A review on some classes of algebraic systems

Vctor Ayala and Heriberto Romn-Flores

Instituto de Alta Investigacin

Universidad de Tarapac
Casilla 7D, Arica, Chile
November 9, 2016

Abstract
In this paper we review some algebraic control system. Precisely, linear
and bilinear control systems on Euclidean spaces and invariant an linear
control systems on Lie groups. All of them has a common issue: there
exists and associated subgroup. From this object we survey the control-
lability property for these class of systems. Specially, from those coming
from our own contribution to the theory.

1 Introduction
This review was intended as an attempt to motivate researchers to take attention
on some special classes of algebraic control systems on Lie groups. In this paper
we analyze the following categories of control systems

1. Lin (Rd ) : Linear systems on the Euclidean space Rd

2. Bil (Rd ) : Bilinear systems on the Euclidean space Rd
3. Inv (G) : Invariant systems on a Lie group G
4. Lin (G) : Linear systems on a Lie group G

All of them has a common issue: there exists a semigroup associated to

the accessible set of the system from the identity element. Furthermore, the
controllability properties of these classes of systems are strict related with this
algebraic object.
Controllability property of control systems is one of the most relevant prob-
lems in control theory. It has been a subject of huge interest and has generated
an enormous activity in research for dierent classes of control system. How-
ever, still there is no general criterion to determine this important property for
Supported by Proyecto Fondecyt n 1150292. Conicyt, Chile.
Supported by Proyecto Fondecyt n 1151159. Conicyt, Chile.

1
a given control system, even for systems with strong dierentiable and algebraic
structures.
Therefore, after to introduce the associated semigroup of a specic class of
algebraic system we also mention some controllability properties associated to
the semigroup. Specially, those coming from our own contribution to the theory.
As we will see, in the classical class of linear systems and when the control are
unbounded, the accessible set from the origin is a subspace, which determines
controllability as soon as this subspace has non empty interior, see [49]. On
the other hand, when the controls are restricted, in [23] the authors show the
existence and uniqueness of a control set with non empty interior. Essentially,
a control set C is a subset of the state space where the system is approximately
controllable on the boundary C and it is controllable on its interior int(C).
Controllability of bilinear systems has been a source of research since more
than 40 years, but still there is not a complete characterization. In this case the
origin is a global singularity, so the accessible set from the identity is trivial.
However as we will see, there exists a semigroup associated to the system. On
the one side there exists an algebraic approach, [31], specially when the controls
are unrestricted. On the other one, for restricted control in [23] the authors
give an analytic dynamic approach that can be applied to the bilinear case.
The main ingredients are the projection of the system on the sphere (actually
on the projective space), the notion of chain control set and the Morse spectrum.
For concrete application we suggest to see [51] and [57].
Linear and bilinear control systems on the Euclidean space Rd are relevant
from practical and theoretical point of views. But, there are many applications
coming from mechanical or physical problems where the state space is not the
vector space Rd but a Lie group, see for instance [22], [30],[40],[45],[46] and [51].
Furthermore, bilinear system can be also approached by Lie theory. In fact, the
Lie algebra generated by the matrices of the system play an important role. See
also some preliminary result on the equivalence between bilinear systems, [6],
[10] and [4].
For the class of invariant systems on Lie groups there are many controllability
results which are obtained on specic state spaces. For example, Hilgert and
Lawson [38] and Ayala in [3] characterize the nilpotent case. Sachkov extends in
[60] and [61] the result in [3] from nilpotent to completely solvable Lie groups.
Furthermore, Jurdjevic and Sussmann [47], study the problem on compact Lie
groups. Gauthier in [33] and Jurdjevic and Kupka in [48] on semi-simple Lie
groups and their homogeneous spaces. L. San Martin and P. Crouch, [65] on
principal bre bundles with compact structure group, and recently in [29] for
complex Lie groups, etc. For a nice and complete survey on the topic see
references [61] and [62].
For the rst and third categories of systems, the accessible set from the iden-
tity A(e) is a semigroup. Here, A(e) stands for the set of points that can be
reached from e through all admissible trajectories in positive time. Unfortu-
nately, for the class of linear systems on Lie groups this is not longer true. In
fact, in [13] the authors give an example on the connected semisimple Lie group
G = SL(2, R) where the system is local controllable around the identity but not

2
controllable at all. Here, G stands for the set of trace zero real matrices of order
two.
With respect to the fourth class we analyze both: the unrestricted and
the restricted case. Actually, we explicitly mention results coming from the
very beginning of this class of systems in [12], [13],[42],[43] and also more recent
results relatives to the restricted case and the existence and uniqueness of control
sets, [8] and [7]. Despite the fact that the accessible set A from the identity is
not a semigroup we associate to the linear system a semigroup S which depends
on A, [7].
According to every particular class of systems, we include controllability
results. Specially those coming from our own contribution to the theory. Except
to the rst class we do not prove every result. However, when possible we explain
the meaning of theorems and notions through examples. We start the review in
a more general set up.

2 Some general facts of systems on manifolds

Let us consider an ordinary dierential equation X on a connected dierentiable
manifold M of dimension n which model a concrete dynamic system. A control
system on M allows to modify the behavior of X according to dierent strate-
gies denominated controls. From the mathematical side, in our case from the
dierential geometry point of view, a control system = (M, D) can be stated
as a family of vector elds D coming from a family of controlled dierential
equations
m

x(t) = X(x(t)) + uj (t)Y j (x(t)),
j =1

where X is the drift, i.e., the dynamic to be controlled. The vector elds Y j ,
j = 1, ..., m, are dened on M and are weighted by the class of admissible locally
integrable functions

u U = {u : R Rm , u is locally integrable}

is the set of the admissible controls with closed, convex and 0 int().
The set U of admissible controls is closed under concatenation, that is, if u
and v belong to U, then the function w = u v dened by

u(t), t [0, T ),
w(t) =
v(t T ), t [T, )

belongs to U as well.
The system should warranty the existence and uniqueness of a solution
(x, u, t) (denoted also by ut (x)), of any dierential equation associated to a
specic control u an arbitrary initial state x of M and t Dom((x, u, )). In
our particular classes of control systems any vector eld is complete, means that
the solution (x, u, t) is well dened at any t R.

3
 Given a control system , it appears some natural mathematical problems
which are hard to answer. For instance, the controllability property. Let x be a
x initial state of M and y M arbitrary, does y can be reached from x in non
negative time through an admissible solution of ? If the answer is positive,
we say that y is accessible from x through . A control system is said to be
controllable from x if the accessible set A(x) of from x dened by

A(x) = {y M : u U, t 0 : (x, u, t) = y}

is the whole space. And is said to be controllable if it is controllable from any

point of the manifold.
On the other hand, the set of controllable points to x is dened by

A (x) = {y M : u U, t 0 : (y, u, t) = x}.

Furthermore, a state x M is said to be accessible from x in T 0 units of

time (in exactly T 0 units of time) if there exists u U, and t [0, T ] such
that (y, u, t) = x, (u U : (y, u, T ) = x). We denotes by A(x, T ) and
E(x, T ) the accessible set from x in T units of time and in exactly T units of
time, respectively. Of course,

A(x) = t 0 A(x, t) = T 0 E(x, T ).

Obviously, not any arbitrary system is controllable and it is of interest to

know whether a system is controllable or not. It is worth to point out that it
is a really challenge to characterize this property for general systems, specially
when is a proper subset of Rm . In this more realistic way, some times it is
possible to characterize a special subset of M with non empty interior where
the system is approximately controllable, in the following sense.
Let = (M, D) a control system. A nonempty set C M is called a control
set if for every x C

i) there exists u U such that (t, x, u) C, for any t R

ii) C cl(A(x)), where cl denotes the closure, and
iii) C is maximal with respect to conditions (i) and (ii).

See [9],[18],[23],[66] and [67].

In this review we consider a semigroup associated to some classes of algebraic
control systems. Means, systems with algebraic structure on the manifold and
on the dynamics. It turns out that on a connected Lie group, a semigroup
with non empty interior containing a neighborhood of the identity element e
generates all the group. Therefore, when the accessible set from the identity is
a semigroup, local controllability from e implies global controllability from the
identity. Unfortunately, in [13] it is shown that for a linear control systems on the
Lie group SL(2, R) the accessibility set from e is not a semigroup. Furthermore,
for a transitive linear system on a Lie group G (see Remark below), in [42] the
author shows that, A(e) is a semigroup if and only if it coincides with G.

4
 For a general system = (M, D), a fundamental results is the Orbit Theo-
rem [69] which allows to reduce the state space M of any initial condition x
to its orbit, i.e., the dierentiable manifold determined by

G (x) = Zt11 ... Ztkk (x) : Z j D, k N and tj R M.

where (Zt )t R stands for the ow of the vector eld Z, in other words, a 1-
parameter group of M-dieomorphisms. Just observe that

G = Zt11 ... Ztkk : Z j D, k N and tj R

is a group of global dieomorphism and

S = Zt11 ... Ztkk : k N, Z j D and tj 0

is a semigroup. Furthermore, if

SpanLA (D) = SpanLA X, Y 1 , ..., Y m < ,

it turns out that G (e) is a nite dimensional Lie group.

Except for linear control systems on Euclidean spaces, we always assume
that any system satisfy the Lie algebra rank condition, (larc), which means that

SpanLA (D) = g.

In this review we consider linear and bilinear systems on the Euclidean space
Rd and invariant and linear systems on a connected Lie group G with Lie algebra
g. For all of these systems, the Lie algebra generated by the corresponding vector
elds D is nite dimensional. In fact, as we will see, the normalizer

n = {Z X (G) : [Z, g] g}

of g in the set X (G) of smooth vector elds on G, contains for any piecewise
constant control every dynamic of our four classes of systems. Actually, in [12]
it is shown that n is dieomorphic to the semidirect product between g and
the Lie algebra aut(G) of the Lie group Aut(G) of all G-automorphism, i.e.,
n= gs aut(G) which is nite dimensional with dimension dim(g) + ((dim(g))
2

at most. And
SpanLA (D) n.
Except for Lin (Rd ) and without lost of generality, we assume that any system
is transitive, which means that any two points can be connected through a -
solution considering positive and negative times. In other words, G (e) = Gor

equivalently SpanLA (D) = g.

Recall that the Lie bracket [X, Y ] between two vector elds is a new vector
eld dened by
d
[X, Y ] (x) = ( )t = 0+ Yt Xt t Yt Xt (x).
dt

5
We notice that the bracket depends on negative times, so in general SpanLA (D)
is related to the controllability property but not to much. Actually, according
to the orbit theorem, in our particular classes of algebraic system we consider, it
is possible to recover de orbit just by a derivation processes. In fact, the tangent
space of any orbit is given by

Ty G(x) = SpanLA (D)(y).

Just to x some ideas, in our context the Lie group G could be a real vector
space Rd , a sphere S n , i.e., when n = 1, 3, 7, a torus T n = S 1 S 1 ... S 1 ,
the Heisenberg Lie group, any connected matrix group, such as GL+ (n, R) the
invertible real matrices of order n, or their subgroups SL(n, R) and SO(n, R),
the matrix groups of determinant 1 and orthogonal, respectively.

3 Linear control systems on Euclidean spaces

Let us start with a very simple example which contains several relevant ingre-
dients of the theory.

Example 1 In the Pontryagin book, [58] the authors establish the following
optimal problem

how to stop a train at the station at minimum time?

To solve it the authors apply the maximum principle (appears for the rst time
on this book) on a linear control system on the plane.

Consider the ideal case of a straight railway line. For any t 0, denotes
by x(t) the distance from the train to the station, that we consider as the
origin. According to the Newton law, the force is giving by F = ma. Here, a =

y(t) where y(t) is the velocity. Without loss of generality we consider m = 1.

Therefore, we get on R2

x(t) = y(t), y(t) = u(t), u U , with

U = {u : R [1.1] Rm , u is locally integrable}

called the set of the admissible restricted control. The train is controlled by the
acceleration u = a.
We can write this system as a restricted linear control system on R2 as
follows
x(t)
0 1 x(t) 0
Lin (R2 ) : = + u(t), u U.
y(t)
0 0 y(t) 1

Geometrically, for a given point (x0 , y0 ) R2 we need to nd a solution

of the system transferring the initial condition to the origin (0, 0) at minimum
time. First, it is necessary to check controllability, means to know if it is possible

6
to connect any initial condition to the origin of the systems in non negative time.
By some elementary computation

rank(BAB) = 2 and Spec(A) = {0} .

As we will see and according to Theorem 8, the system is controllable. Therefore,

there exists at least one control connecting (x0 , y0 ) to (0, 0). Thus, the problem
is well posed. In fact, by applying the Pontryagin Maximal Principal [58], it
turns out that the optimal control u exists and it is bang-bang, which means
in this particular case that u takes it optimal values in the boundary of , i.e.,
= {1, 1} .
The solutions of u = 1 and u = 1, are parabolas and we denote by u1 the
intersection of the integral curve through the origin determined by the control
u = 1 with the half plane y > 0. In the same way we denote by u1 the
intersection of the integral curve through the origin determined by u = 1 with
the half plane y < 0. So, the maximal breaking (u1 ) and acceleration (u1 )
optimal curves arriving to the origin, allow us to solve the problem starting at
any point (x0 , y0 ) and reaching the origin with at most one change: from 1
to 1 or conversely. Just observe that both curves together separate the plane
in two connected components. For instance, to reach the origin in optimal way
starting from (2, 1) you rst travel with u = 1 up to reach the curve u1 which
will transport the train to the origin at minimum time. We observe that the
projection of the intersection point between both curves gives to the driver the
exact position where he needs to change the strategy and to break the train
with u = 1.
It is worth to mention that the Pontryagin Maximum Principle got the Lenin
Prize in Russia recognizing its great contribution to the Society.
The classical linear system Lin (Rd ) = (Rd , D) on the Euclidean space Rd
is determined by the dynamics of D coming from
m
x(t)
= Ax(t) + uj (t) bj , bj Rd and u U
j =1

with is a closed and convex subset of Rm with 0 int(). Here A gl(d, R),
the Lie algebra of the real matrices of order d. The classical cost matrix B is
built with the columns vectors bj , called the control vectors, j = 1, ..., m.
Essentially, this system depends on two kind of dynamics. The linear and
the invariant ones. In fact, the linear dierential equation

x(t)
= Ax(t) on Rd

is controlled by m invariant constant vector elds bj , j = 1, ..., m. On the

other hand, any vector b in the Abelian Lie algebra Rd induces by translation
a vector eld Z b (x) = b on the commutative Lie group Rd which is obviously
bi-invariant.

7
 Given an initial condition x0 Rd and u U, it is possible to completely
describe the solution of the system as
t
ut (x0 ) = etA x0 + e A B u () d
0

which satisfy the Cauchy problem with initial value:

x = Ax + Bu, x (0) = x0 .

Thus, ut (x0 ) with t R describes a curve in Rd such that starting from x0

the elements on the curve are reached from x0 forward and backward through
the specic dynamic of the linear system determined by the controls.
Next, we show that in the unrestricted case, i.e., when = Rm , the acces-
sibility set of a linear control system on Rd is a semigroup. Means, a vector
subspace of (Rd , +). Since the proof is direct we include it here. We follow our
reference [18].

Theorem 2 Let Lin (Rd ) be an unrestricted ( = Rm ) linear control system.

Then,

1. For any x0 Rd and T > 0

E (x0 , T ) A (x0 , T ) A (x0 ) and eT A x0 + A (0, T ) A (x0 , T )

2. If 0 T1 T2
E (0, T1 ) A (0, T2 )

3. For any T > 0 the sets E (0, T ) and A (0) are vector subspaces

Proof. Property (1) comes directly from the solution shape. To prove (2) just
observe that it is possible to rest at the origin with u = 0. Finally, by denition
T
E (0, T ) = uT (0) = eT A e A Bu () d : u U .
0

Under the hypothesis = Rm it follows that U is a vector space. On the

other hand

uT1 (0) + uT2 (0) = uT1 +u2 (0) and u u

T (0) = T (0) .

Hence, E (0, T ) is a vector space. As the union of a increasing chain of

subspaces is a subspace, the proof is done.
Furthermore, it is possible to characterize A (0) in an algebraic way. Denotes
by A, B the small A-invariant subspace of Rd containing Im(B), the image of
the linear transformation B.

8
Proposition 3 Let Lin (Rd ) be an unrestricted linear control system. Then,
for any T > 0
E (0, T ) = A, B .
k
Proof. For A gl(n, R) the exponential map is given by eA = k 0 Ak! . Since
A, B is an A-invariant subspace it follows that e A is also A-invariant for any
R. Hence, for any control u the trajectory starting from the origin
T
eT A e A Bu ( ) d : R
0

it remains at A, B . Hence, A (0) A, B .

Reciprocally, consider the intersection of subspaces

J = E (0, T )
T >0

The Euclidean space Rd is nite dimensional, so there exists T0 > 0 such that

0 < T < T0 implies E (0, T ) = J .

Let u . For every T [0, T0 ], uT (0) J . Therefore, if

T
d
0 T T0 ( eT A e A Bu ( ) d ) = Bu J .
dT T 0

For any T (0, T0 ) dene the control uT : [0, +) R as follows

u 0tT u ut (0) 0tT

uT (t) = to get t T (0) = tA u
0 T <t e t (0) T <t

If T and t are suciently near to the origin we obtain ut T (0) J . By taking

partial derivatives for any k = 0, 1, 2, ... we get

k ut T (0)
( )T =0 = Ak Bu J .
tk t=0 T

Hence, A, B J. Consequently, if 0 < T < T0 and T < T 1

E (0, T ) E 0, T 1 A, B E (0, T )

and the proof is complete.

The previous analysis allows to characterize the controllability property of
this class of systems, as following.

Proposition 4 Let Lin (Rd ) be an unrestricted linear control system. The

following conditions are equivalent

1. Lin (Rd ) is controllable

9
 2. There exists x0 Rd such that Lin (Rd ) is controllable from x0
3. Lin (Rd ) is controllable from the origin
4. A (0, T ) = Rd , for any T > 0.
5. E (0, T ) = Rd , for any T > 0
6. Rd = A, B .
The next and the previous results due to Kalman, [49] provides a criterion
for testing controllability in an algebraic easy way.
Theorem 5 (Kalman rank condition) Let Lin (Rd ) be an unrestricted linear
control system. Then,
Lin (Rd ) is controllable rank(K) = n
where K = B AB A2 B...An1 B is a n(n m) real matrix.
Proof. At present we will merely show that
A, B = Ak bj : j = 1, 2, ..., m , k = 0, 1, 2, ..., n 1 .
But, the Calley-Hamilton theorem said that it is not necessary to compute Ak bj ,
for j = 1, 2, ..., m, when k n.
Remark 6 As we proved, when = Rm the accessible sets from the origin is
a subspace. But its interior could be empty. Equivalently, dim A(0) < n.
Now, what happens if Rm ? It turns out that in few cases A(0) is still
a subspace. Precisely, in [23] the authors show that
Theorem 7 Let Lin (Rd ) be a restricted linear control system which satisfy
the Kalman condition. Hence, there exist one and only one control set with non
empty interior containing the origin and it is given by
C = cl(A(y)) A (y).
Furthermore, the system is controllable at the open set int(C).
We suggest the readers to see an example on the book [23]. Finally, as a
consequence, we obtain
Theorem 8 Let Lin (Rd ) be a restricted linear control system which satisfy the
Kalman condition. Therefore,
Lin (Rd ) is controllable on Rd Spec(A)Ly R = {0} .
Hence, in this case controllability is equivalent to the fact: all the eigenvalues of
A belong to the line iR. Here Spec(A)Ly means the Lyapunov spectrum of the
matrix A, i.e., the set of the real parts of the eigenvalues in Spec(A). In this
particular case the accessibility set A(0) is also a subspace and equals to Rd .

10
4 Bilinear control systems on Euclidean spaces
Bilinear control systems are relevant from both, theoretical and practical point
of view, with a huge range of applications, [22],[23],[30],[31],[40],[45], [46],[51]
and [57]. See also [4],[6] and [10].
By denition a bilinear control system Bil (Rd ) in Rd is determined by the
family of dierential equations

m
x(t)
= A + uj (t)Bj x(t), t R, x(t) Rd . (bilinear)
j=1

Here A, B1 , . . . , Bm gl(d, R) are real matrices of order d and u U

U = {u : R Rm , u is locally integrable}

is the set of the admissible controls with closed, convex and 0 int().
At once you can see that the Lie algebra generated by the matrices

= SpanLA {A, B1 , . . . , Bm }

is a Lie subalgebra of gl(d, R). In fact, the Lie bracket

[A, B] = AB BA gl(d, R), for any A, B gl(d, R).

In particular, there exists a connected Lie subgroup G GL(d, R) with Lie

algebra , [72].
On the other hand, let us consider an ane control system as in 2. If has
a global singularity let say at x0 , then it is possible to linearize the vector elds
X and Y j , j = 1, ..., m of the system obtaining a bilinear system Lin (Rd , x0 ).
Global information of Lin (Rd , x0 ) gives local information of , exactly like in
the classical approach of Hartman-Grobman type theorems, [36].
Contrary to the linear system on Euclidean spaces, the control vectors of
Bil (Rd ) have directly inuence on the state. This situation allows to use
systematically the Lie theory. In fact, by considering the family of piecewise
constant control as admissible, we obtain that for any constant control u, the
bilinear system determines a linear dierential equation. Thus, D is a family of
matrices. Just observe that the dierential equation

x(t)
= Ax(t), x(0) = x0 , x(t) Rd

has the solution x(t) = etA x0 . And, at the same time the solution of the matrix
dierential equation

X(t) = A(t), X(0) = Id, X(t) GL+ (d, R)

is given by X(t) = etA . As we will see in the next section, it is possible to know
the behavior of the control system Bil (Rd ) at the point x0 through the action
of an invariant control system Inv (G) dened on a special Lie subgroup G

11
of the invertible matrix group GL(d, R) of order d. Actually, G is built as the
connected subgroup with Lie algebra
SpanLA (D).
Controllability of bilinear systems has been a source of research since more
than 40 years, but still there is not a complete characterization. On the one
side there exists an algebraic approach, [31], specially when the controls are
unrestricted. And on the other hand, for restricted control in [23] the authors
give an analytic approach on ber bundles that can be applied to Bil (Rd ).
For the following denitions and results we refer to [23], Chapters 7 and 12.
From the Sussmann Orbit Theorem, [69] without lost of generality we assume
that the system satises the Lie algebra rank condition at any x Rd -{0} , i.e.,
m
dim(SpanLA {A + ui Bi : u U}(x)) = dim Rd . (1)
i=1

The control system Bil (Rd ) has the following associated systems:
a) The angle system P which is dened by the projection of Bil (Rd ) onto
the projective space Pd1 ,
m
P : s(t)
= h(A, s(t)) + ui (t)h(Bi , s(t)), s Pd1 , (2)
i=1

here h(A, s) = (A sT AsI)s, where I is the identity matrix and u U,

and
b) The radial system dened on R+ by
r(t) = ||(t, x, u)|| , (3)
where denotes the Euclidean norm in Rd .
The solutions of the projected system (2) are denoted by P(t, s, u) for the
initial value P(0, s, u) = s Pd1 . Recall that the real projective space Pd1
is determined by a quotient manifold from the sphere S d1 as follows: Pd1 =
S d1 / , where the the antipodal dierentiable equivalence relation is dened
by x y y = x.
Since for any control u U the origin is a singularity of any bilinear system,
we need to dene what we understand for controllability in this case. Let us
introduce the associated semigroup to Bil (Rd ) as follows

S = et1 (A+u1 B) etk (A+uk B) : tj 0, j = 1, ..., k and uj R .

Denition 9 A bilinear control system Bil (Rd ) is said to be controllable on

Rd -{0} if given any two points x, y Rd -{0} there exists an admissible control
u U and t 0 transferring x to y at t units of time. In other words,
SBil (Rd ) : y = (x).

12
 There are many results controllability results to this class of control system.
In this paper we concentrate in the specic case of dimension two. In particular,
we show some contribution that we have done, [16] and [5].
First, we start with a fundamental results on this theory due to Colonius
and Kliemann, [23].

Theorem 10 Consider the bilinear control system Bil (Rd ) and its projected
system (2) satisfying (1). We assume that the control range Rm is compact.
Then the following statements are equivalent:

1. Bil (Rd ) is controllable in Rd -{0}

2. a) The projected system P ( 2) is controllable on Pd1 , and
b) 0 ( , ).

In this context, for a solution (t, x, u) with x != 0 and u U the Lyapunov

exponent is dened as
1
(u, x) = lim sup log (t, x, u()) . (4)
t t
The Lyapunov spectrum consists of all Lyapunov exponents, i.e.

Ly = {(u, x) : (u, x) U Rd - {0}}.

Extremal Lyapunov exponents are those exponential growth rates that are de-
ned globally as follows

= inf inf (u, x), = sup sup (u, x).

u U x = 0 uUx=0

Since we are assuming compact < < . It turns out that [23].

Theorem 11 Consider the bilinear control system and its projected system (2)
satisfying (1). We assume that the control range Rm is compact and the
projected system (2) is completely controllable on Pd1 . Then the spectra satisfy,

[ , ] = Ly .

Theorem 11 also involve the Floquet spectrum, but we will not comment
this notion here. In the spirits of Theorem 11, in the sequel we mention two
particular cases: the single control and the multiple controls. We also include
an example, coming from [5].

Theorem 12 (Single control case) Consider the unrestricted bilinear control

system with d = 2, = R and assume that the Lie algebra rank condition (1)
holds. Then the projected angle system is controllable on P1 if and only if there
exists a constant control u R such that the matrix A + uB has a complex
eigenvalue.

13
 The next theorem use the projection map P : Rd {0} Pd1 dened by

P(x) = x (= {x})
Theorem 13 (Multiple controls case) For the system (2) under the Lie algebra
rank condition (1) the following statements are equivalent:
1. The system is controllable
2. The system admits a controlled rotation
m
3. The matrix set M := {A + i=1 ui Bi : u } satises at least one of the
following conditions:

(a) M contains a matrix with a pair of complex eigenvalues, or

(b) M contains a matrix A(u0 ) with double real eigenvalue and only one
eigendirection E 0 , and a second matrix A(u1 ) whose ow moves in
the same direction as {P(t, , u0 ), t 0} in a neighborhood of PE 0 ,
or
(c) M contains a nite sequence {A(ui ), i = 1, ..., n} of matrices with
distinct real eigenvalues whose eigenspaces satisfy either
PE1 (u2 ) < PE2 (u1 ) < PE1 (u3 ) < PE2 (u2 ) < ... < PE2 (un ) < PE1 (u2 ) < PE2 (u1 )
in case of a counterclockwise controlled rotation, or
PE1 (u2 ) > PE2 (u1 ) > PE1 (u3 ) > PE2 (u2 ) > ... > PE2 (un ) > PE1 (u2 ) > PE2 (u1 )
in case of a clockwise controlled rotation.

Let us explain these results through a couple of examples.

Example 14 First we consider the bilinear single control system

x1 2 0 2 2 x1
= + u(t) .
x2 0 1 0 1 x2

The projected angular system on S1 satises the dierential equation

= [(1 + u(t) cos() 2u(t) sin()] sin().
Since = 0 is a common xed point it does not satisfy larc. In particular, the
projected system is not controllable on P1 . The same situation happens if you
consider the system

x1 2 0 1 0 x1
= + v(t) .
x2 0 1 2 1 x2
But, if you combine both systems as

x1 2 0 2 2 1 0 x1
= + u(t) + v(t)
x2 0 1 0 1 2 1 x2

14
The projected angle system on S1 reads as

f ((u(t), v(t)), ) = (1 + u(t) sin() + 2(u(t) + v(t)) cos() + 2(u(t) v(t))

and one can check easily that (1) is satised. Computing the eigenvalues for a
constant control (u, v) we get
3 3 1
(u, v) = + u+v (1 + u)2 + 16uv.
2 2 2
By taking u = 1 and v = 1 we obtain a pair of complex eigenvalues. Hence
the system is controllable on P1 .

Remark 15 Necessary and sucient conditions for an unrestricted bilinear

system = Bil (Rd )

: x(t)
= (A + uB) x(t), t R, x(t) R2 (5)

where given in [19],[41]. In this case u U with = Rm , and A, B gl(2, R).

Theorem 16 Assume X, Y sl(2, R). Then, the unrestricted bilinear system

: x(t)
= Ax(t) + uBx(t), u R

is controllable in R2 -{0} if and only if det [A, B] < 0.

We would like to show that the algebraic condition in the previous theorem
is actually a very interesting and nice geometric condition. In the sequel we
follow [16] and [17]. For the Lie theory concept, see next section.
Given a Lie algebra g and X g, the adjoint map ad(X) : g g is dened
by ad(X)(Y ) = [X, Y ] . The Cartan-Killing form

k(X, Y ) = tr(ad(X) ad(Y ))

is a multiple of the trace tr(XY ), which induced the quadratic non degenerate
form Q(Z) = tr(Z 2 ). Through this form it is possible to identify the Lie algebra
sl(2, R) with R3 with axis

1 0 0 1 0 1
H= , S= , A= .
0 1 1 0 1 0

corresponding to the hyperbolic, symmetric and skew-symmetric basis elements.

The set C = Q(0) is a double cone with axis the line generated by A. Let us
denote by

Cint = {Z : Q(Z) < 0} and Cext = {Z : Q(Z) > 0} .

The elements of C are nilpotent matrices. While Cint contain all the matrices
with imaginary eigenvalues the elements of Cext are the diagonal real matrices.

15
Theorem 17 Assume det [A, B] < 0. Therefore, the restricted bilinear system

: x(t)
= Ax(t) + uBx(t), u = [1, 1]

is controllable if and only if the straight line

l = {A + uB : u = [1, 1]}

meet the interior of C, i.e.,

l int(C) != .

In other words, under the condition SpanLA {A, B} = sl(2, R), the restricted
system is controllable if and only if there exists a control u0 such that A + u0 B
has an imaginary eigenvalue.

More general, for a x non zero element x0 R2 we consider the map

X : [0, ) P(R2 ) dened by X (a) = Aa (x0 )

the accessible set of the system with range control a and where P(R2 ) denotes
the family of subsets of R2 . In [16] it is also shown the following results

Theorem 18 Consider a 0 and assume det [A, B] < 0. Therefore, the con-
trollability of the restricted bilinear system

a : x(t)
= Ax(t) + uBx(t), u a

with restricted range a = [a, a] is given by the relative position of the segment

la = {A + uB : u a }

as follows

1. If det(A) 0 then la Cint != and the system is controllable for any

positive real number a,

X (a) = R2 {0} for any a > 0.

2. If det(A) < 0, there are the possibilities

i) If det [A, B] < 0 then the line A + uB : u R cross the interior of C
and the system a is controllable if and only if la Cint != . The only
discontinuity (bifurcation) point of X is determined by

a = inf {a : la Cint = } .

In fact, for a < a , X (a) is strict contained in R2 -{0} . On the other hand,
for a > a , X (a) = R2 -{0} .

16
 ii) If det [A, B] > 0, the system a is not controllable for any a 0 and X
is a continuous map on (0, ) when you consider the Haudor metric on
P(R2 ).

Next, we show a controllable bilinear control system on the plane.

Example 19 Let us consider the bilinear system

: x(t)
= Ax(t) + uHx(t), u

with range and where A and H are basis elements of sl(2, R) as before. We
have,

1. det(A) = 1 > 0
2. det [A, H] = 4 < 0.

According to the previous theorem, the bilinear system is controllable for

any positive real number a.

Remark 20 It is worth point out that this specic class of control systems
in dimension two has served a model for Ledzewick and Shattler through the
optimal problem control for a two compartment model for cancer chemotherapy
with quadratic objective, [51].

Finally, we would like to mention that some eort was made in order to
study the equivalence problem of this class of systems. In fact, in [4],[6] and
[10].

5 An elementary review on Lie theory

In this section we shortly introduce the main ingredients of the Lie theory
we need for the next two sections which include invariant and linear control
systems on Lie groups. We mention a couple of references on this subject,
[26],[37],[64],[71] and [72].

Denition 21 A Lie group G is an analytic manifold such that the group op-
erations : G G G : (g, h) gh and : G G : g g 1 are analytic.

Example 22 The following sets are Lie groups under obvious multiplication

1. The Euclidean space Rn

c
2. The set GL(d, R) = det1 (0) and GL+ (d, R) the connected component
of the identity
3. The torus T d = S 1 .... S 1 (d-times circle)
4. The orthogonal group O(d) = {A GL(d, R ) | AAt = Id}

17
 5. The special orthogonal group SO(d) = {A O(d) | det(A) = 1}
6. The special linear group SL(d, R ) = {A GL(d, R ) | det(A) = 1}.
7. The Heisenberg group (R3 , ), with

(x, y, z) (a, b, c) = (x + a, y + b, z + c + xb).

Given a Lie group G the analytical maps

Rg , Lg : G G dened by Rg (x) = xg and Lg (x) = gx

called the right and the left translations on G respectively, are dieomorphisms.
The Lie algebra of G comes from the notion of invariant vector elds as
follows. We denote by X (G) the set of C -vector elds on G.

Denition 23 We say that X X (G) is a right invariant vector eld if

X Rg = (Rg ) (X) for every g G.

Here, (Rg ) or (dRg )e denotes the dierential of Rg at the identity e. Analogue

for left invariant vector eld.

Remark 24 It is not dicult to prove that given two right invariant vector
elds X, Y the Lie bracket [X, Y ] between them is also a right invariant vector
eld. Furthermore, observe that X g is completely determined by its value at
the identity. In other words, g is isomorphic to the tangent space Te G.

The set of right invariant vector elds on G is called the Lie algebra g of G
which satisfy

i) [X, Y ] = [Y, X] (skew-symmetric)

ii) [X, [Y, Z]] + [Z, [X, Y ]] + [Y, [Z, X]] = 0, (Jacobi identity).

A subspace V g is said to be a subalgebra if [V, V ] V and an ideal if

[V, g] V.

Example 25 We mention the Lie algebras corresponding to some Lie groups.

Recall that the Lie algebra g of a Lie group G is isomorphic to its tangent space
Te G at the identity element

1. T0 Rd = Rd
2. TId GL+ (d, R) = gl(d, R), the set of real matrices of order d
3. TId S d = Rd
4. TId O(d, R) = o(d) = {A GL(d, R ) | A + At = 0}, the skew symmetric
matrices

18
 5. TId SO(d, R) = so(d) = o(d)

To x ideas we consider the Lie algebra o(3) = X 1 , X 2 , X 3 of the rota-

tional group SO(3) in R3 . Its Lie algebra is computed by taking the derivatives
at t = 0 of the following curves through the identity matrix

1 0 0 cos t 0 sin t cos t sin t 0
L1 (t) = 0 cos t sin t , L2 (t) = 0 1 0 , L3 (t) = sin t cos t 0 ,
0 sin t cos t sin t 0 cos t 0 0 1

to obtain a basis of o(3)

0 0 0 0 0 1 0 1 0
X 1 = 0 0 1 , X 2 = 0 0 0 , X 3 = 1 0 0
0 1 0 1 0 0 0 0 0

That is, the set of skew-symmetric matrices in gl(3, R )

0
o(3) = 0 : , , R = SpanLA {X 1 , X 2 }.

0

In fact, [X 1 , X 2 ] = X 3 .

6. The trace zero matrices sl(d, R ) = {A gl(d, R) | tr(A) = 0}

7. The Heisenberg Lie algebra (R3 , +, [, ]), has the basis X 1 , X 2 , X 3 such
that X 1 , X 2 = X 3 is the only non vanish bracket. In fact, the Heisenberg
group has the matrix representation

1 x1 x3 : g (x ,x ,x )
R3 .
1 2 3
G = g = 0 1 x2 : x1 , x2 , x3 R

0 0 1

For instance, the curve : R R 3 , (t) = (t, 0, 0), determines X 1 .

Therefore, we get

0 1 0 0 1 0 0 1 0
X1 = 0 0 0 , X2 = 0 0 0 , X3 = 0 0 0 .
0 0 0 0 0 0 0 0 0
In this paper we just consider matrix Lie group, so the exponential map is
dened by

1 k
exp : gl(d, R) GL+ (d, R), exp A = A with A0 = Id.
k=0
k!

19
 Since d(exp)0 = Id there exists a neighborhood V G of e such that exp
is a local dieomorphism. Furthermore, for nilpotent and simply connected Lie
groups exp is a global dieomorphism, which means V = G.
The relation between the product exp X exp Y and the exponential of X +Y
is given by the Campbell-Baker-Haussdorf Formula in [72] as follows :

t2 t3
exp tX exp tY = exp(t(X +Y )+ [X, Y ]+ ([[X, Y ], Y ] [[X, Y ], X])+O(t4 ))
2 12
A C group homomorphism between two Lie groups G and H is called
a Lie group homomorphism. A bijective Lie group homomorphism of G with
itself is called a Lie group automorphism. When G is connected, the set of all
automorphisms Aut(G) of G forms a Lie group with Lie algebra aut(G), [72].

Remark 26 An important relation between a Lie group homomorphism :

G H, and its derivative (d)e : TeG TeH is given by (exp X) =
exp(d(X)), which comes from the commutative diagram
(d)e
g h
expg exph
G H

As a particular case etrA = det(exp A), A gl(d, R ), is obtained through

tr
gl(d, R) R
exp e
det
GL+ (d, R) R+

Denition 27 A Lie algebra g is say to be

1. Abelian if for any X, Y g the bracket [X, Y ] is zero

2. Nilpotent if there exists k 1 such that its descendent central series
stabilizes at the origin

0 = adk (g) = adk1 (g), ad1 (g) ... ad1 (g)

3. Solvable if there exists k 1 such that its derivative series stabilizes at

the origin,

0 = ad(k) (g) = ad(k1) (g), ad(k1) (g) ... ad1 (g)

4. Completely solvable if it is solvable and the operator ad() has only real
eigenvalues, i.e.,

Spec(ad(Y )) R, for every Y g

20
 4. Simple if g it is not Abelian and contains non proper ideals
5. Semisimple if it solvable radical, r(g), the bigger solvable subalgebra of g,
is null.

A Lie group is said to be Abelian, nilpotent, solvable, completely solvable,

simple, semisimple, if its Lie algebra is Abelian nilpotent, solvable, completely
solvable, simple, semisimple, respectively.

Example 28 Here we mention the type of some Lie groups

1. The Euclidean space Rd is Abelian

2. The torus T n = S 1 .... S 1 (n-times) is Abelian and compact. On the
other hand, any Abelian group has the form Rd T n for some d, n N.
3. The Heisenberg group (R3 , ) is nilpotent
A y
4. The Ane group : A GL(d, R), y Rd is solvable
0 1
5. The upper triangular group T (d) is completely solvable

T (d) = {A = (aij ) GL(d, R) : aij < 0 for j < i}

6. The orthogonal group SO(d, R) is compact and simple for d != 4

SO(d, R) = {A GL(d, R) | AAt = Id}

7. The orthogonal group SO(4, R) is compact and semisimple

8. The special linear group SL(d, R) is semisimple

SL(d, R) = {A GL(d, R) | det(A) = 1}.

6 Invariant control systems on Lie groups

An invariant control system Inv (G) on a connected Lie group G is determined
by a family D of dierential equations given by

m
D= X+ uj Y j : u U .

j=1

The drift vector eld X and the control vectors Y 1 , . . . , Y m here are elements of
the Lie algebra g of G which we think as the set of right invariant vector elds.
We consider U as before as the set of the admissible class of control.
In this introductory paragraph we follow our reference [11]. It is well known
that the class of invariant control systems is really relevant both from theoretical

21
and practical point of view. In fact, since the beginning of the 1970s many
people has been working in this kind of systems. We mention the rst work
in the subject by Brockett, R. [22]. Then, several mathematician started to
study this system on dierent classes of Lie groups: Abelian, compact, nilpotent,
solvable, completely solvable, simple, semisimple, etc. We mention some of
them [3],[29],[33], [38], [47],[48],[60],[65]. For acomplete survey on the topic see
references [61].
For the rst and third categories of systems, the accessible set from the
identity A(e) is a semigroup. Here, A(e) stands for the set of points that can
be reached from e through all admissible trajectories in positive time. Unfor-
tunately, for the class of linear systems on Lie groups this is not longer true.
In fact, in [13] the authors give an example on G = SL(2, R) where the linear
system is locally controllable from the identity but not controllable at all.
As appointed by Professor Jurdjevic, optimal control on Lie groups is a nat-
ural setting for geometry and mechanics, see [45],[46] and [59]. As a consequence,
dierential systems on Lie groups and their homogeneous spaces deserves to be
developed. For instance, the Dubins problem [30], the brachistochrona problem
[70], the control of the altitude of a satellite in orbit [40], etc., are described by
an appropriate invariant control system on some particular class of Lie group.
As usual, through all this section we assume that any invariant system satisfy
the Lie algebra rank condition, which in this case means that

SpanLA X, Y 1 , ..., Y m = g.

Under this natural condition and by the Sussmann Theorem the orbit of the
identity element coincides with the group, .i.e., GInv (G) (e) = G. The system
has an associated semigroup, in fact

Theorem 29 SInv (G) (e) = A(e) is a semigroup of G

Proof. We just sketch the prove. Since, the vector elds of the system generate
the Lie algebra it turns out that

A(e) = {exp(t1 Z1 ) exp(t2 Z2 )... exp(tj Zj ) : tj 0, Zj D, j = 1, ..., k} .

Thus, given g1 , g2 A(e) the product g1 g2 is an element of A(e).

In the sequel, we establish some controllability results for = Inv (G). We
start on a nilpotent Lie group. In [17] the authors state:
"if there exists a strictly increasing real function f on the positive trajectories
of , then the system can not be controllable".
Actually, assuming controllability it turns out that there exists a loop-
trajectory of the system starting and ending on the identity element, actually at
any element of G. In particular, an strictly increasing function f can not exists
along the loop. To develop this idea they introduce the notion of symplectic
vectors via the co-adjoint representation. Next, in [3] the author search for nd-
ing algebraic conditions to determine the existence of the symplectic vectors on
nilpotent Lie algebras. They state the following results

22
Proposition 30 Let G be a nilpotent simply connected Lie group with lie al-
gebra g and let h be an ideal such that g/h is not an Abelian Lie algebra. If
: g g/h is the canonical projection and there exists Z g such that
(Z) z(g/h) belongs to the center of g/h is a non null vector eld, there-
fore there exits a symplectic vector for Z.

Through this proposition it is possible to characterize the controllability

property of this system. In fact, let us dene the Lie algebra of the control
vectors by
h = SpanLA Y 1 , Y 2 , ..., Y m
and the centralizer of h by

z(g) = {W g : [W, Y ] = 0, for any Y h}

then we proved

Proposition 31 Let Inv (G) be an invariant control system on a nilpotent con-

nected and simply connected Lie group G. Assume that z(g) is properly contained
in z(h). Therefore, Inv (G) can not be controllable.

Finally, by applying the propositions before the article establish an impor-

tant results which decide controllability in a trivial algebraic way. Furthermore,
it allows to compute all the controllable systems on G.

Theorem 32 Let Inv (G) be an invariant control system on a nilpotent con-

nected and simply connected Lie group G. Therefore,

is controllable if and only if g = h.

The main idea in the proof use the co-adjoint representation to build several
homogeneous spaces and then to nd algebraic conditions on these spaces to
dene symplectic vectors.
For the class of invariant control systems on a solvable Lie group G the
controllability property can be characterized, as follows, see [38]:

Theorem 33 is controllable if and only if

i) SpanLA (D) = g, and

ii) D is not contained in a half space of g bounded by a subalgebra.

In order to have a more direct way to check controllability for connected and
simply connected solvable Lie groups, in [60] the author introduce the notion of
a completely solvable Lie algebra.

Theorem 34 Let be an invariant control system on a completely solvable Lie

group G. Then,
is controllable if and only if h = g.

23
 If g is a nilpotent Lie algebra, for any Z g the set Spec(ad(Z)) reduces to
zero. So, g is a completely solvable Lie algebra. In particular, Theorem 34 is a
perfect generalization of our result obtained in [3], 1 .

Example 35 Let us consider the Heisenberg Lie group of dimension three.

Therefore, the system

D = X 3 + u1 X 1 + u2 X 2 : u U .

is controllable on G. In fact, X 3 = X 1 , X 2 .

Example 36 Let G be the Lie group T (d) of all d d upper triangular ma-
trices with positive diagonal entries. T (d) is connected, simply connected and
completely solvable Lie group with Lie algebra t(d) consisting of all d d upper-
triangular matrices. By Theorem 34 any transitive invariant system is con-
trollable on G if and only if h = g.

Remark 37 We could continue with many relevant theorems on the controlla-

bility property for invariant systems = (G, D) for dierent classes of group.
However, in this subject there exists a very nice review by Sachkov, Y. [59].
Instead of that we concentrate on a special kind of controllability problem and
for more theoretical result and applications we send the readers to the references
given at the beginning of the chapter.

Denition 38 For a system = (G, D), a set A G is said to be Isochronous

if there exist T > 0 such that for any two arbitrary elements x, y A there exits
= Zt1 Zt2 ... Ztr S with ri=1 ti = T and (x) = y. In this case, T is
said to be an isochronal time to A. The system is said to be controllable at
uniform time if G is isochronous.

In [47] the authors proved

Theorem 39 Let be an invariant control system on a connected, compact

and semisimple Lie group G. If satisfy the Lie algebra rank condition, then
is controllable at uniform time.

In [15] we give an alternative proof. In fact, to prove it is enough to show the

existence of a time s+ such that the accessibility set from the identityA(e, s+ )
at exact time s+ coincides with G. Next we sketch the proof.
1 It was in 1993 when I have the opportunity to present Theorem 32 in the Conference

"Geometry in Nonlinear Control and Dierential Inclusions" in the Banach Center, Warsaw,
Poland.
Professor Andrei Agrachev attend the talk and ask me for the paper. After a year I re-
ceived a letter from Prof. Yuri Sachkov who send me many thanks because Theorem 34, a
generalization of Theorem 32 was included in his Ph D thesis. Then by letter I ask to him:
how do you like my paper. After one month the answer comes: what paper?
So, from the very beginning Andrei was sure that with the technique I used, Yuri will not
be able to generalizes the result. So, he decides to say nothing about the proof, just to talk
about the algebraic condition of the theorem.

24
 Let H be the normal Lie subgroup of G with Lie algebra given by the ideal

h = idealg Y 1 , Y 2 , ..., Y m

generated by the control vectors of the system. Since the group is compact, for
any positive time t
A(e, t) exp(tX)H.
Furthermore, related to the topology of this submanifold intA(e, t+ ) != . It
follows that for every positive time t, the right translation of H by exp(tX) is
contained in a submanifold of co-dimension 0 or 1, [44].
On the other hand, the Lie algebra is semisimple, thus g does not contains
ideals of co-dimension one, [71]. So, for any positive time t, exp(tX)H = G.
Therefore, we already proved the existence of a time s+ for such that starting
from e it is possible to reach any point of G at exact time s+ . On the other
hand, we consider the invariant transitive control system on G. Again, there
exists a positive time s such that through it is possible to reach e from any
point of the Lie group G at the exact time s . Take s = s+ + s , therefore
any two arbitrary points of G can be connected in exactly time s and is
controllable at uniform time. In a more general set up we obtain

Theorem 40 Let G a connected semisimple Lie group. Let = (G, D) be a

right invariant control system. If is controllable, then G can be covered by
a sequence of isochronous subsets Vn of G with nonvoid interior. Also if G is
compact G itself is an isochronous set.

Remark 41 If is controllable at uniform time it doesnt implies that you can

reach any point of G from e in arbitrary time. But, in some particular cases the
uniform time could be arbitrary. For instance, it happens in the homogeneous
case, i.e., when the system does not has drift vector eld, i. e., X = 0. For a
more general results in this direction see [50].

Example 42 Consider the Torus T 2 = S 1 S 1 and the invariant control system

= (T 2 , D) with
D = {X u = X + uY : u U } .

Here, X = ( x , 0) and Y = (0, y ). The non null component of X and Y are
invariant vector elds on the sphere S 1 . Since T is commutative, h has dimen-
sion 1. So, the subgroup H has co-dimension 1. Thus, for any positive time
t, the accessibly set E(t, e) at exact time t is contained in the one dimensional
submanifold exp(tX)H of T 2 . Therefore, in this situation, we can not expect
uniform controllability.

Example 43 From the general theory of Lie groups we know that any Abelian
Lie group G has the form
G = T d Rn
for some non negative integers numbers d, . Then, for an invariant system on
an Abelian Lie group we can not expect the uniform time property.

25
Remark 44 The proof of Theorem 34 strongly depends on the existence of a
state in the interior of an accessible set of and simultaneously. In
particular, the proof doesnt show that you can reach any point of G from e
in arbitrary time, see an example in our paper [15]. But, the uniform time
could be arbitrary. For instance, when the Lie algebra generated by the control
vectors
h = SpanLA Y 1 , Y 2 , ..., Y m
coincides with g. In fact, in this special case,

A(t, e) = G, for every t > 0

see [50].

Next, we use a relationship between Inv (G) with Bil (Rd ) to apply the
previous results for bilinear systems.

Example 45 Let us consider a bilinear control system

m
Bil (Rd ) : x(t)
= (A + uj (t)Bj ) x(t)
j =1

determined by the matrices A and Bj gl(d, R), j = 0, 1, ..., m, and u U. Let

G be the connected Lie group with Lie subalgebra

g = SpanLA {A, B1 , ..., Bm } gl(d, R).

Therefore, the control system

m

X(t) = (A + uj (t)Bj )X(t), X(t) G, u U
j=1

induced by the bilinear system is invariant on G. Assume that G is compact

and semisimple. Thus, the ane system
m

x(t) = (A + uj (t)Bj ) x(t), x(0) = x0
j=1

dened on the orbit

G(x0 ) = x Rd : x = gxo with g G

is controllable at uniform time on G(x0 ).

Example 46 We consider the bilinear control system on Rd

x(t)
= (A + u(t)B)x(t), u U

26
 where A0 and A1 generate the Lie algebra

g = SpanLA {A, B} = so(d, R)

of the skew symmetric matrices of order d. The associated compact Lie group
is G = SO(d, R). Recall that for each natural number d the group SO(d, R) is
simple, except the semisimple case SO(4, R). It turns out that the ane system
is controllable at uniform time on the manifold M = S d1 .

7 Linear control systems on Lie groups

A generalization of the notion of a linear system from the Euclidean space Rd
to a specic matrix group was given by [55]. After that, Ayala and Tirao in
[12] comes for a denitive denition of the subject on an abitrary connected
Lie group G. The authors involve the concept of normalizer of a Lie algebra g
(considered as the set of right invariant vector elds), in the Lie algebra X (G)
of all smooth vector elds on G, as follows

n = normX (G) (g) = {X : [X, Y ] g for every Y g} .

After characterize
n = g s aut(G)
as the semiproduct between g and the Lie algebra aut(G) of the Lie group of all
automorphisms Aut(G) of G, the authors introduce the following notion

Denition 47 A linear control system Lin (G) on G is determined by the data

m
x(t)
= X (x(t)) + uj (t)Y j (x(t)), x(t) G, u U
j=1

where, X is a linear vector eld, by denition X n with X (e) = 0. Further-

more, X j g and U denotes the admissible class of control function

U = L1loc (R, Rm )

with values on a closed and convex subset of Rm with 0 int().

Remark 48 Why this class of control is relevant? First, it contains the follow-
ing categories of systems on a Lie group G with Lie algebra g
m

x(t) = X(x(t)) + uj (t)Y j (x(t)) + Y, x(t) G, u U
j =1

1. Linear system on Euclidean spaces,

G = Rd , X = A aut(Rd ) = gl(d, R), Y j = bj g = Rd , Y = 0 Rd , [49]

27
 2. Ane systems on Euclidean spaces,

G = Rd , A aut(Rd ), Y j = bj g = Rd and Y Rd , [23]

3. Bilinear systems on Euclidean spaces,

G = Rd and A, B j aut(Rd ), [31]

4. Invariant systems on Lie groups,

X, X j g, [22]

And, by denition includes the linear systems on G. Just observe that

the drift and the control vectors of any system in the list are elements of the
normalizer. For instance,

in (2) : A = 0 + A g s aut(Rd ) and in (4) X = X + 0 g s aut(G).

Finally, we would like to introduce the notion of a normalizer system n

m

x(t) = X(x(t)) + uj (t)Y j (x(t)) + Y, x(t) G, X, Y j , Y n and u U.
j =1

Here, any dynamic of n is an element of the normalizer. In particular, all the

categories mentioned before are special cases. This control system is really a
very general algebraic system and it is a challenge even to start to work on it.
On the other side, in [43] the author shows that linear control systems on
Lie groups are also important for a dierent reason. Actually, he proved an
equivalence theorem as follows

Theorem 49 Let be a transitive ane control system

m

x(t) = X(x(t)) + uj (t)Y j (x(t)), u U
j =1

on a manifold M. Then, is dieomorphic to a linear system on a Lie group

or a homogeneous space if and only if the vector are complete and generate a
nite dimensional Lie algebra.

In particular, this results allows to apply all the theory of linear systems
on concrete applications. In fact, giving a system which satisfy the Jouans
condition, it is possible to obtain information of the initial system on a manifold
by looking at the linear one on the associated group, or on some homogeneous
space of the group.
The main aim of this section is to show some basic properties of Lin (G),
proving some results and/or illustrating through examples the fundamental the-
orems of the theory, [12],[13],[42] and [43] and also more recent results relatives

28
to the restricted case and the existence and uniqueness of control sets, [7] and
[8]. As we explain before, in the very well known paper by Jurdjevic and Suss-
man [47], the authors proved that the positive orbit of the identity element e
of G is a semigroup. Hence, for a connected Lie group controllability on G is
equivalent to the local controllability from e.
For linear systems the situation is totally dierent. In fact, we wiil show
here that if satises the ad-rank condition then it is locally controllable at
e, [12]. However, in [13], the authors show an example of a locally controllable
linear control system at e on the group G = SL(2) such that the positive orbit
of the identity is not a semigroup.
We start by showing a characterization of the drift vector eld X , see [42].
Let X be a vector eld on a connected Lie group G. The following conditions
are equivalent:

Theorem 50 1. X is a linear vector eld

2. X is an innitesimal automorphism
3. X satises

X (gh) = (dLg )h X (h) + (dRh )g X (g), for all g, h G. (6)

If we denote by (t )tR the ow associated to the drift X , by denition an

innitesimal automorphism is a vector eld such that

{t : t R} is a subgroup of Aut(G).

Certainly, the vector eld X is complete. Furthermore, one can associate to

X a derivation D of g dened by

DY = [X , Y ](e), for all Y g.

In fact, the Jacobi identity assures that

D[X, Y ] = [DX, Y ] + [X, DY ].

The relation between t and D is given by the formula, [72]

(dt )e = etD for all tR (7)

which implies that

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

On the other hand, if the group is simply connected any derivation D g

has an associated linear vector eld X = X D through the same formula above.
For connected Lie groups the same is true when D aut(G) the Lie algebra of
Aut(G) the Lie group of G-automorphism (see [12]). As a mater of fact,

aut(G) g and aut(G) = g G is simply connected.

29
 A particular class of such dynamics comes from inner automorphisms. More
precisely, consider an element W g. Since W is complete its ow

Wt (z) = expG (tW )z, z G

denes by conjugation a 1-parameter group of inner automorphisms on G

t (x) = Wt (e) x Wt (e), x G.

Therefore, t Aut(G) for any t R. In this case, the associated derivation

D : g g is again dened by D(Y ) = ad(X )(Y ) = [W, Y ] for any Y in g.

Example 51 Let us consider the solvable, but not complete solvable, group G =
E(2) of the plane Euclidean motions

1 0 0
G = x a b : (x, y) R2 and a2 + b2 = 1

y b a

with Lie algebra

0 0 0 0 0 0 0 0 0
g = Span Y 1 = 1 0 0 , Y 2 = 0 0 0 , Y 3 = 0 0 1 .

0 0 0 1 0 0 0 1 0

For an arbitrary z G and Y 1 g we obtain the linear vector eld

d
X (z) = exp(tY 1 )z exp(tY 1 )
dt t=0

1 0 0 0 0 0
d x + t at a
= b = 1 a 0 0 .
dt t=0 y + bt b a b 0 0

Remark 52 The inner derivations is a subclass which is far from determining

all the elements in g. Actually, there could exist a big dierence of cardinality
between derivations and inner derivations. In fact, for G = Rd any real matrix
of order d is a derivation, so dim Rd = d2 . However for a semisimply Lie
group G any derivation is inner, which means that dim g = d. To show a case
between these extremum, we give he following example

Example 53 Consider the simply connected Heisenberg Lie group H of dimen-

sion three
1 x z
G = 0 1 y : x, y, z R

0 0 1

30
 with Lie Algebra g generated by the basic elements

0 1 0 0 0 0 0 0 1
X = 0 0 0 , Y = 0 0 1 , Z = 0 0 0 .
0 0 0 0 0 0 0 0 0

As a matter of fact [X, Y ] = Z and the other brackets vanish. As invariant

vector elds in coordinates we get the representation

X= , Y = +x , Z= .
x y z z

Any g-derivation D written in the basis (X, Y, Z) has the form

a b 0
D = c d 0 .
e f a+d

In this case dim(h) = 6. In the other direction, there are just two independent
inner derivations: ad(X) and ad(Y ). For a derivation D the associated linear
vector eld X = X D is explicitly given in [42] by:

1 1
X (g) = (ax + by) + (cx + dy) + (ex + f y + (a + d)z + cx2 + by2 )
x y 2 2 z
which is determined by 6 independent parameters.

7.1 Local controllability

In this section we show some controllability results, some of them without proofs.
We follow the reference [12] and we start by the natural transitivity property.
Let Lin (G) be a linear control system. By the Sussmann orbit theorem,
Lin (G) is transitive if and only if the Lie algebra rank condition is satised.
i.e.,

dim SpanLA {Y j , adi (X )(Y j ) | 1 j m and 0 i dim(G)} = dim(g).

Denotes by h = Span Y 1 , ..., Y m , the Lie algebra generated by the control

vectors and by < X | h > the smallest ad(X )-invariant Lie subalgebra containing
h. Just observe that the larc condition is not equivalent to < X | h > = g. But,
for Abelian groups they agree.
For the classical unrestricted linear system the algebraic object associated to
controllability is a subspace, (in particular a semigroup) dened by the smallest
A-invariant subspace of Rd containing the subspace generated by the columns
vectors . And we know that, controllability is equivalent to

dim Span{b1 , ..., bm , Ab1 , ..., Abm , ..., Ad1 b1 , ..., Ad1 bm } = n

31
In fact, Y j = bj , ad1 (X )(Y j ) = A, bj = Abj , ad2 (X )(Y j ) = A, A, bj =
A2 bj , etc, for any bj Rm , 1 j m.
As usual, we assume larc for Lin (G). In order to go further, we need to
consider a more strong condition.

Denition 54 The ad-rank condition is determined by the requirement

dim Span{Y j , adi (X )(Y j ) : 1 j m and 0 i n 1} = dim(g).

We notice that in order to reach the dimension of G the brackets

Y j , adi (X )(Y j ) with 1 j m and 0 i n 1

are forbidden. With this extra assumption we obtain

Theorem 55 Let Lin (G) be a transitive linear control system. If satises

the ad-rank condition, then the system is locally controllable from the identity
element e.

Proof. We follow [12]. Assume

dim Span{Y j , adi (X )(Y j ) : 1 j m and 0 i n 1} = dim(g).

We denote by x(t, u), the solution of

m
.
x = X (x) + uj Y j (x), x G
j=1

through the identity element associated to the control u U. For every non-
negative t consider the innitely dierential endpoint map

Et : u U Et (u) = x(t, u) G.
In a neighborhood B of the control u 0 its dierential d(Et )0 at the control
zero is given by
t m
d(Et )0 (u()) = e(ts)ad(X ) ( uj (s) Yej ) ds, [1]
0 j=1

Suppose this linear map is not surjective. There exists a co-vector in the dual
space Te G
= g of the tangent space Te G
= g, such that

< , d(Et )0 (u()) > = 0 for every u() B.

By the bilinearity property of < , > we get

32
 t m t m
(ts)ad(X )
< , e ( uj (s) Yej ) ds > = < , e(ts)ad(X ) (Yej ) > uj (s) ds.
0 j=1 0 j=1

Since this expression is true for every piecewise constant function u : [0, t] Rm
we can conclude that

< , e(ts)ad(X ) (Yej ) > = 0, s [0, t].

Dierentiating the last expression with respect to t at t = 0 we obtain

< , adi (X )(Yej ) > = 0 for each i 0 and j = 1, 2, ..., m.

which is in contradiction with the ad-rank condition assumption.
Hence, the linear map d(Et )0 is surjective. By the implicit function theorem
the map d(Et )0 is locally onto on G. That is, there exists a neighborhood V of
the identity element e in G such that d(Et )0 is onto on V G. As a consequence,
Lin (G) is locally controllable at e.

Next, we show an example of not local controllable transitive linear system.

Example 56 On the Heisenberg Lie group H of dimension 3 we consider a
system Lin (G) dened by

.
x(t) = X (x(t)) + uY 2 (x(t)), u R
where the drift X gives rise to the derivation

0 1 0
D = 1 0 0 g.
0 0 0
In coordinates the system reads
.
x1 (t) = x2 (t) + 12 x3 (t) + 12 x1 (t) 14 x2 (t)x3 (t)
.
: x2 (t) = x1 (t) 12 x2 (t)x3 (t) + u , uR
.
x3 (t) = 0
A computation shows that [X , Y 1 ] = Y 2 and [X , Y 2 ] = Y 1 . So, we have
ad(X )( Y 1 ) = Span{Y 1 , Y 2 } g
=

while the Lie algebra rank condition is satised. i.e.,

SpanLA {Y 1 , Y 2 } = Span{Y 1 , Y 2 , Y 3 } = g.
The system is not locally controllable although it is transitive. Geometrically
speaking, no one integral curve of the system can leave the x1 x2 -plane.
In the sequel, we establish some controllability results according to the un-
restricted and restricted admissible class of control under consideration. And,
also depending of the type of the group.

33
7.2 Controllability: the unrestricted case
In general, controllability property is a very exceptional issue. We can not
expect something dierent for the class in study. However, like in the classical
linear systems on Euclidean spaces, the same Kalman rank condition determines
controllability for any unrestricted linear systems on an Abelian or on a compact
semisimple Lie group.

Theorem 57 Let Lin (G) be linear control system. Assume

1. the group G is Abelian, then

Lin (G) is controllable Lin (G) is transitive

2. the Lie algebra g of G is semisimple and G is compact, then

Lin (G) is controllable Lin (G) is transitive

Proof. To prove the rst claim, we observe that in this very particular case
the positive orbit of the identity A(e) is a semigroup. On the other hand, since
the system is transitive and the group is Abelian, the rank condition and the
ad-rank condition coincides, which implies local controllability from e on some
neighborhood Ve . Since G is connected and A(e) is a semigroup we obtain

Ve A(e) G = nN (Ve )n A(e)

and Lin (G) is controllable from e. The same argument is possible to apply on
the negative system Lin (G) which is exactly Lin (G) but with the drift X .
Thus, Lin (G) is also controllable from the identity. So, for any element x G
there exists a control u and a positive time tu such that x can be reached from
e through Lin (G) in tu units of time. Equivalently, x can be transferred to
e from Lin (G) with the same control and with the same time. Now, let us
consider an arbitrary element y G. There are controls u and v such that y
can be reached from x. In fact, after to reach the identity from x we continue
with a control v transferring e to y at tv units of time. So, y can be reached by
x at tu + tv units of time. Therefore, the system is controllable from x for each
x G.
We just show the main ideas of the proof of 2. Since G is compact, we note
that the Haar measure is nite. On the other hand, since the group is semisimple
any derivation is inner, which implies that the linear vector eld has the form
t (x) = exp(tW ) x exp(tW ), for some W g and any x G. In this case,
for any constant control the dynamic of the system combine conjugation with
invariant, both preserving the Haar measure. Since G is compact, it is closed.
So for more arbitrary control the argument works for the limit. Furthermore,
according to a well know fact, the system is controllable if and only if satisfy the
larc condition. Actually, the invariance argument together with the Poincar
recurrence theorem, nish the proof, [54].

34
Remark 58 Next we show the rst example in the literature of a non control-
lable system which is locally controllable from the identity. Since the group is
connected it turns out that

"for a linear system on groups the reachable set from the identity A(e) can
not be a semigroup"

Example 59 On G = SL(2, R) consider the derivation D = ad(W ) where

1 0
W = and the linear vector eld X associated to D. Dene the linear
0 1
control system Lin (SL(2, R)) by
.
x = X (x) + uY (x), x G

1 1
where Y = . A simple computation shows that
1 1

Span {Y, [X , Y ] , [X , [X , Y ]]} = g.

Therefore, the system is local controllable at e. For an argument of reversibility

(that scape from the scope of the paper), the authors show that the system can
not be controllable.

As we saw, in general the reachable set is not a semigroup. Actually, under

the larc condition Jouan prove the following result

Theorem 60 Let Lin (G) be a transitive linear control system on G. Hence,

A(e) is a semigroup A(e) = G.

In this section we follow reference [7]. The next Proposition state the main
properties of the reachable sets.

Proposition 61 It holds:

1. 0 1 2 A 1 A 2
2. A (g) = A (g), for any g G
3. , 0
a) A + = A (A ) = A (A ), and inductively
b) A 1 1 (A 2 ) 1 + 2 (A 3 ) n1
i (A n ) =A n
i=1 i , for any positive
i=1
real numbers 1 , . . . , n
4. u U, g G and t 0 t,u (A(g)) A(g)
5. e int A A is open.

35
 The proof of items 1. to 3. can be found in [42], Proposition 2. The items
4. and 5. in [28] Proposition 2.13.
Remark 62 We notice that the item 4. of the above proposition together with
the fact that 0 int shows us in particular that A is invariant by the ow t
in positive time, that is, t (A) A for any t 0.
The positive orbit A of a linear control system Lin (G) is in general not a
semigroup, in this section we associate to = Lin (G) a new algebraic object
S which turns out to be a semigroup. In particular, S enable us to pass from
the control theory of linear systems to the theory of semigroups. Furthermore,
controllability of is equivalent to S = G
As before, we denote by (t )tR the 1-parameter group of automorphisms
associated to X . We dene in [8]
S = tR t (A)
Since t (e) = e for all t R and e A it follows that S != .
Proposition 63 With the previous notations it holds
1. S is the greatest -invariant subset of A
2. For any 0 0
S = t0 t (A)
3. x S if and only if t (x) A for all t 0
4. S is a semigroup
Proof.
1. We start by proving the -invariance of S . Let R, then
(S ) = (tR ) = tR (t (A)) = tR +t (A) = S
Now, let C be a -invariant subset of A. It holds that
C = t (C) t (A), for all t R C tR = S
showing that S is the greatest -invariant subset of A.
2. By the -invariance of A in positive time, we get
0 t 0 0 t (A) A 0 (A) t (A)
and consequently
0 (A) = tt0 t (A).
Therefore,
S = t>t0 t (A) tt0 t (A) = t>t0 t (A) 0 (A) = tt0 t (A)
as desired.

36
 3. We have

x S x t (A) for all t 0 t (x) A for all t 0.

4. Let x, y S with xt = t (x) and yt = t (y). From 3., we just need

to show that

xt yt = t (x)t (y) = t (xy) A for any t 0.

By hypothesis xt A, st > 0 : xt Ast . But y S so st (yt ) =

st t (y) A and st > 0 with st (yt ) Ast showing that S is a
semigroup. In fact,

xt yt = xt st (st (yt )) Ast st (Ast ) = Ast +st A.

Denition 64 S is called the semigroup of the system = Lin (G).

In order to study the controllability property of it is enough to study the

semigroup S .

Theorem 65 A = G if and only if S = A.

Proof. If A = G then t (A) = t (G) = G for all t R. Hence,

G = t>t0 t (A) = S .

Conversely, if S = A then A is a semigroup. However, Proposition 7 of [42]

assures that A = G. Then, the desired conclusion follows.

Corollary 66 A = G A is -invariant

Proof. If A = G, A is certainly -invariant. Conversely, if t (A) = A for any

t R we get A S which implies S = A. So, A = G.
In [28] the authors prove a general controllability result for linear systems

Theorem 67 Assume the reachable set A is open. Then G+,0 S . Moreover,

is controllable G S .

Based in this theorem, in [28] Da Silva extends Corollary 67 at follows

Theorem 68 Let Lin (G) a linear control systems on G. Therefore,

1. If G is solvable

e int(A) and Spec(A)Ly R = {0} Lin (G) is controllable.

37
 2. If G is nilpotent
e int(A) and Spec(A)Ly R = {0} Lin (G) is controllable.

In order to extend the previous theorem, in [7] the authors introduce the
following notion
Denition 69 Let G be a connected Lie group. We say that the Lie group G
has nite semisimple center if all semisimple Lie subgroups of G have nite
center.
We notice that several classes of Lie groups satisfy Denition 69. For ex-
ample, any Abelian, nilpotent and solvable Lie group has the nite semisimple
center property. Furthermore, any semisimple Lie group with nite center and
any direct or semidirect product of these classes of groups have semisimple nite
center.
Theorem 70 Let G be a connected Lie group with nite semisimple center.
Hence,
e int A 0 and Spec(A)Ly R = {0} Lin (G) is controllable.

7.3 Controllability: the restricted case

In this section we follow are reference [9]. Considering the controllability be-
havior of an unrestricted linear system Lin (G), we now approach the problem
in a more realistic way. We go back to the notion of a control set, a maximal
subset C where approximately controllability holds, i.e.,
x int(C) int(C) A(x) and for x C, xn int(C) : xn x.

Like in the classical linear system it is possible to characterize the existence

of a control set with nonemtpy interior around the identity. Of course, many
topological properties of C are intrinsically connected with the eigenvalues of
the associated derivation D to the drift X .
The following decomposition is crucial in the study of controllability of re-
stricted linear systems. It was used for the rst time by Da Silva in [28]. For
a derivation D : g g there exists a special decomposition through its Lya-
punov exponents, [64]. For any eigenvalue of D there exists the -generalized
eigenspace determined by
g = {X g : (D )n X = 0 for some n 1}.
It turns out that for if is also an eigenvalue of D then
[g , g ] g+ when + is an eigenvalue of D

and zero otherwise. Hence, g decomposes as

g = g+ g0 g .

38
The subspaces g+ , g0 and g dened by

g+ = g , g0 = g and g = g .
: Re()>0 : Re()=0 : Re()<0

are Lie algebras. And, g+ , g are nilpotent. Denote by G+ , G0 , G , G+,0 and

G,0 the connected Lie subgroups of G with Lie algebras g+ , g0 , g , g+,0 and
g,0 respectively. By Proposition 2.9 of [28] the subsets G+ , G0 , G , G+,0 and
G,0 are (t )tR -invariant closed subgroups of G.

Theorem 71 Let Lin (G) be a linear system where G is a solvable or G0 G

is compact. Then,

1. The only control set with non empty interior is given by

C = cl(A) A

2. Furthermore,

C is bounded G0 , cl(AG ) and cl(AG+ ) are compact sets

where
AG = AG and AG+ = A G+ .

3. For a nilpotent Lie group G more is true

C is bounded cl(AG ) and cl(AG+ ) are compacts and D is hyperbolic.

References
[1] A. Agrachev, R. V. Gamkrelidze and A. V. Sarychev, Local Invariants
of Smooth Control Systems, Acta Applicandae Mathematicae 14, 191-237,
1989.
[2] V. Ayala and P. Jouan, Almost Remannian Geometry on Lie Groups. SIAM
Journal on Control and Optimzation, Vol. 54, No. 5, pp. 29192947.
[3] V. Ayala, Controllability of Nilpotent Systems, Banach Center Publications,
Polish Academy of Sciences, Vol. 32, pp. 35-46, 1995.
[4] V. Ayala, F. Colonius and W. Kliemann, Dynamical characterization of
the Lyapunov form of matrices, Linear Algebra and its Applications 402
(2005) 272290.
[5] V. Ayala, E. Cruz, L. Laura and W. Kliemann, Controllability Properties of
Bilinear Systems in Dimension 2, Journal of Mathematics and Computer
Science. In press, 2016.

39
 [6] V. Ayala, F. Colonius and W. Kliemann, On Topological Equivalence of
Linear Flows with Applications to Bilinear Control Systems, Journal of
Dynamical and Control Systems, Vol. 13, No. 3, July 2007, 337362.
[7] V. Ayala and A. Da Silva, Controllability of Linear Control Systems on
Lie Groups with Semisimple Finite Center. In Press at SIAM Journal on
Control and Optimization, 2016.
[8] V. Ayala and A. Da Silva, A Semigroup associated to a linear
control system on a Lie group. Systems & Control Letters, DOI:
10.1016/j.sysconle.2016.09.021, 2016.
[9] V. Ayala and A. Da Silva, Control sets of linear systems on Lie groups.
Nonlinear Dierential Equations and Applications NoDea. Submitted in
2016.
[10] V. Ayala and C. Kawan, Topological Conjugacy of Real Projective Flows.
Journal of theLondon Mathematical Society, (2), 90, (2014), pp. 49-66.
[11] V. Ayala and E.Kizil, The covering semigroup of invariant control systems
on Lie groups, In press at Kybernetika.
[12] V. Ayala, J. Tirao, Linear Control Systems on Lie Groups and Controlla-
bility. American Mathematical Society, Series: Symposia in Pure Mathe-
matics, Vol. 64, (1999), 4764.
[13] V. Ayala and L. San Martin, Controllability Properties of a Class of Control
Systems on Lie Groups. Lectures Notes in Control and Information Science,
Vol. 1, N 258, (2001), 8392.
[14] V. Ayala, L. San Martin and R. Ribeiro, Controllability on Sl(2,C) with
restricted controls. SIAM Journal on Control and Optimization, Vol 52, n
4, pp. 2548-2567, 2014.
[15] V. Ayala, W. Kliemann and F. Vera, Isochronous sets of invariant control
systems on Lie groups. Systems & Control Letters 60 (2011) 937942.
[16] V. Ayala and L. San Martin, Controllability of two-dimensional bilinear sys-
tems: restricted controls and discrete time, Proyecciones Journal of Math-
ematics 18 (1994), 207-223.
[17] V. Ayala and L. Vergara, Co-adjoint representation and controllability.
Proyecciones Math Journal of Mathematics, 11, (1992), 37-48.
[18] V. Ayala and R. Zegarra, Controllability of linear systems on group of
matrices, Cubo Journal of Mathematics, 2001, Vol 3, N 2, 171-212.
[19] C. J. Bragas Barros, J. Ribeiro Goncalves, O. do Roco and L. A. B. San
Martin, Controllability of two-dimensional bilinear system. Rev Proyec-
ciones 15, pp. 111-139, 1996.

40
[20] B. Bonnard, V. Jurdjevic, I. Kupka and G. Sallet, Transitivity of families
of invariant vector elds on semi-direct product of Lie groups, Trans. Amer.
Math. Soc. 271: 521-535.
[21] B. Bonnard, V. Jurdjevic, I. Kupka and G. Sallet, Controllabilite sur le
produit semi-direct dun groupe compact par un E.V. reel, Trans. Amer.
Math. Soc., 1982.
[22] R. Brockett, System theory on groups and coset spaces. SIAM J. Control.
10:265-284.
[23] F. Colonius and W. Kliemann, The Dynamics of Control, Birkhuser,
Boston, 2000.
[24] F. Colonius and W. Kliemann, Morse decompositions and spectra on ag
bundles, J. Dynam. Di. Equations, 14 (2002), pp. 719741.
[25] F. Colonius and M. Spadini, Uniqueness of local control sets, J. Dynamical
and Control Systems, 9 (2003), pp. 513 530.
[26] M. L. Curtis, Matrix Groups, Springer Verlag, 1979.
[27] D. DAlessandro, Small time controllability of systems on compact Lie
groups and spin angular momentum. Journal of Mathematical Physics, Vol
42, n 9, (2001).
[28] A. Da Silva, Controllability of linear systems on solvable Lie groups, SIAM
J. Control Optim., 54(1), 372390
[29] A. I. Dos Santos and L. A. B. San Martin, Controllability of control systems
on complex simple lie groups and the topology of ag manifolds, Journal of
Dynamic and Control Systems, April 2013, Volume 19, Issues 2, pp.157-171
[30] L. Dubins, On curves of minimal lengths with a constrains on average cur-
vature and with prescribed initial and terminal positions and tangents. Am.
J. of Mathematics.
[31] D. L. Elliott, Bilinear Control Systems: Matrices in Action, Springer 2009.
[32] T. Gayer, Control sets and their boundaries under parameter variation, J.
Di. Equations, 201 (2004), pp. 177200.
[33] J. P. Gauthier, I. Kupka and G. Sallet, Controllability of right invariant
systems on real simple Lie groups. Systems Control Lett. 5 (1984), 187190.
[34] J. Guckenheimer and P. Holmes, Nonlinear Oscillations, Dynamical Sys-
tems and Bifurcation of Vector Fields, Springer-Verlag, Berlin, 1986.
[35] W. Hahn, Stability of Motion, Springer-Verlag, 1967.
[36] Hartman, Philip (August 1960). "A lemma in the theory of structural sta-
bility of dierential equations". Proc. A.M.S. 11 (4): 610620

41
[37] S. Helgason, Dierential Geometry, Lie groups and Symmetric Spaces. Aca-
demic Press, New York 1978.
[38] J. Hilgert, K. Hofmann, J. Lawson, Controllability of systems on a nilpotent
Lie group, Beitrage Algebra Geometrie, 20, 185-190 (1985).
[39] D. Hinrichsen and A. Pritchard, Mathematical Systems Theory, Springer-
Verlag, 2005.
[40] A. Isidori, Nonlinear Control Systems, Springer-Verlag, 1998.
[41] I. Jo and N. M. Tuan, On controllability on bilinear systems II, (con-
trollability on two dimension). Ann. Univ. Sci. Budapest 35, pp. 217-265,
1992.
[42] Jouan, Ph., Controllability of Linear Systems on Lie group, Journal of
Dynamics and Control Systems, 17 (2011) 591-616.
[43] Ph. Jouan, Equivalence of Control Systems with Linear Systems on Lie
Groups and Homogeneous Spaces, ESAIM: Control Optimization and Cal-
culus of Variations, 16 (2010) 956-973
[44] V. Jurdjevic and H. J. Sussmann, Controllability of Nonlinear Systems,
Journal of Dierential Equations 12, (1972), 95116.
[45] V. Jurdjevic, Geometric control theory, Cambridge University Press, 1997.
[46] V. Jurdjevic, Optimal control problem on Lie groups: crossroads between
geometry and mechanics. In geometry of feedback and optimal control. Ed.
B. Jakubczyk and W. Respondek. New York: Marcel Dekker.
[47] V. Jurdjevic and H. Sussmann, Control systems on Lie groups, J. Di.
Equat, 12, 313-329 (1972).
[48] V. Jurdjevic and I. Kupka, Control systems on semi-simple Lie groups and
their homogeneous spaces, Ann. Inst. Fourier, Grenoble 31, No. 4, 151-179
(1981).
[49] R. Kalman, Y. Ho and K. Narendra, "Controllability of Linear Dynamical
Systems, Contrib. to Di. Equations 1 (2), pp. 189-213, 1962.
[50] H. Kunita, Support of diusion processes and controllability problems. Pro-
ceedings International Symposium on Stochastic Dierential Equations,
edited by K. Ito, Wiley, New York, (1978), 163185.
[51] U. Ledzewick and H. Shattler, Optimal controls for a two compartment
model for cancer chemotherapy with quadratic objective. Proceedings of
MTNS, 2006, Kyoto, Japan.
[52] F. Silva Leite, Uniform controllable sets of left-invariant vector elds on
compact Lie groups. Systems and Control Letters 7 (1986), 213216.

42
[53] F. Silva Leite, Uniform controllable sets of left-invariant vector elds on
non compact Lie groups. Systems and Control Letters 6 (1986), 329335.
[54] C. Lobry, Controlabilit des systemes non lineaires, SIAM J. Control and
Optim. 8: 573-605.
[55] L. Markus, Controllability of Multi-trajectories on Lie Groups, Proceedings
of Dynamical Systems and Turbulence, Lecture Notes in Mathematics 898,
pp. 250-265, Warwick, 1980.
[56] D. Mittenhuber, Controllability of systems on solvable Lie groups: the
generic case, J. Dynam. Control Systems, 7, No. 1, 61-75 (2001).
[57] R. Mohler, Bilinear control processes, Mathematics in Science and Engi-
neering, Vol. 106.
[58] L. S. Pontryagin, V. G. Boltyanskii, R. V. Gamkrelidze and E. F.
Mishchenko. (1962). The Mathematical Theory of Optimal Processes. Eng-
lish translation. Interscience. ISBN 2-88124-077-1.
[59] Y. Sachkov, Control Theory on Lie Groups. Lecture Notes SISSA, 2006.
[60] Y. Sachkov, Controllability of right-invariant systems on solvable Lie
groups, J. Dyn. Control Syst., 3, No. 4, 531-564 (1997).
[61] Y. Sachkov, Controllability of Invariant Systems on Lie Groups and Ho-
mogeneous Spaces, in: Progress in Science and Technology, Series on Con-
temporary Mathematics and Applications, Thematic Surveys, Vol. 59.
[62] Y. Sachkov, Survey on Controllability of Invariant Control Systems on
Solvable Lie Groups, AMS Proceedings of Symposia in Pure Mathemat-
ics, Vol.64, 297-317, 1999.
[63] L. A. B. San Martin and P. Tonelli, Semigroup actions on homogeneous
spaces, Semigroup Forum, 14, 1-30 (1994).
[64] L. A. B. San Martin, Algebras de Lie. Editorial UNICAMP, Campinas, SP,
1999.
[65] L. A. B. San Martin and P. Crouh, Controllability on principal bre bundles
with compact structure group, Systems and Control Letters, 5 (1984), 35-
40.
[66] L. A. B. San Martin, Invariant control sets on ag manifolds, Math. Control
Signals Systems, 6, 41-61, 1993
[67] J. Selgrade, Isolated invariant sets for ows on vector bundles, Trans. Amer.
Math. Soc., 203 (1975), pp. 259390..
[68] E. D. Sontag, Mathematical Control Theory, Springer-Verlag, 1998. Second
Edition (First Edition, 1990).

43
[69] H. J. Sussmann, Orbits of families of vector elds and integrability of dis-
tributions, Trans. Amer. Math. Soc., 180 (1973), pp.171-188.
[70] H. Sussmann and C. Willems, 300 Years of Optimal Control: From the
Brachystochrone to the Maximum Principle, IEEE Control Systems, His-
torical Perspectives, 1997.
[71] V. Varadarajan, Lie Groups, Lie Algebras and their Representations, Pren-
tice Hall, Inc., 1974.
[72] F. W. Warner, Foundations of Dierential Manifolds and Lie Groups. Scott
Foreman, Glenview, 1971.
[73] S. Wiggins, Introduction to Applied Nonlinear Dynamical Systems and Ap-
plications, Springer-Verlag, 1996.

44