Semisimple Lie Algebras and the Root Space

Decomposition

Ebrahim

May 1, 2015

This document will develop just the material needed to describe a semisimple Lie algebra in terms of its root
space decomposition. This takes place in section 5, which is our central focus. The major theorems that we
often cite are put off until section 6, and some critical linear algebra machinery is is left for section 7, which
still needs to be filled in.

1 Lie Algebra Basics

Let k be a field. A Lie algebra g over k is a k-vector space equipped with an antisymmetric k-bilinear map
[·, ·] : g × g → g that satisfies the Jacobi identity:
(1) (antisymmetry) [x, x] = 0
(2) (Jacobi identity) [[x, y], z] + [[y, z], x] + [[z, x], y] = 0

This bilinear map is called the Lie bracket of the Lie algebra. A subalgebra S of a Lie algebra is a vector
subspace which is closed under the bracket (i.e. [S, S] ⊆ S). An ideal I of a Lie algebra is a vector subspace
which is more strongly closed under the bracket in that [g, I] ⊆ I. A Lie homomorphism is a k-linear map
φ : g → h that preserves the brackets:
φ([x, y]g ) = [φ(x), φ(y)]h
The kernel of a Lie homomorphism is always an ideal. An important example is the adjoint representation
map:
ad : g → gl(g) .
That it is a Lie homomorphism is precisely the statement of the Jacobi identity. Another exact restatement
of the Jacobi identity is contained in the fact that ad takes values in the subalgebra Der(g) of derivations.
The notion of a Lie algebra is meant to be an abstraction of the additive commutators of associative algebras.
While lie algebras are almost never associative as algebras, they have the very property that commutators
of associative algebras satisfy by taking place in an associative setting- the Jacobi identity.

In this spirit we often say that x, y ∈ g “commute” when [x, y] = 0. The ideal Z(g) of elements of g that
commute with everything in g is called the center of g. It is the kernel of the adjoint representation map.
We also say g is “abelian” in case Z(g) = g.

1
The sum, intersection, and bracket of Lie ideals gives a Lie ideal. (Note: the bracket [S, T ] of two subspaces
S, T ⊆ g of a Lie algebra is taken to be the subspace generated by the elementwise brackets). If I ⊆ g is
an ideal, then there is a well-defined induced Lie bracket on the quotient vector space g/I and a canonical
surjective Lie homomorphism g → g/I. We obtain standard homomorphism theorems just as in the theory
of modules:

1. If φ : g → h is a Lie homomorphism then it induces an isomorphism g/ker φ ∼

= im φ.
2. (S + I)/I ∼
= S/S ∩ I for a subalgebra S and an ideal I of g.
3. For an ideal I ⊆ g, there is a correspondence between ideals of g/I and ideals of g that contain I.

4. (g/I)/(J/I) ∼
= g/J for ideals I ⊆ J ⊆ g.

We define the direct sum of Lie algebras g and h to be the vector space g ⊕ h with a bracket that kills
cross-terms:
[a + x, b + y] = [a, b]g + [x, y]h for a, b ∈ g and x, y ∈ h.
Even though we use “⊕” to denote this construction, it is only a product and not generally a coproduct in
the category of Lie algebras.

A Lie algebra g is said to be simple if it is nonabelian and has no proper nontrivial ideals.

2 Solvability

The derived series of a Lie algebra g is the sequence of ideals

g ⊇ [g, g] ⊇ [[g, g], [g, g]] ⊇ · · ·

defined recursively by g(0) = g and g(i+1) = [g(i) , g(i) ]. We call g solvable if its derived series hits rock
bottom; that is, g is solvable if g(n) = 0 for some n. For example abelian lie algebras are solvable, and simple
lie algebras are not.

Theorem 1: Let g be a Lie algebra.

(1) If g is solvable then so are all its subalgebras and homomorphic images.
(2) If I ⊆ g is a solvable ideal then: g/I solvable ⇒ g solvable.
(3) The sum of solvable ideals is solvable.

Proof: If S ⊆ g is a subalgebra then S (i) ⊆ g(i) . Similarly if φ : g → h is an epimorphism, then h(i) = φ(g(i) )
(by induction). This proves (1). To prove (2) suppose that (g/I)(n) = 0 and that I is solvable. Apply the
formula used in (1) to the canonical epimorphism g → g/I to see that g(n) ⊆ I. Then by part (1) again g(n)
is solvable, so g is obviously solvable as well.

(3): Now suppose I, J ⊆ g are solvable ideals. Then by (1) I/(I ∩ J) is solvable, and so we may say the
same of (I + J)/J ∼
= I/(I ∩ J). It follows from (2) that I + J is solvable.

2
Part (3) shows that a finite-dimensional Lie algebra g must have a unique maximal solvable ideal. We call
this ideal the radical of g and denote it by Rad(g). We say g is semisimple if Rad(g) = 0. Semisimplicity
is then the opposite extreme condition to solvability. The word “solvable” sounds like it has particularly
positive connotations, but really the two extreme cases of solvability and semisimplicity are both open to
classification for Lie algebras over a sufficiently nice field. The radical of a Lie algebra, the “solvable part,”
ends up being the barrier to obtaining a decomposition into simple pieces.

Exercises: Show that a simple Lie algebra is semisimple. Show that a Lie algebra is semisimple iff it has
no nonzero abelian ideals. Show that g/Rad(g) is semisimple.

This last fact suggests that we can try to understand all finite dimensional Lie algebras g by understanding
all the solvable ones (like Rad(g)) and all the semisimple ones (like g/Rad(g)). Lie’s theorem (section 6.2)
shows that, over an algebraically closed field of characteristic 0, every solvable finite-dimensional Lie algebra
is a subalgebra of an upper triangular matrix algebra. The remaining task is then to understand the structure
of semisimple Lie algebras.

3 Semisimple Lie Algebras

In section 2 we defined a semisimple Lie algebra to be one with trivial radical. We will see that this is
the same as saying that a particular bilinear form is nondegenerate, and that it is also the same as saying
that there is a decomposition in terms of simple ideals. Throughout this section let g be an arbitrary finite
dimensional lie algebra over an algebraically closed field k of characteristic 0.

Define the Killing form κg : g × g → k by

κg (x, y) := Tr(ad(x)ad(y))
This gives a symmetric bilinear form on g which is associative in the sense that
κg ([x, y], z) = κg (x, [y, z]) .
The radical of a symmetric bilinear form β : V × V → k on a k-vector space V is defined by
Rad(β) = {x ∈ V 0 = β(x, ·) : V → k} .
Nondegenaracy of β is then equivalent to Rad(β) = 0. Note that the radical of an associative symmetric
bilinear form on a lie algebra is an ideal.

Theorem 2: Let I be an ideal of g. Then κI = κg |I×I .

Proof: This follows easily from the linear algebra fact that if W is a subspace of the finite dimensional
vector space V , and if φ : V → V maps V into W , then the trace of φ agrees with the trace of φ|W : W → W .

Theorem 3: g is semisimple iff κg is nondegenerate.

Proof: Suppose that Rad(g) = 0. Let s = Rad(κg ). If x ∈ s and y ∈ [s, s] then

Tr(ads (x)ads y) = κs (x, y) = κg (x, y) = 0 .

3
By Cartan’s criterion (section 6.3) it follows that s is solvable. Thus, s ⊆ Rad(g) = 0.

Conversely, suppose that s = 0. We will show that g has no nontrivial abelian ideals. Let I be an abelian ideal
of g. Consider any x ∈ I and y ∈ g. We have that ad(x)ad(y) maps g into I, and that (ad(x)ad(y))2 = 0 since
it maps g into [I, I] = 0. Since nilpotent linear endomorphisms are traceless, we have 0 = Tr(ad(x)ad(y)) =
κ(x, y). Since y was arbitrary, we get x ∈ s. Therefore I ⊆ s = 0.

Theorem 4: Suppose that g is semisimple. Then it is the direct sum of (all) its simple ideals.

Proof: Our assumption implies that κg is nondegenerate. It can get us the required direct sum splitting.

Claim: If I ⊆ g is an ideal, then g = I ⊕ I ⊥ , and I ⊥ is an ideal as well.

Pf: If x ∈ I ⊥ and y ∈ g then for all z ∈ I we have

κg ([x, y], z) = κg (x, [y, z]) = 0 .

Therefore I ⊥ is an ideal. If x ∈ I ∩ I ⊥ and y ∈ [I ∩ I ⊥ , I ∩ I ⊥ ] then

0 = κg (x, y) = κI∩I ⊥ (x, y) = Tr(adI∩I ⊥ (x)adI∩I ⊥ (y)),

so it follows from Cartan’s criterion (section 6.3) that I ∩ I ⊥ is solvable. Then since g is semisimple
we get I ∩ I ⊥ = 0. Since κg is nondegenerate, this is enough to get g = I ⊕ I ⊥ as a vector space. This
is, however, automatically a direct sum in the sense of Lie algebras, for [I, I ⊥ ] ⊆ I ∩ I ⊥ = 0.

We continue the main proof by induction on the dimension of g. At low dimensions, g has no proper nontrivial
ideals and is therefore simple. Assume this is not the case for g, and assume that the theorem holds for
lower dimensions. Let g1 be a minimal proper nontrivial ideal of g. It is simple (it has no proper nontrivial
ideals, and it is nonabelian by an exercise from section 2). By the claim we have g = g1 ⊕ g⊥ 1 . Applying the
induction hypothesis to g⊥ 1 , we see that g splits into a direct sum of simple ideals: g = g1 ⊕ · · · ⊕ gt . Do all
simple ideals of g appear in this decomposition?

If I is any simple ideal of g then [I, g] is an ideal of I. It is nonzero because Z(g) ⊆ Rad(g) = 0. We thus
have:
I = [I, g] = [I, g1 ⊕ · · · ⊕ gt ] = [I, g1 ] ⊕ · · · ⊕ [I, gt ]
(it helps to notice that [I, gi ] ⊆ gi to see why the direct sum remains in the last equality). Since I is simple
all summands must be trivial except, say, the ith one. Then I = [I, gi ] ⊆ gi . But gi is simple too so we get
I = gi .

An easy consequence of this is that a semisimple g has [g, g]. Another consequence is that all ideals and
quotients of a semisimple g are semisimple, due to:

Theorem 5: Any ideal of g is a direct sum of certain of the simple ideals of g.

Proof: Let I ⊆ g be an ideal, with g = g1 ⊕ · · · ⊕ gt the decomposition into simples. Notice that for each
gi , since [I, gi ] is an ideal of gi and of I, it is either 0 or gi . So for each i we have either [I, gi ] = 0 or gi ⊆ I.
We will show that I is spanned by the gi ’s that it contains; this will complete the proof.

4
 P
Consider any x ∈ I. Write x as xi with xi ∈ gi . Consider an i such that [I, gi ] = 0 instead of gi ⊆ I.
Suppose, by way of contradiction, that xi 6= 0. Choose a y ∈ gi so that [xi , y] 6= 0 (using Z(gi ) = 0). Then
0 = [x, y] = [xi , y] 6= 0.

4 Representations of sl2 (k)

Let k be an algebraically closed field of characteristic 0.

Results from this section are needed for the root space decomposition to come. We begin with a very quick
overview of representations, and then we classify the finite-dimensional irreducible representations of sl2 (k).

4.1 Overview

A representation of a lie algebra g over k is a Lie homomorphism φ : g → gl(V ) for some vector space V
over k. Since gl(V ) is really just End(V ) as a set, we view a representation as an action of a Lie algebra on
a vector space. Instead of the usual associativity rule for actions, we have

[x, y].v = x.(y.v) − y.(x.v)

With this viewpoint we often refer to a representation of g as a g-module. There are the usual notions
of submodule, g-homomorphism, quotient module, simple (irreducible) module, and completely reducible
module. There are also notions of direct sum, tensor product, dual, and Hom for g-modules. We also get
the usual isomorphism theorems, and a version of Schur’s lemma.

An important example is the adjoint representaiton adg : g → gl(g) which makes g into a g-module. In this
case submodules are ideals, simple modules are simple Lie algebras, etc.

According to Weyl’s theorem (section 6.4), understanding the finite-dimensional representations of a semisim-
ple Lie algebra amounts to understanding the irreducible ones.

4.2 The Classification for sl2 (k)

01


Let φ :
sl2 (k) → gl(V ) be

an irreducible finite-dimensional representation of sl2 (k). Define x = 00 ,
y = 01 00 , and h = 10 −1
0
, a basis for sl2 (k) in which

[h, x] = 2x
[h, y] = −2y
[x, y] = h .

Since h is semisimple as an element of gl2 (k), it’s usual Jordan-Chevalley decomposition (as a matrix) is h+0.
This must (see section 7) also be the abstract Jordan-Chevalley decomposition, so φ(h) + φ(0) = φ(h) + 0 is
the decomposition for φ(h) ∈ gl(V ). Thus, φ(h) is semisimple. Since k is algebraically closed, it follows that h

5
acts diagonally on V ; we obtain a decomposition of V as a direct sum of eigenspaces Vλ = {v ∈ V h.v = λv}
of the action of h. We refer to the eigenvalues of φ(h) as weights of h, and we refer to the eigenspaces of the
action of h as weight spaces of h.

Suppose that v ∈ Vλ is a vector of weight λ for h. What can we say about x.v or y.v? Well,

h.(x.v) = x.(h.v) + [h, x].v

= λ(x.v) + (2x).v = (λ + 2)x.v

h.(y.v) = y.(h.v) + [h, y].v

= λ(y.v) + (−2y).v = (λ − 2)y.v .

We see that x.v ∈ Vλ+2 and y.v ∈ Vλ−2 when v ∈ Vλ . So x and y shift us up and down the eigenspaces of h,
with x bumping up eigenvalues by 2 and y bumping them down by 2. (At this point we only know that the
eigenvalues are elements of k, so there is not really a notion of “up.” But we can make sense of it later).

Starting with a particular eigenvector v for h, we can iteratively bump up its eigenvalue by applying x.
Finite-dimensionality dictates that eventually this process hits 0. We obtain a vector of maximal “height”,
a nonzero vector v0 ∈ Vλ for some λ with x.v0 = 0. Let’s collect all the vectors “below”, with a scale factor
for convenience:
1 i
vi := y .v0 for i ∈ N.
i!
(Here the y i is a notation for an i-fold application of y). Define also v−1 := x.v0 = 0 and similarly v−2 = 0.
We obtain:

h.vi = (λ − 2i)vi
y.vi = (i + 1)vi+1
x.vi = (λ − i + 1)vi−1

for all i ≥ −1. The first two lines are obvious and the third can be shown by induction. The bumping
down process also has to hit 0 eventually, so let m be the smallest natural number with vm+1 = 0. Of
course vm+1 , vm+2 , · · · must all vanish. The formulas above show that khv0 , v1 , · · · i = khv0 , · · · , vm i is an
sl2 (k)-submodule of V . It’s nontrivial because it contains v0 , so by irreducibility of V we see that v0 , · · · , vm
is a basis for V . Also since
0 = x.vm+1 = (λ − m)vm ,
we get λ = m. So weights for h are always integers. In the basis v0 , · · · , vm of V , the endomorphsisms
φ(h), φ(x), and φ(y) look like
  0 m   
m 0
 m−2    0 m−1 
  1 0 

m−4
  . .  
,  2 0


 , 

0 .  

 . .. 
 

..

   . .. . .. 
  . 1  
−m 0 m 0

respectively. We therefore have that there is at most one irreducible sl2 (k)-module, up to isomorphism,
for each dimension m + 1. And we see that there is precisely one for each dimension if we simply use the
matrices above to construct them.

6
5 The Root Space Decompositon

Let g be a nontrivial semisimple finite-dimensional Lie algebra over an algebraically closed field k of charac-
teristic zero. In this section we put together many of the powerful tools developed throuhgout the document
and we use them to study g.

5.1 Toral Subalgebras

Definition: A subalgebra h of g is called toral if all its elements are ad-semisimple.

First notice that g must have an element which is not ad-nilpotent, for otherwise Engel’s theorem would
imply that g is nilpotent and therefore solvable. The semisimple part of this element would generate a one-
dimensional subalgebra of g consisting purely of ad-semisimple elements. In other words, nontrivial toral
subalgebras exist.

Theorem 6: A toral subalgebra of a semisimple Lie algebra is abelian.

ad
Proof: Suppose that h ⊆ g is a toral subalgebra and x ∈ h. Then g −→ g is a semisimple endomorphism,
adh
so its restriction h −−→ h is too. Since k is algebraically closed, this means that adh (x) is diagonalizable.
Therefore we can show that adh (x) = 0 by showing that it has no nonzero eigenvalues.

Suppose, by way of contradiction, that [x, y] = ay with a ∈ k \ {0} and y ∈ h. By the same argument laid out
above, y is adh -semisimple. Thus x can be written as a linear combination of eigenvectors of adh (y) : h → h.
So adh (y)(x) is a linear combination of eigenvectors of y with nonzero eigenvalue. But our assumption was
that
adh (y)(x) = [y, x] = −ay,
an eigenvector of adh (y) with eigenvalue 0.

Recall that
the trick to classifying representations of sl2 (k) revolved around the eigenspace decomposition of
h = 10 −10
. The magic that happened there was that the other basis elements of sl2 (k) ended up behaving
like “ladder” operators; that is, their behavior could be understood as simply shifting eigenvectors of h up
and down the ladder of eigenvalues. For representations of more general Lie algebras we search for something
analogous to “h.” The key is to search for toral subalgebras!

Notation: If L ⊆ gl(V ) is a linear Lie algebra and v ∈ V is a common eigenvector for the elements of L, then
v has an eigenvalue α(x) for each x ∈ L. The resulting map α : L → k turns out to be a linear functional
α ∈ L∗ , and we call it the “eigenvalue” or, more correctly, the weight of v. There might be additional
independent vectors with the same weight; we collect the all the vectors of weight α into a subspace which
we call Vα .

7
5.2 The Decomposition

Now let h be a maximal toral subalgebra of our semisimple Lie algebra g. We’ve established that h is nontrivial
and abelian. That it is abelian implies adg (h) ⊆ gl(g) consists of commuting semisimple endomorphisms of
g. It follows that they are simultaneously diagonalizable, and we obtain a (vector space) splitting of g into
weight spaces: M
g= gα ,
α∈h∗

where
gα := {x ∈ g (∀h h ∈ h ⇒ [h, x] = α(h)x) } .
Notice that g0 is really just the centralizer of h in g, which gives

h ⊆ g0

since h is abelian. Define

Φ = {α ∈ h∗ \ {0} gα 6= 0},
a finite set of nonzero weights which we call the set of roots of g relative to h. The decomposition above can
be written as
M
g = Cg h ⊕ gα
α∈Φ

and it is called the root space decomposition of g (relative to a maximal toral subalgebra h). We will soon
see that in fact Cg h = h. First, let us explore the interaction of the Lie algebra structure with this vector
space decomposition. I will box the important properties we collect along the way.

For convenience we shall often denote brackets using the notation of actions (the adjoint action). Suppose
that x ∈ gα and y ∈ gβ . For h ∈ h we have

h.[x, y] = [h.x, y] + [x, h.y] = (α(h) + β(h))[x, y] = (α + β)(h)[x, y] ,

so we have
[gα , gβ ] ⊆ gα+β for α, β ∈ h∗ .
Thus, if x ∈ gα and α 6= 0 then ad(x) : g → g is nilpotent.

Now suppose that α, β ∈ h∗ with α + β 6= 0. Choose an h ∈ h so that (α + β)(h) 6= 0. Consider any x ∈ gα

and y ∈ gβ . Let κ be the Killing form of g (a symmetric associative bilinear for on g which is nondegenerate
because g is semisimple). Then:

(α + β)(h)κ(x, y) = α(h)κ(x, y) + β(h)κ(x, y)

= κ(α(h)x, y) + β(h)κ(x, β(h)y)
= κ(h.x, y) + β(h)κ(x, h.y)
= κ([h, x], y) + β(h)κ(x, [h, y])
= −κ([x, h], y) + β(h)κ(x, [h, y]) = 0

so we get κ(x, y) = 0. Thus,

for α, β ∈ h∗ with α + β 6= 0, gα is κ-orthogonal to gβ .

For example g0 is orthogonal to gα for α ∈ Φ. An easy consequence is that κ restricts to a nondegenerate

form on g0 .

8
We’ll have to pause momentarily for a rather technical proof. It is reasonable to skip this proof and return
to it later. For now just note that this is where the maximality of the toral subalgebra h comes in handy.

Theorem 7: Cg h = h

Proof: We only need to prove ⊆. Let C = Cg h.

Suppose that x ∈ C. Then h is contained in the kernel of adg (x). If x = s + n is abstract JC decomposition
(section 7), then adg (x) = adg (s) + adg (n) is the usual JC decomposition. Since ad(s) and ad(n) are both
expressible as k-polynomials evaluated at ad(x) with zero constant term, they both also have h in their
kernel. Thus we have shown that whenever x ∈ C, its semisimple and nilpotent parts are also in C.

Suppose that s ∈ C and s is ad-semisimple. Then h + khsi is an abelian subalgebra of g which is also toral,
so by the maximality of h we obtain s ∈ h. Thus we see that the semisimple things in C make it into h.

Suppose that h ∈ h and that κg (h, h) = 0. We’d like to show that h = 0. Consider any c ∈ C, and let
c = s + n be its abstract JC decomposition. We saw that in this situation we have s, n ∈ C and that we
then further get s ∈ h. Then

κ(h, c) = Tr(ad(h)ad(s)) + Tr(ad(h)ad(n))

= κ(h, s) + Tr(a nilpotent endomorphism) = 0 .

(The first term vanishes by the assumption κg (h, h) = 0, and the second vanishes because the composite
of commuting endomorphisms is nilpotent when one of them is nilpotent). Since c was arbitrary, we get
κ(h, C) = 0. But we argued above that κ was nondegenerate on g0 = C, so h = 0. Thus we have shown that
κ is nondegenerate on h.

Suppose that x ∈ C with abstract JC decomposition x = s + n. We saw that s ∈ C and therefore s ∈ h. In

that case adC (s) = 0. Therefore adC (x) = adC (n) is nilpotent, and this holds for arbitrary x ∈ C. Thus we
have shown that C is nilpotent (by Engel’s theorem, section 6.1).

Suppose that h ∈ h and c, c0 ∈ C. Then

κg (h, [c, c0 ]) = κg ([h, c], c0 )

= κg (0, c0 ) = 0

so h is κg -orthogonal to [C, C]. Since we already saw that κg is nondegenerate on h, it follows that h∩[C, C] =
0.

We now cite the fact that a nontrivial ideal of a nilpotent Lie algebra nontrivially intersects its center (this
can be shown using the first claim in my proof of Engel’s theorem; view the ideal as a module over the entire
Lie algebra and find a nonzero element that is killed by the action). Suppose, by way of contradiction, that
[C, C] 6= 0. Since C is nilpotent, there is a nonzero z ∈ [C, C] ∩ Z(C). Then z ∈
/ h since h ∩ [C, C] = 0. So z
cannot be semisimple, for that would give z ∈ h. Let n be the nilpotent part of z; now we know that n 6= 0.
We have n ∈ C. We also have n ∈ Z(C), because ad(n) is a k-polynomial evaluated at ad(z). It follows that
κg (n, C) = 0, since its elements are traces of nilpotent maps. This contradicts the nondegeneracy of κg on
C. Thus we have proven that [C, C] = 0.

Finally, suppose that C 6= h and choose an x ∈ C \ h. Let x = s + n be its JC decomposition. Then n 6= 0,

since otherwise x = s ∈ h. And n ∈ C. Since [C, C] = 0, ad(n) commutes with ad(y) for any y ∈ C. Then
for any y ∈ C we get κ(n, y) = Tr(ad(n)ad(y)) = 0, contradicting the nondegeneracy of κ on C.

9
 

In light of this fact and the comment above, we see that

κg restricts to a nondegenerate form on h .

∼
=
→ h∗ . For α ∈ h∗ , let tα ∈ h denote
The restriction of κ to h therefore establishes a linear isomorphism h −
the image of α under the inverse of this isomorphism. In other words, for each α ∈ h∗ , tα is the unique
element of h with
α(h) = κ(h, tα ) .

5.3 A First Look at the Roots

The set of roots of g relative to h then corresponds to a certain finite set of nonzero vectors in h:

{tα α ∈ Φ} ⊆ h .

Another fact that will come in handy is that

Φ spans h∗ ,

otherwise, since span(Φ) would be a proper subspace of h∗ , it would correspond to a nontrivial subspace of
h. The “corresponding” subspace of h is
\ \
ker(f ) = ker(α) .
f ∈span(Φ) α∈Φ

So we would have a nonzero h ∈ h with α(h) = 0 for all α ∈ Φ. But then 0 = x.h = [x, h] for all x ∈ gα , for
any α ∈ Φ. And we already have [h, h0 ] = 0 for h0 ∈ g0 = h, so this places h in Z(g). Impossible!

Now suppose that α ∈ Φ. Then if −α ∈ / Φ we would get that gα is orthogonal to all of g, which contradicts
the nondegeneracy of κ. So Φ is closed under negatives:

α ∈ Φ ⇒ −α ∈ Φ .

Next we express brackets in terms of κ and Φ. Suppose that α ∈ Φ, x ∈ gα , and y ∈ g−α . Then for arbitrary
h ∈ h we have

κ(h, [x, y]) = κ([h, x], y)

= α(h)κ(x, y)
= κ(h, tα )κ(x, y)
= κ(h, κ(x, y)tα )

so we get
[x, y] = κ(x, y)tα for α ∈ Φ, x ∈ gα , y ∈ g−α

from our earlier observation that κ is nondegenerate on h. So each [gα , g−α ] ⊆ g0 = h is at most one-
dimensional. In fact,
for α ∈ Φ, [gα , g−α ] = khtα i is one-dimensional .

10
To see that it is nonzero: Choose a nonzero x ∈ gα and notice that κ(x, g−α ) 6= 0 (because we would
otherwise get that x is orthogonal to all of g).

Next, let’s show that the tα themselves have nonzero “norm”:

α(tα ) = κ(tα , tα ) is nonzero for α ∈ Φ .

There are x ∈ gα and y ∈ g−α with κ(x, y) nonzero, since otherwise we would have gα orthogonal to all of g.
After rescaling we may assume that κ(x, y) = 1. Suppose, by way of contradiction, that α(tα ) = 0. Then:

[tα , x] = tα .x = α(tα )x = 0
[tα , y] = tα .y = −α(tα )y = 0
[x, y] = κ(x, y)tα = tα .

Let s = khx, y, tα i; the relations we’ve found show that this is a 3 dimensional solvable subalgebra. By Lie’s
theorem (section 6.2), s stabilizes a flag in g and hence acts by upper triangular matrices. Since tα ∈ [s, s], it
must act by a strictly upper triangular matrix. Therefore adg (tα ) is a nilpotent endomorphism of g. But it is
also semisimple since tα ∈ h and h is toral. It follows that adg (tα ) = 0, and then that tα = 0, a contradiction.

For every α ∈ Φ and every nonzero x ∈ gα we can find a little copy of sl2 (k) containing x. First use the fact
that κ(x, g−α ) 6= 0 (by nondegeneracy of κ) to obtain a y ∈ g−α with κ(x, y) 6= 0. Rescale y to ensure that
2
κ(x, y) =
κ(tα , tα )

and define h by
2tα
h= .
κ(tα , tα )
We then obtain the relations
2tα
[x, y] = κ(x, y)tα = =h
κ(tα , tα )
2
[h, x] = h.x = α(h)x = α(tα )x = 2x
κ(tα , tα )
−2
[h, y] = h.y = α(h)y = α(tα )y = −2y .
κ(tα , tα )

Summarizing,

for α ∈ Φ and x ∈ gα nonzero, there are y ∈ g−α and h ∈ h so that khx, y, hi ∼

= sl2 (k) .

Throughout the next subsection fix a particular α ∈ Φ, and let sα denote the copy of sl2 (k) that we just
plucked out of g.

5.4 The Action of sα

The next stage in understanding the root space decomposition will be to understand how sα acts on g, which
is an sα -module via
ad
sα ,→ g −→ gl(g) .

11
Define the subspace m ⊆ g by M M
m= gcα = h ⊕ gcα
c∈k c∈k×

and notice that m is an sα -submodule of g. What are the weights of h on m (i.e. what are the eigenvalues
of adm (h) : m → m)? There’s 0, on elements of h. Then, if c ∈ k × and cα ∈ Φ, there’s the weight 2c:
 
2tα
h.z = cα(h)z = cα z = 2cz for nonzero z ∈ gcα .
κ(tα , tα )

By Weyl’s theorem (section 6.4) and the fact that sα is semisimple, every sα -module is a direct sum of
irreducible sα -modules. So m is a direct sum of irreducble sα -modules, and we know exactly what those
look like- there’s one for each dimension, and an (m + 1)-dimensional one has the following list of weights
for h: m, m − 2, m − 4, · · · , −m. In particular the weights are all integers. So the weights of h on m are all
integers. Therefore if c ∈ k with cα ∈ Φ, then c must be a half-integer.

Notice that m, m − 2, · · · , −m all have the same parity, so any even weight for h in m would be for an
eigenvector of adm (h) that lives in an irreducible sα -submodule of m that also contains something with
weight zero for h. Can we account for all the irreducible submodules with zero-weight-for-h vectors in m?

Claim: All z ∈ m with h.zL = 0 lie in h and therefore in the sα -submodule h + sα .

Pf: By construction m = c∈k gcα is the eigenspace decomposition for h (i.e. for adm (h)), with gcα
being the eigenspace corresponding to 2c. So g0 = h indeed contains everything that is killed by h.
To see that h + sα is a submodule, notice that h = ker(α) ⊕ khhi (as a vector space) and that sα acts
trivially on ker(α):

for z ∈ ker(α), h.z = 0

x.z = −[z, x] = −α(z)x = 0
y.z = −[z, y] = α(z)y = 0

It follows that h + sα = ker(α) + sα is a submodule.

The irreducibles that this submodule decomposes into are precisely sα and a bunch of one-dimensional
submodules that live in ker(α), and these are apparently the only irreducible submodules in the decomposition
of m that house zero-weight-for-h vectors. Any even weight for h in m then corresponds to a vector which
either (1) lives in sα or (2) lives in one of these one-dimensional submodules in ker(α) (in which case the
weight is 0). So any nonzero even weight for h in m corresponds to a vector which lives in sα , where we know
the weights for h are 2, 0, −2. The only even weights for h in m are 2, 0, −2, and the vectors corresponding
to ±2 live in sα itself.

One consequence of this is that 2α ∈ / Φ, for otherwise 4 would arise as a weight for h in m. Of course α was
arbitrary, so this is a general principle: twice a root is never a root! Since α is by assumption a root, 21 α
cannot be. Therefore 1 never occurs as a weight for h in m.

Knowing this, we have accounted for all the sα -submodules of m in its decomposition into a direct sum of
irreducibles:
• Any irreducible sα -submodule of even dimension must have 1 as a weight of h on it, so no even-
dimensional irreducibles appear.
• Any irredicble sα -submodule of odd dimension greater than 3 would have 4 as a weight of h on it, so
none of those appear either.

12
 • Only one 3-dimensional irreducible appears: sα . This is because the only vectors in m corresponding
to a weight of ±2 for h are the ones in sα .
• A 1-dimensional sα -submodule of m is really just spanned by a vector that h kills. Any such vector
lies in g0 = h.

All this implies that

m = h + sα .
From this we also get that gα = khxi, that g−α = khyi, and that gcα = 0 for c ∈ k \ {1, 0, −1}. This really
holds for arbitrary α:

if α ∈ Φ then −α ∈ Φ and dim(g±α ) = 1, but no other multiples of α are in Φ .

At this point we understand pretty well how sα acts on m. What about the rest of g? Fix a β ∈ Φ \ {±α},
so that β is a root which is not a multiple of α. Define the following sα -submodule of g:
M
mβ = gβ+iα .
i∈Z

By the assumption on β we are guaranteed that β + iα is never 0. So each summand above has dimension
1 or 0. If z ∈ gβ+iα with i ∈ Z, then
h.z = (β + iα)(h)z = (β(h) + 2i)z
so the weights of h on mβ are all of the form β(h) + 2i with i ∈ Z. These are all distinct weights and each
occurs at most “once” (in the sense that each has a weight space of dimension 1 or 0).

Claim: mβ is an irreducible sα -module

Pf: By Weyl’s theorem mβ is a direct sum of irreducibles. If there were more than one irreducible
in the direct sum, then one of two impossibilities would obtain: (a) 0 or 1 occurs more than once as
a weight for h or (b) 0 and 1 both occur as weights for h (this could not be if the weights are all
in {β(h) + 2i i ∈ Z}). Case (a) occurs when mβ has two odd-dimensional or two even-dimensional
irreducible summands in its decomposition, and case (b) occurs when mβ has an irreducible summand
of odd dimension and one of even dimension.

Now the structure of mβ as an sα -module is completely pinned down by one number- its dimension. Let
q ∈ Z be the largest integer for which β + qα ∈ Φ, and let r ∈ Z be the largest integer for which β − rα ∈ Φ
(note that r, q ≥ 0). Then β(h) + 2q, β(h) − 2r are respectively the highest and lowest weights for h on mβ .
We have a couple of ways of listing all the weights of h on mβ :
β(h) + 2q, β(h) + 2q − 2, · · · , −(β(h) + 2q) or
β(h) + 2q, β(h) + 2q − 2, · · · , β(h) − 2r .
Equating the two ways of writing the lowest weight yields r = β(h) + q. We also get a nice list of all the
roots of the form β + iα:
β + qα, β + (q − 1)α, · · · , β, · · · , β − rα ∈ Φ

Interestingly, we have shown that β(h) is an integer (namely r − q). This is true for arbitrary β ∈ Φ \ {±α},
but of course we also have that ±α(h) = ±2 is an integer. So in general for arbitrary roots α, β ∈ Φ we have
that  
2tα
β
κ(tα , tα )

13
is an integer. These are called Cartan integers. In our case since β(h) = r − q, we see that −r ≤ −β(h) ≤ q
and consequently β − β(h)α makes it into the list above (i.e. into Φ). This is true for arbitrary β ∈ Φ \ {±α},
but we also have
±α ∓ α(h)α = ±α ∓ 2α = ∓α ∈ Φ .
So in general for arbitrary α, β ∈ Φ we have that
 
2tα
β−β α∈Φ.
κ(tα , tα )

5.5 The Structure of (h∗ , κ, Φ)

We can turn κ into a nondegenerate symmetric bilinear form on h∗ via

hγ, δi = κ(tγ , tδ ) .

Since Φ spans h∗ , we can get a k-basis ∗

P i of h consisting of roots α1 , · · · , αl ∈ Φ. If β ∈ Φ, then there are
1 l
unique c , · · · , c ∈ k so that β = c αi .

Claim: c1 , · · · , cl lie in the primeP

field Q of k.
Pf: For any 1 ≤ j ≤ l, hβ, αj i = ci hαi , αj i. Each hαj , αj i = κ(tαj , tαj ) is nonzero. We have

l
2hβ, αj i X 2hαi , αj i i
= c
hαj , αj i i=1
hαj , αj i

Note that the left hand side and the coefficients of the ci on the right hand side are all Cartan integers.
Define an l × l matrix A ∈ Ml (k) by
2hαi , αj i
Aji = ,
hαj , αj i
~ ~c ∈ k l by
and note that A in fact lies in Ml (Q). Define β,

2hβ, αj i
β~j =
hαj , αj i

and ~ci = ci . Above we see that β~ = A~c, and β~ in fact lies in Ql . A is invertible in Ml (k) because
α1 , · · · , αl are a k-linearly independent. Therefore A is invertible in Ml (Q) and ~c = A−1 β~ lies in Ql .

This establishes that

EQ := QhΦi
is Q-spanned by α1 , · · · , αl . Therefore

dimQ (EQ ) = dimk (h∗ ) .

Claim: h·, ·i is positive-definite and rational-valued on EQ .

14
 Pf: For λ, µ ∈ h∗ we have

hλ, µi = κ(tλ , tµ )
= Tr(adg (tλ )adg (tµ ))
X
= α(tλ )α(tµ )
α∈Φ
X
= hα, λihα, µi
α∈Φ

so for β ∈ Φ we have X
hβ, βi = hα, βi2 .
α∈Φ

We saw earlier that hβ, βi = κ(tβ , tβ ) is nonzero. Multiplying by 1/hβ,βi2 ,

1 X hα, βi2
=
hβ, βi hβ, βi2
α∈Φ

Comparing the right hand side to the form of the Cartan integers, we see that 1/hβ,βi is a sum of
nonnegative rational numbers. Therefore hβ, βi is a positive rational. This proves positive-definiteness
(we already had nondegeneracy). For α, β ∈ Φ we have 2hα,βi/hβ,βi ∈ Z, so since hβ, βi ∈ Q we get
hα, βi ∈ Q as well.

Now EQ is an “inner product” space over Q. We turn it into an actual real inner product space by defining
E to be the real vector space
E := R ⊗Q EQ
with an inner product extending h·, ·i from EQ . Note that

dimR E = dimQ EQ = dimk h∗ = dimk h .

Let us restate some of the things we’ve shown throughout this section, but this time in terms of the real
inner product space E and the finite subset1 Φ:

• Φ spans E and 0 ∈
/ Φ.
• If α ∈ Φ, then so is −α and no other scalar multiples of α make it into Φ.
• If α, β ∈ Φ, then so is
tα 2hα, βi
β − β( )α = β − .
κ(tα , tα ) hα, αi
2hα,βi
• If α, β ∈ Φ, then hα,αi ∈ Z ⊆ R.

Such an object (E, Φ) is called a root system. We have forged a path g (E, Φ).

. . . Does it depend on the choice of maximal toral subalgebra h?

1 Let us refer to the image of Φ under EQ ,→ E as Φ also.

15
6 Major Theorems

6.1 Engel’s Theorem

Theorem 8: Let g be a finite-dimensional lie algebra over F . If every element of g is ad-nilpotent then g
is nilpotent.

Proof: The case g = 0 is trivial, so assume g 6= 0.

Claim: If L ⊆ gl(V ) is a linear lie algebra over F with V 6= 0 finite dimensional and if every element
of L is nilpotent as an endomorphism of V , then there is a nonzero x ∈ V so that `(x) = 0 for all ` ∈ L.
Pf: Use induction on dim(L). The cases dim(L) = 0 (or even 1) are obvious. Assume the claim for
dimensions lower than dim(L).

Claim: A proper subalgebra of L cannot be self-normalizing.

Pf: Suppose K ( L is a subalgebra. Let gl(L; K) denote the subalgebra of gl(L) consisting
of endomorphisms that stabilize K. There is a natural map α : gl(L; K) → gl(L/K ) and it
is a lie algebra homomorphism. ad : L → gl(L) is a lie algebra homomorphism, and since K
is a subalgebra of L we have ad(K) ⊆ gl(L; K). So we obtain a lie algebra homomorphism
α ◦ (ad |K ) : K → gl(L/K ); let K̄ denote its range. Recall that a nilpotent endomorphism is
always ad-nilpotent, so K satisfies the hypotheses of the claim and dim(K̄) ≤ dim(K) < dim(L).
By induction then we obtain a nonzero vector in L/K which is killed by K̄. In other words, we
get an x ∈ L with x ∈/ K so that [k, x] ∈ K for all k ∈ K. Thus, NL (K) properly contains K.

L has a maximal proper subalgebra, say K. From the claim we have NL (K) = L, so K is an ideal. K
must have codimension 1, for otherwise we could a take a proper nontrivial subalgebra of L/K and pull
it back to L so as to contradict the maximality of K! (L/K is a lie algebra, and abelian one dimensional
subalgebras are available). Since dim(K) < dim(L), the induction shows that

W := {x ∈ V k(x) = 0 ∀k ∈ K}
is nonempty.

Claim: W is stabilized by every ` ∈ L.

Pf: Suppose x ∈ W and ` ∈ L. For any k ∈ K we have
k(l(x)) = l(k(x)) + [k, l](x) = 0
because K is an ideal. So `(x) ∈ W .

Now choose some z ∈ L \ K. Since z stabilizes W , we have a nilpotent endomorphism z |W of W . Get

an eigenvector x ∈ W \ {0} of z |W with eigenvalue 0. Since K has codimension 1, L = K + F hzi. It
follows that everything in L kills x.

Claim: Z(g) 6= 0.
Pf: By hypothesis, the range of ad : g → gl(g) is a linear lie algebra of nilpotent endomorphisms, and
we assumed that g 6= 0. So from the claim we obtain a nonzero x ∈ g which is killed by ran(ad). So
x ∈ Z(g) \ {0}.

We prove Engel’s theorem by induction on dim(g). The base case is automatic; asssume Engel’s theorem for

16
lie algebras of dimension less than dim(g). Well g/Z(g) has smaller dimension than g and all its elements are
ad-nilpotent, so g/Z(g) is nilpotent. It follows that g is nilpotent.

6.2 Lie’s Theorem

Theorem 9: Suppose F is an algebraically closed field of characteristic 0, V is a nontrivial finite dimensional

vector space over F , and L ⊆ gl(V ) is a lie subalgebra. If L is solvable, then L stabilizes some flag in V .

Proof: The case dim V = 1 is clear (there is only one flag and it works). The following claim will drive an
induction on dim V .

Claim: Under the hypotheses of the theorem if L ⊆ gl(V ) is solvable, then there is in V an eigenvector
common to all ` ∈ L.
Pf: The case dim L = 0 is trivial. Assume that dim L > 0 and assume that the claim holds for lower
dimensions. Our approach will be to find a codimension 1 ideal to which we can apply the induction
hypothesis, and then to show that L stabilizes the resulting space of eigenvectors common to elements
of the ideal.

Claim: L has an ideal of codimension 1.

Pf: Since L is solvable and dim L > 0, L(1) is a proper ideal of L. Since L/L(0) is abelian, all its
subspaces are ideals. Pull back any codimension 1 subspace to get a codimension 1 ideal of L.

Get a codimension 1 ideal K ⊆ L. K is solvable, so by the induction we obtain an eigenvector v ∈ V

common to all k ∈ K. Define λ : K → F to be the eigenvalue function; it is given by k(v) = λ(k)v for
k ∈ K. Observe that λ is linear. Define

W = {w ∈ V k(w) = λ(k)w ∀ k ∈ K} .

We already know W 6= 0 because v ∈ W .

Claim: Every ` ∈ L stabilizes W .

Pf: If w ∈ W and ` ∈ L, then for k ∈ K we have

k(`(w)) = `(k(w)) − [`, k](w) = λ(k)`(w) − λ([`, k])w

(since K is an ideal), so we need only prove the following claim:

Claim: For ` ∈ L and k ∈ K, λ([`, k]) = 0.

Pf: We will get a handle on the behavior of λ on K by deriving a formula that relates it
to a trace. Let n be minimal so that {v, `v, · · · , `n v} is linear dependent. For 0 ≤ i ≤ n,
let Wi = F hv, · · · , `i−1 vi. We’ve arranged things so that ` leaves Wn invariant, but it will
take some work to understand how elements of k operate on Wn .

Claim: For k 0 ∈ K and 0 ≤ i < n, k 0 (`i v) ≡ λ(k 0 )`i v modulo Wi .

Pf: The case i = 0 is obvious. Suppose that 0 < i < n and assume the claim below

17
 i. Then:

k 0 (`i v) = `(k 0 (`i−1 v)) − [`, k 0 ](`i−1 v)

≡ `(λ(k 0 )`i−1 v) − λ([`, k 0 ])`i−1 v (mod Wi−1 )
0 i
≡ λ(k )` v (mod Wi )

where we’ve applied the induction hypothesis to both terms and used the fact that
K is an ideal.

It follows that each k 0 ∈ K leaves Wn invariant and, when restricted to an endomorphism

Wn → Wn , has a matrix representation with λ(k 0 ) along the diagonal, with respect to the
basis {v, `v, · · · , `n−1 v}. So for k 0 ∈ K, we get tr(k 0 |Wn ) = nλ(k 0 ). Applying this to
[`, k] ∈ K we get

nλ([`, k]) = tr([`, k] |Wn )

= tr((`k − k`) |Wn )
= tr(` |Wn k |Wn −k |Wn ` |Wn ) = 0

where we’ve used the fact that ` also restricts to an endomorphism of Wn . Since charF = 0,
we conclude that λ([`, k]) = 0.

Choose an ` ∈ L \ K. Now ` stabilizes W and F is algebraically closed, so the endomorphism ` |W :

W → W has an eigenvector w ∈ W . We have that L = K + F h`i, that w (being in W ) is a common
eigenvector for the elements of K, and that w is an eigenvector of `. Therefore w is a common
eigenvector for the elements of L.

Assume that n = dim V > 1 and that Lie’s theorem holds for lower dimensions. Get a common eigenvector
w ∈ V for the elements of L. Since the elements of L leave F hwi invariant, L ⊆ gl(V ; F hwi) (the lie algebra
of endomorphisms of V that stabilize F hwi). Let α : gl(V ; F hwi) → gl(V /F hwi) be the natural map (a lie
algebra homomorphism). Apply the induction hypothesis to α(L) ⊆ gl(V /F hwi) to obtain an α(L)-invariant
flag
0 = V̄0 ⊆ V̄1 ⊆ · · · ⊆ V̄n−1 = V /F hwi .
Pull this back to V to obtain an L-invariant flag

0 = V0 , F hwi = V1 , V2 , · · · , Vn = V .

18
6.3 Cartan’s Solvability Criterion

Theorem 10: Let L ⊆ gl(V ) be a lie subalgebra, with V a finite dimensional vector space over F , an
algebraically closed field of characteristic 0. Suppose that
(∀k, ` (k ∈ [L, L] ` ∈ L) ⇒ tr(k`) = 0).


Then L is solvable.

Proof: Let M = {x ∈ gl(V ) [x, L] ⊆ [L, L]}.

Claim: For any k ∈ [L, L], x ∈ M we have tr(kx) = 0.

Pf: It suffices to check a generator [`, `0 ] ∈ [L, L] (where `, `0 ∈ L). We have for x ∈ M :
tr([`, `0 ]x) = tr(`[`0 , x]) = tr([`0 , x]`).
Since x ∈ M , [`0 , x] ∈ [L, L] (this may look obvious, but keep in mind that the endomorphism x need
not lie in L). So by the main hypothesis tr([`0 , x]`) = 0.

The general situation of the claim above will allow us to conclude that every k ∈ [L, L] is nilpotent, and
hence ad-nilpotent. It will follow from Engel’s theorem that [L, L] is nilpotent, and thus solvable, and thus it
will follow that L is solvable. It remains only to show that the elements of [L, L] are nilpotent. The following
claim should accomplish this and complete the proof.

Claim: Suppose that A ⊆ B ⊆ V are subspaces of V and define

M = {x ∈ gl(V ) [x, B] ⊆ A} = {x ∈ gl(V ) adgl(V ) (x)(B) ⊆ A}.
If k ∈ M satisfies tr(kx) = 0 for all x ∈ M , then k is nilpotent.
Pf: Suppose that k ∈ M and tr(kx) = 0 for x ∈ M . Since F is algebraically closed we obtain a
Jordan-Chevally decomposition k = s + n, where s is semisimple and n is nilpotent. Further, we know
that s is a polynomial over F evaluated at k. We also know that ad(s) is the semisimple part of ad(k)
and is therefore also a polynomial over F evaluated at ad(k). Let {v1 , · · · , vn } be a basis for V in
which s is diagonal, and let a1 , · · · , an ∈ F be those diagonal entries. The goal is to show that s = 0.
Since char F = 0, the prime field of F is Q and F is a vector space over Q. Let E be the subspace of
F Q-spanned by a1 , · · · , an . The following claim then shows that s = 0.

Claim: E ∗ = 0.
Pf: Suppose that f ∈ E ∗ (so f : E → Q is a Q-linear map). Let s0 ∈ gl(V ) be the endomorphism
with matrix diag(f (a1 ), · · · , f (an )) in the {v1 , · · · , vn } basis. Let {eij }i,j be the basis of gl(V )
corresponding to {v1 , · · · , vn }. In this basis ad(s) has the matrix diag({ai −aj }i,j ) and ad(s0 ) has
the matrix diag({f (ai ) − f (aj )}i,j ). By Lagrange interpolation there is some polynomial over F
whose value at (ai − aj ) is (f (ai ) − f (aj )) for all i, j (the linearity of f allows us to avoid conflict).
Evaluating this polynomial at ad(s) yields ad(s0 ). Since ad(s0 ) is a polynomial evaluated at ad(s),
which in turn is a polynomial evaluated at ad(k), we have that ad(s0 ) is a polynomial evaluated
at ad(k). We assumed that k ∈ M , so ad(k)(B) ⊆ A. It follows that ad(s0 )(B) ⊆ A and so
s0 ∈ M . Then by our hypothesis concerning k we have
n
X
0 = tr(ks0 ) = f (ai )ai .
i=1
Pn
Applying f we get i=1 f (ai )2 = 0. It follows from this that each f (ai ) vanishes, and so f
vanishes, and so E ∗ = 0.

19
 

6.4 Weyl’s Theorem

Theorem 11: Suppose that g is a finite dimensional Lie algebra over an algebraically closed field k of
characteristic zero. If g is semisimple, then every finite dimensional representation V of g is completely
reducible.

Proof: Let φ : g → gl(V ) be a representation. We will prove complete reducibility by showing that any
proper nontrivial g-submodule of V is a direct summand (i.e. has a complementary submodule in V ). First
we tackle the case of irreducible submodules of codimension one, then arbitrary codimension one submodules,
then arbitrary submodules. Assume g 6= 0 to avoid the trivial case.

Claim: Suppose that W ⊆ V is an irreducible submodule of codimension one. It is a direct summand.

Pf: We may safely assume that φ is a faithful representation: If φ were not faithful then ker φ, being
an ideal of a semisimple lie algebra, is a direct sum of some of the simple ideals of g. So it has a
complement in g: g = ker φ ⊕ g0 . Then φ|g0 : g0 :→ gl(V ) is a faithful representation of g0 , and it is
easy to check that g-submodules of V correspond exactly to g0 -submodules of V . So just assume that
φ is faithful.

Let cφ : V → V be the Casimir element of φ. Recall that cφ is a g-module homomorphism, so ker cφ is

a submodule. Will it do for a complement to W ?

Notice that g acts trivially on the quotient g-module V /W . This follows from the facts that g is
semisimple and that dim(V /W ) = 1: If φ̄ : g → gl(V /W ) is the induced representation then

φ̄(g) = φ̄([g, g]) = [φ̄(g), φ̄(g)] ⊆ [gl(V /W ), gl(V /W )] = sl(V /W ) = 0 .

In other words, any x ∈ g sends V into W . So cφ , being a sum of products of things in im φ, also
sends V into W . By Schur’s lemma, cφ |W : W → W is a scalar multiple of 1W . If that scalar were 0,
then we would have c2φ = 0 (because cφ (V ) ⊆ W and cφ (W ) = 0). This would give tr(cφ ) = 0, which
contradicts the calculation that tr(cφ ) = dim g. So cφ |W : W → W is an isomorphism.

It is now apparent that V = W ⊕ ker cφ : We have W ∩ ker cφ = ker(cφ |W ) = 0. And ker cφ is one
dimensional because im cφ = W has codimension one.

Claim: Suppose that W ⊆ V is a submodule of codimension one. It is a direct summand.

Pf: Use induction on dim V ; the base case dim V = 1 is trivial. Assume that codimension 1
submodules of g-modules of lower dimension are direct summands. The previous claim covers the
case in which W is irreducible, so suppose that W contains a proper nontrivial submodule W 0 .

Now W/W 0 is a codimension one g-submodule of the quotient g-module V /W 0 , which has stricly lower
dimension than dim V . We obtain a complementary submodule W f
/W 0 so that
V /W 0 = W/W 0 ⊕ W
f
/W 0 .

Since dim(W
f
/W 0 ) = 1, W 0 is a codimension one submodule of W
f , which has stricly lower dimension
than dim V . We obtain a complementary submodule X ⊆ W to W 0 .
f

20
 X should do the trick. We get W ∩ X = 0: if x ∈ W ∩ X then x + W 0 ∈ W/W 0 ∩ W f
/W 0 = 0, so
0
x ∈ W ∩ X = 0. And X is one dimensional so dim X + dim W = dim V . Thus we get V = W ⊕ X.

Claim: Suppose that W ( V is a proper submodule. It is a direct summand.

Pf: Since V and W are g-modules, so is Homk (V, W ). Consider the subspace

V := {f ∈ Homk (V, W ) f |W = λ1W for some λ ∈ k}

and the smaller subspace

W := {f ∈ V f |W = 0}.
Observe that g sends V into W: For f ∈ V (say f |W = λ1W ), x ∈ g, and w ∈ W we have

(x.f )(w) = x.(f (w)) − f (x.w) = λ x.w − λ x.w = 0,

so (x.f )|W = 0. This shows that V, W are submodules of Homk (V, W ).

Now W has codimension one in V. One way to see this is to define a linear surjection F : V → k that
takes f ∈ V with f |W = λ1V to λ, and note that its kernel is W. Thus the previous claim applies and
we obtain a one dimensional complementary submodule to W in V; let f ∈ V span it. By rescaling f
we may ensure that f |W = 1W .

To say that g kills an element of Homk (V, W ) is precisely to say that that element is a g-module
homomorphism. Well g does kill f , since one dimensional representations of semisimple lie algebras
are always trivial (see argument in first claim). So f is a g-module homomorphism, and its kernel is
then the submodule we need! We have W ∩ ker f = ker(f |W ) = 0. And we have W + ker f = V because
any v ∈ V can be written as f (v) + (v − f (v)). So V = W ⊕ ker f .

7 The Jordan-Chevalley Decomposition

[still need to fill in proofs in this section]

Let V be a finite-dimensional vector space over an algebraically closed field k of characteristic 0, and consider
an x ∈ End(V ). Then:
1. There are unique xs , xn ∈ End(V ) such that
• xs is semisimple
• xn is nilpotent
• xs , xn commute
• x = xs + xn
2. There are p, q ∈ k[T ] with a constant term of 0 such that

• xs = p(x)
• xn = q(x)

21
An abstract version of this decomposition can take place inside a semisimple Lie algebra. If g is a semisimple
Lie algebra over k, then every x ∈ g can be written uniquely as x = s + n with s being ad-semisimple, with
n ad-nilpotent, and with [s, n] = 0.

In the case that L ⊆ gl(V ) is a linear semisimple Lie algebra, we have two versions of the decomposition for
x ∈ L (the abstract decomposition and the decomposition of x as an endomorphism). The two notions do
end up agreeing, but the proof of this fact is not simple.

A very useful consequence of this is that representations of a semisimple Lie algebra are compatible with
the abstract decomposition: If g is a finite-dimensional semisimple Lie algebra over k and φ : g → gl(V ) is a
finite-dimensional representation, then for x ∈ g

the abstract Jordan-Chevalley decomposition x=s+n in g

gives the usual Jordan-Chevalley decomposition φ(x) = φ(s) + φ(n) in gl(V ).

In other words what is true by definition for the representation ad : g → gl(g) turns out to be true for more
general representations.

References
[1] Humphreys, J.E. (1972) Introduction to Lie Algebras and Representation Theory. Springer Verlag, New
York.

22