Essays on the structure of reductive groups Structure constants
Bill Casselman
University of British Columbia [email protected]
The principal goal of this essay is to explain Tits’ analysis of the structure constants of semi-simple Lie algebras. I’ll begin in the first part with a survey of the standard approach to the topic, that found originally in [Chevalley:1955] and later explained more accessibly, for example in [Carter:1972]. I’ll include here a reasonably explicit algorithm for computing the structure constants, taken largely from [Cohen-Murray-Taylor:2004]. Then in the second part I’ll try to explain the approach of [Tits:1966]. Tits’
paper is often cited, but usually rather vaguely, and I suspect that few have understood it.
The structure constants of a Lie algebra are the coefficients occurring in the formulas for the Lie bracket in terms of a given basis. For a semi-simple algebra there is a natural direct sum decomposition into eigenspaces with respect to a Cartan subalgebra. One of the eigenspaces is the null eigenspace, the Cartan algebra h itself. This contains elements h
λ, one for each root λ. For h we have the basis h
αparametrized by simple roots α. The other eigenspaces are parametrized by roots λ, and each has dimension one, say with basis e
λ. Given e
λfor λ > 0 there exists a unique e
−λwith [e
λ, e
−λ] = h
λ. We have these formulas for the Lie bracket:
[h
α, h
β] = 0
[h
α, e
λ] = hλ, h
αie
µ[e
λ, h
α] = −hλ, h
αie
µ[e
λ, e
µ] =
h
λif µ = −λ N
λ,µe
λ+µif λ + µ is a root
0 otherwise.
for certain constants N
λ,µ. The main result of Chevalley is that for a suitable choice of the x
λwe have N
λ,µ= ±(p
λ,µ+ 1), where p
λ,µis the greatest p such that µ − pλ is a root. For this basis the structure constants are integral, and this ultimately leads to Chevalley’s construction of split semi-simple groups over arbitrary fields. Chevalley’s formula is well known, but to me still a bit mysterious. As I hope to show, the derivation of [Tits:1966] is better motivated than the standard one.
The next problem is to figure out the signs of the structure constants. The standard procedure is straightforward in logic, but depends on various arbitrary choices, and this also seems to me rather mysterious. Again here, Tits’ argument seems more enlightening.
Tits’ idea is to relate these questions to some other questions about the structure of the normalizer N
G(T ) of a maximal torus T in a split semi-simple group G. There is an exact sequence
1 −→ T −→ N
G(T ) −→ W −→ 1 ,
where W is the Weyl group. This hardly ever splits, but it is at least an extension derived from a much smaller extension
1 −→ T (2) −→ N −→ W −→ 1 ,
where T (2) is the subgroup of 2-torsion in T and N is a subgroup of N
G(T ). Several choices of N
are possible, and one of Tits’ observations is that each one determines an integral structure on G and
g. Another is that the problem of signs is also related to problems concerning N . In Tits’ account some puzzles remain, but far fewer than in the standard one.
This essay contains no really new results of a mathematical nature—I have simply reorganized Tit’s own exposition.
Contents
I. The standard approach
1. The data defining a simple Lie algebra 2. The first main Theorem
3. Chains
4. Proof of the main result 5. Chevalley’s integral constants
6. Computation of structure constants—mathematics 7. Computation of structure constants—programming II. Tits’ analysis
8. The extended Weyl group
9. References
Part I. The standard approach
1. The data defining a simple Lie algebra
Suppose g to be a simple, semi-simple Lie algebra over C. It contains an abelian subalgebra h that is its own centralizer, called a Cartan algebra of g. The adjoint action of h on g decomposes g into a direct sum of eigenspaces. The 0-eigenspace is h itself. All the others are one-dimensional, with respect to characters of h called roots of the pair (g, h). The restriction of the Killing form of g to h is positive definite, and this allows us to identify the roots with elements of h itself. Define for each root λ the element
h
λ= 2λ kλk
2in h. These span a lattice that spans h. The dot products λ
•h
µlie in Z, and the orthogonal root reflections s
λ: h 7−→ h − (h
•h
λ)λ
take Σ into itself. The Weyl group W is the group generated by these. The closure of each of the connected components of the complement in h of the root hyperplanes v
•λ = 0 is a fundamental domain for W . Let C be one of these, and let Σ
+= Σ
+Cbe the set of λ in Σ with λ
•C > 0, ∆ the set of α in Σ
+that vanish on a wall of C. Then Σ is the union of Σ
+and −Σ
+. Roots in ∆ are called simple roots . These form a basis of h, and every λ in Σ
+is a non-negative integral combination of simple roots. The set Σ is the smallest subset of h containing ∆ that is stable under W .
Because g is simple, the root system is irreducible in the sense that ∆ is not the union of two orthogonal subsets. Equivalently, its Dynkin diagram is connected.
Choose e
λ6= 0 in g
λ, for λ > 0. There exists a unique e
−λin g
−λsuch that [e
λ, e
−λ] = h
λ. All in all, we
have [h
λ, h
µ] = 0
[h
λ, e
µ] = hµ, h
λie
µ[e
µ, h
λ] = −hµ, h
λie
µ[e
λ, e
µ] =
h
λif µ = −λ N
λ,µe
λ+µif λ + µ is a root
0 otherwise.
for suitable constants N
λ,µ.
For a given λ, the elements h
λ, e
±λspan a Lie algebra isomorphic to that of SL
2.
The first problem to be dealt with now is straightforward. How to compute the structure constants N
λ,µ, given the Cartan matrix defining the root system? Later on, I’ll discuss a second natural question:
Given a root system, can we construct the corresponding Lie algebra? Defining it by naming a basis is straightforward, according to what we have already seen, and previous material will also determine candidate structure constants, so the question is, can we prove that what we define is in fact a Lie algebra?
Complete and explicit formulas for the structure constants are somewhat complicated, but the book [Carter:1972], which I follow, presents a relatively simple explanation of what’s involved. Carter also presents an algorithm for calculating them explicitly, and this has been made more explicit in [Cohen- Murray-Taylor:2004]. I deviate only in details from that last paper. Another, in many ways more natural bit still complicated, approach to both questions can be found in [Tits:1966].
So to start, under the assumption that the complex semi-simple Lie algebra exists, I’ll explain how to
obtain its structure constants. But afterwards I’ll also look at the converse problem—constructing the Lie
algebra from the structure constants. That is to say, given a root system, the algorithm at hand produces
a set of structure constants. A natural question is whether one can prove directly that they determine a
Lie algebra—that is to say, whether one can prove directly that the bracket operation defined by these structure constants is indeed a Lie bracket. I’ll obtain partial results towards answering this.
Let me give a little background to this problem. Constructing the Lie algebra g from its root system—i.e.
showing that every root system does indeed give rise to a Lie algebra—is not trivial. Originally, this was done on a case-by-case analysis, but different cases required vastly different techniques. The exceptional Lie algebras have always caused some trouble, and the Lie algebras e
n(n = 6, 7, 8) as well as the groups E
nhave always remained rather abstract entities. These groups are certainly complicated, and nobody should expect to carry out computations in E
8as readily as in the groups GL
n, but one might hope that a computer will make things seem a bit less abstract. In this context, defining the Lie algebra in terms of the computed structure constants seems a worthwhile project, and a bit less tedious than might otherwise appear.
There are several methods that have been applied uniformly (as opposed to case-by-case) to construct the Lie algebra from the root system. One is that found in [Jacobson:1962], which is due, to what extent independently I do not know, to Chevalley and Harish-Chandra. A second relies on a later result of [Serre:1966] describing the Lie algebra by generators and relations. As far as a direct construction, there have been two approaches. One is in [Tits:1966] and the other, more recent, in [de Graaf:2000]. But I’d like to see an approach even more direct.
Both the first two methods have been applied to construct Kac-Moody algebras from a generalized Cartan matrix, but in this situation things are not entirely satisfactory. Here, there is a important problem that has not been resolved—the methods of Jacobson on the one hand and Serre on the other are not known to produce the same algebra in all cases. For Kac-Moody algebras the so-called ‘imaginary roots’ cause trouble, since the corresponding eigenspaces may have dimension greater than 1, and I am not aware that anyone has proposed even a good description of a basis, much less the corresponding structure constants. Although I do not make any contribution to its solution, it is this problem that has been my principal motivation in writing this note.
2. The first main Theorem
Suppose C to be a Cartan matrix (c
α,β) with rows and columns indexed by a finite set ∆. This means that (a) the entries of C are integers, (b) c
α,α= 2, (c) c
α,β≤ 0 for α 6= β, and (d) DC is positive definite for some diagonal matrix D (necessarily with positive entries). Let h be a complex vector space, h
∨is dual. Assume ∆ embedded as a basis of h
∨. Then there exists a map α 7→ h
α(sometimes written as α
∨) from ∆ to h such that
hα, h
βi = c
α,β. The linear map
s
α: v 7−→ v − hα, vih
αis a reflection. It will be orthogonal with respect to an inner product u
•v if and only if
h
α•h
α2
hα, β
∨i = h
α•h
βfor all α, β in ∆. Therefore we may define this inner product by the formula h
α•h
β= d
αc
α,β.
I assume that C is irreducible, which means that ∆ cannot be partitioned into two mutually orthogonal subsets.
Let W be the group of reflections of h generated by the s
α. By duality, it may also be identified with a
group of linear transformations of h
∨. Let Σ be the smallest subset of h
∨stable under W and contating
∆. This is the set of roots defined by C. There exists a unique map from Σ to h taking λ = wα to h
λ= wh
α.
The roots Σ span a full lattice in h
∨. Let kλk = h
•h.
For each λ in Σ, let g
λbe a one-dimensional space with generator e
λ, and let g be the direct sum of h and the g
λ. To each pair of roots λ, µ such that λ + µ is a root assign a non-zero complex number N
λ,µ. If λ + µ is not a root, let N
λ,µ= 0. Then define a candidate bracket operation [u, v] on g by the specification:
[h
λ, h
µ] = 0
[h
λ, e
µ] = hµ, h
λie
µ[e
µ, h
λ] = −hµ, h
λie
µ[e
λ, e
µ] =
0 if µ = λ h
λif µ = −λ N
λ,µe
λ+µotherwise.
The span of e
±λand h
λis a copy of the Lie algebra of sl
2.
The problem at hand is to find conditions on the constants N
λ,µthat insure the bracket operation defines a Lie algebra—to be precise, that the bracket operation is anti-symmetric and satisfies the Jacobi identity.
These are both immediate from its definition in any bracket involving an element of h.
Proposition 2.1. The definition of [u, v] makes g into a Lie algebra if and only if the constants N
λ,µ [structure-equations]satisfy the following relations:
(a) N
µ,λ= −N
λ,µ; (b) if λ + µ + ν = 0
N
λ,µkνk
2= N
µ,νkλk
2= N
ν,λkµk
2; (c) N
λ,µN
−λ,λ+µ= (p
λ,µ+ 1)q
λ,µ;
(d) whenever λ + µ + ν + ρ = 0 and no pair is opposite N
λ,µN
ν,ρkλ + µk
2+ N
µ,νN
λ,ρkµ + νk
2+ N
ν,λN
µ,ρkν + λk
2= 0 .
As we shall see, the ultimate reason for the appearance here of the norms is the identity of h
λwith 2λ/ kλk
2.
If λ and µ are any two roots, there exist p = p
λ,µand q = q
λ,µsuch that rλ + µ is a root if and only if
−p ≤ r ≤ q. (I’ll recall in the next section how this works.) The sequence
−pλ + µ, . . . , rλ + µ, . . . , qλ + µ
is called the λ -chain through µ. The significance of chains is that the sum of subspaces g
µas µ varies over a λ chain is a representation of the copy of sl
2correspondng to λ.
in the circumstances, condition (c) is equivalent to Chevalley’s remarkable observation that N
λ,µN
−λ,−µ= −(p
λ,µ+ 1)
2but is more directly related to the Jacobi identity. Later, we shall see how this can be used to compute
structure constants from a few that can be set arbitrarily.
3. Chains
If λ and µ are roots, then p = p
λ,µis the greatest m with −mλ + µ a root and q = q
λ,µthe greatest m with mλ + µ a root. In other words, p measures the distance to the back end of the chain and q to the forward one. It is 0 at one end. Elementary properties:
q
λ,µ= p
−λ,µp
λ,λ+µ= p
λ,µ+ 1 q
λ,λ+µ= q
λ,µ− 1 p
−λ,−µ= p
λ,µwhen these equations make sense.
Lemma 3.1. Suppose λ and µ to be roots with λ 6= ±µ. Then
[chains-a]
(a) if hµ, h
λi < 0, µ + λ is a root;
(b) if hµ, h
λi > 0, µ − λ is a root;
(c) the linear combination rλ + µ is a root if and only if −p
λ,µ≤ r ≤ q
λ,µ; (d) we have
hµ, h
λi = p
λ,µ− q
λ,µ.
Proof. If hµ, h
λi < 0 then the classification of root systems of rank two implies that either hµ, h
λi or hλ, h
µi is −1. In the first case s
λµ = µ − hµ, h
λiλ = µ + λ is a root, and in the second s
µλ = λ + µ is one. Item (b) is similar.
As for (c), it means neither more nor less than that no chain has any gaps in it. But since hµ + nλ, h
λi = hµ, h
λi + 2n
the function hµ + nλ, h
λi is an increasing function of n. If there existed a gap in a chain, by (a) and (b) there would exist some µ in it such that hµ, h
λi ≥ 0 but hµ + nλ, h
λi ≤ 0, a contradiction.
The reflection s
λpreserves each λ-chain, and must therefore reflect the extremity −pλ + µ to the other qλ + µ. Thus
µ + qλ = s
λ(µ − pλ)
= µ − pλ − hµ − pλ, h
λiλ
= µ + (p − hµ, h
λi)λ which implies (d).
Since |hλ, h
µi| ≤ 3 in all cases, this tells us that a chain can have at most 4 roots in it. This can be deduced
also from the fact that any two roots are contained in a root system of rank two. We can classify easily
all possible cases:
Looking carefully at the possible cases will tell us that p
λ,µcan be calculated explicitly in terms of simple data. First of all, q
λ,µcan be calculated from p
λ,µand hµ, h
λi. Hence if hµ, h
λi > 0 we may replace µ by r
λ(µ). As for the cases where hµ, h
λi ≤ 0, here is a list of possibilities, sorted by the value of hµ, h
λi:
(1) hµ, h
λi = 0: p
λ,µ= 0 except when kλk = kµk = √
2.
µ λ C
2p
λ,µ= 1 hµ, h
λi = 0
kµk
2= 1 kλk
2= 1
µ
λ C
2p
λ,µ= 0 hµ, h
λi = 0
kµk
2= 2 kλk
2= 2
µ
λ G
2p
λ,µ= 0 hµ, h
λi = 0
kµk
2= 3 kλk
2= 1
µ
λ G
2p
λ,µ= 0 hµ, h
λi = 0
kµk
2= 1 kλk
2= 3
(2) hµ, h
λi = −1: p
λ,µ= 0 except when kλk = kµk = 1 and the system is G
2.
µ λ A
2p
λ,µ= 0 hµ, h
λi = −1
kµk
2= 1 kλk
2= 1
µ λ C
2p
λ,µ= 0 hµ, h
λi = −2
kµk
2= 2
kλk
2= 1
µ λ G
2p
λ,µ= 0 hµ, h
λi = −1
kµk
2= 3 kλk
2= 3
µ λ G
2p
λ,µ= 0 hµ, h
λi = −3
kµk
2= 3 kλk
2= 1
µ λ G
2p
λ,µ= 0 hµ, h
λi = −1
kµk
2= 1 kλk
2= 3
(3) hµ, h
λi < −1: p
λ,µ= 0.
µ λ C
2p
λ,µ= 0 hµ, h
λi = −1
kµk
2= 1 kλk
2= 2
µ λ G
2p
λ,µ= 1 hµ, h
λi = −1
kµk
2= 1 kλk
2= 1
I point out the obvious fact that this observation allows us to tell simply whether λ + µ is a root or not.
The following can also be deduced from the case-by-case list, but has a more direct proof:
Proposition 3.2. (Geometric Lemma) If λ and µ are roots such that λ + µ is also a root, then
[geometric-lemma]
kλ + µk
2kµk
2= p
λ,µ+ 1 q
λ,µ= p
λ,λ+µq
λ,µ.
Proof. There are three basic types of configurations for µ and µ + λ.
(1) They lie at the center of the λ-chain through µ. Here, µ and λ + µ are of the same length, since s
λswaps them, and p
λ,µ+ 1 = q
λ,µ. Both sides of the equation to be verified are 1.
µ λ
λ + µ
p + 1 = q
(2) They lie on the left. A scan of possible systems of rank two shows that there are two possibilities, and in both the equation can be verified visually.
µ λ + µ
1
√ 2
p + 1 = 1, q = 2
µ λ + µ
√ 1 3
p+ 1 = 1, q = 3
(3) They lie to the right of center. In this case, the equation is equivalent to the near mirror-image—p + 1 and q are swapped, and so are kµk
2and kλ + µk
2.
µ λ + µ µ λ + µ
This is not a difficult result, but a bit mysterious, since case-by-case seems to be the simplest proof. Well . . . it is after all only two cases.
Lemma 3.3. If λ, µ and λ + µ are all roots then
[p-symmetric]
p
λ,µ= p
µ,λ.
Proof. By definition, p
λ,µ= 0 if and only if −λ + µ is not a root, and p
µ,λ= 0 ifa nd only if µ + λ is not a root. But −λ + µ is a root if and only if its negative −µ + λ is one, so p
λ,µ= 0 if and only if p
µ,λ= 0.
If λ + µ is a root and p
λ,µ6= 0 then the diagrams above show that kλk = kµk, and it is well known that
in an irreducible root system the Weyl group acts transitively on roots of the same length. This can also
be deduced from the diagrams below, since every pair of roots is contained in a set of roots of rank two.
Since every pair of roots is contained in a subsystem of rank two, this shows that p
λ,µ= p
µ,λwhen these are not 1.
One consequence is an observation of [Tits:1966] (equation (37) of §3.3):
Corollary 3.4. If λ + µ + ν = 0 then
[mystery]
p
λ,µ+ 1
kνk
2= p
µ,ν+ 1
kλk
2= p
ν,λ+ 1 kµk
2.
Proof. Proposition 3.2(structure) implies that
♣
[geometric-lemma]p
λ,µ+ 1
kνk
2= q
λ,µkµk
2, so that by symmetry it suffices to show that q
λ,µ= p
−λ−µ,λ+ 1. But
p
−λ−µ,λ+ 1 = p
−λ,λ+µ+ 1 = p
−λ,µ= q
λ,µ.
A somewhat more elegant proof is offered by diagrams similar to those used to prove Proposition
♣
[geometric-lemma]3.2(structure) . This result will not be used, but in light of later results it is suggestive. It will appear later on that N
λ,µis, up to sign, the same as p
λ+1, but the proof is not straightforward. The parallelism between N
λ,µand p
λ,µ+ 1 ought to be more transparent.
4. Proof of the main result
We must show that the bracket is anti-symmetric and satisfies the Jacobi identity if and only if Proposition
♣
[structure-equations]2.1(a)–(d) hold.
(a) Since
[e
λ, e
µ] = N
λ,µe
λ+µ[e
µ, e
λ] = N
µ,λe
λ+µ, the bracket is anti-symmetric if and only if N
λ,µ= −N
µ,λ. The rest of the Proposition is related to the Jacobi identity
[[x, y], z] + [[y, z], x] + [[z, x], y] = 0 .
Some repetition can be avoided if we keep in mind that it holds for all permutations of x, y, z if and only if it holds for one. We know that we only have to examine the cases where all of x, y, z are among the e
λ. So we have to consider the condition that
[e
λ, [e
µ, e
ν]] + [e
µ, [e
ν, e
λ]] + [e
ν, [e
λ, e
µ]] = 0
for each set of three roots. There are several cases: (i) λ + µ + ν = 0; (ii) λ + µ + ν is neither 0 nor a root;
(iii) λ + µ + ν is one of the three, say ν, which means that λ = −µ; (iv) λ + µ + ν is a root different from any of the three.
Case (ii) is trivial, since in that case all three of the terms in Jacobi’s identity vanish.
(b)(i) Suppose λ + µ + ν = 0. We need only prove this:
Lemma 4.1. Suppose that λ, µ, and ν are three distinct roots that satisfy
[jacobi-1]
λ + µ + ν = 0 .
The Jacobi identity
[e
λ, [e
µ, e
ν]] + [e
µ, [e
ν, e
λ]] + [e
ν, [e
λ, e
µ]] = 0 holds if and only if
N
λ,µkνk
2= N
µ,νkλk
2= N
ν,λkµk
2.
Proof. We have
[e
λ, [e
µ, e
ν]] + [e
µ, [e
ν, e
λ]] + [e
ν, [e
λ, e
µ]]
= N
µ,ν[e
λ, e
−λ] + N
ν,λ[e
µ,−µ] + N
λ,µ[e
ν, e
−ν]
= N
µ,νh
λ+ N
ν,λh
µ+ N
λ,µh
ν= N
µ,ν2λ kλk
2+ N
ν,λ2µ kµk
2+ N
λ,µ2ν kνk
2= N
µ,νkλk
2λ + N
ν,λkµk
2µ + N
λ,µkνk
2ν
= N
µ,νkλk
2− N
λ,µkνk
2λ + N
ν,λkµk
2− N
λ,µkνk
2µ .
Since any two of the roots must be linearly independent, the Jacobi identity holds if and only if N
µ,νkλk
2− N
λ,µkνk
2= N
ν,λkµk
2− N
λ,µkνk
2= 0 . or, equivalently,
N
λ,µkνk
2= N
µ,νkλk
2= N
ν,λkµk
2. (b)(iii) Suppose now that λ is a root and Π a λ-chain of roots.
Lemma 4.2. The Jacobi identity
[jacobi-2]
[e
λ, [e
−λ, e
µ] + [e
−λ, [e
µ, e
λ] + [e
µ, [e
λ, e
−λ] = 0 for all µ in Π if and only if
N
λ,µN
−λ,λ+µ= (p
λ,µ+ 1)q
λ,µfor all µ in Π.
Proof. We have
[e
λ, [e
−λ, e
µ]] + [e
−λ, [e
µ, e
λ]] + [e
µ, [e
λ, e
−λ]]
= N
−λ,µ[e
λ, e
−λ+µ] + N
µ,λ[e
−λ, e
µ+λ] + [e
µ, h
λ]
= N
−λ,µN
λ,−λ+µe
µ− N
λ,µN
−λ,λ+µe
µ− hµ, h
λie
µ= N
−λ,µN
λ,−λ+µ− N
λ,µN
−λ,λ+µ− hµ, h
λi e
µ. So the Jacobi identity holds if and only if
N
λ,−λ+µN
−λ,µ− N
λ,µN
−λ,λ+µ= hµ, h
λi for all λ, µ.
In other words, it suffices to prove:
Lemma 4.3. We have
[carters-lemma]
N
λ,−λ+µN
−λ,µ− N
λ,µN
−λ,λ+µ= hµ, h
λi for all µ in a given λ-chain if and only if
N
λ,µN
−λ,λ+µ= q
λ,µ(p
λ,µ+ 1) . for all µ in the chain.
Proof. Let C
λ,ν= N
λ,νN
−λ,λ+ν. Choose µ to be the ‘left’ end of the chain for which −λ + µ is not a root.
Thus N
−λ,µ, p
λ,µ, and C
λ,−λ+µare all 0.
Since hµ, h
λi = −q
λ,µin these circumstances, the first sequence therefore reads
−C
λ,µ= −q
λ,µC
λ,µ− C
λ,λ+µ= −(q
λ,µ− 2) C
λ,λ+µ− C
λ,2λ+µ= −(q
λ,µ− 4) C
λ,2λ+µ− C
λ,3λ+µ= −(q
λ,µ− 6) while the first is
C
λ,µ= q
λ,µ· 1 C
λ,λ+µ= (q
λ,µ− 1) · 2 C
λ,2λ+µ= (q
λ,µ− 2) · 3 C
λ,3λ+µ= (q
λ,µ− 3) · 4 One is obtained from the other by summing or taking differences.
(b)(iv) Suppose λ + µ + ν + ρ = 0 but no two cancel. We have [[e
λ, e
µ], e
ν] + [[e
µ, e
ν], e
λ] + [[e
ν, e
λ], e
µ]
= N
λ,µ[e
λ+µ, e
ν] + N
µ,ν[e
µ+ν, e
λ] + N
ν,λ[e
ν+λ, e
µ]
= N
λ,µN
λ+µ,ν+ N
µ,νN
µ+ν,λ+ N
ν,λN
ν+λ,µe
λ+µ+νso the Jacobi identity is valid if and only if
0 = N
λ,µN
λ+µ,ν+ N
µ,νN
µ+ν,λ+ N
ν,λN
ν+λ,µ= N
λ,µN
−ν−ρ,ν+ N
µ,νN
−λ−ρ,λ+ N
ν,λN
−µ−ρ,µ= N
λ,µN
ν,ρkρk
2kν + ρk
2+ N
µ,νN
λ,ρkρk
2kλ + ρk
2+ N
ν,λN
µ,ρkρk
2kµ + ρk
2= N
λ,µN
ν,ρkν + ρk
2+ N
µ,νN
λ,ρkλ + ρk
2+ N
ν,λN
µ,ρkµ + ρk
2= N
λ,µN
ν,ρkλ + µk
2+ N
µ,νN
λ,ρkµ + νk
2+ N
ν,λN
µ,ρkν + λk
2according to (b)(i). This concludes the proof of the Theorem.
5. Chevalley’s integral constants The equation
N
λ,µN
−λ,λ+µ= (p
λ,µ+ 1)q
λ,µmay be rewritten as
N
λ,µN
−λ,λ+µ= (p
λ,µ+ 1)p
−λ,µ= (p
λ,µ+ 1)(p
−λ,λ+µ+ 1) which makes plausible the following:
Proposition 5.1. (Chevalley’s integral formula) We can choose the structure constants so that
[chevalley]
|N
λ,µ| = p
λ,µ+ 1 for all roots λ, µ with λ + µ a root.
This might have also been suggested by a comparison of Corollary 3.4(structure) and (b)(iii) above.
♣
[mystery]There are two parts to the proof. Here is one:
Lemma 5.2. In any semi-simple Lie algebra
[chevalley-equation]
N
λ,µN
−λ,−µ= −(p
λ,µ+ 1)(p
−λ,−µ+ 1) = −(p
λ,µ+ 1)
2.
Proof. Since ( −λ) + (−µ) + (λ + µ) = 0, (b)(i) tell us that N
−λ,−µkλ + µk
2= N
λ+µ,−λkµk
2= − N
−λ,λ+µkµk
2but by the Geometric Lemma
kλ + µk
2kµk
2= p
λ,µ+ 1 q
λ,µso that N
λ,µN
−λ,−µ= −(p
λ,µ+ 1)
2if and only if
N
λ,µN
−λ,λ+µ= −q
λ,µ(p
λ,µ+ 1) .
This reasoning is adequate, but I have to confess I find it rather unsatisfactory. Chevalley’s equation is very simple, but its proof is not so simple as one might hope for. The basis elements e
λmay be chosen arbitrarily, but any two differ by a constant. If we change e
λto ce
λ, however, we change e
−λto c
−1e
−λand each N
λ,µby a factor of (c
λc
µ/c
λ+µ). The point at the moment is that the product N
λ,µN
−λ,−µdoes not change. It is thus something necessarily intrinsic—independent of the basis choice, and therefore likely to be interesting. There should be a more direct way to interpret it.
At any rate, this allows us to apply another observation due to Chevalley. He has shown that there are certain choices of basis that are much better than others. There exists an involution θ of g, called the opposition involution , that acts as −1 on h, hence swapping root spaces for λ and −λ. We can then scale e
λso that θ(e
λ) = −e
−λ. With this choice of basis, we have
N
λ,ν= ±(p
λ,µ+ 1)
for all roots λ, µ.
6. Computation of structure constants—mathematics
Start by assigning a linear order to the simple roots ∆. Extend this to one of all positive roots such that λ < µ if
HT(λ) <
HT(µ). I recall that
HT
X
∆