EE363 Winter 2008-09

Lecture 6
Invariant subspaces
invariant subspaces
a matrix criterion
Sylvester equation
the PBH controllability and observability conditions
invariant subspaces, quadratic matrix equations, and the ARE
61
Invariant subspaces
suppose A R
nn
and V R
n
is a subspace
we say that V is A-invariant if AV V, i.e., v V = Av V
examples:
{0} and R
n
are always A-invariant
span{v
1
, . . . , v
m
} is A-invariant, where v
i
are (right) eigenvectors of A
if A is block upper triangular,
A =
_
A
11
A
12
0 A
22
_
,
with A
11
R
rr
, then V =
__
z
0
_

z R
r
_
is A-invariant
Invariant subspaces 62
Examples from linear systems
if B R
nm
, then the controllable subspace
R(C) = R
_
[B AB A
n1
B]
_
is A-invariant
if C R
pn
, then the unobservable subspace
N(O) = N
_
_
_
_
C
.
.
.
CA
n1
_
_
_
_
is A-invariant
Invariant subspaces 63
Dynamical interpretation
consider system x = Ax
V is A-invariant if and only if
x(0) V = x(t) V for all t 0
(same statement holds for discrete-time system)
Invariant subspaces 64
A matrix criterion for A-invariance
suppose V is A-invariant
let columns of M R
nk
span V, i.e.,
V = R(M) = R([t
1
t
k
])
since At
1
V, we can express it as
At
1
= x
11
t
1
+ +x
k1
t
k
we can do the same for At
2
, . . . , At
k
, which gives
A[t
1
t
k
] = [t
1
t
k
]
_
_
x
11
x
1k
.
.
.
.
.
.
x
k1
x
kk
_
_
or, simply, AM = MX
Invariant subspaces 65
in other words: if R(M) is A-invariant, then there is a matrix X such that
AM = MX
converse is also true: if there is an X such that AM = MX, then R(M)
is A-invariant
now assume M is rank k, i.e., {t
1
, . . . , t
k
} is a basis for V
then every eigenvalue of X is an eigenvalue of A, and the associated
eigenvector is in V = R(M)
if Xu = u, u = 0, then Mu = 0 and A(Mu) = MXu = Mu
so the eigenvalues of X are a subset of the eigenvalues of A
more generally: if AM = MX (no assumption on rank of M), then A and
X share at least Rank(M) eigenvalues
Invariant subspaces 66
Sylvester equation
the Sylvester equation is AX +XB = C, where A, B, C, X R
nn
when does this have a solution X for every C?
express as S(X) = C, where S is the linear function S(X) = AX +XB
(S maps R
nn
into R
nn
and is called the Sylvester operator )
so the question is: when is S nonsingular?
S is singular if and only if there exists a nonzero X with S(X) = 0
this means AX +XB = 0, so AX = X(B), which means A and B
share at least one eigenvalue (since X = 0)
so we have: if S is singular, then A and B have a common eigenvalue
Invariant subspaces 67
lets show the converse: if A and B share an eigenvalue, S is singular
suppose
Av = v, w
T
B = w
T
, v, w = 0
then with X = vw
T
we have X = 0 and
S(X) = AX +XB = Avw
T
+vw
T
B = (v)w
T
+v(w
T
) = 0
which shows S is singular
so, Sylvestor operator is singular if and only if A and B have a common
eigenvalue
or: Sylvestor operator is nonsingular if and only if A and B have no
common eigenvalues
Invariant subspaces 68
Uniqueness of stabilizing ARE solution
suppose P is any solution of ARE
A
T
P +PA+QPBR
1
B
T
P = 0
and dene K = R
1
B
T
P
we say P is a stabilizing solution of ARE if
A+BK = ABR
1
B
T
P
is stable, i.e., its eigenvalues have negative real part
fact: there is at most one stabilizing solution of the ARE
(which therefore is the one that gives the value function)
Invariant subspaces 69
to show this, suppose P
1
and P
2
are both stabilizing solutions
subtract AREs to get
A
T
(P
1
P
2
) + (P
1
P
2
)AP
1
BR
1
B
T
P
1
+P
2
BR
1
B
T
P
2
= 0
rewrite as Sylvester equation
(A+BK
2
)
T
(P
1
P
2
) + (P
1
P
2
)(A+BK
1
) = 0
since A+BK
2
and A+BK
1
are both stable, A+BK
2
and (A+BK
1
)
cannot share any eigenvalues, so we conclude P
1
P
2
= 0
Invariant subspaces 610
Change of coordinates
suppose V = R(M) is A-invariant, where M R
nk
is rank k
nd

M R
n(nk)
so that [M

M] is nonsingular
A[M

M] = [AM A

M] = [M

M]
_
X Y
0 Z
_
where
_
Y
Z
_
= [M

M]
1
A

M
with T = [M

M], we have
T
1
AT =
_
X Y
0 Z
_
Invariant subspaces 611
in other words: if V is A-invariant we can change coordinates so that
A becomes block upper triangular in the new coordinates
V corresponds to
__
z
0
_

z R
k
_
in the new coordinates
Invariant subspaces 612
Revealing the controllable subspace
consider x = Ax +Bu (or x
t+1
= Ax
t
+Bu
t
) and assume it is not
controllable, so V = R(C) = R
n
let columns of M R
k
be basis for controllable subspace
(e.g., choose k independent columns from C)
let

M R
n(nk)
be such that T = [M

M] is nonsingular
then
T
1
AT =
_

A
11

A
12
0

A
22
_
, T
1
B =
_

B
1
0
_

C = T
1
C =
_

B
1


A
n1
11

B
1
0 0
_
in the new coordinates the controllable subspace is {(z, 0) | z R
k
};
(

A
11
,

B
1
) is controllable
Invariant subspaces 613
we have changed coordinates to reveal the controllable subspace:
u
x
1
x
2
1/s
1/s

A
11

A
12

A
22

B
1
roughly speaking, x
1
is the controllable part of the state
Invariant subspaces 614
Revealing the unobservable subspace
similarly, if (C, A) is not observable, we can change coordinates to obtain
T
1
AT =
_

A
11
0

A
21

A
22
_
, CT =
_

C
1
0

and (

C
1
,

A
11
) is observable
Invariant subspaces 615
Popov-Belevitch-Hautus controllability test
PBH controllability criterion: (A, B) is controllable if and only if
Rank [sI A B] = n for all s C
equivalent to:
(A, B) is uncontrollable if and only if there is a w = 0 with
w
T
A = w
T
, w
T
B = 0
i.e., a left eigenvector is orthogonal to columns of B
Invariant subspaces 616
to show it, rst assume that w = 0, w
T
A = w
T
, w
T
B = 0
then for k = 1, . . . , n 1, w
T
A
k
B =
k
w
T
B = 0, so
w
T
[B AB A
n1
B] = w
T
C = 0
which shows (A, B) not controllable
conversely, suppose (A, B) not controllable
change coordinates as on p.615, let z be any left eigenvector of

A
22
, and
dene w = (0, z)
then w
T

A = w
T
, w
T

B = 0
it follows that w
T
A = w
T
, w
T
B = 0, where w = T
T
w
Invariant subspaces 617
PBH observability test
PBH observability criterion: (C, A) is observable if and only if
Rank
_
sI A
C
_
= n for all s C
equivalent to:
(C, A) is unobservable if and only if there is a v = 0 with
Av = v, Cv = 0
i.e., a (right) eigenvector is in the nullspace of C
Invariant subspaces 618
Observability and controllability of modes
the PBH tests allow us to identify unobservable and uncontrollable modes
the mode associated with right and left eigenvectors v, w is
uncontrollable if w
T
B = 0
unobservable if Cv = 0
(classication can be done with repeated eigenvalues, Jordan blocks, but
gets tricky)
Invariant subspaces 619
Controllability and linear state feedback
we consider system x = Ax +Bu (or x
t+1
= Ax
t
+Bu
t
)
we refer to u = Kx +w as a linear state feedback (with auxiliary input
w), with associated closed-loop system x = (A+BK)x +Bw
u
x
w
A
K
B
1/s
Invariant subspaces 620
suppose w
T
A = w
T
, w = 0, w
T
B = 0, i.e., w corresponds to
uncontrollable mode of open loop system
then w
T
(A+BK) = w
T
A+w
T
BK = w
T
, i.e., w is also a left
eigenvector of closed-loop system, associated with eigenvalue
i.e., eigenvalues (and indeed, left eigenvectors) associated with
uncontrollable modes cannot be changed by linear state feedback
conversely, if w is left eigenvector associated with uncontrollable
closed-loop mode, then w is left eigenvector associated with uncontrollable
open-loop mode
in other words: state feedback preserves uncontrollable eigenvalues and the
associated left eigenvectors
Invariant subspaces 621
Invariant subspaces and quadratic matrix equations
suppose V = R(M) is A-invariant, where M R
nk
is rank k, so
AM = MX for some X R
kk
conformally partition as
_
A
11
A
12
A
21
A
22
_ _
M
1
M
2
_
=
_
M
1
M
2
_
X
A
11
M
1
+A
12
M
2
= M
1
X, A
21
M
1
+A
22
M
2
= M
2
X
eliminate X from rst equation (assuming M
1
is nonsingular):
X = M
1
1
A
11
M
1
+M
1
1
A
12
M
2
substituting this into second equation yields
A
21
M
1
+A
22
M
2
= M
2
M
1
1
A
11
M
1
+M
2
M
1
1
A
12
M
2
Invariant subspaces 622
multiply on right by M
1
1
:
A
21
+A
22
M
2
M
1
1
= M
2
M
1
1
A
11
+M
2
M
1
1
A
12
M
2
M
1
1
with P = M
2
M
1
1
, we have
A
22
P +PA
11
A
21
+PA
12
P = 0,
a general quadratic matrix equation
if we take A to be Hamitonian associated with a cts-time LQR problem, we
recover the method of solving ARE via stable eigenvectors of Hamiltonian
Invariant subspaces 623