In this paragraph, V denotes a real vector space of finite dimension l and W denotes a finite subgroup of GL(V), generated by reflections and essential (§ 3, no. 7). Provide V with a scalar product (x∣y) invariant under W. Denote by H the set of hyperplanes H of V such that the corresponding orthogonal reflection sH belongs to W.
1. Definition of Coxeter transformations
An ordered chamber relative to W is a pair consisting of a chamber C determined by H and a bijection i↦Hi from {1,2,…,l} onto the set of walls of C (cf. § 3, no. 9, Prop. 7).
Definition 1
The Coxeter transformation determined by an ordered chamber* (C,(Hi)1≤i≤l)is the elementc=sH1sH2⋯sHlof W.
Proposition 1
All Coxeter transformations are conjugate in W*.
Since W permutes the chambers determined by H transitively (§ 3, no. 2, Th. 1), we are reduced to proving the following: let (C,(Hi)1≤i≤l) be an ordered chamber and π∈Sl; then sH1sH2⋯sHl and sHπ(1)sHπ(2)⋯sHπ(l) are conjugate in W. Taking into account § 4, no. 8, Prop. 8, this will follow immediately from the following lemma:
Lemma 1
Let X be a finite forest, and* x↦gxa map from X to a groupΓsuch thatgxandgyare conjugate wheneverxandyare not linked in X. LetTbe the set of total orderings on X. For allξ∈T, letpξbe the product inΓof the sequence(gx)x∈Xdefined byξ. Then the elementspξare conjugate inΓ.
1) We proceed by induction on n=CardX. The case n=1 is immediate, so assume that n⩾2. There exists in X a terminal vertex a (Chap. IV, Appendix, no. 3, Prop. 2). Let b∈X−{a} be a vertex linked to a if one exists; if a is not linked to any vertex in X−{a}, let b in X−{a} be arbitrary. In all cases, ga commutes with gx for x=b. Let η∈T be such that a is the largest element of X and b the largest element of X−{a}; we let ξ∈T and prove that pξ,pη are conjugate.
2) Assume first that, for ξ, a is the largest element of X and b the largest element of X−{a}. Let X′ be the full subgraph X−{a}, which is a forest. Define a map x↦gx′ from X′ to Γ by putting gx′=gx if x=b, gb′=gbga. Let ξ′,η′ be the restrictions of ξ,η to X′. The induction hypothesis applies, so pξ′ and pη′ are conjugate. But it is clear that pξ′=pξ,pη′=pη, proving the lemma in this case.
3) Assume that a is the largest element of X for ξ. Let X1 (resp. X2) be the set of elements of X−{a} strictly larger (resp. smaller) than b; let ξi be the restriction of ξ to Xi. Then
pξ=pξ1gbpξ2ga=pξ1gbgapξ2,
and this element is conjugate to pξ2pξ1gbga. We are thus reduced to case 2).
4) In the general case, let X3 (resp. X4) be the set of elements of X strictly larger (resp. smaller) than a; let ξi be the restriction of ξ to Xi. Then pξ=pξ3gapξ4, and this element is conjugate to pξ4pξ3ga. We are thus reduced to case 3).
It follows from Prop. 1 that all the Coxeter transformations are of the same order h=h(W). This number is called the Coxeter number of W.
Remark
Let W1,…,Wm be essential finite groups acting in the spaces V1,…,Vm and generated by reflections. Let Cj be a chamber relative to Wj. Let W be the group W1×⋯×Wm acting in the space V1×⋯×Vm. Then C1×⋯×Cm is a chamber relative to W. The Coxeter transformations of W defined by C are the products c1c2…cm, where cj is a Coxeter transformation of Wj defined by Cj.
2. Eigenvalues of a Coxeter transformation
Since all Coxeter transformations are conjugate (no. 1, Prop. 1), they all have the same characteristic polynomial P(T). Let h be the Coxeter number of W. Then
P(T)=j=1∏l(T−exph2iπmj),
where m1,m2,…,ml are integers such that
0≤m1≤m2≤⋯≤ml<h.
Definition 2
The integers* m1,m2,…,mlare called the exponents of W.
Let C be a chamber determined by H,H1,…,Hl its walls, and put si=sHi. Denote by ei the unit vector orthogonal to Hi and on the same side of Hi as C. By Prop. 2 of the Appendix of Chap. IV, we can assume that the Hi are numbered so that e1,e2,…,er are pairwise orthogonal and er+1,er+2,…,el are pairwise orthogonal. Then s′=s1s2…sr is the orthogonal symmetry with respect to the subspace
V′=H1∩H2∩⋯∩Hr,
s′′=sr+1sr+2…sl is the orthogonal symmetry with respect to the subspace
V′′=Hr+1∩Hr+2∩⋯∩Hl,
and c=s′s′′ is a Coxeter transformation. Since (e1,…,el) is a basis of V, V is the direct sum of V′ and V′′.
We deduce first that 1 is not an eigenvalue of c. For if x∈V is such that c(x)=x, then s′(x)=s′′(x), so x−s′(x)=x−s′′(x) is orthogonal to V′ and V′′, and hence is zero. Thus, x=s′(x)=s′′(x)∈V′∩V′′={0}.
Consequently,
0<m1≤m2≤⋯≤ml<h.
(1)
The characteristic polynomial of c has real coefficients. Thus, for all j, the power of T−exph2iπmj in P(T) is equal to that of T−exph2iπ(h−mj). Hence
mj+ml+1−j=h(1≤j≤l).
(2)
Adding the equalities (2) we obtain
m1+m2+⋯+ml=21lh.
(3)
Lemma 2
Assume that* W is irreducible and thatdimV≥2. With the preceding notation, there exist two linearly independent vectorsz′,z′′such that (i) the plane P generated byz′,z′′is stable under s' and s''; (ii) s'|P and s''|P are orthogonal reflections with respect toRz′andRz′′; (iii) z′,z′′∈C, andP∩Cis the set of linear combinations ofz′,z′′with coefficients > 0.
Let (e1,…,el) be the basis of V such that (ei∣ej)=δij. Then C is the open simplicial cone determined by the ei (§ 3, no. 9, Prop. 7). It is clear that V′ is generated by er+1,…,el and V′′ by e1,…,er. Let q be the endomorphism of V such that q(e1)=e1,…,q(el)=el. Its matrix with respect to (e1,…,el) is Q=((ei∣ej)). We have (ei∣ej)≤0 for i=j (§ 3, no. 4, Prop. 3). Since W is irreducible, there does not exist any partition {1,2,…,l}=I1∪I2 such that (ei∣ej)=0 for i∈I1 and j∈I2. Thus (§ 3, no. 5, Lemma 4) Q has an eigenvector (a1,…,al) all of whose coordinates are > 0; let a be the corresponding eigenvalue. Put
z=a1e1+⋯+alel,
z′′=a1e1+⋯+arer∈V′′∩C,
z′=ar+1er+1+⋯+alel∈V′∩C,
and let P be the plane generated by z′ and z′′. Then P∩C is the set of linear combinations of z′ and z′′ with coefficients >0. The relation q(z)=az gives ∑j=1lajej=∑j=1laajej; scalar multiplying by ek (where k≤r) gives
ak+j=r+1∑laj(ej∣ek)=aak; thus
(a−1)z′′=k=1∑r(j=r+1∑laj(ej∣ek))ek
=j=r+1∑laj(k=1∑r(ej∣ek)ek)
=j=r+1∑laj(−ej+k=1∑l(ej∣ek)ek)
=−z′+j=r+1∑lajej.
Thus, (a−1)z′′+z′ is orthogonal to e1,…,er, that is, to V′′. Hence, s′′ leaves stable the plane generated by z′′ and (a−1)z′′+z′, that is, P. Similarly, s′ leaves P stable. Since z′∈P∩V′ and z′′∈P∩V′′, s′∣P and s′′∣P are the reflections with respect to Rz′ and Rz′′.
Theorem 1
Assume that W is irreducible. Then:* (i) m1=1,ml=h−1. (ii) Card(H)=21lh.
We retain the preceding notation. The restriction of c=s′s′′ to P is the rotation with angle 2(z′′,z′) (§ 2, no. 5, Cor. of Prop. 6). Since c is of order h, the h elements 1,c,…,ch−1 of W are pairwise distinct; the elements s′,s′c,…,s′ch−1 are thus pairwise distinct, and are distinct from the preceding elements since ci∣P is a rotation and s′cj∣P is a reflection. The set
{1,c,…,ch−1,s′,s′c,…,s′ch−1}
is the subgroup W′ of W generated by s′ and s′′, and induces on P the group W′′ generated by the orthogonal reflections with respect to Rz′,Rz′′. The transform of C by an element of W′ is either disjoint from −C or equal to −C. Thus, the transform of P∩C by an element of W′′ is either disjoint from −(P∩C) or equal to −(P∩C). Hence, for a suitable orientation of P, there exists an integer m>0 such that (z′′,z′)=mπ (§ 2, no. 5, Cor. of Prop. 7). Moreover, the sets g′(C), for g′∈W′, are pairwise disjoint; the sets g′′(P∩C), for g′′∈W′′, are thus pairwise disjoint; so W′′ is of order 2h. Hence m=h. By definition, c∣P is a rotation with angle m2π, and so has eigenvalues exph2iπ, exph2iπ(h−1). This proves that m1=1,ml=h−1.
The transforms of Rz′ and Rz′′ by W′ are h lines D1,…,Dh of P, and the points of P−(D1∪⋯∪Dh) are transforms by the elements of W′ of points of P∩C. Thus, a hyperplane of H necessarily cuts P along one of the lines Di, and consequently is a transform by an operation of W′ of a hyperplane of H containing Rz′ or Rz′′.
Now, any H∈H that contains Rz′ is one of the hyperplanes H1,…,Hr. Indeed, let eH be the unit vector orthogonal to H and on the same side of H as C. Then eH=λ1e1+⋯+λlel with the λi all ≥0 (§ 3, no. 5, Lemma 6, (i)). Now 0=(eH∣z′)=λr+1ar+1+⋯+λlal, so
λr+1=⋯=λl=0,andeH=λ1e1+⋯+λrer.
Suppose that two of the λi were non-zero, for example λ1 and λ2; since e1,…,er are pairwise orthogonal, we would have
s1(eH)=−λ1e1+λ2e2+⋯+λrer
and the coordinates of s1(eH) would not be of the same sign, which is absurd (loc. cit.). Thus, eH is proportional to one of the vectors e1,…,er, which proves our assertion. Similarly, any H∈H that contains Rz′′ is one of the hyperplanes Hr+1,…,Hl.
The number of elements of H containing Rz′ or Rz′′ is therefore l. If h is even, Card(H) is thus equal to 2hl. If h is odd, Card(H) is equal to 2h−1l+r, and also to 2h−1l+(l−r); hence r=l−r, so r=2l, and Card(H)=2h−1l+2l=2hl.
Remark
Retain the notation of the preceding proof. Let c′ be the C-linear extension of c to V⊗RC, and c′′ the restriction of c′ to P⊗RC. From our study of c∣P, c′′ has an eigenvector x corresponding to the eigenvalue exph2iπ, and this eigenvector does not belong to any of the sets D⊗RC, where D denotes a line of P (since D is not stable under c). Now, for any H∈H, we have seen that H∩P is a line; thus, x∈/H⊗RC.
Corollary
Let R0 be the set of unit vectors of V orthogonal to an element of H. If W is irreducible,
u∈R0∑(x∣u)2=h(x∣x)
for all x∈V.
Put f(x)=∑u∈R0(x∣u)2. It is clear that f is a positive quadratic form invariant under W, and non-degenerate since the ei form a basis of V. Since
W is irreducible, there exists a constant β such that f(x)=β(x∣x) (§ 2, no. 1, Prop. 1). If (xi)1≤i≤l is an orthonormal basis of V for the scalar product (x∣y), then
If W is irreducible and h is even, the unique element of W that transforms C to −C is ch/2*.
We use the notation in the proof of Th. 1. Since c∣P is a rotation through an angle h2π, ch/2 transforms z′ to −z′, z′′ to −z′′, and hence z′+z′′=z to −z. Now z∈C, so the chamber ch/2(C) is necessarily −C.
Proposition 3
Assume that W is irreducible. Let u1,…,ul be homogeneous elements of the symmetric algebra S=S(V), algebraically independent over R and generating the algebra of elements of S invariant under W (§ 5, no. 3, Th. 3). If pj is the degree of uj, the exponents of W are p1−1,…,pl−1*.
Put V′=V⊗RC, S′=S(V′)=S⊗RC, and extend the scalar product on V to a hermitian form on V′. If c is a Coxeter transformation of W, there exists an orthonormal basis (Xi)1≤i≤l of V′ consisting of eigenvectors of c⊗1 (Algebra, Chap. IX, § 7, no. 3, Prop. 4); moreover, we can assume that, for 1≤j≤l, Xj corresponds to the eigenvalue exph2iπmj of c⊗1. It is clear that S′ can be identified with the algebra C[X1,…,Xl], and that we can write uj⊗1=fj(X1,…,Xl), where fj is a homogeneous polynomial of degree pj in C[X1,…,Xl]. Put Dj=∂Xj∂ and J=det(Dkfj). Recall (§ 5, no. 4, Prop. 5) that J(X1,…,Xl) is proportional to the product in S′ of Card(H) vectors yk of V each of which is orthogonal to a hyperplane of H. Since we can assume that X1∈/H⊗C for all H∈H (Remark), the X1 component of each of the vectors yk is non-zero, so J(1,0,…,0)=0. The rule for expanding determinants now proves the existence of a permutation σ of {1,2,…,l} such that (Dσ(j)fj)(1,0,…,0)=0 for all j. Since Dσ(j)fj is homogeneous of degree pj−1, the coefficient of X1pj−1Xσ(j) in fj(X1,…,Xl) is non-zero. Now fj(X1,…,Xl) is invariant under c⊗1, and
This proves that pj−1+mσ(j)≡0 (mod. h). Now h−mσ(j) is an exponent (formula (2)). Permuting the uj if necessary, we can assume that pj−1≡mj (mod. h) for all j. Since pj−1≥0 and mj<h, we have pj−1=mj+μjh with μj an integer ≥0. By § 5, Prop. 3, we see that
Card(H)=j=1∑l(pj−1)=j=1∑lmj+hj=1∑lμj.
Taking into account formula (3) and Th. 1 (ii), we obtain h∑j=1lμj=0, so μj=0 for all j, and finally pj−1=mj for all j.
Corollary 1
If* (mi)1≤i≤lis the increasing sequence of exponents of W, the order of W is equal to(m1+1)(m2+1)…(ml+1).
This follows from the relations mj+1=pj and § 5, Cor. of Th. 3.
Corollary 2
If* c is a Coxeter transformation of W,
exp(h2iπ)andexp(h−2iπ)
are eigenvalues of c of multiplicity 1.
Otherwise, there would exist two non-proportional homogeneous invariants of degree 2 in S, and hence two non-proportional quadratic forms on V* invariant under W, contrary to § 2, no. 1, Prop. 1.
Corollary 3
The homothety with ratio* −1 of V belongs to W if and only if all the exponents of W are odd. In that case, h is even andch/2=−1for any Coxeter transformation c of W.
The first assertion follows from § 5, no. 3, Prop. 4. Assume that the exponents of W are odd. Then h is even by formula (2), and
(exph2iπmj)h/2=exp(iπmj)=−1;
thus ch/2=−1 since c is a semi-simple automorphism of V (Algebra, Chap. IX, § 7, no. 3, Prop. 4).
Exercises
In the exercises below (with the exception of Exerc. 3), the assumptions and notation are those of § 6.