BN-pairs

Arun Ram
Department of Mathematics and Statistics
University of Melbourne
Parkville, VIC 3010 Australia
aram@unimelb.edu.au

Last updated: 27 November 2014

BN-pairs

A group with BN-pair is a quadruple (G,B,N,S) where G is a group,
B is a subgroup of G,
N is a subgroup of G,
S is a subset of the coset space N/(B∩N),
and the following axioms are satisfied:

(T1) B∪N generates G,
B∩N is a normal subgroup of N,
(T2) S generates the group W=B/(B∩N) and
each s∈S has order 2,
(T3) For each s∈S and w∈W, (BsB)(BwB)⊆BwB∪BswB,
(T4) For each s∈S, (BsB)(BsB)=B∪BsB.

Every w∈W is of the form w=n(B∩N), n∈N. By BwB is meant the double coset BnB. Since for n∈N, the double coset BnB depends only on n modulo B∩N, this is well defined. Similarily we can define wB, Bw, w-1Bw by, respectively, nB, Bn, n-1Bn.

Taking inverses shows axiom (T3) is equivalent to:

For each s∈S and w∈W, (BwB)(BsB)⊆BwB∪BwsB.

The elements s∈S are called simple reflections. Every element in w∈W can be written as a product, w=s1⋯sq, of simple reflections.

(Length) For w∈W, its length, denoted l(w), is the minimal q such that w=s1⋯sq, si∈S.

(Bruhat Decomposition) The group G is a disjoint union over w∈W of the double cosets BwB. G=⋃w∈WBwB, (disjoint union) is called the Bruhat decomposition.

Proof.

G=⋃w∈WBwB will be proved below.

We show here that Bw1B=Bw2B implies w1=w2.

We prove this by induction on q, assuming l(w1)≥l(w2)=q.

First note BwB=B implies w=1. This settles the case q=0.

Now suppose Bw1B=Bw2B with l(w1)≥l(w2)=q≥1. Then we can find an s∈S such that l(sw2)=q-1. (If we write w2=s1⋯sq as a product of q simple reflections, s=s1 will do.)

So l(w1)>l(sw2)=q-1. Hence by induction Bw1B and Bsw2B are different, so disjoint.

But (Bsw2B)⊆(BsB)(Bw2B)=(BsB)(Bw1B)⊆(Bw1B)∪(Bsw1B).

Hence Bsw2B=Bs1wB, l(sw1)≥l(w1)-1≥l(w2)=q-1.

By induction sw1=sw2, and we deduce w1=w2.

□

Let s1,s2,…,sq and w∈W. Then (Bs1s2⋯sqB)(BwB) ⊆⋃1≤i1<⋯<ip≤q (Bsi1⋯sipwB) Note: The union is over all substrings of 1<2<⋯<q, including the empty string. Thus the union includes the double coset BwB.

Proof.

The proof is by induction on q using the third BN pair axiom.

The case q=0 is trivial.

Suppose q≥1. Then using BxyB⊆(BxB)(ByB), (Bs1⋯sqB) (BwB)⊆(Bs1B) (Bs2⋯sqB) (BwB). So by induction, (Bs1⋯sqB) (BwB) ⊆ (Bs1B) ( ⋃2≤j1<⋯<jp≤q (Bsj1⋯sjpwB) ) = ⋃2≤j1<⋯<jp≤q (Bs1B) (Bsj1⋯sjpwB).

The third BN-pair axiom (T3) says for s∈S, (BsB)(BwB)⊆ (BwB)∪(BswB) and thus we get (Bs1⋯sqB) (BwB) ⊆ ⋃2≤j1<⋯<jp≤q ( (Bs1sj1⋯sjpwB)∪ (Bsj1⋯sjpwB) ) = ⋃1≤i1<i2<⋯<ip≤q (Bsi1⋯sipB) which proves the lemma.

□

For I⊆S set WI=⟨s | s∈I⟩.

PI=BWIB=⋃w∈WIBwB is a subgroup of G. In particular G=BWB.

Proof.

It is clear that if x∈PI then x-1∈PI, since B and WI are both groups.

Suppose x and y are elements of PI. Then for some s1,…,sq∈S and w∈WI, x∈Bs1⋯sqB andy∈BwB. Then xy∈(Bs1⋯sqB)(BwB). So by the lemma xy∈⋃1≤i1<⋯<ip≤q (Bsi1⋯sipwB) ⊆PI. Hence PI is a group.

If I=S then PI=BWB is a group containing both B and N. Therefore PS=G.

□

Let I,J⊆S and let PI=BWIB and PJ=BWJB be as above. Then, for w∈W, PIwPJ=BWI wWJB.

Proof.

BWIwWJB⊆PIwPJ because BWI⊆PI and WJB⊆PJ.

We now derive the opposite inclusion. Let t1,…,tp∈I and s1,…,sq∈J. Then by the lemma (and its analogous right hand version), (Bt1⋯tpB)(BwB)(Bs1⋯sq) is contained in a union of double cosets, Bw1wW2 with w1∈PI and w2∈PJ.

Hence PIwPJ⊆BWIwWJB.

□

If w has length q, and w=s1,s2,…,sq, then BwB=(Bs1B)(Bs2B)⋯(BsqB).

Proof.

By induction on q. The case q=1 is trivial.

For q>1, s2⋯sq and s1⋯sq-1 each have length q-1. By induction Bs2⋯sqB = (Bs2B)⋯ (BsqB) Bs1⋯sq-1B = (Bs1B)⋯ (Bsq-1B) Hence by BN-pair axiom (T3), (Bs1B) (Bs2B)⋯ (BsqB) = (Bs1B) (Bs2⋯sqB) ⊆ Bs1⋯sqB∪ Bs2⋯sqB, (Bs1B)⋯ (Bsq-1B) (BsqB) = (Bs1⋯sq-1B) (BsqB) ⊆ Bs1⋯sqB∪ Bs1⋯sq-1B. If we can show Bs2⋯sqB∩Bs1⋯sq-1=∅, then we deduce (Bs1B)⋯(BsqB)⊆Bs1⋯sqB, and since the opposite inclusion is always true, that (Bs1B)⋯(BsqB)=Bs1⋯sqB.

But by Bruhat decomposition, Bs2⋯sqB∩Bs1⋯sq-1≠∅ if and only if s2⋯sq=s1s2⋯sq-1 which on left multiplication by s1 gives s1s2⋯s1=s2⋯sq-1, contradicting l(w)=q.

□

For w∈W and s∈S there are a priori three possibilities for the length of sw or ws: either l(w), l(w)+1 or l(w)-1. In fact the first possibility never occurs, as the following proposition, which strengthens the previous, one shows.

Suppose w∈W and s∈S.
Then l(sw)≥l(w) if and only if (BsB)(BswB) =BswB, l(sw)≤l(w) if and only if (BsB)(BswB) =BswB∪BwB,
Similarly for multiplication of w by s on the right.

Proof.

We proceed by induction on q=l(w), the case q=0 being trivial.

Suppose l(sw)≥l(w)=q.

Write w=w′s′, l(w′)=q-1, s′∈S,

Then by induction (BsB)(Bw′B)=Bsw′B and (Bw′B)(Bs′B)=Bw′s′B=BwB.

Thus we have (BsB) (BwB) = (BsB) (Bw′B) (Bs′B) = (Bsw′B) (Bs′B) ⊆ (Bsw′s′B)∪ (Bsw′B)(axiom) = BswB∪Bsw′B We cannot have (BsB)(BwB)=BswB∪BwB unless BwB=Bsw′B. But then, by Bruhat decomposition we deduce, sw=w′, giving l(sw)=q-1, contradicting l(sw)≥l(w)=q.

Hence (BsB)(BswB)=BswB.

Now consider the case l(sw)≤l(w).

This is equivalent to l(ssw)=l(w)≥l(sw). Then by the result above, BwB=BsswB= (BsB)(BswB). Hence we deduce (BsB)(BwB) = (BsB) (BsB) (BwB) = (B∪BsB) (BwB)(axiom) = BwB∪(BsB) (BwB) ⊇ BwB∪BswB = BwB∪BswB. This together with (T3) shows (BsB)(BwB)=BwB∪Bsw in this case.

□

(Cancellation Properties)

(a) Suppose w=s1s2⋯sq, l(w)=q, and l(sw)<l(w).
Then for some i, sw=s1⋯sˆi⋯sq.
Similarity if l(ws)<l(w), ws=s1⋯sˆi⋯sq for some i.
(b) Suppose we express w∈W as a word in S: w=s1s2⋯sq.
Then if l(w)<q, w=s1⋯sˆi⋯sˆj⋯sq, for some i,j.

Proof.

(a) Set wi=s1⋯si. Then l(wi)=i.
Let i be minimal such that l(swi)<l(wi). Such i exist as this is the case for i=q. Hence l(swi-1)=i>i-1=l(wi-1) and l(swi)<l(wi). By the proposition above, (BsB)(Bwi-1B) =Bswi-1Band (BsB)(Bwi)B= BswiB∪BwiB. We calculate (BsB)(Bwi-1B) (BsiB)= { (Bswi-1B) (BsiB)⊆ BswiB∪ Bswi-1B, (BsB)(BwiB)= BswiB∪BwiB. Thus we must have equality in the first case and Bswi-1B=BwiB. By Bruhat decomposition we deduce, swi-1=wi. This gives sw=ss1⋯si-1 si⋯s1=s1⋯ si-1sisi⋯s1 =s1⋯sˆi⋯sq.
(b) Put wj=s1s2⋯sj. Let j≥1 be minimal such that l(wj)<j. (There is at least one such j, j=q, as l(w)=l(wq)<q). Then l(sjwj-1)=l(wj)≤j-1=l(wj-1). By the cancellation lemma there is an i such that s1⋯sj-1sj=s1⋯sˆi⋯sj-1. Postmultiplication by sj+1⋯sq, gives w=s1⋯sˆi⋯ sˆj⋯sq.

□

We deduce immediately from the cancellation results:

If w=s1⋯sp and l(w)=q, then w=si1⋯siq for some sequence, 1≤i1<⋯<iq≤p.

Since the elements of any WI are products of si∈I we deduce:

If w∈WI, I⊆S, has length l(w)=q then w can be expressed w=s1⋯sq, all si∈I.

This means that the length of any w∈WI, defined relative to WI, i.e. with respect to the generators s∈I of WI, is the same as its length defined with respect to the set of generators S of W.

In particular as l(w)=1 only if w=s∈S, we deduce:

We have s∈WI if and only if s∈I.

This generalizes.

Suppose l(w)=q and w=s1⋯sq. Then w∈WI, I⊆S, if and only if each si∈I.

Proof.

The if part follows from the definition of WI.

For the converse we work by induction on q=l(w). The case q=1 is the corollary above.

For q>1, s2⋯sq has length q-1.

By the previous corollaries we can write w=t1⋯t1, all ti∈I.

Then using the cancellation results: for some i, s2⋯sq=s1t1⋯tq=t1⋯tˆj⋯tq∈WI.

Hence by induction s2,…,sq∈I. From this we deduce s1∈WI, which we know implies s1∈I, also.

□

From the above, given w=s1⋯sq of length q, we have w∈WI∪WJ if and only if each si∈I∩J. This together with the definition PK=BWKB, K⊆S proves:

(a) WI∩WJ=WI∩J.
(b) PI∩PJ=PI∩J.

page history