Page 1
arXiv:0901.3764v2 [math.OC] 3 Mar 2009
Electronic Journal of Differential Equations, Vol. 2009(2009), No. 37, pp. 1–32.
ISSN: 1072-6691. URL: http://ejde.math.txstate.edu or http://ejde.math.unt.edu
ftp ejde.math.txstate.edu
CONTROLLABILITY, OBSERVABILITY, REALIZABILITY, AND
STABILITY OF DYNAMIC LINEAR SYSTEMS
JOHN M. DAVIS, IAN A. GRAVAGNE, BILLY J. JACKSON, ROBERT J. MARKS II
Abstract. We develop a linear systems theory that coincides with the ex-
isting theories for continuous and discrete dynamical systems, but that also
extends to linear systems defined on nonuniform time scales. The approach
here is based on generalized Laplace transform methods (e.g. shifts and con-
volution) from the recent work [13]. We study controllability in terms of the
controllability Gramian and various rank conditions (including Kalman’s) in
both the time invariant and time varying settings and compare the results.
We explore observability in terms of both Gramian and rank conditions and
establish related realizability results. We conclude by applying this systems
theory to connect exponential and BIBO stability problems in this general
setting. Numerous examples are included to show the utility of these results.
1. Introduction
In this paper, our goal is to develop the foundation for a comprehensive linear
systems theory which not only coincides with the existing canonical systems theories
in the continuous and discrete cases, but also to extend those theories to dynamical
systems with nonuniform domains (e.g. the quantum time scale used in quantum
calculus [9]). We quickly see that the standard arguments on R and Z do not go
through when the graininess of the underlying time scale is not uniform, but we
overcome this obstacle by taking an approach rooted in recent generalized Laplace
transform methods [13, 19]. For those not familiar with the rapidly expanding area
of dynamic equations on time scales, excellent references are [6, 7].
We examine the foundational notions of controllability, observability, realiz-
ability, and stability commonly dealt with in linear systems and control theory
[3, 8, 22, 24]. Our focus here is how to generalize these concepts to the nonuni-
form domain setting while at the same time preserving and unifying the well-known
bodies of knowledge on these subjects in the continuous and discrete cases. This
generalized framework has already shown promising application to adaptive control
regimes [16, 17].
2000 Mathematics Subject Classification. 93B05, 93B07, 93B20, 93B55, 93D99.
Key words and phrases. Systems theory; time scale; controllability; observability; realizability;
Gramian; exponential stability; BIBO stability; generalized Laplace transform; convolution.
c ?2009 Texas State University - San Marcos.
Submitted January 23, 2009. Published March 3, 2009.
Supported by NSF Grants EHS#0410685 and CMMI#726996.
1
Page 2
2J. M. DAVIS, I. A. GRAVAGNE, B. J. JACKSON, R. J. MARKS IIEJDE-2009/37
Throughout this work, we assume the following:
• T is a time scale that is unbounded above but with bounded graininess,
• A(t) ∈ Rn×n,B(t) ∈ Rn×m,C(t) ∈ Rp×n, and D(t) ∈ Rp×mare all rd-
continuous on T,
• all systems in question are regressive.
The third assumption implies that the matrix I + µ(t)A(t) is invertible, and so
on Z, the transition matrix will always be invertible. We are therefore justified in
talking about controllability rather than reachability which is common [8, 14, 25]
since the transition matrix in general need not be invertible for T = Z.
In the following sections, we begin with the time varying case, and then proceed
to treat the time invariant case. We will get stronger (necessary and sufficient)
results in the more restrictive time invariant setting relative to the time varying
case (sufficient conditions). Although some of the statements contained in this work
can be found elsewhere [4, 5, 15], in each of these cases proofs are either missing,
are restricted to time invariant systems, or are believed to be in error [15] when
T has nonuniform graininess. Moreover—and very importantly—the methods used
here are rooted in Laplace transform techniques (shifts and convolution), and thus
are fundamentally different than the approaches taken elsewhere in the literature.
This tack overcomes the subtle problems that arise in the arguments found in [15]
when the graininess is nonconstant.
2. Controllability
2.1. Time Varying Case. In linear systems theory, we say that a system is con-
trollable provided the solution of the relevant dynamical system (discrete, continu-
ous, or hybrid) can be driven to a specified final state in finite time. We make this
precise now.
Definition 2.1. Let A(t) ∈ Rn×n, B(t) ∈ Rn×m, C(t) ∈ Rp×n, and D(t) ∈ Rp×m
all be rd-continuous matrix functions defined on T, with p,m ≤ n. The regressive
linear system
x∆(t) = A(t)x(t) + B(t)u(t),x(t0) = x0,
y(t) = C(t)x(t) + D(t)u(t),
(2.1)
is controllable on [t0,tf] if given any initial state x0 there exists a rd-continuous
input u(t) such that the corresponding solution of the system satisfies x(tf) = xf.
Our first result establishes that a necessary and sufficient condition for con-
trollability of the linear system (2.1) is the invertibility of an associated Gramian
matrix.
Theorem 2.2 (Controllability Gramian Condition). The regressive linear system
x∆(t) = A(t)x(t) + B(t)u(t),x(t0) = x0,
y(t) = C(t)x(t) + D(t)u(t),
is controllable on [t0,tf] if and only if the n × n controllability Gramian matrix
given by
?tf
is invertible, where ΦZ(t,t0) is the transition matrix for the system X∆(t) =
Z(t)X(t), X(t0) = I.
GC(t0,tf) :=
t0
ΦA(t0,σ(t))B(t)BT(t)ΦT
A(t0,σ(t))∆t,
Page 3
EJDE-2009/37CONTROLLABILITY, OBSERVABILITY, REALIZABILITY3
Proof. Suppose GC(t0,tf) is invertible. Then, given x0and xf, we can choose the
input signal u(t) as
u(t) = −BT(t)ΦT
and extend u(t) continuously for all other values of t. The corresponding solution
of the system at t = tf can be written as
A(t0,σ(t))G−1
C(t0,tf)(x0− ΦA(t0,tf)xf),t ∈ [t0,tf),
x(tf) = ΦA(tf,t0)x0+
?tf
t0
ΦA(tf,σ(t))B(t)u(t)∆t
= ΦA(tf,t0)x0
?tf
= ΦA(tf,t0)x0− ΦA(tf,t0)
?tf
= ΦA(tf,t0)x0− (ΦA(tf,t0)x0− xf)
= xf,
−
t0
ΦA(tf,σ(t))B(t)BT(t)ΦT
A(t0,σ(t))G−1
C(t0,tf)(x0− ΦA(t0,tf)xf)∆t
×
t0
ΦA(t0,σ(t))B(t)BT(t)ΦT
A(t0,σ(t))∆tG−1
C(t0,tf)(x0− ΦA(t0,tf)xf)
so that the state equation is controllable on [t0,tf].
For the converse, suppose that the state equation is controllable, but for the sake
of a contradiction, assume that the matrix GC(t0,tf) is not invertible. If GC(t0,tf)
is not invertible, then there exists a vector xa?= 0 such that
?tf
?tf
and hence
xT
aΦA(t0,σ(t))B(t) = 0,
However, the state equation is controllable on [t0,tf], and so choosing x0= xa+
ΦA(t0,tf)xf, there exists an input signal ua(t) such that
0 = xT
aGC(t0,tf)xa=
t0
xT
aΦA(t0,σ(t))B(t)BT(t)ΦT
A(t0,σ(t))xa∆t
=
t0
?xT
aΦA(t0,σ(t))B(t)?2∆t,(2.2)
t ∈ [t0,tf).(2.3)
xf= ΦA(tf,t0)x0+
?tf
t0
ΦA(tf,σ(t))B(t)ua(t)∆t,
which is equivalent to the equation
xa= −
?tf
t0
ΦA(t0,σ(t))B(t)ua(t)∆t.
Multiplying through by xT
tion. Thus, the matrix GC(t0,tf) is invertible.
aand using (2.2) and (2.3) yields xT
axa= 0, a contradic-
?
Since the controllability Gramian is symmetric and positive semidefinite, Theo-
rem 2.2 can be interpreted as saying (2.1) is controllable on [t0,tf] if and only if
the Gramian is positive definite. A system that is not controllable on [t0,tf] may
become so when either tf is increased or t0is decreased. Likewise, a system that
is controllable on [t0,tf] may become uncontrollable if t0is increased and/or tf is
decreased.
Page 4
4 J. M. DAVIS, I. A. GRAVAGNE, B. J. JACKSON, R. J. MARKS IIEJDE-2009/37
Although the preceding theorem is strong in theory, in practice it is quite lim-
ited since computing the controllability Gramian requires explicit knowledge of the
transition matrix, but the transition matrix for time varying problems is generally
not known and can be difficult to approximate in some cases. This observation
motivates the following definition and our next theorem.
Definition 2.3. If T is a time scale such that µ is sufficiently differentiable with the
indicated rd-continuous derivatives, define the sequence of n × m matrix functions
K0(t) := B(t),
Kj+1(t) := (I + µ(σ(t))A(σ(t)))−1K∆
j(t) −
?
(I + µ(σ(t))A(σ(t)))−1(µ∆(t)A(σ(t))
+ µ(t)A∆(t))(I + µ(t)A(t))−1+ A(t)(I + µ(t)A(t))−1?
j = 0,1,2,....
Kj(t),
A straightforward induction proof shows that for all t,s, we have
∂j
∆sj[ΦA(σ(t),σ(s))B(s)] = ΦA(σ(t),σ(s))Kj(s),
Evaluation at s = t yields a relationship between these matrices and those in
Definition 2.3:
∂j
∆sj[ΦA(σ(t),σ(s))B(s)]
This in turn leads to the following sufficient condition for controllability.
j = 0,1,2,...
Kj(t) =
???
s=t,j = 0,1,2,...
Theorem 2.4 (Controllability Rank Theorem). Suppose q ∈ Z+such that, for
t ∈ [t0,tf], B(t) is q-times rd-continuously differentiable and both of µ(t) and A(t)
are (q − 1)-times rd-continuously differentiable. Then the regressive linear system
x∆(t) = A(t)x(t) + B(t)u(t),x(t0) = x0,
y(t) = C(t)x(t) + D(t)u(t),
is controllable on [t0,tf] if for some tc∈ [t0,tf), we have
rank?K0(tc)
where
∂j
∆sj[ΦA(σ(t),σ(s))B(s)]
K1(tc) ...Kq(tc)?= n,
Kj(t) =
???
s=t,j = 0,1,...,q.
Proof. Suppose there is some tc∈ [t0,tf) such that the rank condition holds. For
the sake of a contradiction, suppose that the state equation is not controllable on
[t0,tf]. Then the controllability Gramian GC(t0,tf) is not invertible and, as in the
proof of Theorem 2.2, there exists a nonzero n × 1 vector xasuch that
xT
aΦA(t0,σ(t))B(t) = 0,
If we choose the nonzero vector xbso that xb= ΦT
t ∈ [t0,tf).
A(t0,σ(tc))xa, then (2.4) yields
(2.4)
xT
bΦA(σ(tc),σ(t))B(t) = 0,t ∈ [t0,tf).
In particular, at t = tc, we have xT
to t,
bK0(tc) = 0. Differentiating (2.4) with respect
xT
bΦA(σ(tc),σ(t))K1(t) = 0,t ∈ [t0,tf),
Page 5
EJDE-2009/37CONTROLLABILITY, OBSERVABILITY, REALIZABILITY5
so that xT
bK1(tc) = 0. In general,
dj
∆tj
?xT
bΦT
A(σ(tc),σ(t))B(t)????
xT
b
?K0(tc)
t=tc= xT
bKj(tc) = 0,j = 0,1,...,q.
Thus,
K1(tc)...Kq(tc)?= 0,
which contradicts the linear independence of the rows guaranteed by the rank con-
dition. Hence, the equation is controllable on [t0,tf].
?
When T = R, the collection of matrices Kj(t) above is such that each member is
the jth derivative of the matrix ΦA(σ(t),σ(s))B(s) = ΦA(t,s)B(s). This coincides
with the literature in the continuous case (see, for example, [3, 8, 24]). However,
while still tractable, in general the collection Kj(t) is nontrivial to compute. The
mechanics are more involved even on Z, which is still a very “tame” time scale.
Therefore, the complications of extending the usual theory to the general time
scales case are evident even at this early juncture.
Furthermore, the preceding theorem shows that if the rank condition holds for
some q and some tc ∈ [t0,tf), then the linear state equation is controllable on
any interval [t0,tf] containing tc. This strong conclusion partly explains why the
condition is only a sufficient one.
2.2. Time Invariant Case. We now turn our attention to establishing results
concerning the controllability of regressive linear time invariant systems. The gen-
eralized Laplace transform presented in [13, 19] allows us to attack the problem in
ways that simply are not available in the time varying case.
First we recall a result from DaCunha in order to establish a preliminary technical
lemma.
Theorem 2.5. [12] For the system X∆(t) = AX(t), X(t0) = I, there exist scalar
functions {γ0(t,t0),...,γn−1(t,t0)} ⊂ C∞
representation
n−1
?
Lemma 2.6. Let A,B ∈ Rn×nand u := ux0(tf,σ(s)) ∈ Crd(T,Rn×1). Then
??tf
Proof. Let {γk(t,t0)}n−1
nential matrix as guaranteed by Theorem 2.5. This collection forms a linearly
independent set since it can be taken as the solution set of an n-th order sys-
tem of linear dynamic equations. Apply the Gram-Schmidt process to generate an
orthonormal collection {ˆ γk(t,t0)}n−1
?γ0(t,t0)
rd(T,R) such that the unique solution has
eA(t,t0) =
i=0
Aiγi(t,t0).
span
t0
eA(s,t0)Bux0(tf,σ(s))∆s
?
= span{B,AB,...,An−1B}.(2.5)
k=0be the collection of functions that decompose the expo-
k=0. The two collections are related by
γn−1(t,t0)?
γ1(t,t0)···
=?ˆ γ0(t,t0)ˆ γ1(t,t0)···ˆ γn−1(t,t0)?
p11
0
...
0
p12
p22
...
0
···
···
...
···
p1n
p2n
...
pnn
,
Page 6
6J. M. DAVIS, I. A. GRAVAGNE, B. J. JACKSON, R. J. MARKS II EJDE-2009/37
where the matrix on the right is the triangular matrix obtained from the QR fac-
torization of the vector consisting of the functions {γk(t,t0)}n−1
Using the QR factorization, we can write the matrix exponential as
k=0on the left.
eA(t,t0) =
n−1
?
n−1
?
k=0
γk(t,t0)Ak
=
k=0
?ˆ γ0(t,t0)ˆ γ1(t,t0) ···ˆ γn−1(t,t0)?pkAk,
where pkis the k-th column vector of the triangular matrix R. It is worth recalling
here that the entries on the diagonal of this matrix are norms of nonzero vectors
and are thus positive. That is, pii> 0 for all i.
Rewriting the integral from (2.5),
?tf
t0
eA(s,t0)Bux0(tf,σ(s))∆s
=
?tf
n−1
?
n−1
?
t0
n−1
?
k=0
γk(s,t0)AkBux0(tf,σ(s))∆s
=
k=0
AkB
?tf
?tf
t0
γk(s,t0)ux0(tf,σ(s))∆s
=
k=0
AkB
t0
?ˆ γ0(s,t0)ˆ γ1(s,t0) ···ˆ γn−1(s,t0)?pkux0(tf,σ(s))∆s.
Let
yk=
?tf
t0
?ˆ γ0(s,t0)ˆ γ1(s,t0)···ˆ γn−1(s,t0)?pkux0(tf,σ(s))∆s,
k = 0,1,...,n − 1. We will show that span{y0,y1,...,yn−1} = Rn. That is, there
exists some u ∈ Crd(T,Rn×1) so that for any arbitrary but fixed collection of vectors
{z0,z1,...,zn−1} ⊂ Rn×1, the system
?tf
t0
p11ˆ γ0(s,t0)ux0(tf,σ(s))∆s := z0=
z00
z01
...
z0(n−1)
?tf
t0
(ˆ γ0(s,t0)p12+ ˆ γ1(s,t0)p22)ux0(tf,σ(s))∆s := z1=
z10
z11
...
z1(n−1)
...
?tf
t0
(ˆ γ0(s,t0)p1n+ ··· + ˆ γn−1(s,t0)pnn)ux0(tf,σ(s))∆s := zn−1=
z(n−1)0
z(n−1)1
...
z(n−1)(n−1)
Page 7
EJDE-2009/37 CONTROLLABILITY, OBSERVABILITY, REALIZABILITY7
has a solution.
To accomplish this, we use the fact that the collection ˆ γk(s,t0) is orthonormal
and search for a solution of the form
ux0(tf,σ(s)) = (uj) =
?n−1
i=0
?
βj
iˆ γi(s,t0)
?
.
Starting with u0, the equations become
?tf
t0
ˆ γ0p11
n−1
?
i=0
β0
iˆ γi∆s = z00
?tf
t0
( ˆ γ0p12+ ˆ γ1p22)
n−1
?
i=0
β0
iˆ γi∆s = z10
...
?tf
t0
(ˆ γ0p1n+ ˆ γ1p2n+ ··· + ˆ γn−1pnn)
n−1
?
i=0
β0
iˆ γi∆s = z(n−1)0.
Since the system ˆ γkis orthonormal, we can simplify the equations above using the
fact that the integral of cross terms ˆ γiˆ γj, i ?= j, is zero. After doing so, the system
becomes a lower triangular system that can be solved by forward substitution. (The
observation that pii?= 0 is crucial here, since this is exactly what allows us to solve
the system.) For example, the first equation becomes
?tf
p11. Using this value for β0
t0
ˆ γ0p11
n−1
?
i=0
β0
iˆ γi∆s =
?tf
0in the second equation,
t0
ˆ γ02β0
0p11∆s = β0
0p11= z01,
so that β0
0=z00
?tf
t0
(ˆ γ0p12+ ˆ γ1p22)
n−1
?
i=0
β0
iˆ γi∆s =
?tf
p11z01+ β0
= z10,
t0
(ˆ γ0p12+ ˆ γ1p22)?β0
1p22
0ˆ γ0+ β0
1ˆ γ1
?∆s
=p12
so that β0
by using forward substitutions to find β0
turn yield u0=?n−1
We are now in a position to establish the following analogue of the Controllability
Rank Theorem (Theorem 2.4).
1=
1
p22z11−
p12
p11p22z01. We can continue solving the system in like manner
jfor all j = 0,1,...,n − 1, which will in
i=0β0
correct linear combinations of ˆ γkto solve the system, and so the claim follows.
iˆ γi. Repeating this process for u1,u2,...,un−1, we find the
?
Theorem 2.7 (Kalman Controllability Rank Condition). The time invariant re-
gressive linear system
x∆(t) = Ax(t) + Bu(t),x(t0) = x0,
y(t) = Cx(t) + Du(t),
is controllable on [t0,tf] if and only if the n × nm controllability matrix
?B AB ··· An−1B?
Page 8
8J. M. DAVIS, I. A. GRAVAGNE, B. J. JACKSON, R. J. MARKS II EJDE-2009/37
has rank n.
Proof. Suppose the system is controllable, but for the sake of a contradiction that
the rank condition fails. Then there exists an n × 1 vector xasuch that
xT
k = 0,...,n − 1.
Now, there are two cases to consider: either xT
Suppose xT
axf?= 0. Then for any t, the solution at time t is given by
?t
= eA(t,0) ∗ Bu(t) + eA(t,0)x0
= Bu(t) ∗ eA(t,0) + eA(t,0)x0
?t
where we have written the solution as a (time scale) convolution and appealed to
the commutativity of the convolution [13, 19]. Choose initial state x0= By, where
y is arbitrary. Then, again by commutativity of the convolution and Theorem 2.5,
aAkB = 0,
axf= 0 or xT
axf?= 0.
x(t) =
t0
eA(t,σ(s))Bux0(s)∆s + eA(t,t0)x0
=
t0
eA(s,t0)Bux0(t,σ(s))∆s + eA(t,t0)x0,
xT
ax(t) = xT
a
?t
n−1
?
t0
eA(s,t0)Bux0(t,σ(s))∆s + xT
aeA(t,t0)x0
=
?t
t0
k=0
γk(s,t0)xT
aAkBux0(t,σ(s))∆s +
n−1
?
k=0
γk(t,t0)xT
aAkBy
= 0,
so that xT
Now suppose xT
Similar to the equation above,
ax(t) = 0 for all t. This is a contradiction since xT
axf = 0. This time, we choose initial state x0= e−1
ax(tf) = xT
axf?= 0.
A(tf,t0)xa.
xT
ax(t) =
?t
aeA(t,t0)e−1
t0
n−1
?
k=0
γk(s,t0)xT
aAkBux0(t,σ(s))∆s + xT
aeA(t,t0)e−1
A(tf,t0)xa
= xT
A(tf,t0)xa.
In particular, at t = tf, xT
Thus in either case we arrive at a contradiction, and so controllability implies
the rank condition.
Conversely, suppose that the system is not controllable. Then there exists an
initial state x0∈ Rn×1such that for all input signals u(t) ∈ Rm×1, we have x(tf) ?=
xf. Again, it follows from the commutativity of the convolution that
ax(tf) = ?xa?2?= 0, another contradiction.
xf?= x(tf) =
?tf
?tf
?tf
t0
eA(tf,σ(s))Bux0(s)∆s + eA(tf,t0)x0
=
t0
eA(s,t0)Bux0(tf,σ(s))∆s + eA(tf,t0)x0
=
t0
n−1
?
k=0
γk(s,t0)AkBux0(tf,σ(s))∆s + eA(tf,t0)x0.
Page 9
EJDE-2009/37CONTROLLABILITY, OBSERVABILITY, REALIZABILITY9
In particular,
n−1
?
k=0
AkB
?tf
t0
γk(s,t0)ux0(tf,σ(s))∆s ?= xf− eA(tf,t0)x0.
Notice that the last equation holds if and only if there is no linear combination of
the matrices AkB for k = 0,1,...,n − 1, which satisfies
n−1
?
The fact that there is no such linear combination follows from Lemma 2.6 once we
realize that an argument similar to the one given in the proof of this result holds if
m < n. Thus, the matrix
?B AB ··· An−1B?
cannot have rank n, and so we have shown that if the matrix has rank n, then it is
controllable by contraposition.
k=0
AkBαk= xf− eA(tf,t0)x0.
?
The preceding theorem is commonly called the Kalman Rank Condition after
R. E. Kalman who first proved it in 1960 for the cases T = R and T = Z (see
[18, 20]). Therefore our analysis has unified the two cases, but we have also extended
these results to an arbitrary time scale with bounded graininess. However, it is
important to point out that the proof here is not the one that Kalman gave, which
is the one classically used for R and Z (see [24] for example). In these two special
cases, an observation about the particular form of the matrix exponential on R
and Z (due to the uniform graininess) allows one to arrive at the result in a more
straightforward manner. The general time scale case requires another argument
altogether as demonstrated above.
We now look at an example illustrating Theorem 2.7.
Example 2.8. Consider the system
x∆(t) =
?−8
45
1
30
−1
45
−1
y(t) =?3
10
?
x(t) +
?2
4?x(t).
?2
1
?
u(t),x(0) =
?5
2
?
,
It is straightforward to verify that
rank?B AB?= rank
−29/90
−13/901
?
= 2,
so that the state equation is controllable by Theorem 2.7.
The next theorem establishes that there is a state variable change in the time
invariant case that demonstrates the “controllable part” of the state equation.
Theorem 2.9. Suppose the controllability matrix for the time invariant regressive
linear system
x∆(t) = Ax(t) + Bu(t),x(t0) = x0,
y(t) = Cx(t),
satisfies
rank?B AB···An−1B?= q,
Page 10
10J. M. DAVIS, I. A. GRAVAGNE, B. J. JACKSON, R. J. MARKS IIEJDE-2009/37
where 0 < q < n. Then there exists an invertible matrix P such that
ˆA11
0(n−q)×q
whereˆA11is q × q,ˆB11is q × m, and
rank?ˆB11
Proof. We begin constructing P by choosing q linearly independent columns p1,
p2,..., pq, from the controllability matrix for the system. Then choose pq+1,...,pn
as n × 1 vectors so that
P =?p1
is invertible. Define G so that PG = B. Writing the j-th column of B as a linear
combination of the linearly independent columns of P given by p1,p2,...,pq, we
find that the last n−q entries of the j-th column of G must be zero. This argument
holds for j = 1,...,m, and so G = P−1B does indeed have the desired form.
Now set F = P−1AP, yielding
P−1AP =
?
ˆA12
ˆA22
?
,P−1B =
?
ˆB11
0(n−q)×m
?
,
ˆA11ˆB11
···
ˆAq−1
11
ˆB11
?= q.
···pq
pq+1
···pn
?
PF =?Ap1
Ap2
··· Apn
?.
The column vectors Ap1,...,Apqcan be written as linear combinations of p1,...,pn
since each column of AkB,k ≥ 0 can be written as a linear combination of these
vectors. As for G above, the first q columns of F must have zeros as the last n−q
entries. Thus, P−1AP has the desired form. Multiply the rank-q controllability
matrix by P−1to obtain
P−1?B
=?G
=
0
AB···An−1B?=?P−1BP−1AB ···P−1An−1B?
ˆAn−1
11
ˆB11
0
FG···Fn−1G?
···
...
?ˆB11
ˆA11ˆB11
0
?
.
Since the rank is preserved at each step, applying the Cayley-Hamilton theorem
gives
rank?ˆB11
Next, we use the preceding theorem to prove the following.
ˆA11ˆB11
···
ˆAq−1
11
ˆB11
?= q.
?
Theorem 2.10. The time invariant regressive linear system
x∆(t) = Ax(t) + Bu(t),x(t0) = x0,
y(t) = Cx(t),
is controllable if and only if for every scalar λ the only complex n × 1 vector p
satisfying pTA = λpT, pTB = 0 is p = 0.
Proof. For necessity, note that if there exists p ?= 0 and a complex λ such that the
equation given is satisfied, then
pT?B
=?pTB
so that the n rows rows of the controllability matrix are linearly dependent, and
hence the system is not controllable.
AB···An−1B?=?pTBpTAB···
···
pTAn−1B?
λn−1pTB?,
λpTB
Page 11
EJDE-2009/37CONTROLLABILITY, OBSERVABILITY, REALIZABILITY11
For sufficiency, suppose that the state equation is not controllable. Then by
Theorem 2.9, there exists an invertible P such that
ˆA11
ˆA12
0(n−q)×q
ˆA22
P−1AP =
?
?
,P−1B =
?
ˆB11
0(n−q)×m
?
,
with 0 < q < n. Let pT=?01×q
pT
qˆA22= λpT
q
?P−1, where pqis a left eigenvector forˆA22.
??ˆB11
ˆA12
ˆA22
Thus, for some complex scalar λ, pT
q, pq?= 0. Then p ?= 0, and
?
P−1=?0
pTB =?0
??ˆA11
pT
q
0
= 0,
pTA =?0pT
q
0
?
λpT
q
?P−1= λpT.
Thus, the claim follows.
?
The interpretation of Theorem 2.10 is that in a controllable time invariant sys-
tem, A can have no left eigenvectors that are orthogonal to the columns of B. This
fact can then be used to prove the next theorem.
Theorem 2.11. The time invariant regressive linear system
x∆(t) = Ax(t) + Bu(t),x(t0) = x0,
y(t) = Cx(t),
is controllable if and only if rank?zI − A
Proof. By Theorem 2.10, the state equation is not controllable if and only if there
exists a nonzero complex n × 1 vector p and complex scalar λ such that
pT?λI − A
But this condition is equivalent to rank?λI − A
3. Observability
B?= n for every complex scalar z.
B?p ?= 0.
B?< n.
?
Next, we turn our attention to observability of linear systems. As before, we
treat the time varying case first followed by the time invariant case.
3.1. Time Varying Case. In linear systems theory, when the term observability
is used, it refers to the effect that the state vector has on the output of the state
equation. As such, the concept is unchanged by considering simply the response of
the system to zero input. Motivated by this, we define the following.
Definition 3.1. The regressive linear system
x∆(t) = A(t)x(t),x(t0) = x0,
y(t) = C(t)x(t),
is observable on [t0,tf] if any initial state x(t0) = x0is uniquely determined by the
corresponding response y(t) for t ∈ [t0,tf).
The notions of controllability and observability can be thought of as duals of
one another, and so any theorem that we obtain for controllability should have an
analogue in terms of observability. Thus, we begin by formulating observability in
terms of an associated Gramian matrix.
Page 12
12 J. M. DAVIS, I. A. GRAVAGNE, B. J. JACKSON, R. J. MARKS IIEJDE-2009/37
Theorem 3.2 (Observability Gramian Condition). The regressive linear system
x∆(t) = A(t)x(t),x(t0) = x0,
y(t) = C(t)x(t),
is observable on [t0,tf] if and only if the n × n observability Gramian matrix
?tf
is invertible.
GO(t0,tf) :=
t0
ΦT
A(t,t0)CT(t)C(t)ΦA(t,t0)∆t,
Proof. If we multiply the solution expression
y(t) = C(t)ΦA(t,t0)x0,
on both sides by ΦT
A(t,t0)C(t) and integrate, we obtain
?tf
t0
ΦT
A(t,t0)CT(t)y(t)∆t = GO(t0,tf)x0.
The left side of this equation is determined by y(t) for t ∈ [t0,tf), and thus this
equation is a linear algebraic equation in x0. If GO(t0,tf) is invertible, then x0is
uniquely determined.
Conversely, if GO(t0,tf) is not invertible, then there exists a nonzero vector xa
such that GO(t0,tf)xa= 0. But then xT
aGO(t0,tf)xa= 0, so that
t ∈ [t0,tf).C(t)ΦA(t,t0)xa= 0,
Thus, x(t0) = x0+xayields the same zero-input response for the system as x(t0) =
x0, and so the system is not observable on [t0,tf].
?
The observability Gramian, like the controllability Gramian, is symmetric posi-
tive semidefinite. It is positive definite if and only if the state equation is observable.
Once again we see that the Gramian condition is not very practical as it requires
explicit knowledge of the transition matrix. Thus, we present a sufficient condition
that is easier to check for observability. As before, observability and controllability
can be considered dual notions to one another, and as such, proofs of corresponding
results are often similar if not the same. Any missing observability proofs missing
below simply indicates that the controllability analogue should be consulted.
Definition 3.3. If T is a time scale such that µ is sufficiently differentiable with the
indicated rd-continuous derivatives, define the sequence of p × n matrix functions
L0(t) := C(t),
Lj+1(t) := Lj(t)A(t) + L∆
j(t)(I + µ(t)A(t)),j = 0,1,2,...
As in the case of controllability, an induction argument shows that
∂j
∆tj[C(t)ΦA(t,s)]
With this, an argument similar to the one before shows the following:
Lj(t) =
???
s=t.
Theorem 3.4 (Observability Rank Condition). Suppose q ∈ Z+is such that, for
t ∈ [t0,tf], C(t) is q-times rd-continuously differentiable and both of µ(t) and A(t)
are (q − 1)-times rd-continuously differentiable. Then the regressive linear system
x∆(t) = A(t)x(t),x(t0) = x0,
Page 13
EJDE-2009/37CONTROLLABILITY, OBSERVABILITY, REALIZABILITY13
y(t) = C(t)x(t),
is observable on [t0,tf] if for some tc∈ [t0,tf), we have
rank
L0(tc)
L1(tc)
...
Lq(tc)
= n,
where
Lj(t) =
∂j
∆sj[C(t)ΦA(t,s)]
???
s=t,j = 0,1,...,q.
3.2. Time Invariant Case. Like controllability, observability has equivalent con-
ditions that become necessary and sufficient in the time invariant case. We begin
with a Kalman rank condition for observability.
Theorem 3.5 (Kalman Observability Rank Condition).
gressive linear system
The time invariant re-
x∆(t) = Ax(t),x(t0) = x0,
y(t) = Cx(t),
is observable on [t0,tf] if and only if the nm × n observability matrix
C
CA
...
CAn−1
has rank n.
Proof. Again, we show that the rank condition fails if and only if the observability
Gramian is not invertible. Thus, suppose that the rank condition fails. Then there
exists a nonzero n × 1 vector xasuch that
CAkxa= 0,k = 0,...,n − 1.
This implies, using Theorem 2.5, that
GO(t0,tf)xa=
?tf
?tf
t0
eT
A(t,t0)CTCeA(t,t0)xa∆t
=
t0
eT
A(t,t0)CT
n−1
?
k=0
γk(t,t0)CAkxa∆t
= 0,
so that the Gramian is not invertible.
Conversely, suppose that the Gramian is not invertible.
nonzero xasuch that xT
Then there exists
aGO(t0,tf)xa= 0. As argued before, this implies
CeA(t,t0)xa= 0,t ∈ [t0,tf).
Page 14
14J. M. DAVIS, I. A. GRAVAGNE, B. J. JACKSON, R. J. MARKS II EJDE-2009/37
At t = t0, we obtain Cxa= 0, and differentiating k times and evaluating the result
at t = t0gives CAkxa= 0, k = 0,...,n − 1. Thus,
and the rank condition fails.
C
CA
...
CAn−1
xa= 0,
?
The proof of the preceding result demonstrates an important point about con-
trollability and observability in the arbitrary time scale setting: namely, proofs
of similar results for the two notions are often similar, but can sometimes be very
different. Indeed, comparing the proof of the Kalman condition for controllability
with the proof of the Kalman condition for observability highlights this contrast.
The following example uses Theorem 3.5.
Example 3.6. Consider the system
x∆(t) =
?−8
45
1
30
−1
45
−1
y(t) =?3
10
?
x(t) +
?2
4?x(t).
1
?
u(t),x(0) =
?5
2
?
,
From Example 2.8, recall that the system is controllable. We claim the system is
also observable. This follows from Theorem 3.5 since
?C
The following three theorems concerning observability have proofs that mirror
their controllability counterparts, and so will not be given here.
rank
CA
?
= rank
?
34
−28
45
−3
10
?
= 2.
Theorem 3.7. Suppose the observability matrix for the time invariant regressive
linear system
x∆(t) = Ax(t) + Bu(t),x(t0) = x0,
y(t) = Cx(t),
satisfies
rank
C
CA
...
CAn−1
= ℓ,
where 0 < ℓ < n. Then there exists an invertible n × n matrix Q such that
Q−1AQ =
ˆA21
?ˆA11
0
ˆA22
?
, CQ =?ˆC11
0?,
whereˆA11is ℓ × ℓ,ˆC11is p × ℓ, and
rank
ˆC11
ˆC11ˆA11
...
ˆC11ˆAℓ−1
11
= ℓ.
Page 15
EJDE-2009/37CONTROLLABILITY, OBSERVABILITY, REALIZABILITY 15
The state variable change Theorem 3.7 is constructed by choosing n − ℓ vectors
in the nullspace of the observability matrix, and preceding them by ℓ vectors that
yield a set of n linearly independent vectors.
Theorem 3.8. The time invariant regressive linear system
x∆(t) = Ax(t) + Bu(t),x(t0) = x0,
y(t) = Cx(t),
is observable if and only if for every complex scalar λ, the only complex n×1 vector
p that satisfies Ap = λp, Cp = 0 is p = 0.
Again, Theorem 3.8 can be restated as saying that in an observable time invariant
system, A can have no right eigenvectors that are orthogonal to the rows of C.
Theorem 3.9. The time invariant regressive linear system
x∆(t) = Ax(t) + Bu(t),x(t0) = x0,
y(t) = Cx(t),
is observable if and only if
rank
?
C
zI − A
?
= n,
for every complex scalar z.
4. Realizability
In linear systems theory, the term realizability refers to the ability to characterize
a known output in terms of a linear system with some input. We now make this
precise.
Definition 4.1. The regressive linear system
x∆= A(t)x(t) + B(t)u(t),x(t0) = 0,
y(t) = C(t)x(t),
of dimension n is a realization of the weighting pattern G(t,σ(s)) if
G(t,σ(s)) = C(t)ΦA(t,σ(s))B(s),
for all t,s. If a realization of this system exists, then the weighting pattern is
realizable. The system is a minimal realization if no realization of G(t,σ(s)) with
dimension less than n exists.
Notice that for the system
x∆(t) = A(t)x(t) + B(t)u(t),x(t0) = 0,
y(t) = C(t)x(t) + D(t)u(t),
the output signal y(t) corresponding to a given input u(t) and weighting pattern
G(t,σ(s)) = C(t)ΦA(t,σ(s))B(s) is given by
?t
When there exists a realization of a particular weighting response G(t,σ(s), there
will in fact exist many since a change of state variables will leave the weighting
pattern unchanged. Also, there can be many different realizations of the same
y(t) =
t0
G(t,σ(s))u(s)∆s + D(t)u(t),t ≥ t0.
Page 16
16 J. M. DAVIS, I. A. GRAVAGNE, B. J. JACKSON, R. J. MARKS IIEJDE-2009/37
weighting pattern that all have different dimensions. This is why we are careful to
distinguish between realizations and minimal realizations in our definition.
We now give equivalent conditions for realizability: as before, we begin with the
time variant case and then proceed to the time invariant case.
4.1. Time Varying Case. The next theorem gives a characterization of realizable
systems in general.
Theorem 4.2 (Factorization of G(t,σ(s))). The weighting pattern G(t,σ(s)) is
realizable if and only if there exist a rd-continuous matrix H(t) that is of dimension
q × n and a rd-continuous matrix F(t) of dimension n × r such that
G(t,σ(s)) = H(t)F(σ(s)),
for all t,s.
Proof. Suppose there exist matrices H(t) and F(t) with G(t,σ(s)) = H(t)F(σ(s)).
Then the system
x∆(t) = F(t)u(t),
y(t) = H(t)x(t),
is a realization of G(t,σ(s)) since the transition matrix of the zero system is the
identity.
Conversely, suppose that G(t,σ(s)) is realizable. We may assume that the system
x∆(t) = A(t)x(t) + B(t)u(t),
y(t) = C(t)x(t),
is one such realization. Since the system is regressive, we may write
G(t,σ(s)) = C(t)ΦA(t,σ(s))B(s) = C(t)ΦA(t,0)ΦA(0,σ(s))B(s).
Choosing H(t) = C(t)ΦA(t,0) and F(t) = ΦA(0,σ(t))B(t), the claim follows.
?
Although the preceding theorem gives a basic condition for realization of linear
systems, often in practice it is not very useful because writing the weighting pattern
in its factored form can be very difficult. Also, as the next example demonstrates,
the realization given by the factored form can often be undesirable for certain
analyses.
Example 4.3. Suppose T is a time scale with 0 ≤ µ ≤ 2. Under this assumption,
−1/4 ∈ R+(T). Then the weighting pattern
G(t,σ(s)) = e−1/4(t,σ(s)),
has the factorization
G(t,σ(s)) = e−1/4(t,σ(s)) = e−1/4(t,0)e⊖(−1/4)(σ(s),0).
By the previous theorem, a one-dimensional realization of G is
x∆(t) = e⊖(−1/4)(t,0)u(t),
y(t) = e−1/4(t,0)x(t).
This state equation has an unbounded coefficient and is not uniformly exponentially
stable (note that e⊖(−1/4)(t,0) = e1/(4−µ)(t,0) is unbounded since 1/(4 − µ) > 0).
However, the one-dimensional realization of G given by
x∆(t) = −1
4x(t) + u(t),
Page 17
EJDE-2009/37CONTROLLABILITY, OBSERVABILITY, REALIZABILITY 17
y(t) = x(t),
does have bounded coefficients and is uniformly exponentially stable.
Before examining minimal realizations, some remarks are in order. First, note
that the inverse and σ operators commute:
P−1(σ(t)) = P−1(t) + µ(t)(P−1(t))∆
= P−1(t) + µ(t)(−P(σ(t)))−1P∆(t)P−1(t)
= P−1(t) − (P(σ(t))−1(P(σ(t)) − P(t))P−1(t)
= (P(σ(t)))−1.
Second, it is possible to do a variable change on the system
x∆(t) = A(t)x(t) + B(t)x(t),
y(t) = C(t)x(t),
so that the coefficient of x∆(t) in the new system is zero, while at the same time
preserving realizability of the system under the change of variables.
Indeed, set z(t) = P−1(t)x(t) and note that P(t) = ΦA(t,t0) satisfies
(P(σ(t)))−1A(t)P(t) − (P(σ(t)))−1P∆(t) = 0.
If we make this substitution, then the system becomes
z∆(t) = P−1(σ(t))B(t)u(t),
y(t) = C(t)P(t)z(t).
Thus, in terms of realizability, we may assume without loss of generality that A(t) ≡
0 by changing the system to the form given above.
It is important to know when a given realization is minimal. The following
theorem gives a necessary and sufficient condition for this in terms of controllability
and observability.
Theorem 4.4 (Characterization of Minimal Realizations). Suppose the regressive
linear system
x∆(t) = A(t)x(t) + B(t)x(t),
y(t) = C(t)x(t),
is a realization of the weighting pattern G(t,σ(s)). Then this realization is minimal
if and only if for some t0 and tf > t0 the state equation is both controllable and
observable on [t0,tf].
Proof. As argued above, we may assume without loss of generality that A(t) ≡ 0.
Suppose the n-dimensional realization given is not minimal. Then there is a lower
dimension realization of G(t,σ(s)) of the form
z∆(t) = R(t)u(t),
y(t) = S(t)z(t),
where z(t) has dimension nz< n. Writing the weighting pattern in terms of both
realizations produces C(t)B(s) = S(t)R(s) for all t,s. Thus,
CT(t)C(t)B(s)BT(s) = CT(t)S(t)R(s)BT(s),
Page 18
18J. M. DAVIS, I. A. GRAVAGNE, B. J. JACKSON, R. J. MARKS II EJDE-2009/37
for all t,s. For any t0and tf > t0, it is possible to integrate this expression with
respect to t and then with respect to s to obtain
GO(t0,tf)GC(t0,tf) =
?tf
t0
CT(t)S(t)∆t
?tf
t0
R(s)BT(s)∆s.
The right hand side of this equation is the product of an n × nz matrix and an
nz×n matrix, and as such, it cannot have full rank since the dimension of the space
spanned by the product is at most nz < n. Therefore, GO(t0,tf) and GC(t0,tf)
cannot be simultaneously invertible. The argument is independent of the t0and tf
chosen, and so sufficiency is established.
Conversely, suppose that the given state equation is a minimal realization of the
weighting pattern G(t,σ(s)), with A(t) ≡ 0. We begin by showing that if either
?tf
or
?tf
is singular for all t0and tf, then minimality is violated. Thus, there exist intervals
[ta
let t0 = min{ta
observability and controllability Gramians yields that both GC(t0,tf) and GO(t0,tf)
are invertible.
To show this, we begin by supposing that for every interval [t0,tf] the matrix
GC(t0,tf) is not invertible. Then, given t0 and tf there exists an n × 1 vector
x = x(t0,tf) such that
GC(t0,tf) =
t0
B(t)BT(t)∆t,
GO(t0,tf) =
t0
CT(t)C(t)∆t,
0,ta
f] and [tb
0,tb
0,tb
f] such that GC(ta
0} and tf = max{ta
0,ta
f) and GO(tb
f,tb
0,tb
f) are both invertible. If we
f}, then the positive definiteness of the
0 = xTGC(t0,tf)x =
?tf
t0
B(t)BT(t)x∆t.
Thus, xTB(t) = 0 for t ∈ [t0,tf).
We claim that there exists at least one such vector x that is independent of t0
and tf. To this end, note that if T is unbounded from above and below, then for
each positive integer k there exists an n × 1 vector xkwith
?xk? = 1,
Thus, {xk}∞
strauss Theorem, it has a convergent subsequence since T is closed. We label this
convergent subsequence by {xkj}∞
xT
0B(t) = 0 for all t, since for any given time ta, there exists a positive integer Ja
such that ta∈ [−kj,kj] for all j ≥ Ja, which in turn implies xT
j ≥ Ja. Hence, xT
Now let P−1be a constant, invertible, n×n matrix with bottom row xT
P−1as a change of state variables gives another minimal realization of the weighting
pattern, with coefficient matrices
xT
kB(t) = 0,t ∈ [−k,k].
k=1is a bounded sequence of n × 1 vectors and by the Bolzano-Wier-
j=1and denote its limit by x0= limj→∞xkj. Note
kjB(ta) = 0 for all
0satisfies xT
0B(ta) = 0.
0. Using
P−1B(t) =
?ˆB1(t)
01×m
?
,C(t)P =?ˆC1(t)
ˆC2(t)?,
Page 19
EJDE-2009/37CONTROLLABILITY, OBSERVABILITY, REALIZABILITY19
whereˆB1(t) is (n − 1) × m, andˆC1(t) is p × (n − 1). Then a straightforward
calculation shows G(t,σ(s)) =ˆC1(t)ˆB1(σ(s)) so that the linear state equation
z∆(t) =ˆB1(t)u(t),
y(t) =ˆC1(t)z(t),
is a realization for G(t,σ(s)) of dimension n − 1. This contradicts the minimality
of the original n-dimensional realization. Thus, there must be at least one ta
one ta
f) is invertible.
A similar argument shows that there exists at least one tb
such that GO(tb
shows that the minimal realization of the state equation is both controllable and
observable on [t0,tf].
0and
f> ta
0such that GC(ta
0,ta
0and one tb
0} and tf = max{ta
f> tb
f,tb
0
0,tb
f) is invertible. Taking t0 = min{ta
0,tb
f}
?
4.2. Time Invariant Case. We now restrict ourselves to the time invariant case
and use a Laplace transform approach to establish our results. Instead of consid-
ering the time-domain description of the input-output behavior given by
y(t) =
?t
0
G(t,σ(s))u(s)∆s,
we examine the corresponding behavior in the z-domain. Laplace transforming the
equation above and using the Convolution Theorem [13] yields Y (z) = G(z)U(z).
The question is: given a transfer function G(z), when does there exist a time
invariant form of the state equation such that
C(zI − A)−1B = G(z),
and when is this realization minimal?
To answer this, we begin by characterizing time invariant realizations. In what
follows, a strictly-proper rational function of z is a rational function of z such that
the degree of the numerator is strictly less than the degree of the denominator.
Theorem 4.5. The p×q transfer function G(z) admits a time invariant realization
of the regressive linear system
x∆(t) = Ax(t) + Bu(t),
y(t) = Cx(t),
if and only if each entry of G(z) is a strictly-proper rational function of z.
Proof. If G(z) has a time invariant realization, then G has the form G(z) = C(zI−
A)−1B. We showed in [13, 19] that for each Laplace transformable function f(t),
F(z) → 0 as z → ∞, which in turn implies that F(z) is a strictly-proper rational
function in z. Thus, the matrix (zI − A)−1is a matrix of strictly-proper rational
functions, and G(z) is a matrix of strictly-proper rational functions since linear
combinations of such functions are themselves strictly-proper and rational.
Conversely, suppose that each entry Gij(z) in the matrix is strictly-proper and
rational. Without loss of generality, we can assume that each polynomial in the
denominator is monic (i.e. has leading coefficient of 1). Suppose
d(z) = zr+ dr−1zr−1+ ··· + d0
Page 20
20 J. M. DAVIS, I. A. GRAVAGNE, B. J. JACKSON, R. J. MARKS II EJDE-2009/37
is the least common multiple of the polynomials in denominators. Then d(z)G(z)
can be decomposed as a polynomial in z with p × q constant coefficient matrices,
so that
d(z)G(z) = Pr−1zr−1+ ··· + P1z + P0.
We claim that the qr-dimensional matrices given by
A =
0q
0q
...
0q
Iq
0q
...
0q
...
...
...
...
...
0q
0q
...
0q
−d0Iq
−d1Iq
−dr−1Iq
,B =
0q
0q
...
0q
Iq
,C =
P0
P1
...
Pr−1
,
form a realization of G(z). To see this, let
R(z) = (zI − A)−1B,
and partition the qr × q matrix R(z) into r blocks R1(z),R2(z),...,Rr(z), each
of size q × q. Multiplying R(z) by (zI − A) and writing the result in terms of
submatrices gives rise to the relations
Ri+1(z) = zRi,i = 1,...,r − 1, (4.1)
zRr(z) + d0R1(z) + d1R2(z) + ··· + dr−1Rr(z) = Iq.
Using (4.1) to rewrite (4.2) in terms of R1(z) gives
(4.2)
R1(z) =
1
d(z)Iq,
and thus from (4.1) again, we have
R(z) =
1
d(z)
Iq
zIq
...
zr−1Iq
.
Multiplying by C yields
C(zI − A)−1B =
1
d(z)
?P0+ zP1+ ··· + zr−1Pr−1
?= G(z),
which is a realization of G(z).
?
The realizations that are minimal are characterized in the following theorem,
which is repeated here for completeness sake in this easier case of Theorem 4.4.
Theorem 4.6. Suppose the time invariant regressive linear system
x∆(t) = Ax(t) + Bx(t),
y(t) = Cx(t),
is a realization of the transfer function G(z). Then this state equation is a minimal
realization of G(z) if and only if it is both controllable and observable.
Proof. Suppose the state equation is a realization of G(z) that is not minimal. Then
there is a realization of G(z) given by
z∆(t) = Pz(t) + Qz(t),
y(t) = Rz(t),
Page 21
EJDE-2009/37CONTROLLABILITY, OBSERVABILITY, REALIZABILITY 21
with dimension nz< n. Thus,
CeA(t,0)B = ReP(t,0)Q,t ≥ 0.
Repeated differentiation with respect to t, followed by evaluation at t = 0 yields
CAkB = RFkQ,k = 0,1,...
Rewriting this information in matrix form for k = 0,1,...,2n − 2, we see
which can be rewritten as
CB
...
CAB
...
CAnB
···
...
···
CAn−1B
...
CA2n−2B CAn−1B
=
RQ
...
RPQ
...
RPnQ
···
...
···
RPn−1Q
...
RP2n−2Q RPn−1Q
,
C
CA
...
CAn−1
?B AB ··· An−1B?=
R
RP
...
RPn−1
?Q PQ ··· Pn−1Q?.
However, since the right hand side of the equation is the product of an nzp × nz
and an nz× nzm matrix, the rank of the product can be no greater than nz.
Thus, nz< n, which implies that that the realization given in the statement of the
theorem cannot be both controllable and observable. Therefore, by contraposition
a controllable and observable realization must be minimal.
Conversely, suppose the state equation given in the statement of the theorem is
a minimal realization that is not controllable. Then there exists an n × 1 vector
y ?= 0 such that
yT?B AB ··· An−1B?= 0,
which implies yTAkB = 0 for all k ≥ 0 by the Cayley-Hamilton theorem. For
P−1an invertible n × n matrix with bottom row yT, then a variable change of
z(t) = P−1x(t) produces the state equations
z∆(t) =ˆAz(t) +ˆBu(t),
y(t) =ˆCz(t),
which is also an n-dimensional minimal realization of G(z). Partition the coefficient
matrices of the state equation above as
ˆA = P−1AP =
?ˆA11
ˆA12
ˆA22
ˆA21
?
,
ˆB = P−1B =
?ˆB1
0
?
,
ˆC = CP =?ˆC1
CA?,
whereˆA11is (n−1)×(n−1),ˆB1is (n−1)×1, andˆC1is 1×(n−1). From these
partitions, it follows from the construction of P thatˆAˆB = P−1AB has the form
ˆAˆB =
?ˆA11ˆB1
ˆA21ˆB1
?
=
?ˆA11ˆB1
0
?
.
Since the bottom row of P−1AkB is zero for all k ≥ 0,
ˆAkˆB =
?ˆAk
11ˆB1
0
?
,k ≥ 0.
Page 22
22J. M. DAVIS, I. A. GRAVAGNE, B. J. JACKSON, R. J. MARKS II EJDE-2009/37
But,ˆA11,ˆB1,ˆC1give an (n − 1)-dimensional realization of G(z) since
ˆCeˆ A(t,0)ˆB =?ˆC1
ˆC2
?
?
∞
?
∞
?
k=0
ˆAkˆBhk(t,0)
=?ˆC1
=ˆC1eˆ A11(t,0)ˆB1,
ˆC2
k=0
?ˆAk
11ˆB1
0
?
hk(t,0)
so that the state equation in the statement of the theorem is in fact not minimal,
a contradiction. A similar argument holds if the system is assumed not to be
observable.
?
We now illustrate Theorem 4.5 and Theorem 4.6 with an example.
Example 4.7. Consider the transfer function
G(z) =
9(37 + 300z)
5 + 75z + 270z2.
G(z) admits a time invariant realization by Theorem 4.5 since G(z) is a strictly-
proper rational function of z. The form of G(z) indicates that we should look for a
2-dimensional realization with a single input and single output. We can write
G(z) =?34??
zI −
?−8
45
1
30
−1
45
−1
10
??−1?2
1
?
,
so that a time invariant realization of G(z) is given by
x∆(t) =
?−8
45
1
30
−1
45
−1
y(t) =?3
10
?
x(t) +
?2
1
?
u(t),x(0) = x0,
4?x(t).
We showed in Example 2.8 that this realization is in fact controllable, and we showed
in Example 3.6 that it is also observable. Thus, Theorem 4.6 guarantees that this
realization of G(z) is minimal.
5. Stability
We complete our foray into linear systems theory by considering stability. In
[23], P¨ otzsche, Siegmund, and Wirth deal with exponential stability. DaCunha
also deals with this concept under a different definition in [10, 11] and emphasizes
the time varying case. We begin by revisiting exponential stability in the time
invariant case and then proceed to another notion of stability commonly used in
linear systems theory.
5.1. Exponential Stability in the Time Invariant Case. We start this section
by revisiting the notion of exponential stability. We are interested in both the
time invariant and time varying cases separately since it is often possible to obtain
stronger results in the time invariant case.
We have already noted that if A is constant, then ΦA(t,t0) = eA(t,t0). In what
follows, we will consider time invariant systems with t0= 0 in order to talk about
the Laplace transform.
DaCunha defines uniform exponential stability as follows.
Page 23
EJDE-2009/37CONTROLLABILITY, OBSERVABILITY, REALIZABILITY 23
Definition 5.1. [11] The regressive linear system
x∆= A(t)x(t),x(t0) = x0,
is uniformly exponentially stable if there exist constants γ,λ > 0 with −λ ∈ R+
such that for any t0and x(t0), the corresponding solution satisfies
?x(t)? ≤ ?x(t0)?γe−λ(t,t0),t ≥ t0.
With this definition of exponential stability, we can prove the next theorem.
Theorem 5.2. The time invariant regressive linear system
x∆(t) = Ax(t),x(t0) = x0,
is uniformly exponentially stable if and only if for some β > 0,
?∞
t0
?eA(t,t0)?∆t ≤ β.
Proof. For necessity, note that if the system is uniformly exponentially stable, then
by [10, Theorem 3.2],
?∞
0
?eA(t,0)?∆t ≤
?∞
0
γe−λ(t,0)∆t =γ
λ,
so that the claim follows.
For sufficiency, assume the integral condition holds but for the sake of contra-
diction that the system is not exponentially stable. Then, again by [10, Theorem
3.2], for all λ,γ > 0 with −λ ∈ R+, we have ?eA(t,0)? > γe−λ(t,0). Hence,
?∞
In particular, if we choose γ > βλ, then
0
?eA(t,0)?∆t >
?∞
0
γe−λ(t,0)∆t =
γ
−λe−λ(t,0)
???
∞
0
=γ
λ.
?∞
0
?eA(t,0)?∆t >βλ
λ
= β,
a contradiction.
?
Now consider the system
x∆(t) = Ax(t),x(0) = I.
Transforming this system yields
X(z) = (zI − A)−1,
which is the transform of eA(t,0). This result is unique as argued in [13, 19]. Note
that this matrix contains only strictly-proper rational functions of z since we have
the formula
(zI − A)−1=adj(zI − A)
det(zI − A).
Specifically, det(zI − A) is an nth degree polynomial in z, while each entry of
adj(zI − A) is a polynomial of degree at most n − 1. Suppose
det(zI − A) = (z − λ1)ψ1(z − λ2)ψ2···(z − λm)ψm,
Page 24
24 J. M. DAVIS, I. A. GRAVAGNE, B. J. JACKSON, R. J. MARKS IIEJDE-2009/37
where λ1,λ2,...,λnare the distinct eigenvalues of the n×n matrix A, with corre-
sponding multiplicities ψ1,ψ2,...,ψm. Decomposing (zI −A)−1in terms of partial
fractions gives
m
?
where each Wkj is an n × n matrix of partial fraction expansion coefficients given
by
1
(ψk− j)!
If we now take the inverse Laplace transform of (zI−A)−1in the form given above,
we obtain the representation
(zI − A)−1=
k=1
ψk
?
j=1
Wkj
1
(z − λk)j,
Wkj=
dψk−j
dzψk−j
?(z − λk)ψk(zI − A)−1????
z=λk.
eA(t,0) =
m
?
k=1
ψk
?
j=1
Wkjfj−1(µ,λk)
(j − 1)!
eλk(t,0), (5.1)
where fj(µ,λk) is the sequence of functions obtained from the residue calculations
of the jth derivative in the inversion formula. For example, the first few terms in
the sequence are
f0(µ,λk) = 1,
f1(µ,λk) =
?t
??t
??t
+
0
1
1 + µλk
∆τ,
f2(µ,λk) =
0
1
1 + µλk
1
1 + µλk
2µ2
(1 + µλk)3∆τ,
∆τ
?2
?3
−
?t
0
?t
µ
(1 + µλk)2∆τ,
µ
(1 + µλk)2∆τ
f3(µ,λk) =
0
∆τ
− 3
0
?t
0
1
1 + µλk
∆τ
?t
0
...
Notice that if µ is bounded, then each fj(µ,λk) can be bounded by a “regular”
polynomial of degree j in t, call it aj(t). That is, fj can be bounded by functions
of the form aj(t) = ajtj+ aj−1tj−1··· + a0. This observation will play a key role
in the next theorem. P¨ otzsche, Siegmund, and Wirth do prove this result in [23],
but our proof differs from theirs in that we use new transform results to obtain
it, while they use other techniques. Note, however, that in the theorem we do use
their definition of exponential stability rather than the one given by DaCunha. For
completeness, we remind the reader by restating their definition here.
Definition 5.3. [23] For t,t0∈ T and x0∈ Rn, the system
x∆= A(t)x,x(t0) = x0,
is uniformly exponentially stable if there exists a constant α > 0 such that for every
t0∈ T there exists a K ≥ 1 with
?ΦA(t,t0)? ≤ Ke−α(t−t0)for t ≥ t0,
with K being chosen independently of t0.
Page 25
EJDE-2009/37CONTROLLABILITY, OBSERVABILITY, REALIZABILITY 25
Recall that DaCunha’s definition of uniform exponential stability of a system will
imply that the system is uniformly exponential stable if we use P¨ otzsche, Siegmund,
and Wirth’s definition of the concept, but the converse need not be true in general.
Thus, DaCunha’s definition is weaker in this sense.
Theorem 5.4 (Spectral Characterization of Exponential Stability). Let T be a
time scale which is unbounded above but has bounded graininess. The time invariant
regressive linear system
x∆(t) = Ax(t),x(t0) = x0,
is uniformly exponentially stable (in the sense of Definition 5.3) if and only if
spec(A) ⊂ S(T), the regressive set of exponential stability for T, given by
1
T − t0
Proof. Suppose the eigenvalue condition holds. Then, appealing to Theorem 5.2
and writing the exponential in the explicit form given above in terms of the distinct
eigenvalues λ1,λ2,...,λm, we obtain
S(T) := {λ ∈ C : limsup
T→∞
?T
t0
lim
sցµ(t)
log|1 + sλ|
s
∆t < 0}.
?∞
0
?eA(t,0)?∆t =
?∞
m
?
m
?
m
?
m
?
0
???
m
?
k=1
ψk
?
j=1
Wkjfj−1(µ,λk)
(j − 1)!
?∞
?∞
?∞
?∞
eλk(t,0)
???∆t
???∆t
≤
k=1
ψk
?
ψk
?
ψk
?
ψk
?
j=1
?Wkj?
0
???fj−1(µ,λk)
|aj−1(t)eλk(t,0)| ∆t
(j − 1)!
eλk(t,0)
≤
k=1
j=1
?Wkj?
0
≤
k=1
j=1
?Wkj?
0
aj−1(t)e−αt∆t
≤
k=1
j=1
?Wkj?
0
aj−1(t)e−αtdt < ∞.
Note that the last three lines hold by appealing to Definition 5.3. Thus, by Theo-
rem 5.2 the system is uniformly exponentially stable.
Now, for the sake of a contradiction, assume that the eigenvalue condition fails.
Let λ be an eigenvalue of A with associated eigenvector v, with λ / ∈ S(C). Direct
calculation shows that the solution of the system x∆= Ax, x(0) = v, is given by
x(t) = eλ(t,0)v. From [23], if λ / ∈ S(C), then limt→∞eλ(t,0) ?= 0, so that we arrive
at a contradiction.
?
5.2. BIBO Stability in the Time Varying Case. Besides exponential stability,
the concept of bounded-input, bounded-output stability is also a useful property
for a system to have. As its name suggests, the notion is one that compares the
supremum of the output signal with the supremum of the input signal. Thus, we
define the term as follows.
Definition 5.5. The regressive linear system
x∆(t) = A(t)x(t) + B(t)u(t),x(t0) = x0,
Page 26
26J. M. DAVIS, I. A. GRAVAGNE, B. J. JACKSON, R. J. MARKS II EJDE-2009/37
y(t) = C(t)x(t),
is said to be uniformly bounded-input, bounded-output (BIBO) stable if there exists
a finite constant η such that for any t0and any input u(t) the corresponding zero-
state response satisfies
sup
t≥t0
?y(t)? ≤ η sup
t≥t0
?u(t)?.
Note that we use the word “uniform” to stress that the same η works for all t0
and all input signals.
The following characterization of BIBO stability is useful.
Theorem 5.6. The regressive linear system
x∆(t) = A(t)x(t) + B(t)u(t),x(t0) = x0,
y(t) = C(t)x(t),
is uniformly bounded-input, bounded-output stable if and only if there exists a finite
constant ρ such that for all t,τ with t ≥ τ,
?t
Proof. Assume such a ρ exists. Then for any t0 and any input signal, the corre-
sponding zero-state response of the state equation satisfies
?t
Replacing ?u(s)? by its supremum over s ≥ t0, and using the integral condition,
we obtain
?t
Thus, the system is BIBO stable.
Conversely, suppose the state equation is uniformly BIBO stable. Then there
exists a constant η so that, in particular, the zero-state response for any t0and any
input signal such that supt≥t0?u(t)? ≤ 1 satisfies supt≥t0?y(t)? ≤ η. For the sake
of a contradiction, suppose no finite ρ exists that satisfies the integral condition.
Then for any given ρ > 0, there exist τρand tρ> τρsuch that
?tρ
In particular, if ρ = η, this implies that there exist τη, with tη > τη, and indices
i,j such that the i,j-entry of the impulse response satisfies
?tη
With t0= τηconsider the m × 1 input signal u(t) defined for t ≥ t0as follows: set
u(t) = 0 for t > tη, and for t ∈ [t0,tη] set every component of u(t) to zero except
for the j-th component given by the piecewise continuous signal
τ
?G(t,σ(s))?∆s ≤ ρ.
?y(t)? =
???? ????
t0
C(t)ΦA(t,σ(s))B(s)u(s)∆s
???? ????≤
?t
t0
?G(t,σ(s))??u(s)?∆s,t ≥ t0.
?y(t)? ≤ sup
t≥t0
?u(t)?
t0
?G(t,σ(s))?∆s ≤ ρ sup
t≥t0
?u(t)?,t ≥ t0.
τρ
?G(tρ,σ(s))?∆s > ρ.
τη
|Gij(tη,σ(s))|∆s > η.
uj(t) :=
1,
0,
−1,
Gij(tη,σ(t)) > 0,
Gij(tη,σ(t)) = 0,
Gij(tη,σ(t)) < 0.
Page 27
EJDE-2009/37CONTROLLABILITY, OBSERVABILITY, REALIZABILITY 27
This input signal satisfies ?u(t)? ≤ 1 for all t ≥ t0, but because of the integral con-
dition above, the i-th component of the corresponding zero-state response satisfies
yi(tη) =
?tη
t0
Gij(tη,σ(s))uj(s)∆s =
?tη
t0
|Gij(tη,σ(s))|∆s > η.
Since ?y(tη)? ≥ |yi(tη)|, we arrive at a contradiction that completes the proof.
We now wish to give conditions under which the notions of exponential stability
and BIBO stability are equivalent. To this end, we begin with the following.
?
Theorem 5.7. Suppose the regressive linear system
x∆(t) = A(t)x(t) + B(t)u(t),x(t0) = x0,
y(t) = C(t)x(t),
is uniformly exponentially stable, and there exist constants β and γ such that
?B(t)? ≤ β and ?C(t)? ≤ α for all t. Then the state equation is also uniformly
bounded-input, bounded-output stable.
Proof. Using the bound implied by uniform exponential stability, we have
?t
τ
?G(t,σ(s))?∆s ≤
?t
τ
?C(t)??ΦA(t,σ(s))??B(s)?∆s
?t
?t
≤αβγ
λ
τ
=αβγ
λ
≤αβγ
λ
≤ αβ
τ
?ΦA(t,σ(s))?∆s
≤ αβ
τ
γe−λ(t,σ(s))∆s
?t
(1 − e−λ(t,τ))
λ
1 − µ(s)λeλ/(1−µλ)(s,t)∆s
.
By Theorem 5.6, the state equation is also bounded-input, bounded-output stable.
?
The following example illustrates the use of Theorem 5.7.
Example 5.8. Let T be a time scale with 0 ≤ µ <1
?−2
2. Consider the system
x∆(t) =
1
−1−sin(t) − 2
y(t) =?1
?
x(t) +
?cos(t)
sin(t)
?
u(t),x(t0) = x0,
e−1(t,0)?x(t),
where here, sin(t) and cos(t) are the usual trigonometric functions and not their
time scale counterparts. DaCunha shows that the system is uniformly exponentially
stable by applying [10, Theorem 4.2] (also found in [11, Theorem 3.2]) with the
choice Q(t) = I. For t ≥ 0, we have ?B(t)? = 1 and ?C(t)? =
√2, since p = −1 ∈ R+from our assumption on T. By Theorem 5.7, the state
equation is also uniformly bounded-input, bounded-output stable.
?1 + (e−1(t,0))2≤
Page 28
28J. M. DAVIS, I. A. GRAVAGNE, B. J. JACKSON, R. J. MARKS II EJDE-2009/37
For the converse of the previous theorem, it is known on T = R and T = Z that
stronger hypotheses than simply having the system be BIBO stable are necessary to
establish exponential stability (see [1, 2, 22, 21]). At present, we lack an analogue
of this result for an arbitrary time scale in the time varying system case. We will
see that the time invariant case does allow for the equivalence of the two notions
in the general time scale case under certain conditions.
5.3. BIBO Stability in the Time Invariant Case. In order to extend the
definition of BIBO stability to the time invariant case, we first need the following
definitions.
Definition 5.9. [13, 19] Let u(t) ∈ Cprd-e2(T,R). The shift of u(t) by σ(s), denoted
by u(t,σ(s)), is given by
u(t,ξ) = L−1{U(z)eσ
⊖z(s,0)},
where U(z) := L{u(t)}(z), and L, L−1denote the generalized Laplace transform
and its inverse.
Definition 5.10. [13, 19] For f,g ∈ Cprd-e2(T,R), the convolution f ∗g is given by
?t
Definition 5.11. For any shift u(t,σ(s)) of the transformable function u(t), the
time invariant system
(f ∗ g)(t) =
0
f(τ)g(t,σ(τ))∆τ.
x∆(t) = Ax(t) + Bu(t),x(t0) = x0,
y(t) = Cx(t),
is said to be uniformly bounded-input, bounded-output stable if there exists a finite
constant η such that the corresponding zero-state response satisfies
sup
t≥0?y(t)? ≤ η sup
t≥0sup
s≥0?u(t,σ(s))?.
Note that Definitions 5.5 and 5.11 are different: one deals with the time varying
case and the other with the time invariant case. The modified definition in the time
invariant case says that the output stays bounded over all shifts of the input.
Theorem 5.12. The time invariant regressive linear system
x∆(t) = Ax(t) + Bu(t),x(t0) = x0,
y(t) = Cx(t),
is bounded-input, bounded-output stable if and only if there exists a finite β > 0
such that
?∞
Proof. Suppose the claimed β > 0 exists. For any time t, we have
0
?G(t)?∆t ≤ β.
y(t) =
?t
0
CeA(t,σ(s))Bu(s)∆s =
?t
0
CeA(s,0)Bu(t,σ(s))∆s,
Page 29
EJDE-2009/37CONTROLLABILITY, OBSERVABILITY, REALIZABILITY29
since y(t) is a convolution. Hence,
?y(t)? ≤ ?C?
?t
?∞
0
?eA(s,0)??B? sup
0≤s≤t?u(t,σ(s))?∆s
≤ ?C?
0
?eA(s,0)?∆s?B? sup
s≥0?u(t,σ(s))?,
which implies
sup
t≥0?y(t)? ≤ ?C?
?∞
0
?eA(s,0)?∆s?B? sup
t≥0sup
s≥0?u(t,σ(s))?.
If we choose η = ?C?β ?B?, the claim follows.
Conversely, suppose that the system is bounded-input bounded-output stable,
but for the sake of a contradiction that the integral is unbounded. Then,
sup
t≥0?y(t)? ≤ η sup
t≥0sup
s≥0?u(t,σ(s))?,
and
?∞
0
?G(t)?∆t > β for all β > 0.
In particular, there exist indices i,j such that
?∞
0
|Gij(t)|∆t > β.
Choose u(t,σ(s)) in the following manner: set uk(t,σ(s)) = 0 for all k ?= j, and
define uj(t,σ(s)) by
uj(t,σ(s)) :=
1,
0,
−1,
if Gij(s) > 0,
if Gij(s) = 0,
if Gij(s) < 0.
Choose β > η > 0. Note supt≥0sups≥0?u(t,σ(s)? ≤ 1, so supt≥0?y(t)? ≤ η.
However,
sup
t≥0?y(t)? = sup
t≥0
???
???
?t
?t
0
G(s)u(t,σ(s))∆s
???
= sup
t≥0
0
Gj(s) · uj(s)∆s
???
≥ sup
t≥0
?∞
> β > η,
?t
0
|Gij(s)|∆s
=
0
|Gij(s)|∆s
which is a contradiction. Thus, the claim follows.
?
The next theorem demonstrates the equivalence of exponential and BIBO sta-
bility in the time invariant case. Recall that this is a notion we currently lack in
the time varying case.
Page 30
30 J. M. DAVIS, I. A. GRAVAGNE, B. J. JACKSON, R. J. MARKS II EJDE-2009/37
Theorem 5.13 (Equivalence of BIBO and Exponential Stability). Suppose the
time invariant regressive linear system
x∆(t) = Ax(t) + Bu(t),x(t0) = x0,
y(t) = Cx(t),
is controllable and observable. Then the system is uniformly bounded-input, bounded
output stable if and only if it is exponentially stable.
Proof. If the system is exponentially stable, then
?∞
by Theorem 5.2.
Conversely, suppose the system is uniformly bounded-input, bounded output
stable. Then
?∞
which implies
lim
t→∞CeA(t,0)B = 0.
Using the representation of the matrix exponential from (5.1), we may write
0
?CeA(t,0)B?∆t ≤ ?C??B?
?∞
0
?eA(t,0)?∆t ≤ η,
0
?CeA(t,0)B?∆t < ∞,
(5.2)
CeA(t,0)B =
m
?
k=1
ψk
?
j=1
Nkjfj−1(µ,λk)
(j − 1)!
eλk(t,0), (5.3)
where the λkare the distinct eigenvalues of A, the Nkjare constant matrices, and
the fj(µ,λk) are the terms from the residue calculations. In this form,
d
∆tCeA(t,0)B
m
?
If this function does not tend to zero as t → ∞, then using (5.3) and (5.2), we
arrive at a contradiction. Thus,
?d
where the last equality holds by noting that if A is constant, then A and eA(t,0)
commute. Similarly, it can be shown that any order time derivative of the expo-
nential tends to zero as t → ∞. Thus,
lim
=
k=1
?
Nk1λk+
ψk
?
j=2
?f∆
j−1(µ,λk)(1 + µ(t)λk)
(j − 2)!
+λkfj−1(µ,λk)
(j − 1)!
??
eλk(t,0).
lim
t→∞
∆tCeA(t,0)B
?
= lim
t→∞CAeA(t,0)B = lim
t→∞CeA(t,0)AB = 0,
t→∞CAieA(t,0)AjB = 0, i,j = 0,1,...
It follows that
lim
t→∞
C
CA
...
CAn−1
eA(t,0)?B AB ···An−1B?= 0. (5.4)
But, the system is controllable and observable, and so we can form invertible ma-
trices Ga
Oby choosing n independent columns of the controllability matrix
and n independent rows of the observability matrix, respectively. Then, by (5.4),
Cand Ga
Page 31
EJDE-2009/37CONTROLLABILITY, OBSERVABILITY, REALIZABILITY 31
limt→∞Ga
follows from the arguments given in Theorem 5.4.
OeA(t,0)Ga
C= 0. Hence, limt→∞eA(t,0) = 0 and exponential stability
?
We use the preceding theorem in the following example.
Example 5.14. Suppose T is a time scale with 0 ≤ µ ≤ 4. The system
x∆(t) =
4530
−1
y(t) =?3
is controllable by Example 2.8 and observable by Example 3.6. The eigenvalues of
A are λ1= −1
the stability region of T. Thus, by Theorem 5.4, the system is exponentially stable.
Theorem 5.13 then says that the system is also BIBO stable.
?−8
1
45
−1
10
?
x(t) +
?2
1
?
u(t),x(0) = x0,
4?x(t).
9and λ2= −1
6. Note that the assumption on T implies λ1,λ2∈ S(C),
As we have seen, the Laplace transform can be a useful tool for analyzing stability
in the time invariant case. With this in mind, we desire a theorem that determines if
a system is BIBO stable by examining its transfer function. The following theorem
accomplishes this.
Theorem 5.15 (Transfer Function Characterization of BIBO Stability). The time
invariant regressive linear system
x∆(t) = Ax(t) + Bu(t),x(t0) = x0,
y(t) = Cx(t),
is bounded-input, bounded-output stable if and only if all poles of the transfer func-
tion G(z) = C(zI − A)−1B are contained in S(C).
Proof. If each entry of G(z) has poles that lie in S(C), then the partial fraction
decomposition of G(z) discussed earlier shows that each entry of G(t) is a sum of
“polynomial-times-exponential” terms. Since the exponentials will all have sub-
scripts in the stability region,
?∞
and so the system is bounded-input, bounded-output stable.
Conversely, if (5.5) holds, then the exponential terms in any entry of G(t) must
have subscripts in the stability region by using a standard contradiction argument.
Thus, every entry of G(z) must have poles that lie in the stability region.
0
?G(t)?∆t < ∞, (5.5)
?
References
[1] B. D. O. Anderson and J. B. Moore. New results in linear system theory. SIAM Journal
Control Optim. 7 (1969), 398–414.
[2] B. D. O. Anderson and L. M. Silverman. Controllability, observability, and stability of linear
systems. SIAM Journal Control Optim. 6 (1968), 121–130.
[3] P. J. Antsaklis and A. N. Michel. Linear Systems. Birkh¨ auser, Boston, 2005.
[4] Z. Bartosiewicz,¨U. Kotta, and E. Pawluszewicz. Equivalence of linear control systems on
time scales. Proc. Est. Acad. Sci. Eng. 55 (2006), 43–52.
[5] Z. Bartosiewicz, E. Piotrowska, M. Wyrwas, Stability, stabilization and observers of linear
control systems on time scales, Proc. IEEE Conf. on Decision and Control, New Orleans,
LA, December 2007, 2803–2808.
[6] M. Bohner and A. Peterson. Dynamic Equations on Time Scales: An Introduction with
Applications. Birk¨ auser, Boston, 2001.
Page 32
32 J. M. DAVIS, I. A. GRAVAGNE, B. J. JACKSON, R. J. MARKS IIEJDE-2009/37
[7] M. Bohner and A. Peterson. Advances in Dynamic Equations on Time Scales. Birkh¨ auser,
Boston, 2003.
[8] F. M. Callier and C. A. Desoer. Linear System Theory. Springer-Verlag, New York, 1991.
[9] P. Cheung and V. Kac. Quantum Calculus. Springer-Verlag, New York, 2002.
[10] J. J. DaCunha. Lyapunov Stability and Floquet Theory for Nonautonomous Linear Dynamic
Systems on Time Scales. Ph.D. thesis, Baylor University, 2004.
[11] J. J. DaCunha. Stability for time varying linear dynamic systems on time scales. J. Comput.
Appl. Math. 176 (2005), 381–410.
[12] J. J. DaCunha. Transition matrix and generalized exponential via the Peano-Baker series. J.
Difference Equ. Appl. 11 (2005), 1245–1264.
[13] J. M. Davis, I. A. Gravagne, B. J. Jackson, R. J. Marks II, and A.A. Ramos. The Laplace
transform on time scales revisited. J. Math. Anal. Appl. 332 (2007), 1291–1306.
[14] J. C. Engwerda. Control aspects of linear discrete time-varying systems. Internat. J. Control
48 (1988), 1631–1658.
[15] L. Fausett and K. Murty. Controllability, observability, and realizability criteria on time scale
dynamical systems. Nonlinear Stud. 11 (2004), 627–638.
[16] I. A. Gravagne, J. M. Davis, J. J. DaCunha, and R. J. Marks II. Bandwidth reduction for
controller area networks using adaptive sampling. Proc. Int. Conf. Robotics and Automation
(ICRA), New Orleans, LA, April 2004, 5250-5255.
[17] I. A. Gravagne, J. M. Davis, and R. J. Marks II. How deterministic must a real-time controller
be? Proceedings of 2005 IEEE/RSJ International Conference on Intelligent Robots and
Systems, (IROS 2005), Alberta, Canada. Aug. 2–6, 2005, 3856–3861.
[18] Y. C. Ho, R. E. Kalman, and K. S. Narendra. Controllability of linear dynamical systems.
Contributions to Differential Equations 1 (1963), 189–213.
[19] B. J. Jackson. A General Linear Systems Theory on Time Scales: Transforms, Stability, and
Control. Ph.D. thesis, Baylor University, 2007.
[20] R. E. Kalman. Contributions to the theory of optimal control. Bol. Soc. Mat. Mexicana 5
(1960), 102–119.
[21] P. P. Khargonekar and R. Ravi. Exponential and input-output stability are equivalent for
linear time-varying systems. Sadhana 18 (1993), 31–37.
[22] A. N. Michel, L. Hou, and D. Liu. Stability of Dynamical Systems. Birkh¨ auser, Boston, 2008.
[23] C. P¨ otzsche, S. Siegmund, and F. Wirth. A spectral characterization of exponential stability
for linear time-invariant systems on time scales. Discrete Contin. Dyn. Syst. 9 (2003), 1223–
1241.
[24] W. Rugh. Linear System Theory, 2nd edition. Prentice Hall, New Jersey, 1996.
[25] L. Weiss. Controllability, realization, and stability of discrete-time systems. SIAM Journal
Control Optim. 10 (1972), 230–251.
John M. Davis
Department of Mathematics, Baylor University, Waco, TX 76798, USA
E-mail address: John M Davis@baylor.edu
Ian A. Gravagne
Department of Electrical and Computer Engineering, Baylor University, Waco, TX
76798, USA
E-mail address: Ian Gravagne@baylor.edu
Billy J. Jackson
Department of Mathematics and Computer Science, Valdosta State University, Val-
dosta, GA 31698, USA
E-mail address: bjackson@valdosta.edu
Robert J. Marks II
Department of Electrical and Computer Engineering, Baylor University, Waco, TX
76798, USA
E-mail address: Robert Marks@baylor.edu
Download full-text