~, r c ~. Tm~oREM 1. Problem (A) has a solution if and only if the f ~ are defined over a separably algebraic extensiol~ of /co atgd satisfy the following conditions : (i)
f , , p = f ~ . , , o f , , p for all p, ~, r i~t ~ ;
(ii)
f ...... ~ (f,.~)~ for all a, r in ~ and all automorphisms r of ko
over ~o.
Moreover, wheat that is so, the solution is unique, up ~o a birational transformation oft Vo, defined over ~co. The conditions are obviously necessary. I f the problem has two solutions this is a birational correspondence between Vo and Vo', defined over /C. Writing that (Vo, f) and (Vo',f') are solutions of (A), we find F ~ F ~ for all ~, r; thus, F is invariant under all automorphisms of ~o over /Co; therefore it is defined over /Co; this proves the unicity assertion in Theorem 1.
(Vo, f) and (Vo',f'), put F ~ f ' - ~ o f ;
[19561
293 TI-IE F I E L D
O]~~ D E F I N I T I O N
OF
A VARIETY.
511
Now assume that (i), (if) are fulfilled; then, if r r are in 3 and is an automorphism of ~o over the compositum of k ~ and k ~, (if) shows that f,.r is invariant under ~; if f~.r is defined over a separably algebraic extension of ko, this implies that it is defined over the compositum of k ~ and k ~. Let x be a generic point of V over k ; for each ~ e $t, put xo f,~.,(x), c being the identity automorphism of k. If we put p ~ r in (i), we see that fr162is the identity mapping of Vr therefore we have x~ ~ x. Let K be the compositum of the fields/cG for all a e ,~ ; this is a Galois extension of ko; call P its Galois group. Take any ~ e P ; as x and xe~ are generic points of V and V% respectively, over K, there is one and only one isomorphism o)* of K(x) onto K(x~) which induces ~o on K and maps x onto x~. As f .... is a birational correspondence, defined over K, we have K(x) ~K(x~,~), so that ~* is an automorphism of K(x). Applying (if) to anv extension of ~ to an automorphism of ~o, we get:
putting p : r and a : e o ~ in (i), we find that the right-hand side of this relation is x ~ ; therefore ~* maps x~ onto x ~ for all ~. From this it immediately follows that the mapping o~-->o~* is a homomorphism of P into the group of all automorphisms of K(x), and more precisely an isomorphism of P onto a group F* of automorphisms of K(x). Call ko(y) the field consisting of those elements of K(x) which are invariant under r * ; it is finitely generated over/~o ( [ 4 ] , App., Prop. 3). As K(x) is regular, hence separable, over K, and K is separable over ko, K(x) is separable over ko (Bourbaki, Alg., Chap. V, w 7, no. 4, Prop. 7) ; hence ko(y) is separable over ko. Any element of the algebraic closure of ko in ko(y) must be in the algebraic closure of K in K(x), which is K since K(x) is regular over K ; as such an element is invariant under P*, it must then be invariant under P, and so it must be in ko. Thus we have proved that k,)(y) is regular over ko. Call Vo the locus of y over ko. I f an element ~ of F induces the identity on k, o~* leaves x invariant; as P* is the Galois group of K(x) over ko(y), this implies that k(x) C k(y), so that we may write x ~ f ( y ) , where f is a mapping of Vo into V, defined over k. We have K(x) CK(y), hence K ( x ) ~ K ( y ) since K(y) is contained in K(x) by definition. This shows that f is a birational correspondence between Vo and V. Transforming the relation x ~ f ( y ) by any o,* in P*, and calling a the isomorphism of k induced on k by ~*, we get x,,~f~(y), hence f,,,=fr by (i), this shows that (Vo, f) is a solution of our problem.
294
[1956] 51o~
ANDRI~ WEIL.
2. I n the introduction, we formulated, in addition to the problem ( P ) , a more precise problem ( P ' ) . We may modify the problem (A) similarly, by requiring f to be everywhere biregular; call (A p) this modified problem. For (A') to have a solution, it is obviously necessary that the f , ~ should be everywhere biregular and satisfy the conditions in Theorem 1. Assume that this is so; let (Vo, f) be a solution of (A). Then, if (Vo',f') is a solution of (A'), the unicity of'the solution of (A) shows that we must have fr ~ f o F -1, where F is a birational correspondence between Vo and Vop, defined over/co. Thus problem (A') may be reformulated as follows. (B) Let k and ko be as in problem ( A ) ; let V and Vo be varieties, respectively defined over k and over ko; let f be a birational correspondence, defined over k, between Vo and V. Find a variety Vo' and a birational correspondence F between Vo and Vo', both defined over ko, such that the birational correspondence f o F ~ between Vo' and V ks everywhere biregular. I t is obvious that, if (B) has a solution, this is unique up to a koisomorphism; therefore the'same is true for (A'). I f (B) has a solution, then (,~ being defined as before) the birational correspondence f - o (f~)-i between V ~ and V ~ must be everywhere biregular for all a, r in ~. We will prove that this condition is also sufficient, at any rate if V is a k-open subset of a projective variety, defined over k. This will be an immediate consequence of the following result. PR01"OSlTIO.'r 1. Let k, ko, ,.~ be as in (A). Let Vo be a variety, defined over ko; let V be a projective (resp. a~ne) variety, defined over Is; let f be a birational correspondence, defined over Is, between Vo and V. Then there is a projective (resp. affine) variety W and a birational correspondence F between Vo and W, both defined over 1% such that Fof-~ is biregular at every point of V where the mappings f~ o f-~ are defined for all a e ,~. Let S be the ambient space of V, projective or affine; f may be regarded asia mapping of Vo into S. Call a l ~ , a 2 , " 9 ",an the elements of ,~, and put F~ ~ (f%. 9 -, f~,,) ; this is a mapping of Vo into the product S )< 999)< S of n factors equal to S, and is defined over the compositum K of the fields k ~. It is clear that F~ o f-~ is defined wherever all the f~ o f-~ are defined. Let x be a generic point of Vo over /Co; let W1 be the locus of Fz(x) over K ; put u = F I ( ~ ?) = (Xl,"" ',X~Z). AS ffl=tE, we have x ~ f ( x ) , so that the image of u by the mapping f o F~ -~ is x~; this shows that f o F~ -~ is the mapping induced on W~ by the projection of the product S X" " ")< S onto its first factor, and is therefore everywhere defined. Thus the birational correspon-
[1956]
295 TIIE
FIELD
OF
DEFINITION
OF
A
VARIETY.
513
dence F1 o f-~ between V and W1 is biregular wherever all the f~of-~ are defined. Let zl," ' ' , z ~ be ~ points of S ; if S is the projective m-space, p u t z ~ (z~o,- 9 ", z~,~) ; and let z' be the point, in a projective space of suitable dimension, whose homogeneous coordinates are all the monomials z~,lz2~,. 99 z,,~, with 0 ~ m ~ m for every i. I f 8 is the affine m-space, put zt ~ (z~l,. - . , z ~ ) , put Z~o~ 1 for 1 --~ i _< n, and let z' be the point, in an affme space of suitable dimension, whose coordinates are the same monomials as before. I n either case, put z ' ~ , t , ( z , . . . , z ~ ) ; it is well-known t h a t ~ is an everywhere biregular m a p p i n g of S X " ' " X S onto its image in projective (resp. affine) space. P u t now F2 ~ o F~; then F_. is a birational correspondence between Vo and W2 ~ r and F 2 o f - 1 is biregular wherever all the f,rof-~ are defined. i f S is projective, let (1, f ~ ( x ) , . . . , f , , ( x ) ) be a set of homogeneous coordinates for f ( x ) ; the f~ are functions on Vo~ defined over b. P u t fo ~ 1. Then we have F , ~ (go," 9 ",g~), where the gp are all the monomials
I f ~ is an automorphism of K over ~, gp~ is again one of the gp, which we m a y write as g~(p); the m a p p i n g ~ - - > ~ ( p ) determines a representation of r (the Galois group of K over k) as a group of permutations on the gp. F o r a given p, let -/p be the subgroup of r determined by co(p) ~ p ; then, for o e F, o)(p) takes a number of distinct values equal to the index dp of 7p in r . I f Kp is the subfield of K consisting of the elements of K invariant under ~,p, gp is defined over Kp ; therefore, if ( ~ , " 9 ", ~ p ) is a basis of Kp over bo, we may write gp ~ ~ ~hp~, where the hp~, for 1 ~ v <--dp, are functions on Vo, 1s
defined over /co . Then we have, for all r e P : y
If, in this relation, we take for (o a set of representatives of the dp cosets of ),p in F, we get a linear substitution expressing the dp distinct functions g~(p) in terms of the dp functions h ~ ; and, since K~ is separable over /co, t h a t substitution is invertible. F r o m this it follows immediately that, if we call F ( x ) the point whose homogeneous coordinates are all the functions h ~ (where p runs through a set of representatives for the classes of equivalence determined by the p e r m u t a t i o n group r on the set {0, 1," 9 ",r}, and where, for each p, we take l ~ - - ~ d p ) , F is of the form ~I, oF2, where 9 is an automorphism of the ambient projective space of W2. I f S is affine, we p u t
296
[1956] 514
ANDRs WEIL.
f ~ ( f ~ , ' ' ",f,~), f o ~ l , and we define F by the same formulas as in the projective case, but regard it as a mapping of Vo into an affine space; then we have again F ~ ~I, oF2, ~I, being now an automorphism of the ambient afflne space of W2. In either case, the mapping F is defined over ko; if W is the locus of F ( x ) over ko, W and F have the properties required by our proposition. S. Before applying this to the problems {A') and (B), we need a general lemma : LEMMA ]. Let f be a birational correspondence between two varieties U and V ; let k be a field of definition for U, V, f. Then the sets of points where f and f-1 are respectively biregular are k-open, and f determines a kisomorphism between them. Call U' the set of points of U where f is defined, U'" the set of points of U where it is biregular; call V" the set of points of V where f-1 is defined, W' the set where it is biregular. By [4], App., Prop. 8, U' and V" are/c-open. Call f' the restriction of f to U ' X V' (i.e. the birational correspondence between U' and V' whose graph is the set-theoretic intersection of the graph of f with U ' } < V ' ) . If f is biregular at a point a of U, it is defined at a, so that a e U ' ; and, if we put b = f ( a ) , f-~ is defined at b, so that b e V ' ; therefore (a, b) is on the graph of f', and f' is defined at a. Conversely, let a be a point of U' where f' is defined; put b = f ' ( a ) ; then b is in V', so that f-~ is defined at b; as U' and V' are open, f is then defined at a, and we have b = f ( a ) ; thus f is biregular at a. This shows that U" is the set of points of U' where f' is defined; similarly V " is the set of points of V' where f'-~ is defined; this implies that they are/c-open. If f " is the restriction of f to U" }( V", it is everywhere biregular by definition (i. e., f " is biregular at every point of U", and f"-~ is so at every point of V"). In order to formulate our results on problems (A') and (B), we will say that a variety U, defined over a field k, is projectively (resp. a~nely) embeddable over k if it is k-isomorphic to a k-open subset of a projective (resp. affine) variety, defined over /C. TtlEOREM 2. Problem (B) has a solution (Vo',F') provided V i.~ projectively embeddable over/c and the birational correspondence f~ o f ~ between V and V ~ is everywhere biregular for every isomorphism cr of k over ~'o into #o. When that is so, Vo' is projectively embeddable over It(;; it is a~n~ly embeddable over k o if V i~v so over k.
[1956]
297 THE
FIELD
OF D E F I N I T I O N
OF A V A R I E T Y .
515
We may assume V to be a /c-open subset of a projective (resp. affine) variety, defined over /C. Take W and F as in Proposition 1; then F o f - ~ is biregular at every point of V; therefore, by Lemma 1, it is a /c-isomorphism between V and the /c-open subset Vo' of W where f o F -~ is biregular. As the f~ o f-~ are everywhere biregular Vo' is also the subset of W where f~ o F -~ is biregular, for every (r; therefore it is invariant under all automorphisms of ~o over /co, so that it is /co-open, by [4], App., Prop. 9. Then, if F ' is the restriction of F to Vo )< Vo', (Vo',F') is a solution of (B). TI~EORE~ 3. Problem (A') has a solution, i.e., problem (A) has a ,~'olution (Vo, f) for which f is everywhere biregula~, provided V is pro]eetively ernbeddable over /C and the f~., are everywhere biregular and satisfy the conditions in Theorem 1. When that is so, Vo is projectively embeddable over/co; it is a]~nely ernbeddable over /co if V is so over /C. The solution is unique up to a ~co-iSomorphism.
Section II. Regular Extensions of the Groundfield. 4. Let l~ow /,~ denote the groundfield. Let T be a variety, defined over /C; let t be a generic point of T over k. When we denote by Vt a variety, defined over k ( t ) , we will agree, whenever t' is also a generic point of T over /c, to denote by Vt, the transform of Vt by the isomorphism of k ( t ) onto k(t') over /C which maps t onto t'. Similarly, if a mapping, defined over k(t)~ is denoted by re, ft' will denote its transform by the same isomorphism; if t, t', t" are three independent generic points of T over k, and ft',t is a mapping, defined over /c(t, t'), we denote by ft",e' the transform of ft'.t by the isomorphism of /c(t,t') onto /c(t',t") over /C which maps ( t , t ' ) onto
(t', t") ; etc. Let Vt be a variety, defined over k ( t ) ; assume that there is a variety V, defined over /C, and a birational correspondence ft, defined over k ( t ) , between V and Vt; then ft' oft -1 is a birational correspondence between Vt and Vv. We therefore modify problem ( P ) of the introduction as follows: (C) Let T be a variety, defined over cb field k , let t, t' be independent generic points of T over k. Let Vt be a variety, defined over k ( t ) ; let ft',t be a birational correspondence, defined over /c(t,t'), between Vt and Vt,. Find a variety V, defined over /C, and a birational correspondence ft, defined over /c(t), between V and Vt, such that ft, t ~ f t , oft -1. THEOREM 4.
the condition :
Problem (C) has a solution if and only if ft'.t satisfies
298
[19561 516
(i)
ANDRI%' WEIL.
:,,,.,~f,,,,,,oft,.,
where t'" is a generic point of T over l~(t,t'). When that is so, the solution is unique, up to a birational transformation on V, defined over to. The condition is obviously necessary. The proof for the unicity of the solution, when one exists, is quite similar to the proof of the corresponding statement in Theorem 1. Now, assuming (i) to be fulfilled, we shall construct a solution of ( C ) . We may replace T by any birational transform of T over k, and so we may assume that T is an affine variety. S i m i l a r l y we m a y assume that Vt is an aftine variety; and, taking x to be a generic point of Vt over k ( t ) , we may replace x by ( x , t ) and Vt by the locus of (x,t) over / c ( t ) ; after that is done, Vt is still an affine variety, and we have k(t) C / r from now on, assume that this is so, and assume that x has been taken generic oi1 Vt over k ( t , t', t"). By [4], App., Prop. 1, /c(x) is a regular extension of /~; call X the locus of x over k. P u t X'~ft, t(x) ; this is a generic point of Vt, over k(t, t', t " ) ; by the definition of Vt,, this implies that there is an isomorphism of k ( t , x) onto k(t',x') over /c, m a p p i n g t onto t' and x onto x ' ; therefore we have lc(t') C lc(x'), hence k(x,t') C k(x,x'). As the definition of x' shows k(x,x') to he contained in k(t',t,x), i.e. in k(x, t'), it follows that we have k(x,x') ~ k(x, t') ; therefore x' has a locus W, over k ( x ) . Let k ( v ) be the smallest field of definition containing/~ for W , ; as k(v) C k(x), tc(v) is a regular extension of k. Call V the locus of v over k; we may write v ~ G(x), where G is a m a p p i n g of X into V, defined over /c. I f we put x"=ft.,.t(x), W, is also the locus of x" over /c(x) ; as the fields lr x') and k,(x,x") are respectively the same as lc(x, t') and k(x, t") and are therefore algebraically i n d e p e n d e n t over k(x), W, is also the locus of ~J' over k(x,x'). But (i) may be written XtP~ft,.t'(XI); therefore W,. is the same as W~. This implies that the isomorphism of It(x) onto k(x') over k which maps x onto x' leaves invariant all the elements of the smallest field of definition of W~, hence also all the elements of k ( v ) , so that we have
G(x) = G ( x ' ) . On the other hand, let K be an overfield of t% algebraically independent from k(x,x') over /c; if q~ is any function on X, defined over K, it will induce on W~ a function which is defined over K ( v ) ; if O(x) ~ q ~ ( x ' ) , that function is a constant, so that its constant value must be in K(v). This shows that K ( v ) is the subfield of K ( x ) consisting of the elements of K ( x ) which are invariant under the isomorphism of K ( x ) onto K(x') over K m a p p i n g x onto x'. Now the relation S ' ~ ft",t (X) shows that x" is rational over ~:(t", t, x),
[1956]
299 THE
FIELD
OF I ) E F 1 N I T I O I ' q OF A V A R I E T Y .
517
i. e. over k ( t " , x), so t h a t we may write x " ~ 4t,,(x), where (bt" is a m a p p i n g of X into Vt,,, defined over /c(t"). The relation x " ~ f t , , t , ( x p) m a y then be written as x " ~ S~t,,(x'). A p p l y i n g to the field K ~ lc(t") and to the function 4~r' what we have proved above, we conclude from this t h a t /c(x") C k(t", v). As we have G ( x ) = G(x'), hence also G ( x ) ~ G ( x " ) , the isomorphism of k ( x ) onto /c(x") over /c which maps x onto x" leaves v ~ G(x) i n v a r i a n t ; applying the inverse of that isomorphism to the relation ~ ( x " ) C k ( t " , v ) , we get lc(x) C k ( t , v ) , hence k ( x ) ~ l c ( t , v ) since lc(t) and lc(v) are both contained in /c(x). Also, since k ( x ) and k ( t ' ) are algebraically independent over lc, the same is true of /c(v) and k ( t ' ) ; as the isomorphism of ts(x') onto k ( x ) over k which maps x' onto x maps t' onto t and v onto itself, this implies that /c(v) and /c(t) are algebraically independent over k. As the relation lc (x) ~ / s (t, v) can also be written lc (t, x) ~ lc (t, v), we conclude t h a t Vt and V are birationally equivalent over /c(t), so t h a t we may write x ~ f t ( v ) , where ft is a birational correspondence between V and Vt, defined over /e(t). Then we have x ' ~ f t , ( v ) . Therefore (V, ft) is a solution of our problem. We also see that X is birationa]ly equivalent to T }( V over /c. 5. J u s t as in Section 1, we consider the problem (C') which consists in finding a solution (V, ft) of (C) such t h a t ft is everywhere biregular. F o r such a solution to exist, it is necessary that ft'.t should be everywhere biregular; it will be shown that this is sufficient. As in Section I, if we make use of Theorem 4, we see t h a t (C') may be reformulated as follows:
Let ]c, T and t be as in (C) ; let V and Vt be varieties, respectively defined over k and over /c(t) ; let ft be a birational correspondence, defined over k ( t ) , between V and Vt. Find a variety V" and a birational correspondet~ce F between V and V', both defined over t~, such that the birational correspondence ft o F 1 between V" and Vt is everywhere biregular. (D)
I n order to solve ( D ) , we need some p r e l i m i n a r y results. LEMMA 2. Let F and H be mappings of a variety X into two varieties W, T, all the~'e bei~tg defined over a field /c; x being a generic point of X over Is, assume that t ~ H ( x ) is generic over 7c on 7' and float x has a locus Vt over k ( t ) . Let Ft be the mapping of Vt into W induced by F on Vr. Then F is defined (~t every point of Vt where Ft and H are both defined. I t is clearly enough to treat the case in which X is an affiue variety and W i-s the affine line. Then Ft is the function on Vt, defined over ~ ( t ) ,
300
[1956] 518
ANDRI~ WEIL.
such that F t ( x ) ~ F ( x ) . I f F t is defined at a point a of Vt, we can write it as F t ( x ) ~ P ( x ) / Q ( x ) , where P, Q are polynomials with coefficients in k ( t ) , such that Q(a)~=0. More explicitly, we have
P ( X ) = 2~ A , ( t ) M , ( X ) ,
Q ( X ) -~- ] ~ t q ( t ) N j ( X ) ,
where tile h,, /~1 are functions on T, defined over k, and the Mi, N i are monomials in the indeterminates ( X ) ; and we have
(1)
Xt~(t)Nj(a) # o. J
Then we have F ( x ) = , ~ ( x ) / q z ( x ) ,
(2)
o(x) ~XXdH(x))M,(x), *
with
*(x) = X , j ( H ( x ) ) N s ( x ) . f
As a is on Vt, (t,a) is a specialization of ( t , x ) over k ; if H is defined at a, we must have H ( a ) ~ t. As t is generic on T over k, the functions A,, t~j are defined at t ; therefore the functions ~,, o H , / ~ j o H are defined at a on X, with the values ) q ( t ) , /~j(t). That being so, the relations (1), (2) show that F is defined at a on V. PROPOSITION 2. Let k, T, t, t" be as in (C) ; let V be a variety, defined over k; let Vt be a variety, defined and projectively (resp. affinely) embeddable over k ( t ) ; let ft be a birational correspondence, defined "over k ( t ) , between, V and Vt. Then: (i) if a is a point of Vt where ft, oft -1 is biregular, there is an affine variety W and a birational correspondence F between V and W, both defined over k, such that F oft -1 is biregular at a; (ii) if ft, oft -~ is everywhere biregular, there is a variety W, defined and projectively (resp. affinely) embeddable over k, and a birational correspondence F between V and W, defined over k, such that F o ft -1 is everywhere biregular. We may assume that Vt is a k ( t ) - o p e n subset of a variety, defined over k ( t ) , in a projective (resp. affine) space S. We m a y also assume that T is a projective (resp. affine) v a r i e t y ; let S ' be its ambient space. I f S, S' are affiue, S X S' is an affme space ; if they are projective, call 9 the well-known biregular embedding of S X S ' into a projective space S " of suitable dimension. L e t v be generic on V over k(t, t'), and p u t x ~ f t ( v ) . We may replace Vt by a suitable k ( t ) - o p e n subset of the locus of ( x , t ) over k ( t ) in the affme case, of r over k ( t ) in the projective case; after t h a t is done,
[1956]
301 TIfE
FIELD
OF DEFINITION
OF A VARIETY.
519
we have /c(t) C k ( x ) , and therefore k(x) ~ k ( t , x ) ~ k ( t , v ) , so that x has a locus X over k, birationally equivalent to T X V, and that we may write t ~ H ( x ) , where H is a mapping of X into T, defined over /~; moreover, the mapping H is everywhere defined on X. Now, since X is birationally equivalent to T X V over lc, and V is birationally equivalent to Vt over lc(t), X is birationally equivalent to T X Vt over k(t). More explicitly, if we put x ' ~ f t , ( v ) , x' is generic over k ( t ) on X, and we have k ( x ' ) = l c ( t ' , v ) , hence k ( t , x ' ) ~ l c ( t , t ' , x ) , so that we may write x r ~ gt (t', x), where g t is a birational correspondence, defined over k ( l ) , between T X Vt and X. We have t ' ~ H ( x ' ) , and we may write x ~ c k t ( x ' ) , where qSt is a mapping of X into Vt, defined over k ( t ) ; then (H, q6t) is the mapping of X into T X Vt, inverse to gt. The mapping gt induces on the subvariety t ' X Vt of T X Vt the mapping (t',x)----> x ' ~ f t , ( f t - l ( x ) ) ; and qst induces on Vt, the mapping x'---> x, i. e. the mapping ft o fr -1. Applying L e m m a 2, we see that gt is defined at ( t ' , a ) whenever a is a point of Vt where ft' oft -~ is defined, and that ~bt is defined at every point of Vt, where ft oft, -~ is defined. Therefore gt is biregular at (if, a) whenever a is a point of Vt where fr oft -~ is biregular. Now let Ao be the k(t)-closed subset of T X Vt where gt is not biregular ; and assume first that a is a point of Vt with the property stated in (i). Then (t'~ a) is not in Ao, so that T >( a is not contained in Ao ; let Ax be the (non-dense) k(t, a)-closed subset of T consisting of those points t~ such that ( t , a ) eAo. By [4], App., Prop. 12, there is a ~-closed subset A2 of T containing all ~-closed subsets of T contained in A1; in particular, every point of A~ which is algebraic over k must be in A~. Let As be the union of the components of A2 and of their conjugates over /r put T ' = T - - A 3 ; this is a k-open subset of T such that, if tl is any algebraic point over k in T', gt is biregular at (tx, a). On the other hand, assume, as in (ii), that ft, oft -~ is everywhere biregular. Then gt is biregular at every point of t' X Vt, so that Ao has no point in common with t' X Vt. This implies that the projection of Ao on T is non-dense in T, so that, if we call A~' the closure of that projection, it is a (non-dense) /c(t)-closed subset of T. Let A2' be the maximal J~-closed subset of T contained in A~'; let Aa' be the union of the components of A~' and of their conjugates over k ; put T " ~ T - - AJ. Then T " is/c-open on T ; and, if t~ is any algebraic point over /~ in T", gt is biregular at every point of t~ X Vt. Now let t~ be a separably algebraic point over k in T' (resp. T " ) ; if k is finite, we may take for t~ any algebraic point over 1c in T ' (resp. T " ) ,
302
[1956] 520
ANDRI~
WEIL.
since in that case every algebraic extension of k is separable; if k is infinite, we apply [4], App., Prop. 13. Let ti," 9 ", t~ be the distinct conjugates of tl over k. As they are in T' (resp. T ' ) , gt is biregular at (h,a) (resp. at every point of h X Vt), and a fortiori at (h, x), for 1--~ i -~- ~; therefore it induces on t~ X Vt a birational correspondence g~ between Vt and the locus V~ of the point g ~ ( x ) ~ g t ( h , x ) over l~(t,t~) in the projective (resp. affine) ambient space of X ; and g~ is biregular at a (resp. at every point of Vt). But, as we have already observed, the relation/~ (x) ~ k (t, v) shows that X is birationally equivalent to T X V; we may write x ~ f ( t , v), where f is a birational eorrespondence between T X V and X, defined over ]c; then we have x ' ~ f ( t ' , v ) ; and f is the product of Yt and of the birational correspondence (t', v) --> (t', x) between T X V and T X Vt. As the latter correspondence is biregular at (h, v), and gt is biregular at (t~, x), for I ~ i ~ n, we see that f is biregular at (h, v), and that we have
g~(~) = g, ( t,, ~) = i ( h, ~). As the point f ( G v ) has the same locus over ]~(h) as over k(t, h), this shows that Vt is defined over k ( h ) . As every automorphism of ~ over k can be extended to an automorphism of ~(v) over k ( v ) , this also shows that V~ is the transform of V~ by the isomorphism of ]a(t~) onto ]~(h) over k which maps t~ onto h. Also, if f~ is the mapping of V into V~, defined over k ( h ) , which is such that f ~ ( v ) ~ f ( h , v ) , we have f ~ g ~ o f t ; and f~ is the tran~form of f~ by the isomorphism of ]~(t~) onto ~ ( h ) over ]a which maps t~ onto t~. Now apply Proposition 1 to the variety V, defined over the groundfield k, to the variety V~, defined over k(t~), and to the birational correspondence f~ ; this gives a projective (resp. afflne) variety W and u birational eorrespondence F between V and IF, both defined over k, such that F o f~-~ is biregular wherever all the f~ o f - 1 are defined, i.e. wherever all the gi o gt -~ are defined, Now, in ease (i), all the g, are biregular at a, so that all the g~og~-~ are biregular at the point gJa) ; therefore F o ft-% which is the same as (F o f~-~) o gl, is biregular at a; as this involves merely a local property of W at the image of a by that mapping, we may replace W, in the projective ease, by one of its afflne representatives. Thus we have solved our problem in ease (i). I n ease (ii), g, is biregular at every point of Vt; as we have just shown, this implies that F o f t -1 is biregular at every point of Vt, so that it determines an isomorphism of Vt onto a k(t)-open subset W' of W. The assumption in (ii) implies that 1'11' is invariant under the isomorphism of ]~(t) onto k ( t ' ) over /~ which maps t onto t'. From this and from [4], App., Prop. 9, it follows easily that W' is k-open; thus (IV', F) is a solution of our problem.
[19561
303 T:HE F I E U D OF D E F I N I T I O N
OF A V A R I E T Y .
5~1
COROLLARY. Let t~, T, t and t' be as in (C) ; let V be a variety, defined over lc; let Vt be a variety, defined over k ( t ) ; let ft be a birational correspondence between V and Vt, defined over k ( t ) and such that ft, oft -~ is everywhere biregular . Then, if a is any point of Vt, there is an alfine variety W and a birational correspondence F between V and W, both defined over k, such that F o f t -~ is biregular at a. We may assume that t" has been taken generic on T over k ( t , a ) ; take t" generic on T over k ( t , t ' , a ) . Call a', a" the images of a by ft, oft -~ and by ft" o ft-% respectively. The isomorphism of /~(t, a, t') onto k(t, a, t") over /~,(t,l~) which maps t' onto t" maps a" onto a"; therefore, if Vt,~ is a representative of the (abstract) variety Vt, on which a' has a representative a,', the point a" of Vt,, has a representative a,~" on Vt,,a. Let ft', be the hirational correspondence between V and Vt,~ which is determined by ft'. As ft,, o ft, -~ is everywhere biregular and maps a' onto a", ft"a oft,, -~ is biregular at a~'. Applying Proposition 2(i) to V, Vt,~ and ft'~, we get a solution ( W , F ) of our problem. 6.
~ow we can deal with problems (D) and (C').
THEOREM 5. Problem (D) has a solution if and only if ft, oft -1 is everywhere biregular for t' generic over It(t) on T. The condition being obviously necessary, assume that it is fulfilled. By the corollary of Proposition 2, there is, to every point a of Vt, an affine variety W~ and a birational correspondence Fa between V and We, both defined over /c, such that F~oft-~ is biregular at a; call ~ the /~(t)-open subset of Vt where F~ oft -~ is biregular, and call Wa' its image on Wa by F~ o ft-% which is a /~(t)-open subset of Wa. Then WJ is the subset of W~ where ft o F~-~ is biregular ; as in the proof of Proposition 2, this implies that Wa" is invariant under the isomorphism of /~(t) onto /~(t') over ~ which maps t onto t', and we again conclude from this that W~' is /c-open. As we have a ~ ~2~ for every a e Vt, the open sets r form a covering of Vt; by the well-known " c o m p a c t o i d " property of open sets in the Zariski topology, there must be finitely many points as on V such that the sets ~a~ cover Vt. Then the k-open subsets W~' of the afflne varieties W~o, together with the birational correspondences Fap o F~. -~ between them, define an abstract variety, which, together with the obvious birational correspondence between it and V, solves our problem. THEOREI~[ 6.
Problem (C') has a solution, i.e., problem (C) has a
304
[1956] 522
ANDRs WEIL.
solution (V, ft) for which ft is e~,erywhere biregular, if and only if ft',t is everywhere biregular and satisfies eondHiott (i) in Theorem 4. The solution 'is unique up to a k-isomorpl~is.m. This is an immediate consequence of Theorems 4 and 5. 7.
As to the projective or affine embeddability of the solution of prob-
lems (D) and (C'), we have the following result. THEOREM 7.
Let V be a variety, defined over a field k, and projectively embeddable over an overfield K of k. Then V is projectively ( resp. a]finely ) (resp. a~nely) embeddable over lc provided (i) K is separable over k or (it) 1; is everywhere normal with reference to 1,'. The assumption means that there is a birational correspondence f, defined over K and biregular at every point of V, between V and a subvariety of a projective (resp. affine) space; if we regard f as a mapping of V into that space, it has a smallest field of definition ~' containing k; we may replace K by k ' ; after that is done, K is finitely generated over k. I f K is separable over k, it is a regular extension k l ( t ) of the algebraic closure kl of k in K, and kl is a separably algebraic extension of k of finite degree. Proposition 2(it) shows that V is then projectively (resp. affinely) embeddable over k~; by Theorem 2, this implies that the same is true over ~; this completes the proof in case (i). I f K is not separable over k, let ]c* be the union of the fields kP-'~, for n = 1 , 2 , . . . ; then the compositum K* of K and k* is separable over k*, so that, by what we have just proved, V is projectively (resp. affinely) embeddable over k*. In order to deal with case (it), it is therefore enough to prove our theorem in the case in which V is everywhere normal with reference to k, and K is purely inseparable over ]~; I owe the proof for this to T. Matsusaka; it is as follows. We may again assume that K is finitely generated over /c; as it is purely inseparable, it is contained in some field ] d ~ lWq, where q is a power of the characteristic. Then there is a mapping f' of V into a projective (resp. affine) space, defined over Z", such that f' determines a birational correspondence, biregular at every point of V, between V and the closure W' of its image by f'; then W' is a projective (resp. affine) variety, defined over .~', and f' determines a /d-isomorphism between V and a It'-open subset of W'. Call the automorphism ~---)~q of the mfiversal domain; put W ~ W'~; W is then a projective (resp. affine) variety, defined over k. Let x be a generic point of V over /c; then W' is the locus of the point y ' ~ f f ( x ) over /d, and W is
[1956]
305 TItE
FIELD
OF D E F I N I T I O N
OF A V A R I E T Y .
523
the locus of the point y~y"~ over k. As y' is rational over k'(x), y is so over k ( x ' ) ; we may write y ~ g ( x ) , where g is a mapping of V into W, defined over k; as we have k ' ( y ' ) ~ / c ' ( x ) , we have k ( y ) ~ I c ( x ' ) , which implies that k(x) is purely inseparable over k(y). I n the projective case, let U be the projective variety derived from W by normalization in the field k ( x ) ~ ; U is birationally equivalent to V over k; let z be the point of U which corresponds t~ x on V. I n the affine case, we take for z a point in a suitable affinc space such that k[z] is the integral closure of the ring k[y] in the field k(x), and for U the locus of z over /C. I n either case we may write z ~ f ( x ) , where f is a birational correspondence between V and U, defined over k. By definition, U is everywhere normal with reference to k, and the mapping h ~ g o f-~ of U into W is everywhere defined and such that the (settheoretic) inverse image of every point of W for that mapping consists of finitely many points of U. Let a be any point of V; let (a, b) be a specialization of (x,z) over x---)a with reference to k,; then, as h is defined at b, (a,b,h(b)) is a specialization of (x,z,y) over k. As f' is defined at a, g is also defined there, so that we must have h ( b ) ~ g ( a ) ; therefore b is one of the finitely many points of U whose image by h is g(a). As V is normal at a by assumption, with reference to /c, this implies that f is defined at a, and that we have b = f ( a ) . We have g ( a ) ~ f ' ( a ) G hence f ' ( a ) = g ( a ) ~ - ' ; as g(a) = h ( b ) , this shows that f ' ( a ) is the unique specialization of y' over z---) b with reference to k"; as f' is biregular at a, f'-~ is defined at f'(a), and therefore x has no other specialization than a over z---)b with reference to U, hence also with reference to k by F-II~, Prop. 3. As U is normal at b, with reference to /C, this implies that f ~ is defined at b ~ f ( a ) . We have thus shown that f is biregular at every point of V, so that it is a k-isomorphiim between V and a /c-open subset of U. As a special ease (already contained in Proposition 2), we see that, in problem (D), V' is projectively (resp. affinely) embeddable over k if Vt is so over/c(t) ; similarly, in problem (C'), V is projectively (resp. affinely) embeddab]e over /C if Vt is so over /c(t). 8. In [4], the construction carried out in Nos. 7-9 can be advantageously replaced by the application of our Theorem 6 to the situation described in No. 6 of that paper. The application is entirely straightforward, so that no further details need he given; this shows that the recourse to the Lang-Weil 1U is the "derived normal mt~del of W in tile field k(x)" according to Zariski's delinition (!5], pp. 69-70); of. also [3].
306
[1956] 57)~4
ANDRt~ WEIL.
Theorem, J.e.. in substance, to the so-called "Riemann hypothesis" in the case of a finite groundfield (loc. cir., p. 374) was unnecessary; so is the assumption of normality in the final result (loc. cir., p. 375) ; normality had to be assumed there merely because of the use made of the Chow point in the construction on p. 370, whereas in the present paper a different device was adopted (in the proof of Proposition 1). Of course, in the main theorem of [4] (p. 375), parts (i) and (ii) remain unchanged. For the sake of completeness, we give here the improved result by which part (iii) of that theorem may now be replaced: PROPOSITION" 3. Let G be a group and W a chunk of transformation-space with respect to G, both defined over t~. Then there is a transformation-space S with respect to G, and a birational correspondence f between W and S, both defined over lc, with the following properties: (a) f is biregular at every point of W; (b) for every s e G and a e W such that sa is defined, we have f(sa) = s f ( a ) ; (c) every point of S can be written in the form sf(a), with s e G and ae W. Moreover, S is uniquely determined by these properties up to a k-isomorphism compatible with the operations of G. UNIVERSITY OF CHIOAG0.
REFERENCES.
[1] W. L. Chow, "Abelian varieties over function-fields," Transactions of the American M a t h e m a t i c a l Society, vol. 78 (1955), pp. 253-275. [2] S. Lang, "AbeIian varieties over finite fields," Proceedings of the National A c a d e m y of Sciences, vol. 41 (1955), pp. 174-176. [3] T. M a t s u s a k a , "A note on m y paper ' S o m e t h e o r e m s on abelian varieties '," N a t . Se. Rep. Ochanomizu Univ., vol. 5 ( ] 9 5 4 ) , pp. 21-23. [4] A. Weil, " On algebraic g r o u p s of t r a n s f o r m a t i o n s , " A m e r i c a n J o u r n a l of Mathematics., vol. 77 (1955), pp. 355-391. [5] O. Zariski, " Theory and a p p l i c a t i o n s of holomorphic f u n c t i o n s on algebraic varieties over a r b i t r a r y groundfields," M e m o i r s of the A m e r i c a n M a t h e m a t i c a l Society, no. 5 (1951), pp. 1-90.
[ 1957a] Zum Beweis des Torellischen Satzes
Vorgelegt v o n t t e r r n 1~I. D e u r i n g
in d e r S i t z u n g v o m 8. F e b r u a r 1957
Die yon Siegel geschaffene Theorie der hSheren Modulfunktionen kann als Theorie der Moduln ffir abelsche FunktionenkSrper angesehen werden; ihre Anwendung auf das klassische Problem der Moduln algebraischer Kurven beruht auf einem Satz yon 1~. Torelli, der ungefi~hr besagt, dab eine algebraische Kurve durch die Normalperioden der abelschen Integrale 1. Gattung auf der Kurve v611ig bestimmt ist. Der Beweis, der von Torelli selbst ffir diesen Satz gegeben wurde (Rend. Ace. Lincei [V] 22 [1914], p. 98), ist im wesentlichen algebraisch-geometrischer Natur. Eine moderne, vSllig einwandfreie und lfickenlose Darstellung dieses Beweises liegt nieht vor; dasselbe grit auch ffir den durchaus klassisch formulierten und an mehreren Stellen ziemlich skizzenhaften Beweis yon A. Andreotti (Mdm. Acad. Belg. 27 [1952], fase. 7). In der vorliegenden Arbeit soll der Torellische Satz in einer solehen Fassung formuliert und bewiesen werden, dab er auch im ,,abstrakten Fall" (beliebiger Charakteristik) seine Gfiltigkeit behalf. Die ganz einfache Beweisidee, die mit dem Andreottischen Ansatz eng verwandt ist, ist leider durch teehnische Schwierigkeiten etwas entstellt, welche es nStig machen, die F~lle niedrigeren Gesehlechtes besonders zu behandeln. I. Zur abstrakten Formulierung des Torellischen Satzes braucht man den Begriff einer ,,polarisierten" abelschen Mannigfaltigkeit (vgl. meinen Vortrag ,,On complex multiplication", Tokyo Symposium on Numbertheory, 1955, sowie eine demn~chst erscheinende Arbeit yon T. Matsusaka im American Journal o/ Mathematics). Zun~chst sei A eine abelsche Mannigfaltigkeit im klassischen Fall, d.h. fiber dem komplexen Zahlk6rper; der zugehSrige FunktionenkSrper sei der KSrper aller periodischen meromorphen Funktionen in einem n-dimensionalen Vektorraum R fiber dem komplexen Zahlk6rper mit einem vorgegebenen Periodengitter G vom Rang 2 n. Damit G tats~chlieh Periodengitter eines solchen KSrpers der Dimension n sei, ist bekanntlieh notwendig und hinreiehend, dal3 es eine auf G • G ganzzahlige alternierende Reprinted from G6ttingen Nachrichten 1, 1974, by permission of Akademie der Wissenschafien GOttingen.
307
308
[1957a] Andrd Weil
34
Bilinearform B gebe, die Imaginiirteil einer im Raum R positiv definiten Hermiteschen Bilinearform ist. Nun sagt man, dab jede solche Bilinearform B eine Polarisierung der abelsehen Mannigfaltigkeit A ----RIG bestimmt, wobei zwei Formen B, B' dann und nur dann dieselbe Polarisierung von A bestimmen, wenn sie sich nur um einen konstanten Faktor unterseheiden. Man sagt, dab A polarisiert ist, wenn man auf A eine bestimmte Polarisierung gew/~hlt hat. Wenn s/~mtliehe Elementartefler der Form B einander gleich sind, sagt man, dab die dureh B polarisierte Mannigfaltigkeit A zu der Hauptschar der polarisierten abelsehen !~annigfaltigkeiten der Dimension n geh6rt. Wenn eine abelsche Mannigfaltigkeit keine komplexen Multiplikationen (d. h. keine nieht-trivialen Endomorphismen) besitzt, dann ist sie nur auf eine Weise polarisierbar. Man kann sieh leieht iiberzeugen, dab es sieh in der klassischen Theorie der abelsehen Funktionen (z.B. in den Arbeiten von Hermite und Humbert) meistens um polarisierte abelsohe Mannigfaltigkeiten handelt; nur dann, wenn von solehen Mannigfaltigkeiten die Rede ist, durf man von ihren Moduln spreehen. Z.B. ist die Siegelsehe Theorie der hSheren Modulfunktionen niehts underes als eine Theorie der l~oduln ftir die Hauptschar der polarisierten abelschen l~annigfultigkeiten einer gegebenen Dimension, indem ein System soleher !~Ioduln dureh einen Punkt im Fundamentalbereieh der Modulgruppe bestimmt ist. Im ubstrukten Fall sei X ein nicht-ausgearteter positiver Divisor auf einer abelsehen Mannigfaltigkeit A ; darunter versteht man einen positiven Divisor X, der hSehstens durch endlieh viele Trunslutionen in einen dazu linear /~quivalenten Divisor verwandelt wird. Dann sagt man, dub X eine Polarisierung von A bestimmt, wobei zwei solehe Divisoren X, X' dann und nur dunn dieselbe Polurisierung von A bestimmen, wenn es zwei positive ganze Zahlen m, m' gibt, ftir welehe m X und m'X" algebraiseh /~quivalent sind; jeder positive Divisor X" mit dieser Eigensehaft heiBe ein Polardivisor ftir die dureh X polarislerte abelsehe Mannigfaltigkeit A. Aus dem Hauptsatz der Theorie der Thetufunktionen folgt sofort, daft im klassischen Fall dieser Polarisierungsbegriff mit dem oben definierten zusammenf/~llt. Insbesondere sei J die Jacobische Mannigfaltigkeit einer algebraischen Kurve C; die Bezeichnungen seien dieselben wie in meinem Bueh Varidtds abdliennes et courbes algdbriques, Paris 1948 (zitiert als VA). Es sei ~ die (bis auf eine Translation eindeutig bestimmte) kanonisehe Abbfldung von C in J. Wenn a----~aiMi ein Divisor auf O ist, werde ich zur Abkiirzung ~(a)---i
~ a ~ ( M ~ ) sehreiben (statt S[~(a)] wie in VA). Wie in VA wird dureh W~ diejenige Untermannigfaltigkeit yon J bezeiehnet, die aus allen Punkten der Form ~0(a) besteht, wenn a die Gesamtheit der positiven Divisoren vom Grad r auf C durehl/iuft. Start Wg-1 wird meistens 0 gesehrieben; werm f ein Divisor der kanonisehen Klasse auf C ist, besteht O ~ W~_x aueh aus den Punkten (~) - - ~0(m), wenn m die Gesamtheit der positiven Divisoren vom Grad g - - 1
[1957a]
309 Zum Beweis des Torellischen Satzes
35
durchlauft; O wird also durch die Abbfldung u -+ ~ ( ~ ) - u auf sich selbst abgebildet. Start Wg_~ werden wir W schreiben; es sei W* das Bild von W bei der Abbildung u --> ~ (1) - - u yon J auf sich selbst. Wenn X ein Zyklus (und insbesondere eine Mannigfaltigkeit) auf J und a ein Punkt auf J ist, so wird (wie in VA) durch Xa der aus X durch die Translation a entstehende Zyklus bezeiehnet. Wir wollen unter der kanonischen Polarisierung von J diejenige verstehen, welche von O oder (was auf dasselbe hinausl~uft) yon einem beliebigen unter den Divisoren Oa bestimmt wird. Diese Polarisierung ist offenbar invariant bei jeder Abbfldung u --> • u + a v o n J auf sich selbst (dasselbe gilt iibrigens fiir jede Polarisierung einer abelschen Mannigfaltigkeit, wie z.B. sofort aus VA-73, prop. 31 folgt). Im klassisehen Fall gehSrt jede kanoniseh polarisierte Jacobische Mannigfaltigkeit zu der Hauptsehar der abelsehen Mannigfaltigkeiten derselben Dimension. 2. Der eigentliehe Inhalt des Torellischen Satzes besagt nun, daB eine polarisierte abelsche Mannigfa~tigkeit im wesentlichen h6chstens auf eine Weise die kanoniseh polarisierte Jacobisehe Mannigfaltigkeit einer Kurve sein kann. Genauer li~Bt er sich wie folgt formulieren: tIauptsatz. Es seien C, C' zwei Kurven vom gleichen Geschleoht g ~ 1 ; es sei (bzw. q~') die kanonische Abbildung yon C (bzw. C') in ihre Jacobische Mannig]altigkeit J (bzw. J'). Es gebe einen Isomorphismus ~ der kanonisch polarisierten Mannig/altig]ceit J au] die kanonisch polarisierte Mannig/altiglceit J'. Dann gibt es einen Isomorphismus ] yon C au/ C' mit der Eigenscha/t, daft ~ o q~ -~ • qJ o ] + a ist mit konstantem a; dabei sind ], a und das Vorzeichen + dutch ~ eindeutig bestimmt, wenn C und C" nicht hyperelliptisch sind; bei hyperelIiptischen C und C" sind / u n d a dutch ~ und das Vorzeichen -4- eindeutig bestimmt, wobei das letztere beliebig gewahlt werden kann. Der Beweis ergibt sieh dadurch, dab man eine explizite Konstruktion ftir C und ~ angibt, wenn J als gegeben gedacht wird. Zuni~chst wird gezeigt, dab die Sehar {O~} der aus O durch Translation entstehenden Mannigfaltigkeiten durch die Polarisierung yon J eindeutig bestimmt ist; das folgt im klassischen l~all aus bekannten S~tzen aus der Theorie der Thetafunktionen und im abstrakten l~all aus folgendem Satz: Satz 1. E i n positiver Divisor Z au/ J laflt sich dann und n u t dann dutch eine Translation in 0 aber/i~hren, wenn Z ~ 0 ist (wobei X ~ O wie in VA-57 dureh das Bestehen der linearen .~quivalenz X, ~ X fiir alle t definiert ist). Es ist ni~mlich O~ -~ O fiir alle a. Es sei Z ~ O m i t positivem Z; dann gibt es nach VA-62, th. 32, cor. 2 ein a, ftir welches Z ~ O~; indem wir Z durch Z_~ ersetzen, brauehen wir also nur zu zeigen, daB, wenn Z ~ O und Z positivist, Z mit O zusammenfallen muB. Es sei x ein generischer Punkt auf J (in bezug auf einen gemeinsamen Definitionsk5rper ftir alle in Betracht kommenden Mannigfaltigkeiten); dann ist naeh VA-41, th. 20 ~ - 1 ( O ~ ) ~ - ~ P ~ , wo die
310
[1957a] A n d r d Weil
36
P~ (1 ~ i < g) P u n k t e auf C sind mit der Eigenschaft ~ ( P ~ )
~- ~(t) -t- x;
i
die P, mfissen also unabh~ngige generische P u n k t e auf C sein, woraus folgt, dab l ( ~ P,) ~- 1. A u s Z ~ O folgt, dab ~-I(Z~) ein mit ~ Pi linear ~quivalenter positiver Divisor auf C sein muB; wegen l ( ~ P,)--~ 1 fallt er also mit ~ Pi zusammen; demnaeh liegt jeder P u n k t T(Pi) auf einer Komponente yon Z~, und der P u n k t g
auf einer Komponente yon Z. Da dieser Punkt, wie letztere Gleichung zeigt, auf O generisch ist, so mul~ O Komponente yon Z sein; also ist Z - - O positiv und ~ 0, woraus folgt, dab Z - O ~ 0 ist, w.z.b.w. N u n wird jedem Divisor X auf J durch die in VA-45 und VA-48 definierte Abbildung X -~ 8~ ein Endomorphismus ~ yon J zugeordnet; dabei ist 8o der identische Automorphismus 6j yon J ; nach VA-57, th. 30 ist dann und nur dann 8x ~ Or, wenn X ~ Y. Wenn X Polardivisor auf der kanoniseh polarisierten Mannigfaltigkeit J ist, so gibt es nach der Definition zwei positive ganze Zahlen m, m', ffir welche m X mit m ' O algebraisch ~quivalent ist; um so mehr besteht dann die Relation m X ~- m ' O , also m(5'X ~- m'(~j; dann mul3 aber (z.B. nach VA-65, th. 33, cor. 1) n ---- m ' / m ganzzahlig sein, woraus folgt, dab 8x----nOj und X - ~ n O mit ganzzahligem n gilt. Sagen wir, daB ein Polardivisor X auf J minimal ist, wenn es zu jedem Polardivisor Y auf J eine ganze Zahl r gibt, fiir welche Y -~ r X . J e t z t folgt sofort aus Satz 1, dab ein Polardivisor dann und nur dann minimal ist, wenn er der Schar {O~} der aus 0 durch Translation entstehenden Divisoren angeh6rt. I)amit ist diese Schar in invarianter Weise a u f der polarisierten Mannigfaltigkeit J gekennzeichnet. 3. Wir sind jetzt schon imstande, den l~all g--~ 2 zu erledigen. In diesem Fall ist n~mlich O ~ W~ ~ ~ (C). I m Hauptsatz kSnnen sich dann die Kurven a(T(C)), ~'(C') auf J ' hSehstens dutch eine Translation unterscheiden. Da ~, ~' die K u r v e n C, C' isomorph auf ~0(C), ~'(C') abbilden, so muB es also / und a geben, ffir welche ~ o ~ --~ ?' o / + a; dab / und a dabei eindeutig bestimmt sind, ergibt sich daraus, dab (z.B. nach VA-62, th. 32, cor. 2) O durch keine Translation in sich selbst fibergeffihrt wird. Bekanntlich ist jede Kurve vom Geschlecht 2 hyperelliptisch; auf einer solchen Kurve C hat ni~mlich die kanonische Vollschar (bestehend aus allen positiven Divisoren, die mit einem beliebigen kanonischen Divisor ~ linear ~quivalent shad) den Grad 2 und die Dimension 1. Es gibt also einen Automorphismus h yon C derart, dab ftir jeden P u n k t M auf C der Divisor M + h ( M ) der kanonischen Vollschar angehSrt; dann ist ~ ( h ( M ) ) = ~(~) - - T ( M ) . Wenn nun ~, ~', a, ], a dieselbe Bedeutung haben wie oben und /1-----] o h , a l = ~ ( q ~ ( ~ ) ) - - a gesetzt wird, so ist ~ o ~ - - - - - - r o / ~ + a ~ ; dabei sind/~, a~ dureh diese Gleichung eindeutig bestimmt, da sonst / ----/~ o h t
t
[1957a]
311 Zum Beweis des Torellischen Satzes
37
und a durch die Gleichung ~ o ~ : ~ ' o ] + a nicht eindeutig bestimmt w~ren. Damit ist der Hauptsatz im Fall g = 2 vollsti~ndig bewiesen. Die in der Literatur h~ufig vorkommende Behauptung, dab im klassischen l~all eine der Hauptsehar angehSrige abelsehe Mannigfaltigkeit der Dimension 2 ,,im allgemeinen" Jacobisehe Mannigfaltigkeit einer Kurve ist, l~Bt sich jetzt leicht auch im abstrakten Fall in folgender pr~ziseren Fassung formulieren und beweisen: Satz 2. E s sei A eine polarisierte abelsche Mannig/altigkeit der Dimension 2. E s gebe au] A einen Polardivisor X , derart, daft deg (X 9X~) ~- 2 sei ]i~r generisches u. D a n n ist entweder X eine Kurve vom Gesehlecht 2, A die kanonisch polarisierte Jacobische Mannig]altigkeit yon X und die kanonische Abbildung yon X in A die identische; oder A ist das Produlct 1" • 1" zweier elliptischer Kurven 1", 1", und X ist von der F o r m 1" • a' + a • I ~, wo a, a' Punlcte au] 1" bzw. 1" sind. Zur bequemeren Formulierung des Beweises wollen wir (in Erweiterung einer Definition in VA-22) folgendes verabreden: Wenn A eine beliebige abelsche Mannigfaltigkeit und X eine Untermannigfaltigkeit von A ist, so werden wir unter der durch X erzeugten abelsehen Untermannigfaltigkeit yon A die kleinste abelsehe Untermannigfaltigkeit B von A verstehen, die si~mtliche Punkte x - - x" mit x und x' in X enth~lt. Aus VA-21 folgt, dab B die dureh diese Punkte erzeugte Untergruppe von A ist; wenn x ein beliebiger Punkt yon X ist, so ist auch B die im Sinne von VA-22 durch X_~ erzeugte abelsche Untermannigfaltigkeit von A. Z.B. wird die Jacobische Mannigfaltigkeit J einer Kurve C durch die Kurve ~ (C) erzeugt, wenn ~ die kanonische Abbildung yon C in J ist; also wird auch J durch jede der Mannigfaltigkeiten W~ erzeugt. Wegen VA-20, th. 9 folgt daraus, dal3 W~ ftir r ~ g mit keiner abelschen Mannigfaltigkeit birational ~quivalent sein kann. Nach dieser Bemerkung werden wir Satz 2 zun~chst in dem Fall beweisen, wo der Polardivisor X eine Kurve ist, d. h. wo X aus einer einzigen Komponente mit dem Koeffizienten 1 besteht. Es sei i die Injektion, d . h . die identische Abbildung, von X in A. Das Geschlecht g yon X kaim weder 0 noeh 1 sein; im ersteren Fall mfiBte n~mlieh i nach VA-19, th. 8 konstant sein; im letzteren Fall w~re i naeh VA-20, th. 9 bis auf eine Translation ein Homomorphismus, also w~re X bis auf eine Trans]ation eine abelsche Untermannigfaltigkeit von A, und es w~re deg (X. Xu) ~ 0. Es sei J die Jacobische Mannigfaltigkeit yon X; es sei ~ die kanonisehe Abbildung yon X in J ; k sei ein Definitionsk6rper ffir A, X, J, ~. Setzen wir w : ~ (M) + ~ (N) und x ~- i(M) + i(N), wo M, N zwei unabh~ngige generische Punkte yon X fiber k sind. Der Oft (,,locus") von w fiber k ist dann nach unseren iiblichen Bezeiehnungen W~; nach VA-39, prop. 15 ist k(w) der KSrper k(M, N)8 der symmetrischen Funktionen von M, N fiber k; also ist k ( x ) in k(w) enthalten, so dab wir x ~ F ( w ) sehreiben k6nnen, wo F eine Abbildung yon W2 in A ist. Es sei X" der aus X
312
[1957a]
38
Andrd Weil
dureh die Abbfldung u - - > - - u laervorgehende Divisor; wegen VA-61, th. 31, eor. 2 folgt aus der Voraussetzung tiber X, dab deg (X. X~) ~ 2; nach VA-13, th. 4, eor. 2 mul~ also x generiseh tiber lc auf A und k (M, N) algebraisch vom Grad 2 tiber k ( x ) sein; wegen k ( x ) ~ k(M, N)s ist demnaeh k ( x ) -~ k ( M , N ) , ~- k (w). Damit ist gezeigt, dal3 F eine birationale Abbfldung yon W~ auf A ist. Wie oben bemerkt, kann W2 ftir g ~ 2 mit keiner abelschen Mannigfaltigkeit birational ~quivalent sein; es ist also g -~ 2 und W2 : J ; nach VA-20, th. 9 ist dann F bis auf eine Translation ein Isomorphismus von J auf A ; andererseits folgt aus VA-18, th. 7, cor., dab ~ und F o i sieh hSchstens durch eine Translation unterscheiden kSnnen. Damit ist Satz 2 in dem Fall, wo X eine Kurve ist, vollst~ndig bewiesen. Falls X keine Kurve ist, kSnnen wir X ~-- ~ X (~ schreiben, wobei X m . . . . , X (~) Kurven sind. Dann ist ftir generisches u: 2
deg (X. X,)
~ deg (X (~ 9X (~
Da jedes Glied auf der rechten SeRe _~ 0 ist, so folgt unmittelbar aus dieser Gleichung (unter Benutzung von VA-22, prop. 6) erstens, dab jede Kurve X (~ bis auf Translation eine abelsche Untermannigfaltigkeit von A sein muB, und zweitens, dab es nur zwei solche Kurven gibt; dann ist deg (X m .X(,2~) ---- 1. Wegen VA-13, th. 4, cor. 2 ist damit der Beweis beendet. I m klassischen Fall gentigt eine polarisierte abelsche Mannigfaltigkeit dann und nur dann der Voraussetzung yon Satz 2, wenn sie der Hauptschar angehSrt; also ist eine solche Mannigfaltigkeit entweder eine Jaeobische Mannigfaltigkeit oder ein Produkt zweier elliptischer Kurven. Es ist mir nicht gelungen, die entsprechende Frage ftir g -~ 3 zu entscheiden. 4. Zur Behandlung des allgemeinen Falles brauchen wir einige Hflfss~tze. Bekanntlich pflegt man zu sagen, dab ein posit~ver Divisor m a u f der Kurve C in der durch einen Divisor a bestimmten linearen Vollschar ~ (bestehend aus allen m i t a linear ~quivalenten positiven Divisoren auf C) enthalten ist, wenn es einen positiven Divisor m' gibt, ftir welchen m + m' der Schar angeh6rt. Unter Benutzung dieser Sprechweise gilt folgender Hflfssatz: Hflfssatz 1. E8 sei a ein Divisor vom Grad 0 au] der Kurve C; es sei a ----- q~(a) O; es sei ~ die dutch f + a bestimmte lineare Vollschar. D a n n besteht die Punktmenge 0 (-~ O~ aus den P u n k t e n ~(m), wenn m die in der Vollschar enthaltenen positiven Divisoren vom Grad g - 1 durchlau/t. Die Vollschar hat dann und n u t dann einen F i x p u n k t P, wenn a yon der F o r m a -~ q~(P - - Q) * ist, wo Q ein P u n k t yon C ist; in diesem Fall ist 0 ~ Oa = W~(p) ~ W_~(~). Die erste Behauptung folgt unmittelbar aus der Tatsache, dab O, O~ aus den Punkten ~(m) bzw. ~(f + a - m') bestehen, wobei m, m' die positiven Divisoren vom Grad g - 1 durchlaufen. Damit P Fixpunkt der Schar ~ sei, ist notwendig und hinreichend, dab l(I + a - - P) -- l(f + a) sei; nach dem Riemann-Rochsehen Satz l~Bt sich diese Gleichung in der Form / ( P - - a )
[1957a]
313 Zum Beweis des Torellischen Satzes
39
-~ l ( - - a) + 1 sehreiben. D a n n ist 1( P - - a) ~ 1, also P - - a m i t einem positiven Divisor ~quivalent; da P - - a v o m Grad 1 ist, gibt es d a n n einen P u n k t Q derart, d a b P--a ~ Q, also a - ~ ~ ( P - - Q ) ist. U m g e k e h r t , w e n n letztere Gleichung b e s t e h t u n d a ~= 0 ist, ist l ( P - a) = l(Q) -~ 1, l ( - - a ) -~ o; wie oben gezeigt wurde, ist d a n n P F i x p u n k t v o n ~ . I n diesem Fall b e s t e h t ~ aus den Divisoren P + ~, wo ~ die durch ~ - - Q b e s t i m m t e Vollschar durchl~uft. Es sei der positive Divisor m v o m G r a d g - - 1 in dieser Schar ~ e n t h a l t e n ; d a n n gibt es einen posiriven Divisor m ' v o m G r a d g - - 1 derart, d a b m + m ' der Sehar ~ angehSrt; also ist P entweder K o m p o n e n t e v o n m oder K o m p o n e n t e von m ' . I m ersten Fall k a n n m a n m ~ P + n schreiben, wo n ein positiver Divisor v o m G r a d 9 - 2 ist, der in der durch ~ - Q b e s t i m m t e n Vollschar e n t h a l t e n ist. Aus d e m R i e m a n n - R o c h s e h e n Satz folgt aber sofort, d a f alle positiven Divisoren v o m G r a d g - 2 in dieser Vollschar e n t h a l t e n sind; jeder Divisor P + n m i t p o s i t i v e m n v o m G r a d g - - 2 ist also in der Sehar ~ enthalten. Zweitens sei m wie oben in ~ e n t h a l t e n u n d P K o m p o n e n t e v o n m ' . D a n n kSnnen wir m ' -~ P + n ' sehreiben m i t einem positiven n' v o m G r a d g - - 2; w e n a u m gekehrt n' ein beliebiger positiver Divisor v o m G r a d g - - 2 ist, so ist wie oben P + n' in ~ enthalten, so d a f es einen positiven Divisor m v o m G r a d g - 1 g i b t derart, d a f m + P + 11' der Schar ~ angeh6rt; d a n n ist m in ~ enthalten, u n d es ist m ~t+a--P--n" ~ f--rr D a m i t ist bewiesen, daft O r~ Oa aus den P u n k t e n ~0(P + n), ~ ( [ - 1 t ' - - Q ) besteht, wenn u, 11' die Menge aller positiven Divisoren v o m G r a d g - 2 durchlaufen. D a m i t ist der Hflfssatz vollst~ndig bewiesen. 5. Ftir g ~ 5 gilt weiter der folgende Hflfssatz: Hflfssatz 2. Die Bezeichnungen seien dieselben wie im Hil/ssatz 1; die Scha~ sei fixpunkt]rei, und das Geschlecht g yon C sei > 5. Welter ],abe 0 r~ Oa mehr als eine Komponente. D a n n gibt es eine Komponente dieser Menge, die dutch unendlich viele Translationen in sich selbst i~berge/ahrt wird. Es sei k ein algebraiseh abgeschlossener DefinitionskSrper ftir C, J , ~ u n d die K o m p o n e n t e n y o n a; M 1 , . . . , M~_~ seien g - - 2 unabh~ngige generische P u n k t e v o n C fiber/c; wir setzen m ----~ M~. Wegen a ~= 0 ist l(~ + a) ----g - 1, v
also l (~ + a - - m) ~-- 1 ; es ist also [ + a - - m ~ 11 m i t einem durch diese Relat i o n eindeutig b e s t i m m t e n positiven Divisor n v o m G r a d g; 11 ist d a a n rational fiber d e m K S r p e r k ( m ) ---- k(M1 . . . . . Mg_~),, z.B. n a e h VA-41, th. 20. J e d e K o m p o n e n t e y o n O ~ O~ h a t die Dimension g - - 2 u n d ist algebraiseh tiber k, also rational tiber k, weft k algebraisch abgesehlossen ist; folglieh h a t ein in Bezug a u f k generiseher P u n k t einer solchen K o m p o n e n t e die Dimension g - - 2 tiber k. N a e h Hflfssatz 1 li~ft sieh ein soleher P u n k t in der F o r m ~ (~) sehreiben, wo ~ ein in ~ e n t h a l t e n e r positiver Divisor v o m Grad g - 1 ist. E s sei ~ ~-- R1 + - . - -t- Rg-1; d a m R ~ 0:) tiber k die Dimension g - 2 habe, mfissen wenigstens g 2 der P u n k t e B~, z.B. R x , . . . , R~_~, fiber k u n a b h a n g i g
314
[1957a] Andrg Weil
40
sein; es gibt also einen Isomorphismus von k(R1 . . . . . Rg_2) auf k(M1 . . . . . Mg-2) fiber k, der Rt in M~ ffir i ---- 1 . . . . . g - - 2 fiberffihrt. Dieser Isomorphismus kann zu einem Isomorphismus von B ( R 1 , . . . , R~-I) in das Universalgebiet erweitert werden; dadureh wird Rg_l in einen P u n k t N von C fibergeffihrt, also ~ in den Divisor m -[- N ; dann ist ~0(m + N) generiseh fiber k auf derselben K o m p o n e n t e von O ('~ O~ wie ~0(~). D a m i t ist gezeigt, dab es zu jeder K o m p o nente y o n O r-~ O~ einen generischen P u n k t der F o r m ~0(m + N) gibt. Dabei muB der Divisor m + N in ~ enthalten sein; da aber die Relation m + n ,~ ~ + a den positiven Divisor rt eindeutig bestimmt, mug N eine der g K o m p o n e n t e n des Divisors n sein. Da II fiber k (m) rational ist, ist u entweder ein rationaler Primdivisor fiber k(m) oder S u m m e yon mehreren solehen Primdivisoren. I m ersten Fall sind samtliehe K o m p o n e n t e n von n zueinander fiber k(m) konjugiert; dasselbe gilt also ffir die P u n k t e ~(m + N), so dab je zwei dieser P u n k t e generische Spezialisierungen voneinander fiber k (m) und u m so mehr fiber k sind; folglieh haben sie denselben Ort fiber k, so dab O ~ O~ nieht mehr als eine K o m p o nente haben k a n n entgegen der Voraussetzung. Demnaeh ist n yon der F o r m rt-----rtl~...+nh, wo die nt rationale Primdivisoren fiber k(m) sind, mit h ~ 2. Es sei vt der Grad von rtt; dann ist ~ vt = g, also vt ~ g - - 1 ffir alle i wegen i
h ~_ 2. Ffir jedes i i s t ~0(ui) rational fiber k(m), also aueh fiber k ( M 1 . . . . . Ma_2) , u n d laBt sieh also (nach VA-18, th. 7, eor.) in der F o r m yJ(M1) + ~v'(M2) + 999 sehreiben, wo % ~v', . . . g - 2 Abbildungen von C in J sind, die (weil ihre Summe eine symmetrische F u n k t i o n von M1 . . . . . Mg_~ ist) sich voneinander h6ehstens durch Addition einer K o n s t a n t e n unterseheiden. Demnaeh k5nnen wir ~(rt~) in der F o r m g--2
(nt) = ~ ~o~(M~) + c~ = ~t (m) + c~ sehreiben, wo ffir jedes i ~ eine Abbfldung y o n C in J und c~ eine K o n s t a n t e bedeutet; dabei ist jedes ~t bis a u f eine K o n s t a n t e eindeutig bestimmt. Ffir jedes i i s t aber ~t bis auf eine K o n s t a n t e v o n d e r F o r m a~ o % wo at ein Endomorphismus yon J ist; wit dfirfen also annehmen, da$ ~Pt = at o ~v; da k algebraiseh abgesehlossen ist, sind dann die at u n d c~ fiber k definiert. D a / ( n ) --~ 1 ist, so ist u m so m e h r l(rtt) ---- 1 ffir jedes i; folglieh ist ffir j e d e s i n~ algebraiseh (und sogar, wie m a n sieh leieht fiberzeugt, rational) fiber k(~ (ut)). W a r e insbesondere ~ (nt) algebraisch, also rational, fiber k, so ware dasselbe ffir jede K o m p o n e n t e von ut der Fall; dann ware aber ein soleher P u n k t F i x p u n k t der Schar ~ entgegen der Voranssetzung, weil m -4- u die Dimension 9 - - 2 fiber k hat und folglieh ein generiseher Divisor der Vollschar ist. D e m n a e h ist at # 0 ffir jedes i. Wenn At die abelsehe Untermannigfaltigkeit von J ist, auf welehe J dureh ~t abgebildet wird, und dt die Dimension yon A~ bedeutet, so ist also d~ > 0 ffir jedes i.
[1957a]
315 41
Z u m Beweis des Torellischen Satzes
Setzen wir fl---- ~j + ~ ~i. D a n n ist: i
woraus folgt, dab fl(~(m)) fiber k rational ist. D a J dureh den O r t W v o n (m) fiber k erzeugt wird, so muB also fl k o n s t a n t sein. Weft fl ein E n d o m o r phismus y o n J ist, ist d e m n a e h fl = 0, d . h . - - x = ~ a i x ffir jedes x a u f J . i
W e n n x generiseh a u f J fiber k gew~hlt ist, h a t ai x die Dimension d~ fiber k, also ~ ~ x hSehstens die Dimension ~ di. D a m i t ist bewiesen, d a b ~ d~ > g. i
i
i
Andererseits wird J dureh ~v(C), also A t dureh ~ (C) erzeugt; d u r c h I n d u k tion n a c h r ffir 1 < r < 9 - - 2 sieht m a n d a n n leieht, dab die Dimension y o n ~ ( M 1 -t- 999 M,) fiber k gleieh r i s t , solange dieser P u n k t a u f A~ nicht fiber k generisch ist; insbesondere ist die Dimension v o n ~i(m) fiber k gleich v~ -~ inf (g - - 2, d~) ; sie ist d a n n u n d n u r d a n n gleieh di, wenn ~ (m) a u f A ~generiseh ist. N u n ist aber v~ ebenfalls die Dimension v o n q (n~) fiber k; d a dieser P u n k t a u f W~ liegt, so ist v'~ < v~; es ist d a n n und n u r d a n n v~ = v~, wenn ~ (n~) a u f W~ generiseh ist; d a J dureh W~ u n d A~ dureh den O r t von yJ~(m) fiber k erzeugt wird, u n d d a sieh ~i(m) und q(rti) n u r u m die K o n s t a n t e ci u n t e r scheiden, so folgt aus v: - - vi, dab J = Ai, also d~ ---- g u n d v~ -----g - 2 ist. N e h m e n wir an, es sei v'i < g - - 2 ffir jedes i; d a n n k a n n nieht v~ ~--vi sein; also ist v~ < v~; zugleieh m u g aueh v~ ---- di sein, also d~ < v~ ffir jedes i, was der Ungleichung ~ d~ > g = ~ vi widersprieht. Folglieh dfirfen wir z.B. i
i
a n n e h m e n , d a b v~ ---- g - - 2 ist; d a n n ist vl > g - - 2, also % < 2. W e n n v~ = 1 w~re, so w~re, wegen 0 < v~ < v~ ffir jedes i, v~ ~-- 1, also v~ ---- v~ -~ g - - 2 u n d g -~ 3 entgegen der Voraussetzung; also ist v~ = 2. W e n n v~ ~-- 2 w/~re, so w~re v~ = g - - 2 u n d g = 4, entgegen der Voraussetzung; also ist v~ = 1 u n d (wegen g > 5) v~ = d~. So ist gezeigt, dab h = 2, v~ = g - 2, v~ = 2, v~ = d~ ---- 1 ist; der Ort y o n ~%(m) ist A~ u n d h a t die Dimension 1; und der Ort B yon q(n~) e n t s t e h t aus A~ dureh die Translation cz. Setzen wir rt~ ~-- N + N ' ; weder N noeh N ' ist algebraisch fiber k; da aber v~ ~-- 1 ist, so ist N" algebraiseh fiber k (N). U n t e r den Mr muB es d a n n g - - 3 P u n k t e geben, z.B. M~, . . . , Mg_a, die fiber k(N, N') u~abh/~ngig sind. D e r Divisor M~ + 999 A- Mg_a + n~ v o m Grad g - 1 h a t d a n n fiber k die Dimension g - - 2 und ist in der Sehar ~ e n t h a l t e n ; also ist, wenn wit t = ~ ( n ~ ) , w=~(M a+...+M~_a) setzen, t + w generischer P u n k t einer K o m p o n e n t e Z v o n O r~ O~. Dabei shad t, w unabh~ngige generische P u n k t e v o n B bzw. Wg_a fiber k; Z wird also dureh jede Translation, die durch einen P u n k t y o n A~ b e s t i m m t ist, in sich selbst fibergeffihrt. D a m i t ist der Hflfssatz bewiesen. U m festzustellen, ob Hflfssatz 2 aueh ffir g ---- 3 u n d g ---- 4 gtiltig bleibt, w~ren ziemlich langwierige Fallunterseheidungen nStig; ich h a b e sie nieht bis zu E n d e durehgeffihrt.
316 42
[1957a] Andrd Well
Ein positiver Divisor c v o m Grad 2 auf einer K u r v e C vom Gesehlecht g ~_ 2 h e i f e hyperelliptisch, wenn l(r = 2 ist; in diesem Fall heiBe ebenfalls die dureh c bestimmte lineare Vollschar der Dimension 1 hyperelliptisch. Wie bekannt ist, gibt es auf einer K u r v e C vom Gesebleeht g ~ 2 h6ehstens eine hyperelliptische Vollschar, und C heil~t hyperelliptiseh, wenn es auf C eine solche Vollsehar gibt. Hflfssatz 3. Keine der Manniq]altiglceiten W1 = q~(C). . . . . W~_~ = W, Wg_l = O, W* wird in sich selbst durch eine Translation (aufler der identischen ) aberge/i~hrt. Es wird dann und n u t dann W dutch eine Translation in W* tlberge]ahrt, wenn C hyperelliptisch ist; in diesem Fall ist W* ~-- W~(c), wenn r ein hyperelliptischer Divisor au] C ist. DaB die erste B e h a u p t u n g ffir O gilt, ist z.B. in VA-62, th. 32, cor. 2 enthalten. Ffir r < g - - 1 ist O der Ort von v + w fiber k, wenn v, w unabhi~ngige generisehe P u n k t e y o n W~ bzw. W~-~-x fiber k sind; wenn also W~ dureh eine Translation in sieh selbst fibergeffihrt wfirde, w~re das aueh der Fall ffir O. Da W* naeh seiner Definition (in Nr. 1 oben) aus W ---- W~_~ dureh die Abbildung u -~ ~ (~) - - u entsteht, so gilt die erste B e h a u p t u n g aueh ffir W*. N e h m e n wir an, die Translation c ffihre W in W* fiber; es sei c ein Divisor v o m Grad 2 mit ~(r = c. Zu jedem positiven Divisor m v o m Grade g - - 2 m u f es dann einen positiven Divisor m ' desselben Grades geben, derart, dab f - - m ~ m' + r mit anderen Worten, es m u f l ( f - c - m) ~ 1 sein. Das Bestehen dieser Ungleiehung ffir generisches m fiber/c (c) ist aber mit der Ungleiehung 1( ~ - - c) g - - 1 gleiehwertig, und umgekehrt folgt aus der letzten Ungleiehung, dab l(f--c--m) _~ 1 ffir jedes positive m v o m Grad g - - 2 . Wegen des RiemannRoehsehen Satzes bedeutet aber l ( f - - r _~ g - - 1 niehts anderes als 1(r ~ 2. Hflfssatz 4. Es eei V der Oft des Punktes qJ( M - - N ) abet k, wenn M , N zwei unabhangige generische Punkte yon C ~ber k sin& Dann hat V die Dimension 2. Es sei ~ die Menge der P u n ~ e a ~= 0 au] J mit der Eigenscha/t, daft 0 r-~ O~ zwei Komponenten hat, wovon keins durch elne nicht-identische Translation in sich selbet abergeht. Dann ist 9~ = V - - {0}, wenn g ~_ 5 und C nicht hyperelliptisch ist; dagegen gibt es, wenn g ~ 5 und C hyperelliptisch ist, eins Kurve Vx au/ V derart, daft 9~ : V - - V1. Es seien P , Q, P ' , Q' irgend vier P u n k t e auf G; aus ~o(P - - Q ) : ~0(P' - - Q ' ) folgt, daft P - - Q ~ P ' - - Q ' , also P + Q ' ~ P'+Q, und welter, wenn C nieht hyperelliptiseh ist, dab entweder P =- P ' , Q : Q' oder P ---- Q, P ' =- Q'. Die Abbfldnng (P, Q) -> ~ 0 ( P - - Q ) yon G>~ C auf V ist also auBerhalb der Diagonale y o n C • G eineindeutig, falls C nieht hyperelliptisch ist; in diesem Fall hat also V die Dimension 2. Wenn C hyperelliptiseh ist, so gibt es einen Automorphismus h von G derart, dab ffir jeden P u n k t P y o n C der Divisor P + h ( P ) hyperelliptiseh ist; dann ist, wenn c irgendein hyperelliptiseher Divisor ist, ~ (h (P)) -~ r (c) - - ~0(P) ffir jeden P u n k t P ; indem dann N ' : h (N) gesetzt wird, sind M, N ' unabhgngige P u n k t e auf C fiber k, und ~o( M - N)
[1957a]
317 Z u m Beweis des Torellischen Satzes
43
nichts anderes als ~(M + N') --~0(c), also g die aus W2 durch die Translation --~(c) entstehende Fl~ehe. Nun sei a ein Punkt in der Menge 9~; nach den Hflfss~tzen 1 und 2 muB jedenfalls a auf V liegen; es ist also a = ~ ( P - Q) ~= 0, und O ~ Oa ist Vereinigungsmenge yon W~(p) und W*_~(Q). Da die letzteren Mannigfaltigkeiten nach Hilfssatz 3 keine Translation in sich selbst zulassen, so gehSrt umgekehrt dann und nur dann ein solcher Punkt der i~Ienge 9~ an, wenn W~(p) mit W*_~(Q)nicht zusammenf~llt, also (nach Hflfssatz 3) wenn P + Q nieht hyperelliptisch ist. Folglich ist wie behauptet 2 ---- V - - {0}, falls C nicht hypereUiptisch ist; im hyperelliptisehen Fall besteht dagegen 2[ aus allen Punkten yon V - - {0}, die sich nieht in der Form ~ (P - - Q) mit hyperelliptischem P + Q schreiben lassen. Wenn nun P - t - Q hypereUiptisch ist, so ist Q -~ h(P), wo h dieselbe Bedeutung hat wie oben; es sei Vt der Ort des Punktes qJ(M - - h ( M ) ) -~ 2~(M) - - ~(c) fiber/c bei generischem M fiber k auf C. Es gibt bekanntlich Punkte P a u f G derart, dab P ~- h (P); also liegt 0 auf der Kurve Vt. Es ist dann 2 ---- V - - V1, W.z.b.w. 6. I m Falle g > 5 ist nun der Hauptsatz eine unmittelbare Folge der oben bewiesenen Hilfss~tze. Oben ist schon die Sehar {Oa} in invarianter Weise auf der polarisierten Mannigfaltigkeit J gekennzeichnet worden. Es sei T = Ot eine beliebig gew~hlte Mannigfaltigkeit aus dieser Schar. Die Menge der Punkte a :4= 0 auf J mit der Eigensehaft, dab T ~ Ta zwei Komponenten hat, wovon keine dureh eine nieht-identische Translation in sich selbst fibergeht, ist dann nichts anderes als die im Hflfssatz 4 definierte Menge 9/; damit ist aueh die Fl~che V als abgeschlossene HiJlle yon 9~ (in der Zariskisehen Topologie anf J) in invarianter Weise definiert. Die Menge aller Komponenten der Durchsehnitte T ~ Ta, wenn a die Punktmenge V - {0} durehl~uft, besteht dann aus allen Mannigt-r wenn P, Q die Kurve C durchlaufen. Es sei faltigkeiten Wt +v(p), W* nun X eine beliebig gewfihlte Mannigfaltigkeit aus dieser Menge; es sei Y die Menge aller Translationen, die X in eine Mannigfaltigkeit derselben Menge fiberftihren. Aus dem Hflfssatz 3 folgt sofort, dab Y eine Kurve ist, die aus der Kurve ~0(C) dureh eine Abbfldung der Form u -~ + ( u - ~(P)) entsteht, wo P ein Punkt yon Gist. Wir haben also eine explizite Konstruktion angegeben, wodurch auf der polarisierten l~Iannigfaltigkeit J die Schar der aus ~ (C) dutch eine Abbfldung u --> • u + a entstehenden Kurven in invarianter Weise gekennzeiehnet ist. Damit ist der Hauptsatz ffir g > 5 vollst~ndig bewiesen. 7. Um die F~lle g --~ 3 und g ~-- 4 zu behandeln, werden wir zuerst die tangentialen linearen Mannigfaltigkeiten zu den Mannigfaltigkeiten W~ bestimmen. Hilfssatz 5. Es sei F eine Abbildun9 einer Mannig]altiglceit U der Dimension n in eine Mannig/altigkeit V; es sei ]c ein Definitionsk6rper /at U, V, F ; es sei u
318 44
[1957a] Andrd Weil
ein generischer Punier yon U abet k; der Punlct v ~ - F ( u ) sei ein]ach au/ V. Unter diesen Umstdnden ist k (u) dann und nur dann separabel algebraisch ~eber k (v), wenn die zu F i m Punkte u tangentiale lineare Abbildung vom Rang n ist. Bekanntlich ist k(u) dann und n u r dann separabel algebraiseh fiber k(v), wenn es keine niehtverschwindende Derivation von k(u) fiber k(v) gibt. Die Derivationen yon k(u) fiber ]c stehen aber in eineindeutiger Beziehung zu denjenigen Tangentialvektoren zu U im P u n k t u, die fiber k(u) rational sind; wenn ~ ein solcher Vektor ist, wird n~mlich durch die Gleichung D[/(u)] ---d / ( u ; t~), wo ] eine beliebige fiber k definierte numerische F u n k t i o n auf U ist u n d d] das Differential yon ] bedeutet, eine Derivation D yon k (u) definiert; und ]ede Derivation yon k(u) fiber k li~Bt sich so definieren. Die in dieser Weise d e m Vektor ~ zugeordnete Derivation D verschwindet dann und nur dann auf k (v), wenn D durch die tangentiale lineare Abbfldung zu F in u auf 0 abgebildet wird. Die B e h a u p t u n g des Hflfssatzes folgt u n m i t t e l b a r daraus. N u n sei A eine abelsehe Mannigfaltigkeit der Dimension n; es sei T der tangentiale V e k t o r r a u m zu A in O; im Siane der Theorie der algebraisehen Gruppen ist T nichts anderes als die Liesehe Algebra von A, oder genauer der dieser Algebra untergeordnete Vektorraum. Wenn a ein beliebiger P u n k t auf A ist, so wird dureh die Translation u -~ u + a der V e k t o r r a u m T auf den zu A in a tangentialen V e k t o r r a u m Ta isomorph abgebildet; wir werden meistens T~ mit T dureh diesen Isomorphismus identifizieren. Bekanntlich wird dadurch der duale V e k t o r r a u m zu T m i t dem Vektorraum der translationsinvarianten Differentiale (oder, was dasselbe ist, der Differentiale 1. Gattung) a u f A identifiziert. Es sei co ein solehes Differential; es sei F eine Abbildung einer Mannigfaltigkeit U in A. Dem Differential w @ird durch die zu F transponierte Abbfldung $'* ein Differential F ' c o auf U zugeordnet, das durch die Gleichung
F*o~(u; ~) = o~(F(u); F ' , ) definiert werden kann, wo u ein generischer P u n k t auf U, D ein Tangentialv e k t o r zu U in u, und F" die zu F in u tangentiale lineare Abbildung bedeuten. D a n n ist bekanatlich F * co ein Differential 1. G a t t u n g auf U; da wir hier diese Tatsache nur im Falle der kanonischen Abbildung einer K u r v e in ihre Jacobische Mannigfaltigkeit benutzen wollen, so brauehen wir auf ihre Begrfindung nieht n~her einzugehen; in diesem Fall ist sie unmittelbar klar. B e t r a c h t e n wir wieder eine K u r v e C v o m Geschlecht g, ihre Jaeobische Mannigfaltigkeit J und ihre kanonische Abbildung ~ in J ; es sei k ein DefinitionskSrper ffir C, J , ~. Es sei M irgendein P u n k t auf C; da ~ in M biregul~r ist, so wird durch die zu ~ in M tangentiale lineare Abbfldung die Tangente zu C in M auf eine Gerade im tangentialen V e k t o r r a u m zu J in ~0(M) abgebildet. Dieser V e k t o r r a u m ist oben mit dem tangentialen Vektorraum zu J in 0 identifiziert worden; damit ist der Tangente zu C in M eine Gerade durch 0 in T zugeordnet, die wir ira folgenden mit t (M) bezeichnen wollen. Es ist
[1957a]
319 Zum Beweis des Torellischen Satzes
45
leicht einzusehen (z.B. durch die Wahl geeigneter affiner Repr/~sentanten ffir die in B e t r a c h t k o m m e n d e n a b s t r a k t e n Mannigfaltigkeiten), daB, wenn M" eine Spezialisierung yon M fiber k ist, t (M') die eindeutig bestimmte Spezialisierung von t (M) fiber M --->M' in bezug auf k ist; also ist t eine fiber k defiuierte Abbfldung (im Sinne der algebraischen Geometrie) von C in die Malmigfaltigkeit der Geraden dureh 0 in T. Es sei P der (g - - 1)-dimensionale projektive R a u m der unendlichfernen P u n k t e zu T ; die P u n k t e yon P stehen in eineindeutiger Beziehung zu den Geraden dureh 0 in T ; wir wollen mit ~ (M) den der Gerade t (M) zugeordneten P u n k t von P bezeichnen; x ist also eine Abbfldung yon C in P. Satz 3. Der Vektorraum der Di~erentiale 1. Gattung au/ J wird dutch die zu transponierte Abbildung q~* au] den Raum der Di~erentiale 1. Gattung au] C isomorph abgebildet. Da beide R/~ume dieselbe Dimension g haben, so brauchen wir nur zu zeigen, dab ~* o~ ~= 0 ist, wenn das Differential 1. G a t t u n g w auf J nicht 0 ist. Es seien M 1 . . . . . Mg g unabh/~ngige generisehe P u n k t e y o n C fiber k; es sei u = (M1 -t- 999 q- Mg). Da k ( M 1. . . . . Ms) fiber k(u) separabel algebraiseh ist, kSnnen wir Hilfssatz 5 auf die Abbildung (M 1 . . . . , Ms) -~ u von G • 999• G auf J anwenden. Daraus ergibt sieh sofort, dab die Geraden t (M1) . . . . . t (Ms) nicht alle in einer echten linearen Untermannigfaltigkeit von T liegen kSnnen. W e n n n u n q0* eo ~- 0 w/~re, so mfiBte (naeh der Definition von ~* oJ) t (M)ffir j eden P u n k t M auf C in der durch o~ -~ 0 b e s t i m m t e n H y p e r e b e n e in T liegen (wobei wir den R a u m der Differentiale 1. G a t t u n g auf J mit dem zu T dualen R a u m identifiziert haben). Dureh eine /~hnliche B e t r a c h t u n g kann m a n folgenden allgemeineren Satz beweisen, den wir hier nur nebenbei erw/~hnen. Es sei F eine Abbildung einer K u r v e C in eine abelsche Mannigfaltigkeit A der Dimension n; es sei k ein DefinitionskSrper ffir C, A, F ; es seien M1 . . . . . M s n unabh~ngige generische P u n k t e von C fiber/c und u ---- _~ (M 1 ~- 999 -b M~). U n t e r diesen Umst/~nden ist ]c (M1 . . . . . Ms) dann und nur dann separabel algebraiseh fiber/r (u), wenn durch F * kein nichtverschwindendes Differential 1. Gattung auf A auf 0 abgebildet wird. J e d e r H y p e r e b e n e durch 0 in T entspricht im projektiven R a u m P ebenialls eine H y p e r e b e n e ; die H y p e r e b e n e in P, die in dieser Weise der durch o~ ----- 0 definierten H y p e r e b e n e in T entsprieht, werde durch H~ bezeichnet; dabei bedeutet oJ wie oben ein nichtverschwindendes Differential 1. Gattung a u f J (bei Identifizierung solcher Differentiale mit den Linearformen in T). Aus den Definitionen folgt sofort, dab H~ dam1 und nur dann den P u n k t ~(M) enth/~lt, wenn ~*eo im P u n k t M versehwindet. Es sei nun (~ol, . . . , ~og) eine Basis ffir den R a u m der Differentiale 1. G a t t u n g auf J, also ffir den zu T dualen R a u m ; die eo~ kSnnen dann als ein System (gew6hnlieher) K o o r d i n a t e n in T und als ein System homogener K o o r d i n a t e n in P b e t r a c h t e t werden.
320
46
[1957a] Angrd Weil
Es ist nach dem obigen klar, dab die Verh~ltnisse der Differentiale ~* w~ im Punkt M yon C, also die Werte ~*co~(M; ~) ffir einen beliebigen nichtverschwindenden Tangentialvektor ~ zu C in M, ein System homogener Koordinaten fiir den Punkt v(M) sind. Die Abbfldung ~ von C in P ist also diejenige, die zur kanonischen linearen Vollschar gehSrt. Bekanntlich folgt unmittelbar aus dem Riemann-Rochschen Satz, dab v ein Isomorphismus von C auf ~(C) ist, wenn C nicht hyperelliptisch ist. I m hyperelliptischen Fall ist, wenn M generisch auf C fiber k gew~hlt wird, k (M) separabel vom Grad 2 fiber k ( ~ ( M ) ) ; der nicht-identische Automorphismus von k(M) fiber k (3 (M)) ist dann derjenige, der oben in Nr. 5 (beim Beweise von Hflfssatz 4) mit h bezeichnet wurde, so dab M + h(M) ein hyperelliptischer Divisor ist. 8. Satz 4. E s 8el a ~ A1 + 999+ Ar ein positiver Divisor vom Grad r ~ g au] C. D a n n und nur dann ist a -~ ~ (a) ein ein/acher P u n k t yon Wr, wenn l (a) 1 ist; in diesem Fall ist die lineare Untermannig/altigkeit yon P, die in P der tangentialen linearen Mannig/altiglceit zu Wr in a entspricht, der Durchschnitt der Hyperebenen H ~ , wenn w die Gesamtheit der Di~erentiale 1. Gattung au] J durchl~u/t, ]t~r welche (q~*o~) ~ a. Zunachst sei a einfaeh auf W~; es sei Tr die tangentiale lineare Mannigfaltigkeit zu W~ in a und L die ( r - 1)-dimensionale lineare Untermannigfaltigkeit yon P, die T~ entsprieht. Wenn M ein generiseher Punkt von C tiber k(A1 . . . . , A~) ist, so enth~lt W~ den Ort yon ~(M + A S + 999+ A~) fiber diesem K6rper; die Tangente zu dieser Kurve im Punkte a ist aber nach dem oben Bewiesenen die Gerade t(A1); also enth~lt T~ diese Gerade, und L enth~lt den Punkt v(A1). Nehmen wir an, es sei l(a) ~_ 2; dann gibt es zu jedem Punkt BI auf C einen Divisor b in der durch a bestimmten linearen Vollsehar derart, da~ b ~ B1; b l~Bt sieh also in der Form B1 + 999+ B~ sehreiben, und es ist ~(l}) -----~(a) -----a. Aus dem oben Bewiesenen folgt nun, dab v(B~) in L liegen muB. Da B~ beliebig ist, ist also die ganze Kurve v(C) in L enthalten, was unmSglieh ist. Demnaeh ist 1(a) ---- 1, wenn a einfaeh ist auf Wr; das Umgekehrte wurde in VA-40, prop. 18 bewiesen. Nun sei l(a) ~-- 1, also a einfach auf W~; wir haben oben bewiesen, da~ L die Punkte v ( A 1 ) , . . . , v(A~) enth~lt. Andererseits ist l ( [ - a) ---- g - r wegen des Riemann-Roehsehen Satzes; m. a. W. bilden die Differentiale co, die der Bedingung (~*r genfigen, einen Vektorraum der Dimension g - - r ; der Durchschnitt L" der Hyperebenen H~, wenn eo diesen Vektorraum dnrehl~uft, hat also dieselbe Dimension r - 1 wie L u n d enth~lt die Punkte ~(Ax), . . . . v (A~). l~alls diese Punkte in keiner linearen Mannigfaltigkeit niedrigerer Dimension liegen, folgt sehon daraus, dab L" mit L zusammenf~Ut. Das ist aber jedenfalls dann der Fall, wenn A~ . . . . . Ar fiber k unabh~ngig sind, da wir beim Beweis von Satz 3 sogar gezeigt haben, dai3 die Geraden t(A~) . . . . . t(Au) in keiner eehten liaearen Untermannigfaltigkeit von T liegen kSnnen, falls A~ . . . . . Ag fiber k unabh~ngig sind.
[1957a]
321 Zum Beweis des ToreUischen Satzes
47
Der allgemeine Fall folgt daraus duroh Spezialisierung. Wenn u ein genefischer Punkt yon W~ fiber k ist, so ist ni~mlieh die tangentiale lineare Mannigfaltigkeit zu W~ in a die eindeutig bestimmte Spezialisierung derjenigen in u fiber u ~ a in bezug auf k. Man braueht also nur zu beweisen, daB, wenn die Behauptung yon Satz 4 ftir die tangentiale lineare Mannigfaltigkeit zu Wr in a = ~ (a) riehtig ist, sie auch ffir jede Spezialisierung a' yon a riehtig bleibt. Es seien nun bl . . . . . bg_~ g - - r fiber k (a, a') unabhi~ngige generisehe positive Divisoren vom Grad g - r - 1; es sei K tier dureh die Komponenten dieser Divisoren fiber k erzeugte KSrper; a' ist noeh Spezialisierung von a fiber K. ' Die Bedingungen (~* to~) ~ a + bQ, (~* we) ~ a ' + bQ, bestimmen dann niehtverschwindende Differentiale %, w~ eindeutig bis auf konstante Faktoren; es ist dana unmittelbar einzusehen, dab bei der Spezialisierung a ~ a' fiber K die Hyperebene H ~ sieh ftir jedes ~ auf H ~ spezialisiert. Wegen der Wahl der b~ sind sowohl die co~ wie die co~ linear unabhangig; der Durehsehnitt der Hyperebenen H ~ wird also auf den Durchsehnitt der H ~ spezialisiert. Damit ist alles bewiesen. 9. Naeh diesen Vorbereitungen kehren wir zum Beweis des Torellisehen Satzes in den F~llen g = 3, g ---- 4 zurfiek. Erstens sei g ---- 3. Naeh Satz 4 ist dann O = W2 im meht-hyperelliptischen Fall singulariti~tenfrei. I m hyperelliptisehen Fall hat dagegen O genau einen singuli~ren Punkt c = ~(c), w o r irgendeinen hyperelliptisehen Divisor auf C bedeutet. Im letzteren Fall ist dana O_r eindeutig dadureh bestimmt, dalt es die einzige Mannigfaltigkeit in tier Sehar {O~} ist, die in 0 einen singuli~ren Punkt hat. Da diese Mannigfaltigkeit aus allen Punkten ~0(M A- N)
--
c
= ~o(M)
--
~
(h (N))
besteht, so ist sie niehts anderes als die im Hilfssatz 4 mit V bezeiehnete F1/~ehe. Der Beweis des Itauptsatzes im hyperelliptisehen Fall g ---- 3 1/Liftsieh darm genauso wie in Nr. 6 dureh Betraehtung der Komponenten der Durehschnitte T c-~ T~ ffir a in V - - {0} zu Ende ffihren, wobei T irgendeine Marmigfaltigkeit aus der Sehar {O,} (z.B. V selbst) bedeutet. Im nicht-hyperelliptisehen Fall sei wieder T = Ot eine beliebig gewKhlte M'annigfaltigkeit aus der Sehar {O~}. Es sei k o ein algebraiseh abgesehlossener gemeinsamer DefinitionskSrper ffir J und T; es sei k ein k o enthaltender DefinitionskSrper ffir C, ~0 und t. Da O dureh die Abbildung u --> ~o(~)- - u anf sich selbst abgebildet wird, so wird T durch die Abbildung u -->F(u) : q~(~) + 2 t - - u auf sich selbst abgebildet. Da O, also aueh T keine nieht-identische Translation in sich selbst zulassen, so ist E die einzige unter den Abbfldungen u -+ c - - u mit konstantem c, die T auf sich selbst abbildet; dadureh ist F nach Wahl yon T in invarianter Weise gekennzeichnet. Es ist klar, dal3 die Tangentialebenen zu T in den Punkten u und F (u) zueinander parallel sind, d. h. dab sie
322
[1957a]
48
Andrg Weil
bei der vorgenommenen Identifizierung der tangentialen Vektorr~ume zu J in u u n d in F (u) zusammenfallen. I m vorliegenden Fall ist 9 ein Isomorphismus yon C auf eine singularit~tenfreie K u r v e 4. Grades in der projektiven Ebene. Es sei u ein generischer P u n k t yon T fiber k; m a n k a n n u in der F o r m u = ~(M1 q- M~) q- t sehreiben, wo M1, M2 zwei unabh~ngige generisehe P u n k t e von C fiber k sind. Die Gerade durch v(M1) und v(M~) ist dann generiseh fiber k, woraus folgt, dab sie zwei weitere S e h n i t t p u n k t e ~ (N1), ~ (N~) mit T (C) hat, und dal~ je zwei ihrer S e h n i t t p u n k t e mit v(C) fiber k unabh~ngig shad; also sind je zwei der P u n k t e M1, M2, N1, N~ 1/nabh~ngige generisehe P u n k t e yon G fiber k; und der Divisor M1 q - M 2 Jr N1 q - N z gehSrt zu der kanonisehen Vollsehar. Es folgt nun unmittelbar aus Satz 4, dab es auf T auBer u genau 5 P u n k t e gibt, in denen die Tangentialebene zu T zur Tangentialebene zu T in u parallel ist; das shad n~mlieh die P u n k t e u" -= q~(N~ -~- N2) + t,
v~j -~ ~ ( M i ~- Nj) -~- t
( i , j ---- 1,2).
Dabei ist u' nichts anderes als E (u). Wenn wir also ffir v ehaen von u und F (u) versehiedenen P u n k t y o n T wahlen, in dem die Tangentialebene zu T zur Tangentialebene zu T in u parallel ist, so kann das nur einer der vier P u n k t e vi~ sein; u n d es ist dann u - v y o n der F o r m ~ ( M - N), wo M, N zwei unabhi~ngige generische P l m k t e yon C fiber k sind. Es ist aber klar, dab ein solcher P u n k t v fiber/co(u ) algebraisch sein muB, weft es sonst unendlieh viele solche P u n k t e g~be; u - v hat also hSchstens die Dimension 2 fiber k0. Da aus Hilfssatz 5 folgt, da~ der Ort von u - v fiber/C die dort mit V bezeichnete Fl~ehe ist, so muB u - - v denselben Ort V fiber Ico haben wie fiber k. Wir haben also die Fl~ehe V als Ort des Punktes u - - v fiber/c o konstruiert; dabei wurde u als generischer P u n k t yon T fiber k gewii~hlt, und v wurde in der oben angegebenen Weise definiert. Es sei nun u 0 ein beliebiger generischer P u n k t y o n T fiber k0; der Isomorphismus yon k o (u) auf/c o (%) fiber/co, der u a u f u 0 abbildet, li~Bt sieh zu einem Isomorphismus der abgesehlossenen Hfille yon /co(u) auf diejenige yon/c0(u0) erweitern. Dadureh ist ersichtlieh, daB es wieder genau vier P u n k t e auBer u0 und F (u0) gibt, in denen die Tangentialebene zu T zur Tangentialebene zu T in u o parallel ist, und daB, wenn v 0 einer dieser vier P u n k t e ist, der Oft von u0 - - Vo fiber/co die Flaehe V ist. Da nun wieder V in vSllig invarianter Weise konstruiert ist, verl~uft der Beweis genauso wie frfiher weiter. 10. Endlieh sei g -~ 4. W e n a C hyperelliptisch ist, so besteht die kanonisehe lineare Vollschar auf C aus den Divisoren der F o r m c + c' -t- c", wo c, c', c" hyperelliptisehe Divisoren shad; folglieh gibt es zu ehaem positiven Divisor a ---- A -{- A' + A " vom Grad 3 nur ehaen zweiten solehen Divisor a' derart, dab
[1957a]
323 Zum Beweis des ToreUischen Satzes
49
a ~ a' kanonisch ist, falls keiner der Divisoren A ~- A ' , A ~ A " , A" ~- A " hyperelliptisch ist. U n t e r Benutzung des Riemann-Roehsehen Satzes folgt dann ans Satz 4, dab die singularen P u n k t e von O genau die P u n k t e ~(r ~ M) sind, wo c einen hyperelliptisehen Divisor und M einen P u n k t auf C bedeuten. Die singularen P u n k t e irgendeiner Mannigfaltigkeit aus der Schar {Oa} bflden also eine Kurve, die sich yon ~ (C) nur dureh eine Translation unterscheidet. D a m i t wird offenbar der hyperelliptische Fall g -= 4 erledigt sein, sobald wir zeigen, wie m a n diesen Fall v o m nicht-hyperelliptischen Fall in invarianter Weise unterscheiden kann. Das wird daraus folgen, da~ im letzteren Fall O n u r einen oder zwei singulare P u n k t e hat. Von nun an setzen wir voraus, dal] C nicht hyperelliptiseh sei. Da v ein Isomorphismus von C auf 9 (C) ist, wollen wir die Bezeiehnungen dadurch vereinfaehen, dal~ wit C mit ~ (C) durch v identifizieren. Es ist dann C eine singularitatenfreie K u r v e v o m Grad 6 im 3-dimensionalen projektiven R a u m P . Aus dem Riemann-Rochschen Satz ist sofort ersiehtlieh, dal~ es wenigstens eine Flache F , vom Grad 2 in P gibt, die C enthalt, sowieeine F l a c h e F a v o m Grad 3, die C aber nicht F , enthalt. Da C v o m Grad 6 ist, so muB dann $'~ die einzige Flaehe v o m Grad 2 sein, die C enthalt; und es ist C ~ F , . F3. W e n n eine Gerade D mit C drei P u n k t e gemeinsam hat, so mug sie in $'3 liegen. Die in F2 enthaltenen Geraden bflden aber eine oder zwei Scharen der Dimension 1, je n a c h d e m F2 ein Kegel ist oder nicht. Folglich kSnnen zwei S e h n i t t p u n k t e einer solehen Gerade mit C n i c h t fiber k unabhangig sein; dabei ist wie frfiher mit k ein Definitionsk6rper ffir C, J , ~ gemeint. Also hat die Verbindungsgerade zweier unabhangiger generischer P u n k t e yon C fiber k keinen dritten Punk~ mit C gemeinsam 1). Die B e s t i m m u n g der singularen P u n k t e von O wird dureh folgenden Hflfssatz geleistet: Hilfssatz 6. E i n positiver Divisor a yore Grad 3 au/ C genCegt dann und nur dann der Bedingung [(a) ~ 2, wenn er sich in der F o r m a : C 9D schreiben lgtflt, wo D eine au/ F2 liegende Gerade ist, und der Schnitt C 9D au/ F~ berechnet wird. Es sei zuerst bemerkt, da~, wenn F2 ein Kegel ist, C den Scheitel von F.~ nicht enthalten k a n n ; wegen C ---- F 2 9~'3 mfil3te namlich sonst dieser Scheitel a u f C singular sein. N u n ist nach dem Riemann-Rochschen Satz die Bedingung l(a) _~ 2 m i t der folgenden gleichwertig: es muB eine Gerade D im R a u m P geben, derart dab fiir jede D enthaltende E b e n e H die Relation H 9C ~ a besteht. D a n n muB offenbar jede K o m p o n e n t e yon a sowohl auf C wie auf D liegen. W e n n D genau zwei P u n k t e A, B m i t F , gemeinsam hatte, so ware D 1 Im Falle der Charakteristik 0 kann bekanntlich nur dann die Verbindungsgerade zweier unabh~ngiger Punkte auf einer Kurve C einen dritten Punkt mit C gemeinsam haben, wenn C in einer Ebene liegt. Im Falle der Charakteristik p ~ 0 gilt dieser Satz nicht mehr; ein Gegenbeispiel wird durch den Ort des Punktes mit den homogenen Koordinaten (1, t~, t~2, t~3) geliefert.
324
50
[1957a] Andrd Weil
zu/~2 in A und in B transversal; eine generische Ebene durch D hatte dann in A und in B hSchstens die Sehnittmultiplizitat 1 mit C; also hatte a hSehstens die zwei Komponenten A, B, jede hSchstens mit dem Koeffizienten 1, was der Voraussetzung widersprieht. Es habe D genau einen Punkt A mit F~ gemeinsam; dann muB A auf C liegen, weft sonst a =- 0 ware; naeh der oben gemachten Bemerkung mull also A ein einfaeher Punkt yon 2' 2 sein. Wenn D nieht die Tangente zu C in A ware, so h~tte wiederum eine generisehe Ebene dureh D die Schnittmultiplizit~t 1 mit C in A, was wie oben zu einem Widersprueh ffihrt. Werm sehlielllieh D die Tangente zu C in A ist, aber nieht in F~ liegt, so nehmen wir fiir H die Tangentialebene zu F~ in A ; dann ist H 9F 2 ~- D' + D " , wo D', D'" zwei Geraden sind; falls F2 ein Kegel ist, ist D' -----D". Nun ist die Sehnittmultiplizit/~t von H mit C in A dieselbe wie die Schnittmultiplizitat yon C mit dem Zyklus D" + D'" in A auf der F1/~ehe F~; sic ist also gleich 2, weft C die Tangente D hat und folglieh sowohl zu D' wie zu D'" in A transversal ist. Demnach kann A hSehstens den Koeffizienten 2 in a haben, was wiederum der Voraussetzung widersprieht. Damit ist bewiesen, dall die Gerade D auf F2 liegen mull. Wenn die Ebene H dureh D geht, so ist t l . F2 = D +4- D', wo D' eine Gerade ist; und H 9C, im Raum P berechnet, ist nichts anderes als der Zyklus (D + D') 9 C, berechnet auf F~. Wenn H generiseh gew~hlt ist, so geht D ' dureh keine Komponente yon a; es mull also D . C ~ a sein; und, wenn es so ist, ist H . C ~ a ftir jede Ebene H dureh D. Wegen C ---- F2 9$'3 ist aber D 9C, auf F~ genommen, nichts anderes als D 9_Fs, bereehnet im Raum P ; D 9C hat also den Grad 3, so dall aus D . C ~ a die Gleiehung D 9C ~-- a folgt. Bei Berfieksiehtigung von Satz 4 zeigt Hilfssatz 6, dall die singularen Punkte yon O die Punkte ~ (D 9C) sind, wenn D die Gesamtheit aller auf F~ liegenden Geraden durehli~uft, und D 9C auf $'2 bereehnet wird. Nun folgt aus dem oben bewiesenen, dall, wenn zwei verschiedene Geraden D, D" auf F2 in einer Ebene H liegen, der Zyklus (D + D ' ) . G, aufF~ bereehnet, derselbe ist wie der Zyklus H 9 C, bereehnet in P ; da die Zyklen der Form H 9C kanonische Divisoren auf C sind, so sieht man, dall unter diesen Umst~nden die Relation D 9C + D" 9C ~ f besteht. Daraus folgt sofort, dall, wenn F2 ein Kegel ist, samtliche Divisoren D 9C miteinander/~quivalent sind; in diesem Fall hat also O genau einen singularen Punkt. Wenn F~ kein Kegel ist, so gibt es auf F 2 zwei Seharen yon Geraden; wenn D und D" nieht derselben Schar angehSren, so liegen sie in einer Ebene. Daraus folgt, dall die Divisoren D . C, wo D eine dieser Scharen durehl~uft, miteinander ~quivalent sind. Also hat dann O entweder einen oder zwei singul/ire Punkte; eine genauere Betraehtung wiirde zeigen, dall es in diesem Fall tatsachlich zwei verschiedene singul/~re Punkte auf 0 gibt. Nun sei H e i n e generisehe Ebene in P in bezug auf k; es sei k (H) ihr kleinster (k enthaltender) DefinitionskSrper; k ( H ) hat dann die Dimension 3 fiber k. Da H generiseh ist, so hat sie bekanntlieh genau sechs versehiedene Schnittpunkte mit C. Es seien M, M ' zwei dieser Schnittpunkte; ihre Verbindungs-
[1957a1
325 Z u m Beweis des Torellischen Satzes
51
gerade ist definiert fiber k(M, M'). Da H dureh diese Gerade geht, so hat k(H) hSehstens die Dimension I fiber k (M, M'); folglieh hat k (M, M') wenigstens die Dimension 2 fiber k; M, M' mfissen also fiber k auf C unabhi~ngig sein. Je zwei der 6 Schnittpunkte yon H mit C sind also fiber k unabh~ngig, woraus folgt, dab keine drei dieser Punkte auf einer Geraden liegen kSnnen. Wenn M, M', M " drei dieser Punkte sind, so bestimmen sie also eine Ebene, die keine andere als H sein kann; demnach mull k(H) in k(M, M', M") enthalten sein, was nur dann mSglich ist, wenn M, M', M " fiber k auf C unabh/ingig sind. Damit ist gezeigt, dall je drei der Schnittpunkte yon H mit C fiber k auf C unabh~ngig sind. Setzen wir H . C ---- ~ M i , und ui~h : 9(Mi + Mj + Mh) ftir jedes i
Tripel (i, j, h). Nach Satz 4 sind die Tangentialhyperebenen zu O in den 20 Punkten ui~-h alle zueinander parallel. Wenn umgekehrt die Tangentialhyperebene zu O in einem Punkt v ---- ~(N + N' + N") zu den Tangentialhyperebenen zu 0 in den Punkten u~j-a parallel ist, so mull naeh Satz 4 H 9C ~ N + N' W N " sein, woraus folgt, d a l l v einer der Punkte ui~h ist. Die Punkte u ~ i h - ucj,h, shad nun die Punkte
O, ~o(M1--M~), q~(MI + M~--Ms--M4), ~o(MI + M2+ M 3 - - M 4 - - M s - - M s ) , sowie diejenigen, die daraus durch eine beliebige Permutation der M~ entstehen. Die Punkte ~ (M, - - M j ) haben die Dimension 2 fiber k; wir werden beweiseh, dall sie dadureh unter den Punkten u i ~ h - ue~,h, charakterisiert sind. Zungchst shad die Punkte uiCh diejenigen, wo die Tangentialhyperebene zn 0 zur Tangentialhyperebene zu k in u~2a parallel ist; da diese Bedingung nur endlieh viele Punkte bestimmt, so sind sie alle algebraiseh fiber k(Ulzs). Da u m gellerisch auf O ist, so haben alle Punkte uijh --ue~,~, h6chstens die Dimension 3 fiber k. Welter ist, weft ~ Mi ein kanonischer Divisor ist: i
~0(M1 ~- M2-~- Ms - - M 4 - M5 - - M s ) -= 2 ~P(M1 -~-M~ + M 3 ) - - q (f) = 2 ux2s--(f); also ist Ulna algebraiseh fiber k ( u l = - u45e); demnach hat der letztere KSrper die Dimension 3 tiber k; dasselbe gilt dann ffir alle daraus dureh Permutation der M, entstehenden Punkte. Der Beweis daftir, dall der Punkt q(M 1 + M ~ - M s - Ma) ebenfalls die Dimension 3 hat, ist etwas umstgndlieher. Er beruht auf folgendem Hilfssatz: Hilfssatz 7. Es gibt eine Spezialisierung (M~) yon (M~) abet k , / a t welch, M~, M'3 abet k unabhangig sind und M~ = M'~ ist. Es seien M~, M~ zwei unabhgngige generische Punkte yon C fiber k. Da Ma, Mi, Ma fiber k unabh/~ngig sind, so kann man sie jedenfalls auf M~, M~, M', spezialisieren und diese Spezialisierung zu einer Spezialisierung (H', M~) yon (H, M~) erweitern0 wo also M~ = M~ ist. Dann ist H' 9C = ~ M'~, also insbesondere H" 9C ~ M~ + M'~, woraus folgt, dall H" dureh die Verbindungs-
326 52
[1957a] Andrd Weil
gerade D yon M~ und M'8 geht, falls diese P u n k t e versehieden sind, und durch die T~ngente D zu C in M~, falls M~ ~ M~ ist; die dadurch definierte Gerade D ist in beiden F~llen rational fiber dem KSrper K = k (M~, M~). N e h m e n wir an, dab die P u n k t e M~, M '3 fiber/r nicht unabh/~ngig seien; dann hat K die Dimension 1 fiber k. Die P u n k t e M'~, und insbesondere M~ und M's, sind alle algebraisch fiber dem kleinsten DefinitionskSrper k (H') ffir H", da M~, M'2 fiber k unabh/~ngig sind, hat demnach k(H') mindestens die Dimension 2 fiber k; H" ist also nieht algebraisch fiber K und ist generisch fiber K in der Schar der durch D gehenden Ebenen. Es seien F ~ 0, F ' = 0 lineare Gleichungen ffir D mit Koeffizienten in K ; die Gleichung ffir H ' 1/~Bt sich in der F o r m F -k tF' = 0 sehreiben, wo t die Dimension 1 fiber K hat. Der P u n k t M~ kann nicht auf D liegen, weil sonst D die Verbindungsgerade der unabh/~ngigen P u n k t e M~, M~ w/~re und keinen dritten P u n k t M~ mit C gemeinsam h/~tte (bzw. nicht die Tangente zu C in M~ sein kSnnte); H" ist dann die durch D und M~ bestimmte Ebene, so dab K(t) in K(M~) enthalten und M' generischer P u n k t von C fiber K ist. Wegen der Gleichung H'.C--M1--M'3=2M~+M~JrM~ folgt daraus sofort, dab M~ fiber K(t) inseparabel algebraisch sein muB, und zwar hSehstens v o m Grad 4. Wenn K (M~) fiber K (t) rein inseparabel (vom Grad 2, 3 oder 4) w/~re, so h/~tten der K 6 r p e r K (M~) und die K u r v e C das Gesehleeht 0. Es muB also K (M~) rein inseparabel vom Grad 2 sein fiber einer separablen Erweiterung K ' yon K (t) vom Grad 2; dann ist aber K ' hyperelliptiseh, und K(M~) ist mit K" fiber K isomorph; also ist C hyperelliptisch, was unserer Voraussetzung widerspricht. Naeh dem Hflfssatz hat also ~ ( M I + M s - M s - M4) die Spezialisierung ~(M~ - - M~) fiber k, mit unabh/~ngigen M~, M 'a fiber/~; letzterer P u n k t hat die Dimension 2 fiber k; wenn der erstere keine gr6Bere Dimension fiber k h/~tte, so w~re die Spezialisierung eine generische; der erstere P u n k t lieBe sieh dann selbst in der F o r m ~ ( N - N') sehreiben mit unabh/~ngigen N, N'; und es w/~re M 1 + M s -k N'~-~ M8 + M4 q- N. Dabei kann nicht l(M 1 -k M2 -k N') > 1 sein; dann mfiBte n~mlich nach Hflfssatz 6 die Verbindungsgerade von M1 und M~ auf F~ liegen, was bei der Verbindungsgerade zweier unabh/~ngiger P u n k t e auf C nicht der Fall sein kann. Es muB also M~ q - M s + N ' - ~ Ms + M4 Jr N sein, was offenbar unm6glich ist. Daraus ergibt sieh folgende invariante Konstruktion ffir die Flgehe V. Man wahle eine beliebige Marmigfaltigkeit T---- Ot aus der Sehar {0~}; es sei ]co ein algebraiseh abgesehlossener Definitionsk6rper ffir J und T ; es sei u o ein generiseher P u n k t von T fiber k 0. U n t e r den 20 P u n k t e n von T, in denen die Tangentialhyperebene zu T zur Tangentialhyperebene zu T in u 0 parallel ist, gibt es 9 P u n k t e mit der Eigenschaft, daB, wenn Ul einer derselben ist, der P u n k t v ~- u~ - - u 0 die Dimension 2 fiber k 0 hat. D a n n ist g der Ort yon v fiber k 0. Das ist n mlich in dem oben Bewiesenen enthalten, falls k o zu
[1957a]
327
Z u m Beweis des Torellischen Satzes
53
gleicher Zeit ein DefinitionskSrper ffir C, q u n d t i s t . W e n n das nicht der Fall ist, w~hlen wir einen k 0 enthaltenden DefinitionskSrper ffir C, q und t. Es i~ndert sich nichts, wenn wir u o durch einen anderen generischen P u n k t yon T fiber k o ersetzen; wir dfirfen also annehmen, d a b uo sogar fiber k generisch ist; m i t denselben Bezeichnungen wie oben kSnnen wir dann u0 = ulna ~- t setzen. Die P u n k t e , wo die T a n g e n t i a l h y p e r b e n e zu T m i t derjenigen in u o parallel ist, sind d a n n die P u n k t e ui~a + t; wenn u~ irgendeiner dieser P u n k t e ist, so ist er algebraisch fiber ko(Uo) ; der P u n k t (uo, u~) h a t also dieselbe Dimension fiber k wie fiber ko, n~mlich 3; folglich h a t er einen Ort fiber k, u n d zwar denselben wie fiber k o; u n d die Dimension yon uo - - u~, sowohl fiber k wie fiber ko, ist nichts anderes als die Dimension des Brides dieses Ortes bei der Abbridung (x, y) -> x - - y y o n J • J a u f J . Oben ist bewiesen worden, dab es 9 P u n k t e u~ gibt, ffir welche dieses Bild die Dimension 2 h a t und dab es dann die Fl~che V ist. D a m i t ist unsere K o n s t r u k t i o n v611ig gerechtfertigt. Den Beweis des H a u p t s a t z e s im vorliegenden Fall k6nnen wir nun genau wie oben fortsetzen. Es b r a u e h t k a u m gesagt zu werden, dab ein einheitlicherer Beweis sehr zu wfinsehen w~re.
[1957b] Hermann Weyl (1885-1955) (avec C. Chevalley)
(~Quand Hermann Weyl et Hella annonc~rent leurs fian~ailles, l'6tonnement rut g~n~ral que ee jeune homme timide et peu loquace, ~tranger aux cliques qui faisaient la loi dans le monde math~matique de GSttingen, efit remport~ le prix convoit~ par tant d'autres. Ce n'est que peu h peu que l'on comprit ~ quel point Hella avait eu raison dans son choix... 1,~ Peut-Stre les v~rit~s math~matiques, comme les femmes, font-elles leur choix entre ceux qu'elles attirent. Est-ce le mieux dou~ qu'elles choisissent, ou le plus s~duisant ? celui qui les d~sire le plus ardemment, ou eelui qui les a l e mieux m~rit~es ? Elles semblent se tromper parfois; souvent il faut du temps pour s'apercevoir qu'elles ont eu raison. Timide, peu loquaee, ~tranger aux cliques, tel apparaissait donc Hermann W e y l / t ses d~buts; tel il devait rester au fond de lui-m~me, en d~pit des succ6s d'une brillante earri~re. Comme beaucoup de timides une fois rompues les barri~res de leur timiditY, il ~tait capable d'enthousiasme et d'~loquence: (~Ce soir-lh, dit-il en raeontant sa premiere rencontre avec celle qu'il devait 6pouser 2, je d~crivis l'incendie d'une grange auquel je venais d'assister; elle me dit plus tard qu'h m'~couter elle s'~tait ~prise de moi aussit6t. ~) Ses propres confidences nous le montrent profond~ment influen~able aussi, jusque dans sa pens~e la plus intime : (~Mon tranquille positivisme 1 E x t r a i t des paroles p r o n o n e e e s p a r C o u r a n t a u x obs~ques de Hella W e y l le 9 s e p t e m b r e i9t~8. Cette c i t a t i o n , e o m m e plusieurs a u t r e s p a r la suite, est tir6e d ' u n e notice inedite eonsaeree p a r H e r m a n n Weu ~t la m 6 m o i r e de Hella Weyl. Nos a u t r e s citations p r o v i e n n e n t des p u b l i c a t i o n s de W e y l . Reprinted from Enseign. Math. IIl, 1957, pp. 157-187, by permission o f the editor.
329
330
[1957b] 158
C. C H E V A L L E Y
ET
A.
WEIL
fut 6bran]6 quand je m'6pris d'une jeune musicienne d'esprit trbs religieux, membre d'un groupe qui s'6tait form6 autour d'un h6g61ien connu... Peu apr6s, j'6pousai une 616re de Husserl; ainsi, ce fur Husserl qui, me d6gageant du positivisme, m'ouvrit l'esprit ~ une conception plus lil~re du monde. )) I1 avail alors vingt-sept ans. C'est ainsi qu'on volt se dessiner, vers l'6poque de son mariage, quelques-nns des prineipaux traits d'une des personnalit6s math6matiques les plus marqnantes et attachantes de la premi6re moiti6 de ce si6cle, mais aussi de l'une des plus difficiles h serrer de pr6s. (~A country lad of eighteen ~, un gars de eampagne de dix-huit ans, ainsi se d6crit-il lui-m~me h son arriv6e h G6ttingen. (~ J'avais choisi cette universit6, dit-il, principalement parce que le directeur de mon lye6e 6tail un cousin de Hilbert et re'avail donn6 pour eelui-ei une lettre de reeommandation. Mais il ne me fallut pas longtemps pour prendre la r6solution de lire et 6tudier tout ce que cet homme avail 6crit. D6s la fin de ma premi6re ann6e, j'emportai son Zahlbericht sous mort bras et passai les vaeances ~ le lire d'un bout ~ l'autre, sans aucune notion pr6alable de th6orie des nombres ni de th6orie de Galois. Ce fnrent les mois les plus heureux de ma vie... 3 )) Un peu plus lard, ce sont les joies de la d6converte: (~Un nouvel 6v6nement fur d6cisif pour moi: je fis une d6couverte math6matique importante. Elle concernait la loi de r6partition des fr6quences propres d'un syst6me continu, membrane, corps 61astique ou 6ther 61ectromagn6tique. Le r6sultat, conjectur6 depuis longtemps par les physiciens, semblait encore bien loin alors d'une d6monstration math6matique. Tandis que j'6tais fi6vreusement occup6 ~ mettre mon id6e au point, ma lampe h p6trole avait commene6 h fumer. Quand je terininai, une 6paisse pluie de flocons noirs s'6tait abattue sur mon papier, sur rues mains, sur mon visage.~) A c e moment, il est d6j& privatdozent /~ GOttingen. Bientbt c'est le mariage, la ehaire ~ Zurich, la 3 ,, De r o u t e rues e x p S r i e n c e s spiritnelles, ~crit-il ailleurs, celles qui m ' o n t c o m b l e de la p l u s g r a n d e joie f u r e n t , e n i 9 0 5 , q u a n d j ' ~ t a i s 5 t u d i a n t , l ' S t u d e d u Zahlberichl ct, en i 9 2 2 , la l e c t u r e de m a i t r e E c k h a r t q u i m e r e t i n t fasein6 p e n d a n t u n s p l e n d i d e h i v e r en E n g a d i n e .
[1957b]
331 HERMANN
WEYL
(1885-1955)
159
guerre. Au bout d'un an de garnison h Sarrebruck (comme simple soldat, pr6cise-t-il), le gouvernement suisse obtient qu'il soit rendu ~ son enseignement /~ l'Ecole polytechnique f6d6rale. <<Je ne puis gu6re me souvenir d'un instant de joie plus intense que le beau jour de printemps, en mai 1916, oh Hella et moi franchimes la fronti6re suisse, puis, arriv6s chez nous, descendimes de nouveau jusqu'au lac ~ t r a v e r s la belie ville paisible. ~ I1 reprend ses travaux. Un cours profess6 h Zurich sur la relativit6 parait en volume, en 1918; c'est 18 c616bre Raum, Zeit, Materie, qui connait cinq 6ditions en cinq ans et, profitant de la vogue extraordinaire du sujet jusque parmi les profanes, r6pand le nora de Weyl bien au-del~ du monde des math6maticiens o~ sa r6putation n'6tait plus h faire. Les ogres de chaires viennent d'un peu p a r t o u t ; celle de Gbttingen en 1922, off il s'agissait de la succession de Klein, fur l'occasion pour lui d'un d6bat de conscience particuli6rement dimcile. A y a n t retard6 sa d6cision t a n t qu'il pouvait, a y a n t encore au dernier m o m e n t pareouru avec sa femme les rues de Zurich en pesant sa r6ponse, il partit enfin au bureau de poste pour t616graphier son acceptation. Arriv6 devant le guiehet, ce fur un refus qu'il t616graphia; il n'avait pu se r6soudre ~ 6changer sa tranquillit6 zurichoise contre les incertitudes de l'Allemagne d'apr6s guerre. < Mais en 1929, quand GOttingen lui ogre la succession de Hilbert, il se laisse tenter. <~Les trois ann6es qui suivirent, dit-il, furent les plus p6nibles que Hella et moi ayons connues. ~>C'est le nazisme, d'abord imperceptible nuage ~ l'horizon, qui grandit r u e d'ceil, s'abat en trombe sur l'Allemagne en d6sarroi, y recouvre t o u t de bone sanglante. Par bonheur pour H e r m a n n Weyl, l'Institute for Advanced Study de Princeton, nouvellement cr66, ogre de le sauver du d6sastre. I1 h6site. I1 aceepte, il refuse. I1 aceepte de nouveau l'ann6e suivante; c'est de Zurich qu'i] envoie sa d6mission ~ G~ttingen en 1933 et qu'il part pour l'Am6rique. I1 n'eut done pas 5 subir ce stage souvent long, parfois humiliant et p6nible, que les circonstances ont impos6 h beaucoup de savants r6fugi6s aux Etats-Unis. La chaire de l'Institute
332
[1957b] 160
C. C t t E V A L L E Y
ET
A.
WElL
lui assura d'embl6e le confort mat6riel et la situation de premier plan darts le monde scientifique am6ricain auxquels tout, certes, lui donnait droit. Ce furent, dit-il, des ann6es heureuses que cel]es qu'il passa ~ Princeton. Sans doute ne s'aecoutuma-t-il jamais ~ porter ais6ment ce qu'il appelle <~le joug d'une ]angue 6trang6re ~>. Mais, grace au respect e t ~ l'affection qui l'entour6rent d~s l'abord, il se sentait enfin chez lui; et on sent percer nouveau le gars de campagne des premiers jours de GSttingen lorsqu'il d6peint le plaisir qu'il 6prouva, en 1938, ~ poss6der son lopin de terre et ~ y bgtir sa maison. Si la mort de sa femme, en i948, le d6chira cruellement, un second mariage, quelque temps apr~s, lui fit retrouver son 6quilibre. A y a n t pris sa retraite l'Institute, il partagea d6sormais son temps entre Princeton et Zurich. Une attaque cardiaque l'emporta h l'improviste, peu apr~s les f6tes de son soixante-dixi6me anniversaire. A son arriv6e en Am6rique, il avait d6jh donn6 en math6matique le meilleur de lui-m~me, et il le savait. Pour tout autre que lui, la tentation efit 6t6 grande de se reposer sur ses lauriers, de s'abandonner ~ un rSle de ~ pontife ~. Combien n'en est-il pas dont toute l'activit6, pass6 un certain Age, eonsiste g aller de commission en commission, pour y discuter gravement des m6rites de t r a v a u x de ~ jeunes ~> qu'ils n ' o n t pas lus, qu'ils ne connaissent que par our'-dire! Hermann W e y l se faisait une bien autre et bien plus haute id6e de son m6tier de professeur. I1 vit que Princeton seul, ~ notre 6poque, peut 6tre ce qu'ont 6t6 autrefois Paris, puis GSttingen: un centre d'6ehanges, un ~ clearing-house ~>des id6es math6matiques qui circulent de par le monde. Rappelant l'intense vie math6matique qui s'6tait d6velopp6e autrefois ~ GSttingen sous l'influence dominante de Hilbert, il a 6crit: <~Les id6es font boule de neige en un pareil point de condensation de la recherche ~>;et il ajoute: (~Nous avons assist6 h quelque chose de semblable ici ~ Princeton pendant les premi6res ann6es d'existence de l'Institute for Advanced Study. ~> S'il en a ~t6 ainsi, c'est en grande partie a lui qu'en revient le m6rite. I1 se donna pour tache prineipale de se maintenir au courant de l'actualit6, de renseigner et 6clairer les chercheurs, de leur servir d'interpr6te, de comprendre mieux qu'eux ce qu'ils faisaient ou essayaient de faire; il s'y consacra en route
[1957b]
333 HEttMANN
WEYL
(1885-1955)
i61
modestie, conscient de faire oeuvre utile, conscient d'y ~tre irrempla~able. Dans sa production, qui, pendant toute cette p~riode, reste abondante et d'une extraordinaire varietY, on retrouve la trace de ses lectures, des s6minaires et discussions auxquels il prenait part, des probl~mes sur lesquels de tous c5t6s on sollicitait ses avis. Parmi ces travaux, il n'en est gu~re qui n'61ucide un point difficile ou ne comble une lacune f~eheuse. Cette activit6 s'est poursuivie jusque dans ses derni~res ann6es. Par une supr6me coquetterie peut-~tre, sa dernibre publication aura 6t6 une ~dition rajeunie, compl~tement refondue, de son premier livre, ]ivre toujours utile, encore actuel, auquel par eette r6vision il a donn6 une vitalit6 nouvelle. Qui de nous ne serait satisfait de voir sa carriSre scientifique se terminer de m6me ?
Un Prot6e, qui se transforme sans eesse pour se d~rober aux prises de l'adversaire, et ne redevient lui-m~me qu'apr~s le triomphe final: telle est l'impression que nous laisse souvent H e r m a n n Weyl. N'est-il pas all6, pouss6 par le milieu sans doute, par l'occasion, mais aussi par (~l'inqui6tude de son g~nie ~), jusqu'~ se muer en logiclen, en physieien, en philosophe ? L'axiome Ne sutor ultra crepidam nous interdit de le suivre si loin en ses mStamorphoses. Mais, dans son oeuvre math6matique m~me, il n'est que trop fr6quent qu'il vous glisse entre les mains lorsqu'on eroit le mieux le saisir; et il faut avouer que la t~che de ses leeteurs n'en est pas faeilit~e. I1 est vrai qu'il appartient une p6riode de transition dans l'histoire des math~matiques et qu'il s'en est trouv6 profond~ment marqu6. Souvent il a pu prendre un plaisir grisant ~ se laisser entrainer ou ballotter par les eourants opposes qui ont agit6 eette 6poque, stir d'ailleurs au fond de lui-m~me (comme lorsqu'il s'abandonna un m o m e n t l'intuitionnisme brouw6rien) que son bon sens foncier le garantirait du naufrage. Son oeuvre a grandement eontribu6 ce changement de vision qui a fait passer de la math~matique elassique, fond6e sur le hombre r6el, ~ la math6matique moderne, fond6e sur la notion de structure. L'emploi syst6matique et tout abstrait du rev~tement universel, la notion de vari6t6 analytique
334
[1957b] 162
C. C H E V A L L E Y
ET
A.
WElL
eomplexe, l'emploi courant et la popularisation, jusque parmi les physiciens, de l'alg6bre vectorielle et du concept d'espaee de repr6sentation d'un groupe, t o u t cela vient a v a n t tout de lui. Mais, s'il 6tait trop 616ve de Hilbert pour ne pas inelure parmi ses outils la m6thode axiomatique, s'il 6tait trop math6matieien aussi pour en d6daigner les sucd6s (son chaleureux 61oge de l'ceuvre d ' E m m y Noether serait lh, si besoin 6tait, pour en faire foi), ee n'6tait pas /~ elle qu'allaient ses sympathies. I1 y v o y a i t <~le filet dans lequel nous nous effor~ons d ' a t t r a p e r la simple, la grande, la divine Id6e ~; mais, darts ce filet, il semble avoir toujours craint que l'on n'attrapfit que des eadavres. A la dissection impitoyable sous le jour cru des projecteurs, il pr6f6fair, en bon romantique, le jeu troublant des analogies, auquel se prdte si bien le langage de la m6taphysique allemande qu'il affeetionnait. Plut6t que de saisir l'id6e brutalement au risque de }a meurtrir, il aimait bien mieux la guetter dans la p6nombre, l'accompagner dans ses 6volutions, la d6crire sous ses multiples aspects, dans sa vivante complexit6. Etait-ee de sa faute si ses lecteurs, moins agiles que lui, 6prouvaient parfois quelque peine le suivre ?
<~Le v6ritable principe de Dirichlet, a dit Minkowski dans un passage que Weyl citait volontiers, ce fur d ' a t t a q u e r les problbmes au moyen d'un m i n i m u m de calcul aveugle, d'un m a x i m u m de r~flexion lucide. ~ E t Weyl a 6crit de son maitre Hilbert : <~Un trait earact~ristique de son oeuvre, c'est sa m6thode d ' a t t a q u e directe; s'affranchissant de t o u t algorithme, il revient toujours au probl~me tel qu'il se pr6sente dans sa puret~ originelle. ~ En deux ou trois occasions, il a atteint pleinement luim~me ~ eet id6al de perfection classique, par exemple dans son travail de 1916 sur l'6gale r@artition modulo 1, et encore darts ses m~moires j u m e a u x sur les fonctions presque pgriodiques et sur les repr6sentations des groupes compacts. Comme il est naturel, ee sont lh, parmi ses travaux, eeux qu'on relit avee le plus de plaisir, ceux dont il est le plus facile aussi de rendre compte. Aussi est-ce par eux que nous commencerons, renongant /~ un ordre logique impossible h suivre lorsqu'il s'agit d'analyser
[1957b]
335 HEBMANN
WE YL
(1885-1955)
163
une oeuvre aussi riche. L'origine du premier, nous dit-il, se trouve dans un travail sur le phgnom~ne de Gibbs, ou s'~tait pr6sent~e incidemment une question d'approximation diophantienne; il s'6tait agi de faire voir que tout nombre irrationnel peut gtre approeh~ par une suite de fractions p~/q~ satisfaisant aux conditions qn = o (n), ] ~ - - pJq~ ] = o (l/n). Un peu plus tard l'attention de Weyl fut attir6e par F. Bernstein sur 18 probl6me du m o u v e m e n t moyen en m6eanique e61este, probl6me qui remontait /~ Lagrange, et dont Bohl s'oceupait alors; il s'agit 1/~ de d6terminer le comportement asymptotique, pour t ~ ~ , de l'argument d'une somme finie d'exponentielles Ea, e (X~ t), les X 6tant r6els ~. Ce fut l'occasion pour lui d'observet d'abord que son lemme diophantien entrainait ais6ment l'6gale r6partition modulo i de la suite (n~) pour a irrationnel, r6sultat qui fut obtenu en m6me temps par Bohl et par Sierpinski. Mais Weyl,/~ l'6eole de Hilbert et surtout par ses propres recherehes sur les valeurs et fonctions propres, avait aequis un sens trop juste de l'analyse harmonique pour s'en tenir 1/~, Convenons de d6signer par M (xn) , pour toute suite (x~), la limite pour n ~ ~ , si elle existe, de la moyenne des nombres x~, ..., x~. Dire que la suite (%~) est 6galement r6partie modulo i t
6quivaut /~ dire qu'on a M [ / ( ~ ) ] = f [ (x) dx pour certaines 0
fonetions p6riodiques partieuli6res, /~ savoir pour les fonctions de p6riode I qui coincident dans l'intervalle [0, 1] avec une fonetion caraet6ristique d'intervalle. Weyl s'aperr que, si eette propri6t6 est v6rifi6e pour les fonctions en question, elle l'est n6cessairement aussi pour route fonction p6riodique de p6riode 1, int6grable au sens de Riemann, et en partieulier pour les earact6res e (nx) du groupe additif des r6els modulo 1; r6eiproquement, si elle l'est pour ces derni6res fonctions, elle l'est aussi, en v e r t u des th6or~mes classiques sur la s6rie de Fourier, pour toute fonction p6riodique int6grable au sens de Riemann, de sorte que la suite (%~) est 6galement r6partie modulo 1; la d6monstration de ces assertions est imm6diate. Le r6sultat sur l'6gale r6partition modulo I de la suite (n~) pour e irrationnel Ici, comme
d a n s t o u t c e q u i s u i t , o n p o s e e (t) ~
e2rdt .
336
[1957b1
16r,
C. C H E V A L L E Y
ET
A.
WElL
d~coule de 1~ aussit6t, sans aucun lemme diophantien; en rempla?ant le groupe des r6els modulo i par un tore de dimension queleonque, on obtient de m~me, et sans ealeul, la forme quantitative des c~l~bres th~or~mes d'approximation de Kroneeker. T o u t eela, si neuf ~ l'@oque du travail de Weyl, nous parait present bien simple, presque trivial. Mais aujourd'hui encore le lecteur reste ~tonn~ de voir eomme Weyl, sans reprendre haleine, passe de 1~ ~ l'~gale r@artition d'une suite (P (n)), off P est un polyn6me queleonque. Cela revient naturellement, d'apr~s ce qui precede, ~ l'~valuation des sommes d'exponentielles Ee(P(n)), probl~me qui avait ~t~ d~j~ l'objet des recherehes de H a r d y et Littlewood. Plus pr~eis~ment, il s'agit de d~montrer la relation N
y , e (P (n)) =
o (N)
n~0
lorsque P e s t un polyn6me off le coefficient du terme de plus haut degr6 est irrationnel. Pour donner une idle de la m6thode de W e y l , qui (avee les perfectionnements qu'y ont apport6s Vinogradov et son ~eole) est rest6e fondamentale en th6orie analytique des nombres, eonsid6rons le cas off P est du second degr~. Posons done: N /j n=0
e 4tant irrationnel. On 6erira alors, eomme dans l'~valuation elassique des sommes de Gauss: N
Isxp = sxs N =
~,
e ( e ( r n = - - n ') + ~ ( m - - n ) )
rrt,n=0 +N
=
Z r=-
e( r= + N
n~I r
off on a substitu5 n + r/~ m, et off I~ d~signe l'interseetion des deux intervalles [0, N] et [ - - r, N - - r]. Si on d6signe par ~ la derni~re somme (eelle qui est 5tendue/t l'intervalle I~), on a donc I srl 2 <~ E levi. Comme % est une somme de N ~- I termes
[1957b]
337 HEBMANN
WEYL
(1885-1955)
165
au plus, on a [ % [ < N + I quel que soit r; comme d'autre part % est somme d'une progression g6om6trique de raison e (2~r), on a aussi: I%'1 ~< I sin ( 2 ~ r ) [
-~ 9
Soit 0 <~ z < 1/2; en vertu de l'6gale r6partition des nombres 2~r modulo 1, ]e nombre des entiers r de l'intervalle [ - - N, + N] qui sont tels que 2~r soit eongru modulo t ~ un nombre de Fintervalle [ - - ~, -k z] est de ]a forme 4zN ~- o (N), et est doric ~< 5zN dbs que N e s t assez grand. Pour chaeun de ces entiers, on a 1% ] ~ N + 1; pour tousles autres, on a ] % I ~ l/sin (~z). On a donc, pour N assez grand: 2N+I sin (= e)
[%[~ < 5~N(N + t) + - -
Pour N assez grand, le second membre sera ~< 6zN~; comme il en est ainsi quel que soit ~, on a bien s N = o (N). Si le degr6 du polyn6me P e s t d -? I avee d > l, la d6monstration se fait de m~me (et non par r@urrence sur d) au moyen d'un lemme sur l'6gale r @ a r t i t i o n modulo I d'une fonetion multilin6aire de d variables. Le r6sultat s'6tend aux fonctions de p variables par r6currence sur p. Avec cet admirable m6moire, Weyl 6tait d6j/~ tr~s prSs des fonctions presque p@iodiques. I1 s'y agissait, en effet, en premier lieu, des sous-groupes cycliques et des sous-groupes ~ un param~tre d'un tore de dimension finie, tandis que la th6orie des fonctions presque p@iodiques traite, dirions-nous, des sousgroupes h u n param6tre d'un tore de dimension infinie. On peut m~me dire que cette th6orie, qui suscita rant d'int@~t pendant une dizaine d'ann6es h la suite des publications de H. Bohr en 1924, eut pour principale utilit6 de m6nager la transition entre le point de vue classique et le point de r u e moderne au sujet des groupes compacts et localement compacts. Au temps m~me o/x Weyl s'occupait h GSttingen d'6gale r @ a r t i t i o n modulo 1, vers 1913, les premi@es id6es sur les fonctions presque p@iodiques y 6talent ((dans Fair ~>. Le probl6me du m o u v e m e n t m o y e n portait sur les sommes d'exponentielles, finies il est vrai, et Weyl en
338
[1957b]
166
C. C H E V A L L E Y
E T A. W E I L
avait trait6 des cas particuliers, sans d'ailleurs approfondir la question, qu'il ne devait r6soudre compl~tement, toujours par la m6me m6thode, que lorsqu'il s'y t r o u v a ramen6 vingt-cinq ans plus tard. Mais H. Bohr, alors 61~ve de Landau, s'occupait de s6ries d'exponentielles en vue de l'6tude de ~ (s) dans le plan complexe, probl~me auquel Weyl va bient6t s'int6resser en passant, d6terminant m6me par sa m6thode le c o m p o r t e m e n t asymptotique de ~ (i -k it). D'autre part, les 6]~ves de Hilbert 6talent accoutum6s ~ consid6rer les termes de la s~rie de Fourier comme fonctions propres, et les coefficients de cette s6rie comme valeurs propres, d'op6rateurs convenablement d6finis. I1 semble donc que les voles fussent toutes pr6par6es dans l'esprit de Weyl, lorsque a p p a r u r e n t les premiere t r a v a u x de H. Bohr sur les fonctions presque p6riodiques, pour reprendre la question du point de vue des ~quations int~grales. Mais il est rare qu'un math6maticien, qu'il s'agisse du plus grand ou du plus humble, parcoure le plus court chemin d'un point ~ un autre de sa trajectoire. A v a n t de revenir aux fonctions presque p6riodiqnes ~ l'occasion d'une conf6rence de H. Bohr h Zurich, Weyl avait men6 ~ bien ses m~morables recherches sur les groupes de Lie et leurs representations, et avait congu l'id~e, d'une audace extraordinaire pour l'6poque, de ~ construire ~ les repr6sentations des groupes de Lie compacts par la complete d6composition d'une representation de degr6 infini. Blas6s que nous sommes par l'exp6rience des trente derni6res ann6es, cette id6e ne nous 6tonne plus; mais son succ~s semble avoir fait l'effet d'un vrai miracle ~t son auteur; ~ c'est lh, r6p~te-t-il ~ maintes reprises, l'une des plus surprenantes applications de la m6thode des 6quations int6grales ~>. D~jh I. Schur avait 6tendu au groupe orthogonal, au moyen de l'Ol~ment de volume invariant dans l'espace de groupe, les relations d'orthogonalit~ entre coefficients des repr6sentations que Frobenins avait d6couvertes pour les groupes finis ; mais il y avait loin de lh/~ un th6or6me d'existence. Weyl n'h6site pas h introduire, sur un groupe de Lie compact, l'alg~bre de groupe, toujours congue chez lui comme alg6bre des fonctions continues par rapport au produit de convolution h = f * g,
h(s) = f [ ( s t -1) g(t)dt,
[1957b]
339 HERMANN WEYL (1885-1955)
167
et dont il fait un espaee pr6hilbertien au m o y e n de la norme (f I] [~ ds)~; les int6grales, naturellement, sont prises au m o y e n de l'616ment de volume i n v a r i a n t que fournit la th6orie de Lie, et que Hurwitz avait sans doute ~t6 le premier h utiliser s y s t 6 m a t i q u e m e n t . Dans cet espace, l'op6rateur / - ~ ~ * ~ * /, off ~ d~signe la fonetion ?~ (s) = ~ (s-l), est hermitien et compl~t e m e n t continu; d'apr~s la th6orie de E. Schmidt, ses valeurs propres forment done un spectre discret, et h chacune correspond un espace de fonetions propres de dimension finie, dont on constate i m m 6 d i a t e m e n t qu'il est i n v a r i a n t par le groupe; c'est done un espaee de repr6sentation de celui-ci. Les th6or~mes de Schmidt fournissent a]ors le d6veloppement de 9 suivant les coefficients des repr6sentations ainsi obtenues, d6veloppement qui converge au sens de la norme. C'est lh une g6n~ralisation directe de la m6thode de Frobenius basSe sur la r~duction de la repr6sentation r6guli~re d ' u n groupe fini; la seule diff6renee, c o m m e l'observe Weyl, e'est l'absence d ' u n 61~ment unit6 dans l'alg~bre d ' u n groupe c o m p a c t ; W e y l y suppl6e par un artifice tir6 de la th6orie des s6ries de Fourier, h savoir l ' a p p r o x i m a t i o n de la masse unit6 plac6e h l'origine p a r une distribution de masses h densit6 continue, concentr6e dans un voisinage de l'origine; la convolution avec celle-ei eonstitue un (~op6rateur r6gularisant)), d'emploi courant aujourd'hui, mais dont c'6tait sans doute la premi6re apparition dans le cadre de la th6orie des groupes de Lie ; Weyl s'en sert pour d6montrer que toute fonction continue p e u t 6tre approch6e, non seulement au sens de la norme, mais m~me uniform6ment, par des combinaisons lin6aires de coefficients de repr6sentations. Bien que le m6moire de W e y l se limit~t n6cessairement aux groupes de Lie, il a v a i t atteint en r6alit~, du premier coup, des r6sultats d~finitifs sur les representations des groupes c o m p a c t s ; apr6s la d~couverte de la mesure de H a a r , il n ' y eut pas un m o t ~ changer ~ son expos6, et, chose rare en m a t h 6 m a tique, il ne vint m6me ~ personne l'id6e de le r6crire. Si, comme nous le raisons aujourd'hui, on consid~re une fonetion presque p6riodique c o m m e d~terminant une representation du groupe additif des r6els darts un groupe compact, et q u ' o n suppose acquise pour celui-ci ]a notion de mesure de H a a r , on d~duit
340
[1957b1 168
C. C H E V A L L E Y
ET
A.
WElL
imm6diatement des r6sultats de Weyl expos6s ci-dessus le d6veloppement de la fonetion en s6rie d'exponentielles. Les outils manquaient h Weyl, en i926, pour adopter ce point de r u e ; il y suppl6e en rempla~ant l'int6grale par une moyenne sur la droite, d6finie comme limite pour T-+ + zr de la valeur moyenne sur l'intervalle It, t ~ 3?] lorsque cette limite est atteinte uniform6ment par rapport au param6tre t. En 1926, il n'allait pas de soi que la th6orie des 6quations int6grales s'appliqu~t h cette moyenne; Weyl est oblig6 de consacrer une bonne partie de son travail A justifier cette application. I1 convient d'observer d'autre part que, sur un groupe compact, la mani6re la plus simple de construire la mesure de Haar consiste j u s t e m e n t h attacher ~ chaque fonction continue une valeur moyenne, par un proc6d6 directement inspir6 de la th6orie des fonctions presque p6riodiques. Que Weyl, en revanche, air cru voir dans la th6orie de Bohr (~le premier exemple d'une th6orie des repr6sentations d'un groupe vraiment non compact )> (par opposition a p p a r e m m e n t avec les groupes de Lie semi-simples dent les repr6sentations, dans son esprit, se ramenaient, par la (( restriction unitaire )), h celles de groupes compacts), cela montre qu'il se faisait encore quelque illusion sur le degr6 de difficult6 des probl~mes qui restaient ~ r6soudre. Ce n'en est pas moins lui qui a ouvert la voie h t o u s l e s progrbs ult6rieurs duns cette direction.
Sur le reste de son oeuvre d'analyste, nous serons beaucoup plus brefs, d ' a u t a n t plus qu'il a lui-m~me excellemment rendu compte d'une honne partie de cette oeuvre dans sa Gibbs Lecture de t948. D6butant, il partieipa activement au courant de recherches qui se proposait d'approfondir et d'appliquer A des probl6mes vari6s d'analyse la th6orie spectrale des op@ateurs sym6triques. Citons particu]igrement, dans cet ordre d'id6es, sa Habilitationsschri/t de 19i0, off il 6tudie un op@ateur diff6rentiel autoadjoint L sur la demi-droite [0, + ~ ]: d (p(t)d~) L (u) = ~t - - q (t) ~,
[1957b]
341 HEBMANN
WEYL
(1885-1955)
~169
Od p, q sont ~ valeurs r6elles et p (t) > O. Sur tout intervalle fini [0, 1], cet op6rateur, soumis aux conditions aux limites du t y p e habituel, (du/dt)o = hu (0), (du/dt)z = h' u (1), rel6ve de la th6orie de Sturm-Liouville ou, en termes modernes, de la th6orie des op6rateurs compl~tement continus; le spectre est r~el et discret et se compose des X pour lesquels l'~quation L u = Xu a u n e solution satisfaisant aux conditions aux limites impos6es. Le passage ~ la limite l ~ -1- w fair apparaitre, non seulement un spectre continu qui peut couvrir tout l'axe r6el, mais encore des ph6nom~nes impr~vus dont la d6couverte est due ~ Weyl. Les plus int6ressants concernent le comportement des solutions pour l-~-t- ~ lorsqu'on donne ~t X une valeur imaginaire fixe; chose remarquable, ils sont ind6pendants du choix de la valeur donn6e ~ X. C'est ainsi que Weyl est amen~ en particulier ~ la distinction fondamentale entre le cas du (~point limite ~ et le cas du <~cercle limite ~>: l'une des propri6t6s caract~ristiques du premier, c'est que l'~quation Lu = Xu y poss+de, quel que soit X imaginaire, une solution et une seule de carr~ sommable sur [0, + w ], tandis que toutes ses solutions le sont, pour X imaginaire, dans le eas du cerele limite. Weyl ~tudie aussi le passage ~ la limite 1-~ + c~ pour les d6veloppements de Sturm-Liouville sur [0, l]; il en tire des formules int6grales o~ apparaissent en g~n~ral des int6grales de Stieltjes, comme on pouvait s'y attendre. Le probl~me des moments de Stieltjes n'est d'ailleurs pas autre chose qne le probl+me aux diff6rences finies, analogue ~ l'6quation Lu ---- Xu sur la demidroite, et Hellinger fit voir par la suite que la m6thode de Weyl s'y transporte prescIue telle qnelle. Mais Weyl put aussi la transposer plus tard ~t un probl~me diff+rentiel o~ le param~tre spectral intervient non lin6airement, ainsi qu'au probl6me aux diff6renees finies correspondant (auqnel il a donn6 le nora de probl6me de Piek-Nevanlinna) ; il apporta m~me ~ cette occasion quelques am61iorations notables ~ son premier expos6. Si celui-ei a donn~ lieu depuis lors ~t des g~n6ralisations assez vari6es, il ne semble pas que la signification v~ritable des r~sultats de Weyl sur les probl6mes ~ param~tre non lin6aire ait jamais 6t6 tir6e au clair. Une autre s6rie de t r a v a u x traite de la r6partition des valeurs propres, dans divers probl~mes de t y p e elliptique. Ils
342
[1957b]
170
C. C H E V A L L E Y
E T A. W E l L
reposent principalement sur un principe qui plus tard fut popularis6 par Courant sous la forme suivante: si A est un 0p6rateur sym6trique compl6tement eontinu dans un espaee de Hilbert H, sa n-i6me valeur propre est la plus petite des valeurs ((~minimum m a x i m o r u m ~) que peut prendre la norme de A, c'est-/~-dire le hombre max (Az, z ) / ( z , z), sur un sousespaee de H de codimension n - - 1. Une lois acquise la th6orie des op6rateurs eomplbtement continus, la v6rifieation de ce principe est d'ailleurs imm6diate. Mais Weyl l'adapte en virtuose routes sortes de situations de physique math6matique. Quant au e o m p o r t e m e n t asymptotique des fonetions propres, il avait, nous dit-il en 1948, eertaines conjectures: ~ mais, n ' a y a n t fair pendant plus de trente-einq ans aucune t e n t a t i v e s6rieuse pour les d6montrer, je pr6f6re, ajoute-t-il, les garder pour moi ~; il aura done laiss6 ee probl6me plus diffieile en h6ritage /~ ses suecesseurs.
C'est en 61~ve de Hilbert encore, et en analyste, que Weyl dut aborder le sujet d'un des premiers eours qu'il professa/~ GSttingen comme jeune privatdozent, la th6orie des fonctions selon Riemann. Le eours terrain6 et r6dig6, il se retrouva g6om6tre, et auteur d'un livre qui devait, exereer une profonde influence sur la pens6e math6matique de son si~cle. Peut-~tre s'6tait-il propos6 seulement de remettre au gofit du jour, en faisant usage des id6es de Hilbert sur le prineipe de Dirichlet, les expos6s traditionnels dont l'ouvrage classique de C. Neumann fournissait le modble. Mais il dut lui apparaitre bient6t que, pour substituer aux constants appels ~ l'intuition de ses pr6d6eesseurs des raisonnements corrects et, eomme on disait alors, ~ rigoureux ~ (et dans l'entourage de Hilbert on n'.admettait pas qu'on trich~t l~-dessus), c'6taient avant t o u t les fondements topologiques qu'il fallait renouveler. Weyl n'y semblait gubre pr6par6 par ses t r a v a u x ant6rieurs. I1 pouvait, dans cette tfiche, s'appuyer sur l'ceuvre de Poincar6, mais il en parle h peine. I1 mentionne, comme l ' a y a n t profond6ment influene6, les recherches de Brouwer, alors darts leur premi6re nouveaut6; en r6alit6, il n'en fait aucun usage. De fr6quents contacts avec Koebe, qui d6s lors
[1957b]
343 HERMANN WEYs
(1885-1955)
171
s'6tait consacr6 tout entier h l'uniformisation des fonctions d'une variable eomplexe, durent lui ~tre d'une grande utilit6, particuli~rement dans la mise au point de ses propres id6es. La premibre 6dition du livre est d6di6e ~ F61ix Klein, qui bien entendu, comme Weyl le dit dans sa pr6face, ne pouvait manquer de s'int6resser h u n travail si voisin des pr6occupations de sa jeunesse ni de donner ~ l'auteur des conseils inspir6s de son temp6rament intuitif et de sa profonde connaissance de l'ceuvre de Riemann. Bien qu'il n'efit jamais connu celui-ci, c'6tait Klein qui, ~ GSttingen, incarnait la tradition riemannienne. Enfin, dans Fun de ses m6moires sur les fondements de la g6om6trie, Hilbert avait formul6 un syst6me d'axiomes fond6 sur la notion de voisinage, en soulignant qu'on trouverait lh le meilleur point de d6part pour <(un traitement axiomatique rigoureux de l'analysis situs ~). De tous ees 616ments si divers que lui fournissaient la tradition et le milieu, Weyl tira un livre profond6ment original et qui devait faire 6poque. Le livre est divis6 en deux chapitres, dont le premier contient la pattie qualitative de la th6orie. Les notions de (~surface ~) (vari~t~ topologique de dimension 2/~ base d~nombrable) et de ((surface de Riemann ~> (vari~t~ analytique complexe h base d~nombrable, de dimension complexe l) y sont d~finies au moyen de syst~mes d'axiomes, inspir6s naturellement de celui de Hilbert, mais qui cette lois (saul une l~g~re omission dans la premiere ~dition) ~taient destines ~ subsister sans retouches, et devaient servir de module ~ Hausdorff pour son axiomatisation de la topologie g~n~rale. Dans la premiere et la deuxi~me ~dition, la condition de base d~nombrable apparait sous forme de condition de triangulabilit~ ; et la triangulation joue un grand rSle dans la suite du volume; elle devait ~tre 61imin~e enti~rement de la troisi~me ~dition. Les questions touchant au groupe fondamental, au rev~tement universel, ~ l'orientation, sont 61ucid6es avec soin dans un esprit tout moderne, ainsi que les rapports entre propri6t6s homologiques et p6riodes des int6grales simples sur la surface. Dans la premi6re et la deuxi~me 6dition, l'auteur va jusqu'h la construction, pour les surfaces orientables compactes, d'un syst6me de <(r6trosections ~), c'est-~-dire essentiellement d'une base privil6gi6e pour le premier groupe d'homologie;
344
[1957b]
172
C. C H E V A L L E Y E T A. W E l L
comme il le dit lui-m&me, il aurait pu, au prix d'un 16ger effort suppl~mentaire, aller jusqu'/~ la repr6sentation de la surface au moyen d'un ~<polyg0ne canonique >>/~ 4g c6t6s (g d6signant le genre), et /t la d6termination explicite du groupe fondamental, et on peut regretter qu'il ne l'ait pas fait. Mais la construction mgme des rgtrosections, n~cessairement bas~e sur la triangulation, disparait dans la troisigme ~dition, au profit d'un traitemerit plus purement homol0gique o/1 n'interviennent que des recouvrements. En tout cas, pour tout l'essentiel, ee chapitre constitue une raise au point /~ peu prbs d6finitive des questions qu'il traite. Les th4orgmes d'existence font l'objet du deuxi~me chapitre. Weyl y donne du principe de Dirichlet une d~monstration sireplifi6e, bas6e naturellement sur l'id~e de Hilbert qui consiste, comme on sait, ~ op@er dans l'espace pr~hilbertien des fonetions diff@entiables a v e c l a norme de Diriehlet; mgme dans la troisigme 6dition, il n'a pas cru devoir suivre la variante qu'il avait pourtant contribu~/~ cr6er lui-mgme, et qui consiste R op6rer par projection orthogonale dans le compl6t6 de l'espaee en question, puis ~ montrer apr~s coup que la solution obtenue est diff@entiable. Une fois acquis le principe de Diriehlel,, l'auteur en tire les principales propri6t6s des int6grales ab61iennes et des fonctions multiplicatives, le thbor~me de I/iemann-I/och, puis le th6or~me de l'uniformisation, c'est-Mdire la representation conforme du rev~tement universel de la surface de t/iemann sur une sph@e, un plan ou un disque. Si on laisse de c6t6 les cas de genre 0 ou 1, le rgsultat peut s'exprimer en disant que route surface de t/iemann compacte, de genre > 1, peut se d~finir comme quotient du plan non-euclidien par un groupe discret de d~plaeements sans point fixe. <~Ainsi, dit Weyl dans la pr6face de la premi@e 6dition, ainsi nous pgngtrons dans le temple off la divinit6 est rendue ~ elle-m~me, d~livr~e de ses incarnations terrestres: le cristal non euclidien, oit l'arch6type de la surface de tliemann se laisse voir dans sa puret~ premiere...>> C'est en songeant sans doute /~ ce passage que Weyl dit plus tard de sa preface que (~plus encore que le livre lui-m~me, elle trahissait la jeunesse de son auteur >). Nous dirions aujonrd'hui qu'on a construit pour la surface de ttiemann un module qui est cano-
[1957b]
345 HEBMANN
WEYL
173
(1885-1955)
nique h u n d6placement pr6s dans le plan non euclidien; autrement dit, on a associ6 canoniquement une structure h une autre. Mais qui saurait mauvais gr6 h Weyl, apr6s avoir achev6 un livre de cette valeur, d'avoir exprim6 d'une mani6re peut-Stre un pen trop romantique son enthousiasme juv6nile ?
C'est en 19t6, pendant la guerre, que Weyl fit paraitre en Suisse son premier m6moire de g6om6trie, sur le c616bre probl6me de la rigidit6 des surfaces convexes. Ici encore, GSttingen lui avait fourni son point de d6part. Sous la direction de Hilbert, Weyl avait collabor6 h la publication des oeuvres compl6tes de Minkowski, off la th6orie des corps convexes tient rant de place. D'autre part, Hilbert avait montr6 comment on peut faire d6pendre les in6galit6s de Brunn-Minkowski de la th6orie des op6rateurs diff6rentiels elliptiques. L'espace t/3 6tant consid6r6 comme espace euclidien, et (x, y ) d6signant le produit scalaire dans R a, soit V un corps eonvexe dans cet espace, d6fini au moyen de la fonction d'appui H; eela veut dire que H satisfait aux conditions II(x + x') ~< H(x) + tt(x'),
H(Xx) ~ ?,H(x)
pour ), >~ 0,
et que V e s t l'ensemble des points y satisfaisant h (x, y> ~< H (x) quel qne soit x. Si on suppose H diff6rentiable en dehors de 0, ]e volume de V est alors donn6 par une formule vol (V) = f H.Q (H) d ~ ,
o6 l'int6grale est 6tendue h la sph6re unit6 S o d6finie par (x, z} ~- i, oh do~ d6signe l'616ment d'aire sur So, et oh Q (H) est une forme quadratique par rapport aux d6riv6es partielles seeondes de H. Soient F, F' deux fonctions, diff6rentiables en dehors de 0, satisfaisant toutes deux h la condition d'homog6n6it6 F (;~x) = kF (x) pour X ) 0; soit B (F, F') la forme bilin6aire sym6trique par rapport aux d6riv6es partielles secondes de F et h celles de F' qui se d6duit de la forme quadratique Q (H) par lin6arisation, c'est-~-dire qni est telle que Q (H) = B (H, H);
346
[1957b] 17~
C. C H E V A L L E Y
ETA.
WEIL
posons aussi LF ( F ' ) = B (F, F'). On v6rifie facilement, au moyen de la formule de Stokes, que l'int6grale I (F, F', F") = j" F".B (F, F')d co, So
o5 F " d6signe une troisi~me fonction satisfaisant aux m~mes conditions que F et F', d6pend sym~triquement de F, F' et F " Cela revient ~ dire que LF, consider6 comme op6rateur diff~rentiel sur les fonctions sur S OprolongSes/~ R 3 par homog6n6itS, est un op~rateur autoadjoint. Si V', V" sont deux corps convexes d~finis par des fonctions d'appui H', H", les formules ci-dessus m o n t r e n t que les <~volumes mixtes ~ associ~s par Minkowski V, V', V" ne sont autres que les nombres I (H, H, H') et I (H, H', H " ) ; de plus, nn calcul simple montre que L H est elliptique. Darts ces conditions, comme le fair voir Hilbert dans ses Grundzi~ge, l'application h L . de la thSorie des op6rateurs autoadjoints elliptiques conduit, pour le cas diffSrentiable, /~ l'inSgalit~ de Brunn-Minkowski. Mais il se trouve que L~ n'est autre qu'un op6rateur qui se pr~sente dans la thSorie de la dSformation infinitSsimale de la surface E fronti~re de V; jointe aux rSsultats de Hilbert, cette observation, due ~ Blaschke, entrainait l'impossibilit6 d'une telle d~formation pour E. Enfin, Hilbert, ~ propos des fondements de la g~om6trie, avait d6montr~ l'impossibilit6 d'appliquer isom~triquement une sphere sur une surface convexe non sph6rique. D'ailleurs, des rSsultats analogues sur les poly~dres convexes avaient 5t5 obtenus jadis par Cauch$: non seulement un poly~dre convexe n ' a d m e t aueune dSformation infinit~simale, mais encore, si P et P' sont deux poly~dres convexes a d m e t t a n t m~me schema combinatoire et a y a n t leurs c6t6s correspondants 5gaux, ils ne peuvent diff~rer Fun de l'autre que par un d~plaeement ou une sym6trie. Tout cela mettait ~ ]'ordre du jour l'extension anx surfaces convexes du second th~or~me de Cauchy. Mais Weyl ne s'arr~te pas 14. I1 consid~re en m~me temps un probl~me d'existence que, faute d'une conception claire de la notion de vari6tg riemannienne abstraite, personne n'avait encore m~me formnl6. I1 s'agit de savoir si <~toute surface
[1957b]
347 HERMANN WEYL
(1885-1955)
175
couvexe ferm6e, donn6e in abstracto, est r6alisable )) on, c o m m e nous dirions m a i n t e n a n t , si route vari6t6 riemannienne compacte, s i m p l e m e u t connexe, de dimension 2, h courbure p a r t o u t positive, a d m e t un plongement isom6trique dans l'espace euelidien Ra; la question d'unicit6, pour ce probl~me d'existence, est alors celle m~me dont Weyl 6tait patti. I n t e r r o m p u dans son travail par sa mobilisation en 1915, il se contenta d'esquisser son id6e de d6monstration, et ne la m e n a jamais ~ terme. I1 p a r t du fair que toute <(surface convexe in abstracto ~) p e u t ~tre repr6sent6e c o n f o r m 6 m e n t sur la sph6re So, donc d6finie p a r un ds 2 donn6 sur S o sous la forme ds 2 = e 2r 2, off d~ est la longueur d'arc (( naturelle )) et (I) une fonction diff6rentiable sur So; soit Z ((I)) la (( surface abstraite )) ainsi d6finie. La condition que Z ((I)) soit h courbure p a r t o u t positive s'exprime par une in6galit6 diff6rentielle K (O) > 0; on constate aussitSt que Fensemble des q) qui y satisfont est convexe; il s'ensuit que Fensemble des surfaces convexes abstraites est connexe. L'id6e de Weyl est alors d ' a p p l i q u e r au probl~me une m6thode de coatinuit6. T o u t revient, I d6signant l'intervalle [0, 1], h d6terminer une application ~ de S o x I dans R 3 de telle sorte que l'application x -~ ~: (x) = ~ (x, v) de S o dans R a applique isom6triquem e n t Z (vq)) sur une surface convexe S~ = ~ (So) , et cela pour t o u t z C I. Pour cela, W e y l consid6re ~9/bv c o m m e une d6formarion infinit6simale de S~, dont la d6termination se ram~ne la solution d'une 6quation Au ~ f, off / est une fonction sur So, d6pendant de S~, et A cst essentiellement l'op6rateur elliptique L . relatif h S~. L'application de la m6thode de Hilbert h cette 6quation donne donc, en principe, une 6quation diff6rentielle fonctionnelle pour ~ ; il s'agit d'en t r o u v e r une solution sur l'intervalle I qui se r6duise pour z = 0 h l'application identique de S o dans R3; et on p e u t esp6rer y p a r v e n i r au m o y e n de l'une quelconque des m6thodes classiques de r6solution des 6quations diff6rentielles. Une d6monstration complSte a 6t6 obteuue r 6 c e m m e n t par Nirenberg en suivant cette vole; les br~ves indications donn6es ici sufilront t o u t au moins ~ faire a p p a r a i t r e l'extr6me hardiesse de l'id6e de H e r m a n n Weyl.
348
[1957b] 176
C. C H E V A L L E Y
ET
A.
WElL
Rentr6 ~ Zurich en 1916, Weyl eut, semble-t-il, quelque vell6it6 de revenir aux surfaces convexes; un m6moire oh il reprend les r6sultats de Cauchy sur les poly6dres pr6sageait peut6tre un mode d'attaque has6 sur des m6thodes moins infinit6simales et plus directes. Mais c'est bient6t la relativit6 qui attire et accapare son attention. Lh encore, il 6tait dans la tradition. Minkowski avait particip6 activement au courant de recherches qui s'6tait d6velopp6 a u t o u r de la relativit6 restreinte. Hilbert suivait de pros les t r a v a u x d'Einstein et cherchait, sans grand succ6s d'ailleurs, h 6claircir les probl~mes de la physique par la m6thode axiomatique. (( I1 faut en physique un autre t y p e d'imagination que celle du math6maticien ~), constate plus tard H e r m a n n Weyl, non sans quelque m61ancolie, dans sa notice sur Hilbert. Sans doute, en 6crivant ces roots, songeait-il aussi $ sa propre exp6rience et h cette (( th6orie de Weyl ~) ~ laquelle, disait-il vers la m~me 6poque, il ne croyait plus depuis longtemps. Mais h partir de 1917, et p e n d a n t plusieurs ann6es, son enthousiasme est d6bordant. E n 1918, il publie son cours de l'ann6e pr6c6dente sur la relativit6 sous le titre Baum, Zeit, Materie. (~A l'occasion de ce grand sujet, 6crit-il dans la pr6face de la premi6re 6dition, j'ai voulu donner un exemple de cette interp6n6tration, qui me tient t a n t h cceur, de la pens6e philosophique, de la pens6e math6matique, de la pens6e physique... )); mais, ajoute-t-il avee une modestie non exempte de naivet6, ((le math6maticien en moi a pris le pas sur le philosophe ~); et ce ne sont pas les math6maticiens qui s'en plaindront. Son ouvrage, dans ses cinq 6ditions successives, fit beaucoup pour r6pandre parmi les math6maticiens et les physiciens les connaissances g6om6triques et les notions essentielles de l'alg~bre et de l'analyse tensorielles. A partir de la troisi6me 6dition, on y trouve aussi un expos6 de la (~th6orie de Weyl )), premier essai d'une (~th6orie unitaire ~ englobant dans un m~me sch6ma g6om6trique les ph6nom~nes 61ectromagn6tiques et la gravitation. Elle 6tait fond6e, dirionsnous h pr6sent, sur une connexion li6e au groupe des similitudes (d6fini au m o y e n d'une forme quadratique de signature (1,3)), au lieu qu'Einstein s'6tait born6 h des connexions li6es au groupe de Lorentz (groupe orthogonal pour une forme de signature
[1957b]
349
HERMANN
WEYL
(1885-1955)
177
(1,3)), et plus pr~cis~ment ~ la connexion sans torsion d~duite canoniquement (par transport parall~le) d'un ds 2 de signature (1,3). Cette th~orie eut du moins le m~rite d'~largir le cadre de la g~om~trie riemannienne traditionnelle et de preparer les voies aux (~g~om~tries g~n~ralis~es ~) de Cartan, c'est-it-dire ~ la th~orie g~n~rale des connexions li~es ~ un groupe de Lie arbitraire. Qnant aux preoccupations philosophiques de Weyl pendant cette p~riode d'intense fermentation, elles ne tard~rent pas (heureusement, serions-nous tent~s de dire) ~ se couler dans un moule plus ~troitement math~matique, l'amenant /~ chercher une base axiomatique aussi simple que possible aux structures g~om~triques sous-jacentes h la th~orie d'Einstein et h la sienne; c'est lh c e qu'il appelle le (~ Raumproblem ~), le probl~me de l'espace; il y consacre plusieurs articles, un cours profess~ h Barcelone et h Madrid, et un opuscule qui reproduit ces lemons. I1 s'agit lfi en r~alit5 de caract~riser le groupe orthogonal (attach4 h une forme quadratique, soit complexe, soit r~elle et de signature quelconque) en rant que groupe lin~aire, par quelques conditions simples au moyen desquelles on puisse rendre plausible que la g~om~trie de l'(~univers ~) est d~finie localement par un tel groupe. Bien entendu, c'est la thdorie des groupes de Lie et de leurs representations qui domine la question; Weyl en donne une esquisse dans un appendice de son livre. De son c6t~, Cartan ne tarda pas h donner, du principal r~sultat math~matique de Weyl sur ce sujet, une d~monstration bas~e sur ses propres m~thodes. I] n'~tait pas dans le temperament de Hermann Weyl, une fois parvenu ainsi au seuil de l'ceuvre de Cartan, de se contenter d'y jeter un coup d'ceil rapide. D'autre part,/~ la suite peut-~tre d'une remarque de Study qui l'avait bless~ au vif, il avait commencd ~ s'intSresser aux invariants des groupes classiques. Study, dans une preface de 1923, iui avait reproeh~, ainsi qu'aux autres relativistes, d'avoir, par leur n~gligence /~ l'~gard de ce sujet, contribu~ ~ (~la mise en jach~re d'un riche domaine culturel ~); il entendait surtout par lh la thSorie des invariants du groupe projectif, dans laquelle il ~tait d'usage de faire rentrer t a n t bien que mal les autres groupes h l'occasion de l'~tude des
350
[1957b]
178
C. C H E V A L L E Y
ET
A.
WElL
covariants simultan6s de plusieurs formes. P a r une r6action bien caract6ristique, Weyl r6pondit ~ Study, avec une promptitude extraordinaire, par un m6moire off il reprend h la base la th6orie classique au moyen d'identit6s alg~briques dues ~ Capelli et indique aussi comment elle s'6tend aux groupes orthogonaux et symplectiques; ce qui ne l'emp~che pas de protester que, ((si m~me il avait connu aussi bien que S t u d y lui*m~me la th6orie des invariants, il n'aurait eu nulle occasion d'en faire usage dans son livre sur la relativit6: chaque chose en son lieu !)~. La synth~se entre ces deux courants de pens6e - - groupes de Lie et invariants - - s'op~re darts son grand m6moire de i926, m~moire divis~ en quatre parties, dont il dit lui-m~me vers la fin de sa vie qu'il repr6sente (( eu quelque sorte le sommet de sa production math6matique ~. L'6tude qu'avait faite Young, vers 1900, de la d~composition des tenseurs en tenseurs irr6ductibles d6finis par des conditions de sym6trie avait abouti en substance h la d6termination de toutes les representations (( simples ,, c'est-h-dire irr~ductibles, du groupe lin6aire sp6cial; mais, enferm~es qu'6~aient ces recherches dans le cadre de la th6orie traditionnelle, il leur ~tait impossible, par d6finition, d'obtenir ce r~sultat sous la forme que nous venons de lui donner. De son c6t6, Cartan, parti de la th6orie g6n6rale des groupes de Lie, avait d6termin6 toutes les representations en question, sans d'ailleurs, semble-t-il, faire le lien entre ses r6sultats et ceux d'Young. D6signons par G le groupe lin6aire sp6cial, et par g son alg~bre de Lie, qui se compose de toutes les matrices de trace 0; soit 4 l'ensemble des matrices diagonales contenues dans g. Une repr6sentation simple de G d6termine une representation simple p de q, donc une repr6sentation de 4. Cartan montre que l'espace V de la repr6sentation p e s t engendr6 par des vecteurs qui sont vecteurs propres de routes les operations p ( H ) , pour H C 4. Soit e l'un de ces vecteurs propres; on a p ( H ) . e = k ( H ) e, off ~ est une forme lin6aire sur ~, qu'on appelle le p o i d s de e; si H est la matrice diagonale de coefficients al, ..., an, il est facile de voir que k (H) est de la forme m I a ~ - ~ . . . - ~ m n an, Ofl les m i sont des entiers d6termin6s l'addition pros d'un m~me entier. Si on ordonne lexicographiquement l'ensemble des syst~mes (mj, ..., ran) de n entiers, on
[1957b]
351
HERMANN WEYL (1885-1955)
i79
obtient donc une relation d'ordre dans l'ensemble des poids des repr6sentations de G. On appelle poids/ondamental d'une repr6sentation simple le plus grand des poids de cette repr6sentation pour la relation d'ordre qu'on vient de d~finir. Cartan avait montr~ que ce poids d6termine compl~tement la repr6sentation (h une ~quivalence pros), qu'il correspond/~ un syst~me d'entiers (mi) tel que m~ ~ ... ~ ran, et que r~ciproquement tout syst~me d'entiers satisfaisant ~ ces in~galit6s appartient au poids fondamental d'une repr6sentation simple de G. Soient de plus p, p' deux representations simples de G, operant respectivement sur des espaces vectoriels V, V'; soient )~, k' leurs poids fondamentaux; soient e, e' des vecteurs de V, V', de poids respectifs )~, k'. Le produit tensoriel p | ~' de p et p' (dit parfois encore (~produit kroneck~rien ~), et not6 le plus souvent p • ~' par Weyl) est une reprSsentation op6rant sur un espace V | V' de dimension ~gale au produit de celles de V e t V', qni est form~ de combinaisons lin6aires d'~l~ments se transformant par G comme les praduits formels xx', avec x ~ V, x'E V'; et, pour cette representation, le vecteur e • e' est de poids k ~- ),'. Soit W le sous-espace de V | V' engendrd par e | e' et ses transform6s par G; il dScoule facilement des r~sultats de Cartau que W ne peut pus se d4composer en somme directe de sous-espaces invariants par les operations de G; et Cartan avait cru pouvoir d~duire de lh que W fournit la repr6sentation simple de poids dominant k -[- )?. Weyl observa que cette conclusion est ill~gitime tant qu'on ne sait pas h priori que les representations de G sent toutes semi-simples (c'est-h-dire compl~tement r~ductibles). A vrai dire, ce dernier r~sultat n'~tait pas indispensable pour se convaincre du fait que la d6composition ~l'Young de l'espace des tenseurs fournit toutes les representations simples de G; Young avait en effet ~tabli l'irr~duetibilit~ des repr6sentations qu'il avait construites, et il suffisait d'~tablir par un calcul facile que leurs poids dominants sent tous ceux pr4vns par la th~orie de fiartan. Mais on n'efit obtenu ainsi que la classification des representations simples. Au contraire, en d~montrant la complete r6ductibilit8 de routes les representations de G, Weyl en obtint du mSme coup (compte tenu des r~sultats de Young et Cartan) la classification d~finitive, qui s'exprime par le fait que
352
[1957b]
t80
C. CHEVALLEY ET A. W E l L
route <~grandeur ]in6aire }~, comme il dit, se d6compose en tenseurs irr6duetibles. On sait aujourd'hui d6montrer le th6or6me de compl6te r6ductibilit6 par des m6thodes alg6briques; c'est 1~ le point de d6part de la th6orie cohomologique des alg6bres de Lie. Mais c'est de consid6rations tout autres que Weyl tire sa d6monstration. I1 observe, comme l'avait d6j~ fait Hurwitz dans son m6moire sur la construction d'invariants par la m6thode d'int6gration, que la th6orie des repr6sentations du groupe lin6aire sp6cial complexe G est 6quivalente ~t celle des repr6sentations du groupe G u form6 des matrices unitaires a p p a r t e n a n t ~ G; en derni6re analyse, cela tient ~ c e que route identit6 alg6brique entre coefficients d'une matrice unitaire reste vraie pour une matrice quelconque. Or Gu poss6de une propri6t6 importante qui n ' a p p a r t i e n t pas a G: il est compact, ce qui permet, comme l'avait fait voir Hurwitz, de construire des invariants pour Gu et par suite pour G par int6gration dans l'espace du groupe Gu au moyen de l'616ment de volume invariant fourni par la th6orie de Lie. La m6thode classique qui permet d'6tablir la compl6te r6ductibilit6 des repr6sentations des groupes finis par construction d'une forme hermitienne, d6finie positive, invariante par les op6rations du groupe, s'6tend alors d'elle-m~me au groupe G~. Ce n'est pas seulement le th6or6me de compl6te r6duetibilit6 pour G que Weyl tire de la restriction au groupe unitaire Gu; il s'en sert aussi pour calculer explicitement les caract6res et les degr6s des repr6sentations simples de G. On voit tout de suite, en effet, que si Z e s t le caract6re d'une repr6sentation de Gu, et si s e s t une matrice diagonale unitaire de d6terminant I et de coefficients diagonaux e (xl) , ..., e (x~), la valeur de X (s) s'exprime comme somme de Fourier finie en xl, ..., x~ et ne change pas par une permutation quelconque des x i. Weyl montre que ces propri6t6s, jointes aux relations d'orthogonalit6 fournies, elles aussi, par la m6thode d'int6gration, suffisent d6jg ~ d6terminer compl6tement les caractbres et ~ en obtenir des expressions explicites. La suite du m6moire de Weyl est consacr6e ~ l'extension des m6thodes ci-dessus aux groupes orthogonaux et symplectiques, puis aux groupes semi-simples les plus g6n6raux. Soit cette
[1957b]
353 HERMANN
WEYL
(1885-1955)
i81
fois .q une alg~bre de Lie semi-simple eomplexe; pour en ~tudier les repr6sentations, Weyl va appliquer la mSthode de restriction unitaire au groupe adjoint G de g, mis sous forme matrieielle relativement ~ une base eonvenable de g. Pour qu'il y air dans G (~assez ~>d'op~rations unitaires, il est nSeessaire que .~ admette ce qu'on appelle aujourd'hui une forme eompaete, ou pour mieux dire une base telle que ]es eombinaisons lin~aires r6elles des ~lSments de cette base forment l'alg~bre de Lie d'un groupe compact. En examinant ehaque groupe simple s@ar~ment, Cartan avait vSrifi~ dans chaque eas l'existence d'une forme compaete; Weyl en donne une d~monstration a priori basSe sur les propri~t~s des eonstantes de structure de g. Cela fait, il introduit ]e groupe G~ des operations unitaires de G, et son alg~bre de Lie g~. Le groupe G~ est compact, et la th~orie des representations de % est 5quivalente ~ celle des reprSsentations de g. Mais iei se pr~sente une difficult~ nouvelle; du fait que G~ peut n'~tre pas simplement connexe, la th~orie des reprSsentations de gu n'est plus entiSrement ~quivalente h celle des reprSsentations de G~. Si on eherehe h r~tablir l'~quivalence en remplagant G u par son rev~tement universel G~, qui, lui, est simplement connexe, il devient n~eessaire de s'assurer que eelui-ei est compact, et aussi d'en faire un groupe, loealement isomorphe G~,. Ce dernier point, qui devait peu apr~s 5tre ~lueid5 par Sehreier, est compl~tement laiss5 de c6t5 dans le mSmoire de Weyl. Mais c'est dans le premier que r~sidait la vSritable diffieult~. La question revient naturellement h faire voir que G~ a un groupe fondamental fini. Pour cela, Weyl introduit un sousgroupe A~ de G~, qui joue le m~me r61e que le groupe des matrices diagonales dans la th~orie du groupe unitaire special. T o u t ~l~m e n t s de G~ est eonjugu5 ~ un ~l~ment de A~; excluant certains 5lSments s, dits singuliers, qui forment un ensemble a y a n t trois dimensions de moins que Gu, s n'est eonjugu5 qu'/~ un hombre fini d'515ments de A u ; de plus, les 516ments de A u qui ne sont pas singuliers forment dans A~ un domaine simplement connexe A. Supposons que s d~erive dans G~ une eourbe fermSe F qui ne reneontre pas l'ensemble des ~lSments singuliers. Si s (t) est le point de param~tre t sur F, on peut dSterminer par eontinuit5 une courbe a (t) dans A~ telle que, pour t o u t t, a (t) soit eonjugu5 ~ s (t).
354
[1957b]
182
C. C H E V A L L E Y
ET
A. W E I L
Quand le point s (t) revient /~ sa position initiale s (1) = s (0), le point a (t) vient en un point a (1) qui est un 616ment de A~ conjugu6 de a (0), ce qui ne laisse pour ee point qu'un nombre fini de possibilit6s. Si on a a (1) = a (0), la courbe d6crite par a (t) est ferm6e, et par suite r6duetible a un point dans A; Weyl montre que P e s t alors elle-m~me r6ductible /~ un point. I1 en r6sulte facilement que le groupe fondamental de l'ensemble des 616ments non singuliers de (3~ est fini. De cela, et du fait que les 616ments singuliers se r6partissent sur des sous-vari6t6s a y a n t an moins trois dimensions de moins que On, Weyl conelut (/~ vrai dire sans d6monstration) que G~ lui-m~me a un groupe fondamental tlni. Ce point ~tabli, la voie est ouverte /~ la g~n~ralisation complete au cas semi-simple des r~sultats obtenus pour le groupe lin6aire sp6eial. Weyl d6montre la complete rdduetibilit6 des representations de g, et d~termine explicitement le caract~re et le degr~ d'nne repr6sentation simple de poids dominant donn~. Ici encore, cette d6termination r6sulte des relations d'orthogonalit~ entre caractbres et des propri6t~s formelles de la restriction Z d'un caract~re au groupe A~ qui reeouvre A~ dans le rev~tement simplement connexe G~ de G~. Ce groupe est un tore; Z e s t une combinaison lin~aire finie de caraetbres de ce tore, invariante par les op6rations d'uu certain groupe fini S d'automorphismes du tore qui g6n6ralise le groupe des permutations de z~, ..., z~ dont il a 6t6 question plus haut/~ propos du groupe unitaire special. Le groupe S, dont les dgveloppements ult6rieurs de la th~orie out montr6 qu'il y joue un r61e fondamental, s'appelle m a i n t e n a n t le groupe de Weyl. Enfin la th6orie s'ach~ve par la d6monstration de l'existence des repr6sentations simples de poids fondamental donn~. Pour les alg~bres simples, cette existence avait 6t~ 6tablie par Cartan par des constructions directes dans ehaque cas partieulier. Weyl, lui, applique au groupe compact G~ la m6thode de d6composition de la <~repr6sentation r6guli~re ~>, obtenue au moyen de la th~orie des 6quations int~grales suivant l'id~e que nous avons expos6e plus haut. Pour eonclure /~ partir de 1/~, il lui faut encore un lemme de nature plus technique, ~none6 seulement dans le m~moire de 1926, et dont Weyl n'a publi6 la d6monstration que
[1957b]
355 HERMA~N
WE YL (1885-1955)
'183
dans son cours de 1934-35 (paru h Princeton sous forme de notes mim6ographi6es, The structure and representations o/ continuous groups).
Beaucoup plus tard, Weyl revint sur la d6termination des repr6sentations des groupes semi-simples dans son ouvrage The classical groups, their in~ariants and representations. L ' e s p r i t de ee livre est assez diff6rent de celui du m6moire de 1926. L ' o b j e t de l ' a u t e u r est m a i n t e n a n t d'une p a r t de d6montrer par des m6thodes p u r e m e n t alg6briques les r6sultats d6j/~ obtenus au sujet des repr6sentations des groupes classiques (groupe lin6aire g6n6ral, groupe lin6aire sp6cial, groupe orthogonal et groupe symplectiqne), et d ' a n t r e p a r t de faire la synthSse entre ces r6sultats et ]a th6orie formelle des invariants qui s'6tait d6velopp6e sons l'influenee de Cayley et Sylvester au cours du x~x e si6cle. Esp6rait-il cette lois se laver d6finitivement du reproehe de S t u d y en r a m e n a n t h la vie cette th6orie qui 6tait sur le point de sombrer dans l'oubli ? I1 nons dit lui-m~me que la d6monstration par Hilbert du th6or6me g6n@al de finitude a v a i t t~presque tu6 le sujet ~>; on peut se d e m a n d e r si Weyl ne lui a u r a pas, en r6alit6, port6 le coup de grace. La situation dans laquelle on se trouve en th6orie des invariants est la suivante. On a uric ou plusieurs repr6sentations lin6aires p, p', ..., d'un groupe G, op6rant sur des espaces r e c t o rids V, V', ... On consid@e des fonetions F (a, x', ...) d6pendant d'un a r g u m e n t a dans V, d'un a r g u m e n t a' dans V', etc. et s'exprim a n t c o m m e polynSmes par r a p p o r t aux eoordonn6es de ces arguments, homog6nes par r a p p o r t aux coordonn6es de chacun d'eux. Une telle fonction s'appelle un i n v a r i a n t si, pour t o u t s dans G, on a F (s.x, s.x', ...) ~- F (x, x', ...) .
Si J~, ..., Jh sour des invariants, t o u t polyn6me en J1, ..., Jh en est un aussi p o u r v u qu'il satisfasse aux conditions d'homog6n6it6 impos6es. Le premier probl6me de ]a th6orie est de t r o u v e r des invariants J~, ..., Jh tels que t o u t autre invariant puisse s'6crire comme polyn6me en les J~; eela fait, on se propose 6galement
356
[1957b]
18~
C. C H E V A L L E Y E T A. W E I L
de d~terminer les relations alg~briques F (J1, ..., Jh) = 0, dites (~syzygies ~, qui lient entre eux les invariants q u ' o n a construits. Pla~ons-nous plus partieuli~rement dans le cas off G a ~t~ identifi0, au m o y e n d'une certaine representation ,ol, avee un sous-groupe du groupe lin~aire/~ n. variables, operant sur l'espace vectoriel V 1 ~- k n, off k est un corps de base q u ' o n suppose de caract~ristique 0. Consid0rons d ' a b o r d le cas off les repr~sentations ~, f , ..., coincident toutes avec 91; on dit alors q u ' o n eherche les invariants d'un certain n o m b r e de ~ vecteurs ~ (on entend par 1~ des vecteurs de V1). R e p r e n a n t sans grand changem e n t son travail de 1924 par lequel il a v a i t r~pondu ~ Study, Weyl montre alors, pour un groupe G unimodulaire, que la d~termination des invariants de vecteurs en n o m b r e quelconque peut se ramener, au m o y e n des identit~s de Capelli, au probl~me analogue pour n I vecteurs. Si G n ' e s t pas unimodulaire, ce r~sultat reste vrai pour les (~invariants relatifs ~ (polyn6mes se multipliant par une puissance du d~terminant de s quand on t r a n s f o r m e tons les vecteurs par s). Weyl d0duit de 1~ la solution des deux probl~mes ci-dessus pour le groupe unimodulaire et pour le groupe orthogonal; et il ~tend cette solution au cas des invariants d~pendant, non seulement d'un certain h o m b r e de vecteurs ~ cogr~dients ~ (se t r a n s f o r m a n t suivant p1), mais aussi d ' u n certain n o m b r e de veeteurs (~contragr~dients ~ (se transf o r m a n t c o m m e les formes lin~aires sur V1). Ensuite il passe aux invariants d~pendant de ~ quantit~s ~> x, x', ... a p p a r t e n a n t des espaces de representation quelconques du groupe ~tudi~; le cas off x, x', ... sont des formes homog~nes p a r r a p p o r t aux coordonn~es d ' u n vecteur ~contragr~dient ~ est celui dont t r a i t a i t plus particuli~rement la th~orie classique. Pour pouvoir aborder ]a question dans ce cadre g~n~ral, il faut a v a n t t o u t connaitre les representations simples du groupe; aussi une partie i m p o r t a n t e du livre est-elle consacr0e ~ la d~termination algabrique des representations (~tensorielles ~ des groupes classiques. Cela fair, Weyl m o n t r e que les i n v a r i a n t s d~pendant de plusieurs ~ quantit~s ~ d'esp~ee quelconque s ' e x p r i m e n t c o m m e polynSmes en un n o m b r e fini d'entre eux; il ~tend ce rSsultat, dans une certaine mesure, au groupe affine. Enfin, il emploie la m~thode d'int~gration pour d~montrer le r~sultat correspondant pour les
[1957b]
357 HERMANN
W E YL (1885-1955)
185
repr6sentations quelconques d'un groupe compact, le corps de base 6rant cette lois le corps des r~els.
Pour f~ter son soixante-dixi~me anniversaire, les amis et 61~ves de Hermann Weyl publi@ent un volume de Selecta extraits de son oeuvre. I1 n'y a peut-6tre pas lieu de se f61iciter de cette mode des morceaux choisis destin6s '~ cS16brer la raise la retraite de math6maticiens 6minents. C'est trop pour les uns; ce n'est pas assez pour les autres. Du moins le volume en question contient-il une bibliographie complete de l'oeuvre de Hermann Weyl, ~tablie par ordre chronologique s, et dont nous avons naturellernent fair grand usage pour r6diger la pr6sente notice. Pour rem6dier en quelque mesure aux inSvitables lacunes de celle-ci, nous donnons ci-dessous une liste des mSmoires de Weyl, classSs par sujet; rien ne peut mieux, croyons-nous, en faire ressortir l'Stonnante vari6t6. Les num@os, bien entendu, renvoient h la liste des Selecta.
I. Analyse. a) Equations int6grales singuli@es: i, 3. b) Probl~mes de valeurs propres et dSveloppements fonctionnels associSs h des 6quations diff@entielles ou aux diff@ences finies: 6, 7, 8, 12, 103.
c) R6partition des va]eurs propres d'op@ateurs complStement continus en physique math6matique: 13, 16, 17, 18, 19, 22.
d) Espace de Hilbert: 4, 5. e)
Ph6nombne de Gibbs et analogues: 10, 11, 14.
/)
Equations diff@entielles li6es h des probl~mes physiques: 36-37 (d~veloppements asymptotiques, apparent6s au ph6nom~ne de Gibbs, au voisinage d'une discontinuit6 dans un
n convient de signaler qu'on n ' a pas fait figurer clans cette bibliographie les notes de cours, publi6es sous formc mim~ographi6e par l ' I n s t i t u t e for A d v a n c e d Study de Princeton, ct qui reproduisent plusieurs des cours qu'il y professa.
358
[/957b] 186
C. C H E V A L L E Y
ET
A.
WElL
probl~me d'61eetromagnOtisme), 123-t24-125 (6tude directe d'une Oquation diff6rentielle li6e /~ un probl6me de eouehe limite).
g) Problbmes elliptiques: 121 (principe de Dirichlet trait6 par la <
Ggomglrie.
a) Surfaces et poly6dres convexes: 25, 27, t06. b) Analysis situs: 24, 26, 57-58-59, 159. c) Connexions, g6om6trie diff6rentielle li6e ~ la relativit6: 30, 3i, 34, 43, 50, 82. d) Volume des tubes: 1t6 (eontient d~jfl, essentiellement, la formule de Gauss-Bonnet pour les vari~t~s plongSes darts un espace euelidien). III.
In~ariants et groupes de Lie.
a) ~ Raumproblem ~: 45, 49, 53, 54. b) Invariants: 60 (1 re partie), 63, 97, i17, 122. c) Groupes de Lie et leurs representations: 6t, 62, 68, 69, 70, 74, 79, 80, 81. IV. [telati~itg. 29, 33, 35, 39, 40, 46, 47, 48, 51, 52, 55, 56, 64, 65, 66, 89, 93, 134, i35.
[1957b]
359 HERMANN WEYL (1885-1955) V.
t87
Thgorie des quanta.
75, 83, 84, 85, 86, 87, 90, 9i, 100, 101, 140, 141. VI. Th~orie des algdbres. a) Matrices de Riemann: 99, t07, 108. b) Questions diverses: 96, 105 (spineurs, en commun avec
R. Brauer), 109, 110, 143. VII.
Th6orie g6om6trique des nombres
(d'apr~s Minkowski et Siegel). 120, 126, 127, 136. VIII. Logique. 9, 32, 41, 60 (2e pattie), 67, 77, 78. IX. Philosophie. t11, 118, 119, 138, 142, 156, 163. X.
Articles historiques et biographiques.
15, 88, 94, 95, 102, 131, 132, 137, 147, 149, 150, 152, 157, 160, 161, 162; et la con%renee <~Erkenntnis und Besinnung~>, Studia Philos., 15 (Basel, 1955) (traduction frangaise dans Rer de Thgol. et Philos., Lausanne, 1955). XI.
Varia.
2, 28, 38, 76, 92, 115, 128, 133, 139, 144, 146, 148, 151, 158.
[1957cl ] R6duction des formes quadratiques, d'apr6s Minkowski et Siegel La th6orie s'ins6re dans le sch6ma suivant, dont on rencontrera d'autres exemples par la suite. Soit G u n groupe de Lie semi-simple; soit F u n sous-groupe discret de G; classiquement, on se propose de trouver, pour F dans G, un " d o m a i n e fondamental", c'est-fi-dire un syst6me de reprdsentants dans G pour G/F possddant des propridt6s sympathiques. Si K est un sous-groupe compact de G, G operera sur l'espace homog6ne H = G / K (notation: on fera op6rer G ~ droite sur H; H est donc l'ensemble des classes/t gauche, K x , suivant K dans G); il est immddiat que, si F est discret darts G, F, en tant que sous-groupe de G, op6re sur H d'une mani6re " p r o p r e m e n t discontinue" (ce qui veut dire: quelle que soit la partie compacte X de H, les 7 ~ F tels que X c~ X7 r 0 sont en nombre fini). En fait, on prend toujours pour K un sous-groupe compact maximal de G; il r6sulte de la th6orie des groupes semi-simples que K est d6fini par l/t d'une mani6re unique, fi un automorphisme intdrieur pr6s de G; autrement dit, H est ddfini d'une mani6re unique fi un isomorphisme pr6s. Sans restreindre essentiellement la gdndralit6 des probl6mes qu'on se propose de traiter, on peut supposer que G n'admet pas de sous-groupe compact invariant; H = G / K est alors, au sens de Cartan, 1' "espace riemannien sym6trique" associ6/t G. Du point de vue thdorique, il est c o m m o d e de prendre G semi-simple au sens strict (non seulement au sens infinitdsimal, mais au sens global), c'est-/L-dire de centre r6duit/t l'616ment neutre (et non pas seulement de centre discret). Du point de vue des calculs explicites, il est souvent c o m m o d e de calculer avec des matrices, ce qui conduit /t 6crire des groupes, infinit6simalement simples ou semi-simples, ayant un centre discret (par exemple le groupe lindaire spdcial sur R, SL(R, n), dont le centre, sin est pair, est _+ 1,, off 1, est la matrice unit6). D'ofi maints abus de langage, dont on s'excuse d'avance. Dans la th6orie de la r6duction des formes quadratiques au sens classique, on part du groupe Go = PL+(R, n) (partie connexe du groupe projectif r6el fi n variables " h o m o g 6 n e s " - quotient du groupe lin6aire L+(R, n)/t n variables, de ddterminant > 0 , par son centre = quotient du groupe lin6aire sp6cial SL(R, n), it n variables, de d6terminant 1, par son centre), et du groupe discret F0, image dans G o du groupe multiplicatif F des matrices de d6terminant _+ 1 /l coefficients dans Z. Un sous-groupe compact maximal de Go est l'image K o dans Go du groupe orthogonal spdcial SO(R, n) (matrices orthogonales de ddterminant 1). Soit P l'espace des formes quadratiques positives non d6g6ndr6es fin variables: n
F(x) = ' x . A . x =
~
aijxixj,
'A = A.
i,j--I
(un vecteur x, en notation matricielle, sera toujours conqu c o m m e matrice ",i n lignes et 1 colonne). On 6crira A >> 0 pour exprimer que la matrice sym6trique A est 360
[1957c]
361
Reduction des formes quadratiques, d'apres Minkowski et Siegel celle d'une forme positive non degendrde. Le groupe L(R, n) opere sur P par la loi (A, X) -~ tX 9A - X, off A ~ P e t X est un element de L(R, n) ecrit comme matrice. Toute forme de P s'ecrivant comme somme de n carrds (de formes lineaires en les x;), le groupe opere transitivement dans P: l'element 1, de P (c'est-fi-dire la forme "standard" ~ x~) etant invariant par le groupe orthogonal O(R, n), P s'identifie l'espace homogene L(R, n)/O(R, n). Dans l'espace vectoriel (de dimension n(n + 1)/2) de routes les formes quadratiques dans R" (ou, ces formes 6rant exprimdes au moyen de la base canonique de R", dans l'espace des matrices symdtriques fi n lignes et n colonnes), Pest ddfini par les inegalites P(x) > 0 (x r 0) et forme donc un c6ne convexe; on verifie immediatement que ce c6ne est ouvert; sa frontiere est l'ensemble des formes positives ddgenerees. Par passage au quotient par la relation d'equivalence dont les classes d'equivalence sont les rayons ("demi-droites") issus de 0, P determine une partie convexe Po d'un espace projectif de dimension n(n + 1)/2 - 1. Par passage au quotient, le groupe projectif Go = P L ( R , n) opere dans Po; il resulte de ce qui precede qu'il y opere transitivement, et que P0 s'identifie fi Go/Ko, espace "riemannien symetrique" associe/t Go. La theorie de Minkowski conduit/t la determination, dans Po, d'un domaine fondamental pour le groupe discret F o, qui est un polyedre convexe (plus exactement, la reunion de l'intdrieur d'un polyedre convexe et d'une pattie convenable de sa frontiere). Soit d'abord n = 2 ("formes binaires"); on ecrit F ( x ) = ax 2 + 2 b x y + cy2;
P e s t le c6ne determine par a > 0, ac - b 2 > 0 (intdrieur de l'une des nappcs d'un cene du second degr6 dans R3); Po est, dans le plan projectif, l'intdrieur de la conique ac - b 2 = 0. C o m m e X --. t X 9A 9X est une representation du groupe lineaire dans l'espace des formes symdtriques A, il s'ensuit, par passage au quotient, que les operations de Go dans Po sont des homographies ou automorphismes du plan projectif ambiant, conservant la conique frontiere de Po; Po 6rant pris comme "modele cayleyien" de la geometrie plane hyperbolique, G induit donc, sur P0, un sous-groupe du groupe des automorphismes de cette geometrie; on verifie sans peine que c'est marne exactement la pattie connexe de ce dernier groupe (c'est-fidire le "groupe des ddplacements non-euclidiens"). La correspondance entre le "modele cayleyien" et le demi-plan de Poincar6 s'obtient comme suit :/t toute forme F e P on fait correspondre, d'une part le point f qu'elle determine dans Po, d'autre part celle des racines de l'equation a z 2 + 2bz + c = 0 dont la partie imaginaire est > 0; la correspondance f ~ z e s t une correspondance biunivoque entre P0 et le demi-plan supdrieur de la variable z; les operations de Go dans ce demi-plan sont 6videmment celles du groupe homographique reel de determinant > 0 (groupe des ddplacements non-euclidiens dans le modele de Poincare). Pour determiner un point de P, on peut, au lieu de se donner une forme quadratique dans R", se donner une forme quadratique F, positive non degenerde, dans un espace vectoriel E de dimension n sur R, et une base (el . . . . . e,) de cet espace. Si (x, y ) est le produit scalaire dans E, associe fi F de la maniere habituelle, les donnees
362
[1957c] Reduction des formes quadratiques, d'apres Minkowski et Siegel en question determinent la forme quadratique dans R" donnee par
Posons A = Hai~ll, air = (ei, ej~; soit X = ]lxijl] un 61ement du groupe lindaire L(R, n); par definition, la transformee de A par X est A' = tX. A - X , ce qui s'ecrit aussi n J A' = Ila'ijll avec aijt = (ei,/ ej), ei! =
xkiek.
2 k
1
Dire que X est ~ coefficients entiers de determinant _+ 1 6quivaut/t dire que les "lattices" (sous-groupes discrets de rang n de l'espace E) engendres par (el . . . . . e,) et par (e'l . . . . . e',) coincident. L'ensemble form6 par A et tous ses transformes par le groupe F est donc l'ensemble des matrices A' = (e~, e:i) lorsqu'on fair parcourir fi (e'1. . . . . e',) l'ensemble de t o u s l e s systemes de n gdndrateurs du lattice engendr6 par el . . . . . e,. Supposons qu'on ait choisi, dans P0, un systeme de reprdsentants Mo pour la relation d'dquivalence ddterminee par les operations du groupe F 0 sur Po; convenons momentandment de dire que la matrice A d'une forme quadratique dans R" est "rdduite" si le point qu'elle determine dans Po appartient /t M0, et aussi qu'une base (e~ . . . . , e,) de l'espace E est "reduite" pour une forme quadratique F dans E si la matrice A des (ei, e j) est reduite. I1 rdsulte de ce qui precede qu'dtant donnds une forme F et un lattice A darts E, il y a au moins un systeme de generateurs de A qui est une base reduite pour E, et que deux tels systemes determinent ndcessairement la m0me matrice A, donc ne different l'un de l'autre que par une transformation du groupe orthogonal de F. Choisir un systeme de representants M 0 revient donc "fi peu pres" fi enoncer une loi qui, /t tout couple form6 d'une forme quadratique F et d'un lattice A, permette d'associer, avec le moins d'ambigu'ite possible, un systeme de generateurs de A. On va, d'apres Minkowski, formuler une telle loi. Soit d'abord n = 2; on prendra pour el un vecteur va0 du lattice A dont la " l o n g u e u r " F(el) ~/2 soit la plus petite possible, puis pour ez un vecteur dont la longueur soit la plus petite possible parmi ceux qui ne sont pas de la forme le~ (t ~ R). I1 est clair qu'il n'y a qu'un hombre fini de vecteurs el satisfaisant/t la 1c condition, et qu'il y e n a au moins deux; des considerations geometriques 61ementaires (et evidentes) font voir qu'il y e n a exactement deux, saufdans les cas suivants : (a) lattice de Gauss (points fi coordonnees entieres dans le plan muni de la forme x2 ~_ y2); (b) lattice hexagonal (engendre par les vecteurs (1, 0) et (89 , / 3 / 2 ) d a r t s le plan muni de x 2 + y2). I1 s'ensuit que, dans tousles cas, les divers choix possibles de el se deduisent les uns des autres par une rotation laissant A invariant (rotation d'angle ~ dans le cas general, d'angle m~/2, resp. m7c/3, avec m entier, dans les cas (a), resp. (b)). Quant au second vecteur e2, on constate non moins aisement qu'il est determine d'une maniere unique au signe pres (une lois q u ' o n a choisi el) saul dans le cas d'un lattice engendre par deux vecteurs (1, 0) ct (2l, y), avec I):1 >x/3/2, dans le plan muni de x 2 + y2. O n levera l'indetermination de signe, autant
[1957c]
363
Rdduction des formes quadratiques, d'apr6s Minkowski et Siegel que possible, en convenant de prendre e 2 tel que ( e l , e2) > 0; il se trouve que, lorsque cette r6gle laisse subsister une ambiguit6, le lattice A admet la droite 0el pour axe de symdtrie, et les choix possibles de e 2 sont symdtriques l'un de l'autre par rapport 5- cet axe. Une base (el, e2) du plan muni d'une forme quadratique F sera dite rOduite si elle poss6de les propridtds 6nonc6es ci-dessus, par rapport au lattice qu'elle engendre et 5- la forme F. Une forme quadratique F = a x 2 "4- 2 b x y + cy 2 dans R 2 sera dite rOduite si la base canonique de R 2, form6e des vecteurs (1, 0) et (0, 1), est r6duite par rapport 5. cette forme. Si on 6crit les indgalitds F((1, 0)) _< F((0, 1)) < F((_+ 1, 1)) on obtient imm6diatement les conditions a < c, 12b[ _< a, qui sont donc ndcessaires pour que F soit r6duite (avec bien entendu a > 0 puisque F doit 6tre positive non d6g6n6r6e). La condition subsidiaire imposde 5. ez s'6crit ici b _> 0. Doric: 0
0<2b
Rdciproquement, ces indgalitds entrainent, d ' a b o r d que F est positive non d6gdn6r6e (c'est clair), puis que F est rdduite (c'est facile 5. v6rifier). Elles d6terminent, dans P0, c'est-5.-dire dans l'int6rieur de la conique ac - b 2 = 0, un triangle Mo, ayant un sommet sur la conique. Puisque, 6tant donn6s un lattice et une forme, la r6gle ci-dessus permet toujours de construire au moins une base rdduite, il s'ensuit que M o contient un syst6me complet de reprdsentants pour le groupe F o opdrant dans P0. Du fait que, pour F et A donnds, les divers choix possibles de la base r6duite se d6duisent tous les uns des autres par une rotation ou une sym6trie conservant A, on conclut que M o constitue marne un tel syst6me de repr6sentants. I1 s'ensuit que P0 est rdunion de M o et de tous ses transformds par Fo (" pavage" du plan non euclidien par les transformds du domaine fondamental). Naturellement, ces transform6s ne sont pas deux fi deux sans point c o m m u n : cela tient justement aux cas d'ambiguit6 possible dans le choix d'une base rdduite, ou, ce qui revient au marne, au fait que F o n'op6re pas sur Po "sans point fixe". L'analyse ddtaill6e du pavage en question est classique, n'offre pas de difficult& e t e s t sans intdr6t pour ce qui suit. O n notera que si, au lieu de partir du groupe F des matrices enti6res de d6terminant +_ 1, on 6tait parti du sous-groupe F + de F form6 des matrices de d6terminant 1, on aurait 6t6 conduit naturellement fi un domaine fondamental deux lois plus grand, correspondant 5. la notion d'"dquivalence propre" des formes quadratiques (au lieu de l'6quivalence " p r o p r e ou impropre" qui a servi ci-dessus); pour l'6quivalence propre, une forme sera dite rdduite si elle satisfait 5. 0 < a < c, 12bl < a; ce sont les conditions classiques de Gauss, correspondant, dans le demiplan de Poincar6, au domaine fondamental classique pour le groupe modulaire (d+termin+ par ]zl > 1, IR(z)[ _< 89 Passons au cas g6ndral. Dans un espace vectoriel E de dimension n sur R, une base (el . . . . . e,), engendrant un lattice A, sera dite r~duite par rapport 5. une forme quadratique F (positive non d6g6ndr6e) si elle satisfait aux deux conditions suivantes : (i) P o u r chaque i, soit E i l'ensemble des vecteurs e 6 A tels que (el . . . . . ei 1, e)
364
[1957c]
Rdduction des formes quadratiques, d'aprds Minkowski et Siegel fasse pattie d'un systeme de n gdndrateurs pour A; e~ est alors un vecteur de longueur minimum parmi ceux de E~. (ii) O n a ( e i , e i + l ) _ > 0 p o u r 1 _ < i _ < n - 1. C o m m e ci-dessus, la condition (ii) sert "a se ddbarrasser, dans la mesure du possible, de rambiguitd de signe inhdrente au choix de e~ quand F et A sont donnds. Si on s'dtait laissd guider par le casn = 2, on aurait, dans (i), pris pour E~ l'ensemble des vecteurs de A qui ne sont pas dans l'espace vectoriel engendrd par e , , . . . , e~ 1 ; avec cette condition plus forte, il n'aurait pas dtd vrai que tout lattice possdde une base rdduite par rapport ~ une forme F donnde, d'ofl ndcessitd de modifier assez fortement l'dnoncd des rdsultats ultdrieurs (il est/t noter que, dans les gdndralisations de la thdorie, par exemple aux corps de nombres algdbriques, on ne peut dviter ces dnoncds d'aspect plus compliqud). On notera que, pour que e e Ei, il faut et il suffit que l'image de e dans le quotient A/At_ ~ de A par le lattice engendr6 par e ~ , . . . , e~_ 1 soit un dldment "primitif" de ce quotient (c'est-~-dire non divisible par un entier > 1). I1 revient au marne de dire que E iest rensemble des vecteurs ~ i xjej Off X l , . . . , X n sont des entiers tels que le plus grand c o m m u n diviseur (x~ . . . . . x,) de x~, x~+ 1. . . . . x, est dgal ~ 1. Par suite, pour qu'une forme quadratique positive non ddgdndrde F ( x ) = Y',u a u x l x j soit rdduite (ce qui veut dire que la base canonique de R" est rdduite par rapport 5, la forme), il faut et il suffit qu'elle satisfasse aux conditions: (i) F ( x l . . . . . x , ) > a , pour tout systdme d'entiers (Xl . . . . , x , ) tels que (xi . . . . .
x.)
=
1;
(ii) ai, i+l -> 0 pour 1 < i _< n - 1. I1 est clair d'ailleurs que la condition (i), jointe /t al 1 > 0 (resp. all -> O) suffit fi entrainer que F est positive non ddgdndrhe (resp. positive). Si on remarque qu'avec les notations ci-dessus on a e~ c E~ et ej-4-_ e~ ~ Ej chaque fois que i < j, on en conclut que (i) entraine les indgalitds: aii <_ ajj
(i < j)
(I)
12aul _< a,,
(i < j).
(II)
et
De plus, un thdordme fondamental, dfi/t Minkowski, affirme qu",i tout n correspond un C, > 0 tel que (i) entraine aussi rindgalitd: alia22 . . . a , ,
<_ C n det(A),
A = I1%11-
(III)
Soit M la partie de P ddfinie par les conditions (i) et (ii); soit Mo son image dans Po- L'ensemble M 6tant ddfini par une infinitd d'indgalitds lindaires et homogdnes par rapport aux a u, il est 6vident que c'est un ensemble convexe "positivement homogene". C o m m e il est dvident que tout lattice A possdde au moins une base rdduite par rapport ~ une forme donnde, Mo contient un systeme de reprdsentants de Po par rapport au groupe F0 . On peut donner "ace sujet des 6noncds beaucoup plus prdcis (voir plus loin), mais ils ont peu d'importance du point de vue des applications. Les rdsultats vraiment importants sont lids ~ l'introduction de deux familles d'ouverts dans P (resp. P0) qu'on va ddfinir maintenant.
[1957c]
365 R6duction des formes quadratiques, d'apr+s Minkowski et Siegel
Pour chaque t > l, soit S(t) la partie de P d6finie par les in6galitds :
aii < t a i + l , i + 1
(1 _< i _< n -- 1) ]
12aijl < ta,
(1 _< i < j ___ n) /
(A)
a l i a 2 2 . . , ann < C,t" det(aij) Le th60r6me de Minkowski montre que M c S(t) pour t > 1 (c'est principalemerit sous cette forme qu'on a fi l'utiliser). 11est clair que les S(t) forment une famille, croissante avec t, de parties ouvertes de P, dont P e s t la r6union. On notera So(t ) l'image de S(t) dans P0. D'autre part, la r6duction de la forme quadratique F ~t une somme de carr6s par la m6thode des alg6bristes babyloniens (m6thode connue 6galement, dans la litt6rature, sous les noms de " m 6 t h o d e de Jacobi" et "m6thode d'orthogonalisation de Schmidt", /t moins qu'il ne convienne plut6t de l'attribuer ",i quelque savant russe ...) permet, d'une mani6re et d'une seule, d'6crire:
F(x) =
~ aiaxix j
i,j 1
= Zn (di xi q- =~i+llijXj)2 i=1
j
ou encore, en termes matriciels, A = tT- D 9 T, off D est la matrice diagonale ayant dl . . . . . d, pour coefficients diagonaux (et 0 partout ailleurs), et T la matrice triangulaire (au sens strict, c'est-fi-dire ayant 1 partout dans la diagonale principa!e) dont les coefficients sont les tii pour i < j, et les bii (1 si i = j, 0 sinon) pour j _< i. Par r6currence, on d6montre ais6ment que les di, tij sont des fonctions rationnelles des a~j, fi d6nominateurs r 0 dans P. Cela pos6, soit S'(u), pour tout u > 1, l'ensemble des points de P de la forme A = tT- D 9 T, off D est une matrice diagonale, fi coefficients diagonaux d l , . . . , d,, et T une matrice triangulaire au sens strict, fi coefficients tlj pour i < j, satisfaisant aux in6galit6s:
0 < di < udi+ 1
(1 _< i _< n -- 1);
]tii[ < u
(1 _< i < j _< n).
(B)
D'apr6s ce qu'on vient de dire, il est clair que les S'(u) forment, eux aussi, une famille croissante d'ouverts de P, dont la rdunion est P. Des calculs triviaux permettent de v6rifier que tout S(t) est contenu dans un S'(u), et r6ciproquement. I1 s'ensuit 6videmment que M est contenu dans S'(u) pour u assez grand. On notera S'o(U) l'image de S'(u) dans P0Orl doit ',i Siegel le tr6s important r6sultat suivant (pour l'6noncer commod6merit, on conviendra, pour toute partie S de P, et toute matrice inversible X, de noter S x le transform6 de S par X, i.e. l'ensemble des ' X . A 9X pour A ~ S): Quels que soient t > 1 et m entier :~0, rensemble des matrices X ?t coefficients entiers, de d~terminant m, telles que S(t) c~ S(t) x ~ O, est fini. I1 s'ensuit naturellement qu'on peut en dire autant pour S'(u). C o m m e M ~ S(t), on en conclut en particulier, en prenant m = + 1, que, pour tout t > 1, S(t) est
366
[1957c] Rdduction des formes quadratiques, d'apr6s Minkowski et Siegel
contenu dans la rdunion de M e t d'un n o m b r e fini de transformds de M p a r des 616ments du g r o u p e F. Enfin, on peut v6rifier, p a r exemple p a r calcul explicite, que, dans P0 consid6r6 c o m m e espace riemannien sym6trique Go/K o, chacun des ensembles (non compacts) So(t) est de volume fini, p o u r le volume " n a t u r e l " ddfini dans l'espace Po (volume qui est naturellement invariant p a r G o, et que cette condition d6termine d'une mani6re unique fi un facteur pr6s). C o m p t e tenu de ce qui pr6c6de, cela 6quivaut n a t u r e l l e m e n t / t dire que M 0 est de volume fini, ou encore que l'espace homog6ne Go/F o est de volume total fini. La d 6 t e r m i n a t i o n du volume de ce dernier espace (l'616ment de volume invariant dans Go 6tant choisi explicitement une fois p o u r routes) est d u e / t Minkowski. P o u r m6moire, on ajoute ce qui suit, d o n t l'intdrat est d ' o r d r e historique et esthdtique mais dorlt en r6alit6 on ne semble gu6re avoir fi faire usage. En premier lieu, l'ensemble convexe M, ddfini p a r les indgalitds (i) et ( i i ) j o i n t e s fi l'indgalit6 al 1 > 0, est en r6alit6 une pyramide convexe; a u t r e m e n t dit, il suffit p o u r le d6finir d ' u n nombrefini d'indgalitds prises p a r m i celles qu'on vient de mentionner. O n en a vu (sans d6monstration, mais la d 6 m o n s t r a t i o n dans ce cas est facile) un exemple, p o u r n = 2; p o u r n = 3 et n = 4, on conna~t un syst6me explicite d'indgalitds ddfinissant M ; rien de tel n'est connu p o u r n quelconque. De plus, M est l'adhdrence, darts P, de son intdrieur. Enfin, non seulement P e s t rdunion de M e t de tous ses transform6s M x p a r les 616ments X du g r o u p e F (ou, plus pr6cis6ment, p a r les transform6s de M p a r un syst6me de reprdsentants, darts F, des classes suivant le centre { + 1,} de F : car ce centre op6re trivialement sur P), mais encore ces transformds forment une triangulation ou un " p a v a g e " de P, au sens suivant: leurs int6rieurs sont disjoints; l'intersection de deux quelconques d'entre eux est une p y r a m i d e convexe de dimension plus petite: tout c o m p a c t n'a de points c o m m u n s qu'avec un n o m b r e fini de ces transform6s.
[1957c2] Groupes des formes quadratiques indefinies et des formes bilindaires alternees 1. Quelques concepts g~n~raux. C o m m e precddemment, soient G un groupe semi-simple non compact, K un sousgroupe maximal de G, F u n sous-groupe discret de G. L'espace homogene G/K est l'espace riemannien symetrique associe 5- G. O n est amene 5- considdrer les proprietds suivantes de F: (I) v(G/F) < + c~ (v designe bien entendu le volume invariant, ou mesure de Haar, sur G; il est invariant 5- droite et 5- gauche, parce que G est semi-simple, et se transporte 5. G/F d'une maniere 6vidente). (II) II existe dans G u n ouvert U de mesure finie (i.e. v(U) < + oc) tel que U F = G (autrement dit, l'image de U dans G/F est G/F), et que U - 1 U n F soit fini (autrement dit, il n'y a qu'un hombre fini d'61ements ~' e F tels que U7 rencontre U). (III) G/F est compact. I1 est clair que (III) ~ (II) ~ (I). Mais, entre (II) et (III), il y a lieu d'inserer une propriet6 de plus; pour l'enoncer, on introduit la notion suivante. Deux groupes F, F' sont dits commensurables si F c~ F' est d'indice fini dans F et dans F'. On demontre sans peine que c'est 15-,pour les sous-groupes d'un groupe G donnd, une relation d'equivalence. II s'ensuit que les x ~ G tels que xFx i soit commensurable "a F forment un groupe F (dit " g r o u p e des transformations" de G/F), et que, si F et F' sont commensurables, les "groupes de transformations" F, F' qui leur sont associes coincident, ce qui implique en particulier que F' c ~. Cela pos6, on s'interesse aussi 5- la propriete: (M) II existe darts G u n ouvert U de mesure finie, tel que KU = U (U est " s a t u r e " par rapport/t K, ou encore est l'image reciproque, par l'application canonique de G sur G/K, d'un ouvert de G/K), que U F = G, et que U - 1U c~ xF soit fini quel que soit x ~ ~. (On observera que, dans le groupe F, toute classe 5- gauche suivant F est contenue dans une reunion finie de classes "a droite, et reciproquement, de sorte que, dans la derniere condition de (M), on peut ecrire Fx au lieu de xF sans rien changer). I1 est clair que (III) ~ (M) ~ (II). Si (II) est satisfaite, Siegel dit que F est "de premiere espece." Si (M) est satisfaite, on propose de dire que F est " m i n k o w s k i e n " dans G. Les theoremes 6nonces dans le premier expose disent essentiellement que, dans le groupe Go = PL+(R, n), le groupe Fo = PL(Z, n) ("groupe modulaire") est minkowskien; dans ce cas, on verifie facilement que le groupe des transformations ['o associe 5-F 0 est P L + ( Q , n); on prend pour ensemble U Fun des ouverts So(t), S'o(u) introduits precedemment; le fait que (M) est verifi6 (et non seulement (II)) est precisement le theoreme de Siegel. Ce meme exemple montre que (M)
367
368
[1957c] Groupes des formes quadratiques ind~finies et des formes bilindaires alterndes
n'entraine pas (III); en revanche, on ignore, (pour parler plus prudemment, le conf6rencier ignore) si (I), (II), (M) sont vraiment distinctes. En dehors des groupes fuchsiens, pour lesquels on poss6de des modes de d6finition g6om6triques (par un "polygone fondamental") et, comme dirait l'autre, fonction-th6or6tiques (rev6tement universel de surfaces de Riemann avec ramifications donn6es), il semble bien que tousles groupes connus satisfaisant 5, (I) soient des groupes/l d6finition arithm6tique, et qu'en vertu des travaux de Siegel on puisse affirmer que tous ces groupes sont minkowskiens. [N.B. On peut donner de tous ces groupes "arithm6tiques" une d6finition unique, comme suit: Soit A une alg6bre semi-simple sur Q, munie d'un antiautomorphisme involutif J; soit ~ un " m o d u l e " dans A (sous-groupe additif de type fini de A, tel que 9J~Q = A). Soit Ar l'extension de A ~ R, munie de l'antiautomorphisme qui 6tend J; soit G le groupe des automorphismes de Ar (munie de J, c'est-~-dire groupe des automorphismes de l'algbbre A r qui commutent avec J); G est semi-simple, et c'est m6me le groupe semi-simple "classique" le plus g6n6ral ("classique" signifiant que G n'a aucun facteur qui soit un des groupes simples exceptionnels). On prend pour F l e plus grand sous-groupe de G qui envoie ~ dans ~R.] I1 est facile de d6montrer que, si (II) est satisfaite, F est engendrO par les 616ments de U c~ U?, donc de typefini. On peut se demander si le 9roupe des relations entre ces g6n6rateurs de F est lui-m6me de type fini; on d6montre assez simplement qu'il en est ainsi, du moins, si KU = U [travailler dans G/K, et se servir du fait connu que G/K est simplement connexe; exercice recommand6 aux lecteurs]. Enfin, il est imm6diat que, si F est minkowskien, il en est de m6me de tout groupe commensurable /t F; on revanche, " o n " ignore s'il peut arriver que, de deux groupes commensurables, l'un soit "de premi6re esp6ce" et l'autre ne le soit pas. C'est marne 15, une des raisons pour lesquelles il est recommand6 de toujours travailler avec (M), plut6t qu'avec (II), malgr6 la complication apparente de la derni6re partie de la condition (M). 2. Formes ind~finies.
Soit, pour commencer, F une forme quadratique inddfinie non-dd9~n~rOe dans un vectoriel E de dimension n sur R Soit G le groupe orthogonal de F (groupe des automorphismes de E qui laissent F invariante). Soit U le dual de E; toute forme bilin6aire sur E d6termine canoniquement, comme on sait, une application lin6aire de E dans E'; en particulier, la forme bilin6aire F(x, y) associ6e/~ F, c'est-/tdire telle que F(x) = F(x, x), d6terminera une application lin6aire f de E sur E', qui est sym6trique (i.e. ~/ --- f ) et de rang n. Soit K un sous-groupe compact de G; il laisse invariante, comme chacun sait, au moins une forme positive non d6g6n6r6e qs; soit ~0 l'application de E sur E' associ6e fi ~P. On peut, par le choix d'une base convenable darts E (cf. Bourbaki, Alg., Chap. IX), mettre F, q5 sous la forme F = Z, d, x2, @ = Zi x~; la matrice de q~ 1.[, pour cette base, sera la matrice diagonale de coefficients dl, . . . , d,. Tout automorphisme de E qui laisse F et @ invariantes commute 6videmment ~i qo-lf ou autrement dit laisse invariants les sous-espaces Ev de E form6s respectivement des
[1957c]
369 Groupes des formes quadratiques inddfinies et des formes bilin6aires altern6cs
vecteurs propres de q~- if par rapport aux valeurs propres de ~0- if, valeurs propres qui ne sont autres que les di (ou plut6t les 616ments distincts parmi ceux-ci). Soit E+ (resp. E ) la somme directe de ceux des E,. qui sont associds fi des valeurs propres > 0 (resp. <0). Alors K est contenu dans le produit des groupes orthogonaux K + , K , des formes induites respectivement darts E+ et dans E par F; K+, K sont compacts, puisque F est positive (resp. n6gative) non ddgdn6r6e sur E+ (resp. E_). Par suite, pour que K soit sous-groupe compact maximal de G, il faut et il suffit qu'on ait K = K+ • K , ou autrement dit que K soit le groupe des automorphismes de E qui laissent invariantes F et une forme positive non d6gdn6rde do telle que q0- i f n'ait pas d'autres valeurs propres que _+ 1, ou, ce qui revient au m6me, telle que (q~ lf)2 = 1. I1 est clair alors que les points de G/K (le riemannien sym6trique associ6 5. G) sont en correspondance biunivoque avec les do possedant cette propridt6; il revient au m~me de dire qu'ils sont en correspondance biunivoque avec les couples (E+, E ) de sous-espaces compldmentaires de E tels que F induise sur E+ une forme positive non d6gdndr6e, et sur E_ une forme n6gative non d6g6n6rde et que de plus E+ et E soient orthogonaux l'un fi l'autre, par rapport F ; ceux-ci se d6terminent r6ciproquement. D'ailleurs, si (p, q) est la signature de F, E+ et E ont ndcessairement les dimensions p, q. Enfin, en vertu de la loi d'inertie, si E+ est un sous-espace de E de dimension p sur lequel F induise une forme positive non d6gdndrde, F induit ndcessairement sur l'orthogonal E_ de E+ une forme ndgative non d6g6n6r6e, et r6ciproquement. Cela permet de repr6senter, canoniquement, G/K comme partie ouverte d'une grassmannienne (fi savoir, soit la grassmannienne des sous-espaces de E de dimension p, soit celle des sous-espaces de E de dimension q); ces ouverts peuvent facilement 6tre ddfinis par des in6galit6s cxplicites. C o m m e prdcddemment, soit P l e c6ne convexe de toutes les formes positives non d6g6ndrdes dans E. A toute forme ind6finie non ddgdndrde F est associ6, d'apr6s ce qui pr6c6de, l'ensemble V(F) des dOe p telles que (~o lf)2 = 1. I1 est clair que les do ~ V(F) ont la propri6t6 do(x) > IF(x)] quel que soit x e E. Ces derni6res in6galit6s ddfinissent une partie convexe de P (ferm6e dans P), dont V(F) est la fronti6re (comme il r6sulte imm6diatement de la possibilit6 de r6duire simultan6ment F et une forme quelconque de P fi la forme diagonale). On a, pour tout x c E, IF(x) l = info~v~f~do(x), d'o/l s'ensuit ais6ment que F est d6termin6e, d'une mani6re unique au signe pr6s, par V(F). Tout cela subsiste d'ailleurs si F est "d6finie", mais devient sans intdr6t, V(F) 6tant alors r6duit/~ {F} ou bien b, { - F}. O n a l'habitude de dire, par abus de langage, que les do 6 V(F) sont les "majorantes" de F (ce sont en rdalit6 les 616ments fronti6res de l'ensemble des majorantes). Le groupe •(E, E) op6re transitivement dans P, comme on a vu, au moyen de do ---, do o X (pour l'application q~ de E sur E', canoniquement associ6e /t do, cela s'6crit ~0 --, tX. q~. X; ainsi en particulier si on 6crit en matrices, apr6s choix d'une base). Bien entendu, V(F o X) est l'ensemble transform6 de V(F) par do --* do ~ X; en particulier, V(F) est invariant par tout 616ment X du groupe orthogonal G de F; la mani6re dont G op6re sur V(F) est celle m~me qu'on obtient en transportant fi V(F), au moyen de la correspondance biunivoque entre G/K et V(F) ddfinie ci-dessus, les opdrations de G sur l'espace homog6ne G/K. Donc, pour 6tudier les propri6tds locales de la bijection G/K ~ V(F), il suffit de les 6tudier au voisinage du point d o = E x ~ 2 , p o u r F d o n n 6 e par F = x 2 + . . . + x p 2- x p +2 l _ _ , . . - x , . 2
370
[1957c] Groupcs des formes quadratiques inddfinies et des formes bilin6aires altern6es
Cela ne fait pas de difficult6; on en conclut que V(F) est une vari6t6 analytique r6elle, plong6e dans P, et que G/K ~ V(F) est un isomorphisme au sens analytique r6el (en particulier, c'est une application diff6rentiable de rang 6gal 5, la dimension pq de G/K). Hermite introduisit les id6es exposdes ci-dessus (qu'on s'est content6 d'assaisonner de sauce bourbachique) en vue de la th6orie arithmdtique ("rdduction") des formes quadratiques ind6finies. La pyramide convexe M ("domaine fondamental de Minkowski") 6tant d6finie dans P c o m m e il a 6t6 dit dans l'expos6 pr6c6dent, on dira que la forme inddfinie F est r~;duite si V(F) rencontre M ; pour que cela air un sens, il faut naturellement q u ' o n soit dans R n, puisque la d6finition de M est relative "fi une base d6termin6e. On dira qu'une forme est ~ coefficients entiers s'il en est ainsi de la matrice de la forme bilindaire associde. I1 est ais6 de voir qu'il
n'y a qu'un nombre fini de Jormes r(duites gt coefficients entiers de dOterminant donnO ( ~ 0 ) . I1 revient au mSme de faire voir que l'ensemble des matrices B symdtriques, h coefficients entiers,de d6terminant donn6 b ~ 0,tellesque V(B) c~ S'(u) ~ 0, est fini, pour chaque valeur donn6e de u. Cela vient de la propri6t6 des S'(u) contenue dans le lemme trivial suivant: L E M M E . - - I I y a une matrice n unimodulaire fi coefficients entiers, telle que, quel que soit u > 1, il y ait un u' > 1 pour lequel 'n. S'(u)- 1. n (ensemble des matrices ~n- A 1. n, pour A ~ S'(u)), soit contenu dans S'(u'). [ O n prendra pour n la matrice de la permutation (1,2 . . . . . n)--*(n,n - 1. . . . . 1), ou autrement dit la matrice (6i,,+ a_ i); le lemme r6sulte alors de ce que, darts le groupe triangulaire, T ~ T - ~ transforme tout compact en un compact (propridt6 qui ne caract6rise nullement le groupe triangulaire).] Cela pos6, V(B) c~ S'(u) est l'ensemble des A ~ S'(u) tels que ( A - IB)2 = 1, i.e. : A = BA
1B = t(n
1B)-(tTc-A-I-n).(n IB).
D o n c la matrice enti6re u - 1B, de d6terminant b, transforme un point de S'(u') en un point de S'(u); si u et par suite u' sont fixds, il n'y a, d'apr6s le thdor6me de Siegel, qu'un n o m b r e fini de matrices susceptibles de faire une pareille chose. C.Q.F.D. [N.B. O n n'a pas suppos6 B ind6finie, donc le cas des formes positives est inclus; ce cas est d'ailleurs facile/t liquider directement.] Ce qui pr6c6de sert principalement b, d6montrer que, darts le groupe orthogonal G d'une forme quadratique ind6finie F, non d6gdn6rde, ',i coefficients entiers, le " g r o u p e des unit6s [arithm6tiques]" de F, intersection de G avec le groupe des matrices de d6terminant _+ 1 sur Z, est minkowskien. P o u r cela, soit B la matrice de F ; d'apr6s ce qui pr6c6de, les matrices Bi 6quivalentes h B (i.e., transform6es de B par des matrices sur Z de d6terminant _+ 1)), relies que V(BI) rencontre S(t), sont en n o m b r e fini (pour un choix, fix6 une lois pour toutes, de t > 1). Pour chacune, soit M~ une matrice sur Z, de ddterminant _+ 1, transformant B dans Bg, i.e. telle que Bg = tM i 9B . Mi. Alors A ~ ' M i ~- A-M~ - t e s t une bijection de V(Bi) sur V(B); soit Ui l'ouvert de V(B), image par cette bijection de S(t) n V(BI); soit U la rdunion des U~. O n montre que U (plus exactement, l'image rdciproque dans G de l'ouvert
[1957c]
371 Groupes des |brmes quadratiques ind6finies et des formes bilinaaires altern6es
de G/K, image de U p a r la bijection ddfinie prdcddemment entre G/K et V(B)) a l e s propridtds 6noncdes darts (M). Q u a n t 5- la premi6re, tout revient 6videmment 5. d d m o n t r e r que, quelle que soit la forme ind6finie F, l'ouvert S(t) r~ V(F) est de mesure finie au sens de l'unique mesure invariante (par r a p p o r t au groupe G) ddfinie dans V(F); cela se fait p a r des m a j o r a t i o n s explicites. P o u r m o n t r e r que U contient un syst6me c o m p l e t de repr6sentants p a r r a p p o r t au sous-groupe (c'est la deuxi6me propridt6 b, v6rifier), soit A e V(B); on peut transformer A en un point t M - A 9M de S(t) au m o y e n d ' u n M de d6terminant +_ 1 sur Z ; celui-ci est alors dans V(B') avec B' = t M . B - M , ce qui implique, p a r d6finition des B i, que B' est l'un des Bi, doric que M i M ~ est une " u n i t 6 " de B, et aussi que le transform6 ~(M. M i 1). A - ( M . M ~ 1) par M . M F 1 est dans Ui, d o n c dans U; autrement dit, A est darts le transform6 de U p a r Mi- M - 1. Q u a n t au dernier point 5- verifier, il r6sulte du thdor6me de Siegel. C o m m e dans la r6duction des formes positives, on peut se p r o p o s e r de calculer (et non pas seulement de majorer) le volume de G/F, l'unit6 de volume 6tant explicitement choisie. LS-, on ne s'en tire pas 5- si bon march6. Le lecteur est pri6 de se r e p o r t e r 5- Siegel. En revanche, on peut, darts ce qui prdc6de, r e m p l a c e r S(t) p a r la p y r a m i d e de M i n k o w s k i ; on obtient alors un " p a v a g e " de G/K p a r un " d o m a i n e f o n d a m e n t a l " et ses transform6s p a r le groupe des unites de B, ce pavage offrant aux a m a t e u r s de jouissances esth6tiques ",ipeu pr6s les m6mes agr6ments que celui dc M i n k o w s k i darts le c6ne P des formes positives. A rioter toutefois qu'fi cause des B~ en n o m b r e fini, le d o m a i n e f o n d a m e n t a l se compose, non pas d'un, mais de plusieurs morceaux, d o n t chacun est une esp6ce de polyadre convexe; en a t t r i b u a n t fi chacun de ces m o r c e a u x une couleur diff6rente, on obtient, p o u r l'ensemble du pavage, des r6sultats fort pittoresques (cf. H. Weyl). A u t a n t qu' " o n " peut savoir, cela ne sert strictement 5- rien. Enfin, ce qui prdc6de permet de decider dans quel cas G/F est c o m p a c t ; il faut et il sutfit p o u r cela, visiblement, que S(t) c~ V(B'), ou, ce qui revient au marne, S'(u) c~ V(B') soit d'adhdrence c o m p a c t e dans le c6ne P p o u r tout B' 6quivalent B. O r c'est l'ensemble des A eS'(u) tels que B' 1AB'-IA = 1,; p o u r qu'il soit compact, il faut et il suffit que l'adh6rence du c6ne de s o m m e t 0 qu'il engendre n'ait aucun p o i n t c o m m u n , autre que 0, avec la fronti6re du c6ne P (i.e., ne contienne aucune forme d6gdn6r6e # 0 ) . Mais ce c6ne est l'ensemble des A e S'(u) tels que B'-IAB' 1A = 2 . 1 , , avec 2 _> 0; si A a p p a r t i e n t 5- l'adhdrence de ce c6ne et cst d6g6n6r6e, on a u r a B'-IAB'-~A = 0. Ecrivant que A est dans l'adh6rence de S'(u) et est # 0, on trouve que A est de la forme
o) avec C non d6gdndrde; de B ' - I A B ' 1A = 0, on conclut alors que forme
(: 0)
B '-1
est de la
372
[1957c] Groupes des formes quadratiques ind6finies et des formes bilin6aircs altern6es
d o n c a au moins un coefficient d i a g o n a l 0. P a r suite, B' 1, d o n c aussi B ' = tB' 9B ' - 1 9B', d o n c aussi B, " r e p r 6 s e n t e n t " 0 (ce qui veut dire que, si F est la forme de matrice B, F(x) = 0 a une solution rationnelle ~ 0). La r6ciproque s'ensuit de m6me. A u t r e m e n t dit, p o u r que G/F soit compact, il faut et il suffit que B ne repr6sente pas 0 (ce qui peut arriver p o u r n = 3 et p o u r n = 4; en revanche, un th6or6me classique de Meyer affirme que toute forme ind6finie fi n _> 5 variables, "fi coefficients entiers, " r e p r 6 s e n t e " 0).
3. Formes altern6es; groupe de Siegel. C'est le couplet suivant de la c h a n s o n ; il se chante sur le marne air. Soit F bilin6aire altern6e non d6g6n6r6e sur E; cela exige, bien entendu, que E soit de dimension paire 2n. Soit f l'application de E sur E' d6finie p a r F. Soit G le g r o u p e des a u t o m o r p h i s m e s de F ; soit K un sous-groupe c o m p a c t de G; il laissc invariante une q~ positive non d6g6n6r6e, "filaquelle a p p a r t i e n t une a p p l i c a t i o n q) de E sur E'. L'adjoint, p a r r a p p o r t fi ~, de l ' a u t o m o r p h i s m e i = q ~ - l f de E, est - ~ ; il s'ensuit que t est " s e m i - s i m p l e " (du p o i n t de vue matriciel, cela veut dire que z peut ~tre r6duit fi la forme diagonale, sinon sur R, en tout cas sur C), b, valeurs propres routes p u r e m e n t imaginaires. Si 1 a au moins deux valeurs propres distinctes et non imaginaires conjugu6es l'une de l'autre, E se d6compose en s o m m e directe de sous-espaces d o n t chacun est invariant p a r tout a u t o m o r p h i s m e de E qui c o m m u t e avec l; on en conclut,/t peu pr6s c o m m e au w qu'alors K ne peut 6tre maximal. P o u r que K soit maximal, il faut et il suttit que t n'ait que deux valeurs p r o p r e s distinctes, imaginaires conjugu6es l'une de l'autre; en multipliant qb p a r un facteur scalaire > 0, on peut supposer que ces valeurs propres sont + i , ce qui revient fi dire que t 2 = - 1. En ce cas, t d6termine une structure complexe sur E (on d6finira dans E la multiplication scalaire p a r les complexes au moyen de (~ + i~)x = :~x + [4lx); on 6crira E, p o u r E muni de cette structure; p o u r celle-ci, il est imm6diat que la forme bilin6aire fi valeurs complexes H = 9 + iF est hermitienne positive non d6g6n6r6e (N.B. lci, et dans ce qui suit, on note indiff6remment p a r ~ , p a r abus de langage, soit la forme q u a d r a t i q u e introduite ci-dessus, soit la forme bilin6aire associ6e.) C o m m e les 616ments de K c o m m u t e n t avec ~, ce sont des a u t o m o r p h i s m e s de E muni de sa structure complexe; il est clair alors que K est le g r o u p e unitaire d6termin6 par la forme hermitienne H. I1 contient donc toujours un centre non discret, form6 des multiples e"- 1 de l ' a u t o m o r p h i s m e identique (cela, au sens de la structure complexe). Dans le cas des formes q u a d r a t i q u e s ind6finies de signature (p, q), le centre du sous-groupe c o m p a c t m a x i m a l est non discret si p = 2 ou q = 2, et dans ce cas seulement. O n d6montre que l'existence d ' u n tel centre est n6cessaire et suffisante p o u r qu'il y air sur G/K une structure complexe invariante p a r G; on v a l e v6rifier dans le cas pr6sent. [N.B. La suftisance de la condition se justifie en g6n6ral c o m m e suit: K op6re darts G/K, avec un point fixe qui est le point de G/K qui c o r r e s p o n d fi K lui-m6me; il op6re donc sur l'espace des vecteurs tangents fi G/K en ce p o i n t ; dans cet espace, chacun des deux 616ments d ' o r d r e 4 du centre de K d6finit un a u t o m o r p h i s m e de cart6 - 1, et p e r m e t doric de d6finir une structure complexe, invariante p a r K. O n peut en faire a u t a n t en chaque point; on a ainsi une structure presque complexe;
[1957c]
373 Groupes des formcs quadratiques ind6finies et des formes bilineaires altern6es
reste ~, montrer qu'elle est int6grable. On peut le voir par exemple (d'apr6s Ehresmann) en remarquant qu'en g6n6ral, pour une structure presque complexe, le "d6faut d'int6grabilit6" s'exprime par un "tenseur mixte," celui qui donne les coefficients des co/~J~., dans l'expression des diff6rentielles d~)~ des formes ~i)~ de type (1, 0); en exprimant que ce tenseur est invariant par le centre de K, on trouve qu'il s'annule.] En d6finitive, on voit que G/K a 6t6 mis en correspondance biunivoque avec l'ensemble des structures complexes sur E pour lesquelles F est la partie imaginaire d'une forme hermitienne positive non d6gdndrde H = 4 ) + iF, et aussi avec l'ensemble V(F) des parties r6elles (I) de telles formes; comme au w V(F) est une sous-varidt6 du c6ne P des formes positives non ddgdn6rees sur E, et G/K V(F) est une bijection analytique rdelle de G/K sur V(F). Si on est dans R 2", et qu'on suppose F donn6e par une matrice fi coefficients entiers, on ddmontre, exactement c o m m e au w que le groupe des "unitds arithmdtiques" de F est minkowskien darts le groupe des automorphismes de F. Dans un expos6 ultdrieur, on ddfinira, d'une mani&e plus ou moins explicite, un ouvert U de G/K satisfaisant 5 la condition (M); ce sera fait, du moins, pour le " g r o u p e de Siegel" (ou " g r o u p e modulaire d'ordre n") proprement dit, qui est celui des unites si F est donnde dans R 2" par une matrice alternde de dOterminant 1. Si " o n " a du vice, " o n " ddfinira m~me, dans ce dernier cas, un " d o m a i n e fondamental" qui, avec ses transform6s, fournit un beau pavage de l'espace G/K. [N.B. I1 est connu que, par un choixtconvenable de 2n gdndrateurs pour le sous-groupe Z z" des vecteurs fi coordonndes enti6res dans R 2", toute formc alternde /t coefficients entiers peut s'6crire Y,i di(xiy,,+~ - x,+~yi), off les di sont des entiers, les "diviseurs 616mentaires", dont chacun est multiple du pr6c6dent. I1 n'y a donc pas besoin de la th6orie de la rdduction, dans ce cas, pour montrer qu'il n'y a, pour un d6terminant donn6, qu'un nombre fini de formes non 6quivalentes deux 5 deux. De plus, toutes ces formes sont 6quivalentes sur Q. Or il est facile de voir que les groupes d'unit6s arithm6tiques de deux formes 6quivalentes sur Q sont toujours commensurables; cela est vrai aussi, bien entendu, pour les formes quadratiques; mais ici on peut en conclure que les groupes de toutes les formes alterndes fi coefficients entiers sont commensurables au groupe de Siegel.] O n va s'occuper maintenant de structure complexe. Pour cela, on introduit le "complexifi6" de E, qu'on notera Ec (pour raison typographique, au lieu de la notation canonique Ec; c'est, c o m m e on sait, E | C muni de sa structure vectorielle sur C; on consid6re E c o m m e plong6 dedans de la mani6re 6vidente). Tout automorphisme t de E, de carr6 - 1 , se prolonge 5. Ec en un automorphisme analogue, qui d6termine une d6composition de Ec en somme directe des sousespaces Vi, V_~ formds des vecteurs propres relatifs aux valeurs propres i resp. - i de z; Vii, V_i sont sous-espaces de Ec sur C, donc sont espaces vectoriels sur C, de dimension n; on a d'ailleurs V_~ = ~, off, suivant l'usage, la barre denote l'imaginaire conjugu6 (d6fini dans Ee de la mani6re 6vidente). Si iE ddsigne l'ensemble des vecteurs "imaginaires purs" de E~. (image de E par x ~ ix), Ec = E @ iE est une somme directe; si ~ ("partie rdelle") est le projecteur de E~ sur E qu'elle d6termine, il est imm6diat que 9l induit sur Vii un isomorphisme de la structure complexe de Viisur la structure complexe de E qui est d6termin6e par l, c'est-',i-dire
374
[1957c] Groupes des formes quadratiques indefinies et des formes bilineaires alternees
sur celle de E,. D o n c ~est c o m p l e t e m e n t ddtermin6 p a r la d o n n e e de Vii.Rdciproquement, soit V~un sous-espace de Ec de d i m e n s i o n n sur C; p o u r que ~ induise sur V~une bijection de V, sur E, il faut et il suffit q u ' o n air l'une des relations 6quivalentes V~ r~ iE = {0}, Vi r~ E = {0}, Vi r~ V~ = {0}; lorsqu'il e n e s t ainsi, V~ p e r m e t doric de ddfinir sur E, p a r t r a n s p o r t de structure au moyen de N, une structure complexe, donc un a u t o m o r p h i s m e t de E de carr6 - 1 ; si alors on 6tend t it E,, V~sera l'espace des vecteurs p r o p r e s de l relatifs it la valeur p r o p r e i. Soit (5 la grassmannienne complexe des sous-espaces de Ec de dimension n sur C; d a n s (5, soit ,3 l'ensemble des sous-espaces d o n t l'intersection avec E se rdduit it {0} ; c'est un ouvert dans (5; d'apres ce qui precede, il s'identifie avec l'ensemble des structures complexes sur E. Si on s'est d o n n e c o m m e precedemment, sur E x E, une forme bilindaire alternde non degeneree F, celle-ci peut s'etendre it une forme Fc sur Ec • Ec; de m e m e p o u r l'extension q~c d'une forme symetrique qb. Si on a, sur E • E, qS(x, y) = F ( - tx, y) (ce qui equivaut it la relation z = q~- l f e c r i t e au debut de ce numero), la relation a n a l o g u e sera vraie p o u r les extensions de F, 0), t, it Ec; en particulier, sur V~, Fc et q5c induisent une forme alternee F ' et une forme symetrique qb' telles que l'on air @' = - iF', ce qui exige evidemment F ' = 0. A u t r e m e n t dit, V~est alors un espace isotrope maximal de F~ ( " i s o t r o p e " signifie j u s t e m e n t que F~ induit 0 sur Vii; on trouve alors, par exemple par le choix d ' u n e base convenable, que Vii, etant de d i m e n s i o n n, est isotrope maximal parce que F c e s t non ddgeneree). Ces espaces forment une sous-variet6 (analytique complexe) ~9 de la grassmannienne (g; on verifie sans difficulte que le g r o u p e des a u t o m o r p h i s m e s de E~ qui laissent F, invariante (le " c o m p l e x i f i e " du g r o u p e G) opere transitivement sur ~3. Reciproquement, soit V ~ r~ ~3; p u i s q u ' o n a E c = Vii| V i, on peut, quel que soit z~E~, 6crire z = u + v, u e Vii, v e V-i, et on a alors zz = iu - iv; si de meme on a z' = u' + v', avec u' ~ V/, v' e I/ i, on a u r a Fc(u, u') = 0 et Fc(v, v') = 0 parce que V/et p a r suite V~ = ~ sont isotropes p o u r F,.; cela p e r m e t de calculer F~(--lz, z') et de voir que cette forme bilineaire est symetrique. I1 faut exprimer de plus que @ est positive non degeneree sur E, ce qui equivaut b. F ( - tx, x) > 0 quel que soit x ~ 0 dans E. Or, l ' i s o m o r p h i s m e de E, sur V/, inverse de r i s o m o r p h i s m e de V/sur E, induit sur V~p a r 9~, s'ecrit x ~ z = x - i - t x (verification immediate); en tenant c o m p t e de ce que Vii, l / ~ = ~ sont isotropes p o u r Fr l'inegalite precddente s'ecrit encore iFc(z, ~) < 0 quel que soit z r 0 dans V~. I1 est clair que les V~satisfaisant it cette condition forment un ouvert f~ dans ~ ; ce qui precede implique que cet ouvert n'est pas vide (puisque tout point du riemannien symetrique G/K determine justement un point de f~). I1 est clair aussi que Vii ~ f~ implique Vii~ ~ ; sinon, en effet, il y aurait un z 4= 0 dans V~ r~ E, et on aurait z = ~, iF(z, z) < 0, ce qui est idiot. De ce qui precede resulte d o n c que G/K s'identifie it fL qui est une variete (analytique complexe) plongee dans (5. D'ofl la structure complexe de G/K. Le g r o u p e G des a u t o m o r p h i s m e s de F (i.e. des a u t o m o r p h i s m e s de E qui laissent F invariante) opere sur ~9 et sur f2 d'une maniere evidente, y laisse invariante la structure a n a l y t i q u e complexe, et sa maniere d ' o p e r e r sur fl est celle qui resulte, p a r t r a n s p o r t de structure, de son o p e r a t i o n sur G/K. Satisfaction gendrale. Profitant de l'absence de Dieudonne, on va traduire qa en matrices, p o u r faire le j o i n t avec Siegel et p o u r se mettre en etat de calculer q u a n d on ne peut pas faire
[1957c]
375 Groupes des formes quadratiques inddfinies et des formes bilin6aires altern6es
a u t r e m e n t (~a arrive encore quelquefois). O n prend une base (el . . . . R p o u r laquelle la matrice de F soit
,
e2n ) de E sur
soient E', E" les sous-espaces engendr6s sur R, respectivement, p a r (el . . . . . e,) et (e,+~ . . . . . e2,); ce sont des sous-espaces de E isotropes m a x i m a u x p o u r F. Si V~E f l , et que l, q), etc., aient le m6me sens que ci-dessus, @ sera positive non d6g~n6r6e sur E, d o n c sur U , ct on p o u r r a choisir dans E' n vecteurs o r t h o n o r m a u x p o u r q~; ils le seront alors aussi, dans E,, p o u r la forme hermitienne H = 9 + iF (puisque E' est isotrope p o u r F); ils formeront d o n c une base de E, sur C, ce qui entraine que E' et rE' sont suppldmentaires dans E. Alors Vii est transversal au complexifi6 E',, de E'; en effet, E'~ est l'ensemble des x' + iy', avec x' e E', y' c E'; si un tel p o i n t est dans Vii,on a y' = - zx' e E' r~ zE', donc x' = y' = 0. De marne Viiest t r a n s v e r s a l / t E',[. Choisissons darts Vi n vecteurs formant une base de Vi (sur C); 6crivons-les c o m m e " c o l o n n e s " (matrices ~ 2n lignes et 1 colonne) au moyen de (el . . . . , e z n ) pris c o m m e base de Ec sur C; cela d o n n e une matrice ~t 2n lignes et n colonnes (sur C), q u ' o n peut 6crire
V ' o/1 U, V sont deux matrices/t n lignes et n colonnes;
si on change les vecteurs de base choisis dans V~, cela revient/L multiplier U, V ~ droite p a r une m~me matrice carr6e inversible. Puisque V~est transversal/L E'~' la matrice U est de rang n, c'est-fi-dire inversible. Ecrivons que V~ est isotrope p o u r F ; cela s'exprime p a r la formule
:0, ou a u t r e m e n t dit ' U - V = tV. U. D e m~me, 6crivons que iFc(z, 2) < 0 p o u r tout z r 0 dans Vi; cela signifie (1/i)(~U 9V - ' V - U) > 0 (le premier m e m b r e est visiblement une matrice hermitienne). P o s o n s Z = V U - 1, matrice qui est ind6pendante de la base choisie dans Vii.La premi6re des relations ci-dessus s'6crit ' Z = Z ; Z est sym6trique. La seconde s'6crit (en m u l t i p l i a n t / t droite p a r U - ~, fi gauche p a r ' U - ~, ce qui ne modifie pas le fait que le premier m e m b r e est hermitien positif non d~g6n6r6) ( 1 / i ) ( Z - ~Z) > 0; a u t r e m e n t dit, si on 6crit Z = X + i Y avec X, Y sym6triques r6els, Y doit 6tre positive non ddg6ndr6e. O n a ainsi obtenu une bijection de fL donc en d6finitive de G/K, sur l'espace de Siegel ~ , form~ des matrices sym6triques Z = X + i Y sur C telles que Y > 0; ~ peut ~tre consid6r6 c o m m e un ouvert de C N, avec N = n(n + 1)/2, muni de la structure complexe induite p a r celle de C ~. L ' o p 6 r a t i o n de G sur ~ s'~crit imm6diatement. En effet, avec les notations ci-dessus, un 616ment de G s'6crira sous forme d ' u n e matrice carr6e
376
[1957c] Groupes des formes quadratiques inddfinies et des formes bilineaires altern6es
off A, B, C, D sont des matrices "5. n lignes et n colonnes sur R; cette matrice doit satisfaire fi la condition
qui exprime qu'elle laisse F invariante. Cette matrice op6re sur V~, d+finie par les matrices U, V, au moyen de la formule
ou autrement dit:
(U, V)--+ (AU + BV, CU + DV); elle opere donc sur Z par la formule
Z ~ (C + DZ). (A + B Z ) - I Enfin, on peut donner de l'op6ration sur ~ du groupe modulaire (sous-groupe de G form6 des matrices "a coeff• entiers) une interpr6tation int6ressante, et meme importante. En effet, F 6tant donnde dans E, les points de ~ sont, d'apr6s ce q u ' o n a vu, en correspondance biunivoque avec les structures complexes E, q u ' o n peut mettre sur E, pour lesquelles F est partie imaginaire d'une forme hermitienne H >> 0. Supposons donn6 en meme temps dans E un lattice A tel que F soit/t valeurs enti6res sur A x A ; E/A est alors un tore de dimension (r6elle) 2n, sur lequel F d6termine une classe de cohomologie enti6re de dimension (r6elle) 2. Toute structure complexe sur E d6termine sur E/A une structure de tore complexe (de dimension complexe n); pour que celui-ci soit une vari6t6 ab61ienne, il faut et il suffit qu'il existe dans E une forme hermitienne >>0 dont la partie imaginaire soit b. valeurs enti6res sur A x A; et, lorsqu'il e n e s t ainsi, il y a sur E/A un "diviseur positif" appartenant 'a la classe de cohomologie d6termin6e par cette partie imaginaire; muni de cette classe, E/A s'appelle alors une vari~t~ ab~lienne polaris~e. O n voit donc qu'/t tout point de ~ correspond sur E/A une structure de varidt6 ab61ienne polaris6e par F; pour qu'/t deux points corresponde la m6me structure, il faut et il suffit qu'ils se d6duisent l'un de l'autre par un automorphisme de E qui laisse invariants la forme F et le lattice A, donc un 616ment du groupe discret F des automorphismes de F qui sont fi coefficients entiers lorsqu'on prend pour base un syst6me de gdndrateurs de A. Autrement dit, les points de ~ / F (quotient de par la relation d'6quivalence d6finie dans | par le groupe discret F) sont en correspondance biunivoque avec les structures de vari6t6 ab61ienne polaris6e par F q u ' o n peut ddfinir sur E/A. Lorsque F est de d6terminant 1 sur A (c'est-fi-dire a tous ses diviseurs 616mentaires sur A 6gaux/~ 1), le groupe F qu'on obtient est le groupe modulaire de Siegel proprement dit; les vari6tds ab61iennes correspondantes sont dites "vari6t6s abdliennes polaris6es de la famille principale" (toute jacobienne est une telle variet6).
[1957c]
377 Groupes des formes quadratiques inddfinies et des formes bilindaires altern6es
Bibliographie Minkowski, Hermann. Geometrie der Zahlen. Leipzig und Berlin, B. G. Teubner, 1910; New York, Chels,ea, 1953. Minkowski, Hermann. Diskontinuit~itsbereich ffir arithmetische Aquivalenz, J. ffir reine und angew. Math., t. 129, 1905, p. 220-274; Gesammelte Abhandlungen, Band 2, Berlin, B. G. Teubner, 1911, p. 53 100. Siegel, Carl Ludwig. Einffihrung in die Theorie der Modulfunktionen n-ten Grades, Math. Annalen, t. 116, 1939, p. 617-657. Siegel, Carl Ludwig. Einheiten quadratischer Formen, Abh. math. Sem. Hamburg Univ., t. 13, 1940, p. 209 239. Siegel, Carl Ludwig. Discontinuous groups, Annals of Math., t. 44, 1943, p. 674 689. Siegel, Carl Ludwig. Symplectic geometry, Amer. J. Math., t. 65, 1943, p. 1 86. Weyl, Hermann. Theory of reduction for arithmetical equivalence, 1., Trans. Amer. math. Soc., t. 48, 1940, p. 126 164; II., Trans. Amer. math. Soc., t. 51, 1942, p. 203-231. Weyl, Hermann. Fundamental domains for lattice groups in division algebras, Comment. Math Helvet., t. 17, 1944/45, p. 283-306.
[ 1958a] Introduction h l'6tude des vari6t6s k ihl6riennes (Pr6face) L a th6orie des vari6t6s k~ihl6riennes a pris un g r a n d essor depuis un q u a r t de sibele. Ces vari6t6s f u r e n t d6finies p o u r la p r e m i b r e fois, semble-t-il, dans une n o t e de K~ihler de 1933 (Hamb.Abh. 9, p. 173). Mais leur i m p o r t a n c e n ' a p p a r u t q u ' h la suite des p r e m i e r s t r a v a u x de H o d g e et s u r t o u t de l'expos6 d ' e n s e m b l e q u ' i l en d o n n a au c h a p i t r e V de son livre The theory and applications o[ harmonic integrals ( C a m b r i d g e 1941), off, i n d 6 p e n d a m m e n t de Kfihler, il e n t r e p r e n d l ' 6 t u d e s y s t d m a t i q u e de la m 6 t r i q u e (( kfihl6rienne )) q u ' o n p e u t d6finir sur t o u t e varidt6 alg~brique sans p o i n t m u l t i p l e plong6e dans un espace projeetif, et en tire des cons6quences tr+s i m p o r t a n t e s p o u r la g6omdtrie alg6brique. C'est p r i n c i p a l e m e n t dans la direction ainsi inaugurde p a r lui que les recherches se s o n t poursuivies depuis lots. On ne t r o u v e r a pas iei une m o n o g r a p h i e de ce sujet, mais, c o m m e l ' i n d i q u e le t i t r e , une simple i n t r o d u c t i o n , bas6e sur des cours professds h Chicago et h G S t t i n g e n d a n s les derni~res ann6es. Ce v o l u m e a u r a a t t e i n t son b u t s'il facilite au l e c t e u r l'~tude des t r a v a u x rdcents sur la question, et particuli~r e m e n t de ceux de K o d a i r a et ses dl~ves. J e n ' a i pu ndanmoins rdsister h la t e n t a t i o n d ' i n s d r e r un c h a p i t r e t r a i t a n t de la thdorie des fonctions t h ~ t a et des vari~t~s ab~liennes sur le corps des complexes. Cette th~orie p e u t ~tre considdrde c o m m e celle d ' u n t y p e p a r t i c u l i e r de s t r u c t u r e s kfihleriennes (les s t r u c t u r e s i n v a r i a n t e s sur un tQre), et c ' e s t m~me ce p o i n t de vue qui c o n d u i t h la d d m o n s t r a t i o n la plus n a t u r e l l e d ' u n des p r i n c i p a u x th~or+mes d'exist e n c e de la th~orie (le th~or~me d i t (( d ' A p p e l l - H u m b e r t )9. I1 n ' a p a s dt~ donnd de b i b l i o g r a p h i e ; on en t r o u v e de fort c o m p l e t e s dans plusieurs o u v r a g e s r~cents. J e me suis efforcd de rdduire au m i n i m u m la s o m m e de connaissances exigde du lecteur : quelques r~sultats dl~mentaires Reprinted from Introduction (~ l'~tude des variet~s kiihl~riennes, Hermann, Paris, by permission of Hermann, 6diteurs des sciences et des arts.
379
380
[1958a] l0
INTRODUCTION A L ' E T U D E DES VARIE,TES K,~HLERIENNES
d'analysc et de thdorie des fonctions ; quelques notions d'alg~bre et de topologie gdndrale (pour lesquelles il sera en gdndral renvoyd aux Eldmenls de N. Bourbaki) ; les d~finitions essentielles de la th~orie des vari~tds diff~rentiables (qu'on trouvera dans le volume de G. de R h a m , Varidlds di[]~rentiables, Paris, HerInann 1955, paru dans cette m~me collection) ; la ddfinition des op~rateurs . , 8 , A de la th~orie des formes harmoniques (expos~e dans le m~me volume) ; et, en quelques points, quelques notions sur la cohomologie enti~re. Encore ai-je, dans la mesure du possible, rappel~ les d~finitions et r~sultats dont il aura h ~tre fait usage. L'indication (( de Rham, w ~, renverra toujours au volume q u ' o n vient de citer ; les renvois Bourbaki seront faits sous la forme canonique. Des thdor~mes d'existence de la thdorie des formes harmoniques, qui jouent bien entendu un r61e essentiel dans le present volume, le lecteur n'a fi connaitre que l'existence des opdrateurs H e t G de de R h a m avec les propridtds formelles dnoncdes au n ~ 1 du Chapitre IV. Pour le cas particulier des tores, une ddmonstration directe de ces rdsultats, inddpendante de celle qui est donnde dans le volume de de R h a m pour le cas gdn~ral, est exposde au n ~ 2 du Chapitre IV, ce qui permettrait, si on le ddsirait, d'aborder la thdorie des fonctions th~ta sans s ' a p p u y e r sur la th~orie g~n~rale des formes harinoniques.
Paris, le 31 mai 1957.
[1958b] On the moduli of Riemann surfaces 1 TO E m i l Artin on his sixtieth birthday The p u r p o s e of the following pages is partly to clarify my own ideas on an interesting topic, at a stage when they are still unripe for publication, but chiefly to be present by p r o x y at Artin's b i r t h d a y celebration. In speaking of these ideas as " m y own", my intention is not to claim originality for them. They are little m o r e than a c o m b i n a t i o n of those of Teichmfiller with the ideas on the variation of complex structures, recently introduced by K o d a i r a , Spencer and others into the t h e o r y of moduli. The first concept to be elucidated is that of reinforcement of structure. Let S o be an oriented c o m p a c t surface of genus g > 1, given once for all. Let S be a Riem a n n surface of genus g, i.e. a surface of genus g, p r o v i d e d with a c o m p l e x - a n a l y t i c structure (or, what a m o u n t s to the same, an oriented surface of genus g with a conformal structure, or again an oriented surface of genus g with a class of conformally equivalent ds2). By a "class of m a p p i n g s " of S o into S, we shall u n d e r s t a n d a class of c o n t i n u o u s mappings, equivalent under h o m o t o p y ; a class will be called " a d m i s sible" if it contains at least one orientation-preserving h o m e o m o r p h i s m of S o onto S. A R i e m a n n surface S, p r o v i d e d with the a d d i t i o n a l structure defined on it by assigning one admissible class of m a p p i n g s of S o into S, will be called a Teichmtiller surface. P e r h a p s the most r e m a r k a b l e of Teichmfiller's results is the following: when p r o v i d e d with a rather obvious " n a t u r a l " topology, the set O of all classes of m u t u a l l y isomorphic Teichmtiller surfaces is h o m e o m o r p h i c to an open cell of real dimension 6g - 6. This global result will neither be used nor discussed in the following pages, the chief p u r p o s e of which is to consider the local p r o p e r t i e s of O and to define on it a " n a t u r a l " c o m p l e x - a n a l y t i c structure, of complex dimension 3g - 3, and a " n a t u r a l " H e r m i t i a n metric. The above "reinforcement of s t r u c t u r e " can be modified in various ways. Instead of admissible classes of mutually h o m o t o p i c mappings, one might wish to introduce classes of m u t u a l l y isotopic orientation-preserving h o m e o m o r p h i s m s of S o onto S; unless I am mistaken, k n o w n results in surface-theory imply that this w o u l d not actually differ from what we have done. O n the other hand, one can also consider a weaker type of reinforcement, in which two m a p p i n g s of S o into S are considered equivalent if they induce the same h o m o m o r p h i s m of the onedimensional h o m o l o g y group of S ~ H1(S~ into H i ( S ) ; as a t e m p o r a r y terminological prop, let us call a "Torelli surface" the surface S with the a d d i t i o n a l structure d e t e r m i n e d by an admissible class of m a p p i n g s for this wider concept of equivalence.
Part of this was done, I am somewhat ashamed to say, while on contract with the Air Force. The opinions of the Air Force do not necessarily coincide with mine.
381
382
[1958b1 On the moduli of Riemann surfaces
Call ~z~ the fundamental group of S ~ with an origin a ~ chosen once for all ; it is the group generated by 2g generators A ~ with the well-known relation, which we shall write as R(A ~. . . . . A~ = 1. Let f , f ' be two mappings ofS ~ into S; they induce h o m o m o r p h i s m s h, h' of ~o into ~(S,f(a~ and ~(S,f ,( a 0) ) , respectively. By considering the universal coverings of S o and S, one sees easily thatfandf' are homotopic if and only if h' can be derived from h by moving the origin of the fundamental group of S along a suitable path f r o m f ( a ~ t o f ' ( a ~ If, by an obvious "abuse of language", we agree to speak of " t h e " fundamental group ~(S) of S (which is then defined intrinsically only up to an inner automorphism), we may say that f a n d f ' are h o m o t o p i c if and only if they induce h o m o m o r p h i s m s of ~0 into ~(S) which are equivalent under inner automorphisms of ~z(S). In particular, an admissible class of mappings of S o into S will define a class of isomorphisms of ~o onto it(S), equivalent under inner automorphisms of ~(S); such a class will be called admissible; and the images of the A ~ under an isomorphism in such a class will be called an admissible set of generators of ~(S). Thus a Teichmfiller surface is defined by selecting on S an admissible set of generators of ~(S), provided two such choices are considered equivalent whenever they can be derived from each other by an inner automorphism. We shall denote by _S any Teichmtiller surface with the underlying Riemann surface S. Similarly, a Torelli surface _Swill be defined by selecting on S an admissible set of generators for the h o m o l o g y group Hi(S); here we have no equivalence relation between such sets. By the definition of an admissible set, the intersection-matrix for an admissible set of generators is
This implies in the well-known manner that we can define on the Torelli surface S_ a normalized set of differentials of the first kind, for which the period-matrix (for the given set of generators) has the form II1,Z(S)II, where Z(S) is a g x g matrix. More precisely, if ~ is the Siegel space of symmetric g x g matrices with positivedefinite imaginary part, Z(_S) is a point of ~ ; if S' is another Torelli surface with the same underlying Riemann surface S, Z(S_') will be a transform of Z(_S) by an element of Siegel's modular group. Thus we have a mapping _S ~ Z(_S) of the set Z of all classes of mutually isomorphic Torelli surfaces into the Siegel space ~. The set Z has an involutory a u t o m o r p h i s m S ~ h(S), viz. the one which leaves the underlying Riemann surface S of S unchanged and induces on Hi(S) the automorphism x --, - x (it is easily seen, e.g. by considering a hyperelliptic surface of genus g, that this automorphism changes one admissible set of generators of H1(S) into another). It is clear that Z(h(S_)) = Z(S), i.e. that _S and h(S) have the same image in the Siegel space. Moreover, Torelli's classical theorem asserts that two Torelli surfaces S, S' have the same image Z(S) = Z(S') in the Siegel space if and only if _S' = _S or S' = h(S), and that S = h(S) if and only if S is hyperelliptic. Call Z1 the image of Z by S_ --, Z(_S); we see that the inverse image of a point of Z 1 by that mapping consists of one or two points according as the corresponding
[1958b]
383 On the re|
of Riemann surl:aces
Riemann surface is hyperelliptic or not. Combining this mapping with the obvious " n a t u r a l " mapping of | onto E, we get a mapping of | onto El, which we write as =S-~ Z(S). Finally, if ~JJl is the set of all classes of mutually isomorphic Riemann surfaces of genus g, we have a natural mapping of 21 onto ~ . It will be seen that, when O is provided with its natural complex structure, S ~ Z(=S) is a holomorphic mapping of O into the Siegel space. Actually 21 is an analytic subvariety of ~, whose points are all simple except those corresponding to hyperelliptic Riemann surfaces (this, with an important additional statement concerning sets of local coordinates in the neighborhood of a simple point of Y~I, is Rauch's theorem). As to ~JJ~,there is virtually no doubt that it can be provided with a structure of algebraic variety (non-complete, of course, and with multiple points), the "variety of moduli", so that the natural mapping of | onto ~ is holomorphic. Let _S be a Teichmfiller surface; this consists of a Riemann surface S with a preferred choice of generators A1 . . . . . A2o for ~(S), these being defined only up to an inner automorphism. We can represent conformally the universal covering of S onto the upper half-plane H = { z l J ( z ) > 0}; it is well-defined up to a hyperbolic displacement. This determines a representation of ~r(S) as a discrete group F of hyperbolic motions, generated by 2g elements z--* aiz = (c~iz + fli)/(?iz + (~i), with :tg6i - fizT~ = 1 and R(al . . . . , a2o) = 1. An inner automorphism of ~(S), or a hyperbolic displacement in FI, will merely transform F by such a displacement; conversely, if two Teichmfiller surfaces determine two sets (al), (a'i) which differ merely by an inner automorphism of the hyperbolic group, they are isomorphic. As remarked by Siegel (Math. Ann. 133), one can select that automorphism in a unique manner so as to get f12o = 72g = 0, ~2g > l, fig = ~ O ; as the matrices for the a~ are determined only up to a factor + 1, one can further normalize them by taking fi~ > 0 for 1 < i _< 2g - 1. In that normalization, all the a~ are uniquely determined by the 69 - 6 numbers (ct~, fl~, 6i) for 1 _< i _< g - 1 and g < i _< 2g - 1 ; more precisely, the relations ~6~ - ~ i ~ i = 1, e(al, . . . , 0"20 ) = 1 together with those which express the normalization, determine the 7~ and the coefficients of o"g, cr2o by means of a set of algebraic equations with non-vanishing functional determinant in the neighborhood of any point of the coordinate space ~6r corresponding to a group such as F. Let O~ be the subset of N6~-6 consisting of the points which can be obtained in this manner; the above remarks show how to define a bijection of O onto 01. It will be seen that this bijection is a real-analytic isomorphism when | is provided with its natural complex-analytic structure; this implies that O1 is an open subset of ~ 6 g 6. Moreover, we shall give explicit formulas to determine the complex structure on O1 which can be derived from that of O by that bijection. In order to justify the statements that have been made so far, we shall make use of the Kodaira-Spencer technique of variation of complex structures. This can be introduced in an elementary manner in the case of complex dimension 1, which alone concerns us here; this, in fact, had already been done by Teichmfiller; but he had so mixed it up with his ideas concerning quasi-conformal mappings that much of its intrinsic simplicity got lost. Perhaps the worst feature of his treatment, in the eyes of the differential geometer, is that his extremal mappings are destructive
384
[1958b] On the moduli of Riemann surfaces of the differentiable structure; this corresponds to the fact that his metric on | is almost certainly not to be defined by a d s 2, even though it is presumably a Finsler metric. Let __So be a Teichmfiller surface. In order to avoid the use of coordinate neighborhoods on the underlying Riemann surface So, we represent So as l-I/F, where FI is the upper half-plane and F is a discrete subgroup of the hyperbolic group; _SO will then be defined by a preferred set of generators (~i) for F, which we may assume to be in normalized form. We consider a small "variation of structure", depending differentiably upon some real or complex parameters u; this can be defined as the conformal structure on H / F determined by a complex-valued differential form ~ = dz + # d~, where # is a complex-valued function of z and of the parameters u, such that any element a of the group F merely multiplies ~ by a scalar factor; this amounts to saying that # dS/dz is formally invariant under F, a property which we also express by saying that g is of type ( - 1, 1) for the usual complex structure of FI/F. When # is so given, the complex structure of the varied surface Su underlying the varied Teichmfiller surface S,, is the one for which ~ is a differential form of type (1, 0) at every point; for this to have a meaning, we must have ]#[ < 1 for all z e F I . We assume that # = 0 for u = 0, i.e. that, for u = 0, S u is no other than _So.The universal covering of S, is FI with the conformal structure determined by ~; for each u, this can be conformally mapped upon H with its natural conformal structure; call F , the differentiable h o m e o m o r p h i s m of H onto itself which realizes this conformal mapping, i.e. which is such that (for fixed u) dF,, differs from ~ only by a scalar factor, i.e. is of the form dF,, =j'~, w h e r e f i s a complex-valued function, everywhere r 0 in II. The conformal mapping F, can be normalized in various ways, e.g. by prescribing that a given point of I1 and a given direction through that point remain fixed under F,. A respectable firm of ellipticians, w h o m I consulted concerning the properties of F,, has assured me that it depends differentiably on the parameters u, and that it is real-analytic in the u's if this is assumed of #. N o w assume that u is a single real parameter; then (c~/~u),_ o is an "infinitesimal variation" in the sense of Kodaira-Spencer; this operator will also be denoted by a dot. It is determined by
which is again a complex-valued function of type ( - 1 , 1). The infinitesimal variation is trivial if and only if the varied structure can be obtained from the initial structure (that of So) by an infinitesimal deformation of the surface. The latter will be defined by a vector-field, i.e. by a complex-valued fuction ~ of type ( - 1, 0); and one finds at once that such a vector-field ~ determines the infinitesimal variation of structure given by v = ~/0~. Therefore we introduce the Kodaira-Spencer space for So, which we define as the quotient of the space of all functions v of type ( - 1, 1) by the space of functions v = 0~/~5 with ~ of type ( - i, 0), both being considered as vector-spaces over C. We shall denote by D = D(v) the element of that space determined by a given infinitesimal variation v = / i . Let ~o be a quadratic differential on So; this can be written as (~) = q. dz 2,
[1958b]
385
On the moduli of Riemann surfaces where q is holomorphic of type (2, 0). Notations being as above, consider the integral SS qv d5 dz, taken over So; this has a meaning, since the integrand is invariant under F. Stokes's theorem shows at once that this is 0 if v = ~?~/c~5with of type ( - 1, 0); therefore it depends only upon co and D = D(v), and m a y be written as ((~), D). For any v, one can solve the equation v = &p/(?~; the solution will be uniquely determined modulo a holomorphic function of z in Fl. If we again assume that v is of type ( - 1, 1), any solution ~0 of that equation will be such that, for every ~ e F, the function ~/J~ = (p -
dz
- ~o'~
(1)
d(~z)
(where (S stands for the function defined by (p~(z) = q)(az)) is holomorphic in H. These functions satisfy the relations dz
~'o, = 0 ; d ( ~
+ ~
(2)
for all a, z in F. Conversely, given a "cocycle" (~9,), i.e. a system of holomorphic functions ~ in 11, satisfying (2) for all a, z in F, let F operate on H x C by the formula 0 % t) =
o-z, d z
(t -
~(~))
;
then the quotient (11 • C)/F is a complex-analytic fibre-bundle with base So, the fibre being the plane C (with the group t ~ at + b). By a well-known theorem on fibre-bundles (which, in the present case, can also easily be verified directly), this has a differentiable cross-section t = ~0(z); this means that (1) has a solution ~0, which is a complex-valued, real-differentiable function. Another general theorem on fibre-bundles (Serre, GAGA), which could also, in this case, be verified directly without difficulty, tells us that (H x C)/F is an algebraic bundle over So; therefore it has a meromorphic cross-section, so that (1) has a meromorphic solution ~o'. We conclude that the Kodaira-Spencer space can be defined in any one of the following manners : (a) as above, as the quotient of the space of functions v of type ( - 1 , 1) by the space of functions v = c?~/c?z with ~ of type ( - 1, 0); (b) as the space of functions qo such that v = ~?q)/c~5is of type ( - 1, 1), divided by the sum of the space of all holomorphic functions and of the space of functions of type ( - 1, 0); (c) as the space of holomorphic cocycles ( ~ ) , i.e. of systems of holomorphic functions satisfying (2), divided by the space of trivial cocycles, i.e. of those for which (1) has a holomorphic solution (p; (d) as the space of meromorphic functions q/ such that all the functions ~o' - (dz/daz)tp '~ are holomorphic, divided by the sum of the space of holomorphic functions and of the space of meromorphic functions satisfying ~0' = (dz/daz)(p '~ for all a.
386
[1958b] On the moduli of Riemann surfaces
In (d), since we are taking the meromorphic functions modulo the holomorphic functions, it is only the principal parts (at the poles) that must be considered; the condition on q~' implies that these are determined once they are given in a fundamental domain. Now, notations being as above, we can write
(co, D)= f f q v d S d z =
~qqodz,
(3)
where ~ means the integral over S 0, or, what amounts to the same, over a fundamental domain for F in I1, and ~ means the integral over the positively oriented contour of such a fundamental domain. The latter may be taken as a polygon with 4(J sides, corresponding (in that order) to al, ao+ 1, ~ ~, a~+~l, a2 . . . . . ~2g~. It is clear that .~ q2 dz is 0 whenever 2 is a continuous function, defined only along the contour of integration, such that, formally, 2/dz takes the same values along corresponding sides of the contour; for then the same is true of q)~ dz, so that the integrals along corresponding sides cancel each other. In particular, we can take 2 = q) - q)', where ~0' is defined as above ; this shows that (co, D) can be written as qq/dz, i.e. that it is the sum of the residues of the meromorphic differential (~o'/dz)co over S 0. It is a well-known consequence of Riemann-Roch that this sum, for a given ~o', will be 0 for all quadratic differentials co if and only if there is on So a " m e r o m o r p h i c differential of degree - 1" q<'/dz having the same principal parts as (p'/dz at every point, and also that, given any quadratic differential co, one may choose the principal parts of ~o'/dz in the fundamental domain, and hence everywhere in H, so that ~ (q)'/dz)co # O. This proves that the space of quadratic differentials on So is the dual of the Kodaira-Spencer space for S 0, the duality between them being given by the bilinear form (co, D). One would have been led to the same conclusion by combining the general results of the Kodaira-Spencer theory with Serre's duality theorem, and specializing these to the case of complex dimension 1. N o w observe that the conformal mapping F , must transform 17 into another group F,, of hyperbolic motions, consisting of the elements or, = F~-1 o ~ o F,; the transforms a~(u) of the a~ are then a set of preferred generators for S,. This shows that the a~(u) depend differentiably upon u. The ai(u) m a y not be normalized ; in order to put them into normalized form, it may be necessary to modify F,,; it is easily seen that this will not affect the differentiability of F,, and therefore of the a,. Differentiation of the relation cr = F~ 1 o a o F, gives, for u = 0: = F a(az)
ko"
dz
dF,, for fixed u, differs from dz + I~ d5 only by a scalar factor, we have l~ = (c~F,/c~)/(c?Fu/~?z);differentiating this with respect to u for u = 0 gives v =
But, since
ti = (?P/c~5. Thus P is a solution of v = c~q~/(?~; therefore, by (1), we can take ~ = (dz/d~z)& This gives:
~,~ = (Tz + ~)(~z + [7) - (~z +/~)(fiz + 3).
[1958b ]
387 On the moduli of Riemann surfaces
N o w apply (3), and calculate ~ in the usual manner by combining together the integrals along corresponding sides of the fundamental polygon. We get:
,1~i
i=1
This is a bilinear expression in the di, fii, 9i, 3i on the one hand, and, on the other hand, in the periods of the vector-valued differential
dz qz dz q dz
qz 2
f'2 =
which has the property that f ~ = M~f~ for every a e F, M , being the constant matrix ~2
2~fl
fi2
72
276
32
In particular, if the d~i, fli, 9i, 75i are all 0, ((o, D) is 0 for all ~0, and therefore D = 0. All this applies equally well if the small variation depends upon real parameters in any number. In particular, take vl . . . . , v3o 3 such that the D(vi) are a basis of the Kodaira-Spencer space, and then take # = ~ i wivi, with complex wl . . . . , w30 3- The preceding calculations show that the ~ , fl~, 7~, 6~ will be dift'erentiable functions of the real and imaginary parts of the w~, with a non-vanishing jacobian. This proves our assertion that the image | of O in N6.o-6 is an open set, and that the real vector-space underlying the Kodaira-Spencer space may be identified with the tangent vector-space to ~60 6 at the point corresponding to =So. Loosely speaking, the "realization" | of | has the right differentiable structure. Our remarks even show (always on the basis of the ellipticians' assurances) that it has the right real-analytic structure. It is clear, too, that, if we select a basis col . . . . . , o39-3 for the quadratic differentials on So, the (COl, D) define 39 - 3 complex covectors which determine an almost complex structure on O, or on 01. It remains to be seen that this is a complex structure. This can be done in two ways. The simpler method has just been suggested to me by L. Bers, who observed that the 3g - 3 complex parameters w~ introduced above can in fact be used as complex local coordinates. As we have already shown that their real and imaginary parts can be used as real local coordinates, it only remains for us to show that, for every small value ofw = (wl . . . . . w3g 3), the complex vectors D~ in the Kodaira-Spencer space for Sw, defined by the differential operators O/~?w> are all 0. More generally, we shall show that this is the case whenever # is a real-differentiable function of z and of complex parameters wi in any n u m b e r and is holomorphic in all the w~. It is obviously enough to consider the case of one complex parameter w. Assume therefore that # = #(z, w) depends real-differentiably upon z and w and that it is holomorphic in w. For a given value w* of w, consider the vector D in the Kodaira-Spencer space for Sw,, defined by (~?/c?~)w-w*. Notations being the
388
[]958b]
On the moduli of Riemann surfaces same as before, the conformal mapping of the universal covering of Sw. onto gl is given by z* = Fw.(z), where Fw. is a differentiable h o m e o m o r p h i s m of II onto itself, such that dz ~ = dFw. is of the form f(z) 9(dz + tl(z, w*)d~), with a scalar factor f ( z ) . In terms of the variable z*, the conformal structure of Sw for any w may be defined by a differential form ~* = dz* + #*(z*, w)d~*
which has to differ from dz + I~(z, w)d~ only by a scalar factor. Then, for each quadratic differential co* = q*(dz*) 2 on Sw., (co*, D) is given by:
aJ
\~/w=w.
But a trivial calculation shows that /t* is again holomorphic in w; therefore ~?/~*/c?~ is identically 0, and D is 0. The second method depends upon the consideration of the integrals of the first kind. We again call in the elliptical engineer, who tells us the following: since the conformal structure of S. can obviously be derived from a d s 2 which depends differentiably upon the parameters u (viz. d s 2 = y 2 Idz + ~ d z I 2 with y = ..~(z), which is invariant under F), the space of real-valued harmonic differentials on S. has a basis consisting of differentials which depend differentiably upon the parameters. Take again the case of a single real parameter u, and denote (O/c?u). = 0 by a dot, as before. The differentials of the first kind on S. are those linear combinations of the harmonic differentials with constant complex coefficients which differ from dz + # d2 only by a scalar factor; clearly the space of such differentials is generated by those which depend differentiably upon u. Let q. = h.(dz + # dz) be one of these; it must satisfy dr/. = 0 and is invariant under F; for u = 0, it reduces to a differential of the first kind r/o = ho dz for So. As before, put v = / i and call (0 a solution of v = (?~0/t?2; differentiating r/. with respect to u for u = 0, we get /l = h " dz + h o v d~ = r . dz + d(ho ~o),
where we have put r = h - 0(h0 ~o)/c~z. This, too, must be invariant under F and must satisfy dO = O. Call pi(u) the periods of r/. along the cycles crl for 1 _< i _< 29; put p~ = p~(0); /~i is then the period of 0 along a~, Let r/~ be any differential of the first kind for So, with the periods p'~. We may write r/; = dJ; w h e r e f i s a holomorphic function in gl. Then r/or/; is a quadratic differential for So; and, if we put D = D(v) as before, we have:
r
: ff ho.
ff o
- f
Combining integrals along corresponding sides of the fundamental polygon, we get : 9
(r/or/'o, D) = ~ (P'o+iPi -- P'il}o+i). i=1
In particular, call r/J, for 1 < j < g, the normalized differentials of the first kind for =S.; by definition, they are such that I p / ( u ) l , for 1 _< j < g, 1 < i < g, is the unit-
[1958b]
389
On the moduli of Riemann surfaces matrix; and then, by definition, we have:
z(__x.) = IIp~+i(u)ll
(1 _< j _< g;1 _< i < g).
This gives II~ill = 0, II~+gll = Z. Substituting r/~, r/~ for r/0, rl• in the above formula, we get:
2 = -II(rt~rl~, D)ll. This proves that S --. Z ( S ) is a holomorphic mapping of O into 6. More precisely, it shows, not only that the coefficients po~+kof Z(=S) are holomorphic on O, but also that any 3g - 3 of them will be local coordinates in the neighborhood of a given point of _So if and only if the corresponding quadratic differentials ~/~tl~ are linearly independent. This is Rauch's theorem. It implies that Z(_S0) is a simple point of the image Y~I of O by Z if and only if the products ~/~r/~ generate the space of quadratic differentials, i.e. if and only if S o is not hyperelliptic. Onc can prove quite similarly that the subset of 521 consisting of the points which correspond to hyperelliptic Riemann surfaces is a non-singular analytic subvariety of ~ of complex dimension 29 - 1. It is clear now that the almost complex structure defined on O is a complex structure in the neighborhood of all the points which do not correspond to a hyperelliptic Riemann surface; by using well-known general theorems on analytic varieties, one can then extend this result even to the points which correspond to hyperelliptic surfaces. Finally, in order to define a " n a t u r a l " Hermitian metric on O, it is only necessary to define a Hermitian structure on the Kodaira-Spencer space for each So, or, what amounts to the same, on the dual of that space, i.e. on the space of quadratic differentials for So. This is done by putting, for any two quadratic differentials co = q d z 2, o)' = q' dz 2 for So: ((~), u/) = [ [ - q~fy2 dz d5 o ,/ with y = J ( z ) . In fact, the integrand can be formally written as e)~)/dS, where dS = y 2 dz dz is the hyperbolic area-element in H; it is therefore invariant under F, and the integral has an intrinsic meaning. This raises the most interesting problems of the whole theory: is this a Kfihler metric? has it an everywhere negative curvature? is the space O, provided with its complex structure and with this metric, a homogeneous space? It would seem premature even to hazard any guess about the answers to these questions.
[1958c] Final report on contract AF 18(603)-57 Within the framework of the project originally submitted to AFOSR, I eventually decided to concentrate on two lines of investigation: (I) The classical problem of the moduli of algebraic curves over complex numbers; (II) A study of the K;,ihler varieties topologically identical with the nonsingular quartics in projective 3-space (henceforward called K3 surfaces). In both directions, my results are still very fragmentary and incomplete; and I have had to postpone the arithmetical considerations which provided the original motivation for the whole project, in order to deal first with the function-theoretic aspects of the above questions. In both problems, the ideas of Kodaira and Spencer on the variation of complex structures have proved fundamental. I have much benefited from repeated consultation with them during my stays in Princeton, in January and February and again in June. It also turned out that Professor L. Bers had been engaged in a parallel investigation of problem (I); consultation with him on this topic has been very fruitful. On the other hand, various aspects of problem (II) have recently engaged the attention of Professor L. Nirenberg, Professor A. Andreotti, and Dr. Atiyah; I have learned a great deal from communications, written and oral, from all of them. In order to give, in what follows, a coherent account of these topics, it will be necessary to include much of the work of my colleagues, and it would be unpractical to try to unravel in detail what may belong to me and what belongs to each one of them. It should be understood that they deserve a large share of the credit for the work described in this report.
I. Moduli o f algebraic curves We consider curves of a given genus 9 -> 1. One basic concept is that of a Teichmtiller structure. If S, S' are two oriented surfaces of genus 9, we say that a class of mappings of S into S' in the sense of homotopy (or, briefly, a class C(S, S')) is admissible if it contains at least one orientation-preserving homeomorphism of S onto S'. Let So be an oriented surface of genus g, given once for all. By a Teichmfiller surface, we understand a Riemann surface of genus 9 (i.e. a surface of genus g, provided with a complex-analytic structure and oriented accordingly), together with an admissible class C(So, S). Isomorphism being defined for these in the obvious manner, we introduce, with Teichmfiller, a space T (the "Teichmfiller space"), whose points correspond in one-to-one manner to all the classes of mutually isomorphic Teichmfiller surfaces of genus 9. Teichmtiller's chief contribution was
390
[1958c]
391 Final report on contract AF 18(603)-57
to define on T a certain topology, the " n a t u r a l " one in a sense described below, and then to prove that T, with this topology, is h o m e o m o r p h i c to an open cell of real dimension 69 - 6. So far, I have mainly been concerned with the local propperties of the Teichm/iller space and of its " n a t u r a l " complex-analytic structure. The definition of the latter depends upon ideas introduced by Teichmiiller himself, but which do not appear to have been fully understood until K o d a i r a and Spencer attacked similar problems for higher dimensions. In order to describe them, it is convenient to substitute for the above definition o f a Teichmiiller surface the following one. Let A1 . . . . , A20 be a set of generators, fixed once for all, of the fundamental group G o of So (with a given origin), satisfying the relation
R(A 1. . . . .
A2o ) = A1AzAllA2
1 ...
A2o~ = 1.
A set at . . . . . a20 of linear-fractional substitutions on z will be called admissible if it has the following properties: (a) it generates a discrete group G of hyperbolic substitutions acting on P; (b) S = PIG is a compact surface of genus 9; (c) there is an admissible class C(S o, S), mapping G o onto G considered as the fundamental group of S, which maps A i onto ai for 1 _< i _< 2 9. Each Teichmfiller surface S has a universal covering which can be mapped conformally onto P, and can therefore be represented as PIG ; the class C(So, S) which belongs to it defines an isomorphism of G o onto G; therefore, to each Teichmtiller surface, there corresponds at least one admissible set (o~ . . . . . 0029). Conversely, it is easy to see that two such sets will define isomorphic Teichmiiller surfaces if and only if they can be transformed into one another by a linear-fractional substitution. Using this fact, it is possible to normalize admissible sets in such a way that there is a one-to-one correspondence between classes of Teichmtiller surfaces (i.e. points of the space T) and normalized admissible sets (00i); this gives a one-to-one mapping of T onto a subset of the coordinate space R 6~ 6. It turns out that the latter subset is open, and that the mapping and its inverse are both indefinitely differentiable (and even, presumably, real-analytic) if T is provided with its " n a t u r a l " differentiable structure; the proof of the latter fact is due to L. Bers. N o w introduce a "variation of structure" as follows. Let/~ be any indefinitely differentiable complex-valued function in P such that I#l < 1; let Pu denote the upper half-plane with the modified complex structure for which dz + # dz is a differential form of type (1, 0); if the latter structure is invariant under G, i.e. if l~ dS/dz is formally invariant under G in an obvious sense, the complex structure of P~ can be projected onto a complex structure P~/G, making the latter into a Riemann surface S u, or rather a Teichmtiller surface if we keep the ai as the distinguished generators of G. Let F , , in that case, be the conformal mapping of P onto Pu; the Teichmfiller surface S~ is then the one defined by the admissible set 00i(/~) = Fj~-100IF~; S~ is isomorphic to S if and only if F , (which is defined only up to a fractional-linear substitution) can be chosen so that it commutes with all the 00~; in that case, the "variation of structure" defined by # will be called trivial. F r o m this, we get an "infinitesimal variation" if we take/~ to depend (differentiably) upon a real parameter t, so that/~ = 0 for t = 0; then v = (dlz/dt)t-o is called an infinitesimal variation of structure; it is clear that any function v in P,
392
[1958c]
Final report on contract AF 18(603)-57 such that v dUdz is formally invariant under G, defines such a variation. The variation v will be called trivial if the finite variation # is tangent, for t = 0, to a trivial one, i.e. if there is a trivial variation y , depending upon t, such that d#/dt = d#'/dt for t = 0. It is easily seen that a necessary and sufficient condition for this is that there should exist a function { in P such that v = c3{/c3~and that {/dz is formally invariant. The infinitesimal variations v can be considered as the elements of a vectorspace V (of infinite dimension) over the complex numbers; the trivial variations make up a subspace V' of V. Standard procedures in c o h o m o l o g y theory and the theory of fibre-bundles (which, in a case like this one, depend merely upon elementary facts such as Stokes's theorem and the theorem of Riemann-Roch) show that V/V' is of finite dimension 3g - 3 and can be "canonically" identified with the dual space to the space of quadratic differentials of the first kind on S. As to the latter assertion, let co = q dz 2 be such a quadratic differential; in other words, we take for q a holomorphic function in P, such that q dz 2 is formally invariant under G. Then, for v in V, vq dz d5 is formally invariant under G, so that we may integrate it over P/G; Stokes's theorem shows that the integral is 0 for v in V', i.e. it depends only upon the class D of v modulo V'; denoting it by (co, D), one finds that this bilinear form defines a duality between V/V' and the space of quadratic differentials, as asserted above. At this point, one must make use of the fact (proved by Bers) that, if # depends differentiably upon some real parameters, the same will be true of the mapping function Fu and hence also of the coefficients in the substitutions ai(#)- it is now easy to calculate the effect of an infinitesimal variation on the coefficients of the cri, i.e. to calculate dai(#)/dt for t = 0, in terms of v = (d#/dt),= o. One can also calculate the effect of a given infinitesimal variation on the periods of the normalized integrals of the first kind on S. The conclusions one can derive from this are as follows. It is possible to provide T with a complex-analytic structure, of complex dimension 39 - 3, such that, whenever # depends holomorphically upon some complex parameters w~, the point of T which corresponds to S u depends holomorphically upon the w~; this observation is due to Bers, who also found, more precisely, that, if we take # = ~ Wiy2qi, with y = Im(z) and (ql . . . . , q3o 3) such that qi dz2 are a basis of the space of quadratic differentials of the first kind on S, then the wi can be taken as local complex coordinates in T in a neighborhood of the point corresponding to S. Furthermore, the quadratic differentials of the first kind on S can be identified with the covectors on T at that point ; and Petersson's hermitian metric, in the space of those differentials (which is no other than the space of automorphic forms of degree - 4 for the group G) defines an intrinsic Hermitian metric on T, which turns out to be a K~ihler metric. The facts concerning the mapping of T into R 6g 6 by means of the coefficients of the ~ri have already been stated. Finally, the periods of the normalized integrals of the first kind on S define a mapping of T into the Siegel space of symmetric g x g matrices with positive-definite imaginary part; the image of T under that mapping is a complexanalytic variety W, whose singular points are those corresponding to hyperelliptic Riemann surfaces; and one obtains a new proof for Rauch's theorem, stating which of the periods of the normalized integrals of the first kind can be used as local coordinates in the neighborhood of any given simple point of W.
[1958c]
393 Final report on contract AF 18(603)-57
H. The K3 surfaces We may start here from the observation (made independently, I believe, by Atiyah and myself) that, when a non-singular surface S in projective 3-space acquires a node, i.e. a conical double point, and the latter is desingularized by a standard dilatation, this process gives a surface S' which is homeomorphic to S. It was easy to surmise that the same is true when a surface acquires any number of distinct nodes; this, in fact, or rather a much more precise theorem, was proved by Atiyah. It shows, in particular, that the non-singular quartic in 3-space, the double plane with a non-singular sextic branch curve, and the desingularized K u m m e r surface, are all homeomorphic. Such surfaces will be called K3 ; they had already occurred in the work of the Italian geometers, and, more recently, in that of Kodaira. The Italians, in fact, had discovered an infinite sequence of families F, (n = 1, 2 . . . . ) of regular surfaces with P0 = 1 ; F 1 consists of double planes with a sextic branch curve; F2, of quartics in 3-space; F, consists of surfaces of degree 2n in projective (n + D-space, whose hyperplane sections are canonical curves of genus n + 1. There are very plausible arguments to indicate that all such surfaces are of type K3, although no complete proof for this seems to have been given yet. For K3 surfaces, the intersection matrix of two-dimensional cycles has the signature (19, 3), the determinant - 1, and is even (i.e. the self-intersection of every cycle is even); hence there is exactly one double differential of the first kind ; if we call it tl, then we must have dq = O, qz = 0 and t/f/>_ 0; if we assume (as seems very likely) that all K3 varieties (algebraic or not) constitute only one connected family, then the canonical class must be 0, so that ~/ 4= 0 and ~/f/> 0 everywhere; this implies that the complex structure is entirely determined by q. Conversely, let there be given, on a differentiable manifold of that nature, a complex-valued differential form t/ of degree 2, satisfying drl = 0, q2 = 0, and ~/f/> 0 everywhere; this determines a complex structure. It seems very plausible (but not at all easy to prove) that two such forms with the same periods must determine complex structures which can be transformed into one another by a difl'erentiable homeomorphism, homotopic to the identity; that all such structures are K~hlerian; and that the periods of rl do not have to satisfy any other condition than those which are implicit in the relations f]2 = 0, /7/7/ • 0. These conjectures (which have also been made independently by Andreotti) may also be expressed as follows. Let S be a class of such structures, two structures being put into the same class if and only if they can be transformed into each other by a differentiable homeomorphism, homotopic to the identity. Let (al . . . . . a22 ) be a minimal set of generators for the two-dimensional homology group with integral coefficients; iD/is as described above, let the Pl be its periods corresponding to the cycles ai; let P be the point with the homogeneous coordinates (Pa . . . . . P 2 2 ) in the complex projective space of dimension 21. Let F(x, y) be the symmetric bilinear forms in x = (xl ..... ")r Y = (Yl . . . . . Y22) whose matrix is the intersection-matrix of the cycles ag. Then the conditions q2 = 0, r/F/ > 0 imply that P is in the open subset H of the quadric F(x, x) = 0 which is determined by the inequality F(x, ~) > 0; H is a homogeneous space of complex dimension 20 for the orthogonal group determined by the real form F(x, x). Now, if we assign to each
394
[1958c] Final report on contract AF 18(603)-57
class S of structures of the given type the point P ~ H, we have a mapping of the set of all such classes into H; and Nirenberg, by an argument combining the Kodaira Spencer technique with techniques derived from the theory of elliptic equations, has proved that the image of that set in H must be open. The conjectures stated above would mean that the mapping is a one-to-one mapping of that set onto H. Furthermore, by a fundamental theorem of Kodaira, a given K3 structure will define an algebraic variety if and only if it has a K~hler metric whose fundamental form has integral periods, i.e. belongs to an integral cycle a; we must have F(a, a) < 0. For such a structure, the point P defined above must belong to the linear variety L~ defined by F(a, x) = 0, and therefore to the set Ha = H c~ L,. It is easy to see that for each integral a such that F(a, a) < O, Ha can be identified with the Riemannian symmetric space belonging to the orthogonal group of the quadratic form F o of signature (19, 2) induced by F on La. Again, one is led to conjecture that this establishes a one-to-one correspondence between such structures and Ha. Now it may happen that two classes S, S' may be distinct and still define isomorphic structures; this will be so when structures belonging to these classes can be transformed into one another by a differentiable homeomorphism, not homotopic to the identity. The latter will induce an automorphism of the homology group, and therefore a unit U of the quadratic form F(x, x), i.e. a matrix with integral coefficients belonging to the orthogonal group of F; let G be the group of all the units of F which can be obtained in this manner. Assuming the truth of the conjectures stated above, we see that two points P, P' of H will determine isomorphic structures if and only if they are equivalent under the group G. A similar statement will hold for Ha; here G has to be replaced by the subgroup Go of G consisting of the elements which leave a invariant. This shows that it is important to determine G, and in particular to determine whether G coincides with the group G of all the units of F. My results on this, too, are still incomplete. However, by applying the theory of theta functions to the Kummer surface considered as a model for K3 surfaces, it has been possible to reduce part of the problem to a purely arithmetical question which has been recently solved by M. Kneser. The result is that G is, at any rate, of finite index in G. Now, since H is not the Riemannian symmetric space for the orthogonal group of F, it follows that G does not act upon H in a properly discontinuous manner; hence there can be no theory of the moduli, in the ordinary sense, for the postulated connected family of K3 surfaces. This is as expected, and is analogous to the wellknown fact that there is no theory of moduli for complex toruses, but only for "polarized" abelian varieties. The situation is quite similar here. For, if one restricts oneself to the family of algebraic K3 surfaces polarized by assigning a K~ihler form belonging to a given integral cycle a, then it follows from what we have said that Go acts in a properly discontinuous manner on Ho, and is of finite index in the group of all units of the quadratic form Fo, so that HjGo is of finite measure. It is therefore to be expected that the automorphic functions in Ha, for the group G o, make up an algebraic function-field, the field of the moduli for the K3 surfaces of the given family. One interesting feature here is the occurrence, in a problem of moduli, of the automorphic functions belonging to the group of units of a quadratic form of signature (n, 2) (with n = 19 in the present case). This is believed to be
[1958 c ]
3 95
Final report on contract AF 18(603)-57 the first time that such a group has appeared in such context. Of course, before this can be more thoroughly investigated, it will be necessary to obtain full proofs for the conjectures stated above. After that is done, analogies with the theory of abelian varieties and of their fields of moduli (given by Siegel's modular functions) will undoubtedly suggest a number of further problems, of a function-theoretic and also of a number-theoretic nature; most fascinating, perhaps, are the possibilities suggested by the theory of complex multiplication. But this is still too remote to be discussed here.
[1958d] Discontinuous subgroups of classical groups (Notes by A. Wallace)
Introduction The object of this course is to prepare the way for a study of certain types of discrete subgroups of the real classical groups and the corresponding quotient spaces. The classical groups will be constructed in a rather special way which actually yields all these groups with only a small number of exceptions. The method consists in taking a semi-simple algebra A over the rationals with an involution or, extending A to an algebra A R over the real numbers, and considering the connected component G of the group of automorphisms of A R which commute with or. G is, in a natural way, an algebraic matric group, and a subgroup Gz of matrices in G whose elements are rational integers is a discrete subgroup. Discrete subgroups obtained in this way are to form the main object of study. An illustration of the kind of theorem to be studied is given in w where conditions for the compacity of G/Gz are worked out. The method of study of G/Gz involves introducing a second involution on AR which is positive (definitions in w and studying the set P(A~) of positive symmetric elements of A~ with respect to this involution. It turns out that G/Gz can be related to a subset of these positive symmetric elements, and that a special type of set W in P(AR) (an M-domain) covers the image of G/Gz, in a certain sense, only a finite number of times. Attention can then be transferred to the set W, which is arithmetically defined. The study of the set W depends on a study of P(A~) which generalizes classical results of Minkowski on the theory of positive definite quadratic forms and their equivalence under transformation by integral matrices. w167 of these notes are concerned with this theory. The next two sections give a list of the classical groups which can be obtained as indicated above. In w some results are obtained on algebras with involutions, and in w these are applied, along with the earlier results, to the construction and study of M-domains. It may appear that the results obtained in this way will depend on the particular way in which the group G is written as a matric group, since the definition of Gz certainly depends on this. However, the properties which are to be of interest eventually are only those which are invariant under commensurability; this can be defined as follows: Two discontinuous subgroups F and F' of a group G are said to be commensurable if F r~ F' is of finite index in each of them. Now if the group G is an algebraic matric group overQ, and is represented as a matric group in two different ways, then the two subgroups F' and F" of integral matrices in these two representations are commensurable. To prove this let G', G"
396
[1958d]
397 Discontinuous subgroups of classical groups
be the two representations of G as matric groups and write x' = 1, + (x'ij) for an element of G', x" = 1,, + (x~,) for an element of G". The isomorphism between G' and G" is expressed by equations xij = Pi~(x~,,), x~, = Q,u(xlj) where the Pij and Q , , are rational functions and in fact can be taken to be polynomials over Q with zero constant terms. If N is a c o m m o n denominator for all the coefficients in the P;i and Qa~,, then, for x;{, -= 0 (mod N), the corresponding x'ii will be integral. Thus F' contains F~, the subgroup of F" consisting of matrices -= 1,, (mod N). Similarly F" D F;e. N o w F~ is of finite index in F", and so the larger group F' c~ F" is of finite index in F"; and similarly in F'. Thus F', F" are commensurable. The result shows that, as far as properties invariant under commensurability are concerned, no generality is lost by the special method used here of constructing Gz. In particular it is easy to see that the compacity of G/Gz discussed in w is such a property.
[1959a] Ad61es et groupes alg6briques
On d6signera toujours par k, soit un corps de nombres algdbriques, soit un corps de fonctions alg6briques de dimension 1 sur un corps de constantes fini. On d6signera par k~ le compl6t6 de k par rapport 5. une valuation v; si une valuation est discr6te, on la ddsignera le plus souvent par un symbole tel que p, et on ddsignera par rp l'anneau des entiers p-adiques dans kp. O n ddsignera par S tout ensemble fini de valuations de k, contenant l'ensemble So des valuations non discr6tes (pour lesquelles le compldt6 est R ou C); bien entendu So est vide si k est un corps de fonctions. Si V e s t une varidt6 alg6brique, ddfinie sur un corps k, on identifiera, suivant l'usage, V avec l'ensemble des points de V sur le domaine universel, et on notera Vk l'ensemble des points de V rationnels sur k. Cette convention s'appliquera notamment si V e s t une varidt6 de groupe, une varidt6 d'alg6bre, une vari6t6 de corps, etc.; on dira par exemple que V e s t une varidt6 d'alg6bre de dimension n, ddfinie sur k, si, en tant que varidt6, c'est un espace a n n e de dimension n, sur lequel on s'est donn6, outre la structure additive usuelle, une structure multiplicative, c'est-'~-dire une application bilindaire (toujours suppos6e associative) de V x V dans V, ddfinie sur k; Vk est alors une alg6bre sur k au sens usuel; si Vk est un corps (donc, au sens usuel, une extension de k de degr6 n), on dit que Vest une vari6t6 de corps de dimension n sur k. Cette mani6re de parler est conforme fi l'usage ancien de Kronecker, Hilbert, etc. (qui ne se g~naient pas pour parler de l'616ment g6n6rique d'un corps). 1. Soit V une varidt6 ddfinie sur k; Vk, peut ~tre munie, d'une mani6re 6vidente, d'une topologie qui la rend localement compacte, et compacte si Vest compl6te (on commence par le cas oh V e s t une vari6t6 a n n e , Vkv 6tant alors considdrde comme partie fermde d'un espace vectoriel de dimension finie sur k~; on passe de lfi d'une mani6re 6vidente au cas d'une vari6t6 abstraite quelconque). Soit p une valuation discr6te de k. Si Vest une varidt6 a n n e , on notera V~ l'ensemble des points de Vkp dont les coordonndes sont dans r~; il est immddiat que c'est une pattie compacte de Vkp. Plus gdndralement, soit V une varidt6 abstraite, d6finie sur k; on sait que V admet toujours un recouvrement fini par des ouverts isomorphes/t des vari6tds affines V (i), ddfinies aussi sur k; autrement dit, on peut 6crire V = Ui Ji(V(~ oh less sont des isomorphismes (ddfinis sur k) des vari6t6s affines V (i) sur des ouverts de V (au sens de la topologie de Zariski, bien entendu). Pour toute valuation discr6te p de k, posons: = U i
c'est 15. une pattie compacte de Vk~. Posons maintenant, pour tout ensemble fini S
398
[1959a]
399 Adeles et groupes algebriques
de valuations de k contenant l'ensemble S O des valuations non discretes:
Vs = I1 vk, X [I veS
pr
off le second produit est 6tendu fi toutes les valuations de k (necessairement discretes) qui n'appartiennent pas fi S; c'est l• une pattie localement eompaete de l ~ Vk,~,et on a Vs = Vs, pour S ~ S'. On d~signera par VAk la limite inductive des Vs, c'est-~-dire la reunion des Vs munie de la topologie pour laquelle un systeme fondamental d'ouverts est form6 par la reunion de l'ensemble des ouverts dans t o u s l e s Vs. Cette notion est justifOe par le Jait que VA~ est dbfinie d'une manikre intrinsbque, c'est-'~-dire ne depend pas de la maniere dont on a 6crit V comme reunion finie d'images isomorphes de varietes affines; la verification de cette assertion est 616mentaire. On appelera VA~ l'espace adOlique associ6 fi V. Au lieu de A , on 6crira souvent A quand aucune confusion n'est possible. L'espace adelique associ6 fila droite affine n'est autre que l'ensemble des adeles (dits aussi "repartitions" ou "valuation-vectors") du corps k, avee sa topologie usuelle: cet ensemble (muni de sa structure topologique, et de sa structure d'anneau) sere note Ak, OU simplement A. La notion d'espace adelique poss6de des proprietes fonctorielles raisonnables. S i f e s t une application partout definie d'une variet6 V darts une variet6 W, definie sur k, on en deduit d'une maniere evidente une application continue de VA dens WA, Si par exemple on s'est donne sur V une loi de groupe algebrique, on en deduit une loi de groupe sur VA, et VA s'appellera le groupe adelique associe fi V. Par exemple, si Vest le groupe multiplicatifGm ?2 une variable sur k, le groupe adelique correspondant est le groupe des idbles de k au sens usuel. Si l'application f de V dens W e s t surjective, il n'en est pas de meme, en general, de l'application de VA dans WA qui lui est associee. Cependant, si Vest un fibre de base W, localement trivial sur k, au sens de la geometrie algebrique, et s i f e s t la projection de V sur sa base W, alors l'applieation correspondante de VA darts WA est surjective. En particulier, si G est un groupe et g u n sous-groupe de G, tous deux definis sur k, et q u ' o n pose H = G/g, l'application de GA darts HA, associee 5. l'application canonique de G sur H, n'est pas surjective en general; mais, si G est (sur le corps de base k) fibrO localement trivial sur H = G/g , alors on peut identifier canoniquement H A avec G A/gA . 2. Soit K une extension separable de k, de degr6 fini d; soit V une varidte affine ou projective de dimension n, definie sur K; on va indiquer comment on peut associer /tces donnees une varidte W de dimension nd, definie sur k, et qui, sur le domaine universel, est isomorphe au produit de V e t de ses conjuguees sur k. P o u r fixer les iddes, considerons le cas affine. Alors Vest definie par des equations: P,(X~ . . . . , XN) = 0
(1 _< /~ < m),
5. coefficients dans K. Soit (a~ . . . . . ad) une base de K sur k; posons: d
x~ = y~ a~ Yj~ j
1
(1 _< i _< N)
400
[1959a] Ad61es et groupes alg6briques
les Yji 6tant de nouvelles ind6termin6es; alors les P~, deviennent des polyn6mes fi coefficients dans K par rapport aux Yji, et peuvent donc s'6crire: d
Pu(X1 . . . . . X u ) = ~ ah Qnu(Yll . . . . . Ydu), h=l
off les Qhu sont des polyn6mes ~ coefficients dans k. Dans ces conditions, W est la vari6t6 d6finie par les 6quations Qhu(Yll . . . . . YdN) = 0
(1 < h _< d, 1 _< # < m)
dans l'espace affine de dimension dN. En effet, une v6rification facile montre qu'apr6s un changement de coordonn6es linOaire, ~ coefficients dans le corps compos6 de K et de tous ses conjugu6s sur k, ces 6quations sont pr6cis6ment celles qui d6finissent le produit de Vet de ses conjugu6es sur k. C'est d'ailleurs seulement cette derni6re v6rification qui exige que K soit s6parable sur k; s'il ne l'6tait pas, les d6finitions ci-dessus auraient encore un sens, mais d6finiraient une op6ration ayant des propri6t6s assez diff6rentes de celles qui nous int6ressent ici. On dira, dans les circonstances ci-dessus, que W e s t d6duite de V par restriction du corps de base de K h k, et on 6crira W = ~ ( V ) . I1 est imm6diat qu'il y a correspondance biunivoque (canonique !) entre VK et Wk. O n v6rifie aussi, sans difficult6, qu'il y a correspondance biunivoque (non moins canonique) entre VA,, et WAk. Cela permet, par exemple, dans t o u s l e s probl6mes relatifs aux espaces ad61iques sur les corps de nombres, de se ramener, si l'on veut, au cas off le corps de base est Q. 3. Soit G une vari6t6 de groupe de dimension n, d6finie sur k. I1 existe sur G une forme diff6rentielle ~o de degr6 n, invariante ~ gauche, d6finie sur k, et ~o est unique un facteur pr6s (facteur qui doit 6tre dans k); si xl . . . . . x, sont des fonctions sur G, d6finies sur k, qui soient des coordonn6es locales dans un voisinage de l'616ment neutre e (et, pour fixer les id6es, nulles en e), on pourra 6crire o~ = y dxl 9 .. dx,, off y est une fonction sur G, d6finie sur k, et finie en e. O n va montrer comment, si vest une valuation quelconque de k, on peut associer ~ Cn une mesure de Haar bien d6termin6e sur Gk,,, mesure qu'on notera ]~o I,,. P o u r cela on distinguera trois cas: (a) k~ = R; alors les xi peuvent servir de coordonn6es locales sur Gk,,au voisinage de e; au voisinage de e, y est fonction analytique r6elle de x ~. . . . . x,; dans ces conditions, y d x ~ . . , dx. peut s'interpr6ter c o m m e une forme diff6rentielle au sens usuel sur Gko au voisinage de e; c o m m e il est clair qu'elle est invariante ~ gauche, elle d6finit par translation une mesure de Haar sur Gk,,, q u ' o n note [~o [,. (b) k~ = C: les xi sont alors des coordonn6es locales complexes sur Gk~ au voisinage de e, et y est fonction analytique complexe des x~ au voisinage de e; on prendra [~o[~ = i"y~ dxl dx, ... dx, dx,. '(c) Soit p une valuation discr6te; Gkp est une vari6t6 analytique sur le corps valu6 complet kp; c o m m e pr6c6demment, x~ . . . . , x, peuvent fitre consid6r6es c o m m e coordonn6es locales au voisinage de e sur Gkp, c'est-~-dire qu'elles d6terminent un h o m 6 o m o r p h i s m e d'un voisinage de e dans Gkp sur un voisinage de 0
[1959a]
401 Ad61es et groupes alg6briques
dans l'espace k~,. Convenons, sur le groupe additif kp, de noter Idx Ip la mesure de Haar, norm6e par la condition que rp (anneau des entiers p-adiques) soit de mesure n 1; la mesure de Haar dans kp, produit des mesures Idxlp sur les n facteurs, sera notde Ddxl ... dx, Ip- On posera alors, au voisinage de e: I~o1~ = lyl~[dx~ ... dx, l~, off lYl~ est la valeur absolue p-adique de la valeur de y au point considdr6 (y est, c o m m e pr~c6demment, fonction analytique de xl . . . . , x, au voisinage de e, ce qui veut dire qu'on peut l'6crire comme s6rie convergente de puissances de x 1. . . . . x,). Naturellement, la valeur absolue p-adique est norm6e de la mani6re usuelle, c'est-fi-dire de fagon que l'automorphisme x ---, ax du groupe additif de k, multiplie la mesure de Haar par le facteur l a Iv. Les d6finitions ci-dessus se justifient du fait qu'elles sont ind6pendantes du choix des coordonn6es locales xl . . . . . x, au voisinage de e; c'est 6vident darts les cas (a) et (b), en vertu de la formule de changement de variables dans les int6grales multiples en analyse classique, et cela rdsulte immddiatement, dans le cas (c), de la formule correspondante en analyse p-adique (qui se d6montre encore plus facilement qu'en analyse classique). De ce qui pr6c6de, on peut, dans certain cas, d6duire la d6finition d'une mesure de Haar sur le groups addlique GA attach6 ~t G. En se reportant au paragraphe 1, il est 6vident que cela est possible chaque fois que, sur le produit
O < X ~0r %
vsS
il existe une mesure produit des mesures [colv; car il est 6vident que les mesures ainsi ddfinies sur Gs et sur Gs,, pour S ~ S', coincident sur Gs. Or, pour que la mesure en question soit ddfinie, il faut et il suffit que le produit infini
6tendu g toutes les valuations discr6tes de k, soit absolument convergent. Lorsqu'il en est ainsi, on notera ~],. Ice Iv la mesure ainsi obtenue. On observera que celle-ci ne d@end pas du choix de co; en effet, si on remplace co par cco, avec c e k • elle se multiplie par 1Iv Icl~, qui est 1 ("formule du p r o d u i t " d'Artin; rappelons que celle-ci se ddmontre comme suit: x ~ cx est un automorphisme du groupe additif Ak, qui multiplie la mesure de Haar par [Iv Iclv, et qui d'autre part induit un automorphisme sur k consid6r6 c o m m e sous-groupe de Ak, donc, par passage au quotient, d6termine un automorphisme du groupe compact Ak/k et par suite laisse invariante toute mesure de H a a r sur ce dernier; k 6tant discret dans A k, A k et Ak/k sent localement isomorphes, donc les mesures de Haar y prennent le m~me facteur). O n dira que G a la propri~t~ de convergence si 1 ~ [co I~ Y est ddfini; le cas off le produit infini 6crit ci-dessus, sans ~tre absolument convergent, est convergent lorsque les p sent ordonnds "naturellement" (c'est-fi-dire par ordre de grandeur croissante des normes) est int6ressant aussi; lorsqu'il en est ainsi, on dira que G a la propri~tO de convergence relative. Les exemples assez vari6s qu'on a trait6s
402
[1959a] Ad61es et groupes alg6briques
jusqu'ici rendent assez plausibles les conjectures suivantes: pour que G a i t la propri6t6 de convergence relative, il faut et il suffit que G n'admette pas d ' h o m o morphisme sur G,,, ddfini sur k; pour que G a i t la propri6t6 de convergence, il faut et il suffit que G n'admette pas d ' h o m o m o r p h i s m e sur Gin, ddfini sur le domaine universel. En particulier, t o u s l e s groupes unipotents, et t o u s l e s groupes semi-simples qu'on a pu traiter de ce point de vue, ont la propri6t6 de convergence. Le groupe Gm n'a pas la propridt6 de convergence relative. Le groupe addlique attach6 au groupe additif G, fi une variable n'est autre que le groupe additif de Ak; ce groupe a 6videmment la propri6t6 de convergence. Nous poserons : k/k
(on convient, une lois pour toutes, d'identifier d'une mani&e 6vidente une mesure de Haar sur un groupe localement compact F avec celle qu'elle d6termine par passage au quotient sur l'espace homog6ne F/7, lorsque y est un sous-groupe discret quelconque de F). On voit facilement que #k = IAI ~/2, o5 A est le discriminant de k, lorsque k est un corps de nombres alg6briques, et que #k = qO 1 si k est un corps de fonctions de genre .q sur un corps de constantes "~ q 616ments. Soit alors G u n groupe de dimension n sur k, ayant la propridt6 de convergence; nous poserons:
nG=
H Io 1 , L'
et nous dirons que c'est la mesure de Tamayawa sur G A. I1 r6sulte de ce qui pr6c6de que cette mesure est d6termin6e d'une mani6re unique (elle ne d~pend pas du choix de co). Soit K une extension s6parable de k, de degr6 rink Soit G u n groupe d6fini sur K ; soit G' le groupe ~K.k(G), d6duit de G par restriction du corps de base de K k; on a d6j/t observ6 que GAK s'identifie canoniquement/t G~k; on v6rifie facilement que Get G' ont simultan6ment la propri6t6 de convergence, et que les mesures de T a m a g a w a ~2G (sur GA,~) et ~2~, (sur G~,~) co'fncident lorsqu'on identifie ces groupes. C'est, entre autres, pour qu'il en soit ainsi qu'on a introduit le facteur ps dans la d6finition de ~G. 4. Toujours avec les m6mes notations, Gk s'identifie d'une mani&e ~vidente avec un sous-groupe discret de GA; on dit que c'est le groupe des ad61es principaux de G. O n s'est aper~;u, dans ces derniers temps, qu'une bonne partie des rdsultats les plus importants de l'arithm6tique classique pouvait s'exprimer en 6 n o n , a n t des propri6t6s de GA/Gk pour des groupes alg6briques G convenables; cette observation, faite d ' a b o r d par Chevalley, c o m m e chacun sait, pour le cas particulier du groupe des id~les (groupe Gin), est d'une importance capitale; le m6rite semble en revenir principalement "~ O n o et Tamagawa. Le cas off G est commutatif est celui de la th6orie classique des corps de nombres alg6briques; celui off G est le groupe orthogonal est celui de la th6orie des formes quadratiques; il semble donc qu'on touche au m o m e n t off ces deux th6ories, confondues/t leurs d6buts (la th6orie des formes quadratiques binaires ne diff6rant pas de celle des corps quadratiques),
[1959a]
403 Ad&les et groupes alg6briques
vont enfin se fondre de nouveau en une seule, 5. savoir la thdorie arithm6tique des groupes alg6briques. I1 n'est pas difficile de ddmontrer (cf. O n o [1]) que GA/Gk est compact lorsque G est unipotent. I1 en est de marne, d'autre part, pour certains groupes semisimples; en voici deux cas typiques: a. G est le groupe des 616ments de norme 1 sur le centre darts une varidt6 de corps non commutatif; b. G est le groupe orthogonal d'une forme quadratique ne repr6sentant pas 0. Marne lorsque GA/Gk n'est pas compact, il peut arriver que cet espace soit de mesure finie (pour une mesure de Haar quelconque sur GA); it semble plausible qu'il en soit ainsi pour les m~mes groupes dont on a conjectur6 plus haut qu'ils ont la propri6t6 de convergence relative. En tout cas, au moyen de la r6duction des formes quadratiques, on a pu ddmontrer cette propri6t6 pour un grand nombre de groupes semi-simples, et Ramanathan a annonc6 qu'il poss6dait une d6monstration s'appliquant fi tous les groupes semi-simples "classiques". Pla~ons-nous dans le cas off G a la propri6t6 de convergence; nous poserons: r(G) = ~
flc,,
d GA/Gk
et nous appellerons z(G), lorsqu'il est fini, le nombre de Tamagawa de G. Le m6rite essentiel de T a m a g a w a consiste 5- avoir ddfini T(G), d'abord dans le cas particulier ot~ G est le groupe orthogonal d'une forme quadratique (sur le corps des rationnels, puis sur un corps de nombres alg6briques), et 5- avoir reconnu les fairs suivants: a. P o u r ce groupe (plus pr6cisdment, pour la composante connexe de l'616ment neutre darts ce groupe), on a r(G) = 2; b. La formule T(G) = 2 est enti6rement 6quivalente 5. l'ensemble des r6sultats des trois c616bres m6moires de Siegel sur les formes quadratiques (C. L. Siegel, Uber die analytische Theorie der quadratischen F o r m e n [2], soit 194 pages, qui ne contiennent m~me pas une d6monstration compl6te pour le cas g6n6ral des formes de signature quelconque sur un corps quelconque). En fait, une lois qu'on a eu l'idde d'exprimer les choses dans ce langage, l'6quivalence de la formule z(G) = 2 avec les r6sultats de Siegel n'est pas trop difficile 5. vdrifier. Naturellement, T a m a g a w a ne s'est pas arr6t6 l~t. I1 a d ' a b o r d cherch6 5. r6diger, darts ce m~me langage, la d6monstration m~me du th6or6me de Siegel (l'id6e de traduire celle-ci dans le langage des id61es 6tait ddjs- venue 5- M. Kneser il y a quelques ann6es, mais celui-ci n'avait rien publi6 sur ce sujet). Un avantage essentiel de la nouvelle mdthode consiste en ce que le thdor6me 5-d6montrer est, du fair marne de son 6nonc6, birationnellement invariant, alors que chez Siegel (et, avant lui, chez Minkowski) il apparaissait comme lid 5. un "genre" de formes quadratiques. En particulier, en vertu des isomorphismes bien connus entre groupes classiques, les cas n = 3 et n = 4 se ram6nent ainsi ~t des questions analogues sur les alg6bres de quaternions, qu'on sait traiter directement; or c'6tait justement les cas qui donnaient le plus de difficult6 chez Siegel. Ces cas 6tant acquis, tout le reste de la d6monstration peut maintenant se pr6senter fort simplement, et presque sans calculs. D'autre part, on a commenc6 "aexaminer d'autres groupes semi-simples: groupe
404
[1959a ] Adeles et groupes algdbriques
spin (Tamagawa), groupes lin6aire sp6cial et projectif sur une alg6bre simple, etc. D a n s t o u s l e s cas q u ' o n a su traiter, on a trouv6 que ~(G) est un entier, et qu'il est dgal ~ 1 lorsque G est un groupe semi-simple "simplement connexe" (au sens alg6brique, c'est-g,-dire que tout g r o u p e isog6ne ~ G est une image h o m o m o r p h e de G). I1 en est bien ainsi, p a r exemple, si G est le g r o u p e des 616ments de n o r m e 1 sur le centre dans une vari6t6 d'alg6bre simple sur un corps de n o m b r e s ou bien sur un corps de fonctions.
Bibliographie 1. Ono, Takashi. Sur une propri6t6 arithm6tique des groupes alg+briques commutatifs, Bull. Soc. math. France, t. 85, 1957, p. 307 323. 2. Siegel, Carl Ludwig. Ober die analytische Theorie der quadratischen Formen, I : Annals of Math., t. 36, 1935, p. 527 606; II: t. 37, 1936, p. 230 263; III: t. 38, 1937, p. 212 291. 3. Tamagawa, Tsuneo, M6moire ',i para~tre aux Annals of Mathematics.
[1959b] Y. Taniyama (lettre d'Andr~ Weil) II est impossible, pour un math6maticien fran~ais de m o n ~ge, d'6crire sur T a n i y a m a sans songer aussit6t "~ Herbrand. Celui-ci aussi restera dans la m6moire de ceux qui l'ont connu c o m m e l'une des plus fortes personnalit6s parmi les math6maticiens de sa g6n6ration. Herbrand est mort h 23 ans dans un accident de montagne; au dire de ses compagnons, la prudence n'6tait pas en montagne sa qualit6 dominante; il semble que la vie n'importe plus beaucoup h ceux qui ont franchi certaines fronti6res de l'intelligence. Pour quiconque a connu Herbrand, il est difficile de croire que, s'il avait v6cu, notre math6matique, et particuli6rement notre th6orie des nombres, aurait tout h fait l'aspect qu'elle pr6sente aujourd'hui; il 6tait de ceux dont on attend, non seulement qu'ils r6solvent tel ou tel probl6me avant les autres, mais qu'ils enrichissent la science d'id6es que d'autres n'auraient point. T a n i y a m a aussi 6tait de ceux-l~. Sa personnalit6 rut pour moi l'une de celles qui domin6rent le colloque de T o k y o - N i k k o en 1955, et dont je conservai la plus forte impression a la suite du s6jour au Japon que je fis ~ cette 6poque; s'il 6tait clair qu'il s'y m61ait des 616ments dissonants, en conflit les uns avec les autres et avec le m o n d e ext6rieur, il n'y avait pas de raison de voir 1~ autre chose que le bouillonnement d'un temp6rament jeune qui n'a pas encore r6alis6 son harmonie interne; et j'attendais beaucoup, pour moi au moins autant que pour lui, du s6jour qu'il avait 6t6 invit6 ~ faire "~ Princeton. Je n'ai pas besoin de dire ici m o n 6motion et m o n chagrin quand j'ai appris queje ne devais plus le revoir. Du moins, plus heureux q u ' H e r b r a n d (dont le nom, en dehors de son oeuvre logique, reste attach6 seulement ~ deux ou trois r6sultats tr6s fins, tr6s en avance sur leur 6poque, et qui n'ont trouv6 leur explication que r6cemment), T a n i y a m a nous a laiss6 un travail de premier ordre, compl6tement achev6 dans le cadre qu'il s'6tait fix6, mais qui ouvre de vastes perspectives sur les plus importants probl6mes de l'arithm6tique moderne; c'est bien entendu de son m6moire de 1957 sur les fonctions L, paru darts le Journal of the Mathematical Society of Japan, que j'entends parler; trop modestement, il dit qu'il y suit "rues m6thodes," alors qu'en r6alit6, prenant pour point de d6part quelques observations que j'avais eu l'occasion de faire, et les joignant "~ ses propres r6sultats, il y d6veloppe des m6thodes enti6rement neuves et d'une grande port6e. Sans entrer ici dans des commentaires d6taill6s, j'observerai seulement que, pour bien apercevoir les id6es directrices du m6moire, il convient de commencer par le dernier chapitre. Lh il est montr6, au moyen de la r6duction modulo p, que, si A est une vari6t6 ab61ienne d6finie sur un corps de hombres alg6briques k, les repr6sentations l-adiques du groupe de Galois de k sur k (off/~ est la cl6ture alg6brique de k), d6termin6es par les groupes de points d'ordre / N sur A, ne sont pas ind6pendantes les unes des autres, mais sont li6es entre elles et avec la fonction z6ta de A par des relations tr6s pr6cises; et l'id6e peut-6tre la plus originale de T a n i y a m a a 6t6 de voir que ces relations, telles qu'il les formule, m6ritent d'fitre 6tudi6es en elles-m~mes, ind6pendamment de la vari6t6 A d'ofi
405
406
[1959b] Y. Taniyama on les a tirees. C'est cette etude qu'il a menee ~t bien, par une analyse des plus fines et ingenieuses, dans le cas "abdlien" off les representations/-adiques en question engendrent des algebres commutatives (cas qui se presente justement dans la multiplication complexe); ses resultats comprennent comme cas particulier ses thdoremes anterieurs sur les fonctions zeta des varietes abeliennes fi multiplication complexe; en meme temps, il 6claire par lfi d'un jour tout nouveau la theorie des caracteres "de type (A0)" dont j'avais seulement signal6 l'existence au cours du colloque de Tokyo-Nikko. Mais c'est bien entendu le cas non abelien dont l'etude l'attirait, sans qu'il air pu, semble-t-il, rien entreprendre ~ ce sujet; il est inutile d'ajouter que ce probleme ne semble abordable par aucune des methodes dont nous disposons actuellement. Pendant longtemps encore, sans doute, ce travail fournira des sujets de meditation aux arithmeticiens. J'ignore, au m o m e n t off j'dcris, s'il s'est retrouv6 parmi les papiers de T a n i y a m a des traces de recherches plus recentes. Mais, quand meme il n'en serait pas ainsi, son memoire de 1957 suffirait fi lui assurer dans l'histoire de notre science une place durable. A ceux qui l'ont connu personnellement, il laisse, avec un inoubliable souvenir, l'amer regret de n'avoir rien su ou rien pu faire pour le retenir parmi eux.
[ 1960a] De la m&aphysique aux math6matiques (h propos d' un colloque re'cent)
Les math6maticiens du XVlIIe si~cle avaient coutume de parler de la (( m6taphysique du calcul infinit6simal )), de la (( mdtaphysique de la thdorie des ~quations )). Ils entendaient par lh u n ensemble d'analogies vagues, difficilement saisissables et difficilement formulables, qui n~anmoins leur semblaient jouer u n r61e i m p o r t a n t hun m o m e n t donn6 dans la recherche et la d~couverte math~matiques. Calomniaient-ils la (( vraie )) mdtaphysique en e m p r u n t a n t son n o m pour d6signer ce qui, dans leur science, 6tait le moins clair? Je ne chercherai pas h 61ucider ce point. E n tout cas, le mot devra 6tre e n t e n d u ici en leur sens; h la (( vraie )) mdtaphysique, je me garderai bien de toucher. Rien n'est plus f6cond, tousles mathdmaticiens le savcnt, que ces obscures analogies, ces troubles reflets d ' u n e th6orie h une autre, ces furtives caresses, ces brouilleries inexplicables; rien aussi ne donne plus de plaisir au chercheur. U n jour v i e n t off l'illusion se dissipe; le pressentiment se change en certitude; les th6ories jumelles r6v~lent leur source commune a v a n t de disparaitre; comme l'enseigne la G~tgt on a t t e i n t h la connaissance et h l'indiffdrence en m~me temps. La m6taphysique est devenue math~matique, prate h former la mati~re d ' u n trait6 dont la beaut6 froidc nc saurait plus nous 6mouvoir. Ainsi nous savons, nous, ce que cherchait ~ deviner Lagrange, q u a n d il parlait de mdtaphysique h propos de ses t r a v a u x d'alg~bre; c'est la thdorie de Galois, qu'il touche presquc du doigt, h travers u n 6cran qu'il n'arrive pas k percer. Lh off Lagrange voyait des analogies, nous voyons des th6orbmes. Mais ceux-ci ne p e u v e n t s'dnoncer q u ' a u moyen de notions et de (c structures )) qui pour Lagrange n'dtaient pas encore des objets math6matiques : groupes, corps, isomorphismes, automorphismes, tout cela avait besoin d'&tre con~u et d6fini. T a n t que Lagrange ne fair que pressentir ces notions, r a n t qu'il s'efforce en vain d'atteindre h leur unit6 substantielle h travers la multiplicitd de leurs incarnations changeantes, 52
Reprinted by permissionof Hermann, 6diteurs des sciences et des arts
4O8
[1960a]
409 mat
h~matiques
il reste pris dans la m6taphysique. Du moins y trouve-t-il le fil eondueteur qui lui permet de passer d ' u n probl~me ~ l'autre, d'amener les mat6risux k pied d'oeuvre, de tout mettre en ordre en vue de la th6orie g6n6rale future. Grace h la notion d6eisive de groupe, tout eela devient math6matique ehez Galois. De m~me eneore, nous voyons les analogies entre le ealeul des diff6renees finies et le ealeul diff6rentiel servir de guide h Leibniz, h Taylor, h Euler, au eours de la p6riode h6roique durant laquelle Berkeley pouvait dire, avee autant d'humour que d'h-propos, que les ~ eroyants ~ du ealeul infinit6simal 6taient peu qualifi6s pour eritiquer l'obseurit6 des myst~res de la religion ehr6tierme, eelui-l~ 6t~nt pour le moins aussi plein de myst~res que eelle-ei. Un peu plus t~rd, d'Alembert, ennemi de toute m6taphysique en math6nmtique eomme ailleurs, soutint dans ses articles de l'Encyclop~die que la vraie m6taphysique du ealeul infinitesimal n'6tait pas autre chose que la notion de limite. S'il ne tira pas lui-m~me de eette id6e tout le patti dont elle 6tait susceptible, les d6veloppements du si~ele suivant devaient lui donner raison; et rien ne saurait 8tre plus elair aujourd'hui, ni, il faut bien le dire, plus ermuyeux, qu'un expos6 correct des 616ments du ealeul diff6rentiel et int6gral. Heureusement pour les ehereheurs, h mesure que les brouillards se dissipent sur un point, e'est pour se reformer sur un autre. Une grande pattie du eolloque de Tokyo s'est d6roul6e sous le signe des analogies entre la th6orie des nombres et la th6orie des fonetions alg6briques. Lh, nous sommes eneore en pleine m6taphysique. C'est de ees analogies, paree que j'en ai quelque exp6rienee personnelle, que je voudrais parler iei, avee l'espoir, vain peut-~tre, de donner aux leeteurs ~ honnStes gens ,~ de eette revue quelque id6e des m6thodes de travail en math6matique. D~s l'enseignement 616mentaire, on fait voir aux 61~ves que la division des polynSmes (~ une variable) ressemble beaueoup h la division des entiers et eonduit des lois toutes semblables. Pour les uns eomme pour les autres, il y a un plus grand eommun diviseur, dont la d6termination se fait par division sueeessive. A la d6eomposition des hombres entiers en faeteurs premiers correspond la d6eomposition des polynSmes en faeteurs irr6duetibles; aux nombres rationnels correspondent les fonetions rationnelles, qui, elles aussi, peuvent toujours se mettre sous forme de fractions irr6duetibles; eeUes-ei s'ajoutent par r6duetion au plus petit eommun d6nominateur, ete. I1 est done tout naturel de penser qu'il y a analogie entre les hombres alg~brlques (raeines d'6quations dont les eoeffieients sont des nombres entiers) et les ]onctions algdbriques d'une variable (raeines d'6quations dont les eoeffieients sont des polynSmes ~ une variable). Le fondateur de la doute 6t6 Galois s'fl qu'on trouve sur ee mort, d'ofl on peut
th6orie des fonetions alg6briques d'une variable aurait sans avait v6eu; e'est ee que permettent de penser les indications sujet dans sa e61~bre lettre-testament, 6erite k la veille de sa eonelure qu'il touehait d6jk h quelques-unes des prineipales 511
410
[1960a1 m
a
t
h
~
m
a
t
i
q
u
e
s
d~eouvertes de Riemann. Peut-6tre aurait-il donnd h cette thdorie une allure algdbrique, conforme ~ l'esprit des t r a v a u x contemporains d ' A b e l et de ses propres reeherches d'alg~bre pure. A u eontraire, Riemann, l'un des moins alg6bristes sans doute parmi les grands math6maticiens du xIxe si~ele, m i t la th6orie sous le signe du (( t r a n s c e n d a n t , (mot qui, pour le math6matieien, s'oppose t~ (calg6brique ,, et ddsigne t o u t ee qui a p p a r t i e n t en propre au continu). Les mdthodes tr~s puissantes mises en oeuvre p a r R i e m a n n amen~rent presque du premier coup la th6orie un degrd d'aeh~vement qui n ' a gu~re 6t6 d6pass6. Mais elles ne tiennent aucun eompte des analogies avee les hombres alg6briques, et ne p e u v e n t 6tre transposdes telles quelles en vue de l'6tude de eeux-ci, 6rude qui relive traditionnellement de l'arithmdtique ou th6orie des hombres, et qui, du v i v a n t ddjh de Riemann, 6tait en voie de d6veloppement rapide. C'est Dedekind, ami intime de Riemann, mais algdbriste eonsomm6, qui d e v a i t le premier tirer p a r t i des analogies en question et en faire un instrument de recherche. I1 appliqua avec succ~s, aux probl~mes trait6s p a r R i e m a n n p a r voie transeendante, les mdthodes qu'il a v a i t lui-m~me cr6~es et raises au point en vue de l'6tude arithm6tique des nombres alg6briques; et il fit voir qu'on peut retrouver ainsi la partie p r o p r e m e a t alg6brique de l'oeuvre de Riemann. A premiere vue, les analogies ainsi mises en 6videnee restaient superficielles, et ne paraissaient pas pouvoir porter sur les probl~mes les plus profonds de l'une ni de l ' a u t r e thdorie. H i l b e r t alia plus loin dans cette voie, ~ ce qu'il semble; mais, s'il est probable que ses 61~ves subirent l'influence de ses iddes sur ce sujet, il n'en est rest~ quelque trace que dans un eompte rendu obscur qui n'a m~me pas dt6 reproduit dans ses (Euvres complktes. Les lois non dcrites de la math6matique moderne interdisent en effet de publier des r u e s mdtaphysiques de eette esp~ce. Sans doute est-ce mieux ainsi; a u t r e m e n t on serait accabl6 d'articles encore plus stupides, sinon plus inutiles, que tous ceux qui encombrent h prdsent nos pdriodiques. Mais il est dommage que les iddes de H i l b e r t n'aient 6t6 ddveloppdes par lui nulle part. I1 y a v a i t loin encore, cependant, de l'arithm6tique, off r~gne le discontinu, h la th~orie des fonctions au sens classique. Or, en disant que les fonetions algdbriques sont raeines d'dquations dont les coefficients sont des polyn6mes, j ' a i volontairem e n t omis un point i m p o r t a n t : ces polynSmes eux-m6mes ont des coefficients; mais eeux-ci, quels sont-ils? Lorsqu'on traite de la division des polyn6mes darts l'enseignement 616mentaire, il v a sans dire que les coefficients sont des ((nombres ,~ : nombres (( rdels ,~ (rationnels ou non, mais donnds en tout eas, si l'on veut, p a r un d6veloppement d~eimal), ou, h u n niveau un peu plus 61ev6, nombres(( r6els ou imaginaires ,, ou, comme on dit, ((nombres complexes ~. C'est exclusivement de nombres complexes qu'il s'agit dans la th6orie riemannienne. Mais, du point de vue de l'alg6briste pur, t o u t ee qu'on demande a u x (( hombres ~ en question, c'est qu'ils se laissent combiner entre eux au moyen des quatre op6rations (ce que l'algdbriste exprime en disant qu'ils forment un ,~ corps ,~). Si on 54
[1960a]
411 m
a
t
h 6 m a t i
q
u
e
s
n ' e n suppose pas plus sur leur eompte, on obtient une thdorie des fonetions alg6briques, fort riche ddj~t (comme en t6moigne le volume r6cent et ddj~ classique q u ' a publi6 Chevalley sur ce sujet), mais qui ne l'est pas assez pour que les analogies avec les nombres algdbriques puissent btre poursuivies j u s q u ' a u bout. Heureusement il s'est trouv6 u n domaine intermddiaire entre l'arithmdtique et la thdorie riemannienne, et qui possbde, avec chacune de ces deux derni~res thdories, des ressemblances beaucoup plus 6troites qu'elles n ' e n ont entre elles; il s'agit des fonctions algdbriques (( sur u n corps fini )). Comme on le savait depuis Gauss, s'il ne s'agit que de pouvoir faire les quatre opdrations, il suffit d ' u n nombre fini d'dldments. I1 suffit par exemple d'en avoir deux, q u ' o n nommera 0 et 1, et pour lesquels on posera par convention la table d'addition et la table de multiplication que voici : 0+0=0, 0+1=1+0=1, 1+1=0; 0 • 0=0, 0 • 1~1 • O~ 1 X 1=1. Quelque paradoxale que puisse paraltre au profane la rSgle 1 -4- 1 = 0, quelque t e n t a n t qu'il soit de dire que c'est 1~ u n pur jeu de l'esprit qui ne rdpond ~ aueune (( rdalitd ,, u n tel syst$me est monnaie eourante pour le mathdmaticien; et Galois en ~tendit beaucoup l'usage en construisant les ~( imaginaires de Galois ,. P r e n a n t done les coefficients de nos polynSmes dans u n (~ corps de Galois ~, on construit des fonetions alg6briques dont la thdorie remonte ~ Dedekind mais s'est particuli~rement d~velopp6e depuis la th~se d'Artin. Pour dire en quoi elle consiste, il faudrait entrer darts des d~tails beaucoup trop techniques qui n ' a u r a i e n t pas leur place icio Mais on peut, je erois, en donner une idde imagde en disant que le mathdmatieien qui 6tudie ces probl~mes a l'impression de d~chiffrer une inscription trilingue. Darts la premiere colonne se trouve la thdorie riemannienne des fonctions alggbriques au sens classique. La troisi&me colonne, c'est la th$orie arithmdtique des nombres algdbriques. La eolonne du milieu est celle dont la ddeouverte est la plus rdeente; elle eontient la thdorie des fonetions alg6briques sur u n corps de Galois. Ces textes sont l'unique source de nos connaissances sur les langues dans lesquels ils sont dcrits; de ehaque colonne, nous n'avons bien entendu que des fragments; la plus complete et celle que nous lisons le mieux, encore ~ prgsent, c'est la premi&re. Nous savons qu'il y a de grandes diff6rences de sens d ' u n e colonne ~ l'autre, mais rien ne nous en avertit ~ l'avance. A l'usage, on se fait des bouts de dictionnaire, qui permettent de passer assez souvent d'une colonne ~ la colonne voisine. C'est ainsi q u ' o n avait ddchiffrd depuis longtemps, dans la derni~re colonne, le ddbut d ' u n paragraphe intituld (~ fonetion z6ta ~). Vers la fin de ee paragraphe, on croit lire une phrase tr~s mystdrieuse; elle dit que t o u s l e s zdros de la fonction se t r o u v e n t sur une certaine droite. Jamais on n ' a pu savoir s'il en est bien ainsi, ou s'll y a eu erreur de lecture. C'est le c~l~bre probl~me de 1' (( hypoth~se de R i e m a n n )), qui dans quelques mois sera tout juste centenaire. 55
412
[1960a1 ~
at
h ~ m a ~
q
ue~
La prineipale ddcouverte d'Artin, dans sa thtse, c'est qu'il y a, dans la seconde colonne, un paragraphe intittd6 aussi (( fonetion z6ta ~,, et qui est t~ peu de chose pros une traduction de celui qu'on connaissait d6jh; notre dietionnaire s'en est trouv6 beaueoup enrichi. Artin apergut aussi, dans eette colonne, la phrase sttr l'hypoth~se de Riemann; elle lui parut tout aussi mystArieuse que rautre. Ce nouveau probl~me, t~ premi$re rue, ne semblait pas plus faerie que le pr6e~dent. Ign rfiahtd, nous savons maintenant que la premitre colonne contenait ddjh tous les 616ments de sa solution. II n'dtait que de traduire, d'abord en th~orie (( abstraite ~ des fonctions algdbriques, puis dans le langage (( galoisien ~ de la seconde eolonne, des rfisultats obtenus depuis longtemps par Hurwitz en (( riemarmien ,,, et que les g~om~tres italiens avaient ensuite traduits dans leur propre langage. Mais les meilleurs sp6eiahstes des th6ories arithm~tique et (( galoisienne ~ ne savaient plus lire le riemannien, ni h plus forte raison l'italien; et il fallut vingt arts de recherches avant que la traduction f6t rnise au point et que la d6monstration de l'hypoth~se de Riemann dans la seeonde eolonne ffit complttement ddehiffr6e. Si notre dietionnaire 6tait suftisamment c0mplet, nous passerions aussitx3t de I~ la troisi~me colonne, et l'hypoth~se de Riemann, la vraie, se trouverait d~montrde, ere aussi. Mais nos cormaissanees n'atteignent pas jusque lh; bien des d~ehiffrements patients seront encore ndcessaires avant que la traduction puisse 6tre faite. Au tours du colloque auquel il a $t6 fair allusion plus haut, il a 6t6 beaucoup diseut6 de (( m~taphysique ~ tr propos de ces probltmes; un jour eelle-ci fera place h une thdorie mathdmatique dans le cadre de laquelle ils trouveront leur solution. Peutgtre, comme c'6tait le cas pour Lagrange, ne nous manque-t-il, pour franehir ee pas d6cisif, qu'une notion, un concept, une (( structure ~. D'ing6nieux philologues ont bien trouv~ le secret des archives de Nestor et de celles de Minos. Combien de temps faudra-t-il encore pour que notre pierre de Rosette, h nous autres arithm~tieiens, rencontre son Champollion?
56
[ 1960b] Algebras with involutions and the classical groups
IT has been knowr~ for a long time t h a t there is a close connection between semisimple algebras with involutions and the classical semisimple Lie groups and Lie algebras. B u t the precise degree of generality of this relationship does not seem to have been ascertained anywhere, at least explicitly, in the printed literature on this subject. I n the first part of the present paper, it will be shown how this can be done, at a n y rate over a groundfield of characteristic 0, b y borrowing some elementary techniques from m o d e r n algebraic geometry. Then, taking for our groundfield the field R of real numbers, we shall give, for the l~iemannian s y m m e t r i c spaces a t t a c h e d to the classical groups, an i n t e r p r e t a t i o n which rests u p o n the use of algebras with involution. This was already implicit in Siegel's f u n d a m e n t a l work on discontinuous groups, and it is hoped t h a t our results will help to achieve a b e t t e r u n d e r s t a n d i n g of t h a t work and to clarify also some of its arithmetical aspects.
PART
I
~EMISlMPLE GROUPS OVER A FIELD OF CHARACTERISTIC 0. 1.
I n P a r t I, all spaces, varieties, groups are to ba understood
in the sense of algebraic geometry; b y this we mean t h a t t h e y are all allowed to have points in the universal domain, and not only in their field of definition. I f V is a v a r i e t y (for instance a group), t Work.supported (in part) by the O.U.I%.P.A.F. The author is greatly indebted to his colleague Mr. M. for a counterexample and other helpful remarks, to Mr. P. (the famous winner of many cocycle races) for the main idea of Part I, to others for conversations totally unrelated to the problems considered here, and to the Indian iV~atheraatieal Society for kindly allowing this paper to rest for over two years in their editorial officesbefore letting it take its flight into the world. Reprinted by permission of the editors of J.
Ind. Math. Soc.
413
414
[1960b] 5'90
ANDRE WEIL
d e f n e d over a field k, we shall denote b y Vk the set of points of V with coordinates in k (in group-theory, this differs from the usage of Chevalley, according to which a group G, defined over an infinite field/~, is always identified with the set which we call G~). We assume the universal domain to be of characteristic 0 ; without restricting the generality, one could take it to be the field of complex numbers. Within the universal domain, we select a groundfield k and a normal algebraic extension K of k, of finite degree d; we call g the Galois group of K over /~, and we write ~ for the imago of an element ~ of K u n d e r an automorphism a ~ g ; we have (~a), = ~a~ for all a, ~ in g. I f V is a vector-space of dimension n (over the universal domain), defined over k, Vk and VK are vector-spaces of dimension n over k and over K, respectively. We have Vk a VK, and we m a y identify VK with the tensor-product V~ | K t a k e n over k; g operates in an obvious m a n n e r on VK, and Vk consists of the elements of VK which are invariant under g. I f V' is the dual space of V (over the universal domain), and if T is a n y tensor-product (V | V | ...) | | ...), over the universal domain, of f~etors identical either with V or with V', t h e n T~ is the tensor-product similarly built up from V~ and its dual V~ over/~; and a similar s t a t e m e n t holds for T K. B y a c o c y d e , we shall u n d e r s t a n d a mapping a - + $'~ of g into the group of automorphisms of the vector-space V, such t h a t : (a) for each a E g, Fo is defined over K ; (b) for all a, ~ in ~, we have Fa~ = ( F o y o F~. Let (Fo) be such a cocycle; for e v e r y x ~ VK, p u t :
O n e s e e s at once t h a t x [~] = (xt~])[~], which means t h a t the group g can also be made to operate on VK b y (x, a ) - + x E~ These operations are k-linear, b u t not K-linear; we have (~x) t~] = ~~ t~] for x e VK and ~ e K. F r o m this, it follows t h a t the set W of those elements of VK which are invariant under all operations x - + x ["], i.e. which satisfy x ~' = F o ( x ) for all a, is a vector-space over k. I f ql . . . . . a~ are linearly independent over k in W, it is easy to see, b y induction on m, t h a t t h e y are linearly independent over K in
[1960b]
415 ALGEBRAS W I T H
INVOLUTIONS
59l
VK; for otherwise, because of the induction assumption, there would be a relation E ~ i a ~ = 0 , with the ~ in K and not all in k, a n d ~m = 1 ; a p p l y i n g to this the operation x - + x [~], we get Z i ~ a~ = 0, hence Z i ( ~ - - ~) a~ = 0, and therefore, because of the induction assumption, ~ = a s s u m p t i o n on the a~. Now k; t a k e a basis (~1 . . . . . ~ ) elements Z~ ~ x [~l are in W ~
~ x [~]
c:~B
~i for all i a n d ~, which contradicts the t a k e for al, ..., a m a basis for W o v e r of K over }; for e v e r y x e V~, the for i = 1 . . . . . d, so t h a t we can write : ~
(1~i
d)
~=1
with % ~ k. I t is w e l l - k n o w n t h a t the d e t e r m i n a n t [ ~ [ is not O, and therefore these equations can be solved for the x ["1, yielding expressions for t h e m as linear combinations of the a with coefficients in K ; in particular, x itself is such a linear combination. We h a v e t h u s shown t h a t a~ . . . . , am is a basis for IrK o v e r K , so t h a t in particular we m u s t h a v e m = n. T a k e now a basis b1.... , b, of Vk over k ; let (I) be the a u t o m o r p h i s m of V which m a p s bi onto a i for 1 ~< i ~< n ; it is defined over K. Combining the relations a i = O(b~), a~ : F~(ai), b~ = bi, we get $'~(O(bi) ) = O~(b~) for all i, hence F~ o 9 : 9 ~ or F~ = (I)~ r (which can be expressed b y saying t h a t "all eocycles are trivial " ). N o w lot t, t', ... be elements of " t e n s o r - s p a c e s " T, T ' , .... i.e. of tensor-products built u p f r o m factors identical with V or V'. More precisely, a s s u m e that t e T k, t' ~ T ' k, . . . , and that all the teasers t, t', ... are i n v a r i a n t u n d e r every one o f the a u t o m o r p h i s m s $',; b y this we m e a n of course t h a t t is i n v a r i a n t under the canonical extension of F , to T, etc. Similarly, write (I) for the canonical extension of ~) to T, T ' , ... and p u t tl = (I)-l(t), etc. We have: t~ = ( r
= r
= tl
for all a, hence t 1 e T~, and similarly t' 1 e T'~, etc. The m a i n application which we have in view concerns the case of algebras a n d of algebras with involution. B y an algebra A , we u n d e r s t a n d a vector-space V with the additional structure deter-
416
[1960b] 592
ANDRE W E I L
mined on it b y a bilinear mapping of V • V into V, or, what amounts to the same, b y an element t of the tensor-space T = V' | V' | V; if, in addition to this, we prescribe an endomorphism ~ of V, or, what amounts to the same, an element t' of T ' ~- V' | V, such t h a t is an i n v o l u t o r y antiautomorphism (or, as we shall say more briefly, an involution) of the algebra A, t h e n we speak of V, with the structure determined on it b y t and t', as the algebra with involution (A, ,). All algebras will be assumed to be associative and to have a unit-element, usually denoted b y 1. We say t h a t the algebra A, with the underlying vector-space V and the multiplicative structure determined b y the element t of V' | V' | V, is defined over k if V and t are defined over k; t h e n A k is an algebra over k in the usual sense, and A K is the algebra derived from A k b y extending the groundfield from k to K. The same holds for algebras with involution. As our universal domain is assumed to be of characteristic 0, it is known t h a t A, is somisimple if and only if A (as an algebra over the universal domain) is so. The center Z of a semisimple algebra A, defined over k, is a commutative semisimple algebra, also defined over k; Z k is then the center of A k. We say t h a t A is absolutely simple if it is simple as an algebra over the universal domain ; then it is isomorphic to a m a t r i x algebra Mn over the universal domain. On the other hand, we say t h a t a semisimple algebra A, defined over k, is simple over k if it is not a direct sum of subalgebras of A, all defined over k; this will be so ff and only ff A k is simple as art algebra over k, or also if and only if the center Z of A is simple over k, or again ff and only if Z k is a field. I t is clear t h a t the groups of automorphisms of algebras and of algebras with involution are algebraic groups. As a special case of the results p r o v e d above, we have now the following theorem : THEOREM 1. Let A be an algebra (resp. an algebra with involution) defined over a field k; let K be a Galois extension of k with the Galois group g. Let (2'0) be a cocycle of fi, consisting of automorphisms of A. Then there is an algebra (rosp. an algebra with involution) A 1, defined
[]960b]
417
ALGEBRAS WITH INVOLUTIONS
593
over k, and an isomorphism r of A 1 onto A , defined over K , such that F a ---- r o Cp-:for all (~ Gg.
In,fact, let V be the underlying vector-space of A; and let t be the tensor (resp. let t, t' be the pair of tensors) on V which defines the structure of A. Define t: and (I) (resp. t:, t~ and (I)) as above by means of the cocycle (F~); and define A 1 as the algebra (resp. the algebra with involution) defined on V by t: (resp. by tl, t~). These will satisfy all the conditions in our theorem. 2. It is our purpose to show, by means of Theorem 1, that, with few exceptions, the classical semisimple groups over any field of characteristic 0 can be represented as groups of automorphisms of semisimple algebras with involution ; in fact, it will turn out that there is almost a one-to-one correspondence between these two classes of objects. We begin by discussing the classical simple groups over the universal domain. Consider first the projective linear group PL(n) in n variables, i.e. the factor-group of GL(n) by its center (to be consistent with our notation, we omit any mention of the underlying field when the latter is the universal domain). Let A be the direct sum of two algebras, both isomorphic to the matrix algebra M~ of order n (i.e. consisting of all n • n matrices); on A, consider the involution defined by (x, Y) - + (t y, tx) ' where X , Y are two matrices of order n, and tX, t y are their transposes. Using the classical theorem of Skolem-2qoethor, one sees immediately that the automorphisms of the algebra with involution A (i.e., the automorphisms of the algebra A which commute with the given involution) make up an algebraic group G, consisting of two connected components G0, G:; GO consists of all the automorphisms of A which leave each component M , of A invariant, and more precisely of the automorphisms: (X, Y ) ---+ (X, y),(M) = ( M - 1 X M ,
tM. Y. t M - 1 )
where M is an arbitrary invertible matrix ; the mapping M - + r is then a homomorphism of GL(n) onto Go whose kernel is
418
[1960b] 594
ANDRE WEIL
the center of GL(n), so that GO may be identified with P L ( n ) . As to (71, it consists of those automorphisms of A which exchange its two components, and may also be defined as the coset of (7o in (7 which contains the automorphism (X, Y ) - + (Y, X). The inner automorphisms of G induce automorphisms on Go ~hieh are either inner automorphisms of Go or products of such automorphisms with the automorphism induced on G o by (X, Y ) - + ( Y , X); the latter may be written as r r and it is easy to see that, for n > 3, this is not an inner automorphism It is well known that these are all the automorphisms of G0 = PL(n). As our calculation also shows that no automorphism of A, other than the identity, induces the identical isomorphism on Go for n > 3, it follows that, for n > 3, every automorphism ofG 0 can be derived, in one and only one way, from an automorphism of A. This implies that, for n > 3, if A ' is an algebra with involution, isomorphic to A, and G' o is the connected component of the identity in the group of automorphisms of A', every isomorphism of G0 onto G'0 can be derived, in one and only one way, from an isomorphism of A onto A'. In order to obtain in a similar manner the orthogonal and sympleotio groups, or rather their quotients by their centers, take A = Mn; sine0 X - + t X is an involution on A, the most general antiautomorphism of A is of the form X - - + . F - I . t X . . F , which is involutory if and only if t F = h F , with ~ in the center; this implies ~9.= 1, so that our involution is given by X - + F -1. t X . . F with F invertible and t.F - 4- .F. An automorphism X - + M - 1 X M of the algebra A commutes with that involution if and only if .F = t M . . F . M ; let G be the group consisting of such automorphisms. I f t $ , = _ F, n must be even, and the matrices M satisfying .F -= t . M . F . M are of determinant 1 (as follows from the consideration of the pfaffian) and make up a connected algebraic group, the symplectic group Sp(n) ; G is the quotient PSp(n) of that group by its center. It is known that in that case (7 has only inner automorphisms ; and we see, just as above, that every automorphism of G
[1960b]
419 ALGEBRAS W I T H I N V O L U T I O N S
595
can be derived in one a n d only one w a y f r o m an a u t o m o r p h i s m of the given algebra with involution. T a k e now the case tF ~ F ; as our underlying field is the universal domain, it would be no restriction to t a k e for F the u n i t - m a t r i x I n. The matrices M for which F = t M . F . M m a k e u p the o r t h o g o n a l group O(n), with two connected c o m p o n e n t s O + ( n ) , O - ( n ) consisting of the matrices M in the group with the d e t e r m i n a n t + I a n d - - 1 , respectively; a n d the connected c o m p o n e n t GO of the identity in 0 is the quotient PO+(n) of O+(n) b y its center. [f n is odd and ~ 3, the center of O(n), consisting of :t: 1~, h a s one element in each one of those components; therefore G is connected v~nd m a y be identified with O+(n). I t is k n o w n that, also in this c~se, G has only inner a u t o m o r p h i s m s , a n d our f u r t h e r conclusions are the same as before. Finally, t a k e the case in which n is even a n d ~ 4. T h e n the center of O(n), consisting again of 4- ln, is contained in O + ( n ) ; therefore G has two connected components. Hero it is k n o w n t h a t the g r o u p of inner a u t o m o r p h i s m s of G o is of index 2 in the g r o u p of all a u t o m o r p h i s m s of Go, e x c e p t for n =: 8, in which case it is of index 6. On the o t h e r hand, it is easily seen t h a t the inner a u t o m o r p h i s m s of (7 induced by elements of G~ determine on Go a u t o m o r p h i s m s which are not inner ones of G0. Therefore, if we leave aside the exceptional ease n = 8, we can again conclude t h a t e v e r y a u t o m o r p h i s m of G O can be derived in one a n d only one w a y f r o m an a u t o m o r p h i s m of the algebra w i t h involution A. N o w observe t h a t e v e r y semisimple algebra A o v e r the universal d o m a i n is a direct sum of m a t r i x algebras, a n d t h a t e v e r y involution of A m u s t either leave a c o m p o n e n t of A i n v a r i a n t or interchange it with a n o t h e r one. Thus, o v e r the universal d o m a i n , e v e r y semisimple algebra with involution is, in an obvious senso~ the direct s u m of algebras with involution of one o f the t y p e s discussed above. On the o t h e r hand, ff Go is the connected c o m p o n e n t O f the i d e n t i t y in the g r o u p of a u t o m o r p h i s m s o f the algebra with involution A, it is clear t h a t the a u t o m o r p h i s m s in Go m u s t t r a n s f o r m each c o m p o n e n t of A into itself. F r o m this
420
[1960b] 596
ANDRE WEIL
it follows immediately t h a t GO must be a direct product of groups of the various types considered above, and t h a t conversely e v e r y such p r o d u c t can be obtained in this manner. Some groups, however, are obtained in this m a n n e r more t h a n once, because of the well-known isomorphisms between groups of the various families ; those are as follows : (a)
M2 admits an inv~176
with J = ( --10 ~ )
which is invariant under all automorphisms and antiautomorphisms of M 2 ; this follows for instance from the fact that, for any invertible m a t r i x M in M2, we have
J-l.tM.J
----dot(M). M -1.
Therefore SL(2) is identical with Sp(2), hence PL(2) with PSp(2). (b)
P O + ( 3 ) is isomorphic with PSp(2).
(c)
PO +(4) is isomorphic with the product of itself.
(d)
PO +(5)
(e) PO +(6)
PO +(3)
with
is isomorphic with PSp(4). is isomorphic with PL(4).
In view of these circumstances, let us restrict our list of groups and algebras to the following: (I) (~roups: all semisimple groups, with center reduced to the neutral element, which, when decomposed into a direct product of simple groups, contain no factor isomorphic either to one of the exceptional groups or to P0+(8). (II) Algebras with involution: all semisimple algebras with involution which, when decomposed into a direct sum, consist of summands isomorphic to one of the following : (a) M~, $ M~ for n ~ 3, with an involution exchanging the two summands ; (b) M2n for n >~ 1, with the involution X - + j - a . eX. j determined b y an invertible alternating m a t r i x J ; (c) M s with the involution X - + tX, f e r n :=- 7 or n ~ 9.
[1960b]
421 ALGEBRAS WITH INVOLUTIONS
597
Then it follows from what we have p r o v e d t h a t each elm of the groups in ore. list is isomorphic to the connected component of the identity in the group of automorphisms of one of our algebras with involution, and that, if A and A' are two such algebras and G and G' are the corresponding groups, a n y isomorphism between G and G' is induced b y a uniquely determined isomorphism between A and A'. The latter s t a t e m e n t holds in particular for A == A', (7 -- (7'. 3. L e t (7 be a connected algebraic group, defined over the groundfield 1~. Lot us assume t h a t (7 is semisimplo, has a center reduced to the neutral element, and that, when G is decomposed into a p r o d u c t of simple factors over the universal domain, none of these factors is isomorphic to one of the five exceptional groups or to PO+(8). F r o m the results ill w2 it follows t h a t there is an algebra with involution A0, defined over the prime field, such t h a t the connected component of the identity Go in the group of automorphisms o f A 0 is isomorphic to G over the universal domain. Let f be an isomorphism of (7 onto Go; take for K a normal algebraic extension of/c, of finite degree, over which f is defined. Notations being now the same as in w1, f~ o f - 1 is an automorphism o f Go, defined over K. B y the results of w2, there is a uniquely determined automorphism $'~ of A o which induces the automorphism f " o f - 1 on Go; it is therefore invariant u n d e r all automorphisms of the universal domain over K ; the characteristic being 0, this implies t h a t it is defined over K. I t is obvious t h a t (F~) is a cocycle. Therefore, b y T h e o r e m 1, there is an algebra with involution A d e f n e d over b, and an isomorphism O, defined over K, of A onto A 0, such t h a t F , = (Pr (I)-1 for all a. Let G' be the connected component of tlm identity in the group of automorphisms of A ; r determines an isomorphism, which we again call (I), of G' onto G o ; t h e n ~ ----(I)-1 o / i s an isomorphism, defined over K, of (7 onto G'. Moreover, we have (I)~ o r ___f~ o f - 1 for all a. This can also be written as ~ = ~ . Therefore ~ is defined over b, and we m a y use it to identify G with G'. We have thus proved t h a t any group G
of the given type can be represented as the connected component of the
422
[1960b] 598
ANDRE W E I L
identity i~ the group of automorphisms of a 8emisimple algebra with involution defined over k. 4. We shall now determine when the connected component Go of the identity in the group of automorphisms of a semisimple algebra with involution (A, ~) is not a semisimple group. Over the universal domain, this is immediately a p p a r e n t from the the results of w3. I n fact, P L ( n ) is simple except for n = 1, in which ease it is reduced to 1; PSp(n) is defined only if n is even, and is always simple; and P O + (n) is semisimple except for n - - 1 and n = 2. Therefore Go is semisimple, and has a center reduced to the identity, provided no component of (A, ~) is isomorphic to M~ with the involution X - + tX. We shall say t h a t a semisimple algebra with involution is non-degenerate if it has no such component and if at the same time it has no c o m m u t a t i v e component. Then we have the following t h e o r e m : THEOREM 2. Let (A, ~) be a non-degenerate semisimple algebra with involution; let Go be the connected component of the identity in its group of automorphisms, and let Uo be the connected component of 1 in the multiplicative group of the elements u of A such that u'u----1. Then Go is a semisimple group, isomorphic to the quotient of Uo by its center; and the center of Go is reduced to the identity. We have already shown t h a t Go is semisimple a n d has a center reduced to the identity. As the group of automorphisms of the center Z of A is finite, e v e r y element o f Go must induce the i d e n t i t y o n Z and is therefore (by the classical theorem of Skolom-Noothor) a n inner automorphism x--+ v - 1 xv, with an invertible v in A. C_,allj(v) this automorphism ; ff we write t h a t it commutes with ~, we find t h a t z = v'v m u s t be in Z; it must then be in the multiplicative group H consisting of those invertible elements o f Z which are even for ~. Call V the group of those elements v o f A for which v'v is in H, and U the group of those elements u of A for which u'u = 1. Consider the homomorphism (h, u) -+ hu o f / / • U into V; one sees at once t h a t it maps H • U onto the group of those elements v of A for which v~v is i n / / 3 ; but, over the universal domain,
[1960b]
423 ALGEBRAS W I T H
INVOLUTIONS
599
we h a v e H ~ --- H (since H is c o m m u t a t i v e a n d the characteristic is not 2), a n d therefore t h a t h o m o m o r p h i s m m a p s H • U onto V. F r o m this, it follows t h a t U a n d Y h a v e the same image j(U) = j ( V ) u n d e r j. As we h a v e seen t h a t G o is contained in j ( V ) , our conclusion follows. Incidentally, we observe t h a t the classical Cayley t r a n s f o r m a t i o n m a y be used to s t u d y the g r o u p U o of T h e o r e m 2. L e t us say t h a t an e l e m e n t x of an algebra with involution (A, ~) is even (:for ~) if x' = x, and odd (for t)if x ' = x; as the characteristic is not 2, A is the direct s u m of the spaces A +, A - of even a n d odd elements for e. N o w let u be a generic element of U 0 over a field of definition k for (A, ~) ; write w = (1 - - u). (1 + u ) - l ; t h e n w is an odd element of A for ,. Conversely, if w is a generic element of A - over ]r the formula u = (1 - - w).(1 ~- w) -1 defines a generic element o f U 0 o v e r k. Thus these formulas define a birational correspondence between U 0 and A S. E v e r y algebra with involution o v e r a field ]c can be w r i t t e n as (A~, ~), where (A, ~) is, in the sense explained above, a n algebra with involution, defined o v e r It; we h a v e already observed t h a t , if the latter is semisimple, the f o r m e r is semisimple, a n d conversely. We shall say t h a t the f o r m e r is non-degenerate if the latter is so. T h e relation between the structures of these two algebras will now be briefly discussed, particularly in order to find when (A k, ~) is non-degenerate. To decompose the semisimple algebra A~ over ]c into simple c o m p o n e n t s is the s a m e as decomposing A o v e r k, i.e. writing it a s a direct s u m o f sub-algebras, defined o v e r k, in such a w a y t h a t none o f the s u m m a n d s can be split a n y f u r t h e r in the same manner. E a c h s u m m a n d is t h e n t r a n s f o r m e d into itseff or interchanged with a n o t h e r one b y the involution L. T h u s it is enough to consider the case in which (Ak, ~) is simple, which m e a n s t h a t A~ is either simple or the direct sum of two simple algebras Bk, Uk interchanged
424
[1960b]
600
ANDRE WEIL
by e. I n the latter case, let K be the center of Bk; call d the degree of K over/c; if we consider B k as a vector-space over K, its dimension must be of the form n 2. Then A splits into the direct sum of two algebras B, C, which m a y be regarded as the tensor-products o f B~ and C~ with the universal domain over ]c, and which are interchanged b y ~. Over the universal domain, B is the direct sum of d algebras isomorphic to M s. Degeneracy can only occur for n -- 1, i.e. when A~ is commutative. Next, assume t h a t A k is simple; let K be its center, and K + the set of all even elements of K for ~; K + is a field; let d be its degree over ]c. I f K is not the same as K +, it is an extension of K + of degree 2; call n 2 the dimension of A~ as a vector-space over K. Then A is, over the universal domain, the direct sum of 2d algebras isomorphic to Ms; b y considering the center of A, one sees at once t h a t none of these 2d components is invariant u n d e r ~. Therefore degeneracy occurs only for n = 1, i.e. again when A~ is commutative. Finally, assume t h a t A k is simple and t h a t ~ induces the identity on the center K of A k, so that, with the above notation, we have K = K +. Let d be the degree of K over ]c, and n ~" the dimension of A~ as a vector-space over K ; t h e n A is the direct sum of d algebras isomorphie to M~, each of which is invariant under ~; it is nondegenerate whenever n ~ 3. I f n = 1, A k is c o m m u t a t i v e and degenerate. I f n = 2, A k is a quaternion algebra over K (which m a y be isomorphic to M s (K)); let 8 be the dimension of A / a s a vectorspace over K. Then one sees at once that, in each one of the d simple components of A over the universal domain, the odd elements for e make up a vector-space of dimension 3; comparing this with the results of w2, we find t h a t we have 3 = 3 or 3 = 1 according as c, in each one of these components, is of the t y p e X ' ~ j - a . tX. j with t j = _ j or of the t y p e X ' = tX. Thus degeneracy occurs if and only if ~ : 1. One verifies easily t h a t this is so if and only if A~, as an algebra over K, can be generated b y an e v e n element u and an odd
element v such t h a t
uv =
-- vu, u ~ ~ K,
v 2 ~ K.
Thus a necessary and sufficient condition for the non-degeneracy of a simple algebra with involution (Ak, e) over /c is t h a t it should
[1960b]
425 ALGEBRAS WITH INVOLUTIONS
601
not belong to the t y p e just described and t h a t A~ should not be commutative. A necessary and sufficient condition for the nondegeneracy of a semisimple algebra with involution (Ak, ,) over/c is then t h a t none of its simple components should be of either one of these types. II.
ALGEBRAS WITH INVOLUTION OVER THE REAL FIELD
6. F r o m now on, we shall not need a universal domain, as we shall be operating with algebras over a fixed groundfield, mostly the field R of real numbers. I f A is an algebra over any ground field/r we denote b y ' t r ' the trace of the regular representation of A. In other words, if L u is the ondomorphism x - + ux of the underlying vector-space to A, tr(u) is the trace of L u. Then tr(xy) is a symmetric bilinoar form on A • A ; according to a well-known criterion, it is non-degenerate if and only if A is absolutely semisimplo (i.e. if it is semisimple and remains so under a n y extension of the groundfiold). The trace tr(x) is invariant u n d e r all automorphisms of A ; if A is semisimplo, the right-hand and left-hand regular representations are equivalent, and t h e n the trace is also invariant under all antiautomorphisms o f A, or, as we shall say more briefly, all antimorphisms of A. I f u is a n y element of A, its " minimal polynomial " P is the polynomial o f smallest degree, with coefficients in the groundfield k, such t h a t P ( u ) = 0; P is also the minimal polynomial for the endomorphism L~ of the underlying vector-space to A. The " s p e c t r u m " Su of u is the set of all the distinct roots of P in the algebraic closure of k. L e t us say t h a t u is semisimple if L is so (or, what a m o u n t s to the same, if u generates an absolutely semisimple subalgebra of A) ; this will be the ease if and only if all roots of P are simple. Assume t h a t u is semisimple, and t h a t its spectrum S, is contained in k; let f be a n y k-valued function, defined on k or on a subset o f k containing S u. T h e n there is a polynomial Q, with coefficients in k, coinciding with f on S~; moreover, Q is uniquely determined modulo P ; therefore the element Q(u) o f A does not depend upon the choice of Q ; this will be denoted by f(u) ; f --+ f(u)
426
[1960b] 602
ANDRE WEIL
is clearly an isomorphism of the ring of k-valued functions on Su onto the subalgebra of A g e n e r a t e d b y u. I f f is a k-valued function, defined on Su, the s p e c t r u m of f(u) is f(Su); and, ff g is a k-valued function, defined on f(Su), we h a v e g(f(u)) = h(u), with h == g o f. The t r a n s f o r m u ~ of art element u of A b y an a u t o m o r p h i s m or a n t i m o r p h i s m A of A has the same s p e c t r u m S u a n d the same m i n i m a l polynomial as u, and is somisimplo if u is so. Therefore, if S u is contained in k a n d f is a k-valued function on S u, we h a v e f ( u ~) = f(u) ~. L o t u be a n invertible element of an algebra A. We shall denote b y j(u) the inner a u t o m o r p h i s m determined b y u, i.e. defined b y the formula X ~
X j(u) "-- % - - l x u .
I f u, v are b o t h iuvertible, we h a v e j(uv) = j(u)j(v). I f A is a n y a u t o m o r p h i s m or a a t i m o r p h i s m of A, A-lj(u)A is the inner autom o r p h i s m j(u') with u ' = u ~ if ~ is an a u t o m o r p h i s m a n d u'---(u~) -1 if it is an a n t i m o r p h i s m . We shall write e(2) = 1 w h e n e v e r is an a u t o m o r p h i s m , e()t) = - - 1 whenever it is an a n t i m o r p h i s m , a n d define a s y m b o l u [~] b y the formula
so t h a t we have, in all cases
A-1 j(U) ~ =j(u[;~]). L e t A be a semisimple algebra, a n d let ~ be an a n t i m o r p h i s m of A. T h e n tr(x ~ y) is a non-degenerate bilinear f o r m on A x A. We h a v e (x ~ y)~ = y~ x ~, and hence tr(x ~ y) = tr(y ~ x~), which shows t h a t the bilinoar f o r m tr(x ~ y) is s y m m e t r i c if a n d only if ;~2 __ 1, i.e. if a n d only ff ~ is an involution. More generally, let A be a n a n t i m o r p h i s m of A; t a k e a e A, b e A, a n d consider the bilinear f o r m tr(ax ~ by). Obviously, it is non-degenerate if a n d only ff a, b are invortible, i.e. if t h e y are not zero-divisors. Assume now t h a t ~ is an involution, and t h a t a a n d b are invortiblo ; the formula
[1960b]
427 ALGEBRAS WITH
I/~/VOLUTIONS
603
tr(ax ~ by) ~ tr(x "~bya) = tr(y ~ b~ xa ~) shows t h a t the f o r m tr(ax ~ by) is s y m m e t r i c if a n d only if, for all y, we h a v e bya = b ~ ya ~, i.e. y = b -1 b ~ ya ~ a -1 ; this is so if a n d only if we h a v e b ~ ~- bz - i , a ~ = az, where z is an invertible element in the center Z of A; as a~-----az implies a = z~ a ~, z m u s t satisfy z z ~ - 1 . L e t , be an involution of the semisimple algebra A. As tr(x' y) is a. non-degenerate s y m m e t r i c bilinoar f o r m on A • A, we can use it to a t t a c h an " a d j o i n t " L ' to e v e r y e n d o m o r p h i s m L of the underlying v e c t o r - s p a c e ; this is defined, as usual, b y the f o r m u l a : tr((Lx)' y) : t r ( x ' ( L ' y)). I n particular, the formula tr((ux) ~y) = tr(x~(u ~y)) shows t h e n t h a t the adjoint o f L~ is L ~ , a n d in particular t h a t L~ is self-adjoint if and only if u = u '. LEMMA 1. Let (A, ~) be a semisim21e algebra with involution over a field k of characteristic O. Let A +, A - be the subspaces of even and of odd elements of A f o r ~. T h e u A + and A - are the orthogonal complements of each other for each one of the symmetric bilinear f o r m s tr(xy) and tr(x' y). The formula t r ( x y ) = tr(y ~x ~) : tr(x' y~) shows at once t h a t , if one of the two elements x, y is in A + a n d the o t h e r in A - , 2tr(xy) a n d therefore tr(xy) m u s t be 0. Conversely, assume for instance t h a t x is orthogonal, with respect to tr(xy), to all vectors y such t h a t y~-=- 9 where e is + 1 or - - 1. T h e n it is orthogonal to y + e y' for all y ~ A, so t h a t we h a v e 0 = tr(x(y + 9 y~)) ----tr((x ~- e x~)y) for all y c A , a n d therefore x + 9 0 since tr(xy) is non-degenerate. The p r o o f for tr(x ~y) is quite similar. Of course L e m m a 1 remains valid for every groundfield of characteristic other t h a n 2, provided tr(xy) is non-degenerate, i.e. provided A is assumed to be absolutely somisimple.
428
[1960b] 604
ANDRE W E I L
7. F r o m now on, we shall deal exclusively with semisimple algebras o v e r R; such a n algebra is isomorphic to the direct s u m of m a t r i x algebras o v e r R, o v e r C (the field of complex numbers) and over K (the division-algebra of quaternions). I f A is a n y algebra o v e r R, we say t h a t a n involution ~ on A is positive if tr(x~x) > 0 for all x other t h a n 0 in A ; tho existence of such an involution implies t h a t tr(x~y) is non-degenerate and therefore t h a t A is semisimple. PI~O1,OSITION 1.
Let A be a semisimple algebra over R; then there exists at least one positive involution on A; and all positive involutions on A coincide on the center Z of A. W r i t e A as a direct s u m of simple c o m p o n e n t s A i. A positive involution :r on A m u s t t r a n s f o r m each A i into itseff; for, if it t r a n s f o r m e d Ai, say, into Aj with j ~=i, x~x would be 0 for all x in A s. F r o m this one concludes a t once t h a t it is enough to p r o v e our proposition in the case o f a simple algebra A, which we can write as a m a t r i x algebra Mn(D ) over a division algebra D which can b e R , C o r K . I f D i s R o r K , the center Z i s R , a n d e v e r y a u t o m o r p h i s m or a n t i m o r p h i s m of A induces the identity on Z. I f D = C, we h a v e Z = D, a n d e v e r y a u t o m o r p h i s m or a n t i m o p r h i s m of A m u s t induce on Z either the identity or the a u t o m o r p h i s m z - + ~ , where ~ is the i m a g i n a r y conjugate of z; as we h a v e tr(z) = n 2 (z-]-z) for e v e r y z e Z, a n y positive involution :r on A m u s t induce z - + z on Z, fur, if it induced the identity, one would h a v e tr(z~z) = n~(z ~ + ~ ) , a n d this is < 0 for z = i. This proves the second assertion in our proposition. As to the first one, observe t h a t x~ > 0 for all x :/: 0 in D i f w e write x for x if D ----R, for the i m a g i n a r y conjugate of x if D = C, a n d for the q u a t e r n i o n conjugate of x if D = K ; also, in all three cases, x - + ~ is an involution on D, and X - + t 2 ~ is as usual, denotes the transpose p u t r(X)----Z~xi~; if, as always, regular r e p r e s e n t a t i o n of M~(D) seen t h a t we h a v e
an of we as
involution on M~(D) ff t~, X. F o r X----(x~j) in M~(D), denote b y t r the trace of the a n algebra over R, it is easily
[1960b]
429 ALGEBRAS ~VITH I N V O L U T I O N S
605
tr(X) = ~, [~-(x) + ~(x)]
with ~ equal to n/2 i f D = R , to n i f D = C , Therefore, if X -- (a0), we have
and t o 2 n i f D = K .
tr(t~. X) = 2y . ~ xox~, which shows that X--+tX is a positive involution on Mn(D ). This completes the proof. COROLLARY. I f ~, fl are two positive involutions on an al9ebra A over R, or is an inner automorphism of A. The assumption implies that A is semisimplo and that a-1fl is an automorphism of A; by prop. 1, it induces the identity on th0 center; b y the Skolem-Noethor theorem, it must therefore be an inner automorphism. If A is any semisimplo algebra over R, the automorphism of the center Z of A induced on it by all positive involutions of A will be denoted by z - + ~ ; if Z~ is any simpl0 component of Z, z --~ ~ induces the identity on it if it is isomorphic to R, and the imaginary conjugate if it is isomorphic to C. Let ~ be a positive involution on the algebra A over R. We have seen above that, ff u----u =, L~ is seff-adjoint for tr(x~y), i.e. for the quadratic form tr(x~x); as the latter is positive, it follows from well-known theorems that u is then semisimplo and has a real spectrum. Now assume that an element a of A is such that the bilinear form tr(x=ay) is symmetric, non-degenerate and positive; as it is symmetric and non-degenerate, a must be invertible and even for ~; that being assumed, the formula
tr(x=ay) = tr((L~x)~y) shows that the positivity of the bilinear form tr(x~ay), i.e. that of the quadratic form tr(x=ax), is the same thing as the positivity of the seff-adjoint operator L~ for the form tr(x~x); and this is positive if and only if all the characteristic roots of La(i.o., all the elements of its spectrum S~) are > 0. When that is so, we say that a is positive for ~; and we denote by P(~) the set of all such
430
[1960b1 606
ANDRE WE.IL
elements. F r o m the results of no. 6, it follows t h a t , if f is a n y real-valued function, defined on the set of real n u m b e r s > 0, f(a) is defined for all aEP(~), a n d satisfies f(a) ~ = f(a); this will be so, in particular, for f(t) = t p, for a n y p c R, a n d for f(t).= log t. I f f(t) > 0 for all t > 0, ~hen f(a) is in P(~) for all a eP(~), since the s p e c t r u m of f(a) is the imago u n d e r f of the s p e c t r u m of a. Thus, for e v e r y p ~ R , we h a v e a m a p p i n g a - + a p of P(~) into itself. F o r all p , p ' i n R , wohaveaP+P'=aPa r a n d (aP) r = a pC. I n particular, we note that, for a eP(~), the only b ~P(a) such t h a t b 9" = a i s
b
=
a 1/2.
PROPOSITION 2. Let o~ be a positive involution on an algebra A over R. Then the set P(a) of positive elements for r162is a convex open subset of the vector-space of even elements for r162 ; and the group A* of invertible elements of A operates on P(~) by (x,a) ~ x~ax. I t is clear t h a t P(~) is convex. T a k e a eP(~) ; if p is the smallest element o f the s p e c t r u m S~ of a, i.e. the smallest charactreistie value of La(with respect to the quadratic form tr(x~x)), we h a v e p > 0 and
tr(x~ax) ~ P tr(x~x) for all x E A. Now, in the space of even elements o f ~ in A, i.e. of elements u of A such t h a t u ~ = u, t a k e a neighborhood U of 0 such t h a t , for e v e r y u E U, all characteristic values of L,~ are > - - p a n d < p ; then, for e v e r y u ~ U, a q- u is in P(~); this shows t h a t P(~) is open in the space of even elements. The last assertion in our proposition is obvious. P~OPOSITION 3. Let ~ be a positive involution on an algebra A over R ; let ~ be any automorphism or antimorphism of A. Then ~-1 ~ is a positive involution on A; and ~ maps P(cr onto P(h-I~2). The first assertion is obvious. F u r t h e r m o r e , we h a v e a e P(a) if a n d only if tr(x ~ ay) is s y m m e t r i c and positive. As the trace is i n v a r i a n t u n d e r 2, this is the same as to say, if ~ is an a u t o m o r p h i s m , t h a t tr(x ~ a ~ y~) is s y m m e t r i c a n d positive ; ~his p r o p e r t y will be unaltered if we replace x, y b y x ~-1, y~-l, a n d is therefore equivalent
[1960b]
431 ALGEBRAS V~ITH INVOLUTIONS
607
to the s y m m e t r y and positivity of tr(x ~-1~ a ~ y), which proves our proposition in this ease. Similarly, if ;~ is an antimorphism, a 9 P(r162 is equivMent to the s y m m e t r y and positivity o f tr(y a a a x~); replacing x, y b y y~-l~, x~-l~, we get our conclusion as before. 8. We can now determine as follo~s the set of all positive involutions on a semisimple algebra A over R. PROPOSITION 4. Let cr be a positive involution on an algebra A over R. Then, for every a e P(r162 the formula fl
:
~j(a)
:
j(a -1/~) (zj(a 1/2)
determines a positive involution fl on A . Conversely, i f fl is any positive involution on A , it can be expressed by that formula with an a 9 P(~). For every a e P(:r the element b = a 11~is also in P(r162 and the formulas of no. 6 give j(a) = j ( b ~) and r162j(b-1) :r = j ( b ) ; this gives at once ~j(a) = j ( b -1) r162 ; as this is the transform of ~ by the automorphism j(b), it is a positive invohttion. The p r o o f of the converse will be derived from the following lemma : Lv,~MA 2. On every semisimple algebra A over R, there is a positive involution ~ such that, i f a and b are any two elements of A , the following properties are equivalent: (i) the bilinear form tr(ax ~ by) in x, y is no~-dege~erate, symmetric, and positive ; (ii) there is an element z of the center Z of A such that z~z = l, z a e P ( ~ ) , z - t b 9 P(r162
One can see at once t h a t (ii) implies (i) whenever ~ is a positive involution on A. I n fact, as tr(ax: by) does not change if we replace a, b b y za, z -1 b with z E Z, we m a y replace the assumption (ii) by the assumption t h a t a, b are in P(~). We have already seen
432
[1960b] 608
ANDRE W E I L
in no. 6 t h a t tr(ax ~ by) m u s t t h e n be n o n - d e g e n e r a t e a n d s y m m e t r i c ; p u t t i n g n o w u ---- a 1/9, v = b 1/2, we h a v e t r ( a x ~ bx) = t r ( u 2 x ~ v 2 x) = t r ( ( v x u ) ~ v x u ) • 0 since u ~ : u, v ~ = v, w h i c h p r o v e s t h e positivity. Now, in order t o c o n s t r u c t a n i n v o l u t i o n for w h i c h t h e c o n v e r s e is true, it is clearly e n o u g h to consider the case in w h i c h A is simple, i.e. of the
f o r m M n ( D ) with D = R ,
the involution X-+tX
C or K ; we shall s h o w t h a t
has t h e n t h e required p r o p e r t y .
I n fact,
a s s u m e t h a t A, B are t w o m a t r i c e s in Mn(D), such t h a t t r ( A t X JBY) is n o n - d e g e n e r a t e , s y m m e t r i c a n d positive; as we h a v e seen in no. 6, t h e first t w o a s s u m p t i o n s i m p l y t h a t A, B are invertible a n d t h a t t ~ = z - 1 1~, A = z A , where z is an invertible e l e m e n t o f the center, i.e. w h e r e z is a n o n - z e r o scalar, a n d a real one if D is R or K. Moreover, z m u s t satisfy z.~ -- 1, w h i c h implies z = 4- 1 if D = R or K. I f D ~ C, t a k e ~ e C such t h a t ~2 = z, a n d p u t A 1 = CA, B1 = ~ - I B ; t h e n A 1, B 1 also satisfy (i), a n d we h a v e t ~ 1 = A I , tJB1 = B1; therefore, after so m o d i f y i n g A, B in t h e case D----C if necessary, we m a y a s s u m e t h a t t ~ __ 9 A, t/~ = 9 B, with 9 = 4- 1 for D = R or K a n d 9 = 1 for D ----C. N o w t h e p o s i t i v i t y a s s u m p t i o n in (i) m e a n s t h a t t r ( A t X B X ) :> 0 for all X :/= 0; p u t t i n g t = r ( A t X B X ) , w h e r e v, as above, d e n o t e s t h e usual t r a c e o f a m a t r i x , this can be w r i t t e n as t + t > 0. B u t , p u t t i n g W = A t X B X , t = ~(tW) = T(tX. 9 B . X .
we h a v e
9 A ) = t,
so t h a t o u r p o s i t i v i t y a s s u m p t i o n a m o u n t s t o v ( A t X B X ) > 0 fo X ~ 0. T a k i n g t w o c o l u m n - v e c t o r u = (u~), v = (vi), p u t :
f ( u ) = q~Au, g(v) = ~vBv ; these are m a t r i c e s o f o r d e r 1, i.e. scalars (in D). T h e p o s i t i v i t y a s s u m p t i o n , applied to t h e m a t r i x X ----v. tu ---- (vi ~j), gives
9 (A ~ X B X ) = ~(f(u) g(v)) =Au) g(v)> 0 for all u ~: 0, v r 0. I f D = R a n d 9 -- - - 1, we h a v e f ( u ) = O, g(v) = 0 for all u, v; therefore this case c a n n o t occur. I f 9 ---- 1, f ( u ) a n d g(v) are real ; t h e y m u s t h a v e t h e s a m e sign for all u r 0, v ~ 0, so t h a t ,
[1960b]
433
ALGEBRAS WITH INVOLUTIONS
609
after replacing A, B b y -- A , - - B if necessary, we can write our assumption as f ( u ) > 0 for all u ~ O, g(v) > 0 for all v ~ 0; as one sees immediately, this implies t h a t A, B are positive for the involution X --+ *X. Finally, if D = K and r ------ 1, p u t q = f ( u ) , r = g(v); t h e n we have q = -- q va 0, 7 ---- -- r ~ 0; applying the positivity assumption to u, 0v with 0 e K, we see t h a t we must have q 0r 0 > 0 for a l l 0 va 0; as one can always find 0 such t h a t 0 r this is impossible.
0------q,
We can now complete the proof of Prop. 4 for the ease of an involution 0r having the p r o p e r t y stated in L o m m a 2. L e t fl be a n y positive involution on A. B y the corollary of Prop. 1, r162 fl is an inner a u t o m o r p h i s m j ( a ) of A, so t h a t we m a y write fl ~ ~j(a) with an invertible a. As we have seen in no. 6, the antimorphism fl = ~j(a) is an involution if and only if tr(x~y) is symmetric ; therefore it is a positive involution if and only if the bilinear form tr(x~y) = t r ( a - a x~,ay) is non-degenerate, symmetric and positive, i.e. if (a -a, a) satisfies condition (i) of the lemma. B u t then, b y our assumption on a, there is z e Z such t h a t z - a a is in P(:r Replacing a b y z - a a does not change j(a) ; we have therefore proved t h a t fl is of the form ~j(a) with a e P(~), hence also of the form j(b) - 1 ccj(b) w i t h b = a 11~. This proves Prop. 4 for the particular :r which we have been considering. I t also proves t h a t all positive involutions on A are transforms of this involution ~ b y inner automorphisms. Since the p r o p e r t y of expressed b y Prop. 4 is obviously invariant under automorphisms, our proof is thus complete. One m a y observe t h a t the p r o p e r t y of ~ expressed in L e m m a 2 is also invariant u n d e r automorphisms ; therefore (i) and ( i i ) a r e equivalent whenever ~ is a positive involution. 9. B y Prop. 4, if ~ is a positive involution on A, the mapping a - + ~j(a) maps P(~) onto the set of all positive involutions on A ; this shows in particular t h a t the latter set is connected if it is provided with its natural topology as a d o s e d subset of the space of all endomorphisms of the underlying vector-space to A.
434
[1960b] 610
ANDRE
WEIL
One can also define as follows a one-to-one mapping, and more precisely a h o m e o m o r p h i s m , of a closed subset of P(cr onto the set of positive involutions on A. We observe first t h a t , if ~, fl are two positive involutions, a n d if we write fi as the t r a n s f o r m of ~ b y a n inner a u t o m o r p h i s m j ( b ) of A, then, b y Prop. 3, P(fl) is the imago of P(~) u n d e r j(b). I n particular, if Z is the center of A, and if we write P ( Z ) = Z c~ P(~), P ( Z ) is independent of the choice o f ~. I f A is the direct sum of the simple algebras Ai, and if, for each i, ei is the unit-element of A i, it is easily seen t h a t P ( Z ) consists of the elements Z~ ti ei where all the ti are real a n d > 0.
Let ~r be a positive involution on the algebra A over R; let a, a' be two elements of P(~r Then we have j(a) -~j(a') if and only if a - la' is in P(Z). LEMMA 3.
P u t b = a 1/2 a n d z---- a - 1 a'. I f z is in Z, we h a v e az-= b~zb ; b y Prop. 2, if this is in P(a), z m u s t be in P(~), hence in P(Z). The converse is obvious. Now, calling again A i the simple components of A, a n d ei the unit-element of A i for each i, write N~ for the n o r m of the regular r e p r e s e n t a t i o n of A i o v e r R ; if d~ is the dimension of A i o v e r R, we have, for t e R , Ni(tei) -- t% F o r e v e r y x i e A i , write vi(xi)-([Ni(xi) ll/di)ei; and, for e v e r y x = Y , i x i i n A, with x i e A i for all i, write v(x) : Z i vi(xi). Then v is a m a p p i n g of A into P(Z), inducing the identity on P ( Z ) a n d such t h a t v(xy) = v(x)v(y) for all x, y in A. Therefore, for e v e r y invertible x in A, there is one a n d only one element z of P ( Z ) such t h a t v(zx) = 1, viz. z = v(x) -1. F u r t h e r m o r e , if a is a semisimple elemenr of A with a positive s p e c t r u m (e.g. if a is in P(~) for some positive involution cr we h a v e v(a~) : v(a) p for all p e R ; for this ,is t r u e if p is a n integer, hence if it is rational, a n d therefore b y continuity in the general case. I f ~ is a n y a u t o m o r p h i s m or a n t i m o r p h i s m of A, we h a v e v(x ~) = v(x) ~ for all x e A. We shall denote b y PI(~.) the sot of all elements a of P(~) such t h a t v(a) ---- 1. Combining Prop. 4 with L e m m a 3 a n d some trivial topological considerations, we get t h e following 9
[1960b]
435 ALGEBRAS WITH INVOLUTIONS
611
I~OPOSITIO~ 5.
Let :r [J be two positive involutions on the algebra A over R. Then there is one and only one element a of PI(~) such that fl = r162 ; and an element a' of P(~) is such that fl = :r i f and only i f it is of the form az with z e P ( Z ) . Moreover, the mapping a----~:cj(a) induces on P1(:r a homeomorphism of PI(U) onto the set of all positive involutions on A .
10. W e shall d e n o t e b y ~ t h e g r o u p o f all a u t o m o r p h i s m s a n d a n t i m o r p h i s m s o f the semisimple a l g e b r a A o v e r R. F o r )~ e ~ , t h e n o t a t i o n e(•), u [~] will be used in t h e sense explained in no. 6. L~MA 4. Let r162 be a positive involution on A ; let ~ be an element of f~ commuting with ~. Then, for a ~ P(~), the involution ~j(a) commutes with ~ i f and only i f a TM ----za with z ~ P ( Z ) ; for a E Pi(~), ~ j(a) commutes with ~ i f and only if a TM = a. Clearly, aj(a) c o m m u t e s w i t h ~ if a n d o n l y i f j ( a ) does so, i.e., b y t h e f o r m u l a s o f no. 6, ff a n d o n l y if j ( a TM) is t h e same as j(a) ; this a m o u n t s t o s a y i n g t h a t a TM m u s t be o f t h e f o r m za w i t h z E Z. B u t , as ~ c o m m u t e s w i t h ~, it t r a n s f o r m s P(~) into itseff (by P r o p . 3), so t h a t a TMis in P(~) w h e n e v e r a is in P(m). A p p l y i n g L e m m a 3, we get t h e first assertion in o u r l e m m a . I f v(a) = 1, we h a v e v(a TM) = 1, hence u(z) = 1 for a TM -~ za ; for z ~ P(Z), this implies z = 1. LEMMA 5.
I f tWO positive involutions ~, fl commute with each other,
they coincide. I n L e m m a 4, t a k e )~ = ~, a n d w r i t e fl as ~j(a) with a E P I ( ~ ) ; w e get a -1 --~ a, i.e. a ~ -~ 1, hence a ---- 1. I t is clear t h a t e v e r y element )~ o f fr t r a n s f o r m s a positive i n v o l u tion ~ into a positive i n v o l u t i o n 2 ~ - ~ , to w h i c h we can a p p l y t h e results p r o v e d a b o v e ; in particular, we can write it as ~j(a(h)) w i t h a (~) e PI(~). W e shall n e e d t h e following p r o p e r t y o f t h e m a p p i n g - + a(%) :
Let ~ be a positive involution on A . For every ~ ~ ~ , let a(h) be the element of P~(a) such that %:r = =j(a(%)). LEMMA 6.
436
[1960b] 612
ANDRE WEIL
Then, for all ~, ix in ~, we have a(A~) = aC~). a(~)t~-11. Call a ' t h e r i g h t - h a n d side o f t h e f o r m u l a to be p r o v e d . B y the definition o f a(~), a(~), we h a v e
~(galx-1) A-1 = }t~}t-~. A j(a(~)) Z -1 = ~j(a(A)) j(a(/x) t~-ll) = aj(a') so t h a t we m u s t h a v e a(Atx) = za', w i t h z in t h e c e n t e r Z. particular, if we replace t L b y A- I , we see t h a t we m u s t h a v e
In
a(~ -~) = z~ a(A)-E~l with z, ~ Z. P u t a = a(A), b = al/2; mj(a) is t h e same as j(b) -~ o:j(b), a n d therefore t h e definition o f a = a(~) can be expressed b y s a y i n g t h a t j(b)A c o m m u t e s w i t h ~ ; therefore it t r a n s f o r m s P(~) into itself. I n p a r t i c u l a r , a j(b)~ must" be in P(~) ; this is no o t h e r t h a n a *. F r o m this it follows a t once t h a t a -[~l is in P~(~) ; in t h e a b o v e f o r m u l a , z, m u s t therefore be equal to 1, a n d we h a v e : a(A -1) = a()l)-[~], which is n o t h i n g else t h a n t h e special case /x = A-1 o f t h e f o r m u l a to be p r o v e d . N o w we go b a c k to t h e g e n e r a l case. W e h a v e s h o w n t h a t a(A/x) : za', w i t h z ~ Z. L e t us w r i t e t h a t za' is in P(~) ; this a m o u n t s to s a y i n g t h a t t h e bilinear f o r m F(x, y) = tr(x ~ za()t) a(/x) [~-11 y) = ~r(a(2)x ~ - 1
za(/x) [~-11 y)
is s y m m e t r i c , n o n - d e g e n e r a t e a n d positive. As t h e t r a c e is i n v a r i a n t b y 2, F(x, y) can be w r i t t e n , if )t is a n a u t o m o r p h i s m , as
F(x, y) = tr(a(A) ~ x ~ z ~ a(/x) y~). B u t in t h a t case a(A) ~ is t h e s a m e as a(~l-1) -1 a n d is therefore in P(~) ; w r i t i n g it as c * w i t h c e P(~), we get
F(x, y) -= tr((x * c) ~ z ~ a(/x) (y~ c)) ; t o s a y t h a t this is s y m m e t r i c , n o n - d e g e n e r a t e a n d positive is to s a y t h a t z ~ a(/~) is in P(~) ; as a(/x) is in P ( a ) , z ~ m u s t therefore be in
P(Z). Similarly, if )t is a n a n t i m o r p h i s m , we use t h e fact t h a t a(2) ~
[1960b]
437
ALGEBRAS WITH INVOLUTIONS
613
is the same as a()t -1) in order to write it as c 2, with c e P(~), and t h e n write F as F ( x , y) = t r ((y~ e) ~ a(/i) -1 z ~ (x ~ c)), again with the same conclusion as before, viz. z ~ E P ( Z ) , i.e. z ~ P ( Z ) . Since obviously v(z) is 1, this implies z = 1, which completes our proof. 11. T~EORE~ 3. Let A be a semisimple algebra over R ; let K be a compact subgroup of the group of all automorphisms and antimorphisms of A . T h e n there is at least one positive involution on A which is invariant under K , i.e. which commutes with every element
elK. Assume first t h a t K is a group of automorphisms of A ; and choose on A a positive involution ~. As in L e m m a 6, we call a()t), for )~ ~ K, the elemen~ of Pi(a) such t h a t Ag~t- i -- ~ j(a(A)).
B y Prop. 5, this is uniquely defined, and A - + .a(2) induces a continuous mapping of K into P(~) ; b y L e m m a 6, this satisfies the relation a()l/i) = a (2). a (/i) ~-i for all A,/~ in K. Now p u t a : [ a(/~) d/z, ,1
K
where d/~ denotes the H a a r measure on K. As P(~) is open and convex in the vector-space of symmetric elements for a, a is in P(a). If, in the integral which defines a, we replace /z b y A/~ with a fixed A e K, and apply the above formula for a ()t/~), we get a -~ a (~). a ~'-1. On the other hand, we have ,~:r
A- i -~ r162
) )~j(a) )~-i : ~j(a(h)a a-i) ;
these two formulas, taken together, show t h a t the positive involution ~j(a) commutes with ~. As A is a r b i t r a r y in K, this proves our conclusion in the present case.
438
[1960bl 614
ANDRE W E I L
N o w assume t h a t K contains a t least one a n t i m o r p h i s m p; then, if K o is the group of all a u t o m o r p h i s m s of A contained in K , we h a v e K = K 0 u p K 0, p 2 ~ K o a n d K o p - - p K o. B y w h a t we h a v e p r o v e d above, we can choose a positive involution ~ w h i c h ' c o m m u t e s with e v e r y element of K-o. I f we write, as before, ~ - 1 in the form aj(a(2) ), with a()t) e Pl(a), for e v e r y A c K , we h a v e a(2) = 1 for A e K 0. P u t a = a ( p ) ; b y L e m m a 6, we h a v e a ( p A ) = a for all A E K 0 , i . e . a ( / ~ ) = a for all / z r 0. As we h a v e p K 0 = K op, this gives, for ~ ~ K0 :
a "= a(~ p) = a(,~), a(p) ~-1 = a ~-1, a n d therefore a = a ~. Similarly, again b y L e m m a 6, we have 1 = a(p 2) = a.(aP) - 1,
a n d therefore a -- a p. Now p u t b -- a 1/2 and fl = aj(b). B~ L e m m a 4, f~ o o m m u t e s with all elements of K 0. Moreover, we h a v e
p fl p - 1 = ~j(a). p j(b) p - ~ = aj(a) j(b[p-ll). As a ==-a p, we h a v e b = bp ; the r i g h t - h a n d side of the last formula is therefore equal to ~j(b), i.e. to ft. This shows t h a t fl c o m m u t e s with all elements of K . CogencY. Let (A, t ) b e a 8emisimple algebra with involution over R. Let K be a compact subgroup of the group of automorphisms of (A, t). Then there is a positive involution on A which commutes with t and with all elements of K. Our a s s u m p t i o n s i m p l y t h a t K U t K is a c o m p a c t subgroup o f the g r o u p of all a u t o m o r p h i s m s and a n t i m o r p h i s m s of A ; our assertion is therefore an i m m e d i a t e consequence of T h e o r e m 3. 12. F r o m now on, we shall deal with a somisimplo algebra with involution (A, ~) o v e r R ; a n d we shall denote b y G its g r o u p of a u t o m o r p h i s m s . B y T h e o r e m 2 of P a r t I, the connected c o m p o n e n t o f the i d e n t i t y in G is semisimple if (A, t) is non-degenerate; and it h a s been explained in P a r t I hi w h a t sense one m a y say t h a t " a l m o s t all " somisimple real groups can be obtained in this manner.
[1960b]
439 ALGEBRAS WITH INVOLUTIONS
615
The corollary of Theorem 3, applied to the case in which K is reduced to the identity, shows t h a t there is at least one positive involution on A which commutes with ~. F r o m the results proved above, we can also deduce at once the following : PROPOSI~ON 6. Let (A, ~) be a semisimple algebra with involution over R. Let ~, fl be two positive involutions on A, both commuting with ~. Then there is one and only one element a of PI(~) such that fl = ~j(a), and it is such that a ' ~ - a - l ; conversely, for every such element a, aj(a) is a positive involution commuting with ~. Furthermore, we have fl-~j(a-1/~):cj(al/~); and the mapping p-+j(aP), for p E R~ is an isomorphism of the additive group R onto a one-parameter group of automorphisms of (A, ~). This is in fact an immediate consequence of Prop. 5 and L e m m a 5. The set of all positive involutions of A commuting with ~, provided with its " n a t u r a l " topology (as explained in no. 9) will be denoted b y R ; the group (7 operates on it b y the law (~,~)-+~-~.~
()~e(7, ~ e R ) ,
and Prop. 6 shows t h a t it operates on it transitively. We shall denote b y K(~) the subgroup of G consisting of the elements of G which commute with ~; R is therefore isomorphic to G/K(~). I t will be shown t h a t K(~) is a maximal compact subgroup of G, so t h a t R is essentially the Riemannian symmetric space attached to G. We first show t h a t R is homcomorphie to an open cell. This will be done b y defining a " t r a c e operator ", corresponding to the " n o r m o p e r a t o r " v defined in no. 9. Write once more A as the direct sum of the simple algebras A i ; call Z~ the center o f A i and ni 2 the dimension of A i as a vector-space over Z i ; denote b y S i the trace in A t over Z~, i.e. the trace of the regular representation of A i considered as an algebra over Z i. Lot x be a n y element of A ; write it as x --- Zixi, with x, e A i for all i; we p u t :
a ( x ) : a ( ~ . , x i ) = ~ n~-~Si(xi) ;
440
[1960b] 616
ANDRE WEIL
a is then a linear mapping of A into Z, which induces the identity on Z ; and, if A is a n y a u t o m o r p h i s m or antimorphism of A, we have a(x ~) ~- a(x) ~ for all x in A. Furthermore, we write
~0(x) = ~[~(x) + ~(x)] where z - + ~ is as defined in no. 7. I f u is a n y element of A, and if e~ is defined as Z u~/n !, (ro(u) is nothing t h a n log v(e u) ; we need only a special case of this, which we formulate as a lemma :
LEMMA 7. Let u be a semisimple element of A with real spectrum ; then we have a o (u) ----log v(e~). I t is clearly enough to consider the case in which A is simple ; its center Z can be identified with R or (2. Let d be the dimension of A over R ; d. a 0 is then the trace of the regular representation of A over R, while, for a n y x e A, d. log ,(x) is the same as log I N (x)[, where N is the n o r m o f the regular representation of A over R. I f u is a semisimple element of A with real spectrum, we can, b y choosing a suitable basis for A, write L~ as a diagonal m a t r i x ; let r 1, ..., r a be its diagonal coefficients. Then, if v = e~, L~ is the diagonal m a t r i x with the diagonal elements e% I n the regular representation of A over R, the trace of u is Zr~ and the n o r m of v is IIe% This proves the lemma. I n particular, let ~ be a positive involution on A ; e v e r y even element u for ~ is semisimple with real spectrum, and, if we put, for such an element, v = e~, v is in P(~) ; conversely, u is given in terms of v by u = log v. Then L e m m a 7 shows that, for such a pair u, v, the relations % ( u ) = 0, v(v) : 1 are equivalent. On the other hand, if u, v is such a pair, and if ~ is a n y involution on A, the relations u ' = - - u , v ~ - - v -1 are equivalent. Combining this with Prop. 6, wo got the following "decomposition t h e o r e m " : THEOREM 4. .Let G be the group of automorphisms of a semisimple algebra with involution (A, ~) over R . Let ~ be a positive involution on A , commuting with e ; let K(r162 the subgroup of G consisting of the elements of G which commute with r162; let W be the vector-space of the elements w of A such that w ~ ~ w, %(w) -~ O, w' -=- -- w. T h e n
[1960b]
441 ALGEBRAS WITH INVOLUTIONS
817
the mapping (w, ~) -+j(ew)~,for w E W, ,~ ~ K(~), is a homeomorphism of W • K(a) onto G. F o r w e W, p u t c -- ew ; f r o m w h a t we h a v e seen above, it follows t h a t t h e m a p p i n g w - + c = ew is a h o m e o m o r p h i s m o f W o n t o t h e set o f all elements c o f Pa(~) such t h a t c r = e -1. Then, b y L e m m a 4, j(e w) is in G, so t h a t j(ew)2 is in G for e v e r y )~ in K(~). L e t n o w /Z be a n y e l e m e n t o f G ; t h e n /Z g/z- 1 is a positive involution c o m m u t i n g w i t h ~, a n d can t h e r e f o r e be w r i t t e n as ~j(a) with a e PI(~), a ' = a - ~. P u t t i n g c = a - ~/2, we g e t
/Z ~ /Z-1 : o~j(a) =: j(e):cj(c) - ~, w h i c h s h o w s t h a t ~ = j(e)-I/Z c o m m u t e s with g ; as /Z a n d j(c) b o t h c o m m u t e w i t h ,, 2 is t h u s in K(~). P u t t i n g w = log c, we get /Z =j(ew)2. This d e c o m p o s i t i o n o f /Z is u n i q u e ; f o r , if we h a v e / Z ----j(c') ?t' with c' e PI(~), c ' ' : - c ' - 1 , ~' e K(~), then, w r i t i n g t h a t )~' c o m m u t e s w i t h ~, we get
/z ~ / z - ~ = j ( ~ ' ) ~j(c')-~ =
~j(c '-~)
which, c o m p a r e d with t h e f o r m u l a w r i t t e n above, gives c ' - * = a a n d t h e r e f o r e c' = a-1/2, b y P r o p . 5. W e h a v e t h u s p r o v e d t h a t t h e m a p p i n g in o u r p r o p o s i t i o n is bijective. Trivial topological considerations will t h e n show t h a t it is bicontinuous. COrOLLArY. Let (A, e) be a semisimple algebra with involution over R. Let R be the set of all the positive involutions on A which commute with ~. Then R, provided with its natural topology, is homeomo~Thic to an open cell. As we h a v e o b s e r v e d above, P r o p . 6 implies t h a t identified w i t h G/K(:r
R can be
if ~. is a n y positive i n v o l u t i o n c o m m u t i n g
w i t h ~ on A. L e t f be t h e canonical m a p p i n g o f G o n t o G/K(a). T h e o r e m 4 s h o w s t h a t t h e m a p p i n g w-+f(j(ew)) is a h o m e o m o r p h i s m o f W o n t o G/K(a). I f the i n v o l u t i o n , induces the i d e n t i t y on the center Z o f A, it is e a s y t o see t h a t t h e relation a ' - - - - a - a , for a n e l e m e n t a o f P(~), implies v ( a ) = 1; a n d similarly t h e relation w ' - - - - - - w, for
442
[]960b] 618
ANDRE WEIL
an even element w for ~, implies ao(W) -~ O. I n t h a t case, it is frequently advantageous, instead of the transcendental mapping w = log c, to use the Cayley transformation t -- (1 -
~). (1 + c) - 1 ,
c -
(1 -
t ) . (1 + 0 - 1 ,
where t satisfies t = t ~ = - - t ' and the inequalities t tr(x ~ tx) i < tr(x~x) for all x r 0 in A ; the set T of these elements t of A is a convex open subset of the space determined b y t = t ~ = - t', and one sees, just as in the proof of Theorem 4 and its corollary, t h a t the formulas written above determine a homeomorphism between T a n d R. 13. W h e n dealing with algebraic groups over R, one must be careful to distinguish between connected components in the algebraic and in the topological sense ; for the example of GL(n, R) shows t h a t the group of points with real coordinates in an algebraic group, defined over R, which is irreducible and therefore connected in the algebraic sense, need not be connected in the topological sense. F o r a similar example where the group is the group of automorphisms G o f a semisimple algebra with involution over R, one m a y take the algebra M ~ + z(R) with the involution X ' = S - ~. tX.S, S being the m a t r i x S----
0
--
?)
'
the algebraic connected component of the i d e n t i t y in G is PO + (S), and it is easy to see t h a t PO+(S, R) is not topologically connected. On the other hand, it is well-known t h a t the set of real points on a n y algebraic variety, defined over R, consists at most of a finite n u m b e r of connected components in the topological sense; this must t h e n be the case, in particular, for all the groups which we have considered so far. As above, let G be the group of automorphisms of a semisimple algebra with involution (A, ~) over R ; for e v e r y positive involution commuting with ~ on A, denote again b y K(a) the group of the elements of G which commute with ~. We shall denote b y G' and b y
[1960b]
443 ALGEBRAS W I T H
INVOLUTIONS
619
K'(a), respectively, the connected components of the identity in G and in K(~) in the topological sense.
Notations being as above, G'/K'(~) is isomorphic to G/K(~), and we have K'(~) -= G' n K(~). LEMMA 8.
P u t K" = G' n K(~) ; call f the canonical mapping of G onto G/K(:r Theorem 4 shows t h a t f(G') is G/K(~); therefore it is simply connccted. But trivial topological considerations show t h a t f(G') is the same as G'/K". On the other hand, K'(~) is obviously the topological connected component of the identity in K". Therefore G'/K'(~) is a covering space of G'/K" ; as the latter is simply connected, this implios t h a t K'(~) = K". T~EOREM 5. Let (A, L) be a semisimple algebra with involution over R ; let G be its group of automorphisms, and G' the topological connected component of the identity in G; let R be the space of all positive involutions commuting with ~ on A ; and let G act on R by (~,~)-+A-l~fori~G,~ ~ R. For each ~ R , let K(~) be the group of those elements of G which commute with ~, and let K'(:r be the connected component of the identity in K(~). Then, for each ~ R, K(g) (resp. K'(~)) is a maximal compact subgroup of G (resp. of G') ; conversely, all maximal compact subgroups of G (resp. of G') are of that form, and are transforms of one another by inner automorphisms of G (resp. of G'). Moreover, R is isomorphic both to G/K(~) and to G'/K'(~) for each ~ ~ R. We have Mready seen that R is isomorphic to G/K(:r ; therefore, by L e m m a 8, it is also isomorphic to G'/K'(~.). I t is clear t h a t K(~) and K'(~) are closed subgroups of the group of all automorphisms of the vector-space underlying A ; as the elements of K(~) and K'(~) commute with ~, t h e y leave invariant the positive quadratic form tr(x ~ x ) ; therefore these groups are compact. Assume t h a t K'(~) is not a maximal compact subgroup of G' ; then it is properly contained in a compact subgroup K of G'. B y the corollary of Theorem 3, K must be contained in K(fi) for some fi ~ R, and therefore also in K'(fl), so that K'(~) is properly contained in K'(fi). By Prop. 6, fi is the transform of ~ b y some inner automorphisra
444
[1960b] 620
ANDRE W E I L
j(b) of A, belonging to G' ; this implies at once t h a t K'(fl) is the image of K'(~) under the inner automorphism of G' determined by j(b), and therefore that these two groups have the same dimension ; as t h e y are topologically connected Lie groups, this shows t h a t K'(~) cannot be properly contained in K'(fi). Similarly, assume t h a t K(~) is not a maximal compact subgroup of G ; then, just as before, we see t h a t K(g) must be properly contained in some K(fi), which implies t h a t K'(~) is contained in K'(fl) ; therefore, as shown above, K'(~) must be the same as K'(fi). As before, we see t h a t K(fl) is the image of K(~) under an inner automorphism of G ; therefore these groups have the same n u m b e r of connected components, this n u m b e r being finite as we have seen before. Since the connected component of the identity in K(~.) and in K(fl) is the same, viz. K'(~), this shows that K(~) cannot be properly contained in K(fi). On the other hand, the corollary of Theorem 3 shows t h a t e v e r y compact subgroup of G is contained in some group K(~) ; therefore, if it is maximal, it must be of the form K(~) ; and a similar proof holds for G'. 14. In order to prove our last result, we have to consider the Lie algebras of the groups discussed above. With the notations of Theorem 5, the Lie algebras of G and of K(~) will be denoted b y g and b y ~(~), respectively. I f Z is the center of the semisimple algebra A, the group of automorphisms of Z is finite ; this implies t h a t any automorphism of A, sufficiently close to the identity, induces the identity on Z and is therefore an inner automorphism. F r o m this, one easily deduces the well-known fact t h a t the Lie algebra of the group of automorphisms of A consists of the inner derivations
for all u e A. I f u and v are in A, we have D~ = D~ if and only if u--v is in Z ; from this, it follows t h a t e v e r y inner derivation can be written in one and only one way as D u with a(u) = 0, a being as defined in no. 12 ; one m a y therefore identify the Lie algebra of the
[1960b]
445 ALGEBRAS ~VITH I N V O L U T I O N S
621
g r o u p of a u t o m o r p h i s m s of A with the s u b s p a c e of A d e t e r m i n e d b y a(u) = O. I f • is a n y a u t o m o r p h i s m or a n t i m o r p h i s m of A, a n d if e()~) h a s the s a m e m e a n i n g as in no. 6, one sees a t once t h a t 2 - 1 Du 2 is the inner d e r i v a t i o n Du, with u' = e()t) u ~. I n p a r t i c u l a r , .Du c o m m u t e s with ~ if a n d only if z - - u - - e ( ) t ) u z i s in Z ; w h e n t h a t is so, we h a v e z ~ ~(z) = ~ ( u ) -
e(~) ~ ( u ) ~,
which shows t h a t z m u s t t h e n be 0 if a(u) -- 0, or also if ~ is a n inner a u t o m o r p h i s m o f A ; in b o t h t h e s e eases, therefore, D , c o m m u t e s w i t h u if a n d o n l y if u = e(2) u z.
L~MMA 9. A s s u m e that (A, ~) is non-degenerate, and let ~, fl be two elements of R . T h e n ~(~) = t(fl) implies ~ = ft. F r o m the results p r o v e d a b o v e , it follows t h a t t h e Lie a l g e b r a g of G consists of t h e inner d e r i v a t i o n s D~ for a(u) = 0, u' = - - u; let U be t h e v e c t o r - s p a c e d e t e r m i n e d b y t h e l a t t e r conditions. L e t V(~), W(~) be t h e s u b s p a c e s of U consisting of the e l e m e n t s o f U which a r e o d d for ~ a n d e v e n for ~, r e s p e c t i v e l y , a n d let V(fl), W(fi) be defined similarly. T h e n ~(cr ~(fl) consist o f t h e inner d e r i v a t i o n s D v for v e V(~) a n d for v ~ V(fl), r e s p e c t i v e l y . B y o u r a s s u m p t i o n , D~ c o m m u t e s w i t h fl for e v e r y v E V(~.) ; as a(v) -~ 0 for v e V(~), this implies t h a t v ~ = v a n d t h e r e f o r e v e V(fl). T h e r e f o r e our a s s u m p t i o n can be e x p r e s s e d as V ( ~ ) = V(fl). N o w t a k e w e W(~) ; as w is e v e n for ~r L e m m a 1 of no. 6 shows t h a t , w i t h r e s p e c t to t h e bilinear f o r m tr(xy), w is o r t h o g o n a l to all t h e o d d e l e m e n t s for ~, a n d in p a r t i c u l a r to all the e l e m e n t s o f V(~), i.e. of V(fl). On t h e o t h e r h a n d , as w is o d d for ~, it is o r t h o g o n a l to all e v e n e l e m e n t s for ~. N o w let B be t h e space of o d d e l e m e n t s for fl ; for a n y b e B, w r i t e : b0 = 89 -~- b~), b 1 = 89 - - b'), b~. = ~(bl), b~ = b 1 - - b 2. T h e n we h a v e b = b o -}- b~ + b 8 ; b 0 is e v e n for ~. As fl c o m m u t e s with ~, b 1 is o d d for fl ; t h e r e f o r e b~ is in V(fl). Moreover, b~ is o d d for fl a n d is in t h e center, so t h a t it is o d d for ~ (by P r o p . 1 of no 7).
446
[1960b] 622
ANDRE WEIL
Thus bo, b~ a n d b a are all orthogonal to w. This proves t h a t w is orthogonal to e v e r y b ~ B ; b y L e m m a 1 of no. 6, w m u s t therefore be even for ft. Thus W(~) is contained in W(fl); exchanging with fl, we see t h a t W(cr = W(fl). As the odd a n d the even elements for u a n d for fi in U are the same, it follows t h a t ~ and fl coincide on U, or in other words t h a t ~ - 1 fi induces the identity on the Lie algebra g of G. Now, b y Prop. 6, ~ - 1 fl is an inner a u t o m o r p h i s m of G' ; since it induces the identity on the Lie algebra of G, it m u s t therefore be in the center of the algebraic connected component of the identity G Oin G. B u t we h a v e a s s u m e d t h a t (A, c) is non-degenerate; and it follows f r o m T h e o r e m 2 of P a r t I that, when t h a t is so, G o is semisimple with a center r~duced to the identity. This completes our proof.
Notations and assumptions being as in Theorem 5, assume also that (A, ~) is non-degenerate. Then, for ~ e R, K'(~) is its own normalizer in G' ; K(cr is the normalizer of K'(~) and is its own normalizer in G; and the mappings ~ - + K'(cr ~-+ K(~) are bijeetions of .R onto the sets of maximal compact subgroups of G' and of G, respectively. T~EOREM 6.
The latter s t a t e m e n t follows at once f r o m L e m m a 9. Now, for e R a n d A ~ G, p u t fi --- A-1 :cA ; A t r a n s f o r m s K ' ( a ) into K'(fl) ; therefore, if A t r a n s f o r m s K'(~) into itself, L e m m a 9 shows t h a t we m u s t have ~ ~ fi, i.e. t h a t A m u s t be in K(~). This proves the theorem. 15. I f (A, ~), when decomposed into simple components, has no s u m m a n d on which ~ induces a positive involution, the group G' has no c o m p a c t factor, so t h a t the space R -- G'/K'(~) is the R i e m a n n i a n s y m m e t r i c space a t t a c h e d to the g r o u p G'. I n fact, it follows f r o m the results of P a r t I t h a t one can obtain in this m a n n e r all the R i e m a n n i a n s y m m e t r i c spaces a t t a c h e d to the semisimple groups which h a v e no c o m p a c t factor and no factor isomorphic to a n exceptional Lie group. Unfortunately, the latter are (for the t i m e being, a t least) still beyond the scope of the m e t h o d discussed in this paper.
[1960b]
447 ALGEBRAS W I T H I N V O L U T I O N S
623
Finally, we o b s e r v e that, if a positive i n v o l u t i o n ~ is chosen in t h e space R, a n d all o t h e r p o i n t s o f R are w r i t t e n as fl = ~j(a) w i t h a e PI(~), a' = a -1, the i n v a r i a n t m e t r i c in R is e x p r e s s e d b y
ds ~ = t r ( a - 1 da. a - 1 da), a n d t h e geodesic joining ~ to fl = ~j(a) ~ j ( a p) for p e R. The Institute for Advanced Study Princeton, :N'ow Jersey
consists o f t h e p o i n t s
[ 1960c] On discrete subgroups of Lie groups
1. L e t G be a topological g r o u p and F an a r b i t r a r y group; one m a y t h i n k o f F as being provided w i t h the discrete topology. Consider the space G (r) of all m a p p i n g s of F into G; this is the same as the product lIve~Gv, w h e r e Gv is the same as G for e v e r y 7 e F, and will be provided w i t h t h e usual p r o d u c t topology. The set .~ = ~ ( F , G) of all r e p r e s e n t a tions of F into G m a y be described as the subset of G (r~, consisting of all t h e m a p p i n g s r of F into G which s a t i s f y r(77') ~ r(~/)r(7 ') for e v e r y pair 7, 7' of e l e m e n t s of F; this is a closed subset of G (F) and will be provided w i t h the topology induced on it by t h a t of G(~); w i t h t h a t topology (the so-called " t o p o l o g y of pointwise c o n v e r g e n c e " ) , ~ will be called the space of r e p r e s e n t a t i o n s of F into G. If F is g e n e r a t e d by a family of e l e m e n t s (%)~ea, indexed by a set A, a r e p r e s e n t a t i o n r of F into G is uniquely det e r m i n e d by the e l e m e n t s r(%), so t h a t t h e r e is a one-to-one correspondence b e t w e e n .9t and a c e r t a i n subset of the set G (A) of all mappings of A into G. More precisely, F is t h e n a homomorphic image of the f r e e group F' w i t h the g e n e r a t o r s (7")~ea; let ~ be the homomorphism of F' onto F which maps 7" onto % for e v e r y a e A; let A' be the kernel of q~; the elem e n t s of A', being e l e m e n t s of F', are " w o r d s " w(7') in the 7", and we have then, for e v e r y such " w o r d " , w(7) = e, w h e r e z is the n e u t r a l elem e n t of F; these are the " r e l a t i o n s b e t w e e n the g e n e r a t o r s % of F " . The space 9~' = ~ ( F ' , G) is t h e n in an obvious one-to-one correspondence with G (A), and it is a trivial m a t t e r to v e r i f y t h a t this is a homeomorphism w h e n G (~) is provided w i t h t h e p r o d u c t topology in a m a n n e r similar to t h a t described above. T h e n ~ is in an obvious one-to-one correspondence w i t h t h e closed subset ~, ef ~', consisting of all the r e p r e s e n t a t i o n s of F' into G which m a p A' into the n e u t r a l e l e m e n t e of G; and it is again a trivial m a t t e r to v e r i f y t h a t this is a homeomorphism. We will i d e n t i f y .~' w i t h G <~), and ~ w i t h ~ , by m e a n s of t h e s e correspondences. L e t us a s s u m e f u r t h e r t h a t (wo)ae~ is a family of e l e m e n t s of A', such t h a t A' is g e n e r a t e d by the w~ and by t h e i r images u n d e r all i n n e r a u t o m o r p h i s m s of F'; if we w r i t e each such e l e m e n t as a " w o r d " w~(7), we say t h e n t h a t the relations b e t w e e n t h e % are " g e n e r a t e d " by the " f u n d a m e n t a l relat i o n s " w~(7) = ~; and ~ , as a subset of ~ ' = G (~), is t h e n the set of the e l e m e n t s (g~) of G c~) which s a t i s f y t h e relations w ~ ( g ) = e. P a r t i c u l a r i n t e r e s t a t t a c h e s to those r e p r e s e n t a t i o n s r of F into G which 369 Reprinted from Ann. of Math. 72, 1960, by permission of Princeton University Press.
449
450 370
[1960c] ANDRE WEIL
are injective (or, as one also says, faithful) and such t h a t r(F) is a discrete subgroup of G with compact quotient space G/r(F); of course G has then to be locally compact. Clearly a representation r has these properties if and only if there are a neighborhood U of e in G such t h a t r - l ( U ) = {z} and a compact s u b s e t / ~ of G such t h a t G = K . r ( F ) . Let us denote by ~0=~0(F, G) the set of all such representations, considered as a subset of :~. It has been conjectured by A. Selberg that, if G is a semi-simple Lie group, .~0 is an open subset of ~J~'. This will now be proved for all Lie groups. More precisely, the following theorem will be proved:
Let G be a connected Lie group; let F be a discrete group, and ro an injective representation of I' into G, such that r0(F) is discrete in G with compact quotient-space G/ro(F). Then there are a neighborhood U of e in G, a compact subset K of G, and a neighborhood 1I of ro in the space ~ of all representations o f f in G, such that r - l ( U ) = {~} and G = K . r ( F ) for all r c ~. Moreover, L' has then a finite set of generators with a finite set or f u n d a m e n t a l relations. The last s t a t e m e n t is included only for the sake of completeness, since it is an immediate consequence of the fact t h a t the fundamental group of any compact manifold has the property in question (incidentally, a proof for this is included in what follows) and t h a t the same is t rue of the fundamental group of any connected Lie group. In view of these facts, !~ can be described as has been done above, using a finite set of generators ( % ) a n d a finite set of fundamental relations (we); then the space denoted above by ~ ' is the product of finitely many Lie groups, isomorphic to G, and is therefore a connected real-analytic manifold; ~{ is the subset of .~' defined by the finitely many real-analytic relations w~(g) = e, and is t h e r ef o r e a real-analytic subset of ~ '. The theorem stated above implies, as we have said, t h a t the subset of ~ which has been denoted by ~o is open on ~.)~;if G is the hyperbolic group (the quotient of SL(2, R) by its center), it is known t ha t ~J~0is actually a manifold; the question naturally arises w h e t h e r this is true for ~0, or r a t h e r for each connected component of ~0, whenever G is a Lie group. 2. As will be seen in no. 11, the general case of our theorem can be reduced to the case when G is simply connected. When G is such, G/ro(F) is a compact manifold V whose f u n d a m e n t a l group is isomorphic to F. We first deal with some purely topological aspects of this situation. In nos. 2-5, we will denote by V any connected manifold (it would be enough for our purposes to assume t h a t V is connected and locally simply connected). L e t V be the universal covering of V, with the projection
[1960c]
451 ON DISCRETE SUBGROUPS
371
p onto V; let P be the f u n d a m e n t a l group of V, considered as a g r o u p of a u t o m o r p h i s m s of l?, so t h a t V can be identified with V/17. We w r i t e x7 for the image of a point x of l? u n d e r an e l e m e n t 7 of I', and e for the n e u t r a l e l e m e n t of P. Assume t h a t (U~)~e~ is a family of connected open subsets of 17, indexed by a finite set I, with the following properties: (a) the sets U~7, f o r i c I, 7 ~ [', are a covering of V (i.e., t h e i r union is I?); (b) f o r every p a i r (i, j) there is at most one element 7 o f [" such that
U~7 meets
~7~.
F o r each i, p u t U~ = p(U~); applying (b) to 4 = j, we see t h a t U~Teannot m e e t U~ unless 7 = s; this means t h a t p induces on U~ a bijeetive mapping, and t h e r e f o r e a homeomorphism, of U~ onto U~. B y (a), t h e sets U~ are a covering of V; call N t h e n e r v e of t h a t covering; this is the set of all subsets J of I, such t h a t Flje~U~ is not e m p t y . In particular, we have {i, j} e N if and only if p(U~) m e e t s p ( U 0 , i.e., if and only if t h e r e is 7 in P such t h a t U~7 meets U~ by (b), w h e n t h a t is so, 7 is uniquely d e t e r m i n e d by t h a t condition; this e l e m e n t 7 will be d e n o t e d by%~. I t i s e l e a r t h a t T ~ = e for all i i n I , and t h a t % 7 ~ = z for all {i, j} e N. A subset J of I is in N if and only if, for some j in J, t h e r e are a point x in Uj and e l e m e n t s 7~ of P such t h a t x is in [7,~%,, for all h in J, and t h e n the same is t r u e for all choices of j in J . B u t t h e n we m u s t have %, = 7,,~j for all h in J . T h e r e f o r e J is in N if and only if, for some j in J, all pairs {j, h} w i t h h in J are in N, and the sets UhT,~, for h e J, have a n o n - e m p t y intersection; and t h e n the same is t r u e for all j in J . In particular, t a k e J = {i, j, k} ; this is in N if and only if {i, j} and {i, k} are in N and t h e r e are points u~ in U~, u, in ~ , u~, in U~, such t h a t u~ = ujT~ = u~%~; b u t t h e n u is in /J.j and i n g~Tk~7~.~, SO t h a t we have As V is connected, the n e r v e N i s connected; this means t h a t I c a n n o t be w r i t t e n as t h e union of two disjoint n o n - e m p t y subsets I', I", such t h a t no pair {i', i"}, with i' in I ' and i " in I", belongs to N. In f a c t , if this w e r e not so, the unions of the U , and of the U < would be disjoint n o n - e m p t y open subsets of V. We select some e l e m e n t i0 of I as " o r i g i n " , and, for each i in I, we select a " c h a i n " C~, i.e., a sequence 4o, i1,... , i ~ = i , w i t h the origin i0 and the end-point i~ = i, such t h a t , for 1 < ~ < m, {4~_, i~} is in N. We p u t S~ = 7~0qTq% . . . 7%~_1~ for the chain C~. We t a k e for C~0 the chain io, io, so t h a t S~ = s. Now, for e v e r y {i, j} in N, p u t a~:j = S~%~S2h These elements s a t i s f y the relations ( ~ = G ~ j ~ when-
452
[1960c] 372
ANDRE
WEIL
ever {i, j, k} is in N (which implies, for i = j = k, t h a t a . = e for all i, and, for i = k, t h a t a ~ a ~ = e for all {i, j} in N), and, for every chain C~ = (i0, . . . , i~), the relation a~0~ " ' " a ~ _ ~ = e. It will now be shown t h a t the a u generate F, and that the relations we have j u s t w r i t t e n generate all the relations between them. 3. As to the first assertion, we use the f a c t t h a t the nerve N-of the covering of V by the U,7 m u s t be connected. Now a pair {(i, 7), (J, 7')} is in N if and only if U~7 meets U~7', i.e., if and only if {i, j} is in N and 7 ---- 7 ~ 7 ' . L e t 7 be a n y element of F; there m u s t be a chain of N , with the origin (io, 7) and the end-point (i0, e); if this chain consists of the elements (i~, 7~), with 0 -_< l~ _-< m, we m u s t therefore have i0 -- i~, 70 = 7, 7~ = e, and, for 1 __