B2.

3 Lie Algebras
Lecture notes

Kevin McGerty1

Hilary 2025

1
Please contact me at [email protected] if you find any errors or ambiguities in these notes. Last up-
dated: 17 February 2025.
Contents

Index of Notation iii

Background iv

1 Lie algebras: Definition and basic notions 1

1.1 Definitions and Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Ideals and isomorphism theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Representations of Lie algebras 7

2.1 Definition and examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Subrepresentations, quotients, duals, and composition series . . . . . . . . . . . . . . . 9
2.3 Tensor products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3 Classifying Lie algebras 16

3.1 Classification by composition factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2 Exact sequences of Lie algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

4 Gluing 𝔤𝔩1 : Solvable and nilpotent Lie algebras 21

4.1 Solvable Lie algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.2 Nilpotent Lie algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.3 Representations of solvable Lie algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.4 Representations of nilpotent Lie algebras . . . . . . . . . . . . . . . . . . . . . . . . . . 29

5 Cartan subalgebras and the Cartan criteria. 32

5.1 Nilpotent Lie algebras as measurements . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.2 Trace forms and Cartan’s criterion for Solvability . . . . . . . . . . . . . . . . . . . . . . 34
5.3 Semisimple Lie algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.4 The representation theory of 𝔰𝔩2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

6 The structure of semisimple Lie algebras 45

6.1 The Killing form and the Cartan decomposition . . . . . . . . . . . . . . . . . . . . . . 45
6.2 Root systems and the classification of semisimple Lie algebras . . . . . . . . . . . . . . . 48
6.3 Abstract root systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

i
 6.4 Cartan matrices and isomorphisms of root systems . . . . . . . . . . . . . . . . . . . . 54

Appendices 1
I Primary Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
II Tensor Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Note: All sections, proofs, and individual Remarks which are marked with an asterisk (*) are non-
examinable.

ii
Index of Notation

(𝐷𝑘 (𝔤))𝑘≥0 the derived series of a Lie algebra 𝔤.

ℕ.𝑋 for any subset 𝑋 of an abelian group 𝐴 (e.g. a vector space) this denotes the set of all sums
of the form ∑𝑦∈𝑌 𝑛𝑦 .𝑦 where 𝑌 is a finite subset of 𝑋 and 𝑛𝑦 ∈ ℕ.

𝔤𝔩𝑉 the Lie algebra of all endomorphisms of a vector space 𝑉 equipped with the commutator
bracket.

𝐼E𝔤 𝐼 is an ideal in 𝔤, that is, 𝐼 is a linear subspace and for all 𝑥 ∈ 𝔤 and 𝑎 ∈ 𝐼, we have [𝑎, 𝑥] ∈ 𝐼

𝜅 = 𝜅𝔤 the Killing form, an invariant symmetric bilinear form on a Lie algebra 𝔤 given by:

𝜅(𝑥, 𝑦) = tr𝔤 (ad(𝑥)ad(𝑦)).

(𝐶𝑘 (𝔤))𝑘≥0 the lower central series of a Lie algebra 𝔤.

[., .] an alternating bilinear map satisfying the Jacobi identity, known as a Lie bracket.

Mat𝑛,𝑚 (k) the space of 𝑛 × 𝑚 matrices with entries in a field k.

Mat𝑛 (k) the space of 𝑛 × 𝑛 matrices with entries in a field k.

ℕ the natural numbers.

𝑁𝔤 (𝔞) the normalizer of a subalgebra 𝔞 in a Lie algebra 𝔤.

rad(𝐵) the radical of a symmetric bilinear form 𝐵, consisting of all 𝑣 ∈ 𝑉 which satisfy 𝐵(𝑣, 𝑤) = 0
for all 𝑤 ∈ 𝑉.

rad(𝔤) the maximal solvable ideal in a Lie algebra 𝔤.

(𝑉, Φ) an abstract root system.

𝔰𝔩(𝑉) the Lie subalgebra {𝛼 ∈ 𝔤𝔩𝑉 ∶ tr(𝛼) = 0} of traceless endomorphisms of 𝑉, known as the
special linear Lie algebra associated to 𝑉.

⟨𝑋⟩𝐹 if 𝑋 is a subset of a k-vector spaces 𝑉 and 𝐹 ≤ k is a subfield of k, then the 𝐹-linear span of
𝑋 in 𝑉, i.e. the intersection of all 𝐹-subspaces of 𝑉 containing 𝑋 is denoted by either ⟨𝑋⟩𝐹
or span𝐹 (𝑋). If the subscript 𝐹 is omitted the field is understood to be k.

iii
*Background

From group actions to group representations

In mathematics, group actions give a way of encoding the symmetries of a space or physical system. For-
mally these are defined as follows: an action of a group 𝐺 on a space1 𝑋 is a map 𝑎 ∶ 𝐺 × 𝑋 → 𝑋, written
(𝑔.𝑥) ↦ 𝑎(𝑔, 𝑥) or more commonly (𝑔, 𝑥) ↦ 𝑔.𝑥 which satisfies the properties

1. 𝑒.𝑥 = 𝑥, for all 𝑥 ∈ 𝑋, where 𝑒 ∈ 𝐺 is the identity;

2. (𝑔1 𝑔2 ).𝑥 = 𝑔1 .(𝑔2 .𝑥) for all 𝑔1 , 𝑔2 ∈ 𝐺 and 𝑥 ∈ 𝑋.

Natural examples of group actions are that of the general linear group GL𝑛 (ℝ) on ℝ𝑛 , or the action of the
group of rigid motions SO3 on 𝑆2 , the unit sphere {𝑥 ∈ ℝ3 ∶ ||𝑥|| = 1} in ℝ3 .
Whenever a group acts on a space 𝑋, there is a resulting linear action (a representation) on the vector
space of functions on 𝑋. Indeed if Fun(𝑋) denotes the vector space of real-valued functions on 𝑋, then
the formula
𝑔(𝑓)(𝑥) = 𝑓(𝑔−1 .𝑥), ∀ 𝑔 ∈ 𝐺, 𝑓 ∈ Fun(𝑋), 𝑥 ∈ 𝑋,

defines a representation of 𝐺 on Fun(𝑋). If 𝑋 and 𝐺 have more structure. e.g. that of a topological space
or smooth manifold, then this action may also preserve the subspaces of say continuous, or differentiable
functions.

Remark. If one defines the action of 𝐺 on Fun(𝑋) by the more natural-looking formula 𝑔(𝑓)(𝑥) = 𝑓(𝑔.𝑥)
then one finds that

𝑔1 (𝑔2 (𝑓))(𝑥) = 𝑔2 (𝑓)(𝑔1 .𝑥) = 𝑓((𝑔2 𝑔1 ).𝑥) = ((𝑔2 𝑔1 ).𝑓)(𝑥), ∀ 𝑥 ∈ 𝑋.

In other words, the resulting action on Fun(𝑋) is a right-action. Since inversion 𝜄(𝑔) = 𝑔−1 gives an
isomorphism of groups 𝜄 ∶ 𝐺 → 𝐺op (where 𝐺op is the group obtained by using the binary operation
𝑔1 ⋆ 𝑔2 = 𝑔2 𝑔1 ), by precomposing with 𝜄 we may turn a right action into a left action, and in this course
our convention will be to do this where necessary to ensure all actions are left action unless the contrary
is explicitly stated.

Infinitesimal symmetries

In this section I use some material, like multivariable analysis, which is not necessary for the main body of the
course, but if you know it, or are happy to rely on notions from Prelims multivariable calculus for which you have
not been given a rigorous definition, it will help to put the material of this course in a broader context. For those
worried about such things, fear not, it is non-examinable.
1
I’m being deliberately vague here about what a “space” is, 𝑋 could just be a set, but it could also have a more geometric
nature, such as a topological space or submanifold of ℝ𝑛 .

iv
 Lie algebras arise as the “infinitesimal version” of group actions, which loosely speaking means they
are what we get by trying to differentiate group actions. When working in ℝ𝑛 we will denote the partial
derivative of 𝑓 in the direction of the 𝑖-th standard basis vector by 𝜕𝑖 𝑓 (in preference to the notation 𝜕𝑓/𝜕𝑥𝑖
you may have seen more often).
In order for a group 𝐺 to have a Lie algebra associated to it, it must have additional structure – for
example when a group is also a metric space in a natural way, and the group operations are continuous
functions, then is an example of a topological group. To possess a Lie algebra, it must make sense to talk
about differentiable functions on 𝐺, not just continuous functions. In general this requires developing the
theory of manifolds, but some important examples already arose already in the Prelims Geometry course.
The orthogonal group O𝑛 of linear isometries of an 𝑛-dimensional real inner product space, is naturally
identified with the subset {𝐴 ∈ Mat𝑛 (ℝ) ∶ 𝐴.𝐴⊺ = 𝐼𝑛 } of 𝑛 × 𝑛 matrices whose columns form an
orthonormal basis of ℝ𝑛 . As such, a smooth function on O𝑛 is just the restriction of a smooth function on
Mat𝑛 (ℝ𝑛 ) (an 𝑛2 -dimensional real vector space) isometries of Euclidean space. Equally, a smooth curve
on O𝑛 is just a smooth curve in Mat𝑛 (ℝ) whose image is contained in O𝑛 .
The action of O𝑛 ℝ𝑛 is linear and hence certainly smooth. To obtain a “derivative” of the group action,
we can consider smooth curves 𝛾 ∶ (−1, 1) → 𝐺 which pass through the identity 𝐼𝑛 at 𝑡 = 0, and then given
a function on ℝ𝑛 we can consider the action of the curve 𝛾 on a function 𝑓 ∶ ℝ𝑛 → ℝ 𝑓(𝑥) ↦ 𝑓(𝛾(𝑡)−1 .𝑥),
and, since our curve is smooth, it makes sense to take the derivative at 𝑡 = 0, that is, we set
𝑑
𝜈𝛾 (𝑓) = (𝑓(𝛾(𝑡)−1 .𝑥)|𝑡=0
𝑑𝑡
In this way, we obtain a family of operators on the space of functions on ℝ𝑛 . While it appears that the
family is indexed by smooth curves through 𝐼𝑛 ∈ O𝑛 , since we take the derivative at 𝑡 = 0, the operator
𝜈𝛾 only depends on the derivative of 𝛾 at 𝑡 = 0, that is, the velocity of the curve as it passes through the
identity element in O𝑛 .
Note that we only consider curves on O𝑛 as they pass through the identity element, rather than at
all points on the group. This is because if measure the infinitesimal action on ℝ𝑛 near a group element
𝐴 ∈ O𝑛 using a path 𝐴(𝑡) with 𝐴(0) = 𝐴, then setting 𝛾(𝑡) = 𝐴−1 𝐴(𝑡) we have, for 𝑥 ∈ ℝ𝑛 , 𝐴(𝑡)−1 .𝑥 =
𝛾(𝑡).𝐴−1 .𝑥, and thus we will simply get the infinitesimal action of the curve 𝛾 at the point 𝑦 = 𝐴−1 𝑥 ∈ ℝ𝑛 .
Knowing the infinitesimal action by curves through the identity at all points in ℝ𝑛 thus already captures
that given by the velocity of curves as they pass through other elements of the group.
As det(𝐴) = det(𝐴⊺ ), if 𝐴 ∈ O𝑛 then det(𝐴)2 = 1 and hence det ∶ O𝑛 → {±1} is a group homomor-
phism to which is easily seen to be surjective. Its kernel SO𝑛 , the group of orientation preserving linear
isometries of ℝ𝑛 , is thus an index 2 subgroup which is the connected component containing the identity
element. The infinitesimal action associated to an action of O𝑛 will thus only depend on the action of SO𝑛 .
Example. When 𝑛 = 2 the group SO2 is just the group of rotations about the origin.
cos(𝑡) −sin(𝑡)
The map 𝑡 ↦ 𝑔(𝑡) = �
sin(𝑡) cos(𝑡) �
is smooth (because each of its components is) and a group
homomorphism, that is, 𝑔(𝑡1 + 𝑡2 ) = 𝑔(𝑡1 )𝑔(𝑡2 ). Indeed since 𝑔(𝑡 + 2𝜋) = 𝑔(𝑡) the parametrization given
by 𝑔 gives us another description of the smooth functions on SO2 – they are simply the smooth functions
on ℝ which are periodic with period 2𝜋.
The action of SO2 on ℝ2 induces an infinitesimal action on the smooth functions on ℝ2 . Since SO2 is
1-dimensional, we need only consider the curve through 𝐼2 given by 𝑔(𝑡). Since 𝑔(𝑡) is a group homomor-
phism, the operator 𝜈 = 𝜈𝑔 defined above becomes in this case
𝑑 𝑑
𝜈(𝑓)(𝑥) = �𝑔(𝑡)(𝑓)(𝑥)� = �𝑓(𝑔(−𝑡).𝑥)�
𝑑𝑡 |𝑡=0 𝑑𝑡 |𝑡=0

Now
cos(𝑡) sin(𝑡) 𝑥1 𝑥1 cos(𝑡) + 𝑥2 sin(𝑡)
𝑔(−𝑡).𝑥 = � =
−sin(𝑡) cos(𝑡) � � 𝑥2 � � −𝑥1 sin(𝑡) + 𝑥2 cos(𝑡) �

v
hence
𝑑 −𝑥1 sin(𝑡) + 𝑥2 cos(𝑡) 𝑑 𝑥2
(𝑔(−𝑡).𝑥) = � � and (𝑔(−𝑡).𝑥)|𝑡=0 = � −𝑥 �
𝑑𝑡 −𝑥 1 cos(𝑡) + −𝑥 2 sin(𝑡) 𝑑𝑡 1

Thus by the chain rule, since 𝑔(0).𝑥 = 𝑥 we have

𝑥2 𝑥
𝜈(𝑓) = 𝑑𝑓𝑥 � � = � 𝜕1 𝑓 𝜕 2 𝑓 � � 2 � = 𝑥 2 𝜕1 𝑓 − 𝑥 1 𝜕2 𝑓
−𝑥1 −𝑥1

Remark 0.0.1. If we instead write 𝑔(𝑡) = exp(𝑖𝑡) and view ℝ2 as ℂ with 𝑧 = 𝑥1 + 𝑖𝑥2 , then we may also
compute 𝜈 as follows:
𝑑
𝜈(𝑓) = (𝑓(exp(−𝑖𝑡)𝑧))|𝑡=0 = 𝑑𝑓𝑥 (−𝑖𝑧) = 𝑑𝑓𝑥 (𝑥2 − 𝑖𝑥1 ) = 𝑥2 𝜕1 𝑓(𝑥) − 𝑥1 𝜕2 𝑓(𝑥)
𝑑𝑡

The operator we obtained in this example, 𝜈 = 𝑥2 𝜕1 − 𝑥1 𝜕2 is a 𝒞 ∞ (ℝ2 )-linear combination of 𝜕𝑥

and 𝜕𝑦 . Operators of this form encode “infinitesimal symmetries”. The next definition formalises its key
properties.

Definition. For any positive integer 𝑛, an ℝ-linear operator 𝜈 ∶ 𝒞 ∞ (ℝ𝑛 ) → 𝒞 ∞ (ℝ𝑛 ) is said to be a
derivation if, for any 𝑓1 , 𝑓2 ∈ 𝒞 ∞ (ℝ𝑛 ) it satisfies

𝜈(𝑓1 .𝑓2 ) = 𝜈(𝑓1 ).𝑓2 + 𝑓1 .𝜈(𝑓2 ). (†)

The next Lemma (which follows readily from a version of Taylor’s theorem for functions on ℝ𝑛 for
example) shows that the previous, somewhat formal, definition, actually results in a class of objects with
a very concrete description.

Lemma. If 𝜈 ∶ 𝒞 ∞ (ℝ𝑛 ) → 𝒞 ∞ (ℝ𝑛 ) is a derivation, then there exist unique smooth functions 𝑎1 , … , 𝑎𝑛 ∈
𝑛
𝒞 ∞ (ℝ𝑛 ) such that 𝜈 = ∑𝑗=1 𝑎𝑗 𝜕𝑗 , that is, for all 𝑓 ∈ 𝒞 ∞ (ℝ𝑛 ) we have
𝑛
𝜈(𝑓) = � 𝑎𝑗 𝜕𝑗 (𝑓).
𝑗=1

Proof. (Sketch.) Let 𝜈 be a derivation. Since 12 = 1 ∈ ℝ, we must have 𝜈(1) = 𝜈(1).1 + 1.𝜈(1), that is,
2𝜈(1) = 𝜈(1) so that 𝜈(1) = 0, and hence by linearity 𝜈(𝜆) = 0 for all 𝜆 ∈ ℝ, i.e. any derivation vanishes
on the subspace of constant functions ℝ ⊆ 𝒞 ∞ (ℝ𝑛 ).
The characterisation of derivations then follows from the fact that for any 𝑐 ∈ ℝ𝑛 , the maximal ideal
𝔪𝑐 ∶= {𝑓 ∈ 𝒞 ∞ (ℝ𝑛 ) ∶ 𝑓(𝑐) = 0} of the ring 𝒞 ∞ (ℝ𝑛 ) is generated by {(𝑥𝑖 − 𝑐𝑖 ) ∶ 1 ≤ 𝑖 ≤ 𝑛} 

Thus to give a derivation is the same as to give an 𝑛-tuple of functions (𝑎1 , … , 𝑎𝑛 ), or in other words a
smooth function 𝑎 ∶ ℝ𝑛 → ℝ𝑛 .

Definition. A vector field on 𝑋 = ℝ𝑛 is a (smooth) function 𝜈 ∶ ℝ𝑛 → ℝ𝑛 . Let Θ ℝ𝑛 denote the vector space
𝑛
of all vector fields on ℝ𝑛 . To any vector field 𝜂 = (𝑎𝑖 )𝑛𝑖=1 one can associate the derivation 𝜃𝜂 = ∑𝑗=1 𝑎𝑗 𝜕𝑗
which one can think of as giving the infinitesimal direction of a flow (e.g. of a fluid, or an electric field say)
and the previous Lemma shows that this gives an equivalence between the space of vector fields and the
space of derivations of 𝒞 ∞ (ℝ𝑛 ).
Note that if we compose two derivations 𝜈1 ∘ 𝜈2 we again get an operator on functions, but it is not
given by a vector field, since the presence of second-order derivatives prevents it from satisfying (†). How-
ever, it is easy to check that if 𝜈1 , 𝜈2 are derivations, then the commutator [𝜈1 , 𝜈2 ] = 𝜈1 ∘ 𝜈2 − 𝜈2 ∘ 𝜈1
does obeys the identity (†) and hence is again a derivation. Thus the space of vector fields Θ ℝ𝑛 inher-
its a nonassociative product given by the commutator of the corresponding derivations. Explicitly, if

vi
𝜂1 = (𝑎𝑖 )𝑛𝑖=1 , 𝜂2 = (𝑏𝑖 )𝑛𝑖=1 ∈ Θ ℝ𝑛 and [𝜃(𝜂1 ), 𝜃(𝜂2 )] = ∑𝑖 (∑𝑗 (𝑎𝑗 𝜕𝑗 𝑏𝑖 − 𝑏𝑗 𝜕𝑗 𝑎𝑖 )𝜕𝑖 , so that [𝜂1 , 𝜂2 ] =
(∑ (𝑎𝑗 𝜕𝑗 𝑏𝑖 − 𝑏𝑗 𝜕𝑗 𝑎𝑖 ))𝑛𝑖=1 .
𝑗
The space of derivations equipped with the commutator bracket [., .] (or equivalently the space of
vector fields with the corresponding operation) is an (infinite dimensional) example of a Lie algebra. The
derivatives of a group action give subalgebras of the algebra Θ 𝑋 : the fact that the commutator product
preserves them is a sort of infinitesimal remnant of the group multiplication.2

Heuristically, we think of the infinitesimal version of a group action as the collection of derivations on
smooth functions obtained by “differentiating the group action at the identity element”. (For the circle
the collection of vector fields we get are just the scalar multiples of the vector field 𝜈, but for actions of
larger group this will yield a larger space of derivations).

Example. Consider the action of SO3 (ℝ) on ℝ3 . This is the group of orientation-preserving linear isome-
tries of ℝ3 . It is well-known that any element of 𝑔 ∈ SO3 (ℝ) is a rotation by some angle, say 𝜃, about an
axis 𝐿 through the origin. Then there is a smooth path 𝛾 in SO3 (ℝ) from the identity to 𝑔 which, for
𝑡 ∈ [0, 1] is the rotation by 𝑡.𝜃 about that axis.
Since 𝛾 is smooth and its definition extends to 𝑡 in an open interval containing 𝑡 = 0, it makes sense
𝑑
to associate to it the operator 𝜈𝛾 on 𝒞 ∞ (ℝ3 ) given, just as for the circle, by 𝑓 ↦ 𝑑𝑡 (𝑓(𝛾(−𝑡)(𝑥)), so that
𝜈𝛾 (𝑓) = 𝑑𝑓(−𝛾′ (0)𝑥). If we pick an orthonormal basis ℱ = (𝑣1 , 𝑣2 , 𝑣3 ) of ℝ3 with 𝑣1 spanning the
⎛ ⎞
⎜⎜ 1 0 0 ⎟⎟
⎜ ⎟
axis of the rotation, then 𝛾(𝑡) has matrix 𝐴 = ⎜⎜⎜ 0 cos(𝑡) −sin(𝑡) ⎟⎟⎟, and a calculation almost identical
⎝ ⎠
0 sin(𝑡) cos(𝑡)
to the case of the circle discussed above shows that, in the coordinates (𝑧1 , 𝑧2 , 𝑧3 ) given by ℱ, 𝜈𝛾 (𝑓) =
𝑧3 𝜕2 − 𝑧2 𝜕3 . Now 𝜕𝑖 is just the directional derivative 𝜕𝑣𝑖 and 𝑧𝑖 (𝑥) = 𝑣𝑖 ⋅ 𝑥, thus 𝜈𝛾 corresponds to the
vector field (𝑣3 ⋅ 𝑥)𝑣2 − (𝑣2 ⋅ 𝑥)𝑣3 , that is, −𝑣1 ∧ 𝑥, where ∧ denotes the vector product. Thus it follows that
the space of all vector fields obtained in this way is the 3-dimensional subspace of Θ ℝ3 which is spanned
by {𝑒1 ∧ 𝑥, 𝑒2 ∧ 𝑥, 𝑒3 ∧ 𝑥} or in terms of derivations it is the span of {𝑥3 𝜕2 − 𝑥2 𝜕3 , 𝑥1 𝜕3 − 𝑥3 𝜕1 , 𝑥2 𝜕1 − 𝑥1 𝜕2 }.

2
To be a bit more precise, it comes from the conjugation action of the group on itself.

vii
Chapter 1

Lie algebras: Definition and basic notions

1.1 Definitions and Examples

The definition of a Lie algebra is an abstraction of the example of the product on vector fields given. It is
purely algebraic, so it makes sense over any field k. We begin, however, with an even more basic definition:

Definition 1.1.1. Let 𝑅 be a commutative ring1 . An 𝑅-algebra is a pair (𝐴, ∗) consisting of an 𝑅-module 𝐴
and an 𝑅-bilinear map ∗ ∶ 𝐴 × 𝐴 → 𝐴, that is, for all 𝑎1 , 𝑎2 , 𝑏1 , 𝑏2 ∈ 𝐴 and 𝑟 ∈ 𝑅, the operation ∗ satisfies:

(𝑟.𝑎1 + 𝑎2 ) ∗ 𝑏1 = 𝑟.(𝑎1 ∗ 𝑏1 ) + (𝑎2 ∗ 𝑏1 ),

𝑎1 ∗ (𝑟.𝑏1 + 𝑏2 ) = 𝑟.(𝑎1 ∗ 𝑏1 ) + (𝑎1 ∗ 𝑏2 ).

We say that (𝐴, ∗) is unital (or has a unit) if there is an element 1𝐴 ∈ 𝐴 such that 1𝐴 ∗ 𝑎 = 𝑎 ∗ 1𝐴 = 𝑎
for all 𝑎 ∈ 𝐴. Note that if it exits, the multiplicative unit is unique. We say that (𝐴, ∗) is associative if
𝑎 ∗ (𝑏 ∗ 𝑐) = (𝑎 ∗ 𝑏) ∗ 𝑐 for all 𝑎, 𝑏, 𝑐 ∈ 𝐴. When 𝐴 is associative, we will normally suppress the operation ∗
and so, for any 𝑎, 𝑏 ∈ 𝐴, write 𝑎𝑏 rather than 𝑎 ∗ 𝑏 for the value of the bilinear map on the pair (𝑎, 𝑏).
Note that an associative ℤ-algebra (i.e. letting 𝑅 = ℤ the integers) is just a ring. In this course we
will usually assume that 𝑅 is a field, which we will denote by k.

Definition 1.1.2. A Lie algebra over a field k is a k-algebra (𝔤, [., .]𝔤 ) which satisfies the following axioms:

1. [., .]𝔤 is alternating, i.e. [𝑥, 𝑥]𝔤 = 0 for all 𝑥 ∈ 𝔤.

2. The Lie bracket satisfies the Jacobi Identity: that is, for all 𝑥, 𝑦, 𝑧 ∈ 𝔤 we have:

[𝑥, [𝑦, 𝑧]𝔤 ]𝔤 + [𝑦, [𝑧, 𝑥]𝔤 ]𝔤 + [𝑧, [𝑥, 𝑦]𝔤 ]𝔤 = 0.

Remark 1.1.3. 1. Note that by considering the bracket [𝑥+𝑦, 𝑥+𝑦]𝔤 it is easy to see that the alternating
condition implies that for all 𝑥, 𝑦 ∈ 𝐿 we have [𝑥, 𝑦]𝔤 = −[𝑦, 𝑥]𝔤 , that is [., .]𝔤 is skew-symmetric. If
char(k) ≠ 2, the alternating condition is equivalent to skew-symmetry.

2. If 𝔤 is a Lie algebra, then the alternating property implies that 𝔤 cannot have a unit. It is also al-
most never the case that an associative product will satisfy the conditions to be a Lie bracket. Thus,
viewed a Lie algebra is a non-commutative, non-unital and (usually) non-associative algebra.2

3. We will normally write [., .] for the Lie bracket on any Lie algebra and decorate it only for emphasis
or where there is the potential for confusion.
1
All commutative rings in this course will have a multiplicative identity.
2
This makes them sound awful. However, as we will see this is not the way to think about them!

1
Definition 1.1.4. Let (𝔤1 , [., .]1 ) and (𝔤2 , [., .]2 ) be Lie algebras. A k-linear map 𝜙 ∶ 𝔤1 → 𝔤2 is said to be a
homomorphism of Lie algebras if it respects the Lie brackets. That is:

𝜙([𝑎, 𝑏]1 ) = [𝜙(𝑎), 𝜙(𝑏)]2 ∀ 𝑎, 𝑏 ∈ 𝔤1 .

An isomorphism of Lie algebras is a bijective homomorphism, since, just as for group homomorphisms
and linear maps, the (set-theoretic) inverse of a Lie algebra homomorphism is automatically itself a Lie
algebra homomorphism.

Example 1.1.5. i) If dimk (𝔤) = 1, then the alternating condition forces the Lie bracket to van-
ish. Thus, up to isomorphism, there is a unique 1-dimensional Lie algebra over k, that is, any 1-
dimensional Lie algebra 𝔤 is isomorphic to k equipped with the zero Lie bracket.

ii) If 𝔞 is any vector space then setting the Lie bracket [., .] to be zero, i.e. setting [𝑎, 𝑏] = 0 for all
𝑎, 𝑏 ∈ 𝔞, we get a (not very interesting) Lie algebra. Such Lie algebras are called abelian Lie algebras.

iii) If 𝐴 is an (associative) k-algebra, then 𝐴 can be given the structure of a k-Lie algebra, where if
𝑎, 𝑏 ∈ 𝐴 then we set [𝑎, 𝑏] = 𝑎.𝑏 − 𝑏.𝑎, the commutator of 𝑎 and 𝑏. The commutator bracket is clearly
alternating, and checking the Jacobi identity is a fundamental calculation. Indeed we have

[𝑥, [𝑦, 𝑧]] = 𝑥(𝑦𝑧 − 𝑧𝑦) − (𝑦𝑧 − 𝑧𝑦)𝑥 = 𝑥𝑦𝑧 − 𝑥𝑧𝑦 − 𝑦𝑧𝑥 + 𝑧𝑦𝑥
= (𝑥𝑦𝑧 − 𝑦𝑧𝑥) + (𝑧𝑦𝑥 − 𝑦𝑧𝑥)

where,in the final expression, we have paired terms which can be obtained from each other by cy-
cling 𝑥, 𝑦 and 𝑧. Since the terms in these pairs have opposite signs, it is then clear that adding the
three expressions obtained by cycling 𝑥, 𝑦 and 𝑧 gives zero. We will write 𝔤𝐴 for the Lie algebra
(𝐴, [., .]) obtained from an associative algebra in this way.

iv) An important special case of the previous example is that where 𝐴 is the space of 𝑛-by-𝑛 matrices
Mat𝑛 (k) with entries in k. The Lie algebra one obtains by equipping this space with the commutator
Lie bracket
[𝑋, 𝑌] = 𝑋.𝑌 − 𝑌.𝑋, ∀ 𝑋, 𝑌 ∈ Mat𝑛 (k),
is denoted 𝔤𝔩𝑛 (k), or, if the field k is clear from context, simply 𝔤𝔩𝑛 . Slightly more abstractly, if 𝑉 is
a k-vector space, then we will write 𝔤𝔩𝑉 for the Lie algebra obtained from the associative algebra
Endk (𝑉) by equipping it with the commutator bracket.3

v) If dim(𝑉) = 1 then 𝔤𝔩𝑉 = 𝔤𝔩1 = k: the action of scalars gives an injective map k → End(𝑉) for any
nonzero vector space 𝑉 which is an isomorphism if dim(𝑉) = 1. We will therefore write 𝔤𝔩1 for k
viewed as a Lie algebra with zero Lie bracket.

vi) If 𝔤 is a Lie algebra and 𝔰 ≤ 𝔤 is a k-subspace of 𝔤 on which the restriction of the Lie bracket takes
values in 𝔰, so that it induces a bilinear operation [., .]𝔰 ∶ 𝔰 × 𝔰 → 𝔰, then (𝔰, [., .]𝔰 ) is clearly a Lie
algebra, and we say 𝔰 is a (Lie) subalgebra of 𝔤. If 𝔰 is a Lie subalgebra of 𝔤 then the inclusion map
𝑖 ∶ 𝔰 → 𝔤 is a homomorphism of Lie algebras.
𝑛
Let tr ∶ 𝔤𝔩𝑛 → k denote the trace map, so that if 𝑋 = (𝑥𝑖𝑗 ) ∈ 𝔤𝔩𝑛 then tr(𝑋) = ∑𝑖=1 𝑥𝑖𝑖 . Then tr is a
homomorphism of Lie algebas: indeed

tr([𝑋, 𝑌]) = tr(𝑋𝑌) − tr(𝑌𝑋) = 0 = [tr(𝑋), tr(𝑌)].

It follows that if 𝔰𝔩𝑛 ≔ {𝑋 ∈ 𝔤𝔩𝑛 ∶ tr(𝑋) = 0} then 𝔰𝔩𝑛 is a Lie subalgebra of 𝔤𝔩𝑛 (even though it is
not a subalgebra of the associative algebra Mat𝑛 (k) provided 𝑛 > 1). Similarly we define 𝔰𝔩𝑉 to be
the Lie subalgebra of 𝔤𝔩𝑉 constisting of endomorphisms of trace 0. These are called special linear Lie
algebras. More generally we say any Lie subalgebra of 𝔤𝔩𝑉 for a vector space 𝑉 is a linear Lie algebra.
3
If it is not clear from context which field k the vector space 𝑉 is over, we will write 𝔤𝔩k (𝑉).

2
 vii) If 𝔤 is a k-Lie algebra and 𝑥 ∈ 𝔤, then the map 𝜙𝑥 ∶ 𝔤𝔩1 (k) → 𝔤 given by 𝜙𝑥 (𝑡) = 𝑡.𝑥 is a Lie alge-
bra homomorphism, because the alternating property means that a Lie bracket vanishes on any 1-
dimensional subspace of a Lie algebra. This gives a bijection between Lie algebra homomorphisms
𝜙 ∶ 𝔤𝔩1 (k) → 𝔤 and the elements of 𝔤 where if 𝑥 ∈ 𝔤 we let 𝑥 ↦ 𝜙𝑥 ∶ 𝔤𝔩1 (k) → 𝔤 as above, while
given 𝜙 ∶ 𝔤𝔩1 (k) → 𝔤 we associate to it 𝜙(1) ∈ 𝔤.

viii) If 𝔤1 , 𝔤2 are Lie algebras, then we may form their direct sum 𝔤1 ⊕ 𝔤2 , which is the direct sum of 𝔤1
and 𝔤2 as a vector space, with Lie bracket given by [(𝑥1 , 𝑥2 ), (𝑦1 , 𝑦2 )] = ([𝑥1 , 𝑦1 ], [𝑥2 , 𝑦2 ]) for all
𝑥1 , 𝑦1 ∈ 𝔤1 , 𝑥2 , 𝑦2 ∈ 𝔤2 . We may define the direct sum of 𝑘 ≥ 2 Lie algebras in the same way.

ix) If 𝔞 is an abelian Lie algebra then if we chose a basis {𝑒1 , … , 𝑒𝑘 } of 𝔞, then we obtain an isomorphism
𝑘
𝜃 ∶ 𝔤𝔩1 (k)⊕𝑘 → 𝔞 where 𝜃(𝑡1 , … , 𝑡𝑘 ) = ∑𝑖=1 𝑡𝑖 𝑒𝑖 . Indeed the Lie bracket on both 𝔞 and 𝔤𝔩1 (k)⊕𝑘 is
zero, hence we need only check that 𝜃 is an isomorphism of vector spaces, which is clear by con-
struction.

The following definition should be understood as the infinitesimal analogue of an automorphism of a

k-algebra.

Definition 1.1.6. Generalising the example of vector fields in the previous chapter, if 𝐴 is a k-algebra and
𝛿 ∶ 𝐴 → 𝐴 is a k-linear map, then we say 𝛿 is a k-derivation if it satisfies the Leibniz rule, that is, if:

𝛿(𝑎.𝑏) = 𝛿(𝑎).𝑏 + 𝑎.𝛿(𝑏), ∀ 𝑎, 𝑏 ∈ 𝐴.

It is easy to see by a direct calculation that if Derk (𝐴) denotes the k-vector space of k-derivations on 𝐴,
then Derk (𝐴) is stable under taking commutators, that is, if

[𝛿1 , 𝛿2 ] = 𝛿1 ∘ 𝛿2 − 𝛿2 ∘ 𝛿1 .

then [𝛿1 , 𝛿2 ] ∈ Derk (𝐴). Indeed

(𝛿1 ∘ 𝛿2 − 𝛿2 ∘ 𝛿1 )(𝑎.𝑏) = 𝛿1 (𝛿2 (𝑎).𝑏 + 𝑎.𝛿2 (𝑏)) − 𝛿2 (𝛿1 (𝑎).𝑏 + 𝑎𝛿2 (𝑏))
= 𝛿1 𝛿2 (𝑎).𝑏 + 𝛿2 (𝑎).𝛿1 (𝑏) + 𝛿1 (𝑎).𝛿2 (𝑏) + 𝑎.𝛿2 (𝛿1 (𝑏))
− 𝛿2 𝛿1 (𝑎).𝑏 − 𝛿1 (𝑎).𝛿2 (𝑏) − 𝛿2 (𝑎).𝛿1 (𝑏) − 𝑎.𝛿2 𝛿1 (𝑏))
= [𝛿1 , 𝛿2 ](𝑎).𝑏 + 𝑎.[𝛿1 , 𝛿2 ](𝑏).

Example 1.1.7. i) If 𝐴 is an associative k-algebra, then if 𝑎 ∈ 𝐴 the operation of taking commutator

with 𝑎 is a derivation. That is, if 𝛿𝑎 ∶ 𝐴 → 𝐴 is given by 𝛿𝑎 (𝑏) = [𝑎, 𝑏] for any 𝑏 ∈ 𝐴, then 𝛿𝑎 ∈
Derk (𝐴). Indeed

𝛿𝑎 (𝑏).𝑐 + 𝑏.𝛿𝑎 (𝑐) = (𝑎𝑏 − 𝑏𝑎)𝑐 + 𝑏(𝑎𝑐 − 𝑐𝑎) = 𝑎.(𝑏𝑐) − (𝑏𝑐).𝑎 = 𝛿𝑎 (𝑏.𝑐)

The map Δ ∶ 𝔤𝐴 → Derk (𝐴) given by Δ(𝑎) = 𝛿𝑎 is a homomorphism of Lie algebras, that is,
Δ([𝑎, 𝑏]) = [𝛿𝑎 , 𝛿𝑏 ]. In fact slightly more is true: if 𝜕 ∈ Derk (𝐴) and 𝑏 ∈ 𝐴 then [𝜕, 𝛿𝑏 ] = 𝛿𝜕(𝑏) .
(Applying this to 𝜕 = 𝛿𝑎 gives the compatibility with commutators). Indeed for all 𝑐 ∈ 𝔤 we have

[𝜕, 𝛿𝑏 ](𝑐) = 𝜕(𝑏𝑐 − 𝑐𝑏) − (𝑏𝜕(𝑐) − 𝜕(𝑐).𝑏) = 𝜕(𝑏).𝑐 − 𝑐.𝜕(𝑏) = 𝛿𝜕(𝑏) (𝑐).

ii) Given a Lie algebra 𝔤 we let Derk (𝔤) = {𝜙 ∈ 𝔤𝔩𝔤 ∶ 𝜙([𝑥, 𝑦]) = [𝜙(𝑥), 𝑦] + [𝑥, 𝜙(𝑦)]}. It is a Lie
subalgebra of 𝔤𝔩𝔤 (indeed the proof above that Derk (𝐴) is a Lie algebra only requires the product
on 𝐴 to be bilinear). Note that if 𝐴 is an associative algebra, then Derk (𝐴) ⊆ Derk (𝔤𝐴 ) since if
𝛿 ∈ Derk (𝐴) then 𝛿(𝑎𝑏 − 𝑏𝑎) = 𝛿(𝑎).𝑏 + 𝑎.𝛿(𝑏) − 𝛿(𝑏).𝑎 − 𝑏𝛿(𝑎) = [𝛿(𝑎), 𝑏) + [𝑎, 𝛿(𝑏)].

3
 iii) One way of interpreting the Jacobi identity is that, assuming the alternating property, it is equiva-
lent to the condition that, for any 𝑥 ∈ 𝔤, the operation ad(𝑥) ∈ 𝔤𝔩𝔤 given by ad(𝑥)(𝑦) = [𝑥, 𝑦] lies in
Derk (𝔤). Indeed
ad(𝑥)([𝑦, 𝑧]) = [ad(𝑥)(𝑦), 𝑧] + [𝑦, ad(𝑥)(𝑧)]
⟺ [𝑥, [𝑦, 𝑧]] = [[𝑥, 𝑦], 𝑧] + [𝑦, [𝑥, 𝑧]]
⟺ [𝑥, [𝑦, 𝑧]] − [𝑦, [𝑥, 𝑧]] − [[𝑥, 𝑦], 𝑧] = 0
⟺ [𝑥, [𝑦, 𝑧]] + [𝑦, [𝑧, 𝑥]] + [𝑧, [𝑥, 𝑦]] = 0
where the equivalence between the third and fourth equalities follows from the alternating prop-
erty of a Lie bracket.
iv) The Jacobi identity is also equivalent, again assuming the alternating property, to the fact that
ad ∶ 𝔤 → 𝔤𝔩𝔤 is a homomorphism of Lie algebras: Indeed, for all 𝑥, 𝑦, 𝑧 ∈ 𝔤 we have

[ad(𝑥), ad(𝑦)](𝑧) = [𝑥, [𝑦, 𝑧]] − [𝑦, [𝑥, 𝑧]]

= [𝑥, [𝑦, 𝑧]] + [𝑦, [𝑧, 𝑥]]
= −[𝑧, [𝑥, 𝑦]]
= ad([𝑥, 𝑦])(𝑧).
where the second and fourth equality uses the alternating property, and the third the Jacobi iden-
tity.
v) If (𝐴, ∗) is any k-algebra then 𝐴op is the k-algebra with product ∗op , where 𝑎 ∗op 𝑏 = 𝑏 ∗ 𝑎. If 𝐴 is
commutative, then 𝐴op is isomorphic to 𝐴. In the case of a Lie algebra 𝔤 then 𝔤op is the Lie algebra
(𝔤, −[., .]). In fact 𝔤 is canonically isomorphic to 𝔤op : if we let 𝑚 ∶ 𝔤 → 𝔤op be the map 𝑚(𝑥) = −𝑥,
then
𝑚([𝑥, 𝑦]) = −[𝑥, 𝑦] = [𝑦, 𝑥] = [𝑥, 𝑦]op = [−𝑥, −𝑦]op = [𝑚(𝑥), 𝑚(𝑦)], 𝑥, 𝑦 ∈ 𝔤.
hence 𝑚 ∶ 𝔤 → 𝔤op is an isomorphism from 𝔤 to 𝔤op .
*Remark 1.1.8. Combining 𝑖𝑖𝑖) and 𝑖𝑣) in the above example we see that the adjoint representation 𝑥 ↦
ad(𝑥) is in fact a Lie algebra homomorphism from 𝔤 to Derk (𝔤). This is, in a sense, where the Jacobi identity
comes from: very roughly, the conjugation action of 𝐺 on itself yields a group homomorphism 𝐺 → 𝔤𝔩𝔤
(since conjugation preserves the identity 𝑒 ∈ 𝐺) whose image lies in Aut(𝔤). The adjoint representation of
𝔤 is then the derivative of this action yields the adjoint representation ad which hence should have image
in Derk (𝔤).

1.2 Ideals and isomorphism theorems

As one might expect if a Lie algebra is suppose to be an “infinitesimal” version of a Lie group, most notions
for groups have analogues in the context of Lie algebras. It might be worth noting, however, that the linear
structure of a Lie algebra comes from the basic properties of the derivative: it is the Lie bracket which
reflects the “infinitesimal” versions of properties of a group. The existence of both the linear structure and
the Lie bracket means that many of the notions we consider for a Lie algebra also have natural analogues
for a ring (which is an algebra object equipped with an addition and an (associative) multiplication.
Definition 1.2.1. An ideal in a Lie algebra (𝔤, [., .]𝔤 ) is a subspace 𝔞 such that for all 𝑥 ∈ 𝔤 and 𝑎 ∈ 𝔞 we
have [𝑎, 𝑥]𝔤 ∈ 𝔞. It is easy to check that if 𝜙 ∶ 𝔤1 → 𝔤2 is a homomorphism, then

ker(𝜙) = {𝑎 ∈ 𝔤1 ∶ 𝜙(𝑎) = 0}
is an ideal of 𝔤1 . We will write 𝐼E𝔤 to indicate that 𝐼 is an ideal in 𝔤.

4
Remark 1.2.2. Notice that because a Lie bracket is alternating, the condition that, for all 𝑥 ∈ 𝔤 and 𝑎 ∈ 𝔞
one has [𝑎, 𝑥] ∈ 𝔞, is equivalent to the condition that [𝑥, 𝑎] ∈ 𝔞 for all 𝑥 ∈ 𝔤, 𝑎 ∈ 𝔞. Thus, similarly to
commutative rings, the notions of a left, right or two-sided ideal in a Lie algebra are all the same.

Just as for rings, in fact any ideal is the kernel of a Lie algebra homomorphism:

Theorem 1.2.3. (The universal property of quotients & the 1st isomorphism theorem): Let 𝔞 be an ideal in a Lie
algebra 𝔤, and let 𝑞 ∶ 𝔤 → 𝔤/𝔞 be the quotient map of vector spaces. Then there is a unique Lie bracket on 𝔤/𝔞 with
respect to which 𝑞 is a homomorphism of Lie algebras, that is, for all 𝑥, 𝑦 ∈ 𝔤

[𝑞(𝑥), 𝑞(𝑦)] = 𝑞([𝑥, 𝑦]), i.e. [𝑥 + 𝔞, 𝑦 + 𝔞] = [𝑥, 𝑦] + 𝔞.

Moreover, if 𝜑 ∶ 𝔤 → 𝔨 is a Lie algebra homomorphism such that 𝜑(𝔞) = 0, then 𝜑 induces a homomorphism
𝜑 ∶ 𝔤/𝔞 → 𝔨 such that 𝜑 ∘ 𝑞 = 𝜑, so that ker(𝜑) = ker(𝜑)/𝔞. In particular, if we set 𝔞 = ker(𝜑) then we see that
𝜑 induces an isomorphism 𝜑 ∶ 𝔤/ker(𝜑) → im(𝜑).

Proof. The proof is almost identical to the proof in the case of rings. The key point is to see that the coset
[𝑥, 𝑦] + 𝔞 is independent of the choice of representative for the cosets 𝑥 + 𝔞, 𝑦 + 𝔞, and the condition that
𝔞 is an ideal ensures this. 

Definition 1.2.4. If 𝑉, 𝑊 are subspaces of a Lie algebra 𝔤, then write [𝑉, 𝑊] for the linear span of the
elements {[𝑣, 𝑤] ∶ 𝑣 ∈ 𝑉, 𝑤 ∈ 𝑊}. Notice that if 𝐼, 𝐽 are ideals in 𝔤 then so is [𝐼, 𝐽]. Indeed to check this,
note that by part 8) of Example 1.1.5, if 𝑧 ∈ 𝔤, 𝑥 ∈ 𝐼, 𝑦 ∈ 𝐽 then we have

[𝑧, [𝑥, 𝑦]] = ad(𝑧)([𝑥, 𝑦]) = [ad(𝑧)(𝑥), 𝑦] + [𝑥, ad(𝑧)(𝑦)] ∈ [𝐼, 𝐽]

since ad(𝑧)(𝑥) = [𝑧, 𝑥] ∈ 𝐼 if 𝑥 ∈ 𝔤, and similarly ad(𝑧)(𝑦) = [𝑧, 𝑦] ∈ 𝐽.

Remark 1.2.5. If 𝐼 and 𝐽 are ideals in a Lie algebra 𝔤 then it is easy to check that their intersection 𝐼 ∩ 𝐽
is again an ideal in 𝔤, and we have [𝐼, 𝐽] ⊆ 𝐼 ∩ 𝐽. (Thus [𝐼, 𝐽] is the Lie algebra analogue of the product of
ideals in a commutative ring.) For example, since 𝔤 is an ideal in 𝔤, its product with itself, [𝔤, 𝔤] is an ideal
𝐷𝔤 ⊆ 𝔤, known as the derived subalgebra of 𝔤4 . The quotient 𝔤/𝐷𝔤 is the largest abelian quotient of 𝔤, in
the sense that if 𝜙 ∶ 𝔤 → 𝔞 where 𝔞 is abelian, then 𝜙 factors through the quotient map 𝑞 ∶ 𝔤 → 𝔤/𝐷𝔤 (we
will prove this later in Lemma 4.1.4).
Similarly, it is easy to see that the linear sum 𝐼 + 𝐽 of 𝐼 and 𝐽 is also an ideal5 .

Definition 1.2.6. Let 𝔤 be a Lie algebra and let 𝔞 ≤ 𝔤 be a subalgebra. The normalizer of 𝔞 in 𝔤 is

𝑁𝔤 (𝔞) ≔ {𝑥 ∈ 𝔤 ∶ ad(𝑥)(𝔞) ⊆ 𝔞} = {𝑥 ∈ 𝔤 ∶ ad(𝑎)(𝑥) ∈ 𝔞, ∀ 𝑎 ∈ 𝔞}.

This is a subalgebra of 𝔤, as one can check using the formulation of the Jacobi identity given in Definition
1.2.4. It is the largest subalgebra of 𝔤 within which 𝔞 is an ideal.

Definition 1.2.7. If a nontrivial Lie algebra has no nontrivial ideals we say that it is almost simple. It it is
in addition not abelian, i.e. the Lie bracket is not identically zero, then we say that it is simple.

Just as for groups and rings, one can deduce the usual stable of isomorphism theorems from the first
isomorphism theorem.

Theorem 1.2.8. i) If 𝔥 is a subalgebra of 𝔤 and 𝐼 is an ideal in 𝔤 then 𝔥 + 𝐼 is a subalgebra of 𝔤 (containing

𝐼 as an ideal) 𝔥 ∩ 𝐼 is an ideal in 𝔥, and

(𝔥 + 𝐼)/𝐼 ≅ 𝔥/(𝔥 ∩ 𝐼).

4
Even though it is, as we have already noted, an ideal, not just a subalgebra.
5
Note however that the linear sum of two subalgebras is not necessarily a subalgebra.

5
 ii) If 𝐽 ⊆ 𝐼 ⊆ 𝔤 are ideals of 𝔤 then we have:

(𝔤/𝐽)�(𝐼/𝐽) ≅ 𝔤/𝐼.

Proof. The proofs are identical to the corresponding results for groups. We give a proof of 𝑖𝑖) as an example.
Since 𝐽 ⊆ 𝐼 the quotient map 𝔤 ∶ 𝔤 → 𝔤/𝐼, which has kernel 𝐼, induces a map 𝑞̄ ∶ 𝔤/𝐽 → 𝔤/𝐼. The kernel of
this map is by definition {𝑥 + 𝐽 ∶ 𝑥 + 𝐼 = 𝐼}, that is, 𝐼/𝐽. The result follows. 

6
Chapter 2

Representations of Lie algebras

Just as for finite groups (or indeed groups in general) one way of studying Lie algebras is to try and under-
stand how they can act on other (usually more concrete) objects. For Lie algebras, since they are already
vector spaces over k, it is natural to study their action on linear spaces, or in other words, their represen-
tations.

2.1 Definition and examples

Definition 2.1.1. A representation of a Lie algebra 𝔤 is a vector space 𝑉 equipped with a linear action of 𝔤,
that is, a homomorphism of Lie algebras 𝜌 ∶ 𝔤 → 𝔤𝔩𝑉 . In other words, 𝜌 is a linear map such that

𝜌([𝑥, 𝑦]) = 𝜌(𝑥) ∘ 𝜌(𝑦) − 𝜌(𝑦) ∘ 𝜌(𝑥)

where ∘ denotes composition of linear maps. We also refer to a representation of 𝔤 as a 𝔤-representation

or 𝔤-module. A representation is faithful if ker(𝜌) = 0. When there is no danger of confusion we will
normally suppress 𝜌 in our notation, and write 𝑥(𝑣) rather than 𝜌(𝑥)(𝑣), for 𝑥 ∈ 𝔤, 𝑣 ∈ 𝑉.
If (𝑉, 𝜌) and (𝑊, 𝜎) are 𝔤-representations, we say that 𝜙 ∶ 𝑉 → 𝑊 is a 𝔤-homomorphism (or homo-
morphism of 𝔤-representations) if 𝜙 ∘ 𝜌(𝑥) = 𝜎(𝑥) ∘ 𝜙 for all 𝑥 ∈ 𝔤. We will write Rep(𝔤) for the collection
of all representations of 𝔤.1

We will study representation of various classes of Lie algebras in this course, but the following give
some basic examples.
Example 2.1.2. i) If 𝑉 is a k-vector space, then the identity map 𝔤𝔩𝑉 → 𝔤𝔩𝑉 gives a representation of
𝔤𝔩𝑉 on 𝑉, which is known as the vector representation. Clearly any subalgebra 𝔤 of 𝔤𝔩𝑉 also inherits
𝑉 as a representation, where then the action map 𝜌 is just the inclusion map.
ii) Let 𝔞 be an abelian Lie algebra. If (𝑉, 𝜌) is a representation of 𝔞, then the image 𝜌(𝔞) of 𝔞 generates
a commutative subalgebra of Endk (𝑉): if 𝑎, 𝑏 ∈ 𝔞 then 0 = 𝜌([𝑎, 𝑏]) = 𝜌(𝑎)𝜌(𝑏) − 𝜌(𝑏)𝜌(𝑎), so that

𝜌(𝑎)𝜌(𝑏) = 𝜌(𝑏)𝜌(𝑎), ∀ 𝑎, 𝑏 ∈ 𝔞.

iii) Given an arbitrary Lie algebra 𝔤, there is a natural representation ad of 𝔤 on 𝔤 itself known as the
adjoint representation. The homomorphism ad ∶ 𝔤 → 𝔤𝔩𝔤 from 𝔤 to 𝔤𝔩𝔤 is given by

ad(𝑥)(𝑦) = [𝑥, 𝑦], ∀ 𝑥, 𝑦 ∈ 𝔤.

1
Abstractly, Rep(𝔤) is an example of a category. A category is a mathematical structure consisting of a collection of objects
together with, for each pair of objects 𝑥, 𝑦, a collection of “morphisms” mor(𝑥, 𝑦) between 𝑥 and 𝑦. We can view Rep(𝔤) as a
category whose objects are representations of 𝔤 and whose morphisms are 𝔤-homomorphisms. (Note that rhe term “collection”
is used above because of set-theoretic subtleties which we can essentially ignore in this course.)

7
 Indeed, as noted in iv) of Example 1.1.7, the fact that this map is a homomorphism of Lie algebras is
just a rephrasing2 of the Jacobi identity. Note that while the vector representation is clearly faithful,
in general the adjoint representation is not. Indeed the kernel is known as the centre of 𝔤:

𝔷(𝔤) = {𝑥 ∈ 𝔤 ∶ [𝑥, 𝑦] = 0, ∀ 𝑦 ∈ 𝔤}.

Note that if 𝑥 ∈ 𝔷(𝔤) then for any representation 𝜌 ∶ 𝔤 → 𝔤𝔩(𝑉) the endomorphism 𝜌(𝑥) commutes
with all the elements 𝜌(𝑦) ∈ End(𝑉) for all 𝑦 ∈ 𝔤.

iv) If 𝔤 is any Lie algebra, the pair (k, 0) consisting of the vector space k together with the zero map
0 ∶ 𝔤 → 𝔤𝔩1 is a 𝔤-representation. This representation is called the trivial representation. It is the Lie
algebra analogue of the trivial representation for a group (which send every group element to the
identity map 1𝑉 ∈ GL(𝑉)).

v) If (𝑉, 𝜌) is a 𝔤-representation, then the sum of all subrepresentations of 𝑉 isomorphic to (k, 0) is a

natural subrepresentation

𝑉 𝔤 = {𝑣 ∈ 𝑉 ∶ 𝜌(𝑥)(𝑣) = 0, ∀ 𝑥 ∈ 𝔤}

known as the 𝔤-invariants in 𝑉.

vi) If (𝑉, 𝜌) is a representation of a Lie algebra 𝔤 and 𝜃 ∶ 𝔥 → 𝔤 is a homomorphism of Lie algebras, then
we define the pull-back of (𝑉, 𝜌) to the representation of 𝔥 given by (𝑉, 𝜌 ∘ 𝜃). The most common
example of a pull-back is restriction, when 𝔥 is a subalgebra of 𝔤 (and thus 𝜃 is the inclusion map
from 𝔥 to 𝔤).

The following definitions are useful when studying the the structure of Lie algebra representations:

Definition 2.1.3. A representation is said to be irreducible if it has no proper non-zero subrepresentations,

and it is said to be completely reducible if it is isomorphic to a direct sum of irreducible representations. A
representation 𝑉 is said to be indecomposable if, whenever we have 𝑉 = 𝑈1 ⊕ 𝑈2 with 𝑈1 , 𝑈2 subrepre-
sentations, either 𝑈1 = 𝑉 or 𝑈2 = 𝑉 (and 𝑈2 = 0, 𝑈1 = 0 respectively).
It is easy to show (by induction on dimension) that any finite-dimensional representation is a direct
sum of indecomposable subrepresentations.

Remark 2.1.4. A representation (𝑉, 𝜌) is said to be semisimple if, given any subrepresentation 𝑈 of 𝑉
there is a complementary subrepresentation 𝑊 of 𝑉, that is, such that 𝑉 = 𝑈 ⊕ 𝑊. A finite-dimensional
semisimple representation is completely reducible, and any representation which is the sum of its irre-
ducible subrepresentations is semisimple, hence in particular, any completely reducible representation is
semisimple.

Example 2.1.5. In this example, we classify the representations of the simplest Lie algebra 𝔤𝔩1 : a rep-
resentation (𝑉, 𝜌) of 𝔤𝔩1 is given by a Lie algebra homomorphism 𝜌 ∶ 𝔤𝔩1 → 𝔤𝔩𝑉 . But we saw in vii)
of Example 1.1.5, that there is a natural bijection between such homomorphisms and elements 𝑥 ∈ 𝔤𝔩𝑉
given by 𝜌 ↦ 𝜌(1). Through this correspondence the problem of classifying 𝔤𝔩1 -representations up to
isomorphism becomes the problem of classifying vector spaces equipped with an endomorphism up to
conjugacy.
If we assume k is algebraically closed this classification of linear endomorphisms is given by the Jordan
canonical form. It is a useful exercise to translate statements about linear maps into statements about
representations of 𝔤𝔩1 . For example, the irreducible representations of 𝔤𝔩1 are the one-dimensional ones,
and correspond to eigenvectors of 𝜌(1). What do the indecomposable representations correspond to?
2
It’s also (for some people) a useful way of remembering what the Jacobi identity says.

8
Example 2.1.6. Now suppose that 𝔤 is any finite-dimensional Lie algebra and that (𝐿, 𝜆) is a one-
dimensional representation of 𝔤. The canonical identifications 𝔤𝔩𝐿 ≅ 𝑘 = 𝔤𝔩1 (k) given in v) of Example
1.1.5 identifies 𝜆 with a homomorphism [𝜆] ∶ 𝔤 → 𝔤𝔩1 . But since k = 𝔤𝔩1 has the zero Lie bracket, this is
just an element of 𝔤∗ which vanishes on {[𝑥, 𝑦] ∶ 𝑥, 𝑦 ∈ 𝔤}, that is, 𝜆 is an element of (𝐷𝔤)0 , the annihilator
of 𝐷(𝔤) = [𝔤, 𝔤], the derived subalgebra of 𝔤 introduced in Remark 1.2.5.

2.2 Subrepresentations, quotients, duals, and composition series

There are a number of standard ways of constructing new representations from old, all of which have their
analogues in the context of group representations. We begin with some definitions.

Definition 2.2.1. Let 𝑉 be a k-vector space and 𝑈 ≤ 𝑉 a subspace. Write 𝑖 ∶ 𝑈 → 𝑉 for the inclusion
map and 𝑝 ∶ 𝑉 → 𝑉/𝑈 for the quotient map. Let

𝔟𝑈 = {𝑥 ∈ 𝔤𝔩𝑉 ∶ 𝑥(𝑈) ⊆ 𝑈} = {𝑥 ∈ 𝔤𝔩𝑉 ∶ 𝑝 ∘ 𝑥 ∘ 𝑖 = 0}.

Then we have linear maps 𝑖∗ ∶ 𝔟𝑈 → End(𝑈) and 𝑝∗ ∶ 𝔟𝑈 → End(𝑉/𝑈). Here 𝑖∗ (𝑥) = 𝑥 ∘ 𝑖 is simply the
restriction of 𝑥 to 𝑈. For the map 𝑝∗ first note that if 𝑥 ∈ 𝔟𝑈 then 𝑈 ⊆ ker(𝑝 ∘ 𝑥), and so by the universal
property of a quotient map 𝑝 ∘ 𝑥 induces a map 𝛼̄ ∶ 𝑉/𝑈 → 𝑉/𝑈 which satisfies 𝑝 ∘ 𝑥 = 𝑥̄ ∘ 𝑝, and we set
𝑝∗ (𝛼) = 𝛼.̄

Lemma 2.2.2. If 𝜌 ∶ 𝔤 → 𝔤𝔩𝑉 is a 𝔤-representation and 𝜌(𝔤) ⊆ 𝔟𝑈 for some subspace 𝑈 ≤ 𝑉, then 𝑈 and 𝑉/𝑈
become 𝔤-representations with action maps 𝑖∗ ∘ 𝜌 and 𝑝∗ ∘ 𝜌 respectively.

Proof. It is clear that 𝔟𝑈 is an associative subalgebra of Endk (𝑉) and both 𝑖∗ and 𝑝∗ are homomorphisms
of associative algebras, hence they are also homomorphisms of the associated Lie algebras. The Lemma
follows immediately. 

It will be useful later to have the following definition:

Definition 2.2.3. Let 𝑉 be a vector space, and let ℱ = (𝐹𝑖 )𝑘𝑖=0 be a flag in 𝑉, that is

ℱ = (𝑉 = 𝐹0 ⊃ 𝐹1 ⊃ 𝐹2 ⊃ … ⊃ 𝐹𝑘 = 0)

is a nested sequence of subspaces with dim(𝐹𝑖+1 ) < dim(𝐹𝑖 ) for 1 ≤ 𝑖 ≤ 𝑘. If ℱ 1 and ℱ 2 are flags in 𝑉
then we say that ℱ 2 is a refinement of ℱ 1 if every subspace in ℱ 1 occurs in ℱ 2 . If dim(𝐹𝑖 ) = 𝑖 for all 𝑖
(so that dim(𝑉) = 𝑘) then ℱ is called a complete flag (as it cannot be refined any further). It is clear (since
any linearly independent set can be extended to a basis) that any flag can be refined to a complete flag.
We let 𝔟ℱ = ⋂1≤𝑖≤𝑘−1 𝔟𝐹𝑖 = {𝑥 ∈ 𝔤𝔩𝑉 ∶ 𝑥(𝐹𝑖 ) ⊆ 𝐹𝑖 }. This is an associative subalgebra of End(𝑉), and
hence a Lie subalgebra. If (𝑉, 𝜌) is a 𝔤-representation, the elements of the flag are subrepresentations of
𝑉 if and only if 𝜌(𝔤) ⊆ 𝔟ℱ .

Definition 2.2.4. If 𝑉 is a k-vector space and 𝑎 ∈ 𝔤𝔩𝑉 , then 𝑎 induces a linear map 𝑎⊺ ∶ 𝑉 ∗ → 𝑉 ∗ which
we call the adjoint (or transpose) of 𝑎, given by 𝑎⊺ (𝑓)(𝑣) = (𝑓 ∘ 𝑎)(𝑣). However, if 𝑎, 𝑏 ∈ 𝔤𝔩𝑉 , 𝑣 ∈ 𝑉,

(𝑎𝑏)⊺ (𝑓)(𝑣) = (𝑓 ∘ (𝑎𝑏))(𝑣) = ((𝑓 ∘ 𝑎) ∘ 𝑏)(𝑣) = 𝑏⊺ (𝑓 ∘ 𝑎)(𝑣) = 𝑏⊺ ∘ 𝑎⊺ (𝑓)(𝑣),

so that 𝑎 ↦ 𝑎⊺ is an algebra anti-homomorphism from Endk (𝑉) to Endk (𝑉 ∗ )op (see 1.1.5 (v))), so that
it is also a Lie algebra homomorphism 𝔤𝔩𝑉 → 𝔤𝔩𝑉 ∗ op . But, again by 1.1.5 (v)) the map 𝑥 ↦ −𝑥 is an
isomorphism from 𝔤 → 𝔤op for any Lie algebra, it follows that 𝑥 ↦ −𝑥⊺ is an isomorphism of Lie algebras
from 𝔤𝔩𝑉 to 𝔤𝔩𝑉 ∗ . For 𝑥 ∈ 𝔤𝔩𝑉 , we will write 𝑥∗ = −𝑥⊺ . Moreover, if (𝑉, 𝜌) is a 𝔤-representation, then we
may define 𝜌∗ (𝑥) = (𝜌(𝑥))∗ , so that 𝜌∗ ∶ 𝔤 → 𝔤𝔩𝑉 ∗ is a Lie algebra homomorphism and hence (𝑉 ∗ , 𝜌∗ ) is a
𝔤-representation, the dual representation to (𝑉, 𝜌).

9
 Recall that the annihilator of a subspace 𝑈 ≤ 𝑉 is the subspace of 𝑉 ∗ given by

𝑈 0 = {𝑓 ∈ 𝑉 ∗ ∶ 𝑓(𝑢) = 0, ∀ 𝑢 ∈ 𝑈}.

It is clear from the definition that the correspondence 𝑈 ↦ 𝑈 0 is order-reversing for containment, that
is, if 𝑈1 ≤ 𝑈2 then 𝑈20 ≤ 𝑈10 . By considering a basis of 𝑉 and the corresponding dual basis of 𝑉 ∗ , one can
also check that dim(𝑈) + dim(𝑈 0 ) = dim(𝑉), and

Lemma 2.2.5. If (𝑉, 𝜌) is a 𝔤-representation with dual representation (𝑉 ∗ , 𝜌∗ ), then the map 𝑈 ↦ 𝑈 0 gives
an order-reversing correspondence between the subrepresentations of 𝑉 and 𝑉 ∗ respectively. Since 𝑉 is finite-
dimensional, (𝑉 ∗ )∗ is canonically isomorphic to 𝑉, and via that canonical identification, this correspondence is an
involution, that is, (𝑈 0 )0 = 𝑈. Moreover, 𝑈 ∗ ≅ 𝑉 ∗ /𝑈 0 and (𝑉/𝑈)∗ ≅ 𝑈 0 as 𝔤-representations.

Proof. Let 𝜄 ∶ 𝑈 → 𝑉 be the inclusion map of a subspace 𝑈 and let 𝑝 ∶ 𝑉 → 𝑉/𝑈 be the quotient map.
Then we have seen that 𝔟𝑈 = {𝑥 ∈ 𝔤𝔩𝑉 ∣ 𝑥(𝑈) ⊆ 𝑈} = {𝑥 ∈ 𝔤𝔩𝑉 ∶ 𝑝 ∘ 𝑥 ∘ 𝜄 = 0}. Now the adjoint of the
inclusion map 𝑖⊺ ∶ 𝑉 ∗ → 𝑈 ∗ clearly has kernel 𝑈 0 and is, moreover, surjective (as one can check using
dual bases for example). Moreover, the adjoint of the quotient map 𝑝⊺ ∶ (𝑉/𝑈)∗ → 𝑉 ∗ is clearly injective
with image 𝑈 0 .
It follows that 𝔟𝑈 0 = {𝑦 ∈ 𝔤𝔩𝑉 ∗ ∣ 𝑦(𝑈 0 ) ⊆ 𝑈 0 } = {𝑦 ∈ 𝔤𝔩𝑉 ∗ ∣ 𝜄⊺ ∘ 𝑦 ∘ 𝑝⊺ = 0}. But if 𝑥 ∈ 𝔤𝔩𝑉 and
𝑥∗ = −𝑥⊺ then
𝑖⊺ ∘ (𝑥∗ ) ∘ 𝑝⊺ = −(𝑖⊺ ∘ 𝑥⊺ ∘ 𝑝⊺ ) = −(𝑝 ∘ 𝑥 ∘ 𝑖)⊺ .
Thus the isomorphism of Lie algebra from 𝔤𝔩𝑉 to 𝔤𝔩𝑉 ∗ given by 𝑥 ↦ 𝑥∗ = −𝑥⊺ identifies 𝔟𝑈 with 𝔟𝑈 0 ,
and 𝑈 is a subrepresentation of 𝑉 if and only if 𝑈 0 is a subrepresentation of 𝑉 ∗ . Since the correspondence
𝑈 ↦ 𝑈 0 is clearly order-reversing, the first sentence of the Lemma follows, and since the maps 𝑝⊺ and 𝜄⊺
are clearly 𝔤-homomorphisms when 𝑈 is a subrepresentation of 𝑉, the final sentence is also immediate.
Finally, the canonical isomorphism 𝑑 ∶ 𝑉 → 𝑉 ∗∗ clearly maps 𝑈 into (𝑈 0 )0 , and the dimension formula
dim(𝑈)+dim(𝑈 0 ) = dim(𝑉) shows they have the same dimension, hence 𝑑(𝑈) = (𝑈 0 )0 as required. 

Remark 2.2.6. Note that the previous Lemma shows that the dual of an irreducible representation is
irreducible.

2.2.1 Exact sequences of representations

Definition 2.2.7. A sequence of maps of 𝔤-representations

𝛼 𝛽
𝑈 /𝑉 /𝑊
is said to be exact at 𝑉 if im(𝛼) = ker(𝛽). A sequence of maps
𝛼 𝛽
0 /𝑈 /𝑉 /𝑊 /0
is called a short exact sequence if it is exact at each of 𝑈, 𝑉 and 𝑊, so that 𝛼 is injective and 𝛽 is surjective and
im(𝛼) = ker(𝛽). If 𝑉 is the middle term of such a short exact sequence, it contains a subrepresentation
isomorphic to 𝑈, such that the corresponding quotient representation is isomorphic to 𝑊, and hence,
roughly speaking, 𝑉 is built by gluing together 𝑈 and 𝑊. Just as for Lie algebras, an exact sequence
𝛼 𝛽
0 /𝑈 /𝑉 /𝑊 /0
is said to be split if 𝛽 admits a right inverse 𝑠 ∶ 𝑊 → 𝑉, that is, a 𝔤-homomorphism 𝑠 such that 𝛽 ∘ 𝑠 = id𝑊 .

Lemma 2.2.8. Suppose that 𝔤 is a Lie algebra and

𝛼 𝛽
0 /𝑈 /𝑉 /𝑊 /0
is a short exact sequence of 𝔤-representations. Then the sequence is split if and only if 𝑉 is isomorphic to 𝑈 ⊕ 𝑊.

10
Proof. If the short exact sequence is split, then 𝑠(𝑊)∩𝛼(𝑈) = {0}, since 𝛽 is injective on im(𝑠) and im(𝛼) =
ker(𝛽), as if 𝑣 ∈ 𝑉, then for any 𝑣 ∈ 𝑉 we have 𝑣 = (𝑣 − 𝑠 ∘ 𝛽(𝑣)) + 𝑠 ∘ 𝛽(𝑣)), where 𝑠 ∘ 𝛽(𝑣) lies in the image
of 𝑠, and since 𝛽 ∘ 𝑠 = id𝑊 , 𝑣 − 𝑠 ∘ 𝛽(𝑣) ∈ ker(𝛽) = im(𝛼). It follows that 𝑉 = 𝛼(𝑈) ⊕ 𝑠(𝑊), and since 𝛼
and 𝑠 are injective, the result follows. 

Example 2.2.9. To see a non-split extension, let 𝔤 = 𝔫2 be the one-dimensional Lie algebra, thought of
as the (nilpotent) Lie algebra of 2 × 2 strictly upper triangular matrices. Then its natural 2-dimensional
representation on k2 given by the inclusion 𝔫2 → 𝔤𝔩2 (k) gives a non-split extension
/k 𝑖 / k2 /k /0
0 0 0

where k0 is the trivial representation, and 𝑖 ∶ k0 → k2 is the inclusion 𝑡 ↦ (𝑡, 0). The extension cannot be
trivial, because the image of 𝔫2 is non-zero. It is fact it’s easy to see using linear algebra that for 𝔤𝔩1 (k) = 𝔫2 ,
an extension of one-dimensional representations k𝛼 and k𝛽 automatically splits if 𝛼 ≠ 𝛽 while there is,
up to isomorphism, one non-split extension of k𝛼 with itself (𝛼, 𝛽 ∈ (𝔤𝔩1 (k))∗ ).

The notion of a composition series, which was used to study finite groups in Part A Group Theory, has
an analogue for representations of a given Lie algebra 𝔤.
Definition 2.2.10. Let 𝑉 be a 𝔤-representation. A nested sequence 𝒞 = (𝑉 = 𝐹0 ⊃ 𝐹1 ⊃ … ⊃ 𝐹𝑑 = 0)
is said to be a composition series for 𝑉 if each 𝐹𝑖 is a subrepresentation, and the subquotients 𝐹𝑖−1 /𝐹𝑖 are
irreducible (for each 𝑖 ∈ {1, … , 𝑑}). Thus a composition series admits no proper refinement. It is easy to
show (say by induction on dimension) that any flag of subrepresentations can be refined to a composition
series (and so, in particular, composition series exist). The isomorphism classes of irreducibles which
occur as one of these irreducible subquotients are called the composition factors, and the multiplicity with
which such an irreducible, say 𝑆, occurs is known as its composition multiplicity. We write [𝑆 ∶ 𝒞 ] for the
multiplicity with which 𝑆 occurs as a composition factor of 𝒞.
Remark 2.2.11. A composition series can also be viewed as the vestige of how the representation 𝑉 was
built up from its composition factor 𝑆𝑖 = 𝐹𝑖−1 /𝐹𝑖 . Indeed for each 𝑘 ∈ {1, … , 𝑑} we have
𝛼 𝛽
0 /𝐹 /𝐹 /𝑆 /0
𝑘 𝑘−1 𝑘

Thus starting with 𝑆1 one constructs 𝐹0 /𝐹2 by extending it by 𝑆2 . One obtains 𝐹0 /𝐹3 by extending 𝐹0 /𝐹2
by 𝑆3 and so on, until finally we get 𝑉 by extending 𝐹0 /𝐹1 by 𝐹1 = 𝑆1 to obtain 𝑉 itself!
Proposition 2.2.12 (Jordan-Hölder theorem). Let 𝑆 be an irreducible finite-dimensional 𝔤-representation and
let (𝑉, 𝜌) be any finite-dimensional 𝔤-representation. The multiplicity with which a given isomorphism class 𝑆 of
simple module occurs in a composition series for 𝑉 is independent of the compositions series, hence we may denote
it by [𝑆 ∶ 𝑉].

Proof. We prove the proposition by induction on the minimal length ℓ(𝑉) of a composition series for a
representation 𝑉. If ℓ(𝑉) = 1 then 𝑉 is irreducible, and hence it has a unique composition series.
Now suppose that ℓ(𝑉) = 𝑘 > 1 and that 𝒞 = (𝐹𝑖 )𝑘𝑖=0 and 𝒞 ′ = (𝐺𝑗 )𝑙𝑗=0 are composition series for
𝑉 (where 𝒞 is of minimal length). Consider 𝐹𝑘 < 𝑉, a simple submodule. Consider the intersections
(𝐺𝑗 ∩ 𝐹𝑘 )𝑙𝑗=0 . These form a weakly decreasing sequence of submodules of 𝐹𝑘 , and hence as 𝐹𝑘 is simple,
there is a unique 𝑠 such that 𝐺𝑗 ∩ 𝐹𝑘 = 𝐹𝑘 for 𝑗 ≤ 𝑠 and 𝐺𝑗 ∩ 𝐹𝑘 = 0 for 𝑗 > 𝑠. But then 𝐹𝑘 ⊆ 𝐺𝑠 but
𝐹𝑘 ∩ 𝐺𝑠+1 = {0}, so that by the second isomorphism theorem
≅
𝐹𝑘 = 𝐹𝑘 /(𝐹𝑘 ∩ 𝐺𝑠+1 ) −
→ (𝐹𝑘 + 𝐺𝑠+1 )/𝐺𝑠+1 ↪ 𝐺𝑠 /𝐺𝑠+1 .

Since 𝐺𝑠 /𝐺𝑠+1 is simple, it follows that 𝐹𝑘 ≅ 𝐺𝑠 /𝐺𝑠+1 and hence 𝐺𝑠 = 𝐹𝑘 ⊕ 𝐺𝑠+1 . But then if 𝑞 ∶ 𝑉 →
𝑉/𝐹𝑘 is the quotient map, it restricts to an isomorphism 𝑞|𝐺𝑠+1 ∶ 𝐺𝑠+1 → 𝐺𝑠 /𝐹𝑘 , and hence (𝑞(𝐺𝑗 ))0≤𝑗≤𝑙,𝑗≠𝑠
is a composition series for 𝑉/𝐹𝑘 . Now clearly (𝑞(𝐹𝑖+1 ))𝑘−1
𝑖=0 is also a composition series for 𝑉/𝐹𝑘 which,
moreover, has length 𝑘−1, so that ℓ(𝑉/𝐹𝑘 ) ≤ 𝑘−1, and thus by induction, the composition series (𝐹𝑖 /𝐹𝑘 )𝑘−1
𝑖=0
and (𝑞(𝐺𝑗 ))𝑗≠𝑠 have the same composition multiplicities.

11
 Since these differ from the composition multiplicities of 𝒞 and 𝒞 ′ respectively only in having one
fewer copy of the isomorphism class of the simple module 𝐹𝑘 , it follows that 𝒞 and 𝒞 ′ have the same
composition multiplicities as required.


Composition factors are compatible with duality for representations. More precisely, we have the
following:

Lemma 2.2.13. Suppose that 𝑉 is a 𝔤-representation and that {𝑆𝑖 ∶ 1 ≤ 𝑖 ≤ 𝑘} are its composition factors. Then
𝑉 ∗ has composition factors {𝑆∗𝑖 ∶ 1 ≤ 𝑖 ≤ 𝑘} and moreover [𝑆𝑖 ∶ 𝑉] = [𝑆∗𝑖 ∶ 𝑉 ∗ ].

Proof. We use induction on ℓ(𝑉) the length of a composition series for 𝑉, the result being trivial if ℓ(𝑉) =
0. that is, if 𝑉 = {0}. If ℓ(𝑉) = 𝑑 and 𝒞 = (𝐹𝑖 )𝑑𝑖=0 is a composition series for 𝑉, then by taking annihilators
0 0
we obtain a flag 𝒞 ∗ = (𝐹𝑑−𝑘 )𝑑𝑘=0 of subrepresentations of 𝑉 ∗ . By Lemma 2.2.5 𝐹𝑑−1 ≅ (𝑉/𝐹𝑑−1 )∗ and
0
𝑉/𝐹𝑑−1 ≅ 𝐹𝑑−1∗
. Since ((𝑉/𝐹𝑑−1 )/(𝐹𝑖 /𝐹𝑑−1 ))∗ ≅ (𝑉/𝐹𝑖 )∗ ≅ 𝐹𝑖0 , the result then follows by applying induction
to 𝑉/𝐹𝑑−1 and its composition series (𝐹𝑘 /𝐹𝑑−1 )𝑑−1 𝑘=0 . 

Although we will not use “characters” of Lie algebra representations (that is, the function on 𝔤 given
by 𝑥 ↦ tr𝑉 (𝜌(𝑥)), we will instead use “trace forms” and the next Lemma will imply that the trace form
only depends on the composition factors of the representation.

Lemma 2.2.14. Suppose that 𝑉 is a finite-dimensional k-vector space and let ℱ = (𝐹𝑘 )𝑑𝑘=0 is a flag of subspaces
of 𝑉, 𝑉 = 𝐹0 > 𝐹1 > … 𝐹𝑑−1 > 𝐹𝑑 = {0}. If 𝔟ℱ = {𝑎 ∈ End(𝑉) ∶ 𝑥(𝐹𝑖 ) ⊆ 𝐹𝑖 ∶ 0 ≤ 𝑖 ≤ 𝑑} then any 𝑎 ∈ 𝔟ℱ
induces a linear map 𝑎𝑖 ∶ 𝑆𝑖 → 𝑆𝑖 where 𝑆𝑖 ≔ 𝐹𝑖 /𝐹𝑖+1 , and we have
𝑑−1
tr𝑉 (𝑎) = � tr𝑆𝑖 (𝑎𝑖 )
𝑖=0

Proof. We use induction on 𝑑 the length of the composition series. If 𝑑 = 0 then there is nothing to prove.
Otherwise, 𝐹𝑑−1 ≠ 0 and 𝑥(𝐹𝑑−1 ) ⊆ 𝐹𝑑−1 = 𝑆𝑑−1 . and hence 𝑥 induces a linear map 𝑥̄ ∶ 𝑉/𝐹𝑑 → 𝑉/𝐹𝑑 , and
if 𝐵1 is a basis of 𝐹𝑑−1 , extended to a basis of 𝑉 by adding the vectors in a set 𝐵2 . Then if 𝑞 ∶ 𝑉 → 𝑉/𝐹𝑑 is
the natural quotient map, 𝑞(𝐵2 ) is a basis of 𝑞(𝑉), so that it is clear that tr𝑉 (𝑥) = tr𝑆𝑑−1 (𝑎)̄ + tr𝑉/𝐹𝑑−1 (𝑎),
̄
𝑑−1
and then by induction applied to 𝑉/𝐹𝑑−1 and the flag (𝐹𝑖 /𝐹𝑑−1 )𝑖=0 gives the result. 

Definition 2.2.15. If 𝑉 is a 𝔤-representation, we let 𝑉 𝑠 = ∑𝑆≤𝑉 𝑆 the socle of 𝑉, be the sum of all irre-
ducible subrepresentations of 𝑉. This is a semisimple subrepresentation of 𝑉 and hence it can be written
as the direct sum of irreducible subrepresentations of 𝑉. It is maximal among semisimple subrepresen-
tations of 𝑉 in the partial order given by containment.

2.2.2 Direct sums and Hom-spaces

Now suppose that 𝑉 = 𝑉1 ⊕ 𝑉2 is a k-vector space. For 𝑗 = 1, 2, we have natural inclusion maps 𝜄𝑗 ∶ 𝑉𝑗 →
𝑉 and projection maps 𝑝𝑗 ∶ 𝑉 → 𝑉𝑗 (with kernel 𝑉3−𝑗 ). We claim, for any vector space 𝑈, we have natural
isomorphisms

𝑖) ℒ(𝑉, 𝑈) ≅ ℒ(𝑉1 , 𝑈) ⊕ ℒ(𝑉2 , 𝑈)

(2.2.1)
𝑖𝑖) ℒ(𝑈, 𝑉) ≅ ℒ(𝑈, 𝑉1 ) ⊕ ℒ(𝑈, 𝑉2 ).

In the case of 𝑖), the map simply takes the restriction of 𝜙 ∈ ℒ(𝑉, 𝑈) to 𝑉1 and 𝑉2 respectively. In
terms of our inclusion and projection maps, for 𝑟 = 1, 2 we have 𝜙|𝑉𝑟 = 𝜙 ∘ 𝜄𝑟 . To see that this map is an
isomorphism, note that any 𝑣 ∈ 𝑉 can be written uniquely as 𝑣 = 𝑣1 + 𝑣2 with 𝑣1 ∈ 𝑉1 and 𝑣2 ∈ 𝑉2 .

12
 2
Indeed 𝑣𝑟 = 𝜄𝑟 ∘ 𝑝𝑟 (𝑣), hence 𝜙(𝑣) = 𝜙(𝜄1 ∘ 𝑝1 (𝑣) + 𝜄2 ∘ 𝑝2 (𝑣)) = ∑𝑟=1 (𝜙 ∘ 𝜄𝑟 ) ∘ 𝑝𝑟 (𝑣). In other words, the
2
inverse to 𝜙 ↦ (𝜙 ∘ 𝜄𝑟 )𝑟=1,2 is the map (𝜓𝑟 )𝑟=1,2 ↦ ∑ 𝜓𝑟 𝑝𝑟 . 𝑟=1
In the case of 𝑖𝑖), the morphism simply takes, for 𝜙 ∈ ℒ(𝑈, 𝑉), the components of 𝜙(𝑢) in 𝑉1 and 𝑉2
respectively, that is 𝜙 ↦ (𝑝𝑠 ∘ 𝜙)𝑠=1;2 . Clearly the inverse of this map is given by (𝜂𝑠 )𝑠=1,2 ↦ ∑𝑠=1,2 𝜄𝑠 ∘ 𝜂𝑠 .
Now consider End(𝑉) = ℒ(𝑉, 𝑉) where 𝑉 = 𝑉1 ⊕ 𝑉2 . We may use 𝑖) and 𝑖𝑖) (twice) to obtain
2 2
⎛ 2 ⎞ 2
⎜⎜ ⎟⎟
ℒ(𝑉, 𝑉) ≅ � ℒ(𝑉𝑟 , 𝑉) ≅ � ⎜⎜⎝� ℒ(𝑉𝑟 , 𝑉𝑠 )⎟⎟⎠ = � ℒ(𝑉𝑟 , 𝑉𝑠 )
𝑟=1 𝑟=1 𝑠=1 𝑟,𝑠=1

by 𝜙 ↦ (𝜙𝑠𝑟 ) where 𝜙𝑠𝑟 = 𝑝𝑠 ∘ 𝜙 ∘ 𝜄𝑟 ∈ ℒ(𝑉𝑟 , 𝑉𝑠 ). This decomposition is just that of a matrix into block
submatrices, so it can be useful to arrange it in that form:

𝜙11 𝜙12 Homk (𝑉1 , 𝑉1 ) = 𝔤𝔩𝑉1 Homk (𝑉2 , 𝑉1 )

End(𝑉) = 𝔤𝔩𝑉 ∋ 𝜙 ↦ � ∈
𝜙21 𝜙22 � � Homk (𝑉2 , 𝑉2 ) = 𝔤𝔩𝑉2 �
(2.2.2)
Homk (𝑉1 , 𝑉2 )

This shows that 𝔤𝔩𝑉1 ⊕ 𝔤𝔩𝑉2 is naturally isomorphic to a subalgebra of 𝔤𝔩𝑉 = 𝔤𝔩𝑉1⊕𝑉2 , and hence
𝑉1 ⊕ 𝑉2 is a representation of 𝔤𝔩𝑉1 ⊕ 𝔤𝔩𝑉2 . More interestingly, the summands ℒ(𝑉1 , 𝑉2 ) and ℒ(𝑉2 , 𝑉1 )
are clearly all preserved by the action of 𝔤𝔩𝑉1 ⊕ 𝔤𝔩𝑉2 , so that in particular, ℒ(𝑉1 , 𝑉2 ) is a representation
of 𝔤𝔩𝑉1 ⊕ 𝔤𝔩𝑉2 , where if 𝑥 = (𝑥1 , 𝑥2 ) ∈ 𝔤𝔩𝑉1 ⊕ 𝔤𝔩𝑉2 and 𝜙 ∈ ℒ(𝑉1 , 𝑉2 ), then 𝑥(𝜙) = 𝑥2 ∘ 𝜙 − 𝜙 ∘ 𝑥1 .
It follows that if (𝑉1 , 𝜌1 ) and (𝑉2 , 𝜌2 ) are 𝔤-representations, then the composition
Δ 𝜌1 ⊕𝜌2
𝔤 /𝔤⊕𝔤 / 𝔤𝔩 ⊕ 𝔤𝔩 / 𝔤𝔩(ℒ(𝑉 , 𝑉 ))
𝑉1 𝑉2 1 2

gives ℒ(𝑉1 , 𝑉2 ) the structure of a 𝔤-representation, where Δ ∶ 𝔤 → 𝔤 ⊕ 𝔤 is the diagonal map Δ(𝑥) =
(𝑥, 𝑥). Explicitly, if 𝑥 ∈ 𝔤 and 𝜙 ∈ ℒ(𝑉, 𝑊) then 𝑥(𝜙) = 𝜌2 (𝑥) ∘ 𝜙 − 𝜙 ∘ 𝜌1 (𝑥).
Remark 2.2.16. Note that the previous example actually includes that of dual spaces: if 𝑉 is a k-vector
space then 𝑉 ∗ = ℒ(𝑉, k) becomes a 𝔤𝔩𝑉 ⊕𝔤𝔩1 -representation, and then simply using the inclusion 𝔤𝔩𝑉 →
𝔤𝔩𝑉 ⊕ 𝔤𝔩1 we see that 𝑉 ∗ becomes a 𝔤𝔩𝑉 -representation, and indeed we obtain the same action: for any
𝑓 ∈ 𝑉 ∗ and 𝑥 ∈ 𝔤𝔩𝑉 we have 𝑥(𝑓) ≔ (𝑥, 0)(𝑓) = 0 ∘ 𝑓 − 𝑓 ∘ 𝑥 = −𝑥⊺ (𝑓).

2.3 Tensor products

Just as with group representations, tensor products of 𝔤-representations can be given the structure of a
𝔤-representation. The key for this is the following general Lemma:
Lemma 2.3.1. Let 𝔤1 , 𝔤2 be Lie algebras over k and let 𝔤 = 𝔤1 ⊕ 𝔤2 be their direct sum (so each of 𝔤1 , 𝔤2 is an
ideal in 𝔤). If (𝑈, 𝜌) is a representation of 𝔤, and we set 𝜌𝑖 = 𝜌|𝔤𝑖 , then each 𝜌𝑖 is a representation of 𝔤𝑖 for 𝑖 = 1, 2,
and [𝜌1 (𝑥), 𝜌2 (𝑦)] = 0 for any 𝑥 ∈ 𝔤1 , 𝑦 ∈ 𝔤2 . Conversely, if 𝜌𝑖 ∶ 𝔤𝑖 → 𝔤𝔩𝑈 are Lie algebra homomorphisms
for 𝑖 = 1, 2 and [𝜌1 (𝑥), 𝜌2 (𝑦)] = 0 for all 𝑥 ∈ 𝔤1 , 𝑦 ∈ 𝔤2 , then 𝜌(𝑥, 𝑦) = 𝜌1 (𝑥) + 𝜌2 (𝑦) is a Lie algebra
homomorphism from 𝔤 to 𝔤𝔩𝑈 .

Proof. Given a representation (𝑈, 𝜌) of 𝔤, the asserted properties of 𝜌1 , 𝜌2 are immediate. For the con-
verse, note that if (𝑥1 .𝑥2 ), (𝑦1 , 𝑦2 ) ∈ 𝔤, where 𝑥1 .𝑦1 ∈ 𝔤1 and 𝑥2 , 𝑦2 ∈ 𝔤2 , then
[𝜌(𝑥1 , 𝑥2 ), 𝜌(𝑦1 , 𝑦2 )] = [𝜌1 (𝑥1 ) + 𝜌2 (𝑥2 ), 𝜌1 (𝑦1 ) + 𝜌2 (𝑦2 )]
= [𝜌1 (𝑥1 ), 𝜌1 (𝑦1 )] + [𝜌1 (𝑥1 ), 𝜌2 (𝑦2 )] − [𝜌1 (𝑦1 ), 𝜌2 (𝑥2 )] + [𝜌2 (𝑥2 ), 𝜌2 (𝑦2 )]
= [𝜌1 (𝑥1 ), 𝜌1 (𝑦1 )] + [𝜌2 (𝑥2 ), 𝜌2 (𝑦2 )]
= 𝜌((𝑥1 , 𝑥2 ), (𝑦1 , 𝑦2 ))
so that 𝜌 is a homomorphism as required. 

13
 Now 𝔤𝔩𝑉 and 𝔤𝔩𝑊 are naturally subalgebras of 𝔤𝔩𝑉⊗𝑊 , via the embeddings 𝑖𝑉 and 𝑖𝑊 respectively, where
𝑖𝑉 (𝛼) = 𝛼 ⊗ 1𝑊 and 𝑖𝑊 (𝛽) = 1𝑉 ⊗ 𝛽 respectively. Since for any 𝛼 ∈ 𝔤𝔩𝑉 , 𝛽 ∈ 𝔤𝔩𝑊 we have
𝑖𝑊 (𝛽) ∘ 𝑖𝑉 (𝛼) = (1𝑉 ⊗ 𝛽) ∘ (𝛼 ⊗ 1𝑊 ) = 𝛼 ⊗ 𝛽 = (𝛼 ⊗ 1𝑊 ) ∘ (1𝑉 ⊗ 𝛽) = 𝑖𝑉 (𝛼) ∘ 𝑖𝑊 (𝛽)
it follows by Lemma 2.3.1 that 𝑑 ∶ 𝔤𝔩𝑉 ⊕ 𝔤𝔩𝑊 → 𝔤𝔩𝑉⊗𝑊 given by
𝑑(𝑥, 𝑦) = 𝑖𝑉 (𝑥) + 𝑖𝑊 (𝑦) = 𝑥 ⊗ 1𝑊 + 1𝑉 ⊗ 𝑦
is a Lie algebra homomorphism, and hence 𝑉 ⊗ 𝑊 is naturally a 𝔤𝔩𝑉 ⊕ 𝔤𝔩𝑊 -representation. It follows
immediately that if (𝑉, 𝜌) is a representation of 𝔤1 and (𝑊, 𝜎) is a representation of 𝔤2 then 𝑉 ⊗ 𝑊 is a
representation of 𝔤1 ⊕ 𝔤2 via 𝑑 ∘ (𝜌 ⊕ 𝜎) and if 𝔤1 = 𝔤2 = 𝔤 then 𝑉 ⊗ 𝑊 is a representation of 𝔤 via 𝑑 ∘ Δ,
where Δ(𝑥) = (𝑥, 𝑥) ∈ 𝔤 ⊕ 𝔤. More explicitly, if (𝑉, 𝜌) and (𝑊, 𝜎) are 𝔤-representations then 𝑉 ⊗ 𝑊 is a
𝔤-representation with
𝑥(𝑣 ⊗ 𝑤) = 𝜌(𝑥) ⊗ 𝑤 + 𝑣 ⊗ 𝜎(𝑤), ∀ 𝑣 ∈ 𝑉, 𝑤 ∈ 𝑊. (2.3.1)

2.3.1 Tensoring with one-dimensional representations

If (𝑉, 𝜌) is any 𝔤-representation, then by Example II.7, we have an isomorphism of vector spaces 𝑉 ⊗k𝜆 →
𝑉 given by the map 𝑣 ⊗ 𝜆 ↦ 𝜆.𝑣. Via this map, one can think of the 𝔤-representation 𝑉 ⊗ k𝜆 as the same
vector space 𝑉 but now equipped with a new action 𝜌𝜆 of 𝔤, where 𝜌𝜆 (𝑥) = 𝜌(𝑥)+𝜆(𝑥).𝐼𝑉 (where we write
𝐼𝑉 for the identity map.) Note that, in particular, if 𝜆, 𝜇 ∈ 𝐷(𝔤)0 then this shows that k𝜆 ⊗ k𝜇 ≅ k𝜆+𝜇 .

2.3.2 Homomorphisms, 𝔤-homomorphism, and tensor products

The properties asserted of the maps described in this section are proved in detail in Appendix II.
Examining the formula (2.3.1) for the action on a tensor product of representations given above we see
that, just as for group representations, if 𝑉 and 𝑊 are 𝔤-representations, then the isomorphism 𝜎 ∶ 𝑉 ⊗
𝑊 → 𝑊 ⊗ 𝑉 given by 𝜎(𝑣 ⊗ 𝑤) = 𝑤 ⊗ 𝑣, (𝑣 ∈ 𝑉, 𝑤 ∈ 𝑊) is compatible with the action of 𝔤 and hence
induces an isomorphism of 𝔤-representations. In the case 𝑉 = 𝑊, 𝜎 becomes an involution on 𝑉 ⊗ 𝑉
commuting with the 𝔤-action. In other words, 𝑆2 , the symmetric group on two letters acts on 𝑉 ⊗ 𝑉 and
the isotypic decomposition of 𝑉 ⊗ 𝑉 under this action, (equivalently the (+1)- and (−1)-eigenspaces of 𝜎)
shows that 𝑉 ⊗ 𝑉 is the direct sum of the subrepresentations of symmetric tensors and skew-symmetric
tensors, that is 𝑉 ⊗ 𝑉 = Sym2 (𝑉) ⊕ Alt2 (𝑉) where
1
Sym2 (𝑉) = spank � (𝑣1 ⊗ 𝑣2 + 𝑣2 ⊗ 𝑣1 ) ∶ 𝑣1 , 𝑣2 ∈ 𝑉� ,
2
1
Alt2 (𝑉) = spank � (𝑣1 ⊗ 𝑣2 − 𝑣2 ⊗ 𝑣1 ) ∶ 𝑣1 , 𝑣2 ∈ 𝑉� .
2

Let 𝑉 and 𝑊 be k-vector spaces. There is a natural linear map 𝜃 ∶ 𝑉 ∗ ⊗ 𝑊 → Hom(𝑉, 𝑊), given by
𝜃(𝑓 ⊗ 𝑤) = 𝑓.𝑤 where (𝑓.𝑤)(𝑣) = 𝑓(𝑣).𝑤 for all 𝑣 ∈ 𝑉, 𝑓 ∈ 𝑉 ∗ and 𝑤 ∈ 𝑊. This map is injective, and its
image is precisely the space of finite-rank linear maps3 from 𝑉 to 𝑊. In particular, if dim(𝑉) < ∞ then
we have End(𝑉) ≅ 𝑉 ∗ ⊗ 𝑉. Similarly, there is a natural map 𝑚 ∶ 𝑉 ∗ ⊗ 𝑊 ∗ → (𝑉 ⊗ 𝑊)∗ , where
𝑚(𝑓 ⊗ 𝑔)(𝑣 ⊗ 𝑤) = 𝑓(𝑣).𝑔(𝑤), ∀ 𝑣 ∈ 𝑉, 𝑤 ∈ 𝑊, 𝑓 ∈ 𝑉 ∗ , 𝑔 ∈ 𝑊 ∗ .

The map 𝑚 is also injective and hence, by considering dimensions, it is an isomorphism when 𝑉 and 𝑊 are
finite-dimensional. This tensor product description of End(𝑉) = Hom(𝑉, 𝑉) gives a natural description
of the trace map: Notice that we have a natural bilinear map 𝑉 ∗ × 𝑉 → k given by (𝑓, 𝑣) ↦ 𝑓(𝑣).
By the universal property of the tensor product, this induces a linear map 𝜄 ∶ 𝑉 ∗ ⊗ 𝑉 → k. Under the
identification with Hom(𝑉, 𝑉) this map is identified with the trace of a linear map.
3
That is, the linear maps from 𝑉 to 𝑊 which have finite-dimensional image.

14
Remark 2.3.2. It is worth noticing that this gives a coordinate-free way of defining the trace, and also
some explanation for why one needs some finiteness condition in order for the trace to be defined.

Remark 2.3.3. If 𝔤 is a Lie algebra and 𝑉 and 𝑊 are 𝔤-representations, then it is also easy to check from
the definitions that the natural map 𝜃 ∶ 𝑉 ∗ ⊗ 𝑊 → Hom(𝑉, 𝑊) defined in Lemma II.10 is also a map of
𝔤-representations, as is the contraction map 𝜄 ∶ 𝑉 ∗ ⊗ 𝑉 → k, where we view k as the trivial representation
of 𝔤. For example, for 𝜄 we have:

𝜄(𝑥(𝑓 ⊗ 𝑣)) = 𝜄�𝑥(𝑓) ⊗ 𝑣 + 𝑓 ⊗ 𝑥(𝑣)� = −𝑓(𝑥(𝑣)) + 𝑓(𝑥(𝑣)) = 0, ∀ 𝑥 ∈ 𝔤, 𝑣 ∈ 𝑉, 𝑓 ∈ 𝑉 ∗ .

Thus all the maps between tensor products of vector spaces discuss in Appendix II yield maps of 𝔤-
representations.

The following example will be very useful in a number of places later in the course.

Example 2.3.4. If 𝔤 is a Lie algebra and (𝑉, 𝜌) is a 𝔤-representation, then 𝜌 induces a natural bilinear map
𝑎𝜌 ∶ 𝔤 × 𝑉 → 𝑉, namely (𝑥, 𝑣) ↦ 𝜌(𝑥)(𝑣). By the universal property of tensor products this yields a linear
map 𝑎𝜌̃ ∶ 𝔤 ⊗ 𝑉 → 𝑉. We claim this map is a homomorphism of 𝔤 representations (where 𝔤 is viewed
as the adjoint representation). To see this, first notice that the bilinear map 𝑎𝜌 ∶ 𝔤 × 𝑉 → 𝑉 is equal to
𝑎𝑉 ∘ (𝜌 × 1𝑉 ) where 𝑎𝑉 ∶ 𝔤𝔩𝑉 × 𝑉 → 𝑉 is the natural action of 𝔤𝔩𝑉 on 𝑉, (𝜙, 𝑣) ↦ 𝜙(𝑣). Thus it suffices to
check the claim for 𝔤𝔩𝑉 and its vector representation 𝑉. Let 𝑎𝑉̃ ∶ 𝔤𝔩𝑉 ⊗ 𝑉 → 𝑉 be the linear map induced
by 𝑎𝑉 . Then if 𝑥, 𝑦 ∈ 𝔤𝔩𝑉 and 𝑣 ∈ 𝑉 we have 𝑥(𝑎𝑉̃ (𝑦 ⊗ 𝑣)) = 𝑥(𝑦(𝑣)), while

𝑎𝑉̃ (𝑥(𝑦 ⊗ 𝑣)) = [𝑥, 𝑦] ⊗ 𝑣 + 𝑦 ⊗ 𝑥(𝑣) = (𝑥𝑦 − 𝑦𝑥)(𝑣) + 𝑦𝑥(𝑣) = 𝑥𝑦(𝑣) = 𝑥(𝑎𝑉̃ (𝑦 ⊗ 𝑣)) (2.3.2)

hence as 𝑦 ⊗ 𝑣 was arbitrary we have 𝑎𝑉̃ ∘ 𝑥 = 𝑥 ∘ 𝑎𝑉̃ for all 𝑥 ∈ 𝔤𝔩𝑉 so that 𝑎𝑉̃ ∈ ℒ𝔤𝔩𝑉 (𝔤𝔩𝑉 ⊗ 𝑉, 𝑉) as
required.

Remark 2.3.5. In fact one can also deduce that 𝑎𝜌̃ is a homomorphism of 𝔤-representations by observing
that under the identification 𝔤𝔩𝑉 ≅ 𝑉 ∗ ⊗𝑉 the map 𝑎𝜌̃ corresponds to the linear map 𝜄13 ∶ 𝑉 ∗ ⊗𝑉⊗𝑉 → 𝑉,
that is, the linear map which is the identity on the 2nd tensor factor and is the contraction map 𝜄 on the
first and third factors:4 so that 𝜄(𝑓1 ⊗ 𝑣2 ⊗ 𝑣3 ) = 𝑓1 (𝑣3 ).𝑣2 where 𝑣2 , 𝑣3 ∈ 𝑉, 𝑓1 ∈ 𝑉 ∗ . Since 𝜄 is a
homomorphism of 𝔤-representations, it follows 𝑎𝜌̃ is also.

4
Here, as usual, we are also identifying k ⊗ 𝑉 with 𝑉 equipped with the scalar multiplication map, 𝑠 ∶ k × 𝑉 → 𝑉, that is
𝑠(𝜆, 𝑣) = 𝜆.𝑣. (Recall that a tensor product 𝑉 ⊗ 𝑊 is a vector space and a bilinear map 𝑉 × 𝑊 → 𝑉 ⊗ 𝑊.)

15
Chapter 3

Classifying Lie algebras

The goal of this course is to study the structure of Lie algebras, and attempt to classify them. The most
ambitious “classification” result would be to give a description of all finite-dimensional Lie algebras up to
isomorphism. In very low dimensions this is actually possible: For dimension 1 clearly there is a unique
(up to isomorphism) Lie algebra since the alternating condition demands that the bracket is zero. In di-
mension two, one can again have an abelian Lie algebra, but there is another possibility: if 𝔤 has a basis
{𝑒, 𝑓} then we may set [𝑒, 𝑓] = 𝑓, and this completely determines the Lie algebra structure. All two-
dimensional Lie algebras which are not abelian are isomorphic to this one (check this). It is also possible
to classify three-dimensional Lie algebras, but it becomes rapidly intractable to do this in general as the
dimension increases.
This reveals an essential tension in seeking any kind of classification result for mathematical objects:
a classification result should describe all such objects (or at least those in a natural, and likely reason-
ably “large” class) up to some notion of equivalence. Clearly, using a stricter notion of equivalence will
mean any classification theorem you can prove will provide finer information about the objects you are
studying, but this must be balanced against the intrinsic complexity of the objects which may make such
a classification (even for quite small classes) extremely complicated. Hence it is likely reasonable to ac-
cept a somewhat crude notion of equivalence in order to be have any chance of obtaining a classification
theorem which has a relatively simple statement.

3.1 Classification by composition factors

The proofs in this section are non-examinable, and may not be discussed in detail in lectures. They are included
here for completemess.
Our approach will follow the strategy often used in finite groups: In that context, the famous Jordan-
Hölder theorem shows that any finite group can be given by gluing together finite simple groups, in the
sense that we may find an decreasing chain of subgroups

𝐺 = 𝐺0 B 𝐺1 B … 𝐺𝑛−1 B 𝐺𝑛 = {𝑒},

where, for each 𝑖, (1 ≤ 𝑖 ≤ 𝑛), the subgroup 𝐺𝑖 is a normal in 𝐺𝑖−1 and 𝑆𝑖 = 𝐺𝑖−1 /𝐺𝑖 is simple. That
such a filtration of 𝐺 exists is easy to prove by induction. The non-trivial part of the theorem is that, for
any fixed finite simple group 𝐻, the number of 𝑆𝑖 which are isomorphic to 𝐻 is independent of the choice
filtration. This is usually phrased as saying that the multiplicity with which a composition factor 𝑆𝑖 occurs
in the sequence {𝐺𝑖−1 /𝐺𝑖 ∶ 1 ≤ 𝑖 ≤ 𝑛} is well-defined.
One can thus give a somewhat crude classification of finite groups, where one considers two finite
groups to be equivalent if they have the same composition factors, by giving a classification of finite simple
groups. But even the question of classifying finite simple groups is not at all obviously tractable, and

16
answering it was one of the spectacular mathematical achievements of the second half of the twentieth
century.
For Lie algebras, we can attempt something similar. In fact, it turns out that, at least in characteristic
zero, we obtain a far more complete answer about the structure of an arbitrary finite-dimensional Lie
algebra than one could hope to obtain in a Part C course on finite group theory. One aspect of this finer
information will reveal a sharp distinction between 𝔤𝔩1 and the non-abelian Lie algebras which have no
proper ideals, which is one reason for the following definition:
Definition 3.1.1. A nonzero Lie algebra 𝔤 is said to be almost simple1 if it has no non-trivial proper ideals. If
𝔤 is almost simple and dim(𝔤) > 1 then we say that 𝔤 is simple. Equivalently, an almost-simple Lie algebra
is simple if it is non-abelian. Thus the only almost-simple Lie algebra which is not simple is 𝔤𝔩1 .
Definition 3.1.2. A flag or 𝔤-flag in a finite-dimensional Lie algebra 𝔤 is a nested chain

𝒞 = (𝔤 = 𝔤0 D 𝔤1 D … D 𝔤𝑑 = 0)

of subalgebras where, for 0 ≤ 𝑖 ≤ 𝑑 − 1, the subalgebra 𝔤𝑖+1 is an ideal in 𝔤𝑖 and the quotient 𝔠𝑘 = 𝔤𝑘 /𝔤𝑘+1
is called a factor of the flag. A flag is strict if the containments 𝔤𝑘+1 ⊆ 𝔤𝑘 are all proper, or equivalently, if
the factors 𝔠𝑘 are all nonzero. A flag 𝒞 ′ = (𝔥𝑘 )𝑙𝑘=0 is said to be a refinement of 𝒞 if, for each 𝑘 ∈ {1, … , 𝑑 − 1}
there exists 𝑗 ∈ {1, 2, … , 𝑙 − 1} such that 𝔤𝑘 = 𝔥𝑗 . Any flag has an associated strict flag which is obtained
by omitting repetitions, that is, if 𝑆 = {𝑘 ∈ {0, 1, … , 𝑑} ∶ 𝔤𝑘 ⊋ 𝔤𝑘+1 } and 0 ≤ 𝑖0 < 𝑖1 < … < 𝑖𝑡 ≤ 𝑑 are the
elements of 𝑆 in increasing order, then the strict flag associated to 𝒞 is 𝒞 ̃ = (𝔤𝑖𝑠 )𝑡𝑠=0 .
If, for each 𝑘 ∈ {0, 1, … , 𝑑 − 1}, 𝔠𝑘 is almost simple, then 𝒞 is a composition series and in this case the
factors 𝔤𝑖 /𝔤𝑖+1 are then called the composition factors of the composition series. If 𝔰 is an almost simple Lie
algebra, we define the multiplicity of 𝔰 in the composition series 𝒞 to be

𝑚𝒞 (𝔰) ∶= #�𝑖 ∈ {1, … 𝑟} ∶ 𝔰 ≅ 𝔤𝑖−1 /𝔤𝑖 �

It is easy to check by induction on dimension that any strict flag in a Lie algebra can be refined to
a composition series (one needs to use isomorphism theorems in the same manner as the proof in the
context of finite groups). In particular, all finite-dimensional Lie algebras possess composition series.
Proposition 3.1.3. Suppose that 𝔤 has a composition series 𝒞 = (𝔤 = 𝔤0 B 𝔤1 B … B 𝔤𝑛 = 0).

i) Let 𝔞 is an ideal in 𝔤, and let 𝑞 ∶ 𝔤 → 𝔤/𝔞 be the quotient map. The strict flag 𝒞 ′ associated to the 𝔞-flag
(𝔞 ∩ 𝔤𝑘 )𝑑𝑘=0 and the strict flag 𝒞 ″ associated to the 𝔤/𝔞-flag (𝑞(𝔤𝑘 ))𝑑𝑘=0 are composition series for 𝔞 and 𝔤/𝔞
respectively, and for any almost simple Lie algebra 𝔰 we have 𝑚𝒞 = 𝑚𝒞 ′ + 𝑚𝒞 ″ .

ii) If 𝔰 is an almost simple Lie algebra, the multiplicities 𝑚𝒞 are independent of the choice of composition series
and we set [𝔰 ∶ 𝔤] = 𝑚𝒞 (𝔰) for any composition series 𝒞 of 𝔤.

Proof. Let 𝔞𝑖 ∶= 𝔞 ∩ 𝔤𝑖 (0 ≤ 𝑖 ≤ 𝑛), and let 𝑝𝑖+1 ∶ 𝔤𝑖 → 𝔤𝑖 /𝔤𝑖+1 denote the quotient map from 𝔤𝑖 to 𝔤𝑖 /𝔤𝑖+1 .
Now the terms of the sequence (𝔞𝑖 )𝑛𝑖=0 , while nested, need not be strictly decreasing. Since 𝔞 is an ideal in
𝔤, the intersection 𝔞𝑖 = 𝔞 ∩ 𝔤𝑖 is an ideal in 𝔤𝑖 and, by the second isomorphism theorem, its image under
the quotient map 𝑝𝑖+1 ∶ 𝔤𝑖 → 𝔤𝑖 /𝔤𝑖+1 is

𝔞𝑖 /𝔞𝑖+1 ≅ (𝔞𝑖 + 𝔤𝑖+1 )/𝔤𝑖+1 = 𝑝𝑖+1 (𝔞𝑖 ) (3.1.1)

Similarly, we may consider the sequence (𝑞(𝔤𝑖 ))𝑛𝑖=0 , where 𝑞 ∶ 𝔤 → 𝔤/𝔞 is the quotient map, then 𝑞(𝔤𝑖+1 ) is
an ideal in 𝑞(𝔤𝑖 ), and by the second isomorphism theorem 𝑞(𝔤𝑖 ) ≅ 𝔤𝑖 /𝔤𝑖 ∩ 𝔞 = 𝔤𝑖 /𝔞𝑖 . Under this identifica-
tion, 𝑞(𝔤𝑖+1 ) is isomorphic to (𝔤𝑖+1 + 𝔞𝑖 )/𝔞𝑖 , and hence

𝑞(𝔤𝑖 )/𝑞(𝔤𝑖+1 ) ≅ (𝔤𝑖 /𝔞𝑖 )�(𝔤𝑖+1 + 𝔞𝑖 )/𝔞𝑖 ≅ 𝔤𝑖 /(𝔤𝑖+1 + 𝔞𝑖 ). (3.1.2)

1
This is not standard terminology, but it is convenient to use here.

17
Thus since 𝔤𝑖 /𝔤𝑖+1 is almost simple and 𝔤𝑖+1 ⊆ 𝔤𝑖+1 + 𝔞𝑖 ⊆ 𝔤𝑖 , we must either have 𝔤𝑖+1 + 𝔞𝑖 = 𝔤𝑖 , in which
case Equations (3.1.1) and (3.1.2) show that 𝑞(𝔤𝑖 ) = 𝑞(𝔤𝑖+1 ) and 𝔞𝑖 /𝔞𝑖+1 ≅ 𝔤𝑖 /𝔤𝑖+1 , or 𝔤𝑖+1 + 𝔞𝑖 = 𝔤𝑖+1 , in
which case 𝔞𝑖 = 𝔞𝑖+1 and 𝑞(𝔤𝑖 )/𝑞(𝔤𝑖+1 ) ≅ 𝔤𝑖 /𝔤𝑖+1 .
Thus removing repetitions from the sequences (𝔞𝑖 ) and (𝑞(𝔤𝑖 )) yields composition series 𝒞 ′ and 𝒞 ″
for 𝔞 and 𝔤/𝔞 respectively, where the composition factors of 𝒞 correspond to a composition factor of pre-
cisely one of 𝒞 ′ or 𝒞 ″ .
For the final part, we use induction on the minimal length ℓ(𝔤) of a composition series for 𝔤. Suppose
that 𝒞 = (𝔤0 B 𝔤1 B … B 𝔤𝑑 = {0}) is a composition series for 𝔤 with 𝑑 = ℓ(𝔤). Let 𝒟 = (𝔤 = 𝔥0 B
𝔥1 … B 𝔥𝑙 = {0}) be another composition series for 𝔤. Then we may apply the previous part to 𝒞 and 𝒟
with 𝔞 = 𝔤1 to obtain composition series 𝒞 ′ , 𝒟 ′ of 𝔤1 and 𝒞 ″ , 𝒟 ″ of 𝔤/𝔤1 . Now 𝒞 ′ = (𝔤𝑘 )𝑑𝑘=1 and since
any composition series of 𝔤1 of length 𝑟 immediately yields one for 𝔤 with length 𝑟 + 1, we must have
ℓ(𝔤1 ) = 𝑑 − 1, and hence by induction 𝑚𝒞 ′ = 𝑚𝒟 ′ . Moreover, since 𝔤/𝔤1 is almost simple, it has a unique
multiplicity function, so that 𝑚𝒞 ″ = 𝑚𝒟 ″ . The result then follows immediately from part 𝑖). 

Remark 3.1.4. Later in this course, we will, at least in characteristic zero, establish a finer description of
the structure of semisimple Lie algebras that the above result provides: we will show that any Lie algebra 𝔤
contains an ideal 𝔯 whose only composition factor is 𝔤𝔩1 (and therefore the multiplicity of that composition
factor is just dim(𝔯) while the quotient, 𝔤/𝔯 is a direct sum of non-abelian simple Lie algebras.

3.2 Exact sequences of Lie algebras

Definition 3.2.1. We say that the sequence of Lie algebras and Lie algebra homomorphisms
𝑖 𝑞
𝔤1 /𝔤 /𝔤
2

is exact at 𝔤 if im(𝑖) = ker(𝑞). A sequence of maps

𝑖 𝑞
0 /𝔤 /𝔤 /𝔤 /0
1 2

is called a short exact sequence if it is exact at each of 𝔤1 , 𝔤 and 𝔤2 , so that 𝑖 is injective, 𝑞 is surjective and
im(𝑖) = ker(𝑞). In this case, we say that 𝔤 is an extension of 𝔤2 by 𝔤1 . The existence of a composition series
for a finite-dimensional Lie algebra shows that any such Lie algebra is constructed through successive
extensions by almost simple Lie algebras.

Definition 3.2.2. An extension of Lie algebras

𝑖 𝑞
0 /𝔤 /𝔤 /𝔤 /0
1 2

is said to be split if there is a homomorphism of Lie algebras 𝑠 ∶ 𝔤2 → 𝔤 such that 𝑞 ∘ 𝑠 = id𝔤2 .

Notice that in this case the image 𝑠(𝔤2 ) of the splitting map 𝑠 is a subalgebra of 𝔤 which is isomorphic
to 𝔤2 and is complementary to 𝑖(𝔤1 ), in the sense that 𝔤 = 𝑖(𝔤1 ) ⊕ 𝑠(𝔤2 ) as vector spaces. Indeed the
homomorphism 𝑠 is determined by 𝑠(𝔤2 ) its image, because it is the inverse of 𝑞|𝑠(𝔤) , the restriction of 𝑞 to
that image. Moreover, since 𝑖(𝔤1 ) is an ideal of 𝔤, the adjoint action of 𝔤 preserves 𝑖(𝔤1 ), and so it restricts
to give an action of 𝑠(𝔤2 ) on 𝑖(𝔤1 ). This completely describes the Lie bracket on 𝔤: For any 𝑥, 𝑦 ∈ 𝔤, there
are unique 𝑥1 , 𝑦1 ∈ 𝔤1 and 𝑥2 , 𝑦2 ∈ 𝔤2 such that 𝑥 = 𝑖(𝑥1 ) + 𝑠(𝑥2 ), 𝑦 = 𝑖(𝑦1 ) + 𝑠(𝑦2 ). Then

[𝑥, 𝑦] = [𝑖(𝑥1 ) + 𝑠(𝑥2 ), 𝑖(𝑦1 ) + 𝑠(𝑦2 )]

= 𝑖([𝑥1 , 𝑥2 ]) + ad(𝑠(𝑥2 ))(𝑖(𝑦1 )) − ad(𝑠(𝑦2 ))(𝑖(𝑥1 )) + 𝑠([𝑥2 , 𝑦2 ]).

This motivates the following definition:

Definition 3.2.3. Suppose that 𝔤, 𝔥 are Lie algebras, and we have a homomorphism 𝜙 ∶ 𝔤 → Derk (𝔥), the
Lie algebra of derivations2 on 𝔥. Then it is straight-forward to check that we can form a new Lie algebra
2
Recall that the derivations of a Lie algebra are the linear maps 𝛼 ∶ 𝔥 → 𝔥 such that 𝛼([𝑥, 𝑦]) = [𝛼(𝑥), 𝑦] + [𝑥, 𝛼(𝑦)].

18
𝔥⋊𝔤, the semi-direct product 3 of 𝔤 and 𝔥 by 𝜙 which as a vector space is just 𝔤⊕𝔥, and where the Lie bracket
is given by:
[(𝑥1 , 𝑦1 ), (𝑥2 , 𝑦2 )] = ([𝑥1 , 𝑥2 ] + 𝜙(𝑦1 )(𝑥2 ) − 𝜙(𝑦2 )(𝑥1 ), [𝑦1 , 𝑦2 ]),
where 𝑥1 , 𝑥2 ∈ 𝔥, 𝑦1 , 𝑦2 ∈ 𝔤. The Lie algebra 𝔥, viewed as the subspace {(𝑥, 0) ∶ 𝑥 ∈ 𝔥} of 𝔥 ⋊ 𝔤, is clearly an
ideal of 𝔥⋊𝔤. Since it does not intersect 𝔥, the quotient map 𝑞 ∶ 𝔥⋊𝔤 → (𝔥⋊𝔤)/𝔥 induces an isomorphism
𝔤 → (𝔥 ⋊ 𝔤)/𝔥, hence 𝔥 ⋊ 𝔤 is a split extension of 𝔤 by 𝔥. It is not difficult to check that any split extension
is of this form.

Example 3.2.4. Let 𝔰2 be a 2-dimensional Lie algebra with basis {𝑥, 𝑦}.Clearly 𝐷(𝔰2 ) is spanned by [𝑥, 𝑦].
Thus either 𝐷(𝔰2 ) = 0 and 𝔰2 is abelian, in which case any 1-dimensional subspace is an ideal, or 𝔰2 has a
unique nonzero proper ideal, namely 𝐷(𝔰2 ). For any 1-dimensional ideal 𝐼 in 𝔰2 , if we pick 𝑧 ∈ 𝐼\{0} and
𝑤 ∉ 𝐼, so that {𝑧, 𝑤} is a basis of 𝔰2 , then setting 𝑖(𝜆) = 𝜆.𝑧 and 𝑞(𝜆.𝑧 + 𝜇.𝑤) = 𝜇, we obtain a short exact
sequence:
𝑖 𝑞
0 / 𝔤𝔩 /𝔰 / 𝔤𝔩 /0
1 2 1

Now clearly this sequence is split, because if we define 𝑠 ∶ 𝔤𝔩1 → 𝔰2 by 𝑠(𝜆) = 𝜆.𝑤, for all 𝜆 ∈ k then
𝑞 ∘ 𝑠 = 1𝔤𝔩1 and 𝑠 is a Lie algebra homomorphism by vii) of 1.1.5. It follows that all two-dimensional Lie
algebras can be realised as a semi-direct product of 𝔤𝔩1 with itself.
Now such a semidirect product is given by a homomorphism of Lie algebras 𝜙 ∶ 𝔤𝔩1 → Derk (𝔤𝔩1 ). But
if 𝔞 is an abelian Lie algebra, Derk (𝔞) = 𝔤𝔩𝔞 , since any linear map is a derivation for the zero Lie bracket.
Thus Derk (𝔤𝔩1 ) = 𝔤𝔩𝔤𝔩1 = 𝔤𝔩1 , and 𝜙 ∶ 𝔤𝔩1 → 𝔤𝔩1 is just a scalar. It is then easy to check that if 𝜙 is zero
we obtain the 2-dimensional abelian Lie algebra, while of 𝜙 ≠ 0 we obtain Lie algebras which are all
isomorphic, as can be checked by, for example, rescaling the map 𝑖.

Example 3.2.5. Generalising the previous example, notice that any short exact sequence of the form
𝑖 𝑞
0 /𝔤 /𝔤 / 𝔤𝔩 /0
1 1

is automatically split. Indeed if we pick any 𝑥 ∈ 𝔤 with 𝑞(𝑥) = 1 ∈ 𝔤𝔩1 (k) then setting 𝑠(𝜆) = 𝜆.𝑥 it is
immediate that 𝑞 ∘ 𝑠 = id, and, as in the previous example, 𝑠 ∶ 𝔤𝔩1 → 𝔤 is automatically a homomorphism
of Lie algebras. It follows that 𝔤 is a semidirect product 𝔤1 ⋊ 𝔤𝔩1 (k). It is important to note that, as this
example shows, a splitting map can be far from unique: the set of 𝑥 ∈ 𝔤 with 𝑞(𝑥) ≠ 0 all provide different
splitting maps for 𝑞 in the above short exact sequence.

Remark 3.2.6. There is a close analogy with the notion of a short exact sequence of groups which you
have seen in a previous course: here one has a sequence
𝑖 𝑞
1 /𝐺 /𝐺 /𝐺 /1
1 2

where we write 1 for the trivial group (rather than 0 for the trivial Lie algebra). Exactness at 𝐺 means that
im(𝑖) = ker(𝑞), and similarly at 𝐺1 and 𝐺2 , so that 𝑖 is injective and 𝑞 is surjective. In Part A Groups you
show that this sequence is split, that is, there exists a splitting map 𝑠 ∶ 𝐺2 → 𝐺 such that 𝑞 ∘ 𝑠 = id𝐺2 , if
and only if 𝐺 ≅ 𝐺1 ⋊ 𝐺2 .

Remark 3.2.7. There will be one other kind of extension which will play an important role in this course.
This is an extension
𝑖 𝑞
0 /𝔷 /𝔤 /𝔤 /0
1 2

where the image of 𝔷1 is assumed to lie in 𝔷(𝔤) = {𝑥 ∈ 𝔤 ∶ [𝑥, 𝑦] = 0, ∀ 𝑦 ∈ 𝔤}, the centre of 𝔤. Such
extensions are (unsurprisingly) called central extensions.
Split and central extensions are in a loose sense complementary to each other: An extension of 𝔤2 by
𝔤1 which is both central and split is just the direct sum 𝔤1 ⊕ 𝔤2 , where 𝔤1 ≅ 𝔤𝔩1 ⊕𝑘 and 𝑘 = dimk (𝔤1 ).
3
This is the Lie algebra analogue of the semidirect product of groups, where you build a group 𝐻 ⋊ 𝐺 via a map from 𝐺 to
the automorphisms (rather than derivations) of 𝐻.

19
Example 3.2.8. Let 𝔫3 = {𝑥 = (𝑥𝑖𝑗 ) ∈ 𝔤𝔩3 ∶ 𝑥𝑖𝑗 = 0∀ 𝑖 ≤ 𝑗} be the Lie algebra of strictly upper triangular
3 × 3 matrices. If we set 𝑥 = 𝐸12 , 𝑦 = 𝐸23 and 𝑧 = 𝐸13 then [𝑥, 𝑦] = 𝑧 and 𝑧 is central. Indeed 𝔷(𝔫3 ) = k.𝑧,
and 𝔫3 /𝔷(𝔫3 ) is a 2-dimensional abelian Lie algebra, and thus
𝑖 𝑞
0 / 𝔷(𝔫 ) /𝔫 / 𝔤𝔩 ⊕2 /0
3 3 1

It follows that 𝔫3 is a central extension of 𝔤𝔩1 ⊕2 , the two-dimensional abelian Lie algebra. Note that this
short exact sequence does not split: indeed if 𝑠 ∶ 𝔤𝔩1 ⊕2 → 𝔫3 is a Lie algebra homomorphism and 𝑠(𝑞(𝑥)) =
𝑥1 , 𝑠(𝑞(𝑦)) = 𝑥2 where 𝑥1 = 𝑎1 𝑥 + 𝑏1 𝑦 + 𝑐1 𝑧 and 𝑥2 = 𝑎2 𝑥 + 𝑏2 𝑦 + 𝑐2 𝑧, then [𝑥1 , 𝑥2 ] = [𝑎1 𝑥 + 𝑏1 𝑦 +
𝑎 𝑎
𝑐1 𝑧, 𝑎2 𝑥 + 𝑏2 𝑦 + 𝑐2 𝑧] = (𝑎1 𝑏2 − 𝑎2 𝑏1 )𝑧, so that [𝑥1 , 𝑥2 ] = 0 if and only if the matrix � 1 2 � is singular,
𝑏1 𝑏2
but then 𝑞(im(𝑠)) is at most 1-dimensional, thus 𝑞 ∘ 𝑠 cannot be equal to the identity on 𝔤𝔩1 ⊕2 .
On the other hand, if we let 𝔞 = k.𝑥 ⊕ k.𝑧 then 𝔷(𝔫3 ) ⊂ 𝔞 ⊂ 𝔫3 is complete flag of ideals in 𝔫3 (i.e. a
nested chain of ideals one of each dimension between 1 and dim(𝔫3 )) so that 𝔫3 is an iterated semidirect
product (k.𝑧 ⋊ k.𝑥) ⋊ k.𝑥

20
Chapter 4

Gluing 𝔤𝔩1 : Solvable and nilpotent Lie

algebras

Conventions: From this point onwards in these notes, we will assume that all Lie algebras and all representations
are finite-dimensional over the field k, unless the contrary is explicitly stated, and from §4.3 onwards, k will be
algebraically closed of characteristic zero.

We now begin to study particular classes of Lie algebras. The first class we study, solvable Lie algebras,
in terms of the discussion on classification of Lie algebras in the previous section, can be given as the
class of Lie algebras which can be built using only 𝔤𝔩1 , the simplest Lie algebra1 which possesses only the
structure of the base field k and the trivial Lie bracket.

4.1 Solvable Lie algebras

4.1.1 Definition and basic properties

Definition 4.1.1. A Lie algebra 𝔤 is solvable if its only composition factor is 𝔤𝔩1 (k). This is equivalent to the
condition that 𝔤 has a nested sequence of subalgebras

𝔤 = 𝔤0 ⊋ 𝔤1 ⊋ … ⊋ 𝔤𝑑 = {0},

where 𝔤𝑘+1 is an ideal in 𝔤𝑘 and 𝔤𝑘 /𝔤𝑘+1 is abelian for each 𝑘 (0 ≤ 𝑘 ≤ 𝑑 − 1). Indeed if such a sequence
of subalgebras exists, any refinement of it to a composition series will have 𝔤𝔩1 (k) as its only composition
factor, and conversely, a composition series with 𝔤𝔩1 (k) as its only composition factor is an example of
such a sequence of subalgebras.
If 𝔤 = 𝔤0 ⊃ 𝔤1 ⊃ … ⊃ 𝔤𝑛 = {0} is a composition series for 𝔤 with 𝔤𝑘 /𝔤𝑘+1 ≅ 𝔤𝔩1 for each 𝑘 ∈
{0, 1, … , 𝑛 − 1}, so that dim(𝔤) = 𝑛, then we have 𝔤𝑛−1 ≅ 𝔤𝔩1 , and, for each 𝑘 ∈ {0, 1, … , 𝑛 − 1}, we have a
short exact sequence
𝜄𝑘+1 𝑞𝑘
0 /𝔤 /𝔤 / 𝔤𝔩 /0
𝑘+1 𝑘 1

where 𝜄𝑘+1 is the inclusion map and 𝑞𝑘 the quotient map. Thus 𝔤𝑘−1 is an extension of 𝔤𝔩1 by 𝔤𝑘 . By Re-
mark 3.2.6, this short exact sequence must split, and so 𝔤𝑘 is a semidirect product of 𝔤𝑘+1 by 𝔤𝔩1 (k), and so
solvable Lie algebras are precisely the Lie algebras one obtains from the zero Lie algebra by taking iterated
semidirect products with 𝔤𝔩1 (k).

Example 4.1.2. Example 3.2.4 shows that 𝔰2 , the 2-dimensional non-abelian Lie algebra, is solvable.
1
Hence starting with nothing...

21
 We can rephrase the condition that a Lie algebra 𝔤 is solvable in terms of a decreasing sequence of
ideals in 𝔤:

Definition 4.1.3. Recall from Remark 1.2.5 that the derived subalgebra 𝐷(𝔤) of 𝔤 is defined to be the ideal
[𝔤, 𝔤]. We define the nested decreasing sequence of ideals (𝐷𝑘 (𝔤))𝑘≥0 by setting 𝐷𝑘 (𝔤) = 𝔤 for 𝑘 = 0 and
𝐷𝑘 (𝔤) = 𝐷(𝐷𝑘−1 (𝔤)) = [𝐷𝑘−1 (𝔤), 𝐷𝑘−1 (𝔤)] for 𝑘 > 0. The sequence of ideals (𝐷𝑘 (𝔤))𝑘≥0 is called the
derived series of 𝔤.

Lemma 4.1.4. Let 𝔤 be a Lie algebra. Then 𝐷(𝔤) is the smallest ideal in 𝔤 such that 𝔤/𝐷(𝔤)is abelian. In partic-
ular, 𝔤 is solvable precisely when the derived series (𝐷𝑘 (𝔤))𝑘≥1 satisfies 𝐷𝑘 (𝔤) = 0 for sufficiently large 𝑘.

Proof. For the first claim, suppose that 𝐼 is an ideal for which 𝔤/𝐼 is abelian. Then, for all 𝑥, 𝑦 ∈ 𝔤, we must
have [𝑥, 𝑦] ∈ 𝐼, and hence 𝐷(𝔤) ⊆ 𝐼. Since this also shows 𝔤/𝐷(𝔤) is abelian, the claim follows.
Next note that we have a short exact sequence
0 / 𝐷(𝔤) /𝔤 / 𝔤/𝐷𝔤 /0

that is, 𝔤 is an extension of the abelian Lie algebra 𝔤/𝐷(𝔤) by 𝐷(𝔤). It follows that if 𝐷𝑘 (𝔤) = {0} for
some 𝑘, then 𝔤 has a filtration by ideals for which the subquotients are abelian, so it is certainly solvable.
Conversely, if 𝔤 is solvable, so that we have a nested sequence of subalgebras 𝔤 = 𝔤0 ⊃ 𝔤1 ⊃ … ⊃ 𝔤𝑛 = {0},
where 𝔤𝑘+1 is an ideal in 𝔤𝑘 and 𝔤𝑘 /𝔤𝑘+1 is abelian. But then 𝐷(𝔤𝑘 ) = [𝔤𝑘 , 𝔤𝑘 ] ⊆ 𝔤𝑘+1 , and so since 𝔤 = 𝔤0 ,
by induction it follows that 𝐷𝑘 (𝔤) ⊆ 𝔤𝑘 , and hence for 𝑘 ≥ 𝑛 we have 𝐷𝑘 (𝔤) = 0.


Remark 4.1.5. Because the terms of the derived series are ideals in 𝔤, it follows that if 𝔤 is solvable, then
there is a filtration of 𝔤 whose terms are ideals in 𝔤 not just subalgebras each of which is an ideal in the
previous term of the filtration. In particular, if 𝔤 is solvable, it follows 𝔤 has an non-trivial abelian ideal,
since the last non-zero term of the derived series must be such an ideal.

Definition 4.1.6. If 𝔤 is an arbitrary Lie algebra, we set

𝐷∞ (𝔤) = � 𝐷𝑘 (𝔤)
𝑘≥0

the “limit” of the derived series (𝐷𝑘 (𝔤))𝑘≥0 . Since [𝐷𝑘 (𝔤), 𝐷𝑘 (𝔤)] = 𝐷𝑘+1 (𝔤) it follows that 𝐷(𝐷∞ (𝔤)) =
⋂ 𝐷𝑘 (𝔤) = 𝐷∞ (𝔤).
𝑘≥1

If dimk (𝔤) < ∞, then the sequence (𝑑𝑘 ) where 𝑑𝑘 = dimk (𝐷𝑘 (𝔤)) is weakly decreasing and bounded
below by 0, hence {𝑚 ∈ ℕ ∶ 𝑑𝑚 < 𝑑𝑚−1 } is finite and if we let 𝑁 be its maximal element then 𝐷𝑁 (𝔤) =
𝐷𝑁+1 (𝔤) = 𝐷𝑚 (𝔤) for all 𝑚 ≥ 𝑁, and so 𝐷∞ (𝔤) = 𝐷𝑁 (𝔤), i.e. the derived series stabilizes to its limit at
𝑘 = 𝑁.
We say a Lie algebra 𝔤 is perfect if it is nonzero and 𝔤 = 𝐷(𝔤), so that its derived series stabilizes
immediately. If 𝔤 is a finite-dimensional Lie algebra then since 𝐷∞ (𝔤) = 𝐷(𝐷∞ (𝔤)) we see that if 𝔤 is not
solvable then the limit of its derived series 𝐷∞ (𝔤) is a perfect ideal in 𝔤. But if 𝔞 ⊆ 𝔤 is a perfect subalgebra
in 𝔤 then 𝐷(𝔤) ⊇ 𝐷(𝔞) = 𝔞 and hence by induction 𝐷𝑘 (𝔤) ⊇ 𝔞 for all 𝑘, so that 𝔞 ⊆ 𝐷∞ (𝔤). It follows that
𝐷∞ (𝔤) is the largest perfect subalgebra in 𝔤 and hence we see that

𝔤 not solvable ⟺ ∃ a perfect ideal 𝐼 E 𝔤 ⟺ ∃ a perfect subalgebra 𝔥 ⊆ 𝔤

Recall the notion of a flag ℱ = (𝑉 = 𝐹0 > … 𝐹𝑑 = {0}) from Definition 2.2.3

Lemma 4.1.7. Let 𝑉 be a finite dimensional vector space and let ℱ = (𝑉 = 𝐹0 ⊃ 𝐹1 ⊃ … ⊃ 𝐹𝑑 = {0}) be a
flag in 𝑉, and set 𝐹𝑛 = {0} if 𝑛 ≥ 𝑑. Let, for any 𝑟 ∈ ℤ≥0 ,

𝔟𝑟ℱ = {𝑥 ∈ 𝔤𝔩𝑉 ∶ 𝑥(𝐹𝑖 ) ⊆ 𝐹𝑖+𝑟 , ∀ 𝑖, 0 ≤ 𝑖 ≤ 𝑑}.

22
 (i) If 𝑘, 𝑙 ≥ 0, then [𝔟𝑘ℱ , 𝔟𝑙ℱ ] ⊆ 𝔟𝑘+𝑙
ℱ .

(ii) If ℱ is a complete flag, and 𝔟ℱ = 𝔟0ℱ , then 𝐷(𝔟ℱ ) ⊆ 𝔟1ℱ and moreover 𝔟ℱ is solvable.

Proof. First note that 𝔟𝑟ℱ ⊆ 𝔟𝑠ℱ if 𝑟 ≥ 𝑠, and that if 𝑥 ∈ 𝔟𝑘ℱ , 𝑦 ∈ 𝔟𝑙ℱ , then clearly 𝑥 ∘ 𝑦 and 𝑦 ∘ 𝑥 lie in 𝔟𝑘+𝑙
ℱ .
It follows that the 𝔟ℱ form a descending sequence of associative subalgebras of Endk (𝑉), where the 𝔟𝑟ℱ
𝑟

for 𝑟 > 0 are two-sided ideals in 𝔟ℱ = 𝔟0ℱ , since 𝔟𝑟ℱ .𝔟𝑠ℱ ⊆ 𝔟𝑟+𝑠ℱ . But this immediately implies (i), that is,
𝑟 𝑠 𝑟+𝑠
[𝔟ℱ , 𝔟ℱ ] ⊆ 𝔟ℱ .
If ℱ is a complete flag, and 𝑥, 𝑦 ∈ 𝔟ℱ , then for any 𝑖, (1 ≤ 𝑖 ≤ 𝑑), 𝑥 and 𝑦 induce linear maps on
𝐹𝑖 /𝐹𝑖+1 , and, since ℱ is complete, dim(𝐹𝑖 /𝐹𝑖+1 ) = 1, so that 𝔤𝔩𝐹𝑖/𝐹𝑖+1 is abelian, and thus the map induced
by [𝑥, 𝑦] on 𝐹𝑖 /𝐹𝑖+1 is zero. But this exactly says that [𝑥, 𝑦] ∈ 𝔟1ℱ , and hence 𝐷(𝔟ℱ ) ⊆ 𝔟1ℱ . But then
𝑘−1
𝐷𝑘 (𝔟ℱ ) ⊆ 𝐷𝑘−1 (𝔟1ℱ ), and using (i) and induction 𝐷𝑘−1 (𝔟1ℱ ) ⊆ 𝔟2ℱ , which is {0} if 2𝑘−1 ≥ 𝑑 = dim(𝑉),
and hence 𝔟ℱ is solvable.


We will see shortly that, in characteristic zero, any solvable linear Lie algebra 𝔤 ⊂ 𝔤𝔩𝑉 , where 𝑉 is
finite dimensional, is a subalgebra of 𝔟ℱ for some complete flag ℱ. We next note some basic properties of
solvable Lie algebras. We establish them using the characterization of solvability in terms of the derived
series, but it is also straight-forward to show them using composition series.2
Lemma 4.1.8. Let 𝔤 be a Lie algebra, 𝜙 ∶ 𝔤 → 𝔥 a homomorphism of Lie algebras.

(i) We have, for all 𝑘 ≥ 0

𝜙(𝐷𝑘 (𝔤)) = 𝐷𝑘 (𝜙(𝔤)) ⊆ 𝐷𝑘 (𝔥).
In particular 𝜙(𝔤) is solvable if 𝔤 of 𝔥 is and thus any quotient or subalgebra of a solvable Lie algebra is
solvable.
(ii) If im(𝜙) and ker(𝜙) are solvable then so is 𝔤. Thus if 𝐼 is an ideal and 𝐼 and 𝔤/𝐼 are solvable, so is 𝔤.

Proof. For the claims about the terms of the derived series we use induction on 𝑘, the case 𝑘 = 0 being
trivial. If the result is known for 𝑘 then we have
𝜙(𝐷𝑘+1 (𝔤)) = 𝜙([𝐷𝑘 (𝔤), 𝐷𝑘 (𝔤)]) = [𝜙(𝐷𝑘 (𝔤)), 𝜙(𝐷𝑘 (𝔤))]
= [𝐷𝑘 (𝜙(𝔤)), 𝐷𝑘 (𝜙(𝔤))]
⊆ [𝐷𝑘 (𝔥), 𝐷𝑘 (𝔥)] = 𝐷𝑘+1 (𝔥).
where we use the inductive hypothesis for the equality on the second line and the containment in the
third. Since [𝐷𝑘 (𝜙(𝔤)), 𝐷𝑘 (𝜙(𝔤))] = 𝐷𝑘+1 (𝜙(𝔤)). The assertions in the final sentence of 𝑖) now follow by
applying the compatibility of the derived series to the case where 𝜙 a quotient map and the inclusion of
a subalgebra respectively.
For 𝑖𝑖) note that if im(𝜙) is solvable then for some 𝑘1 ≥ 0 we have 𝜙(𝐷𝑘1 (𝔤)) = 𝐷𝑘1 (𝜙(𝔤)) = {0} and
hence 𝐷𝑘1 (𝔤) ⊆ ker(𝜙). But then as ker(𝜙) is solvable we have 𝐷𝑘2 (ker(𝜙)) = {0} and hence

𝐷𝑘1+𝑘2 (𝔤) = 𝐷𝑘2 (𝐷𝑘1 (𝔤)) ⊆ 𝐷𝑘2 (ker(𝜙)) = {0}

and hence 𝔤 is solvable as required.


2
If one uses composition series to prove this Lemma, note that if 𝒞 = (𝔤𝑘 )𝑑𝑘=0
is a composition series for 𝔤 then in general if
𝔞 is a subalgebra of 𝔤, while the sequence 𝒞𝔞 = (𝔞 ∩ 𝔤𝑘 )𝑑𝑘=0 is a nested sequence of subalgebras of 𝔞 such that 𝔞 ∩ 𝔤𝑘 is contained
in the normalizer of 𝔞 ∩ 𝔤𝑘+1 , it need not be a composition series of 𝔞, as 𝔞 ∩ 𝔤𝑘+1 /𝔤𝑘+1 ⊆ 𝔤𝑘 /𝔤𝑘+1 is only a subalgebra, and not
an ideal, of the almost simple Lie algebra 𝔤𝑘 /𝔤𝑘+1 , and hence need not itself be almost simple. If all the composition factors are
equal to 𝔤𝔩1 however, then by considering dimensions, we see that (𝔞 ∩ 𝔤𝑘 )𝑑𝑘=0 , once one removes repetitions, is a composition
series of 𝔞.

23
4.2 Nilpotent Lie algebras

In this section we continue our study of Lie algebras which are built from 𝔤𝔩1 , but now by using central
extensions rather than arbitrary extensions.

Definition 4.2.1. A Lie algebra 𝔤 is said to be nilpotent if it can be obtained from 0, the trivial Lie algebra,
by iterated central extensions. If 𝔤 can be obtained by precisely 𝑘 iterated extentions, we say 𝔤 is 𝑘-step
nilpotent. Thus, for example, a Lie algebra is 1-step nilpotent if and only if it is abelian.

To make this more concrete, suppose that 𝔤 is a nilpotent Lie algebra. Then, for some 𝑘 ≥ 0 there are
Abelian Lie algebras (𝔠𝑖 )𝑘𝑖=0 and, for each 𝑖 ≥ 1 a short exact sequence
𝑝𝑖 𝑞𝑖
0 /𝔠 /𝔤 /𝔤 /0
𝑘+1 𝑘 𝑘−1

where 𝔤0 = 𝔠0 and 𝔠𝑖 ⊆ 𝔷(𝔤𝑖 ), that is, 𝔤𝑖 is a central extension of 𝔤𝑖−1 by 𝔠𝑖 . and 𝔤 = 𝔤𝑘 . It follows that
𝑞𝑘 ∶ 𝔤 = 𝔤𝑘 → 𝔤𝑘−1 , and if we set 𝑄𝑖 = 𝑞𝑖+1 ∘ 𝑞𝑖+1 ∘ … ∘ 𝑞𝑘 , then 𝑄𝑖 ∶ 𝔤 → 𝔤𝑖 exhibits 𝔤𝑖 as a quotient
of 𝔤. Set 𝔮𝑖 = ker(𝑄𝑖 ), so that if we set 𝔮0 = 𝔤, then (𝔮𝑖 )𝑘𝑖=0 gives a descending sequence of ideals in
𝔤, and 𝔮𝑖 /𝔮𝑖−1 ≅ 𝔠𝑖 is central in 𝔤/𝔮𝑖−1 . The sequence of central extensions constructing 𝔤 can thus be
reconstructed from the sequence of ideals (𝔮𝑖 )𝑘𝑖=0 .

Definition 4.2.2. For 𝔤 a Lie algebra, let 𝐶0 (𝔤) = 𝔤, and 𝐶𝑖 (𝔤) = [𝔤, 𝐶𝑖−1 (𝔤)] for 𝑖 ≥ 1. This sequence of
ideals of 𝔤 is called the lower central series of 𝔤.

Remark 4.2.3. Notice that 𝐶1 (𝔤) = [𝔤, 𝔤] is the derived subalgebra 𝐷(𝔤) though subsequent terms of the
lower central series will normally be distinct from the terms of the derived series.

Proposition 4.2.4. Suppose that 𝔤 is nilpotent and (𝔮𝑖 )𝑘𝑖=0 the sequence of ideals associated to a realization of 𝔤
as an iterated sequence of central extensions. Then

(i) For each 𝑖 ≥ 0 we have 𝐶𝑖 (𝔤) ⊆ 𝔮𝑖 and hence 𝐶𝑘 (𝔤) = 0.

(ii) Conversely, if 𝔤 is such that, for some 𝑁 ≥ 0 we have 𝐶𝑁 (𝔤) = 0, then 𝔤 is at most 𝑁-step nilpotent.

Proof. Suppose 𝔤 is any Lie algebra, and 𝔟 ⊆ 𝔞 are ideals in 𝔤. If 𝔞/𝔟 is central in 𝔤/𝔟, then for any 𝑥 ∈ 𝔤 and
𝑦 ∈ 𝔞 we must have [𝑥, 𝑦] ∈ 𝔟 and hence [𝔤, 𝔞] ⊆ 𝔟. Since 𝔞/[𝔤, 𝔞] is certainly central in 𝔤/[𝔤, 𝔞] it follows
that [𝔤, 𝔞] is the smallest ideal of 𝔤 contained in 𝔞 for which 𝔞 becomes central in the quotient algebra.
Applying this observation to 𝐶𝑖 (𝔤) inductively yields (i). For (ii), the converse, observe that the previ-
ous paragraph also shows that
0 / 𝐶𝑘 (𝔤)/𝐶𝑘+1 (𝔤) / 𝔤/𝐶𝑘+1 (𝔤) / 𝔤/𝐶𝑘 (𝔤) /0

exhibits 𝔤/𝐶𝑘+1 (𝔤) is a central extension of 𝔤/𝐶𝑘 (𝔤). It follows that if 𝐶𝑁 (𝔤) = 0 for some 𝑁 then 𝔤 is at
most 𝑁-step nilpotent. 

Lemma 4.2.5. Let 𝔤 be a Lie algebra and let 𝜙 ∶ 𝔤 → 𝔥 be a homomorphism of Lie algebras. Then

i)
𝜙(𝐶𝑘 (𝔤)) = 𝐶𝑘 (𝜙(𝔤)) ⊆ 𝐶𝑘 (𝔥)
In particular, any subalgebra or quotient of a 𝑘-step nilpotent Lie algebra is 𝑙-step nilpotent for some 𝑙 ≤ 𝑘.

ii) If 𝔤 is nilpotent then 𝔷(𝔤) ≠ {0} and 𝔤/𝔷(𝔤) is nilpotent if and only if 𝔤 is nilpotent.

Proof. For part 𝑖) We use induction on 𝑘, where the case 𝑘 = 0 is trivial. Now

𝜙(𝐶𝑘+1 (𝔤)) = 𝜙([𝔤, 𝐶𝑘 (𝔤)]) = [𝜙(𝔤), 𝜙(𝐶𝑘 (𝔤))] = [𝜙(𝔤), 𝐶𝑘 (𝜙(𝔤))] ⊆ [𝔥, 𝐶𝑘 (𝔥)] = 𝐶𝑘+1 (𝔥)

24
where we use the inductive hypothesis in the third equality and the containment.
For part 𝑖𝑖) it is clear that 𝔷(𝔤) is always non-zero for 𝔤 a non-zero nilpotent Lie algebra, indeed if 𝔤 is
𝑘-step nilpotent then [𝔤, 𝒞 𝑘−1 (𝔤)] = 𝒞 𝑘 (𝔤) = {0} and hence 𝒞 𝑘−1 (𝔤) ⊆ 𝔷(𝔤). Moreover, this shows that
𝔤/𝔷(𝔤) is a quotient of 𝔤/𝐶𝑘−1 (𝔤) and hence is (𝑘 − 1)-step nilpotent. 

Remark 4.2.6. Notice that if 𝔞 is an arbitrary ideal in 𝔤, and 𝔞 and 𝔤/𝔞 are nilpotent it does not follow that
𝔤 is nilpotent. Indeed recall from Example 3.2.4 the non-abelian 2-dimensional Lie algebra 𝔰2 , with basis
{𝑥, 𝑦} where [𝑥, 𝑦] = 𝑦. Then k.𝑦 is a 1-dimensional ideal in 𝔰2 but it is not central. Indeed 𝔷(𝔰2 ) = 0 so 𝔰2
is not nilpotent, even though the ideal k.𝑦 and the quotient 𝔰2 /k.𝑦 are (since they are both abelian). Note
that this shows that 𝔰2 cannot be written as a central extension of 𝔤𝔩1 by itself.

Remark 4.2.7. The characterisation of the property of nilpotence in terms of the lower central series is
similar to the characterisation of solvable Lie algebras in terms of the derived series. This is one reason
it is commonly used. There is, however, another nature nested sequence of ideals which can be used to
characterize nilpotence: If 𝔤 is any Lie algebra, set 𝑍0 (𝔤) = 𝔤, and, assuming 𝑍𝑘 (𝔤) is defined, let 𝑞𝑘 ∶ 𝔤 →
𝔤/𝑍𝑘 (𝔤) be the quotient map, and set 𝑍𝑘+1 (𝔤) = 𝑞−1 𝑘 (𝔷(𝔤𝑘 )). This process yields an increasing sequence of
ideals of 𝔤 known as the upper central series. If it exhausts 𝔤, that is, if for some 𝑛 ≥ 0 we have 𝑍𝑘 (𝔤) = 𝔤
for all 𝑘 large enough, then 𝔤 is nilpotent. If 𝔤 is not nilpotent, the upper central series will stabilize at a
maximal nilpotent ideal of 𝔤. Note that if 𝔤 is 𝑘-step nilpotent, then 𝑍𝑘−1 (𝔤) ⊊ 𝑍𝑘 (𝔤) = 𝔤 but the ideals
which occur in the upper central series need not coincide with those which occur in the lower central
series.
In terms of the adjoint representation, the centre of a Lie algebra 𝔤 can be viewed as ker(ad), the kernel
of the adjoint action, but it can also be viewed as the invariants in 𝔤, that is

𝔤𝔤 = {𝑧 ∈ 𝔤 ∶ 𝑎𝑑(𝑥)(𝑧) = 0, ∀ 𝑥 ∈ 𝔤}.

Using either the upper or lower central series, it is easy to see that the only composition factor of (𝔤, ad)
is the trivial representation.

We now wish to show that the notion of a flag in a vector space gives us a large supply of nilpotent Lie
algebras. In the next Lemma we use the notation of Lemma 4.1.7.

Lemma 4.2.8. Suppose that ℱ is a (not necessarily complete) flag in a finite-dimensional vector space 𝑉. Then
the Lie algebra 𝔫ℱ = 𝔟1ℱ ⊆ 𝔤𝔩𝑉 is nilpotent.

Proof. By (i) of Lemma 4.1.7, [𝔫ℱ , 𝔫ℱ ] ⊆ 𝔟2ℱ , and by induction [𝔫ℱ , 𝐶𝑘 (𝔫ℱ )] ⊆ 𝔟𝑘+1 𝑘 𝑘
ℱ , so that 𝐶 (𝔤) ⊆ 𝔟ℱ ,
and hence 𝐶𝑘 (𝔫) = 0 if 𝔟𝑘ℱ = 0, which is true whenever 𝑘 ≥ 𝑑 = dim(𝑉). 

Example 4.2.9. When ℱ is a complete flag, so that dim(𝑉) = 𝑑, if we pick a basis {𝑒1 , 𝑒2 , … , 𝑒𝑑 } of 𝑉 such
that 𝐹𝑘 = ⟨{𝑒𝑘+1 , 𝑒2 , … , 𝑒𝑑 }⟩k , then the matrix 𝐴 representing an element 𝑥 ∈ 𝔫ℱ with respect to this basis
is strictly upper triangular, that is, 𝑎𝑖𝑗 = 0 for all 𝑖 ≥ 𝑗. But then if 𝔫𝑑 ⊆ 𝔤𝔩𝑑 denotes the space of strictly
upper-triangular matrices, it is easy to see that dim(𝔫𝑑 ) = �𝑑2�. When 𝑑 = 2 we just get the 1-dimensional
Lie algebra 𝔤𝔩1 , thus the first nontrivial case is when 𝑛 = 3 and in this case 𝔫3 is the 3-dimensional 2-step
nilpotent Lie algebra we constructed previously as a central extension.
On the other hand, if 𝔟𝑑 ⊆ 𝔤𝔩𝑑 denotes the upper-triangular matrices, i.e. 𝔟𝑑 corresponds to the sub-
algebra 𝔟ℱ of 𝔤𝔩𝑉 , and we set 𝔱𝑑 to be the set of diagonal matrices in 𝔟𝑛 , then it is straight-forward to show
by considering the subalgebra 𝔱𝑛 of diagonal matrices in 𝔟𝑛 that [𝔱𝑑 , 𝔫𝑑 ] = 𝔫𝑑 , so that, as 𝔟𝑑 = 𝔱𝑑 ⊕ 𝔫𝑑 , it
follows that 𝔟𝑑 is not nilpotent.

Remark 4.2.10. Note that in Example 4.1.7, unlike in Lemma 4.2.8, it is essential that ℱ is a complete
flag. If ℱ is not a complete flag the corresponding subalgebra 𝔟ℱ will not be solvable (since, for example,
if dim(𝐹𝑖 /𝐹𝑖+1 ) > 1, then there is a surjective homomorphism 𝔟ℱ → 𝔤𝔩𝐹𝑖/𝐹𝑖+1 , which is not solvable.)

25
Remark 4.2.11. Note that the subalgebra 𝔱𝑑 ⊂ 𝔤𝔩𝑑 of diagonal matrices is abelian, and hence nilpotent,
but the only nilpotent endomorphism of k𝑑 that lies in 𝔱𝑑 is 0. Thus a nilpotent linear Lie algebra need not
consist of nilpotent endomorphisms. It turns out that, in some sense, the example of 𝔱𝑑 is the only way in
which a nilpotent Lie algebra 𝔫 ⊆ 𝔤𝔩𝑑 can fail to consist of nilpotent endomorphisms. We will make this
precise in 4.4.

4.2.1 Nilpotent representations

Definition 4.2.12. Let 𝔤 be a Lie algebra and (𝑉, 𝜌) a representation of 𝔤. We say that (𝑉, 𝜌) is nilpotent if,
for all 𝑥 ∈ 𝔤, the endomorphism 𝜌(𝑥) ∈ 𝔤𝔩𝑉 is a nilpotent linear map (that is, for some 𝑛 ≥ 1, 𝜌(𝑥)𝑛 = 0).
Lemma 4.2.13. Let 𝐴 be an associative algebra, and suppose 𝑎, 𝑏 ∈ 𝐴 are nilpotent i.e. for some 𝑛 > 0, we have
𝑎𝑛 = 𝑏𝑛 = 0. Then if 𝑎 and 𝑏 commute, 𝑎 + 𝑏 is also nilpotent.

Proof. This follows from the binomial theorem: Indeed we have

𝑚 𝑚
(𝑎 + 𝑏)𝑚 = � 𝑘 𝑚−𝑘
� �𝑎 𝑏 .
𝑘=0 𝑘

But now if 𝑚 ≥ 2𝑛, then we must have either 𝑘 ≥ 𝑛 or 𝑚 − 𝑘 ≥ 𝑛, hence in either case, each of the
terms on the left-hand side vanishes, hence so does the right-hand side, and hence 𝑎 + 𝑏 is nilpotent as
required. 

Lemma 4.2.14. Suppose 𝔤 is a Lie algebra and (𝑉, 𝜌) and (𝑊, 𝜎) are representation of 𝔤.

i) If 𝑥 ∈ 𝔤 is such that both 𝜌(𝑥) and 𝜎(𝑥) are nilpotent, then the action of 𝑥 on 𝑉 ⊗ 𝑊 is also nilpotent.
Moreover, the action of 𝑥 on 𝑉 ∗ is also nilpotent. Thus if 𝑉 and 𝑊 are nilpotent, so are 𝑉 ∗ , 𝑉 ⊗ 𝑊 and
ℒ(𝑉, 𝑊) ≅ 𝑉 ∗ ⊗ 𝑊.

ii) If 𝑉 is nilpotent, then any subrepresentation and any quotient representation of 𝑉 is also nilpotent.

Proof. By definition, the action of 𝑥 on 𝑉 ⊗ 𝑊 is given by 𝜌(𝑥) ⊗ 1𝑊 + 1𝑉 ⊗ 𝜎(𝑥). Since the two terms in
this sum commute, the claim follows from Lemma 4.2.13 (taking 𝐴 = End(𝑉 ⊗ 𝑊).)
To see that 𝑥 acts nilpotently on 𝑉 ∗ , note that if 𝑓 ∈ 𝑉 ∗ , then

𝑥𝑛 (𝑓)(𝑣) = (−1)𝑛 𝑓(𝜌(𝑥)𝑛 )(𝑣)) = ±𝑓(0) = 0, ∀ 𝑣 ∈ 𝑉, 𝑓 ∈ 𝑉 ∗ .

For part 𝑖𝑖) if 𝑈 is a subrepresentation of 𝑉 then, as in Lemma 2.2.2, 𝜌(𝔤) ⊆ 𝔟𝑈 = {𝑥 ∈ 𝔤𝔩𝑉 ∶

𝑥(𝑈) ⊆ 𝑈}. But 𝔟𝑈 is an associative subalgebra of End(𝑉) and the maps 𝑖∗ and 𝑝∗ from 𝔟𝑈 to End(𝑈)
and End(𝑉/𝑈) are also maps of associative algebras, hence if 𝜌(𝑥) ∈ 𝔟𝑈 is nilpotent, its images in End(𝑈)
and End(𝑉/𝑈) are also. 

The next proposition is the key result in this section. For the proof we will need the notion of the
normalizer 𝑁𝔤 (𝔞) of a subalgebra 𝔞 of a Lie algebra 𝔤 given in Definition 1.2.6. We have

𝑁𝔤 (𝔞) = {𝑥 ∈ 𝔤 ∶ [𝑥, 𝑎] ∈ 𝔞, ∀ 𝑎 ∈ 𝔞},

so that 𝑁𝔤 (𝔥) is the largest subalgebra of 𝔤 in which 𝔞 is an ideal.

Proposition 4.2.15. Let 𝔤 be a Lie algebra, and let (𝑉, 𝜌) be a nilpotent representation of 𝔤.

i) The invariant subspace

𝑉 𝔤 = {𝑣 ∈ 𝑉 ∶ 𝜌(𝑥)(𝑣) = 0, ∀ 𝑥 ∈ 𝔤}
is non-zero.

26
 ii) There is a complete flag ℱ in 𝑉 such that 𝔤 ⊆ 𝔫ℱ . In particular, the image 𝜌(𝔤) is a nilpotent Lie algebra.

Proof. To prove 𝑖), we use induction on 𝑑 = dim(𝔤), the case 𝑑 = 1 being clear. Now if 𝜌 is not faithful,
i.e. ker(𝜌) ≠ 0, then dim(𝜌(𝔤)) < dim(𝔤), and we are done by induction applied to the image 𝜌(𝔤), hence
we may assume 𝜌 gives an embedding of 𝔤 into 𝔤𝔩𝑉 as a subalgebra, and we may thus identify 𝔤 with its
image in the rest of this proof.
Now let 𝒮 = {𝔟 ⊊ 𝔤 ∶ 𝔟 is a proper subalgebra of 𝔤} denote the set of proper subalgebras of 𝔤, and
pick 𝔞 ∈ 𝒮. Now by Lemma 4.2.14, 𝔞 ⊊ 𝔤 ⊆ 𝔤𝔩𝑉 = 𝑉 ∗ ⊗ 𝑉 are all nilpotent representations of 𝔞, since the
restriction of (𝑉, 𝜌) to 𝔞 is. But then, by the same Lemma, 𝔤/𝔞 is also a nilpotent representation, and since
dim(𝔞) < dim(𝔤), it follows by induction that the 𝔞-invariants (𝔤/𝔞)𝔞 form a non-zero subrepresentation.
Let 𝑥 ∈ 𝔤 be such that 0 ≠ 𝑥 + 𝔞 ∈ (𝔤/𝔞)𝔞 . Then ad(𝑎)(𝑥) ∈ 𝔞 for all 𝑎 ∈ 𝔞, or equivalently, since
ad(𝑎)(𝑥) = −ad(𝑥)(𝑎), for all 𝑎 ∈ 𝔞, we have ad(𝑥)(𝑎) ∈ 𝔞, that is, 𝑥 ∈ 𝑁𝔤 (𝔞). Thus the normalizer of 𝔞 is a
subalgebra of 𝔤 which is strictly larger than 𝔞.
Thus if we take 𝔞 ∈ 𝒮 of maximal dimension, we must have 𝑁𝔤 (𝔞) = 𝔤, that is 𝔞 is an ideal in 𝔤. But
then if 𝑧 ∈ 𝔤\𝔞, it is easy to see that k.𝑧 ⊕ 𝔞 is a subalgebra3 of 𝔤, hence again by maximality, we must have
𝔤 = k.𝑧 ⊕ 𝔞. By induction, we know that 𝑉 𝔞 = {𝑣 ∈ 𝑉 ∶ 𝑎(𝑣) = 0, ∀ 𝑎 ∈ 𝔞} is a nonzero subspace of 𝑉. We
claim that 𝑧 preserves 𝑉 𝔞 . Indeed
𝑎(𝑧(𝑣)) = [𝑎, 𝑧](𝑣) + 𝑧(𝑎(𝑣)) = 0, ∀ 𝑎 ∈ 𝔞, 𝑣 ∈ 𝑉 𝔞 ,
since [𝑎, 𝑧] ∈ 𝔞. But the restriction of 𝑧 to 𝑉 𝔞 is nilpotent, so the subspace 𝑈 = {𝑣 ∈ 𝑉 𝔞 ∶ 𝑧(𝑣) = 0} is
nonzero. Since 𝑈 = 𝑉 𝔤 we are done.
For 𝑖𝑖), let 𝒞 = (𝑉 = 𝐹𝑚 > 𝐹𝑚−1 > … > 𝐹1 > 𝐹0 = {0}) be a composition series for 𝑉. It suffices
to show that each of the composition factors are trivial. But if 1 ≤ 𝑘 ≤ 𝑚, then 𝐹𝑘 is a subrepresentation
of 𝑉 and hence it is nilpotent. Similarly 𝑄𝑘 = 𝐹𝑘 /𝐹𝑘+1 , as a quotient of 𝐹𝑘 must be nilpotent. But then by
part (1), its invariants 𝑄𝔤𝑘 are a non-zero subrepresentation of 𝑄𝑘 , and since 𝑄𝑘 is simple it follows that
𝑄𝑘 is the trivial representation as required.


Corollary 4.2.16. (Engel’s theorem.) A Lie algebra 𝔤 is nilpotent if and only if ad(𝑥) is nilpotent for every 𝑥 ∈ 𝔤,
i.e the adjoint representation is nilpotent.

Proof. If 𝔤 is nilpotent, then since by definition ad(𝑥)(𝐶𝑖 (𝔤)) ⊆ 𝐶𝑖+1 (𝔤), we see that ad(𝑥)𝑘 = 0 for all
𝑥 ∈ 𝔤 if 𝔤 is 𝑘-step nilpotent. Now suppose that ad(𝑥) is nilpotent for all 𝑥 ∈ 𝔤. Then (𝔤, ad) is a nilpotent
representation, and hence by part 𝑖𝑖) of Proposition 4.2.15, we see that ad(𝔤) is nilpotent. But since ad(𝔤) ≅
𝔤/𝔷(𝔤) it follows that 𝔤 is nilpotent as required. 

4.3 Representations of solvable Lie algebras

In this section we will assume that our field k is algebraically closed of characteristic zero.

4.3.1 Lie’s theorem

Our first goal is the following theorem:

Theorem 4.3.1. (Lie’s theorem) Let 𝔤 be a solvable Lie algebra and 𝑉 is a 𝔤-representation. Then there is a
homomorphism 𝜆 ∶ 𝔤 → 𝔤𝔩1 (k) and a nonzero vector 𝑣 ∈ 𝑉 such that 𝑥(𝑣) = 𝜆(𝑥).𝑣 for all 𝑥 ∈ 𝔤. Equivalently,
any finite-dimensional irreducible representation of a solvable Lie algebra is one-dimensional.
3
One way to see this is to note that k.𝑧 ⊕ 𝔞 is a line– i.e. one-dimensional subspace– of 𝔤/𝔞 and any such subspace is a
subalgebra, because, by the alternating property, the Lie bracket vanishes on lines. Note in particular that the direct sum is one
of vector spaces, not Lie algebras.

27
 We first explain the equivalence asserted in the last sentence of the statement. Note that the exis-
tence of a non-zero 𝑣 ∈ 𝑉 such that 𝑥(𝑣) = 𝜆(𝑥).𝑣 for all 𝑥 ∈ 𝔤 is equivalent to the assertion that the
line k.𝑣 is a subrepresentation of 𝑉. Thus the statement of the theorem shows that any representation
contains a one-dimensional subrepresentation, and hence any irreducible representation must itself be
one-dimensional. Since any representation contains an irreducible representation, the equivalence fol-
lows.
The crucial observation that is needed to prove Lie’s theorem is given in the following Lemma:
Lemma 4.3.2. (Lie’s Lemma) Let 𝔤 be a Lie algebra, let 𝐼 ⊂ 𝔤 be an ideal, and let 𝑉 be a finite dimensional
𝔤-representation. Suppose that 𝜆 ∶ 𝐼 → 𝔤𝔩1 (k) is a homomorphism of Lie algebras for which the subspace 𝑉𝜆,𝐼 =
{𝑣 ∈ 𝑉 ∶ ℎ(𝑣) = 𝜆(ℎ).𝑣, ∀ ℎ ∈ 𝐼} is nonzero. Then 𝜆 vanishes on [𝔤, 𝐼] ⊂ 𝐼, and 𝑉𝜆,𝐼 is a 𝔤-subrepresentation of
𝑉.

Proof. Fix 𝑥 ∈ 𝔤 and 𝑣 ∈ 𝑉𝜆,𝐼 \{0}. For each 𝑚 ∈ ℕ, let 𝑊𝑚 = ⟨{𝑣, 𝑥(𝑣), … , 𝑥𝑚 (𝑣)}⟩k . The 𝑊𝑚 form an
increasing sequence of subspaces of 𝑉 with dim(𝑊𝑚+1 /𝑊𝑚 ) ≤ 1, from which it is easy to see that there is
some 𝑑 with 𝑊−1 ∶= {0} < 𝑊0 < 𝑊1 < … < 𝑊𝑑−1 = 𝑊𝑑 , where 𝑥(𝑊𝑑 ) ⊆ 𝑊𝑑 .
Claim:
ℎ(𝑥𝑚 (𝑣)) ∈ 𝜆(ℎ).𝑥𝑚 (𝑣) + 𝑊𝑚−1 , ∀ ℎ ∈ 𝐼, 𝑚 ≥ 0

Proof of claim: This obvious for 𝑚 = 0 and if 𝑚 ≥ 1, since [ℎ, 𝑥] ∈ 𝐼, by induction we have

ℎ𝑥𝑚 (𝑣) = 𝑥ℎ𝑥𝑚−1 (𝑣) + (ℎ𝑥 − 𝑥ℎ)𝑥𝑚−1 (𝑣)

∈ 𝑥 �𝜆(ℎ).𝑥𝑚−1 (𝑣) + 𝑊𝑚−2 � + 𝜆([ℎ, 𝑥])𝑥𝑚−1 (𝑣) + 𝑊𝑚−2 (4.3.1)
⊆ 𝜆(ℎ)𝑥𝑚 (𝑣) + 𝑊𝑚−1 ,

since 𝑥(𝑊𝑚−2 ) ⊆ 𝑊𝑚−1 and [ℎ, 𝑥] ∈ 𝐼.

It follows that 𝑊𝑑 is stable under the action of 𝐼 and 𝑥, and that if ℎ ∈ 𝐼, then the matrix of the action of
ℎ on 𝑊𝑑 with respect to the basis {𝑣, 𝑥(𝑣), … , 𝑥𝑑−1 (𝑣)} is upper-triangular with each diagonal entry equal
to 𝜆(ℎ). In particular, tr(ℎ|𝑊𝑑 ) = 𝜆(ℎ).dim(𝑊𝑑 ) = 𝜆(ℎ).𝑑. But then

0 = tr([ℎ, 𝑥]) = 𝜆([ℎ, 𝑥]).𝑑,

and since 𝑑 > 0 and char(k) = 0, it follows that 𝜆([ℎ, 𝑥]) = 0. But now considering (4.3.1) with 𝑚 = 1 we
see that 𝜆([ℎ, 𝑥]) = 0 implies that ℎ𝑥(𝑣) = 𝜆(ℎ).𝑥(𝑣), so that 𝑥(𝑣) ∈ 𝑉𝜆,𝐼 if 𝑣 ∈ 𝑉𝜆,𝐼 as required.


Completion of the proof of Lie’s Theorem:

Use induction on dim(𝔤). Since 𝔤 is solvable, 𝐷(𝔤) is a proper ideal in 𝔤, and by induction the theorem
holds for 𝐷(𝔤). But then let 𝜆 ∶ 𝐷(𝔤) → k be a homomorphism of Lie algebras such that

𝑊 = {𝑣 ∈ 𝑉 ∶ ℎ(𝑣) = 𝜆(ℎ).𝑣, ∀ ℎ ∈ 𝐷(𝔤)} ≠ {0}.

But then by Lemma 4.3.2, 𝑊 is a 𝔤-subrepresentation, and if we let 𝜌 ∶ 𝔤 → 𝔤𝔩𝑊 , then 𝜌(𝔤) ⊆ 𝔤𝔩𝑊 has
𝜆(𝐷(𝔤)).𝐼𝑊 = 𝜌(𝐷(𝔤)) = 𝐷(𝜌(𝔤)). But 𝐷(𝔤𝔩𝑊 ) = 𝔰𝔩(𝑊), and since char(k) = 0, k.𝐼𝑊 ∩ 𝔰𝔩(𝑊) = {0}, so
that as 𝐷(𝔤1 ) ⊆ 𝔰𝔩(𝑊) ∩ k.𝐼𝑊 = {0}. It follows that the action of 𝔤 on 𝑊 factors through 𝔤/𝐷(𝔤), which
is abelian, and the result is then clear, since commuting linear maps on a non-zero vector space always
have a common eigenvector.
Remark 4.3.3. The proof of Lemma 4.3.2 relies on a trick which permeates the course, namely that one
can often compute a trace in two different ways to obtain important information. One way will be by
observing that one is computing the trace of a commutator, which is therefore zero. The other will, in one
fashion or another, follow from consideration of the generalised eigenspaces of the linear map in question.

28
Corollary 4.3.4. Let 𝔤 be a solvable Lie algebra and let (𝑉, 𝜌) be a 𝔤-representation. Then there is a complete
flag ℱ = (𝑉 = 𝐹0 ⊃ 𝐹1 ⊃ … ⊃ 𝐹𝑑 = {0}) where each 𝐹𝑖 is a 𝔤-subrepresentation, so that 𝜌(𝔤) ⊆ 𝔟ℱ . In
particular, if 𝔤 is solvable, then it has a composition series each of whose terms is an ideal in all of 𝔤.

Proof. Take any composition series ℱ for 𝑉. Since Lie’s theorem shows that the irreducible representa-
tions of 𝔤 are all one-dimensional, the resulting chain of subrepresentations will form a complete flag and
𝜌(𝔤) ⊆ 𝔟ℱ . The final sentence follows by applying this to the adjoint representation (𝔤, ad), since 𝐼 ⊆ 𝔤 is
an ideal if and only if it is a subrepresentation of the adjoint representation. 

Definition 4.3.5. Recall from Example 2.1.6 that the elements of (𝔤/𝐷(𝔤))∗ = 𝐷(𝔤)0 naturally parame-
terise the isomorphism classes of one-dimensional representations of a Lie algebra 𝔤. Indeed a homo-
morphism 𝜆 ∶ 𝔤 → 𝔤𝔩1 is just a linear map 𝜆 ∶ 𝔤 → k which vanishes on 𝐷(𝔤). Recall that we write k𝜆
for the representation (k, 𝜆). We will refer to an element of 𝐷(𝔤)0 (equivalently, an isomorphism class of
1-dimensional 𝔤-representations) as a weight of 𝔤. In the case where 𝔤 is solvable, Lie’s theorem shows
that the weights are exactly the isomorphism classes of irreducible 𝔤-representations.

4.4 Representations of nilpotent Lie algebras

In this section we assume that k is an algebraically closed field of characteristic zero.

Definition 4.4.1. Let 𝔤 be a Lie algebra and let 𝒮 be a set of irreducible representation of 𝔤. Let

Rep𝒮 (𝔤) = {𝑉 ∈ Rep(𝔤) ∶ [𝑇 ∶ 𝑉] > 0 if and only if ∃𝑆 ∈ 𝒮 , 𝑇 ≅ 𝑆}

Rep𝒮 (𝔤, 𝑉) = {𝑊 ≤ 𝑉 ∶ 𝑊 ∈ Rep𝒮 (𝔤)}.

If 𝒮 = {𝑆} then we will write Rep𝑆 (𝔤), Rep𝑆 (𝔤, 𝑉) rather than Rep{𝑆} (𝔤), Rep{𝑆} (𝔤, 𝑉) respectively. If
𝑉 ∈ Rep𝒮 (𝔤) then it is easy to see that any subrepresentation or quotient representation of 𝑉 also lies in
Rep𝒮 (𝔤).

Proposition 4.4.2. Let 𝔤 be a Lie algebra and (𝑉, 𝜌) a representation of 𝔤. If 𝒮 is a set of irreducible 𝔤-
representation then Rep𝒮 (𝔤, 𝑉) has a unique element 𝑉𝒮 which is maximal with respect to containment, that
is 𝑉𝒮 ∈ Rep𝒮 (𝔤, 𝑉) and if 𝑈 ∈ Rep𝒮 (𝔤, 𝑉) then 𝑈 ≤ 𝑉𝒮 .

Proof. First note that if it exists, such a maximal element is automatically unique, since if 𝑊1 , 𝑊2 are both
maximal with respect to containment we must have 𝑊1 ≤ 𝑊2 ≤ 𝑊1 and hence 𝑊1 = 𝑊2 .
Next note that if 𝑉1 , 𝑉2 ∈ 𝒱𝑆 then 𝑉1 + 𝑉2 ∈ 𝒱𝑆 . Indeed by the second isomorphism theorem,
(𝑉1 + 𝑉2 )/𝑉1 ≅ 𝑉2 /(𝑉1 ∩ 𝑉2 ), so that any composition factor of 𝑉1 + 𝑉2 must be a composition factor of
𝑉1 or of 𝑉2 /(𝑉1 ∩ 𝑉2 ), and hence is a composition factor of 𝑉1 or 𝑉2 . Now pick 𝑊 ∈ 𝒱𝑆 with dim(𝑊) ≥
dim(𝑈) for all 𝑈 ∈ 𝒱[𝑆] (such a 𝑊 exists if 𝑉 is finite-dimensional, as we always assume). We claim
that 𝑊 is maximal for containment. Indeed if 𝑈 ∈ 𝒱𝑆 then we have just shown that 𝑊 + 𝑈 ∈ 𝒱𝑆 ,
hence dim(𝑊) ≤ dim(𝑊 + 𝑈) ≤ dim(𝑊) by our choice of 𝑊, and hence 𝑈 ≤ 𝑊 and 𝑊 is maximal for
containment as required. Thus 𝑊 = 𝑉𝒮 is the unique maximal subrepresentation in 𝒱𝒮 . 

Example 4.4.3. If 𝔤 = 𝔤𝔩1 and (𝑉, 𝜌) is a 𝔤-representation, let 𝛼 = 𝜌(1). The simple representations of 𝔤
are the lines spanned by eigenvectors of 𝛼, and are captured up to isomorphism by their eigenvalue. If we
set 𝒮 = {𝜆} to be a single irreducible 𝔤-representation then 𝑉𝜆 is the generalized eigenspace for 𝛼 with
eigevalue 𝜆. (For more details on the generalised eigenspace decomposition of a vector space equipped
with a linear map, see I.)

29
Definition 4.4.4. Recall that the isomorphism classes of 1-dimensional representations of 𝔤 can be iden-
tified with 𝐷(𝔤)0 ⊆ 𝔤, and given 𝜆 ∈ 𝐷(𝔤)0 , we write k𝜆 for the 1-dimensional representation (k, 𝜆).
Given a 𝔤-representation (𝑉, 𝜌), we will write 𝑉𝜆 and Rep𝜆 (𝔤, 𝑉) instead of 𝑉k𝜆 and Repk (𝔤, 𝑉). When
𝜆
𝜆 ∈ 𝐷(𝔤)0 we will refer to 𝑉𝜆 as the generalised 𝜆-weight space of 𝑉. If 𝑉 is a finite-dimensional represen-
tation of a Lie algebra 𝔤, let
Ψ𝑉 = {𝜆 ∈ 𝐷(𝔤)0 ∶ 𝜆 is a composition factor of 𝑉}
and let 𝑚𝑉 ∶ 𝐷(𝔤)0 → ℤ≥0 be the function which assigns to 𝜆 ∈ 𝐷(𝔤)0 the multiplicity of 𝜆 as a com-
position factor of 𝑉. Thus Ψ𝑉 is the finite set of the one-dimensional representations 𝜆 of 𝔤 for which
𝑚𝑉 (𝜆) ≠ 0. If 𝔤 is solvable and char(k) = 0 then by Lie’s Theorem Ψ𝑉 contains all the composition
factors of 𝑉.
If 𝜑 ∶ 𝔤1 → 𝔤2 and (𝑉, 𝜌) is a representation of 𝔤2 , then (𝑉, 𝜑∗ (𝜌)) is a representation of 𝔤1 , where
𝜑∗ (𝜌) = 𝜌∘𝜑. Since 𝜑(𝐷(𝔤1 )) ⊆ 𝐷(𝔤2 ), the transpose 𝜑⊺ ∶ 𝔤∗2 → 𝔤∗1 restricts to give a map 𝜑⊺ ∶ 𝐷(𝔤2 )0 →
𝐷(𝔤1 )0 . Clearly, the irreducible 𝔤1 -representation (k, 𝜆) restricts via 𝜑 to the irreducible representation
(k, 𝜑⊺ (𝜆), and hence Ψ𝜑∗(𝑉) ⊇ 𝜑⊺ (Ψ𝑉 ).
More precisely, 𝑚𝜑∗ (𝑉) (𝜇) ≥ ∑𝜆∈Ψ ∶𝜑⊺(𝜆)=𝜇 𝑚𝑉 (𝜆). In the case where Ψ𝑉 is the complete set of
𝑉
composition factors of 𝑉, such as when 𝔤1 is solvable, this inequality is an equality. Now if 𝑥 ∈ 𝔤 and
𝑖𝑥 ∶ 𝔤𝔩1 → 𝔤 is the homomorphism 𝑖𝑥 (𝑡) = 𝑡.𝑥 (∀ 𝑡 ∈ k = 𝔤𝔩1 ), then if 𝜆 ∈ 𝐷(𝔤)0 , we have 𝑖⊺
𝑥 (𝜆) = 𝜆(𝑥).
4

The weights of the 𝔤𝔩1 -representation 𝜌 ∘ 𝑖𝑥 are just the eigenvalues of 𝜌(𝑥), as in Example 2.1.5, it fol-
lows that the eigenvalues of 𝜌(𝑥) are {𝜆(𝑥) ∶ 𝜆 ∈ Ψ𝑉 }, and the 𝜇-generalised eigenspace of 𝜌(𝑥) is
⨁ 𝑉𝜆 .
𝜆∈Ψ𝑉 ∶𝜆(𝑥)=𝜇

Lemma 4.4.5. Suppose that 𝔤 is a Lie algebra and 𝜆, 𝜇 ∈ 𝐷(𝔤)0 are weights of 𝔤. If 𝑉 and 𝑊 are 𝔤-
representations then

i) 𝑉𝜆 ⊗ 𝑊𝜇 ⊆ (𝑉 ⊗ 𝑊)𝜆+𝜇 .
ii) If 𝜙 ∶ 𝑉 → 𝑊 is a homomorphism of 𝔤-representation, then 𝜙(𝑉𝜆 ) ⊆ 𝑊𝜆 .

Proof. For part (i), we may assume that 𝑉 = 𝑉𝜆 and 𝑊 = 𝑊𝜇 , hence there are composition series (𝐹𝑘 )𝑟𝑘=0
and (𝐺𝑙 )𝑠𝑙=0 , where 𝐹𝑘 /𝐹𝑘+1 ≅ k𝜆 for each 𝑘, and 𝐺𝑙 /𝐺𝑙+1 ≅ k𝜇 , for all 𝑙 ∈ {0, 1, … , 𝑟} and 𝑘 ∈ {0, 1 … , 𝑠}.
Pick bases {𝑒𝑖 ∶ 0 ≤ 𝑖 ≤ 𝑟 − 1} and {𝑓𝑗 ∶ 0 ≤ 𝑗 ≤ 𝑠 − 1} of 𝑉 and 𝑊 respectively such that 𝐹𝑘 = ⟨{𝑒𝑖 ∶ 𝑖 ≥ 𝑘}⟩k
and 𝐺𝑙 = ⟨{𝑓𝑗 ∶ 𝑗 ≥ 𝑙}⟩k . If we set 𝐻𝑘 = ∑𝑟+𝑠=𝑘 𝐹𝑟 ⊗ 𝐺𝑠 , then 𝐻𝑘 is a subrepresentation of 𝑉 ⊗ 𝑊 and we
have

𝑥(𝑒𝑘 ) ⊗ 𝑓𝑙 ∈ 𝜆(𝑥)𝑒𝑘 ⊗ 𝑓𝑙 + 𝐹𝑘+1 ⊗ 𝐺𝑙 ;

𝑒𝑘 ⊗ 𝑥(𝑓𝑙 ) ∈ 𝜇(𝑥).𝑒𝑘 ⊗ 𝑓𝑙 + 𝐹𝑘 ⊗ 𝐺𝑙+1 (4.4.1)
⟹ 𝑥(𝑒𝑘 ⊗ 𝑓𝑙 ) = 𝑥(𝑒𝑘 ) ⊗ 𝑓𝑙 + 𝑒𝑘 ⊗ 𝑥(𝑓𝑙 ) ∈ (𝜆 + 𝜇)(𝑥)(𝑒𝑘 ⊗ 𝑓𝑙 ) + 𝐻𝑘+𝑙+1
dim(𝐻𝑘 )−dim(𝐻𝑘+1 )
and thus 𝐻𝑘 /𝐻𝑘+1 ≅ k𝜆+𝜇 . It follows 𝑉 ⊗ 𝑊 has k𝜆+𝜇 as its unique composition factor.
For part (ii), since 𝑉𝜆 ∈ Rep𝜆 (𝔤), and 𝜙(𝑉𝜆 ) ≅ 𝑉𝜆 /ker(𝜙|𝑉𝜆 ) is isomorphic to a quotient of 𝑉𝜆 , it lies
in Rep𝜆 (𝔤, 𝑊) and so by the maximality of 𝑊𝜆 it follows that 𝜙(𝑉𝜆 ) ⊆ 𝑊𝜆 . 

The adjoint representation of a nilpotent Lie algebra 𝔤 has the trivial representation as its only com-
position factor, that is, 𝔤 = 𝔤0 . This has the following important consequence:
Proposition 4.4.6. Let 𝔤 be a nilpotent Lie algebra, 𝔥 ⊆ 𝔤 be a subalgebra of 𝔤, and (𝑉, 𝜌) a representation of
𝔤. Then if 𝜇 ∈ (𝔥/𝐷(𝔥))∗ ≅ 𝐷(𝔥)0 /𝔥0 ⊆ 𝔤∗ /𝔥0 ≅ 𝔥∗ is a weight of 𝔥, and 𝑉𝜇 is the 𝜇-isotypic subrepresentation
of Res𝔤𝔥 (𝑉), the restriction of 𝑉 to 𝔥, then 𝑉𝜇 is a 𝔤-subrepresentation of 𝑉. In particular, taking 𝔥 = k.𝑥 for
𝑥 ∈ 𝔤\{0}, any generalised eigenspace 𝑉𝜇,𝑥 of 𝜌(𝑥) is a 𝔤-subrepresentation.
4
The map 𝑖⊺𝑥 ∶ 𝔤∗ → 𝔤𝔩1 ∗ = k∗ , but k∗ = k via 𝜆 ↦ [𝑡 ↦ 𝜆𝑡], ∀ 𝑡 ∈ k.

30
Proof. Since 𝔤 is nilpotent, we have 𝔤 = 𝔤0 as an 𝔥-representation. But then by Lemma 4.4.5, we have
𝔤 ⊗ 𝑉𝜇 = 𝔤0 ⊗ 𝑉𝜇 ⊆ (𝔤 ⊗ 𝑉)𝜇 , and since the map 𝑎𝜌̃ ∶ 𝔤 ⊗ 𝑉 → 𝑉 given by 𝑎𝜌̃ (𝑥 ⊗ 𝑣) = 𝜌(𝑥)(𝑣) is a
homomorphism of 𝔥-representations by Example 2.3.4, it follows that 𝑎𝜌̃ (𝔤 ⊗ 𝑉𝜇 ) = 𝜌(𝔤)(𝑉𝜇 ) ⊆ 𝑉𝜇 , that
is, 𝑉𝜇 is a 𝔤-subrepresentation as required. 

Definition 4.4.7. Let 𝔤 be a nilpotent Lie algebra and let (𝑉, 𝜌) be a representation of 𝔤. Say 𝑥 ∈ 𝔤 is
𝑉-generic if, for all 𝜆, 𝜇 ∈ Ψ𝑉 we have 𝜆(𝑥) = 𝜇(𝑥) if and only if 𝜆 = 𝜇.
It 𝐷𝑉 = {𝜆 − 𝜇 ∶ 𝜆, 𝜇 ∈ Ψ𝑉 }\{0}, then 𝑥 is 𝑉-generic if and only if 𝑥 ∉ ⋃𝜈∈𝐷 ker(𝜈). If k is infinite,5
𝑉
any nonzero vector space over k is not a union of finitely many hyperplanes, hence 𝑉-generic elements of
𝔤 exist for any finite-dimensional 𝔤-representation 𝑉.

Theorem 4.4.8. Let 𝔤 be a nilpotent Lie algebra and (𝑉, 𝜌) a finite-dimensional representation of 𝔤. For each
𝜆 ∈ (𝔤/𝐷𝔤)∗ , let

𝑊𝜆 = � 𝑉𝜆(𝑥),𝑥 , 𝑉𝜆(𝑥),𝑥 = {𝑣 ∈ 𝑉 ∶ ∃𝑛 > 0 such that (𝜌(𝑥) − 𝜆(𝑥))𝑛 (𝑣) = 0}.

𝑥∈𝔤

If 𝑥0 ∈ 𝔤 is 𝑉-generic, then we have 𝑉𝜆(𝑥0),𝑥0 = 𝑉𝜆 = 𝑊𝜆 and hence 𝑉 = ⨁ 𝑉𝜆 is the direct sum of its
𝜆
generalised weight spaces.

Proof. Since 𝔤 is nilpotent, it is solvable, hence for any 𝔤-representation (𝑈, 𝜎) its composition factors all
lie in Ψ𝑈 and, as in Definition 4.4.4, if 𝑥 ∈ 𝔤 then 𝜎(𝑥) has spectrum {𝜆(𝑥) ∶ 𝜆 ∈ Ψ𝑈 }. In particular, taking
𝑈 = 𝑉𝜆 we see that 𝜌(𝑥)|𝑉𝜆 has 𝜆(𝑥) as its sole eigenvalue, that is, 𝑉𝜆 ⊆ 𝑉𝜆(𝑥),𝑥 . It follows that 𝑉𝜆 ⊆ 𝑊𝜆 .
Now if 𝑥 ∈ 𝔤, we have 𝑉 = ⨁ 𝑉𝜆(𝑥),𝑥 and by Proposition 4.4.6 each 𝑉𝜆(𝑥),𝑥 is a 𝔤-
𝜆(𝑥)∶𝜆∈Ψ𝑉
subrepresentation of 𝑉, hence taking 𝑈 = 𝑉𝜆(𝑥),𝑥 we see that if k𝜈 is a composition factor, then 𝜈(𝑥) =
𝜆(𝑥). It follows that if we take 𝑥0 to be 𝑉-generic, the generalised eigenspace 𝑉𝜆(𝑥0),𝑥0 has 𝜆 as its unique
composition factor, so that 𝑉𝜆(𝑥0),𝑥0 ⊆ 𝑉𝜆 . Hence 𝑉𝜆(𝑥0),𝑥0 = 𝑉𝜆 = 𝑊𝜆 and 𝑉 = ⨁ 𝑉𝜆 .
𝜆∈Ψ𝑉


5
Any field k with char(k) = 0 contains a copy of ℚ and so is infinite. Alteratively, any algebraically closed field is infinite –
e.g. take the 𝑛-th roots of some 𝜇 ∈ k× where 𝑛 is taken coprime to char(k).

31
Chapter 5

Cartan subalgebras and the Cartan

criteria.

5.1 Nilpotent Lie algebras as measurements

In this section we work over an algebraically closed field k. In particular, k is infinite.

In the previous Chapter, we saw that the representation theory of a nilpotent Lie algebra 𝔫 is a nat-
ural generalization of the theory for 𝔤𝔩1 : any representation (𝑉, 𝜌) decomposes 𝑉 = ⨁ 𝑉𝜆 into a
𝜆∈Ψ𝑉
direct sum of subrepresentations 𝑉𝜆 , where Ψ𝑉 ⊆ 𝐷(𝔫)0 is the set of composition factors of 𝑉 as an 𝔫-
representation, and the subrepresentation 𝑉𝜆 is precisely the vectors which are simultaneous generalised
eigenvectors for every 𝜌(𝑥), (𝑥 ∈ 𝔤), with eigenvalue 𝜆(𝑥). In this Chapter, we will use this result to study
the structure of arbitrary finite-dimensional Lie algebras.
Thus suppose 𝔤 is an arbitrary Lie algebra and let 𝔫 be a nilpotent subalgebra (since any 1-dimensional
subspace is isomorphic to 𝔤𝔩1 as a Lie subalgebra, such subalgebras always exist).
Lemma 5.1.1. Let 𝔤 be a Lie algebra, 𝔫 ⊆ 𝔤 a nilpotent subalgebra, and (𝑉, 𝜌) a 𝔤-representation. Then the
𝔫-representations 𝔤 and 𝑉 decompose into direct sums
𝔤 = � 𝔤𝛼 , 𝑉 = � 𝑉𝜆 ,
𝛼∈Φ0 𝜆∈Ψ𝑉

where the 𝔤𝛼 , 𝑉𝜆 is the 𝛼 (resp. 𝜆) isotypical subrepresentation of 𝔤 (resp. 𝑉) as an 𝔫-representation. Moreover,

𝑖) 𝔫 ⊆ 𝔤0 ; 𝑖𝑖) [𝔤𝛼 , 𝑉𝜆 ] ⊆ 𝑉𝛼+𝜆 , (∀ 𝛼 ∈ Φ0 , 𝜆 ∈ Ψ𝑉 ).
In particular, using 𝑖𝑖) with 𝑉 = (𝔤, ad) in 𝑖𝑖) we see that 𝔤0 is a subalgebra and hence each 𝔤𝛼 , 𝑉𝜆 is a 𝔤0 -
subrepresentation of 𝔤 and 𝑉 respectively.

Proof. Restricting the adjoint representation of 𝔤 to 𝔫 we obtain the associated isotypical decomposi-
tion ⨁ 𝔤 . Since the adjoint representation of a nilpotent Lie algebra has the trivial representa-
𝛼∈Φ0 𝛼
tion as its only composition factor, 𝔫 ⊆ 𝔤0 . Finally, since 𝑎 ̃ ∶ 𝔤 ⊗ 𝑉 → 𝑉 is a homomorphism of 𝔤-
representations, it follows from Lemma 4.4.5 that [𝔤𝛼 , 𝑉𝜆 ] ⊆ 𝑉𝛼+𝜆 . Taking 𝑉 to be the adjoint rep-
resentation, this shows that [𝔤𝛼 , 𝔤𝛽 ] ⊆ 𝔤𝛼+𝛽 , which immediately implies that 𝔤0 is a subalgebra and
𝔤=⨁ 𝔤𝛼 , 𝑉 = ⨁ 𝑉𝜆 decompositions into 𝔤0 -subrepresentations. 
𝛼∈Φ0 𝜆∈Ψ𝑉

5.1.1 Cartan subalgebras

Now the fact that 𝔤 = ⨁ 𝔤 is a decomposition of 𝔤 as a 𝔤0 -representation suggests that to under-

𝛼∈Φ0 𝛼
stand the structure of 𝔤 using this decomposition we may need to understand the structure of representa-

32
tions of (not necessarily nilpotent) Lie algebra 𝔤0 . Since we already understand the representation theory
of 𝔫, and 𝔫 ⊆ 𝔤0 , one might hope that it was possible to arrange that 𝔤0 = 𝔫, so that the decomposition
𝔤=⨁ 𝔤𝛼 , so that the 𝔤𝛼 are not representations of any larger subalgebra of 𝔤.
𝛼∈Φ0

Lemma 5.1.2. 𝔫 = 𝔤0 if and only if 𝑁𝔤 (𝔫) = 𝔫.

Proof. First that 𝔤/𝔫 is a 𝔫-representation, and, since 𝔫 ⊆ 𝔤0 we see that 𝔤/𝔥 ≅ 𝔤0 /𝔥 ⊕ ⨁ 𝔤 (where we
𝛼∈Φ 𝛼
write Φ for Φ0 \{0}) and hence 𝔤0 /𝔥 = (𝔤/𝔥)0 is the isotypical summand of 𝔤/𝔥 corresponding to the trivial
representation of 𝔫. But now if 𝑥 + 𝔥 spans a copy of the trivial representation of 𝔫 in 𝔤/𝔫 then [ℎ, 𝑥] ∈ 𝔫
for all ℎ ∈ 𝔫, that is, 𝑥 ∈ 𝑁𝔤 (𝔫).
Thus we see that 𝔫 will be all of 𝔤0 if and only if 𝔫 = 𝑁𝔤 (𝔫), that is, if and only if 𝔫 is its own normalizer
in 𝔤. 

This motivates the following definition:

Definition 5.1.3. A subalgebra 𝔥 is said to be a Cartan subalgebra if it is

i) nilpotent and

ii) self-normalizing, that is, 𝑁𝔤 (𝔥) = 𝔥.

We will call a pair (𝔤, 𝔥) a Cartan pair if 𝔥 is a Cartan subalgebra of a Lie algebra 𝔤, and the the isotypical
decomposition 𝔤 = ⨁ 𝔤𝛼 of 𝔤 as an 𝔥-representation the Cartan decomposition of the pair. By Lemma
𝛼∈Φ0
5.1.2, 𝔥 = 𝔤0 and hence the decomposition is usually written 𝔤 = 𝔥 ⊕ ⨁ 𝔤 where Φ = Φ0 \{0}. The
𝛼∈Φ 𝛼
weights in Φ are called the roots of 𝔤.

Remark 5.1.4. It is not clear from this definition whether a Lie algebra necessarily contains a Cartan sub-
algebra. We will for the moment assume this result, in order to show how they provide a powerful tool to
study the structure of an arbitrary finite-dimensional Lie algebra. In fact, however, when k is algebraically
closed with char(k) = 0 it is known that the set of all Cartan subalgebras of a k-Lie algebra 𝔤 form a sin-
gle orbit under the group of inner automorphisms of 𝔤. This shows that the Cartan Decomposition of 𝔤 is
unique up to automorphisms of 𝔤.

5.1.2 Existence of Cartan subalgebras

Note that, if 𝔥 is a Cartan subalgebra of 𝔤, then the Theorem 4.4.8 shows that there is an 𝑥0 ∈ 𝔥 such that
𝔥 = 𝔤0,𝑥0 . This motivates the following definition:

Definition 5.1.5. If 𝑥 ∈ 𝔤, let 𝔤0,𝑥 be the generalized 0-eigenspace of ad(𝑥). Note that we always have
𝑥 ∈ 𝔤0,𝑥 so that dim(𝔤0,𝑥 ) ≥ 1. We say that 𝑥 ∈ 𝔤 is regular if 𝔤0,𝑥 is of minimal dimension.

Proposition 5.1.6. Let 𝔤 be a Lie algebra over a field k.

i) If 𝑥 ∈ 𝔤 is any element, then 𝔤0,𝑥 is a self-normalizing subalgebra of 𝔤.

ii) If 𝑥 ∈ 𝔤 is a regular element, then 𝔤0,𝑥 is a nilpotent and so a Cartan subalgebra of 𝔤.

Proof. Part 𝑖) is straight-forward: It follows immediately from Lemma 4.4.6 applied to the adjoint repre-
sentation that 𝔥 = 𝔤0,𝑥 is a subalgebra of 𝔤. To see that 𝔥 is a self-normalizing in 𝔤. Indeed if 𝑧 ∈ 𝑁𝔤 (𝔥) then
[𝑥, 𝑧] ∈ 𝔥 (since certainly 𝑥 ∈ 𝔥), so that for some 𝑛 we have ad(𝑥)𝑛 ([𝑥, 𝑧]) = 0, and hence ad(𝑥)𝑛+1 (𝑧) = 0
and 𝑧 ∈ 𝔥 as required.

33
 To establish part 𝑖𝑖), assume that 𝑥 is regular, and let 𝔥 = 𝔤0,𝑥 . To see that 𝔥 is nilpotent, by Engel’s
theorem it suffices to show that, for each 𝑦 ∈ 𝔥, the map ad(𝑦) is nilpotent as an endomorphism of 𝔥. To
see this, we consider the characteristic polynomials of ad(𝑦) on 𝔤, 𝔥 and 𝔤/𝔥: Since 𝔥 is a subalgebra of 𝔤,
the characteristic polynomial 𝜒𝑦 (𝑡) ∈ k[𝑡] of ad(𝑦) on 𝔤 is the product of the characteristic polynomials of
𝑦 𝑦
ad(𝑦) on 𝔥 and 𝔤/𝔥, which we will write as 𝜒1 (𝑡) and 𝜒2 (𝑡) respectively.
𝑛
We may write 𝜒𝑦 (𝑡) = ∑𝑘=0 𝑐𝑘 (𝑦)𝑡𝑘 , where 𝑛 = dim(𝔤). Pick {ℎ1 , ℎ2 , … , ℎ𝑟 } a basis of 𝔥 (so that
𝑟
dim(𝔥) = 𝑟). Then if we write 𝑦 = ∑𝑖=1 𝑦𝑖 ℎ𝑖 , the coefficients {𝑐𝑘 (𝑦)}𝑛𝑘=0 of 𝜒𝑦 (𝑡) are polynomial functions
of the coordinates {𝑦𝑖 ∶ 1 ≤ 𝑖 ≤ 𝑟}. Similarly we have
𝑟 𝑛−𝑟
𝑦 𝑦
𝜒1 (𝑡) = � 𝑑𝑖 (𝑦)𝑡𝑖 , 𝜒2 (𝑡) = � 𝑒𝑗 (𝑦)𝑡𝑗
𝑖=0 𝑗=0

𝑛
where the 𝑑𝑖 , 𝑒𝑗 ∈ k[𝑥1 , … , 𝑥𝑛 ] are polynomials and 𝑑𝑖 (𝑦) = 𝑑𝑖 (𝑦1 , … , 𝑦𝑛 ) where 𝑦 = ∑𝑖=1 𝑦𝑖 ℎ𝑖 . Since
ad(𝑥)(𝑥) = 0, we have 𝑥 ∈ 𝔤0,𝑥 . But ad(𝑥) is invertible on 𝔤/𝔥, since all its eigenvalues are non-zero on 𝔤/𝔥,
hence 𝜒𝑥2 (𝑡) has 𝑒0 (𝑥) ≠ 0, and thus the polynomial 𝑒0 is nonzero.
𝑦 𝑟−𝑠
Now let 𝑠 = min{𝑖 ∶ 𝑑𝑖 (𝑥1 , … , 𝑥𝑛 ) ≠ 0}. Then we may write 𝜒1 (𝑡) = 𝑡𝑠 ∑𝑘=0 𝑑𝑠+𝑘 (𝑦)𝑡𝑘 , and hence

𝜒𝑦 (𝑡) = 𝑡𝑠 (𝑑𝑠 + 𝑑𝑠+1 𝑡 + …)(𝑒0 + 𝑒1 .𝑡 + …) = 𝑡𝑠 𝑑𝑠 𝑒0 + … ,

For any endomorphism of a vector space, the dimension of its 𝜆-generalised eigenspace is the largest
power of (𝑡 − 𝜆) dividing its characteristic polynomial. In particular this implies that, for any 𝑦 ∈ 𝔥, we
have dim(𝔤0,𝑦 ) = min{𝑖 ∶ 𝑐𝑖 (𝑦) ≠ 0}. But since 𝑒0 .𝑑𝑠 ∈ k[𝑥1 , … , 𝑥𝑛 ] is nonzero, there is some 𝑧 ∈ 𝔥 such
that 𝑑𝑠 (𝑧).𝑒0 (𝑧) ≠ 0, and hence dim(𝔤0,𝑧 ) = 𝑠. Now by definition 𝑠 ≤ 𝑟 = dim(𝔤0,𝑥 ), hence since 𝑥 is
𝑦
regular, we must have 𝑠 = 𝑟, and hence 𝜒1 (𝑡) = 𝑡𝑟 , for all 𝑦 ∈ 𝔥. Hence every ad(𝑦) is nilpotent on 𝔥, so
that 𝔥 is a Cartan subalgebra as required.


In the course of the proof of the above Proposition we used the fact that the coefficients of the charac-
teristic polynomial were polynomial functions of the coordinates of 𝑦 ∈ 𝔥 with respect to a basis of 𝔥. This
was crucial because, whereas the product of two arbitrary nonzero functions may well be zero, the prod-
uct of two nonzero polynomials (over a field) is never zero. Lemma I.4 in Appendix I establishes a slightly
more general statemetn which applied to 𝑉 = 𝔤, 𝐴 = 𝔥 and 𝜑 = ad gives a proof of this polynomial
property.1

5.2 Trace forms and Cartan’s criterion for Solvability

In this section we introduce certain symmetric bilinear forms, which will play an important role in the rest
of the course. A brief review of the basic theory of symmetric bilinear forms2 is given in §II.2 in Appendix
1 of these notes.

5.2.1 Invariant bilinear forms

Let Bil(𝑉) be the space of bilinear forms on 𝑉, that is,

Bil(𝑉) = {𝐵 ∶ 𝑉 × 𝑉 → k ∶ 𝐵 bilinear}.
1
If this all seems overly pedantic then feel free to ignore it.
2
Part A Algebra focused more on positive definite and Hermitian forms, but there is a perfectly good theory of symmetric
bilinear forms over an arbitrary field k. When k is algebraically closed, the theory is also straight-forward!

34
From the definition of tensor products it follows that Bil(𝑉) can be identified with (𝑉⊗𝑉)∗ . The involution
𝜎 ∶ 𝑉 × 𝑉 → 𝑉 × 𝑉 given by (𝑣, 𝑤) ↦ (𝑤, 𝑣) induces an involution (which we will also denote by 𝜎) on
Bil(𝑉) and on 𝑉 ⊗ 𝑉. We say that a bilinear form 𝐵 is symmetric if 𝐵 ∘ 𝜎 = 𝐵, that is, if 𝐵(𝑣, 𝑤) = 𝐵(𝑤, 𝑣)
for all 𝑣, 𝑤 ∈ 𝑉.
If 𝑉 is a 𝔤-representation, the identification of Bil(𝑉) with (𝑉 ⊗ 𝑉)∗ shows that Bil(𝑉) also has the
structure of 𝔤-representation: explicitly, if 𝐵 ∈ Bil(𝑉), then it yields a linear map 𝑏 ∶ 𝑉 ⊗ 𝑉 → k by the
universal property of tensor products, and if 𝑦 ∈ 𝔤, it acts on 𝐵 as follows:

𝑦(𝐵)(𝑣, 𝑤) = 𝑦(𝑏)(𝑣 ⊗ 𝑤)
= −𝑏(𝑦(𝑣 ⊗ 𝑤))
= −𝑏(𝑦(𝑣) ⊗ 𝑤 + 𝑣 ⊗ 𝑦(𝑤))
= −𝐵(𝑦(𝑣), 𝑤) − 𝐵(𝑣, 𝑦(𝑤)).

Notice that the involution 𝜎 ∈ End(𝑉 ⊗ 𝑉) commutes with the action of 𝔤 (this is a special case of the
fact that, for any two 𝔤-representations, the map 𝜏 ∶ 𝑉 ⊗ 𝑊 → 𝑊 ⊗ 𝑉 given by 𝜏(𝑣 ⊗ 𝑤) = 𝑤 ⊗ 𝑣 is a
𝔤-homomorphism). It follows that the action of 𝔤 preserves the space 𝑆2 (𝑉) of symmetric bilinear forms.

Definition 5.2.1. We say that a bilinear form 𝐵 ∈ Bil(𝔤) is invariant if it is an invariant vector for the
action of 𝔤 on Bil(𝔤) ≅ (𝔤 ⊗ 𝔤)∗ , that is, if 𝐵(ad(𝑥)(𝑦), 𝑧) = 𝐵(𝑦, −𝑎𝑑(𝑥)(𝑧)) = 0 for all 𝑥, 𝑦, 𝑧 ∈ 𝔤. This is
often written as
𝐵([𝑥, 𝑦], 𝑧) = 𝐵(𝑥, [𝑦, 𝑧]), ∀ 𝑥, 𝑦, 𝑧 ∈ 𝔤.

Remark 5.2.2. If (𝑉, 𝜌) is a 𝔤-representation and 𝐵 ∈ Bil(𝑉) is a bilinear form, then it defines a linear
map 𝜃 ∶ 𝑉 → 𝑉 ∗ where 𝜃(𝑣)(𝑤) = 𝐵(𝑣, 𝑤) (∀ 𝑣, 𝑤 ∈ 𝑉). If 𝐵 is invariant, that is 𝐵 ∈ Bil(𝑉)𝔤 , then
we have 𝜃(𝜌(𝑥)(𝑣))(𝑤) = 𝐵(𝜌(𝑥)(𝑣), 𝑤) = 𝐵(𝑣, −𝜌(𝑥)(𝑤)) = 𝜌∗ (𝑥)(𝜃(𝑣))(𝑤) for all 𝑣, 𝑤 ∈ 𝑉, hence
𝜃(𝜌(𝑥)(𝑣)) = 𝜌(𝑥)∗ (𝜃(𝑣)), that is, 𝜃 ∈ ℒ(𝑉, 𝑉 ∗ )𝔤 = ℒ𝔤 (𝑉, 𝑉 ∗ ) is a homomorphism of 𝔤-representations.
If 𝜃 is an isomorphism, we say 𝐵 is nondegenerate and in that case, for any linear map 𝛼 ∈ End(𝑉) we
may define 𝛼∗ = 𝜃−1 ∘ 𝛼⊺ ∘ 𝜃 ∈ End(𝑉), the adjoint of 𝛼 with respect to 𝐵. If 𝑉 is a 𝔤-representation
and 𝐵 is nondegenerate, then the condition that 𝐵 is invariant can be expressed as 𝜌(𝑥)∗ = −𝜌(𝑥) for all
𝑥 ∈ 𝔤, where 𝜌 ∶ 𝔤 → 𝔤𝔩𝑉 is the action map, that is, 𝜌(𝔤) consists of skew-adjoint endomorphisms of with
respect to the bilinear form 𝐵.

Definition 5.2.3. If 𝛼 ∶ 𝔤1 → 𝔤2 is a homomorphism of Lie algebras, and 𝐵 is a bilinear form on 𝔤2 , then

we may “pull-back” 𝐵 using 𝛼 to obtain a bilinear form on 𝔤1 . Indeed viewing 𝐵 as an element of (𝔤2 ⊗𝔤2 )∗ ,
we obtain an element 𝛼∗ (𝐵) of (𝔤1 ⊗ 𝔤1 )∗ given by 𝛼∗ (𝐵)(𝑥, 𝑦) = 𝐵(𝛼(𝑥), 𝛼(𝑦)). It is immediate from the
definitions that if 𝐵 is an invariant form for 𝔤2 , then 𝛼∗ (𝐵) is an invariant form for 𝔤1 .

It follows that if we can find an invariant form 𝑏𝑉 on a general linear Lie algebra 𝔤𝔩𝑉 , then any rep-
resentation 𝜌 ∶ 𝔤 → 𝔤𝔩𝑉 of a Lie algebra 𝔤 on 𝑉 will yield an invariant bilinear form 𝑡𝑉 = 𝜌∗ (𝑏𝑉 ) on 𝔤.
The next Lemma shows that there is in fact a very natural invariant bilinear form, indeed an invariant
symmetric bilinear form, on a general linear Lie algebra 𝔤𝔩𝑉 :

Lemma 5.2.4. Let 𝑉 be a k-vector space. The trace form 𝑏𝑉 ∶ 𝔤𝔩𝑉 ⊗ 𝔤𝔩𝑉 → k given by

𝑏𝑉 (𝑎, 𝑏) = tr(𝑎.𝑏), ∀ 𝑎, 𝑏 ∈ 𝔤𝔩𝑉 ,

is a nondegenerate invariant symmetric bilinear form on 𝔤𝔩𝑉 .

Proof. Let 𝜗 ∶ 𝑉 ⊗ 𝑉 ∗ → 𝔤𝔩𝑉 be give by 𝜗(𝑤 ⊗ 𝑓) = 𝑓.𝑤 ∈ 𝔤𝔩𝑉 , where (𝑓.𝑤)(𝑣) = 𝑓(𝑣).𝑤 for all 𝑣 ∈ 𝑉.
For 𝑉 finite-dimensional, this map is an isomorphism and if 𝜄 ∶ 𝑉 ⊗ 𝑉 ∗ → k is the natural “contraction”
map induced by the evaluation map (𝑣, 𝑓) ↦ 𝑓(𝑣), then tr(𝜗(𝑣 ⊗ 𝑓)) = tr(𝑓.𝑣) = 𝑓(𝑣) = 𝜄(𝑣 ⊗ 𝑓) (see
Lemma II.10 for details).

35
 Moreover, the isomorphism 𝜗 ⊗ 𝜗 ∶ 𝑉 ⊗ 𝑉 ∗ ⊗ 𝑉 ⊗ 𝑉 ∗ → 𝔤𝔩𝑉 ⊗ 𝔤𝔩𝑉 identifies the composition map
(𝑎, 𝑏) → 𝑎 ∘ 𝑏 with the contraction map on the 2nd and 3rd factors. Indeed for any 𝑣1 , 𝑣2 ∈ 𝑉, 𝑓1 , 𝑓2 ∈ 𝑉 ∗
we have

𝜗(𝑣1 ⊗ 𝑓1 ) ∘ 𝜗(𝑣2 ⊗ 𝑓2 ) = (𝑓1 .𝑣1 ) ∘ (𝑓2 .𝑣2 ) = 𝑓1 (𝑣2 ).(𝑓2 .𝑣1 ) = 𝜗(𝜄23 (𝑣1 ⊗ 𝑓1 ⊗ 𝑣2 ⊗ 𝑓2 ))

Thus we see that (𝑎, 𝑏) ↦ tr(𝑎𝑏) corresponds under 𝜗 ⊗ 𝜗 to the map 𝜄14 ⊗ 𝜄32 ∶ 𝑉 ⊗ 𝑉 ∗ ⊗ 𝑉 ⊗ 𝑉 ∗ → k,
where we write 𝜄𝑘𝑙 for the contraction map acting on the 𝑘-th and 𝑙-th tensor factors if the 𝑘th is 𝑉 and
the 𝑙th is 𝑉 ∗ . The composition (𝑎, 𝑏) ↦ 𝑎𝑏 gives the contraction 𝜄23 and then taking trace corresponds to
the contraction 𝜄14 . Taking tr(𝑏𝑎) gives the same value since (𝑎, 𝑏) ↦ tr(𝑏𝑎) simply contracts the factors
in the opposite order, so that tr(𝑎𝑏) = tr(𝑏𝑎) and thus 𝑏𝑉 is a symmetric bilinear form. To show it is
invariant, since 𝑏𝑉 = 𝜄14 ⊗ 𝜄23 it suffices to check that 𝜄 is invariant. But this is clear, since 𝑥(𝑣 ⊗ 𝑓) =
𝑥(𝑣) ⊗ 𝑓 − 𝑣 ⊗ (𝑓 ∘ 𝑥), thus 𝜄(𝑥(𝑣 ⊗ 𝑓)) = 𝑓(𝑥(𝑣)) − (𝑓 ∘ 𝑥)(𝑣) = 0, while 𝑥(𝜄(𝑣 ⊗ 𝑓)) = 𝑥(𝑓(𝑣)) = 0, since 𝔤𝔩𝑉
acts by 0 on k, the trivial representation. We leave it as an exercise to check the nondegeneracy of 𝑏𝑉 . 

Remark 5.2.5. One can also of course check the invariance property by a direct calculation: for 𝑎, 𝑏, 𝑐 ∈
𝔤𝔩𝑉 we have

tr([𝑎, 𝑏].𝑐) = tr((𝑎𝑏 − 𝑏𝑎).𝑐) = tr(𝑎.(𝑏𝑐)) − tr(𝑏.(𝑎𝑐))

= tr(𝑎.(𝑏𝑐)) − tr((𝑎𝑐).𝑏)
= tr(𝑎.(𝑏𝑐 − 𝑐𝑏)
= tr(𝑎, [𝑏, 𝑐]).

where going from the first to the second line we used the symmetry property of tr to replace tr(𝑏.(𝑎𝑐)) with
tr((𝑎𝑐).𝑏).

Definition 5.2.6. If 𝔤 is a Lie algebra, and let (𝑉, 𝜌) be a representation of 𝔤. we may define a bilinear
form 𝑡𝑉 ∶ 𝔤 × 𝔤 → k on 𝔤, known as a trace form of the representation (𝑉, 𝜌), to be 𝜌∗ (𝑏𝑉 ). Explicitly, we
have
𝑡𝑉 (𝑥, 𝑦) = tr𝑉 (𝜌(𝑥)𝜌(𝑦)), ∀ 𝑥, 𝑦 ∈ 𝔤.

In fact the trace form of a representation depends only on the composition factors of the representa-
tion:

Lemma 5.2.7. Let (𝑉, 𝜌) If {𝑆𝑖 ∶ 1 ≤ 𝑖 ≤ 𝑘} are the composition factors and [𝑉 ∶ 𝑆𝑖 ] denotes the multiplicity of
the composition factor 𝑆𝑖 , then
𝑘
𝑡𝑉 = �[𝑉 ∶ 𝑆𝑖 ]𝑡𝑆𝑖
𝑖=1

Proof. Pick a composition series 𝒞 = (𝐹𝑘 )𝑑𝑘=0 for 𝑉 and let 𝔟𝒞 = {𝑥 ∈ 𝔤𝔩𝑉 ∶ 𝑥(𝐹𝑘 ) ⊆ 𝐹𝑘 , ∀ 𝑘, 0 ≤ 𝑘 ≤ 𝑑}.
Then 𝜌(𝔤) ⊆ 𝔟𝒞 and since 𝔟𝒞 is an associative subalgebra of End(𝑉), for any 𝑥, 𝑦 ∈ 𝔤, the composition
𝜌(𝑥)𝜌(𝑦) ∈ 𝔟𝒞 and the result then follows from Lemma 2.2.14 where we set 𝛼 = 𝜌(𝑥)𝜌(𝑦). 

Definition 5.2.8. The Killing form 𝜅 ∶ 𝔤 × 𝔤 → k is the trace form given by the adjoint representation,
that is:
𝜅(𝑥, 𝑦) = tr(ad(𝑥)ad(𝑦)).

Note that if 𝔞 ⊆ 𝔤 is a subalgebra, the Killing form of 𝔞 is not necessarily equal to the restriction of that
of 𝔤. We will write 𝜅𝔤 when it is not clear from context which Lie algebra is concerned.

If 𝔞 is an ideal in 𝔤, then in fact the Killing form is unambiguous, as the following Lemma shows.

36
Lemma 5.2.9. Let 𝔞 be an ideal of 𝔤. The Killing form 𝜅𝔞 of 𝔞 is given by the restriction of the Killing form 𝜅𝔤 on
𝔤, that is:
𝜅𝔤|𝔞 = 𝜅𝔞 .
Moreover, the subspace orthogonal to 𝔞, that is, 𝔞⟂ = {𝑥 ∈ 𝔤 ∶ 𝜅(𝑥, 𝑦) = 0, ∀ 𝑦 ∈ 𝔞} is also an ideal.

Proof. If 𝑎 ∈ 𝔞 we have ad(𝑎)(𝔤) ⊆ 𝔞, thus the same will be true for the composition ad(𝑎1 )ad(𝑎2 ) for any
𝑎1 , 𝑎2 ∈ 𝔞. Thus if we pick a vector space complement 𝑊 to 𝔞 in 𝔤, the matrix of ad(𝑎1 )ad(𝑎2 ) with respect
to a basis compatible with the subspaces 𝔞 and 𝑊 will be of the form
𝐴 𝐵
� 0 0. �

where 𝐴 ∈ End(𝔞) and 𝐵 ∈ Homk (𝔞, 𝑊). Then clearly tr(ad(𝑎1 )ad(𝑎2 )) = tr(𝐴). Since 𝐴 is clearly given
by ad(𝑎1 )|𝔞 ad(𝑎2 )|𝔞 , we are done. To see that 𝔞⟂ is an ideal, we must check that for any 𝑥 ∈ 𝔤 and 𝑦 ∈ 𝔞⟂
we have [𝑥, 𝑦] ∈ 𝔞⟂ . But if 𝑎 ∈ 𝔞 then 𝜅(𝑎, [𝑥, 𝑦]) = 𝜅([𝑎, 𝑥], 𝑦) = 0 since [𝑎, 𝑥] ∈ 𝔞.


5.2.2 Cartan criterion for solvable Lie algebras

For the rest of this section k is an algebraically closed field of characteristic zero.
We now wish to show how the Killing form yields a criterion for determining whether a Lie algebra
is solvable or not. For this we need a couple of technical preliminaries. Recall that, if (𝑉, 𝜌) is a repre-
sentation of a nilpotent Lie algebra 𝔥, then it decomposes as the direct sum 𝑉 = ⨁ 𝑉𝜆 , where
𝜆∈Ψ𝑉
Ψ𝑉 ⊆ 𝐷(𝔥)0 denotes the set of one-dimensional representations of 𝔥 which occur as composition factors
of 𝑉, and 𝑉𝜆 is the maximal subrepresentation of 𝑉 whose only composition factor is k𝜆 . When (𝔤, 𝔥) is
a Cartan pair and we view 𝔤 as a 𝔥 representation via the inclusion 𝔥 ⊆ 𝔤, the set Ψ𝔤 = {0} ∪ Φ, where
𝔤0 = 𝔥.
Definition 5.2.10. Let (𝔤, 𝔥) be a Cartan pair and let 𝔤 = 𝔥 ⊕ ⨁ 𝔤 be the associated decomposition
𝛼∈Φ 𝛼
of 𝔤. For each 𝛼 ∈ Φ we set 𝔥𝛼 = [𝔤𝛼 , 𝔤−𝛼 ] ⊆ 𝔥. Note that if −𝛼 ∉ Φ, then 𝔥𝛼 = {0}.
Lemma 5.2.11. Let (𝔤, 𝔥) be a Cartan pair and let Φ ⊆ 𝐷(𝔥)0 be the roots of 𝔤 associated to its Cartan de-
composition. Let (𝑉, 𝜌) be a 𝔤-representation and let 𝑉 = ⨁ 𝑉𝜆 be its decomposition into its isotypical
𝜆∈Ψ𝑉
summands as an 𝔥-representation. If 𝛼 ∈ Φ and 𝔥𝛼 ⊆ 𝔥 is as in Definition 5.2.10, then for any 𝜆 ∈ Ψ𝑉 , there is
an 𝑟𝜆 ∈ ℚ such that 𝜆|𝔥𝛼 = 𝑟𝜆 .𝛼|𝔥𝛼 .

Proof. The set of weights Ψ is finite, thus there are positive integers 𝑝, 𝑞 such that 𝑉𝜆+𝑡𝛼 ≠ 0 only for
integers 𝑡 with −𝑝 ≤ 𝑡 ≤ 𝑞; in particular, 𝜆−(𝑝+1)𝛼 ∉ Ψ and 𝜆+(𝑞+1)𝛼 ∉ Ψ. Let 𝑀 = ⨁ 𝑉𝜆+𝑡𝛼 .
−𝑝≤𝑡≤𝑞
If 𝑧 ∈ [𝔤𝛼 , 𝔤−𝛼 ] is of the form [𝑥, 𝑦] where 𝑥 ∈ 𝔤𝛼 , 𝑦 ∈ 𝔤−𝛼 then by Lemma 5.1.1,
𝜌(𝑥)(𝑉𝜆+𝑞𝛼 ) ⊆ 𝑉𝜆+(𝑞+1)𝛼 = {0}, 𝜌(𝑦)(𝑉𝜆−𝑝𝛼 ) ⊆ 𝑉𝜆−(𝑝+1)𝛼 = {0}
we see that 𝜌(𝑥) and 𝜌(𝑦) preserve 𝑀. Thus the action of 𝜌(𝑧) on 𝑀 is the commutator of the action of
𝜌(𝑥) and 𝜌(𝑦) on 𝑀, and so tr(𝜌(𝑧), 𝑀) = 0. On the other hand, we may also compute the trace of 𝜌(𝑧)
on 𝑀 directly: for any ℎ ∈ 𝔥, 𝜌(ℎ) acts on an isotypical summand 𝑉𝜇 with unique eigenvalue 𝜇(ℎ), hence
tr𝑉𝜇 (𝜌(ℎ)) = dim(𝑉𝜇 ).𝜇(ℎ). Applying this to 𝜌(𝑧) we find

0 = tr(𝜌(𝑧), 𝑀) = � tr(𝜌(𝑧), 𝑉𝜆+𝑡𝛼 ) = � dim(𝑉𝜆+𝑡𝛼 )(𝜆 + 𝑡𝛼)(𝑧)

−𝑝≤𝑡≤𝑞 −𝑝≤𝑡≤𝑞
⎛ ⎞
⎜⎜ ⎟⎟
= dim(𝑀).𝜆(𝑧) + ⎜⎜⎝ � 𝑡.dim(𝑉𝜆+𝑡𝛼 )⎟⎟⎠ 𝛼(𝑧)
−𝑝≤𝑡≤𝑞

37
Since dim(𝑀) ≥ dim(𝑉𝜆 ) > 0, this can be rearranged to give 𝜆(𝑧) = 𝑟𝜆 .𝛼(𝑧) as required. 

Definition 5.2.12. Let 𝔤 be a Lie algebra over a field k. We say that 𝔤 is perfect if it satisfies 𝔤 = 𝐷(𝔤) =
[𝔤, 𝔤]. A perfect Lie algebra therefore has no nontrivial abelian quotients.
Proposition 5.2.13. Let 𝑉 be a finite-dimensional k-vector space and let 𝑏𝑉 ∶ 𝔤𝔩𝑉 × 𝔤𝔩𝑉 → k, be the trace form,
𝑏𝑉 (𝑥, 𝑦) = tr𝑉 (𝑥𝑦), for all 𝑥, 𝑦 ∈ 𝔤𝔩𝑉 . If 𝔞 is a non-zero perfect subalgebra of 𝔤𝔩𝑉 then there is an 𝑥 ∈ 𝔞 such that
𝑏𝑉 (𝑥, 𝑥) ≠ 0, so that 𝑏𝑉 does not vanish identically on 𝔞.

Proof. Suppose that 𝔥 is a Cartan subalgebra of 𝔞 so that that 𝔞 = ⨁ 𝔞 is the associated Cartan
𝜆∈Φ∪{0} 𝜆
decomposition, where 𝔥 = 𝔞0 . Now 𝔞 = 𝐷𝔞 so that 𝔞 is not nilpotent, thus Φ ≠ ∅. Now if the trivial
representation was the only composition factor of 𝑉, the same would be true for 𝑉 ∗ and 𝔤𝔩𝑉 , hence if we
let 𝑉 = ⨁ 𝑉𝜇 be the decomposition of 𝑉 into generalised 𝔥-weight spaces as in Theorem 4.4.8, there
𝜇∈Ψ
must be some non-zero 𝜆 ∈ Ψ𝑉 . Next observe that

𝔞 = 𝐷𝔞 = [𝔞, 𝔞] = � � 𝔞𝜆 , � 𝔞𝜇 � = �[𝔞𝜆 , 𝔞𝜇 ].
𝜆∈Φ∪{0} 𝜇∈Φ∪{0} 𝜆,𝜇

Since we know that [𝔞𝜆 , 𝔞𝜇 ] ⊆ 𝔞𝜆+𝜇 , and moreover 𝔥 = 𝔞0 , it follows that we must have

𝔥 = [𝔥, 𝔥] + �[𝔞𝛼 , 𝔞−𝛼 ] = 𝐷(𝔥) + � 𝔥𝛼

𝛼 𝛼

where the sum runs over those roots 𝛼 such that −𝛼 ∈ Φ. But by definition, 𝜆 vanishes on 𝐷𝔥, so that
there must be some 𝛼 ∈ Φ with 𝜆(𝔥𝛼 ) ≠ 0. For such an 𝛼, let 𝑥 ∈ 𝔥𝛼 = [𝔞𝛼 , 𝔞−𝛼 ] be such that 𝜆(𝑥) ≠ 0.
Then we have
𝑏𝑉 (𝑥, 𝑥) = tr(𝑥2 ) = � dim(𝑉𝜇 )𝜇(𝑥)2 .
𝜇∈Ψ

But now by Lemma 5.2.11 for each 𝜇 ∈ Ψ there is an 𝑟𝜇 ∈ ℚ such that 𝜇(𝑥) = 𝑟𝜇 .𝛼(𝑥) for all 𝑥 ∈ [𝔞𝛼 , 𝔞−𝛼 ].
In particular, 0 ≠ 𝜆(𝑥) = 𝑟𝜆 𝛼(𝑥) so that 𝑟𝜆 ≠ 0 and 𝛼(𝑥) ≠ 0. Hence we see that
⎛ ⎞
⎜⎜ ⎟⎟
𝑡𝑉 (𝑥, 𝑥) = ⎜⎜⎝ � dim(𝑉𝜇 )𝑟2𝛼,𝜇 ⎟⎟⎠ 𝛼(𝑥)2 .
𝜇∈Ψ

Since the terms in the sum are nonnegative, and the term corresponding to 𝜆 is positive, we conclude
𝑡𝑉 (𝑥, 𝑥) ≠ 0 are required.


Recall that for any finite dimensional Lie algebra, the derived series stablizes to an ideal which we
denote as 𝐷∞ (𝔤). It has the property that it is equal to its own derived subalgebra, i.e. 𝐷∞ (𝔤) is perfect.
Theorem 5.2.14. Let 𝔤 be a Lie algebra and let (𝑉, 𝜌) be a 𝔤-representation. Then if 𝑡𝑉 vanishes on 𝐷(𝔤) then
𝜌(𝔤) is solvable, or equivalently, 𝐷∞ (𝔤) ⊆ ker(𝜌).

Proof. Since 𝜌(𝐷(𝔤)) = 𝐷(𝜌(𝔤)), replacing 𝔤 by its image 𝜌(𝔤), we may assume that 𝔤 ⊆ 𝔤𝔩𝑉 and 𝑏𝑉 van-
ishes on 𝐷(𝔤). We must show that 𝔤 is solvable, that is 𝐷∞ (𝔤) = {0}. But if this is not the case, then
setting 𝔞 = 𝐷∞ (𝔤) it follows that 𝔞 is a non-zero perfect subalgebra of 𝔤𝔩𝑉 . But then the previous Propo-
sition shows there is some 𝑥 ∈ 𝔞 for which 𝑏𝑉 (𝑥, 𝑥) ≠ 0. But by assumption 𝑏𝑉 vanishes identically on
𝐷(𝔤) ⊇ 𝔞, which gives a contradiction. 

Corollary 5.2.15. (Cartan’s Criterion for Solvability) Let 𝔤 be a (finite-dimensional) k-Lie algebra and let 𝜅
denote its Killing form. Then the following are equivalent:

i) 𝜅(𝐷(𝔤), 𝐷(𝔤)) = 0,

38
 ii) 𝔤 is solvable,

iii) 𝜅(𝔤, 𝐷(𝔤)) = 0, that is, 𝐷(𝔤) ⊆ rad(𝜅).

Proof. Clearly 𝑖𝑖𝑖) ⟹ 𝑖) so it suffices to show 𝑖) ⟹ 𝑖𝑖) and 𝑖𝑖) ⟹ 𝑖𝑖𝑖).

For 𝑖) ⟹ 𝑖𝑖) note that 𝑖) is the hypothesis of Theorem 5.2.14 for the representation (𝔤, ad), and
hence the theorem shows that 𝐷∞ (𝔤) ⊆ ker(ad) = 𝔷(𝔤). But since 𝐷∞ (𝔤) is perfect, this shows that
𝐷∞ (𝔤) = 𝐷(𝐷∞ (𝔤)) ⊆ 𝐷(𝔷(𝔤)) = {0} and hence 𝑖𝑖) holds.
For 𝑖𝑖) ⟹ 𝑖𝑖𝑖) note that if 𝔤 is solvable, by Lie’s theorem if 𝒞 = (𝑉 = 𝐹0 > 𝐹1 > … > 𝐹𝑑 = {0}) is a
composition series for (𝔤, ad) then each subquotient 𝐹𝑖 /𝐹𝑖+1 must be one-dimensional, i.e. 𝒞 is a complete
flag. Pick a basis 𝐵 = {𝑒1 , … , 𝑒𝑑 } of 𝔤 such that 𝐹𝑘 = ⟨𝑒𝑖 ∶ 𝑖 ≤ 𝑑 − 𝑘⟩k , and let 𝛼 ↦ [𝛼]𝐵 denote the
isomorphism 𝔤𝔩𝔤 → 𝔤𝔩𝑑 (k) where [𝛼]𝐵 = (𝛼𝐵𝑖𝑗 )1≤𝑖,𝑗≤𝑑 . This isomorphism identifies 𝔟𝒞 ⊃ 𝐷(𝔟𝒞 ) = 𝔫𝒞 with
𝔟𝑑 ⊃ 𝔫𝑑 the space of upper-triangular matrices and strictly upper triangular matrices in 𝔤𝔩𝑑 respectively.
𝑑
In particular, for any 𝑥, 𝑦 ∈ 𝔟𝒞 we have tr𝔤 (𝑥𝑦) = ∑𝑖=1 𝑥𝐵𝑖𝑖 .𝑦𝐵𝑖𝑖 , and in particular tr𝔤 (𝑥𝑦) = 0 if 𝑥 ∈ 𝔟𝒞 , 𝑦 ∈
𝔫𝒞 . But as 𝒞 is a composition series for 𝔤, we have ad(𝔤) ⊆ 𝔟𝒞 , and hence ad(𝐷(𝔤)) = 𝐷(ad(𝔤)) ⊆
𝐷(𝔟𝒞 ) = 𝔫𝒞 . It follows immediately that 𝜅(𝔤, 𝐷(𝔤)) ⊆ tr𝔤 (𝔟𝒞 , 𝔫𝒞 ) = 0 as required. 

5.3 Semisimple Lie algebras

In this section we assume that our field k is algebraically closed of characteristic zero, and all representations are
assumed to be finite dimensional over k.

5.3.1 The solvable radical

Suppose that 𝔤 is a Lie algebra, and 𝔞 and 𝔟 are solvable Lie ideals of 𝔤. It is easy to see that 𝔞 + 𝔟 is again
solvable (for example, because 0 ⊆ 𝔞 ⊆ 𝔞 + 𝔟, and 𝔞 and (𝔞 + 𝔟)/𝔞 ≅ 𝔟/(𝔞 ∩ 𝔟) are both solvable). It
follows that if 𝔤 is finite dimensional, then it has a largest solvable ideal 𝔯. Note that this is in the strong
sense: every solvable ideal of 𝔤 is a subalgebra of 𝔯 (c.f. Definition 4.4.1 where the same strategy was used
to define the subrepresentation 𝑉𝑆 of a 𝔤-representation given an irreducible representation 𝑆 of 𝔤).

Definition 5.3.1. Let 𝔤 be a finite dimensional Lie algebra. The largest solvable ideal 𝔯 of 𝔤 is known as
the (solvable) radical of 𝔤, and will be denoted rad(𝔤). We say that 𝔤 is semisimple if rad(𝔤) = 0, that is, if 𝔤
contains no non-zero solvable ideals.

Lemma 5.3.2. The Lie algebra 𝔤/rad(𝔤) is semisimple, that is, it has zero radical.

Proof. Suppose that 𝔰 is a solvable ideal in 𝔤/rad(𝔤). Then if 𝔰′ denotes the preimage of 𝔰 in 𝔤, we see that
𝔰′ is an ideal of 𝔤, and moreover it is solvable since rad(𝔤) and 𝔰 = 𝔰′ /rad(𝔤) as both solvable. But then by
definition we have 𝔰′ ⊆ rad(𝔤) so that 𝔰′ = rad(𝔤) and 𝔰 = 0 as required. 

Example 5.3.3. The Lemma shows that any Lie algebra 𝔤 contains a canonical solvable ideal rad(𝔤) such
that 𝔤/rad(𝔤) is a semisimple Lie algebra. Thus we have a short exact sequence:
0 / rad(𝔤) /𝔤 / 𝔤/rad(𝔤) / 0,

so that any Lie algebra is an extension of the semisimple Lie algebra 𝔤/rad(𝔤) by the solvable Lie algebra
rad(𝔤).

In characteristic zero, every Lie algebra 𝔤 is built out of rad(𝔤) and 𝔤/rad(𝔤) as a semidirect product.

Theorem 5.3.4. (Levi’s theorem) Let 𝔤 be a finite dimensional Lie algebra over a field k of characteristic zero, and
let 𝔯 be its radical. Then there exists a subalgebra 𝔰 of 𝔤 such that 𝔤 ≅ 𝔯 ⋉ 𝔰. In particular 𝔰 ≅ 𝔤/𝔯 is semisimple.

39
5.3.2 Cartan’s Criterion for semisimplicity

The Killing form gives us a way of detecting when a Lie algebra is semisimple. Recall that, given a sym-
metric bilinear form 𝐵 ∶ 𝑉 × 𝑉 → k, the radical of 𝐵 is

rad(𝐵) = {𝑣 ∈ 𝑉 ∶ ∀ 𝑤 ∈ 𝑉, 𝐵(𝑣, 𝑤) = 0} = 𝑉 ⟂ .

The form 𝐵 said to be nondegenerate if rad(𝐵) = {0}. We first note the following simple result.
Lemma 5.3.5. A finite dimensional Lie algebra 𝔤 is semisimple if and only if it does not contain any non-zero
abelian ideals.

Proof. Clearly if 𝔤 contains an abelian ideal, it contains a solvable ideal, so that rad(𝔤) ≠ 0. Conversely, if
𝔰 is a non-zero solvable ideal in 𝔤, then the last term in the derived series of 𝔰 will be an abelian ideal of
𝔤. 

We have the following simple characterisation of semisimple Lie algebras.

Theorem 5.3.6. A Lie algebra 𝔤 is semisimple if and only if the Killing form is nondegenerate.

Proof. Let 𝔤⟂ = {𝑥 ∈ 𝔤 ∶ 𝜅(𝑥, 𝑦) = 0, ∀ 𝑦 ∈ 𝔤}. Then by Lemma 5.2.9 𝔤⟂ is an ideal in 𝔤, and clearly the
restriction of 𝜅 to 𝔤⟂ is zero, so by Cartan’s Criterion, and Lemma 5.2.9 the ideal 𝔤⟂ is solvable. It follows
that if 𝔤 is semisimple we must have 𝔤⟂ = {0} and hence 𝜅 is non-degenerate.
Conversely, suppose that 𝜅 is non-degenerate. To show that 𝔤 is semisimple it is enough to show
that any abelian ideal of 𝔤 is trivial, thus suppose that 𝔞 is an abelian ideal, and pick 𝑊 a complementary
subspace to 𝔞 so that 𝔤 = 𝔞 ⊕ 𝑊. With respect to this decomposition, if 𝑥 ∈ 𝔤 and 𝑎 ∈ 𝔞, we have

𝑥1 𝑥2 0 𝑎2 ℒk (𝔞, 𝔞) ℒk (𝑊, 𝔞)
ad(𝑥) = � , ad(𝑎) = � ∈ .
0 𝑥3 � 0 0 � � ℒk (𝔞, 𝑊) ℒk (𝑊, 𝑊) �

0 𝑥1 𝑎2
But then we see that ad(𝑥) ∘ ad(𝑎) = � , and hence tr(ad(𝑥)ad(𝑎)) = 0. It follows that 𝔞 ⊆ 𝔤⟂ =
0 0 �
{0} as 𝜅 is non-degenerate and hence 𝔞 = {0} as required. 

Remark 5.3.7. It is worth noting that the proof of the previous theorem establishes two facts: first, that
𝔤⟂ is a solvable ideal in 𝔤 for any Lie algebra 𝔤, and secondly, that any abelian ideal of 𝔤 is contained in
𝔤⟂ . Combined with the previous Lemma this shows that 𝔤⟂ = {0} ⟺ rad(𝔤) = {0}, but in general the
containment 𝔤⟂ ⊆ rad(𝔤) need not be an equality.

5.3.3 Simple and semisimple Lie algebras

Definition 5.3.8. Recall from Definition 3.1.1 that a Lie algebra 𝔤 is said to be almost simple if it has no
non-trivial proper ideals. We say that 𝔤 is simple if it is nonabelian and has no nontrivial proper ideal,
i.e. 𝔤 is almost simple and nonabelian. We now show that this notion is closed related to our notion of a
semisimple Lie algebra.
Lemma 5.3.9. Let 𝑉 be a k-vector space equipped with a symmetric bilinear form 𝐵. Then for any subspace 𝑈
of 𝑉 we have

i) dim(𝑈) + dim(𝑈 ⟂ ) ≥ dim(𝑉),

ii) the restriction 𝐵|𝑈 is non-degenerate if and only if 𝑈 ⊕ 𝑈 ⟂ = 𝑉.

Proof. See Lemma II.20 in Appendix II for a proof. 

40
Proposition 5.3.10. Let 𝔤 be a Lie algebra, and let 𝐼 be an ideal of 𝔤.

i) If 𝔤 is semisimple then 𝔤 = 𝐼 ⊕ 𝐼 ⟂ , and both 𝐼 and 𝐼 ⟂ are semisimple, hence any ideal and any quotient of
𝔤 is semisimple.

ii) 𝐼 is semisimple if and only if 𝔤 = 𝐼 ⊕ 𝐼 ⟂ .

Proof. For part 𝑖) consider 𝐼 ∩ 𝐼 ⟂ . The Killing form 𝜅 of 𝔤 vanishes identically on 𝐼 ∩ 𝐼 ⟂ by definition, and
since it is an ideal, the Killing form of 𝐼 ∩ 𝐼 ⟂ is just the restriction of the Killing form of 𝔤. It follows from
Cartan’s Criterion that 𝐼∩𝐼 ⟂ is solvable, and hence since 𝔤 is semisimple we must have 𝐼∩𝐼 ⟂ = 0. But then
by part i) of Lemma 5.3.9 we must have 𝔤 = 𝐼 ⊕ 𝐼 ⟂ . Since this is evidently an orthogonal direct sum, the
Killing form must be nondegenerate on both 𝐼 and 𝐼 ⟂ , and since they are ideals, Cartan’s criterion then
implies they are both semisimple. Since the quotient map induces an isomorphism 𝐼 ⟂ ≅ 𝔤/𝐼 it follows
that any quotient of 𝔤 is also semisimple.
For 𝑖𝑖), note that by part 𝑖𝑖) of Lemma 5.3.9, 𝜅 is non-degenerate on 𝐼 if and only if 𝔤 = 𝐼 ⊕ 𝐼 ⟂ . But the
restriction of 𝜅𝔤 to 𝐼 is the Killing form of 𝐼, and so Cartan’s criterion completes the proof. 

Corollary 5.3.11. Let 𝔤 be a semisimple Lie algebra and let Derk (𝔤) be the Lie algebra of derivations of 𝔤. Then
ad ∶ 𝔤 → Derk (𝔤) is an isomorphism, so that in particular any derivation of 𝔤 is inner.

Proof. Suppose that 𝛿 ∈ Derk (𝔤). Then we may form 𝔤1 = 𝔤 ⋊𝛿 𝔤𝔩1 , the semi-direct product3 of 𝔤 and 𝔤𝔩1 .
Now 𝔤 is a semisimple ideal in 𝔤1 , so by Proposition 5.3.10, 𝔤1 = 𝔤 ⊕ 𝔤⟂ . But then [𝔤⟂ , 𝔤] ⊆ 𝔤⟂ ∩ 𝔤 = {0}.
and so if 𝑎 ∈ 𝔞 is such that (𝑎, −1) ∈ 𝔤⟂ then for all 𝑥 ∈ 𝔤 we have

0 = [(𝑎, −𝑡), (𝑥, 0)] = ad(𝑎)(𝑥) − 𝛿(𝑥),

and hence 𝛿 = ad(𝑎) is an inner derivation as required.



Theorem 5.3.12. Let 𝔤 be a semisimple Lie algebra.

i) There exist ideals 𝔤1 , 𝔤2 , … 𝔤𝑘 ⊆ 𝔤 which are simple Lie algebras and for which the natural map:

𝔤1 ⊕ 𝔤2 ⊕ … ⊕ 𝔤𝑘 → 𝔤,

is an isomorphism. In particular, if 𝔤 is semisimple it is perfect, i.e. 𝐷(𝔤) = 𝔤..

ii) Any simple ideal 𝔞 ∈ 𝔤 is equal to some 𝔤𝑖 (1 ≤ 𝑖 ≤ 𝑘). In particular the decomposition in part 𝑖) is unique
up to reordering.

Proof. For part 𝑖) we use induction on the dimension of 𝔤. Let 𝔞 be a minimal non-zero ideal in 𝔤. If 𝔞 = 𝔤
then 𝔤 is simple, so we are done. Otherwise, we have dim(𝔞) < dim(𝔤). Then 𝔤 = 𝔞 ⊕ 𝔞⟂ , and by induction
𝔞⟂ is a direct sum of simple ideals. It follows that 𝔤 = 𝔞 ⊕ 𝔞⟂ , hence any ideal in 𝔞 is also an ideal in 𝔤, thus
since 𝔞 is a minimal ideal, it must be simple, and so 𝔤 is a direct sum of simple ideals as required. Since a
simple Lie algebra is trivially seen to be perfect, each 𝔤𝑖 is perfect and hence so is 𝔤.
For part 𝑖𝑖), suppose that 𝔤 = 𝔤1 ⊕ 𝔤2 ⊕ … ⊕ 𝔤𝑘 is a decomposition as above and 𝔞 is a simple ideal of
𝔤. Now as 𝔷(𝔤) = {0}, we must have 0 ≠ [𝔤, 𝔞] ⊂ 𝔞, and hence by simplicity of 𝔞 it follows that [𝔤, 𝔞] = 𝔞.
But then we have
𝑘
𝔞 = [𝔤, 𝔞] = [� 𝔤𝑖 , 𝔞] = [𝔤1 , 𝔞] ⊕ [𝔤2 , 𝔞] ⊕ … ⊕ [𝔤𝑘 , 𝔞],
𝑖=1
3
Recall that a semidirect product of Lie algebras 𝔞 ⋊ 𝔟 requires, in addition to the two Lie algebras, a homomorphism 𝜑 ∶ 𝔟 →
Derk (𝔞). If 𝔟 = 𝔤𝔩1 however, 𝜑 is determined by 𝜑(1) ∈ Derk (𝔞), i.e. we only need to specify the derivation by which ad(1) acts
on 𝔞 in the semidirect product.

41
(the ideals [𝔤𝑖 , 𝔞] are contained in 𝔤𝑖 so the last sum remains direct). But 𝔞 is simple, so direct sum decom-
position must have exactly one nonzero summand and we have 𝔞 = [𝔤𝑖 , 𝔞] for some 𝑖 (1 ≤ 𝑖 ≤ 𝑘). Finally,
using the simplicity of 𝔤𝑖 we see that 𝔞 = [𝔤𝑖 , 𝔞] = 𝔤𝑖 as required.


5.4 The representation theory of 𝔰𝔩2 .

0 1 1 0 0 0
Definition 5.4.1. The standard basis for 𝔰𝔩2 (k) is 𝑒 = � � ,ℎ = � � and 𝑓 = �
1 0 �
. The
0 0 0 −1
relations obeyed by {𝑒, ℎ, 𝑓} in 𝔰𝔩2 are [𝑒, 𝑓] = ℎ, [ℎ, 𝑒] = 2𝑒 and [ℎ, 𝑓] − 2𝑓. If 𝔤 is an arbitrary Lie algebra,
an 𝔰𝔩2 -triple is a triple {𝐸, 𝐻, 𝐹} of elements of 𝔤 which obey the same relations, so [𝐸, 𝐹] = 𝐻, [𝐻, 𝐸] = 2𝐸
and [𝐻, 𝐹] = −2𝐹. Such a triple determines a homomorphism 𝜃 ∶ 𝔰𝔩2 → 𝔤 where 𝜃(𝑒) = 𝐸, 𝜃(ℎ) = 𝐻 and
𝜃(𝑓) = 𝐹. Since 𝔰𝔩2 is simple, such a homomorphism is determined by its image up to a scalar.
In particular, in this section, when 𝑉 is a representation of 𝔰𝔩2 we will write the action in terms of an
𝔰𝔩2 -triple {𝐸, 𝐻, 𝐹} rather than a homomorphism 𝜌 ∶ 𝔰𝔩2 → 𝔤𝔩𝑉 , i.e., we will write 𝐸 rather than 𝜌(𝑒) etc..

The proofs of the results up to Lemma 5.4.6 are exercises for Problem Sheet 3.
Lemma 5.4.2. Let 𝐴 be an associative algebra and suppose that {𝐸, 𝐻, 𝐹} is an 𝔰𝔩2 -triple in 𝐴. Then we have:

𝑖) [𝐻, 𝐸𝑘 ] = 2𝑘𝐸𝑘 , 𝑖𝑖) [𝐻, 𝐹𝑘 ] = −2𝑘𝐹𝑘 , 𝑖𝑖𝑖) [𝐸, 𝐹𝑘 ] = 𝑘𝐹𝑘−1 (𝐻 − (𝑘 − 1))

Lemma 5.4.3. Let 𝑉 be a representation of 𝔰𝔩2 , let 𝑉 = ⨁ 𝑉𝜆,𝐻 be the generalised eigenspace decomposi-
𝜆∈k
𝑠
tion of 𝑉 with respect to 𝐻, and let 𝑉𝜆,𝐻 ⊆ 𝑉𝜆,𝐻 denote the 𝜆-eigenspace of 𝐻. Then we have
𝑠 𝑠 𝑠 𝑠
𝐸(𝑉𝜆,𝐻 ) ⊆ 𝑉𝜆+2,𝐻 , 𝐸(𝑉𝜆,𝐻 ) ⊆ 𝑉𝜆+2,𝐻 , 𝐹(𝑉𝜆,𝐻 ) ⊆ 𝑉𝜆−2,𝐻 , 𝐹(𝑉𝜆,𝐻 ) ⊆ 𝑉𝜆−2,𝐻 ..

Definition 5.4.4. Since char(k) = 0, the rational numbers ℚ is the minimal subfield of k. We may thus
partially order k by setting 𝜆 < 𝜇 if 𝜆 − 𝜇 ∈ ℚ>0 . If 𝑉 is a finite dimensional 𝔰𝔩2 -representation then the
spectrum 𝑆(𝐻) of 𝐻 (that is, its set of eigenvalues) is finite, and hence contains at least one eigenvalue
𝜆 which is maximal for the partial order. By Lemma 5.4.3 we see that 𝐸(𝑉𝜆,𝐻 ) ⊆ 𝑉𝜆+2,𝐻 = {0}, so that
𝑉𝜆,𝐻 ⊆ ker(𝐸). A vector 𝑣 ∈ 𝑉 is said to be an highest weight vector if it is an 𝐻-eigenvector and 𝐸(𝑣) = 0.
𝑠
Any representation contains a highest weight vector because the elements of 𝑉𝜆,𝐻 are highest weight
vectors whenever 𝜆 is maximal in 𝑆(𝐻).
Proposition 5.4.5. Let 𝑉 be an irreducible representation of 𝔰𝔩2 , where the action is given by the 𝔰𝔩2 -triple
{𝐸, 𝐻, 𝐹} ∈ 𝔤𝔩𝑉 . Then if dim(𝑉) = 𝑛 + 1, 𝑉 is generated by a highest weight vector 𝑣𝑛 of weight 𝑛, the ac-
tion of 𝐻 is diagonalisable with spectrum {−𝑛 + 2𝑘 ∶ 0 ≤ 𝑘 ≤ 𝑛}. Moreover, there is a unique irreducible
representation of 𝔰𝔩2 of dimension 𝑛 + 1 for every 𝑛 ∈ ℤ≥0 .
1
Lemma 5.4.6. Let 𝑉 be a representation of 𝔰𝔩2 and let 𝐶 = 𝐶𝑉 = 𝐸𝐹 + 𝐹𝐸 + 2 𝐻 2 be the Casimir operator.

i) The Casimir operator is an 𝔰𝔩2 -endomorphism of 𝑉.

ii) If 𝑉 = 𝑉(𝑑) is the irreducible representation of dimension 𝑑 + 1, then 𝐶 acts on 𝑉(𝑑) by the scalar 𝜆(𝑑) =
1
2
𝑑(𝑑 + 2).

iii) If 𝑉 is an arbitrary 𝔰𝔩2 -representation and the generalised eigenspace decomposition of 𝑉 with respect to
𝐶 is given by 𝑉 = ⨁ 𝑉 , then if 𝑉𝜆,𝐶 ≠ 0 there is a unique 𝑑 ∈ ℤ≥0 such that 𝜆 = 𝜆(𝑑) and
𝜆∈𝑆(𝐶) 𝜆,𝐶
𝑉𝜆,𝐶 ∈ Rep𝑉 (𝔰𝔩2 ).
𝑑

Corollary 5.4.7. If (𝑉, 𝜌) is a finite-dimensional representation of 𝔰𝔩2 and 𝑊 is a proper subrepresentation with
dim(𝑊) = dim(𝑉) − 1, then 𝑊 has a complement in 𝑉.

42
Proof. By part 𝑖𝑖𝑖) of Lemma 5.4.6 if 𝑉 = ⨁ 𝑉 is the decomposition of 𝑉 into the generalised
𝑑∈𝐷 𝜆(𝑑),𝐶
eigenspaces of the Casimir 𝐶𝑉 , where 𝐷 is the subset of ℤ≥0 for which 𝜆(𝑑) is an eigenvalue of 𝐶𝑉 , the
subspaces 𝑉𝜆(𝑑),𝐶 are 𝔰𝔩2 -subrepresentations with 𝑉(𝑑) their only composition factor. Moreover, 𝑊 =
⨁ 𝑉 ∩ 𝑊 (since it is a 𝐶𝑉 -stable subspace). Since dim(𝑉/𝑊) = 1, we must have 𝑉/𝑊 ≅ 𝑉(0),
𝑑∈𝐷 𝜆(𝑑),𝐶
so that 𝑉/𝑊 ≅ 𝑉0,𝐶 /𝑊 ∩ 𝑉0,𝐶 , thus we may assume that the only composition factor of 𝑉 is the trival
representation. Let 𝒞 = (𝑉 = 𝐹0 > 𝐹1 > … > 𝐹𝑑 = {0}) be a composition series for 𝑉. Then by
assumption 𝒞 is a complete flag and 𝜌(𝔰𝔩2 ) ⊆ 𝔫𝒞 But then 𝜌(𝔰𝔩2 ) must be nilpotent, and as 𝔰𝔩2 is simple
this implies 𝜌(𝔰𝔩2 ) = {0} as required. 

Theorem 5.4.8. Let (𝑉, 𝜌) be a representation of 𝔰𝔩2 . Then 𝑉 is semisimple (and hence completely reducible).

Proof. Suppose that 𝑈 ≤ 𝑉 is a subrepresentation. Now 𝔤𝔩𝑉 ≅ 𝑉 ∗ ⊗ 𝑉 is a 𝔰𝔩2 representation, and if we

let

𝑀1 = {𝜙 ∈ 𝔤𝔩𝑉 ∶ 𝜙(𝑉) ⊆ 𝑈, ∃𝑐 ∈ k, 𝜙|𝑈 = 𝑐.1𝑈 }, 𝑀0 = {𝜙 ∈ 𝔤𝔩𝑉 ∶ 𝜙(𝑉) ⊆ 𝑈, 𝜙(𝑈) = 0}.

then 𝑀0 ⊆ 𝑀1 with dim(𝑀1 /𝑀0 ) = 1, and if 𝜙 ∈ 𝑀1 , 𝑥 ∈ 𝔰𝔩2 we have

[𝜌(𝑥), 𝜙](𝑣) = 𝜌(𝑥)(𝜙(𝑣)) − 𝜙(𝜌(𝑥)(𝑣)) ∈ 𝑊, ∀ 𝑣 ∈ 𝑉,

and if 𝑣 ∈ 𝑊 then 𝜙(𝑣) = 𝑣 and 𝜙(𝜌(𝑥)(𝑣) = 𝜌(𝑥)(𝑣) so that [𝜌(𝑥), 𝜙] ∈ 𝑀0 . Thus 𝑀1 is an 𝔰𝔩2 -
subrepresentation of 𝔤𝔩𝑉 with 𝑀0 a subrepresentation with 𝑀1 /𝑀0 one-dimensional. But then by the
Corollary 5.4.7 it follows that 𝑀0 has a complement, 𝑆 say. Now dim(𝑆) = 1, so that 𝑆 must be a copy
of the trivial representation, and we may pick 𝜙0 ∈ 𝑆 such that 𝜙0|𝑊 = 1𝑊 . But then [𝜌(𝑥), 𝜙0 ] = 0
for all 𝑥 ∈ 𝔰𝔩2 says precisely that 𝜙0 ∈ End𝔰𝔩2 (𝑉) and since 𝜙0 (𝑉) ⊆ 𝑊 with 𝜙0|𝑊 = 1𝑊 , it is an 𝔰𝔩2 -
homomorphism and a projection map to 𝑊, hence 𝑉 = 𝑊 ⊕ ker(𝜙0 ) and ker(𝜙0 ) is a complementary
subrepresentation to 𝑊 as required. 

We end this section with an application of the semisimplicity of finite-dimensional 𝔰𝔩2 -

representations:

Theorem 5.4.9. (Levi’s theorem for 𝔰𝔩2 ). Let 𝔤 be a Lie algebra and let 𝔯 = rad(𝔤). If 𝔤/𝔯 ≅ 𝔰𝔩2 , then there is a
subalgebra 𝔰 of 𝔤 isomorphic to 𝔰𝔩2 so that 𝔤 is the semidirect product of 𝔯 ⋊ 𝔰.

Proof. (Non-examinable.) We consider three steps, with the core of the proof being step 2.
Step 1: [𝔯, 𝔤] = 0, that is, 𝔯 is the centre of 𝔤. Then you can check (see Problem Sheet) that 𝔤 is a represen-
tation of 𝔰𝔩2 and complete reducibility of 𝔤 will provide a splitting.
Step 2: [𝔯, 𝔤] ≠ 0 and 𝔯 has no proper ideals. Since 𝔯 is solvable, this implies that 𝐷𝔯 = 0 and hence 𝔯 is
abelian. Now ad ∶ 𝔤 → 𝔤𝔩𝔤 and 𝔤𝔩𝔤 acts on itself via its adjoint representation, hence their composition
ad𝔤𝔩𝑔 ∘ ad𝔤 gives 𝔤𝔩𝔤 the structure of a 𝔤-representation. Let 𝑀1 = {𝜙 ∈ 𝔤𝔩𝔤 ∶ 𝜙(𝔤) ⊆ 𝔯, 𝜙|𝔯 = 𝑐.1𝔯 } and
𝑀0 = {𝜙 ∈ 𝔤𝔩𝔤 ∶ 𝜙(𝔤) ⊆ 𝔯, 𝜙(𝔯) = 0}, so that 𝑀1 ⊇ 𝑀0 and both contain ad(𝔯).
Now if 𝑥 ∈ 𝔤, 𝜙 ∈ 𝑀1 then for all 𝑦 ∈ 𝔤 we have 𝑥(𝜙)(𝑦) = [𝑥, 𝜙(𝑦)] − 𝜙([𝑥, 𝑦]) ∈ 𝔯 and if 𝑦 ∈ 𝔯
then 𝑥(𝜙)(𝑦) = 0 since 𝜙(𝑦) = 𝑐𝑦, 𝜙([𝑥, 𝑦]) = 𝑐[𝑥, 𝑦]. Moreover, if 𝑥 ∈ 𝔯 then 𝑥(𝜙)(𝑦) = −𝜙([𝑥, 𝑦]) =
−𝑐.ad(𝑥)(𝑦) so that 𝔯(𝑀1 ) ⊆ ad(𝔯). It follows that 𝔤(𝑀1 ) ⊆ 𝑀0 and 𝔯(𝑀1 ) ⊆ ad(𝔯). Thus 𝑀1 /ad(𝔯) is a
𝔤/𝔯 ≅ 𝔰𝔩2 -representation. It follows that 𝑀0 /ad(𝔯) has a complement in 𝑀1 /ad(𝔯), and since dim(𝑀0 ) =
dim(𝑀1 ) − 1, this complement is k.𝜙0 + ad(𝔯), a copy of the trivial representation of 𝔰𝔩2 , where by scaling
𝜙0 appropriately we may assume that 𝜙0|𝔯 = −1𝔯 , that is, 𝜙0 is a projection to 𝔯.
Consider 𝔰 = {𝑥 ∈ 𝔤 ∶ 𝑥(𝜙0 ) = 0}. Then certainly 𝔰 is a subalgebra of 𝔤, and If 𝑥 ∈ 𝔯 then as we saw
above 𝑥(𝜙0 ) = ad(𝑥) = 0, so that k.𝑥 is an ideal in 𝔯, contradicting our assumption that 𝔯 had no proper
ideals. Thus 𝔰 ∩ 𝔯 = {0}. But if 𝑥 ∈ 𝔤, then 𝜙0 (𝑥) ∈ 𝔯, and 𝑦(𝜙0 ) = ad(𝑦) for all 𝑦 ∈ 𝔯, so that

(𝑥 − 𝑥(𝜙0 )(𝜙0 )(𝑦) = 𝑥(𝜙0 )(𝑦) − 𝑥(𝜙0 )(𝜙0 )(𝑦) = 𝑥(𝜙0 )(𝑦) − 𝑥(𝜙0 )(𝑦) = 0,

43
so that 𝑥 = (𝑥 − 𝑥(𝜙0 )) + 𝑥(𝜙0 ) ∈ 𝔰 + 𝔯. Since 𝔰 ∩ 𝔯 = {0}, it follows that 𝔤 = 𝔰 ⊕ 𝔯 as vector spaces, so
that 𝔤 = 𝔰 ⋊ 𝔯 as required.
Step 3: Now suppose that 𝔤 is arbitrary, and prove the theorem by induction on dim(𝔯). If dim(𝔯) = 0
there is nothing to prove, so suppose dim(𝔯) > 0. If 𝔯 has no proper nonzero ideals then [𝔯, 𝔤] = 𝔯 or {0},
and we are done by steps 1 and 2 respectively. Otherwise 𝔯 has a proper nonzero ideal, 𝐼 say, and if we let
𝔤1 = 𝔤/𝐼 then rad(𝔤1 ) = 𝔯/𝐼 has strictly smaller dimension. But then by induction there is a subalgebra
𝔰1 ⊆ 𝔤/𝐼 which is isomorphic to 𝔰𝔩2 , with 𝔰1 ⊕ 𝔯/𝐼 = 𝔤1 . But then if 𝑠 ̃ = {𝑠 ∈ 𝔤 ∶ 𝑠 + 𝐼 ∈ 𝔰1 }, then rad(𝔰)̃ = 𝐼
and dim(𝐼) < dim(𝔯) so that by induction again we may find a subalgebra 𝔰 of 𝔰̃ with 𝔰 ⊕ 𝐼 = 𝔰,̃ and
̃ ≅ 𝔰𝔩2 as required.
𝔰 ≅ 𝔰/𝐼 

44
Chapter 6

The structure of semisimple Lie algebras

If (𝔤, 𝔥) is a Cartan pair, the amount of information captured by the Cartan decomposition 𝔤 = 𝔥 ⊕
⨁ 𝔤 depends on 𝔤. At one extreme, when 𝔤 is nilpotent, we have the trivial decomposition 𝔥 = 𝔤.
𝛼∈Φ 𝛼
The semisimple case is in some sense at the opposite extreme: the Cartan subalgebra 𝔥 turns out to be
abelian and the decomposition of 𝔤 is as fine as possible – 𝔤 is a semisimple 𝔥-representation.

6.1 The Killing form and the Cartan decomposition

Proposition 6.1.1. Let (𝔤, 𝔥) be a Cartan pair and let 𝔤 = ⨁ 𝔤 be the associated Cartan decomposition
𝜆∈Φ0 𝜆
of 𝔤, where Φ0 = {0} ∪ Φ, 𝔤0 = 𝔥, and let 𝜅 denote the Killing form of 𝔤.

i) We have 𝜅(𝔤𝜆 , 𝔤𝜇 ) = 0 unless 𝜆 + 𝜇 = 0.

ii) For any 𝜆 ∈ Φ0 the restriction of 𝜅 to 𝔤𝜆 × 𝔤−𝜆 gives a linear map 𝜃𝜆 ∶ 𝔤𝜆 → 𝔤∗−𝜆 . The Killing form 𝜅 is
nondegenerate, and hence 𝔤 semisimple, if and only if, for every 𝜆 ∈ Φ the linear map 𝜃𝜆 is an isomorphism.
In particular:

a) the restriction of 𝜅 to 𝔥 = 𝔤0 is nondegenerate,

b) dim(𝔤𝜆 ) = dim(𝔤−𝜆 ) and so 𝜆 ∈ Φ if and only if −𝜆 ∈ Φ.

Proof. Let 𝜃 ∶ 𝔤 → 𝔤∗ be the map given by 𝜃(𝑥)(𝑦) = 𝜅(𝑥, 𝑦) for 𝑥, 𝑦 ∈ 𝔤. By 2.2.2, 𝔤∗ ≅ ⨁ 𝔤∗ as an

𝜆∈Φ0 𝜆
𝔥-representation. Now 𝔤𝜆 is the 𝜆-generalised weight space of 𝔤, hence its only composition factor is k𝜆 .
By Lemma 2.2.13 it follows that 𝔤∗𝜆 has k−𝜆 as its unique composition factor, and hence 𝔤∗ = ⨁ 𝔤∗𝜆
𝜆∈Φ0
gives the 𝔥-weight isotypical decomposition of 𝔤∗ where (𝔤∗ )𝜆 ≅ (𝔤−𝜆 )∗ .
Since 𝜅 is invariant, 𝜃 ∈ ℒ(𝔤, 𝔤∗ )𝔤 = ℒ𝔤 (𝔤, 𝔤∗ ), i.e. it is a homomorphism of 𝔤-representations. In
particular, it is an 𝔥-homomorphism, so that 𝜃(𝔤𝜆 ) ⊆ (𝔤∗ )𝜆 = 𝔤∗−𝜆 . It follows that 𝜅(𝔤𝜆 , 𝔤𝜇 ) = 𝜃(𝔤𝜆 )(𝔤𝜇 ) =
0 unless 𝜇 = −𝜆 as claimed. Moreover, if 𝔤 is semisimple, then 𝜅 is nondegenerate, i.e. 𝜃 is an isomor-
phism. But as 𝜃 is an 𝔥-homomorphism this forces 𝜃𝜆 = 𝜃|𝔤𝜆 ∶ 𝔤𝜆 → 𝔤∗−𝜆 to be an isomorphism for all 𝜆
as required. 

Remark 6.1.2. When 𝔤 is semisimple, by Proposition 6.1.1, 𝜅 induces an isomorphism 𝜃0 ∶ 𝔥 → 𝔥∗ . Given

𝜆 ∈ 𝔥∗ , we will write 𝑡𝜆 ∈ 𝔥 for 𝜃−1
0 (𝜆), so that 𝑡𝜆 is uniquely determined by the condition that 𝜅(𝑡𝜆 , ℎ) =
𝜆(ℎ) for all ℎ ∈ 𝔥.
Lemma 6.1.3. If (𝔤, 𝔥) is a Cartan pair and 𝜅 is the Killing form of 𝔤, then we have

𝜅(ℎ1 , ℎ2 ) = � dim(𝔤𝛼 )𝛼(ℎ1 )𝛼(ℎ2 ), ∀ ℎ1 , ℎ2 ∈ 𝔥. (6.1.1)

𝛼∈Φ

45
If 𝔤 is semisimple, it follows that

i) ⟨Φ⟩k = 𝔥∗
ii) 𝔥 is abelian
iii) Recall 𝔥𝛼 = [𝔤𝛼 , 𝔤−𝛼 ] ⊆ 𝔥. We have 𝔥𝛼 ∩ ker(𝛼) = {0}, and hence dim(𝔥𝛼 ) ≤ 1.

Proof. Since 𝔤 = ⨁ 𝔤 as an 𝔥-representation, 𝜅(ℎ1 , ℎ2 ) = ∑𝜆∈Φ 𝑡𝔤𝜆 (ℎ1 , ℎ2 ). Equation (6.1.1) thus
𝜆∈Φ0 𝜆 0
follows immediately by applying Lemma 2.2.14 to the 𝔥-representations 𝔤𝜆 . (Note that ad(ℎ1 )ad(ℎ2 ) lies
in the associative algebra of linear maps which preserve a composition series of 𝔤𝜆 even though it may
not lie in ad(𝔥).)
Now suppose that 𝔤 is semisimple and let 𝑆 = ⟨Φ⟩k . Since Φ ⊆ 𝐷(𝔥)0 ⊆ 𝔥∗ , clearly 𝑆 ⊆ 𝐷(𝔥)0 . Let
𝑆0 = Φ 0 = {ℎ ∈ 𝔥 ∶ 𝛼(ℎ) = 0, ∀ 𝛼 ∈ Φ}, where, since 𝔥 is finite-dimensional, we view 𝑆0 as a subspace of
𝔥 via the canonical isomorphism (𝔥∗ )∗ ≅ 𝔥. It is clear from (6.1.1) that 𝑆0 ⊆ rad(𝜅|𝔥 ), but if 𝔤 is semisimple,
part 𝑖𝑖) of Proposition 6.1.1 shows that this is {0}. But then 𝑆 = (𝑆0 )0 = 𝔥∗ , so that Φ spans 𝔥∗ establishing
𝑖). But 𝑆 ⊆ 𝐷(𝔥)0 hence 𝐷(𝔥)0 = 𝔥∗ and hence 𝐷(𝔥) = {0} and 𝔥 is abelian.
Finally, by Lemma 5.2.11 applied to the adjoint representation of 𝔤, if 𝛽 ∈ Φ, then we have 𝛽|𝔥𝛼 = 𝑟𝛽 .𝛼|𝔥𝛼
for some 𝑟𝛽 ∈ ℚ. But then if 𝑧 ∈ 𝔥𝛼 ∩ ker(𝛼) it follows 𝛽(𝑧) = 𝑟𝛽 .𝛼(𝑧) = 0 for all 𝛽 ∈ Φ, and hence
𝑧 ∈ 𝑆0 = {0} as required. Since dim(ker(𝛼)) = dim(𝔥) − 1, clearly dim(𝔥𝛼 ) ≤ 1. 

6.1.1 The roots of a semisimple Lie algebra

Definition 6.1.4. Recall that, as in Definition 2.2.15, if 𝑊 is an 𝔥-representation then we write 𝑊 𝑠 for
the socle of 𝑊, that is, the sum of all irreducible 𝔥-subrepresentations of 𝑊. If (𝔤, 𝔥) is a Cartan pair and
𝔤 = ⨁ 𝔤𝜆 , is the decomposition of 𝔤 as an 𝔥-representation, then the 𝔥-representation 𝔤𝜆 has k𝜆
𝜆∈Φ0
as its only composition factor and hence 𝔤𝑠𝜆 = {𝑥 ∈ 𝔤𝜆 ∶ [ℎ, 𝑥] = 𝜆(ℎ).𝑥, ∀ ℎ ∈ 𝔥} must be non-zero,
that is, if 𝑥 ∈ 𝔤𝑠𝜆 and 𝑥 ≠ 0 then k.𝑥 ≅ k𝛼 as 𝔥-representations. For example, 𝔤𝑠0 = 𝔤𝔥 , i.e. 𝔤𝑠0 is the
subrepresentation of 𝔥-invariants in 𝔤.

The following Lemma is the first step in constructing a copy of 𝔰𝔩2 for each pair {±𝛼} of roots in a
semisimple Lie algebra.
Lemma 6.1.5. Let (𝔤, 𝔥) be a Cartan pair such that 𝔤 is semisimple, and let 𝔤 = 𝔥 ⊕ ⨁ 𝔤 the associated
𝛼∈Φ 𝛼
𝑠 𝑠
Cartan decomposition. If 𝑥 ∈ 𝔤𝛼 and 𝑦 ∈ 𝔤−𝛼 then [𝑥, 𝑦] = 𝜅(𝑥, 𝑦).𝑡𝛼 . Moreover if 𝑒𝛼 ∈ 𝔤𝛼 is nonzero, then
ad(𝑒𝛼 )(𝔤−𝛼 ) = k.𝑡𝛼 = 𝔥𝛼 so that 𝔥 = 𝔥𝛼 ⊕ ker(𝛼). In particular, if ℎ𝛼 ∈ 𝔥𝛼 is given by 𝛼(ℎ𝛼 ) = 2, there is an
𝑓𝛼 ∈ 𝔤−𝛼 such that [𝑒𝛼 , 𝑓𝛼 ] = ℎ𝛼 .

Proof. For 𝑖) take any 𝑥 ∈ 𝔤𝑠𝛼 , 𝑦 ∈ 𝔤−𝛼 and ℎ ∈ 𝔥. Then for all ℎ ∈ 𝔥,
𝜃0 ([𝑥, 𝑦])(ℎ) ≔ 𝜅(ℎ, [𝑥, 𝑦]) = 𝜅([ℎ, 𝑥], 𝑦) = 𝜅(𝑥, 𝑦)𝛼(ℎ)
Thus [𝑥, 𝑦] = 𝜃−1
0 (𝜅(𝑥, 𝑦)𝛼) = 𝜅(𝑥, 𝑦)𝑡𝛼 . By part 𝑖𝑖𝑖) of Lemma 6.1.3, we know 𝔥𝛼 ∩ ker(𝛼) = {0}. Hence
𝔥𝛼 =⊆ k.𝑡𝛼 . But since 𝜃𝛼 ∶ 𝔤𝛼 → 𝔤∗−𝛼 is an isomorphism, for any non-zero 𝑒𝛼 ∈ 𝔤𝑠𝛼 we may find 𝑦 ∈ 𝔤−𝛼
with 𝜅(𝑒𝛼 , 𝑦) ≠ 0. Hence there is an 𝑓𝛼 ∈ k.𝑦 with 𝜅(𝑒𝛼 , 𝑓𝛼 ) = 2𝛼(𝑡𝛼 )−1 so that [𝑒𝛼 , 𝑓𝛼 ] = ℎ𝛼 and
moreover 𝔥𝛼 ⊇ ad(𝑒𝛼 )(k.𝑓𝛼 ) = k.ℎ𝛼 hence 𝔥𝛼 = k.ℎ𝛼 = k.𝑡𝛼 .

Definition 6.1.6. Given a root 𝛼 ∈ Φ, the element ℎ𝛼 is known as the coroot associated to 𝛼. It charac-
terised by the conditions that it 𝑖) lies on the line k.𝑡𝛼 and 𝑖𝑖) 𝛼 takes the value 2 on ℎ𝛼 . Indeed it follows
that ℎ𝛼 = 2𝛼(𝑡𝛼 )−1 .𝑡𝛼 . Using the definition of 𝑡𝛼 , we have 𝛼(𝑡𝛼 ) = 𝜅(𝑡𝛼 , 𝑡𝛼 ), hence we may also write
ℎ𝛼 = 2𝑡𝛼 /𝜅(𝑡𝛼 , 𝑡𝛼 ).
We write Φ ∨ for the set of coroots, that is Φ ∨ = {ℎ𝛼 ∶ 𝛼 ∈ Φ}

46
Proposition 6.1.7. Let 𝔤 be a semisimple Lie algebra and 𝔥 a Cartan subalgebra with Cartan decomposition
𝔤=𝔥⊕⨁ 𝔤𝛼 . Then
𝛼∈Φ

i) The root spaces 𝔤𝛼 are one-dimensional, and if 𝛼 ∈ Φ, 𝑐 ∈ ℤ, then 𝑐.𝛼 ∈ Φ if and only if 𝑐 = ±1.

ii) 𝔰𝔩𝛼 = 𝔤𝛼 ⊕ 𝔥𝛼 ⊕ 𝔤−𝛼 is a subalgebra of 𝔤 isomorphic to 𝔰𝔩2 (k).

iii) 𝔤 is a semisimple 𝔥-representation, so that for all ℎ ∈ 𝔥, ad(ℎ) is semisimple, and hence 𝔥 consists of semisim-
ple elements, i.e. the Jordan decomposition of ℎ ∈ 𝔥 is ℎ = ℎ𝑠 .

Proof. Fix 𝛼 ∈ Φ, and let {𝑒𝛼 , 𝑓𝛼 , ℎ𝛼 } be as in Lemma 6.1.5, so that {𝑒𝛼 , ℎ𝛼 , 𝑓𝛼 } ⊆ 𝑀 where

𝑀 = � 𝑀𝑘.𝛼 , 𝑀𝛼 = k.𝑒𝛼 , 𝑀𝑝.𝛼 = 𝔤𝑝.𝛼 , ∀ 𝑝 ≤ 0.

𝑘≤1

We claim that 𝑀 is a subalgebra. Since 𝑀𝑝.𝛼 ⊆ 𝔤𝑝.𝛼 for all 𝑝 ∈ ℤ, it suffices to check that [𝑀𝑘.𝛼 , 𝑀𝑙.𝛼 ] ⊆
𝑀(𝑘+𝑙).𝛼 for 𝑘 ≤ 𝑙 ≤ 1. For 𝑘 + 𝑙 ≤ 0 this is clear because 𝑀(𝑘+𝑙)𝛼 = 𝔤(𝑘+𝑙)𝛼 while if 𝑘 = 0 it is equivalent to
𝑀𝑙.𝛼 being an 𝔥-subrepresentation which we have already checked unless 𝑙 = 1, but as 𝑒𝛼 ∈ 𝔤𝑠𝛼 , 𝑀𝛼 ≅ k𝛼 .
Finally if 𝑘 = 𝑙 = 1, then [𝑀𝛼 , 𝑀𝛼 ] = k.[𝑒𝛼 , 𝑒𝛼 ] = 0 = 𝑀2𝛼 .
It follows that ad(ℎ𝛼 ) = [ad(𝑒𝛼 ), ad(𝑓𝛼 )] acts on 𝑀 with trace zero. But 𝑀 = ⨁ 𝑀𝑘.𝛼 is the
𝑘≤1
decomposition of 𝑀 into ad(ℎ𝛼 ) generalised eigenspaces where 𝑀𝑘𝛼 = 𝑀2𝑘,ℎ𝛼 , hence

0 = � 2𝑝.dim(𝑀𝑝.𝛼 ) = 2 − � 2𝑝.dim(𝔤−𝑝.𝛼 ) = 2.�1 − � 𝑝.dim(𝔤−𝑝.𝛼 )

𝑝≤1 𝑝≥1 𝑝>0

If 𝑝 > 1 then 𝑝.dim(𝔤−𝑝.𝛼 ) > 1 unless dim(𝔤−𝑝.𝛼 ) = 0 hence dim(𝔤−𝑝.𝛼 ) = 0 for all 𝑝 > 1 so that dim(𝔤−𝛼 ) =
1. Since dim(𝔤𝜆 ) = dim(𝔤−𝜆 ) it follows dim(𝔤𝑐.𝛼 ) = 0 if |𝑐| > 1 and dim(𝔤𝛼 ) = dim(𝔤−𝛼 ) = 1, which proves
part 𝑖). It follows that 𝔰𝔩𝛼 is three-dimensional with basis {𝑒𝛼 , ℎ𝛼 , 𝑓𝛼 }. But [𝑒𝛼 , 𝑓𝛼 ] = ℎ𝛼 by our choice
of 𝑓𝛼 , and as 𝔤𝛼 , 𝔤−𝛼 are 1-dimensional, 𝔤±𝛼 ≅ k±𝛼 , and hence as 𝛼(ℎ𝛼 ) = 2 we have [ℎ𝛼 , 𝑒𝛼 ] = 2𝑒𝛼 ,
[ℎ𝛼 , 𝑓𝛼 ] = −2𝑓𝛼 . Thus {𝑒𝛼 , ℎ𝛼 , 𝑓𝛼 } is an 𝔰𝔩2 -triple, so that 𝔰𝔩𝛼 ≅ 𝔰𝔩2 , establishing part 𝑖𝑖).
Finally, since 𝔥 is abelian, as an 𝔥-representation it is isomorphic to k0
dim(𝔥)
thus the Cartan decom-
position 𝔤 = 𝔥 ⊕ ⨁ 𝔤𝛼 exhibits 𝔤 as a direct sum of irreducible 𝔥-representations, hence ad(ℎ) acts
𝛼∈Φ
diagonalisably on 𝔤 for all ℎ ∈ 𝔥, which proves part 𝑖𝑖𝑖). 

Definition 6.1.8. Given a root 𝛼 ∈ Φ, the element ℎ𝛼 is known as the coroot associated to 𝛼. We will write
Φ ∨ = {ℎ𝛼 ∶ 𝛼 ∈ Φ} ⊂ 𝔥 for the set of coroots.

Recall that we write Φ0 = Φ ∪ {0} for the set of 𝔥-weights of the adjoint representation.

Proposition 6.1.9. Let 𝛼, 𝛽 ∈ Φ and set 𝑆𝛼 (𝛽) = {𝛽 + 𝑘𝛼 ∶ 𝑘 ∈ ℤ} ∩ Φ0 . Then if 𝑝, respectively 𝑞 ∈ ℤ≥0 are
maximal among positive integers 𝑘 with the property that 𝛽 − 𝑘𝛼 ∈ Φ0 (respectively 𝛽 + 𝑘𝛼 ∈ Φ0 ) then

i) 𝑆𝛼 (𝛽) = {𝛽 + 𝑗𝛼 ∶ −𝑝 ≤ 𝑗 ≤ 𝑞},

ii)
2𝜅(𝑡𝛼 , 𝑡𝛽 )
𝛽(ℎ𝛼 ) = 𝜅(ℎ𝛼 , 𝑡𝛽 ) = = 𝑝 − 𝑞.
𝜅(𝑡𝛼 , 𝑡𝛼 )
In particular 𝛽 − 𝛽(ℎ𝛼 ).𝛼 ∈ Φ.

iii) If 𝛽 = 𝑐.𝛼 for some 𝑐 ∈ k, then 𝑐 = ±1.

47
Proof. If 𝛽 ∈ ℤ.𝛼 then all of the assertions of the Proposition follow from Proposition 6.1.7. Thus we may
assume that 𝛽 ∉ ℤ.𝛼 and hence for all 𝑘 ∈ ℤ 𝛽 + 𝑘𝛼 ≠ 0. Let

𝑀𝛽 = � 𝔤𝛾
𝛾∈𝑆𝛼 (𝛽)

Clearly 𝑀𝛽 is an 𝔰𝔩𝛼 -subrepresentation of 𝔤 on which ℎ𝛼 has eigenvalues {𝛽(ℎ𝛼 ) + 2𝑘 ∶ 𝛽 + 𝑘𝛼 ∈ Φ}, each

with a 1-dimensional eigenspace. It follows that 𝑀𝛽 must be an irreducible representation of 𝔰𝔩𝛼 , and if
𝑞 = max{𝑘 ∈ ℤ ∶ 𝛽 + 𝑘𝛼 ∈ Φ, and dim(𝑀𝛽 ) = 𝑙 + 1, then (𝛽 + ℤ.𝛼) ∩ Φ = {𝛽 + 𝑘𝛼 ∶ 𝑞 − 𝑙 ≤ 𝑘 ≤ 𝑞}, and
moreover we must have 𝛽(ℎ𝛼 ) + 2(𝑞 − 𝑙) = −𝛽(ℎ𝛼 ) − 2𝑞, so that 𝛽(ℎ𝛼 ) = 𝑙 − 2𝑞. It follows that if we set
𝑝 = 𝑙 − 𝑞 ≥ 0, then 𝑆𝛼 (𝛽) ∶= (𝛽 + ℤ.𝛼) ∩ Φ = {𝛽 + 𝑘𝛼 ∶ −𝑝 ≤ 𝑘 ≤ 𝑞} and 𝛽(ℎ𝛼 ) = 𝑝 − 𝑞. Moreover, since
𝛽(ℎ𝛼 ) is an eigenvalue of ℎ𝛼 on 𝑀𝛽 so is −𝛽(ℎ𝛼 ) and hence 𝛽 − 𝛽(ℎ𝛼 ).𝛼 ∈ 𝑆𝛼 (𝛽). We have thus established
parts 𝑖) and 𝑖𝑖).
Now suppose that 𝛽 = 𝑐.𝛼 for some 𝑐 ∈ k, 𝑐 ∉ ℤ. Then 𝛽(ℎ𝛼 ) = 𝑝 − 𝑞 = 2𝑐 is an odd integer
(since 0 is not a weight of 𝑀𝛽 as 𝛽 ∉ ℤ.𝛼) so that 1 is an eigenvalue for the action of ℎ𝛼 on 𝑀𝛽 and hence
1 1
2
𝛼 ∈ 𝑆𝛼 (𝛽) ⊆ Φ. But since 𝛼 = 2( 2 𝛼), this contradicts part 𝑖) of Proposition 6.1.7. 

6.2 Root systems and the classification of semisimple Lie algebras

6.2.1 Rational form of 𝔥 and inner product spaces

Recall that since 𝜅|𝔥 is non-degenerate, it gives an isomorphism 𝜃 ∶ 𝔥 → 𝔥∗ . For 𝜆 ∈ 𝔥∗ , we write

𝑡𝜆 for 𝜃−1 (𝜆), so that 𝜅(𝑡𝜆 , ℎ) = 𝜆(ℎ), (∀ 𝜆 ∈ 𝔥∗ , ℎ ∈ 𝔥). Given a root 𝛼 ∈ Φ, we have seen that 𝔰𝔩𝛼 =
𝔤𝛼 ⊕ 𝔤−𝛼 ⊕ [𝔤𝛼 , 𝔤−𝛼 ] ⊆ 𝔤 is a subalgebra isomorphic to 𝔰𝔩2 . Indeed each summand is 1-dimensional, and
[𝔤𝛼 , 𝔤−𝛼 ] = k.𝑡𝛼 ⊆ 𝔥. We set ℎ𝛼 = 2𝛼(𝑡𝛼 )−1 .𝑡𝛼 , so that 𝛼(ℎ𝛼 ) = 2.
Definition 6.2.1. Let 𝔥ℚ = {ℎ ∈ 𝔥 ∶ 𝛼(ℎ) ∈ ℚ, ∀ 𝛼 ∈ Φ}. Clearly 𝔥ℚ is a ℚ-vector space (a subspace
of 𝔥 viewed as a ℚ-vector space1 ). Recall also that a symmetric bilinear form (., .) on a ℚ-vector space 𝑉
is said to be positive definite if (𝑣, 𝑣) ≥ 0 with equality holding precisely when 𝑣 = 0. Such a form is also
commonly called an inner product.
Lemma 6.2.2. The ℚ-vector space 𝔥ℚ has the following properties

i) If 𝜅ℚ denotes the restriction of the Killing form to 𝔥ℚ , then 𝜅ℚ is a ℚ-valued positive definite symmetric
bilinear form on 𝔥ℚ .

ii) 𝔥ℚ = ⟨Φ ∨ ⟩ℚ = ⟨{𝑡𝛼 ∶ 𝛼 ∈ Φ}⟩ℚ , and dimℚ (𝔥ℚ ) = dimk (𝔥). In particular, 𝜃 identifies the dual (𝔥ℚ )∗
of 𝔥ℚ with ⟨Φ⟩ℚ .

Proof. For part 𝑖), since we have shown in Proposition 6.1.7 that dim(𝔤𝛼 ) = 1 for all 𝛼 ∈ Φ, we may simplify
the expression for 𝜅|𝔥 given in (6.1.1) to obtain 𝜅(ℎ1 , ℎ2 ) = ∑𝛾∈Φ 𝛾(ℎ1 )𝛾(ℎ2 ). It is thus immediate from
the definition of 𝔥ℚ that the Killing form is ℚ-valued on 𝔥ℚ . Moreover, if ℎ ∈ 𝔥ℚ , since 𝛾(ℎ)2 ≥ 0 for all
𝛾 ∈ Φ, we have 𝜅ℚ (ℎ, ℎ) ≥ 0, with equality if and only if 𝛾(ℎ) = 0 for all 𝛾 ∈ Φ, and since ⟨Φ⟩ = 𝔥∗ , this
holds only if ℎ = 0. Thus 𝜅ℚ is positive definite as claimed.
For part 𝑖𝑖), by Proposition 6.1.9, the roots are ℤ-valued on set of coroots Φ ∨ , hence 𝔥ℚ ⊇ ⟨Φ ∨ ⟩ℚ .
Moreover ℎ𝛼 = 𝑐𝛼 .𝑡𝛼 where 2 = 𝑐𝛼 .𝜅(𝑡𝛼 , 𝑡𝛼 ), hence 𝜅(ℎ𝛼 , ℎ𝛼 ) = 𝑐2𝛼 𝜅(𝑡𝛼 , 𝑡𝛼 ) = 2.𝑐𝛼 ∈ ℤ. Thus 𝑡𝛼 =
2𝜅(ℎ𝛼 , ℎ𝛼 )−1 .ℎ𝛼 ∈ ℚ.ℎ𝛼 , and hence clearly ⟨{𝑡𝛼 ∶ 𝛼 ∈ Φ}⟩ℚ = ⟨Φ ∨ ⟩ℚ .
Now since Φ spans 𝔥∗ , we may find a subset 𝐵 = {𝛾1 , … , 𝛾𝑙 } ⊆ Φ which is a basis of 𝔥∗ . Let 𝐵′ =
𝜃−1
0 (𝐵) = {𝑡𝛾𝑖 ∶ 1 ≤ 𝑖 ≤ 𝑙} ⊆ ⟨Φ ∨ ⟩ℚ . Since it is a k-basis of 𝔥, 𝐵′ is linearly independent over k and hence
1
Since k has characteristic zero, it contains a canonical copy of ℚ – it is the intersection of all of the subfields of k)

48
over ℚ. It follows dimℚ (⟨Φ ∨ ⟩ℚ ) ≥ 𝑙. But if 𝜂 ∶ 𝔥ℚ → ℚ𝑙 is given by 𝜂(ℎ) = (𝛾𝑖 (ℎ))𝑙𝑖=1 , then 𝜂 is injective
because 𝐵 is a basis of 𝔥∗ . Hence dim(𝔥ℚ ) ≤ 𝑙. It follows that 𝔥ℚ = ⟨Φ ∨ ⟩ℚ and dimℚ (𝔥ℚ ) = 𝑙 = dimk (𝔥).
Since span{𝑡𝛼 ∶ 𝛼 ∈ Φ} = 𝔥ℚ , the final statement in 𝑖𝑖) is now also clear. 

Definition 6.2.3. Let (−, −) denote the bilinear form on 𝔥∗ which is obtained by identifying 𝔥∗ with 𝔥:
that is
(𝜆, 𝜇) = 𝜅(𝑡𝜆 , 𝑡𝜇 ).
Clearly it is a nondegenerate symmetric bilinear form, and via the previous Lemma, for all 𝛼, 𝛽 ∈ Φ we
have (𝛼, 𝛽) = 𝜅(𝑡𝛼 , 𝑡𝛽 ) ∈ ℚ, so that it restricts to a ℚ-valued symmetric bilinear form on 𝔥∗ℚ which is
positive definite.

*Remark 6.2.4. The group 𝑌 generated by {ℎ𝛼 ∶ 𝛼 ∈ Φ} is a finitely generated abelian group which is
a subgroup of a ℚ-vector space, and so is torsion-free. It follows from the structure theorem for finitely
generated abelian groups2 that 𝑌 is therefore actually a free abelian group. Moreover, the inner product
restricts to an integer-valued positive definite form on 𝑌. A finitely generated free abelian group with such
a form is called a lattice. (See the final problem sheet for some discussion of these.) Note that any basis
𝐵 for 𝑌 is also a ℚ-basis of 𝔥ℚ but not conversely – this gives at least some motivation for the notion of
a base we will see shortly, in that some subsets of Φ may yield only a ℚ-basis of 𝔥ℚ , whereas others may
yield a ℤ-basis of 𝑌.

6.3 Abstract root systems

In this section we study the geometry which we are led to by the configuration of roots associated to
a Cartan decomposition of a semisimple Lie algebra. These configurations will turn out to have a very
special, highly symmetric, form which allows them to be completely classified.
We will work with rational inner product spaces 𝑉, that is, ℚ-vector spaces 𝑉 equipped with a
positive-definite symmetric bilinear form3 which we will denote by (., .). Such vector spaces have, in ad-
dition to a notion of length given by, for any 𝑣 ∈ 𝑉, the norm ‖𝑣‖ = (𝑣, 𝑣)1/2 , a notion of angle: by the
Cauchy-Schwarz inequality there is a unique 𝜃 ∈ [0, 𝜋] with

(𝑣1 , 𝑣2 )
cos(𝜃) = ∈ [−1, 1].
‖𝑣1 ‖.‖𝑣2 ‖

The group of orthogonal linear transformations of 𝑉 is

O(𝑉) = {𝑔 ∈ GL(𝑉) ∶ (𝑔(𝑣), 𝑔(𝑤)) = (𝑣, 𝑤), ∀ 𝑣, 𝑤 ∈ 𝑉}.

Definition 6.3.1. A reflection is a nontrivial element of O(𝑉) which fixes a subspace of codimension 1 (i.e.
dimension dim(𝑉) − 1). If 𝑠 ∈ O(𝑉) is a reflection and 𝑊 < 𝑉 is the +1-eigenspace, then 𝐿 = 𝑊 ⟂ is a line
preserved by 𝑠, hence the restriction 𝑠|𝐿 of 𝑠 to 𝐿 is an element of 𝑂(𝐿) = {±1}, which since 𝑠 is nontrivial
must be −1. In particular 𝑠 has order 2. If 𝑣 is any nonzero element of 𝐿 then it is easy to check that 𝑠 is
given by
2(𝑢, 𝑣)
𝑠(𝑢) = 𝑢 − 𝑣.
(𝑣, 𝑣)
Given 𝑣 ≠ 0 we will write 𝑠𝑣 for the reflection given by the above formula, and refer to it as the “reflection
in the hyperplane perpendicular to 𝑣”.

We now give the definition which captures the geometry of the root of a semisimple Lie algebra.
2
which you may have seen in a previous algebra course...
3
Such forms only make sense over ordered fields, such as ℚ or ℝ.

49
Definition 6.3.2. A pair (𝑉, Φ) consisting a rational inner product space 𝑉 and a finite subset Φ ⊂ 𝑉\{0}
is called an (abstract) root system if it satisfies the following properties:

i) Φ spans 𝑉;
ii) If 𝛼 ∈ Φ, 𝑐 ∈ ℚ, then 𝑐𝛼 ∈ Φ if and only if 𝑐 = ±1;
iii) If 𝛼 ∈ Φ then 𝑠𝛼 ∶ 𝑉 → 𝑉 preserves Φ;
iv) If 𝛼, 𝛽 ∈ Φ and we define
2(𝛼, 𝛽)
⟨𝛼, 𝛽⟩ = , (6.3.1)
(𝛼, 𝛼)
then ⟨𝛼, 𝛽⟩ ∈ ℤ. We say ⟨𝛼, 𝛽⟩ is a Cartan integer.

This definition is, unsurprisingly, motivated by the following result.

Lemma 6.3.3. Let (𝔤, 𝔥) be a Cartan pair where 𝔤 is semisimple, and let 𝔤 = 𝔥 ⊕ ⨁ 𝔤 be the associated
𝛼∈Φ 𝛼
∗ ∗ ∗
Cartan decomposition and let 𝔥ℚ be the ℚ-span of Φ in 𝔥 . Then (𝔥ℚ , Φ) is an abstract root system.

Proof. Let (−, −) denote the symmetric bilinear form on 𝔥∗ℚ induced by the restriction of the Killing form
𝜅|𝔥 as in §6.2.1. Lemma 6.2.2 shows that this restriction is positive definite. Property 𝑖) for an abstract root
system follows immediately from the definitions, and the remaining properties follow from Proposition
6.1.9: part 𝑖𝑖) of that Proposition establishes property 𝑖𝑖), while part 𝑖) establishes properties 𝑖𝑖𝑖) and 𝑖𝑣).


Remarkably, the finite set of vectors given by a root system has both a rich enough structure that it
captures the isomorphism type of a semisimple Lie algebra, but is also explicit enough that we can com-
pletely classify them, and hence classify semisimple Lie algebras.
Definition 6.3.4. Let (𝑉, Φ) be a root system. Then the Weyl group of the root system is the group 𝑊 =
⟨𝑠𝛼 ∶ 𝛼 ∈ Φ⟩. Since its generators preserve the finite set Φ and these vectors span 𝑉, it follows that it is a
finite subgroup of O(𝑉).
Example 6.3.5. Let 𝔤 = 𝔰𝔩𝑛 , Then let 𝔡𝑛 denote the diagonal matrices in 𝔤𝔩𝑛 and 𝔥 the (traceless) diagonal
matrices in 𝔰𝔩𝑛 . As you saw in the problem sets, 𝔥 forms a Cartan subalgebra in 𝔰𝔩𝑛 . Let {𝜀𝑖 ∶ 1 ≤ 𝑖 ≤ 𝑛}
be the basis of 𝔡∗𝑛 dual to the basis {𝐸𝑖𝑖 ∶ 1 ≤ 𝑖 ≤ 𝑛} of 𝔡𝑛 in 𝔤𝔩𝑛 . The Cartan decomposition of 𝔰𝔩𝑛 is
𝔥⊕⨁ k.𝐸𝑖𝑗 , where ad(ℎ)(𝐸𝑖𝑗 ) = (ℎ𝑖 − ℎ𝑗 )𝐸𝑖𝑗 , where ℎ𝑘 = 𝜖𝑘 (ℎ) for 𝑘 ∈ {1, … , 𝑛}. Thus
1≤𝑖≠𝑗≤𝑛
𝑛 𝑛 𝑛
𝔥ℚ = {� ℎ𝑖 𝐸𝑖𝑖 ∶ ℎ𝑖 ∈ ℚ, � ℎ𝑖 = 0} and 𝔥∗ℚ = �� 𝑐𝑖 𝜀𝑖 ∶ 𝑐𝑖 ∈ ℚ� �ℚ.(𝜀1 + … + 𝜀𝑛 ),
𝑖=1 𝑖=1 𝑖=1

where the roots in 𝔥∗ℚ are the (images of the) vectors {𝜀𝑖 −𝜀𝑗 ∶ 1 ≤ 𝑖, 𝑗 ≤ 𝑛, 𝑖 ≠ 𝑗}. Moreover, the Killing form
for 𝔰𝔩𝑛 is 𝜅(𝑥, 𝑦) = 2𝑛.tr(𝑥𝑦), so the {𝐸𝑖𝑖 ∶ 1 ≤ 𝑖 ≤ 𝑛} are an orthogonal basis of 𝔥 with 𝜅(𝐸𝑖𝑖 , 𝐸𝑖𝑖 ) = 2𝑛. The
Weyl group 𝑊 in this case is the group generated by the reflections 𝑠𝛼 which, for 𝛼 = 𝜀𝑖 − 𝜀𝑗 interchange
the basis vectors 𝜀𝑖 and 𝜀𝑗 , so it is easy to see that 𝑊 is just the symmetric group on 𝑛 letters.

6.3.1 Root systems and bases

If (𝑉, Φ) is an abstract root system then, since the set of roots Φ spans 𝑉, it certainly contains (many)
subsets which form a basis of 𝑉. The key to the classification of root systems is to show that there is a
special class of such bases which capture enough of the geometry of the set of roots that the entire root
system can be recovered from the bases of this form. This section is devoted to determining the “correct”
notion of a special basis or “base” and establishing that this class of basis is always nonempty.
We first record some useful notation:

50
Definition 6.3.6. Given a set of vectors 𝑋 in a vector space 𝑉, we will write

ℕ.𝑋 = �� 𝑐𝑠 .𝑠 ∶ 𝑌 ⊆ 𝑋 finite, 𝑐𝑠 ∈ ℕ� ⊆ ℤ.𝑋 = �� 𝑐𝑠 .𝑠 ∶ 𝑌 ⊆ 𝑋 finite , 𝑐𝑠 ∈ ℤ� .

𝑠∈𝑌 𝑠∈𝑌

The set ℕ.𝑋 is closed under vector addition and multiplication by elements of ℕ.

Recall that the invariance of the Killing form together with the fact that it is nondegenerate in the
semisimple case implies that the set of roots is a union of pairs {±𝛼} consisting of a root and its negative.
This suggests that it may be useful to have a notion of positivity for vectors in a ℚ-vector space analogous
to that in ℚ, namely, a set of “positive” vectors 𝒫 in 𝑉.

Definition 6.3.7. Let 𝑉 be a ℚ-vector space. A positive system 𝒫 in 𝑉 is a subset 𝒫 ⊆ 𝑉\{0} such that

• For each 𝑣 ∈ 𝑉\{0}, exactly one of 𝑣 or −𝑣 lies in 𝒫.

• if 𝑣1 , 𝑣2 ∈ 𝒫 and 𝜆 ∈ ℚ>0 then 𝜆.𝑣1 , 𝑣1 + 𝑣2 ∈ 𝒫.

If 𝒫 is a positive system, then we define a total order < on 𝑉 by 𝑣1 < 𝑣2 if and only if 𝑣2 − 𝑣1 ∈ 𝒫.

Example 6.3.8. If 𝑉 is a ℚ-vector space and 𝐵 is a basis for 𝑉 then we will write 𝐵 ⃗ = (𝑒1 , … , 𝑒𝑙 ) for an
ordering of 𝐵, that is, a bijection 𝑖 ↦ 𝑒𝑖 from {1, 2, … , 𝑙} to 𝐵. The ordering equips the basis of 𝑉 ∗ dual to
𝐵 with an ordering, and we denote this ordered basis by 𝐵⃗∗ = (𝛿1 , … , 𝛿𝑙 ) so that for any 𝑣 ∈ 𝑉 we have
𝑙
𝑣 = ∑𝑖=1 𝛿𝑖 (𝑣).𝑒𝑖 . Let 𝑠𝐵⃗ ∶ 𝑉\{0} → {1, … , 𝑙} be given by 𝑠𝐵⃗ (𝑣) = min{𝑘 ∶ 1 ≤ 𝑘 ≤ 𝑙, 𝛿𝑘 (𝑣) ≠ 0}, and, if
𝑣 ∈ 𝑉\{0} has 𝑠𝐵⃗ (𝑣) = 𝑘, let 𝑝𝐵⃗ (𝑣) = 𝛿𝑘 (𝑣). Then

𝑙
⃗ = {𝑣 ∈ 𝑉\{0} ∶ 𝑝⃗ (𝑣) > 0} = �� 𝜆𝑖 𝑒𝑖 ∶ 𝑘 ∈ {1, 2, … , 𝑙}, 𝜆𝑘 > 0�
𝒫 (𝐵) 𝐵
𝑖=𝑘

is a positive system.

Though we will not need the result, the following Lemma shows that all positive systems in a real
inner product space arise via the real analogue of the above construction.

Lemma 6.3.9. Let 𝒫 be a positive system in a real inner product space. Then there is a basis 𝐵 of 𝑉 such that
⃗ of 𝐵.
𝒫 = 𝒫𝐵⃗ for some ordering 𝐵

Proof. Since 𝒫 and −𝒫 are convex subsets of a Euclidean vector space there is a linear functional 𝜆 ∶ 𝑉 →
ℝ which separates 𝒫 and −𝒫. Since 𝑉\{0} = 𝒫 ⊔ −𝒫, by changing the sign of 𝜆 if necessary we may
assume that 𝒫 ⊆ {𝑣 ∈ 𝑉 ∶ 𝜆(𝑣) ≥ 0} = 𝐻 + and −𝒫 ⊆ {𝑣 ∈ 𝑉 ∶ 𝜆(𝑣) ≤ 0} = 𝐻 − . It follows that
int(𝒫 ) = 𝐻 + , and if 𝐻 = {𝑣 ∈ 𝑉 ∶ 𝜆(𝑣) = 0} then 𝐻 is an inner product space and 𝒫𝐻 = 𝒫 ∩ 𝐻 is a
positive system in 𝐻, so that theorem follows by induction on dim(𝑉). 

If (𝑉, Φ) is an abstract root system and 𝒫 is a positive system on Φ, then the first property of 𝒫 listed
+
above shows that Φ𝒫 ∶= Φ ∩ 𝒫 must contain exactly on element of each pair {±𝛼} ⊆ Φ and we call it the
+
set of (𝒫-) positive roots. When there is no danger of confusion we will write Φ + in place of Φ𝒫 . Note
that if 𝐵 ⊆ Φ is a basis of 𝑉, then ±𝐵 ∩ 𝒫 must also be a basis of 𝑉, and hence any set of positive roots
Φ + contains a basis of 𝑉.

*Remark 6.3.10. As already noted in Remark 6.2.4, ℤΦ is a finitely generated free abelian group. It there-
fore possesses a basis B, which, as noted above, may be chosen to consist of positive vectors. It is not clear
that such a basis must lie in Φ, but this turns out to be the case. This hopefully motivates the following
definition:

51
Definition 6.3.11. Let (𝑉, Φ) be a root system, and let Δ be a subset of Φ. We say that Δ is a base (or a set
of simple roots) for Φ if

i) Δ is linearly independent.

ii) If ΦΔ± = ±ℕ.Δ ∩ Φ are the Δ-positive/negative roots, then Φ = ΦΔ+ ⊔ ΦΔ− .

Note that 𝑖𝑖) implies that 𝑉 = ⟨Φ⟩ℚ ⊆ ℚ.Δ so that by 𝑖) Δ is a basis of 𝑉. Indeed by 𝑖𝑖) we have ℤ.Δ = ℤΦ
so that the existence of a base implies that ℤΦ is a free abelian group.

Condition 𝑖𝑖) in the definition of a set of simple roots implies that such a set has an associated set of
positive roots ΦΔ+ , and the key to establishing the existence of sets of simple roots is to show that one can
recover Δ from this positive set. We do this using the following definition:
+ +
Definition 6.3.12. Given a set of positive roots Φ𝒫 , we say that 𝛼 ∈ Φ𝒫 is decomposable if 𝛼 = 𝛽 + 𝛾 for
+
some 𝛽, 𝛾 ∈ Φ𝒫 . A root is indecomposable if it is not decomposable. Let Π 𝒫 be the set of indecomposable
+
roots in Φ𝒫 .

Proposition 6.3.13. Let (𝑉, Φ) be an abstract root system.

i) Suppose that 𝒫 is a positive system. Then the set of indecomposable roots Π = Π 𝒫 is a base of (𝑉, Φ).
+
In particular, every root system has a base and the map Φ𝒫 ↦ Π 𝒫 assigning to a set of positive roots its
indecomposable elements is bijective with inverse Π 𝒫 ↦ ℕ.Π𝒫 ∩ Φ.

ii) Every base of (𝑉, Φ) is of the form Π 𝒫 for some positive system 𝒫, thus the map Δ → ΦΔ+ gives a bijection
+
between set of all bases of (𝑉, Φ) and the set {Φ𝒫 ∶ 𝒫 a positive system} of all choices of a subset of positive
+ +
roots. Indeed for any 𝒫 with Φ𝒫 = ΦΔ we have Π 𝒫 = Δ.

Proof. We establish part 𝑖) in three steps:

+
Claim 1: Φ𝒫 ⊆ ℕ.Π.
Proof : Note that the set Φ + is totally ordered by the order given by 𝒫, so if Φ + is not contained in ℕ.Π,
we may consider 𝛼 ∈ Φ + , the minimal element of Φ + not contained on ℕ.Π. Clearly 𝛼 ∉ Π, hence 𝛼 is
decomposable and there are 𝛾1 , 𝛾2 ∈ Φ + with 𝛼 = 𝛾1 + 𝛾2 . But then 𝛾1 , 𝛾2 < 𝛼, so by minimality of 𝛼,
𝛾1 , 𝛾2 ∈ ℕ.Π. But ℕ.Π is closed under addition, so 𝛼 = 𝛾1 + 𝛾2 ∈ ℕ.Π, giving a contradiction. Hence
+
Φ + ⊆ ℕ.Π as required, and Φ𝒫 ↦ Π 𝒫 is bijective.
+
Claim 2: If 𝛼, 𝛽 ∈ Φ𝒫 are distinct roots such that (𝛼, 𝛽) > 0, then at least one of them must be decompos-
able.
Proof : First note that ⟨𝛼, 𝛽⟩, ⟨𝛽, 𝛼⟩ > 0. Since 𝛼, 𝛽 ∈ Φ + are distinct, they are linearly independent, and
hence by Cauchy-Schwarz, ⟨𝛼, 𝛽⟩.⟨𝛽, 𝛼⟩ ∈ {1, 2, 3}. It follows one of ⟨𝛼, 𝛽⟩ or ⟨𝛽, 𝛼⟩ = 1. By symmetry we
may assume ⟨𝛼, 𝛽⟩ = 1, and hence 𝑠𝛼 (𝛽) = 𝛽 − 𝛼 ∈ Φ. But then one of 𝛼 − 𝛽 or 𝛽 − 𝛼 lies in Φ + . But as
𝛼 = (𝛼 − 𝛽) + 𝛽 and 𝛽 = (𝛽 − 𝛼) + 𝛼, it follows that one of 𝛼 or 𝛽 must be is decomposable as required.
Claim 3: Let 𝑆 be any subset of 𝒫 with the property that if 𝑠, 𝑡 ∈ 𝑆 are distinct, then (𝑠, 𝑡) ≤ 0. Then 𝑆 is
linearly independent.
Proof : Suppose that ∑𝑠∈𝑆 𝑐𝑠 .𝑠 = 0 is a linear dependence. Let 𝑆± = {𝑠 ∈ 𝑆 ∶ ±𝑐𝑠 > 0}, and set 𝑧 =
∑ + 𝑐𝑡 .𝑡 = ∑ − (−𝑐𝑠 ).𝑠 and note that 𝑧 ∈ 𝒫 unless both 𝑆+ and 𝑆− are empty. But
𝑡∈𝑆 𝑠∈𝑆

(𝑧, 𝑧) = ( � 𝑐𝑡 .𝑡, � (−𝑐𝑠 ).𝑠) = � 𝑐𝑡 (−𝑐𝑠 )(𝑠, 𝑡) ≤ 0

𝑡∈𝑆+ 𝑠∈𝑆− 𝑠,𝑡

so that we must have 𝑧 = 0, and hence 𝑆+ = 𝑆− = ∅ and hence 𝑐𝑠 = 0 for all 𝑠 ∈ 𝑆 and so 𝑆 is linearly
independent as required.

52
 +
Claim 1 shows that Φ𝒫 ⊆ ℕΠ and claim 2 shows that Π satisfies the hypothesis of the set 𝑆 in Claim 3,
hence it is linearly independent and therefore a base as required.
⃗ = {𝛼1 , … , 𝛼𝑙 }. Then Δ
For part 𝑖𝑖), given a base Δ, pick an arbitrary ordering Δ ⃗ is an ordered basis of 𝑉
+
and hence gives a positive system 𝒫 = 𝒫Δ⃗ as in Example 6.3.8. Now ΦΔ = ℕ.Δ ∩ Φ ⊆ 𝒫 since Δ ⊆ 𝒫,
1
and hence ΦΔ+ ⊆ Φ ∩ 𝒫 = Φ𝒫 +
, and since ΦΔ+ and Φ𝒫
+
both have 2 |Φ| elements, ΦΔ+ = Φ𝒫
+
. (Thus the
+
set of positive roots Φ𝒫 is independent of ordering of Δ which we chose to obtain 𝒫). Now since Δ is a
basis of ℤΦ we may define ℎ ∶ ℤΦ → ℤ by setting ℎ(𝛼) = 1 for all 𝛼 ∈ Δ. Then ℎ(𝛾) ≥ 1 for all 𝛾 ∈ ΦΔ+
and so ℎ(𝛼) ≥ 2 if 𝛼 is decomposable. It follows that Δ ⊆ Π 𝒫 and since |Δ| = |Π 𝒫 | = dim(𝑉) we must
have Δ = Π 𝒫 as required.


Lemma 6.3.14. Let (𝑉, Φ) be an abstract root system, and suppose that 𝛼 ∈ Φ.
+ +
1. There is a positive system 𝒫 for which 𝛼 is the minimal element of Φ𝒫 , In particular, 𝛼 ∈ Φ𝒫 is indecom-
posable, and so belongs to Π 𝒫 , a base of (𝑉, Φ).

2. If Δ is a base and 𝛼 ∈ Δ, then if 𝛽 ∈ ΦΔ+ and 𝛽 ≠ 𝛼, then 𝑠𝛼 (𝛽) ∈ Φ𝒫

+
. Hence if we write Φ + (𝛼) = Φ + \{𝛼}
then we have 𝑠𝛼 (Φ + (𝛼)) = Φ + (𝛼) whenever 𝛼 ∈ Δ.

Proof. We use the notation of Example 6.3.8. Pick a basis 𝐵 ⊂ Φ of 𝑉 containing 𝛼 and let 𝐵 ⃗ = {𝛾1 , … , 𝛾𝑙 }
be an ordering of it in which 𝛼 = 𝛾𝑙 . Let 𝐷 = {𝛿1 , … , 𝛿𝑙 } be the corresponding dual basis of 𝑉 ∗ and let
+
𝒫 = 𝒫𝐵⃗ . Then if 𝛽 ∈ Φ𝒫 has 𝛽 ≤ 𝛼, since 𝛽 ∈ 𝒫 and 𝑠𝐵⃗ (𝛼) = 𝑙, we must have 𝑠𝐵⃗ (𝛽) = 𝑙 also, and hence
𝛽 = 𝑐.𝛼 for some 𝑐 > 0. But since 𝛼, 𝛽 ∈ Φ this implies 𝛽 = 𝛼. It follows that 𝛽 is minimal and hence it
must be indecomposable. But then 𝛽 ∈ Π 𝒫 which we have seen is always a base.
Next if Δ is a base containing 𝛼, we may take 𝐵 as in part 𝑖) to be Δ, so that 𝒫 = 𝒫Δ⃗ . If Δ ⃗ 1 = 𝑠𝛼 (Δ),
⃗
then 𝒫Δ⃗ = 𝑠𝛼 (𝒫 ). But 𝑠𝛼 (𝛾𝑖 ) = 𝛾𝑖 − ⟨𝛾𝑙 , 𝛾𝑖 ⟩.𝛾𝑙 , hence the dual basis to Δ 1 = 𝑠𝛼 (Δ) is 𝐷1 = 𝑠𝛼 (𝐷) = ⊺
1
{𝛿1 , … , 𝛿𝑙−1 , −𝛿𝑙 − ∑𝑖<𝑙 ⟨𝛾𝑙 , 𝛾𝑖 ⟩𝛿𝑖 }. It follows that 𝑝Δ (𝑣) = 𝑝Δ1 (𝑣) for all 𝑣 ∈ 𝑉\ℚ.𝛼, hence 𝑠𝛼 (𝒫 ) = 𝒫Δ⃗
1
and 𝒫 \ℚ.𝛼 = 𝒫Δ⃗ \ℚ.𝛼 = 𝑠𝛼 (𝒫 )\ℚ.𝛼, so that (Φ\{±𝛼}) ∩ 𝒫 = (Φ\{±𝛼} ∩ 𝒫1 as required. 
1

It turns out that we can recover the entire root system provided we know a base for it. Before we can
show this, we first show that any two bases of Φ are conjugate under the action of 𝑊. Let Δ 0 be a fixed
choice of base for Φ and let 𝑊0 = ⟨𝑠𝛼 ∶ 𝛼 ∈ Δ 0 ⟩ be the subgroup of 𝑊 generated by the reflections 𝑠𝛼 for
𝛼 in the base Δ 0 .
Proposition 6.3.15. Suppose that Δ 1 is any base of (𝑉, Φ) and let Φ1+ be the corresponding set of positive roots.
Then there is some 𝑤 ∈ 𝑊0 such that 𝑤(Φ1+ ) = Φ0+ , and hence 𝑤(Δ 1 ) = Δ 0 .

Proof. We prove this by induction on 𝑑 = |Φ0+ ∩ Φ1− |. If this 𝑑 = 0 then Φ0+ = Φ1+ and hence Δ 0 = Δ 1
(hence we may take 𝑤 = 𝑒 the identity element of 𝑊0 ). Next suppose that 𝑑 > 0. Let 𝒫1 be a positive
system such that Φ1+ = Φ ∩ 𝒫1 . If Δ 0 ⊆ Φ1+ , then since any element of Φ0+ is a positive integer combina-
tion of Δ 0 , it follows Φ0+ ⊆ 𝒫1 ∩ Φ = Φ1+ and hence Φ0+ ∩ Φ1− = ∅, which contradicts the assumption
that 𝑑 > 0. Thus there is some 𝛼 ∈ Δ 0 such that 𝛼 ∈ Φ1− . But then using the notation of Lemma 6.3.14
we see that

|Φ0+ ∩ 𝑠𝛼 (Φ1− )| = |𝑠𝛼 (Φ0+ ) ∩ Φ1− | = | �{−𝛼} ∪ Φ0+ (𝛼)� ∩ Φ1− | = |Φ0+ (𝛼) ∩ Φ1− | = 𝑑 − 1

where the first equality holds because 𝑠2𝛼 = 1𝑉 , the second equality follows from Lemma 6.3.14, and the
third from the fact that Φ1− contains 𝛼 and hence not −𝛼. But then by induction there is a 𝑤 ∈ 𝑊0 with
𝑤𝑠𝛼 (Φ1 )+ = Φ0+ . Since 𝑤𝑠𝛼 ∈ 𝑊0 we are done. 

Corollary 6.3.16. Suppose that 𝛽 ∈ Φ. Then there is a 𝑤 ∈ 𝑊0 and an 𝛼 ∈ Δ 0 such that 𝑤(𝛽) = 𝛼. In
particular, 𝑊 is generated by the reflections {𝑠𝛾 ∶ 𝛾 ∈ Δ 0 }, that is, 𝑊 = 𝑊0 .

53
Proof. For the first claim follows from the fact that every root lies in a base for (𝑉, Φ), shown in part 𝑖) of
Lemma 6.3.14, together with Proposition 6.3.15.
For the final claim, note that if 𝛽 ∈ Φ then we have just shown that there is a 𝑤 ∈ 𝑊0 such that
𝑤(𝛽) = 𝛾 for some 𝛾 ∈ Δ 0 . But then clearly 𝑠𝛽 = 𝑤−1 𝑠𝛾 𝑤 ∈ 𝑊0 , and so since 𝑊 = ⟨𝑠𝛽 ∶ 𝛽 ∈ Φ⟩ it follows
that 𝑊 ≤ 𝑊0 . Since 𝑊0 ≤ 𝑊 by definition, it follows 𝑊 = 𝑊0 as required.


Remark 6.3.17. In fact 𝑊 acts simply transitively on the bases of (𝑉, Φ), that is, the action is transitive
and, if Δ is a base and 𝑤 ∈ 𝑊 is such that 𝑤(Δ) = Δ, then 𝑤 = 1. The proof (which we will not give)
consists of examining the minimal length expression for 𝑤 in terms of these generators {𝑠𝛼 ∶ 𝛼 ∈ Δ 0 }.

6.4 Cartan matrices and isomorphisms of root systems

First let us formulate the notion of an isomorphism of root systems:

Definition 6.4.1. If (𝑉, Φ) and (𝑉 ′ , Φ ′ ) are root systems, we say that a linear map 𝜙 ∶ 𝑉 → 𝑉 ′ is an
isomorphism of root systems if it is a isomorphism of vector spaces such that 𝜙(Φ) = Φ ′ and

⟨𝜙(𝛼), 𝜙(𝛽)⟩ = ⟨𝛼, 𝛽⟩, ∀ 𝛼, 𝛽 ∈ Φ.

Note that 𝜙 need not be an isometry: if 0 < 𝑐 < 1, then (𝑉, 𝑐.Φ) is a root system which is not isometric to
(𝑉, Φ), but 𝜙(𝑥) = 𝑐.𝑥 is an isomorphism from (𝑉, Φ) to (𝑉, 𝑐.Φ).

Definition 6.4.2. Let (𝑉, Φ) be a root system. The Cartan matrix associated to (𝑉, Φ) is the matrix

𝐶 = 𝐶Δ = (⟨𝛼𝑖 , 𝛼𝑗 ⟩)𝑙𝑖,𝑗=1 .

where {𝛼1 , 𝛼2 , … , 𝛼ℓ } = Δ is a base of (𝑉, Φ). Since the elements of 𝑊 are isometries, and 𝑊 acts tran-
sitively on the set of bases of Φ, the Cartan matrix is independent4 of the choice of base (though clearly
determined only up to orderings of the base Δ, ∀ 𝛼, 𝛽 ∈ Δ.

Remark 6.4.3. It should be noted that the Cartan matrix is probably better denoted (⟨𝛼, 𝛽⟩)𝛼.𝛽∈Δ , that is,
as a Δ × Δ array. If 𝛼 ∶ 𝑉 → 𝑊 is a linear map and 𝐵1 and 𝐵2 are bases of 𝑉 and 𝑊, then if, for 𝑏1 ∈ 𝐵1 we
have 𝛼(𝑏1 ) = ∑𝑏 ∈𝐵 𝑎𝑏2,𝑏1 𝑏2 then the scalars 𝑎𝑏2,𝑏1 assemble to give a function 𝑎 ∶ 𝐵2 ×𝐵1 → k, (𝑏2 , 𝑏1 ) ↦
2 2
𝑎𝑏2,𝑏1 or equivalently, a 𝐵2 × 𝐵1 array of scalars. This array encodes 𝛼 without requiring any order to be
imposed on either basis: if 𝑣 ∈ 𝑉 then 𝑣 = ∑𝑏 ∈𝐵 𝑥𝑏1 𝑏1 , and
1 1

𝛼(𝑣) = � 𝑦𝑏2 𝑏2 , where 𝑦𝑏2 = � 𝑎𝑏2,𝑏1 𝑥𝑏1 .

𝑏2 ∈𝐵2 𝑏1 ∈𝐵1

Note that such arrays can be composed: if 𝛽 ∶ 𝑊 → 𝑈 and 𝐵3 is a basis of 𝑈 then if 𝛽(𝑏2 ) = ∑𝑏 ∈𝐵 𝑐𝑏3,𝑏′ 𝑏3 ,
3 3
then 𝛽 ∘ 𝛼 is represented by the 𝐵3 × 𝐵1 array (𝑑𝑏3,𝑏1 ) where 𝑑𝑏3,𝑏1 = ∑𝑏 ∈𝐵 𝑐𝑏3,𝑏2 𝑎𝑏2,𝑏1 .
2 2

In other words, matrix algebra does not require the rows/columns of a matrix to be ordered, but rather
that they should be appropriately labelled. As a Δ×Δ array of integers, the “Cartan matrix” is independent
of the base chosen, in that, if Δ and Δ′ are two bases of a root system, then there is a bijection 𝑏 ∶ Δ → Δ′
such that ⟨𝛼, 𝛽⟩ = ⟨𝑏(𝛼), 𝑏(𝛽)⟩.

Theorem 6.4.4. Let (𝑉, Φ) be a root system. Then (𝑉, Φ) is determined up to isomorphism by the Cartan matrix
associated to it.
4
This might appear to overlook something: While it is true that 𝑊 acts transitively on the set of bases, so we may pick an
arbitrary base in order to compute the Cartan matrix, if 𝑊Δ = {𝑤 ∈ 𝑊 ∶ 𝑤(Δ) = Δ} then 𝑊Δ acts on 𝐶Δ , and thus could yield
constraints on the possible structure of the Cartan matrix. In fact 𝑊Δ is trivial, so we are not missing anything!

54
Proof. Given root systems (𝑉, Φ) and (𝑉 ′ , Φ ′ ) with the same Cartan matrix, we may certainly pick a base
Δ = {𝛼1 , … , 𝛼ℓ } of (𝑉, Φ) and a base Δ′ = {𝛽1 , … , 𝛽ℓ } of (𝑉 ′ , Φ ′ ) such that ⟨𝛼𝑖 , 𝛼𝑗 ⟩ = ⟨𝛽𝑖 , 𝛽𝑗 ⟩ for all
𝑖, 𝑗, (1 ≤ 𝑖, 𝑗 ≤ ℓ). We claim the map 𝜙 ∶ Δ → Δ′ given by 𝜙(𝛼𝑖 ) = 𝛽𝑖 extends to an isomorphism of
root systems. Since Δ and Δ′ are bases of 𝑉 and 𝑉 ′ respectively, 𝜙 extends uniquely to an isomorphism
of vector spaces 𝜙 ∶ 𝑉 → 𝑉 ′ , so we must show that 𝜙(Φ) = Φ ′ , and ⟨𝜙(𝛼), 𝜙(𝛽)⟩ = ⟨𝛼, 𝛽⟩ for each
𝛼, 𝛽 ∈ Φ.
Let 𝑠𝑖 = 𝑠𝛼𝑖 ∈ O(𝑉) and 𝑠′𝑖 = 𝑠𝛽𝑖 ∈ O(𝑉 ′ ) be the reflections in the Weyl groups 𝑊 = 𝑊(𝑉, Φ)
and 𝑊 ′ = 𝑊(𝑉 ′ , Φ ′ ) respectively. Then from the formula for the action of 𝑠𝑖 it is clear that 𝜙(𝑠𝑖 (𝛼𝑗 )) =
𝑠′𝑖 (𝛽𝑗 ) = 𝑠′𝑖 (𝜙(𝛼𝑖 )), so since Δ is a basis it follows 𝜙(𝑠𝑖 (𝑣)) = 𝑠′𝑖 (𝜙(𝑣)) for all 𝑣 ∈ 𝑉. But then since the 𝑠𝑖 s and
𝑠′𝑖 s generate 𝑊 and 𝑊 ′ respectively, 𝜙 induces an isomorphism 𝑊 → 𝑊 ′ , given by 𝑤 ↦ 𝑤′ = 𝜙∘𝑤∘𝜙−1 .
But by Corollary 6.3.16 we have

𝜙(Φ) = 𝜙(𝑊.Δ) = (𝜙𝑊𝜙−1 )(𝜙(Δ)) = 𝑊 ′ 𝜙(Δ) = 𝑊 ′ .Δ′ = Φ ′

Finally, fixing 𝛼 ∈ Δ, clearly the linear functionals given by 𝑣 ↦ ⟨𝛼, 𝑣⟩ and 𝑣 ↦ ⟨𝜙(𝛼), 𝜙(𝑣)⟩ (𝑣 ∈ 𝑉)
agree if 𝑣 ∈ Δ, hence by linearity they are equal. Hence ⟨𝛼, 𝛽⟩ = ⟨𝜙(𝛼), 𝜙(𝛽)⟩ if 𝛼 ∈ Δ and 𝛽 ∈ Φ. But
since 𝑊 and 𝑊 ′ act by isometries

⟨𝑤(𝛼), 𝑤(𝛽)⟩ = ⟨𝛼, 𝛽⟩ = ⟨𝜙(𝛼), 𝜙(𝛽)⟩ = ⟨𝑤′ (𝜙(𝛼)), 𝑤′ (𝜙(𝛽))⟩ = ⟨𝜙(𝑤(𝛼)), 𝜙(𝑤(𝛽))⟩,

so that since 𝑊.Δ = Φ, it follows that ⟨𝛼, 𝛽⟩ = ⟨𝜙(𝛼), 𝜙(𝛽)⟩ for all 𝛼, 𝛽 ∈ Φ.


Thus to classify root systems up to isomorphism it is enough to classify Cartan matrices. It turns out
that there is a more combinatorial way to encode the information given by a Cartan matrix, because the
possible entries of the Cartan matrix are heavily constrained, as the next Lemma shows:

Lemma 6.4.5. Let (𝑉, Φ) be a root system and let 𝛼, 𝛽 ∈ Φ be such that 𝛼 ≠ ±𝛽. Then the Cartan integer
⟨𝛼, 𝛽⟩ ∈ {0, ±1, ±2, ±3} Moreover, the angle between 𝛼 and 𝛽 and the ratio ‖𝛼‖2 /‖𝛽‖2 are determined by the
pair ⟨𝛼, 𝛽⟩, ⟨𝛽, 𝛼⟩, as the table below shows:

⟨𝛼, 𝛽⟩ ⟨𝛽, 𝛼⟩ 𝜃 ||𝛼||2 /||𝛽||2

0 0 𝜋/2 undetermined
1 1 𝜋/3 1
−1 −1 2𝜋/3 1
1 2 𝜋/4 2
−1 −2 3𝜋/4 2
1 3 𝜋/6 3
−1 −3 5𝜋/6 3

Proof. By assumption, we know that both ⟨𝛼, 𝛽⟩ and ⟨𝛽, 𝛼⟩ are integers with the same sign. By the Cauchy-
Schwarz inequality, if 𝜃 denotes the angle between 𝛼 and 𝛽, then:

(𝛼, 𝛽)2
⟨𝛼, 𝛽⟩⟨𝛽, 𝛼⟩ = 4 = 4cos(𝜃)2 < 4. (6.4.1)
‖𝛼‖2 .‖𝛽‖2

Noting that cos(𝜃)2 determines the angle between the two vectors (or rather the one which is less than
𝜋) and (if (𝛼, 𝛽) ≠ 0) ⟨𝛽, 𝛼⟩/⟨𝛼, 𝛽⟩ = ||𝛼||2 /||𝛽||2 (where we write ||𝑣||2 = (𝑣, 𝑣)), it is then easy to verify the
table given above by a case-by-case check. 

Definition 6.4.6. As the previous Lemma shows, if 𝐶 = (𝑐𝑖𝑗 ) is a Cartan matrix, its entries 𝑐𝑖𝑗 are highly
constrained: indeed 𝑐𝑖𝑖 = 2 and if 𝑖 ≠ 𝑗, 𝑐𝑖𝑗 ∈ {0, −1, −2, −3} and {𝑐𝑖𝑗 , 𝑐𝑗𝑖 } = {−1, −𝑐𝑖𝑗 𝑐𝑗𝑖 } so that 𝑐𝑖𝑗 is

55
determined by the product 𝑐𝑖𝑗 .𝑐𝑗𝑖 and the relative lengths of the two roots (set out in the table above).
As a result, the matrix can be recorded as a kind of graph: the vertex set of the graph is labelled by the
base {𝛼1 , … , 𝛼𝑙 }, and one puts ⟨𝛼𝑖 , 𝛼𝑗 ⟩.⟨𝛼𝑗 , 𝛼𝑖 ⟩ edges between 𝛼𝑖 and 𝛼𝑗 , directing the edges so that they
go from the larger root to the smaller root. Thus for example if ⟨𝛼𝑖 , 𝛼𝑗 ⟩ = −2 and ⟨𝛼𝑗 , 𝛼𝑖 ⟩ = −1 so that
||𝛼𝑗 ||2 > ||𝛼𝑖 ||2 , that is, 𝛼𝑗 is longer than 𝛼𝑖 , we record this in the graph as:

𝛼𝑖 • ks •𝛼𝑗

The resulting graph is called the Dynkin diagram.

Definition 6.4.7. We say that a root system (𝑉, Φ) is reducible if there is a partition of the roots into two
non-empty subsets Φ1 ⊔ Φ2 such that (𝛼, 𝛽) = 0 for all 𝛼 ∈ Φ1 , 𝛽 ∈ Φ2 . Then if we set 𝑉1 = span(Φ1 )
and 𝑉2 = span(Φ2 ), clearly 𝑉 = 𝑉1 ⊕ 𝑉2 and we say (𝑉, Φ) is the sum of the root systems (𝑉1 , Φ1 ) and
(𝑉2 , Φ2 ). This allows one to reduce the classification of root systems to the classification of irreducible
root systems, i,e. root systems which are not reducible. It is straight-forward to check that a root system
is irreducible if and only if its associated Dynkin diagram is connected.

Definition 6.4.8. (Not examinable.) The notion of a root system makes sense over the real, as well as
rational, numbers. Let (𝑉, Φ) be a real root system, and let Δ = {𝛼1 , 𝛼2 , … , 𝛼𝑙 } be a base of Φ. If 𝑣𝑖 =
𝛼𝑖 /||𝛼𝑖 || (1 ≤ 𝑖 ≤ 𝑙) are the unit vectors in 𝑉 corresponding to Δ, then they satisfy the conditions:

1. (𝑣𝑖 , 𝑣𝑖 ) = 1 for all 𝑖 and (𝑣𝑖 , 𝑣𝑗 ) ≤ 0 if 𝑖 ≠ 𝑗,

2. If 𝑖 ≠ 𝑗 then 4(𝑣𝑖 , 𝑣𝑗 )2 ∈ {0, 1, 2, 3}. (This is the reason we need to extend scalars to the real num-
bers – if you want you could just extend scalars to ℚ( √2, √3), but it makes no difference to the
classification problem).

Such a set of vectors is called an admissible set.

It is straightforward to see that classifying ℚ-vector spaces with a basis which forms an admissible
set is equivalent to classifying Cartan matrices, and using elementary techniques it is possible to show
that that the following are the only possibilities (we list the Dynkin diagram, a description of the roots,
and a choice of a base):

• Type 𝐴ℓ (ℓ ≥ 1):
• •⋅⋅⋅• •
ℓ
𝑉 = {𝑣 = � 𝑐𝑖 𝑒𝑖 ∈ ℚℓ ∶ � 𝑐𝑖 = 0}, Φ = {𝜀𝑖 − 𝑒𝑗 ∶ 1 ≤ 𝑖 ≠ 𝑗 ≤ ℓ}
𝑖=1
Δ = {𝜀𝑖+1 − 𝜀𝑖 ∶ 1 ≤ 𝑖 ≤ ℓ − 1}

• Type 𝐵ℓ (ℓ ≥ 2):
• •⋅⋅⋅• +3 •

𝑉 = ℚℓ , Φ = {±𝜀𝑖 ± 𝜀𝑗 ∶ 1 ≤ 𝑖, 𝑗 ≤ ℓ, 𝑖 ≠ 𝑗} ∪ {𝜀𝑖 ∶ 1 ≤ 𝑖 ≤ ℓ},

Δ = {𝜀1 , 𝜀𝑖+1 − 𝜀𝑖 ∶ 1 ≤ 𝑖 ≤ ℓ − 1}

• Type 𝐶ℓ (ℓ ≥ 3):
• • ⋅ ⋅ ⋅ • ks •
𝑉 = ℚℓ , Φ = {±𝜀𝑖 ± 𝜀𝑗 ∶ 1 ≤ 𝑖, 𝑗 ≤ ℓ, 𝑖 ≠ 𝑗} ∪ {2𝜀𝑖 ∶ 1 ≤ 𝑖 ≤ ℓ},
Δ = {2𝜀1 , 𝜀𝑖+1 − 𝜀𝑖 ∶ 1 ≤ 𝑖 ≤ ℓ − 1}

56
 • Type 𝐷ℓ (ℓ ≥ 4):
iiiii
•
• •⋅⋅⋅• • iUUUUUU
•
𝑉 = ℚℓ , Φ = {±𝜀𝑖 ± 𝜀𝑗 ∶ 1 ≤ 𝑖, 𝑗 ≤ ℓ, 𝑖 ≠ 𝑗},
Δ = {𝜀1 + 𝜀2 , 𝜀𝑖+1 − 𝜀𝑖 ∶ 1 ≤ 𝑖 ≤ ℓ − 1}

• Type 𝐺2 .
• _jt •
Let 𝑒 = 𝜀1 + 𝜀2 + 𝜀3 ∈ ℚ3 , then:

𝑉 = {𝑣 ∈ ℚ3 ∶ (𝑣, 𝑒) = 0}, Φ = {𝜀𝑖 − 𝜀𝑗 ∶ 𝑖 ≠ 𝑗} ∪ {±(3𝜀𝑖 − 𝑒) ∶ 1 ≤ 𝑖 ≤ 3}

Δ = {𝜀1 − 𝜀2 , 𝑒 − 3𝜀1 }

• Type 𝐹4 :
• • +3 • •
𝑉 = ℚ4 ,
1
Φ = {±𝜀𝑖 ∶ 1 ≤ 𝑖 ≤ 4} ∪ {±𝜀𝑖 ± 𝑒𝑗 ∶ 𝑖 ≠ 𝑗} ∪ { (±𝜀1 ± 𝜀2 ± 𝜀3 ± 𝜀4 )}
2
1
Δ = {𝜀2 − 𝜀3 , 𝜀3 − 𝜀4 , 𝜀4 , (𝜀1 − 𝜀2 − 𝜀3 − 𝜀4 )}.
2
• Type 𝐸𝑛 (𝑛 = 6, 7, 8).

• •

• • • • • • • • • • •

• • • • • • •
These can all be constructed inside 𝐸8 by taking the span of the appropriate subset of a base, so we
just give the root system for 𝐸8 .

1 8 8
𝑉 = ℚ8 , Φ = {±𝜀𝑖 ± 𝜀𝑗 ∶ 𝑖 ≠ 𝑗} ∪ { �(−1)𝑎𝑖 𝜀𝑖 ∶ � 𝑎𝑖 ∈ 2ℤ},
2 𝑖=1 𝑖=1
1
Δ = {𝜀1 + 𝜀2 , 𝜀𝑖+1 − 𝜀𝑖 , (𝜀1 + 𝜀8 − (𝜀2 + 𝜀3 + … + 𝜀7 )) ∶ 1 ≤ 𝑖 ≤ 6}.
2

Note that the Weyl groups of type 𝐵ℓ and 𝐶ℓ are equal. The reason for the restriction on ℓ in the types
𝐵,𝐶,𝐷 is to avoid repetition, e.g. 𝐵2 and 𝐶2 are the same up to relabelling the vertices.

Remark 6.4.9. I certainly don’t expect you to remember the root systems of the exceptional types, but
you should be familiar with the ones for type 𝐴, 𝐵, 𝐶 and 𝐷. The ones of rank two (i.e. 𝐴2 , 𝐵2 and 𝐺2 ) are
also worth knowing (because for example you can draw them!)

57
6.4.1 The Classification of Semisimple Lie algebras

Only the statements of the theorems in this section are examinable, but it is important to know these statements!
Remarkably, the classification of semisimple Lie algebras is identical to the classification of root sys-
tems: each semisimple Lie algebra decomposes into a direct sum of simple Lie algebras, and it is not hard
to show that the root system of a simple Lie algebra is irreducible. Thus to any simple Lie algebra we may
attach an irreducible root system. By the conjugacy of Cartan subalgebras (see Remark 5.1.4) this gives
a well-defined map from simple Lie algebras to irreducible root systems. Then the following theorem
shows that its image classifies simple Lie algebras up to isomorphism.

Theorem 6.4.10. Let 𝔤1 , 𝔤2 be semisimple Lie algebras with Cartan subalgebras 𝔥1 , 𝔥2 respectively, and suppose
now k is of characteristic zero. Then if the root systems attached to (𝔤1 , 𝔥1 ) and (𝔤2 , 𝔥2 ) are isomorphic, there is an
isomorphism 𝜙 ∶ 𝔤1 → 𝔤2 taking 𝔥1 to 𝔥2 .

Thus to obtain a classification of simple Lie algebras, it remains to determine which irreducible root
systems are associated to a simple Lie algebra. In fact all of them are!

Theorem 6.4.11. There exists a simple Lie algebra corresponding to each irreducible root system.

Thus Theorem 6.4.10 says each irreducible root system is associated to at most one simple Lie algebra
(up to isomorphism) and so is a kind of uniqueness theorem, while Theorem 6.4.11 shows each irreducible
root system comes from a simple Lie algebra, so is an existence theorem. Theorem 6.4.10 is not difficult
given the machinery we have developed in this course. The existence statement is the more substantial
result, but we developed enough machinery to see the existence in most cases: the four infinite families
𝐴, 𝐵, 𝐶, 𝐷 correspond to the classical Lie algebras 𝔰𝔩ℓ+1 , 𝔰𝔬2ℓ , 𝔰𝔭2ℓ and 𝔰𝔬2ℓ+1 – their root systems can be
computed directly (as you say in the Problem Sheets). This of course also requires checking that these
Lie algebras are simple (or at least semisimple) but this is also straight-forward with the theory we have
developed.
It then only remains to construct the five ”exceptional” simple Lie algebras. This can be done in a
variety of ways – given a root system where all the roots are of the same length there is an explicit con-
struction of the associated Lie algebra by forming a basis from the Cartan decomposition (and a choice of
base of the root system) and explicitly constructing the Lie bracket by giving the structure constants with
respect to this basis (which, remarkably, can be chosen for the basis vectors corresponding to the root
subspaces to lie in {0, ±1}). This gives in particular a construction of the Lie algebras of type 𝐸6 , 𝐸7 , 𝐸8
(and also 𝐴ℓ and 𝐷ℓ though we already had a construction of these). The remaining Lie algebras can be
found by a technique called “folding” which studies automorphisms of simple Lie algebras, and realises
the Lie algebras 𝐺2 and 𝐹4 as fixed-points of an automorphism of 𝐷4 and 𝐸6 respectively.
There is also an alternative, more a posteriori approach to the uniqueness result which avoids showing
Cartan subalgebras are all conjugate for a general Lie algebra: one can check that for a classical Lie algebra
𝔤 ⊂ 𝔤𝔩𝑛 as above, the Cartan subalgebras are all conjugate by an element of Aut(𝔤) ∩ GL𝑛 (k). This then
shows the assignment of a root system to a classical Lie algebra is unique, so it only remains to check the
exceptional Lie algebras. But these all have different dimensions, and the dimension of the Lie algebra is
captured by the root system, so we are done.5

5
This is completely rigorous, but feels like cheating (to me).

58
Appendices

I Primary Decomposition

Throughout this section, k is an algebraically closed field unless the contrary is explicitly stated.

Definition I.1. If 𝑉 is a k-vector space and 𝑥 ∈ Endk (𝑉) and 𝜆 ∈ k, the generalized eigenspace for 𝑥 with
eigenvalue 𝜆 is
𝑉𝜆,𝑥 = {𝑣 ∈ 𝑉 ∶ ∃𝑛 ≥ 0, (𝑥 − 𝜆)𝑛 (𝑣) = 0},

Thus 𝑉𝜆,𝑥 ≠ {0} if and only if 𝑥 has an eigenvector 𝑣 ∈ 𝑉\{0} with eigenvalue 𝜆. The set of eigenvalues
of 𝑥 is called the spectrum of 𝑥, denoted 𝜎(𝑥) = {𝜆 ∈ k ∶ dim(𝑉𝜆,𝑥 ) > 0}. The subspaces 𝑉𝜆,𝑥 are clearly
invariant under the action of 𝑥, that is 𝑥(𝑉𝜆 ) ⊆ 𝑉𝜆 .

The following basic result is proved in Part A Linear Algebra. We provide a proof for the sake of com-
pleteness.

Proposition I.2. Let 𝑥 ∶ 𝑉 → 𝑉 be a linear map. There is a canonical direct sum decomposition

𝑉 = � 𝑉𝜆,𝑥 ,
𝜆∈k

of 𝑉 into the generalized eigenspaces of 𝑥. Moreover, for each 𝜆, the projection to 𝑎𝜆 ∶ 𝑉 → 𝑉𝜆,𝑥 (with kernel
the remaining generalized eigenspace of 𝑥) can be written as a polynomial in 𝑥, and hence preserve any subspace
preserved by 𝑥.

Proof. Let 𝑚𝑥 ∈ k[𝑡] be the minimal polynomial of 𝑥. Then if 𝜙 ∶ k[𝑡] → End(𝑉) given by 𝑡 ↦ 𝑥 denotes
𝑘
the natural map, we have k[𝑡]/(𝑚𝑥 ) ≅ im(𝜙) ⊆ End(𝑉). If 𝑚𝑥 = ∏𝑖=1 (𝑡 − 𝜆𝑖 )𝑛𝑖 where the 𝑆(𝑥) = {𝜆𝑖 ∶
1 ≤ 𝑖 ≤ 𝑘} is the spectrum of 𝑥, then the Chinese Remainder Theorem and the first isomorphism theorem
shows that
𝑘
im(𝜙) ≅ k[𝑡]/(𝑚𝑥 ) ≅ � k[𝑡]/(𝑡 − 𝜆𝑖 )𝑛𝑖 ,
𝑖=1
It follows that we may write 1 ∈ k[𝑡]/(𝑚𝑥 ) as 1 = 𝑒1 + … + 𝑒𝑘 according to the above decomposition. Now
clearly 𝑒𝑖 𝑒𝑗 = 0 if 𝑖 ≠ 𝑗 and 𝑒2𝑖 = 𝑒𝑖 , so that if 𝑈𝑖 = im(𝑒𝑖 ), then we have 𝑉 = ⨁ 𝑈 . Moreover, each
1≤𝑖≤𝑘 𝑖
𝑒𝑖 can be written as polynomials in 𝑥 by picking any representative in k[𝑡] of 𝑒𝑖 (thought of as an element
of k[𝑡]/(𝑚𝑥 )). Note in particular this means that each 𝑈𝑖 is invariant under im(𝜙).
Now the characteristic polynomial of 𝑥|𝑉𝜆 is clearly just (𝑡−𝜆𝑖 )𝑑𝑖 where 𝑑𝑖 = dim(𝑉𝜆𝑖,𝑥 ), and evidently
𝑖
𝑘
this divides 𝜒𝑥 (𝑡) the characteristic polynomial of 𝑥 ∈ End(𝑉). But since 𝑉 = ⨁ 𝑈𝑖 we must have
𝑖=1
𝜒𝑥 (𝑡) =∏𝑘 (𝑡 − 𝜆𝑖 )𝑚𝑖 , where 𝑚𝑖 = dim(𝑈𝑖 ) and hence 𝑑𝑖 ≤ 𝑚𝑖 . Since 𝑈𝑖 ⊆ 𝑉𝜆𝑖,𝑥 we also have 𝑚𝑖 ≤ 𝑑𝑖 ,
𝑖=1
and hence they must be equal, so 𝑉𝜆𝑖 = 𝑈𝑖 as required. 

1
 Recall that 𝔤𝔩1 is the unique 1-dimensional Lie algebra (with Lie bracket identically zero). Its irre-
ducible representations are all 1-dimensional and parameterised by k, where if 𝜆 ∈ k then (k, 𝜆) is the
𝔤𝔩1 -representation given by the Lie algerba homomorphism 𝑚𝜆 ∶ 𝔤𝔩1 → 𝔤𝔩1 given by 𝑚𝜆 (𝑡) = 𝜆.𝑡. We will
write 𝜆 for the isomorphism class of irreducible containing (k, 𝜆).

Corollary I.3. Let (𝑉, 𝜌) be a finite-dimensional representation of 𝔤𝔩1 . Then 𝑉 = ⨁ 𝑉𝜆 Ψ𝑉 is the set of
𝜆∈Ψ𝑉
composition factors of 𝑉 and 𝑉𝜆 denotes the 𝜆-isotypic subrepresentation of 𝑉, that is, the largest subrepresenta-
tion of 𝑉 whose only composition factor is 𝜆.

Proof. The equivalence of 𝔤𝔩1 -representations with the set of pairs (𝑉, 𝑇) of a vector space 𝑉 together with
a linear endomorphism 𝑇 ∶ 𝑉 → 𝑉 together with Proposition I.2 applied to the linear map 𝑇 yields the
result. 

The next Lemma is included for completeness – it readily implies that the coefficients of the charac-
teristic polynomial 𝜒𝑎 of an element 𝑎 ∈ 𝐴 of a subspace 𝐴 ⊆ 𝔤𝔩𝑉 are polynomial functions of the coor-
dinates of 𝑎 given by taking a basis of the subspace. This is used in the proof of the existence of Cartan
subalgebras.

Lemma I.4. Suppose that 𝑉 and 𝐴 are finite-dimensional vector spaces, 𝜑 ∶ 𝐴 → End(𝑉) is a linear map, and
{𝑎1 , 𝑎2 , … , 𝑎𝑘 } is a basis of 𝐴. Let
𝑑
𝜒𝑎 (𝑡) = � 𝑐𝑖 (𝑎)𝑡𝑖 ∈ k[𝑡]
𝑖=0
𝑘
be the characteristic polynomial of 𝜑(𝑎) ∈ 𝐴. Then if we write 𝑎 = ∑𝑖=1 𝑥𝑖 𝑎𝑖 , the coefficients 𝑐𝑖 (𝑎) (1 ≤ 𝑖 ≤ 𝑑)
are polynomials in k[𝑥1 , 𝑥2 , … , 𝑥𝑘 ].

Proof. Pick a basis of 𝑉 so that we may identify End(𝑉) with Mat𝑛 (k) the space of 𝑛 × 𝑛 matrices. Then
𝑗𝑘 𝑘
each 𝜑(𝑎𝑖 ) is a matrix (𝑎𝑖 )1≤𝑗,𝑘≤𝑛 , and if 𝑎 = ∑𝑖=1 𝑥𝑖 𝑎𝑖 , we have

𝑘
𝜒𝑎 (𝑡) = det(𝑡𝐼𝑛 − � 𝑥𝑖 𝜑(𝑎𝑖 )),
𝑖=1

which from the formula for the determinant clearly expands to give a polynomial in the 𝑥𝑖 and 𝑡, which
yields the result. 

2
II Tensor Products

II.1 Definition and construction

Tensor products were studied in Part B, Introduction to Representation Theory. We review their basic
properties here.

Definition II.1. If 𝑋 and 𝑌 are sets we write 𝑌𝑋 for the set of all functions from 𝑋 to 𝑌. If 𝑌 = k is a field,
and 𝑉 a k-vector spaces, then 𝑉 𝑋 is a k-vector space under pointwise addition and scalar multiplication.
For any 𝑥 ∈ 𝑋 let 𝛿𝑥 ∈ k𝑋 be the function given by 𝛿𝑥 (𝑦) = 1 if 𝑥 = 𝑦 and 0 otherwise. We define k𝑋
to be
k𝑋 = spank {𝛿𝑥 ∶ 𝑥 ∈ 𝑋}
Equivalently, k𝑋 is the space of k-valued functions on 𝑋 which vanish outside of a finite subset of 𝑋. Note
that the set {𝛿𝑥 ∶ 𝑥 ∈ 𝑋} is clearly linearly independent in k𝑋 and so they form a basis of k𝑋 and hence we
have dim(k𝑋) = |𝑋| where |𝑋| denotes the cardinality of 𝑋. . If 𝑓 ∈ k𝑋, then ∑𝑥∈𝑋 𝑓(𝑥).𝛿𝑥 and we will
often write ∑𝑥∈𝑋 𝑓(𝑥).𝑥, replacing 𝛿𝑥 with 𝑥. As a result, k𝑋 is usually referred to as the space of (finite)
“formal linear combinations of elements of 𝑋”.

Definition II.2. If 𝑉1 , 𝑉2 , … , 𝑉𝑘 and 𝑈 are vector spaces over a field k, let

ℳ (𝑉1 , … 𝑉𝑘 , 𝑈) = {𝜃 ∶ 𝑉1 × … × 𝑉𝑘 → 𝑈 ∶ 𝜃 is 𝑘-linear}

be the vector space of all 𝑘-(multi-)linear maps on 𝑉1 × … × 𝑉𝑘 taking values in a vector space 𝑈. Here
we say that a function 𝜃 ∶ 𝑉1 × 𝑉2 × … × 𝑉𝑘 → 𝑈 is a 𝑘-linear if it is linear in each component separately,
that is, if for any 𝑘-tuples of vectors (𝑣𝑖 )1≤𝑖≤𝑘 and (𝑢𝑗 )1≤𝑗≤𝑘 ∈ 𝑉1 × … × 𝑉𝑘 and any 𝜆 ∈ k, we have for each
𝑖 ∈ {1, 2, … , 𝑘},

𝜃(𝑣1 , … , 𝜆.𝑣𝑖 + 𝑢𝑖 , … 𝑣𝑘 ) = 𝜆.𝜃(𝑣1 , … , 𝑣𝑖 , … 𝑣𝑘 ) + 𝜃(𝑣1 , … , 𝑢𝑖 , … 𝑣𝑘 ),

Pick a basis 𝐵𝑖 of 𝑉𝑖 for each 𝑖 (1 ≤ 𝑖 ≤ 𝑘), and let 𝐵∗𝑖 denote the corresponding dual basis of 𝑉𝑖∗ . If
𝑏 ∈ 𝐵𝑖 , let 𝛿𝑏 denote the corresponding element of the dual basis 𝐵∗𝑖 , so that 𝐵∗𝑖 = {𝛿𝑏 ∶ 𝑏 ∈ 𝐵𝑖 }. Let
B = 𝐵 1 × 𝐵2 × … × 𝐵 𝑘

Proposition II.3. In the notation given above, the restriction to B gives an isomorphism

𝑟B ∶ ℳ (𝑉1 , … , 𝑉𝑘 ; 𝑈) → 𝑈 B

from the space of all 𝑘-multilinear maps taking values in 𝑈 to the space of all 𝑈-valued functions on B. Indeed 𝑟B
has inverse given explicitly by

ℱB (𝑓)(𝑣1 , … , 𝑣𝑘 ) = � 𝛿𝑏1 (𝑣1 ) … 𝛿𝑏𝑘 (𝑣𝑘 )𝑓(b).

b=(𝑏1 ,…,𝑏𝑘 )∈B

Note that since, for each 𝑖 ∈ {1, … , 𝑘} we have 𝑣𝑖 = ∑𝑏∈𝐵 𝛿𝑏 (𝑣𝑖 ).𝑏, 𝛿𝑏 (𝑣𝑖 ) ≠ 0 for only finitely many 𝑏 ∈ 𝐵𝑖 ,
𝑖
hence the sum on the right-hand side has only finitely many nonzero terms.

Proof. First note that if b = (𝑏1 , … , 𝑏𝑘 ) ∈ B then the product 𝛿b = 𝛿𝑏1 .𝛿𝑏2 … 𝛿𝑏𝑘 is a 𝑘-linear map (since
multiplication distributes over addition). Moreover, it is clear that 𝛿b (b′ ) = 𝛿b,b′ (that is, is zero un-
less b = b′ in which case it is equal to 1), it is immediate that 𝑟B (ℱB (𝑓)) = 𝑓, so we must show that
ℱB (𝑟B (𝜃)) = 𝜃 for any 𝜃 ∈ ℳ (𝑉1 , … , 𝑉𝑘 ; 𝑈). Explicitly, we must show that

𝜃= � 𝛿b 𝜃(b) (II.1)
b∈𝐵1 ×…×𝐵𝑘

3
Indeed applying 𝜃 to a 𝑘-tuple b ∈ 𝐵1 × … 𝐵𝑘 , we see that the coefficients on the right-hand side are
uniquely determined, so it remains to show the products 𝛿b of dual basis vectors do indeed span.
The case 𝑘 = 1 is simply the standard argument that the functions {𝛿𝑏1 }𝑏1∈𝐵1 are indeed a basis of
𝑉1∗ :if 𝑣1 ∈ 𝑉1 then we may write 𝑣1 = ∑𝑏 ∈𝐵 𝜆𝑏1 𝑏1 for unique scalars 𝜆𝑏1 ∈ k. By the definition of
1 1
the functions 𝛿𝑏1 , it then follows that 𝛿𝑏1 (𝑣1 ) = 𝜆𝑏1 , so that 𝑣1 = ∑𝑏 ∈𝐵 𝛿𝑏1 (𝑣1 ).𝑏1 . Applying 𝜃 gives
1 1
𝜃(𝑣1 ) = ∑𝑏 ∈𝐵 𝛿𝑏1 (𝑣1 ).𝜃(𝑏1 ). But as this holds for all 𝑣1 ∈ 𝑉1 , it follows that 𝜃 = ∑𝑏 ∈𝐵 𝜃(𝑏1 ).𝛿𝑏1 , as
1 1 1 1
required.
The general case then follows by an easy induction: Indeed for any 𝑘-tuple of vectors (𝑣𝑖 )1≤𝑖≤𝑘 with
𝑣𝑖 ∈ 𝑉𝑖 , using the case 𝑘 = 1, we may write 𝑣1 = ∑𝑏 ∈𝐵 𝛿𝑏1 (𝑣1 ).𝑏1 . But then if 𝜃 is 𝑘-linear we have
1 1

𝜃(𝑣1 , … , 𝑣𝑘 ) = 𝜃( � 𝛿𝑏1 (𝑣1 ).𝑏1 , 𝑣2 , … , 𝑣𝑘 ) = � 𝛿𝑏1 (𝑣1 ).𝜃(𝑏1 , 𝑣2 , … , 𝑣𝑘 ).

𝑏1 ∈𝐵1 𝑏1 ∈𝐵1

But for each 𝑏1 ∈ 𝐵1 , the map (𝑣𝑖 )2≤𝑖≤𝑘 ↦ 𝜃(𝑏1 , 𝑣2 , … , 𝑣𝑘 ) is a (𝑘 − 1)-linear map from 𝑉2 × … 𝑉𝑘 to k,
hence the result follows by induction. 

Remark II.4. Note that, for 𝑘 = 1, this says that a linear map is uniquely determined by its values on a
basis of 𝑉1 , and the statement should be thought of as saying that a 𝑘-linear map is similarly determined
“by its values on bases” where the statement of the question gives the precise meaning to the vague phrase
in quotation marks.

The previous Proposition gives one way of constructing the tensor product: If 𝑉 and 𝑊 are k-vector
spaces and we pick bases 𝐵𝑉 and 𝐵𝑊 of 𝑉 and 𝑊 respectively, then by the Proposition, if we set 𝐵 =
𝐵𝑉 × 𝐵𝑊 , then for any vector space 𝑈, we have

ℳ (𝑉, 𝑊; 𝑈) ≅ 𝑈 𝐵 ≅ ℒk (k𝐵, 𝑈), (II.2)

where k𝐵 is, as above, the vector space with basis 𝐵, that is, the space of finite formal linear combinations
of elements of 𝐵. The first isomorphism above is a direct consequence of the Proposition where we take
𝑘 = 2 and 𝑉1 = 𝑉, 𝑉2 = 𝑊, while the second is the case 𝑘 = 1 of the proposition with 𝑉1 = k𝐵.
Now taking 𝑈 = k𝐵 in (II.2), the identity linear map from k𝐵 to itself corresponds to a bilinear map
𝑡 ∶ 𝑉 × 𝑊 → k𝐵.
Lemma II.5. The bilinear map 𝑡 ∶ 𝑉 × 𝑊 → k𝐵 has the universal property, so that the pair (k𝐵, 𝑡) is the tensor
product of 𝑉 and 𝑊.

Proof. This is essentially established in the previous paragraph: if 𝜃 ∶ 𝑉 × 𝑊 → 𝑈 is bilinear, then since
𝜃|𝐵 ∶ 𝐵 → 𝑈 extends to a linear map 𝜃̃ ∶ k𝐵 → 𝑈. Tracking how the isomorphism of Proposition II.3
identifies 𝜃 with the linear map 𝜃̃ it is easy to see that this can be expressed by means of the bilinear map
𝑡 ∶ 𝑉 × 𝑊 → k𝐵 as 𝜃 = 𝜃̃ ∘ 𝑡. 

Remark II.6. Note that there is a natural isomorphism 𝜎 ∶ 𝑉 ⊗ 𝑊 ≅ 𝑊 ⊗ 𝑉 given by 𝑣 ⊗ 𝑤 ↦ 𝑤 ⊗ 𝑣. If

𝑉 ≠ 𝑊, we will normally abuse notation and identify these two spaces and thus write 𝑉 ⊗ 𝑊 = 𝑊 ⊗ 𝑉.
If 𝑉 = 𝑊 however, 𝜎 ∶ 𝑉 ⊗ 𝑉 → 𝑉 ⊗ 𝑉 is an involution on 𝑉 ⊗ 𝑉, and more generally, 𝑉 ⊗𝑛 = 𝑉 ⊗ … ⊗ 𝑉,
the tensor product of 𝑉 with itself 𝑛 times, has an action of 𝑆𝑛 the symmetric group, which permutes the
tensor factors: if 𝜏 ∈ 𝑆𝑛 then 𝜏(𝑣1 ⊗ … ⊗ 𝑣𝑛 ) ∶= 𝑣𝜏(1) ⊗ … ⊗ 𝜏𝜏(𝑛) .
Example II.7. If 𝑉 = k and 𝑊 is an arbitrary k-vector space , then if 𝑠 ∶ k×𝑊 → 𝑊 is scalar multiplication
map given by 𝑠(𝜆, 𝑤) = 𝜆.𝑤, it is clearly bilinear and it is straight-forward to check that it has the universal
property so that k ⊗ 𝑊 ≅ 𝑊.
Note also that ℒ(k, 𝑊) is canonically isomorphic to 𝑊, since if 𝜙 ∈ ℒ(k, 𝑊) then 𝜙(1) ∈ 𝑊, while
if 𝑤 ∈ 𝑊, then 𝑠𝑤 ∶ k → 𝑊 given by 𝑠𝑤 (𝑡) = 𝑡.𝑤 is an element of ℒ(k, 𝑊). Clearly 𝑠𝑤 (1) = 𝑤, so that
𝜙 ↦ 𝜙(1) and 𝑤 ↦ 𝑠𝑤 are inverses of each other,

4
 The construction of the tensor product allows us to replace bilinear maps with linear ones, but one
can also relate bilinear maps to “linear maps to spaces of linear maps” – which is really just the process
of taking a function of two variables and holding one variable fixed in order to obtain a function of one
variable. Formally, we can state:

Lemma II.8. Let 𝑉, 𝑊 and 𝑈 be vector spaces over a field k. Then we have natural isomorphisms

ℒ(𝑉, ℒ(𝑊, 𝑈)) ≅ ℳ (𝑉, 𝑊; 𝑈) = ℒ(𝑉 ⊗ 𝑊, 𝑈) ≅ ℒ(𝑊, ℒ(𝑉, 𝑈))

In particular, if 𝑈 = k, this shows that ℒ(𝑉, 𝑊 ∗ ) ≅ ℳ (𝑉, 𝑊; k) = (𝑉 ⊗ 𝑊)∗ .

Proof. Since there is an obvious identification between ℳ (𝑉, 𝑊; 𝑈) ≅ ℳ (𝑊, 𝑉; 𝑈) it suffices to estab-
lish the first isomorphism. But if 𝜃 ∈ ℒ(𝑉, ℒ(𝑊, 𝑈)), then we let Ψ(𝜃)(𝑣, 𝑤) = 𝜃(𝑣)(𝑤). The fact that
𝜃 is linear shows that Ψ(𝜃) is linear in 𝑉, while the fact that 𝜃(𝑣) lies in ℒ(𝑊, 𝑈) shows that Ψ(𝜃) is
linear in 𝑊. Conversely, if 𝑏 ∈ ℳ (𝑉, 𝑊; 𝑈), then we may define Υ(𝑏)(𝑣) = [𝑤 ↦ 𝑏(𝑣, 𝑤)]. The map
Υ(𝑏) ∈ ℒ(𝑊, 𝑈) because 𝑏 is linear in 𝑊, while the map Υ is linear because 𝑏 is linear in 𝑉. It is clear
that Ψ and Υ are inverse to each other, thus the first isomorphism is established.
The final claim follows immediately. 

Definition II.9. Let 𝑈, 𝑉, 𝑊 be k-vector spaces. If 𝜙 ∈ ℒ(𝑈, 𝑉) and 𝜓 ∈ ℒ(𝑉, 𝑊), then the map
(𝜓, 𝜙) ↦ 𝜓 ∘ 𝜙 given by composition of functions induces a linear map

𝑐 ∶ ℒ(𝑉, 𝑊) ⊗ ℒ(𝑈, 𝑉) → ℒ(𝑈, 𝑊).

Note that if we take 𝑉 = k, then ℒ(k, 𝑊) is canonically isomorphic to 𝑊 via the map which sends 𝜙 ∈
ℒ(k, 𝑊) to 𝜙(1) ∈ 𝑊. Since ℒ(𝑈, k) is just the dual space 𝑈 ∗ of 𝑈, in this case composition yields a map
𝜗 ∶ 𝑉 ∗ ⊗ 𝑊 ≅ 𝑊 ⊗ 𝑉 ∗ → ℒ(𝑉, 𝑊). Specializing again to the case where 𝑈 = 𝑊 one can also compose
in the oppposite order to obtain a map 𝜄 ∶ 𝑈 ⊗ 𝑈 ∗ → k.
Explicitly we have 𝜗(𝑓 ⊗ 𝑤) = [𝑢 ↦ 𝑓(𝑢).𝑤], (∀ 𝑓 ∈ 𝑈 ∗ , 𝑤 ∈ 𝑊) and 𝜄(𝜆 ⊗ 𝑢) = 𝜆(𝑢) (∀ 𝜆 ∈ 𝑈 ∗ , 𝑢 ∈ 𝑈).

Proposition II.10. Let 𝑉 and 𝑊 be vector spaces. The natural map 𝜗 ∶ 𝑉 ∗ ⊗ 𝑊 → ℒ(𝑉, 𝑊) is injective with
𝑓𝑟
image the space ℒ (𝑉, 𝑊) = {𝛼 ∶ 𝑉 → 𝑊 ∶ dim(im(𝛼)) < ∞} of linear maps of finite rank. Moreover, when
𝑉 is finite-dimensional, if 𝜄 ∶ 𝑉 ∗ ⊗𝑉 → k is the contraction map and 𝛼 ∈ Homk (𝑉, 𝑉), then (𝜄∘𝜃−1 )(𝛼) = tr(𝛼).

Proof. First note that if 𝑓 ∈ 𝑉 ∗ , 𝑤 ∈ 𝑊 then 𝜃(𝑓 ⊗ 𝑣) = 𝑓.𝑤 has image contained in k.𝑤, and hence has
𝑚
rank at most 1. It follows that 𝜃(∑𝑘=1 𝑓𝑖 ⊗ 𝑤𝑖 ) has rank at most 𝑚, so that im(𝜃) ⊆ ℒ𝑓𝑟 (𝑉, 𝑊). To see
the reverse containment and that 𝜃 is injective, let 𝑊1 be a finite-dimensional subspace of 𝑊. We claim
that the restriction of 𝜃 to 𝑉 ∗ ⊗ 𝑊1 induces an isomorphism 𝜃1 ∶ 𝑉 ∗ ⊗ 𝑊1 → ℒ(𝑉, 𝑊1 ). Since any
𝑡 ∈ 𝑉 ∗ ⊗ 𝑊 lies in 𝑉 ∗ ⊗ 𝑊1 for some finite dimensional subspace of 𝑊, it follows that ker(𝜃) = {0}, that
is, 𝜃 is injective.
Let 𝛼 ∈ ℒ(𝑉, 𝑊1 ). Pick a basis {𝑤1 , … , 𝑤𝑘 } for 𝑊1 , and let {𝛿1 , … , 𝛿𝑘 } be the corresponding dual
𝑘
basis of 𝑊1∗ so that if 𝑤 ∈ 𝑊1 then 𝑤 = ∑𝑖=1 𝛿𝑖 (𝑤).𝑤𝑖 . But then if 𝑣 ∈ 𝑉, 𝛼(𝑣) ∈ 𝑊1 so that
𝑘 𝑘 𝑘
𝛼(𝑣) = ∑𝑖=1 𝛿𝑖 (𝛼(𝑣)).𝑤𝑖 = ∑𝑖=1 𝛼⊺ (𝛿𝑖 )(𝑣) ⊗ 𝑤𝑖 . Thus it we set 𝜗1 (𝛼) = ∑𝑖=1 𝛼⊺ (𝛿𝑖 ) ⊗ 𝑤𝑖 , it follows
that 𝜃1 (𝜗1 )(𝛼)(𝑣) = 𝛼(𝑣), for all 𝑣 ∈ 𝑉. Moreover, if 𝑓 ⊗ 𝑤 ∈ 𝑉 ∗ ⊗ 𝑊1 , then 𝜃(𝑓 ⊗ 𝑣)⊺ (𝜂) = 𝜂(𝑤).𝑓 and
hence
𝑘 𝑘
𝜗1 ∘ 𝜃1 (𝑓 ⊗ 𝑤) = � 𝛿𝑖 (𝑤)𝑓 ⊗ 𝑤𝑖 = 𝑓 ⊗ � 𝛿𝑖 (𝑤)𝑤𝑖 = 𝑓 ⊗ 𝑤.
𝑖=1 𝑖=1

It follows 𝜗1 = 𝜃−1
1 , and 𝜃1 is an isomorphism as claimed.
Finally, consider the contraction map 𝜄 ∶ 𝑉 ∗ × 𝑉 → k. This is again composition, but now in the
opposite order, so that 𝑣 ∶ k → 𝑉 and 𝑓 ∶ 𝑉 → k compose to give 𝑓(𝑣) ∈ k. Applying the above to

5
𝑊1 = 𝑉 = 𝑊, so that {𝑤1 , … , 𝑤𝑘 } is now a basis of 𝑉 and {𝛿1 , … , 𝛿𝑘 } the corresponding dual basis of 𝑉 ∗ ,
we see that
𝑘 𝑘
𝜄 ∘ 𝜃−1 (𝛼) = 𝜄(� 𝛼⊺ (𝛿𝑖 ) ⊗ 𝑤𝑖 ) = � 𝛿𝑖 (𝛼(𝑤𝑖 )) = tr(𝛼).
𝑖=1 𝑖=1


Remark II.11. Since we only use the cases where 𝑉 and 𝑊 are finite dimensional, the reader is welcome
to ignore the generality the result is stated in and assume throughout that all vector spaces are finite di-
mensional. Here one can be a bit more concrete: if {𝑒1 , … , 𝑒𝑛 } is a basis of 𝑉 and {𝑓1 , … , 𝑓𝑚 } is a basis of 𝑊,
then taking the dual basis {𝛿1 , … , 𝛿𝑛 } of 𝑉 ∗ it is easy to see that the images of 𝛿𝑖 ⊗ 𝑓𝑗 under 𝜗 correspond
to the elementary matrices 𝐸𝑖𝑗 under the identification of Homk (𝑉, 𝑊) given by the choice of bases for 𝑉
and 𝑊, hence 𝜗 is an isomorphism.

Remark II.12. We will usually abuse notation somewhat and write 𝜄 ∶ 𝑉 ⊗𝑉 ∗ → k rather than 𝜄∘𝜎 where
𝜎 ∶ 𝑉 ⊗ 𝑉 ∗ → 𝑉 ∗ ⊗ 𝑉 interchanges the tensor factors.

II.1.1 Linear maps between tensor products.

Let 𝛼 ∶ 𝑉1 → 𝑉2 and 𝛽 ∶ 𝑊1 → 𝑊2 be linear maps. If 𝑣 ∈ 𝑉1 , 𝑤 ∈ 𝑊1 , the map (𝑣, 𝑤) ↦ 𝛼(𝑣) ⊗ 𝛽(𝑤)

from 𝑉1 × 𝑊2 → 𝑉2 ⊗ 𝑊2 is bilinear, and so induces a linear map 𝛼 ⊗ 𝛽 ∈ Hom(𝑉1 ⊗ 𝑊1 , 𝑉2 ⊗ 𝑊2 ). In
fact, the map (𝛼, 𝛽) ↦ 𝛼 ⊗ 𝛽 is itself bilinear, and so we even obtain a map

Hom(𝑉1 , 𝑊1 ) ⊗ Hom(𝑉2 , 𝑊2 ) → Hom(𝑉1 ⊗ 𝑉2 , 𝑊1 ⊗ 𝑊2 ). (II.3)

Moreover, it follows immediately from the definitions that (II.3) also respects composition. In more detail,
if 𝛼2 ∶ 𝑉2 → 𝑉3 and 𝛽2 ∶ 𝑊2 → 𝑊3 are linear maps to any vector spaces 𝑉3 and 𝑊3 , then (𝛼2 ⊗ 𝛽2 ) ∘
(𝛼1 ⊗ 𝛽1 ) = (𝛼2 ∘ 𝛼1 ) ⊗ (𝛽2 ∘ 𝛽1 ). Indeed, if 𝑣 ∈ 𝑉1 , 𝑤 ∈ 𝑊1 , then

(𝛼2 ⊗ 𝛽2 ) ∘ (𝛼1 ⊗ 𝛽1 )(𝑣 ⊗ 𝑤) = (𝛼2 ⊗ 𝛽2 )(𝛼1 (𝑣) ⊗ 𝛽1 (𝑤))

= (𝛼2 ∘ 𝛼1 )(𝑣) ⊗ (𝛽2 ∘ 𝛽1 )(𝑤)
= (𝛼2 ∘ 𝛼1 ) ⊗ (𝛽2 ∘ 𝛽1 )(𝑣 ⊗ 𝑤).

When all the vector spaces 𝑉1 , 𝑉2 , 𝑊1 , 𝑊2 are finite dimensional, the map (II.3) is actually an iso-
morphism, indeed using Lemma II.10 you can check that

Hom(𝑉1 , 𝑊1 ) ⊗ Hom(𝑉2 , 𝑊2 ) ≅ (𝑉1∗ ⊗ 𝑊1 ) ⊗ (𝑉2∗ ⊗ 𝑊2 )

≅ (𝑉1∗ ⊗ 𝑉2∗ ) ⊗ (𝑊1 ⊗ 𝑊2 )
≅ (𝑉1 ⊗ 𝑉2 )∗ ⊗ (𝑊1 ⊗ 𝑊2 )
≅ Hom(𝑉1 ⊗ 𝑉2 , 𝑊1 ⊗ 𝑊2 ),

where the second isomorphism simply permutes the second and third tensor factors.

Example II.13. The map 𝜄𝑉 ∶ 𝑉 ∗ ⊗ 𝑉 → k arises as an instance of composition, but in fact, at least in the
finite-dimensional case, one can express composition of linear maps in terms of the maps 𝜄𝑉 : Suppose
that 𝑈, 𝑉 and 𝑊 are k-vector spaces and that 𝑈 and 𝑉 are finite-dimensional. Because 𝑈 and 𝑉 are
finite-dimensional, we may identify ℒ(𝑈, 𝑉) and ℒ(𝑉, 𝑊) with 𝑈 ∗ ⊗ 𝑉 and 𝑉 ∗ ⊗ 𝑊 respectively, and
ℒ(𝑈, 𝑊) with 𝑈 ∗ ⊗ 𝑊. Composition of linear maps thus corresponds a linear map

𝑐 ∶ (𝑈 ∗ ⊗ 𝑉) ⊗ (𝑉 ∗ ⊗ 𝑊) → 𝑈 ∗ ⊗ 𝑊

To identify it, we consider the composition of two rank-1 linear maps 𝛼 ∶ 𝑈 → 𝑉 and 𝛽 ∶ 𝑉 → 𝑊, say
𝛼 = 𝑓.𝑣, 𝛽 = 𝑔.𝑤 (𝑓 ∈ 𝑈 ∗ , 𝑔 ∈ 𝑉 ∗ , 𝑣 ∈ 𝑉, 𝑤 ∈ 𝑊). Then 𝛽 ∘ 𝛼 = 𝑔(𝑓.𝑣)(𝑤) = 𝑔(𝑣).(𝑓.𝑤), and hence
(𝑓 ⊗ 𝑣) ⊗ (𝑔 ⊗ 𝑤) ↦ 𝑔(𝑣).(𝑓 ⊗ 𝑤). In other words, 𝑐 = 1𝑈 ∗ ⊗ 𝜄𝑉 ⊗ 1𝑊 .

6
Remark II.14. It is sometimes useful to have the following notational convention: Given a tensor product
of more than two vector spaces, such as 𝑈 ∗ ⊗𝑉 ⊗𝑉 ∗ ⊗𝑊, then it can be convenient to write 𝜄32 for the map
which acts via 𝜄 on the third and second factors (that is swapping the second and third factors, applying
𝜄 and the repeating the swap) and by the identity on the remaining tensor factors.

II.1.2 Tensor products and duality

Suppose that 𝑉 and 𝑊 are finite dimensional vector spaces. We wish to understand the relationship be-
tween the tensor product of the dual spaces 𝑉 ∗ ⊗ 𝑊 ∗ and the dual space of the tensor product (𝑉 ⊗ 𝑊)∗ .
But by Proposition II.10, we have an injective map 𝜃 ∶ 𝑉 ∗ ⊗ 𝑊 ∗ → ℒ(𝑉, 𝑊 ∗ ). Now ℒ(𝑉, ℒ(𝑊, k)) ≅
ℒ(𝑉 ⊗ 𝑊, k) = (𝑉 ⊗ 𝑊)∗ by Lemma II.8, and so composing we obtain an embedding 𝑑𝑉,𝑊 ∶ 𝑉 ∗ ⊗ 𝑊 ∗ →
(𝑉 ⊗ 𝑊)∗ . Explicitly,

𝑑𝑉,𝑊 (𝜂 ⊗ 𝜈)(𝑣 ⊗ 𝑤) = 𝜂(𝑣).𝜈(𝑤), ∀ 𝜂 ∈ 𝑉 ∗ , 𝜈 ∈ 𝑊 ∗ , 𝑣 ∈ 𝑉, 𝑤 ∈ 𝑊,

i.e. 𝑑𝑉,𝑊 is given by multiplication of function. Moreover, 𝑑𝑉,𝑊 is an isomorphism provided 𝑊 is

finite-dimensional, and hence by symmetry, 𝑑𝑉,𝑊 is an isomorphism provided one of 𝑉 or 𝑊 is finite-
dimensional.
Next note that if 𝑉 and 𝑊 are finite-dimensional, we have

ℒ(𝑉, 𝑊) ⊗ ℒ(𝑊, 𝑉) ≅ (𝑉 ∗ ⊗ 𝑊) ⊗ (𝑊 ∗ ⊗ 𝑉)
≅ (𝑉 ∗ ⊗ 𝑊 ∗ ) ⊗ (𝑉 ⊗ 𝑊)
→ (𝑉 ⊗ 𝑊)∗ ⊗ (𝑉 ⊗ 𝑊) → k

where we use Proposition II.10 in the first line, permutation of the tensor factors in the second and the
contraction 𝜄𝑉⊗𝑊 in the final line. If 𝛼 = 𝑓.𝑤 and 𝛽 = 𝑔.𝑣 are rank-1 linear maps in ℒ(𝑉, 𝑊) and
ℒ(𝑊, 𝑉) respectively, then the above map sends (𝑓.𝑤) ⊗ (𝑔.𝑣) → 𝑓(𝑣)𝑔(𝑤) = tr𝑊 (𝛼 ∘ 𝛽) = tr𝑉 (𝛽 ∘ 𝛼),
and since rank-1 maps span ℒ(𝑉, 𝑊) provided 𝑊 is finite-dimensional, it follows that the trace pairing
𝑡(𝛼, 𝛽) ≔ tr(𝛽 ∘ 𝛼) = tr(𝛼 ∘ 𝛽) is a nondegenerate bilinear pairing between ℒ(𝑉, 𝑊) and ℒ(𝑊, 𝑉).

Remark II.15. Note that the contraction map 𝜄 ∶ 𝑉 ∗ ⊗ 𝑉 → k, 𝜄(𝑓 ⊗ 𝑣) = 𝑓(𝑣) is an element of (𝑉 ∗ ⊗ 𝑉)∗ .
If 𝑉 is finite-dimensional, then (𝑉 ∗ ⊗ 𝑉)∗ ≅ (𝑉 ∗ )∗ ⊗ 𝑉 ≅ 𝑉 ∗ ⊗ 𝑉 ≅ 𝑉 ∗ ⊗ 𝑉, hence we may identify 𝜄 with
an element of 𝑉 ⊗ 𝑉 ∗ . It is easy to check that this element corresponds to the identity linear map under
the isomorphism 𝜃 ∶ 𝑉 ∗ ⊗ 𝑉 → ℒ(𝑉, 𝑉).

Remark II.16. The isomorphisms of Lemma II.8 come from a very basic bijection: if 𝑋, 𝑌 and 𝑍 are sets
and we write Map(𝑋, 𝑌) for the set of all functions from 𝑋 to 𝑌 then

Map(𝑋 × 𝑌, 𝑍) = Map(𝑋, Map(𝑌, 𝑍))

where the identification is as follows: if 𝜃 ∶ 𝑋 × 𝑌 → 𝑍 then we identify 𝜃 with the function 𝜗 ∶ 𝑋 →

Map(𝑌, 𝑍), where if we let, for any 𝑥 ∈ 𝑋 the function 𝑖𝑥 ∶ 𝑌 → 𝑋 × 𝑌 be given by 𝑖𝑥 (𝑦) = (𝑥, 𝑦), then
𝜗(𝑥) = 𝜃 ∘ 𝑖𝑥 . Equivalently, 𝜗(𝑥) is the function on 𝑌 obtained by restricting 𝜃 the copy of 𝑌 indexed by
𝑥 in 𝑋 × 𝑌. It is trivial to check that the inverse of this map associates to 𝜗 the function 𝜃 ∶ 𝑋 × 𝑌 → 𝑍
given by 𝜃(𝑥, 𝑦) = 𝜗(𝑥)(𝑦).
Note that if we write 𝑌𝑋 for Map(𝑋, 𝑌) (so that, for example, 2𝑋 is set of functions from 𝑋 to the
2-element set {0, 1}, which is readily identified with the power set of 𝑋) then the above bijection gives
a natural identification between 𝑍𝑋×𝑌 and (𝑍𝑌 )𝑋 , that is, it becomes a set-theoretic incarnation of the
power law!

7
II.2 Bilinear forms

Definition II.17. Let 𝑉 be a k-vector space. A bilinear form on 𝑉 is a bilinear map 𝐵 ∶ 𝑉 × 𝑉 → k, that is,
an element of ℳ (𝑉, 𝑉; k). We will denote the vector space of all bilinear forms on 𝑉 as Bil(𝑉). From the
universal property of tensor products, Bil(𝑉) ≅ (𝑉 ⊗ 𝑉)∗ .
There is an action of 𝑆2 = {𝑒, 𝜎}, the symmetric group on two letters, on Bil(𝑉) given by 𝜎(𝐵)(𝑣, 𝑤) =
𝐵(𝑤, 𝑣) (for any 𝑣, 𝑤 ∈ 𝑉). A form 𝐵 is said to be symmetric if 𝐵 = 𝜎(𝐵), that is if 𝐵(𝑣, 𝑤) = 𝐵(𝑤, 𝑣) for all
𝑣, 𝑤 ∈ 𝑉.
If 𝐵 ∈ Bil(𝑉) satisfies 𝐵(𝑣, 𝑣) = 0 we say that 𝐵 is alternating. By expanding 0 = 𝐵(𝑣 + 𝑤, 𝑣 +
𝑤) we see that 𝐵(𝑣, 𝑤) = −𝐵(𝑤, 𝑣), or 𝜎(𝐵) = −𝐵 that is, 𝐵 is skew-symmetric. By our assumption that
char(k) ≠ 2, the condition that 𝐵 is skew-symmetric implies that 𝐵 is alternating6 and moreover, if we
write 𝑆2 (𝑉) for the space of symmetric bilinear forms on 𝑉 and Λ2 (𝑉) for the space of alternating bilinear
forms on 𝑉, then we have Bil(𝑉) = 𝑆2 (𝑉) ⊕ Λ2 (𝑉). Indeed 𝜎 is an involution, hence its action on Bil(𝑉)
is diagonalisable and its eigenvalues are ±1.
We may deal with the symmetric and skew-symmetric cases uniformly (to some extent) by working
with a form 𝐵 which has the property that 𝐵(𝑣, 𝑤) = 𝜖.𝐵(𝑤, 𝑣) for all 𝑣, 𝑤 ∈ 𝑉, where 𝜖 ∈ {±1}.

Remark II.18. The natural map of Lemma II.8

Θ ∶ Bil(𝑉) = ℒ(𝑉 ⊗ 𝑉, k) → ℒ(𝑉, ℒ(𝑉, k)) = ℒ(𝑉, 𝑉 ∗ ),

where if 𝑏 ∶ 𝑉 × 𝑉 → k is a bilinear form then its image Θ 𝑏 under Θ is given by

Θ 𝑏 (𝑣1 )(𝑣2 ) = 𝑏(𝑣1 , 𝑣2 ), ∀ 𝑣1 , 𝑣2 ∈ 𝑉.

Note that the action of 𝜎 ∈ 𝑆2 gives a second isomorphism 1 Θ ∶ Bil(𝑉) → ℒ(𝑉, 𝑉 ∗ ), where 1 Θ = Θ ∘
𝜎, that is, 1 Θ 𝑏 (𝑣1 )(𝑣2 ) = 𝑏(𝑣2 , 𝑣1 ). For symmetric bilinear forms the two maps agree, but for arbitrary
bilinear forms they yield different isomorphisms.

Definition II.19. Given a bilinear form 𝐵, we set

rad(𝐵) = rad𝐿 (𝐵) = ker(Θ 𝐵 ) = {𝑣 ∈ 𝑉 ∶ 𝐵(𝑣, 𝑤) = 0, ∀ 𝑤 ∈ 𝑉}

(here the subscript “𝐿” denotes “left”). Similarly, we set

rad𝑅 (𝐵) = ker(1 Θ 𝐵 ) = {𝑣 ∈ 𝑉 ∶ 𝐵(𝑤, 𝑣) = 0, ∀ 𝑤 ∈ 𝑉}.

If 𝐵 is symmetric or alternating, then rad𝐿 (𝐵) = rad𝑅 (𝐵), but this need not be true otherwise. We say that
𝐵 is non-degenerate if rad𝐿 (𝐵) = {0}. Note that, even though in general rad𝐿 (𝐵) ≠ rad𝑅 (𝐵), it is still the
case that rad𝐿 (𝐵) = {0} if and only if rad𝑅 (𝐵) = {0}.

From now on we will only work with symmetric and alternating bilinear forms. Fix such a form 𝐵 ∈
Bil(𝑉) so that 𝜎(𝐵) = 𝜖.𝐵 for some 𝜖 ∈ {±1}. Then if 𝑈 is a subspace of 𝑉, we define

𝑈 ⟂ = {𝑣 ∈ 𝑉 ∶ 𝐵(𝑣, 𝑤) = 0, ∀ 𝑤 ∈ 𝑈} = {𝑣 ∈ 𝑉 ∶ Θ(𝐵)(𝑣) ∈ 𝑈 0 }.

When 𝐵 is nondegenerate, so that Θ 𝐵 is an isomorphism, this shows that dim(𝑈 ⟂ ) = dim(𝑈 0 ) =

dim(𝑉) − dim(𝑈). The next Lemma shows that this can be refined slightly.
6
If 𝑏 is alternating, then 0 = 𝑏(𝑣 + 𝑤, 𝑣 + 𝑤) = 𝑏(𝑣, 𝑣) + 𝑏(𝑣, 𝑤) + 𝑏(𝑤, 𝑣) + 𝑏(𝑤, 𝑤) = 𝑏(𝑣, 𝑤) + 𝑏(𝑤, 𝑣), so that 𝑏 is skew-
symmetric. If char(k) = 2 however, skew-symmetry is equivalent to symmetry, and a symmetric form need not be alternating,
for example the standard “dot product” ⟨𝑥, 𝑦⟩1 = 𝑥1 𝑦1 + 𝑥2 𝑦2 on k2 is a nondegenerate symmetric bilinear form, and ⟨𝑥, 𝑥⟩1 = 0
if and only if 0 = 𝑥21 + 𝑥22 = (𝑥1 + 𝑥2 )2 , so that (1, 1) spans the line of isotropic vectors for ⟨., .⟩1 . On the other hand if ⟨𝑥, 𝑦⟩2 =
𝑥1 𝑦2 + 𝑥2 𝑦1 , then ⟨𝑥, 𝑥⟩2 = 0 for all 𝑥 ∈ k2 . Note however that the set of vectors 𝑁𝑖 = {𝑥 ∈ k2 ∶ ⟨𝑥, 𝑥⟩𝑖 = 0} is a linear subspace
of k2 for 𝑖 = 1 and 2, and it is easy to see that this is always the case when char(k) = 2.

8
Lemma II.20. Let 𝑉 be a finite-dimensional k-vector space equipped with a symmetric (or alternating) bilinear
form 𝐵. Then for any subspace 𝑈 of 𝑉 we have the following:

i) dim(𝑈) + dim(𝑈 ⟂ ) ≥ dim(𝑉).

ii) The restriction of 𝐵 to 𝑈 is nondegenerate if and only if 𝑉 = 𝑈 ⊕ 𝑈 ⟂ .

Proof. Let 𝑟𝑈 ∶ 𝑉 ∗ → 𝑈 ∗ be given by 𝑓 ↦ 𝑓 ∘ 𝜄 where 𝜄 ∶ 𝑈 → 𝑉 is the inclusion map. (Thus 𝑟𝑈 restricts

the domain of an element of 𝑉 ∗ from 𝑉 to 𝑈.) Given 𝐵 ∈ Bil(𝑉) we set 𝜙𝐵 = 𝑟𝑈 ∘ Θ 𝐵 ∶ 𝑉 → 𝑈 ∗ , that
is 𝜙(𝑣) = Θ 𝐵 (𝑣)|𝑈 . Clearly ker(𝜙) = 𝑈 ⟂ , while im(𝜙) ≤ 𝑈 ∗ and hence dim(im(𝜙)) ≤ dim(𝑈). The
inequality in 𝑖) now follows from rank-nullity.
For the second part, note that 𝐵 is non-degenerate on 𝑈 if and only if 𝑈 ∩ 𝑈 ⟂ = {0}. But then the
inequality in 𝑖) shows that we must have 𝑈 ⊕ 𝑈 ⟂ = 𝑉 for dimension reasons. 

II.3 Classification of symmetric bilinear forms

There is a natural linear action of GL(𝑉) on the space Bil(𝑉): if 𝑔 ∈ GL(𝑉) and 𝐵 ∈ Bil(𝑉) then we set
𝑔(𝐵) to be the bilinear form given by
𝑔(𝐵)(𝑣, 𝑤) = 𝐵(𝑔−1 (𝑣), 𝑔−1 (𝑤)), (𝑣, 𝑤 ∈ 𝑉),
where the inverses ensure that the above equation defines a left action. It is clear the action preserves the
subspace of symmetric bilinear forms.
Since we can find a invertible map taking any basis of a vector space to any other basis, the next lemma
says that over an algebraically closed field there is only one nondegenerate symmetric bilinear form up to
the action of GL(𝑉), that is, when k is algebraically closed the nondegenerate symmetric bilinear forms
are a single orbit for the action of GL(𝑉).
Lemma II.21. Let 𝑉 be a k-vector space equipped with a nondegenerate symmetric bilinear form 𝐵. Then if
char(k) ≠ 2, there is an orthonormal basis of 𝑉, i.e a basis {𝑣1 , … , 𝑣𝑛 } of 𝑉 such that 𝐵(𝑣𝑖 , 𝑣𝑗 ) = 𝛿𝑖𝑗 .

Proof. We use induction on dim(𝑉). The identity7

1
𝐵(𝑣, 𝑤) = �𝐵(𝑣 + 𝑤, 𝑣 + 𝑤) − 𝐵(𝑣, 𝑣) − 𝐵(𝑤, 𝑤)�,
2
shows that if 𝐵 ≠ 0 we may find a vector 𝑣 ∈ 𝑉 such that 𝐵(𝑣, 𝑣) ≠ 0. Rescaling by a choice of square root
of 𝐵(𝑣, 𝑣) (which is possible since k is algebraically closed) we may assume that 𝐵(𝑣, 𝑣) = 1. But if 𝐿 = k.𝑣
then since 𝐵|𝐿 is nondegenerate, the previous lemma shows that 𝑉 = 𝐿⊕𝐿⟂ , and if 𝐵 is nondegenerate on
𝑉 it must also be so on 𝐿⟂ . But dim(𝐿⟂ ) = dim(𝑉) − 1, and so 𝐿⟂ has an orthonormal basis {𝑣1 , … , 𝑣𝑛−1 }.
Setting 𝑣 = 𝑣𝑛 , it then follows {𝑣1 , … , 𝑣𝑛 } is an orthonormal basis of 𝑉 as required. 

Remark II.22. Over the real numbers, for example, there is more than one orbit of nondegenerate sym-
metric bilinear form, but the above proof can be modified to give a classification and it turns out that there
are dim(𝑉) + 1 orbits (“Sylvester’s law of inertia”).
One can also classify alternating forms using essentially the same strategy, except that if 𝐵 is a non-
zero alternating form on a vector space 𝑉, one shows that it contains a two-dimensional space 𝐻 on which
𝐵 is nondegenerate. Then we can choose a basis {𝑒, 𝑓} of 𝑉 with 𝐵(𝑒, 𝑓) = 1 = −𝐵(𝑓, 𝑒), and then since
𝑉 = 𝐻 ⊕ 𝐻 ⟂ one can apply induction. Moreover, in the alternating case, the classification holds over any
field k where char(k) ≠ 2.
7
Note that this identity holds unless char(k) = 2. It might be useful to remember this identity when understanding the
Proposition which is the key to the proof of the Cartan Criterion: it claims that if 𝔤 = 𝐷𝔤 then there is an element 𝑥 ∈ 𝔤 with
𝜅(𝑥, 𝑥) ≠ 0. Noting the above identity, we see this is equivalent to asserting that 𝜅 is nonzero.