REND. SEM. MAX.

UNIV. POLITEC. TORINO

FASC. SPEC. CONTROL THEORY

Edward W. Kamen

THE POLES AND ZEROS OF A LINEAR TIME-VARYING SYSTEM*

Summary. For linear time-varying discrete-time and continuous-time systems, a notion of

poles and zeros is developed in terms of factorizations of operator polynomials with
time-varying coefficients. In the discrete-time case, it is shown that factorizations can
he computed hy solving a nonlinear recursion with time-varying coefficients. In the
continuous-time case, one must solve a nonlinear differential equaton with timevarying coefficients. The theory is applied to the study of the zero-input response
and asymptotic stability. It is shown that if a time-varying analogue of the Vandermonde matrix is invertihle, the zero-input response can he decomposed into a sum of
modes associate d with the poles. Stability is then studied in terms of the components
of the modal decomposition.

1. Introduction
In this paper we develop a notion of poles and zeros for the class of linear
time-varying systems. Our notion is defined in terms of the factorization of
polynomials with time-varying coefficients. Although the idea of factoring
polynomials with time-varying coefficients is not new (e.g., see Ore [1933],
* A prelimrnary version of this paper was presented at the Twentieth Annual Conference
on Information Sciences and Systems, Princeton, NJ, March 1986. This work was supported in part by the National Science Foundation under Grant No. ECS-8400832 and by
the U.S. Army Research Office under Contract No. DAAG29-84-K-0081.

126
Amitsur [1954], Newcomb [1970]), we present a new approach to constructing factoriztions given in terms of the solutions to a nonlinear equation.
We begin in the next section with the study of a linear time-varying
discrete-time system specified by a second-order input/output difference
equation. Various properties of factoriztions are studied in Section 3, and
then in Section 4, we consider a general class of linear time-varying dicretetime systems specified by an nth-order input/output difference equation. In
Section 5 the theory is applied to the study of the zero-input response and
asymptotic stability. The continuous-time case is considered in the last section
of the paper.

2. Poles of a Second-Order System

With Z equal to the set of integers and R equal to the set of real
numbers, let A denote the set of ali functions defined on Z with values in
R . With C equal to the set of complex numbers, let AQ denote the
set of ali complex-valued functions defined on Z . Clearly, the set A can
be viewed as a subset of Ac . Both A and Ac are commutative rings
with pointwise addition and multiplication defined by
(a + b)(k) = a(k) + b(k)
(ab)(k) = a(k)b(k)

where a(k),b(k)
are arbitrary elements of A (or i4 c ) and k is the
discrete-time index.
Now consider the linear discrete-time system given by the second-order
input/output difference equation
(1)

y (k + 2) + t (k)y (k + 1) +a0 (k)y (k) = b(k) u (k)

where y (k) is the output at time k , u (k) is the input at time k , and
o (k),ax (k) are elements of A . We shall write the left side of (1) in
operator form as follows.
For any positive integer i, let z1 denote the z'-step left shift operator
on A defined by

z'7(*)=/(*+/) , feA .
s

127
For any a(k)eA,'.let

a (k)z* denote the operator on A definedby

fe <*) *V>-(*)/(*+0
Then we can write (1) in the operator form
(z2 + ax (*)*- + a0 ih))y (*) = b(k)u (k).

(2)

Suppose that there exist functioris px (k) , p2 (k) belonging to Ac

such that
(3)

(z2+ax

(k)z+a0(k))y(k)

= (z-px

(k))[(z-p2

(k))y(k)].

It follows from (3) that the given system can be viewed as a cascade connection of two first-order subsystems. To see this, let
(4)

v(k) = (z-p2(k))y(k)

so that
(5)

y(k + l)-p2(k)y(k)

= v(k) .

Inserting the expression (4) for v (k) into (3) and using (2), we have that
(z-Pl(k))v(k)

= b(k)u(k)

or
(6)

v (k + 1) -px (k)v {k) = b(k)u (k) .

From (5) and (6) , we have the cascade realization of the system shown in
Figure 1 below.
Given the decomposition of the system into the first-order subsystems
shown in Figure 1, it is tempting to cali px (k) and p2 (k) the poles of
the system. We shall give a formai definition later. First, we want to characterize the cascade decomposition in terms of a polynomial factorization, and
then we consider the existence and construction of factorizations.
Again suppose that there exist px {k),p2 (k) e AQ such that (3) is
satisfied. We want to define a product

128
so that
(7)

[(2 - P ! (*)) o (2 - p 2 (k))]y (k) = (2 -/>, (*)) [( ~P 2 (ft))^ <*)1

Figure 1. Realization of system

Expanding the right side of (7), we obtain
fe-pi

(*))l(*~p2 (k))yWlr(*-pi
,

=y(k +

(k))ly(k + l)-p2

(k)y(k)]

2)-(p1(k)^p2(k^l)yy{k+l)-hpi(k)p2(k)y(k)

=[2 a -(pi(*)+pa(* + l))2+^i(*)pi(*)pi(*)]y (*).''

Inserting this into the tight side of (7), we have that
(8)

( 2 - p j ( * ) ) ( * - p a (^)) = 2 2 - ( p 1 (*)+p 2 (^ + l ) ) 2 + p 1 (k)p2 (*>.

Equation (8) shows that the multiplication is the usuai polynomial multiplication except that
2 0p 2 (fc)=p 2 (+ 1)2

(9)

As a result of the property (9), the multiplication is called a skew polynomial multiplication. In Section 4 we define a ring of polynomials with
the property (9).
We now consider the existence of a skew polynomial factorization of
the form (8). Combining (3), (7), and (8), we have that
(10)

z2-'(pl(k)

+ p2 (fc + D J z + p ! (k)p2 (*) = z2 +*!

(k)z+a0(k).

129
Equating coefficients of 2 in (10), we see that there is a factorization of
the form (8) if and only if
(11)

Pi ( * ) + p 2 ( * + D = --*i (*)

(12)

Pi(k)p2(k)

= a0(k) .

Multiplying both sides of ( l l ) b y p2 (k) and using (12), we obtain

(13)

p 2 ( * + Dp2 (*)+i (k)p2 (*) + * 0 (*) = 0 .

Note that (13) is a nonlinear first-order difference equation with timevarying coefficients. It is also interesting to note that the left side of (13)
looks like the polynomial 22 + a1 (k)z + a0 (k) evaluated at 2=p 2 (k)
with
Z

k-p,(ik) = ? ( * )

and

z2

\z=p2(k)=P2^^^p2(k).

Given the initial value p2 (k0) at initial time k0 , we can compute

p2 (k) for k>k0 by solving (11) and (12) recursively, or by solving
(13) recursively. If p2{k)^0
for ali k>k0 , the solution is unique.
Given p2 (k0) selected at random, the probability is zero that p 2 (k)
will be zero for some value of k > k0 . In other words, for almost ali initial
values p 2 (k0)e C, (13) has a unique solution p 2 (k) with p 2 (^)=^0
for ali k > k0 , Futhermore, px (k) can be computed from (11). We therefore have the following result.
Proposition 1: For amost ali yeC, the operator polynomial 22 4al(k)z+a0(k)
has a unique skew polynomial factorization ( 2 - p t '(k)) o
(2 - p 2 (k)) for k >k0 with p2 (k0)-y '.
^-^it should be stressed that we are asserting the (generic) existence of a unique factorization over the time interval k >k0 with the given initial value
Pi (^o) If w e allow the initial value p 2 (fc0) to range over ali of C, in general we will obtain an infinite collection of skew polynomial factorizations.
If the given system is time invariant for k<k0 (i.e., a0 (k) and
ay (k) are Constant for k <& 0 ) > w e c a n take the initial value p2(k0) to
be equal to one of the initial poles. It the system is not time invariant prior to

130
k=k0 , we could take p2 (k0) to be one of the zeros of the polynomial
z2 +ax (k0)z +a0 (k0).
If px (k), p2 (k) are the solutions to (11) and (12) with p2(k0) =
7 e C , then the complex conjugates px (k), p2 (k) are the solutions to
(11) and (12) with initial value 7 . This can be seen by simply taking the
complex conjugate of both sidesof (11) and (12). Unlike the time-invariant
case, in general ~px (k)i=p2 (k) and /?2 (h)px (k). Thus, in general
there are two different solutions to (11) and(12) with initial value p2 (k0) =
7 or p2 (^o) = 7 We cali the ordered sets (pi (k), p2, (&)) and (pt (k),
Pi ()) Pl e sets on k >k0 with respect to the initial values 7 , 7 - The
elements p2 (k) and p2 (k) are called the right poles on k>k0 .
In the following two examples, the pole sets were computed by solving
(11) and (12) recursively. In particular, given p2 (k0) we used (12) to
compute px {kQ), then (11) to compute p2 (k0 + 1) , then (12) to
compute pi (k0 + 1), and so on.

Example 1: The coefficients of the input/output difference equation (1) are

tf0 () = 0.5 forali keZ,a1(k)=

< - 1 + (l5fc/200), 0 <k < 2 0 0

V 1,5 , k > 200

For

k<0,

z2 + 0 ! (k)z+a0

(k) = z2 -z + 0 . 5 = ( 2 - 0 . 5 - j 0 . 5 ) ( z - 0 . 5 +J0.5),

and as k -+ o
2? -f ^ " ( ^ . Z + t f o (H)-*Z2 + 1 . 5 2 + 0 . 5 - ( 2 + 1 ) ( 2 + 0 . 5 ) .
The poles with p2 (0) = 0.5-jO.5
poles are not complex conjugates.

are shown in Figure 2. Note that the

Example 2: Now suppose that

* 0 () = 0.5 forali K Z , a , ( * ) =

(-1,*<0
j ' - 1 + (2.5fe/50) , 0 <k < 50
l 1.5 , k>50

131
As in Example 1, the initial poles are 0.5 '+ j0.5., 0.5 - jO.5 , and as k - <
z2 +ax (k)z+a0

(k)-+z2 + 1.5z + 0.5 = ( z + l ) ( z + 0.5).

1 Imag

0.5+J0.5

Re

0.5-J0.5

Figure 2. Poles with p2 (0) = 0.5 - j0.5

The main difference between this example and the previous one is that the
rate of change of a1 (k) is much faster in this example. The poles with
p2 (0) = 0.5 -jO.5 are shown in Figure 3. Note the erratic behavior of the
poles as k is increased from the initial value k = 0 . This is a result of the
rapid variation of the coefficient ax (k).
^ Imag
/ :.

Pi(k)

-1 '-.

4-

-0.5

0.5

pi(k)--^,-:/.

.
* ' ,

0.5+]0.5
> - Re
0.5-0.5

Figure 3. Poles with p2 ( 0 ) = 0 . 5 - J0.5

132
3. Properties of Factorizations
Again consider the discrete-time system given by the second-order input/
output difference equation (1). Let px (k) and p2 (k) denote the zeros
of the polynomial z2 + hx (k) z + a0 (k) ; that is,
(&)) = %2 +ax (k)z+a0

(z-px(k))(z~p2

(k) ,

where (z -px (&)) (z -p2 (k)) is the ordinary product of two polynomials.
We can write this product in the form

{z-p, (k))(z~p2 (*))=**-<, (+p a mz+fc (k)p2 (k)

=z2-(px(t)+p2(k+l))z+px(k)p2(k)nMk+l)

If the system is slowly varying in the sense that \p2 (k -f l)p 2 (k)\ is
small for k > kQ , from (14) we have that
(15)

(z -p\

(*)) (z -p\

(*)) - (z - p \ (*)) o (2 - p 2 (*)).

Here = means that the magnitude of the difference of the coefficients of

the polynomials is small for k>k0 . The relationship (15) corresponds to
the well-known result that slowly varying systems can be studied using the
"frozen time apprach".
an
Now suppose that a0(k)-*c0
d #i (&) ~*cx as k-+. Let rx
2
and r2 denote the zeros of z + cx z + c0 . If (px (k), p2 (k)) is a
pole set on k > k0 , it turns out that p 2 (k) will not converge to rx or
r2 in general. In particular, if rx and r2 are complex numbers and
P2 (&o) is a r e a l number, by (13) p 2 (&) is real for ali k>k0 , and
thus p 2 (k) cannot converge to fx or r2 .
As a special case, suppose that
a0 (k)~c0

and #! (k):=cx

forali k>kx

Given the pole set (pj (fc), p 2 (&)) , let A (fc) = p 2 (*) - r 2 ' so that
(16)
s

A(* + l ) = p 2 (* + l ) - r 2

133
By(13),
#n (k)
p2h)
and thus for k > k ,
(17)

(*+ 1 )=

_, --S_ .

Inserting (17) into (16) gives

A f t + 1)= Z(La)L)Zo

Pi (*)'

Now /?2 (&) = A (&) + r2 , and therefore

, x
(18)

.
~ (r\ + Ci r2 + 'o) "" fai + r 2 ) A (&)
2;
A (* + 1.) = 1 - 1 - 2 - L i i - L
.
A

A (k) + r 2
But since rx and r2 a r e t n e zeros of z 2 + Cj z + c 0 , we have that
(19)
r2 4- cx r2 + cQ = 0 and - - - ( ^ 4- r2)-rx
.
Using (19) in (18), we obtain

(20)

r, A (k)
A (* + 1) = T /
; *.->*i. .
A (&) + r 2

We can convert (20) into a linear difference equation by using the transformation
(21)

&)=-*^-,.

134
Combining (20) and (21) gives
(22)

( * + i) = I l [(*) + l ] .
ri

Now (22) can be solved using standard techniques, and once we have determined g (k), we can compute A (k) using(21).If r1:r2
and kx = 0,
the reslt is

(23)

(r1-r2)

A(0)

W'Moy+m-rs-uom,*-'*0-

From (23), we see thatif A (0)i=ry -r2 (i.e., p2 ( 0 ) # r j ) and I r2/rl l> 1,
then A (&) converges to zero as &-<, and hence p2 (k) converges to
r2 as k -* e. Clearly, the condition I r 2 /ri I > 1 requires that rj and
r2 be real (not complex conjugates).
The derivation given above can be generalized to include the case when
#o ( ^ ) - ^ 0 a n d <*i {k)-*cx
as k-*<>. The result is stated below.

Proposition 2. Suppose that a0 (k)-+c0 and ax (k)-*cx

as k -*> and
2
let f^, r 2 be the zeros of z + Cj z 4- c0 . If l^/rj l > 1 , then for any
pole set (pt (k), p2 (k)) with p2 (k0)=r1 , p2 (k) converges to r2 as
k-+<*>.
The hypothesis of Proposition 2 is satisfied in Examples 1 and 2 given in
the previous section, and thus in these cases p2 (k) does converge to
r2 ( = - 1 ) . In the following example, we modify Example 2 so that rt and
r2 are complex.

Example 3 : Suppose that

a0(k) = 0.5 forali

k eZ

(-1
,k<0
, ax (&)= J - 1 + (2.4k/50), 0 <k < 50

As in Example 2, the initial poles are 0.5 + J0.5 , 0.5 - j0.5 , but now .
z 2 +at (k)z +a0 (k)-+z2 4- 1.4z 4-0.5 as &-*oo.

135
In this example, rx = - 0 . 7 - j 0 . 1 and r2 =7i . The poles with p2 (0) =
0.5 -jO.5 are plotted in Figure 4. In this case, the poles pi(k) and p2 {k)
continuously encircle points ih the complex piane as k -* .
Imag

Pi(k)

aa

'. 0.5+J0.5

r'

1
P 2 (k)

-0.5

Re

0.5
0.5 -i0.5

V .
''

Figure 4. Poles with p2 (0) = 0.5 - j0.5 .

To avoid the type of behavior displayed in Example 3, we can reinitialize
the recursion for computing the poles. For example, again letting px {k)
and p2 (k) denote the zeros of z2 +at {k)z + a 0 {k) , if \p2 {k + 1)
p\ {k)\ is small for some range of k<kx
, we could set p2 {k1) = p2 {kx)
and continue with the recursion.
We conclude this section by considering the relationship between two different pole sets. Let {px (k), p2 (k)) and {qt (k),q2(k))
be two pole sets
for the system given by (1) . Then by (11) and (12), we have that
(24)

Pi (k)+p2(k

+ l) = q1 {k) + q2{k + )

(25)

Pi (k)p2 (k) = qx {k)q2 (*)..

Multiplying both sides of (24) by q2 {k) and using (25), we obtain

(26)

p1{k)[q2

{k)-p2

{k)] = q2 {k)[q2 (k + l)-p2

Defining
w{k) = q2 (k)-p2

(k),

(* + l)] .

36
we can write (26) in the form
(27)

Pi(k)w{k)

= q2(k)w(k

+ 1).

If w (k) = 0 for ali k > k0 , from (27) we hve that

(28)

pi (k) =

w (k 4- 1)
w (&)

? 2 (*) , k > k0 .

The relationship (28) shows that it is possible to compute the element px (k)
of the pole set (pi (k), p 2 (k)) from the right ple q2 (&) of the pole set
(<7i (*) #2 (*)) Note that if qx (k) = px (k) and q2 (k) = f2 (k), then
by (28)

, ^ - fi(* + -pa(* + l) - ,.v

Pi (*)"P2 (*)
assuming that p 2 (h)^p2 (k) . This result shows that in general pi (k) is
not the complex conjugate of p2 (k).

4. The nth-Order Case

Let A [z] (resp., AQ [2]) denote the set of ali operator polynomiate
in the left shift operator 2 with coefficients in A (resp., AQ). With the
usuai polynomial addition and with multiplication defined in terms of

2*0 a (k) = a (k +i)z* , a(k)eA

(or

Ac),

A [z] and Ac [z] are skew (noncommutative) rings. These rings have appeared in the work of Hafez [1975], Kamen and Hafez [1975], and Kamen,
Khargonekar, and Poolla [1985].
Now given
a(z, k) = zn + aAk)zjeA
*=o '

[2]

137
and
b(z,k)=

S b.(k)zleA

[z],

consider the linear discrete-time system defined by the input/output equation

a(z,k)y(k)

= b(z,k)u(k)

where as before, y (k) is the output and u (k) is the input.

We cali pn (k) e AQ a right pole of the system if there exists a e (k, z) e
Ac [:z.] such that
(29)
We cali q (k)eAQ
Ac [z] such that

a(z,k) =

e(z,k)o(z-pu(k))'.

a right zero of the system if there exists a h (z, k) e

b(z, k) = b (z,

k)o(z-q(k)).

Suppose that there exist pn (k) and e (k, z) such that (29) holds. Then
writing
e(k,z) = zn~1+n22

e.{k)zi

we have that (29) is equivalent to the equations

(30)

en_2 ik)=pn tt.+ i f - D + tf^ (*)

(31)

e._1 (k) = e. (k)pn (k H ) + a. (k) , i = - 2 , - 3 , . . . , 1

(32).

0 = ^0(*)P.(*)+o(*).

Given the initial values pn (k0 - n + 1), pw (^ 0 - + 2), ..., pB (^0) we can
solve (30)-(32) recursively to compute pn (k) and the e{(k) for
k>k0.
As in the case n = 2 , a unique solution exists for almost ali possible values of
the initial data.
Once we have determined the factorization (29), we can then pul out a
right factor from e (k, z), and so on, until we obtain the pole set (pij (fe),

138
p2 (&)> >/> (^)) I* i s important to note that this is an ordered set. Due to
the noncommutativity of multiplication in the ring AQ [2], a permutation of
the elements in a pole set would not result in anotherpole set (nless of course
the pf(k) are Constant).
As we now show, Equations (30)-(32) can be combined to yield a single
equation for pn (k). First, multiplying both sides of (30) by p (k + n - 2)
and using (31) with i = n - 2 , we obtain
(33)

V 3 ( ^ , ( ^ + ^ ^ + ^ 2 ) ( * ) p

+-2)+^) .

Multiplying both sides of (33) by pn(k+n~3)

i = n 3 gives

and using (31) with

*_4 * >?n <* +*2""1 )P <* + % ~ 2 >P * +n-3)+an_ t(k )pn (k +n-2 )pn (k +n-l ) +
* . - 2 - < * > P . < * + - 3 > + . - 3 (*>
Continuing, we obtain
0^n{k+n-\)pn(k+n-2)...pn(k)+n

ai(k)pn(k+i-l)pn(k^i-2)...pn(k)+a0(k)

(34)
If pn (k)= 0 for ^ > f e 0 - + l , we can rewrite (34) in the form (assuming n > 3)

(35)
flp (fe)

pn(k+n-2)pn(k+n-)...pn{k)

'

Equation (35) can be solved recursively to compute the right pole pn (k) .
We can compute a collection of right poles by solving (35) for different
initial conditions. The calculation of right poles using (35) can be carried out
in parallel. In the next section we will show that the computation of a set of
right poles arises in the derivation of a modal decomposition of the zero-input
response.
We conclude this section by showing that our notion of a zero can be

139
interpreted to be a transmissin blocking zero as .in the time-invariant case.
Let q (k) be a right zero of the system defined above so that
b(z,k)

h{z,k)o(z-q(k)).

Define

(k , k0)=h

(k-l)q

(k-2) ...q(k0),

k>k0

/o

,k<k0

The function (k, kQ) is called the mode associated with the right zero
q (k)'. Note that if q (k) is Constant, so that q (k) = q for k *>k0 , then

^ <M0 )-**"*, *>*o'.

Let q (k) be a right zero with associated mode 0 (k, k0) . With the
input u (k) of the system equal io 0 (k, k0), we want to show that we
can choose the initial output values y (kQ ri), y (k0 n + 1), ...,y (k0l)
so that y (k) = 0 for ali k > k0 . First, we have that
b(z,k)u(k)

h(z,k)[(z~q(k))u(k)]

= b{z,k)[4>q(k
By definition of

+ l t k0)-q

(k)(j>q (k , * 0 )]

0 (k, k0) ,

. ^ ( * + l , * o ) = * ( * ) 0 , , Aio) , * > * o ,
and thus
b(z,.k)u(k)

= 0 for

&>& 0

Returning to the input/output difference equation, we then have that

(36)

y(k+n)

=-

" a.(k)y(k+i),

k>k0

If y (&0 +/) = 0 for i = 0, 1, ..., n - 1 , it follows from (36) that y (&)= 0

forali k>k0

as desired . If d0 (&)=() for & = &0 - 1 ,&0 - 2 , ..., &0 - ,

140
it Is a standard aonstruction to show that the initial values y (k0 - 1 ) ,
y (k0 - 2 ) , ...,y (k0 -n) can be chosen so that y (k0 + i) = 0 for / = 0 , 1 ,
. . . , - ! . Thus we have the following result.

Propositon 3: Suppose that a0 (k) = 0 for k = k0 - 1, k0 - 2, ..., k0 - n .

Then with u fk) == 0 (&, fc0) , for some initial values y (k0 - 1), y (k0 - 2),
y (&o "" n ) t n e output y (k) is identically zero for ali k > &0 .

5. The Zero-Input Response and Stability

In this section we assume that the input u (k) to the system is zero for
ali k > k0 , so that the input/output difference equation reduces to
i

(37)

y{k+r)+

X a.(k)y(k+i)

=0

Using our operator notation, (37) can be written in the form

(38)

a (z, k)y (&) = 0

The initial conditions for (37) (or (38)) are the values y (k0), y (k0 + 1),
..., y(k0 + - l ) .
Let pnl (k), pn2 (k),...,pnn
(k) be n right poles; that is, suppose
that there exist polynomials ei (z , k) e Ac [z] such that
a(z,k)

= e.(z;k)o(z-pn.(k))

for i = l, 2,...,..

Let V (k) denote the time-varying version of the n X n Vandermonde

matrix defined by
'1

1""

K<*) = p . 1 l - * l V ^ 1 ( f t )

_ . , < * + - 2 ) .../>,<*) ... />(*+-2) ... * (*)_

141
Finally, Jet <t> {k', k0)

denote the mode associated with the pole p^-W-

ni

Recali that
\ 1, k ~ kQ
<t>P ..(. *o) = </>,(* - DP,- (* - 2 ) .../., (ho),

k>k0

/o,k<k0
We then have the following result.

Theorem 1: If the determinant of V (k0) is nonzero, there exist constants

clt c2,..., cn such that the solution to a (2, k)y (k)~0 can be written in
the forni
(39)

y(k)=ci

(k,k0)

r= 1

for

k>k0.

Pni

The constants ci in (39) depend on the initial values y (k0), y (k0 -f- 1),...,
3/(^0 + - l ) .

Proof: We will first show that y{k)i-c^

We Have that

(k, k0) is a solution to "a (z,k)y(k)=0.

a (2, k) ce '(k, kQ) = cei <*; 2) [(2 - p ^ , ) ) ^ .. <*, *)]

"tu,

*w.

By definition of the mode

(k, k0) ,

and thus
a (z, k) ce

(k,k0)~0

for

k>k0.

ni

Using linearity, we have that the right side of (39) is a solution t

a (2, k) y (k) = 0 . Again using the definition of the modes 0
(k, k0) ,

142
from (39) we also have
"><*o)

c2

y(k0+i)
=

V(k0)

y(k0+n-l)
Since V (k0) is invertible, we can solve the above matrix equation for the
constants cl) c2, ..., cn . We therefore have the desired result.
Equation (39) is the time-varying version of the well-known modal decomposition of the zero-input response in the time-invariant case.
In fact, if the right poles pni(k) are Constant, so that pni(k) = Pj fr
k > k0 , then (39) simplifies to the modal decomposition in the time-invariant
case given by

y(k)=

jtt c.p***,

k>k0.

The repeated root case is left to another paper (see Kamen [1987]).
We conclude this section by applying the modal decomposition (39) to
the study of asymptotic stability. Recali that the given system is asymptotically stable (a.s.) on k>k0
if the zero-input response y ih) converges
to zero as k-+ for any initial values y (k0 ) , y (k0 + 1), ..., y (k0 + n 1).
The system is uniformly asymptotically stable (u.a.s.) on k > k0 if for any
real number e > 0 , there is a positive integer N such that
\y(k+N)\<e\y(k)\

forali

k>k0.

Theorem 2: Let pnl (k),pn2 (k)y...,pnn (k) be n right poles with associated Vandermonde matrix V (k), and suppose that V (k0) is invertible.
Then the system is a.s. on k>k0
if and only if
(40)

l p ^ - l ) p M t ^ - 2 ) . . . p w ^ 0 ) l ^ 0 as *->oofor i = 1, 2, .... .

The system is u.a.s. on k > k0

if and only if for any real number e > 0 ,

143
there is a positive integer N

such that

\pni(k+Ne-l)pni(k+Nr2)...pJk)\<

for ali k > k0

and i = 1, 2, ..., n.

The prof of Theorem 2 follows easily from the modal decomposition

(39). The details are omitted.
By Theorem 2, we see that testing for a.s. or u.a.s. can be reduced to
testing the stability of the first-order systems corresponding to the components
of the modal decomposition (39).
It follows from (40) that a sufficient condition for asymptotic stability on
k>k0
is
(41)

\pni (k)\<c<tk>k0

;.i = l, 2, .,., n .

It is interesting to note that if the p . (k) in (41) are replaced by the ordinary zeros of a (z, k), then the condition is no longer sufficient (or necessay) for stability.
It should be mentioned that condition (41) is not necessary for asymptotic stability in general. For example, if the p . (k) are periodic for k>kQ
with period two, the system is a.s. if and only if
\pni(k0

+ Dpni(kQ)\<l

for

i=l,2,...,.

6. Continuous-Time Case
Ali of the constructions given in the previous sections have an analogue in
the continuous-time case. A sketch of this case is given below.
With D equal to the derivative operator, consider the linear continuoustime system given by the input/output differential equation
(42)

a (D, t) y(t) = b (D, t) u (t) '

where
a(D, t) = Dn + "S 1 la.(t)D*
*=o
b(D,t)=

b.(t)D\

144.
We assume that the coefficients a: (t) and b. (t) can be differentiated a
suitable number of times.
We begin by considering the second-order case (n = 2 ) , so that
a (D, t)~Dz

+at (t)D+a0

(t) .

Suppose that there exist functions pi (t) and p2 (t) (both of which may
be complex valued) such that
(43)

(D2 + * j (t)-D+a0 (t))y (t)~(D~px

(?)) [(D ~p2 (;t))y (t)} .

Expanding the right side of (43), we obtain

(D-Pl

(t))[(D~p2

(t))y(t))~[D2~(pl

(t) + p2 (t))D +

(44)
Pi(t)p2
where p2 (t) is the derivative of

(t)~p2
p2

(t)]y(t)

(t).

Now we want to define a polynomial multiplication o such that

(45). [ ( D -

Pl

(t))o (D-p2

(t))]y(t)

= (p-Pl

(t))[(D-p2

(t))y(t)]

Comparing (44) and (45), we see that

(D~Pl
(46)

(t))o(D-p2

(t))~D2-(P
Pi (t)p2 (t)~p2

+p
(t) .

(t))D +

Note that
D o p2(t)=p2

(t)D-p2

(t) ,

and thus the multiplication o is noncommutative. Also note that the noncommutativity is different from that in the discrete-time case (see (9)).
If there exist px (t) and p2 (t) such that (43) holds, we cali the ordered set (pi (t), p2 (t)) a pole set, and we cali p2 (t) a right pole.
It follows from (43)-(46) that (pi (t), p2 (t)) is a pole set if and only if
(47)
\

pi W + P 2 ( 0 = - ^ 1 (t)

_
(4.8)

px ( 0 p2 (t)~p2

145

\t) = a0 (t) .

Multiplying both sides of (47) by p2 (t) and using (48), we have that
(49)

pi ( 0 4- p2 (t) + a ! (t) p2 (0 4- a 0 ( 0 = 0 .

Note that (49) is a nonlinear first-order differential equation with timevarying coefficients. In fact, the left side of (49) looks like the polynomial
D2 4-0, (t)D+aQ (t) evaluatedat D~p2 (t) with
D|

D=MO ^

{t)

and

" W,(0

(t) +

^2

(t)

'

We conjecture that (49) has a unique solution p2 (t) for almost ali
initial conditions p2 (t0 ) e C . We will show that this is the case when
tfo (0 = c o a n d ^ i ( 0 ~ c i forali t>t0 .
Let r t and r2 denote the zeros of D2 + c t D f c0 and define
A(t) = p 2 ( 0 - r 2 .
Then
A(f)=p2 (0 = ~p2 (*)-Ci P2 (O-^o

Inserting p2 (t) = A (t) + r2

for

*>'o

into this expression gives

( 0 = - A 2 ( 0 - ( r 2 - r t ) A (0 for t > t 0 .

(50)

Defining g (t) - I/A (t) , we can convert (50) into the linear differential
equation
(51)

() = 0 - 2 - ^ ( 0 + 1

Solving (51) and using the relationship A (?) = \lg (t) , we obtain
A

(52)

,,

A (t)=

(r'i-r a )A(0)

A(0) + ( r , - r 3 ) - A ( 0 )

^^
,

exp ( - ( r t - r 2 ) 0

In deriving (52), we have assumed that rx # r2 and t0 = 0 .

t>0.

146
It is interesting that (52) has exactly the same form as in the discrete-time
case (see (23)), except that the exponential terni is replaced by (r2/ri)k
in
the discrete-time case. From (52), we see that there is a unique solution
p2 (t) = A (t) + r2 for any initial condition p2 (0).
Returning to the nth-order case defined by (42) , we cali pn (t) a right
pole of the system if there is a polynomial e (D, t) such that
a(D,t) =

e(P,t)o(p-pn(t))t

and we cali q (t) a right zero of the system if there exists a polynomial
h (D, t) such that
b(D,t) =

b(D,t)o(D-q(t)).

It is assumed that the coefficients of e (D, t) and h (D, t) belong to some

space of functions that can be differentiated an appropriate number of times.
The precise specification of the space of coefficients will not be considered.
If p (t) is a right pole or a right zero, we define the mode associated
with p (t) by
v

0 , t < t0

exp

i>

( X ) dX

t>t0

We then have the following continuous-time counterpart to Proposition 3.

Proposition 4: Let q (t) be a right zero of the system. Then with the input
(l)
u (t) = (j) (t, t0) , there exist initial conditions y,<0
(t0), i = 0, 1, ..., n - 1
such that the output response is identically zero for t > tQ
The proof of this proposition is very similar to the proof of Proposition 3 and
is therefore ornitted.
Now suppose that pnl (t), pn2 (t),...,/? Wfl (t) are n right poles of
the system. Define the operator
Sp .(f(t)) =
* ut

Pni(t)f(t)+f(t),

147
and let V (t) denote the generalized Vandermonde matrix defined by
1

1
p

Pnl <*>

Pnl

spr 2<P2<'

(p At))

<'>

(p (*))
WK

V(t) =

(t)

O-(f)) s p>- ( t ) )

S"~2(/>

(0

We then have the following counterpart to Theorem 1

Theorem 3: Suppose that the determinant of V (t 0 ) is nonzero. Then for

any initial conditions y^ (t0), i 1, 2, ...,'w 1 , there exist constants
clf c2, ..., cn such that the solution to a (D, t) y (t)~0 can be written in
the form
(53)

y(0=

S e (j>

(t, t0)

for t>t0

Again the proof is very similar to the one given in the discrete-time case, so we
shall omit the details.
If the right poles pni (t) are Constant, so that pni (t)p. for t> t0 ,
the modal decomposition (53) simplifies to the well-known form in the timeinvariant case given by
y (t) = 2 c.exp (p. t) , t>t0

It follows directly from (53) that the system is asymptotically stable

(a.s.) on t > t0 if and only if
(54)

li

(t,t0)\-+0

as t->oo for

= 1,2, . . . , .

"ni

A condition which imples (54) is

(55)

Rcpni(t)<c<0

for t>t0

and 2 = 1,2,...,n

148
If the p . (t) in (55) are replaced by the ordinary zeros of a (D, t) , then
the condition is no longer sufficient for stabili ty.

REFERENCES
[1] O. Ore, Theory of noncommutative polynomiah, Ann. Math., Voi. 36, pp. 480-508,
1933.
[2] A.C. Amitsur, Differential polynomiah and division algebras, Ann Math., Voi. 59,
pp. 245-278, 1954.
[3] R.W. Newcomb, Locai time-variable synthesis, in Proc. Fourth Colloquim on Microware Communications, Budapest, 1970.
[4] K.M. Hafez, New results on discrete-time time-varying linear systems, Ph. D. Dissertation, G'eorgia:Institu*te of Technology, Atlanta, 1975.
[5] E. W. Kamen and K. M. Hafez, Algebraic theory of linear time-varying systems, SIAM
J. Control and Optimization, Voi. 17, pp. 500-510, 1975.
[6] E.W. Kamen, P. P. Khargonekar, and K. Poolla, A transfer fimction approach to linear
time-varying discrete-time systems, SIAM J. Control and Optimization, Voi. 23, pp.
550-565, 1985.
[7] E.W. Kamen, On the factorization of difference and differential polynimials, in preparation.

EDWARD W. KAMEN - Department of Electrical Engineering University of Pittsburgh

Pittsburgh, PA 15261