Root Systems and Elliptic Curves
Root Systems and Elliptic Curves
net/publication/227153612
CITATIONS READS
200 659
1 author:
Eduard Looijenga
Tsinghua University
109 PUBLICATIONS 0 CITATIONS
SEE PROFILE
All content following this page was uploaded by Eduard Looijenga on 15 September 2014.
Introduction
This paper describes a new invariant theory for root systems. We briefly outline
the main geometric applications; for precise results we refer the reader to the
relevant theorems. Let R be a root system, R v its dual and QV the lattice generated
by R v. If E is an elliptic curve over tr, then A : = QV | is an abelian variety on
which the Weyl group W of R acts. It is easily shown that there is a W-invariant
symmetric bilinear form I on QV such that 1 takes the value 2 on the roots of
smallest length. It follows from the theorem of Appell-Humbert that there exists a
line bundle 5O over A of "Chern c l a s s " - I such that the action of W on A lifts to
an action of W on 5 ~ with the property that W leaves the fibre over 0 c A fixed.
Then 5 ~ 1 is ample and the algebra of W-invariant sections
oo
S(E)W,= @ r(so-k) w
k=0
is a polynomial algebra on l + 1 generators, where l denotes the rank of R. The
S(E)W-module of W-anti-invariant sections
oo
S(E)-W:= (~ F(SO-k) - w
k=0
is free on one generator. Moreover, the divisor A of this generator is precisely the
union of reflection-hypertori in A.
The Weyl formula for affine root systems, in its full generality first proved by
Macdonald, plays in this invariant theory a r61e analogous to the r61e played by
the classical Weyl formula in the ordinary exponential invariant theory. In the
last section we give a natural and relatively simple proof for it (at least for the cases
we are considering): the point is to interpret both sides of this formula as theta
functions for A.
This suggests that our results (as well as their proofs) are more naturally
expressed in the framework of affine root systems than in the present somewhat
narrower context.
We were naturally led to these results by our research on the semi-universal
deformations of the simple elliptic singularities. There the line bundle Lf turns up
with R of type E 6 , E 7 or E 8.
H. Pinkham communicated to me that he found an independent proof of the
fact that Spec S(E) w is isomorphic to ~t+~.
Part of this work was carried out in the academic year 74/75, when the author
was a visitor of the University of Liverpool.
1. A Quadratic Form
(1.1) Let R be a reduced irreducible root system in a real vector space Vand let
R v c V* denote its dual. We write Q c V and QV c V* for the lattices generated
by R and R v respectively. Their dual lattices are denoted by PV c V* and P c V
and we have Q c P and QV c pv.
Once and for all we fix a basis ~1 . . . . . ~t for R (1= rank R). Then ~ ' , ..., ~[ is
a basis for R v.
Let ~ R be the highest root. Since (c~, ~ v ) has the same sign as (~, ~ ) , ~
will be a positive root of R ~. In case all roots of R have the same length we even
have equality: ( ~ , ~ v ) = (~, ~ ) and then it follows that ~ is the highest root of
/ l
It is clear that I is positive definite and invariant under the Weyl group W of R.
We show that ! takes integral values on Q~ • Q~.
(1.2) Lemma./(~,~v)=2.
Proof. Clearly, ~ (~, &~)2 = 2 ~, (~, 8c~ )2, where R + denotes the set of positive
~R ~R +
roots. Now, for any ~ R + -{~},(~, ~ ) equals 0 or 1. Hence the sum is equal to
2( Z (c~,~)+4)=2( Z ( ~ , f i v ) +2)
~R + - {~} ~R +
= 2((2p, ~) + 2),
where p stands for the half sum of positive roots. Because (p, a~ ) = 1 for all
2 [1, Ch. VI, w 1, prop. 29], the last term equals 2 ( 2 g x + . . . + 2 g t + 2 ) = 4 g . The
lemma follows.
(1.3) Corollary. I takes integral values on Q~ x Q~.
Root Systems and Elliptic Curves 19
Proof Since ~ is a root of maximal length [1, Ch. VI, w 1, prop. 25], ~ is a root of
R v of minimal length. It follows that for any root eVeR ~, I ( e ~ , ~ ) e { 2 , 4 , 6 } .
Since I is W-invariant, we have
I(~ ~, x*)
(cCx*)=2--
l(~V, ~v)
for any ~ER, x * e V * . So if ~,fleR, then I(~V,flv)= 89 v ) which is
integral since i ( ~ v ~v) is even.
(1.4) Remark. Let S 2 Qv denote the lattice of integral symmetric bilinear forms
on QV. U p to a constant factor there is a unique W-invariant bilinear form on V*.
By (1.3), (S 2 Q~)W4: 0, so (S 2 Q~)W is infinite cyclic. One checks easily that I
generates (S2QV) w unless R is of type C~ (l>3), in which case 89 is a generator.
I determines a h o m o m o r p h i s m V*-+ V which we also denote by I. We set
M = I - I ( P ) and we define d~e V* by I(d~, ~ , ) = 3 ~ , (Kronecker symbol). Then
d 1..... d I generate M, of course.
The basis ~1 . . . . . ~z determines a chamber C', being the set of x * e V* with
( ~ , x * ) > 0 for 2 = 1..... I. The alcove C is the set o f x * e C' satisfying (~, x * ) < 1.
C is an open/-simplex.
(1.5) Lemma. The vertices of C are the origin f o : = 0 and the points f x : = g ~ l da.
Proof. A vertex of C is a point where all but one of the functions ~1 . . . . . at, 1 - ~
vanish. We have
= 2 I(~t~'-da)-I(~ ~, dz)= gz,
(&, d~) i(&v ' &~)
(~, d~)=2 ~ ) - 0 ,
I(~ ' ~u )
so that % vanishes in fa if p 4:,t.
Following Macdonald, we put r = g - 1(d~ + . . . + dl)
(1.6) Corollary. Let k be a nonnegative integer. Then
(a) k - ~ M c ~ C = ~ i f k < g .
(b) g - a m c~ C = {r}.
(c) The map b : k -1 M ~ C-+ ((g + k)- 1M) c~ C which sends m to (g + k)- 1(km + g r)
is a bijection.
Proof. In affine coordinates we have r = g - l ( 1 "fo + g l f l + " " + g l f ) (recall that
1 + g~ + ... + g~= g). Since all coefficients are positive we have r e C.
If m e k - 1 M ~ t~, then (g + k)- ~(km + g r) lies on the line connecting m and r,
but is not equal to m. Since C is convex, this implies that ( g + k ) - ~ ( k m + g r ) e C .
So the m a p b goes to where the lemma says.
Now let m' ek -~ Mc~ C. Then kin'= k ld~ +''' +kzd~ (kxe7l) = k~ gafa + " " +ktg~f~"
Since m ' e C we must have k~gx>O and k > k a g ~ + . . . + k t g ~ . It follows that
k > g l + ... +g~, i.e. k>g. Equality can only occur if k~ . . . . . k~= 1, which means
20 E. LooOenga
m' = r. If k > g, then look at k m' - g r = (kl - 1) g 1.fl + " " + (kz- 1) g~fl. We have
ka___ 1, so the coefficients are >=0. Moreover, their sum (k I - 1) gl + ' " +(kl-- 1) gl
is strictly smaller than k - ( g - 1) and therefore < k - g. It follows that
m'. = (k - g)- l (k m' - g r)c(k - g)- l M c~ C.
It is clear that b maps m to m'.
We shall also need the "strange formula" of Freudenthal-de Vries. We put
this formula in a form which is more convenient for our purpose.
(1.7) The strange jbrmula. 8 9 where h denotes the Coxeter
number of R.
Proof The form 2 g I ( = ~ 6 | is nondegenerate and so induces a symmetric
6eR
bilinear form cb on V This is just the canonical bilinear form [1, Ch. VI, w 1, no. 12].
The strange formula of Freudenthal-de Vries [2, p. 243] asserts that
(1.7.1) ~(p,p)=2~(h+l)I,
where p is as before the half sum of the positive roots. Since 2 g l(r, ~'~) = 2 = (2p, c~),
it follows that 2gI(r, r ) = q~(2p, 2p). Hence (1.7) follows from (1.7.1).
(1.8) We close this section by quoting a result of Macdonald.
The affine Weyl group W, is the group of affine transformations of V* which is
generated by W a n d QV. It appears that W, is a semi-direct product of these groups.
Furthermore, (~' and C are fundamental domains for W and W~ respectively.
An affine linear function ~: V* --, IR is called an affine root if a(x*) = (e, x * ) + n
for some e e R and n e Z . We denote the set of affine roots by R a. It is easily checked
that Wa leaves the set ofaffine roots invariant. There is precisely one affine reflection
s e W, which sends a to - a . An affine root doesn't vanish in a point of C: if (c~,x*)
= n e Z , then x* is fixed under the reflection x * - - , s * ( x * ) + n c t v in IV,. Since C is
a fundamental domain of W,, we cannot have x * e C. We call an affine root a
positive if a is positive on C. The set of positive affine roots is denoted by R +, so
that R is the disjoint union of R + and - R +. More precisely,
(1.8.1) R,+={~+n:a~R+,n>=0}w{~+n:~e-R+,n__>l},
for the righthand side is clearly contained in the lefthand side and for any a e R ,
either a or - a is in the righthand side.
For any we Wa, let R(w) denote the set of affine roots which are positive on C,
but negative on w C. Clearly a e R(w) iff a is positive and the zero set of a ( = re-
flection hyperplane of sa) separates C and w C. It follows that R (w) is a finite set.
Following Macdonald we put
(1.8.2) s(w)= ~ a.
aER(w)
(2.1) Let F denote the lattice of affine linear functions on V* which take integral
values on QV. F contains the subgroup 2~ of constant functions and we have in
fact a natural isomorphism F ~ P@2~ defined by f--~ (f-f(O), f(0)). Let e(F) denote
the subgroup of Z v whose elements are of the form
o= y~ % e(f)
f(r)>=n
for some n E IR. Here e(f) stands for the element of •v which is one on f and zero
on F - { f } . We then call inf{f(r): c i # 0 } the order of 0, usually denoted by v(0),
and
in(O):= ~ cfe(f)
f(r)=v(O)
the initial part of 0.
Note that V* acts on F (by translation) and hence on 7/F. An element Oze(F)
will be called a theta function of degree k ( = 0 , 1, 2 .... ) if for any vcQ v we have
(2.1.1) v*O=e(-kI(v)- 89
This makes sense, for i(~v, a v ) i s even for all ~V~R so that 89 It is not
difficult to see that (2.1.1) is equivalent to the condition that for all ~v ~R v
(2.1.2) av*O=e(- 89
The set of theta functions of degree k is denoted by F a. Observe that any element
of the form
c e(n); no,nCZ,
n~no
lies in F ~ and conversely, any element of F ~ is of this form. Hence F ~ has a natural
ring-structure.
We may view
F:= @F k
k=O
as a subgroup of e(F). Our first aim is to show that F is in a natural way a graded
algebra over F ~ For this we shall need the following
(2.2) Lemma. Let O=~ cye(f)~F k. Then .for any t~IR, the set {f: f(r)=t and
Cf :~=0} is finite, f
Since I is positive definite, the last coefficient is nonzero for only finitely many
w Q v. Hence {f'~O(f):f'(r)=t and cy, 4=0} is finite.
(2.3) Corollary. If 0 ' = ~ c'I e(f) and 0" =~, c~e(f) belong to F k and U respec-
f f
tively, then for any f~F, ~ c'ii c~2 is a finite sum.
fl+f2=f
The Weyl group W acts on V*, hence also on F and 7/v. From the Definition
(2.4.1) we see that O~m=wOm (We W), SO that Wleaves F* invariant. We denote the
set of W-invariant, (resp. W-anti-invariant) theta functions of degree k by (Fk)w
(resp. (Fk)-W). Note that F-w: = ~ ( F k ) - w is a Fw : =~(Fk)W-module.
Root Systems and Elliptic Curves 23
form a F~ of (Fk) w and in (,9v Ore)= e (-- k l (m)). The anti-invariant theta functions
d 0,,: = ~ det (w) Owm
w e [r
0':=0- Z 4J(m)~~ W
O ...Omo=
m~k- l M c~
then v(&(m))> 1.
Proof. The terms of Ore... Ore. that must be considered are those which belong to
the affine functions
We must show that the matrix (~b~(m')) is invertible. It follows from (2.6) that
m'
As for the last clause of (2.7), let mek -~ Mc~ C. By (1.6) we must have k > g and
there is a unique m o e(k - g)- ~M c~ (~ such that k m = (k - g) m o + g r. Then write
'~OmoOA = E d~m'(m) d O m ' '
m' ~k - 1M c~C
To complete the proof of (2.7) we must show that (q$"'(m)) is invertible. This is
done in a way entirely analogous to the method used above and will therefore
be omitted.
In Section 4 we will show that the functions q$,,(m') and qSm(m') are up to a
factor (which is a fractional power of e) the formal expansions of modular functions
at ioo.
We now show that (3.2.1) is just the W-fixed point part of (3.2.4). Let us determine
the fixed point set of W in P| IR/7l). An element x e P | IR) projects
onto a W-invariant element of p| IR/Z) iff (x, ~a ) % = x - sa x e P| ;g)
for 2 = 1 .... ,l. On the other hand ( x , c ~ ) T a e Q Q H ~ ( E , TZ) for 2 = 1 ..... l iff
x 6 P Q H I ( E , 2g). It follows that xF-*~.(x,c~a)% induces an isomorphism
,l
between (pQHI(E,~L/Z)) w and (P/Q)QHX(E, JE). A priori, this isomorphism
depends on the choice of our basis {% ..... ~}. But since W acts transitively on the
set of bases for R and trivially on P/Q, this is not the case. We so obtain (3.2.1).
Because W is finite, (3.2.4) is (noncanonically) split as an exact sequence of
W-modules. Hence (3.2.1) is as the W-fixed point part of (3.2.4) exact.
(3.3) An important element of A is constructed as follows. Any root ~eR deter-
mines a homomorphism A = Q v | 7Z| The kernel of this homomor-
phism is a divisor which we denote by A(c0. Then A..= ~ A(7) is a W-invariant
divisor and thus defines an element of A. ~R +
Now let he be a holomorphic line bundle over A whose isomorphism class [ 5 ~
belongs to A and satisfies c'[50] = - I. The isomorphism between 50and w* 50 (we W)
will be unique if we require that it is the identity on the fibre over the W-fixed point
0eA. We so obtain an action of W on the total space Tot (50) of 50. We write S(E)
for the algebra of sections
oo
| r(50- k),
k=0
S(E) w and S(E) - w for the submodules of invariant, resp. anti-invariant sections.
Note that S(E) w is a subalgebra of S(E) and that S(E)- w is a S(E)W-module. Our
main results are contained in the theorem below.
(3.4) Theorem. With the notations of(3.3) we have
(a) 50-1 is ample,
(b) S(E) w is a polynomial algebra,freely generated by l + 1 homogeneous elements
of degree go(= 1), gx ..... gt,
(c) S(E)-w is a free S(E)W-module on one generator 0 A of degree g,
(d) the divisor of Oa equals A.
The geometric interpretation of (3.4) is given in the following
(3.5) Corollary. (a) W ' , Tot(50)/O-section=Spec S(E) w is isomorphic to affine
l + 1-space.
(b) Let rt: Tot(he)--~ A denote the bundle projection. Then the critical set of the
quotient mapping q: Tot(50)-+ W'-.Tot he/O-section is equal to n-l(IA[) w {0-
section}.
(c) The discriminant of q is the zero-set of a weighted homogeneous polynomial
with weights go . . . . . gt and of degree 2g.
Proof that (3.4) implies (3.5). Property (a) is an immediate consequence of
(3.4a, b). To prove (b), we choose generators O o . . . . . O z of S(E) w homogeneous of
degree go, ..., g, respectively, and consider them as functions on Tot (he). Then the
Root Systems and Elliptic Curves 27
Since (2t(A) is trivial, we actually get a section of 50-g. This section is clearly anti-
invariant with respect to W and non-trivial. It follows from (3.4c, d) that its
divisor is equal to A. This implies that the divisor defined by dO 0/x .../x dO t is
just 7t-1 (A)+(0-section), which proves (b). It also follows that the discriminant of
q is defined by the principal ideal generated by 0 2 in S(E) w and that proves (c).
(3.6) Remarks. What makes property (3.5b) interesting is the fact that in general
W does not act freely on A - IAI. For example, let (~1, ~2, ~3, ~4) be a basis of D 4
such that ~t2 is the branch point and assume that E = C/(Z + z Z). Then the image
of x v +~4)+~r~ in E is not in IAI, but it is clearly a fixed point of
- id ~ W(D4).
Our p r o o f of (3.4) will actually show that the polynomial mentioned in (3.5c)
can be chosen in such a way that its coefficients lie in a finite extension K of Q (JE)
(where JE stands for the j-invariant of E) such that GaI(K/@ (JE)) is isomorphic to
some GL 2 (Z/N). See remark (4.3).
4. Proof of (3.4)
(4.1) Given an ample line bundle 50' over A, then any other bundle with the same
Chern class, i.e. with the same image under c in (3.2.2), is obtained by translating
~o, in A [6]. Since the assertions of (3.4) are clearly invariant under a translation,
it therefore suffices to prove (3.4) for one particular 50.
Let us construct one. We assume that E is the quotient of II~ by a lattice L
generated by 1 and r with I m ( z ) > 0 . Let e: QV| x V* -.112" be defined by
e(v+zv',z)=exp(2rciI(z,v')+~izl(v',v')); v,v'eQ",
where I has been extended to a complex bilinear form on Vfl. This function deter-
mines an action ~ of QV | on II? x V,*:
The orbit space of this action is in a natural way a line bundle over V*/Q"|
Since I is W-invariant, we have t~(wu)(2,wz)=(2, w(z+u)) for any we W. Hence
the action of W on l ~ x V * defined by w(2, z)=(2, wz) induces one on 5 ~ In
particular we have [50] ~A. Following Appell-Humbert and the identifications
made in (3.2) we have c' [50] = - 1, so that 50 is as required.
The fact that c' [50-1] = I is positive definite implies (by a theorem of Lefschetz)
that 50-t is ample.
0 on Vr satisfying
O(z +u)=e(u, z) -k O(z)
= exp 2 n i t k ( - I(v', t - 1 z ) - 89I(v', v')). 6) (z)
with u = v + v' t eQ v | We put 6)'(0: = 0 (t z). Then 6)' is periodic modulo the
lattice k -1 t -1QV and has therefore a Fourier development with respect to this
lattice:
6)'(z)= ~ c m e x p 2 n i t k I ( z , m ).
nlEM
Moreover 6)' satisfies (2.1.1) with e replaced by exp 2zt i r. Conversely, if reek -1 M,
then O m determines a section 0 m of ~ - k defined by
O,.(z):= Z expZ~itk(-I(v,r-lz)+ 89189 9
v~rn+Q v
Since I is positive definite the absolute value of the general term is bounded by
e x p ( - c Im(t)I(v, v)) for some positive constant c, so that the series converges
absolutely on compact sets. It follows from (2.5) that if m runs over a set S of
representatives for k-~ M/Q ~ the { O,.}m~S will be linearly independent. In fact, the
classical theory asserts that the {O,.}m~S determine a basis of F(Lf-k). For our
purpose it is more convenient to replace 8,, by
6),. (z): = exp(~z i z k I(m, m)) O,,(z)
= ~ exp2~irk(-I(v,t-lz)+ 89
vEm+Q v
Note that O monly depends on the class o f m modulo Q~. Clearly any relation of the
form
0.1...0. = Z
m~S
that for generic z, this space is of dimension 2(I+ 1). Since two elliptic curves over
Ir are always isomorphic over 1R, TE@ ~ is of dimension 2(l+ 1) for any E. It
follows that TE is of dimension l + 1. Hence if we pick homogeneous elements
O ~ e F ( S g - ~ ) w ( 2 = 0 ..... l) such that their image in T* is a basis, then O0, ..., O,
generate S(E) w. In other words the graded-algebra homomorphism (of degree 0)
is surjective. Since the two algebra's have the same dimension in each degree it is
injective as well. This proves (3.4b). A similar argument applied to the (S(E) w-
module homomorphism S(E) w ~ S(E)-w, 0 -~ O 0 a with
proves (3.4c).
(4.5) Proofof(3.4d). We first observe that divisor (Oa) > A. Indeed, if ze A(ct), then
the reflection G associated to ct leaves the fibre over z fixed, so that 0 a (z) = G OA(Z)
= -- Oa (G Z) = -- Oa (Z), implying 0 A (z) = O.
Since OaeF(LP-g), the Chern class of 0 A is equal to gl. The Chern class of
A(c~) is cc| v, and so the Chern class of A equals ~ ct|189 ct|
9 ~R + ~R
by (1.1.1). It follows that divisor (OA)--A is a nonnegative divisor of vanishing
Chern class. Hence divisor (OA)= A.
Our aim is to give an analytic proof of the Macdonald-Weyl formula (5.7). From
the previous section essentially the following two facts will be used: (i) up to a
factor 0 A is the only anti-invariant theta function of degree g and (ii) A is the divisor
of Oa.
We begin with studying the behaviour of our theta functions 0,. under the
modular group. Since 0,, also depends on r, we shall write O,,(z;r)instead of Ore(z).
(5.1) Lemma. Let reek - 1 M / Q v. Then
where (Cram,) is a symmetric unitary matrix whose entries are algebraic numbers.
Proof. We have
0 (z;--r-x)= ~ exp[-ltir-XkI(m+v,m+v)-2rcikI(m+v,z)]
oEQ v
Now the sum in the line above is by the generalised Jacobi inversion formula [3]
equal to
( - i t ) 89189 ~ exp[nikI(v,v)+2nikl(v,m+zz)]
vEk-l M
This product is absolute convergent on any compact subset of the upper half plane
and has no zeroes there. ~/ satisfies ~ / ( z + l ) = e x p {2ni]]q(z) and r / ( - z - 1 ) =
\24
~((- iv)rl(z) (see for instance [4]). It follows from (5.2) that
(5.3.1) r / ( - r - 1 ) - / OA(Z; --r-a)=c~/(z) l exp nizgI(z, z). Oa(--zz; z)
and by (1.7) we have
-2nil 9 ~ l
(5.3.2) r/(z + 1)t OA(Z;Z+ 1)=exp (~+ntgI(r, r)) ,(r) O(z; z)
g
(2nihl~
=exp \-~--/q(r) -10(z;r).
(5.4) We construct another theta function for A with the same modular behaviour
as OA. For the elliptic curve ~/(72 + z Z) the fundamental theta function is by
definition
n=l
Root Systems and Elliptic Curves 31
where q(.) stands for exp 27z i(.). The function f ( . ; z) is a theta function with divisor
(0) for E. Its m o d u l a r behaviour is expressed by the following two formulas
(5.4.1) f(u; ~ + 1)=q(~) f(u; r)
and
(5.4.2) f(u; -r-')= c' q(89 r u 2 ) f ( - "cu; r)
where c' is a root of unity. F o r all these remarks, see [4], Ch. 19, w 1.
It follows that F(z; r ) : = [ I ( - f ( ( : r z ) ; z)) is a theta function with divisor A
~R +
for A. N o w (5.4.1) and the fact that [R+[=89 imply
where all additive and product expansions converge absolutely on any compact
subset of the upper half plane. Letting r - * i ~ , we see that the first line of (5.6)
tends to 1, while the last line tends to lim C(r). Since C is invariant under the
~ioo
modular group it follows that C is identically 1. If we substitute (1.8.3) for the first
line of (5.6) we finally obtain
(5.7) Macdonald's Weyl formula:
~3
References