You are on page 1of 43

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/225200046

Semigroup actions on homogeneous spaces

Article  in  Semigroup Forum · December 1995


DOI: 10.1007/BF02573505

CITATIONS READS

83 70

2 authors:

Luiz A. B. Sanmartin Pedro Aladar Tonelli


University of Campinas University of São Paulo
94 PUBLICATIONS   689 CITATIONS    22 PUBLICATIONS   493 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Control Theory View project

All content following this page was uploaded by Luiz A. B. Sanmartin on 19 October 2014.

The user has requested enhancement of the downloaded file.


Semigroup Actions on Homogeneous Spaces
Luiz A. B. San Martin∗
Instituto de Matemática
Universidade Estadual de Campinas
Cx.Postal 6065
13081-970 Campinas SP, Brasil
Pedro A. Tonelli
Instituto de Matemática e Estatı́stica
Universidade de São Paulo
Cx. Postal 20570
01452-990 - São Paulo, S.P, Brasil
November 19, 2005

Abstract
Let G be a connected semi-simple Lie group with finite center and
S ⊂ G a subsemigroup with interior points. Let G/L be a homoge-
neous space. There is a natural action of S on G/L. The relation
x ≤ y if y ∈ Sx, x, y ∈ G/L, is transitive but not reflexive nor sym-
metric. Roughly, a control set is a subset D ⊂ G/L, inside of which
reflexivity and symmetry for ≤ hold. Control sets are studied in G/L
when L is the minimal parabolic subgroup. They are characterized
by means of the Weyl chambers in G meeting int S. Thus, for each
w ∈ W , the Weyl group of G, there is a control set Dw . D1 is the only
invariant control set, and the subset W (S) = {w : Dw = D1 } turns
out to be a subgroup. The control sets are determined by W (S) \ W .

Research partially supported by CNPq grant n◦ 301060/94 − 0

1
The following consequences are derived: i) S = G if S is transitive on
G/H, i.e. Sx = G/H for all x ∈ G/H. Here H is a non discrete closed
subgroup different from G and G is simple. ii) S is neither left nor
right reversible if S 6= G. iii) S is maximal only if it is the semigroup
of compressions of a subset of some minimal flag manifold.

subject classification 22E15, 22E46


keywords: Semigroups, semi-simple Lie groups, flag manifolds, control
sets

1 Introduction
This paper deals with subsemigroups of semi-simple Lie groups and their
actions on certain homogeneous spaces of the Lie groups, the flag manifolds.
The main issue here are the so called control sets for these actions. It is
shown that a clarification of some of the algebraic and geometric aspects of
these sets reveals many global properties of the semigroups themselves.
One of the distinguishing features of semigroup and group actions comes
from the fact that the non existence of inverses in the semigroup turns their
orbits into non return routes. This characteristic of the semigroup orbits
makes them much harder to analyze than the group orbits. For instance,
suppose that S is a semigroup of diffeomorphisms of a differentiable manifold
M . In case S is a group, reasonable assumptions ensure that their orbits are
well behaved submanifolds of M (see e.g. [20]). The proof of this makes heavy
use of the fact that the relation x ≤ y defined by y ∈ Sx = {ϕ(x) : ϕ ∈ S}
is an equivalence relation for group actions. This is not to be expected for
semigroup actions. For these the relation ≤ despite being transitive is not in
general reflexive nor symmetric.
Roughly, a control set for the S-action on M is a subset D ⊂ M inside
of which the relation ≤ is an equivalence relation (see the precise definition
below). This concept was introduced by L. Arnold and W. Kliemann [1]
with the aim of describing dynamical and ergodic properties of the diffusions
generated by stochastic differential equations ([1] considers, more properly,
invariant control sets). To a stochastic equation one associates, quite nat-
urally, a semigroup (control semigroup, c.f. [2]). The invariant control sets
for this semigroup appear then as the supports of the invariant measures for
the diffusion, making it possible to detect the ergodic components by means

2
of the invariant control sets. Through ergodic theorems invariant measures
play an important role in the asymptotic analysis of the diffusion. One of
these ergodic theorems is Oseledec’s Multiplicative Ergodic Theorem, which
detects the Lyapunov exponents of random dynamical systems. In order to
look at the Lyapunov exponents of a stochastic equation, one is led to in-
vestigate invariant control sets for subsemigroups of the general linear group
in their actions on projective space, Grassmannians, and more generally on
flag manifolds (see [2, 16], and references therein). ¿From the way this ques-
tion comes up, the assumption that the semigroup is a subsemigroup with
nonempty interior in a Lie subgroup G of Gl(n, IR), is almost an axiom.
Therefore the policy is to start by looking firstly at G, and distinguish it
according to the different classes it belongs to. At this regard, Levi type
decompositions permit the reduction to semi-simple Lie groups, making the
analysis for this class of groups the core for the question. Actually, there are
two complementary problems. First one should work intrinsically by looking
at invariant control sets of semigroups on homogeneous spaces of G, then
represent them as orbits on the flag manifolds. The intrinsic side of this
problem parallels the works of Guivarc’h and Raugi [8] on random products
on semi-simple Lie groups. These works suggest that, regarding semi-simple
Lie groups, the “correct” homogeneous spaces to look at are the generalized
flag manifolds, the Furstenberg boundaries. This way, we arrive at the sub-
ject matter of this paper which is the action of the subsemigroups with non
empty interior of semi-simple Lie groups on the Furstenberg boundaries of
the groups. The invariant control sets (in fact the only one in each flag man-
ifold) were studied in [17]. Here the same type of methods as those employed
in [17] are used to study control sets in general.
There is a series of recent papers by F. Colonius and W. Kliemann (see
[4, 5, 6], and references therein) relating the control sets to the dynamics of
the shift associated to a control system. For instance, Lyapunov exponents,
sets of recurrence, chaos, etc. for the shift are looked through the glasses of
the control sets.
These results suggest a main feature of the control sets, namely that they
are composed of minimal sets for the diffeomorphisms inside the semigroup
in question (the control semigroup). This aspect of the control sets is al-
ready encountered in [17] for the invariant control sets on the flag manifolds:
The (only) invariant control set is the closure of the set of fixed points of a
certain kind for a certain type of regular elements of the group which can be

3
found inside the semigroup (see Theorem 3.1 below). The regular elements
in question are those which belong to the split component A in some Iwa-
sawa decomposition G = KAN , that is, those regular elements whose adjoint
representation is diagonalizable. Such elements have special fixed points on
the flag manifolds, namely those which have a stable manifold which is open
and dense (Bruhat decomposition). The set of these attracting fixed points
constitute the invariant control set.
In this paper a similar characterization for the other control sets is pro-
vided. For example, in the maximal flag manifold, the above regular elements
have as many fixed points as the Weyl group W of G has elements. Each
element w ∈ W gives rise to a fixed point of the type w (e. g. the identity
of W corresponds to the fixed point with an open and dense stable mani-
fold). Our first job will be to show that fixed points of the same type for the
regular elements inside the semigroup are on the same control set, so that
for each w ∈ W there is defined in a natural way a control set Dw on the
maximal flag manifold. The map w → Dw is not necessarily one-to-one. It is
shown however that there exists a subgroup, say W (S), of the Weyl group so
that the control sets become parametrized by the coset space W (S) \ W . In
particular the elements of W (S) are associated to the invariant control set.
The subgroup W (S) turns out to become an effective tool for further
analysis of the semigroup. In fact, W (S) can also be seen as the subgroup
which fixes the elements with lowest degree of regularity, contained in some
split subgroup, which can be found inside the semigroup. As such, W (S)
is generated by the reflexions defined by some subset of the simple system
of roots, that is to say, it is canonically associated to a parabolic subgroup
of G. From this, many consequences about the semigroup can be drawn by
just looking at the action of W (S) on coset spaces of W . As an example we
mention the fact that proper semigroups are not transitive on large classes of
homogeneous spaces. This requires that the group has finite center because
an argument involving a fibration with compact fiber is needed. Another
application concerns maximal semigroups. It is an easy task to show that
W (S) is maximal if the semigroup is maximal. From this one gets that the
maximal semigroups are those which leave some subset of one of the minimal
boundaries invariant. There are thus a finite number of classes of maximal
semigroups each one associated to a minimal boundary. The classification
of the maximal semigroups of Lie groups is one of the major problems in
the theory of semigroups (see Lawson [13] for comments on this line). The

4
results presented here are not conclusive about this because the subsets of
the minimal boundaries which give rise to maximal subsemigroups are not
arbitrary. An understanding of these sets is still open. Two specific cases
are presented in Theorems 6.10 and 6.11. These cases cover the rank one Lie
groups and the maximal semigroups in Sl(n, IR) associated to the projective
space.
Most of the ideas here come from control theory, whose links with the
theory of semigroups have been extensively treated elsewhere (see for instance
[3, 12]). Some questions in control theory led in a natural way to global
problems in Semigroup theory (see [10, 14]). Maximality and transitivity,
which are treated here fit properly in this context.

2 Control Sets
Let S be a subsemigroup of a Lie group G and M a space provided with
a G-action. Although many of the statements appearing below are valid in
general, for our purposes here we assume that the G-action on M is transitive.
Thus we think of M as a homogeneous space of G where S acts as a semigroup
of diffeomorphisms. Also, we assume throughout that S has interior points
in G.

Definition 2.1 A control set for the S-action on M is a subset D ⊂ M


satisfying
i) int D 6= ∅,
ii) D ⊂ cl(Sx) for all x ∈ D, and
iii) D is maximal with properties i) and ii).

Condition ii) is the central one. It is related to the transitivity of S inside


of D, in the sense that for any two points x, y there exists a g ∈ S with
gx = y. Conditions i) and iii) are made in order to avoid pathologies or
trivialities. For instance without i) one could have one point sets as control
sets.
We catalog now some properties of control sets which will be used later.
The following statement clarifies the transitivity inside of D turning condition
ii) more precise.

5
Proposition 2.1 Let D be a control set for S and put

D0 = {x ∈ D : ∃g ∈ int S with gx = x} (1)


= {x ∈ D : x ∈ (int S)x} (2)

Then
a) D0 = (int S)D ∩ D.
b) D ⊂ (int S)−1 x for any x ∈ D0 if D0 6= ∅.
c) D0 = (int S)x ∩ (int S)−1 x for any x ∈ D0 if D0 6= ∅.
d) For any x, y ∈ D0 there exists g ∈ int S with gx = y.
e) D0 is dense in D provided D0 6= ∅
f) D0 is S-invariant inside D in the sense that

hx ∈ D0 if h ∈ S, x ∈ D0 and hx ∈ D.

g) D0 6= ∅ if SD ⊂ D or S −1 D ⊂ D. In the latter case D0 = D.

Proof a) Take x ∈ (int S)D ∩ D. It satisfies (int S)−1 x ∩ D 6= ∅. Also,


Sx ∩ D 6= ∅ because D contains interior points. Let z ∈ Sx ∩ D. Since
D ⊂ clSz, Sz ∩ (int S)−1 x 6= ∅. This shows that x ∈ (int S)Sz ⊂ (int S)x so
that x ∈ D0 . On the other hand, x ∈ D0 implies that x ∈ (int S)x ∩ D ⊂
(int S)D ∩ D.
b) Take x ∈ D0 and y ∈ D. We have that (int S)−1 x ∩ D 6= ∅. So by
property ii) of the definition Sy ∩ (int S)−1 x 6= ∅, which shows that y ∈
(int S)−1 x.
c) Take y ∈ (int S)−1 x ∩ (int S)x. There are g, h ∈ int S such that y =
gx = h−1 x. This together with the maximality property iii) imply that
y ∈ D. Since y = ghy we have that y ∈ D0 . Reciprocally, if x, y ∈ D0 then
by b) y ∈ (int S)−1 x and x ∈ (int S)−1 y, i.e., y ∈ (int S)−1 x ∩ (int S)x.
d) Follows immediately from c).
e) Take x ∈ D0 . By c) and the fact that both (int S)x and (int S)−1 x are
open we have that

cl D0 = (cl(int S)x) ∩ (int S)−1 x.

6
On the other hand, by b) D ⊂ (int S)−1 x, and D ⊂ cl(Sx) ⊂ cl(int S)x
because Sx ⊂ S(int S)x ⊂ (int S)x. Therefore D ⊂ cl D0 and D0 is dense in
D.
f) In fact, hx = hgx where g ∈ int S satisfies gx = x. So hx ∈ (int S)x.
Now hx ∈ D and b) imply that hx ∈ (int S)−1 x so hx ∈ D0 by c).
g) The first part is a consequence of a). Suppose that S −1 D ⊂ D and
take x ∈ D. We have that (int S)−1 x is open and contained in D, so Sx ∩
(int S)−1 x 6= ∅, that is to say, x ∈ (int S)x and x ∈ D0 . 2
In the sequel we refer to D0 as the set of transitivity inside of D. It is not
guaranteed in advance that D0 is not empty. Those control sets for which
D0 6= ∅ will be named effective control sets. These are the only ones which
are considered throughout the paper. Note that by a) in the proposition,
every control set is effective in case int S is dense in S.
There is a natural ordering between control sets which is given by putting
D1 < D2 iff there exists x ∈ D1 such that cl Sx ∩ D2 6= ∅. This is a partial
order (D1 < D2 and D2 < D1 implies D1 = D2 by maximality) which is not
in general total. With respect to this order an element is maximal provided
it is an S-invariant control set, and it is minimal if it is S −1 -invariant. By
the above proposition these are effective control sets.
Not every semigroup action admits control sets. For instance the orbits
of the semigroup {ϕt (x) = x + t : t ≥ 0} acting on the real line go away from
their initial points so there are no control sets in the sense of Definition 2.1.
It is known however that control sets always exist in case M is compact. In
fact, it is possible to show that cl Sx contains an invariant control set for any
x ∈ K, a compact invariant S-control set (c.f. [2]). These invariant control
sets are finite in number and closed because of the transitivity of G on M and
the existence of interior points in S. The following statement, also proved in
[2] is a refinement of this.
T
Proposition 2.2 Suppose that C = x∈M cl Sx 6= ∅. Then C is the only
invariant control set for S on M .

This statment has the following simple but useful consequence concerning
the transitivity of S on M (S is said to be transitive if Sx = M for all x ∈ M ).
T
Corollary 2.1 Suppose that C = x∈M cl Sx 6= ∅ is the only invariant con-
trol set. Suppose moreover that there is an S −1 -invariant control set, say C − ,
such that (int C) ∩ (int C − ) 6= ∅. Then S is transitive on M .

7
Proof Since C is invariant it is effective. For x ∈ C0 , x ∈ (int S)−1 x and
x ∈ cl Sy for all y ∈ M , so that x ∈ Sy for all y ∈ M and S −1 x = M . Take
x ∈ (int C − ) ∩ C0 , which exists because C0 is dense in C. Then S −1 x = M
and since C − is S −1 -invariant we have that C − = M . It follows that C0− = M
and thus that S −1 is transitive. This implies that S is transitive. 2
Concerning still the existence of control sets there is the following result
proved in [21], which will be used in the sequel.

Proposition 2.3 Suppose that g ∈ int S, and let Ω be a minimal set for the
g-action on M . Then there exists a control set D for S such that Ω ⊂ int D.
In particular g-fixed points are in the interior of control sets.

Another aspect of control sets we record here is their behaviour under


equivariant fibrations. Let L1 ⊂ L2 be closed subgroups of G and form the
homogeneous spaces G/L1 and G/L2 . There is a natural fibration

π : G/L1 → G/L2

which maps the coset gL1 into the coset gL2 . From L1 ⊂ L2 it follows that the
action of G is equivariant with respect to this fibration, that is, π ◦ g = g ◦ π
for all g ∈ G. Of course any equivariant fibration between homogeneous
spaces is obtained in this manner. We have, the following proposition:

Proposition 2.4 Let S, G and π: G/L1 → G/L2 be as above. Suppose that


D ⊂ G/L1 is an effective control set for S on G/L1 , and let D0 be its
set of transitivity. Then there is an effective control set E ⊂ G/L2 with
π(D0 ) ⊂ E0 .

Proof Given x ∈ D0 and g ∈ int S with gx = x, gπ(x) = π(gx) = π(x) so by


Proposition 2.5 π(x) is in some control set E, which is effective. Take y ∈ D0 .
The transitivity of int S inside D0 (d. of Proposition 2.2) implies the existence
of g1 , g2 ∈ int S with g1 y = x, g2 x = y, that is, g1 π(y) = π(x), g2 π(x) = π(y).
Hence π(y) ∈ E otherwise E would not satisfy the maximality property
defining a control set. Actually, π(y) ∈ E0 because there exists g ∈ int S
with gy = y, so gπ(y) = π(y). Therefore π(D0 ) ⊂ E0 as ascertained. 2
For invariant control sets this statement can be made more precise.

8
Proposition 2.5 Let π: G/L1 → G/L2 be an equivariant fibration with G/L1
compact. We have
a) Suppose that C1 ⊂ G/L1 is an S-invariant control set. Then C2 =
π(C1 ) is an invariant control set for S on G/L2 .
b) If C2 ⊂ G/L2 is an invariant control set, there exists an invariant
control set C1 ⊂ G/L1 , with π(C1 ) = C2
c) With C1 and C2 as in a), suppose moreover that for some y ∈ C2 , we
have π −1 {y} ⊂ C1 ; then C1 = π −1 (C2 ).

Proof a) is a consequence of the previous proposition and the invariance of


C1 and of π(C1 ).
b) π −1 (C2 ) is S-invariant and compact, so it contains an invariant control
set C1 . Take x ∈ C1 . Invariance implies π(C1 ) = π(cl Sx) = cl Sπ(x) = C2 .
c) Let y be as in the statement and take z ∈ (C2 )0 . By Proposition 2.2
b) there exists g ∈ int S with gy = z, so that gπ −1 {y} = π −1 {z}. Since
π −1 {y} ⊂ C1 and C1 is invariant we have that π −1 {z} ⊂ C1 . This together
with Proposition 2.2d) applied to (C2 )0 imply that π −1 ((C2 )0 ) ⊂ C1 . Since
(C2 )0 is dense in C2 and C1 is closed we arrive at the result. 2

3 Control sets on the maximal flag manifolds


Throughout this section we let G denote a connected semi-simple Lie group
with Lie algebra g. For our purposes here we do not loose generality in
assuming that G has finite center. Actually, we shall look at semigroup
actions on G-homogeneous spaces which are obtained through the adjoint
action of G on g. So there is no loss of generality even in assuming that G
is in fact the adjoint group of G, and hence without center.
We start by settling some notation concerning the structure of g and G.
We refer to [22, 23, 7] for the theory of semi-simple Lie groups and their
parabolic subgroups.
Let g = k⊕s be a Cartan decomposition of g with k a maximal compactly
embedded subalgebra of g and s its orthogonal complement with respect to
the Killing form. Select a maximal abelian subspace a ⊂ s. It is decomposed
into Weyl chambers. Select one of them say a+ . Associated to it there is
the positive system of roots ∆+ , which is the subset of those roots which are
strictly positive on a+ . We denote by Π the associated set of simple roots.

9
For a root α let

gα = {X ∈ g : ad(H)X = α(H)X, for all H ∈ a}

be its root space. The subalgebra


X
n+ = gα
α∈∆+

is nilpotent and g decomposes, after Iwasawa, as

g = k ⊕ a ⊕ n+

We put A = exp a. This is a split subgroup of G. It is abelian and as a


manifold diffeomorphic to a euclidean space. We put also A+ = exp a+ .
This is a (positive) chamber in G. The global Iwasawa decomposition reads
G = KAN + where K is the compact (if G has finite center) subgroup K =
exp k and N + = exp n+ . We let M denote the centralizer of a (or A) in K,
that is

M = {u ∈ K : Ad(u)H = H for all H ∈ a} (3)


= {u ∈ K : uhu−1 = h for all h ∈ A}, (4)

and

M ∗ = {u ∈ K : Ad(u)a = a} (5)
= {u ∈ K : uAu−1 = A}, (6)

is the normalizer of a (or A) in K. The finite group W = M ∗ /M is the Weyl


group of the pair (g, a).
Let m stand for the Lie algebra of M . The subspace

p = m ⊕ a ⊕ n+

is a subalgebra of g called minimal parabolic subalgebra. It is the Lie algebra


of the Lie group P = M AN + (minimal parabolic subgroup). One has that
P is its own normalizer. Also, it is the normalizer of p in G. The subject of
this section is to study actions (control sets) of subsemigroups of G on the
homogeneous space B = G/P (the maximal boundary of G, in the sense of

10
Furstenberg). Since P is the normalizer of p, G/P can be realized as the orbit
of p under the adjoint action of G on the Grassmannian of those subspaces
of g which have the same dimension as p. G/P can also be realized as the
subset of subgroups of G conjugate to P (the subset of minimal parabolic
subgroups of G). With these realizations we can think of the elements of G/P
as either minimal parabolic subalgebras or minimal parabolic subgroups. The
identification is gP ∈ G/P ↔ gP g −1 ↔ Ad(g)p. Of course, b ∈ G/P is
identified either with its isotropy subgroup or with its isotropy subalgebra.
There are the following special elements in B. Its origin which we denote by b0
(= P ∈ G/P ), and w̃b0 where w ∈ W and w̃ denotes any of its representatives
in M ∗ . Note that the above identifications identify w̃b0 either with M ANw
or m ⊕ a ⊕ nw where

Nw = w̃N + w̃−1 and nw = Ad(w̃)n+ .

By restricting the action of G to K, we obtain a transitive action. As a


K-homogeneous space we have B = K/M , which exhibits the compactness
of B.
The choice of the above subalgebras and subgroups is not unique. In
fact, by conjugating everything by some g ∈ G a new set of subalgebras and
subgroups is obtained. For instance, gA+ g −1 is a chamber which is positive
in gP g −1 , that is, gP g −1 = (gM g −1 )(gAg −1 )(gN + g −1 ) and gN + g −1 is built
from gA+ g −1 in the same way as N + from A+ . Similar statements hold for
Ad(g)a+ and Ad(g)p etc.
In the sequel these conjugations will be extensively exploited. We shall
refer to the previous choices as a canonical setting. Any other setting is
obtained from a canonical one by a conjugation by an element of G.
We discuss next two equivariant fibrations between G-homogeneous spaces,
which are useful for taking the above mentioned conjugations into account.
The product M A is a closed subgroup of G. After forming the homoge-
neous space G/M A, there is the equivariant fibration

G/M A → G/M AN +

which maps an M A coset into the M AN + coset containing it. Geometrically


G/M A can be realized as the set of Weyl chambers in g (or G). In fact, M A
is the subgroup of G which fixes the chamber a+ (A+ ). With this in mind, the
above fibration can be interpreted as the map which to a given chamber in

11
g(G) associates the unique parabolic subalgebra (subgroup) which contains
the given chamber as a positive one. We note that N + is transitive on
the fiber over b0 = M AN + . This means that any chamber positive for
P = M AN + is of the form nA+ n−1 with n ∈ N + . Similar statements hold
for the other points in B with the corresponding parabolic subgroup and
nilpotent constituent.
The product M ∗ A is a closed subgroup of G containing M A. After form-
ing the homogeneous space G/M ∗ A, there is the equivariant fibration

G/M A → G/M ∗ A.

Geometrically the homogeneous space G/M ∗ A is the set of split subalgebras


of g (subgroups of G). In fact, M ∗ A is the normalizer of a (A) in G. With
this in mind the fibration above maps a given chamber in G/M A into the
split subalgebra or subgroup containing it.
Note that M A is normal in M ∗ A so the map G/M A → G/M ∗ A defines
G/M A as a principal bundle over G/M ∗ A. The group of this bundle is
M ∗ A/M A ≈ M ∗ /M = W , the Weyl group. Thus this fibration is a covering
and there is a natural right action of W on G/M A. It is given by (gM A)w =
g w̃M A, g ∈ G, w ∈ W , where w̃ on the right hand side stands for any
representative of w in M ∗ . In terms of the above geometric interpretations,
this right action is given as follows:
Let α+ = gM A ∈ G/M A be a chamber and let α ∈ G/M ∗ A be the
split subgroup containing it. We have α+ = gA+ g −1 and α = gAg −1 . The
normalizer of α in gKg −1 is gM ∗ g −1 . Given w ∈ W let w̃ be any of its
representatives in M ∗ . We have

α+ w = (gM A)w = g w̃M A

and this can also be expressed as

(g w̃)A+ (g w̃)−1 = (g w̃g −1 )(gA+ g −1 )(g w̃g −1 )−1

so w ∈ W acts on α+ as an element of the Weyl group of α, which is obtained


by conjugating M and M ∗ by g. This means that conjugation by g sets an
isomorphism between the Weyl groups of A and α. This isomorphism does
not depend on the specific g which maps A+ into α+ , but it changes if different
chambers in A and α are considered. We stress that the right action of W

12
on G/M A does not coincide with any left action of M ∗ , where M ∗ is given
by a setting chosen in advance. In other words, we can not avoid the above
conjugations.
Our analysis of semigroup actions on G/P is based on the dynamics of the
action on B of elements inside split subgroups. Thus let h ∈ A+ . Viewing
it as a diffeomorphism of B, it has as many fixed points as W has elements.
These are the elements of the orbit of b0 under the left action of M ∗ on
B (or equivalently under the right action of W on B). These fixed points
are hyperbolic. The stable manifold for w̃b0 , w ∈ W is the Bruhat cell
N − w̃b0 . Here N − is the subgroup opposed to N + , that is, N − = exp n−
P
where n− = α∈∆+ g−α or equivalently N − = w− N + w− −1
where w− is the
+ +
only element in W which maps a into −a . These stable manifolds are
disjoint to each other and cover B. There is just one of them which is open
(and hence dense), namely N − b0 . We use the bijection between the Weyl
group and the h-fixed points on B to name them. We say that w̃b0 is the h-
fixed point of type w, where b0 is the attractor one. Of course, conjugates
of h also have a similar behaviour just by conjugating everything by the
same element in G. In particular, the fixed point of type w for ghg −1 , g ∈ G,
h ∈ A+ , is g w̃b0 = (g w̃g −1 )(gb0 ), which is the image of the attractor gb0 under
the element of the Weyl group of gAg −1 identified to w by the isomorphism
defined by g.
Now, let S be a semigroup of G. We assume throughout that S has non
empty interior in G.
By compactness of B there are at least two control sets for S on B, a
maximal one, which is S-invariant and a minimal one, which is S −1 -invariant.
Actually, the existence of an open and dense Bruhat cell permits us to show
that there is just one maximal control set, as well as one minimal (see [17]
Th. 3.1 for details).
Put
Σ = {α+ ∈ G/M A : α+ ∩ int S 6= ∅}.
This set accounts for those elements in int S whose action on B has a fixed
point which attracts an open and dense set. It is not empty. In fact, we have
the following result (proved in [17], c.f. Th. 3.5) which shows that Σ is large
enough so that it permits us to characterize the invariant control set.

Theorem 3.1 Let C be the unique invariant control set for S on B, and let

13
C0 be the set of transitivity inside of C. Let
π: G/M A → G/M AN +
be the canonical projection. Then
C0 = π(Σ).

Thus the invariant control set on B is essentially given by those fixed


points for certain elements in int S which have a dense stable manifold. The
other control sets are given by other classes of fixed points, as the forthcoming
theorems show.

Theorem 3.2 For a given w ∈ W ,


π(Σw)
is entirely contained in a control set for S. (Here π is the projection π :
G/M A → G/M AN + ).

Proof Fix b ∈ C0 and put


ν(b, w) = π{αw ∈ G/M A : α ∈ π −1 {b} ∩ Σ}.
We have π(Σw) = ∪b∈C0 ν(b, w). So in order to prove the theorem we need
to show that
i) ν(b, w) is contained in a control set, say D(b, w)
ii) For b1 , b2 ∈ C0 , D(b1 , w) = D(b2 , w)
We have

Lemma 3.1 i) holds.

Proof Pick b ∈ C0 and assume without loss of generality that b = b0 =


M AN + . Take α1 , α2 ∈ π −1 {b0 } ∩ Σ and put b1 = π(α1 w), b2 = π(α2 w).
We must show that b1 and b2 are in the same control set for S.
Since α1 and α2 are positive chambers in M AN + , there exists n0 ∈ N +
such that α2 = n0 α1 . We have
b2 = π(α2 w) = π(n0 α1 w) (7)
= n0 π(α1 w) (8)
= n0 b1 (9)

14
because of the equivariance of the G action with respect to π. Therefore b1
and b2 are in the same N + -orbit. Take h ∈ α1 ∩ int S. For every integer k,
we have
h−k b2 = (h−k n0 hk )b1 .
Since b1 belongs to the interior of a control set (c.f. Proposition 2.5) and
h−k n0 hk → 1 as k → ∞, we see that the control set containing b2 is attained
from the control set containing b1 . Reverting the argument we get that b1
and b2 are in the same control set and thus the lemma. 2
In order to prove ii) we first need another lemma.

Lemma 3.2 Let b1 , b2 ∈ C0 be given. Then there exist α ∈ π −1 {b1 } ∩ Σ and


g ∈ int S such that gb1 = b2 and gα ∈ Σ.

Proof Let P1 and P2 be the isotropy subgroups at b1 and b2 respectively,


and put
Si = Pi ∩ int S for i = 1, 2.
These are semigroups with nonempty interior in P1 and P2 , because b1 , b2 ∈
C0 . In order to get the lemma it is enough to show the existence of g ∈ int S
with gb1 = b2 such that the semigroups in P2 , int(gS1 g −1 ) and int S2 meet the
same positive chamber, say β, for P2 . In fact, if this is the case, β = gα with
α positive for P1 and α ∩ int S1 6= ∅ so that α ∈ Σ. Moreover, β ∩ int S2 6= ∅
so β = gα ∈ Σ.

To show the existence of such an element g we take first g ∈ int S with

g b1 = b2 . Its existence comes from the fact that b1 , b2 ∈ C0 . Put
∼ ∼ ∼−1
S 1 = g S1 g .
∼ ∼
This is a semigroup with non empty interior in P2 . For u ∈ S1 , g u ∈ int S, g
ub1 = b2 , and
∼ ∼ ∼ ∼−1 ∼ ∼−1 ∼ ∼−1 −1
(g u)S1 (g u)−1 = (g u g )(g S1 g )(g u g ) (10)

= h S 1 h−1 (11)
∼ ∼−1 ∼ ∼
with h =g u g ∈S 1 . So, in order to get a g of the form g =g u, u ∈ S1 , it

is enough to find h ∈S 1 and a chamber β ∈ π −1 {b2 } such that β meets the

interior of h S 1 h−1 and S2 simultaneously.

15

Now, by Theorem 3.1 S1 intersects a chamber positive for P1 . Hence S 1
intersects a chamber which is positive for P2 . Denote this chamber by β1 .
By the same reasoning S2 intersects a chamber, say β2 , positive for P2 . We
assume without loss of generality that β1 = A+ and P2 = M AN + . Take

h1 ∈ β1 ∩ int S 1 and h2 ∈ β2 ∩ int S2 . We have h2 = h0 n for some h0 ∈ β1 and
n ∈ N + , and for integer k
h−k k −k 0
1 h2 h1 = h1 (h n)h1
k
(12)
= h0 (h−k k −k k
1 nh1 ) ∈ h1 S2 h1 . (13)

Also, int S 1 intersects the chamber which contains h0 hence it meets the
chambers which contain points near h0 . Since h−k k
1 nh1 → 1 as k → ∞, we see

that int S 1 and int(h−k k
1 S2 h1 ), for large k, meet the same chamber positive
k
for P2 . So h = h1 is as desired, and the lemma is proved. 2
−1
Now, take b1 , b2 ∈ C0 . By the above lemma there exist α ∈ π {b1 } ∩ Σ
and g ∈ int S such that gα ∈ π −1 {b2 } ∩ Σ. Therefore
gπ(αw) = π(gαw) ∈ ν(b2 , w).
Since π(αw) ∈ ν(b1 , w) we get that the control set containing ν(b2 , w) is at-
tainable from the control set containing ν(b1 , w), i.e., D(b1 , w) < D(b2 , w).
By reverting the argument we have the coincidence of the control sets. This
shows ii) and finishes the proof of Theorem 3.2. 2
In what follows we denote by Dw the only control set which contains
π(Σw). Strictly speaking, the definition of Dw depends on the choice of
the Weyl chamber A+ , so a better notation for it would be Dw (A+ ). We
shall however stick to the simplicity of notation and sepecify which chamber
is being considered when confusion might arise. We stress, nevertheless,
that D1 is the invariant (maximal) control set regardless of the basic Weyl
chamber. This is because the fixed point of type 1 is the attractor for any
split element h. Note that by the very definition, Dw contains points which
are fixed by elements in int S, so its set of transitivity (Dw )0 is not empty.
Also, any control set which contains a point fixed by g ∈ int S with g in some
split subgroup is a Dw for some w ∈ W . Actually, any effective control set D
on B is of the form Dw . In fact by Proposition 2.2 the isotropy group P at b
satisfies P ∩ int S 6= ∅ if b ∈ D0 . This isotropy group is of the form M AN + .
The subset
σ = {m ∈ M : ∃hn ∈ AN + with mhn ∈ int S}

16
is a semigroup with nonvoid interior in M because M normalizes AN + . With
the assumption that G has finite center, M is compact so σ is a subgroup
containing the identity component of M . Hence int S ∩ AN + 6= ∅. Since the
union of the split subgroups in AN + is dense, we conclude that there exists
g ∈ int S with gb = b and g in some split subgroup of G. Thus D is Dw for
some w ∈ W .
It is not immediate however that b ∈ (Dw )0 belongs to some ν(b0 , w), b0 ∈
C0 . The following result complements Theorem 3.2 in this direction giving a
complete characterization of (Dw )0 as the union of the sets ν(b, w), b ∈ C0 .

Theorem 3.3 For w ∈ W let (Dw )0 be the set of transitivity inside the
control set Dw , and take b ∈ (Dw )0 . Then there exists b0 ∈ C0 such that
b ∈ ν(b0 , w).

Proof For any b ∈ (Dw )0 there exists h ∈ int S with hb = b. By the


above discussion we can take h in some split subgroup. Since the result
is independent of the choice of the basic Weyl chamber there is no loss of
generality in assuming that h ∈ A+ and that b = w̃1 b0 , for some w1 ∈ W ,
where b0 = M AN + . We have that b0 ∈ C0 because h ∈ A+ ∩ int S, so
b0 = w̃b0 satisfies b0 ∈ (Dw )0 . Hence there are g, g 0 ∈ int S such that gb0 = b
and g 0 b = b0 . By taking the Iwasawa decomposition of g (resp. g 0 ) with
respect to KANw (respectively KANw1 ), we have

g = w̃1 w̃−1 h1 (w̃n1 w̃−1 ) h1 ∈ A, n1 ∈ N + (14)


g 0 = w̃ w̃1−1 h2 (w̃1 n2 w̃1−1 ) h2 ∈ A, n2 ∈ N + . (15)

This follows from gb0 = w̃1 w̃−1 b0 = b and g 0 b = w̃ w̃1−1 b = b0 . We have


that ghk g 0 ∈ int S for positive integers k. Also,

ghk g 0 = w̃1 (w̃−1 h1 w̃)n1 (w̃−1 hk w̃)(w̃1−1 h2 w̃1 )n2 w̃1−1 .

Using the fact that A normalizes N + , this expression reduces to

ghk g 0 = w̃1 h0 nw̃1−1 , n ∈ N +

where
h0 = (w̃−1 h1 w̃)(w̃−1 hk w̃)(w̃1−1 h2 w̃1 ).

17
By taking k large enough we can assume that h0 ∈ w̃−1 A+ w̃. Note that
ghk g 0 b = b, so ghk g 0 ∈ M ANw1 which is the isotropy at b. With respect to
this decomposition we have
ghk g 0 = h00 n0
where n0 = w̃1 nw̃1−1 ∈ Nw1 and h00 = w̃1 h0 w̃1−1 .
In order to conclude the proof of the theorem we suppose first that n0 = 1.
We have that h00 belongs to the chamber α = w1 w−1 A+ , so α ∈ Σ. Let us
check that π(αw) = b. The right action of w on α is given as a left action
after a conjugation by w1 w−1 is performed. Thus
αw = (w1 w−1 )w(w1 w−1 )−1 α (16)
= w1 A+ = w̃1 A+ w̃1−1 . (17)
Now, by the definition of w1 , we have that π(w̃1 A+ w̃1−1 ) = w̃1 b0 = b, so
b = π(αw) ∈ ν(b0 , w) as desired. The general case reduces to this one by
conjugating everything by n0 ∈ Nw1 where n0 is such that n0 h00 n0 n−1 00
0 = h .
2
The control sets on the other boundaries also admit a description similar
to the one embodied in the above two theorems. This description however
requires a further discussion about which of the control sets Dw are distinct
to each other. We postpone this discussion to a later section. By now we
make some remarks about the way the invariant control set on B fibers over
the invariant control sets on the other boundaries.
We start by setting notations and recalling some facts about the other
parabolic subgroups and subalgebras. We follow the same notation as in [22].
Selecting a simple system of roots Π of the pair (g, a), let Θ ⊂ Π be any
subset. Associated to Θ there is the parabolic subgroup PΘ . Its Lie algebra
is the parabolic subalgebra pΘ . These are built as follows:
P
Let n± (Θ) be the subalgebra of n± generated by g±α , with the sum
P
extended to α ∈ Θ. These subalgebras satisfy n± (Θ) = α g±α with α ∈
hΘi+ , where hΘi+ is the smallest closed subset of ∆+ containing Θ (c.f.).
The parabolic subalgebra pΘ is given as
pΘ = n− (Θ) ⊕ p,
and the parabolic subgroup PΘ is the normalizer of pΘ in G. The Lie algebra
of PΘ is pΘ because pΘ is self normalizer in g. In general, PΘ is not connected.

18
Note that pΘ1 ⊂ pΘ2 if Θ1 ⊂ Θ2 . Also, p∅ = p and pΠ = g, so that
p ⊂ pΘ , P ⊂ PΘ for all Θ ⊂ Π.
There are further decompositions of pΘ : Let a(Θ) be the subspace of a
spanned by Hα , α ∈ Θ, where Hα ∈ a stands for the dual of α : α(·) =
hHα , ·i. Denote by aΘ the orthogonal complement of a(Θ) in a, and by n+Θ
P
the subalgebra of n+ defined by n+
Θ =
+ +
α gα , with α ∈ ∆ − hΘi . We have

pΘ = (m + a(Θ) + n+ (Θ) + n− (Θ)) + (aΘ + n+


Θ) (18)
= mΘ + (aΘ + n+
Θ ), (19)
with both terms of this decomposition subalgebras. Moreover aΘ ⊕ n+ Θ is an
ideal in pΘ , and mΘ centralizes aΘ .
The subalgebra g(Θ) generated by n+ (Θ) + n− (Θ) is semi-simple, and
a(Θ) is contained in g(Θ) as a split subalgebra. Also, g(Θ) is an ideal
in mΘ because m normalizes n+ (Θ) + n− (Θ). Therefore the orthogonal
complement g(Θ)p rp of g(Θ) in mΘ is an ideal in mΘ which satisfies mΘ =
g(Θ) ⊕ g(Θ)⊥ . Since in this decomposition both components are ideals we
have [g(Θ), g(Θ)p erp] = 0, so that g(Θ)⊥ is the centralizer of g(Θ) in mΘ .
We have that g(Θ)⊥ ⊂ m because m is the centralizer of a.
Turning to PΘ , we have that PΘ = M̃Θ AΘ NΘ+ where AΘ = exp aΘ , NΘ+ =
exp n+Θ , and M̃Θ is a closed subgroup with Lie algebra mΘ . In general M̃Θ
is not connected, although the number of its connected components is finite.
We denote by MΘ the identity component of M̃Θ . It is the closed subgroup
MΘ = exp mΘ . Furthermore, M̃Θ normalizes AΘ NΘ+ .
Now, let S ⊂ G be a semigroup with non empty interior. Take the above
canonical setting and suppose that b0 ∈ C0 . Let
π: G/P → G/PΘ
be the canonical equivariant fibration. In G/PΘ there is just one invariant
control set for S, say CΘ . It is given by
CΘ = π(C).
Therefore, if we put ξ0 = PΘ ∈ G/PΘ , we have ξ0 = π(b0 ) so that ξ0 ∈ (CΘ )0 .
With these preliminaires out of the way we can go now into our main
objective here, which is to look at the fiber F = π −1 (ξ0 ) ∩ C0 of C0 over ξ0 .
Put
S̃Θ = S ∩ PΘ .

19
This is a semigroup with non empty interior in PΘ because ξ0 ∈ (CΘ )0 and
PΘ is the isotropy at ξ0 . Also, by the fact that C0 is the set of transitivity
inside C, S̃Θ F = F . Put
SΘ = S̃Θ ∩ (PΘ )0
where (PΘ )0 is the identity component of PΘ . Since PΘ has a finite number
of connected components, SΘ has non void interior in (PΘ )0 . Of course,
SΘ F ⊂ F . We claim that b ∈ SΘ b0 any b, b0 ∈ F . In fact, there exists
h ∈ int S with hb = b and such that h is in some split subgroup of G. We
have that hξ0 = ξ0 so h ∈ int S̃Θ . Changing, if necessary, h by hm , for some
positive integer m, we can assume that h ∈ (PΘ )0 , so h ∈ int SΘ . Now,
for b0 in an open and dense subset of π −1 (ξ0 ), hn b0 → b, as n → ∞, hence
b ∈ cl(SΘ b0 ) and since hb = b with h ∈ int SΘ , we have that b ∈ SΘ b0 as
claimed.
We show now that the closure of F is actually the invariant control set
of a semigroup with non empty interior in a semi-simple Lie group acting on
π −1 (ξ0 ). The subgroup AΘ NΘ+ of (PΘ )0 acts trivially on π −1 (ξ0 ) because it is
normal and fixes b0 . Therefore the action of SΘ on π −1 (ξ0 ) depends only on
the action of ΓΘ = SΘ /AΘ NΘ+ which is a semigroup with non empty interior
in MΘ ≈ (PΘ )0 /AΘ NΘ+ .
Let Z(g(Θ)) be the centralizer of g(Θ) in MΘ . It is a closed normal
subgroup of MΘ because g(Θ) is an ideal in mΘ , the Lie algebra of MΘ .
Moreover, the Lie algebra of Z(g(Θ)), which is the centralizer of g(Θ) in
mΘ , and coincides with g(Θ)⊥ , complements g(Θ) in mΘ . Hence the Lie
algebra of
G(Θ) = MΘ /Z(g(Θ))
is isomorphic to g(Θ), which is semi-simple. Now, Z(g(Θ)) is contained in
M , the centralizer of a, hence Z(g(Θ))b0 = b0 , and since it is normal its
action on π −1 (ξ0 ) is trivial. It follows that the action of SΘ in π −1 (ξ0 ) only
depends on the action of the subsemigroup with non empty interior of G(Θ)
given by
S(Θ) = SΘ /Z(g(Θ)).
Therefore the closure of F is an invariant control set for S(Θ). Actually, it is
the only invariant control set for S(Θ) in π −1 {ξ0 } because this fiber is exactly
the maximal boundary of G(Θ), as follows from the fact that M̃Θ ∩ P is a
minimal parabolic subgroup of the reductive Lie group M̃Θ (see [23], Prop.
20, ch. 6, part II).

20
Summarizing we have

Proposition 3.1 Let C and CΘ be the invariant control sets on G/P and
G/PΘ respectively, and take ξ0 ∈ (CΘ )0 . Let S(Θ) ⊂ G(Θ) be as above. We
have that π −1 (ξ0 ) is the maximal boundary of G(Θ). Put F = π −1 (ξ0 ) ∩ C0 .
Then cl F is the S(Θ)-i.c.s. on π −1 {ξ0 } and F is contained in its set of
transitivity.

4 A Subgroup of the Weyl Group


We study now which of the control sets Dw on the maximal boundary coin-
cide. We start by the invariant control set, that is, by looking at the subset
of those elements w ∈ W such that Σw and Σ project onto the same subset
of B. Put
W (S) = {w ∈ W : Dw is the S − i.c.s}
In principle W (S) is just a subset of the Weyl group. We show below that
it is in fact a subgroup. This is not immediate because W (S) is not defined
as the subset which leaves Σ invariant : For w ∈ W (S) and α ∈ Σ it is not
necessarily true that αw ∈ Σ. We have only that some chamber positive for
π(αw) belongs to Σ. As Dw , W (S) also depends on the choice of a basic
Weyl chamber A+ for its definition. When needed we shall indicate this by
writting W (S, A+ ) instead of just W (S). With A+ fixed, W (S) is a subset of
the Weyl group of A, and w ∈ W (S) iff there exists g ∈ G such that gA+ ∈ Σ
and π(g w̃g −1 A) ∈ C0 where w̃ is a representative of w in M ∗ . From this it
is clear that W (S, A+ + −1
1 ) = gW (S, A )g if A+ +
1 = gA .

Proposition 4.1 Fix a canonical setting and suppose that b0 = M AN +


belongs to C0 . There are the equivalences
i) w ∈ W (S)
ii) N + w̃b0 ⊂ C0 , moreover if A+ ∈ Σ they are also equivalent to
iii) w̃b0 ∈ C0 , where w̃ is any representative of w.

Proof Since b0 ∈ C0 , there are h ∈ A+ , and n ∈ N + such that

hn ∈ int S.

21
Suppose first that n = 1, i.e., h ∈ int S and A+ ∈ Σ. We have that the right
action of w on A+ coincides with the left action of any of its representatives.
Hence π(A+ w) = π(w̃A+ ) = w̃π(A+ ) = w̃b0 , because of the G-equivariance
of π. It follows that w̃b0 ∈ C0 if w ∈ W (S).
Now,the fact that C0 is open implies the existence of a neighborhood U
of the identity in N + such that

U w̃b0 ⊂ C0 .

We have N + = ∪k≥0 hk U h−k . In fact, h−k nhk → 1 as k → ∞, if n ∈ N + . So


for k large enough, h−k nhk ∈ U , i. e., n ∈ hk U h−k . However

(hk U h−k )w̃b0 = hk U w̃b0 ⊂ C0

because U w̃b0 ⊂ C0 and hk ∈ int S. This shows that N + w̃b0 ⊂ C0 .


We consider now the general case. There exists n0 ∈ N + such that

hn = n0 hn0−1 .

We have that n0 A+ n−1 −1


0 ∩ int S 6= ∅. Since n0 w̃n0 is a representative of
+ −1 + −1
w in the normalizer of n0 A n0 , and of course n0 N n0 = N + , the above
arguments yields that

N + n0 w̃n−1 +
0 b0 = N w̃bo ⊂ C0 ,

and hence that (i) implies (ii). As to the converse, note that (ii) entails that
n0 w̃n−1
0 b0 = n0 w̃b0 ∈ C0 , so (i) follows from the definiton of W (S). Finally,
(ii) implies (iii) trivially, and (iii) implies (i) in case A+ ∈ Σ by the very
definition of W (S). 2

Proposition 4.2 W (S) is a subgroup of W .

Proof W (S) 6= ∅ because D1 is the maximal control set.


For the rest of the proof we keep fixed a canonical setting with b0 ∈
C0 . Take w1 , w2 ∈ W (S) and choose representatives w̃1 , w̃2 ∈ M ∗ . By
Proposition 4.1, w̃1 b0 ∈ C0 . Conjugating by w̃1 we get, from w2 ∈ W (S),
that
(w̃1 w̃2 w̃1−1 )(w̃1 b0 ) ∈ C0

22
that is, w̃1 w̃2 b0 ∈ C0 . This shows that W (S) is a semigroup in W . Since W
is finite, W (S) is a subgroup. 2
The group W (S) accounts for those elements w ∈ W for which Dw is the
invariant control set. We shall see later that the other control sets have a
similar description by the right cosets of W (S). Before doing that we need
some further properties of W (S).
For the discussion to come we fix a canonical setting and assume that
A ∩ int S 6= ∅. Put
Λ0 = {H ∈ a : exp tH ∈ int S, for some t > 0}, (20)
Λ = {h ∈ A : ∃n ∈ N + with hn ∈ int S} (21)
and ∼
Λ= {H ∈ a : exp tH ∈ Λ, for some t > 0}.

We have that Λ0 and Λ are convex cones with non empty interior in a (c.f.

[17]). Of course, Λ0 ⊂Λ. Also, it was proved in [17], Lemma 3.3, that

w−1 Λ0 ⊂Λ if w ∈ W (S). Since W (S) is a subgroup of W , we have that

wΛ0 ⊂Λ if w ∈ W (S).
Let Γ = Γ(S) be the convex cone generated by ∪{wΛ0 : w ∈ W (S)}.

Then Γ is W (S)-invariant and Γ ⊂Λ.
There are the following possibilities:

1) Λ= a. Then Ad(S) = Ad(G) and S can be a proper semigroup only
if G has infinite center (see [17], Theorem 4.2). The results to follow do not
contemplate this case.

2) Λ is a proper convex cone in a. Then Γ is also proper, so W (S) leaves
a cone in a invariant. Since W (S) is finite this holds if and only if there
exists a nonzero H ∈ a which is pointwise fixed by W (S), that is, wH = H
for all w ∈ W (S).
Now we choose a point fixed by W (S) in the closure of a chamber which

meets Λ. This is possible because a fixed point is given by
X
H= wH 0
w∈W (S)

where H 0 is any non zero vector in Γ. Since Γ is contained in Λ, we have

that H is in the closure of a chamber, say a+ , which satisfies a+ ∩ Λ6= ∅. We
note that H is in the boundary of a+ unless W (S) = {1}.

23
Associated to a+ there is a simple system of roots Π. As is well known
(see e.g. [22], Th. 1.1.2.8) the group which leaves H pointwise fixed is of
the form WΘ where Θ is a subset of Π and WΘ is the subgroup generated
by the reflexions defined by the roots in Θ. We note that Θ is given by
Θ = {α ∈ π : α(H) = 0}. Thus W (S) ⊂ WΘ for some Θ ⊂ Π. Now we
choose H with maximal regularity, that is, such that the corresponding Θ is
minimal. We claim that with this choice W (S) coincides with WΘ .
The proof of this relies on Proposition 3.6. Let PΘ be the parabolic
subgroup associated to Θ and consider the equivariant fibration

π : G/P → G/PΘ .

Denote by C and CΘ respectively the invariant control sets for S on G/P


and G/PΘ , and assume that b0 ∈ C0 , so that ξ0 = π(b0 ) ∈ (CΘ )0 . Put
Cξ0 = C ∩ π −1 (ξ0 ). Then Cξ0 is the invariant control set of the semigroup
S(Θ), which is a subsemigroup of the semi-simple Lie group G(Θ), whose Lie
algebra is g(Θ). The fiber π −1 (ξ0 ) coincides with the maximal boundary of
G(Θ). Also, WΘ is the Weyl group of G(Θ). By Proposition 4.1, w̃b0 ∈ Cξ0
if w ∈ W (S). Hence W (S) is a subgroup of W (S(Θ)).
Now, the assumption that Θ is minimal implies that W (S) does not
have a non zero fixed vector inside the split subalgebra a(Θ) of g(Θ). Since
W (S) ⊂ W (S(Θ)), we conclude that W (S(Θ)) does not leave proper cones
in a(Θ) invariant. Therefore S(Θ) = G(Θ) and Cξ0 = π −1 (ξ0 ). This implies
that w̃b0 ∈ Cξ0 ⊂ C for all w ∈ WΘ , so by Proposition 4.1, W (S) = WΘ as
claimed.
Summarizing, we have

Theorem 4.1 Fix a canonical setting with b0 ∈ C0 and A+ ∈ Σ.


Then W (S, A+ ) = WΘ for some Θ ⊂ Π, where Π is the simple system of
roots associated to the Iwasawa decomposition defined by b0 . Moreover, let
CΘ be the S-i.c.s. on G/PΘ . Then C = π −1 CΘ where π : G/P → G/PΘ is
the canonical fibration.

Proof The first part comes from the discussion above. The second one
is a consequence of the fact that Cξ0 = π −1 (ξ0 ) (notations as above) and
Proposition 2.7. 2

Corollary 4.1 The following statements are equivalent:

24
i) w ∈ W (S).
ii) There exists some minimal parabolic subgroup P = M AN and g ∈ P ∩int S
such that g = hn, h ∈ A, n ∈ N and wh = h.
iii) There exists some split subgoup A1 and a ∈ A1 ∩ int S such that wa = a.

Remark. As noted before, the definition of W (S) depends on the choice of some
Weyl chamber, a base point in G/M A. For these equivalences, it must be
taken a chamber which contains h in its closure, for ii), and a in its closure,
for iii). Any one of them—in each of these cases—is admissible. As becomes
clear from the proof, for the equivalence between ii) and iii), the isomorphism
between the Weyl groups of A and A1 is provided by a conjugation sending
a chamber containing h in its closure into a chamber containing a in its
closure. ∼
Proof Assuming i), we have by the above theorem that w fixes H ∈Λ. Thus
for some n ∈ N + , hn ∈ int S with h fixed by w so we get ii).
Take g = hn as in ii) and let g = gs gu be its Jordan decomposition with
gs semi-simple, gu unipotent and gs gu = gu gs (c.f. [22], Prop.1.4.3.3). We
have that h and gs are conjugate. Since h is hyperbolic, the same happens
to gs , so it belongs to some split subgroup, say A1 , of G (c.f. [9], Th. IX.2).
The centralizer of gs is the reductive group M̃Θ (notation as in section 3) for
Θ a subset of the simple systems of roots associated to some chamber in A1
containing gs in its closure. Clearly, gu ∈ M̃Θ and since it is unipotent, there
exists a compact element, say b in M̃Θ , with bk = 1 for some integer k, near
enough gu so that gs b ∈ int S (c.f. [17], Lemma 4.1). Therefore, if we put
a = gsk then a = (gs b)k ∈ int S. Of course, an element of the Weyl group of
A1 fixes a if and only if it fixes gs . Now, using the fact that h and gs are
conjugate, it is possible to get u ∈ G satisfying uAu−1 = A1 and uhu−1 = a.
It is clear that u maps chambers bordering h into chambers containing gs
(and hence a) in their closures. Also, multiplying u on the right by elements
of the Weyl group fixing h and on the left by those fixing a it is possible to
prescribe in advance the chambers which are interchanged by u (if h belongs
to the chambers A+ and wA+ then wh = h because the W -orbit of h crosses
wA+ once). Using now the isomorphism between the Weyl groups of A and
A1 provided by u, we get iii).
Finally, if a is as in iii), it belongs to the boundary of some chamber , say
+
A1 , in A1 (otherwise w = 1 and there is nothing to prove). Since a ∈ int S

25
and wa = a, we have that A+ + +
1 and wA1 meet int S. Thus wA1 ∈ Σ and so
w ∈ W (S). 2
Remark. The above theorem and its corollary say that W (S) is the subgroup
which fixes pointwise the diagonalizable element h with highest degree of
irregularity which can be found inside the interior of S. This degree of
irregularity is measured by the amount of one’s which appear in the spectrum
of the adjoint representation of h.
We are now in position to prove that the different control sets of the form
Dw are parameterized by the right cosets in W (S) \ W .

Theorem 4.2 Dw1 = Dw2 if and only if W (S)w1 = W (S)w2 . So the effec-
tive control sets for S on B are in one-to-one correspondence to W (S) \ W .

Proof Suppose that W (S)w1 = W (S)w2 , take a canonical setting with b0 ∈


C0 and assume that A+ ∈ Σ. For w ∈ W we let w̃ be any of its representatives
in M ∗ . We have w̃1 b0 ∈ Dw1 , w̃2 b0 ∈ Dw2 and w̃1 w̃2−1 b0 ∈ C0 . Let us check
that Dw2 < Dw1 . For this, fix a canonical setting adapted to w̃2 w̃1−1 b0 , and
put a prime in its elements in order to distinguish it from the previous one.
We have
w̃1 b0 = ((w̃1 w̃2−1 )w̃2 (w̃1 w̃2−1 )−1 )(w̃1 w̃2−1 b0 )
so that the fixed point of type w2 for the elements in A0+ is w̃1 b0 , and the
fixed points of the same type for the other positive chamber in A0 N 0+ are
of the form nw̃1 b0 with n ∈ N 0+ . Since w̃1 w̃2−1 b0 ∈ C0 , we have that Dw2
contains points of the form nw̃1 b0 , n ∈ N 0+ . Now by using an expansion
argument like in the proof of Proposition 4.1, we get that is possible to steer
Dw2 into Dw1 , that is Dw2 < Dw1 . Similarly Dw1 < Dw2 so they coincide.
Reciprocally, suppose that Dw1 = Dw2 and fix a canonical setting with
b0 ∈ C0 and A+ ∈ Σ. We take W (S) with respect to this chamber, that is
W (S, A+ ). We set

Λ2 = {H ∈ a : ∃n ∈ Nw2 s.t. (exp tH)n ∈ int S for some t > 0}.
∼ ∼
As Λ0 , Λ2 is a convex cone with non void interior in a. Since we are assuming
∼ ∼
that A+ ∈ Σ, we have that Λ2 ∩a+ 6= ∅. Let us show that Λ2 also meets
w2 w1−1 a+ . By hypothesis w̃2 b0 ∈ Dw1 . Actually, w̃2 b0 is in the interior of
the set of transitivity inside Dw1 , because hw̃2 b0 = w̃2 b0 for h ∈ A+ ∩ int S.
Therefore, by Theorem 3.5 there exists g ∈ int S with g w̃2 b0 = w̃2 b0 and

26
such that the image under w1 of the positive chamber which contains g is a
positive chamber for w̃2 b0 . That is to say, g is of the form

g = h0 n

with h0 ∈ w2 w1−1 A+ and n ∈ Nw2 . We have thus that Λ2 ∩w2 w1−1 a+ 6= ∅.

Now, using the convexity of Λ2 there are regular elements H1 , . . . , Hk in
a such that
i) H1 ∈ w2 w1−1 a+ and Hk ∈ a+

ii) The line segment {Hi , Hi+1 } between Hi and Hi+1 is contained in Λ2 ,
for all i = 1, . . . , k − 1,
iii) If a segment {Hi , Hi+1 } crosses a boundary of a chamber then it is
done through a hyperplane which is anihilated by just one root, and
iv) Any segment is contained in the union of at most two chambers.
Let Hi0 be the intersection of {Hi , Hi+1 } with the hyperplane in iii), and
let si ∈ W be such that si Hi0 = Hi0 . Then si interchanges the chambers which
contain {Hi , Hi+1 }, because si takes points near Hi0 into points which are
also near Hi0 . Also, by Corollary 4.4, si ∈ W (S) when W (S) is defined by
means of a chember containing Hi0 in its closure. Clearly, si−1 · · · s1 w2 w1−1 a+
is one of these chambers. Since we are defining W (S) by means of a+ , we
conjugate to have that

(si−1 · · · s1 w2 w1−1 )−1 si (si−1 · · · s1 w2 w1−1 ) ∈ W (S)

However, by construction sk . . . s1 w2 w1−1 a+ = a+ so sk · · · s1 w2 w1−1 = 1 .The


above expression shows then that sk−1 · · · s1 w2 w1−1 ∈ W (S). Proceeding by
induction on j, it follows that sk−j · · · s1 w2 w1−1 ∈ W (S) for j = 0, . . . , k, that
is, w2 w1−1 ∈ W (S). 2

Corollary 4.2 Keeping fixed a basic chamber A+ , we have


W (S −1 , A+ ) = W (S, A+ ).

Proof Let w0 ∈ W be such that w0 a+ = −a+ . Then Dw0 is the minimal


control set for S and hence the set of transitivity of the invariant control set
for S −1 . We can assume without loss of generality that A+ ∈ Σ, so that
b0 ∈ C0 . We have then that s ∈ W (S −1 , A+ ) if and only if sw0 b0 ∈ Dw0 ,
which by the above theorem holds if and only if s = sw0 wo −1 ∈ W (S, A+ ),
so that W (S −1 , A+ ) = W (S, A+ ) 2

27
Remark. We should compare this corollary with Theorem 4.3. There W (S, A+ ) =
WΘ when we choose A+ ∈ Σ. Hence in order to find the subset Θ corre-
−1 −1
sponding to S −1 we must take A+ and Θ ⊂ −Π. Now, W (S −1 , A+ ) =
−1
w0 W (S −1 , A+ )w0−1 where w0−1 A+ = A+ , so that the invariant control set
for S −1 is of the form π −1 Cw−0 Θ , where Cw−0 Θ stands for its invariant control
set on G/Pw0 Θ .

Corollary 4.3 Ad(S) = Ad(G) provided Dw1 = Dw2 and w2 w1−1 does not
have non zero fixed points in a.

Finally, we have the following statement which will be needed later

Proposition 4.3 Let Θ be such that W (S, A+ ) = WΘ , and let CΘ be the S-


i.c.s. on G/PΘ . Put ξ0 = PΘ ∈ G/PΘ , and assume without loss of generality
that ξ0 ∈ (CΘ )0 . Then CΘ ⊂ N − ξ0 .

Proof Suppose that CΘ intersects the complement of N − ξ0 and take h ∈ A+ ∩


int S. We have that the complement of N − ξ0 is closed and is a union of stable
manifolds for h. Therefore C0 contains a point, say ξ1 , in the complement
of N − ξ0 , such that hξ1 = ξ1 . But now ξ1 = wξ0 for w ∈ W − WΘ , because
w0 ξ0 = ξ0 for all w0 ∈ WΘ . This contradicts the fact that W (S) = WΘ . 2

5 The Other Flag Manifolds


It was shown in Proposition 2.6 that through equivariant fibrations effective
control sets are projected inside effective control sets. For the fibrations
between flag manifolds we have the following much more precise statement.

Proposition 5.1 Let G/P be the maximal boundary and π : G/P → G/PΘ
the canonical fibration with PΘ a parabolic subgroup. Let E be an effective
control set for S on G/PΘ . Then there exists w0 ∈ W such that π(Dw )0 = E0
for all w ∈ w0 WΘ . Moreover π(D0 ) = E0 if D is an effective control set
satisfying D0 ∩ π −1 (E0 ) 6= ∅.

Proof Take ξ ∈ E0 . Then there exists a regular h ∈ int S such that hξ = ξ.


In fact, denoting by Q the isotropy at ξ, the fact that ξ ∈ E0 implies that
int S ∩ Q 6= ∅. The existence of such an h follows then easily from the

28
structure of the parabolic subgroup Q, which is conjugate to PΘ (c.f. the
discussion preceding Proposition 3.6).
Fix the canonical setting defined by h. It satisfies b0 ∈ C0 where C is
the invariant control set on G/P . We have that h leaves the fiber π −1 {ξ}
invariant and its fixed points on this fiber are of the form wb0 with w ∈ w0 WΘ
for some w0 ∈ W . We can choose an element in the coset w0 WΘ , say w(ξ),
such that the h-stable manifold of w(ξ)b0 inside π −1 {ξ} is open and dense.
We have, for w ∈ w(ξ)WΘ , that π(wb0 ) = ξ so that ξ ∈ π(Dw )0 and by
Proposition 2.6 π(Dw )0 ⊂ E0 . Now, take η ∈ E0 . There exists g ∈ int S with
gη = ξ, so by the choice of w(ξ) we have that Dw(η) < Dw(ξ) . By reverting the
argument we get that Dw(ξ) = Dw(η) , and hence that W (S)w(ξ) = W (S)w(η).
Therefore, if w ∈ w(ξ)WΘ , w ∈ W (S)w(η)WΘ so w = w1 w(η)s for some
w1 ∈ W (S), s ∈ WΘ . We have Dw = Dw1 w(η)s = Dw(η)s by Theorem 4.5,
and since η ∈ π(Dw(η)s )0 we conclude that η ∈ π(Dw )0 which shows that
π(Dw )0 = E0 if w ∈ w(ξ)WΘ , proving the first part of the statement. The
second one follows immediately from the above arguments and the fact that
any effective control set on G/P is Dw for some w ∈ W . 2
This proposition shows that the effective control sets on G/PΘ are the
projections of those on the maximal boundary. Theorem 4.5 implies then the
following result.

Corollary 5.1 The number of effective control sets on G/PΘ equals the num-
ber of elements in W (S) \ W/WΘ , that is the number of W (S)-orbits in
W/WΘ .

Example 5.1 Take G = Sl(n, IR). A canonical setting is given by taking


a as the algebra of diagonal matrices with zero trace. The roots are αij =
λi − λj where λi (H) = ai if H = diag{a1 , . . . , an }. A simple system of
roots is given by Π = {αi,i+1 : i = 1, . . . , n − 1}, and the Weyl group is
just the group of permutations in n elements. It acts on a by permuting
the entries of the diagonal matrices. An interval in Π is a subset of the
type Π(i, j) = {αr, r+1 : i ≤ r ≤ j}. Any Θ ⊂ Π is a disjoint union
Θ = Π(i1 , j1 ) ∪ . . . ∪ Π(ik , jk ) with j` + 1 < i`+1 for all ` = 1, . . . , k − 1.
With Θ given this way, WΘ becomes the direct product of the permutation
groups of the subsets {i` , . . . , j` + 1}, ` = 1, . . . , k. Also, G/PΘ is realized as
Fn (1, . . . , i1 − 1, j1 + 1, . . . , ik − 1, jk + 1, jk + 2, . . . , n), where Fn (r1 , . . . , rs )
stands for the manifold of flags (V1 ⊂ . . . ⊂ Vs ) with Vi ⊂ IRn a subspace of

29
dimension ri . Since the order of WΘ is |WΘ | = (j1 − i1 + 2)! · · · (jk − ik + 2)!,
Corollary 5.2 tells us that the number of effective control sets on G/PΘ is at
most n!/(j1 − i1 + 2)! · · · (jk − ik + 2)!.
In case G/PΘ is the projective space IRP n−1 , Θ = Π(2, n − 1) so that
there are at most n = n!/(n − 1)! effective control sets on IRP n−1 , for any
S ⊂ Sl(n, IR) with int S 6= ∅. This upper bound occurs only when W (S)
reduces to the identity. It is easy to check that for the other possibilities for
W (S) (see Theorem 4.3) the number of elements in W (S) \ W/WΘ is strictly
less than n. For instance if W (S) = WΘ with Θ = Π(2, n − 1), there are
only two control sets (the maximal and the minimal ones). This is because
WΘ is the subgroup of permutation of {1, . . . , n} which fixes 1. Therefore it is
transitive on {2, . . . , n} so it has exactly two orbits on W/WΘ = {1, . . . , n}.
The above upper bound on the number of effective control sets on projective
space is one of the main results in [4], with the difference that there it is not
assumed that the semigroup has non void interior in Sl(n, IR) but only that
int(Sx) 6= ∅ for any x ∈ IRP n−1 . This assumption amounts to say that S
is a semigroup with non void interior in a subgroup of Sl(n, IR) which acts
transitively on IRP n−1 . As is well known, such a subgroup is reductive hence
it decomposes as the direct product of its center by a semi-simple Lie group.
So the more general result of [4] should be obtained from Corollary 5.2 after
taking projective representations of semi-simple Lie groups, that is by looking
at the extrinsic side of the problem, as mentioned in the introduction.

Example 5.2 As a specific instance of the above example, let G = Sl(3; IR).
There are the three boundaries, the maximal one F3 (1, 2), and two minimal
ones IRP 2 and the Grassmannian Gr2 (3). There are also three possibilities
for W (S):
{1}, WΘ1 = {1, (2, 3)} and WΘ2 = {1, (1, 2)} which are associated,
respectively, to F3 (1, 2), IRP 2 and Gr2 (3). Suppose that W (S) = WΘ1 . On
the maximal flag there are |W (S) \ W | = 3 effective control sets. They are
D1 = D(2, 3) , D(1, 2) = D(1, 3, 2) and D(1, 3) = D(1, 2, 3) . Since (1, 3) maps
the positive chamber into its opposite, D(1, 3) is the minimal control set, and
D(1, 3) < D(1, 2) < D1 .
Let π: F3 (1, 2) → IRP 2 be the projection, and let b0 ∈ F3 (1, 2) be
the canonical flag, that is, b0 = (V1 ⊂ V2 ) where V1 = span{e1 }, V2 =
span{e1 , e2 } and {e1 , e2 , e3 } is the canonical basis. We have π((1, 2)b0 ) =
π((1, 2, 3)b0 ). Therefore π(D(1, 2) ) = π(D(1, 3) ) is the minimal control set on

30
IRP 2 . Because of the duality between IRP 2 and Gr2 (3), changing IRP 2 by
Gr2 (3) or WΘ1 by WΘ2 a similar picture is obtained.

Example 5.3 Let G = Sp(n, IR)(n; IR) be the symplectic real group. Its Lie
algebra sp(n, IR) is realized as the subalgebra of 2n × 2n matrices of the form
à !
A B
C At
with A, B, C n × n matrices, and B and C symmetric. A choice for a is
the subalgebra of diagonal matrices in sp(n, IR), diag{H, −H} with H an
n × n diagonal matrix. A simple system of roots is given by Π = {λ1 −
λ2 , . . . , λn−1 − λn , 2λn } where λi is the i-th entry of H. The Weyl group has
2n n! elements. Let Θ = {λ1 − λ2 , . . . , λn−1 − λn }. The maximal parabolic
subalgebra pΘ is the subalgebra of matrices of the form
à !
A B
0 At

and the minimal boundary G/PΘ is the set of Lagrangian subspaces in IR2n ,
that is, the subspaces where the canonical symplectic form
à !
t 0 −1
Φ(x, y) = y x
1 0
vanishes identically. The subgroup WΘ is the permutation in n elements,
therefore Corollary 5.2 implies that there are at most 2n effective control sets
on the Grassmannian of Lagrangian subspaces for any S ⊂ Sp(n, IR) with
interior points.

6 Applications
6.1 Nontransitivity on Homogeneous Spaces
We used extensively the fact (proved in [17]) that G is the only of its sub-
semigroups with non void interior that is transitive on the maximal boundary
provided G has finite center. We shall now extend this result, by regarding
the action of S on other homogeneous spaces.
We start by showing the following

31
Lemma 6.1 Let G be semi-simple and S ⊂ G a semigroup with nonempty
interior. Let PΘ be a parabolic subgroup and suppose that S is transitive on
G/PΘ . Then W (S) acts transitively on W/WΘ .

Proof Fix a canonical setting with b0 ∈ C0 and A+ ∈ Σ. Let π: G/P →


G/PΘ be the canonical fibration and put ξ0 = π(b0 ). The Weyl group acts
on G/PΘ through M ∗ , and the isotropy at ξ0 is WΘ . Thus the ξ0 -orbit under
W is W/WΘ .
Pick ξ1 = w̃ξ0 , w ∈ W , and let us show that ξ1 = w̃1 ξ0 with w1 ∈ W (S),
if S is transitive on G/PΘ . Take h ∈ A+ ∩ int S. Then h preserves π −1 {ξ1 }
and there exists b1 ∈ π −1 {ξ1 } with hb1 = b1 and such that hn b → b1 as
n → ∞ for b in an open and dense subset of π −1 {ξ1 }. Since hb1 = b1 , there
exists w1 ∈ W such that b1 = w̃1 b0 .
With the assumption that S is transitive on G/PΘ , we get g ∈ int S with
gξ0 = ξ1 . We have that gb0 ∈ π −1 {ξ1 }, so by the choice of b1 we conclude
that b1 ∈ C0 . Hence w1 ∈ W (S), and since w̃1 ξ0 = ξ1 we get the result. 2
We have now

Theorem 6.1 Let G be connected, simple, and with finite center. Let G/PΘ
be any of its boundaries. Then G itself is the only of its subsemigroups with
interior points which is transitive on G/PΘ .

Proof By Theorem 4.3, W (S) = WΘ0 for some subset Θ0 of Π, the simple
system of roots. By the above lemma we need only to check that WΘ0 is not
transitive on W/WΘ unless WΘ0 = W .
In order to show this let H, H 0 ∈ a be such that

Θ = {α ∈ Π : α(H) = 0}

and
Θ0 = {α ∈ Π : α(H 0 ) = 0}
We have
WΘ = {w ∈ W : wH = H}
(c.f. Th. 1.1.2.8 in [24]). Therefore W/WΘ can be identified with the orbit
of H under W . There are two possibilities
0
a) hH, H 0 i = 0. Since WΘ0 leaves invariant the hyperplane H ⊥ orthogonal
0
to H 0 , we have that the orbit of H under WΘ0 is contained in H ⊥ . By the

32
assumption that G is simple, W does not leave invariant proper subspaces of
a. Therefore WΘ0 H 6= W H and so WΘ0 is not transitive on W/WΘ0 .
b) hH, H 0 i 6= 0, e.g. hH, H 0 i > 0. Then WΘ0 H is contained in the half
0
space bounded by H ⊥ and containing H, because WΘ0 H 0 = H 0 . Since W
does not leave invariant this half space we get that WΘ0 , is not transitive on
W/WΘ . 2
Remark. This Theorem is not true in general for non simple groups. This is
due to case a) in the above proof. For non simple groups it is possible that
0
W leaves invariant the hyperplane H ⊥ . On the other hand, if hH, H 0 i 6= 0,
S is not transitive on G/PΘ , even for non simple groups, as follows from b)
above. Examples of this situation are when Θ = ∅, i.e., G/PΘ is the maximal
flag manifold, and when Θ = Θ0 so that H = H 0 . Note that this last case
could be handled directly from Theorem 4.3 and Proposition 2.7c.

Example 6.1 Let G1 and G2 be simple noncompact and let S ⊂ G1 be a


proper semigroup with int S 6= ∅. Let P be a parabolic subgroup of G2 . Then
G1 × P is parabolic in G1 × G2 . The semigroup S × G2 has non void interior
in G1 × G2 and is transitive on (G1 × G2 )/(G1 × P = G2 /P ).

Now we extend this result to arbitrary homogeneous spaces. To begin


with suppose that G/L is realized as an orbit of a projective representation
of G. By such a representation we understand the action of G on a projective
space induced by the linear action of a representation of G on a finite dimen-
sional vector space. It was proved in [18] (c.f. Lemma 1) that transitivity of
S on a projective orbit G/L implies its transitivity on the orbits contained
in the closure of G/L.
Now, it was shown in [17], Theorem 5.4 that every projective orbit of a
semi-simple Lie group contains in its closure a compact orbit, say G/H, whose
isotropy H contains AN + and thus is contained in some parabolic subgroup
if H 6= G. Therefore, through the equivariant fibration G/H → G/P , we get
that S is transitive on G/P provided it is on G/P .
Joining these two results we get an extension of Theorem 6.2 as follows.
Let G/L be a homogeneous space and denote by l the Lie algebra of L.
Let N (l) be the normalizer of l in G. It is a closed subgroup which
contains L, so, through the equivariant fibration G/L → G/N (l), we get
that S is transitive on G/N (l) if S is transitive on G/L. But G/N (l) is a
projective orbit of G. In fact, after taking the adjoint representation of G,

33
we can make G act on any Grassmannian of subspaces of g. The orbit of
the subspace l is given, as a homogeneous space, as G/N (l), and the orbit of
a linear action on a Grassmannian turns into a projective orbit after taking
the representation on some k-fold exterior product. Suppose now that G is
simple and that 0 < dim L < dim G. Then the compact orbit in the closure
of G/N (l) is not a fixed point. From the above arguments we then have the
following result.

Theorem 6.2 Let G be a connected simple Lie group with finite center, and
G/L a homogeneous space of G such that 0 < dim L < dim G. Let S ⊂ G be
a semigroup with int S 6= ∅, and suppose that S is transitive on G/L. Then
S = G.

Remark. Trivial instances rounding the assumptions of this theorem are L = G


and L finite. In the former case every semigroup is transitive on G/L, and in
the latter S = G if int S 6= ∅ and S is transitive on G/L. This follows from
the fact that int S ∩ L is a subgroup of L so 1 ∈ int S. On the other hand if L
is a lattice in G, i.e., L is discrete and there is a G-invariant finite measure on
G/L, then any semigroup with non empty interior is transitive on G/L. This
is a general fact: the presence of a finite invariant measure implies recurrence
of the G-action, which in turn guarantees the G-invariance of the S-orbits.

6.2 Nonreversibility.
A subsemigroup S of a group G is said to be left (resp. right) reversible if
one of the following equivalent conditions are satisfied

i) ∀g ∈ G, gS ∩ S 6= ∅ (resp. Sg ∩ S 6= ∅)

ii) SS −1 (resp. S −1 S) is a subgroup.

These concepts are related to the problem of embedding abstract semi-


groups into groups. In [18] it was proved that no proper subsemigroup S,
with int S 6= ∅, of a semi-simple Lie group with finite center is simultane-
ously left and right reversible. In particular a semi-simple Lie group with
finite center does not admit proper invariant semigroups with interior points
(invariant means gSg −1 ⊂ S, for all g ∈ G). This is because invariant semi-
groups are left and right reversible. This result shows a distinction between

34
semi-simple Lie groups with finite and infinite center: In every semi-simple
Lie group with infinite center there exists at least one invariant semigroup
(see e.g. [10, 24]).
The proof in [18] of the nonreversibility (right and left) of S is based on
the following very simple link between reversibility and transitivity: Suppose
that S has non empty interior. Then S is left (right) reversible if and only
if SS −1 (resp. S −1 S) is G, because G is assumed to be connected. Now, to
say that SS −1 = G amounts to say that the semigroup S × S ⊂ G × G is
transitive from the identity of the homogeneous space G = (G × G)/∆. Here
∆ is the diagonal ∆ = {(g, g) ∈ G × G : g ∈ G} and the action of G × G on
G is given by (g1 , g2 )h = g1 hg2−1 . An analogous statement holds for S −1 S
and S −1 × S −1 = (S × S)−1 . Therefore S is simultaneously left and right
reversible if and only if S × S is transitive on G = G × G/∆. Based on this,
in [18] a whole construction was made in order to prove that transitivity of
S × S on G × G/∆ entails transitivity of S × S on the maximal boundary
of G × G if G is semi-simple. This proof might be simplified with the aid of
Theorem 6.3.
We show however, by a completely different technique that even one-sided
reversibility is not permissible for proper semigroups with non void interior in
semi-simple Lie groups with finite center. The proof is based on the following
Lemma 6.2 Let S ⊂ G be a subsemigroup and G/L a homogeneous space.
Suppose that there are a closed invariant control set C for S on G/L, and
g0 ∈ G such that g0 C ∩ C = ∅. Then S is not left reversible.
Proof For g0 ∈ G with g0 C ∩ C = ∅ we have that g0 S ∩ S = ∅, and hence
the result. In fact, suppose by contradiction that g0 S ∩ S 6= ∅. Then there
are h1 , h2 ∈ S with h1 = g0 h2 . Take b ∈ C. We have
h1 b ∈ C and h1 b = g0 h2 b,
so g0 h2 b ∈ C, which is a contradiction because h2 b ∈ C and g0 C ∩ C = ∅. 2
In order to show that semigroups in semi-simple Lie groups satisfy the
hypothesis of the above lemma we need another lemma.
Lemma 6.3 Let S ⊂ G with int S 6= ∅ and G semi-simple. Denote by C the
S-i.c.s. on the boundary G/Q. Suppose that there exists h regular in some
split subgroup of G, such that C is contained in the open stable manifold of
h. Then there exists g0 ∈ G such that g0 C ∩ C = ∅.

35
Proof Without loss of generality we can assume that h ∈ A+ and Q = PΘ ,
for some Θ. The stable manifold is then N − ξ0 where ξ0 = PΘ . The action
of h in N − ξ0 is linear. In fact, N − is diffeomorphic to its Lie algebra n−
through the exponential mapping. Moreover, the eigenvalues of h on n− are
strictly less than 1. So using the fact that C is compact and that C ⊂ N − ξ0
we have, for any neighborhood U of ξ0 , that hm C ⊂ U for sufficiently large
integers m.
Now, hm C is the invariant control set of the semigroup hm Sh−m . Taking
U small enough, we can find k ∈ K such that kU ∩ U = ∅. In fact, K
acts continuously and transitively on G/P . Therefore k(hm C) ∩ hm C = ∅.
Conjugating back by h−m we get g0 ∈ G with g0 C ∩ C = ∅. 2
We have now,

Theorem 6.3 Let G be a connected semi-simple Lie group with finite center.
Then the only of its subsemigroups which is left or right reversible is G itself.

Proof It is enough to check that S is not left reversible if S is a subsemigroup


with non empty interior. This is because S −1 is left reversible if S is right
reversible, as follows directly from the definition. But this is a consequence
of the above two lemmas and Proposition 4.8. 2

6.3 Maximality
Proposition 6.1 With S ⊂ G as before, let Θ be such that W (S) = WΘ . In
order that S be maximal it is necessary that Θ is a maximal proper subset of
Π, the simple system of roots.

Proof Suppose that there exists maximal Θ0 with Θ ⊂ Θ0 and Θ 6= Θ0 6=


Π. Then WΘ is a proper subgroup of WΘ0 and PΘ ⊂ PΘ0 . Denoting by
π : G/PΘ → G/PΘ0 the canonical fibration, we have CΘ0 = π(CΘ ), where CΘ
and CΘ0 are the invariant control set for S.
Set
S 0 = {g ∈ G : gCΘ0 ⊂ CΘ0 }.
Then S 0 is a proper subsemigroup of G which contains S, so int S 0 6= ∅. Let
us show that W (S 0 ) = WΘ .
Fix a canonical setting with A+ ∩ int S 6= ∅, and let h ∈ cl A+ be such
that it is pointwise fixed by WΘ0 . Let ξ0 and η0 be the origins in G/PΘ and

36
G/PΘ0 respectively. By Proposition 4.8, CΘ ⊂ N − ξ0 , hence CΘ0 = π(CΘ ) ⊂
π(N − ξ0 ) = N − η0 . The action of h in N − η0 is linear, and because of the
maximality of Θ0 it has eigenvalues strictly less than 1. Therefore, by the
compactness of CΘ0 we have that for any neighborhood U of η0 there exists
an integer m with hm CΘ0 ⊂ U . Taking U ⊂ CΘ0 we get that hm CΘ0 ⊂ CΘ0 ,
so hm ∈ int S 0 . Since WΘ0 hm = hm , we have from Corollary 4.4 that WΘ0 ⊂
W (S 0 ). By the maximallity of Θ0 and the fact that S 0 6= G we get that
W (S 0 ) = WΘ0 .
Therefore W (S) is a proper subgroup of W (S 0 ), so S is properly contained
in S 0 and S is not maximal. 2
As a consequence we have that a maximal semigroup with interior points
in a semi-simple Lie group with finite center is the semigroup of contractions
of some subset in one of the minimal boundaries:

Corollary 6.1 Suppose S is maximal. Then there exists a minimal boundary


G/PΘ and a subset C of G/PΘ such that

S = {g ∈ G : gC ⊂ C}.

Moreover, C is closed, cl(int C) = C and C is the invariant control set for S


on G/PΘ

Proof By the above proposition, W (S) = WΘ with G/PΘ a minimal bound-


ary. Let C = CΘ be the S-i.c.s. on G/PΘ . C is closed and its interior is
dense in itself. Moreover S leaves C invariant so S is as claimed. 2
The necessary condition stated in Proposition 6.7 is not sufficient. This is
because not every compact subset of a minimal boundary satisfying the con-
ditions of Proposition 6.7 is the invariant control set of a maximal semigroup.
We shall not give here any general characterization of such invariant control
sets, but we illustrate some possible outcomes in two specific situations. We
need first the following statement, which gives a sufficient condition in order
that a subset is the invariant control set of a maximal semigroup.

Proposition 6.2 Assume that G has finite center and let G/PΘ be a mini-
mal boundary. Let C ⊂ G/PΘ be a proper closed subset with cl(int C) = C.
Suppose that the following conditions are satisfied.
i) For any b ∈ int C there exists h in some split subgroup of G satisfying
hb = b and such that its stable manifold contains C

37
ii) For any b0 ∈ C c there exists b ∈ int C and h as in i) such that hb0 = b0 .
Then
S = {g ∈ G : gC ⊂ C}
is a maximal semigroup with interior points in G. Moreover C is the invari-
ant control set for S in G/PΘ .
Proof Assumption i) ensures that for any b ∈ int C there is an open set U ⊂
C with b ∈ U , and h ∈ G with hb = b, such that hC ⊂ U . By the continuity
of the G-action on G/PΘ , with respect to the compact-open topology, we have
that h ∈ int S, so S has interior points. Also, since b ∈ int C is arbitrary,
int C is contained in the invariant control sets for S on G/PΘ . Therefore C
is the invariant control set because it is S-invariant.
Condition ii) together with Theorem 3.2 ensures that any point outside C
is in the interior of some control set for S. From this we get the maximallity
of S. In fact, suppose that S 0 is a semigroup containing S properly, and let
C 0 be its invariant control set on G/PΘ . Then C ⊂ C 0 and C 6= C 0 because C
is not invariant by S 0 . Therefore W (S 0 ) contains WΘ properly so W (S 0 ) = W
and S 0 = G. That is S is maximal. 2
Remark. A subsemigroup S of a group G is said to be total if G = S ∪ S −1 .
In Lawson [13] (see also [10]) a series of results relating maximal to total
semigroups where derived. Condition ii) above says that G/PΘ is the union
of control sets for S, so it is natural to say that S is total with respect to
the flag manifold. With this terminology, the above proof shows that S is
maximal if S is total w.r.t. G/PΘ and is the semigroup of those elements
which contract the invariant control set.
The first application we give of the previous result is for semigroups in
G, when G has rank one. For these groups there is just one boundary and
W consists of two elements. So for G with finite center W (S) reduces to the
identity in case S is a proper semigroup. We have,
Theorem 6.4 Let G be a rank one semi-simple Lie group with finite center,
and denote by B its boundary. Let C be a proper closed subset of B satisfying
cl(int C) = C. Let
S = {g ∈ G : gC ⊂ C}.
Then int S 6= ∅ and S is a maximal proper semigroup in G. Moreover C is
the invariant control set for S in B and the other control set complements
C.

38
Proof Take a canonical setting with b0 ∈ int C, let W = {1, w} be the Weyl
group and put b− = wb0 . The N + orbit of b− is B − {b0 }. So for any b 6= b0
there exist n ∈ N + with nb− = b. Now, let h be in the positive chamber
and take b outside C. Then nhn−1 satisfies conditions i) and ii) of the above
proposition, because nhn−1 b0 = b0 and nhn−1 b = b. So the result follows. 2
Another application of Proposition 6.9 concerns semigroups in Sl(n, IR).
The minimal boundaries for this group are the Grassmannians of k-planes in
IRn , k = 1, . . . , n − 1. So according to Corollary 6.8 there are n − 1 classes of
maximal subsemigroups of Sl(n, IR), each one associated to a Grassmannian.
The following result clarifies the situation for the class associated to IRP n−1 ,
the projective space. The maximal connected semigroups are those which
leave a pointed cone in IRn invariant. By a cone we mean a closed convex
cone. A pointed cone Γ is a cone which does not contain non zero subspaces,
that is, Γ ∩ (−Γ) = {0}, where −Γ = {x : −x ∈ Γ} (see [10]). We have

Theorem 6.5 Let Γ be a pointed cone with non empty interior in IRn , and
set
S = {g ∈ Sl(n, IR) : g(Γ ∪ −Γ) ⊂ Γ ∪ −Γ}.
Then S is a semigroup with non void interior which is maximal in Sl(n, IR).

Proof Let {e1 , . . . , en } be a basis of IRn , and H the linear map such that in
this basis H = diag(n − 1, −1, . . . , −1). Put h = exp H. The open stable
manifold of h in IRP n−1 is

{π(x1 , . . . , xn ) ∈ IRP n−1 : x1 6= 0},

where π: IRn − {0} → IRP n−1 is the the canonical mapping. Now, put C =
πΓ. Then C is a proper closed subset of IRP n−1 which is the closure of its
interior. Take b = πx ∈ int C, with x ∈ int Γ, and b0 = πx0 outside C. Since
Γ is a pointed cone there exists λ ∈ (IRn )∗ such that λx > 0 and λx0 < 0 (see
[10]). Let {x, x0 , f3 , . . . , fn } be a basis of IRn with λfi = 0, i ≥ 3 and take
H and h as above, with respect to this basis. Then h satisfies conditions i)
and ii) of Proposition 6.9. Therefore the semigroup

SC = {g ∈ Sl(n, IR) : gC ⊂ C}

is maximal in Sl(n, IR). Since this semigroup coincides with the semigroup
which preserves the union of Γ with −Γ we have the result. 2

39
References
[1] Arnold, L., and W. Kliemann Qualitative theory of stochastic sys-
temsProbabilistic Analysis and Related Topics, vol. 3, A. T. Barucha-
Reid, Ed., Academic Press, 1983, 1–79

[2] Arnold, L., W. Kliemann, and E. Oeljeklaus : Lyapunov exponents of


linear stochastic systemsIn: Lyapunov Exponents, L. Arnold, and V.
Wihstutz, Eds., LNM (Springer) 1186 (1986)

[3] Chon, I., and J. D. Lawson : Problems on semigroups and control Semi-
group Forum 41 (1990), 245–252

[4] Colonius, F. and W. Kliemann : Linear control semigroups acting on


projective spaces Journal of Dynamics and Differential Equations 5
(1993), 495–528

[5] Colonius, F.and W. Kliemann : Some aspects of control systems as


dynamical systemsJournal of Dynamics and Differential Equations 5
(1993), 469–494

[6] Colonius, F. and W. Kliemann : Lyapounov Exponents of Control


FlowsIn: Lyapounov Exponents, L. Arnold, C. Eckemann, and H.
Crauel, Eds., LNM (Springer) 1486 (1991)

[7] Duistermaat, J. J., J. A. C. Kolk, and V. S. Varadarajan : Functions,


flows and oscillatory integrals on flag manifolds and conjugacy classes in
real semisimple Lie groupsCompositio Mathematica 49 (1983), 309–398

[8] Guivarc’h, Y., and A. Raugi : Frontière de Furstenberg, propriétés de


contraction et théorèmes de convergence Z. Wahrscheinlichkeitstheorie
Verw. Geb. 69 (1985), 184–242

[9] Helgason, S. : “Differential Geometry, Lie Groups and Symmetric


Spaces,”Academic Press, 1978

[10] Hilgert, J., K. H. Hofmann, and J. D. Lawson : “Lie Groups, Convex


Cones and Semigroups,”Oxford University Press, 1989

40
[11] Hofmann, K. H., J. D. Lawson, and J. S. Pym : “The Analytical and
Topological Theory of Semigroups,” De Gruyter, Expositions in Math-
ematics 1 (1990)

[12] Jurdjevic, V. and I. Kupka : Controllability of right invariant systems


on semi-simple Lie groups and their homogeneous spacesAnn. Institut
Fourier (Grenoble) 31 (1981), 151–179

[13] Lawson, J. D. : Maximal subsemigroups of Lie groups that are totalProc.


Edinburgh Math. Soc. 30 (1987), 479–501

[14] Neeb, K-H. : Conal orders on homogeneous spaces Inventiones Math.


104 (1991), 467–496

[15] Ruppert, W. A. F. : On open subsemigroups of connected


groupsSemigroup Forum 39 (1989), 347–362

[16] San Martin, L. and L. Arnold : A control problem related to the Lya-
punov spectrum of stochastic flowsMat. Aplicada e Computacional 5
(1986), 31–64

[17] San Martin, L. :Invariant control sets on flag manifoldsMath. of Control,


Signals and Systems 6 (1993), (1992), 41–61

[18] San Martin, L.A.B. :Nonreversibility of subsemigroups of semi-simple


Lie groupsSemigroup Forum 44 (1992), 376–387

[19] San Martin, L.A.B. : Two results on maximal subsemigroups of Lie


groupsTechnical Report, University of Campinas 1992

[20] Stefan, P. :Accessible sets, orbits and foliations with singularitiesProc.


London Math. Soc. 29 (1974), 699–713

[21] Tonelli, P. A. : Control sets on homogeneous spacesThesis, University of


Bremen, 1991

[22] Warner, G. : “Harmonic Analysis on Semi-simple Lie Groups”Springer


- Verlag, 1972

[23] Varadarajan, V. S. : Harmonic analysis on real reductive groups LNM


(Springer) 576 (1977)

41
View publication stats

[24] Vinberg, E. B. : Invariant convex cones and orderings in Lie


groupsFunct. Anal. and Appl. 14, pp. 1-13 (1980)

42

You might also like