Jordan decomposition and the recurrent set of flows of automorphisms

Vı́ctor Ayala∗ Adriano Da Silva†

Instituto de Alta Investigación Instituto de Matemática
Universidad de Tarapacá, Arica, Chile Universidade Estadual de Campinas, Brazil
Philippe Jouan
Laboratoire de Mathématiques Raphaël Salem
Université de Rouen, France

August 4, 2020

Abstract
In this paper we show that any linear vector field X on a connected Lie group G admits a Jordan
decomposition and the recurrent set of the associated flow of automorphisms is given as the intersection of
the fixed points of the hyperbolic and nilpotent components of its Jordan decomposition.

Key words: Linear vector fields, recurrent points, Jordan decomposition

2010 Mathematics Subject Classification: 37B20, 54H20, 37B99.

1 Introduction
Many relevant applications are coming from physical problems where the state space is a Lie group. For
instance, the Noether Theorem [15], states that every differentiable symmetry of the action of a physical system
has a corresponding conservation law. And, it is possible to associate symmetry with dynamic through the
notion of invariant vector fields on Lie groups. Furthermore, since the remarkable Brockett’s paper ”Systems
theory on group manifolds and cosset spaces”, in 1972, many people have been working in control theory from
the geometric point of view. Specially, on invariant control systems. On the other hand, in [7] the authors
introduced the class of linear control systems on Lie groups, which is determined by a linear vector field X as
a drift, controled by a number of invariant vector fields. This class of system is a perfect generalization of the
classical linear control systems on Euclidean space, and it is relevant from both, theoretical and practical point
of view (see [10]). Recently, this class have been associated with the almost-Riemannian structures (ARS) (see
[6]). The analysis of any linear system and each ARS on a Lie group, depends strongly of the dynamics of the
drift (see for instance [2, 3, 4, 5, 8]). The main aim of this paper is to give a contribution to understand the
dynamic of X . In order to do that, we first study the recurrent points of X and prove that any linear vector
field admits a mutiplicative Jordan decomposition of its flow, more precisely.
A vector field X on G is said to be linear if its flow {ϕt }t∈R is a 1-parameter subgroup of Aut(G). Associated
to X there is a derivation D : g → g of g that satisfies
(dϕt )e = etD for all t ∈ R.
Through the Lyapunov spectrum of D we consider its well known additive Jordan decomposition into elliptic,
hyperbolic and nilpotent parts. Then, we say X admits a Jordan decomposition if it can be written as a the
sum of commutitive vector fields
X = XE + XH + XN ,
∗ Supported by Proyecto Fondecyt n◦ 1190142. Conicyt, Chile.
† Supported by Fapesp grant no 2018/10696-6

1
where XE , XH and XN are linear vector fields whose associated derivation are the elliptic, hyperbolic and nilpo-
tent parts of D, respectively. Although any linear map admits a Jordan decomposition, the same decomposition
for linear vector fields depends on the integration of automorphisms of g to automorphism of G, which is in
general not always possible. However, if a linear vector field admits a Jordan decomposition and we denote by
{ϕH N
t }t∈R and {ϕt }t∈R the flows of XH and XN , respectively, then

1.1 Theorem: It holds that

R(ϕt ) = fix(ϕH N
t ) ∩ fix(ϕt ).

We build the proof by considering the cases where the toral component T (G) of G is trivial or not. Here T (G)
is the maximal compact, connected subgroup of Z(G)0 . The knowledge of R(ϕt ) and the fact that any linear
vector field on simply connected Lie groups admits a Jordan decomposition allows us to prove that in fact,

1.2 Theorem: Any linear vector field X on a connected Lie group G admits a Jordan decomposition.

In particular, the set of recurrent points of any linear vector field X is completely characterized by the hyperbolic
and nilpotent parts of its Jordan decomposition. Finally, through the so called Arnold’s cat map, we show that
the main result about the recurrent point of a linear vector field is not true for discrete-time flows.
The paper is structured as follows: In Section 2 we introduce all the background needed for the whole paper.
We also prove some results concerning linear flows on vector spaces and some properties of linear vector fields
admiting Jordan decomposition. In Section 3 we prove our main result characterizing the set of recurrent points
of linear vector fields by means of its Jordan decomposition. Section 4 is used to prove that any linear vector
field admits in fact a Jordan decomposition. We finish the paper with an Appendix where we analyze when the
flow of a linear vector field is a flow of isometries for a given almost-Riemannian structure.

Notations
If H ⊂ G is a subgroup, we denote by H0 the connected component of H containing the identity element e ∈ G.
If H = {e} we say that H is a trivial subgroup of G. By Lg and Rg we denote, respectively, the left and
right-translations by g. The conjugation of g is the map Cg := Lg ◦ Rg−1 . The center Z(G) of G is the set of
elements in G that satisfy Cg = idG . If f : G → H is a differentiable map between Lie groups, the differential
of f at x is denoted by (df )x .

2 Preliminaries
Let X be a topological space and {φt }t∈R a flow on X. For any x ∈ X the (positive) orbit of φt is the set

O+ (x, φ) := {φt (x); t ≥ 0} O(x, φ) := {φt (x); t ∈ R}.

The set of fixed points of φt read as

fix(φt ) := {x ∈ X; φt (x) = x, ∀t ∈ R}.

The set of recurrent points of φt is given by

R(φt ) := {x ∈ X; ∃tk → +∞; φtk (x) → x} .

Of course one have that fix(φt ) ⊂ R(φt ). In the next sections we relate the set of fixed and recurrent points
with the Jordan decomposition of flows on vector spaces and more generally on Lie groups.

2
2.1 Dynamics on vector spaces
Let V be a finite real vector space and A : V → V a linear map. Recall that A is semisimple if its extension
to the complexification VC is diagonalizable. We say that A is nilpotent if An ≡ 0 for some n ∈ N and A
is elliptic (hyperbolic) if it is semisimple and its eigenvalues are pure imaginary (real). The (additive) Jordan
decomposition of A reads as

A = AE + AH + AN , where A, AE , AH and AN commute,

AE is elliptic, AH is hyperbolic and AN is nilpotent.

For any given eigenvalue λ of AH let Vλ := {v ∈ V ; AH v = λv} be its associated eigenspace. Define the
A, AE , AH , AN -invariant vector subspaces of V
M M
V+ = Vλ , V 0 = ker AH and V − = Vλ .
λ>0 λ<0

Then V = V + ⊕ V 0 ⊕ V − and for any norm on V there exists λ, t0 > 0 such that

tA −tλ
e|V + ≥ etλ and etA
|V − ≤ e , for t ≥ t0 . (1)

Moreover, there is an inner product h·, ·i in V such that etAE is an isometry for any t ∈ R.
We will use the notation V +,0 and V −,0 for the subspaces V + ⊕ V 0 and V − ⊕ V 0 , respectively.
The next lemma will be important in the context of linear vector fields ahead.

2.1 Lemma: Let γ : R → V be a continuous curve that satisfies

γt+s = γt + etA γs , ∀t, s ∈ R.

If A has no eliptical part and (γtk )k∈N is bounded for some sequence tk → ±∞, then γt ∈ V ∓ for all t ∈ R. In
particular, (γt )t≥0 is bounded.

Proof: By analogy it is only necessary to show that if (γtk )k∈N is bounded for some sequence tk → +∞ then
γt ∈ V − . Moreover, it is enough to consider the cases where A is nilpotent or where it has only eigenvalues
with positive real parts.
In fact, if W, U ⊂ V are A-invariant subspaces such that V = W ⊕ U and assume w.l.o.g. that U = W ⊥ . If
π : V → W is the orthogonal projection onto W and we define γtW := π(γt ), the A-invariance of W and U
implies that
W
γt+s = π(γt+s ) = π(γt + etA γs ) = π(γt ) + π(etA γs ) = π(γt ) + etA π(γs ) = γtW + etA γsW .

Moreover, the fact that |π(v)| ≤ |v| for any v ∈ V gives us that

(γtk )k∈N is bounded =⇒ (γtW )

k k∈N
is bounded.

In particular, the result follows if we show that

(γtW )
k k∈N
is bounded =⇒ γ W ≡ 0 for W = V + and W = V 0 ,

that is, we only have to analyze the cases where A is nilpotent or it has only eigenvalues with positive real parts.
On the other hand, for any k ∈ N and r ∈ (0, 1] we can write tk = rnk + rk with nk ∈ N and rk ∈ [0, 1). Then

|γrnk | = |γtk −rk | = |γ−rk + e−rk A γtk | ≤ |γ−rk | + ke−rk A k|γtk |

and
|γtk | = |γrnk +rk | = |γrk + erk A γrnk | ≤ |γrk | + kerk A k|γrnk |.

3
Since by continuity,
{γt , t ∈ [−1, 1]} and {ketA k, t ∈ [−1, 1]} are bounded,
then
(γtk )k∈N is bounded ⇐⇒ (γrnk )k∈N is bounded,
and so we can assume w.l.o.g. that tk = rnk with nk ∈ N and r ∈ (0, 1]. Moreover, since for any t ∈ R there
exists n ∈ Z and r ∈ (0, 1] such that t = nr, then
n
X
γr = 0 ∀r ∈ (0, 1] =⇒ γt = γnr = ejrA γr = 0 ∀t ∈ R
j=0

and we only have to show that if A is nilpotent or it has only eigenvalues with positive real parts then (γrnk )k∈N
bounded for some sequence nk → +∞ implies γr = 0.

Case 1: A has only eigenvalues with positive real parts;

Since,
nk
X
(1 − erA )γrnk = (1 − erA ) ejrA γr = (1 − e(nk +1)rA )γr ,
j=0

and 1 − erA is invertible, it holds that

(γrnk )k∈N is bounded ⇐⇒ ((1 − e(nk +1)rA )γr )k∈N is bounded.

However,    
(1 − e(nk +1)rA )γr  ≥ e(nk +1)rA γr  − |γr | ≥ e(nk +1)λ |γr | − |γr | ,
   

and hence,
(γrnk )k∈N is bounded ⇐⇒ γr = 0.

Case 2: A is nilpotent;
Let p ∈ N be the greatest integer such that Ap γr 6= 0. Then
   
nk nk  p nk nk
rp

X X (jr) X X
γrnk = ejrA γr = γr + jrAγr + · · · + Ap γr = (nk +1)γr + j  rAγr +· · ·+ j p  Ap γ r .
j=0 j=0
p! j=1 j=1
p!

If p ≥ 1, a simple calculation shows that

 −1  −1  
Xnk nk
X Xnk
lim  j p  (nk + 1) = lim  jp  j q  = 0, if 1 ≤ q ≤ p − 1
k→+∞ k→+∞
j=1 j=1 j=1

and hence, there exists k0 ∈ N such that

 
nk
1 X rp p rp
|γrnk | ≥ jp |A γr | ≥ nk |Ap γr | , k ≥ k0 .
2 j=1 p! 2(p!)

Consequently,
(γrnk )k∈N bounded =⇒ Ap γr = 0, ∀p ≥ 1.
However, Ap γr = 0 for all p ≥ 1 implies necessarily that γrnk = (nk + 1)γr which is bounded if and only if
γr = 0.

4
By the previous cases, γt ∈ V − for all t > 0. By considering 0 < t < s we have that

γ−t+s = γ−t + e−tA γs =⇒ γ−t = γ−t+s − e−tA γs ∈ V − + e−tA V − ⊂ V − ,

therefore γt ∈ V − for all t ∈ R.

For the last assertion, let t ≥ t1 and consider pt ∈ N and rt ∈ [0, 1) such that t = pt t1 + rt . Then,
 
Xpt
γt = γpt t1 +rt = γrt + ert A  ejt1 γt1  ,
j=0

and so, by equation (1), we obtain

 
pt
X 1 − e−(pt +1)t1 λ
|γt | ≤ |γrt | + e−rt λ  e−jt1 λ  |γt1 | = |γrt | + e−rt λ |γt1 | .
j=0
1 − e−t1 λ

Since  
e−rt λ 1 − e−(pt +1)t1 λ ≤ 1 and γ is continuous,

the assertion follows. 

We finish this section with some remarks concerning recurrent sets, fixed points and Jordan decomposition of
the matrix exponential.

2.2 Remark: A slight modification of Proposition 4.2 of [16] gives us that the recurrent set of etA are related
with the fixed points of etAH , etAN and etAH,N by

R(etA ) = fix(etAH ) ∩ fix(etAN ) = fix(etAH,N ),

where AH,N := AH + AN .

2.3 Remark: If V is a finite vector space and A : V → V is a linear map, the commutator of A is defined by

ad(A) : gl(V ) → gl(V ), ad(A)B := [A, B] = BA − AB.

Note that [ad(A), ad(B)] = ad([A, B]) implying that ad(A) and ad(B) commutes when A and B commutes.
Also, a simple calculation shows that
k  
X k
ad(A)k (B) = (−1)k Ak−j BAj
j
j=0

and hence, Ak = 0 implies ad(A)2k−1 = 0. Moreover, if {v1 , . . . , vm } is a basis of V such that Avi = λi vi ,
i = 1, . . . , m and Ei,j ∈ gl(V ) satisfies

vi if k = j
Ei,j vk = =⇒ ad(A)Ei,j = (λi − λj )Ei,j ,
6 j.
0 if k =

it follows that, A diagonalizable implies ad(A) diagonalizable. Consequently, if A = AE + AH + AN is the

Jordan decomposition of A then ad(A) = ad(AE ) + ad(AH ) + ad(AN ) is the Jordan decomposition of ad(A).

5
2.2 Linear vector fields
In this section we introduce the notion of linear vector fields and its main properties. Let G be a connected Lie
group with Lie algebra g identified with the set of left invariant vector fields.
A vector field X on G is said to be linear if its flow {ϕt }t∈R is a 1-parameter subgroup of Aut(G). Associated
to any linear vector field X there is a derivation D of g that satisfies

(dϕt )e = etD for all t ∈ R.

In particular, it holds that

ϕt (exp Y ) = exp(etD Y ), for all t ∈ R, Y ∈ g.

For any connected Lie group G we denote by T (G) the toral component of G, that is, the maximal compact,
connected subgroup of Z(G)0 . As a standard fact we know that if T (G) is trivial, then Z(G)0 is simply connected
and that T (G/T (G)) is trivial (see [17, Proposition 3.3]).
Let {ϕt }t∈R ⊂ Aut(G) be a flow. By maximality, for any t ∈ R we get that ϕt (T (G)) = T (G). However, the
group of the automorphisms of T (G) is discrete and hence ϕt |T (G) = idT (G) . In particular, it follows that

T (G) ⊂ fix(ϕt ).

Let D = DE + DH + DN be the Jordan decomposition of D. By Theorem 3.2 and Proposition 3.3 of [19] we have
that DE , DH and DN are still derivations of g. Moreover, if gλ stands for the eigenspaces of the hyperbolic
part DH , then [gλ , gµ ] ⊂ gλ+µ if λ + µ is an eigenvalue of DH and zero otherwise (see [19] Proposition 3.1). It
turns out that M M
g+ = gλ , and g− = gλ
λ>0 λ<0

are nilpotent Lie subalgebras. Denoting by g = ker DH we obtain the decomposition g = g+ ⊕ g0 ⊕ g− .

0

Now we extend the Jordan decomposition from a linear differential equation to linear vector fields. Let X be a
linear vector field on a connected Lie group G. We say X is elliptic, hyperbolic or nilpotent, respectively, when
its associated derivation D is elliptic, hyperbolic or nilpotent. The Jordan decomposition of X is given by

X = XE + XH + XN , where X , XE , XH , XN commute,

with XE elliptic, XH hyperbolic and XN nilpotent. By the uniqueness of the Jordan decomposition of D and the
connectedness of the group G we get that the Jordan decomposition of X is unique, when such decomposition
exists.
Let G be a connected Lie group and X a linear vector field with associated flow {ϕt }t∈R . If X admits Jordan
decomposition X = XE + XH + XN then

for all t ∈ R, ϕt is a commutative product, ϕt = ϕEt ◦ ϕH N

t ◦ ϕt ,

where {ϕit }t∈R is the flow of Xi for i = E, H, N . Moreover, for any i, j ∈ {E, H, N } with i 6= j we have that
ϕi,j i j j i
t := ϕt ◦ ϕt = ϕt ◦ ϕt is the flow of the linear vector field Xi,j := Xi + Xj whose associated derivation is
Di,j := Di + Dj .

2.4 Remark: For any given linear vector field X on G and any surjective homomorphism π : G → H the
family {ψt }t∈R defined by the relation
π ◦ ϕt = ψt ◦ π, t ∈ R
is a one-parameter group of automorphism of H if and only if ker π is a ϕ-invariant subgroup. In particular,
if Y stands for the linear vector field associated with {ψt }t∈R then Y is elliptic, hyperbolic or nilpotent if X is
elliptic, hyperbolic or nilpotent, respectively.

6
We define the dynamical subgroups of G associated with the hyperbolical part of X by

G0 = fix(ϕH + + − −
t ), G = exp(g ) and G = exp(g ).

These subgroups where first defined in [8] without the use of Jordan decompositions. However, it is straight-
forward to see that their definitions are equivalent. The next proposition states the main properties of these
dynamical subgroups. Its proof can be found at [8, Proposition 2.9].

2.5 Proposition: It holds:

1. G0 normalizes G+ and G− and hence G+,0 := G+ G0 and G−,0 := G− G0 are subgroups of G;

2. G− ∩ G0 = G+ ∩ G0 = G−,0 ∩ G+ = G+,0 ∩ G− = {e};

3. The dynamical subgroups are closed in G;
4. If G is solvable, then G0 is connected and G = G− G+,0 . Moreover, fix(ϕt ) ⊂ G0 .

We will say that G is decomposable if G = G− G+,0 . In particular, if G is decomposable and x ∈ G there exists
unique a ∈ G− , b ∈ G+ and c ∈ G0 such that x = abc.
A Levi subgroup (subalgebra) S ⊂ G (s ⊂ g) is a maximal connected semisimple subgroup (subalgebra). By
Levi’s Theorem, Levi subgroups (subalgebras) always exists and we have the Levi decomposition (see [18,
Chapter I-4])
G = SR, with dim(S ∩ R) = 0,
where R is the solvable radical of G.

2.6 Proposition: There exists a ϕH -invariant Levi subgroup S ⊂ G such that

fix(ϕH H H
t ) = fix(ϕt |S ) fix(ϕt |R ).

Proof: Since {etDH , t ∈ R} is a group of semisimple automorphisms of g, Corollary 5.2 of [14] implies the
existence of a DH -invariant Levi subalgebra. Hence, the connected subgroup S ⊂ G with Lie algebra s is a Levi
subgroup and the DH -invariance of s implies the ϕH -invariance of S.
For the equality of the sets of fixed points, let us show the inclusion fix(ϕH H H
t ) ⊂ fix(ϕt |S ) fix(ϕt |R ) since the
H
opposite one is trivial. Let x ∈ fix(ϕt ) and write it as x = ab with a ∈ S, b ∈ R. Then,

ab = x = ϕt (x) = ϕH H
t (a)ϕt (b) =⇒ S 3 ϕH (a−1 )a = ϕH
t (b)b
−1
∈ R.

Since dim(R ∩ S) = 0 and t 7→ ϕHt (a

−1
)a is a continuous curve in R ∩ S, we must necessarily have that
ϕt (a )a = ϕ0 (a )a = e implying that a ∈ fix(ϕH
H −1 H −1 H
t |S ) and also that b ∈ fix(ϕt |R ) as desired. 

Let ∆ ⊂ g be a subspace. A differentiable curve θ : [0, T ] → G is said to be admissible for ∆ if

θ̇(s) ∈ (dLθ(s) )e ∆, almost everywhere.

The next result implies that we can endow G with a left-invariant Riemannian metric such that {ϕEt }t∈R is a
flow of isometries.

2.7 Theorem: Let ∆ ⊂ g be a D-invariant subspace. If D|∆ is skew-symmetric then {ϕt }t∈R preserves the
arc-lenght of any admissible curve for ∆. In particular, {ϕEt }t∈R is a flow of isometries for some left-invariant
metric on G.

7
Proof: Let us denote by h., .i the left-invariant metric and by l(t) the length of the curve s 7−→ ϕt (θ(s)).
Thanks to the left-invariance on the one hand, and to the fact that (dϕt )e = etD is orthogonal on ∆ for all t on
the other one, we get:
Z T  12
d d
l(t) = ϕt (θ(s)), ϕt (θ(s)) ds
0 ds ds ϕt (θ(s))
Z T  12
d d
= (dϕt )θ(s) θ(s), (dϕt )θ(s) θ(s) ds
0 ds ds ϕt (θ(s))
Z T  12
d d
= (dLϕet (θ(s)) )e (dϕt )e (dLθ(s)−1 )θ(s) θ(s), (dLϕet (θ(s)) )e (dϕt )e (dLθ(s)−1 )θ(s) θ(s) ds
0 ds ds ϕt (θ(s))
Z T  21
d d
= θ(s), θ(s) ds = l(0),
0 ds ds θ(s)

where for the third equality we used that ϕt ◦ Lg = Lϕt (g) ◦ ϕt for any g ∈ G and t ∈ R.
The assertion on {ϕEt }t∈R follows from the fact that there exists an inner product h·, ·i on g such that {etDE }t∈R
is a flow of isometries, or equivalently, DE is skew-symmetric. By extending h·, ·i to a left-invariant metric on
G we have that {ϕEt } preserves the arc-lenght of any curve in G and the result follows. 

2.8 Proposition: It holds:

1. The (positive) orbit of ϕt at x ∈ G is bounded iff the (positive) orbit of ϕH,N

t at x ∈ G is bounded;
2. If x ∈ R(ϕt ) then cl(O(x, ϕ)) = cl(O+ (x, ϕ)).

Proof: 1. Let us denote by % the distance associated with the left-invariant metric such that {ϕEt } is a flow of
isometries. Then,
for all t ∈ R, %(ϕH,N
t (x), e) = %(ϕEt (ϕH,N
t (x)), ϕEt (e)) = %(ϕt (x), e),
which implies the result.
2. Since O+ (x, ϕ) ⊂ O(x, ϕ) it is enough to show that ϕ−t (x) ∈ cl(O+ (x, ϕ)) for any t > 0. However, since
x ∈ R(ϕt ) there exists tk → +∞ such that ϕtk (x) → x as k → +∞ and so
ϕtk −t (x) = ϕ−t (ϕtk (x)) → ϕ−t (x), k → +∞.
Since tk → +∞, there exists k0 ∈ N such that tk − t > 0 for any k ≥ k0 and so ϕtk −t (x) ∈ O+ (x, ϕ) implying
that ϕ−t (x) ∈ cl(O+ (x, ϕ)) as stated. 

3 Recurrent points of linear vector fields

Our aim in this section is to prove the following fact: If a linear vector field admits a Jordan decomposition, the
set of recurrent points coincides with the intersection of the sets of fixed points of the hyperbolic and nilpotent
parts of the mentioned decomposition.
Let us fix a linear vector field X and assume through the whole section that X admits a Jordan decomposition.
We start by showing the result in question on the adjoint group.

3.1 Proposition: Let X be a linear vector field on G with associated derivation D. Then,
Ad(ϕt (g)) = et ad(D) Ad(g), g ∈ G, t ∈ R. (2)
In particular,  
R (Ad ◦ϕt ) = fix Ad ◦ϕH,N = fix Ad ◦ϕH ∩ fix Ad ◦ϕN



t t t ,
where Ad ◦ϕt (Ad(g)) := Ad(ϕt (g)).

8
Proof: Since G is connected, we only need to show equation (2) for g = eX with X ∈ g. However,
tD tD
ad(X)e−tD
Ad(ϕt (exp(X))) = Ad(exp(etD X)) = ead(e X)
= ee = etD ead(X) e−tD = CetD Ad(exp(X)).

Furthermore, since CetD : gl(g) → gl(g) is linear, it holds that

CetD = Ad(etD ) = et ad(D) =⇒ Ad(ϕt (exp(X))) = et ad(D) Ad(exp(X)),

which proves equation (2).

By equation (2) it turns out that Ad ◦ϕt = etD |Ad(G) implying
 
R (Ad ◦ϕt ) ⊂ R et ad(D) ∩ Ad(G).

On the other hand, ad(D) is a linear map of gl(g) with Jordan decomposition ad(D) = ad(DE )+ad(DH )+ad(DN )
which by Remarks 2.2 and 2.3 implies that
       
R et ad(D) = fix et ad(DH,N ) = fix et ad(DH ) ∩ fix ead(DN ) .

Since    
fix et ad(DH,N ) ∩ Ad(G) = fix Ad ◦ϕH,N
t and

fix etDH ∩ fix etDN ∩ Ad(G) = fix Ad ◦ϕH ∩ fix Ad ◦ϕN

t t ,
we get that  
R (Ad ◦ϕt ) ⊂ fix Ad ◦ϕH,N = fix Ad ◦ϕH ∩ fix Ad ◦ϕN



t t t ⊂ R (Ad ◦ϕt ) ,

where the last inclusion follows from the fact that by Theorem 2.7, Ad ◦ϕE = et ad(DE ) is an isometry, for any
t ∈ R. 

Furthermore, it is possible to prove something more

3.2 Proposition: With the previous notations, it holds that

fix(ϕH,N
t ) = fix(ϕH N
t ) ∩ fix(ϕt ).

Proof: Our proof consider the different classes of Lie groups as follows:
• Assume first that G is a semisimple Lie group;
By Proposition 3.1 it follows that

x ∈ fix(ϕH,N ) =⇒ ∀t ∈ R, x−1 ϕit (x) ∈ Z(G), i = H, N .

Since the center of a semisimple Lie group is discrete we must have x−1 ϕit = e for all t ∈ R. Therefore,

fix(ϕH,N ) ⊂ fix(ϕH ) ∩ fix(ϕN ),

which implies the equality.

• Let us assume now that G is a solvable Lie group;
Since the corresponding dynamical subgroups of ϕH,N and ϕH coincides, we have by Proposition 2.5 item 4.
that
fix(ϕH,N ) ⊂ G0 = fix(ϕH ) =⇒ fix(ϕH,N ) ⊂ fix(ϕH ) ∩ fix(ϕN ).

• Finally, let G be an arbitrary Lie group.

9
Let S be a ϕH -invariant Levi subgroup given by Proposition 2.6 and consider π : G → G/R be the canonical
projection. Denote by ϕ̄H,N , ϕ̄H and ϕ̄H the flows induced by ϕH,N , ϕH and ϕN on G/R, respectively. Since
G/R is semisimple, we obtain that

π(fix(ϕH,N
t )) ⊂ fix(ϕ̄H,N
t ) = fix(ϕ̄H N
t ) ∩ fix(ϕ̄t ).

Let then x ∈ fix(ϕH,N

t ) and write it as x = ab with a ∈ S and b ∈ R. By the ϕH -invariance of S and R we get

ϕH H H H
t (x) = ϕt (ab) = ϕt (a)ϕt (b) =⇒ π(ϕH H
t (x)) = π(ϕ (a))

and hence
π(a) = π(x) = ϕ̄H H H
t (π(x)) = π(ϕt (x)) = π(ϕt (a)) =⇒ S 3 a−1 ϕH
t (a) ∈ R.

Since dim(R ∩ S) = 0 and t 7→ a−1 ϕH H

t (a) is a continuous curve, it follows that a ∈ fix(ϕt ). On the other hand,

fix(ϕH,N )0 = fix(ϕH H
=⇒ x−1 ϕH H
=⇒ b−1 ϕH H 0


t t ) ∩ fix(ϕt ) 0 t (x) ∈ fix(ϕt ) t (b) ∈ fix(ϕt ) ∩ R ⊂ R .

Let us write ϕH 0
t (b) = bct with ct ∈ R . Since b ∈ R and R is solvable, Proposition 2.5 implies the existence of
− + 0
unique p ∈ R , q ∈ R and r ∈ R such that b = pqr. Hence,

ϕH
t (pqr) = pqrct =⇒ ϕH H H
t (p) = p, ϕt (q) = q, ϕt (r) = rct .

However, fix(ϕH 0
t |R ) = R implying that p = q = e and consequently that b = r. In particular, we get that
0 H
b ∈ R = fix(ϕt |R ) and by Proposition 2.6 we conclude that

x = ab ∈ fix(ϕH H H
t |S ) fix(ϕt |R ) = fix(ϕt ).

Since fix(ϕH,N
t ) ∩ fix(ϕH N H N
t ) = fix(ϕt ) we get that x ∈ fix(ϕ ) ∩ fix(ϕt ) implying the equality

fix(ϕH,N
t ) = fix(ϕH ) ∩ fix(ϕN
t ),

and concluding the proof. 

Next, we prove our main result concerning recurrent points. We build the proof by considering the cases where
toral component of G is trivial or not.

3.3 Theorem: With the previous notations, it holds that

R(ϕt ) = fix(ϕH,N
t ) = fix(ϕH N
t ) ∩ fix(ϕt ).

Proof: By Proposition 3.2 we just need to show that

R(ϕt ) = fix(ϕH,N
t ).

Let us consider x ∈ fix(ϕtH,N ) and let % be the left-invariant Riemannian metric for such that {ϕEt }t∈R is a flow
of isometries. Therefore,

%(ϕt (x), e) = %(ϕEt (x), e) = %(ϕEt (x), ϕEt (e)) = %(x, e), ∀t ∈ R.

Thus, there exists tk → +∞ such that (ϕEtk (x))k∈N is convergent. In particular, for any ε > 0 there exists
k0 ∈ N such that tk > tk0 implies

%(ϕEtk (x), ϕEtk (x)) < ε =⇒ %(ϕtk −tk0 (x), x) = %(ϕEtk −tk (x), x) = %(ϕEtk (x), ϕEtk (x)) < ε,
0 0 0

showing that x ∈ R(ϕt ) and hence fix(ϕtH,N ) ⊂ R(ϕt ).

Let x ∈ R(ϕt ). Since Ad(R(ϕt )) ⊂ R(Ad ◦ϕt ), Proposition 3.1 shows that

∀t ∈ R, Ad(ϕH,N
t (x)) = Ad(x) or equivalently ∀t ∈ R, x−1 ϕH,N
t (x) ∈ Z(G).

10
Let us define the curve
ζ : R → Z(G), t ∈ R 7→ ζt := x−1 ϕH,N
t (x).
In particular,
x ∈ fix(ϕH,N
t ) ⇐⇒ ζ ≡ e.
Moreover, since ζ is continuous and ζ0 = e it follows that ζt ∈ Z(G)0 for all t ∈ R and we can divide our analysis
in two cases:
Case 1: T (G) is trivial;
Here, Z(G)0 is simply connected and the exponential map exp : z(g) → Z(G)0 is a diffeomorphism. Therefore,
the curve,
γ : R → z(g) given by the formula ζt = exp γt ,
is well defined. Moreover, for any t, s ∈ R obtain
H,N
 −1 H,N 

ζt+s = x−1 ϕt+s (x) = x−1 ϕH,N
t ϕ H,N
s (x) = x ϕt (x) ϕH,N
t x −1 H,N
ϕs (x) = ζt ϕH,N
t (ζs )

which implies

exp γt+s = exp γt ϕtH,N (exp γs ) = exp γt exp etDH,N γs = exp γt + etDH,N γs .

Therefore,
∀t, s ∈ R, γt+s = γt + etDH,N γs .
On the other hand, there exists tk → +∞ such that

ϕtk (x) → x =⇒ (ϕtH,N

k
(x))k∈N is bounded =⇒ (ζtk )k∈N is bounded =⇒ (γtk )k∈N is bounded

where the first implication follows from the fact that ϕEt is an isometry for every t ∈ R. Since DH,N has no
eliptical part, Lemma 2.1 implies that γt ∈ g− for all t ∈ R and that (γt )t≥0 is bounded. Then,

ζt = exp γt ∈ G− , ∀t ∈ R and (ζt )t≥0 is bounded.

Since ζt = x−1 ϕH,N

t (x) we conclude that O+ (x, ϕH,N ) is also bounded. By Proposition 2.8 and the fact that
x ∈ R(ϕt ) we conclude that

cl(O+ (x, ϕH,N )) is bounded =⇒ cl(O+ (x, ϕ)) = cl(O(x, ϕ)) is bounded =⇒ cl(O(x, ϕH,N )) is bounded.

Therefore, (ζt )t∈R is bounded and by applying Lemma 2.1 again we concluded that ζt ∈ G+ for all t ∈ R.
Therefore,
∀t ∈ R, ζt ∈ G+ ∩ G− = {e} =⇒ x ∈ fix(ϕH,Nt ).

Case 2: T (G) is nontrivial.

Since G/T (G) has trivial toral component, we get by the previous case that

R(ϕt ) ⊂ fix(ϕH,N
t )T (G).

However, T (G) ⊂ fix(ϕH,N ) and hence R(ϕt ) ⊂ fix(ϕH,N

t ), finishing the proof. 

As a direct consequence we have the following

3.4 Corollary: With the previous notations, it holds that

fix(ϕt ) = fix(ϕEt ) ∩ fix(ϕH N

t ) ∩ fix(ϕt ).

11
Proof: Since any fixed point is recurrent, Theorem 3.3 implies that fix(ϕt ) ⊂ R(ϕt ) = fix(ϕH N
t ) ∩ fix(ϕt ).
E H,N E
Therefore, if x ∈ fix(ϕt ) we get x = ϕt (x) = ϕt (ϕt (x)) = ϕt (x) implying that
fix(ϕt ) ⊂ fix(ϕEt ) ∩ fix(ϕH N
t ) ∩ fix(ϕt ).

Since the opposite inclusion is trivial, the result follows. 

3.5 Remark: It is important to remark that Theorem 4.1 was proved in [16, Lemma 4.3] in the context of
automorphisms on simply connected connected Lie groups.

The next example show that Theorem 3.3 does not holds for discrete-time flows.

3.6 Example: The 2-dimensional torus T2 = R2 /Z2 is an abelian Lie group whose Lie algebra is R2 . Let
  √  
ln 3+2 5 0
 
2√ 2√
A=   √   and P =
0 ln 3−2 5 1+ 5 1− 5

and consider the hyperbolic derivation D = P AP −1 . Since tr D = 0 we have that

 
1 1
eD = ,
1 2
induces the automorphism ϕ ∈ Aut(T2 ) given by
ϕ ([x, y]) = [x + y, x + 2y], where [x, y] := (x, y) + Z2 .
Such automorphism, known as Arnold’s cat map is the recurrent in the torus and hence R(ϕn ) = T2 (see [1,
Example 1.16]), that is R(ψn ) = T2 , where ϕn is the discrete-time flow
ϕn := ϕ ◦ · · · ◦ ϕ .
| {z }
n−times

On the other hand, a simple calculation shows that fix(ϕn ) = {[0, 0]}.

4 Jordan decomposition
As a relevant consequence of the main result in Theorem 3.3, in this section we prove that any linear vector field
X admits a Jordan decomposition. In particular, Theorem 3.3 holds for any linear vector field on connected
groups.

4.1 Theorem: Any linear vector field X on a connected Lie group G admits a Jordan decomposition.

Proof: Let us first assume that G is simply connected and let D = DE + DH + DN be the Jordan decomposition
of D. Due to the extra topological assumption on G, Theorem 2.2 of [7] assures the existence of linear vector
fields Xi associated with the derivations Di for i = E, H, N . By definition, XE is elliptic, XH is hyperbolic and
XN is nilpotent.

1. Let us start by showing that X = X E + X H + X N . If {ϕEt }t∈R , {ϕH N

t }t∈R and {ϕt }t∈R are the flows of
XE , XH and XN , respectively, it follows that
ϕt (exp X) = exp etD X = exp(et(DE +DH +DN ) X) = exp(etDE etDH etDN X)
= ϕEt (exp(etDH etDN X)) = ϕEt (ϕH
t (exp(e
tDN
X))) = ϕEt (ϕH N
t (ϕt (exp X))).
By the connectedness of G we conclude that
ϕt = ϕEt ◦ ϕH N
t ◦ ϕt , t ∈ R.

Finally, by derivation we get that X = XE + XH + XN as stated.

12
 2. Next, we prove that the vector fields X , XE , XH and XN commutes. We just show that [X , XE ] = 0, since
the other cases are analogous. However, [X , XE ] = 0 is equivalent to ϕs ◦ ϕEt = ϕEt ◦ ϕs , t, s ∈ R, which
follows from the commutativeness of D and DE . In fact, for any X ∈ g it holds that

ϕs (ϕEt (exp X)) = ϕs (exp(etDE )) = exp(esD etDE X)

= exp(etDE esD X) = ϕEt (exp(esD )) = ϕEt (ϕs (exp X)).

The connectedness of G implies ϕs ◦ ϕEt = ϕEt ◦ ϕs , for any t, s ∈ R, as claimed.

By 1. and 2. we obtain that any linear vector field on a connected simply connected Lie group admits Jordan
decomposition. Let G be a connected Lie group and denote by G e the connected, simply connected covering of
G. If D is the derivation associated with X , Theorem 2.2 of [7] assures the existence of a linear vector field Xe
e with associated derivation D. Again, the connectedness of G
on G e gives us

for all t ∈ R, π ◦ ϕ
et = ϕt ◦ π, (3)

where π : G e → G = G/D
e is the canonical projection, D ⊂ Z(G) is a discrete subgroup and {ϕt }t∈R , {ϕ et }t∈R
denote the flows of X , Xe, respectively. Equation (3) implies in particular that ϕ
et (ker π) = ker π and since ker π
is discrete we obtain by continuity that ker π ⊂ fix(ϕ et ).
eit }t∈R
By the simply connected case, Xe admits a Jordan decomposition Xe = XeE + XeH + XeN . If we denote by {ϕ
the flow of Xei for i = E, H, N , we get by Corollary 3.4 that

fix(ϕ eEt ) ∩ fix(ϕ

et ) = fix(ϕ eH eN
t ) ∩ fix(ϕt )

eit ) for i = E, H, N . In particular,

and hence ker π ⊂ fix(ϕ

ϕit : G → G, defined by the relation π ◦ ϕ

eit = ϕit ◦ π, t ∈ R, i = E, H, N

is a one-parameter subgroup of automorphisms of G and its associated vector field Xi is elliptic, hyperbolic or
nilpotent if Xei is elliptic, hyperbolic or nilpotent (see Remark 2.4). Therefore, X = XE + XH + XN is the Jordan
decomposition of X proving the result for the connected case and finishing the proof. 

As a consequence of Theorems 3.3 and 4.1 we get the following result.

4.2 Theorem: The recurrent set of the flow of a linear vector field X on a connected Lie group G is given as
the intersection of the fixed points of the hyperbolic and nilpotent components of its Jordan decomposition.

4.3 Remark: It is important to remark that Theorem 3.3 was proved first in [16, Theorem 4.4] in the context
of automorphisms on simply connected, connected nilpontent Lie groups.
 
0 −1
4.4 Example: Let θ = and define ρt := etθ . Consider the semi-direct product G = R ×ρ R2 .
1 0
Following [2] a linear vector field of G and its associated derivation are given, respectively, by
 
0 0
X (t, v) = (0, Av + Λt ξ), and D = ,
ξ A
 
λ −µ
where ξ ∈ R2 , θ = and Λt = (ρt − 1)θ−1 . If λ2 + µ2 6= 0, the the eliptic, hyperbolic and nilpotent
µ λ
parts of A reads, respectively, as
   
0 −µ λ 0
AE = , AH = and AN = 0.
µ 0 0 λ
Consequently,    
0 0 0 0
DE = , DH = and DN = 0.
AE A−1 ξ AE AH A−1 ξ AH

13
are the corresponding eliptic, hyperbolic and nilpotent parts of D. Hence,

XE (t, v) = (0, AE v + Λt AE A−1 ξ), XH (t, v) = (0, AH v + Λt AH A−1 ξ) and XN = 0.

Since XN = 0 we get
R(ϕt ) = fix(ϕH
t ) = {(t, v) ∈ G; XH (t, v) = (0, 0)}.

However, a simple calculation shows us that

{(t, v) ∈ G; XH (t, v) = (0, 0)} = {(t, v) ∈ G; X (t, v) = (0, 0)},

implying that
R(ϕt ) = fix(ϕt ) = {(t, v) ∈ G; Av + Λt ξ = 0}.
In particular, if ξ = 0 we get R(ϕt ) = R × {0}.

A One-parameter groups of isometries of left-invariant sub-

Riemannian metrics
In what follows G is a n-dimensional connected Lie group and g stands for its Lie algebra. The group is
endowed with a sub-Riemannian, possibly Riemannian, left-invariant structure defined by a set {Y1 , Y2 , . . . , Yp }
of left-invariant vector-fields (with p ≤ n). More accurately the p-uple (Y1 , Y2 , . . . , Yp ) is an orthonormal basis
of the subspace ∆ it generates in g and defines a left-invariant sub-Riemannian structure on G. This structure
is Riemannian if p = n. It is sub-Riemannian if p < n and we assume in that case that the distribution
{Y1 , Y2 , . . . , Yp } is bracket generating. On the other hand let X be a linear field on G. As previously we denote
by D = − ad(X ) the associated derivation of g and by {ϕt }t∈R the flow of X . On the context of sub-Riemannian
geometry, Theorem 2.7 reads:

A.1 Theorem: If ∆ is D-invariant and D|∆ is skew-symmetric then {ϕt }t∈R is a one-parameter group of
isometries of G.

Conversely let {ϕt }t∈R be a one-parameter group of isometries of a left-invariant structure on G. Let us assume
that ϕt is also an automorphism of G, for each t ∈ R. In that case the one-parameter group {ϕt }t∈R is generated
by an infinitesimal automorphism, that is a linear field X , and to this linear field we can associate as usual the
derivation D = − ad(X ).
The first thing to notice is that each ϕt transforms an admissible curve into an admissible curve hence preserves
the distribution ∆. Then ϕt preserves the length of all admissible curves and its differential should preserve the
length of the tangent vectors to admissible curves. Thanks to the fact that

ϕt ◦ Lg = Lϕt (g) ◦ ϕt , ∀t ∈ R, g ∈ G,

this implies that (dϕt )e = etD is orthogonal on ∆ for all t and finally that D is skew-symmetric w.r.t. the inner
product of ∆.
It is not true that all isometries of left-invariant sub-Riemannian structures that fix the origin are automor-
phisms. For instance a counter-example was built by John Milnor on the rototranslation group (see [13]).
However it has been proven by Kivioja and Le Donne that in case when G is nilpotent (and connected) the
group of isometries is a Lie group of affine transformations (see [11]). An affine transformation is by definition
the composition of an automorphism with a left translation. Since all left translations are isometries, we can
state:

A.2 Theorem: Let G be a nilpotent connected Lie group and (∆, h , i) be a sub-Riemannian structure on G,
where ∆ is a left-invariant bracket generating distribution and h , i is a left invariant inner product on ∆. The
group of isometries of G that preserves the origin is the Lie group of automorphisms of G whose Lie algebra is
the set of derivations of g that preserve ∆ and are skew-symmetric on ∆.

14
A.1 The semi-simple case
It is assumed in this section that g is a semi-simple Lie algebra, and that g = l ⊕ p is a Cartan decomposition
of g. Let θ be the related Cartan involution, that is the automorphism of g whose restriction to l is the identity
Il and the restriction to p is −Ip .
We assume in what follows that the inner product defining the sub-Riemannian metric is the restriction to the
left-invariant distribution ∆ of the associated inner product, that is:

hY, Zi = −B(Y, θZ) where B stands for the Killing form.

All the derivations being inner, let D = − ad(X) for some X ∈ g. A straightforward computation using the
properties of the Killing form shows that

∀Y, Z ∈ g hDY, Zi = −hY, (θ ◦ D ◦ θ) Zi.

Let us assume that ∆ is D = − ad(X)-invariant. Then the restriction of D is skew-symmetric on ∆ if and only
if it commutes with θ. But another straightforward computation shows that:

(D ◦ θ)Y = (θ ◦ D)Y and (D ◦ θ)Z = (θ ◦ D)Z =⇒ (D ◦ θ)[Y, Z] = (θ ◦ D)[Y, Z].

Since ∆ is assumed to be bracket generating we obtain that D, and not only its restriction to ∆, must commute
with θ. In other words D cannot be skew-symmetric on ∆ without being skew-symmetric on g.
Let Y ∈ g. Then (D ◦ θ)Y = [θY, X] and (θ ◦ D)Y = θ[Y, X] = [θY, θX]. If these two expressions are equal for
all Y ∈ g then according to the properties of semi-simple Lie algebras we deduce θX = X hence X ∈ l.

A.3 Theorem: Let G be a semi-simple connected Lie group and (∆, h , i) be a sub-Riemannian structure on
G, where ∆ is a left-invariant bracket generating distribution and h , i is a left invariant inner product on ∆.
Let n be the normalizer of ∆ in g. The group of automorphic isometries is isomorphic to the Lie group whose
Lie algebra is equal l ∩ n.

Proof: Indeed the group of isometries that fix the origin is a Lie group ([11]). The group of automorphic
isometries is its intersection with the Lie group of automorphisms of G, hence a Lie group. Its Lie algebra is
the set of derivations D = − ad(X) that preserve ∆, that is X ∈ n, and that are skew-symmetric on ∆ hence
on g, that is X ∈ l. 

References
[1] V. I. Arnold and A. Avez, Ergodic Problems in Classical Mechanics. New York: Benjamin (1968).

[2] V. Ayala and A. Da Silva, Controllability of Linear Control Systems on Lie Groups with Semisimple Finite
Center, SIAM Journal on Control and Optimization 55 No 2 (2017), 1332-1343.
[3] V. Ayala and A. Da Silva, On the characterization of the controllability property for linear control systems
on nonnilpotent, solvable threedimensional Lie groups, Journal of Differential Equations, 266 (2019), 8233-
8257.

[4] V. Ayala, A. Da Silva and G. Zsigmond, Control sets of linear systems on Lie groups. Nonlinear Differential
Equations and Applications - NoDEA 24 No 8 (2017), 1 - 15.
[5] V. Ayala, A. Da Silva, P. Jouan and G. Zsigmond, Control sets of linear systems on semi-simple Lie
groups. J. Differ. Equ. (2020) https://doi.org/10.1016/j.jde.2019.12.010

[6] V. Ayala and P. Jouan, Almost-Riemannian Geometry on Lie Groups. SIAM Journal on Control and
Optimization 54 No 5 (2016), 2919-2947.

15
 [7] V. Ayala and J. Tirao, Linear Control Systems on Lie Groups and Controllability. American Mathematical
Society, Series: Symposia in Pure Mathematics, Vol. 64, pp. 47-64, 1999.
[8] A. Da Silva, Controllability of linear systems on solvable Lie groups, SIAM Journal on Control and
Optimization 54 No 1 (2016), 372-390.

[9] T. Ferraiol, M. Patro and L. Seco, Jordan decomposition and dynamics on flag manifolds, Discrete and
Continuous Dynamical Systems - Series A 26 No 3 (2010), 923-947.
[10] P. Jouan, Equivalence of Control Systems with Linear Systems on Lie Groups and Homogeneous Spaces.
ESAIM: Control Optimization and Calculus of Variations, 16 (2010) 956-973.

[11] V. Kivioja, E. Le Donne, Isometries of nilpotent metric groups. J. École polytechnique, Mathmatiques,
Tome 4 (2017), p. 473-482.
[12] A. W. Knapp, Lie Groups Beyond an Introduction. Second Edition, Birkhäuser, Berlin, (2004).
[13] J. Milnor Curvatures of left invariant metrics on Lie groups. Advances in Math. 21 (1976), no. 3, 293-329.

[14] G. D. Mostow, Fully Reducible Subgroups of Algebraic Groups. American Journal of Mathematics, 78 No
1 (1956), 200-221.
[15] E. Noether, Invariant Variation Problems. Gott. Nachr., vol. 1918, pp. 235-257, 1918. [Transp. Theory
Statist. Phys.1,186(1971)].
[16] M. Patrão, Entropy and its variational principle for non-compact metric spaces. Ergod. Th. & Dynam.
Sys., 30, (2010), 15291542.
[17] M. Patrão, The topological entropy of endomorphisms of Lie groups. Israel Journal of Mathematics, 234
No 1, (2019), 55-80.
[18] A. L. Onishchik and E. B. Vinberg, Lie Groups and Lie Algebras III - Structure of Lie Groups and Lie
Algebras. Berlin: Springer (1990).
[19] L. A. B. San Martin, Algebras de Lie, Second Edition, Editora Unicamp, (2010).

16