Bourbaki

§ 1. Coxeter groups

In this section, W denotes a group written multiplicatively, with identity element 1, and S denotes a set of generators of W such that S = S^{-1} and 1 \notin S. Every element of W is the product of a finite sequence of elements of S. From no. 3 onwards we assume that every element of S is of order 2.

1. Length and reduced decompositions

Definition 1

Let w \in W. The length of w (with respect to S), denoted by l_S(w) or simply by l(w), is the smallest integer q \geq 0 such that w is the product of a sequence of q elements of S. A reduced decomposition of w (with respect to S) is any sequence s = (s_1, \ldots, s_q) of elements of S such that w = s_1 \cdots s_q and q = l(w).

Thus 1 is the unique element of length 0 and S consists of the elements of length 1.

Proposition 1

Let w and w' be in W. We have the formulas:

l(ww′)≤l(w)+l(w′),(1) l(ww') \leq l(w) + l(w'), \tag{1}
l(w−1)=l(w),(2) l(w^{-1}) = l(w), \tag{2}
∣l(w)−l(w′)∣≤l(ww′−1).(3) |l(w) - l(w')| \leq l(w{w'}^{-1}). \tag{3}

Let (s_1, \ldots, s_p) and (s'_1, \ldots, s'_q) be reduced decompositions of w and w' respectively. Thus l(w) = p and l(w') = q. Since ww' = s_1 \ldots s_p s'_1 \ldots s'_q, we have l(ww') \leq p + q, proving (1). Since S = S^{-1} and w^{-1} = s_p^{-1} \ldots s_1^{-1}, we have l(w^{-1}) \leq p = l(w). Replacing w by w^{-1} gives the opposite inequality l(w) \leq l(w^{-1}), proving (2). Replacing w by ww'^{-1} in (1) and (2) gives the relations

l(w)−l(w′)≤l(ww′−1),(4) l(w) - l(w') \leq l(w{w'}^{-1}), \tag{4}
l(ww′−1)=l(w′w−1).(5) l(w{w'}^{-1}) = l(w'w^{-1}). \tag{5}

Exchanging ww and w′w' in (4) gives l(w′)−l(w)≤l(ww′−1)l(w') - l(w) \leq l(w{w'}^{-1}) by (5), proving (3).

Corollary

Let s=(s1,…,sp)s = (s_1, \ldots, s_p) and s′=(s1′,…,sq′)s' = (s'_1, \ldots, s'_q) be two sequences of elements of SS such that w=s1…spw = s_1 \ldots s_p and w′=s1′…sq′w' = s'_1 \ldots s'_q. If the sequence (s1,…,sp,s1′,…,sq′)(s_1, \ldots, s_p, s'_1, \ldots, s'_q) is a reduced decomposition of ww′ww', then ss is a reduced decomposition of ww and s′s' is one of w′w'.

By hypothesis, l(w)≤pl(w) \leq p, l(w′)≤ql(w') \leq q and l(ww′)=p+ql(ww') = p + q. By (1), we must have l(w)=pl(w) = p and l(w′)=ql(w') = q, hence the corollary.

Remark In view of formulas (1) and (2), the formula d(w,w′)=l(ww′−1)d(w, w') = l(w{w'}^{-1}) defines a distance on WW, invariant under right translations.

2. Dihedral groups

Definition 2

A dihedral group is a group generated by two distinct elements of order two.

Example

Let MM be the multiplicative group {1,−1}\{1, -1\}, and let mm be an integer ≥2\geq 2 (resp. m=∞m = \infty). Then MM acts on the group Z/mZ\mathbf{Z}/m\mathbf{Z} (resp. on Z\mathbf{Z}) by (−1).x=−x(-1).x = -x, and the corresponding semi-direct product of MM by Z/mZ\mathbf{Z}/m\mathbf{Z} (resp. of MM by Z\mathbf{Z}) is denoted by DmD_m. The elements of DmD_m are thus the pairs (ε,x)(\varepsilon, x) with ε=±1\varepsilon = \pm 1 and x∈Z/mZx \in \mathbf{Z}/m\mathbf{Z} (resp. x∈Zx \in \mathbf{Z}); the group law on DmD_m is given by the formula

(ε,x).(ε′,x′)=(εε′,ε′x+x′). (\varepsilon, x).(\varepsilon', x') = (\varepsilon \varepsilon', \varepsilon' x + x').
We denote by ι\iota the class of 1 modulo mm (resp. ι=1\iota = 1) and set
ρ=(−1,0),ρ′=(−1,ι),π=(1,ι). \rho = (-1, 0), \quad \rho' = (-1, \iota), \quad \pi = (1, \iota).
Then ρ2=ρ′2=1\rho^2 = {\rho'}^2 = 1 and π=ρρ′\pi = \rho \rho'. The formulas
πn=(1,nι),ρπn=(−1,nι) \pi^n = (1, n\iota), \quad \rho \pi^n = (-1, n\iota)
show that DmD_m is a dihedral group generated by {ρ,ρ′}\{\rho, \rho'\}.

Proposition 2

Assume that SS consists of two distinct elements ss and s′s' of order 2.

(i) The subgroup PP of WW generated by p=ss′p = ss' is normal, and WW is the semi-direct product of the subgroup T={1,s}T = \{1, s\} and PP. Moreover, (W:P)=2(W : P) = 2.

(ii) Let mm be the order (finite or infinite) of pp. Then m≥2m \geq 2 and WW is of order 2m2m. There is a unique isomorphism φ\varphi from DmD_m to WW such that φ(ρ)=s\varphi(\rho) = s and φ(ρ′)=s′\varphi(\rho') = s'.

(i) We have sps−1=sss′s=s′s=p−1sps^{-1} = sss's = s's = p^{-1}, and hence

spns−1=p−n sp^n s^{-1} = p^{-n}

for every integer nn. Since WW is generated by {s,s′}\{ s, s' \}, and hence by {s,p}\{ s, p \}, the subgroup PP is normal. It follows that TPTP is a subgroup of WW, and since TPTP contains ss and s′=sps' = sp, we have W=TP=P∪sPW = TP = P \cup sP. To prove (i), it is therefore enough to show that W≠PW \neq P. If W=PW = P, the group WW would be commutative, so p2=s2s′2=1p^2 = s^2 {s'}^2 = 1. But then the only elements of W=PW = P would be 1 and pp, contradicting the hypothesis that WW contains at least three elements, namely 1, ss and s′s'.

(ii) Since s≠s′s \neq s', we have p≠1p \neq 1 and so m≥2m \geq 2. Since PP is of order mm and (W:P)=2(W : P) = 2, the order of WW is 2m2m. If mm is finite (resp. infinite), there exists an isomorphism φ′\varphi' from Z/mZ\mathbf{Z}/m\mathbf{Z} (resp. Z\mathbf{Z}) to PP taking π\pi to pp. Moreover, there exists an isomorphism φ′′\varphi'' from M={1,−1}M = \{ 1, -1 \} to TT taking −1-1 to ss. The group WW is the semi-direct product of TT and PP. In view of the formulas (9) and ρπnρ−1=π−n\rho \pi^n \rho^{-1} = \pi^{-n}, φ′\varphi' and φ′′\varphi'' induce an isomorphism φ\varphi from DmD_m to WW such that φ(ρ)=s\varphi(\rho) = s and φ(π)=p\varphi(\pi) = p, and hence φ(ρ′)=s′\varphi(\rho') = s'. The uniqueness of φ\varphi follows from the fact that DmD_m is generated by {ρ,ρ′}\{ \rho, \rho' \}.

Remark

Consider a dihedral group WW of order 2m2m generated by two distinct elements ss and s′s' of order 2. Denote by sqs_q (resp. sq′s'_q) the sequence of length qq whose odd (resp. even) numbered terms are equal to ss and whose even (resp. odd) numbered terms are equal to s′s', and let wqw_q (resp. wq′w'_q) be the product of the sequence sqs_q (resp. sq′s'_q). We have

w2k=(ss′)k,w2k+1=(ss′)ks, w_{2k} = (ss')^k, \quad w_{2k+1} = (ss')^k s,
w2k′=(s′s)k=(ss′)−k,w2k+1′=(s′s)ks′=(ss′)−k−1s. w'_{2k} = (s's)^k = (ss')^{-k}, \quad w'_{2k+1} = (s's)^k s' = (ss')^{-k-1} s.

If s=(s1,…,sq)s = (s_1, \ldots, s_q) is a reduced decomposition (with respect to {s,s′}\{ s, s' \}) of an element ww of WW, then clearly si≠si+1s_i \neq s_{i+1} for 1≤i≤q−11 \leq i \leq q-1. Hence, s=sqs = s_q or s=sq′s = s'_q.

If m=∞m = \infty, the elements (ss′)n(ss')^n and (ss′)ns(ss')^n s for n∈Zn \in \mathbf{Z} are distinct. Hence, the elements wqw_q (q≥0q \geq 0) and wq′w'_q (q>0q > 0) are distinct, and if ss is a reduced decomposition of wqw_q (resp. wq′w'_q) we necessarily have s=sqs = s_q (resp. s=sq′s = s'_q). It follows from this that l(wq)=l(wq′)=ql(w_q) = l(w'_q) = q and that the set of reduced decompositions of the elements of WW consists of the sqs_q and the sq′s'_q. Moreover, every element of WW has a unique reduced decomposition.

Suppose now that mm is finite. If q≥2mq \geq 2m, we have wq=wq−2mw_q = w_{q-2m} and wq′=wq−2m′w'_q = w'_{q-2m}; if m≤q≤2mm \leq q \leq 2m, we have wq=w2m−q′w_q = w'_{2m-q}, wq′=w2m−qw'_q = w_{2m-q}. Hence, neither sqs_q nor sq′s'_q are reduced decompositions if q>mq > m. It follows that each of the 2m2m elements of WW is one of the 2m2m elements w0=w0′w_0 = w'_0, wqw_q and wq′w'_q for 1≤q≤m−11 \leq q \leq m-1, and wm=wm′w_m = w'_m. These 2m2m elements are thus distinct and it follows as above that l(wq)=l(wq′)=ql(w_q) = l(w'_q) = q for q≤mq \leq m and that the set of reduced decompositions of elements of WW consists of the sqs_q and the sq′s'_q for 0≤q≤m0 \leq q \leq m. Every element of WW except wmw_m has a unique reduced decomposition; wmw_m has two.

3. First properties of Coxeter groups

Recall that from now on we assume that the elements of S are of order 2.

Definition 3

(W, S) is said to be a Coxeter system if it satisfies the following condition:

(C) For s, s' in S, let m(s, s') be the order of ss' and let I be the set of pairs (s, s') such that m(s, s') is finite. The generating set S and the relations (ss')^{m(s,s')} = 1 for (s, s') in I form a presentation¹ of the group W.

When (W, S) is a Coxeter system, one also says, by abuse of language, that W is a Coxeter group.

Example 1

Let m be an integer ≥2\geq 2 or ∞\infty and let W be a group defined by a set of generators S = {s, s'} and the relations s2=s′2=1s^2 = {s'}^2 = 1 when m=∞m = \infty, s2=s′2=(ss′)m=1s^2 = {s'}^2 = (ss')^m = 1 when m is finite. Consider on the other hand the dihedral group D_m (no. 2, Example) and the elements ρ\rho and ρ′\rho' of D_m defined by (7). Since ρ2=(ρ′)2=1\rho^2 = (\rho')^2 = 1 and (ρρ′)m=1(\rho\rho')^m = 1 when m is finite, there exists a unique homomorphism f from W to D_m such that f(s)=ρf(s) = \rho and f(s′)=ρ′f(s') = \rho'. Since ρρ′\rho\rho' is of order m, it follows that ss' is itself of order m. Hence, (W, S) is a Coxeter system, W is a dihedral group of order 2m2m and f is an isomorphism (Prop. 2).

By transport of structure, it follows that every dihedral group is a Coxeter group.

Example 2

Let Sn\mathfrak{S}_n be the symmetric group of degree n, with n≥2n \geq 2. Let sis_i be the transposition of i and i+1i+1 for 1≤i<n1 \leq i < n, and let S = {s_1, ..., s_{n-1}}. One can show (§ 2, no.4, Example and § 1, Exerc. 4) that (Sn,S)(\mathfrak{S}_n, S) is a Coxeter system.

Example 3

For the classification of finite Coxeter groups, cf. Chap. 4, § 4.

Remark

Suppose that (W, S) is a Coxeter system. There exists a homomorphism ε\varepsilon from W to the group {1, -1} characterized by ε(s)=−1\varepsilon(s) = -1 for all s∈Ss \in S. We call ε(w)\varepsilon(w) the signature of w; it is equal to (−1)l(w)(-1)^{l(w)}. The formula ε(ww′)=ε(w)ε(w′)\varepsilon(ww') = \varepsilon(w)\varepsilon(w') thus translates into l(ww′)≡l(w)+l(w′)mod  2l(ww') \equiv l(w) + l(w') \mod 2.

¹ This means that (W, S) has the following universal property: given a group G and a map f from S to G such that (f(s)f(s′))m(s,s′)=1(f(s)f(s'))^{m(s,s')} = 1 for (s, s') in I, there exists a homomorphism g from W to G extending f. This homomorphism is unique because S generates W. An equivalent form of this definition is the following. Let W‾\overline{W} be a group, f a homomorphism from W‾\overline{W} to W and h a map from S to W‾\overline{W} such that f(h(s))=sf(h(s)) = s and (h(s)h(s′))m(s,s′)=1(h(s)h(s'))^{m(s,s')} = 1 for (s, s') in S and such that the h(s) (for s∈Ss \in S) generate W‾\overline{W}. Then f is injective (and hence is an isomorphism from W‾\overline{W} to W).

Proposition 3

Assume that (W, S) is a Coxeter system. Then, two elements s and s' of S are conjugate2^2 in W if and only if the following condition is satisfied:

(I) There exists a finite sequence (s1,…,sq)(s_1, \ldots, s_q) of elements of S such that s1=s,sq=s′s_1 = s, s_q = s' and sjsj+1s_j s_{j+1} is of finite odd order for 1≤j<q1 \leq j < q.

Let s and s' in S be such that p=ss′p = ss' is of finite order 2n+12n + 1. By (9), sp−n=pnssp^{-n} = p^n s hence

pnsp−n=pnpns−p−1s=s′ss=s′, p^n sp^{-n} = p^n p^n s - p^{-1} s = s' ss = s',

and s′s' is conjugate to s.

For all s in S, let AsA_s be the set of s′∈Ss' \in S satisfying (I). With the hypotheses in (I), the elements sjs_j and sj+1s_{j+1} are conjugate for 1≤j<q1 \leq j < q by the above, hence every element s′s' of AsA_s is conjugate to s.

Let ff be the map from S to M={1,−1}M = \{1, -1\} equal to 1 on AsA_s and to -1 on S−AsS - A_s. Let s′s' and s′′s'' in S be such that s′s′′s's'' is of finite order m. Then f(s′)f(s′′)=1f(s') f(s'') = 1 if s′s' and s′′s'' are both in AsA_s or both in S−AsS - A_s. In the other case, f(s′)f(s′′)=−1f(s') f(s'') = -1, but m is even. Thus (f(s′)f(s′′))m=1(f(s') f(s''))^m = 1 in all cases. Since (W, S) is a Coxeter system, there exists a homomorphism g from W to M inducing f on S. Let s′s' be a conjugate of s. Since s belongs to the kernel of g, so does s′s', hence f(s′)=g(s′)=1f(s') = g(s') = 1 and finally s′∈Ass' \in A_s. Q.E.D.

4. Reduced decompositions in a Coxeter group

Suppose that (W, S) is a Coxeter system. Let T be the set of conjugates in W of elements of S. For any finite sequence s=(s1,…,sq)s = (s_1, \ldots, s_q) of elements of S, denote by Φ(s)\Phi(s) the sequence (t1,…,tq)(t_1, \ldots, t_q) of elements of T defined by

tj=(s1…sj−1)sj(s1…sj−1)−1for 1≤j≤q. t_j = (s_1 \ldots s_{j-1}) s_j (s_1 \ldots s_{j-1})^{-1} \quad \text{for } 1 \leq j \leq q.

Then t1=s1t_1 = s_1 and s1…sq=tqtq−1…t1s_1 \ldots s_q = t_q t_{q-1} \ldots t_1. For any element t∈Tt \in T, denote by n(s,t)n(s, t) the number of integers jj such that 1≤j≤q1 \leq j \leq q and tj=tt_j = t. Finally, put

R={1,−1}×T. R = \{1, -1\} \times T.

Lemma 1

(i) Let w∈Ww \in W and t∈Tt \in T. The number (−1)n(s,t)(-1)^{n(s, t)} has the same value η(w,t)\eta(w, t) for all sequences s=(s1,…,sq)s = (s_1, \ldots, s_q) of elements of S such that w=s1…sqw = s_1 \ldots s_q.

(ii) For w∈Ww \in W, let UwU_w be the map from R to itself defined by

Uw(ε,t)=(ε⋅η(w−1,t),wtw−1)(ε=±1,t∈T). U_w(\varepsilon, t) = (\varepsilon \cdot \eta(w^{-1}, t), wtw^{-1}) \quad (\varepsilon = \pm 1, t \in T).

The map w↦Uww \mapsto U_w is a homomorphism from W to the group of permutations of R.

(2) Recall that two elements (resp. two subsets) of a group W are said to be conjugate if there exists an inner automorphism of W that transforms one into the other.

For s∈Ss \in S, define a map UsU_s from RR to itself by

Us(ε,t)=(ε.(−1)δs,t,sts−1)(ε=±1,t∈T), U_s(\varepsilon, t) = (\varepsilon.(-1)^{\delta_{s,t}}, st s^{-1}) \quad (\varepsilon = \pm 1, t \in T),
where δs,t\delta_{s,t} is the Kronecker symbol. It is immediate that Us2=IdRU_s^2 = \mathrm{Id}_R, which shows that UsU_s is a permutation of RR.

Let s=(s1,…,sq)s = (s_1, \ldots, s_q) be a sequence of elements of SS. Put w=sq…s1w = s_q \ldots s_1 and Us=Usq…Us1U_s = U_{s_q} \ldots U_{s_1}. We shall show, by induction on qq, that

Us(ε,t)=(ε.(−1)n(s,t),wtw−1). U_s(\varepsilon, t) = (\varepsilon.(-1)^{n(s,t)}, wt w^{-1}).
This is clear for q=0,1q = 0, 1. If q>1q > 1, put s′=(s1,…,sq−1)s' = (s_1, \ldots, s_{q-1}) and
w′=sq−1…s1. w' = s_{q-1} \ldots s_1.
Using the induction hypothesis, we obtain
Us(ε,t)=Usq(Usq(ε.(−1)n(s′,t),w′tw′−1))=(ε.(−1)n(s′,t)+δsq,w′tw′−1,wtw−1). \begin{align*} U_s(\varepsilon, t) &= U_{s_q}(U_{s_q}(\varepsilon.(-1)^{n(s',t)}, w't {w'}^{-1})) \\ &= (\varepsilon.(-1)^{n(s',t)+\delta_{s_q,w't {w'}^{-1}}}, wt w^{-1}). \end{align*}
But Φ(s)=(Φ(s′),w′−1sqw′)\Phi(s) = (\Phi(s'), {w'}^{-1} s_q w') and n(s,t)=n(s′,t)+δw′,−1sqw′,tn(s, t) = n(s', t) + \delta_{w', -1} s_q w', t, proving formula (14).

Now let s,s′∈Ss, s' \in S be such that p=ss′p = ss' has finite order mm. Let s=(s1,…,s2m)s = (s_1, \ldots, s_{2m}) be the sequence of elements of SS defined by sj=ss_j = s for jj odd and sj=s′s_j = s' for jj even. Then s2m…s1=p−m=1s_{2m} \ldots s_1 = p^{-m} = 1 and formula (11) implies that

tj=pj−1sfor 1≤j≤2m. t_j = p^{j-1} s \quad \text{for } 1 \leq j \leq 2m.
Since pp is of order mm, the elements t1,…,tmt_1, \ldots, t_m are distinct and tj+m=tjt_{j+m} = t_j for 1≤j≤m1 \leq j \leq m. For all t∈Tt \in T, the integer n(s,t)n(s, t) is thus equal to 0 or 2 and (14) shows that Us=IdRU_s = \mathrm{Id}_R. In other words, (UsUs′)m=IdR(U_s U_{s'})^m = \mathrm{Id}_R. Thus, by the definition of Coxeter systems, there exists a homomorphism w↦Uww \mapsto U_w from WW to the group of permutations of RR such that UsU_s is given by the right-hand side of (13). Then Uw=UsU_w = U_s for every sequence s=(s1,…,sq)s = (s_1, \ldots, s_q) such that w=sq…s1w = s_q \ldots s_1 and Lemma 1 follows immediately from (14).

Lemma 2

Let s=(s1,…,sq)s = (s_1, \ldots, s_q), Φ(s)=(t1,…,tq)\Phi(s) = (t_1, \ldots, t_q) and w=s1…sqw = s_1 \ldots s_q. Let TwT_w be the set of elements of TT such that η(w,t)=−1\eta(w, t) = -1. Then ss is a reduced decomposition of ww if and only if the tit_i are distinct, and in that case Tw={t1,…,tq}T_w = \{ t_1, \ldots, t_q \} and Card(Tw)=l(w)\mathrm{Card}(T_w) = l(w).

Clearly Tw⊂{t1,…,tq}T_w \subset \{ t_1, \ldots, t_q \}. Taking ss to be reduced, it follows that Card(Tw)≤l(w)\mathrm{Card}(T_w) \leq l(w). Moreover, if the tit_i are distinct, then n(s,t)n(s, t) is equal to 1 or 0 according to whether tt does or does not belong to {t1,…,tq}\{ t_1, \ldots, t_q \}. It follows that Tw={t1,…,tq}T_w = \{ t_1, \ldots, t_q \} and that q=Card(Tw)≤l(w)q = \mathrm{Card}(T_w) \leq l(w), which implies that ss is reduced. Suppose finally that ti=tjt_i = t_j with i<ji < j. This gives si=usju−1s_i = us_j u^{-1}, with u=si+1…sj−1u = s_{i+1} \ldots s_{j-1}, hence

w=s1…si−1si+1…sj−1sj+1…sq. w = s_1 \ldots s_{i-1} s_{i+1} \ldots s_{j-1} s_{j+1} \ldots s_q.

This shows that ss is not a reduced decomposition of ww.

Lemma 3

Let w∈Ww \in W and s∈Ss \in S be such that l(sw)≤l(w)l(sw) \leq l(w). For any sequence s=(s1,…,sq)s = (s_1, \ldots, s_q) of elements of SS with w=s1…sqw = s_1 \ldots s_q, there exists an integer jj such that 1≤j≤q1 \leq j \leq q and

ss1…sj−1=s1…sj−1sj. ss_1 \ldots s_{j-1} = s_1 \ldots s_{j-1} s_j.

Let pp be the length of ww and w′=sww' = sw. By the Remark of no. 3, l(w′)≡l(w)+1l(w') \equiv l(w) + 1 mod. 2. The hypothesis l(w′)≤l(w)l(w') \leq l(w) and the relation

∣l(w)−l(w′)∣≤l(ww′−1)=l(s)=1 |l(w) - l(w')| \leq l(w{w'}^{-1}) = l(s) = 1

thus leads to l(w′)=p−1l(w') = p - 1. Choose a reduced decomposition

(s1′,…,sp−1′) (s'_1, \ldots, s'_{p-1})

of w′w' and put s=(s,s1′,…,sp−1′)s = (s, s'_1, \ldots, s'_{p-1}) and Φ(s′)=(t1′,…,tp′)\Phi(s') = (t'_1, \ldots, t'_p). It is clear that s′s' is a reduced decomposition of ww and that t1′=st'_1 = s. The elements t1′,…,tp′t'_1, \ldots, t'_p being distinct by Lemma 2, we have n(s′,s)=1n(s', s) = 1. Since ww is the product of the sequence ss, we have n(s,s)≡n(s′,s)n(s, s) \equiv n(s', s) mod. 2 by Lemma 1, hence n(s,s)≠0n(s, s) \neq 0. Consequently, ss is equal to one of the elements tjt_j of the sequence Φ(s)\Phi(s), hence the lemma.

Remark

The set TwT_w defined in Lemma 2 consists of the elements of the form w′′sw′′−1w''s{w''}^{-1} corresponding to the triples (w′,w′′,s)∈W×W×S(w', w'', s) \in W \times W \times S such that w=w′′sw′w = w''sw' and l(w′)+l(w′′)+1=l(w)l(w') + l(w'') + 1 = l(w).

5. The exchange condition

The "exchange condition" is the following assertion about (W,S)(W, S):

(E) Let w∈Ww \in W and s∈Ss \in S be such that l(sw)≤l(w)l(sw) \leq l(w). For any reduced decomposition (s1,…,sq)(s_1, \ldots, s_q) of ww, there exists an integer jj such that 1≤j≤q1 \leq j \leq q and

ss1…sj−1=s1…sj−1sj. ss_1 \ldots s_{j-1} = s_1 \ldots s_{j-1} s_j.

We assume in this number that (W,S)(W, S) satisfies (E). By Lemma 3, this is so if (W,S)(W, S) is a Coxeter system. The results of this number thus apply to Coxeter systems.

Proposition 4

Let s∈S,w∈Ws \in S, w \in W and s=(s1,…,sq)s = (s_1, \ldots, s_q) be a reduced decomposition of ww. Then one of the following must hold:

a) l(sw)=l(w)+1l(sw) = l(w) + 1 and (s,s1,…,sq)(s, s_1, \ldots, s_q) is a reduced decomposition of swsw.

b) l(sw)=l(w)−1l(sw) = l(w) - 1 and there exists an integer jj such that 1≤j≤q1 \leq j \leq q, (s1,…,sj−1,sj+1,…,sq)(s_1, \ldots, s_{j-1}, s_{j+1}, \ldots, s_q) is a reduced decomposition of swsw and (s,s1,…,sj−1,sj+1,…,sq)(s, s_1, \ldots, s_{j-1}, s_{j+1}, \ldots, s_q) is a reduced decomposition of ww.

Put w′=sww' = sw. By formula (3) of no. 1, we have

∣l(w)−l(w′)∣≤l(s)=1. |l(w) - l(w')| \leq l(s) = 1.
We distinguish two cases:

a) l(w′)>l(w)l(w') > l(w). Then l(w′)=q+1l(w') = q + 1 and w′=ss1…sqw' = ss_1 \ldots s_q, so

(s,s1,…,sq) (s, s_1, \ldots, s_q)
is a reduced decomposition of w′w'.

b) l(w′)≤l(w)l(w') \leq l(w). By (E), there exists an integer jj such that 1≤j≤q1 \leq j \leq q and (16) is satisfied. Then w=ss1…sj−1sj+1…sqw = ss_1 \ldots s_{j-1} s_{j+1} \ldots s_q and hence

w′=s1…sj−1sj+1…sq. w' = s_1 \ldots s_{j-1} s_{j+1} \ldots s_q.
Since q−1≤l(w′)≤qq - 1 \leq l(w') \leq q, it follows that l(w′)=q−1l(w') = q - 1 and that (s1,…,sj−1,sj+1,…,sq)(s_1, \ldots, s_{j-1}, s_{j+1}, \ldots, s_q) is a reduced decomposition of w′w'.

Lemma 4

Let w∈Ww \in W have length q≥1q \geq 1, let D be the set of reduced decompositions of ww, and let F be a map from D to a set E. Assume that F(s)=F(s′)F(s) = F(s') if the elements s=(s1,…,sq);s′=(s1′,…,sq′)s = (s_1, \ldots, s_q); s' = (s'_1, \ldots, s'_q) of D satisfy one of the following hypotheses:

a) s1=s1′s_1 = s'_1 or sq=sq′s_q = s'_q. b) There exist s and s′s' in S such that sj=sk′=ss_j = s'_k = s and sk=sj′=s′s_k = s'_j = s' for jj odd and kk even.

Then F is constant.

A) Let s,s′∈Ds, s' \in D and put t=(s1′,s1,…,sq−1)t = (s'_1, s_1, \ldots, s_{q-1}). We are going to show that if F(s)≠F(s′)F(s) \neq F(s') then t∈Dt \in D and F(t)≠F(s)F(t) \neq F(s). Indeed, w=s1′…sq′w = s'_1 \ldots s'_q and hence s1′w=s2′…sq′s'_1 w = s'_2 \ldots s'_q is of length <q< q. By Prop. 4, there exists an integer jj such that 1≤j≤q1 \leq j \leq q and the sequence u=(s1′,s1,…,sj−1,sj+1,…,sq)u = (s'_1, s_1, \ldots, s_{j-1}, s_{j+1}, \ldots, s_q) belongs to D. We have F(u)=F(s′)F(u) = F(s') by condition a); if j≠qj \neq q we would have F(s)=F(u)F(s) = F(u) for the same reason, and hence F(s)=F(s′)F(s) = F(s') contrary to our hypothesis. Thus j=qj = q and hence t=u∈Dt = u \in D and F(t)=F(s′)≠F(s)F(t) = F(s') \neq F(s).

B) Let s and s′s' belong to D. For any integer jj with 0≤j≤q+10 \leq j \leq q + 1 define a sequence sjs_j of q elements of S as follows:

s0=(s1′,…,sq′) s_0 = (s'_1, \ldots, s'_q)
s1=(s1,…,sq) s_1 = (s_1, \ldots, s_q)
sq+1−k=(s1,s1′,…,s1,s1′,s1,s2,…,sk) s_{q+1-k} = (s_1, s'_1, \ldots, s_1, s'_1, s_1, s_2, \ldots, s_k)
for q−kq - k even and 0≤k≤q0 \leq k \leq q
sq+1−k=(s1′,s1,…,s1,s1′,s1,s2,…,sk) s_{q+1-k} = (s'_1, s_1, \ldots, s_1, s'_1, s_1, s_2, \ldots, s_k)
for q−kq - k odd and 0≤k≤q0 \leq k \leq q.

Denote by (H_j) the assertion "s_j \in D, s_{j+1} \in D \text{ and } F(s_j) \neq F(s_{j+1})". By A), (H_j) \implies (H_{j+1}) for 0≤j<q0 \leq j < q, and (H_q) is not satisfied by condition b). Hence, (H_0) is not satisfied. Since s0=s′s_0 = s' and s1=ss_1 = s, it follows that F(s)=F(s′)F(s) = F(s').

Proposition 5

Let M be a monoid (with unit element 1) and f a map from S to M. For s, s' \in S, let m(s, s') be the order of ss' and put

a(s,s′)={(f(s)f(s′))lif m(s,s′)=2l,l finite(f(s)f(s′))lf(s)if m(s,s′)=2l+1,l finite1if m(s,s′)=∞. a(s, s') = \begin{cases} (f(s)f(s'))^l & \text{if } m(s, s') = 2l, l \text{ finite} \\ (f(s)f(s'))^lf(s) & \text{if } m(s, s') = 2l + 1, l \text{ finite} \\ 1 & \text{if } m(s, s') = \infty. \end{cases}

If a(s,s′)=a(s′,s)a(s, s') = a(s', s) whenever s≠s′s \neq s' are in S, there exists a map g from W to M such that

g(w)=f(s1)…f(sq) g(w) = f(s_1) \ldots f(s_q)
for all w∈Ww \in W and any reduced decomposition (s1,…,sq)(s_1, \ldots, s_q) of w.

For any w∈Ww \in W, let DwD_w be the set of reduced decompositions of w and FwF_w the map from DwD_w to M defined by

Fw(s1,…,sq)=f(s1)…f(sq). F_w(s_1, \ldots, s_q) = f(s_1) \ldots f(s_q).
We are going to prove by induction on the length of w that FwF_w is constant, which will establish Prop. 5. The cases l(w)=0,1l(w) = 0, 1 being trivial, we assume that q≥2q \geq 2 and that our assertion is proved for the elements w with l(w)<ql(w) < q. Let w be of length q and s,s′∈Dws, s' \in D_w; by Lemma 4 it suffices to prove that Fw(s)=Fw(s′)F_w(s) = F_w(s') in cases a) and b) of that lemma.

a) The formula

Fw(s1,…,sq)=f(s1)Fw′′(s2,…,sq)=Fw′(s1,…,sq−1)f(sq) F_w(s_1, \ldots, s_q) = f(s_1)F_{w''}(s_2, \ldots, s_q) = F_{w'}(s_1, \ldots, s_{q-1})f(s_q)
for w′=s1…sq−1w' = s_1 \ldots s_{q-1} and w′′=s2…sqw'' = s_2 \ldots s_q and the induction hypothesis show that Fw(s)=Fw(s′)F_w(s) = F_w(s') if s1=s1′s_1 = s'_1 or sq=sq′s_q = s'_q.

b) Suppose that there exist two elements s and s' of S such that sj=sk′=ss_j = s'_k = s and sk=sj′=s′s_k = s'_j = s' for j odd and k even. It suffices to treat the case s≠s′s \neq s'. The sequences s and s' are then two distinct reduced decompositions of w in the dihedral group generated by s and s'. By the Remark in no. 2, the order m of ss' is necessarily finite and, in the notation of that remark, s=sms = s_m and s′=sm′s' = s'_m. Consequently, Fw(s)=a(s,s′)F_w(s) = a(s, s') and Fw(s′)=a(s′,s)F_w(s') = a(s', s) and hence

Fw(s)=Fw(s′). F_w(s) = F_w(s').

6. Characterisation of Coxeter groups

Theorem 1

(W, S) is a Coxeter system if and only if it satisfies the exchange condition (E) of no. 5.

Lemma 3 of no. 4 shows that any Coxeter system satisfies (E).

Conversely, suppose that (E) is satisfied. Let G be a group and f a map from S to G such that (f(s)f(s′))m=1(f(s)f(s'))^m = 1 whenever s and s′s' belong to S and ss′ss' is of finite order m. By Prop. 5, there exists a map g from W to G such that

g(w)=f(s1)…f(sq) g(w) = f(s_1) \ldots f(s_q)
whenever w=s1…sqw = s_1 \ldots s_q is of length q. To prove that (W, S) is a Coxeter system, it suffices to prove that g is a homomorphism, which is a consequence of the formula
g(sw)=f(s)g(w)for s∈S,w∈W g(sw) = f(s)g(w) \quad \text{for } s \in S, w \in W
since S generates W. By Prop. 4 of no. 5, only two cases are possible:

a) l(sw)=l(w)+1l(sw) = l(w) + 1: if (s1,…,sq)(s_1, \ldots, s_q) is a reduced decomposition of ww, then (s,s1,…,sq)(s, s_1, \ldots, s_q) is a reduced decomposition of swsw, hence (21).

b) l(sw)=l(w)−1l(sw) = l(w) - 1: put w′=sww' = sw; then w=sw′w = sw' and l(sw′)=l(w′)+1l(sw') = l(w') + 1. By a), g(w)=f(s)g(sw)g(w) = f(s)g(sw) and hence f(s)g(w)=g(sw)f(s)g(w) = g(sw) since (f(s))2=1(f(s))^2 = 1.

7. Families of partitions

Suppose that (W, S) is a Coxeter system. For any s∈Ss \in S, let PsP_s be the set of elements ww in W such that l(sw)>l(w)l(sw) > l(w). We have the following properties:

(A) ⋂s∈SPs={1}\bigcap_{s \in S} P_s = \{1\}.

Indeed, let w≠1w \neq 1 be in W and let (s1,…,sq)(s_1, \ldots, s_q) be a reduced decomposition of ww. Then q≥1q \geq 1 and (s2,…,sq)(s_2, \ldots, s_q) is a reduced decomposition of s2ws_2w, so l(w)=ql(w) = q and l(s1w)=q−1l(s_1w) = q - 1. Hence, w∉Ps1w \notin P_{s_1}.

(B) For any s in S, the sets PsP_s and sPssP_s form a partition of W.

Let w∈Ww \in W and s∈Ss \in S. By Prop. 4 of no. 5, we must distinguish two cases:

a) l(sw)=l(w)+1l(sw) = l(w) + 1: then w∈Psw \in P_s. b) l(sw)=l(w)−1l(sw) = l(w) - 1: put w′=sww' = sw so w=sw′w = sw'; then

l(w′)<l(sw′) l(w') < l(sw')
hence w′∈Psw' \in P_s, that is w∈sPsw \in sP_s.

(C) Let s, s' be in S and let w be in W. If w∈Psw \in P_s and ws′∈Psws' \in P_s then sw=ws′sw = ws'.

Let qq be the length of ww. From w∈Psw \in P_s it follows that l(sw)=q+1l(sw) = q + 1; and from ws′∈Psws' \in P_s it follows that l(sws′)=l(ws′)−1≤ql(sws') = l(ws') - 1 \leq q. Since l(sws′)=l(sw)±1l(sws') = l(sw) \pm 1, we have finally that l(ws′)=q+1l(ws') = q + 1 and l(sws′)=ql(sws') = q.

Let (s1,…,sq)(s_1, \ldots, s_q) be a reduced decomposition of ww and sq+1=s′s_{q+1} = s'; then (s1,…,sq,sq+1)(s_1, \ldots, s_q, s_{q+1}) is a reduced decomposition of the element ws′ws' of length q+1q + 1. By the exchange condition, there exists an integer jj with 1≤j≤q+11 \leq j \leq q + 1 such that

ss1…sj−1=s1…sj. ss_1 \ldots s_{j-1} = s_1 \ldots s_j.
(22) If 1≤j≤q1 \leq j \leq q, we would have sw=s1…sj−1sj+1…sqsw = s_1 \ldots s_{j-1} s_{j+1} \ldots s_q contradicting the formula l(sw)=q+1l(sw) = q + 1. Thus j=q+1j = q + 1 and formula (22) can be written sw=ws′sw = ws'.

Conversely, we have the following result:

Proposition 6

Let* (Ps)s∈S(P_s)_{s \in S} be a family of subsets of WW satisfying (C) and the following conditions:

(A') 1∈Ps1 \in P_s for all s∈Ss \in S. (B') The sets PsP_s and sPssP_s are disjoint for all s∈Ss \in S.

Then, (W,S)(W, S) is a Coxeter system and PsP_s consists of the elements ww of WW such that l(sw)>l(w)l(sw) > l(w).

Let s∈Ss \in S and w∈Ww \in W. There are two possibilities:

a) w∉Psw \notin P_s. Let (s1,…,sq)(s_1, \ldots, s_q) be a reduced decomposition of ww and

wj=s1…sj w_j = s_1 \ldots s_j
for 1≤j≤q1 \leq j \leq q; also put w0=1w_0 = 1. Since w0∈Psw_0 \in P_s by (A') and since w=wqw = w_q is not in PsP_s, there exists an integer jj with 1≤j≤q1 \leq j \leq q such that wj−1∈Psw_{j-1} \in P_s and wj=wj−1sjw_j = w_{j-1} s_j does not belong to PsP_s. By (C),
swj−1=wj−1sj. sw_{j-1} = w_{j-1} s_j.
We have thus proved the formula
ss1…sj−1=s1…sj−1sj ss_1 \ldots s_{j-1} = s_1 \ldots s_{j-1} s_j
which implies that sw=s1…sj−1sj+1…sqsw = s_1 \ldots s_{j-1} s_{j+1} \ldots s_q and l(sw)<l(w)l(sw) < l(w).

b) w∈Psw \in P_s: put w′=sww' = sw, so that w′∉Psw' \notin P_s by (B'). By a), we then have l(sw′)<l(w′)l(sw') < l(w'), that is l(w)<l(sw)l(w) < l(sw).

Comparison of a) and b) proves that PsP_s consists of those w∈Ww \in W such that l(sw)>l(w)l(sw) > l(w). The exchange condition follows from this as we have seen in a), hence (W,S)(W, S) is a Coxeter system (Th. 1 of no. 6).

8. Subgroups of Coxeter groups

In this number, we assume that (W,S)(W, S) is a Coxeter system. For any subset XX of SS, we denote by WXW_X the subgroup of WW generated by XX.

Proposition 7

Let ww be in WW. There exists a subset SwS_w of SS such that {s1,…,sq}=Sw\{s_1, \ldots, s_q\} = S_w for any reduced decomposition (s1,…,sq)(s_1, \ldots, s_q) of ww.

Denote by MM the monoid consisting of the subsets of SS with the composition law (A,B)↦A∪B(A, B) \mapsto A \cup B; the identity element of MM is ∅\varnothing. Put f(s)={s}f(s) = \{s\} for s∈Ss \in S. We are going to apply Prop. 5 of no. 5 to MM and ff. We have a(s,s′)={s,s′}a(s, s') = \{s, s'\} for s,s′s, s' in SS if m(s,s′)m(s, s') is finite, hence there exists a map g:w↦Swg : w \mapsto S_w from WW to MM such that g(w)=f(s1)∪⋯∪f(sq)g(w) = f(s_1) \cup \cdots \cup f(s_q), that is Sw={s1,…,sq}S_w = \{s_1, \ldots, s_q\} for any w∈Ww \in W and any reduced decomposition (s1,…,sq)(s_1, \ldots, s_q) of ww.

Corollary 1

For any subset XX of SS, the subgroup WXW_X of WW consists of the elements ww of WW such that Sw⊂XS_w \subset X.

If w=s1…sqw = s_1 \ldots s_q with s1,…,sqs_1, \ldots, s_q in SS, then w−1=sq…s1w^{-1} = s_q \ldots s_1; hence

Sw−1=Sw. S_{w^{-1}} = S_w.

Prop. 4 of no. 5 shows that Ssw′⊂{s}∪Sw′S_{sw'} \subset \{s\} \cup S_{w'} for s∈Ss \in S and w′∈Ww' \in W, which implies the formula

Sww′⊂Sw∪Sw′ S_{ww'} \subset S_w \cup S_{w'}

by induction on the length of ww. By (23) and (24), the set UU of w∈Ww \in W such that Sw⊂XS_w \subset X is a subgroup of WW; we have X⊂U⊂WXX \subset U \subset W_X, hence U=WXU = W_X.

Corollary 2

For any subset XX of SS, we have WX∩S=XW_X \cap S = X.

This follows from Cor. 1 and the formula Ss={s}S_s = \{s\} for ss in SS.

Corollary 3

The set SS is a minimal generating set of WW.

If X⊂SX \subset S generates WW, then W=WXW = W_X and hence X=S∩WX=SX = S \cap W_X = S by Cor. 2.

Corollary 4

For any subset XX of SS and any ww in WXW_X, the length of ww with respect to the generating set XX of WXW_X is equal to lS(w)l_S(w).

Let (s1,…,sq)(s_1, \ldots, s_q) be a reduced decomposition of ww considered as an element of WW. We have w=s1…sqw = s_1 \ldots s_q and sj∈Xs_j \in X for 1≤j≤q1 \leq j \leq q (Cor. 1); moreover, ww cannot be a product of q′<qq' < q elements of X⊂SX \subset S by definition of q=lS(w)q = l_S(w).

Theorem 2

(i) *For any subset XX of SS, the pair (WX,X)(W_X, X) is a Coxeter system.

(ii) Let (Xi)i∈I(X_i)_{i \in I} be a family of subsets of SS. If X=⋂i∈IXiX = \bigcap_{i \in I} X_i, then WX=⋂i∈IWXiW_X = \bigcap_{i \in I} W_{X_i}.

(iii) Let XX and X′X' be two subsets of SS. Then WX⊂WX′W_X \subset W_{X'} (resp. WX=WX′W_X = W_{X'}) if and only if X⊂X′X \subset X' (resp. X=X′X = X').

Every element of XX is of order 2 and XX generates WXW_X. Let x∈Xx \in X and w∈WXw \in W_X with lX(xw)≤lX(w)=ql_X(xw) \leq l_X(w) = q. By Cor. 4 of Prop. 7, we have

lS(xw)≤lS(w)=q. l_S(xw) \leq l_S(w) = q.
Let x1,…,xqx_1, \ldots, x_q be elements of XX such that w=x1…xqw = x_1 \ldots x_q. Since (W,S)(W, S) satisfies the exchange condition (Th. 1 of no. 6), there exists an integer jj such that 1≤j≤q1 \leq j \leq q and xx1…xj−1=x1…xj−1xjxx_1 \ldots x_{j-1} = x_1 \ldots x_{j-1} x_j. Thus, (WX,X)(W_X, X) satisfies the exchange condition and is therefore a Coxeter system (Th. 1 of no. 6). This proves (i).

Assertions (ii) and (iii) follow immediately from Cor. 1 of Prop. 7.

9. Coxeter matrices and Coxeter graphs

Definition 4

Let II be a set. A Coxeter matrix of type II is a symmetric square matrix M=(mij)i,j∈IM = (m_{ij})_{i,j \in I} whose entries are integers or +∞+\infty satisfying the relations

mii=1for all i∈I;(25) m_{ii} = 1 \quad \text{for all } i \in I; \tag{25}
mij≥2for i,j∈I with i≠j.(26) m_{ij} \geq 2 \quad \text{for } i, j \in I \text{ with } i \neq j. \tag{26}

A Coxeter graph of type II is (by abuse of language) a pair consisting of a graph Γ3\Gamma^3 having II as its set of vertices and a map ff from the set of edges of this graph to the set consisting of +∞+\infty and the set of integers ≥3\geq 3. Γ\Gamma is called the underlying graph of the Coxeter graph (Γ,f)(\Gamma, f).

A Coxeter graph (Γ,f)(\Gamma, f) is associated to any Coxeter matrix MM of type II as follows:

the graph Γ\Gamma has set of vertices II and set of edges the set pairs {i,j}\{i, j\} of elements of II such that mij≥3m_{ij} \geq 3, and the map ff associates to the edge {i,j}\{i, j\} the corresponding element mijm_{ij} of MM.

It is clear that this gives rise to a bijection between the set of Coxeter matrices of type II and the set of Coxeter graphs of type II.

To assist the reader in following our arguments, we often represent a Coxeter graph of type II by the diagram used to represent its underlying graph, and write either next to or above each edge {i,j}\{i, j\} the number f({i,j})f(\{i, j\}). We generally omit these numbers if they are equal to 3.

3 See the appendix for the definition and properties of graphs used here.

If (W, S) is a Coxeter system, the matrix M=(m(s,s′))s,s′∈SM = (m(s, s'))_{s, s' \in S}, where m(s,s′)m(s, s') is the order of ss′ss', is a Coxeter matrix of type S which is called the Coxeter matrix of (W, S): indeed, m(s,s)=1m(s, s) = 1 since s2=1s^2 = 1 for all s∈Ss \in S, and m(s,s′)=m(s′,s)⩾2m(s, s') = m(s', s) \geqslant 2 if s≠s′s \neq s' since ss′=(s′s)−1ss' = (s's)^{-1} is then ≠1\neq 1. The Coxeter graph (Γ,f)(\Gamma, f) associated to M is called the Coxeter graph of (W, S). We remark that two vertices ss and s′s' of Γ\Gamma are joined if and only if ss and s′s' do not commute. For example, the Coxeter matrix of a dihedral group of order 2m2m is (1mm1)\begin{pmatrix} 1 & m \\ m & 1 \end{pmatrix} and its Coxeter graph is represented by

∘—m∘ \circ \overset{m}{\text{---}} \circ

when m⩾3m \geqslant 3 (or

∘—∘ \circ \text{---} \circ

if m=3m = 3) and by

∘∘ \circ \qquad \circ

when m=2m = 2. \* The Coxeter graph of the symmetric group Sn\mathfrak{S}_n is represented by

∘—∘—∘⋯∘—∘ \circ \text{---} \circ \text{---} \circ \cdots \circ \text{---} \circ

(n−1n - 1 vertices.) \*

We show later (Chap. V, § 4) that, conversely, any Coxeter matrix is the matrix of a Coxeter system.

A Coxeter system (W, S) is said to be irreducible if the underlying graph of its Coxeter graph is connected (Appendix, no. 2) and non-empty. Equivalently, S is non-empty and there exists no partition of S into two distinct subsets S' and S'' of S such that every element of S' commutes with every element of S''. More generally, let (Γi)i∈I(\Gamma_i)_{i \in I} be the family of connected components of Γ\Gamma (Appendix, no. 2) and let SiS_i be the set of vertices of Γi\Gamma_i. Let Wi=WSiW_i = W_{S_i} be the subgroup of W generated by SiS_i (cf. no. 8). Then the (Wi,Si)(W_i, S_i) are irreducible Coxeter systems (no. 8, Th. 2) called the irreducible components of (W, S). Moreover, the group W is the restricted direct product[^1] of the subgroups WiW_i for i∈Ii \in I. Indeed, this follows from the following more general proposition:

Proposition 8

Let (Si)i∈I(S_i)_{i \in I} be a partition of SS such that every element of SiS_i commutes with every element of SjS_j if i≠ji \neq j. For all i∈Ii \in I, let WiW_i be the subgroup generated by SiS_i. Then WW is the restricted direct product of the family (Wi)i∈I(W_i)_{i \in I}.

It is clear that for all i∈Ii \in I the subgroup Wi′W'_i generated by the union of the WjW_j for j≠ij \neq i is also generated by Si′=⋃j≠iSjS'_i = \bigcup_{j \neq i} S_j. Thus

Wi∩Wi′=W∅={1} W_i \cap W'_i = W_\varnothing = \{1\}
by Th. 2 of no. 8. Since WW is generated by the union of the WiW_i, this proves the proposition.

Exercises

See the exercises for § 1.

[^1]: A group G is the restricted direct product of a family (Gi)i∈I(G_i)_{i \in I} of subgroups if, for any finite subset J of I, the subgroup GJG_J of G generated by the GiG_i for i∈Ji \in J is the direct product of the GiG_i for i∈Ji \in J and if G is the union of the GJG_J. Equivalently, every element of GiG_i commutes with every element of GjG_j for i≠ji \neq j and every element of G can be written uniquely as a product ∏i∈Igi\prod_{i \in I} g_i with gi∈Gig_i \in G_i and gi=1g_i = 1 for all but finitely many indices ii. This last condition is equivalent to saying that G is generated by the union of the GiG_i and that Gi∩GJ={1}G_i \cap G_J = \{1\} for all i∈Ii \in I and all finite subsets J of I such that i∉Ji \notin J.

Transcribed from the printed volume. Not translated.