This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
II(u), of CU onto the set of all subsets of II such that (5.4.4)
d(U)
n d_ =
{
0 d(n*) = -d(n). d( gl)
for all a e II(u)} for all a e II(n)}
5.5. Let Z ~ 1)# C 1)' be the set of all integral linear forms on 1). We recall that p. e Z if and only if 2(p.,
End Vf . a basis of A m and if we put EndAm*@V.I. E Z be defined by (5.1.6)
be the unique, up to equivalence, irreducible representation of G1 having ~ as an extremal weight. Thus if ~u ~2 € Z then !.IP and !.I~2 are equivalent if and only if there exists a € WI such that a~1 = ~2. Now let (5.5.1) m 1 = m n gl so that d(m1) is defined, and
227
GENERALIZED BOREL-WElL THEOREM
351
(5.5.2) is a disjoint union where, as in § 5.3, n = uo. Define (5.5.3) The elements of D, will be called dominant (with respect to m,). One knows that D, is a fundamental domain for the action of W, on Z so that every irreducible representation of G, is equivalent to ).Ii for one and only one g ED,. If).l, is an irreducible representation of G" the unique weight that is both dominant and extremal is called the highest weight of ).I,. Thus for any g E D" g is the highest weight of ).Ii. Also if V, is a representation space for G, we will call a weight vector in V, extremal (resp. highest) if it lies in an irreducible comr;onent of V, and there corresponds to an ex-· tremal (resp. the highest) weight. Now if (5.5.4)
(3: G,
->
EndC
is a representation of G" define Co ~ C as the space of all vectors in C which transform according to the irreducible representation).li of G. Thus (5.5.5) is a direct sum. A representation of a Lie group induces a representation of its Lie algebra. We will, throughout, use the same letter, in this case (3, to· denote this corresponding representation of its Lie algebra. Now put and let
Cin, =
Cm, = {s E CI(3(z)s = 0 for all Z Em,} , Cm, n C". Then one knows that
Cm, = E"ED, Ctit, . and that the set of non-zero elements in Ctit, is the set of all highest weight vectors in Co. Thus Ctit 1 is the weight space for the weight g in the subrepresentation (31 Cli and (5.5.6) dim Ctit, = multiplicity of ).Ii in (3 . The above notation without the subscript 1 refers to the case when 9 is substituted for 9,. Note that D ~ D, and that W, is a subgroup of W. Furthermore if we put then (5.5.7) where (5.5.8)
228
352
BERTRAM KOSTANT
For each cp E a let for any x' E 1}'
1:'1' E
W be the reflection corresponding to cp so that 1:rpx' = x' - z(x', cp) cp •
(5.5.9)
(cp, cp)
We now observe LEMMA 5.5. Let g2 E 1}' be defined by (5.5.8). Then (g2' cp) = 0 for all E ~(gt) so that X U2 E 1} lies in the center of gt. PROOF. It follows immediately from Proposition 5.3 or Proposition 5.4 that ~(gl) + a(n) ~ a(n) .
cp
Thus if cp e ~(gl) it follows from (5.5.9) that ~(n) is stable under 1:'1'. But then obviously 1:rp(g2) = g2. Hence, again by (5.5.9), (g2' cp) = O. q.e.d. REMARK 5.5. Since the elements 1:"" cp e a(gl), generate WI it follows from the proof above that ~(n) is stable under any 1: E WI. 5.6. Let (3 be given by (5.5.4). Let Rfl E End C be the Casimir operator corresponding to the restriction of (g) to gl. Let I x' I denote the length of a vector x' E 1)# with respect to the restriction of ('/) to 1)# • We recall the following well known proposition. PROPOSITION 5.6. Let (3 be given by (5.5.4). Then, for any reduces to the scalar on the subspace C< of C. PROOF. Writing (3(e",)(3(e_rp) Rfl
=
E:~l(3(xn(3(x,)
~e
Du Rfl
= (3(xrp) + (3(e_rp)(3(erp) it follows that
+ 2(3(x Ul) + 2 E",EA(mll (3(e_rp)(3(erp)
,
where x" 1 ~,i ~ l, is an orthonormal basis of 1). One knows that Rfl reduces to a scalar on C< so that to find the scalar it is enough to restrict Rfl to Gtit l. But since (3(e_",)(3(e",) vanishes on Gml the proof follows immediately. q.e.d. 5.7. We now return to the considerations of § 4. We now assume, however, that the representation}') of g is the irreducible representation },)A where A, '= D, so that V = VA. Let U E cVand now let a = n where n is the nilpotent Lie summand uo• Thus 7C = },)A I n. We wish to determine the s)ectral resolution of the laplacian L" on the cochain complex C(n, VA). Now since g,lies in the normalizer of n it follows that A n is stable under the representation e I gl of gl. Let
229
GENERALIZED BOREL-WElL THEOREM
353
(3: g1 --> End C(u, VA) (5.7.1) be the representation of g, on C(u, VA) formed by taking the tensor product ()f ].1.1. [ g1 and the representation of g1 on Au' contragredient to the repre:sentation of g1 on Au defined by restricting e [g1 to Au. For any Z e g1 it is then obvious that, as mappings from C(u, V) into Ag ® V,
{5.7.2) 1J ® l·(3(z) = e~(z).1J ® 1. Since (3 clearly arises from a representation of G1 on C(u, V) we may, :as in § 5.5, form the decomposition (5.5.5). The spectral resolution of the laplacian L" is given in THEOREM 5.7. Let A, e D and let ].1.1. be an irreducible representation of .g, with highest weight A" on a vector space V.I.. Let u be any Lie subalgebra of 9 which contains the maximal solvable Lie subalgebra b of g. Let u be the maximal nilpotent ideal of u (see Proposition 5.3). Let 11: = ].1.1. [ U and let L" be the laplacian on the cochain complex C(u, VA) defined as in § 3.5 with a = u and V = V.I.. Let g1 = un u* (see § 5.3) and let (:3 be the representation of g1 on C(u, VA) defined following (5.7.1). Write, as in § 5.5, C(u, VA) = L:<ED1 C(u, VA)< where C(u, vAy is the set of all vectors in C(u, VA) which transform under
fJ according to the irreducible representation ].Ii of g1. Then L" reduces to a scalar on C(u, vAy and the scalar is ; ([ g
+ A, [2 _
where g =
[g
+L:4'E~+
+ g [2) rp
and Ll+ is defined as in § 5.4. PROOF. Since u is a Lie summand we can apply Theorem 4.4. But by Prorosition 5.3, u-L = u* + g1. Thus we can choose the Zj of Theorem 4.4 :so that for i ~ m, Zj = elP for some rp e Ll(u) and for j > m, Zj either lies in u* or g1. Furthermore if Zj E u* then by Proposition 5.4 we may assume Zj = e- IP for some rp E Ll(u). Now apply Theorem 4.4. Then since e; = e_ 1P it follows from (5.7.2), and the definition of RfJ that for any p e C(u, VA) (1J ® l)L"p
= +(1 ® R'A + L:IPHcnl e~{e_lP)e(elP) -
L:IPHCnl e~(elP)e~(e_IP»)(1J
= ~ (1 ® W
A -
® l(p») - ~1J ® II RfJp)
2e~(Xg2»)1J ® l(p) - ~1J ® l(RfJp) ,
where g2 is given by (5.5.8).
230
354
BERTRAM KOSTANT
But 1 ® R~\ by Proposition 5.6, reduces to the scalar I g + >"1 2 _ I g 1:10 on Ag ® V. Thus since X U2 e g, we can apply (5.7.2) once more and obtain (5.7.3)
L"
= ~(I g + >"1 2- I g 12)1 -
(;3(x U2 )
+ !RIl)
where 1, here, denotes the identity operator on C(n, VA). But XU2 lies in the center of g, by Lemma 5.5. Hence L" reduces to a scalar on C(n, yAy. To determine the scalar it suffices to compute L" on a highest weight vector p e C(n, VA)<. But then since p belongs to the weight t; it follows from Proposition 5.6 that L" reduces, on C(n, yAy, to the scalar
+(1 g + >"1 2- I g 12) - (g2' t;) + (gu t;) + ! = +(1 g + >"1 2- I g + t; 12) , since g = g, + g2' q.e.d.
It;
12) ,
Now, as one easily shows, ;3(z) for any z e gu commutes with both d. and d:. In fact since ;3(z)* = ;3(z*) (by 5.7.2) this is implied by Lemma. 4.2, (4.3.4) and (5.7.2). Thus if we consider the orthogonal direct sum decomposition (see Remark 2.3) (5.7.4)
C(n, VA)
= Imd" + lmd: + Ker L"
,
it follows that each of the three subs paces of C(n, VA) appearing on the right side of (5.7.4) is stable under ;3(z) for all z e g, and hence induces sub-representations of;3. Since d" maps 1m d:, bijectively, onto 1m d" it follows that the sub-representations of ;3 defined by 1m d" and 1m d: are equivalent. Now since ;3(z) commutes with d" for all z e g, it follows that ;3 induces. a representation 13: g, -> End H(n, VA) of g, on the cohomology space H(n, VA). On the other hand it is obvious that 13 is equivalent to the sub-representation of ;3 defined by Ker L". But then since L" is positive semi-definite we obtain, immediately, the following corollary of Theorem 5.7. 5.7. Let t; e D,. Then if the multiplicity of vi in ;3 is positive one must have COROLLARY
Ig+>"I~lg+t;I·
Furthermore if I g + >"1 > I g + t; I then the multiplicity of vi in 13 = 0' and if I g + >..1 = I g + t; I then the multiplicity of vi in 13 = multiplicity' of vi in ;3.
5.7. (A) Another way of expressing the statement in Corollary 5.7 is as
REMARK
231
GENERALIZED BOREL-WElL THEOREM
355
"*
0, then p is a cocycle which is not cohomolofollows. If P E C(n, Vh)<, p gous to zero if and only if I g + )., I = I g + ~ I. If on the other hand p is a cocycle, then p is a coboundary if and only if I g + )., I > I g + ~ I. At a later point we will make important (for us) use of the following fact (contained implicitly in Corollary 5.7). (B) Every irreducible component of jJ is inequivalent to any irreducible -component of the sub-representation of f3 defined by 1m d n • 5.8. Let z+ ~ Z be the semi-group generated by ~+. Writing an element
Y' E Z as a linear combination of simple roots it is clear that Z+ can be characterized by (5.8.1)
Z+
= {'o/ E
Z I (p,
'0/)
~
0 for all P E D} .
Now let)., E D and let ~h denote the set of weights of the irreducible representation ).Ih of g. One knows that if P E Z then a necessary condition for P E a h is that (5.8.2) 'The following lemma is a consequence of this fact. LEMMA 5.8. Let).,H).,2 E D. Let PI
E ~hl,
P2 E
~h2.
Then
(5.8.3)
.and equality holds in (5.8.3) if and only if there exists a
W such that
+ P2 . W be such that T(PI + P2) E D. a(A.1
+ ).,2) =
E
PI
PROOF. Let T E For i = 1, 2, put "fr, = A., - Tp,. Since TP, E ~h£ it follows then from (5.8.2) that "fr, E Z+ and hence "fr E Z+ where "fr = "fr1 + '0/2. Now put P = TPI + TP2 so that P s D. But then ~ + ).,2 = P + "fr. Consequently, since Ipi = I PI + P21, one has
I ).,1 + ).,21 2 = I PI + P2 12 + I"fr 12 + 2(p, "fr) . But by (5.8.1) (p, '0/) ~ O. This proves the inequality (5.8.3). Furthermore if equality holds in (5.8.3) then obviously "fr=0. But since "fr="fr1 +"fr2 and '0/17 "fr2 E Z+, it follows that '0/1 = "fr2 = O. That is, )." = TP .. i = 1, 2. The lemma follows in one direction by putting a = T- 1 • The other direction is obvious. q.e.d. REMARK 5.8. Let the notation be as in Lemma 5.8. Let a E W. Then the proof of Lemma 5.8 also yields the statement (by putting T = a-I) that a(A.I + ).,2) = PI + P2 implies a).,l = PI and a).,2 = 11..
"*
5.9. We recall that an element P E Z is called regular if (p, rp) 0 for all rp E~. One knows that P E Z is regular if and only if ap = p, a E W,
232
356
BERTRAM KOSTANT
implies (] is the identity element of W. We recall that g € D and that g is regular. In fact both of these statements are consequences of the well known relation (5.9.1)
=
(g, a)
(a, a) 2
for any a € II. One obtains (5.9.1) from the easily verified fact that a is the only root in ~+ which "changes sign" under Ta,. That is, T",~_
Consequently REMARK
T",
n ~+ =
(a) .
g = g - a. But by (5.5.9) this is equivalent to (5.9.1).
5.9. Freudenthal has proved (see e.g., [6, 6.1])
(5.9.2)
*
for any A. € D and any P € ~ \ P A.. We observe that, since g is regular and g E D, (5.9.2) follows from Lemma 5.8 by putting A.I = PI = g. We now wish to consider the irreducible representation ),IU of g whosehighest weight is g. Weyl has given a formula for the dimension of a representation in terms of its highest weight. Weyl's formula asserts. that for any A. € D (5.9.3)
This formula generally proves to be quite awkward for computational purposes. However in the special case when A. = g we observe that (5.9.3} immediately yields (5.9.4) where r (= dim m) is the number of roots in ~+. We wish to determine the weights of ),Ig and their multiplicities. For any subset <1> ~ ~+ let <<1» € Z be defined by
<<1» =
E"E(
rp •
Let the elements of ~+ be ordered so that Now observe that if <1> ~ ~+ (5.9.5)
g - <<1»
=
~+
=
{rp,}, i
= 1, 2, ... , r_
1
2(±rpl ± rp2 ± ... ± rp,.)
for some choice of the signs; and that furthermore as <1> runs through all 2" subsets of ~+, then the right hand side of (5.9.5) runs through all 2'choices of signs. It is suggested by (5.9.4) and definition of g that),lu somehow behaves like a spin representation. The analogy is further strengthened by
233
GENERALIZED BOREL-WElL THEOREM LEMMA 5.9. Let feZ. Then f e such that (5.9.6)
~ a if
357
and only if there exists
~ ~+
f = g - (CI».
Furthermore the multiplicity off as a weight of lJU is equal to the number of subsets CI> ~ ~+ satisfying (5.9.6). PROOF. Let 5 ~ Endg be the Lie algebra of all operators on g which are skew-symmetric with respect to (g). Thus s is isomorphic to the Lie algebra of SO(n, C). Since B(z), for z e g, is determined by its restriction to g, we may regard B as a monomorphism mapping g into s. Furthermore we may find a Cartan subalgebra b of s such that B maps ~ into b. Now let B': b' ->1)' be the mapping whose transpose is the restriction of B to 1). It is obvious that if p is a representation of s then B' maps the weights of p into the weights of poB. Now one knows that the non-zero weights of the given representation of s on g are of the form ± \'h i = 1, 2, "', [n/2], where \'11 " ' , \'[n/2] are linearly independent in b'. Furthermore it is clear that we can choose the ordering and signs of the \'1 so that i = 1,2, "', r i = r + 1, "', [n/2].
(5.9.7) Now let
v: 5 - End Vv be the spin representation of s. One knows that dim Vv the weights of v are all elements in b' of the form
t( ± \'1 ± \'2 ± ... ±
=
2[n/2]
and that
\'[n/2]) ,
and that each weight occurs with multiplicity one. Writing n = l + 2r, it follows then from (5.9.5) and (5.9.7) that the weights of v 0 B are all elementsfe 1)' of the form g-(¢-> where CI> ~ ~+ and that the multiplicity of f is equal to 2[1/ 2] times the number of subsets CI> ~ .6.+ such that
f =
g - (4;>.
In particular we note that g is a weight of v 0 B and that its multiplicity is at least 2[1/2]. On the other hand we now observe that every weight vector corresponding to g is necessarily a highest weight vector. To prove this, it suffices to note that if rp e ~+ then g + cp is not a weight of voB. Indeed if it were we would have g + rp = g - (CI» or rp + (CI» = 0 for a subset
234
358
BERTRAM KOSTANT
~ A+. But this is impossible since Z+
the multiplicity of
J.)g
in u 0 0 is at least
n-
2[1/2].
= O.
This proves that But from the identity Z+
it follows from (5.9.4) that J.)g occurs exactly 2[1/2] times in u 0 0 and that no other irreducible representation of 9 occurs in u 0 O. The lemma then follows from the statement above concerning the weights of u 0 O. q.e.d. 5.10. Let a E W. Define the subset <1>" <1>"
=
aA_
~ A+
n A+
by putting
•
It then follows at once that
(5.10.1)
ag
=g
- <<1>,,) •
Since ag E Ag and, being extremal, since it occurs with multiplicity one as a weight of J.)g, it follows from (5.10.1) and Lemma 5.9 that for any subset C A+ (5.10.2) Now one knows (see e.g., [1, 4.9]) that the mapping a . . . . . aA_ is a bijection of Wonto the family of all subsets Ao of A satisfying the two conditions (1) Ao is closed under and (2) A = Ao U -Ao is disjoint union. It follows then that there exists a unique element IC E W such that
+,
(5.10.3) Furthermore one deduces PROPOSITION 5.10. The mapping a -> <1>" of W is a bi;"ectionof Wonto the family of all subsets of A+ which satisfy the condition that and its complement
+.
+.
+.
Ao
=
U -(C) •
Obviously A = Ao U - Ao is a disjoint union. On the other hand it is straightforward to verify that Ao is closed under +. Hence, as noted above,
235
359
GENERALIZED BOREL-WElL THEOREM
aA_ for some unique a € W. But then obviously
=
(5.10.4)
A+ =
is a disjoint union for any a 5.11. For any subset
~
E
W.
A+ (= A(m» denote by
eq,€
Am the element
where
~
A+, form
(5.11.1)
then the elements e_q"
Am*.
be the representation of 1) on Am* @ V.I. obtained by restricting B, 11) to Am* @ V.I.. Let A~ be the set of weights of It is obvious then that if ~ € Z then ~ E A~ if and only if ~ can be witten as
s.
(5.11.2)
where
E
A"'. But then as an immediate corollary to Lemma
5.11. Let ~€ Z. Then ~ E A~ if and only if g +
~
can be written
g+~=f+p.
where f
E
Ag and p.
5.12. For any a
E
E ~.I..
W put ~ 00
(5.12.1)
=
a(g
+ >v) -
g .
Also let SuI<. E V.I. be the extremal weight vector (unique up to a scalar multiple) corresponding to the weight a>v of ).I"'. The following lemma IS the main lemma needed together with Theorem 5.7 to yield the cohomology group H(n, V.I.). LEMMA
5.12. For any
~ E A~
one has
Ig+>VI~lg+~I·
Let a € Wand let ~u be defined by (5.12.1). Then the mapping a ----->~" is a bijection of W onto the set Of all weights ~ of such that
s
236
360
BERTRAM KOSTANT
Furthermore, as a weight of S, EO" occurs with multiplicity one and the weight vector corresponding to EO" is the element of
Am* Q9 V,...
PROOF. It follows immediately from Lemmas 5.8 and 5.11 (putting A.l = g, f1r = f, A.2 = A., f-l2 = f-l) that EO"E d~, that I g + A.I ;S I g + EI for EE d~ and that equality holds if and only if E = EO" for some a E W. Also EO" = Er implies a = T since g + A. is obviously regular. Since e_q,O" Q9 sO"A is obviously a weight vector for EO", to prove the lemma it suffices only to show that the multiplicity of EO" is one. But since we can find a basis of Am* Q9 VA consisting of weight vectors of the form e_q,Q9s,. where f-l E d\ and s". E VA is a corres:r;onding weight vector, it suffices only to show that (5.12.2) imples one has
=
<1>0" and f-l
=
aA.. But now if (5.12.2) is satisfied, then adding g a(g
+ A.) = f + f-l
where f = g - <0")' But then = <1>0" by (5.10.2). q.e.d. REMARK 5.12. A more direct proof of Lemma 5.12 which also does not require the use of a particular case (5.9.4) of Weyl's dimension formula has been found by Cartier. See [4]. The usefulness of such a proof is that it makes the proof of Weyl's character formula and its generalization given in § § 7.4 and 7.5 independent of the particular case (5.9.4). 5.13. Let 1+ E q; and let gi and n be defined as in § 5.3. We isolate a subset WI of W by setting (5.13.1)
WI
= {a E
WI <1>0" ~ A(n)} .
Recalling (5.5.2) and (5.5.3) we observe that the elements of WI can be characterized as follows: REMARK 5.13. Let aE W. Then the following three conditions are equivalent, (1) a E WI, (2) a-I(A(ml )) ~ d+. and (3) a(D) ~ D 1 •
237
GENERALIZED BOREL-WElL THEOREM
361
The following proposition states that WI defines a "cross-section" with respect to the canonical mapping of Wonto the right coset space W1\ W. PROPOSITION 5.13. Every element T € W can be uniquely written where T1 € WI and a € WI. PROOF. Let al> a 2 € WI. Let T1
Remark 5.5, A(n) is stable under
=
T1"
a 1a;;1 and assume T1 € WI. Then by
But this clearly implies
= <1>0"-1 On the other hand the inverse a 2a l 1 also lies in WI. Thus <1>0"-1 2 1 which, by Proposition 5.10 implies a 1 = a2 • Thus no two distinct elements of WI lie in the same right coset of WI. Now let T € W be arbitrary. Let <1>1 = T(A_) n A(m1) and let <1>2 be the complement of <1>1 in A(m1). Then <1>2 = T(A+) n A(m1) so that both <1>1 and <1>2 are closed under Now apply Proposition 5.10 to the case where [gl> g1]' the maximal semi-simple ideal of gl' is substituted for g. It follows then that there exists T1 € WI such that (since A(n*) is stable under T1)
+.
<1>'1
=
<1>1 •
Now put a = TIlT. It is then straightforward to verify a(A_) n A(m1) is empty so that a € WI. q.e.d. It is implicit in the proof above that if T = T 1a is the decomposition given by Proposition 5.13 then (5.13.2) is a disjoint union; the components on the right being also the respective intersections of <1>, with A(m1) and A(n). Now for any a € W put (5.13.3)
n(a)
=
Since, obviously, (5.13.4) note that (5.13.5)
number of roots in <1>0" •
n(a) = n(a- 1)
•
Furthermore if T € Wand T = Ti a is the decomposition given by Proposition 5.13, then it follows from (5.13.2) that (5.13.6)
n(T)
= n(T1) + n(a)
.
REMARK 5.13. Let T € W. We note as a consequence of (5.13.6) that the unique element a € WI in the right coset WIT can be characterized by
238
362
BERTRAM KOSTANT
the statement that n(a) ~ nCr') for all r' e Wr and that equality holds if and only if r' = a. Using (5.13.5) it follows that a similar statement involving the set {a-I}, a e WI, can be made for the left co sets of WI' 5.14. Now for any non-negative integer j put
W(j) = {a e WI n(a) = j} and let
WI(j) = W(j) n WI . Also "let {e~..,}, qJ ~ d(n), be the basis of An' dual to the basis {e..,}, qJ ~ d(n), of An so that by (5.11.1) and (3.2.1) (5.14.1) r;(e~..,) = e_'l> . We can now state THEOREM 5.14. Let u be any Lie subalgebra of 9 which contains the maximal solvable Lie subalgebra {) of 9. Let n be the maximal nilpotent ideal of u (see Proposition 5.3) and let 91 = un u* so that 91 is a reductive (in 9) Lie subalgebra and u = 91 + n is a semi-direct sum (as Lie algebras). Let )., e D and let ).IA be the irreducible representation of 9 on a vector space VA whose highest weight is ).,. Let H(n, VA) be the cohomology group formed with respect to the representation n = ).IA In of n on VA and let lJ be the representation of 91 on H(n, VA) defined as in § 5.7. Now for any [; e DI let H(n, vAy be the space of all classes in H(n, VA) which transform under lJ according to the irreducible representation ).I~ of 91 whose highest weight is [;. Now for any a e W let [;.,. be defined by [;.,. = a(g + ).,) - g .
"*
Then if a e WI one has [;.,. e Dl and for any [; e DIone has H(n, VA)' 0 if and only if [; = [;.,. for some a e WI. Furthermore H(n, VAY.,. is irreducible for all a e WI so that a -> H(n, VAY.,. is a bijection of WI onto the set of all irreducible (under lJ) components of H(n, VA). Moreover degree-wise, for any non-negative integer j Hj(n, VA) = E"'EW11j)H(n, VAY.,. (direct sum) so that for any a e WI, the elements of H(n, VA)'''' are homogeneous of degree n(a). Finally if S"'A e VA is the weight vector for the extremal weight a)., of).lA then the highest weight vector in H(n, VA)'.,. is the cohomology class having e~'l>.,.
®
s"'A
as a representative (harmonic) cocycle.
239
GENERALIZED BOREL-WElL THEOREM
363
*
PROOF. Now by Corollary 5.7 H(n, VA)< 0 if and only if g is a highest weight of an irreducible component of (3 and (5.14.1) Moreover in such a case the multiplicity of ).I; in (3 is the same as its multiplicity in l3. But now the representation (311) of 1) is obviously equivalent to the sub-representation of t (see § 5.11) of 1) defined by the subspace An* Q9 Vof Am* Q9 V. But then by Lemma 5.12 the only weights of (3 which satisfy (5.14.1) are the weights gcr for a e W' and they occur with multiplicity one. Therefore to prove the theorem up to the statement "Moreover ..• ", it suffices only to show that the tT occur as highest weights in the decomposition of (3. But to prove this it is enough to show, for any cp e .1.(m,), a e WI, that gCT + cp is not a weight of (3. Put g = gCT + cpo Then 1g
+ g 12 =
1a(g
+ >v) +
cp 12
=
1g
+ >v 12 +
2(a(g
+
>V), cp)
+
1cp 12 •
But now by Remark 5.13 (3), a(g + >V) e D, so that (a(g + >V), cp) ~ o. But then 1g + g 1 > 1g + >v I. By Lemma 5.12 this implies g is not a weight of t and a fortiori g is not a weight of (3. Now by Lemma 5.12, e'-IP CT Q9 SCTA is the unique (up to scalar multiple) weight vector for the weight gCT of (3. But from above it must be the highest weight vector of an irreducible component of (3. Hence by Theorem 5.7, e'-IPCTQ9sCTA is a harmonic cocycle (element of Ker L~). But then, clearly, its cohomology class is the highest weight vector in H(n, VYCT. Now this class is obviously homogeneous of degreen(a). SinceH(n, VA)
71:
of n on VA be as
E (a(g + >V), cp) ----==--------''------CT€W1IJ) II'7'€alml)
II'7'Ealm , 1 (g,
cp)
PROOF. Let (g), be any non-singular, invariant bilinear form on g and let (1)'), be the bilinear form on 1)' induced by (g),. Now observe that in Weyl's formula (5.9.3) one obtains the same result using (l)'), instead of (1)'). (This is clear since any root corresponds to a simple component of g). Furthermore one need only assume that g is reductive instead of semisimple. But then to determine dimH(n, vAyCT one mayapplyWeyl's formula
240
364
BERTRAM KOSTANT
to the representation )..I;" of 91 using the restriction (91) of (9) to 91< But for any rp € A(m1), (a(g + A,) - g + gu rp) = (a(g + A,), rp) and (gu rp)
since by Lemma 5.5 (g - gu rp) from Theorem 5.14. q.e.d.
=
=
(g, rp) ,
(g2' rp)
= 0. The corollary then follows
5.15. Let w(j) (resp. W1(j» be the number of elements a in W(j) (resp. Wl(j». In general the dimension of H J(n, V") varies with A,. In fact by Corollary 5.14, if m1 =1= 0, by choosing A, properly, it can be made to be arbitrarily large. However if ill1 = 0, that is, if m is substituted for n it was first proved by Bott [2] that dim HJ(m, VA) is constant over all A, € D. In fact he observed that dim HJ(m, VA) = w(j) . (5.15.1) This result of Bott is an immediate consequence of COROLLARY 5.15. Let the notation be as in Theorem 5.14 except that m is substituted for n so that W is substituted for Wl. Then, for any a € W, H(n, VA)"" is one dimensional and in fa~t H(n, VA)""
=
(e~",,,
® s"A»)
where (e~<1>" ® s"A) is the cohomology class defined by the cocycle e~<1>" ® SeTA. PROOF. In the special case of Theorem 5.14 considered here £) plays the role of 91. But since H(n, VA)"" is irreducible under £) and since £) is commutative, it follows that dim H(n, VA)"" is one dimensional. q.e.d.
REMARK 5.15. Observe that a statement generalizing the result (5.15.1) to the case of n involves multiplicity of re;Jresentations rather than dimension. Such a statement is the following: The number of irreducible components in HJ(n, VA) under the action of !:l1 is equal to W1(j) (and consequently is independent of A,). 6. Application I. The generalized Borel.W eil theorem
1. Let u € q; (see § 5.2) and let n (= UD) be the maximal nilpotent ideal in u. Also, as in § 5.3 let 91 = u n u*. Now let U, Nand G1 be the subgroups of G corresponding, respectively to u, n and ~\. The subgroups U and G, we recall, are closed by Remark 5.1. But N is closed also since O(N) is unipotent (see Proposition 5.3). Thus since the center of G1 operates reductivelyon 9 it is clear that G1 n N reduces to the identity and hence
241
GENERALIZED BOREL-WElL THEOREM
365
(6.1.1)
is a semi-direct product. Since n lies in the commutator of u (because ~ c u) it is clear that N maps onto a unipotent linear group under any representation of U. But since N is normal in U it is obvious then that any irreducible representation of U is trivial on N and hence is equivalent to vi, for some t; € Du on G1' Conversely given t; € DI or, more generally, given t; e Z (see § 5.5) the representation vi of GI on VI' extends to an irreducible representation
vi: U -> End V/ of U on VI' by making it trivial on N. Hereafter we will regard vi as so extended. Thus, up to equivalence, all irreducible representations of U are of the form vi for t; € D1' Now, as in § 5.2, let X = G/ U. Then X may be regarded as the base space of a holomorphic fiber bundle with G as total space and U as fiber. Given t; e Z one obtains an associated holomorphic vector bundle E' with fiber Vt as the set of equivalence classes in G x VI' with respect to the equivalence relation (au, s)
==
(a, vi(u)s)
for any a e G, u e U and s e VI'. Let a, beG. If x = bU e X, let a·x e X denote the coset abU. Similarly if veE' is the equivalence class containing (b, s) where s e V/, let a·v e E' denote the equivalence class containing (ab, s). It is clear then that if Xo ~ X is an o,Jen set in X and "fr is a local holomorphic section of Ee defined on a-I·Xo then a("fr), given by a("fr)(x)
= a·"fr(a-I.x)
,
where x e XJ is a local holomorphic section of E' defined. on X)' But now the mapping y. -> a("fr) defines an o)erator p'(a) on H(X, SE') where SEe is the sheaf of local holomorphic sections of Ee and H(X, SE') is the cohomology group over X with coeffbients in SE'. Now from general considerations concerning such cohomology groups one knows that H(X, SE') is finite dimensional. But for any j = 0, 1, ••• , and a € G, it is obvious that HJ(X, SE') is stable under p'(a). We will let pM: G -> End HJ(X, SEe)
be the representation of G (and also g) defined by restricting p«a) to '2 G. It is clear, using Weyl's dimension formula, that a knowledge as to how pJ" decomposes into irreducible representations yields in particular the
HJ(X, SEe) for all a
242
366
BERTRAM KOSTANT
dimension of HJ(X, SEe). We concern ourselves then with the question of decomposing rP. 6.2. Let g, )., € Z. Let PI be the representation of u on Hom (V\ VD. the space of linear mappings from VX into V/, defined by putting (6.2.1) for all y € U and all A € Hom (V'\ VD. Then with respect to this representation one can form the relative cohomology group H(u, gl' Hom (VX, Vl». Concerning this cohomology group and the decomposition of pH, Bott (see [2, 1.6]) has proved PROPOSITION 6.2. Let g € Z. For J. = 0,1, ... let pH be the representation of G on HJ(X, SEe) defined in § 6.1. Then for any)., € Z one has multo of }.ih in pJ'< = dim HJ(u, gH Hom (V,\ VD). In the next section we will put Proposition 6.2 in a somewhat simpler form (Proposition 6.3) expressing it as a reciprocity law.
6.3. Now one knows that for any g € Z the representation }.i~< is equivalent to the representation contragredient to }.ii. Without loss of generality therefore we will, from now on, assume that, for any g € Z, V l -< is in fact the dual space of V/ and }.i~< is the representation contragredient to }.ii. For any g € Z the unique extremal weight of }.ii lying in - Dl will be called the lowest weight of }.if (corresponding weight vectors are called lowest weight vectors). Thus for any g € DIone has that -g is the lowest weight of }.i~<. (In this section it will be convenient to use -Dl (as we may) instead of Dl to index the irreducible representations of Gl). Substituting G for Gl the conventions made above will hold also when }.ih is substituted for }.ii. Let)., € D and g € Dl" Now in the usual manner we may identify Hom (V-'\ Vl-<) with VX 0 v-e. It is clear then that PH for the values -)." -g, is equal to the tensor product of}.ih I u and }.i~<. On the other hand let }.il: u --> End VI be any representation such that }.ill glis completely reducible and letH(u, VI) be defined with respect to }.il I u. Then where the representation
(31: gl --> End H(u, VI) is defined in a manner similar to the definition of (3 in §5.7 and HJ(u, Vlt is the set of all elements in HJ(u, VI) transforming under (3, according to the zero representation of gl it is a simple and well kown fact (see e.g., [2, Corollary 2, p. 223] or [7, p. 603]) that
243
GENERALIZED BOREL-WElL THEOREM
367
(6.3.1) for j = 0, 1, .. " . Now putting V 1 = Hom (V-A, V-<) and ],,11 equal to the tensor product of ]"I~ III and ).11-< and recalling that ]"II < I n is trivial, it follows that (6.3.2)
Hl(U, VI) = (H(n, VA»)
®
V 1-< •
Now let ljJ,~: gi
-->
End HJ(n, V~)
be the representation of gi on HJ(n, V~) defined by restricting (3 (see § 5.7) to HJ(n, V~). It then follows from (6.3.1) and (6.3.2) that (6.3.3)
dim HJ(u,
gH
VI)
= multo of
]"Ii
in (3J,~ .
Substituting - A, for A, and -g for g, Proposition 6.2 becomes the following reciprocity law. PROPOSITION 6.3. Let j be a non-negative integer. Let g € DI and let pJ'-< be the representation of g on HJ(X,. SE-') defined as in § 6.1. Let A, € D and let (3J,~ be the representation of gl on HJ(u, V~) defined above. Then
multo of ).I-~ in pi'-< = multo of
]"Ii
in (3j,~ .
REMARK 6.3. The proof of Proposition (Bott) 6.2 may be simplified considerably. In fact after making a few simple observations the proof of Proposition 6.2 or rather more directly Proposition 6.3, follows almost immediately from a theorem of Dolbeault. We will sketch the arguments. Let K ~ G be the subgroup of G corresponding to f. Let G""(K) be the space of all infinitely differentiable complex valued functions on K. Now let 1J.t and ]"IE be the representations of K on G""(K) defined by
(]"IL(a)f)(b) = f(a-Ib) and
(]"IE(a)f)(b) = f(ba) , where f € G""(K) and a, G € K. Now the representation ]"IR induces (by differentiation) a representation of f on G""(K) and by complexification a representation (6.3.4) of g on G""(K). Now let p: G --> X
244
368
BERTRAM KOSTA NT
be the canonical mapping. One knows (since 9 = f + 11) that p maps K onto X so that p induces a diffeomorphism of K/Kl on X where Kl = Un K. Note that by definition of 91 (see § 5.3) one has that 91 is the complexification of fl where fl is the Lie algebra of K 1 • It follows therefore that if ~ e D" and ).11: 91 -> End (C~(K) ® Vl~') is the representation defined by taking the tensor product of ).IR I 91 and ).1;-', and (C~(K) ® Vl~')O is the set of all elements in C~(K) ® V)~' transforming under).ll according to the zero representation of 9" then (C~(K)® V;-,)O is canonically isomorphic to the space of all C~ cross sections of E~'. This fact leads immediately to the Frobenius reciprocity law. But now one has the following. Let a e K and let qa be the mapping of 9 onto the complex tangent space to K at a induced by ).IR (6.3.4) and let P .. be the mapping, induced by p, of the complex tangent space to K at a onto the complex tangent space to X at p(a). One then observes that the composition Paoqa maps n bijectively onto the set of all anti-holomorphic tangent vectors at p(a); that is, onto the space of all complex tangent vectors at p(a) which are orthogonal to the space of all holomorphic 1-covectors at p(a). It follows then that if 7r: R :
n -> End
(C~(K)
® V
I -,)
is the representation defined by taking the tensor product of).lR I n and the trivial representation, and if (3R: 91 -> End C(ll, C~(K) ® VI') is the representation of 91 on the cochain complex C-< = C(u, C~(K)® VI-i) (formed with respect to 7r:R ) defined in the same way as (3 of § 5.7 (except that).ll replaces).l>" I 91) then, more generally for any j, (CJ,~,)O is canonically isomorphic to the space CO,l(X, E~') of all C ~ differential forms of type (O,}) on X with values in E~'. Here (Cl,~')O is the space of all homogeneous elements of degree j in C~, which transform under (3R according to the zero representation of 91' (We say more generally since if j = 0, this statement is identical with the one made above concerning (C~(K) ® V,-7). Moreover if dR: (Cl,~')O -> (CJ 11,~<)O is the mapping induced by the coboundary operator on C -', then under the isomorphism (CI,~,)O -> CO,i(X, E~<), i = j, j + 1, one also observes (and this is the key observation) that dR corresponds to the usual coboundary operator d" on CO,l(X, E~'). It follows then from the reductive properties of the action of 91 that one obtains an isomorphism
245
369
GENERALIZED BOREL-WElL THEOREM
(6.3.5) where the superscript 0 is defined with respect to (3B, and (3B is defined in a manner similar to (3 of § 5.7. On the other hand by Dolbeault's theorem one has the isomorphism (6.3.6)
HO,J(X, E-<)
->
HJ(X, SE-<)
so that (6.3.5) and (6.3.6) yield the isomorphism (6.3.7)
(HJ(n, C=(K)
® V1-<»)O -> HJ(X, SE-<) •
But now the representation).lL of K on C =(K) extends to a representation ).IL ® 1 of K on C=(K) ® V 1-<. Since).lL ® 1 obviously commutes with ).11 and 7T:B , it induces a representation pf'-<: K -> End (HJ(n, C=(K) ® V1 <»)O . One then observes that under the isomorphism (6.3.6) pf'-< corresponds to the representation pJ,-< I K of K on HJ(X, SE-<) (see § 6.1). Now one proceeds in a manner similar to that used in the proof of the Frobenius reciprocity law. Using the Peter-Weyl decomposition of C=(K) one easily establishes an isomorphism (6.3.8) where if p-A is the representation of K on the summand V-A
® (HJ(n,
VA)
®
V;-
formed by taking the tensor product of ).I-A I K on V-A and the trivial representation of K on (HJ(n, VA) ® V1-<)O, and p is the representation of K on the right hand member of (6.3.8) formed by taking the direct sum of the p-\ then p corresponds to p~,-<. But then Proposition 6.3 follows from the obvious fact, observed before, that dim (HJ(n, VA) ® Vl-<)O = multo of ).I~ in (3J,A • 6.4. Let ~ e Dp One knows that HO(X, SE-<) is just the space of all holomorphic cross-sections of E-<. Now assume that It = b, so that X = Y and D1 = Z. In this case E-< is ,a line bundle over X. It follows from a well known theorem of Kodaira on positive line bundles that if ~ e D then for all j
>0 .
But then applying Hirzebruch's formulation of the Riemann-Roch theorem to the case at hand one obtains that (6.4.1)
246
370
BERTRAM KOSTANT
(note that on the right hand side the subscript 1 is absent). If pO.-f is the representation of G on HO( Y, SE-f) defined as in § 6.1, it is then suggestive from (6.4.1) that pO.-f is equivalent to ),i~<. It is the assertion of the theorem of Borel-Weil that this is in fact the case. Now return to the case where u e q; is arbitrary. (Here again one can still show, without the use of cohomology or sheaf theory, that if ~ e D then the representation pO.~f of G on the space HO(X, SE-f) of holomorphic sections of E-f is equivalent to ),i-f). We now consider the general situation where ~ e DJ is arbitrary (so that the Kodaira theory is not applicable) and the nature of pj.-f is sought for arbitrary i. For any A. e D and a e W let ~(A., a) e Z be defined by putting ~(A.,
a) = a(g
+ A.) -
g .
In § 5.12, since A. was regarded as fixed, we denoted this element, more simply, by ~... We now isolate a special subset D~ of DI (and thereby, by our indexing, isolate a special family of representations of gl). Let Dt be defined by putting Dt = {~ e Dig + ~ is regular (see § 5.9)}. We first observe LEMMA 6.4. Let WI be defined by (5.13.1). Then the mapping D x WI __ Z given by (A., a) -- ~(A., a) ,
where A. e D and a e WI, maps D x WI biiectively onto Di'. PROOF. Let ~ e Dp Since 9 e D ~ DI it is obvious that g + ~ e D I. But now if ~ e Dt then there exists a unique a e W such that a-I(g + ~) e D.
Furthermore by Remark 5.13 (3) it is clear that a e WI. Moreover since a-l(g + ~) e D is regular, it follows from (5.9.1) that A. also lies in D where A. is defined by A. = a-I(g
+ ~) -
g .
But then obviously ~(A., a) =~. Moreover the uniqueness of a obviously shows that ~(A.', a') = ~ implies A. = A.' and a = a' if A.' e D. It suffices only to prove ~(A., a) e Dlo for all A. e D and a e WI. But by Theorem 5.14, ~(A., a) e DI (one can easily give a simpler and more direct proof of this fact). On the other hand if ~ = NA., a) then g + ~ = a(g + A.) and since g + A. is obviously regular, it follows also that g + ~ is regular so that ~ eDt. q.e.d. REMARK 6.4. Lemma 6.4 should perhaps be viewed in the following
247
GENERALIZED BOREL-WElL THEOREM
371
light. If E e Dlo then writing ~ = E(;\" 0') we observe that E, and hence also the representation ),Ii of gu picks out in this way a unique;\, € D and hence a special representation l.I~ of 9 and also a unique 0' E WI and hence, in particular, a special integer n(O'). Note also that D ~ Dt ~ DI and that if EE D then upon writing E= E(;\" 0') one has;\, = Eand 0' is the identity element of W. After applying the Riemann-Roch theorem in the general case considered above, the following generalization of the Borel-Weil theorem was conjectured by Borel and Hirzebruch. It was then later proved by Bott [2, Theorem IV']. THEOREM 6.4. Let
E€ D
I •
Then if E¢ Dlo one has for all ,j = 0, 1, .••
If ~
€ Dlo, then upon writing (uniquely, see Lemma 6.4) E = E(A., 0') where D and 0' E Wl one has H J(X, SE-<) = 0 for all j 1= n(a) and for j = n(O') one has dim Hn(C1')(X, SE-<) = dim V-~
;\, E
where in fact if pn(C1'),-E is defined as in § 6.1, then p"(C1'),-f is equivalent to the irreducible representation l.I-~ of G. PROOF. We have only to apply Proposition 6.3, Theorem 5.14 and Lemma 6.4. That is, if ~ ¢ D,o then by Lemma 6.4 and Theorem 5.14 the multo of ),I; in 13J,A' equals zero for all ;\,' € D and all j. It follows then from Proposition 6.3 that ),I-A' has zero multiplicity in pJ,-f for all j and ;\,' E D. Hence pJ,-E is the zero representation for all j. This proves the first statement. Similarly if E = E(;\', 0') e Dlo then Lemma 6.4 and Theorem 5.14 assert that the multiplicity of l.If in 13J,A' is zero for all ;\,' E D and all j unless both i = n(O') and ;\" = ;\, in which case the multiplicity is one. The theorem then follows from Proposition 6.3. q.e.d. 7. Application II. Weyl's character formula and ita extension to non-connected groups
1. In this section let U be any (not necessarily connected) complex Lie group. Let n be the Lie algebra of a normal connected Lie subgroup of U. Let, for any a E U, (7.1.1) /3~(a) e End Rn' be the inverse transpose to the automorphism of n induced from conjugation by a. Furthermore for j = 0, 1, ... , dim n let
248
372
BERTRAM KOSTANT
(3t: U -> End
Nu'
be the representation of Uon Nu' formed by takingthej'h exterior product of the representation defined by (7.1.1). Now for any a e U put x~jl(a)
= trace (3t(a) ,
and let
E7=o( -1)J x6 jl(a) .
xo(a) =
One, of course, knows that (7.1.2)
Xo(a)
Now for any subset U'
= det (1 -
~
R(U')
(3~(a») •
U, let R( U') ~ U' be defined by
=
{a e U' I Xo(a)
"* O} •
Note that, by (7.1.2), R( U) is the set of all a e U such that (3~(a) has no non-zero fixed vectors. 7.1. Although we make no use of the fact, itcan be easily shown that if R( U) is not empty then u is necessarily a nilpotent Lie algebra. REMARK
7.2. Now let I):
U-End V
be a representation of U on Vand, for any a e U, let x>(a) = trace I)(a) .
Our intention now is to give a formula for the character homology groups defined by u. Let (3J: U -> End
til involving co-
Nu' ® V
be the tensor product of the representations (3t and character of (3J one obviously has
I).
Thus if
t
Jl
is the
(7.2.1)
Let 'lC = j.) I u. Then we recall that Au' ® V is the underlying space of the cochain complex C(u, V) defined by 'lC. Furthermore if d n is the corresponding coboundary operator, then it follows easily that for any a e U (7.2.2)
on CJ(u, V). Since (7.2.2) holds also for j - 1, (3J induces a representation
249
GENERALIZED BOREL-WElL THEOREM
373
'$J: U --> End HJ(n, V)
of U on the cohomology group HJ(n, V). Now let X(J) be the character of '$J and put
X=
E (-l)JX(J)
x=
E(-l)Jt i ).
•
Similarly put It is then a simple and well known fact (the Euler-Poincare principle) using (7.2.2) that for any a e U
(7.2.3)
x(a)
=
x(a) .
Let Xoequal Xfor the case when the identity representation is substituted for 1). It follows then from (7.2.3) that also
(7.2.4)
Xo(a)
= Xo(a) .
But now taking the alternating sum with the expressions in (7.2.1) as summands one obtains (7.2.5)
X"Xo = X •
We have proved, using (7.2.3), (7.2.4) and (7.2.5). PROPOSITION 7.2. Let 1) be a representation of U on a vector space V and X' be its character. Let n be the Lie algebra of a normal Lie subgroup of U and let X(resp. Xo) be the alternating sum of the characters of the representations 13i (resp. 130 of U on HJ(n, V) (resp. HJ(n». Let R(U) be the set of all a e U which, (see Remark 7.1) under the representation of U on n induced by con}ugacy, correspond to operators on n without non-zero fixed vectors. Then if a e R( U) one has Xo(a) =1= 0 and
X'(a) If a ¢ R( U) one has Xo(a)
= ~(a) . Xo(a)
= x(a) = O.
7.3. Let u e V (see § 5.2). We apply Proposition 7.2 to the case where U is the subgroup of G corresponding to u and n (= un) is the maximal nilpotent ideal of u. Also let 1) = 1)" I Uwhere ~ e D so that V = V". Now if ~ e D1let xi be the character of the representation 1)i of Gl' Then if, as in §7.2, XU) is the character of the representation j3i of U on H(n, V") it follows from Theorem 5.14 that for any a e G1 ~ U, (7.3.1)
250
BERTRAM KOSTA NT
374
where, we recall, ~CT = a(g + :\,) - g and WI(j) is given by (§ 5.14). For any a e W let sga, as usual denote the determinant of n. If n(n) is defined by (5.13.3) it is then well known that (7.3.2)
sgn = (_l)n(CT) •
(In fact sjnce there are obviouslyn(n) root "walls" separating, for example, = -1 for any cp e d). But now Proposition 7.2, (7.3.1) and (7.3.2) yield
g and ng, (7.3.2) follows from the fact that SgTg>
PROPOSITION 7.3. Let:\' e D and let X~ be the character of the ir1'educible representation lJ~ of G. Let GI and n be defined as in § 5.3 and let R(GI ) be the set of all a e GI such that O(a)z = z, for zen implies z = O. Here e denotes the ad}oint representation of G on g. Then for any a € R(G1) ~() a
(7.3.3)
=
X
ECTEwlSg nXf(UHH(a) ECTEWIsgnXf(U)-U(a) ,
where for any ~ e DH Xi is the character of the irreducible representation of GI and WI is given by (5.13.1).
lJ;
7.4. Now consider the special case of Proposition 7.3 where u = 0 so that n = m and GI = H where H ~ G is the (Cartan) subgroup corresponding to 'fJ. In this case DI = Z and WI = W. Furthermore if a e H then writing a = exp x for x € 'fJ one has, for any ~ e Z xi(a)
= e«'x)
•
Moreover R(H) is the set of all elements in H that are regular in G. Multiplying numerator and denominator of (7.3.3) by e(g·3) one obtains, as an immediate corollary to the proposition above, PROPOSITION 7.4. (Weyl's character formula). Let Xl. be the character of the representation lJ h • Let a e H be regular in G. Then writing a = exp x one has ~(a)
X
-
"
sgae
.L.iCTEW
-" sgae(CTU.X> .L.iCTEW
•
REMARK 7.4. Let x € 1) and put a = exp x. Note then that the identity Xo(a) = Xo(a), (see (7.2.4», is just the familiar relation
II
g>E6+
(1 -
e-(tp·x» = e-(u,x>."
sgne
.L.iCTEW·
7.5. Let g+ be any reductive complex Lie algebra. Without loss of generality, however, we may assume that 9 is the maximal semi-simple ideal in g+. Now let G+ be any complex Lie group (not necessarily connected) whose Lie algebra is gl.
251
GENERALIZED BOREL-WElL THEOREM
375
Here we will let 0 denote the adjoint representation of G+ on g+. If a e G+, it is clear of course that c, the center of g+, and 9 are both stable under O(a). However, we note that both O(a) I c and O(a) I 9 may be, respectively, outer automorphisms of c and g. Now let
(7.5.1)
H+ = {a e G+ I m and 1) are both stable under O(a)} .
In case G+ is connected it is clear that H+ is a Cartan subgroup of GI. However, if G+ is not connected then, for one thing, the group H+ may not be commutative (and in fact H+ may be quite complicated, especially if the identity component G: of G+ has a non-trivial center and g+ has a large number of isomorphic simple components.). Nevertheless as far as conjugacy and representation theory are concerned, as we now observe, H+ appears to be the natural substitute for H. Now let C denote the Cartan group of 9 operating, like W, in 1) and contragrediently in 1)'. One knows then (since W is transitive on the Weyl chambers) that C= CoW is a semi-direct product where W, the Weyl group, is a normal subgroup of C and Co = {r e C I r: II -> II} . Now let a ---. r(a) be the homomorphism of H I into Cn defined by the condition that for any a e H+, a e IT
(7.5.2)
(O(a)e .. )
We then denote by Cn' Now let
~
=
(erial .. ) •
Co the image of H1- under this homomorphism. ))1:
H+ ---. End VI
be an irreducible representation of H+ on Vi" Since))1 induces a representation of 1) (which clearly also arises from a representation of H) we may consider the set ,6.~1 of weights of ))1 (we ignore the center c of g) and note that ,6.~1 ~ Z. It is then immediate from (7.5.1) that if A. e a~1 then ,6.>1 is given by
(7.5.3) We will now say that
))1
is a dominant representation of HI if a~1 ~
D.
Since D is stable under Ct" observe that by (7.5.3) for at least one A. e ,6.>',
252
))1
is dominant if A. e D
376
BERTRAM KOSTANT
Now let A be an index set for the equivalence classes of all dominant irreducible representations of H+. Now just as the elements of D index both the classes of dominant representations of H and all representations of G we now observe that A is an index set for the classes of all irreducible representations of G+. That is, to each 0 e A there exists a unique (up to equivalence) irreducible representation such that if and }.if: H+ -> End
VIS
is the representation defined by restricting }.is I HI- to V,s (obviously a stable subspace), then}.if is a dominant irreducible representation of H t- belonging to the equivalence class corresponding to o. Furthermore every irreducible representation }.i of G+ is equivalent to }.is for some, necessarily unique, o e A. The proof of the statements above proceeds in the same way as in the classical situation as soon as one observes that G+ = HIG. where G" is the subgroup of corresponding to g. Now let a e H+ and let a e W. Since W is normal in C we can let a' e W defined by the relation
G:
r(a)a = a'r(a) .
Recalling that by definition CP.,. = some scalar x~(a),
a(~_)
n~
I'
we then observe that for
(7.5.4) Similarly if V; ~ VS is defined in the same way as VIS except that replaces ~I we observe that
a~
t-
(7.5.5) It follows therefore that if H: is the subgroup of H+ defined by
H:
=
{a e H+ Jr(a) commutes with a} ,
then V; is stable under }.isJ H: and hence defines a representation }.i! of H:. Let X! be the character of }.i!. REMARK 7.5. If xf is the character of }.if note that xf determines X! for any a e W. In fact if b(a) e Ge is any element which induces, by conjugation, the transformation a on 1) observe that
253
GENERALIZED BOREL-WElL THEOREM x~(a)
377
= xHb(a)ab(a)-l)
for any a € H:. The element b(a) is needed since a itself does not in general operate on H~-. Finally put (7.5.6)
H:.
for any a € An element a € G+ is called regular if the rank of 8(a) - 1 is minimum in the connected component of G+ containing a. In case gl = 9 this definition is the same as that given by Gantmacher, [8, pp. 112, 119]. Since 8(a) I c = 8(b) I c for a, b € G+ lying in the same connected component we may apply the results of [8] to the case at hand. In particular, it follows then from Theorems 12, 23 and 29 in [8] that every regular element is conjugate to an element in H+ and that a € H+ is regular if and only if the kernel of 8(a) - 1 lies in 1)+, the Lie algebra of H+. But the latter clearly implies that a € H+ is regular if and only if 8(a) has no fixed vectors in m. Thus if we apply the considerations of §7.2 to the case where Uis the normalizer in G+ of the subgroup of G. corresponding to m and n = m, it follows that R(H+) is the set of all elements in H+ that are regular in G I. Applying Proposition 7.2 where ).1=).181 U, we obtain the following generalization of Weyl's character formula. THEOREM 7.5. Let g+ be any reductive Lie algebra and let G t be any Lie group (not necessarily connected) whose Lie algebra is gl-. We may assume that 9 is the maximal semi-simple ideal in g+. Now let H+ be defined by (7.5.1) so that there is a one-one relation between all dominant irreducible representations of H+ (indexed by A) and all irreducible representations of G+. Let 8 € A and let be the character of the irreducible representation ).18 defined above. Let a € H+ be regular in G+ (every regular element of Gt is conjugate to an element in H+) and let Wa be the subgroup of W consisting of all a € W which commute with -r(a) (see (7.5.2». Then where Xf(a) and Xf,8(a) are given respectively by (7.5.4) and (7.5.6) one has
t
x 8(a) =
EuewasgaXf,B(a) EuewasgaXf(a)
PROOF. Define H(m, VB) with respect to the representation 7'C = ).18 I m. By decomposing VB into irreducible components under the action of ).181 g, it follows from Corollary 5.15, that the space of cochains (e'--q,.,.)® consists (except for zero) of non-cobounding cocycles and if (e'--q,.,.) ® V:) denotes the corresponding space of cohomology classes, one has the direct sum
V:
254
378
BERTRAM KOSTANT
H(m, VB) = EUEW(e~~) ® V:) .
We have now only to apply Proposition 7.2, (7.5.4) and (7.5.5).
q.e.d.
8. Application III. Symmetric complex spaces X and a generalization of a theorem of Ehresmann
1. Let u € V and let 91 and n be defined as in § 5.3. We continue with the notation of § 5 except that now it is assumed that A, = O. Thus (3 is a representation of 91 on An' and $ is the induced representation of 91 on H(n).
Now let (3* be the representation of 91 on An defined by restricting e I91 to An. Thus (3* is the representation contragredient to (3. Since (3* obviously commutes with the boundary operator aon An it defines a representation
$*:
9 -. End H*(n)
on 91 on the homology group H*(n). It is of course clear that, with respect to the canonical duality between H*(n) and H(n), $* is just the representation contragredient to $. Applying Theorem 5.14, one then immediately obtains COROLLARY 8.1. Let u € V and let $ * be the representation of 91 on the homology group H*(n) defined above. For any g € -Dl let H*(nY be the set of all elements in H*(n) which transform under $* according to the irreducible representation (with lowest weight g) )..Ii of 91' Then for any a € WI one has g - ag € - Dl and for any g € - Dl one has H*(n)< 1= 0 if and only if g = g - ag for some a € Wl. Furthermore H*(n)g-U g is irreducible for all a € Wl so that a -. HAn)U- ug is a bijection of WI onto the set of all irreducible (under $*) components of H*(n). Moreover, degree-wise, for any non-negative integer j,
Hj(n)
=
EuEWI{j)
g H*(n)g-U ,
so that the elements of H*(n)g-U g are homogeneous of degree n(a). Finally the lowest weight vector of H*(n)g-U g is the homology class having e~" as a representative cycle. 8.2. We consider the cases (u € V) when n is commutative. Let n(u) ~ n be defined as in § 5.4 and for any cp € a let the integer n ..(cp), a € n, be defined also in § 5.4. It is then asserted that n is commutative if and only if for every cp E a(n) (8.2.1) Indeed since
E"ETI(U)
~(n)
n",(cp) ~ 1 .
is precisely the set of all cp
255
€ ~
such that the left hand
GENERALIZED BOREL-WElL THEOREM
379
sum of (8.2.1) is ~ 1, it follows that the condition (8.2.1) implies that 11 is commutative. On the other hand if there exists a root such that the left hand sum of (8.2.1) is ~ 2 then since m is generated by the e.. , a € 11, it follows that there exists 1J € A(n) and a € ll(u) such that 1J + a € A. But Rince 1J, a, 1J + a € A(n), this implies that 11 is not commutative. This proves the assertion. An immediate consequence of this and symmetric space theory is PROPOSITION 8.2. Let u € CU and, as in § 5.2, let X = G/ U so that X is a complex compact homogeneous space. Then X is also a symmetric space in the sense of E. Cartan if and only if n, the maximal nilpotent ideal of It, is commutative. PROOF. It is immediate that the condition (8.2.1)is satisfied if and only if no two elements of l1(u) lie in the same connected component (in the sense of Dynkin) of 11; and for any a e l1(u), one has n ..(1J) ~ 1 for all 1J e A. But then the result follows from the structure theory of complex, compact, symmetric spaces (see e.g. [3, 40, p. 260]). q.e.d. But now if n is commutative, the boundary operator on An is zero. Thus H(n) = An. Hence in the symmetric case Corollary 8.1 yields Corollary 8.2 below describing how An decomposes under the action of gl' Corollary 8.2 contains, as a special case, results of Ehresmann asserting how A11 decomposes when X is symmetric and G is a classical group. We will work out the case when X is the grassmannian in § 8.6. COROLLARY 8.2. Let u € CU. Assume that X = G/ U is a symmetric space. Let n be the maximal nilpotent ideal of u and let 13* be the representation of gl on An obtained by restricting e I gl to An. (Recall that e is the adioint representation of 9 on Ag). Now let WI be the subset of the Weyl group defined as in § 5.13. Then for any a e W\ one has g - ag € - Dl and for any ~ e - DII the irreducible representation ).If of fh occurs in 13* if and only if ~ = g - ag for some a € WI. Furthermore if a € W\ then ).If-
Finally the lowest weight vector in (An)U-
256
380
BERTRAM KOSTANT
this let IC e W be defined as in § 5.10. Write (uniquely according to Proposition 5.13) =
IC
IC l lC
l
where ICl e WI and 1C1 e WI. By (5.13.2) it follows that <1>'1 = .o4(m1 ) (see § 5.5). Thus for any Ee - DH one has that 1C1(E) e Dl and ICl(E) is the highest weight of ).Ii. It follows therefore that in the notation of § 5.5 (An)g-o-g
=
(An)
8.3. Let m be a positive integer and let gm be the Lie algebra of all complex m x m matrices regarded as operating on em = V in the usual way. For any y e gm, let YiJ' i, j = 1,2, ... , m be the matrix coefficients of y. Now let am ~ gm be the set of m x m complex matrices of zero trace. We apply the considerations of § 8.2 to the case where 9 = am. We choose the maximal solvable Lie subalgebra b of 9 so that b
=
{y e 9 I Yo;
= 0,
i
> j}
(super-triangular matrices) and the Cartan subalgebra f) of !1 so that f) is the set of all diagonal matrices in g. The corresponding roots are then canonically indexed by all pairs i, j = 1,2, ... , m, i =f=. j, where qJ,} e A is given by (8.3.1) for any x e f) and the corresponding root vectors may be chosen so that e",,} =
V
1
2m e'J
where ei ; is the usual matrix unit. REMARK 8.3. The coefficient (l/V 2m) is necessary to insure the relation (5.1.2). It also insures (5.4.2) where f is chosen to be the set of all skewhermitian matrices in g. Since (
(8.3.2)
VZ --e", = ei }.
I qJij I
ij
But using (5.1.3) this implies that (8.3.3)
257
GENERALIZED BOREL-WElL THEOREM
381
Now it is clear from the choise of b that AI = A(b) = {cptJ}, i < j. Hence the simple positive roots may be indexed so that Il={aJ, i=l, 2"", m-l, where Now let Zm be the set of all m-tuples r, where the r i are integers. It is clear from § 5.5 and (8.3.3) that we can define a mapping
Zm -> Z, r
->
fJ.(r) ,
by letting fl( r) be defined by
+r
2
x 22
+ ... + r ",X mm
for any x e 1). The Weyl group W may be identified with the permutation group on the numbers 1,2, "', m. If we let Woperate on Zm by putting then this identification may be made so that p(ar) = ap(r) for any r e zm. Note then from (8.3.1) that for any root CPiJ (8.3.4) Now let
Dm = {r e Zm I r 1 ;S r 2 ;S ... ;S r m} . It is obvious that Dm is a fundamental domain for the action of Won zm
and by (8.3.3) the mapping r --+ p(r) carries Dm onto D. Now for any r e zm let n(r) = E~=l rio We recall some facts in the representation theory of g"'. Let Gm be the group of all m x m complex non-singular matrices. For each r e Z'" let lJT: gm ->
End
ViL(T)
be the irreducible representation of gm on ViL(T) defined so that lJT I 9 = lJiL(T) and, if 1m is the m x m identity matrix, lJT(l m ) is the scalar n(r) on ViLer'. REMARK 8.3. Note that lJ-r is the contragredient representation to lJT for any r e zm. It is clear that lJT is equivalent to lJU(T) for any a e W. Also one knows that every irreducible representation of gm which arises from an irreducible representation of G'" is equivalent to lJr for one and only one elementr e Dm.
8.4. Now let P be the set of all partitions of all non-negative integers.
258
382
Thus if P
BERTRAM KOSTANT €
P then P is given by a finite sequence P = {PH P2, ••• , PA;}
where PI ~ P2 ~ ••• ~ PA; are positive integers. To each P € P one associates two integers, n(p), where n(p) = E:=IPi
is the number being partitioned and m(p) (= k), the number of parts. Also one associates to P a Young diagram Y(p) which may be regarded as the set of all pairs (i, j) of positive integers where 1 ;£ i ;£ m(p) and j ;£ Pi' Schematically the pairs of Y(p) are represented by the boxes in the figure
~ (8.4.1)
1,2\'
E'" k,l \.
Let P € P. Then one associates with P another partition P (called its conjugate) where m(p) = PI and PJ, for 1 ;£ j ;£ PH is given by One has that n(p) = n(p), m(p) = PH m(p) = PI and schematically the box representation for Y(P) is obtained by transposing (8.4.1) as one would a matrix. Now let pm = {p € Plm(p);£ m},
and let r"': pm
-+
Z'"
be the mapping given by the relation (rm(p»i = Pi for 1 ;£ i ;£ m(p) and = 0 for m(p) < i ;£ m. It is clear that rm is a bijection of pm onto the subset D,{, of Dm consisting of all r € Dm such that r i is non-negative for all i. Let p € pm. We recall (Young theory) how one obtains the representation )./rm(Pl of gm. Let e" i = 1, 2, ... , m, be a basis of V such that e,,(e i ) = e, for all i. For any non-negative integer j let ®J V be the tensor product of V with itself j times and let )./J be the representation of gm on ®J V formed by tensor product of the canonical representation of gm on V with itself j times. (rm(p»;
259
GENERALIZED BOREL-WElL THEOREM
Now if 1
~
383
®J V be alternating tensor defined by e(j) = Lo-sgo-e ® ... ® eo- J
j ~ m, let e cn €
u1
where the sum is over all permutations on the numbers 1,2, "', j. Now let eCp ) € ®nCP) V be defined by where k = m(p) = Pl' Also let VP ~ ®ncP) V be the subspace generated by e Cp ) under the representation lI nCP ). Then the sub-representation of lI n (P) defined by VP is irreducible and is equivalent to lI r"'(P). Moreover eCp) is a highest weight vector of lI n (P) I 9. REMARK 8.4. One knows that for any integer j all the irreducible subrepresentations of ~ are of the form lI r where r € D';, that is, of the form lI r"'(P) where P € pm. It follows therefore that, since this is true for all j, any irreducible sub-representation of the tensor product lI r ' ® lI r " where r', r" € D'; is again of the form lI r where r € D';.
8.5. Let q; be as in § 5.2. Let s be an integer where 1 ~ s ~ m. We apply the considerations of § 8.2 to the case where X is the grassmannian of all s complex planes in V. That is, to the case where u € q; is the set of all matrices Y € 9 of the form Y
=
(A l1 (Y), A 12 (Y») o A 22(y)
where, if t = m - s, An(Y) is an s x s matrix, A 12(y) is an s x t matrix and A 22 (y) is a txt matrix. It is clear that n is the set of all Y € u such that Al1(Y) = A 2lY) = O. Furthermore if f is chosen to be the set of all skew-hermitian matrices in g, then 91 is the set of all Y € u such that A 12 (y) = O. Also note that if x € gl' Y € nand z = [x, y] = f3Ax)y € n then (8.5.1) Now by definition (see (5.13.1» W1 is the set of all 0- € W such that .6.+ and o--l(cp) € .6._ implies qJ € .6.(n). But now since .6.(n) is the set of all qJtlG € A such that 1 ~ i ~ sand s < k ~ m, it is clear from (8.3.4) that 0- € W1 if and only if
qJ €
(8.5.2) and o--l(S REMARK
+ 1) < o--l(S + 2) < ... < o--l(m) •
8.5. It follows immediately from (8.5.2) that an element 0- € W1
260
384
BERTRAM KOSTANT
is characterized by the values a-I(i), i = 1,2, ... , s. Thus there are (~) elements in WI, the elements being in a canonical one-one correspondence with the set of all subsets of s integers between 1 and m. Write t = m - s. Now let a 6 WI. It is obvious from (8.5.2) that for any 1 ~ i ~ sand 1 ~ j ~ t, one has (8.5.3) On the other hand we now observe that one or the other (but obviously not both) of the following inequalities must hold
a- 1(i)
(8.5.4)
~
a--I(s
+ j)
or a-I(s
+ j) < i + j
~
a- (i) . 1
That is, in any event i + j lies between a-I(i) and a- 1(s + j) and is greater than the minimum of the two. This is an immediate consequence of (8.5.2) and the fact that a-I is a permutation. We can now easily compute
=
{IPi.S+J
W"
6 ~
Ii
~
s, i
< a-1(i)
and 1
~
j
~ a-I (i)
- i} .
Now given a 6 it follows immediately from (8.5.3) that we can define a partition p" 6 ps by the relation (8.5.6)
rS(p")
= (a-I(s) - s, ..• , a- 1(1)
-
1) .
But it is then an easy consequence of (8.5.4) that the conjugate partition p" lies in pt and that (8.5.7)
rt(p")
= (s + 1 -
a-I(s
+ 1),
... , m - a- I (m») ,
since the jlh entree in rt(p") is by definition the number of entrees in rS(pIT) which ~ j. Now let q '6 Dm be defined by q
=
(m, m - 1, ",,1) .
It is then clear from (5.9.1) and (8.3.3) that p.(q) note that if a 6 W, then q - aq
= (a- (1) I
Now recall that the element 4PKI = A(mI ) (see Remark 8.2). It follows easily then that
= g.
On the other hand
- 1, "', a-I(m) - m) .
"1 6
WI is characterized by the fact that
(8.5.8)
261
GENERALIZED BOREL-WElL THEOREM
385
where Il' = (1, 8)(2, 8 - 1), ... , and K" = (8 + 1, m)(s + 2, m - 1), .... But then adjoining a t-tuple to an 8-tuple to make an m-tuple, it follows from (8.5.6) and (8.5.7) that for any a € W' {8.5.9) Kr
= {
j)
A I (K'(i),
J €
€
Y(pCT)}
:so that by (8.5.8) 0(8.5.11)
K,(
=
{qJ iln
.
1
}
E A I (i, j) E Y(p")} .
8.6. Now identify H"
= AS EB g'
with the set of all m x m matrices of the form y given in § 8.5 where A,ly) = 0 and A l1 (y) € gS, A 21 (y) € g' are arbitrary. Let r € zrn. We may write, uniquely, r = (r" r2) where r' € zs and r 2 € Zt. We will let J.>r be the irreducible representation of au = GS x G' {)n ViL(j" 0 ViL(r 2 ) given by
It is clear then that every irreducible representation of Gs,t is equivalent to J.ir for some r € Zm and in fact r is uniquely chosen if one insists that r' € DS and r 2 € D'. Now the adjoint representation of AS" on n (see 8.5.1) extends in the usual way to a representation fl •. ,: g8,' ~ End
An
()f gs.t on An. We observe that the representation fls., is obtained as an -extension of the representation fl* of g, on An by defining fl •. ,(1"') = o. We wish to decompose the representation fls., into irreducible components. We first observe, however, that if reD' then-K"r € D' and J.>-K"r is equivalent to the representation contragredient to the representation v r of g'. This is clear from the definition of K". It follows easily therefore from (8.5.1) and Remark 8.4 that any irreducible component of flu is ()f the form J.>t' ,-K"r2) where r' € D~ and r 2 € Dt. Now let Q'" be the set of all pairs (p" p2) where p' € P' and p2 € P' is such that J.>t',-K"r 2) occurs in the complete decomposition of fl •. , if r'=rA(p') a.nd r2 = r'(p2). From the remark above we see that every irreducible
262
386
BERTRAM KOSTANT
component of (3s,t is of this form so that (38,t is determined as soon as the elements of Q8.t are known together with the corresponding multiplicities. The following theorem is due to Ehresmann. See [5, § 5]. THEOREM 8.6. Let Qs.t be the set of pairs of partitions defined above describing the decomposition of the representation (3 s,t of gS E9 gt (gil: is the Lie algebra of all k x k complex matrices) on An where n is isomorphic to the space of all s x t complex matrices. Let WI be defined as in § 5.13 so that here WI is the set of permutations n € W satisfying (8.5.2). If n € WI, let pO' € ps be the partition defined. by rS(pO') = (n-l(s) - s, ... , n-l (l) - 1) . Then n(pO') = n(n)
(8.6.1)
where the left and right sides of (8.6.1) are defined respectively as in § 8.4 and by (5.13.3). Furthermore n --> pfT is a bijection of WI onto the set of all partitions p such that m(p)
~
sand m(p)
~
t
where p is the conjugate partition (that is, the set of all partitions whose Young diagram (block representation) "fits" into an s x t rectangle of blocks). Finally Qs,t is the set of all pairs (pfT, pfT) where n runs through W]. Moreover the irreducible representation of g8,t corresponding to any pair (pfT, pfT) occurs with multiplicity one and the representation induced on gl is lJ~l(g-fTg). Moreover the space of the representation consists of homogeneous n(pfT) vectors and a highest weight vector of lJ K1 (g- O' U) is (in any order)
where II denotes exterior multiplication. PROOF. The equality n(pO') = n(n) follows from (8.5.11) and the other statements about pfT follow from Remark 8.5. To prove the theorem therefore we have only to apply Corollary 8.2 and Remark 8.2, and to determine the element of Qs,t corresponding to representation lJ~l(g-O'U) of gl on the subspace (/\ nYI(g-fT g) of N(fT) n. That is we must find the pair (pI, p2) € QU such that (1) p(r S(pl), -1C"rt(p~» = 1C1(g - ng) and (2) n(pl) = n(n) (since (3s,t(Y) must reduce to the scalar n(n) on N(rT)n if y € gS,t is the element such that A 22(y) = 0 and A ll (y) = 18).
263
GENERALIZED BOREL-WElL THEOREM
387
It is easy to see that (1) and (2) define (pt, p2) uniquely. But by (8.5.9), (8.9.10), and the equality (8.6.1), it follows that (pt, p2) = (per, per). The final statement follows from (8.5.11) and Remark 8.2. q.e.d.
REMARK 8.6. Theorem 8.6 lends some insight into the nature of the weight g - ag, at least for the case at hand. The striking thing is that the partitions pI and p2 of the pair (pI, p2) corresponding to the weight g - ag not only determine each other but are related to the extent that -one is the conjugate of the other. Furthermore except for a limitation on .size, the choice of pI can be made arbitrary by choosing a properly. UNIVERSITY OF CALIFORNIA, BERKELEY REFERENCES 1. A. BOREL and F. HIRZEBRUCH, Characteristic classes and homogeneous spaces, I. Amer. J. Math., 80 (1958), 458-538. 2. R. BOTT, Homogeneous vector bundles, Ann. of Math., 66 (1957) 203-248. 3. E. CARTAN, Sur une classe remarquable d'espaces de Riemann, Bull. Soc. Math. France, 54 (1926), 214-264. 4. P. CARTIER, Remarks on "Lie algebra cohomology and the generalized Borel-Weil theorem", by B. Kostant, Ann. of Math., 74 (1961), 388-390. 5. C. EHRESMANN, Sur la topologie de certains espaces homogenes, Ann. of Math., 35 (1934), 396-443. 6. H. FREUDENTHAL, Zur Berechnung der Charaktere der halbeinfachen Lieschen Gruppen Neder, Akad. Wetensch. Indag. Math., 57 (1954),369-376. 7. G. HOCHSCHILD and J.-P. SERRE, Cohomology of Lie algebras, Ann. of Math., 57 (1953), 591-603. 8. F. GANTMACHER, Canonical representation of automorphisms of a complex semi-simple Lie group, Mat. Sb., 47 (1939), 101-143. 9. J. L. KOSZUL, Homologie et cohomologie des algebres de Lie, Bull. Soc. Math. France, 78 (1950), 65-127. 10. H. C. WANG, Closed manifolds with homogeneous complex structure, Amer. J. Math .• 76 (1954). 1-32.
264
Reprinted from the TRANSACTIONS OF THE AMERICAN MATHEMATICAL SOCIETY Vol. 102, No. 3, March 1962 Pp. 383-408
DIFFERENTIAL FORMS ON REGULAR AFFINE ALGEBRAS BY
G. HOCHSCHlLD, BERTRAM KOSTANT AND ALEX ROSENBERG(I)
1. Introduction. The formal apparatus of the algebra of differential forms appears as a rather special amalgam of multilinear and homological algebra, which has not been satisfactorily absorbed in the general theory of derived functors. It is our main purpose here to identify the exterior algebra of differential forms as a certain canonical graded algebra based on the Tor functor and to obtain the cohomology of differential forms from the Ext functor of a universal algebra of differential operators similar to the universal enveloping algebra of a Lie algebra. Let K be a field, R a commutative K-algebra, TR the R-module of all K-derivations of R, DR the R-module of the formal K-differentials (see §4) on R. It is an immediate consequence of the definitions that T R may be identified with HomR(D R, R). However, in general, DR is not identifiable with HomR(TR, R). The algebra of the formal differentials is the exterior Ralgebra E(D R) built over the R-module DR. The algebra of the differential forms is the R-algebra HomR(E(TR), R), where E(TR) is the exterior R-algebra built over T R and where the product is the usual "shuffle" product of alternating multilinear maps. The point of departure of our investigation lies in the well-known and elementary observation that TR and DR are naturally isomorphic with Ext1-(R, R) and Torr(R, R), respectively, where R6=R@K R. Moreover, both ExtR.(R, R) and TorR·(R, R) can be equipped in a natural fashion with the structure of a graded skew-commutative R-algebra, and there is a natural duality homomorphism h: Exh.(R, R)~HomRCTorR·(R, R), R), which extends the natural isomorphism of TR onto HomR(D R, R). We concentrate our attention chiefly on a regular affine K-algebra R (d. §2), where K is a perfect field. Our first main result is that then the algebra TorR·CR, R) coincides with the algebra E(D R ) of the formal differentials, Exh.CR, R) coincides with E(TR), and the above duality homomorphism h is an isomorphism dualizing into an isomorphism of the algebra ECD R) of the formal differentials onto the algebra HomR(E(TR), R) of the differential forms. In order to identify the cohomology of differential forms with an Ext functor, we construct a universal "algebra of differential operators," VR, Received by the editors May 5, 1961. (1) Written while B. Kostant was partially supported by Contract AF49(638)-79 and A.
Rosenberg by N.S.F. Grant G-9508.
383 B. Kostant, Collected Papers, DOI 10.1007/b94535_14, © Bertram Kostant 2009
265
384
G. HOCHSCHILD, BERTRAM KOSTANT AND ALEX ROSENBERG
[March
which is the universal associative algebra for the representations of the K-Lie algebra TR on R-modules in which the R-module structure and the TR-module structure are tied together in the natural fashion. After establishing a number of results on the structure and representation theory of V R, we show that, under suitable assumptions on the K-algebra R and, in particular, if R is a regular affine K-algebra where K is a perfect field, the cohomology K-algebra derived from the differential forms may be identified with ExtvR(R, R). In §2, we show that the tensor product of two regular affine algebras over a perfect field is a regular ring, and we prove a similar result for tensor products of fields. §§3, 4 and 5 include, besides the proof of the first main result, a study of the formal properties of the Tor and Ext algebras and the pairing between them, for general commutative algebras. In the remainder of this paper, we deal with the universal algebra V R of differential operators. In particular, we prove an analogue of the Poincare-Birkhoff-Witt Theorem, which is needed for obtaining an explicit projective resolution of R as a VR-module. Also, we discuss the homological dimensions connected with V R. We have had advice from M. Rosenlicht on several points of an algebraic geometric nature, and we take this opportunity to express our thanks to him. 2. Regular rings. Let R be a commutative ring and let P be a prime ideal of R. We denote the corresponding ring of quotients by R p • The elements of Rp are the equivalence classes of the pairs (x, y), where x and yare elements of R, and y does not lie in P, and where two pairs (Xl, YI) and (X2' Y2) are called equivalent if there is an element z in R such that z does not lie in P and Z(XlY2 - X2YI) = o. By the Krull dimension of R is meant the largest non-negative integer k (or 00, if there is no largest one) for which there is a chain of prime ideals, with proper inclusions, PoC ... CPkCR. A Noetherian local ring always has finite Krull dimension, and it is called a regular local ring if its maximal ideal can be generated by k elements, where k is the Krull dimension. A commutative Noetherian ring R with identity element is said to be regular if, for every maximal ideal P of R, the corresponding ring of quotients Rp is a regular local ring [2, §4]. It is well known that a regular local ring is an integrally closed integral domain [14, Cor. 1, p. 302]. It follows that a regular integral domain R is integrally closed; for, if X is an element of the field of quotients of R that is integral over R then xER p , for every maximal ideal P of R, which evidently implies that xER. Let K be a field. By an affine K-algebra is meant an integral domain R containing K and finitely ring-generated over K. An affine K-algebra is Noetherian, and its Krull dimension is equal to the transcendence degree of its field of quotients over K, and the same holds for the Krull dimension of everyone of its rings of quotients with respect to maximal ideals [14, Ch.
VII, §7].
266
19621
DIFFERENTIAL FORMS ON REGULAR AFFINE ALGEBRAS
385
THEOREM 2.1(2). Let K be a perfect field, and let Rand S be regular affine K-algebras. Then R®K S is regular.
Proof. Suppose first that R®K S is an integral domain, and let M be one of its maximal ideals. Put Ml = (M(\R) ®K S+R®K (M(\S). Then Ml is an ideal of R®K S that is contained in M, and we have (R®K S)/M1 = (R/(M(\R)) ®K (S/(M(\S»). Now R/(R(\M) and S/(M(\S) are subrings of (R ® K S) / M containing K. Since (R ® K S) / M is a finite algebraic extension field of K, the same is therefore true for R/(M(\R) and S/(M(\S). Since K is perfect, it follows that we have a direct K-algebra decomposition (R®K S)/M1= U+M/M1. Let z be a representative in R®K S of a nonzero element of U. Then z does not belong to M, and zMCM1. Hence it is clear that M(R®K S)M=M1(R®K S)M. Since R is regular, the maximal ideal (M(\R)RMnR of the local ring RMnR is generated by d R elements, where d R is the degree of transcendence of the field of quotients of Rover K. Similarly, (M(\S)SMns is generated by d s elements, where d s is the degree of transcendence of the quotient field of S over K. These dR+d s elements may be regarded as elements of (R®K S)M and evidently generate the ideal M1(R®K S)M. Hence we conclude that the maximal ideal of (R ® K S) M can be generated by d R+ds elements. Since the degree of transcendence of the quotient field of R®K S over K is equal to dR+d s , this means that (R®K S)M is a regular local ring. Thus R®K S is regular. N ow let us consider the general case. Let Q(R) and Q(S) denote the fields of quotients of Rand S. Let KR and KS be the algebraic closures of K in Q(R) and in Q(S), respectively. Since Rand S are integrally closed, we have KR CR and KS CS. Since Q(R) and Q(S) are finitely generated extension fields of K, so are KR and KS. Thus KR and KS are finite algebraic extensions of K. Let M be a maximal ideal of R ® K S. Since K is perfect, we have a direct K-algebra decomposition KR®KKs= U+M1, where M1=M(\(KR®KKS). Hence we have R ®KS
=R
®KR(KR®KKS) ®KSS
=R
®KB U ®KSS
+ M 2,
where the last sum is a direct K-algebra sum, and M2=R®KB M1®KS SCM. Evidently, U may be identified with a subring of the field (R®K S)/M containing K. Hence U is a finite algebraic extension field of K. Identifying KR and KS with their images in U, we may also regard U as a finite algebraic extension field of KR or KS. Since K is perfect, U is generated by a single element over KR or over KS. The minimum polynomial of this element over KB or over KS remains irreducible in Q(R)[x] or in Q(S)[x], because KR is algebraically closed in Q(R) and KS is algebraically closed in Q(S). Hence (") The referee informs us that this result is an immediate consequence of cohomology results obtained by D. K. Harrison in a paper on Commutative algebras and cohomology, to appear in these Transactions.
267
386 G. HOCHSCHILD. BERTRAM KOSTANT AND ALEX ROSENBERG
[March
R ® KR U and U ® KS S are integral domains. Moreover. by the part of the theorem we have already proved, they are regular. Let T denote the field U®KS Q(S). This is a finitely generated extension field of the perfect field KR. Let (tl, ...• t n ) be a separating transcendence base for T over KR, and put To = KR(tl' ... , t n ). We have Q(R) ®KR T = (Q(R) ®KR To) ®To T, and we may identify Q(R) ®KR To with a subring of Q(R) (h •...• t n ), with (tl' ... , t n ) algebraically free over Q(R). Since KR is algebraically closed in Q(R), it follows that KR(h, ... , t n ) is algebraically closed in Q(R)(lI, ... ,tn ) [6, Lemma 2, p. 83]. Now it follows by the argument we made above thatR ®KR Tis an integral domain,so thatR®KR U®KSS is an integral domain. On the other hand, this is the tensor product, relative to the perfect field U, of the regular affine U-algebras R ® KR U and U ® KS S. Hence we may conclude from what we have already proved that R ® KR U ® KS S is regular. Now consider the direct K-algebra decomposition R ®K S
=
R ®KR U ®Ks S
+M
2•
Since M2 C M, the corresponding projection epimorphism R ® K S -R®KR U®KS S sends the complement of M in R®K S onto the complement of Mrl(R®KR U®KS S) in R®KR U®KS S. Moreover, there is an element z in the complement of M such that ZM2= (0). Hence it is clear that the projection epimorphism yields an isomorphism of (R®K S)M onto the local ring over R ® KR U ® KS S that corresponds to the maximal ideal Mrl(R®KR U®KS S). Hence (R®K S)M is a regular local ring. and Theorem 2.1 is proved. THEOREM 2.2. Let K be an arbitrary field, let F be a finitely and separably generated extension field oj K, and let L be an arbitrary field containing K. Then F®K L is a regular ring.
Proof. It is known that the (homological) algebra dimension dim(F), i.e., the projective dimension of F as an F®K F-module is finite; in fact, it is equal to the transcendence degree of F over K [11, Th. 10]. Since dim(F®K L) =dim(F), where F®K L is regarded as an L-algebra [4, Cor. 7.2, p. 177] we have that dim(F®K L) is finite. Since L is a field, this implies that the global homological dimension d(F®KL) is also finite [4. Prop. 7.6, p. 179]. Since F®K L is a commutative Noetherian ring, we have, for every maximal ideal M of F®K L, d«F®K L)M) ~d(F®K L) [4, Ex. 11, p. 142; 1, Th. 1]. Thus each local ring (F®K L)M is of finite global homological dimension. By a well-known result of Serre's [12, Th. 3], this implies that (F®K L)M is a regular local ring. Hence F®K L is a regular ring. Note. Actually. we shall later appeal only to the following special consequence of Theorem 2.2: let F be a finitely separably generated extension field of K; let J be the kernel of the natural epimorphism F®K F-F; then the
268
1962)
DIFFERENTIAL FORMS oN REGULAR AFFINE ALGEBRAS
387
local ring (F®K F)J is regular. This special result can be proved much more easily and directly along the lines of our proof of Theorem 2.1. On the other hand, Theorem 2.1 can be derived more quickly, though less elementarily, from the result of Serre used above. 3. The Tor-algebra for regular rings. Let Rand S be commutative rings with identity elements, and let cp be a unitary ring epimorphism S----'>R. We regard R as a right or left S-module via cp, in the usual way, and we consider TorS(R, R). Since S is commutative, every left S-module may also be regarded as a right S-module, and we shall do so whenever this is convenient. Let H stand for the homology functor on complexes of S-modules, and let U and V be any two S-module complexes. There is an evident canonical homomorphism of H( U) ® s H( V) into H( U ® s V), which gives rise to an algebra structure on TorS(R, R), as follows. Let X be an S-projective resolution of R. With U = V = R ® S X, the canonical homomorphism becomes a homomorphism TorS(R, R) ®s TorS(R, R)
----'>
H«R ®s X) ®s (R ®s X)).
Evidently, (R®sX) ®s(R®sX) maybe identified with (R®sR) ®s(X®sX), and hence with R®s (X®s X). Now X®s X is an S-projective complex over R®s R=R, whence we have the natural homomorphism H(R ®s (X ®s X))
----'>
TorS(R, R).
Composing this with the homomorphism above, we obtain an S-module homomorphism
Tors (R, R) ®s TorS(R, R)
----'>
TorS(R, R).
m
This is the product of [4, p. 211] and it is independent of the choice of the resolution X. Standard arguments on tensor products of complexes and resolutions show that this product is associative and skew-commutative in the sense that a~ = ( -l)pq~a when a is homogeneous of degree p and ~ is homogeneous of degree q. In principle, this product is a product of S-algebras. However, S operates on TorS(R, R) through cp: S----'>R, and we shall accordingly regard TorS(R, R) as an R-algebra. THEOREM 3.1. Let Sand R be Noetherian commutative rings with identity elements, and let cp be a ring epimorphism S----'>R with kernel I. Assume that R is a regular ring and that, jor every maximal ideal M oj S that contains I, the local ring SM is regular. Then TorS(R, R) is finitely generated and projective as an R-module and is naturally isomorphic with the exterior R-algebra constructed over Torf(R, R).
Proof. Let T denote the tensor algebra constructed over the R-module Torf(R, R), let P denote the kernel of the canonical R-algebra homomorphism
269
388 G. HOCHSCHILD, BERTRAM KOSTANT AND ALEX ROSENBERG [March !/t: T-tTorS(R, R), and put Q=TorS(R, R)N(T). Let U denote the 2-sided ideal of T that is generated by the squares of the elements of Torf(R, R). The last assertion of our theorem means that Q = (0) and P = U. The statement Q= (0) is equivalent to the statement RN®R Q= (0), for all maximal ideals N of R. The statement P = U is equivalent to the statement (P + U) / P = (0) and (P+U)/U=(O), or to the statement RN®R (P+ U)/P= (0) and RN®R(P+ U)/ U = (0), for all maximal ideals N of R. This, in turn, is equivalent to the statement that the images of RN ® RP and RN ® RUin RN ® R(P U) coincide with RN®R (P+ U). Since RN is R-flat, these tensor products may be identified with their canonical images in RN®R T; and RN®R P is thereby identified with the kernel of the homomorphism of RN®R T into RN ®R TorS(R, R) that is induced by!/t. Hence it is clear that the statement Q= (0) and P= U is equivalent to the statement that the homomorphism of RN ® R T into RN ® R TorS(R, R) that is induced by!/t is an epimorphism with kernel RN ® R U, for every maximal ideal N of R. Let M be the maximal ideal of S that contains I and is such that M / I = N. Clearly, the epimorphism cf> induces an epimorphism SM-tRN with kernel ISM in the natural fashion. Now let X be an S-projective resolution of R. Since SM is S-flat, SM®S X is then an SM-projective resolution of SM®S R=SM/ISM=RN. Hence we have
+
TorSM(SM/ISM, SM/IS M) = H«SM/ISM) ®SM (SM ®s X)),
On the other hand,
Since RN is R-flat, we have H(RN®R (R®s X)) =RN®R TorS(R, R). Thus RN®R TorS(R, R) is naturally isomorphic with TorSM(SM/IS M, SM/IS M). Similarly, we see that RN ® R T is naturally isomorphic with the tensor algebra constructed over the RN-module TorfM(SM/IS M, SM/IS M). Moreover, it is easily seen that these isomorphisms transport our homomorphism RN ® R T -tRN®R TorS(R, R) into the canonical homomorphism of the tensor algebra over TorfM(SM/IS M, SM/IS M) into TorSM(SM/IS M, SM/IS M). Each Tor~(R, R) is finitely generated as an S-module, and hence also as an R-module. Hence if we show that RN®R Tor~(R, R) is a free RN-module, for every maximal ideal N of R, we shall be able to conclude from a standard result [4, Ex. 11, p. 142] that Tor~(R, R) is a finitely generated projective R-module. In particular, if Torf(R, R) is a finitely generated projective Rmodule, we imbed it as a direct R-module summand in a finitely generated free R-module to show that the exterior algebra constructed over it has nonzero components only up to a certain degree and is a finitely generated projective R-module. From this preparation, it is clear that it suffices to adduce the following
270
1962]
DIFFERENTIAL FORMS ON REGULAR AFFINE ALGEBRAS
389
result (3): let L ( = S M) be a regular local ring and let J ( = ISM) be a prime ideal of L such that the local ring L/ J is regular. Then Torf(L/ J, L/ J) is a finitely generated free L/ J-module, and TorL(L/ J, L/ J) is naturally isomorphic, as an L/ J-algebra, with the exterior algebra constructed over Torf(L/ J, L/ J). To prove this, note first that the assumptions imply that the ideal J can be generated by an L-sequence (aI, ... , aj) of elements of L, i.e., by a system with the property that each ak is not a zero-divisor mod the ideal generated by aI, ... , ak-l [14, Th. 26, p. 303 and Cor. 1, p. 302]. If X is the Koszul resolution of L/ J as an L-module [4, pp. 151-153], constructed with the use of this L-sequence, then X has the structure of an exterior L-algebra over a free L-module of rank j, this algebra structure being compatible with the boundary map, so that it induces the algebra structure on TorL(L/ J, L/ J) via (L/ J) ® LX. Moreover, the boundary map on (L/ J) ® L X is the zero map. Hence it follows immediately that TorfcL/ J, L/ J) is a free L/ J-module of rank j and that TorL(L/ J, L/ J) is the exterior algebra over this module. This completes the proof of Theorem 3.1. 4. Duality between Tor and Ext. Let Rand S be commutative rings with identity elements, and let cp be a ring epimorphism of S onto R. As before, all R-modules are regarded as S-modules via cpo Let X be an S-projective resolution of R, and let A be an R-module. Then Exts(R, A) =H(Homs(X, A». Clearly, we may identify Homs(X, A) with HomR(R®s X, A), so that we may write Exts(R, A) =H(HomR(R®s X, A». Now there is a natural map (a specialization of [4, p. 119, last line]) h: H(HomR(R ®s X, A»
~
HomR(TorS(R, R), A)
defined as follows. Let p be an element of H(HomR(R®s X, A». Then p is represented by an element uEHomR(R®s X, A) that annihilates d(R®s X), where d is the boundary map in the complex R ® s X. Hence, by restriction to the cycles of R®s X, u yields an element of HomR(TorS(R, R), A), and it is seen immediately that this element depends only on p and not on the particular choice of the representative u. Now h(p) is defined to be this element of HomR(TorS(R, R), A). Clearly, h is an R-module homomorphism of Exts(R, A) into HomR(TorS(R, R), A). In degree 0, we have Torg(R, R) =R®s R=R, and Ext~(R, A) = Homs(R, A) = HomR(R, A), and this last identification transports h into the identity map. Thus h is an isomorphism in degree o. Note that Torg(R, R) =R is projective as an R-module, whence the following lemma implies, in particular, that h is an isomorphism also in degree 1. LEMMA 4.1. Let cp: S~R be an epimorphism of commutative rings with identity elements, and regard R-modules as S-modules via cpo Let A be an R-
(3) This is a special case of [13, Th. 4, etc.], which gave the suggestion for our proof of Theorem 3.1.
271
390
G. HOCHSCHILD, BERTRAM KOSTANT AND ALEX ROSENBERG
[March
module, and let k be a positive integer. Assume that Tor~(R, R) is R-projective for aU i < k. Then the map S
i
hi: Exts(R, A) ---+ HomR(Tor. (R, R), A),
obtained by restriction of the map h defined above, is an isomorphism, for all i
~ k.
Proof. Let Zi denote the kernel of d in R®s Xi, and put Bi=d(R®s X i +I), C.=R®s Xi. We have Zo= Co. Suppose that we have already shown, for some i
272
1962]
DIFFERENTIAL FORMS ON REGULAR AFFINE ALGEBRAS
391
r*=-y.
rETR(A) and Thus the map r~r* is an R-module isomorphism of TR(A) onto Ext1(R, A). Regard S as an R-module such that x' (y®z) = (xy) ®z, and let J be the R-submodule of S that is generated by the elements of the form 1 ® (xy) -x®y-y®x. The factor module S/J is the R-module of the formal differentials of R. The map x~dx = the coset of 1 ®x mod J is the usual derivation of
R into the R-module of the formal differentials and, in fact, the definition of these amounts simply to enforcing the rule d(xy) =xdy+ydx. Let DR denote the R-module of these formal differentials. It is easily verified that the map S~1 that sends x®y onto x®y- (xy) ®1 induces in the natural wayan isomorphism of the R-module DR onto 1/12 = Torf(R, R) [3, Exp. 13]. Tracing through our above definitions and identifications, we see immediately that the duality isomorphism hI: ExtHR, A)~HomR(Torf(R, R), A) is transported into the map TR(A)~HomR(DR' A) attaching to rE TR(A) the element r' of HomR(D R, R) given by nL:xdy) = L:x·r(y). Let U be a multiplicatively closed subset of nonzero elements of R containing the identity element. Let Ru denote the corresponding ring of quotients. This is still a K-algebra, and we write Su for RU®K Ru. By an argument almost identical with the localization argument of the proof of Theorem 3.1, we see that RU®R TorB(R, R) is naturally isomorphic with TorBu(R u , Ru). In particular, we have RU®R DR naturally isomorphic with DRu' It is immediate from Theorems 2.1 and 3.1 that, if K is a perfect field and R is a regular affine K-algebra, then TorB(R, R) is a finitely generated projective R-module. We can use the above results to prove the following converse. THEOREM 4.1(4). Let K be a perfect field and let R be an affine K-algebra. Put S = R ® K R and suppose that Torf (R, R) is R-projective. Then R is a regular ring.
Proof. Let Q be the field of quotients of R, and let N be a maximal ideal of R. Since S is Noetherian, D R( =1/12) is finitely generated, and, byassumption, it is R-projective. Hence RN®R DR is a finitely generated projective, and hence free RN-module. Thus DRN is a finitely generated free RN-module. Now Q®RN DRN is isomorphic, as a Q-space, with D Q. We have HomQ(D Q, Q) isomorphic with TQ(Q), and, since Q is a finitely generated separable extension field of K, TQ(Q) is of dimension t over Q, where t is the transcendence degree of Q over K. Hence the dimension of D Q over Q is equal to t, whence we conclude that DRN is of rank t over R N. Write L for RN and M for NR N. Since D L is a free L-module of rank t, the L/ M-space HomL(D L, L/ M) is of dimension t over L/ M. We know from the (4) The essential, local, part of this result is contained in [10, Folgerung, p. 177]. Our proof of the local part is adapted from [3, Exp. 17, Th. 5]. The global theorem has also been obtained by Y. Nakai, On the theory of differentials in commutative rings, J. Math. Soc. Japan 13 (1961), 63-84.
273
392
G. HOCHSCHILD, BERTRAM KOSTANT AND ALEX ROSENBERG [March
above that HomL(D L, LIM) is isomorphic with T L(LI M). Hence T L(LI M) is of dimension t over LIM. Now a derivation of L into LI M must annihilate M2 and hence induces a derivation of LI M2 into LIM. Moreover, every derivation of LI M2 into LIM can evidently be lifted to give a derivation of L into LIM. Hence TL/M2(LI M) is isomorphic, as an LIM-space with TL(LIM), and thus is of dimension t. Now LI M2 is a finite dimensional algebra over the perfect field K with radical M 1M2. Hence we can write L I M2 as a semi direct sum L I M2 = V M 1M2, where V is a subalgebra isomorphic with the field LIM. Since K is perfect, every K-linear derivation of V into LI M must therefore be O. Hence it is clear that the restriction of the elements of T L/M2(LI M) to M 1M2 yields an isomorphism of T L/ M2 (LIM) onto HomL/M(MIM2, LIM). Hence we conclude that the dimension of MI M2 over LIM is equal to t. It follows by a standard argument from this that M can be generated by t elements. Since t is the Krull dimension of L, this shows that L, i.e., R N , is a regular local ring. This completes the proof. 5. Explicit multiplication. Let K be a commutative ring with identity, and let R be a commutative K-projective K-algebra. As before, let S = R 18> K R, and let cf> be the natural epimorphism S---)R. If A and Bare S-modules we regard them as two sided R-modules in the usual way, and we form A 18>R B. This is again a two sided R-module, and hence an S-module. Let X be an S-projective resolution of R. Using that R is K-projective, we see that XI8>R X is S-projective; essentially, this follows from the fact that SI8>R S = R 18> K R 18> K R, with the two sided R-module structure in which a· (u 18> v 18>w) = (au) 18>vl8>w and (ul8>vl8>w) 'a=ul8>vl8> (wa) , so that SI8>R Sis S-projective whenever R is K-projective. Moreover, S is R-projective as a left or right R-module, so that X is an R-projective resolution of R. Hence H(X 18>R X) = TorR(R, R) and therefore has its components of positive degree equal to 0, so that X 18> R X is still an S-projective resolution of R. For two sided R-modules U and V, regard Homs( U, V) as a two sided R-module such that (r -J) (u) = r -J(u)( = J(r· u» and (J·r)(u) = J(u) ·r( = J(u·r». Now the standard S-module homomorphism
+
1/;: Homs(X, A) 18>R Homs(X, B) ---) Homs(X 18>R X, A 18>R B),
where 1/;(f 18> g)(u 18> v) =J(u) 18>g(v) , induces an S-module homomorphi!5m Exts(R, A) 18> R Exts(R, B) ---) Exts(R, A 18> R B).
This is the product V, as given in [4, Ex. 2, p. 229], and it is independent of the choice of the resolution X. In particular, for A =B =R, this defines the structure of an associative and skew-commutative R-algebra on Exts(R, R). In order to make the algebra structures on TorS(R, R) and Exts(R, R) explicit, we use the following well-known resolution Y of R as an S-module. We put Yo=S and we let cf>: S---)R be the augmentation. Generally, let Y" be
274
1962]
DIFFERENTIAL FORMS ON REGULAR AFFINE ALGEBRAS
393
the tensor product, relative to K, of n+2 copies of R. The S-module structure of Y n is defined so that (a ® b)· (xo ® ... ® Xn+l) = (axo) ® Xl ® ... ® Xn ® (xnHb).
The boundary map don Y is given by n
d(xo ® ... ® Xn+l) =
L (-l) ixo ®
... ® (X;Xi-I-l) ® ... ® Xn+l .
• =0
This complex is not only acyclic but it has actually a right R-module homotopy h, where h(xl ® ... ®x n) = 1 ®XI ® ... ®xn. Since R is K-projective, it follows as above for X ®R X that Y is S-projective. Thus Y is an S-projective resolution of R. The complex Y can be given the structure of an associative skew-commutative S-algebra with respect to which d is an antiderivation, as follows. If Xl, ... , Xp and YI, •.. , yq are elements of R let [Xl, ... , Xp; Yl, •.. , Yq] stand for the sum, in the tensor product over K of p+q copies of R, of all terms of the form ±Zl® ... ®zp+q, where Zir.=Xk for some ordered subset (iI, ... , i p) of (1, ... , P+q), and Zh=Yk for the ordered complement VI, ... ,jq), and where the sign is or - according to whether the permutation (iI, ... ,ip,jl, ... ,jq) of (1, ... ,p+q) is even or odd. Then the product in Y is given by the maps Yp®s Yq~Yp+q that send
+
(xo ® ... ® Xp+l) ®s (Yo ® ... ® YqH)
onto (XoyO) ® [Xl, ... , Xp; Yl, ..• , yq] ® (Yq+lXP+l).
It can be verified directly that this is indeed an associative and skew-commutative product and, if a is homogeneous of degree p and (3 arbitrary, one has d(a{3) = d(a){3+( -l)Pad({3); d. [4, pp. 218-219]. This product evidently induces a product in R®s Y, and hence in TorS(R, R). By the nature of the definition of the product on TorS(R, R), as given earlier in the general case, the product induced from that on Y is the standard product (f) on TorS(R, R). Next we shall define a map of the complex Y into the complex Y®B Y which will serve to make the product on Exts(R, R) explicit. We have (Y®R Y)p= Yr®R Y p- r. As an S-module, each Yr®R Y p- r may be identified with the tensor product, relative to K, of p+3 copies of R, i.e., with Y p +!. With this understanding, we define an S-module homomorphism "Yr: Yp~Yr®R Y p- r such that 'Yr(XO ® ... ® Xp+l) = Xo ® ... ® Xr ® 1 ® Xr+l ® ... ® Xp+l.
L:-o
Now the desired map "Y: Y~Y®R Y is defined so that, for uE Y p, the component of "Y(u) in Yr®R Y p- r is "Yr(U). It is somewhat lengthy, but not difficult, to verify that"Y is compatible with the boundary maps on Y and on
275
394 G. HOCHSCHILD, BERTRAM KOSTANT AND ALEX ROSENBERG
[March
Y®R Y. The product V on Exts(R, R) is induced by the product on Homs( Y, R) induced by 'Y. In particular, with aEHoms( Y p, R) and ~EHoms(Yq, R), we have(5)
(a{3) (xo ® ... ® xpt-a+l)
= a(xo ® .. ® Xp ® 0
= xoa(l ® Xl ® .
0
1),8(1 ®
Xpt-I
®
® Xp ® 1),8(1 ®
•
0
••
XJt+1
®
Xpt-q+l)
® .
0
•
® xpt-q ® l)Xpt-q+l.
Consider a formal differential LxdyED R. It is easily verified that the corresponding element of Tor~(R, R) is represented in R ® s Y 1 by the element LX®s (l®y®l). On the other hand, let rETR(R)=TR (say). Then it is easily seen that its image r*EExt1(R, R) is represented in Homs(YI , R) by the element r', where S'(XO®XI®X2) =XOS(XI)X2. Now let rl, .. Sn be elements of T R , and let ri ... S! denote the product in Exts(R, R) of their canonical images S: in Ext1(R, R). Then si ... r! is represented in Homs(Yn , R) by the product sf . r,:, as induced from the above map 'Y. One sees immediately from the formula written above that 0
,
0
(rt ... 5': )(xo ®
0
0
•
®
Xn+l)
= X051(Xl)
0
0
•
•
tn(Xn)Xn+l.
Now let a E Tor~(R, R), (j E Tor~(R, R), and let us compute h(n ... S;+a)(a(j). Choose representatives aER®s Yp and bER®s Yq of a and ~, respectively. Then a(j is represented in R®s Y p+q by the product abo We obtain h(S:'·· r;+a)(a(j) by applying the element of HomR(R®s Y pH , R) that corresponds naturally to S{ ... S;+q to abo Clearly, the result so obtained is the same as the result one would obtain by performing the shuffling involved in forming ab on the sequence S{ , . 5;+q rather than on the arguments Xi and Yj in the product formula for abo Hence we have 0
0
* h(51'
0
•
* 5pt-q)(a{3)
•
,
* ..• 5t(p»)(a)h(5t(pt-l) * * * = '" L..J U(t)h(5t(l) ... 5t(pt-q»)({3), t
where the summation goes over all those permutations t of (1, .. p+q) for which t(l) < ...
0
0
0
,
,
h*(a)(51' •.. ,5p)
= h(5~ •.• 5:)(a).
Put A (TR) = Lp Ap(TR). Then A (TR) is the usual algebra oj the differen(6) This is the value-wise product of cochains, such as was used in
276
[11, Th. 6].
1962]
DIFFERENTIAL FORMS ON REGULAR AFFINE ALGEBRAS
395
tial forms on R, where the multiplication is defined by means of the above shuffling of the arguments ri and summing with the appropriate signatures. Our result is therefore the following. THEOREM 5.1. Let K be a commutative ring with identity, and let R be a K-projective commutative K-algebra with identity. Let S=R ®K R. Then the duality map h of Exts(R, R) into HomR(TorS(R, R), R) and the multiplication in Exts(R, R) yield, in the natural fashion, an R-algebra homomorphism h* of TorS(R, R) into the R-algebra A (T R) of the differential forms on R.
In particular, suppose that R is a regular affine K-algebra, where K is a perfect field. Then, in virtue of Theorem 2.1, the assumptions of Theorem 3.1 are satisfied, and we conclude that TorS(R, R) is finitely generated and projective as an R-module, and may be identified with the exterior R-algebra constructed over Torf(R, R). By imbedding Torf(R, R) as a direct R-module summand in a finitely generated free R-module, we see that HomR(TorS(R, R), R) is isomorphic in the standard fashion (dual of exterior algebra ~ exterior algebra over dual) with the exterior R-algebra constructed over HomR(Torf(R, R), R). Let E(TR) denote the exterior R-algebra constructed over T R. The map r---4h(r*) is an isomorphism of TR onto HomR(Torf(R, R), R). By what we have just remarked, this extends in the standard fashion to an isomorphism p of E(TR) onto HomR(TorS(R, R), R). Let rl, ... , rp be elements of T R, and let rl ... rp stand for their product in E(TR). Then we see from Theorem 5.1 that h(r'i ... r:) =p(rl ... rp). By Lemma 4.1, h is an isomorphism. Hence we conclude that the map r---4r* extends in the natural fashion to an R-algebra isomorphism of E(TR) onto Exts(R, R). Moreover, it is clear that, in the present case, h* is a monomorphism sending TorS(R, R) onto the subalgebra of A (T R) that is generated by the strongly alternating maps, i.e., by the maps that vanish whenever two of the arguments are equal. We may summarize these results as follows. THEOREM 5.2. Let R be a regular affine K-algebra, where K is a perfect field. Then TorS(R, R) is naturally isomorphic with the exterior algebra E(D R) constructed over the R-module DR of the formal differentials, and Exts(R, R) is naturally isomorphic with the exterior algebra E(TR ) constructed over the R-module T R of the K-derivations of R. These isomorphisms transport the duality map h: Exts(R, R)---4HomR(Tor S(R, R), R) into the canonical homomorphism E(TR)---4HomR(E(D R), R), which is an isomorphism because DR is finitely generated and projective as an R-module. The homomorphism h* of Theorem 5.1 becomes an isomorphism of E(D R ) onto the R-algebra of the strongly alternating differential forms.
Now suppose that K is an arbitrary field, and that F is a finitely generated separable extension field of K. Then everything we have said above concerning the regular affine K-algebra R holds equally for F, the only change in the
277
396 G. HOCHSCHILD, BERTRAM KOSTANT AND ALEX ROSENBERG [March
proof being that the appeal to Theorem 2.1 is now replaced with an appeal to Theorem 2.2. Thus we have the following result. THEOREM 5.3. The conclusions of Theorem 5.2 hold also when K is an arbitrary field and R is a finitely generated separable extension field of K.
In his thesis (Chicago, 1956), W. Ballard has obtained a part of Theorem 5.3, namely: Exts(R, R) ~E(TR). It is of interest to observe, in connection with Theorem 5.1, that the weak definition of "alternating" used in describing A (T R), rather than the usual stronger requirement on "alternating" maps, which demands that they vanish whenever two of the arguments are equal, is appropriate, in general. This is shown by the following example. Let K be a field of characteristic 2 and let R=K[a], with a 2 =uEK, but aEEK. Consider the element 1 ®s(l ®a®a®a) + 1 ® s(l ® u (Sa ® 1) ER ® s Y2. One checks immediately that this element is a cycle and thus represents an element aETor~(R, R). We have TR=Rr, where = 1. Now one verifies immediately that h*(a)(r, =a. Thus h*(a) is not alternating in the strong sense. Of course, it is clear from Theorem 5.2 that this phenomenon cannot arise when R is a regular affine algebra over a perfect field. Note that in our present example the element a does not belong to the subalgebra of TorS(R, R) that is generated by Tor~CR, R); indeed, h* must evidently vanish on Tor1CR, R)Tor1CR, R), in the present case. 6. The algebra of differential operators. Let K be a commutative ring with identity, and let R be a commutative K-algebra with identity. Let TR denote the R-module of all K-derivations of R. Clearly, TR has naturally the structure of a Lie algebra over K. We make the directR-module sum R +TR into a Lie algebra over K, defining the commutators by the formula [r1+71, r2+72]=(71h)-72h))+[71, 72], where riER, 7iETR, and [71,72] is the ordinary commutator 7172 - 7271 of the derivations 71 and 72. Let U denote the universal enveloping algebra of the K-Lie algebra R+TR, defined as the appropriate homomorphic image of the tensor K-algebra constructed over the K-module R + T R. Denote the canonical K-module homomorphism of R+TR into U by z-tz'. Let U+ denote the subalgebra of U that is generated by these elements z'. Let P denote the two sided ideal of U+ that is generated by the elements of the form r'z' - (r ·z)" where r ranges over R, z ranges over R+TR, and r·z is the r-multiple of z for the R-module structure of R+TR. We define V R as the factor K-algebra U+/P. I t is clear from this definition that the unitary V R-modules are precisely those Lie algebra modules M for the Lie algebra R+TR on which we have r·(z·m)=(r·z)·m, for all rER, zER+TR, mEM, and on which l·m=m, where 1 is the identity element of R. We shall call such modules regular (R+TR)-modules. In particular, it is clear that R is a regular (R+TR)module in the natural way, and one sees easily that the representation of R+TR on R is faithful. Hence the canonical homomorphism of R+TR into
rea)
n
278
1962]
DIFFERENTIAL FORMS ON REGULAR AFFINE ALGEBRAS
397
V R is actually a monomorphism, and we shall accordingly identify R+TR with its image in VR. Let A and B be any two regular (R+TR)-modules. We define the structure of a regular (R + T R)-module on HomR(A, B) such that, with rER, rETR, aEA, hEHomR(A, B), (r·h)(a) = r·h(a)
and
(T·h)(a) = T·h(a) - h(T·a).
The verification that the above conditions for regularity are satisfied presents no difficulties. Thus, if A and B are any two unitary VR-modules, this defines the structure of a unitary VR-module on HomR(A, B). LEMMA 6.1. Let B be a unitary V R-module, and regard V R as a V R-module in the natural fashion (the operators being the left multiplications). Then the V R-module HomR( V R, B), with the structure defined above, is isomorphic with the V R-module HomR( V R, B) in which the module structure is defined in the usual way from the right multiplications in VR, i.e., in which (u·h)(v) = h(vu) , for all hEHomR(VR, B) and all u, v in VR.
Proof. For every hEHomR(VR, B), define h*EHomR(VR, B) by h*(u) = (u·h)(l), where u·h denotes the transform of h by the element u of V R, for the first VR-module structure of HomR(V R, B). We claim that, for all u, v in VR, (u·h*)(v) = (v·h)(u). This is immediately verified from the definitions when uER. Now suppose that the result has already been established for some u and all v. Let rETR • Then we have (TU· h*) (v) = T· «u· h*) (v)) - (u· h*) (TV) = T· «v· h) (u)) - (TV· h) (u) = (v· h) (TU).
Thus our claim follows inductively for all right R-multiples u of monomials in elements of T R and hence generally for all uE V R. In particular, (u·h*)(l) =h(u), so that h**=h. Thus our map h~h* is an additive (actually K-linear) involution of HomR(V R, B). Now we have (v·h)*(u) = (uv·h)(l) =h*(uv). This means that our involution h~h* transports the VR-module structure defined originally into the VR-module structure given by (v·h)(u) =h(uv). This completes the proof of Lemma 6.1. Let A and B be unitary VR-modules. Using the induced R-module structures on A and B (written on the right or on the left, as the notation requires), we may form the tensor product A ®RB. We define the structure of a regular (R + T R)-module, and hence that of a unitary V wmodule, on A ® RB such that, for rER, aEA, bEB, rE T R , r· (a®b) = (r·a) ®b( =a®(r·b)) and r· (a®b) = (r·a) ®b+a®(r·b). It is verified in a straightforward way that these definitions indeed satisfy the conditions for a regular (R + T R)-module. LEMMA 6.2. Let B be a unitary VR-module, and regard V R as a VR-module by left multiplication and as a right R-module by right multiplication. Then the V R-module B ® R V R, with the module structure defined as above, is isomorphic
279
398 G. HOCHSCHILD, BERTRAM KOSTANT AND ALEX ROSENBERG [March with the V R-module V R ® R B in which the module structure is defined from that of V R alone, i.e., is such that U· (v®b) = (uv) ®b.
Proof. There is an evident VR-module homomorphism 1/;: VR®R B --+B®R V R such that l/;(u®b)=u·(b®l). We shall show that I/; is actually an isomorphism by exhibiting an inverse. For this purpose, we must momentarily return to the definition of V R as the factor algebra U+/P. The map r+T--+r-T is evidently an anti-automorphism of order 2 of the K-Lie algebra R+TR and induces, in the natural way, a K-linear involution u--+u* of U+ such that, for rER and TE T R, (r') * = r' and (1")*= -1", and, for arbitrary elements u and v of U+, (uv)*=v*u*. Our return to U+ is necessitated by the fact that this involution does not send the ideal P into itself, so that it does not induce an involution of V R • Let PI be the two sided ideal of U+ that is generated by the elements of the form r{rl -(rlr2)', with rl and r2 ranging over R. Write W for U+/P I and P' for P/P I • Then the map r--+r', followed by the canonical epimorphism U+--+W, is a homomorphism of R into Wand yields the structure of a two sided Rmodule on W in the natural fashion. Our involution u--+u* of U+ sends PI into itself and hence induces an involution of W which we still denote by w--+w*. If z is an element of R+TR we shall write z' also to denote the canonical image of z in W. By copying the above definitions of the VR-modules B®R V R and VR®R B with W in the place of VR, we define the W-modules B®R Wand W®R B. There is evidently a K-module homomorphism x--+x* of B®R W into W®R B such that (b®u)*=u*®b. Now define the map cp: B®R W--+W®R B such thatcp(b®u)=(u*·(b®l'»*. We claim that cp is a W-module homomorphism. Let rER. Then we have cp(r'·(b ® u»
= cp«r·b) ® u) = (u*·«r·b) ® 1'»* = (u*· (b ® r'l'»* = (u*· (b ® l'r'»* = r'· (u*· (b ® 1'»* = r'·cp(b ® u).
Let TE T R. Then we have cp(T'· (b ® u» = cp«T·b) ® u) = (u*'«1"b) ®
= -
+ cp(b ® (T'U» 1'»* -
«u*1")·(b ®
1'»*
= -
(u*·(b ® 1"1'»*
(u*· (b ® 1'1"»*
= T'·(u*·(b ®
1'»* =
1"'cp(b ® u).
This suffices to establish our claim. N ow we shall show that cp maps the canonical image of B ® R P' in B ® R W into the canonical image of P' ®R Bin W®R B. We recall that P is the two sided ideal of U+ that is generated by the elements of the form r'z' - (rz)', where rER and zER+TR• Actually, P coincides with the right ideal that is generated by these elements. In order to see this, it suffices to show that the
280
1962]
DIFFERENTIAL FORMS ON REGULAR AFFINE ALGEBRAS
399
K-subspace spanned by the elements of the form r's' - (rs)' is stable under commutation with the elements of R+ T R • Now let slER+ T R • Then we have
[z{ , r'z' - (rz)']
[z{, r']z' + r'[z{ , z'] - [z{, (rz)'] = [Zi, rl'z' + r'[Zl, z]' - [Zh rz]' = ([Zi, d'z' - ([Zi, r]z)') + (r'[Zl, z]' - (r[zh z])'), =
and this is indeed of the required form. Next we observe that (r's' - (rs)') . (b ®u) = b ® «r's' - (rz)')u) , as is easily verified directly from the definition of the W-module structure on B®R W. Since q, is a W-module homomorphism, this shows that q, maps the element on the right into the canonical image of P' ®R Bin W®R B. Hence we may now conclude that q, maps the image of B®R P' in B®R Winto the image of P'®R Bin W®R B. Hence q, induces a map 'Y of B®R V R into VR®RB. Since q, is a W-module homomorphism, it is clear that 'Y is a VR-module homomorphism. It follows immediately from this that 'Y 0 1/1 is the identity map on VR®R B. There remains to prove that 1/1 0 'Y is the identity map on B ®R VR. It is immediate that (1/1 0 'Y) (b ®u) = b ®u whenever uER. Suppose that we have already shown that this holds for some uE VR and all bEB. Then we have, with'TETR,
(1/10 'Y)(b ® (TU»
= (1/1 ° 'Y)(T· (b ® u) - (T·b) ® u) = T·(I/Io'Y)(b ® u) - (T·b) ® U =
T·(b ® u) - (T·b) ® U = b ® (TU).
Hence it follows by an evident induction on the "degree" of u, written as a polynomial in elements of T R, that 1/1 'Y is indeed the identity map. This completes the proof of Lemma 6.2. Now let us assume that K is a field and that R is an affine K-algebra. Let Q denote the field of quotients of R. Since every K-derivation of R extends uniquely to a K-derivation of Q, we have a natural injection Tw-~TQ, and hence a canonical Q-linear map Q® R T r-+ T Q. Since R is finitely ring-generated over K, we see immediately that this map is an epimorphism. Since Q is Rfiat, the map Q®R TR~Q®R T Q induced by the injection TR~TQ is a monomorphism. Evidently, the canonical map Q®R TQ~TQ is an isomorphism. Thus the canonical map Q®R TR~TQ is an isomorphism. We shall identify Q®R TR with T Q whenever convenient. The injection R+TR~Q+TQ is both a Lie algebra homomorphism and an R-module homomorphism, and hence extends uniquely to a K-algebra homomorphism VR~ V g • This induces a canonical Q-linear epimorphism Q®R VR~ V Q. Now we note that Q is a unitary VR-module in the natural fashion, so that we may equip Q®R VR with the structure of a unitary V Rmodule, in the manner explained above. The induced R-module structure coincides with the R-module structure induced by the natural Q-module
°
281
400
G. HOCHSCHILD, BERTRAM KOSTANT AND ALEX ROSENBERG
[March
structure of Q®R VR. We define a product on Q®R V R by means of this V Rmodule structure and the natural Q-module structure, the definition being such that
It is not difficult to verify, by the same kind of induction we used in the proof of Lemma 6.2, that this gives the structure of an associative K-algebra on Q®R VR. Moreover, it is then clear that our Q-linear epimorphism Q®R VR -+ V Q is actually a K-algebra epimorphism. For every non-negative integer m, let VB denote the R-submodule of V R consisting of the elements that can be written as sums of products of elements of Rand T R , where each product has at most m factors from T R • If m is a negative integer, let VB= (0). LetG(VR ) be the graded R-algebra VB/VR'-I obtained from this filtration of V R; G( V R) is indeed an R-algebra, and not only a K-algebra, because the commutation with an element of R sends each VB into VB-I. Let S(TR ) denote the symmetric R-algebra built over the Rmodule T R. We have an evident natural R-algebra epimorphism S(TR) -+G(VR). In the same way, we define G(VQ), S(TQ) , and we have the natural Q-algebra epimorphism S(TQ)-+G(VQ).
:Em
THEOREM 6.1. Let R be an affine K-algebra, where K is an arbitrary field, and let Q be the field of quotients of R. Let F be an arbitrary extension field of K, regarded as a K-algebra. Then the canonical epimorphisms S(TF)-+G(VF) and Q®R V R-+ V Q are isomorphisms.
Proof. The first part of the theorem is analogous to the Poincare-BirkhoffWitt Theorem for universal enveloping algebras of Lie algebras and can be proved by the method in [4, Lemma 3.5, p. 272]: if (T;) is an ordered F-basis for T F, it is clear from the definition of V F that every element of V F can be written as an F-linear combination of ordered monomials Til' • . Ti.; il~ ... ~in. Now one shows inductively that S(Tp) can be equipped with the structure of a regular (F+Tp)-module such that the transform of lES(T F) by vE V F is precisely the element of S(T F) that is represented by the above standard expression for v. This evidently implies that the representation of the elements of V p in this standard form is unique and that the canonical map SeT F)-+G( V p) is an isomorphism. In order to prove the second part, let us observe first that the isomorphism Q ® R T R-+ T Q extends canonically to a Q-algebra isomorphism Q ® R S( T R) -+S(TQ), because Q®R S(TR) may be identified with S(Q®R TR)' The natural homomorphism V R-+ V Q induces an R-algebra homomorphism G(VR) -+G(VQ) which, in turn, induces a Q-algebra homomorphism Q®RG (VR) -+G(VQ). Thus we have an exact and commutative diagram of homomorphisms
282
1962]
DIFFERENTIAL FORMS ON REGULAR AFFINE ALGEBRAS
401
o i O~S(TQ)--~)G(VQ)~O
i
i
Q ®R S(TR) - Q ®R G(VR) - 0
i
o which shows that the homomorphism Q®R G(VR)-G(VQ) is an isomorphism. Now let x be an element of Q®n V R such that the image of x in VQ is O. If x~O, let m be the lowest non-negative integer such that x belongs to the canonical image of Q® R VB in Q® R V R. Then x represents a nonzero element of Q®R CV~/V~-l) CQ®R G(Vn ). Since the map Q®R G(VR)-G(VQ) is a monomorphism, this contradicts the assumption that the image of x in V Q is O. Hence the epimorphism Q®n V R- VQ is indeed an isomorphism, and Theorem 6.1 is proved. THEOREM 6.2. In the notation of Theorem 6.1, assume that TR is projective as an R-module. Then the natural homomorphism V n- V Q is a monomorphism, and the canonical epimorphism S( Tn) -G( V n) is an isomorphism.
Proof. Since Tn is R-projective, so is S(T n), as is seen by imbedding TR as a direct R-module summand in a free R-module. Hence the natural homomorphism S(TR)-Q®R S(Tn) =S(TQ) is a monomorphism. Now we consider the exact and commutative diagram of homomorphisms
o -S(TQ) -
G(VQ) - 0
i S(T
i R) -
G(Vn) - 0
i
o This shows immediately that the epimorphism S(Tn)-G(Vn ) is an isomorphism and that the homomorphism G(Vn)-G(VQ) is a monomorphism. The last fact implies as above that the homomorphism V n - VQ is a monomorphism, so that Theorem 6.2 is proved. 7. The Ext functor for the algebra of differential operators. If M is any module for a commutative ring, we denote by E(M) = Lp Ep(M) the exterior algebra constructed over M. We deal with an affine algebra R over a field K, and we let Q denote the field of quotients of R. Consider the natural V Qmodule VQ®QE(TQ), where u'(v®e) = (uv) ®e. This is graded by the submodules VQ®Q Ep(TQ). Exactly as for the analogous situation of the universal
283
402 G. HOCHSCHlLD, BERTRAM KOSTANT AND ALEX ROSENBERG [March enveloping algebra of a Lie algebra [4, Th. 7.1, p. 280], we define a homogeneous VQ-endomorphism d of degree -1 on this module such that d 2 = O. This endomorphism is given by the formula p
d(v ® TO •.• Tp) =
L
(_1)i(VT;) ® TO ..• ~; ... Tp
+L
(-l)rl-· v ® [T" T.]TO •.• ~r
•••
~8
•••
Tp,
r<.
where the T; are arbitrary elements of T Q. The definition enforces this formula in the case where the Ti belong to a given Q-basis of T Q, and then one verifies easily that the formula holds generally. We augment this complex by the natural VQ-module epimorphism VQ®Q EO(TQ) = VQ~Q, where the image of vE V Q in Q is the transform v·l of 1EQ by v, according to the natural V Qmodule structure of Q. Since, by Theorem 6.1, G(VQ) is isomorphic with S(TQ) , the usual filtration argument [4, pp. 281-282] shows that the augmented complex VQ®Q E(TQ) is acyclic and is therefore a VQ-free resolution of the VQ-module Q; a sketch of this argument is included in what follows. Note that this conclusion holds more generally for an arbitrary extension field of K in the place of Q. Now let us assume that TR is R-projective. Then the same holds for ~(TR). By Theorem 6.2, the natural homomorphism VR~ VQ is a monomorphism. Hence the induced homomorphism VR®R E(TR)~ VQ®R E(TR) is a monomorphism. Now VQ®RE(TR) VQ®Q(Q®RE(TR)) = VQ®QE(Q®R T R) = VQ®Q E(TQ). Taking these identifications into due account, we see that our result means that the map VR®R E(TR)~ VQ®Q E(TQ) that is induced by the natural maps VR~ VQ and E(TR)~E(TQ) is a monomorphism. It is seen immediately from the explicit formula for the boundary map d on VQ®Q E(TQ) that the image of VR®R E(TR) is stable under d. Hence d induces a boundary map (still denoted d) on VR®R E(TR), which satisfies the same explicit formula. We have a filtration of this complex by the Rsubcomplexes ~-q®REq(TR) (which, since E(TR) is R-projective, may be identified with their canonical images in VR®R E(TR)). Now the associated graded complex may evidently be identified with the complex G(VR) ®R E(TR) and hence, using Theorem 6.2, with the complex S(TR) ®R E(TR). The boundary operator induced by d (still denoted d) IS given by the formula
=
Lq
p
d(u ® TO ••. Tp) =
L
(_1)i(UT;) ®
TO • • •
~;
•••
Tp.
;=0
Now let F be a free R-module containing TR as a direct R-module summand. Then our complex S(TR) ®R E(TR) is a direct R-complex summand of the usual Koszul complex S(F) ®R E(F). We shall prove from this that the augmented complex VR®R E(TR) has an R-homotopy.
284
DIFFERENTIAL FORMS ON REGULAR AFFINE ALGEBRAS
1962]
403
As is well known, the augmented complex S(F) ® R E(F) has an R-homotopy sending R onto SO(F) ®R EO(F) and each Sp-q(F) ®R Eq(F) into Sp-q-l(F) ®R Eq+l(F) , [8, p. 259]. Since S(TR) ®R E(TR) is a direct R-complex summand of S(F) ®R E(F), the augmented complex S(TR) ®R E(TR) has an induced R-homotopy h sending R onto SO(TR) ®R EO(TR) and each Sp-q(TR)®REq(TR) into Sp-q-l(TR)®REq+l(TR). Let X stand for the augmented complex VR®R E(TR). For p ~O, let Xp stand for the augmented subcomplex L:q V~-g®R Eq(TR), and put Xp= (0), for p
L: Xp/Xp_1 = L: GP(X) p
= G(X).
p
Thus we may regard h as an R-homotopy of G(X) under which each Gp(X) is stable. N ow consider the natural R-module epimorphism X p-tX p/X p-l = Gp(X). Since Xp/X p_1 is isomorphic, as an R-module, with L:q Sp-q(TR) ®R Eq(T R), for p>O, and since X o is isomorphic, as an R-module, with SO(TR) ®R EO(TR) +R=R+R, we know that each Xp/X p_1is R-projective. Hence we can make a direct R-module decomposition X p = X p-l + Y p, and X p is the direct Rmodule sum L:up Yq. If Olp denotes the natural R-module epimorphism Xp-tXp/Xp_1 then the restriction of Olp to Yp is an isomorphism, and we may define an R-module isomorphism 01: X-tG(X) by making Ol=Olp on the component Yp. In particular, it follows that X is projective as an R-module. We have Olpd=dOlp. Moreover, Ol-Olp evidently sends Xp into L:q
'Yd
+ d'Y
- 1 = Ol-lh(Old - dOl)
+ (dOl- 1 -
Ol-ld)hOl.
Hence we find that (-yd+d'Y-1)(Xp) CXp - 1• Furthermore, each Xp is stable under 'Y. Now we have - (-yd+d'Y-1)2='Y'd+d'Y' -1, where 'Y' = 2'Y-'Yd'Y-d'Y2. Hence ('Y'd+d'Y' -1)(Xp) CXp- 2 , and ('Y' -'Y)(Xp) CXp _ 1• Iteration of this process leads to a sequence of R-endomorphisms 'Yk of X such that ('Ykd+d'Yk -l)(Xp) CXp_2k, and ('Yk+1-'Yk)(Xp) CXp_2k. Since X q= (0), for q < 0, 'Yk+r -'Yk annihilates X 2k -1, for all r ~ O. Hence there is an R-endomorphism on X such that coincides with 'Yk on X2k_l' for each k, and we have rd+dr= 1, i.e., r is an R-homotopy of X. Thus we have the following result.
r
r
THEOREM 7.1. Let R be an affine algebra over a field K, and suppose that T R is R-projective. Then the complex VR®R E(TR), as defined above, is a VR-projective resolution of the V R-module R and has an R-homotopy. If F is an arbitrary extension field of K, regarded as a K -algebra, then V F® F E (T ]1') is a V rfree resolution of the V F-module F.
285
404 G. HOCHSCHILD, BERTRAM KOSTANT AND ALEX ROSENBERG [March Note. It is clear from our proof that Theorem 7.1 holds more generally for any integral domain Rover K such that TR is R-projective; one must merely replace T Q with the canonical image of Q®R TR in T Q• As was pointed out to us by A. Shapiro, further generalizations, covering cases where R is not an integral domain (e.g. the algebra of all differentiable functions on a differentiable manifold), are easily obtainable by simultaneous use of a suitably large family of localizations of R. Naturally, these generalizations extend to Corollary 7.1 below. Let P = R, or P = F, as in Theorem 7.1, and let A be a unitary V p-module. By Theorem 7.1, we have Extvp(P, A) =H(Homvp(Vp®p E(Tp), A». We may identify Homvp(Vp®p E(T p), A) with Homp(E(Tp), A). Let ~ be the coboundary map in Homp(E(Tp), A) that is induced by the boundary map don Vp®pE(Tp). Then, writing the elements of Homp(E(Tp) , A) as strongly alternating maps with arguments in T p and values in A, we have P
(6f)(TO, ... , Tp)
=
E (_1)iTi(f(TO, ... , Ti, ... , Tp» i=O + E (-1)>+1([T T.], TO, ... , T r,
.<.
r , ••• ,
T., ... , Tp),
iffEHomp(Ep(Tp) , A). Thus we see that Extvp(P, A) is naturally isomorphic with the usual cohomology space based on the strongly alternating A-valued differential forms. We state this formally, for reference. COROLLARY 7.1. Let P = R, or P = F, as in Theorem 7.1. Then the cohomology K-space based on the strongly alternating differential forms with values in a unitary Vp-module A may be identified with Extvp(P, A).
If we assume that P is either a regular affine K-algebra, where K is a perfect field, or that K is an arbitrary field and P is a finitely generated separable extension field of K, we may appeal to Theorem 5.2, or to Theorem 5.3, respectively, to see that the complex of the strongly alternating P-valued differential forms may be identified with the complex E(Dp) of the formal differentials, whose differential operator is the canonical extension of the map x~dx of Pinto D p. Hence, in this case, the cohomology K-space based on the formal differentials of the K-algebra P may be identified, by Corollary 7.1, with Extvp(P, P)(6). 8. The homological dimension of the algebra of differential operators. THEOREM 8.1. Let F be a finitely generated extension field of an arbitrary field K. Then the global homological dimension d( V F) of V F is equal to the (8) It may be of interest to point out that, in a rather different situation, namely when P is a finitely generated purely inseparable extension field of exponent 1 of K, the cohomology K-algebra of the strongly alternating P-valued differential forms has been determined explicitly by P. Cartier in [5].
286
1962]
DIFFERENTIAL FORMS ON REGULAR AFFINE ALGEBRAS
405
dimension over F oj the space T F oj all K-derivations oj F, and is equal to the projective dimension dVF(F) oj the V rmodule F.
Proof. LetMbe any unitary Vrmodule. Then the resolution VF®F E(T F) of F, as in Theorem 7.1, dualizes into an exact sequence of V F-modules and V F-homomorphisms (0) ~ M
= HomF(F, M)
~
HomF(V F ®F EO(TF), M)
~
HomF(V F ®F EI(T F), M)
~
....
As a VF-module, each HomF(VF®FEp(T F), M) is isomorphic with a direct sum of a finite number of copies of HomF(V F, M), with the module structure defined just above the statement of Lemma 6.1. However, by Lemma 6.1, this VF-module is isomorphic with the VF-module HomF(V F, M) with the module structure given by (u· h)(v) = h(vu). As is well known, and easy to show directly (using that F is a field), this module is V F-injective. Hence we conclude that the above exact sequence is a V F-injective resolution of M. Since Ep(T F) = (0) when p exceeds the F-dimension of T F, it follows that d(V F ) does not exceed the F-dimension of T F • Now let n denote the F-dimension of TF and let A be any unitary V Fmodule. There is a basis fl' ... , fn for TF over F and elements Xl, ... ,Xn in F such that f.(x;) = Oi; [7, Lemma 2.1]. In particular, we have [fi, f;] = 0, for all i and j. If we use the resolution VF®FE(T F) of F, the K-space ExtVF(F, A) appears as the factor space Homp (En(TF) , A)/o(HomF (En-I(T F) , A).
Using the explicit formula for 0, as given below the statement of Theorem 7.1, and taking account of the fact that [fi,f;] = 0, we see that o(HomF(En-I(T F), A) ~ TF"A. On the other hand, HomF(En(T F) , A) ~A. Hence we see that ExtVF(F, A) is isomorphic with A/TF"A. Taking A = V F, we see from this that Exth(F, V F) ~ F~ (0). Hence we have indeed d( V F) =dvF(F) =n. Now let R be an affine K-algebra such that TR is R-projective. In this case, it appears that the relative global homological dimension d( V R, R) is more easily accessible than d(VR ). We refer to [8] for the requisite notions of
relative homological algebra, but we recall that the relative homological notions for (VR, R) are obtained from the corresponding notions for V R simply by replacing "exact sequence of VR-homomorphisms" with "R-split sequence of VR-homomorphisms" throughout. The ordinary projective dimension of a VR-module M will be denoted by dvR(M) , and the relative projective dimension of M will be denoted by d(vR.R)(M). We shall prove the following result. THEOREM 8.2. Let K be a field and let R be an affine K-algebra such that T R is R-projective. Let Q be the field oj quotients oj R, and let n be the Q-dimension oj T Q• Then, jor every unitary VR-module A, the canonical map (induced
287
406
G. HOCHSCHILD, BERTRAM KOSTANT AND ALEX ROSENBERG [March
from maps of the appropriate resolutions) of ExtwR.R)(R, A) into ExtvR(R, A) is an isomorphism. We have dVR(R) =dWR.R)(R) =d(Vn, R) =n. If R is regular then d( V R) ~ n +t, where t is the transcendence degree of Q over K.
Proof. By Theorem 7.1, VR®R E(TR) is both a VR-projective resolution of R and a (VR, R)-projective resolution with R-homotopy. Hence both ExtvR(R, A) and Ext(VR.R)(R, A) can be computed as H(Homv R (VR ®R E(TR),
A»,
which establishes the first assertion of our theorem. We know that Q®R E(TR) is isomorphic with E(TQ). Since E(TR) is R-projective, the natural map E(TR)~Q®R E(T R) is a monomorphism. Hence we have Ep(TR) = (0) for p>n. Hence it is clear from the resolution VR®R E(TR) of R that dVR(R) =d(VR.R)(R) ~n. Now let us consider the same sequence we used at the beginning of our proof of Theorem 8.1: (0)
~
M = HOffiR(R, M)
~
HOffiR(V R ®R EO(T R), M)
~
HOffiR(V R ®R El(T R), M)
~
....
By Theorem 7.1, the resolution VR®R E(TR) of R has an R-homotopy, which evidently induces an R-homotopy of the above dual sequence. On the other hand, we see from Lemma 6.1, as in the proof of Theorem 8.1, that each HomR(VR®R Ep(TR), M) is isomorphic with a direct VR-module summand of a direct sum of VR-modules HomR(V R, M) with the module structure given by (v·h)(u) =h(uv). By [8, Lemma 1], this last VR-module is (VR, R)-injective. Hence the above sequence is a (VR, R)-injective resolution with Rhomotopy of M. Hence it is clear that d( V R, R) ~ n. Now we claim that, for every VQ-module M, ExtvR(R, M) is isomorphic with ExtvQ(Q, M). By Theorem 6.1, V Q is isomorphic with Q®n V R • It is clear from the definition of the algebra structure of Q®R Vn that this isomorphism transports the right Vn-module structure of Q®R VR into the right VR-module structure of VQ obtained from the natural map of V R into V Q. Since Q is R-flat, this implies that, as a right VR-module, VQis Vn-flat. Hence, if X is any VR-projective resolution of R, VQ®v R X is a VQ-projective resolution of VQ®VR R. Using the natural VQ-module structure of Q, we obtain a VQ-module epimorphism VQ®Vn R~Q sending v®r onto v·rEQ. It is easy to verify that this is actually an isomorphism, so that we may identify VQ®VR R with Q, as a VQ-module. Thus VQ®VR X is a VQ-projective resolution of Q. Since HomvQ(VQ®vR X, M) may be identified with Homvn(X, M), this establishes our claim. Now it is clear from Theorem 8.1 (applied to Q) that dVR(R) ~n. We have shown that n~dVR(R) =dWR.R)(R) ~d(VR, R) ~n, so that all but the last statement of Theorem 8.2 is proved. As a by-product of our proof of the existence of an R-homotopy in
288
1962]
DIFFERENTIAL FORMS ON REGULAR AFFINE ALGEBRAS
407
Vn®n E(Tn), we had obtained the result that this complex is R-projective. In particular, VI.: is R-projective as a left R-module; in fact, as an R-module, Vn is isomorphic with G(Vn ) ~S(TR). Similarly, Vn is R-projective as a right R-module. Hence we may apply [9, Th. 1] to conclude that, for every unitary Vn-module N, dVll(N) ~ d(Vll.R)(N)
+ dn(N).
If R is a regular affine K-algebra, we have d(RM) = t, for every maximal ideal M of R, by [2, Ths. 1.9, 1.10]. Since d(R) =maxM(d(R M» [4, Ex. 11, p. 142; 1, Th. 1], we have therefore d(R) =t. Hence the above results give d(Vn) ~d(Vn, R)+d(R) =n+t. This completes the proof of Theorem 8.2. 9. The product for Extvp(P, *). Let K be a field, and let P be a K-algebra which is either an affine K-algebra with Tp P-projective or an arbitrary extension field of K. If A and B are unitary Vp-modules there is a product Extvp(P, A) ®K Extvp(P, B)
~
Extvp(P, A ®P B)
which is defined as follows. Let X be any V p-projective resolution of P. Noting that V p is P-projective, so that X is P-projective, and appealing to Lemma 6.2, we see that the Vp-module X®p X is still Vp-projective. Moreover, since X is also a P-projective resolution of P, we have H(X®p X) = TorP(P, P), whence H(X®p X) has its components of positive degree equal to (0). Hence X®p X is still a Vp-projective resolution of P®p P=P. Hence the natural K-space homomorphism
cf>: Homvp(X, A) ®K Homvp(X, B)
~
Homvp(X ®P X, A ®P B),
where q,(j®g)(u®v) =f(u) ®g(v), induces a product for Extvp(P, *), as indicated above. It is seen as usual that this product is associative and skewcommutative. In order to make this product explicit, we require a map of the complex X into the complex X®pX, when X=Vp®pE(Tp). By imbedding T p as a direct P-module summand in a free P-module, we see easily that there is a P-module homomorphism E"'(Tp)~Ep(Tp) ®P En-p(T p) sending each product of elements of Tp onto
rl ... r.. L, q(t)tl(l) ... tl(p) ® tl(p-tl) ... tt(nh
where the summation goes over all permutations t of (1, ... , n) for which t(l)< ...
V®rl ... s..
v-(E( ~ q(t)(l ® r'(l)· .. tt(p»
® (1 ®t'(p-tl) ... teen»)).
This map is evidently a Vp-module homomorphism. It is rather tedious to verify that it commutes with the boundary maps on X and X ® p X, but no
289
408 G. HOCHSCHILD, BERTRAM KOSTANT AND ALEX ROSENBERG difficulties other than those of notation are encountered in carrying out this verification by induction on the degree in X. On the other hand, it is clear from the definition of this map that our product for Extvp(P, *) is the same product as that obtained from the usual shuffle product of alternating differential forms, via the identification of Corollary 7.1. REFERENCES 1. M. Auslander, On the dimension of modules and algebras. III, Nagoya Math. J. 9 (1955), 67-77. 2. M. Auslander and D. Buchsbaum, Homological dimension in local rings, Trans. Amer. Math. Soc. 85 (1957), 390--405. 3. H. Cartan et C. Chevalley, Seminaire 1955/56, Ecole Norm. Sup. Paris, 1956. 4. H. Cartan and S. Eilenberg, Homological algebra, Princeton Univ. Press, Princeton, N. J., 1956. 5. P. Cartier, Questions de rationalite des diviseurs en Geometrie algebrique, Bull. Soc. Math. France 86 (1958), 177-251. 6. C. Chevalley, Introduction to the theory of algebraic functions of one variable, Math. Surveys, Amer. Math. Soc., Providence, R. I., 1951. 7. G. Hochschild, Double vector spaces over division rings, Amer. J. Math. 71 (1949), 443460. 8. - - - , Relative homological algebra, Trans. Amer. Math. Soc. 82 (1956), 246-269. 9. - - - , Note on relative homological dimension, Nagoya Math. J. 13 (1958), 89-94. 10. E. Kunz, Die Primidealtheiler der Differenten in allgemeinen Ringen, J. Reine Angew. Math. 204 (1960), 165-182. 11. A. Rosenberg and D. Zelinsky, Cohomology of infinite algebras, Trans. Amer. Math. Soc. 82 (1956), 85-98. 12. J .-P. Serre, Sur la dimension des anneaux et des modules noetheriens, Proc. Int. Symp. on Algebraic Number Theory, Tokyo-Nikko, 1955, pp. 175-189. 13. J. T. Tate, Homology of Noetherian rings and local rings, Illinois J. Math. 1 (1957), 14-27. 14. O. Zariski and P. Samuel, Commutative algebra, Vol. II, Van Nostrand, Princeton, 1960. UNIVERSITY OF CALIFORNIA, BERKELEY, CALIFORNIA NORTHWESTERN UNIVERSITY, EVANSTON, ILLINOIS
290
Reprinted from ILLINOIS JOURNAL OF MATHE)'1ATIC'S Vol. 6, No.2, June 1962 Printed in U.S.A.
DIFFERENTIAL FORMS AND LIE ALGEBRA COHOMOLOGY FOR ALGEBRAIC LINEAR GROUPS BY
G.
HOCHSCHILD AND
B.
KOSTANT
1. Introduction In the study of the rational cohomology theory of algebraic linear groups, the differential forms, constructed from the algebra of the rational representative functions on the group, playa major role in providing the link between the group cohomology and the Lie algebra cohomology [5). Moreover, the cohomology of the differential forms has some significance as an algebraic geometric invariant. For instance, it follows from [5, Theorem 4.1) that, if R is the algebra of the rational representative functions on an irreducible algebraic linear group G over a field F of characteristic 0, the cohomology of the differential forms of R is trivial (if and) only if R is an ordinary polynomial algebra over F. Our main purpose here is to extend the theory of these differential forms to the case where the group G is replaced by a "homogeneous space" G/K, K being a fully reducible algebraic subgroup of G. This means that R is replaced by the subalgebra RK of R consisting of the functions that are constant on the cosets of K in G. If the base field F is algebraically closed, the designation of G/K as a homogeneous space is actually justified: we shall see (Theorem 5.1) that G/K has then the structure of an affine algebraic variety, with RK as its algebra of polynomial functions. The connections between the cohomology of the differential forms, the rational cohomology of the group, and the Lie algebra cohomology extend to this case, with the relative Lie algebra cohomology taking the place of the ordinary Lie algebra cohomology (Theorems 3.1, 3.2, :3.:3, 4.2). In order to get the full information here, it was necessary to extend the known tensor product decomposition theory for Lie algebra cohomology to the relative case (Section 4), and this may be of independent interest. For an arbitrary unitary F-algebra P, there are two known constructions giving a complex of "differential forms." One of these is based on the F-derivations of P (Section 3), while the other is quite direct and purely formal K (Section 5). We show that, for P = R , the two complexes thus obtained are naturally isomorphic (Theorem 5.2). Actually, this result holds under more general circumstances; results of Kunz [6) are relevant here. We take this opportunity to thank M. Rosenlicht for his help in some clarifying discussions on the topic of differential forms. 2. The algebra of the rational representative functions Let G be an irreducible algebraic linear group over an infinite field F. We denote by R(G), or simply by R, the F-algebra of the rational representative Received February 17, 1961. 264
B. Kostant, Collected Papers, DOI 10.1007/b94535_15, © Bertram Kostant 2009
291
DIFFERENTIAL FORMS AND LIE ALGEBRA COHOMOLOGY
265
functions on G, i.e., of the F-valued rational functions defined at every point of G and with the property that their translates under the action of G span only a finite-dimensional space of functions. Equivalently, the rational representative functions are those rational functions which remain everywhere defined under arbitrary extensions of the base field F. If G is given concretely as an algebraic group of linear transformations of determinant 1, then R coincides with the algebra of all polynomial functions on G. The field of quotients, Q, of R is the field of the rational functions on G. We consider the action of G by left translations on R and on Q; for feR, and x and y in G, the left translate x-f of fis defined by (x·f)(y) = f(yx). Let @ denote the Lie algebra of G. The action of G on R by left translations induces an action of @ on R and hence on Q by which @ acts as a Lie algebra of F-derivations. In this way, @ becomes identified with the Lie algebra of all F-derivations of R (or of Q, by canonical extension of derivations) that commute with the right translations f ~ f·x, where (f·x) (y) = f(xy). The Q-space of all F-derivations of Q is spanned by @ and in fact is canonically isomorphic with Q ® @. Moreover, the R-module of all F-derivations of R is canonically isomorphic with R ® @ (see [5, Lemma 4.1], and [1]). LEMMA 2.1. Let K be an algebraic subgroup of the irreducible algebraic linear group G, and let QK denote the subjield of Q consisting of the elements left jixed by the left translations from K. Then QK separates the cosets xK of K in G. Furthermore, every element of @ that annihilates QK belongs to the Lie algebra ~ ofK.
Proof. Let I denote the ideal of R that is associated with K. We can evidently find F-linearly independent elements fl , ... , fn of I such that I = Rfl + ... + Rfn and Ffl + ... + Ffn is stable under the action of K by left translations. Let V denote the smallest G-submodule of R (under the action of G by left translations) that contains all the j/s. Let W be the homogeneous component of degree n of the exterior F-algebra built over V. We consider the action of G and @ on W that is induced by the action of G on V. Let w be the exterior product, in W, of fl , ... ,fn. Then Fw is evidently a K-stable I-dimensional subspace of W. If 'Y is any element of the dual space to W, we define the element 'Y/w of R by ('Y/w)(x) = 'Y(x·w), for every x e G. If a and (3 are elements of the dual space to W, and if (3/w ;;e 0, the quotient (a/w)/((3/w) is immediately seen to be an element of QK, as a consequence of the fact that Fw is K-stable. Now let y be an element of G that does not belong to K. Then y·w f. Fw, because otherwise y·I c I, which implies that y e K. Hence we can choose a and (3 from the dual space to W such that a(w) = 1, a(y·w) = 0, (3(w) = 1, (3(y·w) = 1. Then (a/w)/((3/w) is defined at the points 1 and y of G and takes the values 1 and at these points, respectively. Thus QK separates the cosets yK and K, which proves the first part of Lemma 2.1. Now let.l be an element of @ that does not belong to~. Then .I(w) f. Fw, because otherwise .I(I) c I which implies that .I E~. Hence we can choose
°
292
266
G. HOCHSCHILD AND B. KOSTA NT
aand/3fromthedualspacetoWsuchthata(w) = 1,aU"(w)) =O,/3(w) = 1, /3(t(w)) = 1. The transform of (a/w)/(/3/w) by t is 2
((a/t(w)) (/3/w) - (a/w)(/3(t(w)) )/(/3/w) 0/= O.
This establishes the second part of Lemma 2.1. LEMMA.
2.2.
The tran8cendence degree of Q over QK i8 equal to the dimen8ion
ofK. Proof. Q is a finitely generated separable extension of QK, whence the transcendence degree of Q over QK is equal to the dimension of the Q-space of the QK-derivations of Q. This space contains the canonical isomorphic image of Q ® sr whose dimension is equal to the dimension of sr over F, which is equal to the dimension of K. Hence it suffices to show that every QK-derivation T of Q belongs to the canonical image of Q ® sr. Like every F-derivation of Q, T is of the form Li qi ti , with qi ~ Q and ti ~ ®. Subtracting an element of Qsr and multiplying by a nonzero element of R, we obtain a QK_ derivation of Q that has the form Li fi ti , where the 1;'s are elements of R and the t/s are elements of ® that are linearly independent mod sr. Now what we have to show is that the 1;'s are O. We have Ldi ti(q) = 0, for every q ~ QK. Translating from the right with an arbitrary element x of G, and noting that QK· X = QK, we find that Li (f;-X)ti(q) = 0, for every q ~ QK. Let q f QK, and let y be an element of G at which each ti(q) is defined. Let z be an arbitrary element of G. Then we have Ldi(Z)ti(q)(y)
=
(Li (f;-zy-1ni(q)(y) =
o.
Thus we may conclude that Ldi(Z)ti(q) = 0, for every q f QK. By the second part of Lemma 2.1, this implies that Ldi(Z)tifsr, whence each fi(z) = O. This completes the proof of Lemma 2.2. PROPOSITION 2.1. Let G be an irreducible algebraic linear group over the infinite field F, and let K be an algebraic 8ubgroup of G. If K i8 unipotent, or if F i8 of characteri8tic 0 and K i8 fully reducible, then the field of quotient8 of RK coincide8 with QK.
Proof.
We must show that, if q is any nonzero element of QK, (RqnR)K 0/=
o.
This is evidently the case if K is unipotent, because in that case every element of R generates a unipotent K-module. Now suppose that F is of characteristic 0 and K is fully reducible. Let V be a nonzero simple K-submodule of Rq n R. Let V' be the dual K-module HomF( V, F). Let S be the space of the representative functions on K that are associated with V'. There is a K-module monomorphism of V' into the direct sum of a finite number of copies of S. Since V is simple, so is V', and we may conclude that V'is isomorphic, as a K-module, with a submodule
293
DIFFERENTIAL FORMS AND LIE ALGEBRA COHOMOLOGY
267
of S. Now observe that the restriction of functions on G to K yields a K-module epimorphism of R(G) onto R(K). Since K is fully reducible and F is of characteristic 0, R( G) is semisimple as a K-module. Hence there is a K-module monomorphism R(K) ~ R(G) that is inverse to the restriction map. Composing this with a K-module monomorphism V' ~ S, we obtain a K-module monomorphism cp of V' into R. Now let VI, ... , Vn be a basis for V, and let fl , ... , fn be the dual basis for V'. We choose the Vi so that VI (1) = 1, while Vi( 1) = 0, for every i > 1. Put gi = CP(fi). Then, for every x € G, we have Lf~l (gi·X)Vi € (Rq n R)K. Choosing x so that gl(X) ~ 0, we ensure that this function is not o. This completes the proof of Proposition 2.1. PROPOSITION 2.2. Under the assumptions of Proposition 2.1, every F-derivation of RK into R extends to an F-derivation of R.
Proof. Let T be an F-derivation of RK into R. Then T extends canonically to an F-derivation of the field of quotients of RK into Q, i.e., by Proposition 2.1, to an F-derivation of QK into Q. This, in turn, can be extended to an F-derivation of Q, because Q is separable over QK. Let tl, ... , t. be a basis for ~, and extend this to a basis tl, ... , t8+t for ®. There is an F-derivation of Q extending T that has the form L~=l qi t8+i , with qi € Q. By Lemma 2.2, the transcendence degree of QK over F is equal to t. Since Q is finitely generated and separable over F, so is QK, and every finite system of generators for QK over F contains a separating transcendence base for QK over F. Hence there are elements Ul , . . . , Ut in RK that form a separating transcendence base for QK over F. Since Q is separable over QK, we can complete this to a separating transcendence base Ul , . . . , Ut+8 for Q over F. Then no nonzero F-derivation of Q annihilates each Ui, and it follows that the determinant formed with the tj( Ui) is different from o. On the other hand, t j( Ui) = 0 whenever both i ~ t and j ~ s. Hence the determinant formed with the t8+k( Ui), where i and k range from 1 to t, is not equal to O. Let D denote this determinant. Now we have
for each i
~
t.
Hence Dqk € R, for each k. Let J denote the ideal of all f € R such that fqk € R, for each k. We have just shown that D € J. On the other hand, we may evidently replace Ul, .•• , Ut by Ul· x, ... , Ut· x, where x is an arbitrary element of G. This replaces D by D· x, so that we conclude that D·x € J, for every x € G. Hence J has no zero on G. If F is algebraically closed, every proper ideal of R lies in the kernel of some F-homomorphism R ~ F. On the other hand, every such homomorphism is of the form f ~ f(x), with some x € G. Hence we conclude that, if F is algebraically closed, J = R, i.e., qk € R, for each k, so that our extended T actually sends R into R. In the general case, we extend G and K canonically . closure F * of F; let GF* , K F* denote these extended groups. over the algebraIC
294
268
G. HOCHSCHILD AND B. KOSTA NT
Then we have R(G F") = R(G) ® F*
and
(R(GF*»KF* = R(G)K ® F*.
Hence it is clear that Proposition 2.2 for (G F", KF*) implies Proposition 2.2 for (G, K). This completes the proof. Observe that Proposition 2.2 implies that the canonical map of R ® (@jSl') into the R-module of the F-derivations of RK into R is an epimorphism. The argument we made at the end of the proof of Lemma 2.2 shows that this map is also a monomorphism. Thus the R-module of all F-derivations of RK into R is canonically isomorphic with R ® (@jSl'). Regard @jSl' as a K-module via the adjoint representation of G on @, and regard R ® (@jSl') as the tensor product of the K-modules R (by left translation) and @jSl'. Then we have the following result. COROLLARY 2.1. Under the assumptions of Proposition 2.2, the RK-module of all F-derivations of RK is canonically isomorphic with (R ® (@jSl'»K.
Proof. Let Ldi ® Si be an element of R ® (@jSl'), where fi f Rand Si f @jSl'. Let g f RK. Then we have, for every x f K, X· CLdi Si(g»)
=
Li (X-!i)X·Si(g)
= Li (x·fi)(x·si) (x·g) = Li (X-!i) (X·Si)(g). Hence our derivation sends RK into RK if and only if it coincides with the derivation of RK effected by X· CLdi ® Si), for every x f K. By the remark that just precedes the statement of our corollary, this is so if and only if Ldi ® Si e (R ® (@jSl'»K, Q.E.D. 3. The complex of differential forms
Let P be a commutative unitary F-algebra, where F is a field. Let Tp denote the P-module of all F-derivations of P. For every positive integer q, let Aq(Tp) denote the P-module of all (P, q)-multilinear alternating maps from Tp to P, and put AO(Tp) = P. The elements of Aq(Tp) are called the homogeneous differential forms of degree q on P. The weak direct sum of the A q(T p) will' be denoted A (T p) . With every T e T P, we associate a P -linear homogeneous endomorphism of degree -1 of A ( T p), called the contraction with respect to T and denoted CT. This is defined as follows: CT = on P; for a e A q(Tp) and q > 0, cT(a) is the element of Aq-1(Tp) given by cT(a)(T1, ... ,Tq-1) = aCT, T1, . . • ,Tq-1).
°
The natural action of Tp on P is extended to an action of Tp on A(Tp) by homogeneous F-linear endomorphisms tT of degree 0, where tT(a)(T1' ... ,Tq) = T(a(T1' ... ,Tq»
One verifies directly that c~
=
+ Ll=la(T1,
0, [tT' c,,]
=
295
... ,[Ti, T], ... , Tq).
CrT,,,] , and [tT' t,,]
=
t[T,U] '
Next
DIFFERENTIAL FORMS AND LIE ALGEBRA COHOMOLOGY
269
one verifies inductively on the degree that there is one and only one homogeneous F-linear endomorphism a of degree 1 on A (Tp) satisfying 2 aCT + CTa = tT , for all T E T p . Then a commutes with each tT and a = o. One has the familiar explicit formula for a: (aa)(To, ... , Tq)
The complex (A(Tp), a) is called the complex of the differential forms on P. We shall examine this complex in the cases P = Rand P = RK. The case P = R was investigated in [5], and we shall obtain results on RK from a certain canonical map of the complex for RK into the complex for R. We identify TR with R ® ®, and we define an F-linear projection f3: TR ~ ® by f3('Ldi ® ti) = 'Ldi(l)ti. For every t E ®, we define the derivation t* of R by t*(f) (x) = t(x·f)(l), so that t* commutes with the left translations by the elements of G and t*(f) (1) = t(f) (1), for every fER. Clearly, t*(RK) c R K, and we make the convention that where t* occurs as an argument with a differential form on RK it should be replaced with its restriction to RK. The adjoint action of G on ® is extended to the action T ~ X· T of G on T R , where (x· T) (f) = X· T(X-I.f), for every fER. Then we have f3(x· (fT)) = f(X)f3(X·T). Now let a eAq(TRK). Then we define a function pea) on q-tuples of elements of TR and with values in R by setting p(a)(TI, ... ,Tq)(X) = a({3(x·TI)*, ... ,(3(X·T q )*)(X).
Our last remark shows that p( a) is R-multilinear, and hence one sees immediately that pea) EAq(T R), so that we have defined a homogeneous RK-linear map p of degree 0 of A(TRK) into A(TR). We claim that p commutes with a. Since the derivations t*, with t ranging over ®, span TR over R, it suffices to show that (apa)(tci, ... ,t:) = (paa)(tci, ... , t:),
for all ti E ®. Let t E ®. Clearly, X· t* = t*, for every x E G. Furthermore, we have (3U*) = t. In order to see this, write t* = 'Ldi ti, withfi E Rand tiE ®. Let fER and x e G. Then we have t(f)(x) = t(f·x)(l) = t*(f·x)(l) = 'Ldi(l)ti(f·x)(l) = (3(t*)(f-x)(l)
= (3(t*)(f)(x), which proves our assertion. Hence we have (pa) (d , ... , t:) = a(t:', ... , t:), and the similar equality for aa in the place of a. Now the equality to be proved follows at once from the explicit formula for a, noting that [t; , t:] = [t., tr]*. The group G operates on T R and on T RK on the right, as follows: (T·X) (f) = T(f-X- I) ·x.
296
270
G. HOCHSCHILD AND B. KOSTANT
We make Aq(TR) and Aq(TRK) into G-modules, setting (x·O') (Tl , ... , Tq) =
0'(
Tl·X, ... , Tq·X) ·x-1.
We shall show that p is a G-module homomorphism.
We have
(x·p(O'»(Tl, ... , Tq)(Y) = p(O') (Tl·X, ... , Tq·X)(X-1y) = O'((3(X- 1Y·Tl·X)*, ... , (3(X- 1Y·Tq·X)*)(x-1y).
Now one verifies directly that (3(x- 1. T·X) = x- 1.(3( T) and, for S E ®, (x-1·s)* = s*·x. Hence the last expression above is equal to O'((3(Y·Tl)*·X, ... , (3(Y·Tq)*·X)(x-1y) = p(X·O')(Tl, ... , Tq)(Y), Q.E.D. It is seen immediately from the explicit formula that the coboundary operator 0 is also a G-module homomorphism. It is clear that the G-module structure just defined on A ( T R) and A ( T RK) is that of a rational G-module, in the sense of [5]. Hence it induces the structure of a ®-module on our complexes, and it is easily verified that this induced ®-module structure coincides with the ®-module structure given by r--~ tr*. Since
-
cr- 0
+ OCr-
= tr* ,
it follows that the corresponding ®-module structure on the cohomology groups of our complexes is trivial. Hence, if the base field F is of characteristic 0 and G is irreducible, the induced action of G on the cohomology groups of the complexes A (T R) and A (TRK) is trivial. By changing sides in the above definition of the G-module structure on A (T R), we obtain a second G-module structure; we indicate the operations of this structure by 0' ~ x( 0'), where, for 0' E A q( T R), x( 0') is defined by X(O')(Tl, ... , Tq) = x·O'(X-1·Tl, ... , X-1·T q).
One sees immediately that this is the structure of a rational G-module, and that 0 is a G-module endomorphism also for this new G-module structure. Let AK(TR) denote the subset of all 0' EA(TR) such that x(O') = 0', for every x E K, and cr(O') = 0, for every S E~. The ®-module structure of A(TR ) that is induced by our new G-module structure is given by .I ~ tr, for every S E ®. Hence, for each .I E~, tr annihilates A K ( T R ), and the formula OCr + Cr 0 = tr shows that A K( T R) is a subcomplex of A (TR). Evidently, it is also an RK-submodule of A (TR). Finally, one checks easily that A K( T R) is also a G-submodule of A (T R), for the G-module structure 0' ~ X· 0' we defined originally. We shall show that p(A(TRK» c AK(TR). It follows at once from the definitions that x(p(O'» = p(O'), for every O'EA(TRK) and every XEK. There remains to show that Cr 0 p = 0, for all S E~. Let 0' E A q( TRK). Then we have (cr PO')( T2 , ... , Tq)(X) = (PO')(s, T2, ... , Tq)(X)
= O'((x·n*, (3(X·T2)*' ... , (3(X·T q )*)(X).
297
DIFFERENTIAL FORMS AND LIE ALGEBRA COHOMOLOGY
271
Now the map T --7 aCT, (:J(X'T2)*, ... ,(:J(X'Tq)*) is an element of AI(TRK). Hence we see that it suffices to prove that, for every a ~ Al ( T RK) and every x ~ G, we have a( (x· r) *)(x) = O. But a( (x· r) *)(x)
=
a(r* ·x-l ) (x)
=
l
(x- ·a)(r*)(l).
Hence it suffices to show that aU*)(l) = 0, for every a~AI(TRK). Let S denote the field of quotients of R K , and choose a maximal set (rl, ... , rn) of elements of ® such that the derivations of S effected by ri , ... , are linearly independent over S. We have F eSc Q, and Q is finitely generated (as a field) over F. Hence S is finitely generated over F. Hence there are elements fl , ... , fn in RK such that the determinant formed from the fi'(fJ is different from O. Let D be this determinant. If we replace the ri by X-l·ti, with x ~ G, and the fj by frx, D is changed to D·x. Hence we may choose the ri and the fj so that D(1) ,.e O. Now if we consider the system of linear equations
r:
Lf~l Sj f:(fj)
=
a(7),
we find that there are elements gj ~ RK such that Da(7)
=
for each i.
Lf=lgjr7(fj),
On the other hand, there is a nonzero element g in RK such that gt* coincides on RK with an RK-linear combination of the t7. Hence Da(t*)
= Lf~l gj t*(fj) ,
whence D(l)aU*) (1)
so that a(t*) (1) = 0, Q.E.D. PROPOSITION 3.1. Let G be an irreducible algebraic linear. group over the field F of characteristic 0, and let K be a fully reducible algebraic subgroup of G. Then the map p is an isomorphism of A(TRK) onto AK(TR)'
Proof. Regard TR as a K-module via the extended adjoint representation T --7 X· T. Let I be the inverse image of TRK for the restriction map of TR into the R-module of the F-derivations of RK into R. Clearly, I is a K-submodule of T R , and (TR)K C I. Now T R , and hence I, are rational K-modules.
Since K is fully reducible, it follows that I is semisimple as a K-module. With the trivial K-module structure on TRK, the restriction map 1--7 TRK is evidently a K-module homomorphism. By Proposition 2.2, it is an epimorphism. Hence we conclude that the restriction map is an epimorphism of (TR)K onto TRK. Now (TR)K consists precisely of the RK-linear combinations of the derivations t*, where t ~ ®. If a ~ A q( TRK) and tl , ... , tq are elements of ®, we have (pa)(ti , ... ,t:)
=
a(ri , ... ,
Hence it is clear that p is a monomorphism.
298
r:).
272
G. HOCHSCHILD AND B. KOSTANT
Now let us recall from the proof of Lemma 2.2 that the kernel of the restriction map of TR into the R-module of the F-derivations of RK into R is precisely R ® st. It follows that, if 'Y E Ak( T R) and Tl , ... , Tq are elements of T R , then 'Y( Tl , ... , Tq) depends only on the restrictions of the T/s to K RK. Moreover,sincex('Y) ='Y,foreveryxEK,wehave'Y(Tl, ... ,Tq) ER , whenever the T/S belong to (TR)K. Hence the restriction of'Y to q-tuples of elements of (TR)K induces an element a EAq(TRK). We have (pa)(ti , ... ,t:)
=
'Y(ti , ... , t:),
whenever the ti belong to @. It follows that pea) = 'Y, and we have shown that p is an epimorphism. This completes the proof of Proposition 3.1. Let V be a rational G-module, and consider the rational G-module complex A(TR) ® V. If 'Y is an element of Aq(TR) and v is an element of V, then q 'Y ® v defines an element of the space C (@, V) of the alternating q-cochains for @ in V by ('Y ® V)(tl,··· ,tq) = 'Y(tl, ... ,tq)(1)v. Thus we have a map if; of A ( T R) ® V into C (@, V). It has been shown in [5, Section 5] (and is actually easy to verify directly) that the restriction of if; to (A (T R) ® V) G is an isomorphism of the complex (A ( T R) ® V) G onto the complex C(@, V). Now we consider the K-module structure of (A(TR) ® V) G obtained by using the operations a ~ x(a) on A(TR) and the trivial action on V. On Cq (@, V), we introduce the usual K -module structure given by (X·",)(tl, ... ,tq) = x·",(X-1·tl, ... ,x-1·tq).
We claim that, for these K-module structures, the restriction of if; to (A (TR ) ® V) G is a K-module isomorphism. Let x E K, and let 'Y and v be as above. Then we have if;(x('Y ® v)) (tl, ... ,tq) = x('Y) (tl , ... , tq)(l)v
Hence, if a E (A q(TR ) ® V)
G,
=
'Y(X-1·tl, ... ,x-1·tq)(x)v
=
(x-1.'Y) (X-1·tl , ... ,x-1·tq)(1)v.
this gives
if;(x(a))(tl, ... ,tq)
=
1 x·(a(x-1·tl, ... ,x- ·tq)(1)v)
= (x·if;(a)) (tl , ... , tq). Thus our isomorphism of complexes if;:(A(TR) ® V)G ~ C(@, V) is also a K-module isomorphism. Hence we see immediately that if; maps (AK(T R) ® V) G isomorphically onto C(@/st, V)K. Under the conditions of Proposition 3.1, we compose if; with the isomorphism of (A(TRK) ® V) G onto (AK( T R) ® V) G that is induced by the isomorphism p of Proposition
299
DIFFERENTIAL FORMS AND LIE ALGEBRA COHOMOLOGY
273
3.1 to obtain an isomorphism of the complex (A(TRK) ® V)G onto the complex C(@/sr, V)K. Let Kl be the irreducible component of the identity in K. Then C(@/sr, V)K is the K/Kr-fixed part of C(@/sr, V)Kt, which is the sr-annihilated part of C(@/sr, V). The cohomology space of the complex C(@/sr, V)Kl is the relative Lie algebra cohomology space H(@, sr, V) for (@, sr) in V. As a module for the finite group K/Kl , C(@/sr, V)Kl is semisimple. Hence the K/Kr-fixed part of H(@, sr, V) is the cohomology space of the complex C(@/sr, V)K. Hence we have the following result. THEoREM 3.1. Let G be an irreducible algebraic linear group over the field F of characteristic 0, and let K be a fully reducible algebraic subgroup of G. Then, for any rational G-module V, the cohomology space of the complex (A ( T RK) ® V) G is isomorphic, via the maps p and if; defined above, with the K-fixed part H(@, sr, V) K of the relative Lie algebra cohomology space for (@, sr) in V.
In the special case where G is fully reducible, Theorem 3.1 gives a determination of the cohomology space of the complex A (T RK) . In that case, the rational G-module A (T RK) is semisimple. Since the action of G on the cohomology space of the complex A (TRK) is trivial, this implies that the injection A (T RK) G ---> A (T RK) induces an isomorphism of the cohomology spaces. Hence, if we take for V the trivial G-module F, Theorem 3.1 gives the following result. THEOREM 3.2. Let G be an irreducible algebraic linear group over the field F of characteristic O. Suppose that G is fully reducible, and let K be a fully reducible algebraic subgroup of G. Then the cohomology space of the complex A(TRK) is isomorphic with H(@, sr, F)K.
In the general case of Theorem 3.1, there is a spectral sequence linking the tensor product of the cohomology space of the differential forms on RK and the space of the rational cohomology for G in V to the relative Lie algebra cohomology space for (@, sr) in V. In order to derive this result, we need more information on the G-module A(TRK). Under the present assumptions, A(TRK) may be identified with AK(TR), and we shall show that A K( T R) is a direct G-module summand of A (T R) . Since K is fully reducible, we have a direct K-module decomposition @ = sr + 113. For each q > 0, the R-module A q(T R ) is the direct sum of two R-submodules A lr (T R) and A; (T R), consisting of the elements annihilated by the Cr with S- E sr, or of the elements annihilated by the Cr with S- E 113, respectively. Since sr and 113 are K-submodulesof @, it is clear that these two R-submodules are stable under the operations a ---> x(a), with x E K. With reference to this K-action, the K-fixed part of A q( T R ) is therefore the direct sum of Ak( T R ) and the K-fixed part of A;( T R ), and it is clear that
300
274
G. HOCHSCHILD AND B. KOSTANT
this is a direct G-module decomposition for the G-module structure a --> X·a. Hence AK(TR) is a direct G-module summand of the K-fixed part (for the action a --> x( a)) of A (T R)' Since K is fully reducible, the rational K-module A(TR ) (for the action a --> x(a)) is semisimple. Hence the K-fixed part of A(TR) is a direct G-module summand of A(TR)' Thus AK(TR) is indeed a direct G-module summand of A(TR)' It follows at once from [5, Proposition 2.2] and from the form in which A (T R ) is exhibited in [5] that, for every rational G-module V, the G-module A (T R ) ® V is rationally injective, in the sense of [5]. In virtue of what we have just proved, this implies that, under the assumptions of Theorem 3.1, the G-module A (TRK) ® V is rationally injective. q q Put U = Aq(TRK) ® V, U = Lq U . Then U is a rational G-module, and we consider the complex C(G, U) = Lp CP(G, U) of the nonhomogeneous rational representative co chains for Gin U, in the sense of [5, Section 2]. Let oG denote the coboundary operator of this complex, and let Ou denote the coboundary operator of the complex U. Then C(G, U) has the structure of a double complex of rational G-modules, with total co boundary operator 0 = oG + (-l)P ou , on CP(G, U). Since U is rationally injective as a G-module, the rational cohomology groups Hn(G, U) for G in U are (0), for all n > O. A standard argument [3, Proposition 4], shows that the injection U G --> CO(G, U) c C(G, U) induces an isomorphism of Hn(U G, ou) onto Hn(C(G, U), 0), for all n ~ 0, q where the grading on C(G, U) is given by C(G, U) n = Lp+q~n CP(G, U ). We introduce a decreasing filtration (Li) on this double complex, where Li = Lp~i CP ( G, U). The spectral sequence derived from this filtration is the spectral sequence of Cartan-Leray, with the ordinary group cohomology replaced by the rational cohomology of G. The arguments and results of [3, Chapter I, Section 5] apply without change to the present situation, giving G G the following result. Let H ( U ) p denote the subspace of H (U ) whose image in H(C(G, U)) is the image of H(Lp). Then the limit E", of the spectral sequence is given by E!,q = Hp+q(UG)p/HP+Q(UG)P+1; the term E2 of the spectral sequence is given by Ef,q
=
HP(G, Hq(U)).
Now we have Hq(U) = Hq(A(TRK)) ® V, and we have seen earlier that the action of G on Hq(A(TRK)) is trivial. Hence we obtain Ef,q
=
HP(G, V) ® Hq(A(TRK)).
On the other hand, by Theorem 3.1, H( U G ) = H(®, ~, V)K. Thus we have the following result, which is the analogue of a result for Lie groups due to van Est [7, Theorem 3].
301
275
DIFFERENTIAL FORMS AND LIE ALGEBRA COHOMOLOGY
THEOREM 3.3. Let G be an irreducible algebraic linear group over a field F of characteristic 0, and let K be a fully reducible algebraic subgroup of G. Let V be a rational G-module. Then there is a spectral sequence with
E2 = H(G, V) ® H(A(TRK)) and with E", the graded space obtained from a filtration of H(®,
sr,
V)K.
4. Relative lie algebra cohomology While the connection expressed in Theorem 3.3 is rather vague, we have very precise connections of this type in two special cases. One of these is the case where G is fully reducible, which case is covered by Theorem 3.2. The other is the case where K is a maximal fully reducible subgroup of G. In this case, the spectral sequence of Theorem 3.3 collapses, in the sense that E2 coincides with E"" H(A(TRK)) = H\A(TRK)) = F, and the result becomes a superficially different form of the isomorphism of [5, Theorem 5.2]. In order to express the relative Lie algebra cohomology in terms of the cohomology of differential forms on an algebra of representative functions and the rational cohomology of G, we make a reduction to the above special cases on the level of the relative Lie algebra cohomology. The reduction is based on an imbedding of the given fully reducible subgroup of G in a maximal fully reducible subgroup of G, which enables us to use an easy generalization of the tensor product decomposition of the Lie algebra cohomology given by [4, Theorem 13]. THEOREM 4.1. Let ® be a finite-dimensional Lie algebra over the field F of characteristic O. Let sr be a reductive sub algebra of ®, and let ~ be an arbitrary subalgebra of sr. Suppose that the restriction map H(®, ~, F) ~ H(sr, ~, F) is an epimorphism. Let V be a finite-dimensional ®-module that is semisimple as a sr-module. Then, for each n ~ 0, H n (®, ~, V) is isomorphic with
Lp+Q=n HP(sr, ~, F) ® H (®, q
sr, V);
an isomorphism of the second space onto the first is obtained in the natural way from any homogeneous linear monomorphism H(sr, ~, F) ~ H(®, ~, F) inverse to the restriction epimorphism, the canonical map H(®, sr, V) ~ H(®, ~, V), and the cup product H(®, ~, F) ® H(®, ~, V) ~ H(®, ~, V).
The proof is almost identical with the proof of [4, Theorem 12]. quires the following generalization of [4, Theorem 10].
It re-
LEMMA 4.1. Let sr be a finite-dimensional reductive Lie algebra over the field F of characteristic 0, and let M be a finite-dimensional semisimple sr-module such that MIi1 = (0). Then, if ~ is any subalgebra of sr, Hn(sr, ~,M) = (0), for all n ~ o.
Proof. We may evidently suppose that M is simple and that n > O. Let ~ be the center of sr. Then sr is the direct sum of [sr, sr] and ~, and
302
276
G. HOCHSCHILD AND B. KOSTANT
[Sf, Sf] is semisimple. For every l' ~ (i£, 1" M is a Sf-submodule of M, and thus is either (0) or M. Suppose first that I'·M = M. Let j be a relative ncocycle for (Sf, ~) in M. Then coy(f) is a relative (n - l)-cochain for (Sf, ~) in M, and o(coy(f)) = I'I But l' acts as a @-module automorphism Poy on M,andl"j = poyoj. Hencewehavej = p-:;1 ° o(coy(f)) = o(p-:;locoyCf)),and p-:;1 ° coyU) is evidently a relative (n - l)-cochain for (Sf, ~) in M. Hence we may now suppose that (i£·M = (0). Let 'r denote the annihilator of M in sr. Then 'r contains (i£, whence Sf is a direct sum'r + ®, where ® is a semisimple ideal of Sf and the representation of ® on M is faithful. Let if; be the Casimir operator of this representation. If t ~ Pr is the representation of Sf on .M, we have if; = Li Pri ° PTi , where ti and Ti are elements of Sand Li ([a, til ® Ti + ti ® [a, Ti]) = 0, for all a ~ Sf. Now put g = Pr i ° CTi (f) . Then one shows by a familiar computation (see [2, p. 118]) that o(g) = if; oj, so that j = y;-loo(g) = o(y;-l og).
L
Hence there remains only to see that if;-1 ° g is a relative cochain for (Sf, ~) in M. Clearly, cr(if;-1 ° g) = 0, for every t ~~. Also o(cr(if;-log)) = t·(y;-l og ) - cr(o(y;- l og)) = t·(if;-l og ) - cr(f)· Taking t ~~, we see from this that t· (if;-1 ° g) = O. Thus if;-1 ° g is indeed a relative cochain, and the proof of Lemma 4.1 is complete. In order to prove Theorem 4.1, one can now proceed in exactly the same way as in [4], replacing the ordinary cochain complex C(@, V) by the relative complex C(@, ~, V) = C(@j~, V)~. One considers the filtration of this complex that is obtained by intersecting the filtration groups used in [4] with C(@, ~, V). For the corresponding spectral sequence, one shows first that Ef'q = Hq(Sf, ~, CP(@jSf, V)). Here, the only deviation from the proof in [4] is that the projection l' ~ l' * of @ onto sr used in the proof of [4, Theorem 1] must now be chosen so that it is a Sf-module projection, which is possible, because Sf is reductive in @. Next one proceeds to show, as in the proof of [4, Theorem 11 and Corollary], that Ef,q = Hq(Sf,~, F) ® H P(@, Sf, V), replacing the use of [4, Theorem 10] with an appeal to Lemma 4.1 above. The rest of the proof of Theorem 4.1 is exactly as the proof of [4, Theorem 12], where one now ignores the mutiplicative feature of the isomorphism to be established. Now suppose that there is an ideal 9[ in @ such that @ is the semidirect sum Sf + 9[, where Sf is a reductive sub algebra of @. Let ~ be any subalgebra of Sf. Then it is easy to see that the restriction map H(@,
~,
F)
~
H(Sf,
~,
F)
is an epimorphism. Furthermore, it is seen exactly as in [4, p. 603] that H(@, Sf, V) is naturally isomorphic with H(9[, V/'\ for every @-module V
303
DIFFERENTIAL FORMS AND LIE ALGEBRA COHOMOLOGY
277
that is semisimple as a Sf-module. Hence Theorem 4.1 shows that H( @,~, V) is isomorphic with H(Sf, ~,F) ® H(m, V)@. Now let G be an irreducible algebraic linear group over the field F of characteristic o. Let V be a finite-dimensional rational G-module, and let L be a fully reducible algebraic subgroup of G. Let K be a maximal fully reducible subgroup of G that contains L. Then K is irreducible as an algebraic linear group, and G is the semidirect product K· N, where N is the maximum unipotent normal subgroup of G. Now if @, Sf, ~, mare the Lie algebras of G, K, L, N, respectively, all of our above assumptions hold. The algebra of the rational representative functions on K may be identified with RN. Hence Theorem 3.2 shows that H(Sf, ~, F) L is isomorphic with H(A(TRLoN». On the other hand, by [5, Theorem 5.2], the cohomology space H(m, V)@ is isomorphic with the space H(G, V) of the rational cohomology for G in V. Passing to the L-fixed parts in the above isomorphism result for the Lie algebra cohomology, we obtain the result that H(@,~, V) Lis isomorphic with H (G, V) ® H (A (T R L oN) ) . Thus the change from the L algebra RL to the subalgebra R .N closes the gap left by Theorem 3.3, and we have the following result. THEOREM 4.2. Let G be an irreducible algebraic linear group over a field F of characteristic 0, and let K be a fully reducible algebraic subgroup of G. Let V be a finite-dimensional rational G-module, and let N be the maximum unipotent normal subgroup of G. Then H(@, Sf, V)K is isomorphic with R(G, V) ® H(A(TRKoN».
5. The universal differential forms Let P be a commutative unitary F-algebra, where F is a field. Besides the complex A ( T p) which we defined at the beginning of Section 3, one can define a formally similar complex without referring to the derivations of P. This construction, which is well known, is as follows. Regard P ® P as a P-module such that a(b ® c) = (ab) ® c. Let J be the submodule that is generated over P by the elements of the form 1 ® (ab) - a ® b - b ® a. Let D~ denote the factor module (P ® P) / J. The elements of D~ are called the universal differential forms of degree 1 on P. It is immediately verified that the P-module Tp of all F-derivations of P is naturally isomorphic with Homp(D~ , P); the isomorphism is induced by attaching to each T E Tp the map P ® P ~ P that sends a ® b onto aT(b). We define an F-derivation d:P ~ D~ by d(a) = 1 ® a + J. Then d(P) evidently generates D~ over P. Put D~ = P, and let Dp = D~ denote the exterior P-algebra built over D~. It is easily verified that the map d:D~ ~ D~ has a unique extension, still denoted d, to a homogeneous F-linear antiderivation of degree 1 and square 0 on D p • The complex (D p , d) is called the complex of the universal differential forms on P. The dual of the natural P-module isomorphism Tp ~ Homp(D~, P),
Lq
304
278
G. HOCHSCHILD AND B. KOSTANT
preceded by the canonical P-module homomorphism of D~ into its bidual, gives a natural P-module homomorphism
Proof. Let G be given as an algebraic group of linear automorphisms of determinant 1 of a finite-dimensional F-space V. Let P be the algebra of all polynomial functions on the space of all linear endomorphisms of V, and let p denote the restriction epimorphism P ~ R. Let I be the kernel of p. We regard P as a K-module, with K operating by left translation. Then, since K is fully reducible, Pis semisimple as a K-module. By the fundamenK tal theorem of invariants, p is therefore finitely generated as an F-algebra. Since P is semisimple as a K-module, p induces an epimorphism of p K onto R K, whence RK is finitely generated. LEMMA. 5.2. In addition to the assumptions of Lemma 5.1, assume that F is algebraically closed. Then every homomorphism RK ~ F leaving the constants fixed is of the form f ~ f(x), with some x E G.
Proof. Let
+
+
+
+
THEOREM 5.1. Let G be an irreducible algebraic linear group over an algebraically closed field of characteristic o. Let K be a fully reducible algebraic
305
279
DIFFERENTIAL FORMS AND LIE ALGEBRA COHOMOLOGY
subgroup of G. Then the set G/ K of the cosets xK has the structure of an affine algebraic variety, with RK as the algebra of the polynomial functions. LEMMA 5.3. Let G be an irreducible algebraic linear group over the field F of characteristic 0, and let K be a fully reducible algebraic subgroup of G. Then the canonical map rp:D~K ~ AI(TRK) is an isomorphism.
Proof. By extending the base field F to its algebraic closure and using the remarks made at the end of our proof of Proposition 2.2, we see that no generality is lost in assuming F to be algebraically closed, which we shall now do. Let a € D~K, and suppose that rp( a) = O. We may write a
=
IJ=I ai d(b i),
where ai and bi are elements of RK, and the bi are different from O. We can choose elements fI , ... , fn in RK so that they are algebraically independent over F and RK is algebraic over F[fI , ... , fn]. Then there are F-derivations CTI , ••• , CT n of the field of quotients QK of RK such that CT i(jj) = Oij. Since RK is finitely generated over F, we can find a nonzero element f € RK such that each fCTi sends RK into itself. Thus we obtain elements T I , ' " , Tn of TRK such that the determinant formed with the elements Ti(fj) isdifferent from O. Now let Pi be a nonzero polynomial with coefficients in F[fI , ... , fn] and of minimal degree such that pi(b i ) = O. Let Pij denote the polynomial obtained from Pi by differentiating its coefficients with respect to fj , and let P: denote the ordinary derivative of the polynomial Pi. Then we have p:(b i ) d(b i )
Put P
=
+ 2::f=I Pij(b i) d(fj)
p~(bI) .,. p~(bq).
= 0, and p:(bi) ~ O.
Then p ~ 0, and we have
Since rp( a) = 0, we have
o=
paC Tk) = 2::~=I gi Tk(fi),
for each k.
Since the determinant 'formed from the coefficients of the gi here is not 0, this gives gi = 0, for each i, whence pa = O. Make RK ® RK into a G-module such that X· (f ® g) = (!-x-I ) ® (g,x- I). Then the kernel of the canonical epimorphism of RK ® RK onto D~K is evidently G-stable, so that we get an induced G-module structure on D~K. Clearly, the F-space spanned by the transforms of any element of D~K is finite-dimensional. Applying the above argument to a finite number of transforms of a, we see therefore that there is a nonzero element f in RK such that f(x'a) = 0, for every x € G. Hence (!-x)a = 0, for every x € G. Now let I be the ideal of all g € RK such that ga = O. Our last result shows that I has no zero on G. By Theorem 5.1, this means that I = RK, whence a = O. Thus we have shown that rp is a monomorphism.
306
280
G. HOCHSCHILD AND B. KOSTANT
Finally, we observe that the argument that just precedes the statement of Proposition 3.1 shows that, if 'Y is any given element of A I( TRK), the ideal of all g f RK such that g'Y f cp(D~K) has no zero on G, whence we conclude that cp is an epimorphism. This completes the proof of Lemma 5.3. We have a direct K-module decomposition @ = Sf' + 'l3. Hence, as an RK-module, (R ® @)K is the direct sum of (R ® Sf')K and (R ® 'l3)K. Now TRK is isomorphic, as an RK-module, with (R ® @/Sf')K, and hence with (R ® 'l3)K. Thus we see that TRK is isomorphic with a direct RK-module summand of (R ® @)K. On the other hand, (R ® @)K may be identified with RK ® @*. Hence TRK is a direct RK-module summand of a finitely generated free RK-module. In other words, TRK is a finitely generated projective RK-module. If M is any RK-module, let E(M) denote the exterior RK-algebra built K over M. Evidently, A(TRK) may be identified with HomRK(E(TRK) , R ). On the other hand, if M is a free RK-module with a finite RK-basis, the canonical map E(HomRK(M, RK)) ~ HomRK(E(M) , RK) is an isomorphism. Since TRK is a direct RK-module summand of such an M, it follows that the canonical map E(A I( TRK)) ~ HomRK(E( TRK), RK) is also an isomorphism. Thus, using the above identification, the canonical map of E(A \ TRK)) into A(TRK) is an isomorphism. Taking account of Lemma 5.3, we have the following result. THEOREM 5.2. Let G be an irreducible algebraic linear group over a field of characteristic 0, and let K be a fully reducible algebraic subgroup of G. Then the canonical map cp of the complex DRK of the universal differential forms on RK into the complex A(TRK) of the differential forms based on the F-derivations of RK is an ismorphism. Moreover, these complexes are finitely generated projective RK-modules, and the homogeneous components of degree larger than the dimension of @/Sf' are (0).
Only the last statement of this theorem still requires verification. Let q be the dimension of @/Sf'. Then we can find elements fr , ... ,fq in RK such that RK is algebraic over F[fI , ... ,fq]. Let gi , ... , gq+1 be arbitrary elements of RK. By an argument we used in proving Lemma 5.3, we see that there is a nonzero element Pi in RK such that Pi d(gi) can be written in the form ~]=I aj d(fj) , with aj f RK. Hence we find that PI ... Pq+1 d(gI) ... d(gq+l) = O.
Since D~""il is a projective RK-module, this implies that d(gl) ... d(gq+I) = 0, Q.E.D. REFERENCES
1. 2.
Theorie des groupes de Lie, II, Paris, Hermann, 1951. S. ElLENBERG, Cohomology theory of Lie groups and Lie algebras, Trans. Amer. Math. Soc., vol. 63 (1948), pp. 85-124.
C. CHEVALLEY,
C. CHEVALLEY AND
307
DIFFERENTIAL FORMS AND LIE ALGEBRA COHOMOLOGY
281
3. G. HOCHSCHILD AND J-P. SERRE, Cohomology of group extensions, Trans. Amer. Math. Soc., vol. 74 (1953), pp. 110-134. 4. - - - , Cohomology of Lie algebras, Ann. of Math. (2), vol. 57 (1953), pp. 591-603. 5. G. HOCHSCHILD, Cohomology of algebraic linear groups, Illinois J. Math., vol. 5 (1961), pp. 492-519. 6. E. KUNZ, Die Primidealteiler der Differenten in allgemeinen Ringen, J. Reine Angew. Math., vol. 204 (1960), pp. 165-182. 7. W. T. VAN EST, Une application d'une methode de Cartan-Leray, Indag. Math., vol. 17 (1955), pp. 542-544. UNIVERSITY OF CALIFORNIA BERKELEY, CALIFORNIA
308
Reprinted from the BULLETIN OF THE AMERICAN MATHEMATICAL SOCIETY
luly, 1963, Vol. 69. No.4 Pp.518-526
LIE GROUP REPRESENTATIONS ON POLYNOMIAL RINGS! BY BERTRAM KOSTANT2 Communicated by Raoul Bott, February 1, 1963
O. Introduction. 1. Let G be a group of linear transformations on a finite dimensional real or complex vector space X. Assume X is completely reducible as a G-module. Let S be the ring of all complexvalued polynomials on X, regarded as a G-module in the obvious way, and let Jt:;;;.S be the subring of all G-invariant polynomials on X. Now let J+ be the set of allJEJ having zero constant term and let HCS be any graded subspace such that S=J+S+H is a G-module direct sum. It is then easy to see that
(0.1.1)
S = JH.
(Under mild assumptions H may be taken to be the set of all Gharmonic polynomials on X. That is, the set of all JES such that af= 0 for every homogeneous differential operator a with constant coefficients, of positive degree, that commutes with G.) One of our main concerns here is the structure of S as a G-module. Regard S as a J-module with respect to multiplication. Matters would be considerably simplified if S were free as a J-module. One shows easily that Sis J-free if and only if S = J 0 H. This, however, is not always the case. For example S is not J-free if G is the two element group {I, - I} and dim X ~ 2. On the other hand one has EXAMPLE 1. It is due to Chevalley (see [2]) that if G is a finite group generated by reflections then indeed S=J0H. Furthermore the action of G on H is equivalent to the regular representation of G. EXAMPLE 2. S is J-f,ree in case G is the full rotation group (with respect to some Euclidean metric on X. For convenience assume in this example that dim X ~ 3). Note that the decomposition of a polynomial according to the relation S= J0H is just the so-called "separation of variables" theorem for polynomials. This is so because J is the ring of radial polynomials and H is the space of all harmonic polynomials (in the usual sense). Now, for any xEX, let O",r;;;.X denote the G-orbit of x and let S(O,) be the ring of all functions on 0", defined by restricting S to 0",. Since J reduces to constants on any orbit it follows that (0.1.1) inI This research was supported by National Science Foundation grant NSF· G19992. 2 The author is an Alfred P. Sloan fellow.
518
B. Kostant, Collected Papers, DOI 10.1007/b94535_16 , © Bertram Kostant 2009
309
LIE GROUP REPRESENTATIONS ON POLYNOMIAL RINGS
519
duces a G-module epimorphism H~S(O",).
(0.1.2)
Since our major concern is the case where X is a reductive Lie algebra and G is the adjoint group and since the methods used there belong to algebraic geometry we will assume now that X is complex and that G is algebraic and reductive. All varieties considered are over C. If Y has an algebraic structure R( Y) will denote the ring of everywhere defined rational functions in Y. Obviously one always has (0.1.3)
S(O.,) C R(O",).
On the other hand if GxCG is the isotropy group defined by xEX then one has a G-module isomorphism (0.1.4) The significance of (0.1.4) is that one knows the G-module structure of R(G/Gx) completely by a very simple algebraic Frobenius reciprocity theorem (even though Gx may not be reductive). In fact if VA is any irreducible G-module with respect to the representation pI. and VA is the dual module then one has (0.1.5)
multo of
pA
in R(G/c"')
vr
= dim
vt
where is the space of vectors in VA fixed under G"'. Now in Examples 1 and 2 (assume complexified) the following three optimum situations occur: (a) Sis J-free so that S=J®H, (b) the map H~S(O",) is an isomorphism for certain xEX and for those x, (c) R(O",) = S(O",). But one observes that if in any general case (b) and (c) hold then, clearly, upon combining (0.1.4) and (0.1.5) one gets the G-module structure of H. If one gets in addition the "graded" G-module structure of H and knows the structure of J then one gets the full graded G-module structure of S in case (a) also holds. In Example 2 the conditions (b) and (c) hold for any x~O (even if (x, x) =0). In fact, classically, one has exploited (b) and (c) for (x, x) >0 to solve the Dirichlet problem with the sphere as boundary. That is, if i is any continuous function on the sphere one first expands i as a Fourier development of spherical harmonics in. The sphere is O",r'lRm and the in are in R(O",). The equality R(O",) = S(Os) and the isomorphism H~S(O",) then yields the extension of in uniquely as
310
520
BERTRAM KOSTANT
Uuly
harmonic polynomials hn on X. But this yields the desired extension off· In Example 1 the conditions (b) and (c) are satisfied for any "regular" element xEX. Our first concern in this paper is to give criteria for (a), (b) and (c) to hold in general. Since our interest is in the continuous case we will assume G is connected (and hence a variety). Thus Example 2 rather than Example 1 serves as a model. N ow let P ex be the cone of common zeros defined by the ideal J+S in S. Let X* be the dual space to X and let p*eX* be defined in a similar way with the roles of X and X* interchanged. As a criterion to establish (a) and more we prove PROPOSITION 0.1. Assume (1) that J+S is a prime ideal in Sand (2) there exists an orbit Oeep which is dense in P. Then S=J0H. Furthermore if G is a subgroup of the complex rotation group then H may be taken as the space of all G-harmonic polynomials. ]v[oreover H then coincides with the space spanned by all powers fk where fEP*.
It may be observed that the criterion is satisfied in Example 2. An element xEX is called quasi-regular if peCI(C*·O,,). A criterion to establish (b) is given by PROPOSITION 0.2. Assume conditions (1) and (2) of Proposition 0.1 are satisfied. Then the G-module epimorphism H-tS(O,,) is an isomorphism for any quasi-regular element xEX.
It may be observed that in Example 2 every nonzero xEX is quasi-regular. From known facts in algebraic geometry one has the following criterion to insure (c). PROPOSITION 0.3. Let xEX and assume (1) the closure Cl(O,,) is a normal variety and (2) CI(O,,) -0., has a codimension of at least 2 in CI(O.,). Then R(O,,) =S(O.,).
It may be observed that the conditions of Proposition 0.3 are satisfied for every xEX in Example 2. Now assume that X = g is a complex reductive Lie algebra and G is the adjoint group. Here the structure of J is given by a theorem of Chevalley. This asserts that J is a polynomial ring in l (the rank of g) homogeneous generators U;, i = 1, 2, ... ,l, with deg ui=mi+1 where the mi are the exponents of g.
311
LIE GROUP REPRESENTATIONS ON POLYNOMIAL RINGS
521
Now one knows that here P is the set of all nilpotent elements of g [13, Theorem 9.1]. But then by [13, Corollary 5.5], P does contain a dense orbit 0., namely, the set of all principal nilpotent elements in g. Thus to apply Propositions 0.1 and 0.2 one must prove that J+ S is a prime ideal. If n = dim g (all dimensions are over C) then one sees easily that n-l is the maximal dimension of any orbit. Let r= {xEgldim 0 .. =n-l}. Any regular element xEg belongs to r. But also eEr for any principal nilpotent element e. These in fact are extreme cases. PROPOSITION 0.4. Let xEg be arbitrary. Write (uniquely) x=y+z where y is semi-simple, z is nilpotent and [y, z] = O. Let gil be the centralizer of y in g so that gY is a reductive Lie algebra and zEglI. Then xEr if and only if z is principal nilpotent in gll. Let xEg. Consider the values (du;)z of the l differential forms du;, i = 1, 2, ... , I, at x. It is known that these covectors are linearly independent whenever x is regular. (One recalls that the product of the positive roots is an I Xl minor of a suitable n Xl matrix determined by the du;.) But to prove the primeness of the ideal J+ S one needs to know that these covectors are linearly independent if x is a principal nilpotent element. This fact is contained in THEOREM 0.1. Let xEg. Then the (du;)z are linearly independent if and only if x Er. Proposition 0.1 may now be applied. THEOREM 0.2. One has S=J@H where H is the space of all Gharmonic polynomials on g. Furthermore H coincides with the space of all polynomials spanned by all powers of "nilpotent" linear functionals. Since Theorem 0.1 shows also that P is a complete intersection the decomposition S=J@H when combined with [15, Proposition 5, §78] gives, in the notation of FAC, all the sheaf cohomology groups Hi(P, e(m» where P is the projective variety defined by P. Added in proof. Another application of the primeness of J+ S in algebraic geometry is THEOREM 0.3 (Added in proof). The intersection multiplicity of P, at the origin, with any Cartan subalgebra is w, where w is the order of the Weyl group. Next, Proposition 0.2 is put into effect for all orbits of maximal dimension by THEOREM 0.4. The set r coincides with the set of all quasi-regular ele-
312
522
BERTRAM KOSTANT
Ouly
ments in g. (Thus Hand 5(0",) are isomorphic as G-modules for any xEr.) As a consequence of Theorems 0.2 and 0.4 one shows that not only is the ideal J+ 5 prime in 5 but J 15 is prime for any prime ideal J l CJ. Furthermore one gets the following characterization of all the invariant prime ideals in 5 which are generated by elements of J. THEOREM 0.5. Let I~S be any G-invariant prime ideal. Let u~g be the affine variety of zeros of I. Then I is of the form 1= J1S for J l a prime ideal in J if and only if u n r is not empty.
Since R(Ox) = S(Ox) in case Ox is closed and since 0" is closed if x is regular one gets the G-module structure of H by applying Theorem 0.3 and (0.1.5) for x regular. Thus if D denotes the set of dominant integral forms corresponding to a Cartan subgroup A, so that D indexes all the irreducible representations of G as highest weights, then one has multo of
(0.1. 6)
VA
in H
=
Ix
where IA = dim V~ is the multiplicity of the zero weight of VA. In order to determine the G-module structure of 5k, the space of homogeneous polynomials on 9 of degree k, one must know more than (0.1.6). In fact using the relation S= J0H what one wants is the multiplicity of vA in Hi = SinH for any A and j. As it turns out, for this, one needs R(O.) = S(O.) where e is a principal nilpotent element. To show the latter using Proposition 0.3 it is enough to show that P is a normal variety and P - O. has a codimension of at least 2 in P. Let l'J r be a set of all orbits of maximal dimension (n -I). The set l'J r may be parameterized by Cl in the following way. Let U:g~CI
be the morphism given by putting u(x) = (Ul(X) , .•. , Ul(X)) for any xEg. Since u reduces to a constant on any orbit it induces a map l1r:
l'Jr ~ CI.
One has THEOREM
0.6.
l1r
is a bijection.
Thus to each ~E Cl there exists a unique orbit, n-l which correspond to ~ under 1/r. Now let ~ECI so that
Om of dimension
pm =U-l(~)
313
for any
LIE GROUP REPRESENTATIONS ON POLYNOMIAL RINGS
523
g = U P(~) £EC'
is a disjoint union. Note that of Cl. One proves THEOREM
0.7. For any
p(~)
~ECl
P(~)
=p and Om = O. if
~
is the origin
one has
=
Cl(O(~»
so that P(~) is a variety of dimension n -1. Moreover P(~) is a complete intersection and Om coincides with the set of simple points on P(~). Finally P(~) is a finite union of orbits so that Cl(Ox) is a finite union of orbits for any xEg. Since P(~) is a complete intersection and since its singular locus is the complement (a finite union of orbits) of in P(~) one would get the normality of P(~) by a theorem of Seidenberg if one knew the dimension of the other orbits in P(~) were at most n -1- 2. Now it is well known that dim Ox is even (and hence dimR 0" is a multiple of 4) for any semi-simple element xEg. Less known is the following proposition observed independently by the author, Borel, and (most simply proved by) Kirillov.
am
PROPOSITION
0.5. The dimension of Ox is even for any xEg.
Combining Theorem 0.6 and Proposition 0.5 one obtains THEOREM 0.8. Let ~E Cl be arbitrary. Then P(~) is a normal variety and the codimension of P(~) - O(~) in P(~) is at least 2.
Applying Proposition 0.3 one then has THEOREM 0.9. Let xEr. Then R(O,,) = S(O,,). (This implies that all R(Ox) for xEr are isomorphic as G-modules; even though they are not in general isomorphic as rings.) Let ~=u(x). Then R(O,,) (=R(GjGx» is an affine algebra (even though 0" is not necessarily an affine variety) and P(~) is the variety of all maximal ideals of R(O,,). Thus the embedding of GjGx in 9 as Ox is special in that any morphism of GjGx (or 0,,) into any affine variety extends uniquely to a morphism of P(~) = CI (Ox) into the variety. (In particular this holds for 0. and CI(O.) =P.) Finally (using (0.1.5) and the equality R(O",) =S(O,,» one has, for any XED
(0.1. 7)
G%
dim V"
=
lx
so that the left side of (0.1. 7) is independent of xEr. Now let L, Xo, e be a principal S-triple (that is, a "canonical" basis
314
524
BERTRAM KOSTANT
Uuly
of a principal three dimensional simple Lie subalgebra). In particular then e is a principal nilpotent element. Used heavily in the theorems above is the result of [13] which asserts that g" is l-dimensional and has a basis Zi, i = I, 2, . , l, such that
(0.1.8) where, we recall, the mi are the exponents of g. But now since g" = g a' (because ge is commutative) and since (0.1. 7) holds for x = e this suggests a generalization of the notion of exponent. Let V be any finite dimensional G-module with respect to a representation 11. If l. is the multiplicity of the zero weight of 11 then by (0.1.7) one has dim va" = Iv. It follows therefore that there exists a unique nondecreasing sequence of non-negative integers mi(lI) , i = I, 2, ... , 1., such that one has II(XO)Zi
= m.(II)Zj
for a basis Zi of va". If 11 is the adjoint representation the m.(II) are the usual exponents. If II=IIA we will write mJ).) for m.(IIA) and note (because the highest weight has multiplicity one) that
for j = Ix where o(}.,) is the sum of the coefficients of X relative to the simple roots and that this highest value occurs with multiplicity one among the generalized exponents mi(X). (This specializes to the familiar relation ml = 0(1/;) when g is simple and I/; is the highest root.) The following theorem now gives the G-module structure of Hi and hence Sk for any j and k. THEOREM 0.10. Let XED be arbitrary and let H(X) be the set of Gharmonic polynomials which transform under G according to 11\ Let (by (0.1.6)) H(X) = L~':.l Hj(X) be a decomposition into irreducible components so that Hj(X) c;;;.Hn; where nj, j = 1,2, ... ,lA' is a nondecreasing sequence of integers. Then nj= mj(X) for all j. In particular then k = o(X) is the highest degree k such that II}. occurs in Hk. ].,foreover it occurs with multiplicity one for this value of k.
Assume for convenience that g is simple ancllet I/; ED be the highest root. Let Xi, i= 1,2, ... , n, be a basis of g. If the ltjEJ are chosen properly one sees that allj/aXi, i = I, 2, ... , n, is a basis of H;(I/;). One notes then that Theorem 0.10 is a generalization of the result in [13] given by (0.1.8). H. S. Coxeter observed and A. J. Coleman proved in [4] that if W is the Weyl group and
315
LIE GROUP REPRESENTATIONS ON POLYNOMIAL RINGS
525
then the eigenvalues of IJ operating on the Cartan subalgebra are e21fim;/., j = 1, 2, ... , l, where s is order of IJ. Now more generally TV operates on the zero weight space of VX for any AED according (say) to some representation 1l"x of W. As a generalization of the CoxeterColeman theorem one now has 0.11. For any XED the eigenvalues of 1l"X(IJ) are e2ri m;(X)/., j=l, 2, ... , z".• THEOREM
0.2. By applying the Birkhoff-Witt theorem the results above carry over from S to U, the universal enveloping of 9 (U is obviously a Gmodule in a natural way). THEOREM 0.12. Let U be the universal enveloping algebra over 9 and let ZC U be the center of U. Then U is free as a Z-module (under multiplication). In fact
(0.2.1)
U=Z®E
where E is the subspace (and G-submodule) of U spanned by all powers Xk for all nilpotent elements xEg. Moreover E is equivalent to H as a G-module so that every irreducible representation of G occurs with finite multiplicity in E (in fact pX occurs h.. times in E for any AED). Let V be a finite dimensional irreducible U-module so that one has a G-module algebra epimorphism p:
U --t End V.
Since p(Z) reduce to the scalars it follows = End V. Now let Y be any subspace of U. then it is due to Harish-Chandra that there module V such that p is faithful on Y. This dim Y~ 2. However it is true if YCE.
from (0.2.1) that peE) If Y is one-dimensional exists an irreducible Uis not true in general if
THEOREM 0.13. Let YCE be any finite dimensional subspace. Then there exists an irredudble U-module V such that p is faithful on Y.
I would like to express my thanks to C. Chevalley, M. Rosenlicht and A. Seidenberg for helpful conversation about questions in algebraic geometry. In particular to Seidenberg for making me aware of his criterion for normality and to Chevalley for simplifying my proof of the primeness of J+ S. REFERENCES
1. A. Borel and Harish-Chandra, Arithmetic subgroups of algebraic groups, Ann. of Math. (2) 75 (1962), 485-535.
316
526
BERTRAM KOSTANT
2. C. ChevaIley, Invariants of finite groups generated by reflections, Amer. J. Math. 77 (1955), 778-782. 3. - - - , Fondements de las geometric algebrique, Course notes at the Institut Henri Poincare, Paris, 1958. 4. A. J. Coleman, The Betti numbers of the simple Lie groups, Canad. J. Math. 10 (1958), 349-356. S. E. B. Dynkin, Semi-simple subalgebras of semi-simple Lie algebras, Amer. Math. Soc. Transl. (2) 6 (1957), 111-244. 6. F. Gantmacher, Canonical representation of automorphisms of a complex semisimple Lie group, Mat. Sb. 47 (1939), 104-146. 7. Harish-Chandra, On a lemma of F. Bruhat, J. Math. Pures Appl. (9) 3S (1956), 203-210. 8. - - - , On representations of Lie algebras, Ann. of Math. (2) 50 (1949), 900-915. 9. G. Hochschild and G. D. Mostow, Representations and representative functions on Lie groups. III, Ann. of Math. (2) 70 (1959), 85-100. 10. S. Helgason, Some results in variant theory, Bull. Amer. Math. Soc. 68 (1962), 367-371. 11. N. Jacobson, Completely reducible Lie algebras of linear transformations, Proc. Amer. Math. Soc. 2 (1951),105-133. 12. B. Kostant, A formula for the multiplicity of a weight, Trans. Amer. Math. Soc. 93 (1959), 53-73. 13. - - - , The principal three-dimensional subgroup and the Bette numbers of a complex simple Lie group, Amer. J. Math. 81 (1959), 973-1032. 14. M. Rosenlicht, On quotient varieties and the affine embedding of certain homogeneous spaces, Trans. Amer. Math. Soc. 101 (1961), 211-223. 15. J. P. Serre, Faisceaux algebriques coherent, Ann. of Math. (2) 61 (1955), 197278. 16. A. Seidenberg, The hyperplane sections of normal varieties, Trans. Amer. Math. Soc. 64 (1950), 357-386. 17. R. Steinberg, Invariants of finite reflection groups, Canad. J. Math. 12 (1960), 616-618. 18. O. Zariski and P. Samuel, Commutative algebra, Vol. I, Van Nostrand, Princeton, N. J., 1958. 19. - - - , Commutative algebra, Vol. II, Van Nostrand, Princeton, N. J., 1960. MASSACHUSETTS INSTITUTE OF TECHNOLOGY
317
Reprinted from the AMERICAN JOURNAL OF MATHEMATICS Vol. LXXXV, Number 3, July, 1963 Pp. 327-404
LIE GROUP REPRESENTATIONS ON POLYNOMIAL RINGS.*l By BERTRAM KOSTANT. 2
o.
Introduction. 1. Let G be a group of linear transformations on a finite dimensional real or complex vector space X. Assume X is completely reducible as a G-module. Let S be the ring of all complex-valued polynomials on X, regarded as a G-module in the obvious way, and let J C S be the subring of all G-invariant polynomials on X. Now let J+ be the set of all f E J having zero constant term and let H C S be any graded subspace such that S = J+S H is a G-module direct sum. It is then easy to see that
+
(0.1.1)
S=JH.
(Under mild assumptions H may be taken to be the set of all G-harmonic polynomials on X. That is, the set of all f E S such that of = 0 for every homogeneous differential operator f) with constant coefficients, of positive degree, that commutes with G.) One of our main concerns here is the structure of S as a G-module. Regard S as a J-module with respect to multiplication. Matters would be considerably simplified if S were free as a J-module. One shows easily that S is J-free if and only if S = J 0 H. This, however, is not always the case. For example S is not J-free if G is the two element group {I,-I} and dim X >2. On the other hand one has
Example 1. It is due to Chevalley (see [2]) that if G is a finite group generated by reflections then indeed S = J ® H. Furthermore the action of G on H is equivalent to the regular representation of G. Example 2. Sis J-free in case G is the full rotation group (with respect to some Euclidean metric on X. For convenience assume in this example that dimX> 3). Note that the decomposition of a polynomial according to the relation S = J ® H is just the so-called "separation of variables" theorem for polynomials. This is so because J is the ring of radical polynomials and H is the space of all harmonic polynomials (in the usual sense). * Received February 11, 1963. This research was supported in part by National Science Foundation Grant NSF· Gl9992. • The author is an Alfred P. Sloan Fellow. 1
327 B. Kostant, Collected Papers, DOI 10.1007/b94535_17, © Bertram Kostant 2009
318
328
BERTRAM KOSTANT.
N OW, for any x E X, let Ox C X denote the G-orbit of x and let S ( 0 x) be the ring of all functions on Ox defined by restricting S to Ox. Since J reduces to constants on any orbit it follows that (0. 1. 1) induces a G-module epimorphism H~S(Oa;). (0.1.2) Since our major concern is the case where X is a reductive Lie algebra and G is the adjoint group and since the methods used there belong to algebraic geometry we will assume now that X is complex and the G is algebraic and reductive. All varieties considered are over C. If Y has an algebraic structure R(Y) will denote the ring of everywhere defined rational functions in Y. Obviously one always has (0.1.3) On the other hand if Gx eGis the isotropy group defined by x E X then one has a G-module isomorphism (0.1. 4) The significance of (0.1. 4) is that one knows the G-module structure of R (G/G:c) completely by a very simple algebraic Frobenius reciprocity theorem (even though G:c may not be reductive). In fact if V>' is any irreducible G-module with respect to the representation v>' and V>. is the dual module then one has (0.1.5) where V>. G· is the space of vectors in V>. fixed under Gx" Now in Examples 1 and 2 (assume complexified) the following three optimum situations occur: (a) (b) those x (c)
Sis J-free so that S=J®H the map H ~ S ( 0 a;) is an isomorphism for certain x E X and for
R(Oa;) =S(O:c).
But one observes that if in any general case, (b) and (c) hold then, clearly, upon combining (0.1. 4) and (0.1. 5) one gets the G-module structure of H. If one gets in addition the" graded" G-module structure of Hand knows the structure of J then one gets the full graded G-module structure of S in case (a) also holds. In Example 2 the conditions (b) and (c) hold for any x =F 0 (even if (x, x) = 0) . In fact, classically, one has exploited (b) and (c) for (x, x) > 0
319
329
LIE GROUP REPRESENTATIONS.
to solve the Dirichlet problem with the sphere as boundary. That is, if f is any continuous function on the sphere one first expands f as a Fourier development of spherical harmonics fm- The sphere is 03) n Rn and the fm are in R (03)). The equality R (03)) = 8 (03)) and the isomorphism H ~ 8 (03)) then yields the extension of fm uniquely as harmonic polynomials hm on X. But this yields the desired extension of f. In Example 1 the conditions (b) and (c) are satisfied for any" regular" element x E X. Our first concern in this paper is to give criteria for (a), (b) and (c) to hold in general. Since our interest is in the continuous case we will assume G is connected (and hence a variety). Thus Example 2 rather than Example 1 serves as a model. Now let P C X be the cone of common zeros defined by the ideal J+8 in 8. Let X* be the dual space to X and let p* C X* be defined in a similar way with the roles of X and X* interchanged. As a criterion to establish (a) and more we prove PROPOSITION 0.1. Assume (1) that J+8 is a prime ideal in 8 and (2) there exists an orbit Oe C P which is dense in P. Then 8 = J 0 H. Furthermore if G is a subgroup of the complex rotation group then H may be taken as the space of all G-harmonic polynomials. Moreover H then coincides with the space spanned by all powers fk where f E P*.
It may be observed that the criterion is satisfied in Example 2.
An element x E X is called quasi-regular if P C C* . ()~. establish (b) is given by
A criterion to
PROPOSITION 0.2. Assume conditions (1) and (2) of Proposition 0.1 are satisfied. Then the G-module epimorphism H ~ 8(03)) is an isomorphism for any quasi-regular element x E X.
It may be observed that in Example 2 every nonzero x E X is quasiregular.
From known facts in algebraic geometry one has the following criterion to insure (c). PROPOSITION 0.3. Let x E X and assume (1) the closure 03) is a normal variety and (2) 0 IlJ - O:c has a codimension of at least 2 in 0 OJ. Then R(O:c) = 8(03)).
It may be observed that the conditions of Proposition 0.3 are satisfied for every x E X in Example 2.
320
330
BERTRAM KOSTANT.
Now assume that X = g is a complex reductive Lie algebra and G is the adjoint group. Here the structure of J is given by a theorem of Chevalley. This aserts that J is a polynomial ring in l (the rank of g) homogeneous generators Ui, i = 1,2,' . " l with deg Ui = mi 1 where the mi are the exponents of g. Now one knows that here P is the set of all nilpotent elements of g ([13J, Theorem 9.1). But then by [13J, Corollary 5.5, P does contain a dense orbit Oe, namely, the set of all principal nilpotent elements in g. Thus to apply Propositions 0.1 and 0.2 one must prove that J+S is a prime ideal. If n = dim g (all dimensions are over C) then one sees easily that n-l is the maximal dimension of any orbit. Let t={xEgldimO",=n-l}. Any regular element x E g belongs to t. But also e E t for any principal nilpotent element. These in fact are extreme cases.
+
+
PROPOSITION 0.4. Let x E g be arbitrary. Write (uniquely) x = y z where y is semi-simple, z is nilpotent and [y, z J = O. Let gY be the centralizer of y in g so that gY is a reductive Lie algebra and z E gY. Then x E t if and only if z is principal nilpotent in gY.
Let x E g. Consider the values (dui) '" of the l-differential forms dUi, i = 1, 2,' . " l, at x. It is known that these covectors are linearly independent whenever x is regular. (One recalls that the product of the positive roots is the determinant of an l X l minor of a certain n X l matrix determined by the duo-) But to prove the primeness of the ideal J+S one needs to know that these covectors are linearly independent if x is a principal nilpotent element. This fact is contained in THEOREM O. 1. Let x E g. Then the (dui) x is linearly independent if and only if x E t. Proposition 0.1 may now be applied. THEOREM 0.2. One has S = J 0 H where H is the space of all Gharmonic polynomials on g. Furthermore H coincides with the space of all polynomials spanned by all powers of nilpotent n linear functionals. U
Since Theorem 0.1 shows also that P is a complete intersection the decomposition S = J @ H when combined with Proposition 5, § 78, in F AC [15J, gives, in the notation of FAC, all the sheaf cohomology groups Hj (P, (fj (m» where P is the projective variety defined by P. Another application of the primeness of J+S in algebraic geometry is THEOREM 0.3. The intersection multiplicity of P, at the origin, with any Cartan subalgebra is w, where w is the order of the Weyl group.
321
LIE GROUP REPRESENTATIONS.
331
Next, Proposition O. 2 is put into effect for all orbits of maximal dimension by 0.4. The set r coincides with the set of all quasi-regular elements in g. (Thus Hand S (0 aJ) are isomorphic as G-modules for any x E r.) THEOREM
As a consequence of Theorems 0.2 and 0.4 one shows that not only is the ideal J+S prime in S but J 1 S is prime for any prime ideal J 1 C J. Furthermore one gets the following characterization of all the invariant prime ideals in S which are generated by elements of J. THEOREM 0.5. Let I C S be any G-invariant prime ideal. Let u C g be the affine variety of zeros of I. Then I is of the form 1= JIS, for J I a prime ideal in J, if and only if u n r is not empty.
Since R ( 0 II!) = S ( 0 II!) in case 0 II! is closed and since 0 II! is closed if x is regular one gets the G-module structure of H by applying Theorem 0.4 and (0.1. 5) for x regular. Thus if D denotes the set of dominant integral forms corresponding to a Oartan subgroup A, so that D indexes all the irreducible representations of G as highest weights, then one has (0.1.6)
where lx = dim VXA is the multiplicity of the zero weight of vx. In order to determine the G-module structure of Sk, the space of homogeneous polynomials on g of degree k, one must know more than (0.1. 6). In fact using the relation S = J ® H what one wants is the multiplicity of vX in Hj = Sj n H for any A and j. As it turns out, for this, one needs R (Oe) = S (Oe) where e is a principal nilpotent element. To show the latter using Proposition O. 3 it is enough to show that P is a normal variety and P-Oe has a co dimension of at least 2 in P. Let @r be the set of all orbits of maximal dimension (n -l). The set @r may be parametrized by CZ in the following way. Let u:
g~Cl
be the morphism given by putting u (x) = (u l (x),' . " Uz (x» Since u reduces to a constant on any orbit it induces a map
for any x E g.
'YJr: @r~Cz.
It is known that u induces a bijection from the set of all orbits consisting of semi-simple elements onto CZ (for completeness a proof of this fact will be given here). Oombining this with Proposition 0.4 one obtains
322
332
BERTRAM KOSTANT. THEOREM
0.6. 'fJt is a bijection.
Thus to each ~ E C! there exists a unique orbit, Ot (~), of dimension n-l which correspond to ~ under 'fJt. Now let pa) = u- 1 (~) for any ~ E C! so that g=UP(~) ~EC'
is a disjoint union. of C!. One proves THEOREM
0.7.
Note that pa) =p and Ota) =0. if For any
~E
~
is the origin
C! one has
P(~) =
Of(~)
so that pa) is a variety of dimension n-l. Moreover P(~) is a complete intersection and Ora) coincides with the set of non-sing1tlar points on P(~). Finally P (~) is a finite union of orbits so that Oil} is a finite union of orbits for any x E g.
Since P(~) is a complete intersection and since its singular locus is the complement (a finite union of orbits) of Of(~) in pa) one would get the normality of P(~) by a theorem of Seidenberg if one knew the dimension of the other orbits in pa) were at most n-l-2. Now it is well known that dim Oil! is even (and hence dimR Orc is a multiple of 4) for any semi-simple element x E g. Less known is the following proposition observed independently by the author, Borel, and (most simply proved by) Kirillov. PROPOSITION
0.5. The dimension of O{ll is even for any x E g.
Combining Theore:m
o. 7
and Proposition
o. 5 one
obtains
THEOREM 0.8. Let ~ E C! be arbitrary. Then P (~) is a normal variety and the co dimension of P(~) -Ot(~) in P(~) is at least 2.
Applying Proposition 0.3 one then has THEOREM 0.9. Let x E t. Then R(O{ll) = 8(0:1). (This implies that all R(OIl}) for xE t are isomorphic as G-modules; even though they are not in general as rings.) Let ~=u(x). Then R(O{ll) (=R(G/Grc» is an affine algebra (even though Orc is not necessarily an affine variety) and P(~) is the variety of all maximal ideals of R (0(II) • Thus the embedding of G/ Grc in g as Oil! is special in that any morphism of G/Gm (or Om) into any affine variety extends uniquely to a morphism of P(~) = Orc into the variety. (In particular
323
333
LIE GROUP REPRESENTATIONS.
this holds for 0 e and {j m = P. ) Finally (using (0. 1. 5) and the equality R(Oa;) = 8(Ox)) one has, for any A ED
dim VAaz=h
(0.1.7)
so that the left side of (0.1. 7) is independent of x E r.
Now let {e_,xo,e} be a principal 8-triple (that is, a "canonical" basis of a principal three dimensional simple Lie sub algebra ) . In particular then e is a principal nilpotent element. Used heavily in the theorems above is the result of [13] which asserts that ge is l-dimensional and has a basis Zi, i = 1, 2,' . " t, such that (0.1.8) where, we recall, the mi are the exponents of g. The main application of this is the following result: Let a be any subspace of g such that (1) g= a [e_, g] is a direct sum and (2) a is stable under ad Xo (e. g. take a = ge). Then if b is the l-plane defined by the translation b = e_ a one has THEOREM 0.10. The variety b is contained in r. Moreover each orbit in (fj r intersects b in one and only one point. Finally the mapping f ~ fib induces an isomorphism of J onto R (b) .
+
+
Remark. If g is the set of all l X l complex matrices then one shows easily that r is the set of all matrices whose characteristic polynomial is equal to their minimal polynomial. An example of the subvariety b is the set of all "companion" matrices. Here the validity of Theorem 0.10 is a wellknown fact in matrix theory. Now since ge = ga e (because ge is commutative) and since (0.1. 7) holds for x = e this suggests a generalization of the notion of exponent. Let V be any finite dimensional G-module with respect to a representation v. If lv is the multiplicity of the zero weight of v then by (0.1. 7) one has dim Va e = lv. It follows therefore that there exists a unique non-decreasing sequence of non-negative integers mi( v), i = 1, 2,' . " lv, such that one has v ( Xo ) Zi =
mi ( v ) Zi
for a basis Zi of Va e • If v is the adjoint representation the m, (v) are the usual exponents. If v=v A we will write m,(A) for m,(v A ) and note (because the highest weight has multiplicity one) that
mj(A) =O(A) for j=l"ll. where
0
(A) is the sum of the coefficients of A relative to the simple roots and
324
334
BERTRAM KOSTANT.
that this highest value occurs with multiplicity one among the generalized exponents m.(A). (This specializes to the familiar relation ml=o(o/) when g is simple and", is the highest root.) The following theorem now gives the G-module structure of Hi and hence Sk for any j and le. THEOREM 0.11. Let A ED be arbitrary and let H (A) be the set of Gharmonic polynomials which transform under G according to vA. Let (by
(0. 1. 6»
IA
="'2:. Hi (A)
H (A)
be a decomposition into irreducible components
i=1
so that Hj(A)C Hnl where nj, j = 1, 2,· .. , lA, is a non-decreasing sequence of integers. Then nj = mj(A) for all j. In particular then le = 0 (A) is the highest degree le such that vA occurs in Hk. Moreover it occurs with multiplicity one for this value of le.
Assume for convenience that g is simple and let tf; E D be the highest root. Let Xi, i = 1, 2,· .. , n be a basis of g. If the Uj E J are chosen properly one sees that
~Uj , i = 1, 2,· . ., n, is a basis of
uXj
H j (tf;).
One notes then that
Theorem 0.11 is a generalization of the result in [13J given by (0.1. 8). H. S. M. Coxeter observed and A. J. Coleman proved in [4J that if W is the Weyl group and '()" E W is the Coxeter-Killing transformation then the eigenvalues of '()" operating on the Cartan sub algebra are e2 11'iffll/8, j = 1, 2,· .. , l, where s is order of fT. Now more generally W operates on the zero weight space of VA for any A E D according (say) to some representation 7rA of W. As a generalization of the Coxeter-Coleman theorem one now has THEOREM
O. 12.
For any A E D the eigenvalues of
7rA
(fT) are e211'iffll(A)/S,
j= 1, 2,· .. , h. 0.2. By applying the Birkhoff-Witt theorem the results above carry over from S to U, the universal enveloping of g (U is obviously a G-module in a natural way). THEOREM 0.13. Let U be the universal enveloping algebra over g and let Z CUbe the center of U. Then U is free as a Z-module (under multiplication) . In fact (0.2.1) U =Z®E
where E is the subspace (and G-submodule) of U spanned by all powers xk for all nilpotent elements X E g. Moreover E is equivalent to H as a G-module so that every irreducible representation of G occurs with finite multiplicity in E (in fact vA occurs lA times in E for any A E D).
325
LIE GROUP REPRESENTATIONS.
335
Let V be a finite dimensional irreducible U-module so that one has a Gmodule algebra epimorphism p: U~End
V
Since p(Z) reduce to the scalars it follows from (0.2.1) that p(E) = End V. Now let Y be any subspace of U. If Y is one-dimensional then it is due to Harish-Chandra that there exists an irreducible U-module V such that p is faithful on Y. This is not true in general if dim Y > 2. However it is true if YCE. O. 14. Let Y C E be any finite dimensional subspace. Then there exists an irreducible U-module V such that p is faithful on Y. THEOREM
I would like to express my thanks to C. Chevalley, M. Rosenlicht and
A. Seidenberg for helpful conversations about questions in algebraic geometry. In particular to Siedenberg for making me aware of his criterion for normality and to Chevalley for simplifying my proof of the primeness of J+S. l. Consequences of the primeness of J+8 and a dense orbit in P. 1. Let X be a n-dimensional vector space over the complex numbers C. Let S* = S* (X) symmetric algebra over X. One knows that S* may be regarded as the algebra of all differential operators 0 on X which may be put in the form
8=}: ail"""in(8~)il. . (8~ )in where the a;,"""in are complex constants and Zl,' " ., Zn are the affine coordinates of X. Let S* = S* (X) (or just S) be the symmetric algebra over the dual space to X. Then S is just the ring of all polynomials on X. In fact we take the point of view that X is an affine variety (over C) and S is its ring of everywhere defined rational functions. The algebra S (resp. S*) is graded in the obvious way and a subspace L C S (resp. L C S*) will be called graded if it is spanned by its homogeneous components LJ =L n Sj (resp. L n Sj). N ow one knows that a non-singular pairing of S* and S into C is established by putting (1.1.1) <8,f>=of(0) where 8 E S*, f E Sand 8f(0) denotes the value of the function 8f at the origin. In this way Sic is orthogonal to Sj if j =F k and becomes isomorphic to its dual if k = j. It is obvious from (1.1.1) that
(1.1.2)
326
336
BERTRAM KOSTANT.
for any 01, O2 E S* and f E S and hence in particular if by the Taylor expansion
<(~( ,f> =
(1.1.3)
f E Sm
and x E X then
f(x)
where Ox is the element of S1 (X) :::: X corresponding to x. Now assume that G C Aut X is a connected linear reductive algebraic group, i. e., G is the complexification of a connected compact subgroup of Aut X. We regard G as not only operating on X, but by unique extension, as a group of algebra automorphisms of S* (X) and also as a group of algebra automorphisms of S (X). The action on the latter is also uniquely defined by requiring that
= <0, f>
(1.1.4)
for all a E G, 8 E S* and f E S. Note by (1.1. 3) that
(a· f) (x)
(1.1.5)
=
f(a- 1 x)
for any xE X, fE Sand aE G. Now let J C S be the graded subring of G-invariant polynomials in S. That is J = {f E S I a' f = f for all a E G}, and let J+ = {f E J I f(0) = O}. We will often be concerned with the homogeneous ideal J+S in S generated by J+. 1. 'Let L is a direct sum.
PROPOSITION
S = J+S
+L
be any graded subspace of S Then
such that
Proof. We must show Sk C J L for all k. This -is obvious if k = 0 since Assume it is true for Si where j < k. But since one clearly has
So C L.
Sk+1 C J+S(k)
+ L where S(k)
k
=
LSi it is then obviously true for Sk+1. Q. E. D. • =0
Weare interested first of all in the question as to when S is free over J, or more specifically, as to when S is a tensor product of J and L. (Choosing L to be G-stable, such a decomposition of S reduces the study of its G-module structure to that of L.) We first observe that the two conditions are equivalent. The expression linear independence (resp. basis) without any reference
327
337
LIE GROUP REPRESENTATIONS.
to a ring always means linear independence (resp. basis) with respect to C. Furthermore, tensor product without reference to a ring means tensor product over C. LEMMA
1.
The following conditions are equivalent,'
1. Let L be as in Proposition 1. Then the map J0L~S
given by f i8) g ~ fg is an isomorphism. 2. S is free over J. 3. Let M C S be any subspace such that M n J+S = (0). Then for elements in M linear independence is equivalent to linear independence over J. Proof. Obviously (1) => (2) since a C basis of L defines a J basis of S. Assume (2) and let~, i=1,2,·· ., define a J basis of S. Let f1,' . ',h be linearly independent elements of M. For j = 1,' .. , le write fj = L G,;je; i
where we may assume i = 1, 2,' .. , p and G,;j E J. To show that the fj are linearly independent over J it clearly suffices to see that the le X p matrix (G,;j) with entries in J is of rank le. If f' denotes the image of f E S in SjJ+S it is clear that f'j=LG,;j(O)e'i' Since the fj are obviously linearly independent it is clear that the C matrix G,;j (0) is of rank le. Hence not all le X le minors of the J matrix (G,;j) can be zero. This proves that the fj are linearly independent over J. It is trivial that linear independence over J implies linear independence. To obtain (1) from (3) we simply put M =L and apply Proposition 1.
Q.E.D. 1. 2.
Now for any x E X let
00;= {yE X
I y=ax, for
A subset 0 C X is said to be an orbit if 0 For any x E X let
=
some aE G}. Oil! for some x EX.
It is clear that Gx is an algebraic (hence a closed, complex Lie subgroup) subgroup of G. Furthermore if
(1.2.1) is the map given by
/3'x (a)
=
ax then /3'Il! induces a bijection
(1.2.2)
328
338
BERTRAM KOSTANT.
Now let U be universal enveloping algebra of the Lie algebra of G. Then since the representation of G on S induces a representation of its Lie algebra on S it is clear that S becomes a U-module by further extension to U. Obviously Sk is a U-submodule for any k. We denote by P . f E S the effect of applying p to f where p E U and f E S. For any subset P C X and f E S let f I P be the restriction of f to P.
Let ff E S, j = 1, 2,' . " k, and let pi, i = 1, 2,' . " be a basis of U. Oonsider the k-column matrix D = (dij) where dij = Pi' ff. (The matrix thus has entries in S and hence, for any x E X, D (x) = (d ij (x» is a C matrix.) LEMMA
2.
Then if x E X the functions fi I O:c on 0 31, j independent if and only if D (x) has ranle le.
=
1, 2,' . " k, are linearly
Proof. Let 0 be the algebra of all holomorphic functions on the complex homogeneous space GIGa:. It is clear that 0 is a module for the Lie algebra of G since the latter defines holomorphic vector fields on GIG3I. Hence 0 also becomes a U-module and in such a way that if a:
S~O
is the homomorphism given by af = f 0 {3:c one has
a (p . f)
(1. 2. 3)
=
P . af
for any f E S. Furthermore if s E GI Ga: denotes the point corresponding to Ga: and g E a then since the Lie algebra of G spans the holomorphic tangent space at s it follows from the Taylor expansion that g vanishes identically on GIG3I if and only if (p' g) (s) = 0 for all p E U. Now assume that 'rank D(x) < k. Then there exists a non-zero vector (c 1 , ' . ',Ck) ECk such that ~dij(X)Cj=O for all i. Thus if f=~CjfjES j
j
one has (p' f) (x) = 0 for all p E U. Thus by (1. 2. 3) (p' af) (s) =0 for all p E U and hence af vanishes identically on GI Ga:; or equivalently flO:c is zero. Hence the fj I O:c are linearly dependent. Conversely assume that the fj I 031 are linearly dependent so that (~cdj) I O:c is zero for a non-zero vector (c1 , ' . " Ck) E Ck. But then a (~cd,) = 0 and hence for i = 1, 2,' J
~
Cjp• . f,(x)
=
(Pi .
a(~
j
i
cifi» (s)
=0 Thus rank D(x)
Q.E.D.
< le.
329
339
LIE GROUP REPRESENTATIONS.
As a corollary we obtain the following criterion for linear independence over J. LEMMA 3. Let fi E S, i = 1, 2,' . " k. Assume the functions fj I 0 are linearly independent for some orbit 0 ex. Then the fi are linearly independent over J.
Proof.
Assume
~
gifj = 0 where gi E J, j
=
1, 2,' . " k.
Let D be the
j
matrix given in Lemma 2 and let x EO. Then by Lemma 2 there exists a k X k minor of D whose determinant et E S is such that et (x) =F O. But then there exists a neighborhood W of x such that et(y) =F 0 for all yEW. Thus D(y) has rank k and hence the fi lOy are linearly independent for all yE W. But since the gi reduce to constants on any orbit Oy it follows from the relation ~ gifi = 0 that the gj vanish identically on W. This implies that the gj j
vanish on X since the gj are polynomials. 1. 3.
Q. E. D.
For any subset Y C X we will let
I(Y)={fES! f!Y=O} be the ideal in S defined by Y. Now let P C X be the cone (since J+ is homogeneous) given by P = {x E X
I f (x) =
0 for all
f E J +}
Since P is defined by the ideal J+S one knows that I (P) is the radical of J+S and that the cone P is irreducible (in the sense of algebraic geometry) and J+S=I(P) if and only if J+S is prime. We now give a criterion for the conditions of Lemma 1 to be satisfied. PROPOSITION 2. Assume (1) J+S is prime and (2) there exists an orbit such that (5 = P. Then the conditions of Lemma 1, § 1. 1, are satisfied. In particular if S = J+S L is a direct sum where L is a graded subspace of S then the map
o
+
(1.3.1)
given by f 0 g ~ fg, is an isomorphism. Proof. Let M C S be any subspace such that M n J+S = (0). Since J+S = I (P) it is clear that if fj E M, j = 1, 2,' . " k, are linearly independent then the fj I P are linearly independent. But this obviously implies that the fj I 0 are linearly independent since (5 = P. But then the fj are linearly independent over J by Lemma 3 and thus the result follows by (1) and (3) of Lemma 1. Q. E. D.
330
340
BERTRAM KOSTANT.
Remark 1. In the proof we have only used the fact that J+8 = I (P) and not that J+8 is prime. However assumption (2) already implies that the cone P is irreducible (recall that G is connected) so that there is no loss in assuming that J+8 is prime. 1. 4. Now assume that B is a symmetric non-singular G-invariant bilinear form on X. Then, as one knows, B induces a unique G-module ring isomorphism (also written B) B: 8*~8
(1.4.1)
of degree zero where <0"" Boy) = B (x, y) for any x, y E X. 01 , O2 E 8* one has
Obviously for any
(1. 4. 2)
Now let
and let J *+ be the space of elements in J * having zero constant term. It is obvious that
An element
f E8
is called G-harmonic in case of=O
for all 0 E J *+. Let H be the (obviously graded) space of all G-harmonic polynomials in 8. By (1.1. 2) and (1.1. 4) it is clear that H is a G-submodule of 8. One obtains a class of G-harmonic polynomials in the following way: Let pI be the cone in 8 1 defined by putting P'=B{(3)E 8 1
I xE P}
and let Hp C 8 be the (graded) space spanned by all powers zm, m = 0,1,' . " for all z E P'. The following proposition was proved independently by Helgason (see [10]). PROPOSITION 3. One has Hp C H. Furthermore 8 G-module direct sum so that by Proposition 1, § 1. 1,
=
J+8
+H
is a
Proof· Let zE P' so that z=B(o",) for xE P. Let BE J](, where k> O. We wish to show that ozm = O. We may assume that m > k and hence by
331
LIE GROUP REPRESENTATIONS.
(1.1. 2) it suffices to show that
=
0 for all a1 E 8m-k'
341 But by
where f = Ba and fl = Bal' But f E Jk and hence ffl = g E J+8. Thus
=
l,
af>
=0 for all a1 E 8 and aE J *+. I t follows immediately then that H' = H. To prove that 8 = J+8 H is a direct sum therefore it suffices, by dimension, (one restricts to 8 k ) to show that H n J+8 = o. (One already uses that B(J*+8*)=J+8). But to show HnJ+8=0 it suffices to show that B induces a non-singular bilinear form on J*+8*. Now let K be a maximal compact subgroup of G. Then one knows there exists a real subspace X R C X such that (1) X R is stable under K, (2) B is positive definite on X Rand (3) X = X R iXR is a real direct sum. But by (2) it is obvious that B induces a positive definite bilinear form on 8*.R(XR) =8R. But by (1) and (3) J*+ is the complexification of its intersection with 8 R since G is the complexification of K. Hence J*+8* is also the complexification of its intersection with 8 R and consequently B induces a non-singular bilinear form on J*+8*. Q. E. D.
+
+
Now combining Propositions 2, § 1. 3, and 3, § 1. 4, we obtain the following "separation of variables" theorem. PROPOSITION 4. Assume that G leaves invariant a symmetric nonsingular bilinear form on X.
Let 8, Hand J be, respectively, the ring of all polynomials on X, the space of G-harmonic polynomials on X and the ring of invariant polynomials onX. Let P C X be the homogeneous affine variety defined by the ideal J+8 in 8. Now assume (1) that there exists an orbit 0 such that (j =P and (2) the ideal J+8 is prime. Then the mapping (1. 4. 3)
J®H-", 8
332
342
BERTRAM KOSTANT.
given by f0 g~ fg is a G-module isomorphism. Furthermore if P' C 8 1 is the cone corresponding (by means of the bilinear form) to P and Hp is the subspace of 8 generated by all powers zm, z E pI, m = 0, 1,· . ., then
(1.4.4) Proof. Since G operates as algebra automorphisms of 8 it is obvious that (1. 4. 3) is a G-module map. But it is also an isomorphism by Propositions 2 and 3. Now let H* C 8* be the space generated by all powers fJa;m where xEP. Let B be as in (1.4.1). Then clearly B(H*) =H p. Now if Hpm?,= Hm then by dimension and Proposition 3 there exist a non-zero fEHm such that
°
Remark 2. A familiar instance as to when the conclusion of Proposition 4 holds is the case where X = C", n > 2, and G is the full complex rotation group. Here J is the algebra generated by 1 and Z1 2 zn 2 , H is space of all polynomials which satisfy Laplace's equation.
+ ... +
n fJ2f ~-fJ2=0
(=1
Z..
+ ... +
and P is the conic given by Z12 zn2 = 0. 1£ n > 3 one obtains the classical separation of variables theorem as a consequence of Proposition 4 since (1) P=O, where 0 is the set (easily seen to be an orbit) of all vectors xE such that xE P and x ?'= 0, and (2) J+8 is a prime ideal, because it is zn2 and this polynomial is irreducible if n > 3. generated by Z1 2
en
+ ... +
1. 5. For any element x E X besides Om, we may consider the orbit Ocm where c E C* is any non-zero scalar. It is obvious of course that Oca; = cOa;. Now where P C X is the cone defined in § 1. 3 and x E X is arbitrary put
Pa;= U Oca;nP CEC*
It is clear of course that Pa; C P is stable under the action of G. An element x E X will be called quasi-regular if
Pa;=P Remark 3. Note that if P = 0 for an orbit 0 then any element of 0 is quasi-regular. Now for any subset W C X let 8 (W) be the ring of all functions on W of the form g I W where g E S. Note that if W is stable under G then 8(W) is a G-module with respect to the action of G given by (1. 1. 5) where f E 8(W).
333
343
LIE GROUP REPRESENTATIONS.
Let x EX. We are particuiarly interested in the ring S (Oa;) of functions on the orbit Ox. Obviously the map
(1. 5.1) defined by the correspondence
f-7 f I Ox
is a G-module epimorphism.
As in Proposition 1, § 1.1, let L C S be any graded subspace such that S = J+S L is a direct sum. Also for any x E X let PROPOSITION
5.
+
Yx: L-7S(Ox)
be the linear map obtained by restricting (1. 5.1) to L. epimorphism.
Then yx is an
Assume that conditions (1) and (2) of Proposition 2, § 1. 3, are satisfied. Then yx is an isomorphism for any quasi-regular element x EX. In particular if G leaves invariant a non-singular symmetric bilinear form on X and, as in Proposition 4, § 1. 4, L = H is the space of G-harmonic polynomials on X then (1.5.2)
ya;: H -7 S(Ox)
is a G-module isomorphism for any quasi-regular element x E X. Proof. Since J reduces to scalars on any orbit Ox it follows from Proposition 1 thatyx maps L surjectively onto S (Ox), Assume now that conditions (1) and (2) of Proposition 2 are satisfied. Let x be quasi-regular. We must show that yx is injective. Let f E L. Since L is graded we may write
f=
k
~ i=l
Ci/i where fi E L
is homogeneous of degree ni, and the
fi'
i
=
1, 2,' .. , k,
are linearly independent. But now since J+S is the prime ideal corresponding to P it follows that the functions fi I P are linearly independent. But since P = (5 for an orbit 0 one has also that the functions fi I 0 are linearly independent. The argument of Lemma 2, § 1. 2, shows that for any yEO there exists a neighborhood W of y in X such that for any z E W the functions fi I Oz are linearly independent. But now since yEP = P x there exists a non-zero scalar c such that OC!l} = Oz for some z E W. Hence there exists a non-zero scalar c such that the functions fi I OC!l} are linearly independent. Now let
334
1344
BERTRAM KOSTANT.
be the bijection defined by y ~ 1/c· y. If JL* is then the corresponding contravariant isomorphism on functions one has
But then since the fi I Ocx or l/C n'(fi I Ocx) are linearly independent it follows that the fi I Ox are linearly independent. But then if f I Ox is zero it follows that the Ci are all zero and hence f is identically zero. Thus yx is injective. The isomorphism (1. 5. 2) is a G-module map since (1. 5. 1) is a Gmodule map. Q. E. D. Remark 4. In the example of Remark 2, § 1. 4, not that :r E en is quasi-regular if and only if x ~ O. Thus in that example one has that (1.5.3) is an isomorphism for any x alent as G-modules.
~
Hence all S (Om) where x
O.
~
0 are equiv-
1. 6. In order to apply Propositions 3,4 in § 1. 4 or Proposition 5, § 1. 5, one needs to know that J+S is a prime ideal. In general this appears to be difficult to ascertain even if one knows J completely. (Except of course if J has only one ring generator, as in the example of Remark 2.) However we will now observe (Proposition 6, § 1. 6) that in the familial' case when J is a polynomial ring the question of the primeness of J+S reduces to a more manageable one. Throughout much of the remainder of the paper we will need to draw upon techniques and results in algebraic geometry. Our reference for all definitions will be [3] where for us the fixed algebraically closed field is of course C. We recall in particular that by definition, among other things, a variety is irreducible in its Zariski topology. To avoid confusion of terminology we remark here that the words open, closed, closure and denseness, etc. will have their usual Hausdorff topological meaning unless stated otherwise (i. e., unless preceded by "Zariski"). If fi E S, i = 1, 2,' . " l, are arbitrary let (fl" . " fz) denote the ideal in S that they generate. If Y C X is a Zariski closed subvariety of X of dimension n - l then we recall that Y is called a complete intersection in case
I(Y)
=
(fl,' . ',f!)
for some fiE I(Y), i=I,' . ·,l. Now for any fE Sand xE X let (df)1JJ be the value of the differential dt
335
LIE GROUP REPRESENTATIONS.
345
at x. If fi E 8, i = 1,· .. , l, then one knows that the (dfi) iIJ are linearly independent if and only if the n X l matrix (oa:!M (x), j = 1,· . ., n, has rank l where the Xj is any basis of X. The following lemma in one form or another is well known in algebraic geometry. 4.
LEMMA
Let fi E 8, i
=
set given by Y = {xE X
1, 2,· .. , l, and let Y be the Zariski closed
I Mx) =O,i=l,·
.. , l}.
Assume (1) Y is a subvariety of X (that is, assume Y is irreducible) and (2) there exists y E Y such that (dfi)Y, i = 1, 2,· .. , l, are linearly independent. Then Y is a subvariety of dim n -l. F1trthermore (1.6.1) so that (a)
(ft,· . ., fz) is a prime ideal and (b) Y is a complete intersection.
Proof. Let 8 y be the local ring of X at y. Let [ = (fl,· .. , fz). Since the (dfi)Y are linearly independent the fi may be included in a complete system of uniformizing variables at y. Thus by [3], Proposition 3, p. 219, [8 y is a prime ideal of 8 y • Furthermore since [(Y) is the radical of [ in 8 it is clear that [8 y is the ideal of Y at y (that is, [(Y)8 y =[8y ) so that. by the same reference, dim Y = n - l and [8 y n 8 = [ (Y) . To prove (1. 6. 1 ) it suffices to show that [ is primary for [( Y) since in that case [8 y n 8 = I (a primary ideal is equal to the contraction of its extension; see [13], Theorem 19, p. 228). But [ is primary by MacCaulay's theorem (see [19], p. 203) which asserts that there are no embedded primes for [ so that [( Y) is the only associated prime ideal. Q. E. D. Now we recall that G is a connected algebraic reductive group. Hence G has the structure of an affine variety. (It is Zariski closed in AutX but not necessarily Zariski closed in EndX.) Since (1. 2.1) is obviously a morphism it follows that any orbit 0 C X is an irreducible constructible set. In fact since 0 is epais ([3], Proposition 4, p. 95) and G operates transitively on it, it follows that 0 is a subvariety ([3], Theorem 5, p. 68) of X. It follows therefore that its (usual) closure (5 is a Zariski closed subvariety or the same dimension as o. As an application of Lemma 4, § 1. 6, we have 6. Assume J, as a ring, is generated by l homogeneous algebraically independent polynomials ~, i = 1, 2,· .. , l. PROPOSITION
336
346
BERTRAM KOSTANT.
N OW let ~ E c z, ~ = (~l'· .. , ~z), be an arbitrary complex l-tuple and let P(~)
={xEX I Ui(X)
=~.,i=1,2,·
Assume P (~) is not empty and there exists an orbit 0 pa)
(1.6.2)
. ·,l}
a)
such that
=O(~).
Then P (~) is a Zariski closed subvariety (of X) of dimension n-l. Furthermore the ideal (Ul - ~l'· .. , Uz- ~z) in S is prime if and only if there exists y EPa) such that the (du;,)y are linearly independent. In such a case P(~) is a complete intersection and the set P(~)8 of simple points on pa) is given by
(1.6.3)
P (~) 8 = {x E P (~) I (dU;,) "" i = 1, 2,· .. , l, are linearly independent}.
Proof. Since 0 a) is irreducible it follows from (1. 6. 2) that P (~) IS a subvariety of the same dimension as Oa). Now by [3J, Corollary, p. 102, it is clear that dimPa) >n-l.
To prove that dim P (0 = n - l it sufi1ces to show that dimO
(1.6.4)
for any orbit O. Let m be the maximum of the dimensions of all orbits. Let u C End X be the Lie algebra of G and for each x E X let cpx be the homomorphism of u into X given by CPa: (z) = Z (x) . It is then obvious that rank CPx = dim 0 X· By consideration of minors it is then clear that U={xEX I dimOx=m}
a non-empty Zariski open subset of X. Now let V be the Zariski open subset of X consisting of all x E X such that the (du.)", are linearly independent. To see that V n U is not empty it clearly suffices to see that V is not empty. But this is a known consequence of algebraic independence. Indeed if ujE S, j = l 1,· .. , n, are chosen so that Ui, i= 1, 2,· .. , n, is a transcendental basis of S then each element of a coordinate basis zj, j = 1, 2,· .. , n of X is algebraically dependent upon the Ui. Hence on a non-empty Zariski open set each dZ j is in the span of the du.. This proves that V and hence V n U is not empty. Now let x E V n U and let W be the n - l dimensional variety (see [3J, Proposition 3, p. 219) containing x whose prime ideal at x is ISIl! where I = (U1-U1(X),· . ·,uz-uz(x)) and S", is the local ring at x.
IS
+
337
LIE GROUP REPRESENTATIONS.
347
Since, obviously, O:c C Wand dim O:c = m it follows that m < n - l and this proves (1.6.4) and hence dimP(~) =n-l. Now if the ideal (Ul-~l" . ',u!-~!) is prime it must equal I(P(~») and hence (df)w for any xEPa) and fEI(pa» lies in the span of the (d(ui-~i))iJJ= (dui):C' But since dimP(~) =n-l one immediately obtains (1. 6. 3) by Zariski's criterion and since (P(~»)8 is not empty there exists yEP (~) such that the (dui)Y are linearly independent. Conversely if the latter holds (Ul-~l" . ·,U!-~l) is prime by Lemma 4, §1.6, and hence P(~) is a complete intersection. Q.E.D. For us Proposition 4, § 1. 4, will be put into effect by PROPOSITION 7. Assume J, as a ring, is generated by l algebraically independent homogeneous polynomials Ui, i = 1,2,' . " l.
Assume also that there exists an orbit 0 such that P = 0. Then P is a subvariety of dimension n - l . Moreover J+S is prime if and only if there exists yEP such that (dni)Y, i= 1, 2,' . " l, are linearly independent. Proof.
This is just the special case
~=
0 of Proposition 6.
Q. E. D.
2. Normality and the closure of an orbit. 1. If Y is any variety we let R(Y) denote the ring of everywhere defined rational functions on Y. Now let 0 be the ring of all holomorphic functions on G. If f EO, a E G, then the left (resp. right) translate a . f (resp. f . a) of f by a is the function defined by putting (a·f)(b)=f(a-1b) (resp. (f·a)(b)=f(ba- 1 ». It is obvious that a' fE 0 (resp. f· aE 0) for all fE 0, aE G. One knows that R (G) is a sub ring of 0 which in fact may be given by (see [9], Theorem 5.2) R ( G) = {f E 0 I space spanned by all a . f, a E G, is finite dimensional} It is obvious that R ( G) is stable under left and right translations.
Now let D denote the set of equivalence classes of all irreducible rational (equivalently, holomorphic) finite dimensional representations of G. For each AE D choose a fixed irreducible representation.
belonging to A. The dual space to VA will be denoted by V A and the irreducible representation of G on VA contragradient to vA will be denoted by VA' If M is any G-module we will let MA denote the set of all vectors in M which transform according to the irreducible representation vA. Since G is assumed to be reductive one knows that if each vector in M generates a finite
338
348
BERTRAM KOSTANT.
dimensional cyclic G-module then M is in fact a direct sum of the MA. In particular regarding R ( G) as a G- module under left translation one has that
R(G)
=
~
RA(G)
AED
is a direct sum. Since VA ( G) generates End VA one can be very explicit about the structure of RA ( G). In fact let dA be the dimension of V A and let Vi and v'j, i, j = 1, 2,' . " d A, be, respectively, a basis of VA and a basis of its dual space VA' Now let %A be the function on G defined by
(2.1.1) Then one knows that the dA2 functions defined in this way form a basis of RA ( G) . In particular RA ( G) is finite dimensional and in fact
dimRA(G)
=
dA2 •
Now assume that F eGis an algebraic (and hence closed, Lie) subgroup. Then one knows that G/F (space of left coset aF, a E G) has the structure of an irreducible algebraic variety where dim G/F=dim G-dimF and the ring R(G/F) of everywhere defined rational functions on G/F may be identified, in the obvious way, with the set of elements in R (G) that are right invariant under F. Now for any A ED let VAF be the space of all vectors in the dual space VA to VA that are fixed under all transformations on V A of the form VA (a) where aE F. Put
Now it is obvious that R(G/F) is a G-submodule (by left translations) of R (G). It is furthermore obvious that
The following is a special case of an algebraic Frobenius reciprocity theorem. We prove for it for completeness. PROPOSITION
of
VA
and let
W'j
8. For i
1,' . " dA and j
1,' . " dAF let
Vi be a basis be a basis of VAF. Also let hij be the function on G given by
=
hij(b)
=
=
Then kif E RA (G /F) and in fact the dAd AF functions defined in this way are a basis of RA(G/F).
339
LIE GROUP REPRESENTATIONS.
ThtlS so that v>" occurs with multiplicity d>..F in R(G/F). (2.1.2)
is a direct sum.
R(G/F)
=
~
>"€D
Furthermore
R>"(G/F)
Proof. The decomposition (2. 1. 2) is obvious since each element of R ( G) is an element of R ( G) and hence generates a finite dimensional sub8pace under the action (left translation) of G. Furthermore it is also obvious that the d>..d>..F functions hij defined in the proposition are in R>"(G/F) and (see (2.1.1)) are linearly independent. To prove the proposition therefore one simply has to show that every element of R>" ( G) invariant under right translation by elements of F is in the span of the hij • Assume that g E R>" ( G) and g . a = g for all a E F. Let %>.. be as in (2. 1. 1) (a basis of R>" ( G)) . Write g = ~ gi/Cji where cij E C defines a matrix and hence, relative to the basis v'j, a linear transformation Q; of V>... It suffices only to show that 1m Q; C V>..F. But the condition on g implies that (v>.. (a) -1)Q; = 0 for all a E F. This proves 1m Q; C V>..F'. Q. E. D. Remark 5. A case of importance for us is the case where F = A is a Cartan subgroup of G. Here V>..A is just the zero weight subspace, corresponding to A, of V>... To make it independent of A we will put h. = d>..A so that h. = multiplicity of the zero weight of VA (2. 1. 3). Remark 6. Since one knows that the multiplicity of any weight p. for v>.. is equal to the multiplicity of - p. for v>" it follows that h. is also the multiplicity of the zero ,weight of v>". 2. 2. Now we wish to apply the considerations of § 2. 1 to the case where F = Gx for any x E X. See § 1. 2. By Proposition 8 any question as to the complete reduction of R (G/Ga;) as a G-module becomes a question in the finite dimensional representation theory of G and how such representations restrict to Gx. Now, as we observed in § 1. 6, the orbit 03] is a subvariety of X. Furthermore the bijection f3x: G/GJJ~ Ox induced by f3'x is an algebraic isomorphism (this follows easily from the transitivity of G together with [3J, Corollary, p. 53 and Corollary 2, p. 90. (See also [lJ, § 2. 2.) Thus if R(Oa;) is regarded as a G-module, using the action of G in Ox, it follows that f3x induces a G-module and ring isomorphism
(2.2.1)
R(G/Gx)
340
~R(O",).
350
BERTRAM KOSTANT.
Now we recall that 8(0a;) is the ring of functions on Oa; obtained by restricting 8 (the ring of polynomials on X) to Ox. Since {3a; is a morphism one obviously has
for any x E X and in fact it is clear that 8 (0 a;) is a G-submodule of R (0 x). Unlike R (Ox) whose G-module structure is completely determined by Proposition 8 because (2.2.1) is a G-module isomorphism, in the general case it seems (to us) to be quite difficult to describe how 8 ( 0 x) decomposes as a G-module. In many instances, however, 8(0a;) =R(Orc) (and hence, in such cases, one knows the G-module structure of 8 (0 JJ) ) • Indeed, in the general case since (j a; is Zariski closed in X one has (2.2.2) Thus (2.2.3)
Remark 7. In the example of Remark 2, § 1. 4, one depends upon the equality 8(0a;) =R(OgJ for a particular x in order to solve the Dirichlet problem in Rn. Indeed let x E Rn where (x, x) = IX> 0 and let f be a conThe problem tinuous function on the sphere 8 n- 1 = 0", n Rn of radius is to extend f as a harmonic function f' defined in the interior of Sn-l. To do this one expands f
Va.
using some limiting process (e. g., L 2 ), as an infinite sum of spherical harmonics fA. That is, here CA E C and h = gA I 8 n - 1 where
However since R (0 a;) = 8 (0 a;) it follows from (1. 5. 3) that there exists a unique harmonic polynomial hA E H on ,Cn such that hA lOa; = gAo One then puts f' = ~ cAh'A AED
where h'A is the restriction of hA to the interior of 8 n - 1 • Now it is not necessarily true, in general, that 8(0a;) =R(Oa;). For example let X be the m 2 dimensional space of all complex m X m matrices and G is the general linear group Gl (m, C) regarded as operating on X by left matrix multiplication. Then if x is the identity matrix Orc is isomorphic to
341
LIE GROUP REPRESENTATIONS.
351
Gl(m,C). But 8(0",) ¥=R(O",) since in particular if f(a) = (deta)-l for aE G then fE R(Ox) but f¢ 8(0",). The equality 8 (0",) = R(O",) in the example of Remark 7 when (x, x) > 0 may be established either using the fact that 0", is closed (see (2.2.3)) or by applying the Stone-Weierstrass theorem to both 8 (0",) and R (0 a;) restricted to Ox n Rn. 'l'hese methods also work more generally in case (x, x) ¥= o. However, they do not apply to 0", where x¥=O and (x,x) =0. Nevertheless it is still true in this case that R(Oa;) =8(Ox). The more powerful tool (and the one that will be required in § 5. 1) needed to establish the equality for this: case is given in the next proposition. For any x E X let Om be the Zariski closed subset of X defined by taking the complement of Ox in Oa;. If we put codim 0", = dim Oa; - dim Oa; then of course one has co dim O:c > 1. An affine variety Y is called normal in case the ring R (Y) is integrally closed in its quotient field. PROPOSITION 9. Let x EX. Assume (1) that Oa; is a normal variety and (2) that codim Oa; > 2. Then
Proof. If Y is any variety let Q(Y) denote the field of all rational functions on Y. In any f E Q(O:c) let 1 denote its image in Q(OlD) under the canonical isomorphism Q( 0 a;) ~ Q( OlD) defined by extension. Now let fE R(Oa;). 'J,'hen obviously IE Q(OID) is defined at every point of Ox. Thus if T is the set of points of Ox where f is not defined then T C Oa;' Since codim Om > 2 one also must have codim T > 2. But now for a normal affine variety Y one knows (see [3], Proposition 2, p. 166 and 10, p. 134. Also Corollary, p. 135), that if g E Q(Y) then either g E R (Y) or the set of points where g is not defined has co dimension 1. Since Oa: is assumed to be normal it follows that the first alternative must hold for 1. That is, 1 is everywhere defined on Oa;. But then 1, as a function on Ox, is the restriction of a polynomial on X to Ox' (See (2.2.2).) But then this is certainly true of f so thatfE8(Ox). Q.E.D. Remark 8. Proposition 9 is stronger than the criterion O:c = Oa: for insuring 8(0",) =R(O:c). In fact if O:c=Oa; (in which case we may take (2) to be trivially satisfied) then Oa; is empty and Oa; is non-singular. But
342
352
BERTRAM KOSTANT.
since non-singularity implies normality the conditions of Proposition 9 are satisfied in case Ox is closed. The proof that Ox is normal for the example of Remark 2 where (x, x) = 0, x oF 0, and d > 3 follows from a result of Seidenberg (see § 5. 1) .
3. The orbit structure for the adjoint representation. 1. Let g be a complex reductive Lie algebra of dimension n. Then g is a Lie algebra direct sum (3.1.1)
g=3+[g,gJ
where 3 is the center of g. The commutator [g, gJ is, as one knows, the maximal semi-simple ideal in g. A subalgebra a egis said to be reductive in g if the adjoint representation of a on g is completely reducible. Such a subalgebra is necessarily reductive (in itself). Let g"', for any x E g, denote the centralizer of x. An element x Egis called semi-simple if ad x is diagonalizable. One knows that gx is reductive in g for any semi-simple element x E g (see e. g. Theorem 'I in [l1J). An element x Egis called nilpotent in case (1) x E [g, gJ and (2) ad x is a nilpotent endomorphism.
Remark 9. If x E a C g where a is reductive in g then x is semi-simple (resp. nilpotent) with respect to a if and only if it is semi-simple (resp. nilpotent) with respect to g. The proof of these statements are immediate consequences of the representation theory of reductive Lie algebras. Now one knows that the most general element x E g may be uniquely written (3.1.2) where y is semi-simple, z is nilpotent and [y, zJ = 0. We will speak of y and z, respectively, as the semi-simple and nilpotent components of x. See [l1J, Theorem 6.
Remark 10. If x E a C g where a is a subalgebra reductive in g then by Hemark 9 the decomposition (3.1. 2) formed in g is the same as the decomposition (3.1. 2) formed in a. In particular given the decomposition (3.1. 2) one should observe that z is not only nilpotent in g but also in the" reductive in g" subalgebra gil. In particular then (3.1.3)
z E [gll, gYJ.
343
LIE GROUP REPRESENTATIONS.
353
Conversely ir y Egis semi-simple and z is nilpotent in gY and one puts x = y z then y and z are respectively the semi-simple and nilpotent components or x.
+
3.2. We wish to apply the considerations or §§ 1 and 2 to the case where X = g and G C Aut g is the adjoint group or g. Thus not only is G a connected algebraic reductive linear group but in ract G is then a semisimple Lie group whose Lie algebra is isomorphic to [g, g]. In this case we observe that the orbit 01/} defined by any x Egis just the set or elements or g that are conjugate to x. Ii a egis any sub algebra then under the adjoint representation a corresponds to a Lie subgroup A C G. Indeed A is the group generated by all exp ad x where x ranges over a. In this way gl/} clearly corresponds to the identity component or the algebraic subgroup Gill. We recall that an element x Egis semi-simple if and only ir x may be embedded in a Cartan subalgebra (C. S.) or g. Equivalently x Egis semisimple ir and only ir gill contains a C. S. or g. The rollowing lemma is known. We will prove it for completeness and also because, as noted in Remark 11 below, the proor may be used to give a more general result.
5. Assume x Egis semi-simple. Then (1) G'" is connected and (2) 03] is closed in g. LEMMA
Proof. We first show Grc is connected. Let bE Gx. Then by Theorem 2, p. 108, in [6], one knows that b may be uniquely written (3.2.1)
b =aexpady
where a EGIS diagonalizable and y Egis nilpotent and a(y) = y. Put Ct = exp t ad x. Then b = CtbCt-1 = (Ctavt-1) exp ad Ct (y). By the uniqueness of the decomposition (3.2.1) it rollows that a= Ctact-1 and Ct(Y) =y. Hence a E Ga; and y E gill. But then b is "connected" to a in Gill by means or the curve a exp s ad y, s E R. Thus we may assume that b is diagonalizable. But now by Theorem 10, p. 117 in [6], ir gb is the Lie subalgebra or all y such that b (y) = y then gb contains a C. S. g or g and i:f g is any C. S. in gb then b=expadz ror some zE g. But now x E gb and since ad x I gb is semi-simple there exists a C. S. g such that x E g C gb. But b = exp ad z ror some z E g. However since g C gill it rollows that b may be joined to the identity in GI/} by a curve; indeed one uses the curve exp t ad z. Hence Gill is connected.
344
354
BERTRAM KOSTANT.
To show that O:c is closed let f) be a Cartan sub algebra such that x E f). By the Iwasawa decomposition we may write G=KMHo where K and Mare connected Lie groups which are, respectively, compact and unipotent (an endomorphism u is called unipotent if u - 1 is nilpotent; a group is called unipotent if all its elements are unipotent) and Ho is an abelian Lie group corresponding to a sub algebra of f). Since x E f) it follows then that x is fixed under Ho. Thus O:c=KMx. We have proved (unpublished) that any orbit of a connected unipotent Lie group is closed. Rosenlicht [14J has generalized this to the case of a field of arbitrary characteristic. Thus we may use the reference [14J to establish that Mx is a closed subset of g. But since Ox is obtained by applying a compact group to a closed set it follows easily that O:c is closed.
Remark 11. Another proof that 0", is closed if x is semi-simple follows from Theorem 4, § 3. 8. In fact one sees there that 0", is closed if and only if x is semi-simple. This observation was also made in [1]. Note however the proof given above, that Ox is closed when x is semi-simple generalizes and shows that the orbit of any zero weight vector for any representation of G is closed. As a consequence of the connectivity of Gx for x semi-simple one has LEMMA 6. Assume x Egis semi-simple. Then gX is stable under Ga; and the restriction of G'" to g'" is the adjoint group of gX.
Proof. It is trivial that gaJ is stable under Gx. Furthermore as we have observed in the beginning of this section the identity component of GJ} corresponds to g'" under the adjoint representation of g and hence its restriction to g:c is the adjoint gmqp of g"'. But Gx is connected by Lemma 5. Q. E. D. 3.3. Now for the case at hand S is just the symmetric algebra S* (g) over the dual space to g. The well known description of the ring of invariants J given below is due to Chevalley. If l is the rank of g then J is generated by l algebraically independent homogeneous polynomials. That is, there exist homogeneous elements u. E J, i = 1,' . " l, such that if C[Y1 , ' • " YzJ denotes the polynomial ring, over C, in l indeterminates and
(3.3.1) is the homomorphism given by p(Y 1 ," "YZ)~P(Ul'" is an isomorphism. Moreover, if we write deg u. = m.
345
',u!) then (3.3.1)
+ 1 then the integers
355
LIE GROUP REPRESENTATIONS. I
rn,!, called the exponents of g, are those special integers such that II (1 ;=1
is the Poincare polynomial of g. Throughout we will assume that the
We will refer to the Ut, i
=
Ui
+t
2ml 1
+ )
are ordered so that
1, 2,' . " l, as the primitive invariants.
Rernark 12. One knows that the primitive invariants and even the l-dimensional space they span is not unique. However, in § 5.4 in connection with G-harmonic polynomials one normalizes the space they span in a natural way. See Remark 26, § 5.4. We now define a mapping u: g~CI
(3.3.2)
by putting U ( x) =
(u1 ( X ) "
• " Ul ( x)
).
It is obvious that U is a morphism. Now let (fj be the set of all obits 0 C g. Since U obviously maps any orbit into a point it is clear that U induces a map
Now if u egis any subset stable under the action of G it is obvious that u is a union of orbits. Let
and we will let '¥Ju be the restriction of '¥J to (fj u' Let ~ be the set of all, semi-simple elements in g. Obviously under G so that we may consider the case where u = £I.
~
is stable
Now it is easy to see that '¥J is not one-one, that is it does not separate all orbits. One observes, however, that not only does '¥J separate the orbits in .5 but also that '¥J9 is a surjection. The following proposition is no doubt known. We prove it for completeness. PROPOSITION
10.
Let.5 be the set of all semi-sirnple elements tn g.
Then the map
induced by
U
(see (3. 3. 2»
is a bijection.
Proposition 10 permits us to parameterize (fj9 by all complex l-tuples. In
346
356
BERTRAM KOSTANT.
order to prove Proposition 10 we need some further notation and Lemma 7 below. Let f) be a Cartan sub algebra of g regarded as fixed once and for all. Let W be the Weyl group of g regarded as operating in f). Let Ll C 8 1 (f) be the set of roots and let Ll+ C Ll be a system of positive roots fixed once and for all. An element x Egis called regular if g'" is a Cartan subalgebra. If x E f) one knows that x is regular if and only if <x,cf»=FO for all cf>E Ll. Now let ulj:
be the restriction of
U
f)~C!
to f).
LEMMA 7. The map ulj is proper. compact set is compact).
(That is, the inverse image of any
Proof. Let 7r: g ~ End V be a faithful completely reducible representation of g and let m = dim V. For any positive number 7c let ric be a positive m-1
number such that for any monic polynomial ym
+ .=0~ CiYi
=
P (Y) in the
indeterminate Y, where Ci EC, one has 1 Ci 1 < 7c, i = 0, 1,' . ., m -1, implies 1 A 1 < ric for any root A of p (Y). In fact, it suffices to take ric = m7c 1. Now let fi E J be the invariant polynomials defined so that
+
(3.3.3)
m-l
det(Y -7r(x»
=
ym
+ ~ fi(X)Yi i=O
for any x E g.
Now there exist unique polynomials
so that fi=Pi(U , ,'" ·,u!). Thus regarding C[Y l ," nomial ring on Cl it follows that
',Y l] as the poly-
(3.3.4) for any x E f). Now let ECC! be any compact set. We wish to show that ulj-l(E) IS compact. Let
7c = sup 1 Pi(~) I. ~EE
i=O,l, .. , ,m-l
It follows therefore from (3. 3. 4) that 1 fi (x) 1 < 7c for all x E Ulj-l (E). Hence if A is a root of (3.3.3) it follows that 1,\ 1
347
LIE GROUP REPRESENTATIONS.
N ow if a (-71") C 8 1 (f)
357
is the set of weights of 7r and we put
Ix I =
max I tfr (x) I '" t <1(...)
then since 7r is faithful it is clear that I x I defines a norm on the space f). But for any tfrE a(7r), A=tfr(X) is a root of (3.3.3). Hence I x I
det Om,Ui I f)
(3.3.5)
=
c II cp ¢E<1+
where c is a non-zero scalar. (3.3.6)
c
See e. g. [17J. But
II cp (x) ¥o 0 if and only if x is regular.
¢ (<1+
Hence the Jacobian (3.3.5) of ul) does not vanish identically on f) and consequently the Zariski closure of u'f)(f) equals Cl. But since uf} is a morphism ul) (f) contains a Zariski open subset of its closure; that is, a Zariski open subset of Cn. But any Zariski open set is dense in the usual topology. Thus in the usual sense
But now let ~ E Cl and let Yi E f), j = 1, 2,' . " be such that uf} (Yi) converges Since u'f) is proper the set Yi has a cluster point y. Obviously u'f) (y) = ~ and hence uf} is surjective. It follows therefore that 7]9 is surjective since every element of f) is sewi-simple. To show that 7]9 is injective we must show that if x, Y E {l and U (x) = U (y) then x and yare conjugate. Since every element in {l is conjugate to an element in 1) we may assume x, y E f). But now by Lemma 9.2 in [13J (the extension to the reductive case in trivial) if ui (x) = ui (y) for j = 1, 2,' . " l, then x and yare conjugate under the Weyl group and consequently are conjugate with respect to G. One explicitly uses here the well known theorem that, under the mapping 8*(g) ~8*(f) induced by injection f)~g, J maps onto the algebra of Weyl group invariants. Q. E. D. Since 7]9 is a bijection we may invert it. For any ~ E Cl we will let 0 9 (~) be the unique semi-simple orbit 0 such that 7]9(0) =~.
to~.
3.4.
We now wish to look at the orbits of maximal dimension.
348
By
358
BERTRAM KOSTANT.
adding and subtracting the dimension of the center of 9 it is obvious from that for any x E 9 (3.4.1)
dim 0", = dim 9 - dim g'" =n-dimg"'.
11. Let x E 9 be arbitrary. Then g'" contains an l dimensional commutative subalgebra. PROPOSITION
Proof. This is just Theorem 5. 'I in [13]. (Using the grassmannian of all l-planes in 9 the proof is an easy consequence of the fact that the set Q. E. D. of regular elements is dense in g.) As a corollary one has PROPOSITION
12. Let x E g.
Then
(3.4.2)
and the set of x for which equality holds in (3.4.2) is not empty. Proof. The equality in (3.4.2) clearly holds for all regular elements in g. (It also holds for a larger collection of elements. See (3.4.3) and Theorem 2, § 3. 5). The inequality (3.4.2) for all elements follows from (3.4.1) and Proposition 11. Q. E. D. By Proposition 12 n - l is the maximal dimension of any orbit. We now wish to consider the set of elements in 9 which define orbits of this dimension. Put (3.4.3)
r={xE 9 I dimg"'=l}.
It is obvious that, r is stable under G so that we may consider the subset @t C @. In fact @t is the set of orbits given by (3.4.4)
@t={OE @ I dimO=n-l}.
The structure of rand @t will be known as soon as we establish certain facts about principal nilpotent elements (Theorem 1). We will use a different definition of principal nilpotent than that given in [13J. Let .p be the set of all nilpotent elements in g. called principal nilpotent if and only if
An element x Egis
xE.pllr that is, if and only if (1) x is nilpotent and (2) dim g'" = l.
349
359
LIE GROUP REPRESENTATIONS.
Obviously -lJ is stable under G so that we may consider the set of orbits (i}'p' The following was established in [13]. 1. There are only a finite number of elements in (i}'p' Furthermore the set p n r of all principal nilpotent is not empty and is in fact a single orbit 0 E (i}'p' That is, for any e E p n r one has (1) THEOREM
dimOe=n-l and (2)
(3.4.5) Moreover, this orbit is dense in p. That is,
(3.4.6) for any principal nilpotent element e. If 0 E (i}'p is any orbit other than the orbit of principal nilpotent elements one has (3.4.7)
III
Proof. [13].
dim 0
< n-l.
Theorem 1 above is a restatement of Corollaries 5. 3 and 5. 5
Q.E.D.
Remark 13. It follows from (3.4.6) that the set of nilpotent elements -lJ egis an affine variety of dimension n-l. With regard to all the orbits in (i}'p it should perhaps be recalled that in [13] it was shown that excluding the orbit consisting of zero alone they are in a natural one-one correspondence with the conjugacy classes of all 3-dimensional simple sub algebras of g. If x Egis arbitrary and x = y z is the decomposition (3. 1. 2) then by the uniqueness of the decomposition, clearly,
+
(3.4.8) and hence (3.4.9) The subset r may be characterized as follows:
+
PROPOSITION 13. Let x E g be arbitrary. Write x = y z where y and z are, respectively, the semi-simple and nilpotent components of x. Then x E r if and only if z is a principal nilpotent of the reductive Lie algebra gY.
Proof· Now gY is a reductive Lie algebra of rank l, and z is a nilpotent
350
360
BERTRAM KOSTANT.
element of gY. (Remark 10, § 3.1.), Furthermore gY n gZ is exactly the centralizer of z in gY. Thus by definition z is a principal nilpotent element of gY if and only if dim gY n gZ = l. But then by (3.4.9) one has x E r if and only if z is a principal nilpotent element of gY. Q. E. D. 3.5. Now let x E g and let x = y for x. Then as was observed in [13]
f(X)
(3.5.1)
=
+z
be the decomposition (3.1. 2)
fey)
for any invariant polynomial f E J. That is, f(x) depends only upon the semi-simple component of x. Indeed (3.5.1) is an immediate consequence of the fact (see [13], p. 1031) that (3.5.2)
y
+ z and y + cz are
conjugate
for any non-zero complex number c. We can now completely describe tJ t • Proposition 10, § 3. 3, shows that the orbits of semi-simple elements are in a natural one-one correspondence with Cl. We now observe (and this is more important for us) that the set tJ t of all orbits of maximal dimension (n -l) is also in a natural one-one correspondence with Cl. THEOREM
2. The map
(given by the primitive invariant polynomials bijection.
Ul,' . " Ul;
see § 3.3) is a
Proof. Let ~ E Cl. Then by Proposition 10, § 3. 3, there exists a semisimple element y E g sllch that u(y) =~. Now let z be a principal nilpotent element in gY. Then by Proposition 13, and (3.5.1) if x = Y z then x E r and f(x) =f(y) for all fE J. Hence u(x) =~. Thus 7Jt(O",) =~ so that 7Jt is surjective. To show that 7Jt is injective we must show that if Xl, x 2Er and U(Xl) = U(Xi) then Xl and X2 are conjugate. Let x. = y. Zi, i = 1, 2, be the decomposition (3.1.2) for Xi. By (3.5.1) one has U(Yl) =U(Y2)' But then, by Proposition 10, § 3. 3, Yl and Y2 are conjugate. Hence we may assume that Yl = Y2 = y. But since Xl, x 2 E r it follows from Proposition 13 that Zl and Z2 are principal nilpotent elements of gY. But then by Theorem 1 applied to gY it follows from Lemma 6, § 3. 2, that there exists a E Gy such that aZ l = Z2' Thus aX l = X2. Q. E. D.
+
+
351
361
LIE GROUP REPRESENTATIONS.
Since 'YJt is a bijection we may invert it. For any ~E CI let be the unique orbit a of dimension n - l such that 'YJt (a) =~. 3. 6.
at(~) E (fjr;
Let {e>}, cp E A, be a set of root vectors belonging to A.
Let
(3.6.1) be the maximal Lie algebra of nilpotent elements defined by A+. Let m* be the corresponding nilpotent Lie algebra defined by the negative roots A_=-A+ so that g=m*+~+m
(3.6.2) is a linear direct sum.
Let A be the Cartan subgroup of G core ssp on ding to
be the set of simple positive roots. (Here k a¢ E C* be the non-zero scalar defined by aCe»~ =
(3.6.3)
=
~
and let
l - dim 3) . For any cp E A let
a¢e>.
Since G is the adjoint group of g one knows that the mapping
(3.6.4) given by a ~ (a(Y.l,. . ., a(Y..) is an isomorphism. LEMMA
8. Let y E g be semi-simple. Then the center of Gv is connected.
Proof. If x is conjugate to y then Gm is conjugate to Gv and hence it is enough to show the center of Gm is connected for some x E y. Since all Cartan subalgebras are conjugate we may choose x E~. Moreover, by applying an element of the Weyl group W to x, if necessary, we may assume that gm is of the form
a
(3.6.5) where III is some subset of the set of simple roots II and Ai C A is the set all roots cp of the form for integers n(y'.
352
362
BERTRAM KOSTANT.
Let Z be the center of G31 and let a E Z. It is then obvious that av = v for any v E g31. Hence a E A and atX = 1 for all 0: E 111. Since GID is connected (see Lemma 5, § 3. 2) the converse is clearly true. Hence Z = {a E A
Ia
tX
= 1 for all
0:
E Ill}.
Using the isomorphism (3.6.4) it is then obvious that Z is connected.
Q.E.D. Remark 14. The structure of GID is not as simple as Lemma 8, § 3. 6 seems to indicate. In particular even though the center of G31 is connected the subgroup of G:c corresponding to the maximal semi-simple ideal [g"', glD] of g:c may have a non-trivial discrete center. The structure of Gm is analogous to that of the general linear group Gl ( d, C) . Let M and B be respectively the unipotent and Borel subgroups of G corresponding to m and 0 = ~ m. The orbits of maximal dimension (n -l) are uniform in the following respect.
+
14. For any x E t the group Gm is an abelian, connected, algebraic subgroup of G of dimension l - dim 3. (Recall that 3 is the center of g). PROPOSITION
Proof. By Proposition 11, § 3. 4, it is immediate that g:c is a commutative Lie algebra of dimension l. Therefore to prove the proposition it is enough to show that Gm is connected. Let e E 9 be given by
Then by [13], Theorem 5.3, e is a principal nilpotent element of g. We :hst show that Ge is connected. Let hE Ge. But [13], Corollary 5.6, m must be stable under h (this corollary asserts that a principal nilpotent element lies in one and only one nilpotent Lie algebra of the form (3.6.1)). But now since m is stable under h it follows that hE B. This may be proved in the following way. According to the Bruhat decomposition of G (see [7]) we may write h = bS(rT)g where bE B, gEM and s (rT) is in the normalizer of A inducing the element rT E W on~. To prove hE B it suffices to show that rT is the identity of W. But this is obvious since h, b, g and hence s «(I) leaves m stable (only the
353
363
LIE GROUP REPRESENTATIONS.
identity element of W leaves .6.+ stable). aE A and dE M. Now since dE M
h(e) where wE
Em, mJ.
But since h(e)
Write h = da where
da(e) =a(e) +w
=
e and
=
a(e)
Thus a E B.
= a
~ aaea £II
(see (3.6.3» it follows that w=O and a(e) =e. The latter implies that aa = 1 for all ex E II and hence by (3.6.4) a is the identity of G. Thus h = d E M. But then h may be uniquely written h = exp ad v where vErn. But then by well known properties of nilpotent Lie algebras one has v E ge and hence h lies in the identity component of Ge or since h was arbitrary Go is connected. Since all principal nilpotents of g are conjugate it follows that Gz is connected for any principal nilpotent element z E g. Now let x be an arbitrary element of r and let x = y z be the decomposition (3.1. 2) for x. We recall that, by Proposition 13, § 3. 4, z is a principal nilpotent element of the reductive Lie algebra gY. Let F be the adjoint group of gY so that applying what just proved (in the case of g) to gY it follows that Fz is connected. Now by Lemma 6, § 3. 2, the restriction of Gy to gY induces an epimorphism
+
(3.6.6) with kernel Z equal to center of Gy. But now the full inverse image of Fz under the map (3. 6. 6) is just Gy n Gz = GIIJ (see (3.4. 8) ) . But since Fz is connected and Z is connected by Lemma 8, § 3. 6, it follows that GIIJ is connected. Q. E. D.
Remark 15. Note that if x E r the abelian group GID ranges all the way from a reductive group, (in the case where x is a regular element) to a unipotent group (in the case where x is a principal nilpotent element). In the general case GID, for any x E r, is the direct product of a abelian reductive group and an abelian unipotent group. It seems suggestive from Lemma 5, § 3. 2, and Proposition 14, that possibly GIlJ is always connected for any x E g. This, however, is false. If g is the Lie algebra of the exceptional simple Lie group G2 and '" = 3ex 2{3 is the highest root where II = {ex, {3}, then one can show that GIlJ is not connected in case x = ea e",. In fact one sees easily that GIlJ contains the non-
+
+
354
BERTRAM KOSTANT.
364
trivial diagonalizable element a E A where aU = 1 and afl = -1 whereas on the other hand the identity component of G'" is unipotent. One proves the latter statement by using the first line in Table 21, p. 186 in [5] (no Xo term) together with the argument in the proof of Theorem 3, § 3. 7. 3.7. It was pointed out to us by Dixmier that the following proposition is a special case of a more general result of Kirillov (all orbits of any Lie group for the representation contragredient to the adjoint representation are even dimensional). The simple proof given here is due to Kirillov and is easily modified to give the more general result. PROPOSITION 15.
For any x E g one has dim 0", is even.
Proof. Let B be a G-invariant non-singular symmetric bilinear form on g. Now for any x E g let BIl! be the alternating bilinear form on g given by B",(y,z) =B(x, [y,z]). From the invariance of B it is clear that Bm(Y, z) = 0 for all z E g if and only if y E gm. It follows therefore that Bm defines a non-singular alternating bilinear form on g/gm. But since such a bilinear form can only be carried by an even dimensional space it follows that dim g - dim g'" is even. The proposition then follows from (3.4.1). Q. E. D. 3.8. We now consider the cone P C g, defined as in § 1. 3 by J+. That is, P is the set of common zeros for all the polynomials in J+. The following proposition was essentially proved in [13] (it was proved for the case when g is semi-simple). PROPOSITION 16.
The cone P is identical with the set .p C g of all
nilpotent elements in g. Proof. If x E P then clearly ad x is nilpotent. Indeed if f; E 8 is defined by f;(y) =tr(ady); for any yE g and positive integer j one has f;E J+ and hence f; (x) = 0 for all such j implies ad x is nilpotent. On the other hand if x E P then also x E [g, gJ. In fact let x = Xl X 2 be the decomposition of x according to (3.1. 1) where Xl E cr and X 2 E [g, g]. To see that Xl is zero observe that every linear functional (element of 8 1 ) on g which vanishes on [g, g] lies in Jl. But then one must have f(x) = 0 for all such linear functionals. Hence x E [g, gJ and thus P C lJ (see the definition of nilpotent elements). But .pCP by (3.5.2) since y=O when x is nilpotent and f(O) = 0 for any f E J+. Q. E. D. Now Theorem 1 together with Propostiion 16 above implies that (1)
+
355
365
LTill GROUP REPRESENTATIONS.
P has a dense orbit (the set of principal nilpotent elements) (2) P is of dimension n - l and (3) P is a finite union of orbits. On the other hand from the particular structure of J the cone P can be given by considering only the primitive invariants Ui E J. That is, P={xE g I Ui(X) =O,i=1,2,. . ·,l}. We now generalize Theorem 1, and thereby encompass every point of g, by proving a similar theorem after substituting any point in C! for the l-tuple of zeros above. For any ~EC!, ~=(~1'·· .,~!), let pa), as in Proposition 6, §1.6, be the affine variety in g given by
That is, with respect to the map u (see (3.3.2) one has (3.8.1) Thus (3.8.2) is a disjoint union. Now recall (see end of § 3. 3 and § 3. 5) that oa(~) and Ota) are, respectively, the orbit of semi-simple elements and orbit of maximal dimension corresponding to any ~ E C! under 1)a and 1)t. Now obviously P(~) is a union of orbits. In particular oa(~) and Ot(~) are contained in P(~). Furthermore every orbit 0 lies in some P(~). Theorem 1 generalizes in the following way. THEOREM 3. Let ~ E C! be arbitrary. Then pa) is the set of all x E g whose semi-simple component lies in oa (~). P(~)
(3.8.3)
=
Ot(~)
U ... U oa(~)
is a union of a finite number of orbits. Moreover
(3.8.4) and Ot(~) in fact
tS
the unique orbit of maximal dimension (n-l) in dim 0 < (n-l)-2
(3.8.5) for any other orbit in
P(~).
Next (2)
(3.8.6)
356
P(~)
and
366 and
BERTRAM KOSTAN'r.
os(~)
is the orbit of minimal dimension in pa)
(3.8.7)
=
P(~).
Finally
Ot(~)
so that the Zariski closed set P (~) is irreducible and dimP(~)
3.8.8)
=n-l.
Proof. The statement that P(~) is the set of all x E g whose semi-simple component lies in Os(~) is an immediate consequence of (3.5.1) and Proposition 10, § 3. 3. This of course implies (3.8.6). Furthermore using (3.5.2) it follows from (3.8.5) that for any orbit 0 C P(~) one has
(3.8.9) so that
Os(~)
is the orbit of minimal dimension in pa).
Now let y E Os (~) so that gY is a reductive Lie algebra. Let.py denote the set of nilpotent elements of gY. Applying Theorem 1 to g1l there exist k elements Zi E .p1l, i = 1,' . " k, such that under the adjoint group F of g1l every nilpotent element in g1l is conjugate to one and only one of the Zi' Furthermore Theorem 1 asserts that if N is the set of all principal nilpotent elements of gY then N is an orbit under F and N = .p1l. Now put X,,=y+Zi, i=l,' . ',k, so that by Remark 10, §3.1, Y and Zi are, respectively, the semi-simple and nilpotent components of Xi. We now assert that every element in P (~) is conjugate to one and only one of the x". We first show that the Xi lie in different conjugate classes. Assume ax" = Xj for a E G. Then by the uniqueness of the decomposition (3.1. 2) one has ay = y so that a E G1I and az" = Zj. But, by Lemma 6, § 3. 2, Zi is then conjugate to IZj under F and hence i = j. Now let v E P(~) be arbitrary and let wand e be, respectively, its semi-simple and nilpotent compotents. We show that v is conjugate to one of the Xi. By the first statement of the theorem (already proved above) there exists a E G such that aw = y. Hence we may assume w = y. But then e E .p1l and hence by Lemma 6 there exists a E Gy such that ae = Zi for some i so that av = Xi. Thus there are only a finite (k) number of orbits in P (~) . The statement (3.8.4) and the fact that Ot(~) is the unique orbit of maximal dimension in P(~) is just a restatement of Theorem 2. The inequality (3.8.5) then follows by Proposition 15. Finally since N (see above) is dense in .py, by Theorem 1, each Xi is in the closure of Ot(~) so that, clearly, one has (3.8.7) and hence also (3.8.8). Q.E.D.
357
367
LIE GROUP REPRESENTATIONS.
Remark 16. We can be very explicit about the number of orbits in pa) for any ~E 0. Let yE O~(~). By the argument above and Remark 13, § 3.4, if k is the number of orbits in P (~) then k -1 is the number of conjugacy classes of three dimensional simple Lie algebras in g1l. But such classes have been listed by Dynkin (see [5]) for every simple Lie algebra and hence one knows k as soon as one knows the maximal semi-simple ideal [g1l, g1l] in g1l. We note also that (3. 8. 9) together with Lemma 5, § 3.2, implies that the semi-simple orbits are the only closed orbitA 1. Let x E g be arbitrary. Then Oa; is a union of only a finite number of orbits. Moreover if O(]} is the (Zariski closed) complement of Oa; in Oa; then, where co dim Oa; is, as in § 2. 2, defined with respect to Oa;, one has OOROLLARY
codimOa;> 2.
Proof. If u(x) =~ then obviously Oa;CP(~) and since P(~) is composed of only a finite number of orbits the same is true for 0",. But if 0 C 0", then certainly dim Oa;-dim 0 > 2 by Proposition 15 and the fact that dim 0
a
OOROLLARY 2. Let 0 be any orbit and let T be the set (Zariski closed) of all non-simple points of the affine variety 0. Then if codim T is defined with respect to a one has codimT> 2.
a
Proof. One knows that the set of all simple points of is Zariski open in and non-empty. It therefore meets O. Since G is transitive on 0 it follows that all the points of 0 are simple. Thus T C Oa; where 0 = Oa; and hence the result follows from Oorollary 1.
a
Remark 17. It is suggestive from Oorollary 2 that possibly a is a normal variety for any orbit O. This will be proved later (see § 5. 1) for all orbits of maximal dimension. For such orbits it will also be seen that T is exactly the complement of 0 in 0.
4.
The transversal I-plane U. 1. Now we recall that S, the ring of polynomials on g and S*, the symmetric algebra over g (the ring of differen-
358
368
BERTRAM KOSTANT.
tial operators with constant coefficients on g) are G-modules and are paired by (1. 1. 1) . By taking the differential (via the adjoint representation) they become g-modules and by (1.1. 4) one has (4.1. 1)
<x' 8, f> + <8, x . f> =
0
for all x E g,8 E 8* and f E 8. Furthermore any x E 9 operates as a derivation of degree 0 of 8 and 8* and hence, by (4.1.1), its action is completely determined by its restriction to 8 1 , But the latter is given by x'8y =8[IC,y]
(4.1.2)
for any y E g. Note that if 8 E 8* is of the form 8 = x . 81 where x E 9 and 81 E 8* then by (4.1.1) ( 4. 1. 3)
<8, f>
=
0 for all f E J.
This criterion for an element 8 E 8* to be orthogonal to J is especially convenient to use when x equalS a certain element Xo E f), now to be defined. Recall that IT C ~+ is the set of simple positive roots. We now put Xo equal to the unique element in f) n [g, g] such that (see [13], § 5. 2)
<xo, ex> =
1 for all ex E IT.
If ep E ~ is arbitrary and the order
(4.1. 4)
o(ep)
0
(ep) of ep is the integer defined by
=~na(ep) a€II
where (4.1. 5)
~
na ( ep ) ex
a(ll
then clearly (4.1. 6)
<xo, ep> =
0 ( ep )
and hence (4.1. 7)
[xo, e>] =
0 ( cp ) e>.
As usual let Z denote the set of all integers. For every integer j E Z let
8*(J) = <8E 8*
I Xo • 8 =
j8>.
It is obvious that S*U) is a graded subspace of 8* and since
derivation of 8 it follows immediately from (4.1. 7) that 8*=~8*(j) ja
359
Xo
operates as a
369
LIE GROUP REPRESENTATIONS.
is a direct sum and (4.1. 8)
Similarly let gU) be the eigenspace of ad Xo for the eigenvalue j so that g is a direct sum of the gU). Since ad Xo is a derivation of g clearly [g(i), gU)]
(4.1. 9)
C
g(i+j).
The decomposition (3.6.2) is related to
Xo
in the following way.
LEMMA 9. The nilpotent Lie algebras m and m* may be expressed in terms of the eigenspaces gU> of ad Xo as follows,'
Moreover h=g(O)=g"'o (i.e.
Xo
is regular).
Proof. Obvious from (4. 1. 7) and the fact roots cp and negative for negative roots cp. Since S*U) is in the range of the action of by (4.1. 3),
(4.1.10)
<0, f>
=
0 if f E J,
0
(cp) is positive for positive Q. E. D.
Xo
whenever j =1= 0 one has,
aE S*U) where j =1= o.
In the obvious way the symmetric algebra S*(u) over any subspace u C g may be regarded as a subalgebra of S*. Let b be the maximal solvable Lie subalgebra of g given by the direct sum ( 4.1.11)
(resp. put b* = m*
+ f).
One knows that if g = gl (d, C) then f (x) depends only on the diagonal entries of x in case x is a triangular matrix and f E J. More generally one has
+
PROPOSITION 17. Let x E b* so that x = y v where y E f) and v E m*. Then for any fE J one has f(x) =f(y). In particular
u(x) =u(y) where u is the map (3.3.2). Proof· Since J is graded we may assume f E Jk. Then, by (1.1. 3), k!f(x) = (om)\ f> = (Oy ov)\ f> = «Oy)" 0, f> = k!f(y) <0, f> where,
<
< +
+
360
+
370
BERTRAM KOSTANT.
But now by Lemma 9, § 4.1 and
by binomial expansion, 8 E m* . S* (b*). (4.1. 8) it follows that
Q.E.D.
and hence <8,f>=0 by (4.1.10). Thus f(x) =f(y).
4.2. The following simple characterization of the principal nilpotent elements in m was given in [13]. THEOREM
4.
Let e E m. Write e=
~
c.pe.p
¢EA+
then e is principal nilpotent Proof.
if and
only
if
c'" =F 0 for every simple root a E II.
Q.E.D.
This is just Theorem 5.3 in [13].
N ow for every simple a E II let c''" be an arbitrary non-zero complex number (normalized in § 4. 4). We isolate a particular principal nilpotent element (by Theorem 5, after interchanging the roles of .6.+ and .6._) e_ by putting (4.2.1)
e_=
~ c'ae-a. "'Ell
The following lemma gives a very simple method for constructing elements in r (in fact by Lemma 11 and Proposition 10, § 3. 3, at least a representative for every orbit of maximal dimension is constructed in this way). Recall that b is the maximal solvable Lie sub algebra LEMMA
~
+ m.
10. One has the relation
where, we recall, r is the set of all x E g such that dim g'" = l. Proof.
For any j E Z put a<J) =
ge_ n g(j).
Then since e_ E g(-l) it is clear from (4.1. 9) that (4.2.2) and hence (4.2.3) is a direct sum.
361
371
LIE GROUP REPRESENTATIONS.
Since e_ E r one therefore has ~
(4.2.4)
dim aU) = 1.
} EZ
Now filter g by putting
(4.2.5) Thus (1) g}~ gj+l for all j (2) g = gj for j sufficiently small and (3) I1J = 0 for j sufficiently big. Now let vEfJ and x=e_+v. But now by (4.1.9) one has
(4.2.6) for all j. On the other hand since the g} induce a filtration on gil! one has (4.2.7)
~
dim gl'/ g}+l il! = dim g"'.
}EZ
But now if
(4.2.8) is the obvious injective map induced by (4.2.5) it follows immediately from ( 4. 2. 2) and (4. 2. 6) that the image of (4. 2. 8) lies in aU). Thus (4.2.9) for any j. Comparing (4.2.4) and (4.2.7) it follows that dim g'" < 1. But, by (3.4.2), dim gil! > 1. Hence dim gil! = 1 (and hence also the equality holds in (4.2.9) for any j). Q.E:D. LEMMA 11. Let y E ~ be arbitrary. also u(x) =u(y).
Put x = e_
+ y.
Then x E rand
Proof. Since y E fJ one has x E r by Lemma 10. On the other hand since e_ E m* it follows that u(x) = u(y) by Proposition 17, § 4.1. Q. E. D.
4.3. In § 1. 5 we defined the notion of a quasi-regular element x E X for the general case of a linear group G operating on a vector space X. The notion is important for us because of Proposition 5, § 1. 5. The question as to which elements are quasi-regular, for the case at hand, is settled by PROPOSITION 18. The set r (all elements x E g such that dim gil! = 1) is identical with the set of all quasi-regular elements in g. (See § 1. 5.)
362
372
BERTRAM KOSTANT.
Proof. In the case at hand P is the set 1J of all nilpotent elements. If x Egis quasi-regular then by definition P Il! = 1J. Hence in particular if e E 1J is a principal nilpotent element there exists a sequence Xj, j = 1, 2,· .. , such that Xj E OCjll! where Cj E C* and such that Xj converges to e. Now let k = dim gil!. We wish to show that k = l. By Proposition 12, § 3. 4, it suffices to show that k < l. Obviously dim gCjll! = k so that dim gll!j = k for all j. Now consider the Grassmanian (which one recalls is compact) of of all subspaces of dimension k. Let u be a cluster point of the g"'l. It then follows easily (see argument in [13J, p. 1003) that u C ge. But then k < 1 since dim ge = 1 and dim u = k. Thus k = 1 and consequently x E t. Now conversely assume that x E t. We wish to show that x is quasiregular. Since P", is closed and stable under G to prove P", = 1J it suffices by (3.4.6) to show that PIl! contains a principal nilpotent element. Now let ~ = U (x). Then by Proposition 10, § 3. 3, there exists y E f) such that u(y) =~. Put Xl = e_ y. Then by Lemma 11, § 4. 2, and Theorem 2 it follows that Xl E Om. The same argument shows that for any C E C* one has that Xc E Ocm where Xc = e_ cy. (One uses the £act that U (x) = U (y) implies u (cx) = u (cy) ; an immediate consequence of the homogeneity of the Ui.) But now, obviously, Xc ~ e_ as c ~ 0. Hence e_ E P Il!. But since e_ is principal nilpotent this proves X is quasi-regular. Q. E. D.
+
+
4.4. Now for every simple root 0: E IT let c'" E C* be any arbitrary nonzero complex number. Let e+ be the principal nilpotent element (see Theorem 5, § 4. 2) given by e+ =
( 4. 4. 1)
~ c",e",. "'Ell
Since e+ is principal nilpotent one has dim ge+ = l. The following description of ge+ proved in [13J will playa fundamental role in this paper. THEOREM
5.
There exists a basis Zi, i
=
1, 2,· .. , l, of ge+ such that
(see § 4.1) ( 4. 4. 2)
Zi E gem,)
where, we recall, mi is that integer given by degui=mi+ 1 and
E J is the i-th primitive invariant polynomial (see § 3. 3). In particular then
Ui
(4.4.3)
363
LIE GROUP REPRESENTATIONS.
373
Proof. When g is simple the first statement here is just the 2nd and 3rd from the last statements of Theorem 6. 7 in [13] (our Zi here is the Ui of that theorem) together with Corollary 8.7 (which shows that ki = mi) of [13]. If g is semi-simple the first statement is still true since e+ may be written e+= L e+.i
..
where the e+.i are principal nilpotent elements of the various simple components of g and each is of the form (4.4. 1) for the corresponding simple component. One then uses the fact that the exponents (the mi) of g are composed of the exponents of the various simple components of g. In the general case the first statement of the theorem also follows since if a is the center of g then m. = 0 if and only if i < dim 3. But clearly 3 = ge+ n g(O). Since the mi are non-negative integers the relation (4.4.3) follows from Lemma 9, § 4. 1. Q. E. D. Remark 18. Subject only to the conditions of Theorem 5, § 4. 4, it is obvious that the basis of primitive polynomial invariants Ui does not uniquely determine the basis Zi of ge+. However with the further relations uncovered in § 4. 6 we wish to note that the Ui do uniquely determine a basis Zi of ge+.
Now one knows the elements Xa E ~, ex E II, given by Xa = [ea, e-a], form a basis of ~ n [g, g]. Hence we may write
L raXa
Xo =
a eII
where Xo is defined as in § 4. 1. Now define e'_ =
~ aeII
Talca e-a
where the Ca define the principal nilpotent element e+ (see (4.4. 1) ). Then as observed in [13], § 5.2, the elements e'_, Xo and e+ form a basis of a principal three dimensional simple sub algebra 0 0 of g. Furthermore this basis satisfies the commutation relations of an S-triple (see [13], p. 996). It is obvious then that e'_ is conjugate to e+ and hence e'_ is a principal nilpotent element of g. By Theorem 4, §4.2, one must therefore have that Ta =1= 0 for any ex E II and hence we may normalize the c' a of (4. 2. 1) by putting
c'a=raICa so that e_ = e'_. Now if V is any finite dimensional irreducible module for the three
364
374
BERTRAM KOSTANT.
dimensional simple Lie algebra aD with respect to a representation 7r then one knows that V is a direct sum of Ker7r(e+) and Im7r(e_). Since any finite dimensional aD-module is completely reducible the same must be true without the assumption of irreducibility. It follows therefore from (4. 2. 2) and ( 4. 4. 3) that by restriction to Ii we have proved 12. Let Ii be the maximal solvable subalgebra of 9 given by ( 4. 1. 11) . Then LEMMA
(4.4.4) is a direct sum.
+
4. 5. A subset u C 9 will be called a plane if it is of the form u = w a where wE 9 is an element and a is a subspace. (It is called a le-plane if dim a = le.) It is ObVIOUS that ale-plane u is a le-dimensional affine subvariety of g. Furthermore S (u), the restriction of S to u, is the affine algebra of u and in fact if, as above, u = w a then writing an arbitrary x E u in the form
+
k
x=w+~ri(x)Yi where Yi is a basis of a it is clear that the ri are in S(u) i=l.
and define a coordinate system on u.
Moreover one obviously has
(4.5.1) Note also that if
f E Sand g =
flu then
(4.5.2) Now let
be the restriction of u to u.
The plane u will be called transversal if
Since Uu is a morphism it is clear that u is transversal if and only if the functions viE S(u), where Vi=Ui I u, are algebraically independent. But this is the case if and only if the Zariski open subset U o of u given by Uo =
{x E u I (dvi) "" i = 1, 2,' . " l, are linearly independent}
is non-empty (see proof of Proposition 6, §1.6). But by (4.5.2) ti~ is the set of all points x in tI where the leXl matrix (OYjUi) (xL i=l," ',l, j = 1,' . " le, is of rank l. Thus (to make it independent of the basis Yi of a)
365
LIE GROUP REPRESENTATIONS.
375
ir d (u) C S (u) is the space or runctions on u spanned by the determinants or all l X l minors or OYJUi I u (put equal to zero ir k < l) then (4.5.3)
u is transversal ir and only i£ dim d (u) > 1
and (4.5.4)
Uo =
{x E u
I g (x) =1= 0 ror some g E d (u) }.
Now ir u is an l-plane then obviously dim 11 is either 1 or 0 (according as to whether u is transversal or not). In case 11 is a Cartan subalgebra, say I), then we have already observed that 11. is transversal. In £act (4.5.5)
d(l))
=
(
II
cf»
1; E d+
and 1)0 = U o is the set or all x E I) which are regular in g. (See (3. 3. 5).) Although I) is transversal it is not suitable ror our needs, mainly because 1)0 =1= I). On the other hand ir we put 0 equal to the l-plane given by (4.5.6) then not only is 0 transversal but it will also be shown (1) that d (0) = C so that 00 = o. Moreover it will be seen that every element or 0 lies on an orbit or maximal dimension and every such orbit meets 0 in one and only one point. Remark 19. In a sense 0 is to r as I) is to £l, the set or all semi-simple elements in g. However 0 has the advantage in that there are no "W eyl group ambiguities" with regard to conjugation. Furthermore the restriction or J to I) induces only a monomorphism or J into S (I)) whereas (by Theorem 8, § 4. 7) the restriction or J to 0 induces an isomorphism or J onto S (0) . Remark 19' (added in proor). Ii a is any linear complement or ad e_ (m) in '6 which is stable under ad Xo it is clear that a may be substituted ror gO+ in Theorem 5. We now wish to observe that ir a is substituted ror g6+ in the definition or 0 above (see (4.5.6)) then all the results to be proved hencerorth about 0 will still hold true. That is, the only properties or gO+ needed are Theorem 5 and (4.4.4). In this generality the results contain Theorem 0.10 or the Introduction. In particular they apply to the special case or the plane or companion matrices. (See Remark rollowing Theorem 0.10.) In order to show first that 0 is transversal the rollowing obvious ract will be userul. Assume that u and tv are planes and that 1f!: u -7 tv is a morphism defined so that
(4.5.7)
366
376
BERTRAM KOSTANT.
Then clearly (since
Uu (u)
C Urn (ro) )
u is transversal implies ro is transversal.
(4.5.8)
The following proposition asserts among other things that every element of the l-plane e_ f) is conjugate to an element in the I-plane b.
+
+
PROPOSITION 19. For each element x E e_ f) there exists a unique element a:c EM such that a",(x) E b. Furthermore the map
(4.5.9) given by x ~ a", is a morphism. Proof. Let ei, i
1, 2,· . ., m, he a basis of m so that e. E gU) for some
=
j> O. (See Lemma 9, § 4.1.) In fact let r(i) be that positive number such that e.E g(r(i». We may then order the basis ei so that r(i)
and the
Wi
form a basis of ad e_ (m) .
Obviously one has (4.5.10) for the map (a, x)
~
ax.
+
Now for any v E e_ 0 let c.( v), i = 1, 2,· . " m, be the scalar defined so that if v is uniquely written v = e_ V 1 V 2 where V1 E ge+ and V 2 E ad e_(m), according to the decomposition (4.4.4), then
+ + m
V2 =
~
C.(V)W•.
• =1
We now make the following inductive assumption about a positive integer le. There exists lc functions gi E S (e_ f)), i = 1, 2,' . ., lc, such that if zE m where
+
Then one has for any x E e_
+ f),
CJ ( exp
ad z ( x) )
=
0 for all j < lc
367
377
LIE GROUP REPRESENTATIONS.
if and only if b.. = g.. (x)
(4.5.11) for all i < le.
+
We now show that the assumption holds for le 1. Indeed if we compute Ck+l = Ck+l (exp ad z (x» where z satisfies (4. 5. 11) it is straightforward, using ( 4. 1. 9), to see that
(4.5.12)
Ck+1 = bk+1
+ fO(gl(X),'
. " gk(X» z ~f.. (gl(X),· .. ,gk(x»r.. (x)
+
;'=1
+
where f;" i = 0, 1,' . " l, are polynomials in le variables and ri ES (e_ f) is the same as in (4.5.1) with u = e_ f). Now consider the equation C'm = O. Since (4. 5. 12) is linear in bk +1 we can obviously uniquely solve for h+l obtaining bk+1 = gk+l(X) where gk+l E S (e_ f). Thus the induction assumptions hold for le 1. On the other hand (4.5.12) is also valid for le = 0 provided fo = 0 and f.. are constants for i = 1, 2,' . " l. Thus, similarly, the induction assumption holds for le= 1. We have thus proved inductively that given x E e_ f) there exists a
+
+
+
+
unique element
m Z=~gi(x)ei
in m such that e;,(expadz(x» =0 for i=l,
;'=1
+
2,' . " m. That is, such that exp ad z (x) E b = e_ ge+ and that furthermore gi E S ( e_ f) . But if afC = exp ad z this proves the lemma since one knows the map m ~ M given by z ~ exp ad z is an algebraic isomorphism.
+
Q.E.D. PROPOSITION 20.
For every x E e_
+ f) let afll E M be defined as in Proposi-
tion 19. If now is the map given by x ~ afC (x) then p is a morphism.
Proof. Obviously the map (4.5.10) is rational and everywhere defined. But by Proposition 19 so is the map e_+f)~MX
(4.5.13)
(e_+f)
given by x ~ (afC' x). One obtains the proposition by composing (4. 5. 10) with (4.5.13). We can now prove LEMMA
13.
The l-plane b = e_
368
+ ge+ is transversal.
378
BERTRAM KOSTANT.
Proof. Let p be as in Proposition 20. Since p (x) E OII! for any x E e_ it is obvious that
+ f)
+
where u = e_ f). Thus, by (4. 5. 8), to prove tJ is transversal it suffices to show that u is transversal. But if
r:
f)~u
is the map by r(y) =e_+y then of course r is a morphism. hand by Lemma 11, § 4. 2,
On the other
Thus by (4.5.8) to prove u is transversal it suffices to show that f) is transversal. But f) is transversal by (4. 5. 5) . Q. E. D. 4.6. Now let Zi, i = 1, 2,' . ',1, be the basis of ge+ given by Theorem 5, § 4. 4. We recall that Zi E g(m,). On the other hand one has deg Ui = mi 1. Hence if we put
+
(4.6.1) then gi E 8 ' . That is, gi is just a linear functional on g. LEMMA
14. Let 1 < i, j < 1.
Then
whenever mi ¥= mj' Proof. Since gi is a linear function on g one has, by (1. 1. 2) and (1. 1. 3),
(4.6.2) for any Z E g. But since e_ E g(-I) one has (lJ eJm, E 8*(-m,) by (4.1. 8) and, if zE g(k), then also (lJ eJ""lJz E 8*(k-m,). But if k¥=m;, that is, if k-mi¥=O then gi(Z) = 0 by (4.1.10) and (4.6.2). In particular gi(Zj) = 0 if mj ¥= mi.
Q.E.D. The following lemma is crucial. algebra given by (4. 1. 11).
Recall that 0 is the maximal solvable
LEMMA 15. Let 1 < i, j < 1. Then 0. In fact to a constant on e_
+
if
(4.6.3)
369
mi <
mj
the function
IJZjUi
reduces
379
LIE GROUP REPRESENTATIONS.
Furthermore, if mi < mj then the constant is zero. That is, in this case (4.6.4)
OZjUi\ e_+'6=O.
Proof. We first observe that if x E e_ integer k one has
+ '6
then for any non-negative
(4.6.5)
+
Indeed this is clear from (4.1. 8) and Lemma 9, § 4. 1, upon writing x = e_ y where y E b and using binomial expansion. In fact from the binomial expansion it is obvious that (OeJk is the component of (Om)k in 8* (-I(). That is
(Om)k- (oeJk E
(4.6.6)
~
8*(p).
p>-k
Now if (4.1.10)
f EJ
(4.6.7)
it follows, since
Zj
E g(ln j ), that by (1. 1. 2), (4. 1. 8) and
for all 0 E 8*(p) where P =1= - mj; in particular for all p
>-
mj.
But now if k = mi in (4. 6.6) then the sum there is over all p where p>-mi. Hence if m;<mj, so that -m;>-mj, the sum in (4.6.6) is over all p, where p > - mj. Thus, by (4.6.7), (4.6.8)
+
for all x E e_ b whenever mi < mj. We now assert that this implies (4.6.9)
+
for any x E e_ '6 whenever mi < mj. Indeed replace f by Ui and divide by mil in (4. 6. 8) . Recalling that deg OZj Ui = mi the left side of (4. 6. 8) becomes the left side of (4. 6. 9) by (1. 1. 3) . On the other hand by (1. 1. 2) the right side of (4.6.8) becomes the right side of (4.6.9) by (4.6.2). (Recall that 8* is commutative.) This proves (4.6.3). But now if mi < mj then the right side of (4.6.9) vanishes by Lemma 14. Hence one obtains (4.6.4). Q. E. D. We can now show that the Jacobian matrix of functions OZjUi \ b of the map u b takes triangular form and reduces to non-zero constants along the diagonal.
370
380
BERTRAM KOSTANT.
6. There exists a unique basis zj, j 1, 2,' . " l,
THEOREM
that for i
=
(4.6.10)
=
1, 2,' . " l, of oe+ such
gi(zj)=8ij.
Furthermore the basis satisfies the condition of Theorem 5. That is for all j. Furthermore
Zj
E g(ffl l )
(4.6.11) so that not only is b transversal but in fact
(4.6.12) and hence (see § 4. 5)
d(b) =C.
(4.6.13)
Proof. An integer le will be called an exponent if le = mi for some i. Let E be the set of exponents and for any le E E let P k C {1, 2,' . " l} be the set of all i such that mi = le. Now, for any le E E put
It then follows from Lemma 15 that det OzJUi is a constant on e_ in fact
+ V and
+
But since b C e_ V and since b is transversal (Lemma 13) this constant can not be zero. Thus bk¥=O for any leEE. That is, the matrix g.. (Zj), i,jEP k , is non-singular and this holds for any le E E. It follows immediately then from Lemma 14 that a unique basis zJ of ge+ exists so that (4.6.10) is satisfied. It is also clear from L~mma 14 that the Zj necessarily satisfy the condition of Theorem 5. Since b C e_ V the remaining statements follow from Lemma 15. Q. E.D.
+
4. 7. We will assume from here on that the basis Zj of Oe+ is given by Theorem 6. Now let Sj E S (b) be the coordinate functions on b corresponding to the Zj. That is, Sj is such that x = e_ + ~ Sj (x) Zj. We have already noted that S(b) =C[Sl,' . ',sz] (see (4.5.1)). In notational simplicity let
(instead of u b ) denote the restriction of u to b. Thus for any x E b t(x)
=
(t1(x),' . " tz(x))
371
381
LIE GROL"P REPRESENTATIONS.
where
ti =
Ui
I b.
It follows therefore from (4. 5. 2) that
,,= Oti
(4.7.1)
uS;
OZJUi
I b.
Now if u is an arbitrary k-plane in 9 let J
(4.7.2)
~S(u)
be the ring homomorphism obtained by restricting an invariant polynomial to u. Now in general one could hardly expect (4. 7. 2) to be an isomorphism. Indeed if (4. 7. 2) is an epimorphism one must have k > l and if (4. 7.2) is a monomorphism one must have k < l (since the Ui are algebraically independent) . Hence the possibility could only exist if k = l. If u is a Cartan sub algebra the one knows that (4. 7. 2) is a monomorphism and the image is the space of Weyl group invariants. Hence in such a case (4.7.2) is an isomorphism only when 9 is abelian. On the other hand when u = b we have, in general, the following corollary of Theorem 6 THEOREM
7.
If u = b then (4. 7. 2) is an isomorphism. Moreover ill e
map (4.7.3) obtained by restricting U to b is an algebraic isomorphism so that t 1 , · define a global coordinate system on b.
.
.,
tl
Furthermore the relationship between the ti and the linear coordinates S; on b is as follows: For i = 1, 2,· .. , l, there exists polynomials Pi and qi in i - I variables without constant term such that (4.7.4) and (4.7.5) Proof. To prove the theorem observe that it suffices only to prove (4. 7. 4). Indeed using (4. 7.4) we can solve for Si obtaining (4. 7. 5) . It is then immediate that t is one-one, onto and is in fact a biregular birational map. Since the ti generate the image of J in S (b) it is then also obvious that (4. 7. 2 ) is an isomorphism. But now (4.7.4) is immediate from (4.6.11) and (4.7.1). Finally by definition of the coordinate system Si one has Si ( e_) = 0 for all i. On the other hand ti ( e_) = 0 for all i since t ( e_) = U ( e_) = 0 (recall that e_ is nilpotent). Thus the Pi and qi have no constant term. Q. E. D.
372
382
BERTRAM KOSTANT.
Any orbit 0 of semi-simple elements (i. e., 0 E (h) intersects ~ in a finite number but in general more than one point. We now find that any orbit 0 of maximal dimension (i. e. 0 E @t) intersects 0 in one and only one point. THEOREM
8.
One has 0 C r.
j1'urihwrrnore if
(4.7.6)
is the map given by x --?o Om then (4. 7. 6) is a bijection. That is, no two distinct elements of 0 are conjugate and every element in r is conjugate to one and only one element in 0.
+
Proof. Since 0 C e_ fJ one has 0 C r by Lemma 10, § 4. 2. But now if we compose (4.7.6) with the bijection T)t (Theorem 2, §3.5) we obtain the bijection t (see Theorem 7). Hence (4.7.6) must be a bijection. Q. E. D. We can now obtain the following characterization of the set r. THEOREM 9. Let x E g. Then x E r if and only if (dui) x, i are linearly independent.
=
1, 2, ... , l,
Proof. By (4. 6. 12) the matrix (OZ/Ui) (x) is of rank l for any x E 0. Thus (du.) m, i = 1,· .. , l, are linearly independent for any x E 0. But then by Theorem 9 and conjugation the same is true for any x E r. Now let x E g but where x ~ r. We must prove that the (dui).;, i = 1, 2,· . ., l, are linearly dependent. Assume first that x E's (that is, x is semi-simple). Then x is not regular so that gm contains a Cartan subalgebra as a proper subalgebra. It follows therefore that if u is the center of gm and lm = dim u one has lo; < l. Furthermore it is also clear that It is the set of l1xed vectors for the action of Gm on g (recall that Gm is connected. See Lemma 5, § 3. 2) . Thus there exists a non-abelian simple component gl of g of rank, say ll' such that in the notation of § 2. 1 d
< II
where tf; ED and the irreducible representation v ifi is equivalent to the adjoint action of G on gl. (One uses here that yifi is self-contragredient.) But then by Proposition 8 the multiplicity of yifi in the G-module R(G/Gm) or R(Om) is less than ll. But S (Om) C R (Ox) (in fact here S (Ox) = R (Ox) since 0., is closed. See (2. 2. 3» so that the same is true for the G-module S (Ox). Now let g2 be an ideal in g complementary to gl so that g = gl El1 g2 is a direct sum. It is obvious that we may choose the primitive invariants so that for 1 < i 1 < i2 <. . . < il,. < lone has U..(Xl
+ x 2) =Ui.(X1)
373
LIE GROUP REPRESENTATIONS.
383
where xJ E gJ, j = 1, 2, and 7c = 1, 2,· .. , ll.' (The dependence or independence of the (du.,) III obviously does not depend upon the choice of the primitive generators Ui)' That is, for any such 7c, (4.7.7) N ow for any
Oy (Ui.) U
=
0 for any y E g2
E J, Y E g and a E G one clearly has a . OyU = oayu.
(4.7.8)
Now let y be the root vector ep where cp is the highest root of gl. It follows then from (4. 7. 8) that OyU E S"', and hence OyU I OlD E S'" (0 Ill), and in fact, if not zero, these functions are highest weight vectors. But then since the multiplicity of II'" in S (0 Ill) is less than II it follows that OyUi. IOu;, 7c = 1, 2,· . ., ll' must be linearly dependent. Thus there exists scalars Ci., not all zero, such that if U = ~ Ci.U•• then OyU I Ou; = O. But then by (4.7.7) Ie
and (4. 7. 8) one has O"U I 0 III = 0 for all z E g. In particular then (o"u) (x)
=
0 for all z E g.
Thus (du) I1J = 0 and hence the (du.,)"" i = 1, 2,· .. , l, are linearly dependent. N ow assume that x Egis arbitrary where x ¢ r. Since the set of all y such that the (dUi)Y are linearly independent is an open set (see § 1. 6) to prove the theorem it suffices from above to show that there exists a sequence Xk such that Xk~ x where the Xk are semi-simple but not regular. In fact it suffices to show this for the case where x is nilpotent. Indeed if x = y z is the decomposition (3.1. 2) for x then by Proposition 13 z is not a principal nilpotent element if gy. Hence if such a sequence has been shown to exist in the nilpotent case there exists a sequence Yk of non-regular semi-simple elements in gY such that Yk ~ z. Hence Xk = Y Yk converges to x. But clearly Xk is semi-simple and non-regular in g. Hence we may assume that x is a non-principal nilpotent element of g. By conjugation we may also assume x E m so that, by Theorem 4, § 4. 2, x = ~ cpep where there exists a simple root IZ E II such that Ca = O. Let
+
+
so that x E n. Also let Y E f) be such that 3, Y> is positive for /3 E II where f3 oF IZ and
+
374
384
BERTRAM KOSTANT.
+
y x is conjugate to y and hence is semi-simple and non-regular. But we may substitute y/k for y. Hence if Xk = y/k x then Xk~ x where Xk is semi-simple and non-regular. Q. E. D.
+
4.8. Generalizing the definition of J+, let J~, for any ~ E Cl, be the (maximal) ideal in J generated by Ui-~i' i=1,2,· . ·,1. Obviously then J~8= (U1-~1'· . ·,Ul-~Z)
Recall that P (~) egis the set of zeros of
J~8.
We can now prove
cz
THEOREM 10. Let ~E be arbitrary. Then P(~) is a Zariski closed subvariety (of X) of dimension n-l and its ideal I(pa» is given by
I(P(~» =J~8
(4.8.1)
so that (a) P a) is a complete intersection and (b) J~8 is a pl'ime ideal. Furthermore if P a) . is the set of simple points of P (~) then (4.8.2) where or(~) is the unique orbit of dimension n-r in P(~). (8ee Theorem 4, § 3. 8). Moreover the set of non-simple points in P (~) is a finite union of orbits and has a codimension of at least 2 in P (~). Proof. By Theorem 3, § 3. 8, P a) is a Zariski closed subvariety of dimension n - land P (~) n r is the orbit or (~) of dimension n - 1 defined in § 3. 5. But if x E P(~) then (dU;)"" i = 1, 2,· . ·,1, are linearly independent if and only if x E or(~) by Theorem 9, § 4.7. Hence by Proposition 6, § 1. 6, one obtains (4.8.1) and (4.8.2). Furthermore the set of non-simple points of P (0 is a finite union of orbits and have a codimension of at least two in p a) by Theorem 3, § 3. 8. Q. E. D. Remark 20. By (4.8.2) note that p(~) is a non-singular yaridy if and only if or(o =P(~); that is, if and only if or(~) is an orbit of regular elements. Now let B be a non-singular symmetric G-invariant bilinear form on g. (One extends the Cartan-Killing form of [g, gJ in an obvious way.) Let H C 8 be the graded space of G-harmonic polynomials defined as in § 1. 4 so that 8 = J+8 H is a G-module direct sum.
+
Remark 21. By Proposition 16, § 3. 8, note that the subspace Hp C 8 (see § 1. 4) is the space of polynomials spanned by all powers of gk E 8 le, k = 0, 1,· . ., and all linear functionals g E 8 1 corresponding to 0"" under the map (1. 4.1)' where x is an arbitrary nilpotent element of g.
375
LIE GROUP REPRESENTATIONS.
385
Finally for any AED recall that h is the multiplicity the zero weight for the representation Jl x. That is, h = dxA for any Cartan subgroup A C G. (See §2.1.) We can now prove THEOREM 11. Let g be a complex reductive Lie algebra and let 8 be the ring of all polynomials on g. Let G be the adjoint group of g and let J C 8 be the subring of G-invariant polynomials. Then 8 is free as a Jmodule (under multiplication). Furthermore
(4.8.3) and if
(4.8.4) is the map given by u ® h ~ uh then (4. 8.4) is a G-module isomorphism. Also for any ~ E C! the ideal J E8 is prime in 8 and 8=JE8+H
(4.8.5) is a direct sum.
Moreover H is completely reducible as a G-module and for any A E D the irreducible representation JlX of G occurs with multiplicity lx in H (so that H = ~ HX is a direct sum and dim H).. = hdx where dx is the dimension XED
of
JlX).
Let t be the set of all x E g whose corresponding orbit 0$ has maximal dimension (n -l) . Let x E t and let 8 (0$) be the ring of functions on 8 (0 $) obtained by restricting 8 to 0$ and let
(4.8.6)
H~8(0:JJ)
be the map obtained by restricting G-harmonic functions to Om. Then (4.8.6) is a G-module isomorphism so that all 8(0",), for xE t, are isomorphic as Gmodules. Proof. By Theorem 1, § 3. 4, and Proposition 15, § 3. 8, the cone P has a dense orbit and by Theorem 10 J+8 is a prime ideal (case where ~ = 0). One obtains (4.8.3) and (4.8.4) as a consequence of Proposition 4, § 1.4. Moreover the map (4. 8. 6) is an isomorphism by (1. 5. 2) and Proposition 18, § 4. 3. Since J E8 is prime by Theorem 10 one therefore obtains the direct sum decomposition (4.8.5) (using 3.8.7)). Obviously H is a completely reducible G-module. To find the multiplicity of Jlx in H one uses the isomorphism (4. 8. 6) and chooses x to be re:~ular.
376
386
BERTRAM KOSTANT.
In such a case R(O:c) =S(O:c) by (2.2.3) since OI/J=O{/). But the multiplicity of llAo in R(Of/}) is lAo by Proposition 8 since G:c is a Cartan subgroup Q.E.D. of G.
Remark 22. Except for (4. 8. 3) mte that by Proposition 2, § 1. 3, one may replace H in Theorem 11 by any G-stable complement of J+S in S. Also we wish to note that every irreducible representation of G appears with positive multiplicity in H. That is lAo > 1 for any A E D. Indeed let Z C I)' be the discrete subgroup generated by all roots cp E A. Since G is the adjoint group one knows that every weight of llAo can be regarded as an element of Z. On the other hand if D is identified with the subset of all }J. E Z such that }J.(xp) > O,for all cp E~, where Xp E ij is the root normal corresponding to cp, in such a way that A is the highest weight of llAo then it is known that any }J. ED is a weight of llAo if and only if A -}J. is a non-negative integral combination of positive roots. Since}J. = 0 always satisfies this condition it follows that h > 1. Remark 22'. It has been pointed out to us by Serre, as we ourselves have also noticed, one of the conclusions of Theorem 11, namely S = J 0 H, can be obtained directly from the theorem of Chevalley mentioned in Example 1 of the Introduction. (Added in proof.) As used above, the primeness of J+S implies that J+S is the ideal of the variety P. Another application of this fact in algebraic geometry is the following theorem. Clearly P meets the Cartan subalgebra ij (or any Cartan subalgebra) dim ij = dim g. only at the origin and dim P
+
THEOREM 12. Let w be the intersection multiplicity of P and ij at the origin. Then w is the order of the Weyl group.
Proof. Let L be the local ring at the origin (as a point of g) and let I and K, respectively, be the prime ideals of L corresponding to ij and P. Now one knows that w is the alternating sum of the integers dim ToriL(LjI,LjK). But clearly dim Tor.;,L(L/I, L/K) = dim ToriS(S (ij), S (P» We now observe, however, that by Theorem 11 one has ToriS(S (ij), S (P» = 0 for i> O. Indeed since S is J-free and S(P) = SjJ+S there is a spectral sequence converging to TortS (S (ij), S (P» where E p ,q2 =
TorpS(P)
(Tori (S (ij), C), S (P»
(See Cartan and Eilenberg, Homological Algebra, Chapter XVI, § 6, Theorem
377
387
LIE GROUP REPRESENTATIONS.
6.1, p. 349). On the other hand since 8(g) is J-free (See [2]) one has E p ,q2=0 unless p=q=O and EO,02=8(g)jJ+8(g). This verifies the observation and also proves that
w = dim 8 (g) jJ+8(g) .
(4.8.7)
But by [2] (and also from the cohomology theory of the generalized flag manifold) one knows that the right side of (4. 8. 7) is the order of the Weyl Q. E. D. group. 4. 9.
Now let p ( t) be the formal power series 00
pet) =
~
dim Hktk.
k=0
Since 8 is isomorphic to the tensor product J 0 H by Theorem 11 it follows therefore that !
I11-tm, (4.9.1) Now let of g and let
p(
g denote
t)
~i=.=-.l_ _
=
(l-t)n'
the projective space of all one dimensional subspaces
g- (0)
(4.9.2)
~g
be the canonical projection map. If u is a homogeneous Zariski closed subvariety of g let ii C g be the image of u - (0) under (4.9.2) so that ii is a projective variety. In the remaining portion of this section we use the notation of FAC, [15]. Consider the projective variety ~ C Ii defined by the cone 1J of all nilpotent elements in g. The dimensional determination of the sheaf cohomology groups Hi (~, qj (k» for all j, k E Z is given by THEOREM
13.
Let k E Z be arbitrary. Then
Hi (~, qj (k»
(4.9.3)
=
0
where j is any integer other than 0 or n - Z - l .
(4.9. 4)
HO(~, qj (k»
On the other hand
Hn-l-1(p, qj (Z- (2:
=
+ n) ».
(Recall that n - l is even.) Furthermore if q(t) is the formal power series defined by 00
q(i) =
~ dimHO(~, qj ~oo
378
(k) )tk
388
BERTRAM KOSTANT.
then q(t)
=
p(t).
That is q(t) may be given by z
II 1 - tm • (4.9.5)
q(t)
=
i(~_t)n .
Proof. By Proposition 15, § 3. 8, one has .p = P. Hence by Theorem 10, for ~ = 0, one has that ii is a complete intersection so that Proposition 5, §78, in [15] is applicable. This yields (4.9.3) and (4.9.4) since clearly N = l - n. On the other hand in the notation of [15] one has dim HO(.p, @(k» - dim Sk ($). But since J+S is the prime ideal of $ by (4.8.1) and since S=J+S+H is a direct sum it follows that dimHk=dimS k ($). Thus p(t) =q(t) and hence one obtains (4.9.5) from (4.9.1). Q.E.D.
4.10.
One obviously has an isomorphism
f~
S (C Z) of all polynomials on CZ by defining, for any
f* ~
of J onto the ring
E C z)
where p is that polynomial in l variables such that f = p (Ul" . " uz). We recall that the 14 are the primitive invariant polynomials. N ow let 'U be the set of all Zariski closed subvarieties of CZ. We may use 'U to index all the prime ideals in J by defining JU C J for U E 'U to be the prime ideal consisting of all f E J such that f* vanishes on U. If I C S is any prime ideal in S let u(I) C g be the corresponding Zariski closed subvariety of g of all points in g at which I vanishes. It is of course obvious that I is stable under G if and only if u(I) is stable under the action of G on g; that is, if and only if u is a union of orbits. It is clear of course that if I is generated by invariant polynomials then I is G-stable. However this is not a necessary condition. The question arises: how does one characterize all those G-stable Zariski closed subvarieties u of g whose prime ideal I(u) is generated by invariant polynomials? The following theorem asserts that a necessary and sufficient condition is un t should not be empty. Note that since t is a Zariski open subset of g (obvious from its definition. Also see the proof of Proposition 6, § 1. 6) then un t is Zariski dense in u in case un t is not empty. Theorem 14 also generalizes most of Theorem 10 (case where U has only one point). THEOREM 14. Let JU C J, U E'U, be any prime ideal in J. JUS is a prime ideal in Sand
(4.10.1)
u(JUS)
=
U pa)·
~E
U
379
Then
389
LIE GROUP REPRESENTATIONS.
That is, u(JuS) =u-1(U) where u is the map (3.3.2).
(4.10.2)
"Moreover
U~u(JUS)
defines a one-one correspondence between the set of all Zariski closed subvarieties of C! and the set of all G-stable Zariski closed subvarieties u C 9 such that un r is not empty. In particular U =u(u) is in 'U for such a subvariety u C g, u = u(JuS) and
(4.10.3) Let U E 'U and put u = u(JuS).
(4.10.4)
Then
co dim U in C! = codim u in g.
Furthermore if x E u n r then x is a simple point of u if and only if u(x) tS a simple point of U. Finally if r C is the (Zariski closed in g) complement of r in 9 then
(4.10.5)
co dim un r C (in u) > 2.
Proof. Let J' be any ideal in J. It is immediate that J'S is the image of J' ® H under the isomorphism (4. 8. 4) and hence one has
(4.10.6)
J'S
n J =J'.
Now assume that J' is a radical ideal (an ideal equal to its own radical) in J. We will show that J'S is a radical ideal in S. Let J' * be the radical ideal in S (CZ) corresponding to J' under the isomorphism J ~ S (C!) where f ~ f* and let U C C! be the Zariski closed set of all t E Cl at which J' * vanishes. It is obvious that if u is the Zariski closed set, in g, of all x E 9 at which J'S vanishes then U = u- 1 (U) or (4.10.7)
u= U pa)· ~E
U
To prove J'S is a radical ideal it suffices to show that if f E S is assumed to vanish on U then f E J'S. By Theorem 11, § 4.8, we can write f = ~ fihi where fi E J, hi E H and the hi are linearly independent. Let t E U. Then since the fi reduce to constants on P it follows from the isomorphism (4.8.6) and (3.8.7) that since f vanishes on pet) the fi also vanish on pa). Thus the fi are in J' by the nullstellensatz and hence f E J'S so that J'S is a radical ideal. Now let U E 'U so that JU is prime in J. Put J' = JU so that, from above, JUS is a radical ideal in S. To prove that JUS is prime it suffices now only to show that u is irreducible.
a)
380
390
BERTRAM: KOST.ANT.
Let f E S and let U (f) be the set of all H U such that f I P (0 is not zero. Obviously flu is not zero if and only if U (f) is not empty. We first show that in such a case U (f) contains a non-empty Zariski open subset of U. Indeed assume U (f) is not empty and ~ E U (f) . Then f I Ot (~) is not zero by (3. 8. 7). Hence there exists a E G such that (a· f) I Ot (~) n b is not zero, by Theorem 8, § 4. 7, where b is defined as in (4. 5. 6) . Thus (a· f) I b n b. Using the does not vanish on a Zariski subset of b containing Or isomorphism (4. 7. 3) it follows that U (a· f) contains a non-empty Zariski open subset of U. But clearly U (a· f) = U (f). Hence U (f) contains such a subset.
un
Now let fi E S, i = 1, 2, be arbitrary except that fi I u is not zero. To show u is irreducible we must show that fd2 I u is not zero. From above it follows that U (fi) contain a non-empty Zariski open subset of U. But since U is irreducible it follows that U (fl) n U (f2) is not empty. But then fd2 I P (~) is not zero in case ~ E U (h) n U (f2) since pa) is irreducible by Theorem 3, § 3. 8. Thus u is irreducible and hence JUS is prime. The relation (4.10.1) is just (4.10.7). then (4.10.3) follows from (3.8.4).
Furthermore if u = u(JuS)
Moreover, using (4.10.1), it is immediate that the map, given by (4.10.2), from 'U into the set of all Zariski closed G-stable subvarieties u of g such that u n r is not empty is injective. Now assume that u is such a subvariety. We will show that It is in the image of the map defined by (4.10.2). Let the set U C cz be defined by putting U = U (u n r). Since u is Zariski irreducible and u n r is Zariski dense in u it follows that U is 'Zariski irreducible. On the other hand by Theorem 8, § 4.7, it is clear that U corresponds to u n b under the isomorphism (4. 7. 3). But since u n b is Zariski closed in b it follows that U is Zariski closed in C z. Hence U E 'U. But U is Zariski dense in u (u). But this implies u (u) = U since U is Zariski closed. Thus u C u-1 (U). But u- 1 (U) is clearly in the Zariski closure of unr since the relation (4.10.3) obviously holds. Thus u=u-1(U) or u=u(JuS). Now obviously (dUi*)~' i = 1, 2,· . ., l, are linearly independent at any point ~E cz. Since (df)a: is in the span of the (dUi)g; for any fE J and xE g it follows from Theorem 9, § 4. 7, that (df) g; = 0 if and only if (df*) ~ = 0 for any f E J and x E r, where ~ = U (x). It follows in particular that if U E 'U and x E u n r where u = u (JuS) then the dimension ra: of the space spanned the (df) g; for all f E JU is the same as the dimension r~ of the space spanned by all (df*) ~ where f E JU and ~ = U(x) . If r is the co dimension of
381
391
LIE GROUP REPRESENTATIONS.
U in CZ and Us is the set of simple points of U then by the Zariski criterion for all ~E U and r~=r if and only if ~E US. Thus rl1J
r~
+
+
Remark 23. By putting U = g in (4.10.5) note that r C has a dimension of at least 2 in g. On the other hand if q is the set of regular elements in g and v E J is the invariant polynomial such that v I f) is the product of all the roots (positive as well as negative, so that it is a Weyl group invariant) then the complement qe of g is the set of zeros of v and hence has co dimension 1 in g.
5.
The normality of the varieties
PC;)
and the generalized exponents.
The following criterion for normality and its proof is due to Seidenberg. THEOREM 15 (Seidenberg). Let u C g be a Zariski closed subvariety of g. Assume (a) that u is a complete intersection and (b) the set of nonsimple point of u has a codimension of at least two in u. Then u is a normal variety.
Proof. Let r = dim,u. By [16J, Theorem 3, one knows that u is normal if (1) u is free of (r - 1) -dimensional singularities and (2), every principal ideal in the affine algebra of u is unmixed. Since assumption (1) is satisfied (statement (b) in Theorem) it suffices therefore only to show that if I(u) is the prime ideal, in S, corresponding to u and f E S then the ideal (I (u), f) is unmixed. Obviously one may assume that (I (u), f) has dimension r-1. But then the result follows from Macaulay's theorem (see [19J, p. 203) since, by (a), one has that I(u) = (fl" .. ,fn-r) for some fi E I(u). Q. E. D. N ow if V is any finite dimensional G-module and x Egis arbitrary consider va' the subspace of vectors in V that are fixed under all elements of Gx. We now find that the dimension of Va' is the multiplicity of the zero weight in V (and hence is the same) for all x E r. (This, incidentally is not
382
392
BERTRAM KOSTANT.
necessarily true for the covering group of G). Obviously it is enough to show this for irreducible G-modules. THEOREM 16. Let ~ E C! be arbitrary. Then P(~) is a normal variety. That is, if x E r then 0", is a normal variety (see Theorem 4, § 3. 8). Furthermore if R(O",) denotes the ring of everywhere defined functions on the orbit 0", (R(O",) is isomorphic to R(GjG"'» then R(O",) is an affine algebra. (That is, it is finitely generated.) In fact R (0 x) = S (0 x) where S ( 0 x) is the restriction of S to 0$ so that if ~=u(x) then P(~) =Ox is the affine variety of maximal ideals of R (0 x) . ]}foreover R (0",) is a completely reducible G-module and for any A E D the multiplicity of vA in R (0 x) is lx, where lA is the multiplicity of the zero weight of VA (or equally of ]lA), so that R(Ox)' for all xE r, are isomorphic as G-modules. Finally for any x E r
dim VAG"=lx
(5.1.1) where V A and
VA
is defined as in § 2. 1.
Proof. By Theorem 10, § 4.8, P (~) is a complete intersection. Also by has a co dimension of at least Theorem 10 the set of non-simple points of P two in P(~). Hence P(~) is normal by Theorem 15. But now by Corollary 1, § 3. 8, the complement of Oil! in Om has a co dimension of at least two in Oil! for any xEr. Hence R(OIl!) =S(OIlJ) by Proposition 9, §2.2. But in any case S (Ox) is isomorphic to S (Ox), Since S (Ox) = R (O:c) (see (2.2.2» it follows that every element of R (0 x) extends uniquely to an element of R(Ox). This induces an isomorphism
a)
(5.1. 2) so that R(O",) is an affine algebra. Obviously R(O",) is a completely reducible G-module. But by Theorem 11, § 4. 8, the multiplicity of ]lA in S (0 1lJ) is lA' It follows therefore from Proposition 8, § 2.1, and the G-module isomorphism (2.2.1) that dim VAG" = lx. Q. E. D. Let x E r. As a corollary of Theorem 16 we now observe that 0", is distinguished among all affine varieties into which O:c may be embedded. 3. Let x E r be arbitrary so that ~ = u(x) E C! is arbitrary. If we identify Gj Gx with 0$ (ltsing the isomorphism (1. 2. 2» then any morphism of G/Gx into any affine variety X extends uniquely to a morphism of the affine variety P (~) into X. In particular any morphism of the orbit of principal nilpotent elements COROLLARY
383
393
LIE GROUP REPRESENTATIONS.
into an affine variety X extends to a morphism of the variety of all nilpotent elements into X. Proof. U sing the isomorphism (5. 1. 2) this result follows from Corollary Q. E. D. 1, p. 58 in [3J. Remark 24. By Theorem 16 all R (0 a;) for x E r, are isomorphic as Gmodules. However it should be noted that they are not isomorphic as rings. Indeed the corresponding variety of maximal ideals of R (0 a;) is non-singular in case x is regular and, by Theorem 10, § 4. 8, is singular otherwise (for example, in case x is principal nilpotent). 5. 2. Now for any .\ E D consider Homo (VA, S) the space of all G-module maps y of VX into the ring of polynomials S. Obviously y(VX) C SX for any such y. We now observe that any x E g induces a linear map w",:
Homo(VA,S)
~
VxGz
by the relation (5.2.1)
for all v E VX and any y E Homo (VA, S). Since y IS a G-module map it follows immediately from (1.1. 5) that for any a E G
(5.2.2)
vx(a)wa;(y) =w"",(y)
and hence, obviously, Wa; (y) E V XGZ for any x E g. Now by Theorem 11, § 4. 8, the subspace Homo (VA, H) of Homo (VA, S) is of dimension h. On the other hand, by (5.1.1), VxGz is also h-dimensional whenever x E r. As a corollary of Theorem 16 we obtain COROLLARY 4.
(5.2.3)
Let x E r and let .\ E D. Then the map Homa(VA, H)
~
VxGz
defined by restricting Wa; is an isomurphism. Proof· Since both sides of (5. 2. 3) are vector spaces of dimension lx it suffices to show (5.2.3) is a monomorphism. Let y be in the kernel of (5.2.3). Then, by (5.2.2), waa;(y) =0 for all aE G. Now let vE VA. Then, by (5.2.1), (y(v) )(ax) = for all a E G. Thus y(v) I Oa;= 0. But since (4. 8.6) is an isomorphism it follows that y (v) = for all v E VX. Hence y = 0, and consequently (5.2.3) is an isomorphism. Q. E. D.
°
°
5.3. N ow since ad maps g into the Lie algebra of G it is clear that any finite dimensional representation
v:
G~AutV
384
394
BERTRAM KOSTANT.
of G induces, by taking differentials, a representation of g, which we also denote by v. Note that one always has v(a) = 0 where a is the center of g. Now let Z be the subgroup of I/ generated by the set of roots a. If o (p.), the order of p., is defined by
o(p.)
=
<xo,p.>
for any p. E Z it is then clear from (4.1. 6) that 0 (p.) is always an integer. Now since every weight of II is clearly necessarily an element of Z it follows therefore that V = ~ V(k) k€Z
is a direct sum where value k. Obviously
V(k)
is the eigenspace of II(Xo) belonging to the eigen-
(5.3.1) Now e_ be the principal nilpotent element defined as in § 4. 2. For notational simplicity write F = vo e-. Since Ge_ is connected (Proposition 14, § 3. 6) one also has (5.3.2)
F=Kerll(ge-)
and by (5.1.1) dimF=lv where lv is the dimension of the zero weight of v. Now since Xo lies in the normalizer of ge_ it follows from (5. 3. 2) that F is stable under v (xo). But then we observe that there exists a unique sequence of integers mi (II), i = 1, 2,· . ., lv, where m1(V) < ... <mlv(lI)
such that F has a basis
Vi,
i = 1, 2,· . ., l., where
(5.3.3) Remark 25. By applying the inner automorphism which carries e+ into e_ and Xo into - Xo note that we would get the same integers mj (II) by using e+ into instead of e_ and dropping the minus in (5.3. 3). Observe then that the mi(v) generalizes the notion of exponents. Indeed if II is the adjoint representation then lv = l and, by Theorem 5, § 4. 4, mi(V) = mi since F = ge_. Since F C Ker v (e_) it follows from the representation theory of a three dimensional simple Lie algebra (e. g. see [13], § 2. 5) that for any i
(5.3.4)
385
LIE GROUP REPRESENTATIONS.
395
Now, as in Remark 22, § 4. 8, identify D with the (subset of Z) set of all dominant integral (with respect to G) forms on f) so that any A ED is the highest weight of vA. Note then that - A is the lowest weight of VA. When V = VA and v = VA we will write FA for F and miCA) for mi(VA). The miCA), i=1,2,·· ·,lA, will be called the generalized exponents of g (corresponding to A). See Remark 25. Now let VA EVA be the lowest weight (- A) vector. Then since VA, as one knows, is a cyclic module with respect to the universal enveloping algebra of m with VA as cyclic vector it follows that
(V ) (-k)
_
A
-
{O
if k> O(A) (VA) if k=O(A).
Since VA (-o(A» is obviously contained in FA. (One uses the relation ge_ C m* a mentioned in the proof of Theorem 5, § 4. 4). It follows then that (5.3.5) miCA) <mIA(A) =O(A)
+
for 1
H(A)
=
"2:.H(>.Y· j=O
On the other hand by Theorem 11, § 4. 8, the multiplicity of vA in H is h. Hence one has a direct sum (5.4.1) where (1) Hi(A) is an irreducible G-module, (2) Hi(A) is a space of homogeneous polynomials so that for some degree ni(A) one has Hi(>") C Sn,(A) where (3) we may assume the ni(A) are monotone non-decreasing with i. The integers ni(A) are the degrees k such that vA occurs in Hk. The question arises: how does one determine these integers? Since S (>..) is obviously isomorphic to the tensor product J ® H (A) by Theorem 11, § 4. 8, it is clear that such information is needed if one is to determine the formal power series 00
qA ( t)
=
"2:. dim S (A) k
tk
k=O
and, as a consequence, the multiplicity of vA in Sk for any k.
386
396
BERTRAM KOSTANT.
The following theorem asserts that the ni('\) are exactly the generalized exponents mi('\). THEOREM 17. For any ,\ ED and i = 1, 2,· .. , lA, one has ni('\) = mi('\) so that Hi (,\) C Sm,(A). In particular 7c = 0 (,\) is the maximum degree such that H(,\)k#O. Furthermore H(,\)k is irreducible for this value of le. That is,
Moreover the formal power series
qA (t)
(5.4.2)
qA (t)
may be given by
= dA _i7~1_ _ __
II (l-tml) i=l
where d A = dim VA. Proof. Let V'i' i = 1, 2,· . ,h, be a basis of FA such that V'i EVA (-m,(A». Now let c E C* be arbitrary. Let r E C be such that e- r = c and let a E G be defined by putting a = exp r ad Xo. It is then clear that (5.4.3) JIA (a) v'i = Cm,(Alv'i. Also note that (5.4.4) Now by Corollary 14, § 5. 2, there exists a basis Yi, i
=
1, 2,· . ., h, of
Homa(VA,H) such that But now by (5.4.4-5) and (5.2.2) Olle getR the equation (5.4.5)
cm,(A)we_ (Yi)
=
Wce_ (Yi).
Substituting in (5. 2. 1) this implies that for any v E VA
(Yi(V» (ce_)
=
Cm,(A)Yi(V) (e_).
But then, conjugating by G (and using (1. 1. 5)
it also follows that
Yi ( v) (cy) = Cm,(A)Yi ( v) (y)
for all yEO e_. But then since (4. 8. 6) is an isomorphism for x = e_ it follows easily by choosing c, for example, to be positive that
Yi(VA) C Sm,(A). On the other hand by definition of the Yi
387
397
LIE GROUP REPRESENTATIONS.
is a direct sum of irreducible G-modules. By uniqueness of the nt(A) it then follows that ni(A) = mi(A) for i= 1,' .. , h. The second and third statements of the theorem follow immediately from (5.3.5). The equation (5.4.2) is an immediate consequence of the obvious fact that the isomorphism (4.8.4) induces an isomorphism of J®H(A) onto 8 (A). Q. E. D. Remark 26. We observe here that Theorem 17 is a generalization of Theorem 5, § 4. 4, asserting that mi(v) = mi where v is the adjoint representation. Indeed let U be the subspace of all U E J+ such that <0, = 0 where 0 E (J *+) 2. I t follows immediately that dim U = Z and that any homogeneous basis of U is a set of primitive invariants Ui, i = 1, 2,' . ., Z. But if the Ut are so chosen then by definition of U one has OUt E JO = 8 0 (immediate from (1. 1. 2» for any 0 E J *+. It follows immediately then that, for 1 < i < Z,
u>
oxUt E Hm, for every x E g.
(5.4.6)
(Recall that 8* is commutative.) But now if gl is a simple component of g of rank Zl and Ut., k = 1, 2,' .. , Zl, are as in the proof of Theorem 9, § 4.7, then one must have oxUt. E H (if) for any x E gl where if E D is defined as in the proof of Theorem 9. See (4.7.8). (In particular note that if g is simple and Xj is a basis of g then (5.4.7)
OXJUi,
i=I,· . ',Z, j=I,· . ',n is a basis of H(if)
where if is the highest root of g). It follows immediately that nk (if) = mi•. Applying Theorem 17 one then obtains mk(if) = mi. which immediately yields mj(v) =mj' It should also be observed then that the relation 0 (A) = mlA (A) generalizes the well known relation 0 (0/) = mi. See Corollary 8.6 and Lemma 9.1
in [13]. Remark 27. Note that the argument given in the proof of Theorem 17 may be reversed. That is, if L is an arbitrary irreducible G-submodule of H (A) and y E HomG ( VA, H) is such that (5.4.8) then for any integer j > 0 one has (5.4.9) Indeed if L C 8 1, then by (5.2.1), wee) y) = cjwe_ (y) for any c E C*. But then if a EGis defined as in the proof of Theorem 17 one obviously has
388
398
BERTRAM KOSTANT.
We_(Y) E VX(-i) Slllce vx(a)we_(Y) =ciwe_(Y), by (5.4.4) and (5.2.2) and c E C* is arbitrary. The argument for the other direction has been given in the proof of Theorem 17. For any ,.\ E DIet Yi XE Homa (VA, H), i = 1, 2,· .. , lx be fixed so that YiX(VX) = Hi("\). Obviously the y.x are a basis of Homa(VA,H) by (5.4.1). Thet argument in Remark 27 may be used to yield 18. Let x E r and a E a. Assume a(x) Now let ,.\ E D. Then the lx-dimensional (by 5.1. 1» under vx (a) . Furthermore for i = 1, 2,· .. , lx, THEOREM
=
cx for some c E c*. space V xG" is stable
{cm,(X)} are t·he eigenvalues of vx (a) I V XG· and w'" (y.x) is a corresponding basis of eigenvectors. Proof. Since Hi("\) C Smi(X) by Theorem 17 it follows from (5.2.1) that We", (y.X) = cm,(X)w", (YiX). But then vx (a) w'" (y.x) = cm,(X)w", ( yl) by (5. 2. 2) . On the other hand by Corollary 4, § 5. 2, the WI/) (YiX) are a basis of V xG·.
Q.E.D. 5.5. Let V be a finite dimensional a-module with respect to a representation v. Let A C a be the Cartan subgroup of a corresponding to f) so that VA is the zero weight space. We recall that W is the Weyl group of a corresponding to f). Now since VA is obviously stable under the normalizer of A in a it follows that v induces a representation 7r: W~AutVA
of the Weyl group W on VA. One notes that this is a generalization of the usual representation of W on f) (case where v is the adjoint representation). When V = V x and v = VA we will write 7rX for 7r. Now assume in the remainder of this section that g is a non-trivial simple Lie algebra. Let t{I ED be the highest root of g so that v'" is the adjoint representation. Let
+
or equivalently let s = 1 mz. See end of Remark 26, § 5.4. We recall that an element a E W is called a Coxeter-Killing transformamation in [13J, § 8.1, if it can be expressed as the product of the reflections defined by the simple roots (in any order) relative to any system of positive roots. Let a E W be a Coxeter-Killing transformation. It was observed empirically by Coxeter and then proved independently by Steinberg and in [13J
389
LIE GROUP REPRESENTATIONS.
399
that the order of u is s. It was also observed empirically by Coxeter and proved in [4] that the eigenvalues of 7rrp(u) are e21TimJ/8, j=1,2,· . ·,l. We will now generalize this for all ,\ ED. By a theorem of Coleman (see [4]) there exists a regular element z E f) such that u (z) = e21Ti /·z. (5.5.1) In fact, by Corollary 9.2 in [13], z is a cyclic element of g. 19. Let
Proof.
Let a E G be any element of the normalizer of A which defines I V XA=7rX(u). Since a(z) =cz where c=e21Ti /8 the result follows immediately from Theorem 18 since by Coleman's theorem z is regular and hence z E r. Q. E. D.
uE W so that vx(a)
Remark 28. If g is the three dimensional simple Lie algebra we may identify D with the set of non-negative integers where dim VX = 2,\ 1. Here lx=l for all ,\ED and m 1 (,\)=,\ by (5.3.5) since 0('\)=>". Also obviously s = 2. Note then that one recovers from Theorem 19 the well known fact that 71'X (u) = (-1) x for u E W, u # 1.
+
Remark 29. For any k E Z let [k] denote its canonical image in Z8=ZlsZ and for any mE Z let rx(m) be the number of integers 1
+
6. A decomposition theorem fo·r the universal enveloping algebra U of g. 1. Let T be the tensor algebra over g. Then one knows that in a unique way T is a G-module so that G operates as a group of algebra automorphisms extending the action of G on g. Let Q C T denote the G-submodule of all symmetric tensors in T. Also for T = 0,1 let IT be the ideal in T generated by all (x0y-y0x) -T[X, y] where x, y E g. Then since IT is G-stable the algebra TT = T lIT is a G-module
390
400
BERTRAM KOSTANT.
and one knows that the canonical epimorphism T isomorphism 8T: Q~ TT.
~
T T induces a G-module
However, by definition the G-module To is the symmetric algebra S* with the G-module structure of § 1. 1 and the G-module T1 is the universal enveloping algebra U of g with G operating by the usual extension of the adjoint representation. Now if 8* = 81 000- 1 then obviously 0*: S*~ U
is a G-module isomorphism and that furthermore since Q is the subspace of T generated by all tensors of the form x 0· . . ® x where x E g it follows that 0* «O",)k) = Xk for any x E g. Finally then if we compose 8* with the imer;:;e of B (see § 1. 4) one obtains a G-module isomorphism 0:
S~U
such that, for any integer k > 0, (6.1.1)
for every g E Sl. Furthermore, one knows, by the theorem of Birkhoff-WiGt that the filtration of U defined by the G-submodules k
U k =8( }:.Si) i=O
is such that (6.1.2)
8(fg)
=
8(f)8(g) mod Ui +H
for (not necessarily homogeneous) polynomials deg g < j. In particular (6. 1. 2) implies (6. 1. 3)
f
and g where deg f < i and
UiU j C U i+j •
Now let Z CUbe the center of U. By a theorem of Chevalley one knows that Z, like J, is isomorphic to a polynomial ring in l generators. Furthermore since Z is clearly the sub algebra of fixed elements under the action of G on U and since 8 is a G-module isomorphism it follows that (6.1.5)
8(J) =Z.
We now introduce a G-submodule E of U by letting E be the subspace spanned by all elements of the form :ik, k = 0, 1,· .. , where x Egis nilpotent. Remark 30. Note that the universal enveloping algebra U (m) of m is contained in E and (since every nilpotent element is conjugate to an element in m) that in fact E is the subspace of U spanned by all the algebras {U (m') } where m' runs through all the Lie sub algebras of g conjugate to m.
391
401
LIE GROUP REPRESENTATIOKS. THEOREM
20.
One has
8(H) =E where, we recall, H is the space of all G-harmonic polynomials on g. Proof. Obviously 8(P') =P where P is the set of all nilpotent elements in g and P' is defined as in § 1. 4. But since H = H p by (4. 8. 3) the thee rem follows immediately from (6.1. 1). Q. E. D. The filtration on U induces a filtration on E where Ek = En U'e. N ow regard U as a Z-module (with respect to multiplication). The following is our main result on the structure of U. THEOREM 21. Let U be the universal enveloping algebra of a reductive Lie algebra g. Let Z be the center of U. Then U is free as a module over Z. In fact if E is the G-submodule of U defined above (see Remarlc 30) then the map (6.1.6) Z0E~U
given by p 0 q ~ pq is a G-module isomorphism. Furthermore for any A E D the multiplicity of the irreducible representation vA. in E is h (the multiplicity of the zero weight for vA.). 111 oreover the order 0(11.) of A (see § 5.3) is the smallest integer k such that the multiplicity of vA. is Ek is h. Proof. Let f3 denote the (G-module) map (6.1. 6). To show first that f3 is surjective assume inductively that Uj CImf3 (obviously Uo CIm(3 since
U o C E) for some integer j. Let r E Uj+1. Then r = 8 (g) where, by Theorem 11, § 4. 8,
with fi E J, hi E Hand deg fi
+ deg hi < j + 1.
r=(3C~8(fi)
But then
08(hi » modU j
i
by (6.1.2). We have, of course, used (6.1.5) and Theorem 20. But since U j C 1m f3 it follows that r E 1m (3. Hence (3 is surjective. Now let p E Z 0 E where p =1= o. Then p is the image of an element e E J 0 H under the isomorphism J 0 H ~ Z 0 E induced by 8 (using (6. 1. 5) and Theorem 20). Furthermore we may assume that e = L fi 0 hi where i
0=1= fi E J and the hi are homogeneous and linearly independent in H. Now let k = max ( deg fi i
§ 4. 8, that if g =
L fihi
+ deg hi) .
It follows therefore by Theorem 11,
then g =1= 0 and deg g = k.
i
(6.1.7)
8(g)
=1= 0 mod Uk-I.
392
Hence
BERTRAM KOSTANT.
402
one has
S(g)
(6.1.8)
=
f3(p) mod U k - 1 •
Thus f3 (p) # 0 by (6. 1. 'l' -8) and hence f3 is an isomorphism. The remaining statements of the theorem follow immediately from Theorems 11, 1'l' and 20.
Q.E.D. 6.2. If
Let G1 be any algebraic reductive group whose Lie algebra is g.
(6.2. 1)
v: G1~ Aut V
is a representation of G1 on a finite dimensional space V then the corresponding representation of g and U on V will also be denoted by v. An element a E G1 is called unipotent if a is of the form a = exp x where x Egis nilpotent. Let v be as in (6.2.1) and let W C End V be the space spanned by all operators on V of the form v (a) where a E G1 is unipotent. Then W =v(E).
LEMMA 16.
Proof. By definition of E and the exponential formula it is obvious that WCv(E). On the other hand if x is nilpotent then v(exptx) E W for all real numbers t. But, clearly, W also contains all t-derivatives of v (exp tx) . Hence v(Xk) E W for k = 0, 1,· . .. Thus v(E) C W. Q. E. D.
As a corollary of Theorem 21, § 6. 1, we obtain THEOREM 22. (6.2.2)
Assume v (as in (6.2.1)) is irreducible. v(E)
=
Then
End V
or equivalently (by Lemma 16), every operator b on V may be put in the form k
(6.2.3)
b=
L CiV(a;,) i=l
where the Ci are complex scalars and the a;, are unipotent elements of G1 • Proof. Since v is irreducible one has v(U) = End V and v(Z) =
c.
The equation (6.2.2) then follows from the fact that (6.1. 6) is an isoQ. E. D. morphism. The second form of it follows from Lemma 16. Now let D 1, V1~ and V1~' for ~ E D 1, play the same role for G1 as the corresponding notation without the subscript 1 plays for G. (See § 2.1 and Remark 22, § 4. 8). Now End V1~ is a G1-module with respect to the tensor product of V1 e with the representation of G1 contragradient to V1~. But since the center of
393
403
LIE GROUP REPRESENTATIONS.
G1 obviously operates trivially on End V1~ it follows that End V1~ is a Gmodule and in fact the homomorphism of U onto End V1~ induced by V1~ is a G-module epimorphism. Since all modules under consideration are completely reducible it follows therefore by Theorem 21, § 6.1, that V1~ induces a Gmodule epimorphism (6.2.4) for every A ED. Since every representation of G induces a representation of G1 it is clear that D C D 1. One defines a partial ordering on D1 by putting ~ > ~', for ~, f E D 1, whenever ~ - f E D 1. It is easy to see that D1 is a directed set with respect to this ordering. (" Sufficiently large" will mean all ~ E D1 such that ~ > f for some ~' E D 1 . ) 17. For any ~ E D1 and A ED let nJ\ in End V 1 ~ (regarded as a G-module). Then
LEMMA
of
vJ\
a)
denote the multiplicity
(6.2.5)
and (for fixed A) the equality holds for
~
sufficiently large.
Proof. Since (6.2.4) is an epimorphism the inequality (6.2.5) is an immediate consequence of Theorem 21, § 6.1. However a much simpler and more direct proof of the inequality may be given using § 4. 1 in [12 J. Now identifying V 1(@ V1.~ with End V1~ and regarding G-modules as G1 -modules it follows immediately from Schur's lemma, upon forming the triple tensor product VJ\@V1~@V1.~' that nJ\(~) is also the multiplicity of V1~ in VJ\ @ V1( We refer now to [12J, § 4. 4, for the definition as to when A is totally subordinate to~. By Theorem 5.1, (3) in [12J A is totally subordinate to ~ for ~ sufficiently large. :{3ut now by (6) in this theorem (where JL = 0, A2 = A, A1 =~) the multiplicity of V1~ in VJ\ @ V1~ is lx whenever A is totally subordinate to~. Hence nJ\ (~) = lJ\ whenever A is totally subordinate to ~ or when ~ is Q. E. D. sufficiently large. Harish-Chandra proved in [8J that if Y C U is anyone-dimensional subspace there exists ~ E D1 such that V1~ is faithful on Y. This is not true in general for higher dimensional subspaces. For example if p E Z and q E U where q ¥= 0 and p ¢ U o then q and pq span a two dimensional space in U but its image under V1~' for any ~ E D 1 , is at most a one dimensional space. We now observe, however, that the generalization is true provided that Y C E. THEOREM 23. Let Y C E be any finite dimensional subspace. Then the irreducible representation V1~ is faithful on Y for all ~ E D1 sufficiently large.
Proof·
Since Y is finite dimensional there exists k such that Y C E k •
394
404
BERTRAM KOSTANT.
N ow let 0 CD denote the set of all A E D such that vA occurs with positive multiplicity in E k • Since E is finite dimensional it is obvious that 0 is a finite set. N ow since Dl is a directed set it follows that equality holds in (6.2.5) for all A E 0 and all ~ sufficiently large. But then by Theorem 21, § 6.1, the map (6.2.4) is an isomorphism also for all AE C and ~ E Dl sufficiently large. Thus Vl~ is faithful on Ek and hence on Y for all ~ sufficiently large. Q. E. D.
REFERENCES.
[I] A. Borel and Harish-Chandra, "Arithmetic subgroups of algebraic groups," Annals of Mathematics, vol. 75 (1962), pp. 485-535. [2] C. Chevalley, "Invariants of finite groups generated by reflections," American Journal of Mathematics, vol. 77 (1955), pp. 778-782. [3] - - - , Fondements de la geometrie algebrique, Course notes at the Institut Henri Poincare, Paris, 1958. [4] A. J. Coleman, "The Betti numbers of the simple Lie groups," Canadian Journal of Mathematics, vol. 10 (1958), pp. 349-356. [5] E. B. Dynkin, "Semi-simple subalgebras of semi-simple Lie algebras," American Mathematical Society Translations, ser. 2, vol. 6 (1957), pp. I Il-244. [6] F. Gantmacher, "Canonical representation of automorphisms of a complex semisimple Lie group," Matematii5eskii Sbornik, vol. 47 (1939), pp. 104-146. [7] Harish-Chandra, "On a lemma of Bruhat," Journal de Mathematiques Pures et Appliquees, vol. 9, 315 (1956), pp. 203'-210. [8] - - - , "On representations of Lie algebras," Annals of Mathematics, vol. 50 ( 1949), pp. 900-915. [9] G. Hochschild and G. D. Mostow, "Representations and representative functions on Lie groups, III," Annals of Mathematics (2), vol. 70 (1959), pp. 85-100. [10] S_ Helgason, "Some results in invariant theory," Bulletin of the American Mathematical Society, vol. 68 (1962), pp. 367-371. [II] N. Jacobson, "CompJetely reducible Lie algebras of linear transformations," Proceedings of the American Mathematical Society, vol. 2 (1951), pp. 105-133. [12] B. Kostant, "A formula for the multiplicity of a weight," Transactions of the American Mathematical Society, vol. 9 (1959), pp. 53-73. [13] - - - , "The principal three-dimensional subgroup and the Betti numbers of a complex simple Lie group," American Journal of Mathematics, vol. 81 (1959), pp. 973-1032. [14] M. Rosenlicht, "On quotient varieties and the affine embedding of certain homogeneous spaces," Transactions of the American Mathematical Society, vol. 101 (1961), pp. 2Il-223. [15] J. P. Serre, "Faisceaux algebriques cohere;It," Annals of Mathematics, vol. 61 1955), pp. 197-278. [16] A. Seidenberg, "The hyperplane sections of normal varieties," Transactions of the American Mathematical Society, vol. 64 (1950), pp. 357-386. [17] R. Steinberg, "Invariants of finite reflection groups," Canadian Journal of M athematics, vol. 12 (1960), pp. 616·618. [18] O. Zariski and P. Samuel, Commutative Algebra, vol. I, van Nostrand Company, Princeton, 1958. r19] - - - , Commutative Algebra, vol. II, van Nostrand Company, Princeton, 1960.
395
ANNALS OF MATHEMATICS
Vol. 77, No. I, January, 1963
Printed in Japan
LIE ALGEBRA COHOMOLOGY AND GENERALIZED SCHUBERT CELLS By BERTRAM KOSTANT
(Received January 5, 1962)
1. Introduction
1.1. This paper is referred to as Part II. Part I is [4]. The numerical I used as a reference will refer to that paper. A third and final part, Clifford algebras and the intersection of Schubert cycles is also planned. In a word let X be any compact algebraic homogeneous space of positive Euler characteristic. We solve here the problem of § 1, Part I, by constructing on X closed invariant differential forms whose cohomology classes are dual to the Schubert homology classes. These differential forms are defined using irreducible representations of the isotropy group on the homology of a nilpotent Lie algebra (although long suspected, the results are new even in the case of the grassmannian). The method introduces a new type of laplacian. In more detail let 9 be a complex semi-simple Lie algebra and let u be any Lie subalgebra which contains a (fixed) maximal solvable Lie subalgebra of 9. If n is the maximal nilpotent ideal of u the cohomology group, H(n, V), where V is any 9 (and hence n) module, was determined in Part 1. In fact H(n, V) is a 91-module where 91 = u n u* (the *-operation on 9 is defined relative to a fixed compact form f of 9) and the irreducible components were shown to be in a natural one-one correspondence with a subset W1 of the Weyl group Wof 9. See I, Theorem 5.14. (Since u = 91 + n is a Lie algebra semi-direct sum, one could just as well have replaced 91 byujn.) In Part II we will only need the result for the special caseH(n), that is, where V is the trivial module. In fact it is somewhat more convenient to dualize and deal with homology, H*(n), instead of cohomology for n. Now let G and U be a Lie group and subgroup corresponding to 9 and u. If X = Gj U then X is complex compact algebraic homogeneous space and every such space of positive Euler characteristic is of this form. One knows (Chevalley-Borel) that X may be written as a disjoint union
X-u uew1 Vrr where Vrr is the orbit of N (the unipotent subgroup corresponding to n) on X defined by a. Furthermore the Vrr , called Schubert cells (generalizing 72
B. Kostant, Collected Papers, DOI 10.1007/b94535_18 , © Bertram Kostant 2009
396
LIE ALGEBRA COHOMOLOGY
73
terminology from the case where X is the grassmannian) are homeomorphic to cells and define a basis of the homology group of X, and hence define a dual basis x", a E W\ of H(X, C). The problem, posed in § 1 of Part I, is how to relate H(X, C) with the gcmodule HAn) or, more specifically, how to relate, for any a E W\ the class x" E H(X, C) with the irreducible component H*(n)<", ~" = g - ag, of H*(n). Now, since X is compact, it is immediate and well known that H(X, C) is canonically isomorphic with the relative cohomology group H(g, gl). In fact the corresponding cochain complex C(g, gl) may be identified with the space of all K-invariant (K is the maximal compact subgroup of G corresponding to f) complex valued differential forms on X. On the other hand if t is the ortho-complement of gl in g, then t = n + n* so that the glmodule At = An (59 An* , and if superscript gl designates the space of invariants, one clearly has an isomorphism (1.1.1) The first step in dealing with the problem stated above is solving the algebraic problem of relating the gl-module H*(n) with H(g, gl). In case X is a symmetric space this is immediate since first of all n is commutative (see I, Proposition 8.2) and hence H*(n) = An. Since, in any case, the representation of gl on An* is contragredient to that on An, the irreducible components (An)
397
74
BERTRAM KOSTANT
H(C,d) and H(C,a) are in a natural way isomorphic. In fact if we put S = da + ad (the laplacian) then both groups are in canonical isomorphism with Ker S, a space of representative (harmonic) d-cocycles and a-cycles.
One also has a "Hodge decomposition" of C (see I, Proposition 2.1). Now introduce a Lie algebra structure on t = n + n* by retaining the given structures on nand n*, but making [n, n*] = O. (For one motivation of such a structure, see Proposition 6.14.1.) Let at E EndAt be the boundary operator for the corresponding chain complex C*(r). Now put C equal to the left side of (1.1.1). It is immediate that C is stable under at and if a = at IC then one easily has H(C, a)
= (HAn) ® H*(n*»)gl .
Analogously with the symmetric space case, the irreducible components HAnY.,. then define a basis h"', (J E WI, of H(C, a). Now let dE EndC be the operator on C corresponding to the coboundary operator on C(~, gl) under the mapping (1.1.1). Thus d is defined so that (1.1.1) induces an isomorphism (1.1.2) H(C, d) ----> H(g, gl) . Theorem 4.5 then asserts that d and a are disjoint. Thus this proves, for one thing, that H(C, d) and H(C, a) are isomorphic, and in fact (using Ker S, S = da + ad) establishes an isomorphism (1.1.3)
H(C, a)
---->
H(C,d) .
One therefore recovers, in a purely algebraic way, all the known results on the cohomology group H(X, C). That is
*
if p q if p = q (See I, §5.15) . But, more than this, disjointness implies that in each class s E H( C, d) there exists one and only one cocycle 8 E s such that a8 = 0; namely, that (harmonic) cocycle such that 8 E Ker S. Now let s"', (J E WI, be the basis of H(C, d) defined as the image of h", (J E WI, under the map (1.1.3), and let 8" E s'" be the harmonic representative. If oJ" is the K-invariant differential form corresponding to 8" under the map (1.1.1), it is obvious that s" corresponds to [w"'] under (1.1.2). Thus to HAn)'" we have associated a class [w,,] E H(X, C), and a distinguished representative differential form W".
For the proof of disjointness another result is needed. Theorem 4.3 asserts that H(n, 1m an)gl = 0 .
398
LIE ALGEBRA COHOMOLOGY
75
Here an E EndAn is the boundary operator of C*(n) and Iman C;;;;; An is an n-module with respect to the adjoint representation of n on An. The whole cochain complex C(n, 1m an) is easily seen to be a 9cmodule. The relation to the Schubert classes is established by Theorem 6.15. This asserts that up to scalar multiple, [(0""] equals x"". The exact scalar )..,fT will be determined in Part III. In the present paper, however, we give an integral formula for the scalar )..,fT. An essential role in the proof of Theorem 6.15 is the fact that as"" = O. Another application of this property is that if Y = G/ B is the generalized flag manifold and Y - X is the projection defined by inclusion B - U, then harmonic forms on Y, in our sense, go into harmonic forms on X. This is not true for the usual definition of harmonic forms. Other applications of the property as fT = 0 will be made in Part III. Theorem 5.6 gives an explicit formula for computing sfT for any a E WI. Thus, together with Theorem 6.15 and the knowledge of the scalar )..,fT, this formula constructs closed differential forms whose cohomology classes define the dual basis to the Schubert homology classes. 1.2. The definition of t, and the operators d, a and S in the paper are different from that indicated above. They are defined in more general terms. In fact in our definition, t has nothing to do with g. It is simply a complex Lie algebra with a real form tR which itself has an underlying complex structure. It also has a hermitian structure. However to prove the disjointness of d and a we have to assume that t is essentially like grin g. In point of fact we eventually assume it to be the space of complex 1-covectors at the origin of X. The Lie algebra structure on t is then motivated by the fact that tR is in a natural way (real) isomorphic with n. This isomorphism, incidentally, is independent of the choice of the compact form f of 9. In the introduction to Part I, it was remarked that the non-zero eigenvalues of the laplace L" of I, Theorem 5.7, will be needed in Part II. They are in fact used here in Theorem 5.6. However their main use will be in Part III.
2. A family of operators defined by a Lie algebra with a hermitian structure
2.1. We adopt the following conventions. Assume that V is a vector space over C, the complex numbers. Unless called real, a subspace of V will always mean a complex subspace. In case V is graded
399
76
BERTRAM KOSTANT
a subspace Vl ~ V will be said to be graded if it is graded by the intersections V! = VJ n Vl. In case V is, in addition, bi-graded
V = '" L...Jp,q
vp,q
,
the bi-grading will always be consistent with respect to the grading. That is
and a subspace V l ~ V will be said to be bi-graded if it is bi-graded by the intersection Vr q = Vp,q n Vl. In case V l and V 2 are graded (resp. bi-graded) vector spaces, a homomorphism A: V l ---> V 2 will be said to be of degree j (resp. of bi-degree if (s,
t»
for all i (resp. A: VI p q ---> Vr' Ht for all p, q). If V l and V 2 are complex vector spaces, a linear mapping A: V l ---> V2 , unless called real or R-linear, will always mean a complex linear mapping. For convenience we will write End V for End c V. 2.2. If a real vector space is complexified, one defines in a natural way an operation of conjugation on the complexification. If the real vector space is a real Lie algebra, then the complexification carries a Lie algebra structure. However, if the real Lie algebra itself carries a complex structure, then one knows that its complexification decomposes into a direct sum of two commuting ideals, each the conjugate of the other. We will be concerned here with a family of operators on an exterior algebra that arises from this situation when the real Lie algebra carries a positive definite real bilinear form; or, equivalently, when one of the ideals carries a hermitian positive definite inner product. No connection is assumed to exist between the Lie algebra structure and the bilinear form. Most significant for us is the operator St (see (2.8.2).). Let t be a finite dimensional complex vector space. For any U EAt let s(u) E EndAt be the operator of left exterior multiplication by u on At. Thus s(u)v = ul\v for any u,v EAt. A real linear form tR of t is a real subspace tR such that t = tR + itR is a real direct sum. Assume such a space is given once and for all. It is clear then that At = ARtR
+ iARtR
is a real direct sum. Besides being a real linear form of A t, it is clear
400
LIE ALGEBRA COHOMOLOGY
77
that ARtR is a real subalgebra of At. One introduces an operation of conjugation in A t by putting
u
+ iv =
u - iv
for any u, v E A RtR. Conjugation is a conjugate linear automorphism of order 2 and degree zero (with respect to the grading of At) of At. It ind uces a similar operation on End At where if A E End A t then A E End At is defined by Au=Au
where u E At is arbitrary. The fact that conjugation on At preserves multiplication may then be expressed by the relation (2.2.1)
s(u) = s(u)
for any u EAt. If p ~ t is a subset, then :P ~ t is defined by conjugating all the elements in p. It is obvious that :P is a subspace whenever p is a subspace.
2.3. Now assume that j E EndAt is an automorphism of degree zero (and hence determined by its restriction to t) such that (1) j" = -1, where 1 is the identity operator on At, and (2) tR is stable under j. It is clear then that t=a+a is a direct sum where (if i =
vi -1)
0=
{xEtijx
=
ix}.
By putting
A'Mt
= (APa)/\(Aqa) ,
we observe that one thus defines a bi-grading of At. It is also clear that
(2.3.1) 2.4. If V is a complex vector space, then as in Part I, {V} denotes a positive hermitian inner product on V and (V) is a bilinear form on V. About {V}, we always assume that for any j, g E V, {f, g} = {g,j} ,
and for any A. E C, A.{j, g}
= {:\,j,
g} = {j, :\g} .
We now assume that {a} is defined. We then observe that there is a
401
78
BERTRAM KOSTANT
unique {t} such that a1-
(2.4.1)
=a
(the orthogonal complement being taken in t); and for any x, yEt, {x, y}
(2.4.2)
= {y, x}
;
and {a} is the restriction of {t} to a. (We could have defined {t} by starting with a positive definite bilinear form on tR with respect to which j ItR is an orthogonal transformation.) But now {t} defines, in the usual way, a positive definite hermitian inner product {At} on At. That is, with respect to {At}, if p if p
(2.4.3)
*q =
q
REMARK 2.4. It is immediate from (2.4.2) and (2.4.3) that (2.4.2) holds more generally for all x, YEA t. Also observe that (2.4.1) generalizes to AMt and AP',q't are orthogonal
(2.4.4) if (p, q)
* (p', q').
2.5. Now define a bilinear form (At)a on At by putting (u, v)a = {u, v}
for any u, v E At. It is obvious that (At)a is non-singular and (by Remark 2.4) symmetric. We note also that (At)a agrees with {At} on ARtR and hence, in particular, is positive definite there. Now if A E EndAt, denote by AS E EndAt the transpose of A with respect to (At)a, and by A* E EndAt the adjoint of A with respect to {At}. PROPOSITION 2.5. The three operations A ----> A, AS and A* on EndAt are related in the following way: The operations commute with each other, and the composite of any two is the third. That is, these operations together with identity operation on End A t form a group, and this group is isomorphic to the Klein 4-group. PROOF. Since each is of order 2 it suffices only to show that
A* = AS for any A E EndAt. Let u, v E At. Then {Au, v} = (Au, V)B = (u, ABv)B = {u, A8V} = {u, ABv}. But {Au, v} = {u, A*v}. Now let 0 E EndAt be the automorphism of At defined by
402
LIE ALGEBRA COHOMOLOGY au
=
79
(-l)Ju
for any integer j and u E A Jt. Let A E EndAr. If we put At = aA5a ,
(2.5.1)
then At E End A t is the transpose of A with respect to the symmetric bilinear form (At) defined by (2.5.2)
(u, v) = (u, av)s = {u" av} .
REMARK 2.5. Since (At) is non-singular it defines an algebra isomorphism (2.5.3) of At onto its dual and if A E EndAt, then At corrasponds under (2.5.3) to the usual transpose of A formed on At'. We observe that by (2.3.1) and (2.4.4) (2.5.4)
(u, v)
=
(u, v)s
=0
for any u E A p,qt and v E AP',q't unless (p, q) = (q, p). One defines t(u) E EndAt for any u E At by (2.5.5)
t(u)
=
c(u)t ,
and we recall that (see Remark 2.5) t(x) is a derivation of degree -1 for x E t and that (2.5.6)
c(x)t(y)
+ t(y)c(x) = (x, y)l
for any x, yEt. Here 1 E EndAt is the identity operator. 2.6. Now assume that 0 is a (complex) Lie algebra. We then see that there exists a unique Lie algebra structure on t such that (2.6.1)
[x, y]
= [x, 17]
for all x, yEt (which implies that tR is a real Lie subalgebra of t) and (2.6.2)
[0, a]
=
0
so that 0 and a are ideals in t. (Equivalently we could have assumed that tR is a real Lie algebra such that j ItR commutes with the adjoint representation and hence defines a complex structure on tR and that t is the complexification of t R.) Let at E EndAt be the boundary operator of the chain complex C*(r). Thus a~ = 0, and ar(x/\y) = [x, y] for any x, yEt. It is then obvious from (2.6.1) that
403
80
BERTRAM KOSTANT
(2.6.3) Now since ar anti-commutes witn 0, we can define br E EndAt by putting t
(2.6.4)
br = -ar =
S
ax
and note that, by Remark 2.5, br corresponds under (2.4.4) to the coboundary operator of C *(t), and hence br is a derivation of degree 1. By (2.6.3), (2.6.4), and Proposition 2.5, one has (2.6.5) br = ar* , and hence a positive semi-definite hermitian operator Lr E End A t of degree zero is defined by putting
+ brar .
Lr = arb r
(2.6.6)
By (2.6.5) the operator Lr is the laplacian defined by at and (see I, Remark 2.3) its kernel is isomorphic to H*(t).
a; and hence
2.7. For any x E t, let n(x) E EndAt be defined by (2.7.1)
n(x) = c(x)ar
+ arc(x)
.
Then one knows that n: t
----->
EndAt
is the adjoint representation of t on At (see e.g. [5]) . Now let nt: t
----->
EndAt
be the representation defined by putting (2.7.2) for any x E t. Since nt(x) is clearly of degree zero, we can now define a new bracket relation [x, y]~ on t by putting (2.7.3) for any x, yEt. We recall (see I, § 3.6) that any linear mapping t -----> A 2t is the restriction of a unique derivation of degree 1 of At. Since [x, y]" is clearly alternating in x and y, we therefore observe that there exists a unique derivation d r E EndAt of degree 1 of At such that (2.7.4)
(drZ, x/\y) = -(z, [x, y],,)
for all x, y, Z E t. Now let er E EndAt be the derivation of degree 1 of At defined by
404
LIE ALGEBRA COHOMOLOGY (2.7.5)
dt
=
bt
81
+ Ct .
One now finds that Ct is given explicitly by PROPOSITION 2.7. Let Xi be any basis of t, and let y j be the basis of t defined so that (x., yj) = O.i. Then (2.7.6)
Ct
=
EAxi)n(y,) .
PROOF. Let et E EndAt be the operator given by the right side of (2.7.6). Since n(y.) is a derivation of degree zero, it follows immediately that et is a derivation of degree 1. Therefore it suffices only to prove that CtZ = etZ for all Z E t or that (2.7.7) for all x, yEt. But now by (2.6.4) and the definition of Ct, it follows that -cf(x 1\ y)
=
-ei(xl\y)
= (E,nt(Yi)c(X,»)Xl\y = nt(x)y - nt(y)x ,
nt(x)y - nt(y)x .
On the other hand q.e.d.
This proves (2.7.7). 2.8. One of our principal concerns in this paper will be with the operator St E EndAt (or rather its restriction, S, to the subspace C ~ At defined in § 3.9) of degree zero defined by (2.8.1) By (2.7.5) and (2.6.5) it follows that (2.8.2)
St = L t
+ Et
where (2.8.3) The operator E t is given explicitly by PROPOSITION 2.8. Let Xi and Yj be as in Proposition 2.7. Then Et
=
Etn(xi)n(Yj) .
PROOF. We have only to apply Proposition 2.7, the fact that n(y.) commutes with at (follows from (2.7.1», (2.7.1) itself, and of course (2.8.3). 2.9. In this section we give an example (to be used later) of a space t having the structure assumed in §§2.2-2.6. Proposition 2.9 gives an explicit expression for the operator d r for this case.
405
82
BERTRAM KOSTANT
We recall some of the notation and definitions of Part I. First of all, 9 denotes the complex semi-simple Lie algebra considered in Part I, {) is a fixed maximal solvable Lie subalgebra of g, and V is the family of all Lie algebras u such that {)~u~g.
The Cart an-Killing form on 9 is denoted by (g), and its extension to Ag is denoted by (Ag). Given any subspace $ ~ g, the polar (see I, § 4.1) of !3 in 9 with respect to (g) is denoted by !30 • If u E V then one knows (I, Proposition 5.3) that n = U O is the maximal nilpotent ideal in u. A real compact form f of 9 has been fixed once and for all. With respect to t, a *-operation and a positive definite hermitian inner product {Ag} have been defined on Ag. We recall that if gl = un u* for u E V then
+n
u = gl
is a Lie algebra semi-direct sum (with n as ideal) and
(2.9.1)
+ n + n*
9 = gl
is an orthogonal (linear) direct sum. Now put
(2.9.2)
tx = n
+ n*
.
In this section t will equal t x. We will now put structure on Xx so that the assumptions §§ 2.2-2.6 are satisfied. The subscript X is used to distinguish this example from the general case. A real linear form Xx.R is defined by putting
(2.9.3)
tX.R
= Xx n f
,
and conjugation in A tx is defined relative to this real form in the same way as in § 2.2. Recalling the definition of the *-operation in A 9 (see I, § 3.3) note that it follows, for any U E A xx, that
(2.9.4) where Ox E EndAtx is defined in a similar way to O. Obviously then n*
= it.
Now let ix E EndAtx be the automorphism defined so that and - i on n*. Then, here n
ix
= i on n
=0.
Now let {n} be the restriction of {Ag} to n. It follows that if {A xx} is defined as in §2.4, then {Axx} is just the restriction of {Ag} to Axx.
406
LIE ALGEBRA COHOMOLOGY
83
Furthermore recalling (2.5.2), (2.9.4), and (I, (3.3.1», it also follows that (A xx) is just the restriction of (Ag) to A xx. We recall (see §2.6) that Xx is made into a Lie algebra by a Lie algebra structure on n = Q. For the latter we use the structure induced on n as a Lie subalgebra of g. Thus if x, y E X and x = e1 + f1 and y = e2 + f2 where ei , fi E n, i = 1, 2, then denoting the bracket in Xx with the subscript X one has (2.9.5) where the brackets on the right are the usual brackets in g. Let .p be any subspace in g. For any x, y E g let [x, yh be the component of [x, y] in.p according to the decomposition g = .p + .pl.. We now observe LEMMA 2.9.1. Let x, y E Xx. T hen if n is given as in § 2.7 where x = Xx and .p denotes either n or tt one has if x, y E 'p nt(x)y = {O [x, y]P if x E l3 and y E 'p • PROOF. Let
Z
E Xx. By definition
(2.9.6) If x, Y E .p then this expression vanishes since 'p is an ideal in Xx and (A xx) is totally singular on.p. This proves that nt(x)y = O. In case x E:P and y E.p, then nt(x)y can have no component in l3 since (2.9.6) vanishes if Z E.p (because [:P, .p]x = 0). To find its component in .p, it is enough to let Z E j). But then [x, z]x = [x, z]. Hence nt(x)y = [x, y]p by the invariance of the Cartan-Killing form. q.e.d. As an immediate consequence of Lemma 2.9.1 we now observe that if x E:P and y E.p, where .p is either n or tt, then
(2.9.7) Now consider the bracket relation (see (2.7.3» [x, y]" for the case X= XX. LEMMA 2.9.2. Let x, y E Xx. Then [x, y]"
= [x, Y]rx •
PROOF. By definition [x, y]" = [x, Y]x + nt(x)y - nt(y)x. Writing x and y as in (2.9.5) it follows from Lemma 2.9.1 and (2.9.7) that nt(x)y - nt(y)x =
[ell h]r x + L{;., e2]rx' Adding this to (2.9.5) proves the lemma. q.e.d. We recall some further notation of Part I. By definition 'Y E EndAg is the boundary operator for the chain complex GAg) and c = -'Yt so that C E EndAg corresponds to the coboundary operator of G*(g) under the
407
84
BERTRAM KOSTANT
isomorphism Ag -> Ag' induced by (Ag). Now let
r: Ag -> Arx be the orthogonal projection of Ag onto A rx. Lemma 2.9.2 may then be expressed by the relation (2.9.8) for any x, y
r'Y(X/\Y) E
= [x,
y)'"
rx. One may then prove
PROPOSITION 2.9. In the case r d rx
= rx one has = rc
on Arx .
PROOF. This follows immediately from the definition of drx (see (2.7.4», (2.9.8) and the fact that rc is a derivation of degree 1 in Arx (see I, Ftemark 3.8). q.e.d. 2.10. We continue in this section with the example of §2.9. As in Part I, let () denote the adjoint representation of 9 on Ag. Now observe that Arx is stable under (}(x) for all x E gl' Thus Arx becomes a gl-module with respect to the representation (2.10.1)
flrx: gl -> EndAtx
where flrx is the sub-representation of () Igl defined by A t x. The following proposition asserts that the given structure on A rx is "invariant" under flr x' PROPOSITION 2.10. Let x E gl then one has (a) [flrx(x), ix] = 0 , (b) [fltx(x), arx] = 0 , (c) fl r/ x )* = flrx(x*) , (d) flrx(x) = -flrx(x*) .
PROOF. Clearly (a) and (b) follow from the fact that gl lies in the nomalizer of both nand n. The relation (c) follows immediately from I, (3.9.7). Finally (d) follows obviously from the fact that f is a real Lie subalgebra of g. q.e.d. 3. Lie algebra cohomology defined by the adjoint representation
3.1. The operator d r is more easily understood by writing it as a sum d r = d; + d;' where d; and d;' are respectively of bi-degree (1,0) and (0,1). Proposition 3.1 asserts that such a decomposition (necessarily unique)
408
LIE ALGEBRA COHOMOLOGY
85
exists. It will be shown later that d~' (and also d;) is essentially the coboundary operator associated with the co chain complex defined by Aa, regarded as an a module (using the adjoint representation). We begin with PROPOSITION 3.1. Let e E EndAt be any derivation of degree 1 such that (a) e = e, and (b) (3.1.1) e: a - a/\t ,
then e may be uniquely written (3.1.2)
e = e'
+ e" ,
where e', e" E EndAt are, respectively, of bi-degree (1, 0) and (0, 1). Furthermore e' and e" are derivations (of degree 1) and e" = ?
(3.1.3)
PROOF. Since e = e it follows from (3.1.1) that
e: a-t/\o; therefore since e is a derivation,
e: A p,qt -
A P+1,qt
+A
p,q+1 t
for any p, q. But this clearly implies the existence of a unique decomposition (3.1.2). Since e is a derivation, it is obvious that e' and e" are derivations of degree 1. Now if e1 E EndAt is of bi-degree (1,0) it follows immediately from (2.3.1) that e1 is of bi-degree (0, 1). Since e = e = e' + e" it follows therefore from the uniqueness of (3.1.2), that (3.1.3) holds. q.e.d. We now observe that btl Cr and d r satisfy the conditions of Proposition 3.1. LEMMA 3.1. Let e = br , cr or dr. Then e = e and (3.1.4) e:a-a/\t. PROOF. By Proposition 2.5, (2.6.3) and (2.6.4), it follows that lir = br. Now if we conjugate the equation (2.7.1), it follows from (2.2.1) and (2.6.3) that (3.1.5)
1r(x)
=
1r(x)
for any x E t. But now it follows easily from (2.4.2) that (3.1.6)
(x, y)
for all x, yEt. Conjugating from (3.1.5) and (3.1.6) that
Cr ,
=
(x, jj)
as given by (2.7.6), it follows therefore
er = ero
409
86
BERTRAM KOSTANT
This of course implies that dr = dr. We have only to prove (3.1.4). The orthocomplement of a /\ t in A 2t is A20 and the orthocomplement of a in t is o. Since ar = to prove (3.1.4) for e = bn it is enough therefore, to observe that ar maps A 20 into a. But this is obvious since a is a Lie subalgebra of t. If e = Cr. then (3.1.4) is obvious from Proposition 2.7 since a is an ideal in t. This of course proves that (3.1.4) holds also if e = dr. q.e.d.
b:,
c;,
3.2. By Lemma 3.1 we can define derivations of degree 1, b;, d~ and and d;' in accordance with Proposition 3.1. The former are of bidegree (1,0) and the latter are bi-degree (0,1). By (2.7.5) it is obvious that
b;', c;',
(3.2.1)
d't = b'r
+ c'r
and (3.2.2)
d"r
=
b"r
+ c"t '
PROPOSITION 3.2. Let fi be a basis of a, and let {fj be the basis of a defined so that (fi' {fj) = Oij. Then (3.2.3)
c;' = EtC({fi)n-(fi)
and
(3.2.4) PROOF. Let e" and e' be respectively the operators given by the right side of (3.2.3) and (3.2.4). Now recall Proposition 2.7. By letting Yk be the basis fi' {fj of t, it follows from Proposition 2.7 that cr = e' + e". But since a and a are ideals in t it is obvious that e' and elf are respectively of bi-degree (1,0) and (0, 1). Hence e' = c~ and elf = 0;'. q.e.d. Now define operators E EndAt by putting
a;, a;'
(3.2.5)
a; =
-(b~'Y
and
a"t = -(b')t t
•
By (2.5.4) it is obvious that a; and a~' are respectively of bi-degree ( -1,0) and (0, -1). Furthermore by (2.6.4) one has (3.2.6)
at =
a; + a;' . a;
By the uniqueness of this decomposition, note that and given by the property that for any U E Aa, v E Aa one has (3.2.7)
410
ar'
may be
LIE ALGEBRA COHOMOLOGY
87
Note then that (3.2.8) 3.3. Now define L~, E: and S: and also L~', E;', and S:' in the same way as Lx, En and Sx except that all the operators are primed in one case and double primed in the other case. Thus (3.3.1)
S'x =L'x +E'x
and (3.3.2)
S" = L'X X
+ E'x '
and we note that all the operators in (3.3.1) and (3.3.2) are of bi-degree (0,0). Matters are considerably simplified by PROPOSITION 3.3.1. One has 1
E'x = E" x = -Ex. 2 PROOF. Let g E a. By (3.2.7) it is obvious that multiplication (c(g)8) by g. Hence one has
c(g)a;
(3.3.3)
+ a;c(g) =
a; commutes with right
0.
But now obviously for all x E t since 7r(x) is of bi-degree (0, 0) and it commutes with by (3.2.3) one has (3.3.4)
ax'
Thus
c"a' x x + a'xc" x = 0 •
Hence (3.3.5)
E;' = c~'ar
+ al;' = 'E 7r (g;)7r(fi) j
by (3.2.3). Similarly, by conjugating (3.3.4) one obtains (3.3.6)
c~a;'
+ ar'c; = 0
since
f1r
(3.3.7) by Proposition 2.5. Thus
=
a·'r
E; = c;ax + arcr and hence
(3.3.8) by (3.2.4). But also since cr
= c; + c;',
411
one has
Er
= Ex' + E;'.
On the
BERTRAM KOSTANT
88
other hand comparing (3.3.5) and (3.3.8), it follows that x(f.) commutes with XUii). q.e.d. Next one has
E; = E;' since
LEMMA 3.3. Let er be either br , Cr or dr. Then a'relfr + e"a' r r = 0
and a"e' r r
+ e'a" r r = 0 •
PROOF. In case er = Cr these are just the relations (3.3.4) and (3.3.6). If er = br the first relation is obvious from (3.2.7) and (3.2.8). Conjugating gives the second relation. In case er = d t the result follows from (3.2.1) and (3.2.2). q.e.d. As an immediate corollary one obtains PROPOSITION 3.3.2. One has
L r = L'r
+ L"r
and S r = S r '+ SIt r in particular Lr and Sr are operators of bi-degree (0, 0). PROOF. Immediate from Lemma 3.3 and the definitions of the concerned operators. q.e.d.
80
REMARK 3.3. Let ...Arc End A t be the set of the 21 operators ar, bt , Cn dn and Sr' primed, and double primed. We observe here that ...Ar lies in the smallest subalgebra of End A t that contains d;' and is closed under the operations A -> At, A. Lro E r ,
a;
3.4. Let aa E EndAa be the boundary operator for the chain complex C*(a) (recall that a is a Lie subalgebra of t) and let ba, E EndAa', the negative transpose of ar' be the coboundary operator of the cochain complex C(a). If the vector space V is an a-module with respect to a representation Xo:
a -> End V
then we may take V Q9 A a' as the underlying vector space for the cochain complex C(a, V). (For notational convenience here the order of the factors is the reverse of that given in I, §3.1.) Furthermore we recall the coboundary operator bo for this complex may be written (3.4.1)
412
89
LIE ALGEBRA COHOMOLOGY
where (3.4.2) and (3.4.3)
g:
where fi is a basis of a, and is the dual basis of a'. See T, §3.1 and § 3.12. We are particularly interested in the case where V = A a and (henceforth shall assume that)
no: a -> EndAa is the adjoint representation of a on Aa. Now let
YJa': Aa -> Aa' be the linear map defined so that = (u, v) for any u E Aa, v E A a. By (2.5.4) and the non-singularity of (At) it is obvious that YJa' is an isomorphism. Therefore the map
YJ: At -> Aa ® Aa'
(3.4.3) given by (3.4.4)
YJ(u/\ v)
= u ® YJa'V
for U E A a, v E A a is also a (linear) isomorphism. But obviously (3.4.3) sets up an algebra isomorphism EndAt -> End(Aa ® Aa') .
(3.4.5)
Now bo, b1 and b2 are elements of the right side of (3.4.5). What they correspond to in End A t is given in PROPOSITION 3.4. Let Va be the restriction of V to isomorphism (3.4.5) one has
a; -> aa ® 1 (b) b;' -> Va ® b~ = (va (c) c;' -> (va ® 1)b2
Aa. Under the
(a)
® 1)b1
and consequently (3.4.6) PROOF.
Clearly (a) is obvious from (3.2.7), and the fact that
413
90
BERTRAM KOSTANT
aa
= atl Aa
.
Also (b) is similarly obvious from (3.2.8) and the definition of bt • But now (c) is also true by (3.2.3) and (3.4.3) and the obvious fact that under(3.4.5) (3.4.7) for any f Ea. (Note that oa comes in since S(gi) means left multiplication by y•. ) Finally one obtains (3.4.6) by (3.2.2) and (3.4.1). q.e.d. REMARK 3.4. Since oa ® 1 clearly commutes with bo, we note that as a. consequence of Proposition 3.4 one has (d~')2 = O. Similarly (d~)2 = 0 by (3.1.3). However d~ 0, in general, as the example in § 2.9 shows. See Proposition 2.9.
*"
3.5. Let {Aa} be the restriction of {Ax} to Aa. Now v conjugate linear isomorphism Aa ----> Aa' where
---->
v' defines a.
{u, v} = for all u, v E Aa. Define {Aa'} so that {u, v} = {v', u'}
(3.5.1)
for all u, v E An. Since, clearly v' = 1Ja'ov, note by Remark 2.4, that 1Jar is an isometry. Now let {Aa ® Aa'} be defined by {Aa} and {Aa'}. Since a and a are orthogonal in x, it then follows easily from the definition of {Ax} that '1f also is an isometry. Now define laplacians La E EndAa and La' E End A a" by putting
La = La' =
+ a1 aa , ba,b~, + bci,b a,
aaa~
where the ad joints are defined relative to {Aa} and {Aa'} respectively. LEMMA 3.5. Under the map (3.4.5) one has (a) L~ ----> La ® 1 (b) L;' ----> 1 ® La' .
Furthermore (c) La' is the transpose of La. PROOF. We first observe that (3.5.2)
b't = (a')* and b"t = (a")* t t·
Indeed this follows from (2.6.5) and the obvious fact that (a;)* and (a;')* are respectively of bi-degree (1, 0) and (0, 1). (See (2.4.4».
414
91
LIE ALGEBRA COHOMOLOGY
But now since 1} is an isometry it follows, (a) and (b) of Proposition 3.4, that under (3.4.5) a~'
(3.5.3)
--->
oa ® M,
and b~--->a~
® 1.
But then (a) follows immediately from the definition of from these definitions it is clear that (3.5.4)
and
L~'.
But
L" r = (L')' t'
One obtains (c) then easily from the definition of REMARK
L~
1}a"
q.e.d.
3.5. Combining (a) of Proposition 3.4 and (3.5.3) note that
(3.5.5) under the map (3.4.5). 3.6. Assume that C is a finite dimensional bi-graded vector space. If e E End C is of degree ±1 where e2 = 0, and Cl ~ C is a bi-graded subspace stable under e, we put Z(ClJ e)
= Ker e [C
B(ClJ e)
= Ime[Cl
l
and and the homology space H(ClJ e)
=
Z(ClJ e)jB(ClJ e) .
Each of the three spaces is graded in the usual way. Define H:M(ClJ e) = {aEH(Cl , e)[ancp,q"* O}.
In case H(Cl , e) is a direct sum of these subspaces, note that they define a bigrading of H(ClJ e). This is clear since H:M(CII e) ~ HPH(Cl , e), and hence (3.6.1) Furthermore (3.6.1) is a direct sum. When this is the case, we will simply say that H(Cl , e) is bi-graded. REMARK 3.6. In case e is of bi-degree (p, q) where of course p + q = deg e it is obvious that H(Cl , e) is bi-graded. Furthermore Z(ClJ e) and B(Cl , e) are also naturally bi-graded.
3.7. Now assume that dE End C is of degree 1, and aE End C is of degree
415
92
BERTRAM KOSTANT
1 and d 2 = fj2 = O. Recall I, §2.1. We say that d and a are disjoint in case dax = 0 implies x = 0, and ady = 0 implies y = 0 for all x, y E C. In such a case if S is the laplacian (3.7.1)
S = da
+ ad,
then one has the "Hodge decomposition" (see I, Proposition 2.1) (3.7.2)
+ 1m a + 1m d
C = Ker S
.
Morever (3.7.2) is a direct sum and one has (3.7.3)
Ker S
= Ker a n Ker d
and (3.7.4)
1m S = 1m a + 1m d
so that (3.7.5)
+ ImS
C = KerS
is a direct sum. The elements of Ker S are called harmonic (with respect to the disjoint pair d and a). Now obviously since (3.7.6)
+ 1m d + 1m a ,
Ker d = Ker S Ker a = Ker S
the mapping of cycle to homology class induces isomorphisms (3.7.7) (3.7.8)
'td,S: Ker S 'ta.s: Ker S
~
H(C, d) ~ H( C, a) .
Furthermore since S is of degree 0, it is clear that 'td.S and 'ta.s are of degree O. Thus (3.7.9)
'ta.a: H(C, d)
~
H(C, a)
is an isomorphism of degree 0 where
'ta,a = 'ta.s o('t I1.S)-l . REMARK 3.7. The map 'ta.a may be given independently of S by defining 'ta.a(a) = b in case a n b =1= O. This is well defined and is an isomorphism. PROPOSITION 3.7. Assume that S is of bi-degree (0, 0). Then Ker S, H(C, d) and H(C, a) are all bi-graded and furthermore the maps 'ta.s, 'td,S and 'ta,a are of bi-degree (0, 0).
416
LIE ALGEBRA COHOMOLOGY
93
PROOF. Obviously KerSis bi-graded. Let MM be the image of (KerS)M under the map V'a.s. Since the latter is an isomorphism of degree 0 it is obvious that the MM defines a bigrading of H(C, d). It suffices therefore to show that MM = HM(C, d). Obviously M'M ~ HM(C, d) by definition of the latter. Now let U E Z(C, d). Write U = U 1 + U 2 where U 1 E Ker S, U 2 E 1m S. Now clearly U 2 E B(C, d) by (3.7.6). But if U E Cp·q, then both U H U 2 E CM since CM is stable under S. But then U 1 E (Ker S)M and hence HM(C, d) ~ MM. Thus H(C, d) is bi-graded and V'a.s is of bi-degree (0,0). Similarly V'a.s, and hence V'a.a is of bi-degree (0, 0). q.e.d. Now assume bE EndC, b2 = 0, is another operator of degree 1 such that b and a are disjoint. Let L
=
ba
+ 8b
be the laplacian defined by band 8. Now since Ker Land Ker S are both isomorphic to H(C, 8), they are naturally isomorphic to each other. In fact (3.7.10)
V'L.8: Ker S
-->
Ker L
is the isomorphism, and V'L.S is of degree 0 where
V'L.S = (V'a.L)-l°V'a.s . Let (3.7.11)
P: C ---Ker L
by the projection operator of C onto Ker L that vanishes on 1m L. Then the mapping V'L.8 is given by LEMMA 3.7. One has
V'L.8 = PIKer S . Furthermore V'L.8 is oj bi-degree (0, 0) in case both Land S are oj bidegree (0, 0). PROOF. Let x E Ker S. Put y = Px E Ker L. To show that y = V'L.8X, it suffices only to show that x and y define the same class in H(C, 8). But x - Y E 1m L n Ker a = 1m a. Hence x and y do define the same class in
H(C, a). If Land S are of bi-degree (0, 0), then V'L.S is of bi-degree (0, 0) since P is clearly of bi-degree (0, 0) and Ker S is bigraded. q.e.d. 3.8 Our final structure assumptions is that v is some Lie algebra, At is a v-module and certain conditions are satisfied. More specifically assume that v is a Lie algebra with a *-operation (x --> x* is a conjugate linear endomorphism of order 2 and [x, y]* = [y*, x*] for any x, y E tJ) and that
417
94
BERTRAM KOSTANT
(3.8.1) is a representation such that fJt(x) is a derivation of degree 0 for all x E V. Furthermore it is assumed that the v-module structure on A't defined by (3.8.1) is compatible with the given structure on A't to the extent that (3.8.2), (3.8.3), (3.8.4) and (3.8.5) below are satisfied:
=0
(3.8.2)
[fJt(x), j]
(3.8.3)
[fJt(x), ar] = 0
(3.8.4) (3.8.5)
fJt(x)* fJr(x)
= fJb;*) = -fJix*)
for all x E V. PROPOSITION 3.8. The representation fJr is completely reducible. Moreover for any x E v, fJX<x) is ofbi-degree (0,0) and all the (21) operators in Jlr (see Remark 3.3) commute with fJix). PROOF. Condition (3.8.4) insures complete reducibility. In fact by (3.8.4) one may use orthocomplements for v-stable complements. Obviously (3.8.2) implies that fJr(x) is of bi-degree (0, 0). Now let y E 't, and put z = fJX<x)y. Now by (2.7.1) and (3.8.3), it is clear that (3.8.6)
[fJr(x), x(y)]
=
x(z) •
But now by (3.8.4), (3.8.5), Proposition 2.5, and (2.5.1), it follows that (3.8.7)
(fJr(x»)'
=
-fJt(x) •
But then taking the negative transpose of (3.8.6) one obtains (3.8.8)
[fJt(x), x'(y)] = x'(z) •
It follows immediately then from (3.8.3), (3.8.7) and (3.8.8), and the definition of dn that
(3.8.9) But since fJt(x) is of bi-degree (0, 0), it follows from (3.8.9) and (3.8.3) that fJix) commutes with a~ and d~'. But then recalling Remark 3.3, it follows from (3.8.4), (3.8.5) and (3.8.7) that fJt(x) commutes with all the operators in Jlt • q.e.d. 3.9. If a vector space F is a v-module, we denote by F'D the subspace of all elements in F that are annihilated by all x E V. Once and for all we put
C = (A't)'D .
418
LIE ALGEBRA COHOMOLOGY
95
Now it is obvious from Proposition 3.8, that Cis bi-graded. Let {C} and (C) be respectively the restrictions of {At} and (At) to C. Since C is clearly stable under conjugation, it is immediate that (C) is non-singular. Thus the operations A -> At, A *, A and AS are definable in End C in the same way as in EndAt. Furthermore if EnduAt is the subalgebra of all v-endomorphisms on A t, and (3.9.1)
EnduAt -> End C
is the homomorphism defined by A -> A Ic, it is obvious that the operations above commute with (3.9.1). But now Jir ~ EnduAt by Proposition 3.8. We adopt the following notational convention. Recall that all the operators in Jir were notated with the subscript t. We will now denote its image in End C, under the map (3.9.1), by dropping the subscript t. Thus if er E Ji r , then by definition e E End C is defined by (3.9.2)
e
= eriC.
Since the orthocomplement of C in A t is clearly stable under en note that for all er E Jir (3.9.3) 1m e = 1m er n C . Concerning questions of homology we adopt the notation of § 3.6. Our immediate concern is with H(C, 8). 3.10. Assume that the finite dimensional vector space F is a v-module with respect to a completely reducible representation fJ of v. Now let e be a b-endomorphism of F such that e2 = 0 (a b-differential operator) and let H(F) be the homology space defined with respect to e. Then one knows that fJ induces a representation ~: v -> End H(F) ,
and that ~ is also completely reducible. More generally let Fl ~ F2 ~ F be b-submodules of F that are stable under e. Then e induces a differential operator on FN FP and a b-differential operator on the v-module F2/ Fl. The natural map of the former into the latter clearly induces a map (3.10.1) LEMMA 3.10. The map (3.10.1) is an isomorphism. PROOF. Since fJ is completely reducible it is clear the map inducing (3.10.1) defines an isomorphism F 2°/FlU -> (F2! Fl)U • Put D = F 2 /Fl • We must show that H(Db) = H(D)u.
419
96
BERTRAM KOSTANT
By the complete reducibility of f3, we can write uniquely D = Db + Do where the latter is a v-module direct sum. But then it is obvious that the differential operator leaves both summands stable. But then clearly H(Db) = (H(D»b. q.e.d. Now if u E An and v E Acr, it is obvious that ar(ul\v) = arul\v
+ oul\arv
.
If F = At, f3 = f3r and e = ar, it follows therefore by the Kiinneth formula (and an obvious identification) that
(3.10.2) so that ~r is a representation (3.10.3) In fact if f3o, and f3a are the sub-representations of f3r defined by Aa, it is clear (up to an obvious equivalence) that (3.10.4)
f3r = f3o, @ 1
An and
+ 1 @ f3a
and
~r = ~o, @ 1 + 1 @ ~a
.
As an application of Lemma 3.10 we obtain PROPOSITION 3.10. This is a natural identification (3.10.5)
Furthermore H(C, a) is bi-graded and (3.10.6) PROOF. Since C = Fb we have only to identify by means of the isomorphism (3.10.1) where F2 = F, Fl = 0 and H(F) is given by (3.10.2). (For the complete reducibility of f3r see Proposition 3.8.) The relation (3.10.6) is obvious by definition of HP,q(C, a). Since H(C, a) is clearly a direct sum of these spaces it is, by definition, bi-graded. q.e.d. Since b = a* (see (2.6.5» it follows that b and a are disjoint, and L = ba + ab is the corresponding laplacian. But since L is of bi-degree (0, 0) (see Proposition 3.3.2) it follows that Ker L is bi-graded and that furthermore the isomorphism (3.10.6) defined in § 3.6, is, by Proposition 3.7, of bi-degree (0, 0).
420
97
LIE ALGEBRA COHOMOLOGY 4. The disjointness of d and 8
4.1. The structure assumptions in §§2 and 3 put no restrictions whatsoever on the nature of the Lie algebra a (the Lie algebra '0 could just have well been zero). It will be one of the main consequences of this section that, under certain conditions, d" and 8" are disjoint. When, furthermore, it is assumed that d 2 = 0, it will be shown that d and 8 are disjoint as well. But then this establishes a natural isomorphism ('ta.a, see (3.7.9» between H(C, d) and (H*(a) ® H*(a»b = H(C, 8), (see Proposition 3.10). Since in our applications H(C, d) is the complex cohomology of the algebraic homogeneous space GI U on one hand and '0 = g10 and a is isomorphic to n on the other hand, this establishes the relation between the cohomology of GI U and H*(n), as a gcmodule mentioned in the introduction of Part I. Using the results of Part I, which determined the structure of H*(n) as a gl-module, one obtains, for example, the "strange" relation (1.1.1) of Part I. The easy half of the disjointness of d" and 8" (Proposition 4.1.1) will be a consequence of
Condition AI. The Lie algebra a is nilpotent (4.1.1). If.
Aa =
(4.1.2)
~
£..JI=O
V.•
is an orthogonal finite direct sum decomposition of
Aa let
(4.1.3) so that its orthocomplement V(tl in
Aa is given by
(4.1.4) LEMMA 4.1.1. If Condition Al is satisfied, there exists a decomposition (4.1.2) such that for all k (a) for any fE a n(f): V k
->
V(tl ,
(b) V k is stable under the representation (3t of '0. PROOF. Let Vo = O. Assume Vi' i ~ j, has been defined, and VUl is
stable under n(f). Put
E j = {space spanned by n(f)u, all f
E
a, u
E
V(]l}
so that E j ~ V(]l. But, if V(]l =1= 0, then E j =1= VGl since a is nilpotent. Put Vi+1 equal to the orthocomplement of E j in V Gl. This defines a decomposition (4.1.2) and (a) of Lemma 4.1.1 is satisfied. By Proposition 3.8 one
421
98
BERTRAM KOSTANT
knows that An is stable under (Jr' Now, however, by (3.8.6) the subspaces V(t) are also o-submodules. Taking orthocomplements and using (3.8.4), it follows finally that the V k are themselves stable under (Jr. q.e.d. By (b) in Lemma 4.1.1 and (3.10.4), it is clear that V k ® (A a) is stable under (Jr' If we then define (4.1.5) it is obvious that (4.1.6) is an orthogonal direct sum decomposition of C. Let C(l<) and C(t) be defined as in (4.1.3) and (4.1.4) with C replacing F. As a consequence of Lemma 4.1.1 we obtain LEMMA 4.1.2. If Condition Al is satisfied, then for all k (a) e": Ck ----> C(t)
whereas (b) Ck is stable under both a" and b" so that (c) b" = d" on Ck mod C(t) . PROOF. The relation (a) follows immediately from (a) of Lemma 4.1.1 and (3.2.3), whereas (b) follows immediately from the form of a" and b" given by (3.5.3) and (b) of Proposition 3.4. (Vk is obviously stable under 0a). Finally (c) follows from the relation d" = b" + e". q.e.d The following proposition is clearly stronger than "half" of the disjointness of d" and a".
PROPOSITION 4.1.1. Assume Condition Al is satisfied. Let u EAt. Then d"a"u E Im a" implies a"u = O. PROOF. Assume a"u -=1= O. Let k be minimum with the property that a"u ¢ C~). Thus a"u can be written a"u = U 1 + U 2 where 0 -=1= U 1 E Ck and U 2 E C(t). Thus by (a) of Lemma 4.1.2 and the relation d" = b" + e", one has
d"a"u
= b"u1 + Us
,
where by Lemma 4.1.2, one has b"u1 E Ck and by the relation b" = (a")* (see (3.5.2»
{b"ur, u}
C(t). But b"u 1 -=1= 0 since
= {ur, a"u} = {ul, u
I}
-=1=
0 since
But then
{d"a"u, b"u1}
Us E
= {b"ur, b"u -=1=
O.
422
l}
U I -=1=
0.
LIE ALGEBRA COHOMOLOGY
99
This however contradicts the fact that d"a"u E 1m a" since 1m a" and 1m bIT are necessarily orthogonal to each other (because (a")2 = 0). q.e.d. REMARK 4.1. By conjugation it is immediate that Proposition 4.1.1 holds when prime replaces double prime throughout the statement. As a corollary we obtain PROPOSITION 4.1.2. Assume Condition At is satisfied. Let u EAt. Then if S'u = 0, one has d'a'u = a'd'u = 0 and in fact a'u = O. PROOF. If S'u = 0, then d'a'u = -a'd'u. But this implies that d'a'u E 1m a'. By Remark 4.1, one has then that a'u = O. Thus d'a'u = -a'd'u = O. q.e.d. 4.2. In order to get the other half of the disjointness of d" and a" we need Condition A 2. L' = L", which will be seen to have a number of other consequences. Heuristically speaking Condition A2 would seem to imply that C is "small" relative to At. An immediate consequence of the condition is PROPOSITION 4.2. Assume Condition A2 is satisfied. Then (4.2.1)
L' = L" = 2-L ' 2
and (4.2.2)
S' = S" = 2-s . 2
PROOF. The relation (4.2.1) follows immediately from Proposition 3.3.2, and (4.2.2) follows from Proposition 3.3.2 together with Proposition 3.3.1 which asserts that in any case E' = E" = (1/2)E. q.e.d. 4.3. Consider the cohomology space H(a, V) defined by an a-module V. Now if Vt ~ V 2 ~ V are two a-submodules then in the notation of §3.10 it is easy to see that (4.3.1) where F = V® Aa', Fi = Vi ® Aa', i = 1,2 and e is the coboundary operator of C(a, V). Now as in § 3.4, consider the special case where V = Aa and the latter is an a-module by the adjoint representation no. In this case e = boo Now let f3 a,: b -> EndAa' be the representation contragredient to f3 a• Recalling the definition of
423
100
BERTRAM KOSTANT
the isomorphism 1)a': A Q-+ Aa', if then follows immediately from (3.8.7) and Proposition 3.8 that 1)a' is in fact a b-module isomorphism. Furthermore if Aa ® Aa' is regarded as a b-module with respect to the representation fJ defined by (4.3.2)
fJ
=
fJa
® 1 + 1 ® fJa'
,
then one has LEMMA 4.3 The map 1):
At -+ Aa ® Aa'
defined in § 3.4 is a b-module isomorphism. Moreover the co boundary operator bo of the complex C(a, Aa) is a b-endomorphism of Aa ® Aa'. PROOF. Since 1)a' is a b-module isomorphism, the first statement follows by comparing (4.3.2) and (3.10.4). The second statement follows from (3.4.6) and the fact that d~' is a b-endomorphism of At. (See Proposition 3.8.) q.e.d. Now if Vl ~ Aa is an a-submodule and a b-submodule(with respect tofJa) then not only is C(a, Vl) defined, but also the underlying space Fl of C(a, Vl) is a b-module. Moreover since bo is a b-endomorphism, H(Fl ) = H(a, Vl) is also a b-module. We are now particularly interested in the case where V l = 1m aa. Since aa is both an a- and a b-endomorphism of Aa, it is clear that 1m aa and Ker a.:t are each both an a- and a b-submodule of Aa. The following is our first theorem. THEOREM 4.3. Assume that Conditirms Al and A2 are satisfied. (See § 4.1 and 4.2). Then (4.3.3)
H(a, 1m act)\)
=0.
PROOF. We first observe that if Vl ~ V2 ~ A a are both a- and b-submodules then where Fl and F2 are as in (4.3.1), the isomorphism (3.10.1), together with (4.3.1) becomes an isomorphism (4.3.4) where, it may be understood, the underlying space V2/ Vl ® Aa' of C(a, V 2/ Vl) is a b-module by taking the tensor product of fJo.' with the representation on V 2 / V l induced by fJ a• Next since a is nilpotent (Condition A l ) there exists an orthogonal direct sum decomposition 1m aa = E,=o Vi such that, in the notation of (4.1.3) and (4.1.4), (a) and (b) of Lemma
424
101
LIE ALGEBRA COHOMOLOGY
4.1.1 hold. The proof is the exact same as Lemma 4.1.1 with 1m (Ja replacing Aa. But then the subspaces Vlt) of 1m (Ja are both a- and v-submodules of A a. Hence if Fk = Vlt) @ A a', one has by (4.3.4) isomorphisms (4.3.5) But now the Fib are a filtration of Fob there is an isomorphism
= (1m (Ja @ Aa,)b,
and by (4.3.4)
(4.3.6) To prove the theorem, it suffices therefore to prove that H(FoD) = O. But to prove this, it suffices to show that the El terms of the spectral sequence ·defined by the filtration above are identically zero. But the El terms are made up of the left side of (4.3.5) for all k. But then by the isomorphism (4.3.5) it suffices to prove that the right side of (4.3.5) is zero for all k. But now V~-lJ Vlt) is a trivial a-module by (a) of Lemma 4.1.1, and as an v-module it is isomorphic to V k • Thus one has an isomorphism (4.3.7)
H(a, VILIJ Vtt))D -
(Vk @ H(a))b .
It suffices therefore to prove that the right side of (4.3.7) vanishes for all k. Now clearly as subspaces of Aa, Ker ba,
= Ker La' + 1m ba,
is a v-module direct sum decomposition where La' (see §3.5) is the laplacian defined by ba, and its adjoint. But then Ker La' is equivalent to H(a) as a v-module. It therefore suffices to show that (Vk @ Ker La,)b = 0 for all k. Assume not. Then there exists a k and U E (Vk @ Ker La,)b such that U =1= O. But now (4.3.8)
(L a @ l)u =1= 0
since Ker(L a@l) = Ker L a@ Aa' and V k On the other hand obviously
n Ker La = 0, because
V k ~ Im(Ja.
(4.3.9) Now let v EAt be defined by putting v = r;-lu. Then clearly v E C, and by Lemma 3.5, (4.3.8) and (4.3.9), one has L'u =1= 0 and L"u = O. This contradicts the equality L' = L" of Condition A 2. q.e.d. REMARK 4.3. For the proof of Theorem 4.3, note that Condition A2 could have weakened to the assumption that only Ker L' = Ker L". Now in accordance with the notation of § 3.6, H(C1 , e) is defined if C1 is a bi-graded subspace of C that is stable under e where e E EndC is of degree ± 1 and e2 = O.
425
BERTRAM KOSTANT
102
We now wish to consider the case where e = d". Since a' anti-commutes withd" (Lemma 3.3) and a' isofbi-degree( -1,0), it follows that H(lma',d") is defined. As a corollary to Theorem 4.3 (and in fact equivalent to it) we have COROLLARY 4.3. Assume that conditions Al and A2 hold. Then H(lm a', d") = 0 .
(4.3.10)
PROOF. Now by Theorem 4.3 and (4.3.6) one has that H(FoO) = 0 where Fo = ImaaQ9 Aa', where we recall the coboundary operator on Foo is bolFoo. But now by (a) of Proposition 3.4, the space Fo corresponds to 1m a~ ~ At, and hence Foo corresponds to 1m a~ n C under the map r;. But now by (3.9.3) this implies that (4.3.11)
is an isomorphism. But d~' corresponds to (oQ Q91)b o under the map (3.4.5) according to (3.4.6). But obviously H(FoO) also vanishes in case (oa Q91)b o is substituted for boo Hence by (4.3.11) one has H(lm a', d") = O. q.e.d. 4.4. If e E End C is an in § 3.6 and C1 ~ C2~ Care bi-graded subspaces stable under e, then let H(C 2{C1 , e) be the homology group defined by C2{C1 and the differential operator on C2{C1 induced bye. Now H(Ker a', d") is defined as well as H(lm a', d"). Let (4.4.1)
H(Ker a', d")
---+
H(C, d")
be the map (of bi-degree (0, 0), see Remark 3.6) induced by the injection Ker a' ----> C. LEMMA 4.4.1. Assume that Conditions Al and A2 hold. Then the map (4.4.1) is an isomorphism. PROOF. It is clear that a': C{Ker a'
---->
1m a'
is an isomorphism. But since a' anti-commutes with d". it follows immediately from Corollary 4.3 that H(C{Ker a', d") = 0 .
(4.4.2)
Now consider the exact sequence
o
---->
Ker a' ----> C ----> CfKer a' ----> 0 .
This induces, on the level of homology, an exact sequence H(C{Ker a', d")
---->
H(Ker a', d")
---->
H(C, d")
---->
H(C{Ker a', d") ,
where the map in the middle is just (4.4.1). Since the ends vanish by
426
LIE ALGEBRA COHOMOLOGY (4.4.2), this proves the lemma. Next we need
103
q.e.d.
LEMMA 4.4.2. Assume Conditions Al and A2 hold. Then (4.4.3)
Z(Ker a', d") n 1m L = B(Ker a', d") .
PROOF. Let u lie in the left side of (4.4.3). Then since 1m L = 1m L' (by Proposition 4.2), one has
u E Ker a' n 1m L' = 1m a' . But d"u = O. Thus u E Z(lm a', d"). But then by Theorem 4.3, one has u = d"v where v E 1m a' ~ Ker a'. Hence u E B(Ker a', d") so that the left side of (4.4.3) is contained in the right side. Now conversely assume that u E B(Ker a', d"). We have only to prove that u E 1m L. In order to prove this we will first prove that e"; Ker a' -> 1m a'
(4.4.4)
.
To prove this, by (3.2.3) and (3.9.3), we have only to show that
n(f); Ker a~ -> 1m a~
(4.4.5)
for all f Ea. But (4.4.5) is obvious from the relation
.s(f)a;
(4.4.6)
+ a;.s(f) =
n(f)
which one obtains from (2.7.1) and the conjugate of (3.3.3). Hence (4.4.4) is established. But now we assert that d"; Ker a' -> 1m L .
(4.4.7) Indeed d" = b" (4.4.8)
+ e".
Hence by (4.4.4) d"; Ker a'
->
1m b"
+ 1m a' .
But 1mb" ~ ImL" and lma' ~ ImL'. But ImL" = ImL' = ImL, by Proposition 4.2. Thus (4.4.8) implies (4.4.7). But now by assumption u = d"v where v E Ker a'. Thus u Elm L by (4.4.7). q.e.d. Finally we can prove THEOREM 4.4. Assume that Conditions Al and A2 hold. Then the operators d" and a" are dis}oint. Similarly d' and a' are dis}oint. PROOF. Using conjugation, it is enough to prove only that d" and a" are disjoint. Recalling Proposition 4.1.1, we have only to prove that a"d"u = 0 implies d"u = 0 for any u E C. Indeed assume that a"d"u = O. Put v = d"u. Then obviously a"v = d"v = O. Hence S"v = O. But S" = S' by Proposition 4.2. Thus S'v = O. But then by Proposition 4.1.2,
427
104
BERTRAM KOSTANT
one has a'v = O. That is v E Kera'. That is v E Z(Kera', d"). But now v = d"u. That is v defines the zero class in H(C, d"). But then by Lemma 4.4.1, it must have already defined the zero class in H(Ker a', d"). Thus v E B(Kera', d"). But then byLemma4.4.2,itfollowsthatv E ImL = ImL" (by Proposition 4.2). But now recall that a"v = 0 (by assumption). Thus
v E 1m L" n Ker a" = 1m a"
.
That is, there exists w such that v = a"w. But d"v = O. Hence d"v = d"a"w = O. But now by Proposition 4.1.1, this implies that a"w = v = d"u = O. q.e.d. We may substitute a for a' and a" in Theorem 4.4. COROLLARY 4.4.1. Assume that Conditions Al and A2 hold, then d" and a are disjoint. 8imilarly d' and a are disjoint. PROOF. We first observe that by Lemma 3.3 and Proposition 4.2, (1/2)8 is given by either
~8 = d'a
(4.4.9)
2
+ ad'
or
= d"a + ad" .
(4.4.10)
On the other hand by Theorem 4.4, d" and a" are disjoint and hence (1/2)8, their anti-commutator (see Proposition 4.2) is the laplacian they define. A similar statement holds for d' and a'. But then by property (3.7.4) of the laplacian (4.4.11)
1m d'
+ 1m d" + 1m a' + 1m a" =
1m 8
and hence (4.4.12)
1m d
+ 1m a ~ 1m 8
.
Now let u E C, and assume d"au = o. Put v = au so that by (4.4.12) v E 1m 8. But now d"v = av = O. Hence v E Ker 8 by (4.4.10). But since (1/2)8 is the laplacian defined by d" and a", one has Ker 8
n 1m 8 = 0 .
Thus v = o. A similar argument, using 1m d" ~ 1m 8, (see 4.4.11) shows that ad"u = 0 implies d"u = O. Thus d" and a are disjoint. Conjugation shows that d' and a are likewise disjoint. q.e.d. REMARK 4.4.1. To obtain the disjointness of d" and a we used only two properties, (1) the anti-commutator of d" and a is a non-zero multiple of 8, and
428
LIE ALGEBRA COHOMOLOGY
105
(2) 1m d" and 1m a are contained in 1m S. One consequence of Corollary 4.4.1 is that the homology groups H(C, d") and H(C, a) are isomorphic.
COROLLARY 4.4.2. Assume Conditirms Al and A2 hold so that by Corollary 4.4.1, d" and a are disjoint. Then the map 'ta.d" defined as in §3.7 (see (3.7.9» is an isomorphism (4.4.13)
'ta.d": H(C, d")
-+
(H*(a) 0 HAaW
of bi-degree (0, 0). PROOF. We have only to use Propositions 3.7, 3.10 and the fact that S is of bi-degree (0, 0) (see Proposition 3.3.2). q.e.d. REMARK 4.4.2. It is immediate from (4.3.4) and (3.4.6) that H(C, d") is isomorphic to H(a, Aa)'o. It follows therefore by Corollary (4.4.13) that if Conditions Al and A2 hold, there is an isomorphism (4.4.14) However this fact can be obtained directly from Theorem 4.3. An argument similar to the one used in the proof of Lemma 4.4.1 shows directly that (4.4.15) H(a, Aa)'o = (H*(a) 0 H(a»)'o , where the tensor product on the right is a v-module with respect to the representation fJ a 01 + 10 fJ a,. 4.5. In the cases that interest us d 2 = 0, and we will be particularly concerned with H(C, d). We therefore consider the case where Condition Aa: d 2 = 0 is satisfied. One finds that not only is d disjoint from a but also from a' and a". THEOREM 4.5. Assume that Conditions All A2 and Aa hold. (See §§4.1 and 4.2.) Then where ae denotes a, a' or a", one has that d and ae are disjoint and the corresponding laplacian is given by (4.5.1)
dae
+ aed = {~/2)S
if ae = a' or a" if ae = a .
Furthermore H(C, d) is bi-graded, and if the map 'ta.d is defined as in §3.7 (see (3.7.9» then 'ta.d is an isomorphism (4.5.2)
'ta.d: H(C, d)
-+
(H*(a) 0 HAaW
of bi-degree (0, 0). PROOF. The proof of disjointness follows exactly as in Corollary 4.4.1. See Remark 4.4.1. With regard to the latter, (1) is satisfied by (4.5.1),
429
106
BERTRAM KOSTANT
which follows immediately from Lemma 3.3. Moreover (2) is satisfied by (4.4.12). Since S is of bi-degree (0, 0) by Proposition 3.3.2, the last statement follows from Propositions 3.7 and 3.10. q.e.d. 4.6. We shall assume in this section that Conditions AlJ A2 and A3 are satisfied. Note then that the notation of §3.7 (concerning laplacians and the "Hodge decomposition") apply here. We can therefore use freely here all the relations of that section. Now if e = 8 or d and u E Z(C, e) we will let [u]. E H(C, e) be the class determined by u. Now let s E H(C, d). Among the various cocycles which represent s, we will be particularly interested in the unique such (harmonic) cocycle s lying in Ker S. In the notation of §3.7 (4.6.1)
'Vra.s(s)
=s.
Largely because 8s = 0 the harmonic representative will later be seen to enjoy properties making it useful for a number of applications. In our considerations the problem will thereupon arise, how does one construct s "knowing" s. By "knowing" we really mean knowing its image (4.6.2)
'Vra.is)
=
h
in (H*(a) ® H*(a»b under the isomorphism (4.5.2). It should be pointed out that the right side of (4.5.2) will be explicitly known by the results in Part I. But now, because of the relatively simple nature of L, knowing h, one immediately determines the unique cycle 'Vra.~(h)
(4.6.3)
=h
in Ker L such that [hla = h. Thus the problem effectively is, givenh E Ker L, find the unique element s E Ker S such that (4.6.4) (see (3.7.10) for the definition of the isomorphism 'VrL.S). A formula for computing s in terms of h will be given by Theorem 4.6. Now by Lemma 3.7, one has h = Ps where P, we now observe (using the fact that L is a hermitian operator) is the orthogonal projection of Con Ker L. Thus if u E C is defined by (4.6.5) then u (4.6.6)
s=h+u, E
1m L. That is u = Qs
430
LIE ALGEBRA COHOMOLOGY
107
if Q E End C is defined by (4.6.7)
P
+ Q=
1
(and hence is the orthogonal projection of C on 1m L). Now let Lo E End C be the (inverse of L on 1m L) the unique operator which vanishes on Ker L and satisfies (4.6.8) Finally where we recall S = L
+ E,
(4.6.9)
= -LoE.
R
put
REMARK 4.6. The operator E has been given by (3.3.8), and in our applications Lo is known by the results of Part 1. Thus R is given explicitly in our applications (see § 5.6). We now observe LEMMA 4.6. The operator R is nilpotent. PROOF. Let the subspaces Ck ~ C be given by (4.1.5) and Lemma 4.1.1. Because of (4.1.6), it suffices to show that for all k (4.6.10) using the notation of §4.1. But now by Lemma 4.1.2 (b), it is obvious that C k is stable under L" = (1/2)L. Thus since (4.1.6) is an orthogonal decomposition, it follows that Ck is also stable under Lo. Thus to prove (4.6.10) one need only show that E maps Ck into C(tl. But this is obvious from Lemma 4.1.1 and the formula (3.3.8) after one recalls that E" = (1/2)E (see Proposition 3.3.1). q.e.d. Now let p be a positive integer such that RP+l = o. We then have THEOREM 4.6. IJs,hand u are as in (4.6.5) where s = 'Vrs.ih),h E Ker L. (and hence s E Ker S), then (4.6.11)
s
=
(1 - R)-lh
=
E~=oRjh
and (4.6.12)
u = E~=lRjh .
PROOF. Let sland u 1 be defined respectively by the right side of (4.6.11) and (4.6.12). Thus SI = h + u 1 • It is obvious from (4.6.9) and (4.6.12) that U 1 E 1m Lo = 1m L. Thus PSI = h. To prove s = S1> it suffices therefore to prove only that SI E Ker S since 'VrL.8 = PI Ker S is an isomorpism. Next observe that (4.6.13)
[R,8]
=0.
Indeed 8 commutes with Land S since the latter are laplacians defined using 8. Thus 8 commutes with E and also Lo. This proves (4.6.13). But
431
108
BERTRAM KOSTANT
now h E Ker L ~ Ker a. But then by (4.6.13) one clearly has also 81 E Ker a. But by the properties of the laplacian, one knows that 8 maps Ker a into 1m a. Thus to prove that 881 = 0, it suffices to show that L oS8 1
=0
since Lo is non-singular on 1m a c 1m L. But since 8
= L +E
LoS = Q - R = (1- R) - P But then by (4.6.11) L oS8 1 8 1 = 8. q.e.d.
= h - P8 = O. Thus I
by (4.6.7) . 81 E
Ker 8 and hence
5. The case where a = n 5.1. It will be convenient at this time to recall and collect some further (see § 2.9) notation from Part 1. The Cartan subalgebra ~ is fixed once and for all by putting ~ = 0 n 0* (see §2.9 and 1, §§5.1-4). The discrete group Z ~ ~' denotes the spaces of integral linear forms on l) and DI ~ Z is the set of dominant integral linear forms relative to gl (and 0). We recall that G is the group corresponding to g, and G1 is the subgroup corresponding to gl' All the representations of gi considered here induce representations of G1 • If F is any such gcmodule and ~ E Z we denote by F" the subspace of all vectors in F which transform according to the irreducible representation vi of gl' It is clear of course that F" = F"" for any a E WI' the Weyl group of G1 • Now A ~ Z is the set of roots, A+ is the set of positive roots (A+ = A(O» and A(n) ~ A+ is the subset corresponding to n. Now to each a E W, the Weyl group of G, one associates a subset
The subset WI
~
W is then defined by
(5.1.4)
For any }, one lets (5.1.5)
WI(})
=
{a E WI I
and w 1 ( } ) is the number of elements in WI(}). Now if
For convenience we put (this differs by sign from the notation of Part 1)
432
109
LIE ALGEBRA COHOMOLOGY
(5.1.7) for any a E W, and note that (5.1.8)
~CT
= g - ag
where, as usual, (5.1.9) 5.2. In §2.9 we considered the example I = Ix (the orthocomplement of gl in g) satisfying the structure assumptions of §§2.2-2.6. The final structure assumption, that of § 3.8, is satisfied by putting b = gl and (3t = (3tx· To see this we have only to compare Proposition 2.10 with (3.8.2-5). For this case we put Cx = C and the restriction of any operator erx E Jlrx to Cx will be denoted by ex. (See (2.9.2).) In effect we will be dealing with this example throughout the remainder of the paper. However, rather than assume that t = tx (in fact t will eventually be the space of complex 1-covectors at the origin of the algebraic homogeneous manifold X = G/ U) we shall assume (5.2.1) and (5.2.2) is an algebra gcisomorphism of bi-degree (0,0) which preserves all the structure of §§2.2-2.6. That is, Ar is an isometry, it commutes with conjugation, it makes jx correspond to j, and atx correspond to ar . We observe of course that At maps n onto a. Using the results of Part I we can now be more explicit about H(C, a). First of all, the structure of H*(a) as gcmodule (with respect to /3a) is given by I, Corollary 8.1. If ~ E - D then (5.2.3)
if and only if
and furthermore for any a E W l , (5.2.4)
(H*(a»)
Moreover (5.2.5)
433
~ = ~CT
for some a E W l
110
BERTRAM KOSTANT
is a direct sum for any j. The structure of H*(a) as a gl-module is an easy consequence of LEMMA 5.2. Assume that u EAt is a weight vector for the weight \, E Z. Then u is a weight vector for - \'. PROOF. Let x E~. Then f3 t (x)u = -f3t (x*)u by (3.8.5). But the latter vector is - <x*, \,>u. On the other hand, it is easy to see that - <x*, \,> = - <x, \,>. But this proves the lemma. q.e.d. Combining Lemma 5.2 and (3.8.5) one easily obtains the relation (5.2.6)
for any ~ E Z. Now since = morphism, h --> ii,
at an
conjugation in At induces a conjugate linear iso-
(5.2.7)
But clearly by (5.2.6) one has (5.2.8)
for any ~ E Z. It follows therefore that (5.2.3-5) hold if -~ replaces ~. Let (J E W 1 and put (5.2.9)
Hu
=
(H*(a)
a replaces a and
.
We then obtain THEOREM 5.2. One has
Hp,q(C, a) = 0
ifp"* q,
and (5.2.10)
is a direct sum. Furthermore dimHu = 1 80
that dim HP.P(C, a) = w1(p) .
PROOF. The space Hu is one dimensional by (5.2.4) and Schur's lemma. The same argument shows that (H*(a)
in case
(J
=1= T
)gl
= 0
since (5.2.5) is a direct sum. We have now only to apply
434
LIE ALGEBRA COHOMOLOGY
111
(5.2.5) and Proposition 3.10. q.e.d. Up to a scalar multiple, Theorem 5.2 defines a basis of H(C, 8) indexed by WI. That is if (5.2.11) then the elements (normalized later) h", a E Wr, form a basis of H(C,8) and if h" E Ker L is defined so that (5.2.12)
""'I' a,L(h") = h"
,
then the elements h", a E Wr, form a basis of Ker L. 5.3. Now let (5.3.1) and for any
~,,,y E
A put
(5.3.2) Using (5.2.6) note that these subspaces generate At. We draw further upon results of Part I to prove THEOREM 5.3. For any r.. E Z let b(r..) be the scalar defined by
b(r..) = 1g 12 - 1g - r..12 .
(5.3.3)
Then
if~,,,y E
A, one has
(5.3.4)
L~
= 2-b(~) 2
and L~'
= 2.b("y) 2
on the subspace t(~, "y) of At. PROOF. Consider the subspace (Aa')-'" of Aa', the underlying space of the cochain complex C(a). Since -"y E DI we may apply I, Theorem 5.7 (for the case r.. = 0) to assert that
La'
1 = -b("y) 2
on this subspace. But then by Lemma 3.5 (c), one has that
La = 2-b(~) 2
on the subspace (Aa)" of Aa. The result then follows from (a) and (b) of Lemma 3.5. q.e.d. Now for any ~ E A, put We then obtain
435
112
BERTRAM KOSTANT
COROLLARY 5.3.1. One has that (5.3.5)
is an orthogonal direct sum decomposition. Furthermore
L' = L"
(5.3.6)
and (since L = L'
+ L")
(5.3.7)
L
on the subspace Co of C, where
b(~)
=
b(~)
is given by (5.3.3).
PROOF. Using (3.8.4) it is obvious that the subspaces t(~, ofr) define an orthogonal direct sum decomposition of At. But t(~, ofr )'h = 0 if ofr ~ by Schur's lemma. We have now only to apply Theorem 5.3 and Proposition 3.3.2. q.e.d. We have almost proved
*
COROLLARY 5.3.2. Conditions AH A2 and A3 hold (See §§4.1, 4.2 and
4.5). That is (1) a is nilpotent. (2) L' = L", and (3) d 2 = O. PROOF. First of all (1) is obvious since a is isomorphic to n. Second, (2) is just the relation (5.3.6). To prove (3) it suffices, using the isomorphism (5.2.2), to show that d~ = O. But now by definition, d x is restriction of drx to Cx. But then by Proposition 2.9, d x is the restriction of c to Cx where clearly (5.3.8)
Cx
= {u E Ag I8(x)u = t(x)u = 0 for all x E gl} .
(Here c is given by I, (3.8.2) and is not to be confused with Cr IC.) But now it is obvious (and well known in the theory of relative Lie algebra cohomology) that Cx is stable under c. Indeed this follows easily from the relation (5.3.9)
t(x)c
+ ct(x) = 8(x)
(see I, (3.9.5». But since c = 0, this implies d~ = o. q.e.d. We can now use all the results of § 4. Concerning the cohomology of n, we obtain 2
COROLLARY 5.3.3. One has
H(n, 1m (7t1)91 = 0 .
Furthermore
436
LIE ALGEBRA COHOMOLOGY
113
that
so
ifp,* q ifp = q. PROOF. The first statement follows from Theorem 4.3. The second from Remark 4.4.2, Proposition 3.10, and Theorem 5.2. q.e.d. More important for us are the consequences concerning the cohomology group H(C, d). COROLLARY 5.3.4. One has
ifp '* q ifp = q. PROOF. We have only to apply Theorems 4.5 and 5.2. q.e.d. REMARK 5.3. It is obvious from the proof of Corollary 5.3.2 that H(C, d) is isomorphic to the relative Lie algebra cohomology group H(g, 91). Since one knows the latter is isomorphic to the cohomology group H(X, C), where X is the algebraic homogeneous space GI U, one sees that Corollary 5.3.4 recovers the results of Borel, Bott, and Chevalley on the structure of the cohomology groups (not rings) of such spaces. Note that the method here used only linear algebra. Furthermore it explains the identity I, (1.1.1) whose "strangeness" motivated this paper. 5.4. Now using other results from Part I, we can be more explicit about the basis h of Ker L (see (5.2.11». rT
THEOREM 5.4. For any a E Wi one has Moreover if ~ E A then (see (5.3.3» (5.4.5) so
b(~)
°
= if and only if ~ =
~cr
~rT
for some a E Wi
that (by Corollary 5.3.1)
(5.4.7)
is an orthogonal direct sum. Futhermore (5.4.8)
and (5.4.9)
is a generator. PROOF. By definition of A, it is obvious that
437
= g - ag E A (see 5.3.1).
BERTRAM KOSTANT
114 (5.4.10)
is a direct sum of non-zero subspaces. But La' = (1/2)b(~) on the summand (Aa')-< by I, Theorem 5.7 (for the case:'\' = 0) since -~ E Dr. To obtain the first statement, (5.4.5), and (5.4.7) we have then only to apply I, Corollary 5.7; I, Theorem 5.14; and Corollary 5.3.1. The relation (5.4.8) arises from Schur's lemma and the fact that (Aa')-
0' E
WI, and let
U;
be an orthonormal basis 01
(Aa)
(5.4.11)
h rT = arTEtu;I\U i
•
PROOF. By I, Corollary 5.7 and I, Theorem 5.14 (see second to last statement) one knows that (5.4.12) where (5.4.13)
n(O')
= number of elements in
Hence by (2.5.2), (5.4.14) But now by (3.8.7), it follows from (5.2.6) that up to sign the Uj transform under (3r like the dual basis to the u i • Since G
Iv = -I-v
(5.5.2)
for any 't E A(tx). Now we may assume that there is given a lexicographical ordering in ql, the "real" part of q' (see I, § 5.1) so that the elements of A+ are positive. For each subset 'l" C A(tx ) we then let (5.5.3)
Iw
=
Iv 1 1\Iv2 1\ ... I\lv k
438
'
LIE ALGEBRA COHOMOLOGY where'l' = {Vi}, i = 1,2, ••.. , k and Vi>
115
VHI'
REMARK 5.5. It is obvious from the definition that the Iw: form an orthonormal basis of weight vectors of A I. (In fact Iw: is a weight vector for the weight <'l'> E Z). For the case I = Ix, we write ew: for Iw: and note that this possibly differs in sign from the definition given in Part 1. In particular the relation I, (5.11.1) is not necessarily valid. If <1;> ~ A(n) is a subset {g:>} we will put -<1;> = {-g:>}. One important advantage of our definitions is the easily verified relation (5.5.4) for any subset <1;> ~ A(n). The relation (5.5.4) makes sense since elements of even degree commute with each other. Note also that if <1;> ~ A(n) has n elements then _
(5.5.5)
n(n+l)
I", = (-1)-2-1_",
.
This follows easily from (5.5.2). Let a E Wl. By the last statement in I, Corollary 8.1, one knows I"'a- E (An)
ha-=ra-+va-
where
(5.5.9) PROOF. This follows immediately from Corollary 5.4, (5.5.5) and (5.5.6) by choosing (in Corollary 5.4) U l =/", and aU appropriately. q.e.d. Now recall (see I, § 5.3) that m is the maximal nilpotent Lie algebra bO so that
439
116
BERTRAM KOSTANT
(5.5.10) and m1
= m n 91' The following lemma will be needed later on.
LEMMA 5.5. The hyperplane (An)'! in (An)<
(5.5.11)
fJa(x*)f
for all x E mI' This implies by (3.8.4) that fJa(x)u E (An)'! for all u E (An)<
Va.is
The class s
by (4.6.6). Note also that (5.6.4) since 1m L n Ker iJ = 1m iJ and u
440
117
LIE ALGEBRA COHOMOLOGY
for all (J E W1. Note then in particular the elements sO", (J E W\ form an orthogonal basis of C. Case 2. In this example we take up the case associated with the generalized flag manifold. That is, we assume n = n1. Hence gl = 1) and W 1 = W. In this case (An)~O" = 0 for all (J E W since (An)'O" is one dimensional (because 1) is commutative). Thus, by Proposition 5.5, vO" = 0 and hence (5.6.6) 1
)n(O"l
= ( -. 27r1,
f~ I\f-~ U
0"
.
Now to compute sO" we will use the formula given by Theorem 4.6. To be of any practical use we will need to be explicit about R. We first observe that (5.6.7) Indeed this follows from (3.3.8), the relation (f",.!-v) = o",.v for cp, ofr E A+ (see I (5.1.2» and Proposition 3.3.1. Next we observe that for this case the subspace C, ~ C for any [; E A, may be explicitly given by (5.6.8)
C, = {space spanned by all f'Pl\f-'P1
=[; where
cI>, 'II ~ A+} .
But then by Corollary 5.3.1, Theorem 5.4 and (4.6.8) on C, 0" for any (5.6.9)
(J
(J
E
W
on C, where [; =I: [;0" for all
(J
E
W.
THEOREM 5.6. Assume we are in the case where n = n1. Thenfor any W the element sO" E Ker S may be given by the formula
E
(5.6.10)
where R = -LoE and Lo, E are explicitly given by (5.6.7) and (5.6.9) and p is such that Rp+1 = o. PROOF. We have just to apply (5.6.6) and Theorem 4.6. q.e.d. REMARK 5.6. Since the adjoint representation 7r is known in all cases (one just needs the constants of structure of n) formula (5.6.10) establishes a computable process for obtaining the sO". For its practical use, note that whenever Lo is needed in the formula it is always applied to elements in 1m L. Thus, with regard to the two
441
BERTRAM KOSTANT
118
possibilities of (5.6.9) only the second occurs. This is clear since, by induction, ERJrrT Elm 8 for all j. 5.7. Now by definition n = UO where u E CU. In this section we will show that S and the elements srT have. certain "invariance" properties as we vary the choice of u. In particular it will be shown that a knowledge of srT for the example of Case 2 § 5.6, determines the srT for the general case. Hence the formula (5.6.10) can be used in all cases. Now as usual, adding the subscript X to all appropriate notation changes the meaning only in so far as tx is referred to instead of t. Thus for example (5.6.3) becomes s~ = M + u~ when t = t x • Now tx is anyone of 21 possible cases depending on the choice u E CU. We will henceforth write t y to denote the special case when n = m. Furthermore we will use the subscript Y instead of X in all appropriate notation when reference is made to this case. Now obviously (5.7.1) ~
Furthermore since f:)
gI' one necessarily has
(5.7.2) LEMMA 5.7. Let
(J
E WI then
py(sn = h y . PROOF. Let 'Y E EndAg be the boundary operator of C*(g). Then by (3.2.7) one clearly has (5.7.3) for any U E A n and v E A n. Since 'Y is independent of the choice of u E CU, one therefore has (5.7.4) But now s~ = Hence by (5.6.4)
h~
U~
+ u~ and
E Ker 8;'
s~, h~
E Ker 8~ by (5.3.6) and Theorem 4.5.
n 1m Lx = 1m 8;' .
But then by (5.7.4) and (5.7.2) U~
Thus py(un (5.7.5)
E 1m 8~
~
1m L y
•
= 0, and hence py(sn = Py(M) .
442
119
LIE ALGEBRA COHOMOLOGY
Now if u E (An)
by (5.7.4). But now if v E (An)~rr, then vEIma~y
(5.7.6) by Lemma 5.5, since
f3n(X)u
= 7ry(x)u
for all u E (An)
(by 4.4.5». But then by Proposition 5.5, (5.7.6) and (3.2.7) v~ E 1m a;y. But in fact we assert that (5.7.7)
v~
h~
=
r~
+ v~
where
E Ima~ .
Indeed by (3.9.3) it suffices only to see thatv~ E Cy • NowM E Cy by (5.7.2). But r~ E Cy also by (5.6.8). Thus v~ E Cy , and hence (5.7.7) is proved. But since 1m a~ ~ 1m L y, this proves that Py(v~) = 0, and hence
py(sn
= Py(rn .
But obviously r~ = r~ and r~ = hf by (5.6.6). On the other hand hf since hf E Ker L y by definition. q.e.d. We can now prove THEOREM
Py(h~) =
5.7. The space Cx is stable under Sy and moreover
(5.7.8)
Furthermore for any (5.7.9) PROOF.
(J
E
WI, s~=s~.
By (4.5.1)
(5.7.10) On the other hand, as in the proof of Corollary 5.3.6, (5.7.11) where C E EndAg is defined as in Part 1. Now the expressions for d x and a~ given by (5.7.11) and (5.7.3) are in terms of c and 'Y, operators which
443
120
BERTRAM KOSTANT
do not depend on the choice U E cU. This together with (5.7.10) immediately implies the first statement of the theorem and also (5.7.8). Now let a E WI. Consider the formula (5.6.10) given for s~. Since
~(m)
+ ~(n)
C;;;;; ~(n)
(see I, Proposition 5.4) it follows from (5.6.10) that (5.7.13)
s~ E
Arx .
We must next show that s~ E Cx. That is, it must be shown that s~ is a gcinvariant. But now s~ E Ker Sy and hence, from the properties of the laplacian, it follows therefore that s~ E Ker d y • But then by (5.7.11) this implies s~
(5.7.14)
E Ker c .
Now we recall, from relative Lie algebra cohomology theory, that Kercn Arx = Cx . Indeed (5.7.15) follows from (5.3.8) and (5.3.9). But then (5.7.13) and (5.7.14). Hence
(5.7.15)
s~
E Cx by
s~EKerSx
by (5.7.8). We can therefore write (5.7.16)
s~
=
ETEWIaTs~ .
But now Py(sH = h~ by Lemma 3.7. Hence if we apply P y to both sides it follows from Lemma 5.7 that
But since the h~ are a basis of Ker L y , it follows that aT = 0 for aeT = 1. Hence s~ = s~ from (5.7.16). q.e.d.
1:
=1= a
and
REMARK 5.7. One consequence of Theorem 5.7 is that formula (5.6.10) established for the case n = m may be used in the general case to compute the seT (for a E WI). In particular seT = heT of Case 1 may be computed this way. It should also be noted that the in variance properties of S and seT indicated by Theorem 5.7, are not shared by L, E, or heT. In particular h~ =1= h'f for a E WI. In this connection Rx =1= R yICx so that Theorem 4.6 yields a formula other than (5.6.10) for the computation of seT in the general case. However (5.6.10) may be preferred since one always knows hY( = r'f) but notM.
444
LIE ALGEBRA COHOMOLOGY
121
5.8. In this and the next section we determine ser for some simple types of a. If cP E a, we recall that 1:rp E W denotes the reflection defined by cpo Also IT ~ a+ denotes the set of simple positive roots. LEMMA 5.8. Let a E W. Then n(a) = 1, (i.e., a E W(I» if and only if a = 1:", for some a E IT. PROOF. If a E IT then clearly n(1:",) = 1, since (5.8.1) (see I, § 5.9). Now if a E W, then by I, Proposition 3.10, the set
IT(u) = IT n a(u)
as one easily deduces from I, Proposition 5.4, one has PROPOSITION 5.8. The subset WI(l) ~ WI is given by (5.8.3) so that (5.8.4)
dim H1.l(C, d)
= number of elements in IT(u) .
PROOF. The relation (5.8.3) follows immediately from Lemma 5.8, (5.8.1), (5.8.2), and the definition of WI(l). The relation (5.8.4) then follows from Corollary 5.3.4. q.e.d. If cp E a+ and a E IT we recall (see I, (5.4.3» that n",(cp) is the a-coefficient of cp relative to the basis IT of 1)'. THEOREM 5.8. If a E WI and n(a) for some a E U(u), then
= 1, so that (by Lemma 5.8) a =
1:",
(5.8.5) PROOF. Let x E 1). Using the formula given by I, Lemma 3.11, for End A9 (and letting the Zi run through the root vectors) it follows easily that CE
(5.8.6)
445
122
BERTRAM KOSTANT
We next observe that (5.8.7) Indeed first of all by (5.8.6) one obviously has e(x) E C~l. Put Then by (5.7.11) dyu = O. But also 8;u = 0 since obviously
= e(x).
for any v E Cl.J and any } .
8'v = 0
(5.8.8)
U
Thus e(x) = U E Ker Sy by Theorem 4.5. But now dim(Ker Sy)1.1 = l by (5.8.4). Since the right side of (5.8.7) is also an l-dimensional subspace, this proves (5.8.7). Now by (5.8.1) and (5.6.6) one has (5.8.9) Also by (5.4.7) one then has (5.8.10) and hence (5.8.11) But now 8~ is characterized, as an element of Ker Sy, by the property that Py8~ = M. It follows therefore from (5.8.6-11) (excepting (5.8.8)) that 8~ = c(x) where, for any fJ E II, '0
(5.8.12)
<x, fJ> =
\~ 27C~
if fJ
*- a
if fJ
=a .
But if x has the property indicated by (5.8.12) then clearly for any cP E ~+ . This together with (5.7.9) and the fact that cP E ~+, n",(cp) cp E ~(n) (see I, Proposition 5.4) proves (5.8.5). q.e.d.
*-
0, implies
5.8. The above proof together with Corollary 5.3.4 gives new information about c-cocycles in A tv. In fact if U E A l·Jty, then REMARK
eu
=
0 implies
u = 0 {U = ex,
if}
*- 1
x E 1) if} = 1 .
Indeed eu = 0 implies U E Cy by (5.7.15). But then U E KerSy by (5.8.8). However Hl.J( C, d) = 0 for} *- 1 by Corollary 5.3.4. Hence U = 0 if } *- 1. If } = 1 we have only to apply (5.8.7).
446
LIE ALGEBRA COHOMOLOGY
123
5.9. Assume in this section that n = m (Case 2 of § 5.6). Now let CU, and let no = (uo)o. It is clear from I, Proposition 5.4, that a(n o) and its complement in a+ are closed under + and hence by I, Proposition 5.10, there exists a unique a E W such that U oE
(5.9.1) The following proposition gives s" for a satisfying (5.9.1). PROPOSITION 5.9. If a E W is such that (5.9.1) holds then s" = h" 1 )"(")
= ( -. 27r~
fq, /\f-q, • "
"
PROOF. By (5.7.12) with no substituted for n, it follows that 7r (f'P) fq,
6.1. Let M, B, and U be the subgroups of G corresponding to m, 0 and u. Thus (6.1.1)
M~B~U.
The subgroup M is closed since it corresponds to a unipotent group under a faithful representation of G. The subgroups Band U are closed by I, Remark 5.1. For each element a E W, let a(a) E G be such that O(a(a» induces the transformation a on 1). The generalized lemma of Bruhat (see e.g. [3]) then asserts that (6.1.2) is a disjoint union (of double cosets). More generally PROPOSITION 6.1. One has (6.1.3)
is a dis}oint union. PROOF. Using I, Proposition 5.13, it is enough to prove
447
124
BERTRAM KOSTANT
(6.1.4)
for any a E WI. Now let MI and BI be the subgroups corresponding to ml = m n gl and bl = b n gl. The Bruhat decomposition applied to G1 then asserts that (6.1.5)
(The fact that G1 is only reductive and not semi-simple clearly causes no trouble. Note also that a(r") E GH for T E WI since ~ ~ gl). On the other hand by I, Remark 5.13, (2), (6.1.6)
a(a)-IM1a(a)
~
M
for any a E WI. Now let N be the (closed; for the same reason that M is closed) subgroup corresponding to n. Since n is an ideal in both u and b, one has (6.1.7)
and (6.1.8) Thus by (6.1.5-8) Ma(a)-IU
=
Ma(a)-IG1N
=
U TEWI Ma(a)-l M Ia(T)-1B IN
= UTEW Ma(a)-la(T)-IB . 1
However since a(Ta)-1 and a(a)-la(T)-\ both define the same double M, B coset. This proves (6.1.4). q.e.d. 6.2. We need the following fact about nilpotent Lie algebras. LEMMA 6.2. Let mo be a nilpotent Lie algebra. Assume (66.2.1) is a linear direct sum where ml and m2 are Lie subalgebras. Let Mo be a simply connected Lie group corresponding to mo, and let M i , = 1,2, be the subgroups corresponding to mi , i = 1,2. Then the map
t:p: MI x M2 --> Mo , where t:p(a H a 2 ) = al U 2 is a bijection. PROOF. Since Mo is simply connected, the map exp: mo-->Mo is a bijection. It follows therefore that t:p is injective. Now let mi be the sequence of ideals in mo such that ml = mo and [m\ mJ] = mJ +1. Now assume inductively that exp x is in the image of t:p for all x E mJ+I (obviously true if j
448
LIE ALGEBRA COHOMOLOGY
125
is such that mHl = 0). Let Yo E mJ. Write Yo = Yl + Y2 where Yi E mi , i = 1, 2. Put a i = expYi, i = 0,1,2. Then clearly ao = a1·exp x·a 2 where x E m Hl (since [Yl, Y2] = [Yl> Yo] E mHl). But exp x = b1b2 where bi E Mi by the induction assumption. Hence a o is in the image of cp. q.e.d. REMARK 6.2. It is obvious that the image of cp is open in Mo. To prove surjectivity in another way, it suffices to show that the image of cp is closed in Mo. We have proved elsewhere (unpublished), more generally, that even without the condition (6.2.1) MIM2 is always closed in Mo. This fact has been generalized by M. Rosenlicht to the case of an arbitrary algebraic unipotent group (see [8, Th. 2, p. 221]). Now for any a E W let mO" ~ m be that Lie subalgebra, spanned by root vectors, such that (6.2.2)
=
A(meT)
and let MO" ~ M be the corresponding subgroup. Now if that
IC E
W is such
(6.2.3)
is a linear direct sum by I, (5.10.4). Hence by Lemma 6.1, (6.2.4)
and
MO"nMeT< = (1)
(6.2.5)
for every a E W. Now if a E W let M: be the subgroup corresponding to m:. Since A(m:) = -
M: = a(a)Mo--la(a)-l .
(6.2.6)
6.3. Now as in Part I, we let X be the homogeneous space (6.3.1)
X=G/U
of all left cosets a U, a E G. The point corresponding to U is called the origin and is denoted by o. If a, bEG and x = b U E X then a· x will denote abU. Consider the action of M on X. If a E WI, let VO" be the M orbit defined by putting (6.3.2)
and let M(a)
~
M be the isotropy subgroup of M defined by the point
449
126
BERTRAM KOSTANT
a(a)-l·o. The decomposition of X into M orbits is given by PROPOSITION 6.3. One has that (6.3.3)
is a disjoint union. Furthermore if a E Wl, then M(a) = Mer-I. so that by (6.2.4-5) the map b ---> ba(a)-l·o induces a holomorphic diffeomorphism (6.3.4)
Mer-l ---> Ver .
PROOF. The fact that (6.3.3) is a disjoint union is an immediate consequence of Proposition 6.1. Now let a E WI. We first show that
UnM: =
(6.3.5)
(1).
Indeed let 1 * a E Mer*. Then {}(a) * 1 since {}(M:) is a unipotent (hence simply connected) group. Now if x E f;J is regular, then clearly {}(a)x = x + e where 0 e E m:. But since m: n u ~ n* n u = 0, this implies that a ¢ U which proves (6.3.5). We next show that
*
(6.3.6) Indeed where
~_
=
-~+,
a(
= =
a(a-l/C(~_) n ~+) ~+na~+ ~+
n a/C~_
=
and clearly (6.3.6) is an immediate consequence of (6.3.7). Now let (6.3.8) Mer = a(a)Ma(a)-1 . By (6.3.6) and (6.2.6) one has Since Mer. ~ U, it then follows immediately from (6.3.5) that the isotropy subgroup of Mer defined by 0 is just Mer.. But then by (6.3.6) and (6.3.8) one obviously has M(a) = Mer-I.. q.e.d. Since Mer-l is homeomorphic to a cell, it follows from (6.3.3) and (6.3.4) that every orbit of M, and hence every orbit of any subgroup of G conjugate to M, is homeomorphic to a cell. Any such orbit will henceforth be called a Schubert cell. Note that by (6.2.6) and (6.3.4)
450
LIE ALGEBRA COHOMOLOGY
127
dimR V
(6.3.9)
6.4. One may consider the map (6.3.4) and the decomposition (6.3.3) from the point of view of topology and also algebraic geometry. It is an easy consequence that V
n(r) E
< n(a)
Wl(a) .
REMARK 6.4. Chevalley (unpublished) has given a criterion for an element r E W to lie in Wl(a). His characterization is in terms of sequences of elements in W. In Part III, using a different approach, we will give a function theoretic characterization of Wl(a). Now since X
(6.4.4)
H 2i +l(X, Z) = 0 ,
and H 2j (X, Z) is the free abelian group of rank wlU) given by (6.4.5) 6.5. Now consider the cohomology group H(X) of X with complex coefficients. For anyx l E H(X) and x E H(X, Z), we denote by <x\ x> E C the value assigned to Xl and x by the pairing of homology and cohomology. Now let x
E
Wi. Since there is no torsion in H*(X, Z) one has an imbedding
451
128
BERTRAM KOSTANT
of H(X, Z) into H(X). We then observe by (6.5.1) and (6.4.5), that not only is x" E H"(")(X, Z) but the elements x", a E W\ form a basis of the free abelian group H(X, Z). Now using de Rham theory we may take H(X) to be the cohomology group defined by the cochain complex whose underlying space is C(X), the set of all complex-valued C~ differential forms on X, and whose coboundary operator is exterior differentiation. If Q) E Z(X), the subspace of closed C~ forms, we will let [w] E H(X) be the cohomology class defined by w. Note that if we write (6.5.2) then (6.5.3) On the other hand, by a theorem of de Rham [7] and Lelong [6] (since the elements of V" are regular in X,,) one knows PROPOSITION 6.5. Let a E WI and let tion of w to VO' is integrable and
WE
Z2n(,,)(X). Then the restric-
(6.5.4)
where V" is oriented by its complex structure. 6.6. As in Part I, let Y be the special case of X defined by (6.6.1)
Y = GIB
(the generalized "flag manifold"). We substitute Y for X and y for x in the notation of § 6.4-5 when this case is referred to. Now consider the holomorphic surjection (6.6.2)
l.i:
Y
--->
X
defined by the injection B ---> U. LEMMA 6.6. Let a E Wand (6.6.3)
<[l.i *W] , y" > -_
WE
ZJ(X). Then
{<[W], x,,> 0
PROOF. If d *- 2n(a), the result is obvious. Hence we may assume d = 2n(a). Now it follows easily from (6.3.4) that, since l.i commutes with the action of G, it defines by restriction a diffeomorphism (6.6.4)
452
129
LIE ALGEBRA COHOMOLOGY
(the subscripts X and Yare introduced into VO" to distinguish the case X from Y). One then obtains the first relation of (6.6.3) by using (6.5.4). Now assume that a ¢ WI. By I, Proposition 5.13, we can then uniquely write a-I = all!, where a l E WI and!' E WI. But now (6.1.4) defines by restriction a surjection But (6.6.5) by I, (5.13.6), (see also I, Remark 5.13), and (6.3.9) since !'"* 1. But (6.6.5) and (6.5.4) clearly imply the second relation of (6.6.3). q.e.d. As an immediate corollary one has PROPOSITION 6.6. If).i is the map (6.6.2), then ).i*(xO")
(6.6.6)
= yO"
for any a E WI. REMARK 6.6. One effect of Proposition 6.6, is that for most cohomological questions one need consider only the case Y. 6.7. In this and the next section regard X as a real manifold, G as a real Lie group, and 9 a real Lie algebra. Let tR be the tangent space to X at o. Let Tl be the R-linear epimorphism (6.7.1) defined so that Tl(X), x at 0, and let
E
g, is the tangent vector to the curve exprx, r
E
R,
(6.7.2) be the homomorphism which extends Tl. LEMMA 6.7. Let tX,R ~ 9 be the subspace defined by (2.9.3). Then the restriction of T defines an isomorphism (6.7.3) PROOF. Let (6.7.4) so that fl is the set of all x E u such that x* = -x. Since gl follows immediately that fl is a real form of g10 that is
453
= un u*, it
130
BERTRAM KOSTANT
(6.7.5) and hence (6.7.6)
= fl + tX.R
f
is a direct sum. But then by dimension one must have (6.7.7) Since u is the kernel of T\ it follows from (6.7.4) and (6.7.7) that TIl f is an epimorphism with kernel fl. Hence TIl tX.R is an isomorphism by (6.7.6). Thus (6.7.3) is an isomorphism. q.e.d. 6.8. Now once and for all fix tors at o. That is
to be the space of all complex 1-covec-
t
(6.8.1) Next let (6.8.2)
AI: Atx -> At
be the isomorphism (by Lemma 6.7) defined so that (6.8.3)
=
(u, v)
for any U E Atx and v E ARtX.R. Now, using AI' carryover the structure from tx to t so that t satisfies the conditions of §§ 2.2-6, and § 3.8 with b = gl. But then AI obviously satisfies the conditions of § 5.2, and hence the results of § 5 apply. Now let (6.8.4) be linear isotropy representation on the space of all covectors at o. If a E U, and v E g, note then by definition of T
YEt,
(6.8.5)
=
.
We will also let PI denote the corresponding representation of u. REMARK 6.8. It should be observed that PI is not a holomorphic representation of U. That is, in general PI(ix) =1= ipix) for x E u. In fact, if not zero, they must be unequal since ARtR is clearly stable under PI. LEMMA 6.8. One has (6.8.6) PROOF. It suffices to show that (6.8.7)
454
LIE ALGEBRA COHOMOLOGY
131
for all z E fll yEt. But now by (6.8.5) one has (6.8.8) ~for
=
all v E g. On the other hand, if y f3r
=
-
=
Aru, u E tx then
Ar(O(z)u)
by definition of f3 r (see § 2.10). Substituting O(z)u for u in (6.8.3) it then follows from the relation O(z)t = -O(z) that (6.8.9)
= -
in case v E tX.R. But then (6.8.5) and (6.8.9) are equal for all v E tx.R. But then, by Lemma 6.7, this proves (6.8.7). q.e.d. 6.9. Now since K is compact, it follows from (6.7.7) that the projection G - X induces a diffeomorphism (6.9.1) where (6.9.2) Now let C(xt be the space of K-fixed differential forms. As a corollary of Lemma 6.8, we obtain LEMMA 6.9. The subspace
c~
At may be given by
(6.9.3) PROOF. Since X is simply connected, it follows from (6.9.1) that K1 is connected. Hence by (6.7.4) K1 is the subgroup corresponding to fl' It follows therefore that the right side of (6.9.3) equals (At)!l, defined with respect to Pr' But by (6.8.6) this equals (At)B1, defined with respect to f3 r since f3 r is a holomorphic representation of 91' But, by definition, C = (At)B1. q.e.d. Now if UE C let
C(xt be the unique element such that (w")o = u. w'" E
The map u--->w'" then defines
an injection (6.9.4)
C- C(X).
Regarding X as a homogeneous space of a compact Lie group (by (6.9.1» we may apply the Chevalley-Eilenberg theory of cohomology of compact homogeneous spaces (see [1]) to obtain H(X). We recall from this theory and also from the theory of relative Lie
455
BERTRAM KOSTANT
132
algebra cohomology that exterior differentiation on C(X) induces a derivation d l E End C of degree 1 on C, and that furthermore d l may be given by the relation (6.9.5)
for all WE C, V E Af where "/ E EndAg is the boundary operator of the chain complex C*(g). Moreover a theorem of Chevalley-Eilenberg asserts that the map (6.9.4) ind uces an isomorphism (6.9.6)
The significance, then, of the operator dE End C is given by PROPOSITION 6.9. One has d = dlso that (6.9.6) becomes an isomorphism
"'X,d: H(C, d) ~ H(X) .
(6.9.7)
PROOF. It follows from I, (3.8.2) and (5.7.11) that (dxu, v) = -(u, "/v)
(6.9.8)
for all u
E
Cx,
V E
At But if Aru =
W
then clearly (see §5.2)
Ardxu = dw. But then by (6.8.3) and (6.9.8) one has
(6.9.9)
Hence d
= -<w, T"/v> .
= d l by comparing (6.9.5) with
(6.9.9).
q.e.d.
REMARK 6.9. Using the decomposition (given in Part I) of HAn) as a gl-module we have found in §5, that W\ in a natural way, indexes a basis s" of H(C, d). On the other hand using the Schubert cells we have seen that Wl also indexes a basis x" of H(X). It will be one of the main results of the paper to show that up to a scalar multiple ",x,is") equals xO'. 6.10. Let a E Wl. For simplicity we will write (J)O' E C(X) instead of for the image of s" under the map (6.9.4). Also when dealing simultaneously with both X and Y, we will write (J)~ and (J)~ to indicate which of two spaces is being referred to.
(J)sO'
PROPOSITION 6.10. Let
1.1
be the map (6.6.2). Then for any a E WI,
(6.10.1) 80
that by (6.6.3)
(6.10.2)
456
133
LIE ALGEBRA COHOMOLOGY
for any
W 1• Let
T E
PROOF.
ax: Cx
-----+
C(X)
be the map obtained by composing the restriction of At to Cx and (6.9.4). Thus ax(sn = w~ for any a E W1. But now it is clear that the following diagram is commutative Cy~C(Y)
I
v*1
Cx~C(X)
where the left vertical map is defined by injection (see (5.7.2». But then (6.10.1) follows immediately from (5.7.9). q.e.d. 6.11. Let a E W1. The complex Lie group M: is a real Lie group of dimension 2n(a). Furthermore it has a natural orientation. That is, the complex structure picks out a generator ter E Aii'(er)m: unique up to positive multiples. In fact if Xj, j = 1,2, ••. , n(a) is basis of m: (as a complex vector space) then the elements X;, ix;, j = 1, 2, ... , n(a) are R-linearly independent, and ter may be given by
t er =
(6.11.1)
lI
n(er) j=1
X;
1\.
R'l,X;
(see [8, (5) on p. 14]). We now normalize the choice of ter by requiring that (6.11.2) (Since every n x n unitary matrix defines a 2n x 2n orthogonal matrix of determinant 1, it follows that ter, given by (6.11.1), is independent of the X; so long as (6.11.2) is satisfied.) Now let t = tR
+ itR
be the complexification of the tangent space tR so that we can regard At as the dual space to At. Furthermore let t Yr , 'Vr E A(tx), be the basis of t defined so that for all cP, 'Vr E A(tx) if'Vr if'Vr
(6.11.3) Also let tIP E
At,
for
~
*' cP =
rp •
A(tx), be defined in the same way as (5.5.3) with
457
134
BERTRAM KOSTANT
t replacing f. Now since ARm:
~
LEMMA 6.11.1. Let
ARg, note that T(t
E W. Then T(t
(6.11.4)
PROOF. Since the root vectors eip, f;l, it follows from (6.11.1-2) that
E A, form an orthonormal basis of
tp
(6.11.5) Now if x E n*, we assert that for any (6.11.6)
tp
= (eip' x
E A(tx) - x*) •
Indeed let z = x-x* so that x = x* + z. Now x* En C ll. Hence T(x*) = 0, and thus T(x) = T(z). But z* = -z. Hence z E f n tx = tX,R by (2.9.3). But then (6.11.6) follows from (6.8.3). Setting x = e_", and ie_", in (6.11.6) for t E A(m
(6.11.7)
=
L", - t",
and (6.11.8)
T(ie_",)
= i(L", + t",)
•
Thus T(e_"'/\Rie_",)
= 2iL",/\t",
which, by (6.11.5), (5.5.4) and (6.2.2) implies (6.11.4). As an immediate corollary one obtains LEMMA 6.11.2. Let (6.11.9)
0'
q.e.d.
E WI. The for all U E At one has
(2n")"(
=
where r
PROOF. It is clear from (6.11.3) that the isomorphism of At onto At defined by (At) maps f", into t", for all t E A(tx ). Thus by (6.11.4)
= i"(
= (27r)"(
by (2.5.2) q.e.d.
6.12. Now if V is a differentiable manifold, tp : V --> V, is a diffeomorphism, and tp is a k-vectoratpE V. Let tp·t p be the k-vector at tp(p)
458
LIE ALGEBRA COHOMOLOGY
135
defined by the differential of cpo Now if Wp is a k-covector at p, we will let cp·w" be the k-covector at cp(p) defined so that (6.12·1)
<w", t p >
=
for all k-vectors t p at p. Let a E Wl. Now, having fixed t rr , normalize left (equals right since m: is nilpotent) Haar measure on M: by the condition that, for any integrable continuous function f(a) on M:, (6.12.2)
r
JAlb-
f(a)da
=i
JM~
f(a)a,
where a is the left invariant 2n(a) form on M: such that
H
t rr >
=1
(al is the value of a at the identity element of M:) and Mrr* is given its natural orientation.
LEMMA 6.12. Let W be any integrable 2n(a) differential form on Mrr* (oriented by its complex structure). Then (6.12.3)
f
JAl~
W
=f
JM~
where cp(a) is left multiplication of M by a.
PROOF. If f(a) is the integrand on the right side of (6.12.3) then clearly f(a)a = w by (6.12.1). But then the lemma follows from (6.12.2). q.e.d. 6.13. Recall that q (6.13.1)
= if. Now put 1)q = q n 1)
and Oq
= 1)q + m .
These subspaces are clearly real Lie subalgebras of g, and moreover if Hq and Bq are the corresponding subgroups, one knows that Bq is an Iwasawa subgroup of G; (6.13.2)
is a semi-direct product, and (6.13.3)
G = KBq
is an Iwasawa decomposition of G. Now for a E G, let k(a) E K and b(a) E Bq be the unique elements such
459
136
BERTRAM KOSTANT
that (6.13.4)
a
=
k(a)b(a) .
Now since b(a)-l E Bq ~ U obviously Pr(b(a)-l) is defined for any a E G. PROPOSITION 6.13. Let a E WI, and let u
E Z2 n l<TI(C, d) so that W U is a
closed K-invariant differential 2n(a) form on X. Then
(6.13.5) where r
PROOF. Let V:
~
X be the M: orbit defined by o. That is
V: = M:·o.
(6.13.6) Now if (6.13.7)
a:M:->V:
is the map defined by putting a(a) = a·o, it follows from (6.3.5) that a is a diffeomorphism. But now by (6.2.6) and Proposition 6.3, it also follows that the action of a(a) on X induces a diffeomorphism But since, as one knows, a(a) can be chosen to be in K, it follows from the K-invariance of W U and (6.5.4) that
< [WU] , X
r
Jv~
WU •
= a*(wU) then since a is a diffeomorphism, it follows from (6.12.3)
(6.13.8) where cp(a) is left multiplication of M: bya. But by definition of a* and T, it is obvious that (6.13.9) where p(a) denotes the diffeomorphism of X defined by the action of a. Now a·o
by (6.13.4) since b(a)
E
= k(a)·o
U. But then by definition of wU , one has
(wU)a.o
=
p(k(a»).u .
460
LIE ALGEBRA COHOMOLOGY
But since p(a)-lp(k(a»
=
137
p(b(a)-l) by (6.13.4) it follows from (6.13.9) that
<
=
T(t u» .
But if bE U, then p(b)·u = Pr(b)u for any u E At by definition of the representation Pr. But now the result follows from (6.11.9). q.e.d 6.14. In §6.8 we determined Prill. Now consider Prln. It is at this point that the Lie algebra structure on t enters. Let (6.14.1)
fl-: n --> tR
be the map defined by fl-(x) = Aix)
+ Ar(x)
for any x En. Obviously f.1 is a real Lie algebra isomorphism. But much more important, one has PROPOSITION
tation of t on
6.14.1. Let x En. Then where one has
71:
is the adjoint represen-
At
(6.14.2)
For any Z E g, and any subspace .p ~ g, let zll = r~ where r ll is the orthogonal projection of g on.p. Now let Y E t x • Then clearly since XE n, PROOF.
7I:(f.1(x»)A ry
= 7I:(Arx + Arx)ArY = [Arx, ArY[ + [Arx, ArY] = Ar[x, Yn] - Ar[x*, Yn'] ,
since Arx = -Arx* (see (2.9.4» and since Ar is a Lie algebra isomorphism. But then if v E tX,R, <7I:(f.1(x»)Ary, Tv>
(6.14.3)
= ([x, Yn] - [x*, y n.], v) = -(y, [x, V]n' - [x*, V]n)
by (6.8.3)
by I, (3.7.1). On the other hand if (6.14.4)
u = [x, V]n. ,
then clearly u*
= ([x, V]*)n = [v*, x*]n =
[x*, V]n
Thus by (6.14.3) (6.14.5)
(7I:(u(x»A ry, Tv)
=
-(y, u - u*) •
461
since v E f .
138
BERTRAM KOSTANT
On the other hand by (6.8.5) (pt(x)Aty, Tv)
=
-(Aty, T[x, v)) •
[x, v]
=
[x, v]u
But
+u
;
and since T vanishes on u, one has
(6.14.6)
by (6.11.6) since u E n*. Thus by comparing (6.14.5) and (6.14.6) one obtains (6.14.2) on t, and hence on At since both sides are derivations. q.e.d. To prove Theorem 6.15, we need the following corollary of Proposition 6.14.1. PROPOSITION 6.14.2. Let a EN, and let v E Ker at. Then (6.14.7) PROOF. Let x En. Then by Proposition 6.14.1 Pt(x): Ker at --> Im at
,
since 11:(Y) has this property (by (2.7.1» for all yEt. Exponentiating, this clearly implies pt(a)v - v E Im at' q.e.d. 6.15. If hE Hq (see § 6.13) and homomorphism defined so that for h
~E
Z, we will let h --> h< E R be the
= expx, x E 1}q.
LEMMA 6.15. Let a E Wl and hE Hq then (6.15.1)
where r" E At is defined by (5.5.7) and ~" = g - ago PROOF. Let x E u and YEa. We assert that (6.15.2) Indeed let v E g, and let [x, v]u E u and [x, v]n* E n* be defined as in the proof of Proposition 6.14.1. Then since T([x, v]u) = 0, one has by (6.8.8) (which of course holds for any x E u) (6.15.3) Now let
(resp
~) ~
=
t be the subspace spanned by L"", (resp. t"",) for all
462
LIE ALGEBRA COHOMOLOGY ,., E ~(n)
so that t
139
= S + ~ is a direct sum. Now write
(6.15.4) where tl E sand t2 E~, it then follows from (6.11.7) and (6.11.8) (after choosing a E Wl so that mIT = n. See §5.9.) that
= T(i[x, v]n')
T([ix, v]n*)
= itl - it 2
But now since YEa one has
•
= 0 by (6.11.3). Thus substituting ix
=
l)
= i
on Aa.
But then since Hq c Glf one has (by Remark 5.5 and (5.1.7» (6.15.6)
But since
Pt(h)f"'IT = hlilT f"'IT •
A tR is obviously stable under pt(a) for all a E U, one has
(6.15.7) Hence by (5.5.5) one has (6.15.8) But then (6.15.1) follows immediately from (6.15.6,8). q.e.d. We can now prove the main theorem which asserts that up to a scalar multiple sIT and xIT correspond under the isomorphism 'Vrx,a' The scalar which is given here as an integral will be explicitly determined in Part
III. For any a E G let h(a) E Hq and m(a) E M be the unique elements such that (6.15.9)
b(a)
= h(a)m(a)
See (6.13.2).
463
.
140
BERTRAM KOSTANT
THEOREM 6.15. Let seT, a E Wr, be the basis of H(C, d) defined in §5.6 (using (1) the disjointness of d and a, and (2) the irreducible components (determined in Part I) of HAn) as a gcmodule). Furthermore let xeT, a E Wr, be the basis of H(X) (and also H(X, Z) defined in §6.5 (using the Schubert cell decomposition of X. See (6.3.3». Then (6.15.10) where ""'x.a is the isomorphism (6.9.7) and by
(6.15.11)
)...,eT
= (~)n(eT) r 2n
JM~
'A,eT
is the positive scalar given
I h(a)-'eTI2da .
PROOF. By (6.5.3) we have to show that <[weT], x,) =
*'
{~eT
*'
if'r a if'r = a
for all a, 'r E Wl. This of course is obvious if n(a) n('r). Hence we assume n(a) = n('r). Now by (6.10.2) it suffices to prove the theorem for the special case X = Y. We will therefore assume that n = m. (See Case 2, §5.6.) Now by Proposition 6.13, it is enough to prove if'r if'r
(6.15.12)
*' a *' a .
To prove this we first observe that (6.15.13) for all a E Gl • Indeed this is obvious on An by (6.15.5). (See (3.8.3).) By conjugation (see (6.15.7» it follows that Aa also is stable under pia) and (6.15.13) holds on Aa. But since ax
Pr(h(a)-l): 1m ar
--->
1m ar
•
We will next prove that for any a E G, Pr(b(a»)-lseT may be put in the form (6.15.15) where
464
LIE ALGEBRA COHOMOLOGY
(6.15.16)
WE
ImL t
141
•
Indeed by (5.6.3) and (5.6.6) one has
sO'= rO' and uO' E 1m a by (5.6.4). If v
+ uO',
= pih(a)-l)uO', V E
then by (6.15.14)
1m at,
and by (6.15.9) and Lemma 6.15,
pt(b(a)-l)sO' = pt(m(a)-l)(\,rO'
+ v)
.
But now since m(a)-l EM = N, it follows from (6.14.7) that
pt(m(a)-l)v E 1m at and because rO' = hO' E Ker L ~ Ker a, pt(m(a)-l)\,rO' = \'rO'
,
+u
,
where u E 1m at. Hence if W = u + pim(a)-l)v, one obtains (6.15.15-16) since 1m at ~ 1m Lt. But now {w, rt} = 0 since rt E Ker L ~ Ker L t and WE 1m Lt. Thus {pt(b(a)-lsO', rt} = Ih(a)-<0'12{rO', rt} • But if
(6.15.17)
a*-
T
ifa=T. by Theorem 5.4, (5.6.6), and Remark 5.5. Thus the integrand of (6.15.12) vanishes if a integral is \,0'. q.e.d.
*- T,
and otherwise the
REMARK 6.15. We wish to remark that the proof above rested heavily on the fact that there exists a representative cocycle sO', in sO', such that asO' = 0 (we recall such a cocycle exists by the disjointness of d and a). If sO' E sO' were an arbitrary representative cocycle, then even if T *- a, the integrand of (6.15.12) would not in general identically vanish. COROLLARY 6.15. Let a E Wl(j). Then formula (5.6.10) defines a closed differential form roO' on X which vanishes identically on every Schubert cell Vu T E Wl(j), for T *- a and whose integral over VO' is a positive number. PROOF. For the case X = Y, this follows from Theorem 5.6, Theorem
465
142
BERTRAM KOSTANT
6.15, (6.15.17) which shows that the integrand of (6.15.12) vanishes if (J T, and the proof of Lemma 6.12 which shows that the vanishing of the integrand implies that oj'" restricted to V, is zero. For the general case, this follows from (5.7.9), (6.10.1) and the diffeomorphism (6.6.4). q.e.d.
*'
6.16. We will determine )..,rr here for n«(J) = 1 by direct evaluation of the integral (6.15.11). In Part III, )..,rr will be determined for all a E W using different methods. Assume a E Wand n(a) = 1. Then a = Tr» for some a E II (see (5.8.3». Now let x = e_r» and y = ie_r», so that x and y are a real basis of m:. Now coordinatize M: (normal coordinates) so that
a(s, t)
= exp(sx + ty)
for s, t E R. It then follows immediately that 2ds /\ dt is the form a of (6.12.2) since t rr = (1/2)x/\y by (6.11.5). Thus if f(a) is an integrable function on M: and rp(s, t) = f(a(s, t») ,
then by (6.12.2) (6.16.1)
r
J.ar~
f(a)da = 2
rJR2 rp(s, t)dsdt
where R2 is the real plane. LEMMA 6.16. If f(a) is the function on Mrr* given by putting
f(a) = Ih(a)-
aE
M:,
f(a) = {O(a)e"" O(a)e",}-l •
= e
Now a = k(a)h(a)m(a). But restricting consideration to the three dimensional Lie algebra with basis er»' e_r» and [e"" e_",], it follows from the uniqueness of the Iwasawa decomposition that the three factors all lie in the corresponding subgroup of G. But then in particular m(a) lies in the 1dimensional subgroup generated bye",. Hence
466
143
LIE ALGEBRA COHOMOLOGY
O(m(a»)e",
= e...
But then since O(k(a» is unitary with respect to {A g} one has
1h(ayrT 12 =
{O(a)e"" O(a)e",}
by (6.16.4). Inverting one obtains (6.16.3). Now if c = s + it E C then a = a(s, t) = expce_"" and hence
O(a)e", = e..
+ [ce_""
e..] + 2..[ce_"" [ce_"" e",]] 2
(the Taylor series stops with 3-terms since O(e_",)3e.. (5.1.3)
= 0). Thus by I,
(6.16.5) where x.. E 1) is the root normal corresponding to a. But since the three vectors on the right side of (6.16.5) are orthogonal to each other, and since {x"" x",} = <x.. , a) = (a, a), one has by (6.16.3)
f(a)
Since 1c 12
=
S2
PROPOSITION
+ t2,
(1 + IcI2(a, a) + 1:1'(a, a)2r 2 = (1 + 1~12(a, a)r . =
this proves (6.16.2).
1
(q.e.d.)
6.16. If a = '/:'"" a E II, then ).,rT
=
(g, a)-l
where as usual g is one half the sum of the positive roots. PROOF. By (6.15.11)
But then by (6.16.1) and Lemma 6.16, (6.16.6) But since the integrand is a radial function, one converts to integration over (0, co) by introducing the factor 2nrdr where r2 = S2 + t 2 • Thus
,< ~ 2 [ (1 + ¥r.), ~ (a~ a) .
467
144
BERTRAM KOSTANT
But (g, a) = «a, a){2). See I, (5.9.1). REMARK
(6.16.7)
6.16. For any
(J
E
q.e.d.
W\ put 1 -T -1- SCT SC ",CT
and let sf = [sna. The sf are a more natural basis of H(C, d) than SCT since, by Theorem 6.15, they correspond exactly to the integral classes x CT under the map 'Vrx,a. In case (J = T"" a E IJ(u), one has, by Proposition 6.16, sf= (g, a )SCT, and hence (6.16.8) by (5.8.5). This will be needed in Part III to determine
",CT
for all
(J.
MASSACHUSETTS INSTITUTE OF TECHNOLOGY REFERENCES 1. C. CHEVALLEY and S. ElLENBERG, Cohomology theory of Lie groups and Lie algebras Trans. Amer. Math. Soc., 63 (1948), 85-124. 2. C. EHRESMANN, Sur la topologie de certains espaces homogenes, Ann. of Math., 35 (1934), 396-443. 3. HARISH-CHANDRA, On a lemma of Bruhat, J. Math. Pures et Appl. (9), 315 (1956), 203-210. 4. B. KOSTANT, Lie algebra cohomology and the generalized Borel- Weil theorem, Ann. of Math., 74 (1961), 329-387. 5. J. L. KOSZUL, Homologie et cohomologie des algebres de Lie, Bull. Soc. Math. France, 78 (1950), 65-127. 6. P. LELONG, Integration sur un ensemble analytique complexe, Bull. Soc. Math. France, 85 (1957), 239-262. 7. G. DE RHAM, Seminars on analytic functions, Princeton, 1957, vol. 1, 54-64. 8. M. ROSENLICHT, On quotient varieties a.nd the affine embedding of certain homogeneous spaces, Trans. Amer. Math. Soc., 101 (1961), 211-223. 9. A. WElL, VariHes Kahleriennes, Hermann, Paris, 1958.
468
Topology Vol. 3, Suppl. 2, pp. 147-159. Pergamon Press, 1965. Printed in Great Britain
EIGENV ALUES OF A LAPLACIAN AND COMMUTATIVE LIE SUBALGEBRAS BERTRAM KOSTANTt
(Received 29 August 1963) §1. INTRODUCTION
(1.1). IF K
IS a compact semi-simple Lie group and 9 is the complexification of its Lie algebra then one knows that the algebra Q of (Maurer-Cartan) complex-valued left invariant differential forms may be naturally identified with the exterior algebra Ag. Also, one knows then that Ag is stable under the Laplacian defined with respect to the canonical Riemannian metric on K.
Let L be the restriction of the Laplacian to Ag. It is then well known that the minimal eigenvalue of L is 0 and the corresponding eigenspace is the space of all harmonic differential forms on K. What we wish to consider here is the maximal eigenvalue mk of L on Akg and the corresponding eigenspace Mk ~ Ng. Now on the other hand let Ak ~ Akg be the subspace spanned by all non-zero decomposable elements in Akg (an element u E Akg is decomposable if u = Zl/\ ... /\Zk where Zj E g) whose corresponding subspace in 9 (the subspace corresponding to u is the one spanned by the Zj) is a commutative Lie subalgebra of g. If p = max dim a
where a runs over all commutative Lie subalgebras of 9 then obviously Ak # 0 if and only if 1 ::; k ::; p. The determination of the integer p was first made by Malcev [4]. In the case of the classical groups, like the dimension, p is a quadratic expression in the rank of K. For the exceptional groups G2 , F 4 , E 6 , E7 and E8 one has, respectively, p = 3, 9, 16, 27 and 36. The two matters brought up above are related by the following (proved here as Theorem (5), §4.4). THEOREM.
If mk
is the maximal eigenvalue of the Laplacian L on Akg then one always
has k mk <-
-2
Moreover one has mk = k/2
if and only if there exists a commutative Lie subalgebra of 9 having
t The author is an Alfred P. Sloan fellow. 147 B. Kostant, Collected Papers, DOI 10.1007/b94535_19, © Bertram Kostant 2009
469
148
BERTRAM KOSTANT
dimension k; that is, a case one has
if and only !f Ak i= 0, or equivalently, I :::; k :::; p. Furthermore in such Mk =Ak
where Mk is the eigenspace of L on Akg belonging to m k and Ak is the subspace defined, as above, by all k-dimensional commutative Lie subalgebras of g. Finally any decomposable element in Ak necessarily correspondY to a commutative Lie subalgebra of g. It may not be out of place to emphasize that even though it was the Laplacian on K under consideration the result above required use of the commutative subalgebras of the complexification g. The Lie algebra of K has no commutative subalgebras having dimension greater than the rank.
One can give a more accessible description of A k • The algebra Ag is of course a K-module with respect to the adjoint representation. It is immediate that Ak is a K-submodule. (A K-submodule will be referred to as simple if the representation of K it defines is irreducible.) The following, in more detail, is proved in the paper as Theorem (7), §(4.5). THEOREM. Let b be any maximal solvable Lie subalgebra of g. If a is any commutative ideal of b let An be the cyclic K-submodule of Ag generated by the decomposable element u corresponding to 0 (u E Akg if dim 0 = k). Then An is a simple K-module. In fact a -+ An sets up a one-one correspondence from the set of all commutative ideals of b onto the set of all K-simple submodules of A = IA k • Moreover there are only a finite number of commutative ideals of band (1.2.1)
where a runs through this .finite set. Finally, not only is (1.2.1) a decomposition of A as a direct sum of simple K-modules but is in fact the unique such decomposition. (The last statement implies and is implied by the fact that A n, and An2 are inequivalent as K-modules whenever al i= 02') §2. THE LAPLACIAN ON DECOMPOSABLE MULTI-VECTORS
(2.1). Let K be a compact semi-simple Lie group and let f be its Lie algebra. Let 9 = f + if be the complexification of f and let (x, y) denote the value the Cartan-Killing form B assigns to any pair x, y E g. Now let z -+ z* be the *-operation (analogous to Hermitian adjoint; see [2, §3.1] for more details) on 9 defined by z* = -x + iy for any z E 9 where z = x + iyand x, YEt Since - B is positive definite on f it is then clear that a positive definite Hermitian inner product is defined on g by putting {x, y}
[2; §3.3]. Thus if
Xl' ...
,x.
= (x, y*)
is an orthonormal basis of 9 (always taken with respect to
{x, y}) then
(2.1.1)
470
EIGENVALUES OF LAPLACIAN AND COMMUTATIVE LIE SUBALGEBRAS
149
Now let g~
0:
End Ag
be the unique representation of 9 on the exterior algebra Ag extending the adjoint representation of 9 on 9 and such that O(x) is a derivation of Ag for any x E g. Also let Q denote the exterior algebra of left invariant complex-valued (Maurer-Cartan) differential forms on K. Since Q is isomorphic to the algebra of complex covectors at the identity of K it is clear that B induces an algebra isomorphism Ag~Q
(2.1.2)
On the other hand Q is stable under exterior differentiation and if d denotes the restriction of exterior differentiation to Q then one knows that dw(x, Y)e = -w([x, YDe for any I-form WE Q and x, Y E 9 where subscript e means contraction at the identity of K. It follows therefore that if we identify Ag with Q by (2.1.2) then by (2.Ll) and [3, §3] one has 1
for any orthonormal basis by y.
Xi
n
=-
d
2i
L E(Xt)O(x =1
(2.1.3)
i)
of 9 where E(Y) denotes left (exterior) multiplication in Ag
Now let H be the Hermitian positive inner product on Ag given by I
h1/\ ... /\Y p '
_
)
_
ZI/\ ... /\L.qf -
(0 if p "# q Id {. }·f \ et Yi' Zj 1 P = q
(2.1.4)
We recall from [2, (3.9.2) and (3.9.7)] that if A* denotes the Hermitian adjoint for any operator A on Ag one has O(x)*
= O(x*) for any x E 9
(2.1.5)
and d* = () where () is given by D( Y 1 /\
••• /\
Yp)
=
L (- l)i +
j
+ 1 [Yi'
Y j] /\ Y1 ..•
/\
Yi ... /\ Yj
... /\
Yp
i<j
where Yi
E
g. (In particular one notes that o(y /\ z) = [y, z] for y,
Z E
g.)
Now - B is positive definite on f and induces a 2-sided invariant Riemannian metric on K. Recalling the theory of harmonic integrals as it would apply to K with respect to this metric it is clear from left invariance that Q is stable under the operation w --+ *w of taking adjoints (in the sense of harmonic integrals). It follows easily then, using left invariance, that the Hermitian inner product induced on Q is a scalar multiple of H. It follows also that Q is stable under the adjoint of exterior differentiation so that Q is stable under the Laplacian. In fact, then, if L is the restriction of the Laplacian to Q one must have L
= do + od
On the other hand since il commutes with O(x) and one has E(y)il §3-4]) it follows therefore from (2.1.3) and (2.1.6) that L =
t LO(xt) O(Xi)
for any orthonormal basis Xi of g.
471
(2.1.6)
+ OE(Y) =
O(y) (see [3, (2.1.7)
150
BERTRAM KOSTANT
(2.2). Throughout LEMMA
Xl' ... , Xn
we will denote an orthonormal basis of g.
(1). For any y E 9 one has Ly =
lY.
Proof Let Z E 9 be arbitrary. It suffices to show that {Ly, z} = t{y, z}. But now
(Ii 8(xiW(x;)y, z} = Ii {[Xi' y], [Xi' z]} by (2.1.5). But using (2.1.5) again one has 2{Ly, z} = Ii {[z*[y, Xi], Xi} = tr ad z*ady = (y, z*) = {y, z}. Hence {Ly, z} = -!{y, z}. Q.E.D. Another lemma needed is LEMMA
(2). Let y, z
E
9 be arbitrary. Then
I8(xi)y1\ 8(Xi)Z = d[z, y]. i
Proof Let u, v E g. Then by (2.1.5)
I
{8(xi)y, u}· {8(Xi)Z, v} =
i
I
{[y, u*], X;}{Xi' [v, z*]}
i
= {[y, u*], [v, z*]}
= - {y, [u, [v, z*]]} But then if we interchange u and v above and subtract the resulting expression from the one above we obtain {[z, y], [u, v]}. On the other hand by (2.1.4) this difference is exactly {~:i 8(xi)y 1\ 8(xj)z, Ul\v}. But since d* = 0 and o(UI\v) = [u, v] one has also {d[z, y], U1\ v} = {[z, y], [u, v]}. Since Uand v are arbitrary this proves the lemma. Q.E.D. For future reference we note that the last equality in the proof implies that
{d[z, y], yl\z} = -{[z*[z, y]], y} for any z, y
E
(2.2.1)
g.
(2.3). Let a by putting
~
9 be any subspace. We define an operator on Ag, corresponding to a,
La = I 8(zj) 8(z) j
where Zl' ... , Zn is an orthonormal basis of a. It is clear that La is independent of the orthonormal basis chosen. Actually our interest is only in the restriction of La to a itself. Let P a be the orthogonal projection of Ag on a ~ g. Let Sa = P a La P a and define a scalar Sa corresponding to a by putting (2.3.1) Sa = tr Sa LEMMA (3). For any subspace a commutative Lie subalgebra of g.
~
9 one has Sa
~
0 and sa = 0
if and only if a is a
Proof Since Sa is clearly positive semi-definite one has that sa ~ 0 and sa = 0 if and only if Sa = O. But since 8(zj) 8(zj) is positive semi-definite one has then that Sa = 0 if and only if P a 8(zj) 8(zj)Pa = 0 for all j, or more simply, if and only if 8(z)Pa = 0 for allj. But 8(z)Pa = 0 if and only if [Zj' y] = 0 for all yEa. But since the Zj are a basis of a one has that sa = 0 if and only if a is a commutative Lie algebra. Q.E.D.
472
EIGENVALUES OF LAPLACIAN AND COMMUTATIVE LIE SUBALGEBRAS
151
(2.4). Now recall that an element u E Ng is called decomposable if u = Zl /\ ... /\ Zk where Zj E g. If u '# 0 then the Zj are linearly independent spanning a subspace 11 £ 9 of dimension k, which will be referred to as the subspace corresponding to u. (One knows, of course, this sets up a one-one correspondence between all k-dimensional subspaces of 9 and all lines through the origin of non-zero decomposable elements in Akg.) PROPOSITION (1). Let u E Akg be any non-zero decomposable element normalized so that {u, u} = 1. Let 11 £ 9 be the corresponding k-dimensional subspaces of g. Then if L is the Laplacian (2.1.6) one has {Lu, u} =!(k - sa)
where the scalar Sa is given by (2.3.1). Proof. We may write u = Zl/\ ... /\ Zk where the Zj are an orthonormal basis of 11. Let 1 ~ p, q ~ k be arbitrary. We define an element Vp,q E Ng as follows. If p = q put Vp,q = 0 and if p '# q let Vp,q = L Zl ... A(J(xi)zp ... A(J(Xi)Zq ... /\Zk i
(The expression above presupposes that p < q. If q < p then O(xj)zp and O(Xi)Zq should be interchanged.) Then by (2.1.4), if p '# q, {vp, q' u}
= {f O(xi)zp /\ O(Xi)Zq, zp /\ Zq}
since the Zj are orthonormal. But then by Lemma (2) and (2.2.1) one has {vp,q, u} = - {e(z;) O(Zq) zp' zp} for alII ~ p, q ~ k. Thus if v = L~,q=l Vp,q one has (2.4.1)
{v, u} = -Sa On the other hand since O(y) is a derivation of Ag for any y
E
9
k
Lu = v + L Zl ... /\ LZj ... /\ Zk j=l ku
=v+T by Lemma (1). Applying (2.4.1) one therefore has {Lu, u} = !(k -
sJ.
Q.E.D.
More important for our purpose is the following corollary of Proposition (1). CoROLLARY
(Ll). Let u ENg, Land 11 be as in Proposition (1). Then {Lu, u}
and the equality holds
~!k
if and only if 11 is a commutative Lie subalgebra of g.
Proof. This is an immediate consequence of Lemma (3) and Proposition (1). Q.E.D. It is suggestive from Corollary (1.1) that the maximal eigenvalue of the positive semi-
definite operator Lon Akg is at most k12. However we have tested L only on decomposable elements. We get a more complete picture of L in the next section.
473
152
BERTRAM KOSTANT
§3. THE MAXIMAL EIGENVALUE OF THE CASIMIR OPERATOR
(3.1). Let V be a finite dimensional complex vector space which is a K and hence a g-module with respect to some representation v:
g~End
V.
A subspace V1 £; V stable under v(g) will be referred to as a K-submodule and will be called simple if the subrepresentation of K it induces is irreducible. Since K is compact we may assume that a positive definite Hermitian inner product {v, w} is defined on V with respect to which K operates as a subgroup of the unitary group. It follows therefore that v(x)* = v(x*) for any x E n where v(x)* is the Hermitian adjoint ofv(x). Now by the relation (2.1.1) it is clear that the Casimir operator FE End V is given by F
= L v(xt) v(x i) i
One thus observes that F is positive semi-definite. We recall some familiar facts in representation theory. Let I be the rank of g. One knows we may choose a Cartan subalgebra lJ £; 9 so that Yl ... , Yl is an orthonormal basis of 1) and Y; = Yj' j = 1, ... ,T. (i.e. iYj E f) and if ~ denotes the set of roots we may choose root vectors eq" ep E ~, so that the Yj together with e", forms an orthogonal basis of g. Moreover the root vectors may be chosen so that e: = e _'" for any ep E~. (See e.g. [2, §§5.1 and 5.4].) It follows therefore that if ~+ is a system of positive roots then I
F =
L V(Yj)2 + L j= 1
"'el!+
v(e",)v(L",)
+ v(e_",)v(e",)
Since B is non-singular on {) we may identify {) with its dual. One knows then that [eq" e _q,] = ep for anyep E ~+. (See e.g. [2; (5.1.3)].) Thus v(e",) vee _.",) = v(ep) + vee _",) v(e",). It follows therefore that
F=E+N where E = L~= 1 V(y)2 + 2v(g) and 9 = operator given by N = Lq,el!+ vee _",)v(e",).
t
I"'el!+ ep
Remark (1). One notes that for any WE V, Nw
ep
(3.1.1)
and N is the positive semi-definite
= 0 if and only if l'(e",)w = 0 for all
E ~+.
Let Z £; 1) b~ the group of integral linear forms on f). We recall that an element Jl E Z is called a weight of V if there exist WE V, W # 0, such that v(y)w = Jl(Y)w all Y E lJ' The vector W is called a weight vector belonging to Jl. The space spanned by all such w is called the weight space for Jl and its dimension is called the multiplicity of Jl (mult Jl). Now let w be a weight vector belonging to Jl E Z. It is then clear from the expression above given for E that since, as one knows, B = H is real on the real subspace of lJ spanned byZ (3.1.2) Ew = (11g + Jlll 2 - IIgIl2)W where
IIA.II = (A., A.)'h for any A. E Z.
474
EIGENVALUES OF LAPLACIAN AND COMMUTATIVE LIE SUBALGEBRAS
It follows therefore from (3.1.1) that ifw is normalized so that {w, w}
{Fw, w} = Ilg
+ 11112 -- IIgl1 + {Nw, 2
=
153
1
w}
(3.1.3)
Now for any vector WE V let Vw be the cyclic K-submodule of V generated by w. Now a weight vector WE V is called a highest weight vector if Nw = 0 (see Remark (1»). One knows that there is a one-one correspondence between all lines (w) defined by highest weight vectors and all simple K-submodules of V; the correspondence being established by the map (w) ~ V w ' Also if WI' ... , wp are linearly independent highest weight vectors then the sum of simple K-submodules p
LV
w,
is direct
(3.1.4)
i= t
Now if w is a highest weight vector and w belongs to AE Z then A, called the highest weight of V"" occurs with multiplicity one in Vw' Its weight space in Vw is just (w). Now F, as one knows, commutes with the action of K and hence reduces to a scalar on any simple K-submodule of V. Applying (3.1.3) one recovers the following well known fact. Remark (2). Let VI ~ V be a simple K-submodule of V and let A E Z be its highest weight then F = !Ig + A.112 - IIgl1 2 on VI'
(3.2). Now let D ~ Z be the set of all AE Z such that (A, cf» ;:.: 0 for all cf> E Ll+. For each A E D one knows that there exists an irreducible representation VA : 9 ~ End VA whose highest weight is A and every simple K-module is equivalent to V' for one and only one A ED. Thus the structure of any K-module W is known as soon as one knows the nulliplicity of VA in W for any A E D. Our interest here is not in the K-module structure of V, which is arbitrary, but in a submodule now to be considered. Let f be the maximal eigenvalue of F and let M ~ V be the corresponding eigenspace. Obviously M is a K-submodule of V. For each A E Diet M(A) be the set of all vectors WE M which transform according to vi. (that is, all simple K-submodules of Vw are equivalent to VA) and let D(M) be the set of all A. E D such that M(A.) ¥: 0 so that
L
M =
M(A.)
(3.2.1)
AE/}(M)
is a direct sum of (primary) K-submodules. On the other hand define a weight Jl of V to be g-maxim:ll if ilg + Jl i z - :1 g il2 = f and let r be the set of all g-maximal weights of V. (Obviously r is not empty by Remark (2).) (2). One has D(M) = r. Moreover (f A E r then any weight vector for A is a highest weight pector. In fact if I' = mult A and 11'1' ... , H'r is a basis of the weight space for A then PROPOSITION
r
M(A) =
L1 V
w,
(3.2.2)
i=
and (3.2.2) is a decomposition of M(A) as a direct sum of simple K-modules so that mult
VA
in M = mult A
475
(3.2.3)
154
BERTRAM KOSTANT
Proof If A E D(M) then A E r by Remark (2). Now if WE Vand {w,
W} =
1 then since
f is the maximal eigenvalue of F one obviously has {Fw, w} ::; f But if Ji E rand w is a weight vector belonging to Ji it then follows from (3.1.3) that {Nw, w} = O. But then Nw =0 since N is positive semi-definite. Hence w is a highest weight vector and Ji E D(M) by Remark (2). Thus D(M) = r and every weight vector for Ji E r is a highest weight vector. The decomposition (3.2.2) and (3.2.3) is then immediate from (3.1.4) upon considering the Q.E.D. space of highest weight vectors in M(A). For one of our purposes Proposition (2) has a much more convenient corollary. (2.1). Let W1' ••• , Wm be any basis of V consisting of weight vectors. (Recall that vl~ is diagonalizable.) Let the Wi be ordered so that the weights corresponding to the Wi are g-maximal if and only if 1 ::; i ::; q. Then the Wi for 1 ::; i ::; q, are highest weight vectors and (3.2.4) COROLLARY
Furthermore (3.2.4) is a decomposition of M as a direct sum of simple K-modules. Proof This is an immediate consequence of Proposition (2), (3.2.1) and the fact that those weight vectors among the Wi which belong to any weight Ji form a basis of the weight Q.E.D. space for Ji. §4. THE MAXIMAL EIGENVALUE OF THE LAPLACIAN L AND THE
CORRESPONDING EIGENSPACE
(4.1). We apply the considerations of §3 to the case where V = Akg and v is the subrepresentation of () that V defines. Obviously the Laplacian 2L on Ng (see (2.1.7») is just the Casimir operator F.
We wish first, however, to focus attention on a certain class of K-submodules of V. A K-submodule V1 s V will be called decomposably-generated if it is spanned by decomposable elements. Remark (3). If V1 is simple note that it is decomposable generated if and only if it contains at least one non-zero decomposable element.
We wish to note also that not every K-submodule is decomposably generated. One needs only to consider the space of invariants in A 3g when rank gl > 1 for some simple component gi of g. For simple K-submodules one has (3). Let V 1 be a simple K-submodule of V = Ng. Then V1 is decomposably-generated if and only if its highest weight vector w is decomposable. PROPOSITION
Proof If w is decomposable then V 1 is decomposably-generated by Remark (3). Now assume V1 is decomposably-generated. To prove the converse let m be the maximal nilpotent Lie subalgebra of 9 given by m = Lt/>eA/et/» and let 0 be the maximal solvable Lie subalgebra given by the semi-direct sum b=~+m
476
(4.1.1)
EIGENVALUES OF LAPLACIAN AND COMMUTATIVE LIE SUBALGEBRAS
155
Now since [0, 0] = m it follows that a vector v E V is a highest weight vector if and only if the line (v) is stable under o. That is, if B denotes the solvable subgroup of the complexification of K corresponding to 0 and P(V) is the projective space of all lines in V and B is regarded as operating on P(V) then v is a highest weight vector if and only if the point v E P(V) corresponding to v is fixed under B. Now let Us;; P(V) be the subset defined by all points v where v is a non-zero decomposable element in V1 • Obviously U is a complete subvariety of P(V) and U is stable under B. By [1; Proposition 15.5] there exists a fixed point for B in U. By uniqueness of the highest weight vector in V1 this must be w. Hence w is decomposable. (One may give a representation theoretic proof of Proposition (3) without using Borel's theorem but it is somewhat longer.) Q.E.D. (4.2). Our main concern in this paper is with the maximal eigenvalue mk of the Laplacian Lon Ng, k = 1, ... , dim g, and with the corresponding eigenspace M k • (Obviously mk = f/2 in the notation of §3.) As a K-module we now observe PROPOSITION (4). Mk is a decomposably-generated K-submodule of V = Akg. In fact there exists decomposable highest weight vectors W1' ... , Wq such that
(4.2.1) so that (4.2.1) is a decomposition of Mk as a direct sum of decomposably-generated simple K-submodu/es. Proof. Proposition (4) is an immediate consequence of Corollary (2.1) since we may find a basis of Akg consisting of decomposable weight vectors. Indeed if yj' j = 1, 2 ... I, and e>, ¢ E A, are as in §(3.1) then all exterior k products of these basal vectors give such a basis of Ng. Q.E.D.
(4.3). From now on we reserve p for the integer given by p = max dim
where
tl
tl
runs through the class of all commutative Lie subalgebras of g.
To determine the values of p it clearly suffices to know p in the case where 9 is simple. But in this case the values have been determined. (Malcev). (1) For 9 = Alone has p = [(I + 1)2/4]; 9 = B3 one has p = 5 ;for 9 = B I ,I';?4' one has p = /(/ - 1)/2 + 1; 9 = Clone has p = /(/ + 1)/2; 9 = D" / ;?: 4, one has p = /(/ - 1)/2, and finally 9 = G2 , F4 , E 6 , E7 and E8 one has p = 3,9,16,27 and 36 respectively.
THEOREM
(2) (3) (4) (5)
For For For For
Proof. This is proved in [4].
Q.E.D.
(4.4). Now for any k> 1 let Ak S;; Ng be the (K-module) space spanned by all decomposable vectors U E Ng whose corresponding k-dimensional subspace tl S;; 9 is a
477
156
BERTRAM KOSTANT
commutative Lie algebra. Obviously then one has
Ak
=f;
(4.4.1)
0 if and only if I :::; k :::; p
We now have THEOREM (5). Let k ~ I and let m k be the maximal eigenvalue of the Laplar.ian L 011
Akg. Then
k
/11 k
(4.4.2)
<-2
and mk = k /2 if and only if there exists a commutative Lie subalgebra of 9 having dimension k; that is, if and only if 1 :::; k :::; p. If, moreoz'er, Mk denotes the eigenspace of L corresponding to m k then in such a case (l :::; k :::; p) one has
(4.4.3) where Ak is the space spanned by all decomposable elements u corresponding to k-dimensional commutatit)e Lie subalgebras 11 ~ g. Conversely if u E Ak is decomposable then the subspace 11 ~ 9 corresponding to u is necessarily a commutative Lie subalgebra of g. Proof Follows immediately from Corollary (Ll) and Proposition (4) which asserts that Mk is always decomposably-generated. Q.E.D.
The complete description of L on A 1 9 is given by Lemma (l). One can do the same for L on A2g. COROLLARY (5.1). Let A 2 , as defined above, be the subspace of A2n spanned by all bivectors x /\ y where [x, y] = O. Then one has the equality
A2 = {u
E
A 2 glau
= O}
and, consequently, also the direct sum
(4.4.4) where d(g) is the space of coboundary elements in A 2n. Moreover L A2 =f; 0, i.e. rank 9 > 1) L = 1 on A 2 • Proof. Obviously A2n = Ker a
!1
A20
+ d(g)
= ton d(g) and (assuming
is a direct sum. Moreover since
d : 9 -+ d(g)
(4.4.5)
is a K-module isomorphism (follows e.g. from Lemma (1») d(g) is equivalent to 9 as a K-module and hence L =! on d(g) by Lemma (1). Furthermore obviously A2 ~ Ker a!1 A2g. Thus to prove Corollary (5.1) it suffices, by Theorem (5) to show that if U E Ker a!1 A2g then U E A 2 • Write u = Lj Yj /\ Zj where yj' Zj E g. But dau = Lj d[Yj' Zj] = Osince au = o. Thus LLi O(xi)Yj /\ O(xi)Zj = 0 by Proposition (1). But then Lu = LjL(yj) /\ Zj + LjYj /\ L(z) = u by Lemma (1). That is, u is an L-eigenvector for the eigenvalue I. Thus u E A2 by Theorem (5). Q.E.D. Iff: !1 -+ Y is a linear transformation from 9 to any arbitrary vector space Y then the bilinear function g from 9 to Y defined by g(x, y) = f([x. y]) obviously has the property that g(x, y) = 0 if x commutes with y. We now find that the converse is true; every such bilinear function is a linear function (necessarily unique since [g, g] = 9) of the bracket.
478
157
EIGENVALUES OF LAPLACIAN AND COMMUTATIVE LIE SUBALGEBRAS
COROLLARY (5.2). Let Y be an arbitrary complex vector space and 9 a bilinear function from !l to Y. Then 9 is of the form
g(x, y)
= f([x,
y])
for a linear junction jfrom 9 to Y if and only if g(x, y)
=0
I-'o'henever x and y commute.
Proof Assume g(x, y) = 0 whenever [x, y] = O. Then g(x, x) = 0 for all x E 9 and hence 9 is alternating. That is, 9 may be regarded as a linear map A2g --+ Y where g(x, y) = g(x 1\ y). But then by assumption 9 vanishes on A 2 , and hence, using the decomposition (4.4.4), g(x 1\ y) depends only on the component of x 1\ y in d(g). The result then follows since 2a vanishes on A2 and is the inverse (by Lemma (1» of the isomorphism (4.4.5.) on d(g). Q.E.D. Remark (4). It is unknown to us whether or not Corollary (5.2) is true for an arbitrary Lie algebra. (4.5). We would like to be more explicit about determining or constructing, the submodules Ak f; Akg. By Proposition (4) (and Theorem (5» all we need to determine are the decomposable highest weight vectors w1 , •.• ,Wq' We are aided by the fact that a decomposable highest weight vector is of a very simple nature.
v
Let be the maximal solvable Lie algebra of 9 given by (4.1.1). A subspace a will be said to be v-normal if [v, a] f; a
f;
9
Since in particular a is stable under ad f) it follows that there exists a unique set of roots d(a) f; d such that (\ = l)
n a+
L
(e",)
(4.5.1)
","4(a)
+
and (4.5.1) is a direct sum. Moreover if denotes addition of roots whenever the sum is a root then d(a) is clearly restricted so that d+
+ d(a)
f;
d(a)
(4.5.2)
Now let (a) EZ be the linear form defined by (0)
=
L
", ..4(a)
As observed in the proof of Proposition (3) a vector WE Akg is a highest weight vector if and only if (w) is stable under V. Applying this to the case of a decomposable vector one immediately obtains (6). Let WE Ng be decomposable and let a f; 9 be the corresponding k-dimensional subspace. Then w is a highest weight vector if and only if a is b-normal. Moreover (0) is the highest weight of the simple K-module Vw ' PROPOSITION
One thus establishes a one-one correspondence between the set of all decomposablygenerated K-simp/e submodu/es of Akg (see Proposition (3» and all k-dimensional v-normal subspaces of g.
479
158
BERTRAM KOSTANT
It will be convenient to know when two decomposably-generated simple K-submodules of Ag are equivalent, even when they are not of the same degree. In this connexion we will make use of LEMMA (4). Let aI' a2 ~ 9 be two 6-normal subspaces of 9 (not necessarily of the same dimension). Let \}I = ~(al) n ~(a2) and let !/J = }2¢ where the ¢ run through \}I. Then (al) = (a2) if and only if!/J = (al) = (a2)' Proof Let \}Ii> i = 1,2, be the complement of \}I in ~(ai) and let !/Ji be the sum of the roots in \}Ii (taken to be zero if \}Ii is empty). Assume (al) = (a2)' We have only to prove that !/Jl = !/J2 = O. But by assumption !/Jl = !/J2' Hence, for one thing, we are reduced to the case where \}II and \}I2 are both not empty. Let ¢i E \}Ii' i = 1,2. We assert that (¢1' ¢2) :::; O. Indeed if (¢l> ¢2) > 0 then, as one knows ¢1 - ¢2 is a root. By possibly interchanging the order we may furthermore assume that ¢1 - ¢2 = ¢ E ~+. Thus ¢1 = ¢ + ¢2' But then ¢1 E ~(a2) by (4.5.2). Hence ¢1 E \}I. This is a contradiction. Thus (¢1' ¢2) :::; O. But this implies that (!/J1, !/J2) :::; O. However!/Jl = !/J2' Thus 11!/J111 = 0 which implies !/J1 = !/J2 = O. Q.E.D.
Included among the 6-normal subspaces of 9 are all the ideals of 6. Our interest is primarily in those ideals a of 6 where a ~ [6, 6] = m. Since [eq" e _q,] = ¢ E 'f) for any ¢ E ~ one may characterize such 6-normal subs paces by Remark (5). A 6-normal subspace a an'f) = O.
~
9 is an ideal of 6 with a
~
m if and only if
Clearly, by (4.5.1) there are only a finite number of such 6-normal subspaces a and each is characterized by the set ~(a) ~ ~+. Concerning the simple K-submodule they define one has THEOREM (7). Let a1' a2 be any two ideals of6 lying in [6,6] = if and only if a1 = a2'
m. Then (a1) = (a2)
That is, if VI ~ Akg, V 2 ~ Aig are any two decomposably-generated simple K-submodules which correspond (using Proposition (6) to ideals of 6 lying in [6, b] then VI is equivalent to V 2 if and only if VI = V 2. Proof Assume (a1) = (a2)' Let the notation be as in Lemma (4) and its proof. One has, by the latter, !/J1 = !/J2 = O. But since ~(ai) ~ ~+, i = 1,2, one has \}Ii ~ ~+. However, then !/Ji = 0 implies \}Ii is empty. Thus ~(ai) = ~(a2) and hence a1 = a2' The final statement follows from Proposition (6). Q.E.D.
If a is a commutative ideal in 6 then (as easily seen) one automatically has a The following yields a fairly complete description of the maximal L-eigenspaces A k •
~
m.
THEOREM (8). Let k ~ 1 and let «I> = (¢1' ... , ¢k) be any set of k positive (distinct) roots. Let elf) E Akg be the decomposable vector eq,l A ... A eq,k (multiplied in some order) and let alf) = Lq,elf)(eq,) be the corresponding subspace in g. Then one always has the inequality Ilg + ¢1 + ... + ¢k11 2 -11911 2 :::; k (4.5.3)
and the equality holds if and only ifalf) is a commutative ideal of6 (see (4.1.1.)). Moreover every commutative ideal in 6 is (uniquely) of this form.
480
159
EIGENVALUES OF LAPLACIAN AND COMMUTATIVE LIE SUBALGEBRAS
Now let q denote the number of commutative ideals in 6 of dimension k and let <1>1' .,. ,
(4.5.4) is a unique decomposition of Ak as a direct sum of simple K-modules. In fact vary then the V w, are the only simple K-submodules of A = LkAk'
if we
let k
Proof Let w = e
and Corollary (Ll). Furthermore by Theorem (5) the equality holds if and only if 1 :-:;; k :-:;; p and J1. is a g-maximal weight (see §3.2). But since w is decomposable and also 0<1> £;; b this is the case, by Proposition 2, (6) and Theorem (5) if and only if 0<1> is a commutative ideal in 6. But now assuming 1 :-:;; k :-:;; p (or else there is nothing more to prove) so that Ak = Mb then the Wi of Proposition (4) are decomposable vectors corresponding to 6-normal subspaces of 9 which are also commutative Lie subalgebras (by Theorem (5)). We will have proved (4.5.4) therefore (using Proposition (4)) if we can show that every such subspace equals 0<1>, for some 1 :-:;; i :-:;; q. This proves in particular that every commutative ideal of 6 is also of this form. But this is immediate from (4.5.1) and Remark (5) since one must have 0 n q = 0 for any such subspace o. (Indeed if 0 =I- x Eon q there exists cP E Ll+ such that [x, e",l = cp(x)e", where cp(x) =I- O. But then e", E 0 which contradicts the fact that a is commutative.) Finally Vw , is not equivalent to VWj for i =I- j, even assuming the k for Wi and Wj are possibly different, by Theorem (7) since a
REFERENCES 1. A. BOREL: Groupes lineaires algebriques, Ann. Math., Princeton 64 (1956), 20-82. 2. B. KOSTANT: Lie algebra cohomology and the generalized Borel-Weil theorem, Ann. Math, Princeton 74 (1961), 329-387. 3. J. L. KOSZUL: Homologie et cohomologie des algebres de Lie, Bull. Soc. Math. Fr. 78 (1950),65-627. 4. A. I. MALCEV: Commutative subalgebras of semi-simple Lie algebras, Izv. Akad. Nauk SSR, Ser. Mat. 9 (1945),291-300 (Russian); Translation No. 40, Series 1, American Mathematical Society (English).
Massachusetts Institute of Technology, Cambridge, Massachusetts, U.S.A.
481
Reprinted from the Proceedings of the United States-Japan Seminar in Differential Geometry Kyoto, Japan, 1965 Nippon Hyoronsha, Co., Ltd. Tokyo, Japan
Orbits, Symplectic Structures and Representation Theory Bertram
KOSTANT
We introduce a general approach to unitary representations for all Lie groups. An underlying feature is a study of sympletic manifolds X2n (i. e. there exists a closed non-singular 2-form on X). If [w] e H2(X, R) is an integral class there is an associated affinely connected Hermitian line bundle L over X which is unique if X is simply connected. Given a complex involutory totally singular distribution F" on X there is an associated cohomology theory HeLF) for the sheaf L.F of local sections of L which are constant along F. There then exists a Lie subalgebra a of the Lie algebra (under Poisson bracket) C of all smooth functions on X which naturally operates on HeLF)' For an element of C to operate on H(LF) in the case when X is a cotangent space is the inverse operation of taking the symbol of a differential operator. In general it is analogous to what the physicists call quantizing a function. Next one gives a complete classification of all sympletic homogeneous spaces for all Lie groups. They are related to orbits in the dual of the Lie algebra. The correspondence with an orbit maps the Lie algebra of the group into a (see above) and yeilds a representation of the group on H(L.F). The theory thus obtained embraces the Borel-Weil theory for compact groups, the Kirilov theory for nilpotent groups and Harish-Chandra-Gelfand theory for semi-simple groups.
Massachusetts Institute of Technology
71
B. Kostant, Collected Papers, DOI 10.1007/b94535_20, © Bertram Kostant 2009
482
Reprinted from Proceedings of Symposia in Pure Mathematics Volume 9 Algebraic Groups and Discontinuous Subgroups Copyright by the American Mathematical Society, 1966
Groups Over Z BY
BERTRAM KOSTANT 1. Preliminaries.
1.1. Let C be a commutative ring with 1. Let A be a co algebra over C with diagonal map d: A ~ A ®c A (it is assumed A has a counit B : A ~ C) and let R be an algebra over C with multiplication m: R ®c R ~ R (it is assumed R has a unit p : C ~ R). Then one knows that HomdA, R) has the structure of an algebra over C with unit where if f, g E HomdA, R) the product f * g E HomdA, R) is defined by f*g = m c(f@ g) cd. That is, one has a commutative diagram
In particular if we put R = C the dual A' = HomdA, C) has the structure of an algebra. Now assume that A is a Hopf algebra (A is an algebra and coalgebra such that d and B are homomorphisms and Bp is the identity on C). By an antipode on A we mean an element (necessarily unique if it exists) S E HomdA, A) such that I * S = S * I = B where I is the identity on A and * is as above with A taken for R. From now on Hopf algebra means Hopf algebra with antipode. 1.2. Now assume A is a Hopf algebra over C and R is any commutative C-algebra. Then if GR
=
U E HomdA, R)I f
is an algebra homomorphism}
one sees immediately that GR is a group under
* where
for any f
E
GR , a E A.
Thus one has a functor R ~ GR from all commutative algebras over R into groups and the functor is represented by A. Now if C is the set of integers Z then we may drop the word algebra so that R ~ GR is a functor from all commutative rings R to groups. 90 B. Kostant, Collected Papers, DOI 10.1007/b94535_21, © Bertram Kostant 2009
483
GROUPS OVER
Z
91
EXAMPLE. If A = Z[X ij , liD], i,j, = 1,2"", n, where the Xij are indeterminates and D = det(Xij), then A is a Hopf algebra over Z where
dX ij =
L X ik ® X kj , k
so that dD = D ® D. Also c(Xij) = 0 and s(X i ) = (_l)i+ j cofactor XjdD. Here G R = GI(n, R) for any commutative ring R. In the example above if one replaces A by its quotient with respect to the ideal generated by D - 1 then one obtains G R = Sl(n, R) for any commutative ring R. More generally for any semisimple Lie group G we will define a Hopf algebra Z(G) over Z with the following properties: (1) Z( G) is a finitely generated commutative integral domain; (2) for any field k k(G)
= Z(G) ®z k
is an affine algebra defining a semisimple algebraic group over k which is split over k, and is of the same type as G; (3) Q(G) defines Gover Q, where Q is the field of rational numbers. 1.3. From now on C = Z. Let B be a Hopf algebra over Z. An ideal Is; B will be said to be of finite type if BII is a finitely generated free Z-module. If I and I' are of finite type then the kernel I 1\ I' of the composed map
B~B ® B~BII ® BII' is again clearly of finite type defining an operation on the set of all such ideals. A family F of ideals of finite type will be said to be admissible if (1) nIEFI = (0); (2) s(I) E F for all I E F; (3) F is closed under I\. Now given such a family put AF
= {J E Hom(B, Z)IflI = 0 for some IE F}.
It is immediate then that AF has the structure of a Hopf algebra over Z. The multiplication in AF is defined as the transpose of the diagonal map in B. (It exists since F is closed under 1\.) The diagonal map in AF is defined as the transpose of the multiplication in B. (It exists since each f E AF vanishes on an ideal of finite type in B.) The antipode is simply the transpose of the antipode in B. (It exists since F is closed under s.) 1.4. Now let G be a complex semisimple Lie group and let 9 be its Lie algebra. Let V be the universal enveloping algebra of 9 so that V is a Hopf algebra over C
where
dx=x®l+l®x
484
BERTRAM KOSTANT
92
for any x E g. Also c; is given by c;(x) = 0 for any x E 9 and s is the anti-automorphism of U defined by s(x) = - x for any x E g. We will now define a Hopf algebra B over Z where B s U. The family of ideals F will be defined by G and one puts Z(G)
=
AF •
2. The definition and structure of B. Let 1) be a Cartan sub algebra of 9 and let d be the corresponding set of roots. Chevalley has shown (see [1]) the existence of a set of root vectors e"" cP E d, such that if cP, 1/1, cP + 1/1 E d then
[e"" e",l = ±re",+", where r E Z + (the set of nonnegative integers) is the minimum integer such that (ad e_",Ye", = 0 and if h", = [e"" e_",] then
cP(h",) = 2. We fix the e", as above and put gz equal to the Z span of all the e", and h", for cP E d. We recall some facts from [1] which, in fact, are easy to check. Let d+ be a system of positive roots and let II = (1J(1, ... , 1J(1) be the corresponding set of simple roots. Put hi = hai , i = 1,2,···, I, for simplicity. Then one has PROPOSITION 1. The elements h 1 , · · · , hI together with all e"" cP E d form a free Z-basis of gz. REMARK 1. Proposition 1 is of course only really a statement about the Z-span of the h", and the statement is of course well known. Now it is clear that gz is a Lie algebra over Z. Somewhat less obvious is the following fact of [1]: PROPOSITION 2. gz is stable under (aded>t/n! for any cPEd and nEZ+. REMARK 2. If h, e and f is a basis of the Lie algebra of SI(2, C) where [h, e] = 2e, = - 2f and (e,f) = h then Proposition 2 in essence reduces to the following fact: If Vb· .. ,Vk is a basis of an irreducible SI(2, C) module consisting of h-eigenvectors such that [h,f]
then the Z-span of the Vj is stable under em/m! and r/n! for all n, mE Z +. We now define B to be the algebra generated over Z by all elements e~/n! E U for all cPEI1 and nEZ+. 2.2. To prove that B is a Hopf algebra over Z with suitable properties we shall need some multiplication relations in U.
485
93
GROUPS OVER Z If h, e E 9 where [h, e] = Ae for some scalar }. then one easily establishes
(2.1.1) for any mE Z + and polynomial P E C[X]. N ow if U E U is arbitrary and mE Z + put Cum . =
u(u - 1)··· (u - m m.,
+
1)
.
Somewhat less trivial than (2.1.1) is the following useful relation among the generators of the Lie algebra of SI(2, C). LEMMA 1. Let h,e,jEg where [h,e] = 2e, [h,j] = -2fand [e,j] = h. Then for any n,mEZ+ one has k
f n- j
em-j
L (n _ J..),Ch-m-n+2j,j(m _")' J.
j=O
where k is the minimum of n and m. PROOF. One first of all proves directly from the bracket relation that
Lemma 1 is then just an exercise using (2.1.1), the relation above, and induction on m. 2.3. A sequence of C-linear independent elements to u(n) E U, n = 0,1,2,' . " where u(O) = 1, is called a sequence of divided powers in case du(n) =
m
L
u(j) ® u(n- j)
j=O
for all n. It is clear of course that the Z-space of the u(n) is a coalgebra over Z. EXAMPLE. If X E 9 and U
where N = (n b · · · , nk ) E Z"r, the UN over all N E Z"r are C-linearly independent. Let V be the Z-span of all UN' It is then clear that V is a coalgebra over Z and if D = Homz(V, Z) then D, as in §1.1, has the structure of a commutative algebra. But the point is that the algebra structure on D is particularly easy to describe. Let (Xi E D, i = 1,2,"', k, be such that '}'i(UN ) = for all N except Yi(UP») = 1.
°
486
94
BERTRAM KOSTANT
We leave it as an exercise to prove PROPOSITION 3. For any N = (n l , ... , nk ) E Z~ let YN = Y1' ... yk k • Then one has YN(UM) = 0 unless M = Nand YN(U N) = 1 so that D is the ring offormal power series D
=
Z[[Yb ... ,ydJ.
2.4. Now introduce the partial ordering in A where cP < IjJ in case IjJ - cP can be written as a sum of positive roots. Then simply order A+ so that A+ = (CPl, CP2, ... , CPr) where CPi < CPj implies i ~ j. Let n be the complex nilpotent Lie algebra spanned by all e", where cP E A+ and let U(n) ~ U be the universal enveloping algebra of n. In each r-tuple M = (m b · · · , mr) where miE Z+ put
so that the elements eM form a Birkhoff-Witt basis of U(n). Now let E be the Z-algebra in U(n) generated over Z by and nE Z+.
e~/n!
for all cP E A
LEMMA 2. The elements eM' over all MEZr+,for afree Z-basis of E. PROOF. Let El be the Z-span of all eM for ME zr+. Since the eM are independent over C they certainly form a free Z-basis of El and El ~ E. Since El contains the generators of E, to prove El = E we have only to show that El is closed under multiplication. We first observe that for any 1 ~ j ~ r there exists Sj E Homdg, C) such that (1) Sj vanishes on all root vectors e"" (2) sih",) = 1 and (3) Sj takes values in Z on gz. Indeed this is clear from Proposition 1 since any root, e.g., CPj can be embedded in a system of simple roots. Now consider the adjoint representation of n on g. Extending to U(n) one has that 9 is a U(n) module. If F = Homc(U(n), C) and 1 ~ j ~ r let jj E F be defined by
jj(U) = siu . L",) for any UE U(n). If M j = (ml' ... , mr) is defined by mi = 0 for i ¥- j and mj = 1, then clearly jj(e M) = 1, that is, jj(e",) = 1. On the other hand if one orders Z'+ lexicographically it is immediate that jj(e M) = 0 for all M> M j . But now by Proposition 2 jj must take values in Z on E. Now since U(n) is a co algebra F is an algebra over C. For any N = (n l ,· .. ,nr) put fN = ff' ... f~r E F. But now since E is the algebra generated over Z by all e~/n! it follows that dE is in the Z-span of all elements in U(n) ®c U(n) of the form U ® v where Ulv E E. Consequently fN also takes values in Z on E for any N E zr+. But El ~ E and by Proposition 3 one has fN(eN) = 1 and fN(e M) = 0 for all M > N.
487
GROUPS OVER Z
95
Now assume E1 is not an algebra. Then there exists N, M such that eNe M¢ E 1. That is, since the ep, P E Z,"+ are a C-basis of U(n) and one writes
there exists Cp such that Cp ¢ Z. Let L be minimal with this property. But then fL(eNe M)¢ Z. This however contradicts the fact thatfL takes integral values on E. REMARK 3. We note here that Lemma 2 may be strengthened in that the same conclusion is true when we use any ordering in 11+. Indeed iffM is defined in the same way as eM except with respect to a different ordering in 11+ and r
IMI=
L
mi
i= 1
for M = (m b · .. ,mr) then by the Birkhoff-Witt theorem there exists M' E zr+ such that IM'I = IMI and
eM - fM' = LcNe N where the sum is over N such that INI < IMI. But the CN lie in Z by Lemma 2. The result then follows by induction on IMI. 2.5. If X is an indeterminate one knows that C X •n for all n E Z + form a free Z-basis of the Z-ring R of all polynomials p in C[X] such that p(n) E Z for all n E Z. Since CX-m,n E R for any mE Z and is of degree n it is clear that the polynomial CX-m,n is an integral combination of CX,d for 0 ~ j ~ n. Now for any K = (kb ... ,kd E zt+ let hK = C hlok1 ... Ch/,k/' It is then clear that the hK over all K E zt+ is a C-basis of the universal enveloping algebra U(g) of g. On the other hand from above and §2.3 it is also clear that the Z-span H of all hK is a Hopf algebra over Z. Also from above, H contains Ch,-n"k for any kE Z+ and miEZ. We have defined eM E U(n) for any ME zr+. Now similarly define
for any N E zr+. Recall that B is the Z-algebra genetated over Z by all
e~/n!
for ¢
E
11, n E Z +.
THEOREM 1. The elements
fNhKe M for all N,
ME
Z,"+ and K
E
ZI+ form a free Z-basis of B.
PROOF. For convenience put n = 2r + I and for any P E Z"+- write P = (N, K, M) and put bp = fNhKeM' By the Birkhoff-Witt theorem it is clear that the bp form a C-basis of U. Let Uz be the Z-span of all bpo We first show that Uz 5; B. For this it is clearly enough to show that if h = hi' 1 ~ i ~ I and k E Z + then Ch,k E B. Put e = elZ , and f = LlZi so that h, e and f satisfy the conditions of Lemma 1.
488
96
BERTRAM KOSTANT
Assume inductively that Ch,i E B for all j < k. Then by Lemma lone has (ekjk!)(fkjk!) = Ch,k plus terms all involving ePjp!, r/q! and Ch-m,i where p, q E Z +, mE Z and j < k. By induction therefore Ch,k E B so that V z c;; B. With the same definition of h, e,j as above we now show that V z is stable under right multiplication by en In!, f" In! and Ch,n for any n E Z +. For the case of enln! the result is immediate by Lemma 2. For Ch,n the result follows from (2.1.1) since cP(h) E Z for all roots of cP E ~ so that eMCh,n = Ch-m,neM for some mEZ. Finally we want to show bpf"ln! = fNhKeMf"ln! lies in V z for all P. Since the argument above shows that V z is stable under left multiplication by fN and hK it is enough to show that eMfnln! E V z for all M. But by Remark 3 we can change the order of the roots in ~+ without changing E. Order the roots in ~ + so that cPr = (Xi and let S be the set of all ME zr+ where mr = O. We must therefore show
for all MES, m, nEZ+. But by Lemma 1, (emjm!)(fnjn!) can be rewritten as an integral sum of elements of the form
p
ej
-C i! h,k )!'
Hence we have only to show eMf" In! E V z where ME S. But now one knows that the set (cPl"", cPr-l, - (Xi) forms a new system of positive roots (obtained from ~+ by the reflection corresponding to (XJ Hence Lemma 2 and particularly Remark 3 apply to this new system. Thus eMf"ln! can be written as an integral sum of elements of the form (fi/j!)eN where again N E S. But these all lie in V z. Thus V z is stable under right multiplication by enln!,f"ln! and Ch,k' By symmetry the same is true for left multiplication. Now consider the adjoint representation of 9 on V. This extends to V so that V is a V-module and if Vi is the finite dimensional subspace spanned by all products of 9 with itself at most j times then one knows that Vi is a V -submodule. It is also clear that if V~ = V z (\ vj then V~ is a Z-form of V z with a free Z-basis consisting of all b p where !PI ~ j. But if x E 9 and u E V then ad x(u) = xu - ux. Hence
±
n
i
(-lY e ad(en)U = n! i=O (n - i)!
u~. j!
Thus V~ is stable under ad(enln!) and similarly ad(f"ln!) for all n. It follows therefore if 'Tt i (recall e = ea ) is the representation of SL(2, C) on V defined by ad h, ad e and adf and we let (Ji = 'Tti(e12 - ell) where eii' i = 1,2, are the matrix units in Ml(C) then V~ and hence V z is stable under (Ji' If X is the group generated by the (Ji for all i then V z is stable under X and one knows there is a
489
GROUPS OVER Z
97
homomorphism (j -+ (j of X onto the Weyl group W such that (jeq, = ± eiTq, for all (j E X. Since every root is W -conjugate to a simple root it follows therefore that V z is stable under right multiplication by e~/n! for all cjJ E d and n E Z +. This implies V z = B. 2.6. If we regard V(1») as the algebra of all polynomials on the dual space 1)' to 1), then for any f E 1)' one has that
Thus if L s;: h' is the group of all integral linear forms on 1) then hM(f) E Z for all ME ZI+. In fact, using the standard basis of L it follows easily that H is exactly the set of all P E V(1») which take integral values on L. Furthermore (since the same is true for Rand Z; see §2.5) given any finite subset F s;: L and .Ie E F there exists P E H such that p(.Ie) = 1 and p(Jl) = 0 for U E F and Jl #- .Ie. Now assume that Vis an arbitrary finite dimensional V-module. Let d(V) s;: L be the set of weights of V and for each Jl E d(V) let VI' be the corresponding weight space. A Z-form Vz of V (V = Vz Q9 z C) is called admissible if it is stable under B. COROLLARY 1 TO THEOREM 1. There exists an admissible Z-form Uz in V. Moreover if Vz is any admissible Z-form in V and V~ = Vz (\ VI' for Jl E d(V) then Vz
= EB
V~.
I'Et.(V)
PROOF. To prove the existence of an admissible Z-form it is enough to assume V is V-irreducible. Let v be a highest weight vector and put Vz = B . v. Since fM· v#-O for only a finite number of M it is clear that Vz is finitely generated over Z, stable under B and generates V over C. Furthermore Vz is a direct sum of the V~ = Vz (\ VI' for JlE d(V). To prove Vz is a Z-form of V we have only
to show that if C b • .. , Ck E C are independent over Z and VI'···' Vk E V~ are such that L CiVi = 0 then one already has Vi = 0 for all i. Indeed if, say, VI #- 0, there exists pEE (see §2.4) of weight .Ie - Jl (where .Ie is the highest weight of V) such that p. VI #- O. But for all i, p. Vi = miv for some mi E Z since p. Vi is of the form qiV where qi E H by Theorem 1. Hence
contradicting the fact that the Ci are Z-independent since we have m i #- O. Thus Vz is a Z-form of V. Now assume Vz is any Z-form of V. For each Jl E d(V) let PI' E H be such that PI'(Jl) = 1 and ply) = 0 for all Y E d(V), Y #- Jl. But then LI' PI' operates as the identity on Vand if WE VZ and wI' = PI'· W then W = Ll'wl' and wI' E V~. This proves the direct sum decomposition stated in the corollary. It follows from its definition but clearer from Theorem 1 that B is a Hopf algebra over Z.
490
98
BERTRAM KOSTANT
THEOREM 2. If J s V is any ideal offinite codimension in V then 1= B n V is an ideal of finite type in B. PROOF. Put V = VII so that by left multiplication V is a finite dimensional V -module. Since an admissible Z-form exists B is represented by m x m matrices with coefficients in Z where m = dim V. This implies J is an ideal of finite type in B. Now let A index all equivalence classes of finite dimensional modules for G. Regard these as modules for V and let J a S V, for (J. E A, be the corresponding kernels. If Ia = J a n B then it follows easily from Theorem 2 that the la' (J. E A, form an admissible family F of ideals of finite type in B. One puts Z(G) = A F , (see §1.4) defining the Hopf algebra Z(G). If V is anyone of these modules and Vz is an admissible Z-form in V with Z basis Vi and Wj is the dual basis then one always has hj E Z(G) where
hiu) = , hi of gz and the Yi define a dual basis to this basis of gz in the order indicated by Theorem 1. Then (by §2.3) if yp = yfl ... y~n where P = (Pb' .. Pn) one has yp(bQ) = ()PQ. Furthermore Z(G) S Sand S is the ring of formal power senes
S = Z[[YI ... Yn]]. REFERENCE
1. C. Chevalley, Sur certaines groupes simples, T6hoku Math. J. (2) 7 (1955), 14-66.
491
“This page left intentionally blank.”
Kostant’s Comments on Papers in Volume I 1. Holonomy and the Lie Algebra of Infinitesimal Motions of a Riemannian Manifold, Trans. Amer. Math. Soc., 80 (1955), 528–542. This paper develops a relation between the Lie algebra of Killing vector fields on a connected Riemannian manifold and the holonomy Lie algebra s at a point o ∈ M. If X is a vector field (v.f.), let A X be the (1,1) tensor field on M defined so that A X Y = −∇Y X and let a X = (A X )o . If X is a Killing field, then exp ta X tranforms, at time t, the effect of parallel transport of To (M) along the trajectory of o, generated by X , to the flow of To (M) along this trajectory. (The notation A X seems to have been later retained by many authors). One result in [1] is Theorem 1.1. If M is compact, then a X ∈ s. The proof of Theorem 1.1 depends on a novel use of Green’s theorem (see MR0084825 (18,930a) of K. Yano). Theorem 1.1 is cited as Theorem 4.5, p. 247 in [Kobayashi–Nomizu, I]. Let h ⊂ End To (M) be the Lie algebra generated by all a X for X a Killing field on M. The main result (Theorem 4.5) in [1] generalizing E. Cartan’s theorem (for symmetric spaces) is Theorem 1.2. If M is a compact homogeneous space with an invariant Riemannian metric, then h = s. This result was cited as Theorem 4.7, p. 208 in [K-M, 2] and in K. Nomizu’s Bourbaki Seminar No. 98. It also appears in [L] (see (13-1), p.110.) In general, in [L], Lichn´erowicz refers to the group corresponding to h as the Groupe de Kostant.
2 and 3. On the Conjugacy of Real Cartan Subalgebras I, PNAS, 41, No. 11 (1955); and On the Conjugacy of Real Cartan Subalgebras II, (1955). The problem of the classification of the conjugacy classes of Cartan subalgebras in a real simple Lie algebra was solved in these papers. Due to later work of Harish-Chandra the classification took on added significance since there is a “series” of group representations associated with each conjugacy class. Both papers were submitted by NAS member Saunders Maclane for publication in PNAS. The second paper contains the tables of the conjugacy classes. However the PNAS said it could not print the second paper because of the complicated nature of the tables. A preprint of the second paper was widely circulated at
493
494
Kostant’s Comments on Papers in Volume I
the time. See e.g., the letter to the editor by A.J. Coleman in Volume 44 (1997), No. 4 of the Notices of the AMS. At some later point M. Sugiwara independently published a list of the Cartan subalgebras. In a letter to me he verified that our lists were identical. The tables in Part II contain more than just a listing of the Cartan subalgebras. Any Cartan subalgebra decomposes into a direct sum of an elliptic part and a hyperbolic part. The tables list the centralizers of both of these parts. The tables are printed in the present volume. We wish to thank Fokko du Cloux for recently checking the validity of the tables and making (mostly misprint) corrections. 4. On Invariant Skew-Tensors, Proc. Nat. Acad. Sci. USA, 42, No. 3 (1956), 148–151. This paper readily establishes that the holonomy group of n-sphere with an arbitrary Riemannian metric is necessarily the full rotation group S O(n). I.M. Singer had previously proved this result if n is even. Let o ∈ S n and let G be the holonomy group at o. Any invariant of G operating on ∧To∗ (S n ) defines (by parallel transport) a harmonic form on S n . Since the Poincar´e polynomial of S n is 1 + t n it follows that G has no nontrivial invariants in ∧k To∗ (S n ) for 0 < k < n. We show that this alone implies that G = S O(n) (see Corollary 2.2 p.151) except for the possible special case where n = 5 and G = S O(3). Although not shown in the paper this special case cannot occur since G = S O(3) is not transitive on the unit sphere in R5 . Indeed the non-transitivity implies, by M. Berger’s theorem, that S 5 , with the given metric must be locally symmetric with S O(3) as holonomy group. But the simple connectivity of S 5 then implies S 5 = SU (3)/S O(3) which is clearly false.
5. On Differential Geometry and Homogeneous Spaces I, Proc. Nat. Acad. Sci. USA, 42 (1956), 258–261. Assume G is a compact connected Lie group and H ⊂ G a closed subgroup. Let M = G/H be a given a G-invariant Riemannian metric and let o ∈ M be the “point” H . Let g = Lie G and h = Lie H . Let p be an Ad H invariant complement of h in g so that we may identify p = To (M). Let s be the holonomy algebra at o so that s ⊂ End p. Assume that the Levi-Civita connection on M is a canonical affine connection of the first kind, in the sense of Nomizu, with respect to the decomposition g = h + p and let P : g → p be the projection with respect to this decomposition. Let k be the Lie subalgebra of End p generated by P ad x P for all x ∈ g. Then the results of paper #5 imply
494
Kostant’s Comments on Papers in Volume I
495
Theorem 5.1. One has k = s. Let ⊂ Aut p be the full holonomy group at o and let e be the identity component of so that Lie e = k by Theorem 5.1. In general there are not many results about full holonomy groups like . The determination of is the main result (Theorem 2) of paper #5. It is stated here as Theorem 5.2 below. Let σ : H → Aut p be the isotropy representation. Theorem 5.2. One has σ (H ) ⊂ and = σ (H ) e .
6. On Differential Geometry and Homogeneous Spaces II, Proc. Nat. Acad. Sci. USA, 42 (1956), 354–357. Let the notation be as in comments in paper #5 except that there is no assumption of compactness for G. What is referred to as a canonical affine connection of the first kind in the sense of Nomizu in paper #5 is referred to as naturally reductive in [B], See 7.84 Definition, p. 196 in [B]. As referenced as Theorem 7.85 (Kostant) in [B], p. 196 we have established, in paper #6, a necessary and sufficient condition that the affine connection on M corresponding to an Ad H -invariant bilinear form B on p be naturally reductive. The condition is that B should suitably extend to the g ideal g = p + [p, p] so as to be g invariant. This theorem plays a significant role in [D-Z]. See p. 4 in [D-Z].
7. On Holonomy and Homogeneous Spaces, Nagoya Math. Jour., 12 (1957), 31–54. In this paper we deal with the question of holonomic irreducibility for compact Riemannian homgeneous spaces. Let G be a compact connected Lie group and let H be a closed subgoup. Let M = G/H and assume M is given an arbitrary G-invariant Riemmannian structure. First of all we establish a statement similar to that of Theorem 5.2 above without the assumption that the corresponding affine connection is a canonical affine connection of the first kind in the sense of Nomizu. The main result in paper # 7 is Theorem 7.1. Assume χ(M) = 0 where χ (M) is the Euler characteristic. Assume also that G operates faithfully on M. Then M is holonomically irreducible if and only if Lie G is simple.
495
496
Kostant’s Comments on Papers in Volume I
This result extended a theorem of Matsushima and Hano in [M-H] which established only the “if” implication. Contrary to a conjecture of Nomizu, Theorem 7.1 is not true if the condition χ(M) = 0 is omitted. A counterexample to this conjecture is given in §3 of paper #7. See also p. 377 in [K-M,2]. The example gives the existence of an interesting case where Lie G is simple but M is holonomically reducible (e.g., M = S O(7)/SU (3)). The example also proves that Theorem, parts 1o and 3o , p. 1413, in Lichn´erowicz’s paper [L,2] is false. It is also proved in paper #7 that if χ (M) = 0 and Lie G is simple, then the affine connection uniquely determines the metric up to a scalar multiple.
8. A Theorem of Frobenius, a Theorem of Amitsur–Levitski and Cohomology Theory, J. Math. and Mech., 7, No. 2 (1958), 237–264. I came upon the identity known as the Amitsur–Levitski theorem independently while at Berkeley. The following is a brief sketch of my connection with this result. In the late 1950s at Berkeley I became interested in Lie algebra cohomology. My interest in this subject was very much stimulated by Koszul’s beautiful thesis. I also became aware of work of Dynkin on formulas for the primitive cohomology classes for the unitary group. I was also becoming familiar with Weyl’s book, involving Young diagrams, on the classical groups. Frobenius, having determined the characters of the symmetric group, published a paper in 1899 giving the characters of the alternating group. Playing a key role here are the self-dual Young diagrams of size k. The number of such diagrams is equal to the number of partitions of k whose parts are odd and distinct. As established in paper #8 it is no accident that this number is also the k-th Betti number of U (n) for sufficiently large n. Paper # 8 is mainly devoted to establishing the equivalence of the following 3 results: (I) the Amitsur–Levitski theorem, (II) the formula of Dynkin, and (III) a (special case) of a character formula of Frobenius for the alternating group. The failure of S O(2n) to have a primitive cohomology class of degree 4n − 1 gave rise to a new standard identity for 2n × 2n skew-symmetric matrices. Having been encouraged by N. Jacobson, L. Rowen found a simple proof of this new identity in 1982. Paper # 8 also introduced some new trace identities for n × n matrices. These were later rediscovered by C. Procesi. See the comments in the last paragraph on p. 581 in the Review in Bull. Amer. Math. Soc.,Volume 43, No.4, October 2006, by Edward Formanek of a book by Kanel–Belov and L. Rowen. Paper #8 was also the subject of Bourbaki Seminar No. 243 given by J. Dieudonn´e.
496
Kostant’s Comments on Papers in Volume I
497
9. A Characterization of the Classical Groups, Duke Math. Jour., 25, No. 1 (1958), 107–124. The classical groups are given in terms of what might be described as a defining representation. In this paper we deal with the question of characterizing the defining representation from the general point of view of Cartan–Weyl theory. The main theorems assert, in effect, that if g is a reductive Lie algebra and π is an irreducible representation, then g is classical and π is defining if and only if there exists x ∈ g such that the rank of π(x) is sufficiently small (actually 2 or 1 depending on the classical group). Applications are given to the sectional curvature of symmetric spaces. Another theorem algebraically chacterizes the curvature tensor of a symmetric space X in terms of the operator it defines on ∧2 Tp (X ) for p ∈ X . This result was used by J. Simons in his tour de force classification-independent proof (Ph.D. thesis) of M. Berger’s theorem that the holonomy group of a non-symmetric irreducible Riemannian manifold is transitive on the unit tangent space at a point.
10. A Formula for the Multiplicity of a Weight, Trans. Amer. Math. Soc., 93, No. 1 (1959), 53–73. Paper #10 gives a highly referenced solution to a problem in representation theory. Introducing what is now generally referred to as the Kostant Partition Function P, a formula for the multiplicity of a weight of a compact Lie group was written in this paper in terms of P. This initiated a large number of works by many authors to write branching laws in terms of partition functions (e.g., Blattner formula, Steinberg formula, by Kac for symmetrizable Kac–Moody groups and its extension by Shrawan Kumar and Olivier Mathieu to arbitrary Kac–Moody groups). A tensor product formula, in terms of a left ideal of the enveloping algebra of the nilradical of a Borel subalgebra, was also introduced in paper #10. As mentioned by the authors, the proof of Theorem 2.1 in the fundamental paper (PRV), representations of complex semisimple Lie groups and Lie algebras by K. Parthasarathy, R. Rao and V. Varadarajan were based on this idea. Using weights of the form σρ − ρ where ρ has its usual meaning and σ is the Weyl group, another function Q was introduced in paper #10. The function Q seems to have been ignored by most mathematicians. Nevertheless Hans Freudenthal in his review of this paper described the establishment of the equality P = Q as being “really marvelous.”
11. The Principal Three Dimensional Subgroup and the Betti Numbers of a Complex Simple Lie Group, Amer. Jour. of Math., 81 (1959), 973–1032. Paper #11 was my first really incisive penetration into the structure of a simple
497
498
Kostant’s Comments on Papers in Volume I
complex Lie algebra g. At the time # 11 was written, there was a growing literature on the topology of compact Lie groups Borel, Bott, Koszul, etc). Much (Chevalley, (1+t 2m i +1 ) of a corresponding compact centered on the Poincar´e polynomial i=1 simple group. Here is the rank and the m i are called the exponents. I was particularly interested in determining how the root structure encodes the exponents. One ingredient in establishing some of the main results was the existence of a special three-dimensional simple Lie subalgebra (TDS) a of g, called the principal TDS, already introduced by Dynkin and de Siebenthal. Another ingredient was the special element of the Weyl group (unique up to conjugacy) which I called the Coxeter–Killing transformation but it is now referred to as a Coxeter element. Decomposing g into irreducible components under the adjoint action of a principal TDS it is proved in this paper that gi g = ⊕i=1
(11.1)
where dim gi = 2m i + 1. This result gave a general proof of a case-by-case observation of Arnold Shapiro. An element a in the adjoint group G of g which induces the Coxeter element on some Cartan subalgebra is called principal. It is proved that the set of principal elements is a single conjugacy class and that principal elements are regular and have order equal to the Coxeter number h. Furthermore it is proved that any regular element g in Ad g has order ≥ h and that equality occurs if and only if g is principal. Dynkin classified the (finite) set S of conjugacy classes of TDS in g. In #11 we proved that the map C→S (11.2) of the set C of conjugacy classes of nilpotent elements into S, defined by Jacobson– Morosov, is in fact a bijection — proving incidentally that C is finite. This result is often mistakenly attributed to Dynkin. In fact Dynkin did not deal with C. The elements of the nilpotent conjugacy class O which corresponds, via (11.2), to a principal TDS are called principal nilpotent. It is proved in #11 that O is open and Zariski dense in the nilpotent cone of g. Moreover it is shown that O has codimension in g and the centralizer ge of any e ∈ O is an abelian Lie subalgebra of dimension whose elements are nilpotent. Moreover if n is the nilradical of a Borel subalgebra of g it is shown that e ∈ O ∩ n if and only if the coefficients of e corresponding to simple root vectors in n are all nonzero. Let h be a Cartan subalgebra and let z ∈ g be given by z = c−ψ e−ψ +
ci eαi
(11.3)
i=1
where the coefficients ci and c−ψ are nonzero scalars, the eαi are root vectors corresponding to a choice of simple positive roots and e−ψ is a root vector corresponding
498
Kostant’s Comments on Papers in Volume I
499
to the negative of the highest root ψ. An element x ∈ g is called cyclic if x is not nilpotent and p(x) = 0 for all G-invariant polynomial functions p on g where deg p < h. It is proved in #11 that any cyclic element is regular semisimple and, up to a scalar multiple, any two cyclic elements are conjugate. Furthermore it is proved that x ∈ g is cyclic if and only if x is conjugate to an element z of the form (11.3). Let ω = e2 πi/ h . The significance of cyclic elements in #11 goes back to a result of A.J. Coleman which asserts that a Coxeter element has a unique eigenvector with eigenvalue ω and the eigenvector is regular. It is established in #11 that the eigenvector is cyclic and that any cyclic element x is such an eigenvector for a unique Coxeter element σ operating in the unique Cartan subalgebra hx which contains x. For x = z, given by (11.3), a σ -eigenbasis of hz is given (see Theorem 6.7 in No.11) using natural bases of ge and ge˜ , where e and e˜ are certain “dual” principal nilpotent elements associated to h. Paper #11 contains a number of other results, some which have to do with the orbits of a Coxeter element on the set of roots and cross sections to these orbits. It was presented by J-L. Koszul in the Bourbaki Seminar No. 191.
12. A Characterization of Invariant Affine Connections, Nagoya Math. Jour., 16 (1960), 35–50. It is a result of Nomizu that if M = G/H is a reductive homogeneous space, then there exists a G-invariant affine connection on M. I was curious about the opposite direction. Given, say, a simply-connected manifold M with an affine connection A, what is a geometric condition on A so that M admits a transitive Lie group G whose action preserves A. W. Ambrose and I. Singer in a 1950 paper in the Duke Math. Journ. gave a necessary and sufficient condition in the Riemannian case. In this paper we give a generalization of the Ambrose–Singer theorem. Our main result introduces a new idea about a pair of affine connections. If A and B are affine connections on a manifold M then, as one knows, A and B differ by a tensor field S of type (1,2). We then say that A is rigid with respect to B if S is covariant constant with respect to B. Our main theorem then asserts that if M is a simply-connected manifold with an affine connection A, then M is a reductive homogeneous space with respect to a connected Lie group G whose action leaves A invariant if and only if there exists an affine connection B on M such that (1) B is invariant under parallelism (i.e., its curvature tensor R B and torsion tensor T B are B-covariant constant), (2) A is rigid with respect to B, and (3) M is complete with respect to B (i.e., every B-geodesic may be extended for arbitrary large values of its canonical parameter). This result was cited in Volume II of S. Kobayashi and K. Nomizu, Foundations of Differential Geometry, Interscience Publishers, 1969. See p. 376.
499
500
Kostant’s Comments on Papers in Volume I
13. Lie Algebra Cohomology and the Generalized Borel–Weil Theorem, Ann. of Math., 74, No. 2 (1961), 329–387. Let G be a complex semisimple Lie group and B ⊂ G a Borel subgroup. Let X = G/B and let E be a G-homogeneous complex line bundle over X , and let S E be the sheaf of germs of holomorphic sections of E. Motivated by Hirzebruch’s Riemann–Roch theorem, Raoul Bott determined the sheaf cohomology H (X, S E) and explicitly described its structure as a G-module. This result, now referred to as the Bott–Borel–Weil theorem (BBW), is called the Generalized Borel Theorem in paper #13. Let b = Lie B and let n be the nilradical of b. Let W be the Weyl group regarded as acting on b/n. Then Bott drew, as a consequence of BBW, the equality dim H q (n, Vλ ) = Number of elements of W of length q.
(13.1)
Let Vλ be an irreducible G-module with highest weight λ. Bott’s paper [B] was published in the Ann. of Math. 66 (1957), 203–248. In Remark 1 of [B], p. 247, Bott says that his proof of (13.1) is obviously unsatisfactory but that he knows of no direct argument to establish (13.1). The main result of paper #13 is a direct algebraic proof of a generalization of (13.1). If a is a complex finite-dimensional Lie algebra, a is its dual space, and V is a finite-dimensional complex a-module, then there is a coboundary operator of dV on C = ∧a ⊗ V (making the pair (C, dV ) into a cochain complex) whose derived cohomology is the Lie algebra cohomology H (a, V ). The approach to the determination of H (a, V ) in this paper is via a “Laplacian” operator L V . One introduces a Hilbert space structure on C, and then one defines a positive semidefinite operator L V on C by puttting L V = dV dV∗ + dV∗ dV where dV∗ is the Hermitian adjoint of dV . One then has a natural isomorphism Ker L V ∼ = H ∗ (a, V ).
(13.2)
This of course transfers the problem to the determination of Ker L V . In paper #13 we introduce a special class of Lie subalgebras of g. A Lie subalgebra in this class is called a Lie summand. If a ⊂ g is a Lie subalgebra, then a is a Lie summand if its orthogonal complement a0 with respect to the Killing form is again a Lie subalgebra. For any Lie subalgebra a ⊂ g and g module V we write dV as a difference of two simpler operators involving the full g-module structure on V and also involving a choice of a compact real form k of g (see Proposition 3.13). The choice of k also naturally defines a Hilbert space structure on C. But now if a is a Lie summand, one has a neat formula for L V involving a difference of two explicit positive semidefinite operators (see Theorem 4.4.) The nilradical of a parabolic subalgebra is a Lie
500
Kostant’s Comments on Papers in Volume I
501
summand. If u is the nilradical of a standard parabolic subalgebra p and we choose a = u and V = Vλ , then Theorem 4.4 yields an explicit determination of the spectral resolution of L V (see Theorem 5.7). This resolution involves the natural action of a Levi factor g1 of p on H (u, Vλ ). The choice of g1 also defines a now familar decomposition W = W1 W 1 of the Weyl group W . (This notational terminology, especially the superscripted W 1 , introduced in this paper, is now widely used.) Restricting consideration to Ker L V , one obtains Theorem 5.14 which describes H (u, Vλ ) as a g1 -module. For any σ ∈ W 1 let ξσ = σ (ρ + λ) − ρ, regarded ξ as a g1 -dominant weight. If V1 is a g1 -module, let V1 σ denote the primary g1 submodule corresponding to ξσ as a highest weight. Also for j ∈ Z+ let W 1 ( j) be the set of σ ∈ W 1 having length j. Then Theorem 5.14 asserts that H (u, Vλ )ξσ is nonzero and irreducible for any σ ∈ W 1 and in fact σ → H (u, Vλ )ξσ is a bijection of W 1 onto the set of all g1 -irreducible components of H (u, Vλ ). In particular H (u, Vλ ) is multiplicity-free. Degree-wise for j ∈ Z+ one has the direct sum H j (u, Vλ ) = H (u, Vλ )ξσ . (13.3) σ ∈W 1 ( j)
Finally the harmonic (Ker L V ) representative of the highest weight vector in H (u, Vλ )ξσ is explicitly determined and shown to be a completely decomposable element of C. The results above generalize Bott’s formula (13.1) and the latter is derived, by Bott, as a consequence of BBW. However, as pointed out by Wilfried Schmid in his article, The Mathematical Legacy of the Paper “Homogeneous Vector Bundles,” p. 40 in Volume I of Bott’s , Collected Papers, R. MacPherson Editor, Birkh¨auser, 1994, [BC], formula (13.1) is equivalent to BBW. See p. 42 in [BC]. In fact Schmid goes on to point out on p. 42 that my computation of L V “carries over directly to a computation of the Laplace operator on the Dolbeault complex of a homogeneous line bundle E −→ X .” Much of the remainder of paper #13 is devoted to applications of Theorem 5.14. One of the applications is a generalization of a theorem of Ehresmann. Consider the case where u is commutative (i.e., Y = G/P is a Hermitian symmetric space where P ⊂ G corresponds to p). Here, as in all cases, W 1 parameterizes the Schubert classes in Y . Take λ = 0 so that C = ∧u and ξσ = σ ρ − ρ for σ ∈ W 1 . We may identify u with the nilradical of the parabolic subalgebra p “opposite” to p noting that g1 is also a Levi factor of p . In this case dV = 0 so that H (u, Vλ ) = ∧u . Theorem 5.14, among other things, then implies that ∧u is a multiplicity-free g1 -module and σ → (∧u )ξσ
501
502
Kostant’s Comments on Papers in Volume I
is a bijection of W 1 onto the set of all g1 -irreducible components of ∧u. When g is classical this is then a result of Ehresmann. Another application is an extension of Weyl’s character formula to disconnected complex reductive Lie groups. Howard Garland extended the Laplacian idea of paper #13 to the affine Kac– Moody case and consequently gave a beautiful proof of the Kac–MacDonald denominator formula. More than that, the positive semidefinite nature of the Laplacian established remarkable inequalites. Garland’s result is a special case of a much more general result by Shrawan Kumar. In particular Kumar obtained a far-reaching infinite-dimensional analogue of Theorem 5.7 in paper #13.
14. (with G. Hochschild and A. Rosenberg), Differential Forms on Regular Affine Algebras, Trans. Amer. Math. Soc., 102, No. 3 (1962), 383–408. Assume R is a commutative ring over a field K . Let S = R ⊗ K R. Then the Hochschild homology H∗ (R) of R is Tor S (R, R) (using terminology of Cartan– Eilenberg). One then knows that H1 (R) is the R-module of formal differentials of R (the span of f dg , f, g ∈ R, with dc = 0 if c ∈ K ). Also H∗ (R) has a skewcommutative R-algebra structure. Under Noetherian and regularity assumptions, it is shown in Theorem 3.1 that H∗ (R) is projective, finitely generated over R and H∗ (R) is isomorphic to the exterior algebra over H1 (R).
(14.1)
The cohomology H ∗ (R) of R, by definition, is Ext∗S (R, R). Also H 1 (R) is the R-module of K -derivations of R. If E is the exterior algebra of H 1 (R) and (R) = Hom R (E, R), then (R) is the algebra of differential forms defined by R. On the other hand, one has a natural isomorphism h : H 1 (R) → Hom R (H1 (R), R).
(14.2)
Henceforth assume that the conditions imposed on R in Theorem 3.1 are satisfied. But then, by (14.1), this extends to an isomorphism E → Hom R (H∗ (R), R).
502
(14.3)
Kostant’s Comments on Papers in Volume I
503
Dualizing (14.3) one establishes the isomorphism H∗ (R) → (R)
(14.4).
See Theorem 5.20. The isomorphism (14.4) has generally been referred to as the HKR theorem. Subsequently a number of mathematicians have established (14.4) with different assumptions about the ring R. Perhaps most notable is the result of A. Connes proving (14.4) for the case where R is the ring of smooth functions on a manifold. In any event (14.4) has become the focus of active research. The ring (R) (or H∗ (R)) is itself a differential complex leading to the de Rham cohomology Hd R (R). Section 7 in this paper is devoted to obtaining the de Rham cohomology as an Ext functor on the algebra of differential operators of R. See Corollory 7.1. The motivation for this comes from Lie algebra cohomology. However this result seems to have attracted less attention than 14.4. One reason no doubt for this is the neat relationship found by A. Connes between de Rham cohomology and cyclic (co)homology.
15. (with G. Hochschild), Differential Forms and Lie Algebra Cohomology for Algebraic Linear Groups, Illinois Jour. Math., 6 (1962), 264–281. Paper #15 extends earlier results of Hochschild on algebraic groups to the homogeneous case. Let R be the affine ring of an affine algebraic group over an algebraically closed field F. Assume that K is a reductive subgroup of G. Under right translations let R K be the subring of K -invariant functions. It is proved in paper #15 that G/K is an affine variety and R K is its affine ring. See Theorem 5.1. If F has characteristic zero and G itself is reductive, then it is proved in this paper that the deRham cohomology of R K is the same as the relative Lie algebra cohomology. See Theorem 3.2. This says in effect that if, say, F = C, then the algebraic holomorphic differentials on the affine variety G/K “sees” the topological cohomology of G/K . One keeps in mind here that this is true in spite of the fact that there is no Poincar´e lemma for algebraic holomorphic differentials. Later, a general theorem of Grothendieck established this fact for any nonsingular affine variety. Grothendieck cited this paper as an inspiring influence for his very general result. 16. Lie Group Representations on Polynomial Rings, Bull. Amer. Math. Soc., 69, No. 1 (1963), 518–526. Paper #16 is a Bull. Amer. Math. Soc. announcement of some of the results in paper #17.
503
504
Kostant’s Comments on Papers in Volume I
17. Lie Group Representations on Polynomial Rings, Amer. J. Math., 85 (1963), 327–404. Paper #17 establishes a rather large number of results. We will take this opportunity to revisit some of the main results and present them from a more modern perspective. Let g be a complex reductive Lie algebra and let S be the ring of polynomial functions on g. Let G be the adjoint group of g. Using methods of algebraic geometry, this paper is a study of the orbit structure of G on g and, contragrediently, the module structure of G on S. Let n = dim g, = rank g and let J = S G be the ring of polynomial G-invariants in S. One knows (Chevalley) that J is a polynomial ring C[u 1 , . . . , u ] where the u i are homogeneous algebraically independent polynomials, and if we write deg u i = m i + 1, then the m i are the exponents of g. If x ∈ g, let Ox = G · x so that Ox = G/G x where G x is the isotropy group at x. Let gx = Lie G x . Then dim Ox ≤ n − and in current terminology (which we retain here) x is called regular if dim Ox = n − . (At the time paper #17 was written the word regular was reserved for semisimple regular elements. See p. 356. The latter restriction is thus abandoned now.) Let r be the Zariski open set of all regular elements in g. If x ∈ r, then one easily has that gx is a commutative Lie subalgebra of dimension . Let x ∈ g. Then in paper #17 it is proved that x ∈ r ⇐⇒ (du 1 )x , . . . , (du )x are linearly independent. (See Theorem 9). Also let x = y + z be the Jordan decomposition of x where y is semisimple and z is nilpotent so that in particular g y is reductive and z ∈ g y . Then it is proved in paper #17 that x ∈ r ⇐⇒ z is principal nilpotent in g y . See Proposition 13. Now for any ξ = (ξ1 , . . . , ξ ) ∈ C let P(ξ ) = {x ∈ g | u i (x) = ξi , i = 1, . . . , }. In particular note that P(ξ ) is a closed G-stable subvariety of g, and if P ⊂ g is the nilcone, note also that P = P(0). Let O r (ξ ) = P(ξ ) ∩ r and if s is the set of semisimple elements in g, let O s (ξ ) = P(ξ ) ∩ s. In paper #17 we prove Theorem. There are a finite number of G-orbits in P(ξ ) for any ξ ∈ C . In particular the nilcone P has only a finite number of G-orbits. Moreover O r (ξ ) is the unique orbit of maximal dimension in P(ξ ). Furthermore O r (ξ ) is Zariski open in P(ξ ) and P(ξ ) = O r (ξ ),
504
Kostant’s Comments on Papers in Volume I
505
so that P(ξ ) is an irreducible variety of dimension n − . Also O s (ξ ) is the unique closed orbit in P(ξ ) and it is the unique orbit of minimal dimension in P(ξ ). For any x ∈ g one has x ∈ P(ξ ) if and only if the semisimple component of x (relative to its Jordan decomposition) is in O s (ξ ). See Theorem 3 and (3.8.9) in paper #17. Let (h, e, f ) be a standard basis of a principal TDS. Another result in paper #17 is a (now well known) sectioning of the G-action on r. Let a be an ad h-stable subspace complementary to [ f, g]. Then dim a = . An example of a choice of a is ge . Let v be the -dimensional affine plane defined by putting v = f + a. Also let u : g → C be the morphism defined so that u(x) = (u 1 (x), . . . , u (x)). Then the following result is established in this paper. Theorem. One has v ⊂ r. Furthermore the restriction u : v → C is an algebraic isomorphism. Moreover, v meets every regular G-orbit at exactly one point. In fact if ξ ∈ C , then u(O r (ξ ) ∩ v) = ξ. See Theorem 7 and Remark 19 . For any ξ ∈ C let J ξ be the ideal of codimension 1 in J generated by u i − ξi , i = 1, . . . , . One notes that J 0 is the augmentation ideal J + in J . Also J ξ S = (u 1 − ξ1 , . . . , u − ξ ). We prove (see Theorem 10) Theorem. For any ξ ∈ C , the ideal in J ξ S defines the n − -dimensional irreducible variety of P(ξ ) so that (a) P(ξ ) is a complete intersection and (b) J ξ S is a prime ideal. Furthermore O r (ξ ) is the set of simple points of P(ξ ) and the subvariety of nonsimple points of P(ξ ) has codimension of at least 2 in P(ξ ). Let h and b, where h ⊂ b, be a Cartan subalgebra and a Borel subalgebra. Let D ⊂ h∗ be the set of all dominant weights in the root lattice, and for any λ ∈ D let ν λ : G → Aut V λ be an irreducible representation of G with highest weight G λ. For λ ∈ D let λ be the dimension of the zero weight space in V λ . Let Diff+ be the space of all G-invariant constant coefficient differental operators on g with zero constant term and let G H = { p ∈ S | ∂ p = 0, ∀ ∂ ∈ Diff+ }.
505
506
Kostant’s Comments on Papers in Volume I
Then H is G-stable and the polynomials in H are called harmonic. One has (see Theorem 11 and (5.3.5)) Theorem. G-isomorphism
S is free as a J module and in fact multiplication defines a J ⊗ H → S.
Moreover, H is the span of all powers (e )k where e ∈ g∗ corresponds, under the Killing form isomorphism, to a nilpotent element of g . Furthermore H has finite multiplicities as a G-module and in fact the multiplicity of ν λ in H is λ . In addition, H is complementary to the prime ideal J ξ S in S, for any ξ ∈ C , so that H and S have the same restrictions to P(ξ ). Also the restriction of H to P (ξ ) is faithful. In particular the affine rings of P(ξ ), over all ξ ∈ C , even though not equivalent as rings, are all equivalent as G-modules. For any ξ ∈ C , let R(O r (ξ )) be the ring of regular functions on the quasiaffine variety O r (ξ ). Let x ∈ O r (ξ ) so that Ox = O r (ξ ) and hence as a G-module mult ν λ in R(O r (ξ )) = dim (V λ )G . x
(17.1)
Using a criterion for normality due to A. Seidenberg we establish one of the main theorems in paper # 17 (see Theorem 16). Theorem. P(ξ ) is a normal variety for any ξ ∈ C . In particular the nilcone is a normal variety. Furthermore O r (ξ ) “sees” its closure P(ξ ) in the following sense: Any everywhere defined function on O r (ξ ) extends to a regular function on P(ξ ) so that R(O r (ξ )) identifies with the affine ring of P(ξ ) (i.e., R(O r (ξ )) is an affine ring and P(ξ ) is its variety of maximal ideals). In particular if x ∈ r and λ ∈ D, then x dim (V λ )G = λ . (17.2) Also the restriction of functions defines a G-isomorphism H → R(O r (ξ )).
(17.3)
For λ ∈ D let Vλ be the dual space to V λ and let νλ be the representation of G on Vλ which is contragredient to ν λ . As above, let (h, e, f ) be a standard basis of a principal TDS. Since the zero weight spaces of Vλ and V λ are clearly the same dimension, and all other weights e are in the root lattice, it follows from (17.2) that VλG is λ -dimensional and admits a νλ (h/2) eigenbasis with nonnegative integral eigenvalues m i (λ), i = 1, . . . , λ . The m i (λ) are called generalized exponents (the usual exponents are the special
506
Kostant’s Comments on Papers in Volume I
507
case where g is simple and λ is the highest root). Let o(λ) = 12 λ(h). Taking the generalized exponents to be in nondecreasing order, one readily has m i (λ) ≤ o(λ) and equality occurs ⇐⇒ i = λ .
(17.4)
The generalized exponents “see” the degrees in which ν λ occurs harmonically (Theorem 17). Theorem. Let H = ⊕λ∈D H λ be the primary decomposition of H (with H λ corresponding to ν λ ). Further decomposing we can write, as a direct sum, λ Hiλ H λ = ⊕i=1
(17.5)
where the Hiλ are G-irreducible homogeneous components of H λ . Let n i (λ) be the degree of Hiλ and choose the ordering in (17.5) to be nondecreasing. Then, for i = 1, . . . , λ , n i (λ) = m i (λ).
(17.6)
In particular o(λ) is the highest degree in which ν λ occurs harmonically and this maximal occurrence has multiplicity 1. The next result (see Theorem 18) generalizes a theorem of A.J. Coleman which determines the eigenvalues of the Coxeter element in terms of the ordinary exponents. Theorem 18 in paper #17 also gives an expression for the corresponding eigenbasis. Theorem. Let x ∈ r and assume a · x = c x for some a ∈ G and c ∈ C× . x Let λ ∈ D. Then νλ (a) stabilizes and is diagonalizable in VλG . Furthermore the x eigenvalues of νλ (a)|VλG are cm i (λ) , i = 1, . . . , λ . In analogy with the “separation of variables” theorem for S(g), the final section (18) of paper #17 gives a “separation of variables” result for the enveloping algebra U of g. The result (see Theorem 21) asserts that multiplication defines a Gisomorphism Z⊗E →U
(17.7)
where Z = Cent U and E is the span of all powers ek where e ∈ g is nilpotent. Paper #17 was presented in a Bourbaki seminar given by Roger Godement.
507
508
Kostant’s Comments on Papers in Volume I
18. Lie Algebra Cohomology and Generalized Schubert Cells, Ann. of Math., 77 (1963), No. 1, 72–144. Paper #18 is Part 2 of a paper where Part 1 is paper #13. We will freely use here some of the notation in paper #13. The problem which is dealt with in paper #18 is an understanding and proof of the “strange equality” (terminology of Raoul Bott) (18.1) H q (n, M) = H 2q (X ; C) empirically observed as (15.3) in Bott’s paper, “Homogeneous Vector Bundles”. In (18.1) n is the nilradical of a Borel subalgebra of g, M is an irreducible G-module and X = G/B is the G-flag manifold. Since the left side of (18.1) has a basis parameterized by the Weyl group of G for any W , it suffices to consider (18.1) for the case where M is the trivial module. Paper #18 then generalizes the statement of (18.1). Going back to the notation of paper #18, first of all, W is the Weyl group of the pair (g, h). Next u is a standard parabolic subalgebra. That is, u is a Lie subalgebra of g containing b and U ⊂ G is the parabolic subgroup corresponding to u. Now X is the generalized flag manifold given by putting X = G/U . Also g1 is the Levi factor of u containing h and G 1 ⊂ G is the reductive subgroup of G corresponding to g1 . In addition n is now the nilradical of u so that u = g1 ⊕n is the Levi decomposition of u. Furthermore, n∗ is the unique ad g1 -stable complement of u in g so that (18.2) g = g1 ⊕ n ⊕ n∗ is a triangular decomposition of g. As in paper #13, the Weyl group W = W 1 W1 where W1 is the Weyl group for the pair (g1 , h) and if σ ∈ W 1 , then σ is the element of minimal “length” ((σ )) in the W1 left coset σ W1 . The use of the notation n∗ in (18.2) is suggestive in that n∗ identifies with the dual of n using the Killing form and the adjoint action of g1 on n∗ is contragredient to its action on n. On the other hand, recalling the notation of (13.3) above, we have proved in paper #13, using Lie algebra homology instead of cohomology (see Corollary 8.1 in paper # 13) that, as a g1 module, H∗ (n) is multiplicity-free and, for any q ∈ Z+ , Hq (n) is a direct sum of the primary (and irreducible) components, H∗ (n)ρ−σρ for σ ∈ W 1 (q). It then follows from Schur’s lemma that dim(Hq (n) ⊗ Hq (n∗ ))g1 = w1 (q)
(18.3)
where w1 (q) is the cardinality of W 1 (q). In fact Schur’s lemma implies that (Hq (n) ⊗ Hq (n∗ ))G 1 has, up to scalar multiplication, a natural basis hσ , σ ∈ W 1 ( j). On the other hand, recalling the K¨ahler structure on X , one has that H p,q (X ) = 0 if p = q, and using the Schubert classes of dimension p, one also has that dim H q,q (X, C) = w1 (q) and that H q,q (X, C) has a natural Schubert class
508
Kostant’s Comments on Papers in Volume I
509
basis xσ , σ ∈ W 1 (q). So as a far-reaching generalization of the observation (18.1) one should establish a natural isomorphism ψ : H q,q (X, C) → (Hq (n) ⊗ Hq (n∗ ))G 1
(18.4)
where, for any σ ∈ W 1 (q), up to a scalar factor, ψ(xσ ) = hσ .
(18.5)
Paper #18 accomplishes (18.4) and (18.5) in two steps. To deal with the first step assume C is a finite-dimensional Z-graded C vector space with a coboundary (raises degrees by 1) operator d and a boundary (lowers degrees by 1) operator ∂. One then has cohomology H ∗ (C, d) and homology H∗ (C, ∂). We say that d and ∂ are disjoint in case for any u ∈ C one has, d ∂ u = 0 implies ∂ u = 0 ∂ d u = 0 implies d u = 0. In such a case one readily shows that if S = d ∂ + ∂ d, then (a) Ker S is a space of d-cocycles and the corresponding map ψd,S : Ker S → H ∗ (C, d), is a degree preserving isomorphism.
(18.6)
Next (b) Ker S is a space of ∂-cycles and the corresponding map ψ S,∂ : Ker S → H∗ (C, ∂), is a degree preserving isomorphism.
(18.7)
In particular (18.6) and (18.7) then define a degree preserving isomorphism ψ∂,d : H ∗ (C, d) → H∗ (C, ∂).
(18.8)
We apply (18.8) to the following case: let r be the orthocomplement of g1 in g and let C = (∧r)G 1 . (18.9) Then, recalling the definition of relative Lie algebra cohomology, C is a cochain complex with coboundary operator d where H ∗ (C, d) = H ∗ (g, g1 ). In fact if K is a compact form of G, chosen so that if K 1 ⊂ G 1 where K 1 = K ∩ U , then X = K /K 1 and C identifies with the space of K -invariant differential forms on X and d is induced by the exterior derivative of differential forms. Thus (Cartan– Eilenberg–de Rham theory), H ∗ (C, d) = H ∗ (X, C).
509
(18.10)
510
Kostant’s Comments on Papers in Volume I
Now r = n ⊕ n∗ . Introduce a Lie algebra structure on r by retaining the given Lie algebra structures on n and n∗ , but now putting [n, n∗ ] = 0. Then ∧r is a chain complex for the Lie algebra homology H∗ (r) = H∗ (n) ⊗ H∗ (n∗ ).
(18.11)
But G 1 is a reductive group of automorphisms of the Lie algebra r and hence C is stable under the boundary operator. Let ∂ be the restriction of this operator to C so that, by the reductivity of G 1 , one has H∗ (C, ∂) = (H∗ (n) ⊗ H∗ (n∗ ))G 1 . The main problem of paper #18 is to establish the disjointness of d and ∂. This is done in Theorem 4.5 and Corollary 5.3.2 in paper #18 establishing the isomorphism (18.6). In fact ∧r is bigraded and C = ⊕q∈Z+ C q,q .
(18.12)
By Theorem 4.5 one, in fact, has the isomorphism (18.4) above. Note also that S is now a new kind of Laplacian operating on K -invariant differential forms on X and Ker S is the space of a new kind of harmonic forms on X . Let q ∈ Z+ . Then if s ∈ (Ker S)q,q (see Proposition 3.3.2) s is a K -invariant form on X of type (q, q) and, by Theorem 4.5, s ∈ H q,q (X, C) where s = ψd,S (s) (see (18.6)). Let σ ∈ W 1 (q) so that hσ is a “Schur’s lemma” basal element of (Hq (n) ⊗ Hq (n∗ ))G 1 . Let s σ ∈ (Ker S)q,q be defined so that ψ∂,d (sσ ) = hσ
(18.13)
where sσ = ψd,S (s σ ). But now (second step) (18.5) is established as soon as one proves that sσ = xσ (18.14) up to a scalar multiple. But this fact is proved in Theorem 6.15 of paper #18. The Schubert cells Vτ of complex dimension q are parameterized by τ ∈ W 1 (q). The proof of Theorem 6.15 is established by showing that the integral of s σ over Vσ is a positive number and showing that, if σ = τ , then s σ = 0. (18.15) Vτ
On the other hand (18.15) follows from a striking property of the differential form s σ , namely that s σ |Vτ is identically zero. (18.16)
510
Kostant’s Comments on Papers in Volume I
511
(See Corollary 6.15) and this property follows from the fact ∂s σ = 0. See Remark 6.15. The d, ∂-harmonic theory developed in this paper has been extended by Shrawan Kumar to any symmetrizable Kac–Moody setting. Thus, he has extended most of the results in this paper to any symmetrizable Kac–Moody flag varieties. He has used these results to show that any such flag variety is a formal space in the sense of rational homotopy theory.
19. Eigenvalues of a Laplacian and Commutative Lie Subalgebras, Topology, 13 (1965), 147–159. Paper # 19 is a 12-page paper written in 1965 which seems to have spawned a great deal of research activity focused on abelian ideals of a Borel subalgebra of a complex semsisimple Lie algebra, paricularly in the first decade of the twenty-first century. Some of the researchers (besides myself) are D. Peterson, R. Suter, D. Panyushev, G. Rohrle, P. Cellini, P. Papi and N. Kwon. There was a conference in Italy in October 2007 devoted to this subject. Let K be a compact semisimple Lie group and let k = Lie K . Then the negative of the Killing form defines a two-sided K -invariant Riemannian structure on K . Let Lap be the Laplacian with respect to this Riemannian structure and let g be the complexification of k. The exterior algebra ∧g identifies with the space of all complex valued left K -invariant differential forms on K . Furthermore ∧g inherits a natural Hilbert space structure {u, v}. Moreover if d is the operator on ∧g induced by exterior differentiation of forms and ∂ is the Hermitian adjoint of d, then ∧g is stable under Lap and one has Lap| ∧ g = d ∂ + ∂ d. Whereas Hodge theory focuses on Ker Lap | ∧ g, paper # 19 was motivated instead by the question of determining, degree-wise, the maximal eigenvalue of Lap| ∧ g. This quickly becomes a question in representation theory since if Cas is the Casimir operator on ∧g corresponding to the adjoint action of g on ∧g, then one readily has that Cas = 2 Lap| ∧ g. Let k ∈ Z+ . If u ⊂ g is a k-dimensional subspace, then ∧k u is a 1dimensional subspace of ∧k g. Let p be the maximal dimension of an abelian Lie subalgebra of g. The value of p has been determined by Malcev for all semisimple g (for example if g is of type E 8 , then p = 36.). We define the g-submodule Ak ⊂ ∧k g as follows: If k > p, then put Ak = 0. Otherwise let Ak be the span of all the 1-dimensional subspaces ∧k a where a is any k-dimensional abelian
511
512
Kostant’s Comments on Papers in Volume I
subalgebra of g. Let m k be the maximal eigenvalue of Cas| ∧k g. The following theorem is proved in paper #19. Theorem. For any k one has m k ≤ k.
(19.1)
Furthermore one has equality in (19.1) if and only if k ≤ p. Moreover in such a case the corresponding eigenspace for Cas | ∧k g is Ak . Finally if 0 = u ∈ ∧g is of the form u = x1 ∧ · · · ∧ xk for xi ∈ g, then u ∈ Ak if and only if the xi mutually commute. Let h be a Cartan subalgebra of g and let ⊂ h∗ be the set of roots for (g, h). For all roots ϕ let corresponding root vectors eϕ be chosen. Also choose a system of positive roots + thereby defining a Borel subalgebra b containing h. Let n be the nilradical of b. Simply order + and if ⊂ + we write = {ϕ1 , . . . , ϕk }
(19.2)
in increasing order. Let e ⊂ ∧k g be defined by putting e = eϕ1 ∧ · · · ∧ eϕk .
(19.3)
Any ideal v of b which is contained in n defines a subset ⊂ + of the form (19.2) and necessarily k v= C eϕi . (19.4) i=1
In such a case the G-submodule spanned by G · e of ∧k is irreducible and C e is the highest weight space. Thus =
k
ϕi
(19.5)
i=1
is the highest weight. Write = (v). Given two such ideals v1 , v2 , we show in this paper that v1 = v2 ⇐⇒ 1 = 2 (19.6) where for i = 1, 2, we have put i = (vi ). In particular, distinct such ideals define inequivalent irreducible representations. Now let C be the set of all abelian ideals a in b and let C(k) be the set of abelian ideals of dimension k. One has a ⊂ n for any a ∈ C so that the cardinality of C is
512
Kostant’s Comments on Papers in Volume I
513
finite by (19.4). Now for a ∈ C let Aa be the G-module generated by e(a) . Then if a ∈ C(k), (19.7) Aa is an irreducible G-submodule of Ak . ∞ Put A = k=0 Ak . The following result in paper #19 places the set of abelian ideals in b at center stage. Theorem A. A is a multiplicity-free G-module. Furthermore A= Aa
(19.8)
a∈C
is the unique complete reduction of A as a sum of irreducible G-modules. Degreewise, for any k ∈ Z+ , Ak = Aa (19.9) a∈C (k)
is the unique complete reduction of Ak as a sum of irreducible G-modules. The abelian ideals in b are characterized in paper # 19 by the following result. Theorem B. Let ⊂ + . Let k be the cardinality of and let the notation be as in (19.2). Then (with the usual present-day definition of ρ and the usual norm in weight-lattice) one has |ρ + ϕ1 + · · · + ϕk |2 − |ρ|2 ≤ k
(19.10)
and equality occurs in (19.10) if and only if is of the form = (a) for some a ∈ C(k). Interest in the subject matter of paper # 19 was considerably stimulated by the subsequent discovery, due to Dale Peterson, of the following striking result: Card C = 2
(19.11)
where = rank g. Another surprise in Peterson’s proof of (19.11) was the role played by the affine Weyl group. A considerable clarification of (19.11) was provided by P. Cellini and P. Papi. If V is the fundamental alcove in the Weyl chamber, then 2 V is a union of 2 alcoves. Cellini and Papi established a natural bijection of C with these 2 alcoves. The dimension of an abelian ideal associated to an alcove is the number of walls separating the alcove from the fundamental alcove. Beautiful results of D. Panyushev related the maximal abelian ideals with the long roots in + . R. Suter showed that Peterson’s result can be deduced from Theorem B above.
513
514
Kostant’s Comments on Papers in Volume I
Paper # 19 became the basis of later results and conferences. In particular, it has been used by Etingof–Kac and Kumar in the solution of the Cachazo–Douglas– Seiberg–Witten conjecture on the structure of conformal algebras.
20. Orbits, Symplectic Structures and Representation Theory, Proc. U.S.-Japan Seminar in Differential Geometry, Kyoto, Japan, 1965, p. 71. In the early 60s I became interested in Hamiltonian mechanics and its symplectic manifold and Poisson bracket underlying structure. I also thought it was quite mysterious and marvelous that physicists in quantizing classical mechanics converted scalar functions (classical observables) on phase space in some fashion or other to operators on Hilbert space. Particularly striking in this process was that the classical observables were functions of position and momentum, q’s and p’s, whereas the elements in the Hilbert space were “functions” on half the variables (e.g., the q’s or the p’s). It seemed to me it would be very interesting to be able to make this process rigorous. The ideas I developed during the early 60s to do this are now referred to as geometric quantization of Kostant–Souriau theory. It was a fortunate time to think about these matters. For one thing there was the Borel–Weil theorem, and growing out of Hirzebruch’s Riemann–Roch theorem, line bundles and Chern classes were very much in the air. Bott had proved his generalization of the Borel–Weil theorem. There were also new constructions of unitary representation of Lie groups: Kirillov’s complete treatment for nilpotent groups and Gelfand and Harish-Chandra’s construction of such representations for semisimple groups using parabolic induction. The spark which ignited geometric quantization for me was Kirillov’s observation that there is a nonsingular alternating 2-form on Lie group coadjoint orbits. Symplectic manifolds as an object of study were not in vogue at that time, but I soon realized that this 2 form is indeed symplectic and that coadjoint orbits yield a vast supply of symplectic homogeneous spaces. Much more than that I came to the realization that what the physicists were doing and the above construction of representations are in fact manifestations of the same idea. The space C ∞ (X ) of smooth functions on a symplectic manifold X is a Lie algebra under Poisson bracket, and as such is a central extension of the Lie algebra Ham (X ) of Hamiltonian vector fields on X , thereby giving rise to a Lie algebra exact sequence 0 −→ C −→ C ∞ (X ) −→ Ham(X ) −→ 0.
(20.1)
The point of departure, in quantization, was the critical recognition that the symplectic 2-form, ω — constrained only by an integrality condition for the corresponding de Rham class [ω] — should be regarded as the curvature of a line bundle L, with connection, over X . I then found that the Lie algebra C ∞ (X ) operates, via what I
514
Kostant’s Comments on Papers in Volume I
515
called prequantization, on the space (L) of smooth sections of L. So functions become operators. Moreover, in the spirit of the Heisenberg uncertainty principle, under prequantization, the constant function operates as a nonzero scalar operator so that, unlike in classical mechanics, the action does not descend (see (20.1)) to Ham(X ). If a Lie group G with Lie algebra g operates symplectically on X in such a fashion that the action induces a homomorphism σ : g → Ham(X ) (this is always the case if X is simply connected) then one says that X is a Hamiltonian G-space if σ lifts to a homomorphism σ : X → C ∞ (X ). I introduced this terminology but restricted my considerations to the case where G operated transitively on X . It has since become standardized terminology but without the assumption of homogeneity. If X is a Hamiltonian G-space, then the points of X define linear functionals on g giving rise to a map µ : X → g∗ (20.2) now well known as the moment or momentum map. In the homogeneous case (20.2) is a covering of a coadjoint orbit and using this, one of early results was a classification of all symplectic homogeneous spaces for G. For example, if g is semisimple, then the most general symplectic homogeneous space is a covering of a coadjoint orbit. In case G is also compact, then the coajoint orbits are themelves simply-connected so that one obtains a generalization of a theorem of H. C. Wang on the classification of all compact K¨ahler homogeneous spaces for G. To carry out geometric quantization one requires some additional structures, the main one involving a choice of what I called a polarization F of (X, ω). This is a choice of a complex involutory distribution of half the dimension of X whose “leaves” (in a complex sense) are Lagrangian (e.g., a K¨ahler structure). This “explains” the choice of half the variables in constructing the Hilbert space of states for physicists and parabolic induction in representation theory. The term polarization has been widely accepted and is now in common usage. Another ingredient required for geometric quantization (in order to obtain a Hilbert space structure) was the introduction of what I called half-forms. Given a polarization F, and inspired by the Bott–Borel–Weil theorem, one is led to introduce the sheaf S of germs of local sections of L which are constant along the leaves of F and then to consider the sheaf cohomology H (X, S). If X is an integral coadjoint orbit of G and F is invariant under the action of G, then G operates on H (X, S). Although there are many unresolved questions there are still a large number of examples where irreducible unitary representations of G can be extracted from this action. Except for half-forms, I gave a course at MIT in 1965 on the above subject. Notes of these lectures by N. Iwahori were widely distributed. See J. Wolf, Bull. AMS Vol 75 (1969) and Repr´esentations des groupes de Lie r´esolubles, P. Bernat et
515
516
Kostant’s Comments on Papers in Volume I
al, Dunod, vol. 4 (1972) for a reference to these notes. I spoke about this subject at a 1965 conference in Differential Geometry in Kyoto, Japan. An all too brief outline, paper # 20, appears in the 1965 proceedings of this conference published by Nippon Hyoronsha Co. I also presented the material above as Phillips lecturer at Haverford college in 1965. I finally published some of the material in Vol. 170 of the Lecture Notes in Mathematics, Springer, 1970. 21. Groups Over Z, Proc. Symposia in Pure Math., 9 (1966), 90–98. I became interested in the theory of Hopf algebras in the early 60s. I was mainly inspired by a paper of Milnor and Moore. They proved a theorem which asserted that a “connected cocommutative Hopf algebra H over a field of characteristic zero is the universal enveloping algebra U (g) of the Lie algebra of primitive elements in H .” If one discards connectedness, then H may contain elements g with augmentation value 1 such that δ(g) = g ⊗ g where δ is the diagonal homomorphism. I called such elements group-like since the set of elements form a group. This terminology has been adopted and become standardized terminology in Hopf algebra theory. I then went on to prove that the most general cocommutative Hopf algebra H over, say C, is the smash product H = C[G] # U (g)
(21.1)
where C[G] is the group algebra over the group G of group-like elements in H and g is the Lie algebra of primitive elements in H . I did not publish the theorem but it appears in a well-known (and by now classic) book, Hopf Algebras, written by one of my students at that time, Moss Sweedler. See the introduction in Hopf Algebras for the proper citation of this theorem. If H is a Hopf algebra, let H be the space of those linear functionals on H which vanish on an ideal of finite codimension in H . We will say that H is dualizable if H is nonsingularly paired to H . In such a case H is a dualizable Hopf algebra and H ⊂ H .
(21.2)
But now (21.1) and (21.2) provide a possible algebraic device for constructing a group G associated to a Lie algebra g without appealing to the usual Lie theoretic machinery, i.e., G ⊂ H for the case, where under suitable conditions, H = U (g). This was part of the motivation which led to paper #21. In more detail Chevalley in his famous Tohoku paper introduced a group G(F), where F is any field, “modeled” after a complex simple Lie group. If g is a complex simple Lie algebra, Chevalley found a lattice gZ in g and root vectors eϕ ∈ gZ with the property that gZ was stable 1 under n! (ad eϕ )n for any n ∈ Z+ and any root ϕ ∈ where is the set of roots
516
Kostant’s Comments on Papers in Volume I
517
with respect to a Cartan subalgebra h. Tensor product by F replaces gZ by g F and introduces F-parameter groups eϕ (t), where t ∈ F, with an automorphism action on g F . G(F) is the group generated by these F-parameter groups. The key objective of paper # 21 was to do the above in a Hopf algebra context so that hopefully we would get the affine ring of the desired algebraic group as a Hopf dual, construct the hyperalgebra at the identity, and find G(F) in the double dual. The first problem was to replace C by Z and construct a Z-form UZ (g). We defined UZ (g) as the algebra over Z in U (g) generated by all elements of the form 1 n n! eϕ for n ∈ Z+ and ϕ ∈ . Let + be a choice of positive roots and order, + = {ϕ1 , . . . , ϕr } so that if ϕ j − ϕi is a sum of positive roots, then j > i. Let = rank g and if the set of simple roots = {α1 , . . . , α }, let h i = [eαi , e−αi ]. For N , M ∈ Zr+ and K ∈ Z+ put m1 eϕm r h1 h e ϕ1 b(N , K , M) = ··· ··· r ··· n1! n r ! k1 k m 1 ! mr ! n1 e−ϕ 1
nr e−ϕ r
(21.3)
where N = {n 1 , . . . , nr }, K = {k1 , . . . , k } and M = {m 1 , . . . , m r }. Let d = dim g. The main theorem of Theorem in paper # 21 asserts the following. Theorem 1. The elements b(N , K , M) for (N , K , M) ∈ Zd+ are a Z-basis of UZ (g) and also a (PBW) C-basis of U (g) so that U (g) = C ⊗Z UZ (g).
(21.4)
Furthermore the Hopf structure on U (g) induces a Z-Hopf structure on UZ (g). In fact the b(N , K , M) are a d-multisequence of divided powers. In addition if V is a finite-dimensional U (g)-module, then UZ stabilizes a Z-lattice VZ in V . Moreover VZ is the sum of its intersections with the weight spaces in V . The Z-algebra UZ (g) has been referred to as the Kostant Z-form of U (g) and is well known in Lie theory. The last statement in Theorem 1 above implies that UZ (g) has a Hopf dual H , where the definition of the latter is modified so that Z replaces C. If A is any commutative ring, then H A = H ⊗Z A has the structure of a Hopf algebra and the group-like elements G(A) in its dual define a functor, A → G(A), from commutative rings to groups. In case A is an algebraically closed field, I had hoped at some later point to show that G(A) was the Chevalley group, modeled on G, and associated to A, and that H A is the affine ring of G(A). However I did not succeed in doing this. An unsolved problem for me was even to show that H A is Noetherian. (Theorem 3, attributed to Chevalley, in paper # 21 should be ignored since it is a misunderstanding on my part of a statement of Chevalley.). However the result is true and was proved by George Lusztig. See
517
518
Kostant’s Comments on Papers in Volume I
his paper entitled “Study of a Z-form of the coordinate ring of a reductive group”, Jour. AMS, March 31, 2008, posted online. Lusztig also establishes that this Hopf algebra approach to Chevalley theory generalizes to the quantum case.
518