THE EXPONENTIAL MAP

It is a basic fact that the exponential exp : g −→ G is a local diffeomorphism, and its
general features are taken for granted in what follows. Purpose of this note is to prove that
exp is an analytic diffeomorphism in the nilpotent setting. Along the way we uncover the
intimate structure of solvable Lie algebras and other things.

Semidirect sums
Let Der (g) be the algebra of derivations of the LA g. Suppose there is a vector space-
decomposition g = h ⊕ b with h an ideal in g and b a subalgebra
[h, g] ⊆ h, [b, b] ⊆ b.
If B ∈ b, then ad(B)h ⊆ h, so that ad(B) ∈ Der (h). Thus the adjoint representation
ad : b −→ Der (h)
B 7→ ad(B)
is a LA homomorphism describing the brackets [h, b].
This picture goes under the name of the semidirect sum of b and h and sometimes
indicated h ⊕ad b. Note that ad could be replaced by any other derivation π, as well.
On the other hand, take two LAs g0 , g1 and a morphism θ : g1 −→ Der g0 . Then I can
define g = g0 + g1 by the bracket
[(x0 , x1 ), (y0 , y1 )] = ([x0 , y0 ] + θ(x1 )y0 − θ(y1 )x0 , [x1 , y1 ])
and see that this is the semidirect sum of the ideal g0 and the subalgebra g1 of g.
Remark. Putting θ ≡ 0 yields the notion of ‘direct sum’ of the two ideals g0 , g1 .
Elementary sequences
Let s be a LA. An elementary sequence is a nested sequence of subalgebras of s
s = e0 ⊃ e1 ⊃ . . . ⊃ en−1 ⊃ en = 0
in which ek+1 is an ideal in ek for all k’s and dim ek /ek+1 = 1.
Proposition 0.1. A LA s is solvable ⇐⇒ there exists an elementary sequence.
Proof. If s is a SLA, denote by (sm ) its derived series. Take a codimension one subspace
e1 of s containing s′ = [s, s]. Since [e1 , s] ⊆ s′ ⊆ s, I have that e1 is an ideal in s = e0 .
The proof now assumes the inductive step, that is suppose there are subalgebras e1 , . . . , ej
such that:
i) for all i 6 j, ei is a codimension one ideal in ei−1 ;
ii) for all i 6 j there’s a k = ki (i.e. depending on i) with sk−1 ⊇ ei−1 ⊃ ei ⊇ sk .

1
2

Let ej+1 be a codim. 1 subspace of ej containing the algebra bj , where

 k+1
s for ej = sk
bj = k
s otherwise.
As ej = sk would imply
[ej+1 , ej+1 ] ⊆ [ej+1 , ej ] ⊆ [ej , ej ] = [sk , sk ] = sk+1 = bj ⊆ ej+1 ,
while ej 6= sk would imply
[ej+1 , ej+1 ] ⊆ [ej+1 , ej ] ⊆ [ej , ej ] ⊆ [sk−1 , sk−1 ] = sk = bj ⊆ ej+1 ,
I have that ej+1 is an ideal in ej , and a subalgebra. Eventually, define

kj + 1 for ej = sk
kj+1 = ,
kj otherwise
to the effect that skj+1 −1 ⊇ ej ⊃ ej+1 ⊇ skj+1 . In other words (em ) is an elementary
sequence.

Now assume such a sequence exists. Choose Xi ∈ s so that ei = hXi i ⊕ ei+1 . I claim
that si ⊆ ei , so that sn = 0. In fact (induction)
e0 = s0 ,
if si ⊆ ei then ei+1 = [ei , ei ] ⊆ [hXi i ⊕ ei+1 , hXi i ⊕ ei+1 ] ⊆ [hXi i, ei+1 ] + [ei+1 , ei+1 ] ⊆ ei+1 .
This shows that s is a SLA. 
Upshot: Elementary sequences are the totipotent stem cells of solvable LAs Solvable
Lie algebras are those built by semidirect products starting from the zero Lie algebra and
adding one element at a time. Putting
ei = hXi i ⊕ ei+1 ,
as before, I see that hXi i is line in ei , hence a subalgebra, whilst ei+1 is an ideal. In other
words ei = hXi i ⊕ ei+1 is a semidirect sum (with respect to the obvious ‘ad’ action).

The characterisation of SLAs via elementary sequences is the basic ingredient in proving
the theorem of Lie:
Theorem 0.1. Let s be a SLA over the field K, V an n-dimensional vector space (0 <
n < ∞) on K, ρ a representation of s on V . Then there exists a simultaneous eigenvector
v ∈ V \{0} for all endomorphisms in π(s).
The proof relies on induction and sketchily goes as follows:
1) prove in dim one.
2) Using elementary sequences find a codimension one ideal h. Inductive hypothesis on
it give an e0 with π(H)e0 = l(H)e0 for a l : h → K and any H ∈ h. Pick X ∈ / h and let
ep = π(X)p e0 . The span W of these elements is π(X)-invariant.
3) Claim: π(H)ep = l(H)ep , all H ∈ h.
4) W is π(X)-invariant =⇒ v ∈ W eigenvector for π(X). 
 3

Corollary 0.1. Let s be a SLA over K, V an n-dimensional vector space (0 < n < ∞) on
K, ρ a representation of s on V . Then there exists a complete flag V1 ⊂ V2 ⊂ . . . ⊂ Vn−1
in V , each subspace of which is π(s)-invariant.
Proof. If dim V = 1 it is fairly clear. By the previous theorem I call v an eigenvector
for π(s), and V1 its span. The latter being invariant pushes π to the quotient V /V1 (same
letter), whose dimension is n−1, so there is a complete flag W1 ⊂ W2 ⊂ . . . ⊂ Wn−2 ⊂ V /V1
with π(s)Wi ⊆ Wi , all i 6 n − 1. The inverse image of this flag via the projection
p : V → V ?V1 is a complete π(s)-invariant flag in V . Since p is equivariant, I have
p(π(X)Vi+1 ) = π(X)Wi ⊆ Wi , whence π(X)Vi+1 ⊆ p−1 Wi = Vi+1 . 
Remark. Both theorem and corollary can be stated requiring that V be a K-vector space,
but then one needs to assume that the eigenvalues of all π(X) belong in K.

A LA is called split-solvable if there is an elementary sequence

s = e0 ⊃ e1 ⊃ . . . ⊃ en−1 ⊃ en = 0
where ei is an ideal in s (instead of ei−1 ).
Exercises. a) Are the real SLAs
   
 0 θ x   t 0 x 
g = θ 0 y , g̃ = 0 t y  x, y, t, θ ∈ R
   
0 0 0 0 0 0
split-solvable?
b) A LA is split-solvable iff there exists a complete flag invariant under the adjoint
representation
c) A SLA is split-solvable iff the eigenvalues of all endomorphisms ad(g) belong to K.
Upper-triangular matrices
The LA over K
  
 0 ∗ ... ... ∗ 
 ..  

 .. 

 0
 . . 

n(n, K) := 
 .. .. ..  : ∗ ∈ K ⊂ gl(n, K)

  . . . 

   



 0 ∗ 


0
is the prototype of a NLA (convince yourself this is true!). I claim that ad(X) is a nilpotent
endomorphism for any X ∈ t(n, K).

Proof. Suppose g ⊆ gl(V ) with V of finite, positive dimension. Then

p  
p
X p
ad(X) Y = X p−i Y X i .
i
i=0
4

But
p  
p
X p
(X + Y ) = X p−i Y i if [X, Y ] = 0.
i
i=0
2 = 0, I have ad(E)p = 0 for
Let now Eij = (δij ), i < j be basis vectors of t(n, K). Since Eij
p > 3. But the Eij ’s all commute with each other, so [ad(Eij ), ad(Ekl )] = ad[Eij , Ekl ] = 0.
Induction on the number of non-zero coefficients in that basis shows the claim. 
Exercises. Prove that:
a) g NLA iff adg is NLA.
b) g NLA =⇒ z(g) 6= 0
c) Let ω be the 2-form on R2n defined by
 
0
ω(X, Y ) = X t JY,
−In×n
J= In×n 0 .

Define hn as the LA obtained taking R2n × R with

[(X, t), (Y, s)] = (0, −ω(X, Y )).
This is called (2n + 1)-dimensional (real) Heisenberg algebra. Show that it is nilpotent and
find its centre.
The exponential map
A connected Lie group is said nilpotent/solvable if its LA is nilpotent/solvable. Proposition
0.1 carries over to the LG setup:
Proposition 0.2. Let s be a SLA of dimension n. Then
1) there is a simply connected Lie group S whose LA is s (aka Lie’s 3rd theorem)
2) S is diffeomorphic to Rn
3) there is a sequence of closed simply connected subgroups
S = S0 ⊃ S1 ⊃ . . . ⊃ Sn−1 ⊃ Sn = {1}
in which Si is a semidirect product R ×πi Si+1 . 
Since nil =⇒ sol, a simply connected nilpotent LG is like Euclidean space. The diffeo-
morphisms is actually the exponential map
Theorem 0.2. If N denotes the simply connected nilpotent LG with LA n, the map
exp : n −→ N
is an (analytic) diffeomorphism.
Proof. Let’s show that exp is a regular map. The Baker-Campbell-Hausdorff formula can
be written as follows
∞
X (−1)k ad(X)k
(exp∗ )X = (lexp X ∗ )exp X ◦ , (l = left-translation).
(k + 1)!
|0 {z }
1−e−adX
adX
 5

Since for any X ∈ t the endomorphism adX is nilpotent, in a suitable basis it is represented
by an upper-triangular matrix. Thus det(exp∗ )X 6= 0.
Bijectivity is proven by the usual induction. In dimension one the result is clear, since
N = R and exp = I. Suppose the theorem holds in dimension < n = dim N . Use
proposition 0.2 and let n = n0 ⊃ n1 ⊃ . . . ⊃ nn−1 ⊃ nn = 0 be an elementary sequence
of algebras with dim nk /nk+1 = 1. It is not difficult to prove that ‘nil =⇒ split-solvable’
(cf. exercises), so we can assume that each ni is an ideal in n. The corresponding Lie
subgroups are closed and simply connected. Let Z be the LG of z = nn−1 . Clearly z ⊆ z(g)
and Z ⊆ Z(G). Then Z is a normal subgroup and the projection ϕ : N −→ N/Z is a
homomorphism. Since N is simply connected and Z connected, N/Z is simply connected
too. But it is also nilpotent, its LA is n/z which has dimension < n.
Injectivity. Suppose then exp X = exp X ′ . Using ϕ I deduce expn/z(X+z) = expn/z(X ′ +
z), so by induction X + z = X ′ + z. Therefore X − X ′ ∈ z and it commutes in particular
with X ′ . So
exp X = exp(X ′ + X − X ′ ) = exp X ′ exp(X − X ′ ) =⇒ 1 = exp(X − X ′ ) = expz(X − X ′ ).
But Z is simply connected and one-dimensional, so X = X ′ .
Surjectivity. Let x ∈ N and X +z ∈ n/z such that expn/z(X +z) = ϕx. Call x′ = exp X.
Then
ϕx′ = ϕ(exp X) = expn/z(X + z) = ϕx,
hence x = x′ · z with z ∈ Z. Because Z is connected, Abelian, there is an X ′′ ∈ z with
expz X ′′ = exp X ′′ = z. Now [X, X ′′ ] = 0 implies
x = x′ z = exp X exp X ′′ = exp(X + X ′′ ).

Example. The LG n  o
cos 2t sin 2t x
G1 = − sin 2t cos 2t y :t∈R
0 0 1
is solvable. Take the double covering
n G1 0 0
 o
G= 0 cos t sin t :t∈R .
0 − sin t cos t
The map exp : g → G is not onto, for the element corresponding to t = π, x = 1, y = 0
does not belong to the image of exp.