4

Transfer-Function Realizations

It was noted in Chap. 2 that the transfer-function representation provides

an input/output description of linear systems, while the time-domain repre-
sentation through state and output equations contains information about the
input/output behavior as well as the internal structure of linear (and non-
linear) systems. Because of the analytical advantages associated with the
time-domain representations, a large number of synthesis/design tech-
niques are based on state/output equations.* On the other hand, because of
practical/operational problems, t the majority of engineering systems are
described through input/output response/transfer functions. Thus to estab-
lish a bridge between these two complementary aspects of control, a
(ideally one-to-one) correspondence between the two representations is
necessary. In addition to this, it is usually more efficient to simulate com-
plex systems through a set of integro-differential equations rather than
through a transfer-function matrix. t This problem of translating system
representation from G(s) to S(A,B,C,D), known as the realization problem,
has been extensively studied, yielding a number of realization procedures.
Some basic aspects of realization theory and a few selected methods are
considered in this chapter.

4.1 REALIZATION PROBLEM [1-9]

The derivations in Chap. 3 showed that the relationship between state-space

and input/output representation for a proper system is

*A number of synthesis/design techniques based on frequency-domain repre-

sentations are now available; some of these are considered in Chap. 10.
T Although this is changing due to the advent of new sensing and signal-
processing techniques.
tThis situation is also changing due to the availability of a range of high-
level block-diagram-orientated languages on a number of digital computer
systems.

231
232 Transfer-Function Realizations

G(s) = C(sI - A)"7B+ D

Thus for a given set of system matrices A, B, C, and D, the transfer func-
tion matrix G(s) may be uniquely determined. The realization problem is
the inverse problem where the system matrices A, B, C, and D are to be
evaluated for a given G(s). A realization is said to be minimal if the dimen-
sion of the square matrix A [or the state vector x(t)] is minimal. It can be
shown that every rational proper G(s) has minimal realizations, and all
minimal realizations are equivalent. The example below outlines the signif-
icance of minimal realization [1].
Example 4.1: Realization concept —one-input/ one-output system

S(A, bse)t ox(t) =) diag ay,a,5a54

aig} x(t)\+ [b, ,b,, 0, 0] -u(t)
(4.1)
y(t) = [0,c,,0,c,]y(t) + [O]u(t)

The transfer-function matrix of this fourth-order system is

#(s)i =).6(sl=A)s'b
s Siy b;

S-a, b,
= [0,c,,0,c,] S-a,

S-a, 0

2 eee 2 (4.2)

The signal flow graph of the state-space representation Fig. 4.1 shows that
the state-controllable and state-observable (= output controllable) part of
the system may be represented by

X,(t) = a,x, (t) + b, (t)

y(t) = ¢2x,(b) ae
and this defines the passage of control signal from the input terminal to the
output terminal, and has a transfer function c,b,/(s - a,). It can therefore
be concluded that (4.3) is an ''adequate" realization of g(s) in (4.2), and
this is the minimal realization of g(s). The description is (4.1) is a non-
minimal realization of g(s).
Realization Problem 233

Fig. 4.1 Signal flow graph of Example 4.1.

4.1.1 Concepts and Definitions

As the example illustrates, for every realizable transfer function (matrix),

there is an unlimited number of linear time-invariant state and output equa-
tions, and there are direct relationships between controllability and observ-
ability properties of S(A, B,C, D) and minimal realizations. These lead to
the following concepts associated with realization theory [2].

Theorem 4.1

1. The impulse-response matrix of a linear time-invariant system

S(A, B, C, D) depends solely on the controllable and observable
part of the system
t iS)tel NOE: Be nie Cee
he Ea) and is
is given
gi explicitly
ici by

H(t) = C5? a to) Moe cP oe (4.4)

Lt)

where ¢¢,(°) is the state transition matrix corresponding to S,y-

2. Knowledge of the impulse-response matrix H(t, Tt) identifies the
completely controllable and completely observable part of the
system, and this part (Sco) is itself a dynamical system with the
smallest dimension among all realizations of H(t, 7).
3. A realization of S(A,B,C,D) is minimal (= irreducible) if and
only if at all times it consists of S,, alone; thus every irreducible
234 Transfer-Function Realizations

realization of S is completely controllable and completely observ-

able (i.e., completely output controllable).
4. Any two completely controllable and completely observable reali-
zation of S(A, B, C, D) are algebraically equivalent.

Proof of the theorem follows from the canonical decomposition theory

(Theorem 3.5) and Fig. 3.5. A consequence of the theorem is given below.

Definition 4.1 A realization

x(t) Ax(t) + Bu(t)

(4.5)
y(t) Cx(t) + Du(t)

of a rational transfer function matrix G(s) is said to be minimal or irreduc-

ible if not other realization

X(t) = Ax(t) + Bu(t)

y(t) = Cx(t) + Du(t) (4.6)

of G(s) exists with dimension of x(t) less than the dimension of x(t), the
minimal order of realization being nmin = 6 {G(s)}-

With Theorem 4.1 and the derivations in Sec. 3.2, it is apparent that
a minimal realization of G(s) can easily be obtained through the derivation
of the controllable and observable canonical forms of (4.4). This procedure,
although conceptually simple, requires a significant amount of numerical
computation. Before considering the various realization procedures for
multivariable systems, some basic results for single-input systems are
stated below without detailed derivations [2,3,7].

4.1.2 Standard Forms

The various canonical forms of realization of

n-1 n-2
NG) i eee ee
8(s) A(s)
= n n-1
+ a
Ss 1a 's he + Sas

a (4.7)
= g(s) “a

are given here; these are based on the results in Sec. 3.2.
Realization Problem 235

(a) Controllable canonical form

x,(b) 0 1 0 0

x, (6) 0 0 i doo (Q)

x, 1 0 0 1

x0 ale aeise eo ae ad
(4.8)
x, (4)

x, (t)
y(t) = [p> acne eo +» Bo, By] : ae du(t)

x1

x ie)

(b) Observable canonical form

Ono 0 sor x, (t)

i @ 0 FOG, x, (t)

OO 0 mos x1

0 0 1 —a, x (t)
(4.9)

y(t) = [0 0 *++ 0
236 Transfer-Function Realizations

(b)

Fig. 4.2 Canonical-form realization of g(s) in (4.7): (a) controllable

form; (b) observable form.
Realization Problem 237

The controllable and observable canonical form realizations are shown in

Fig. 4.2, and the derivations are illustrated through the following example
[3].
Example 4.2

a4svx 258" tbs 34 2 0.58°

+ 2.58 tl
+2 (4.10)
Ge) SoS eoet ose si Gaaiose
Controllable form:

fie 0
x(t)= | 0° 0 1{x(t)+] 0] u(t); y(t) =[1 2.5 0.5]x(t)
+ 2u(t) (4.11)
Sseeel 0-6 1
Observable form:

0 0 -8 1

x(t) = }]1 O -10) x(t) + | 2.5 Ju(t)s y(t) =[0 0 1)x(t) + 2u(t) (4.12)

OF a6 0.5

(c) Jordan canonical form

The forms above, also known as companion form representations, are con-
venient in that they can be derived directly from the transfer function g(s).
Another method, derived by partial fraction expansion of g(s), is given
below.

Distinct eigenvalues of A: In this case the coordinate transformation

through the modal matrix (Sec. 2.2) x(t) = Mz(t) transforms the transfer
function g(s) into

g(s) = e(sI - A)-"b + d = eM(sI- A)“M“b +d “(A = MAM) (4.13)

ence
=c diag : by +d Be +d, Ke Gilby oooh 8 (4.14)
2 eases Ree
Thus if g(s) can be expanded as

g(s) = Y -
k=1 ©*
+d

the elements of c and b can be directed computed from

238 Transfer-Function Realizations

hykese bute dketiieeee

en (4.15)

This is illustrated below [8].

Example 4.3

Sot Aerino a
g(s) aa g3 + 682 + lls ae 5 (A es -1, -2, -3)

aS ee (4.16)

A normalized representation is therefore

-1 0 S
z(t) = -2 z(t) +/-1] u(t)s y(t) =[2 1 2]z(t) (4.17)
0 -3 S

Repeated eigenvalues of A: The idea of extending the above to the

case of repeated roots is outlined through the expansion of g(s) into

e Co e e e
is (emer
Sai ee See 4° (al (4.18)

where g(s) has poles at A,, A,, A,, Az, and A3.-
By using the transformation x(t) = Jx(t) it can be shown (Sec. 1.4)
that the derivations above would yield the following realization for (4.18):

zZ,(t) Agim ion Ot OREO 0

Z,(t) 0 A», 0 io 0 0
Za =O 0 Ay 0 Olle z(t) + | ae
“HCG (ho |LtOemms
Meleelg wy 1
Zs (t) 0 20n 0 9 ne 1

WW) = |lOr Cp Sa On Callao) Gl (4.19)

The representation of g(s) above is called the Jordan canonical form. This
is shown in Fig. 4.3a. An alternative form of representation (changes in
b and c) is shown in Fig. 4.3b. Realization in Jordan canonical form has
two difficulties. First, the denominator of the transfer function must be
factored, or the poles of g(s) must be computed. Second, if g(s) has com-
plex poles, A, b, c will have complex elements, and some equivalent trans-
formation with complex-conjugate elements may have to be introduced [3].
Realization Problem 239

Fig. 4.3 Two different Jordan-canonical-form realizations of (4.18).

240 Transfer-Function Realizations

In most cases, however, the elements of J, b, and c can be obtained through

suitable numerical algorithms [9].
Before the results above for single-input and single-output systems
are extended to multivariable systems, the following theorems are stated
to form a basis for further analytical developments.
Theorem 4.2 For the single-input/single-output system,

X(t)= Ax(t) + bu(t); y(t) = ex(t) + du(t) (4.20)

where A, b, c, and d are of dimensions n Xn, nX1, 1Xn, and1X1, if
the transfer function is given by

g(s)
E
= e(sI - A)~b+d
pe eN(G) eA) +d= — N(s)
qeuisr ,
A(s) +d (4.21)

then (4.20) is irreducible only if the polynomials N(s) and A(s) do not have
nontrivial common factor (i.e., if they are relatively prime). This leads
to the following theorem.

Theorem 4.3 The dynamical equation (4.20) is irreducible only if the

denominator of the transfer function in (4.21) is equal to the characteristic
polynomial of A and the degree of the denominator of the transfer function
is equal to the dimension of A.

4.2 NONMINIMAL REALIZATIONS (8,4, 6, 10, 13]

The derivations of the previous sections are directly extended here to

multi-input/multi-output systems. Although such realizations are not min-
imal, they provide a simple procedure for realizing transfer-function
matrices, and a basis for more refined methods considered in the following
sections.

4.2.1 Block Forms

Let a(s) be the monic least common multiple (lem) of the denominators of
the elements of the proper rational matrix G(s) of order £ xX m; then, from
(4.4),

eae ue
Gis) = {6,,(s)} su = aigy
G
7 OST - ANB
=
+D (4.22)
where G(s ) ={ei,(s)}, Bij(s) = gi(s), Fie 1, al a ean tec aa
If p=d{a(s)}, ion p<n=eae, G(s) }, with G(s) expressed as
Nonminimal Realizations QA1

uo}te wo
Gs) ll Q n a Q 0)
Me} + + Q> n ae Q (4.23)

where

a(s) = s p Pont LEUGUes a sta

p-1 Pp
and Cx, eyes 0, ---, p-1, are (£ Xm) constant matrices. Then by following
the realization for the single-input/single-output systems in (4.8), the con-
trollable realization of an m-input/£-output system may be derived [3, 6]

x, () 0 Te 0 OGn 0 x,(t) 0

= (t)
Xp 0 0 In sralie 0 Xo (t) 0 :1 (t)

: he may:
0 0 0 rn 0 us)
. -a -a -a -a t
*, pon ellen p-2 m 1m Sa ) an

(4.24)

y,© x, (t)

Ee Sot G,] u(t)

y p(t) x(t)
where

D = lim G(s)
s—oo

The transfer-function matrix of (4.24) can easily be shown to be

Pp “~
p-k
2 Gs

G(s) = p witha, 0 =1 (4.25)

Deon ce:

=0
This form of realization (called direct realization [6]), however, is not
minimal since the order of the A matrix is p X m, which, in most cases,
is likely to be greater than n [the degree of G(s)], where p< n. If the
observable form in (4.9) were used, the resulting multivariable observable
realization would be as follows:
242 Transfer-Function Realizations

OL Oe LOU Eea al G 0
p £ p

UR ee G 0
(wel ° . .
pty,
.
p-| x
Pl; cle=| e
|; D= time)
g—-0o

Ohana Lpanoel, G, i, (4.26)

The dimension of the state vector in the observable form is 2 Xp, which
again is likely to be greater than n, the order of minimal realization (The-
orem 4.1). This method of direct (nonminimal) realization is illustrated
below.

Example 4.4: Direct realization of G(s)

s 1
(So esa?
G(s) = (4.27)
ast1 i
aia i) Garal

The monic least common denominator of the denominators of the elements

of G(s) is

a(s) = s(s'+ 1)* = s?+2s?+5s; 6{a(s)} = p = 3

G(s) may therefore be expressed as, from (4.25),

3
@ ee1 a p-k I esas WE mee a
(s) PY) 2 G KS i= fee
ie [Gs ats G,s + G.] (4. 28)

Then G(s) is expanded to transform it into the form (4.23):

y -s? s(s+1) -s* s?+s

G(s) = =
(28+ 1)(s+1) s(s+1) 2s7+3s+1 s*+5s

aos An OLeneL 0 0
= g* + st
my ih 3 1 i 8

G, G, G,
lim G(s) = [0]
sS—co
Nonminimal Realizations 243

The direct realization in controllable form is therefore, from (4.24):

0 IL, 0

A = 0 0 I.|, 2, =

-2 31, I, -a,1,

One Ge ot 0 0

oa 0 0 0

OR ea 0 0
Aa

i Or “OF 940 Tlie Bear io C=, G, G1]

Or a0 71 0 1 QO @ @ teali

Ore de aaa nel -2 0 eS

Oyo
(Oe
OFS ta sOeeroneml
ss *Ziagee

(4.29a)

The observable form of this direct realization, obtained from (4.26), is

Oh Oye J Te =O)
OPS:

eo 0 On,
O° 1, rsagly
i
aS
SS S&
TSS)
SerSy
OS
SSeS
SS)
SS
=

0
1
G, F
B= Q, alee (4.29b)
G, oA
2 oO
Oo
FEF
FP
PrP

The realization procedure above, although fairly simple in terms of numer-

ical computation, is rather artificial in that the order of the resulting system
is excessively high. The method, however, may be refined to derive reduced
order realizations; these reduced realizations, though nonminimal,
form a realistic basis for minimal realization. The modified method is
based on identifying the least common multipliers of the denominators of
244 Transfer-Function Realizations

the m columns (for controllable form) and the £ rows (for observable form)
of the (£ x m) transfer-function matrix G(s), rather than a single lcm for
all the (2 X m) elements of G(s). The derivations of controllable and observ-
able forms are considered separately [1,11, 14].

4.2.2 Controllable Companion Form [1, 6,10, 12]

Let {Bj‘(s)}, j@ 1, ---, m, denote the m columns of the proper rational

tee G(s), di(s) be the monic lem of the denominators of the elements of
gj(s), and let 6j = deg {al(s)} for j ¢ 1, ..., m. Then G(s) may be expressed
as

Gis) = [g,(s) - 8,(5) * 8,

(6))
and any element 8,5), icl, ..., £, jE1, -.., m, may be expressed as

pisOia ees ee
g..(s) = SL (4.30)
= leo! so) ard
1 6j

Let the coefficients of the numerator and denominator polynomials of (4.30)

be represented through the elements of two column vectors gil and aJ:

0j 0j
es ij
Par aj
pli =| 6j-1 :
O° }@ixs a!j = j-
Pie? 3
(5) <4) (4.31)

Byij aj

By following (4.8) a controllable realization of 84;(S) may be expressed by

S)Ay bs»c.)” where

0 1 0: 0 0
0 0 iis 0 0

Aes 3 (OPO9F bo= | 1(6j Se);

j 0 0 aoe 4) ae j
0
-a) -a) -o) iff
Nonminimal Realizations 245

ley »
c, = [8 (a x 69)
ij ij ij
= (p60. nee eve By

Defining a vector e, as the kth column of an identity matrix (i.e., the kth
element of e, is unity while all others are zero), the matrices above may
be expressed in compact form as

et
2

' sey
A=
tsé
alte [e,)3
b, =[e..J3 c..
{jek
¢, ==[85]4 (4.32)
A3Z

e
6j : Ew Jlgtesh peas
a (i e., C14 7 (B.] ’ O41 el [Pe 4] ’ etc.)

Realization of the entire system may now be completed by combining the

individual realizations are given above by taking j from 1 to m and is given
by S(Acg, Be, C,), where

0 0 0
a
0 0 0
A»
A, = ie =diag {A,}, je 1, +--+, m
0
nal A, (4.33a)

ov 0 0 A_|(N xN )
c (6)

0
by

be
Bee che b.. jie 4, . 1m
=diag?
Cc » J
j (4.33b)

0 b |(N xm)
m Cc
246 Transfer-Function Realizations

hig ee 1j 1m
2 Akuma 2\poteak vx nmap er
Gaal: je (4.338c)
§ it 1512 ij im

eke asi xN
“or oe ©3 earns)
uoT
where any ore = (65) is a (1 X 6j) row vector and N. = aH ee This reali-
zation may be proved to be controllable by extending (4.8) to multi-input/
multioutput systems; the method is illustrated below [6].

Example 4.5: Derivation of controllable form of G(s) = [g,(s) g,(s)]

ms as Bes
(s-1)2 (8s-1)(s+3)
G(s) = (4.34)
4. 6 s-2
(s -1)(s + 3)? (s+3)2

il
(sje ; (s+ 3)?
£,(s) = DPE
a ERE
7 6 Be cian)
(s -1)(s+3)?
1 (s? +6s+9)
3 SE ; Ge, = Z!
4 a ies, td 1
s#+4s3-2s 12st+9 Dgeems

O-- a OP 0 0 9 6
Of) Ov aereeth Rs
ee nscale eo een
1 Or. OF “20h, at a of? F541 Ciesi.
\n.0
=9 412.) of 24 1 0 0

Thus

tier 218%
={9 6 1 OJ; Co, ~ (Be3] =(6 -6 0 0]
Nonminimal Realizations 247

1
SS +
(s -1)(s+3) ee
g3(s) = ees
ae (s - 1)(s +3)?
(s +3)? (s -2)(s - 1)

' st3

eee
$3+ 582 +3s-9 ;
62 i:=3
S3812

0 iL 0 0 3 2
ar, 22
Porefe, OF 0s als) ey 2 0 i Bso = 1 ;
? Bso =| - 3

9 -3 -5 1 0 1

Fa Sd : e220:
C19 [Bo] = [3 1 0); Coo [B59] = (20-3 1}

The complete controllable realization is therefore

Om 104 On
1

(eae ieans O11

|
A, 0 0 0 0 it 1

ae = 1/20 112.) 42° 4 ;

Vy gh eaten * peal oO4 Oo.
1
1 Ok Oo BL
I

fe ae gers
0! (4.35)
|

01
0 |
bi ae Ci "ers 9 126 [Link]
04 20
a 20 Perches col |6ece On.) 2 ort
= aL ! hs (el oS =

0” b;
1
1
1
1
I
!

The order of the controllable realization above is Ng = 61+ 62=7, which

can be shown to be less than the direct realization of (4.34) obtained by
using Sec. 4.2.1.
248 Transfer-Function Realizations

4.2.3 Observable Companion Form [1, 6,10, 12]

The proper rational (£ x m) transfer-function matrix G(s) is now expressed

in terms of its £ rows g;(s), ic 1, -.., &,

g,(s)

G(s) = g (Ss)

g(9)

Let d'(s) be the monic lcm of the denominators of all the elements of any
ith row, and 6i be the degree of d4(s). Then any element 8ij(S) may be ex-
pressed as

eo nae ie
vs : +38 Te aes bays
g,,(8) = ey ic tnirtees (4.36)
1 OL

The coefficients of the numerator and denominator polynomials are assumed

to be the elements of the following column vectors:

ij Fe
Y6i éi
vl at
a ae
y=] loixys at =| "| eixy (4.37)

ij i
v, a

Then by using (4.9) an observable realization of the scalar transfer function

8ij(S) in (4.36) may be described by S(Ai bie), where

000 90 io
6i
pa ere i ; P
A, = 10 0 a] (i X 6i)

ip er 3
1

Sey. GA PQS -a }
Nonminimal Realizations 249

ij : ial 12
b ij = fy aloe3 ices.
-@. bi ta Lys] ’ b 12 ae Ys] ? etc
te.

c. = 10 @O ile le.) (1 X 6i)

where, as before, es; is the unit column vector of dimension 6i for any i.
Therefore, if the realization above were extended to allic 1, ..., 2,
the observable realization of G(s) is obtained, which is given by
S(A, By Cy):

A, G- © ¢ 0
0 Ay 0 O 0
eae te, A, 0 = diag {A,}, i Gl 1 jens, 0 (40885)

0 0 0 0 Ay (Ny x N))

Py Die re Daan
Bai. [OC Way Oo Woy (4.38b)
fe) il i2 ij im

x
a Poe 0 Pini oe

a
“2
C, = 7 c, = diag
di te} ; Cc;ge
Ons diced
ic es Oate

: (4.38c)
cy (LX N))

ek
where each bi; = Ly] is a (6i X 1) column vector,
e, and N,
Be:ay [Link] The

realization may be shown to be observable. The method is illustrated by

using the rows of (4.34).

Example 4.6: Observable realization of

£1 (S)
G(s) = (4.39)
82 (Ss)
250 Transfer-Function Realizations

of 1 if ic1 real
£1(s) ie-1)? (s aril (s =1)7('F3) [s ]

PS pS ate eZ
Ew moa cia a
3
ats
=

eg aie 3 “1
A, 1 0 5 Tha 1
|;° UG at 51
= Pere 3 =) es PY c [
=[0 0 a]

ed Wilgten! 0 6
a 6 s-2 = 1 a a hs
B2(s) = E(s -1)(s+3)° cane SSG? sg te ae.
al
Wesmneraeetg lO s*S3s+2]3 “62=93

On 50) 9 -6 2
21 vy)
A, a 1 O = -83 > Y59 ae 0 ’ Y59 Sat 3 f) co = i © 1 al,

Oa iee Ss 0 1

The observable realization of (4.39) is thus

0 0 -3!
1
Ol oa 0
|
AG 0 Ondeet
IN Bin (ale oem ea eee
) @. A> OF 02 39
1
0 i eR O2=3
1

10 tf 95

3 -1

he, Vi 19 te Pe
ik | 12 Vet Net Oro
B

co)
—

b 5
=
yt ve
=
“6 8 (4.40)
%

Zee 22 62 62 aes

0 1

¢, 0 0) 502 “SI 02m Ose m0

Nonminimal Realizations 251

Order of this realization N, = 6, + 6, = 6 is less than its controllable form.

It can be seen that for the example above, the degree of the monic lem of
the denominators of all elements of G(s) is 4. By Sec. 3.3.4, therefore,
the system is of fourth order. This suggests that neither of the realizations
above is completely controllable and observable. Hence the realizations

(which have orders Ze ONgOr = 6i) are unlikely to be minimal. If, however,

the unobservable modes in the controllable realization Sc(Ag, Be, Ce) [or the
uncontrollable modes in the observable realization Sg(Ag, By, Co)] are iden-
tified and removed, the resulting system can be made completely control-
lable and observable, and hence irreducible. This can be achieved by
applying the canonical transformations of Sec. 3.2 on either S,(Ag, Be, Ce)
or So(Ag, Bp, Co). Thus by using this two-stage procedure, an irreducible
realization of any rational G(s) may be obtained. This is illustrated below
[15].
Example 4.7: Irreducible realization of G(s)

Sel seissmale ab
ae s* %s
G(s) = 4 (4.41)
0 0 =
s

g(s) 2,(s) £3(s)

Stage 1: Controllable form of G(s):

ii aba
, (e+ F role baloacs
= eo ; = 2
81 (8) = 32 . eo ais oe, gel gre 1
61 61-1

12 12
1 Sica! 1 Poo P5o-1
eck =) a ; =2
82(8) = & s*+a%s+ a3 22 | gee 2
62 62-1

13
1 P53

es : staj ge? ee
63

Therefore, the controllable realization of the above is

252 Transfer-Function Realizations

0 1a OD ertt
I 1
Ay 0 sal at | 0 0 01
eee ahem gn
Aas Ay = ano tal = 'O 11
I ! I 1
0 A; 0 | -a% -a7 10 0}
See ne pene es || ee ee eee ae ses
1-9} 0
0;
I

Di
b, 0 ao

Dee by = ial : (4.42)

oi fees
0 b; 10
1
fe

nla 12.7 12.T

[854] [bso] [Bs3] E Onaie.0 |
Cm =
c 21 22 T 237 G00 0071
[sisal Geel “Bes!
Stage 2:

1. Check for controllability: p[B AB] = 5 — controllable.

2. Check for observability: only three linearly independent columns
of C, hence the observability index is 3. The observable form of
(4.41) can be shown to be [15]:

OR eae) ORO mae: Tat ies

AO = 0 0 0 5 B, = il 1 0 ; a = OIRO: i

ORO O Oy al

of order ngog < n. Detailed derivations (using Sec. 3.2) are left
to the reader.

4.2.4 Normal Form

Since the transfer-function matrix represents the controllable and observ-

able part of the system, the method of direct derivation of state-space
representation of Sec. 4.1.2(c) may be extended to multivariable system
as well. This is considered here for distinct eigenvalue case, realization
for systems with repeated eigenvalues (Jordan form) is considered in the
following section.
The transfer-function matrix may in this case be directly extended
from (4.14) as
Nonminimal Realizations 253

G(s) = C lst - A|"B. a0 BI (4.44)

where Cy and By are the normal forms of C and B, and A = diag {rx}
k=1, .-., n, are the roots of A; (4.44) may thus be expressed as

n G n
G(s) — Nis)
ats} _ k +D; :
L Tay d(s) & I (s - A) (4.45)

where

¥ T
GrEIG be (4.46)

Cnk is the kth column of cy, bar is the kth row of B,, and G, is of dimension
2 xm [same as G(s)]. The residue matrix G, in (4.46), being a vector outer
product, has rank 1 and can be derived by

G = lim (s - A,)G(s)
k s—),

with (4.47)
D= lim Gs)
s—oo

Once G,, ke 1, ..., n, are computed, cy, may be derived for an arbitrary

value of Bees or vice versa.* The resulting realization is therefore not

unique, and consequently may not be completely controllable or completely
observable. If, however, the uncontrollable or unobservable modes of the
system are known, the summation in (4.45) need only be taken up to no,
the number of controllable and observable modes. The order of the control-
lable and observable realization will then be of order nog < n. The method
is illustrated below.

Example 4.8: Realization of G(s) through residual matrices

il 2 il
Shar Ab Jena st+4

& 1 iL
G(s) = 0 Per ad. (4.48)

1 2(st3
stl (st+1)(s+2)(s+3)

a
* An alternative method of deriving b ie and c Ae is given in Refs. 3 and 16.
254 Transfer-Function Realizations

The lem of the denominators of all the elements of G(s) = d(s) =

(s + 1)(s + 2)(s + 8)(s + 4).
Full-order realization:

2 iy 1

Guo lind (s+ Gia 0 0 0.) = 1407) [ey 2 00)

pe | —S on
iP ey 0) il pe
Same nl
nl

Similarly,

0 0 il

Goe= 1s Oe Oso eeGe — 1 kOm se et eee eG onal 1a SOO eel]

Sa 3 ——_ 4 Aa
-2 pe 0 be 0 be
aad n2 aes n3 anak n4
n2 n3 n4

The realization is, therefore,

ih 0)
ono ik © @O- 2
IN = diag
7 = =o)
= = -3.-4 g BL = Me ad ike & Cc = OFT AOL (4549)
-

0.0 1 12) 00)

The system is not completely observable (Fig. 4.4).

Fig. 4.4 Signal flow graph of full-order realization of (4.48); x,(t) does not
appear at the output.
Nonminimal Realizations 255

Reduced-order realization: By inspection, it is seen that G3, =

2/(s + 1)(s + 2). This reduces d(s) to (s + 1)(s + 2)(s + 4). The resulting
realization therefore need not contain the pole } = -3 and the third row and
column of By, and C, (Fig. 4.4). Thus the reduced-order realization is

tO) 1
ya = di jag{~— =1,-2,-4}
= = 3 B=)= 06 1 6 ; Ch = 0
Ov: 1 1

which is completely controllable and observable and hence a minimal reali-

zation of G(s).

Example 4.9: Completely controllable and observable realization with

1 1
(s+3)(st+4) st1
G(s) = (4.51)
: 0
st+3

The system has three poles at -1, -3, and -4. The residue matrices are
therefore

CG I lim (s+ 1)G(s) =

s—--1 Oo i=)

10
G,=G = lim (s + 3)G(s) 3
A= -3
s—-3 17 40

-1 0
Gz G._ = lim (s+ 4)G(s) =

Therefore,

al iL i at, if
al iT
Goes be = [O, ds. (G2 = ce be. = = i200
1 nl nl 0 0 0 2 n2 n2 1 0 1

-1 0 i!
Le
= — = =i) 0

CE kes !
256 Transfer-Function Realizations

x,(t)

Fig. 4.5 Signal flow graph of minimal realization of (4.51).

The diagonal realization is therefore

Out
: ; Bey a ee A
ea diag {-1,-38,-4)3 Bobs bos uate {1.05.0 Fale aan Car
=) 0
linia?
oa ipa a
(4.52)

The realization is controllable and observable (Fig. 4.5).

4.3 IRREDUCIBLE REALIZATIONS

The realization methods considered in previous sections, although concep-

tually simple and convenient for hand calculations, have the disadvantage
that they may require further transformations to yield the minimal repre-
sentations. In the following derivations it is shown that the Smith-Mc Millan
form of a rational transfer-function matrix may be constructively used to
obtain a controllable and observable (i.e., minimal) realization. Before
the realization procedure is developed, however, the concept of direct sum
is first established.

4.3.1 Direct Sum

It has been shown earlier that for any m-input/-output nth-order system
S(A, B, C) with repeated eigenvalues, there exists a similarity transforma-
tion which yields the irreducible* form S(A, B, C) defined by

*Controllable and observable (Sec. 3.2.4).

Irreducible Realizations 257

A= SAT. B= TB; C= CT (4.53)

Thus if A has q distinct eigenvalues A1, A2, ..., Ak, .-., AQ, with any Ak

being repeated nk times ee nk = n), then the triple (A, B, C) is given by

(using notations slightly different from those in Sec. 3.1.2).

AS diag {Ik 1K)} a ke ihe See Se q;

Bl

= =k anager ak ad
B=|B : PECL [ C : J px) ;
(4.54)

ad
(nXm)

where a (Ak, nk) is the Jordan block associated with the eigenvalue Ak in A,
kel, ..-, q, and has the form

J ak,nk) = |: (4.58)

Y OY O & Bebe

In the subsequent derivations the Jordan form in (4.54) is assumed to be

irreducible, and consequently the input and the output matrices have full
rank [i-e., p(B) =m and p(C) = £]. The submatrices BK and CK contain the
following m-row vectors (denoted by subscript r) and £-column vectors
(denoted by subscript c) [from (3.38) and (3.39)]:

ak k |__ J jth row of Bo

; (4. 56a)
rj an m-row vector
258 Transfer-Function Realizations

nk
SS ee

ee = ick Oe cK 0 (4. 56b)

en Cj enk! }

=k
jth column of C :
an £-column vector

and

q
n= y, nk (4. 56c)
k=1

Combining (4.54) and (4.55), the transfer-function matrix of the transformed

system may be expressed as

q
G(s) = C(sI = A) 'B = Gis (ee G_( s) (4.57)
iXe
iia k=1
where G(s) is the transfer-function matrix of the kth subsystem 5.(J*- re cy
with 6{G,(s)} =nk. The derivations above show that the vector space x” is
the direct sum* of the q vectors xnl tenegs xnk o.+, x» 4. Thus the com=—
posite state vector

nq
x

according to the definition in Sec. 1.1.4, is the state vector of the composite
system S( - ) formed by the parallel connection of the subsystems Si.( =)
that is,

u(t) = u,(t)
q (4.58)
10) = AO!
al

*In the general definition, neither the subsystems Sk ( *) nor their direct
sum S(- ) need be in Jordan form. The transformation is used for compu-
tional convenience.
Irreducible Realizations 259

Jordan Block

(b)

Fig. 4.6 Schematic representation of (a) Jordan form S(J, B, C) in (4.54);

(b) the kth Jordan block of S (5 B. C.) in (4.55) and (4.56).

The resulting representations are shown in Fig. 4.6; consequently, the

dynamical system represented by (4.57) and (4.58) may also be expressed as

we g (ai, n1) ae

xfeed hs aa nk) = +

x (ng, nq) ||x°4

260 Transfer-Function Realizations

Eyre
(oie Cre on on kee (4.59)

Thus once the transfer function G(s) for any subsystem and the associated
repeated pole Ak are identified, their state-space representations may be
derived independently. Another consequence of the above is that the direct
sum S(°) is controllable (observable) only if each of its constituent sub-
systems Sic( - ) is controllable (observable).

4.3.2 Jordan Form [15-18]

The derivations below establish a correspondence between the foregoing

diagonal representation in the time domain and the transfer function through
the Smith-McMillan form (Sec. 1.5).
It is assumed that the transformed system in (4.57) has a strictly
proper (£ X m) transfer-function matrix*; then it can be expressed as

G(s) = L(s)M(s)R(s) (4. 60)

where L(s) and R(s) are unimodular matrices of dimensions £ X £ and m Xm,
respectively, and M(s) is the Smith-McMillan form of G(s) given by

(8) |
diag 3,8) | 0 |} a, 619, (s)}=nk
!

ay a aT i teat halbe atari

NN rr
qe md
-1
ai} diag {y,(s)}i 0 | |diagig(s)H 0 | }q, kel,...,q
eh gy (ee he 8 Cee etree acres
L-at{ 0 eno | 0 eee tee aog
ae Need — ee Ss
q m-q q m-q (4.61)
eeee a

A(s) T(s)

*This is generally valid, since the transmission matrix D can be computed

directly as D= Tim 5 oo G(s), and the strictly proper part of G(s) easily
identified.
Irreducible Realizations 261

Thus if LCi(s) denotes the jth column of L(s) and R's) the jth row of R(s) for
any jcl,..., £, for L(s) andj =1, ..., m for R(s), then (4.61) becomes

G(s) = L(s)[A(S)]IF(s)]"
Rls) = {Ls }[diag {y, (s) Hldiag{ o, (8) 7 {Rs}
1
= 1L aca (5),(8)> <(5)
Sh) R rj “(s) (4.62)
where

(1) PjK(S) and pa,(s) represent the corresponding diagonal elements

of A(s) and I(s), defined as below:

Y,(8) for jak eae eee

p(s) =
ik 0 fOLIE—" (Qikr 1). rs 0

Op tiore is GI Ae oe al
k
o, (s)
J a ore Tp (Cle ID Goo, sail

Consequently,

Ly, (s) forsjokic 1s28 -... O

x
(2) L's),(8)= (4.63)
0 toy i Sahay Ia come, 100)

and since there is only one pole Ak associated with any o;x(s), ke 1, 2,.-.,q,

nk
(3) 9, (8) = (S - Ak) (4.64)

where 6 {¢;,(s)}> 51,_4(s)} because of the divisibility property of the

invariant polynomials of M(s). Thus, combining (4.62), (4.63), and (4.64),
we have*

L(s)y,,
(8) (s)
G(s) = d eee |

qd Ly (s)R'(s)
= ey Swed, (ey0 (4.65)

. , L1(syR™(s)
ee ne)
where

*The summation being taken over j = 1 to max (m, £).

262 Transfer-Function Realizations

Le) 2 L(syy, (8), j=k=1, 2,.--.,4q

Thus comparison of (4.57) and (4.65) yields*

=—Cj Yj
L ‘(s)R_ “(s)
G(s) = fork=j=1, 2, --., q (4.66
i! ) $,,(8) J )

To establish a correspondence between (4.54) and (4.65), the following

identity is necessary:

[sI - i: Olean po = adj {sI - soak uN

(s - ak)"
(s - Nora (s - ARPES oe i|

ee: 0 (s - ARE s - Ak
$,.(8)
0 0 (eile

1
s -Ak
= mC [ia oak >-:-g-ak 1] | een

(s - Ak)
Since

G(s) = Gist : Tipuerhs BS (4.68)

combination of (4.66) and (4.67) gives

Ss -Ak
k
L(syR"(s) = $,,(8)G,(s) = C eae a eek
eae

row matrix
nk-1
(S - Ak) (4.69)
Se Nai a

column matrix

* Each G(s) is of order £ Xm.

Irreducible Realizations 263

Thus, equating matrices of similar dimensions gives us

7L“(s)=C
Cj 2 . img ‘
(Aah ey eeae (4. 70a)

and

a) (Ce a 1B*; 5 Sk eile 2 eer. (43:7 0b)

since Si (Ik, By, GC.) kel, 2, ..., q, are assumed to be controllable and
observable, the derivation above provides a procedure for irreducible reali-
zation of a given G(s). A few illustrative examples are given below to high-
light the main features of this procedure [15, 18].

Example 4.10: System with multiple poles at the origin

s?-s4+1 il Streea9)
i
G(s) = a 1.5s+1 sal -1.5s-2 (4.71)

s°-9s*-st1 <s*+1 $°=e-s

The degree of G(s) can be seen to be 8, and hence the order of irreducible
realization (nyjn) is 8-

Stage 1—Transformation of G(s) into its Smith-McMillan form:

G(s) = L(s)M(s)R(s) = L(s)A(s)[T(s)]“R(s)

For the system in (4.71),

1
a4 0 0

Mis)=| 0 G1 0

GaaGr pie
s

iL 0 0

L(s) = }1+s il 0

t=s% -2-10eF2s% <1

264 Transfer-Function Realizations

3? a6241 1 os? 522

Ris) = | -s°+st0.5 0 s?-s+0.5
2s3-9s-5s+9 0 -283+9s2+5s-9
ee ee Ce Oe

f t
Therefore,

A(s) = diag tb,(s)} a0

Ts) diag {d(s)} = diag {s*,s°,s}

Stage 2—Identification of the number of Jordan blocks: For this sys-
tem, there are three Jordan blocks, J(0,4), J(0,3), and J(0,1), corre-
sponding to $,(s), ¢,(s), and $,(s) in I(s).

Stage 3—Realization of the individual blocks: In view of (4.69) and

(4.70), the three blocks may be realized independently.

Block 1: J(0,4): nl =4, Al =0

iL

s
Ciel
J ee oe 52 [s? s* sel] = BE
a3

=3 1
34 isz 3 s2 a2 4s) 68Al (4.72)

i RB 4s)
p01 (gy P28)

1 1
where Phe (s) and Ro (s) are the first column and the first row of L(s) and
R(s), respectively, and

Aloe
C {eis}
(4.73)
BPl — 1
tb;
Irreducible Realizations 265

Combining (4.72) and (4.73) gives us

1 1 1 1 1
Ci, C2 «C3 Cg 1
1 1 1 1 SH) cls ks
Ca C22 23 S2ay | i} ~ L (s)=|1*s
1 1 1 1 Ss : 2
Cay Co2 mecosmences 1-s

Comparison of terms of with similar powers of s gives

1t0°
TIO: REO
Chena
hh. 0" 0

Similarly,

bh bi, bi;
Bere: bau bee bes —[s7—se
3 2 tl 3 2
[SwtSacSeel
‘bh bh bls 1 st s45= 2]

bar Ba bg
which yields

ib Oy il

= -l oO 1
S oO, OQ, @
ed ey

Block 2: J,(0,3), n2=38, A2=0. The procedure is as before and

uses the following equations:

at
= _ = r2
([O4le | =i (s) and [s? | 1B¢] = R -()
a2

giving

ch Ch, Chs a a8
Coe Weee mec C5.) le 90 40
Cone eeaChs ET)
266 Transfer-Function Realizations

Sh eee on aolting
Bai=| bee beibsa ha L 20 eek
be bs cbs; 0.5550" 7005

Block 3: J,(0,1), n3=1, AI = 0. The equations are

feiil= bo (a\eands (iio) =m) (3)

which when equated for similar powers of s give

Ca 0
C3 =| c3,|= |] 0} and 5 = [b3, bj, b7,) = [9 0 -9]
ea; 1

Comparison of the various blocks of B? and C! suggest that terms marked f

in R(s) do not appear in B and C. The minimal realization of G(s) is there-
fore given by

OS —1e O20
|

(ete oe1 5
Oer0e On ele
i
‘ 02) 0» 20.7 8047
A = diag {J"(ak,nk)} = |~-tono Whe aie tod (4.74a)
| I
|i I
; |0 0 .
eee ei
a ee calcao

1 0
1 OF a1
1 0 i
<3 0 0 0
B!

ae he iL do Wee
B =| B2]= (4.74b)
:
B
0 0 0
1 Oe H1
=(.5) 1 Omen OLD
9 0 -9
Irreducible Realizations 267

157208 30 0M OA) 100140

= 2 = 4 | I
Coa CeCe C2 yay 1 1 a0 01 1. 440 010 (4.74c)
de GhOmytt © Os2 -10. 219
which can be seen to be controllable [o(B) = 3] and observable [p(C) = 3].
A close observation of the derivation above would suggest that reali-
zation could almost be derived by inspecting the unimodular matrices L(s)
and R(s). This, although a strength of the procedure, may not be possible
when the multiple poles are not at the origin (Ak # 0) and the order of multi-
plicity (nk) is high. The problems of deriving the irreducible realization by
hand calculations are highlighted through the following example.

Example 4.11: Irreducible realization of G(s)

i (s+ 1)(s+3)
(s+ 1)? (s +1)?
G(s) = (4.75)
s(s+3) s(st3) Ss
(s+1)3 (s+1)2 ~ (s+2)2

Stage 1: The Smith-McMillan form of G(s) is given by

st3
eo (s+ 1p 0 ibe» Baral+

G(s)=
s
ye Gh 0 (s+ 2)2
Sa 0 =
1

1 Ol{fs+3 o][(s+n? One tn ent

oe! 0 s 0 (s+ 2)? OC eak

s+3 01](s+1)3 0 Te 1 Jeet

s(st+3) s 0 (s+ 2)2 Or -1
ee era So

L(s) T\(s) R(s)

Stage 2: There are two Jordan blocks, J(-1,3) and J(-2,2), corre-
sponding to $,(s) =(s +1)? and $,(s) =(s + 2)? inI(s).
Stage 3: Realization of blocks 1 and 2.

Block 1: The equations are

268 Transfer-Function Realizations

1
ei| cal Bak eee non BEDS eB ee
(s+ 1)? Gt, + Ohj(sit 1) cp,(8+ 1)7 s(s+8)

and

rl
[(s+1)? s+1 1]B! = R (s)
= [6*1)*bint (8 ti)by + bi. 1(6+ 1)7bi, > (8+ 1b, + biol = (hi Sail
As can be seen, solutions for the elements of C! and B! may not be easy to
obtain without the aid of any numerical algorithm. The solutions for this
case are

: Zwei "0 a
Coe eile ee | aes
0
Block 2: The input/output matrices are derived as above:

Seis? (et
Ons Ali 0
Deiat Ole nye = 0 0 ani (4. 76a)
Bass;TL eete i
1
10 22
0 0
re Ome
ge Lo pe ee (4. 76b)
Be ihoa
0 0
oe Oo Ops OneekO
C.=afes Cs] '= (4. 76c)
—2 1
Irreducible Realizations 269

For the given system 6{G(s)} = 5 =6{I(s)} =myjin, and therefore the
realization above is minimal.

4.3.3 Companion Form [15-18]

The realization procedure in the preceding section, while simple and fairly
convenient for hard calculations, has the limitation that any denominator
polynomial ¢;(s) should contain only one root. The concept of this special
case is extended below to remove any structural constraint on the Smith-
McMillan form of a proper rational G(s). The derivation of a generalized
irreducible algorithm uses the rather complex modulo theory of modern
algebra.* For the sake of brevity, only the definition of modulo and the
basic equations for realization are given, without any detailed derivation
of the abstract theory.
Definition 4.2: Modulo Two polynomials @ = a(s) and B = f(s) are
said to be congruent modulo the polynomial ¢ = ¢(s) [written as
a@ = £ (mod ¢)] if a and 8 have the same remainder after division by ¢.

Similarly, two polynomial matrices P = P(s) and Q = Q(s) are said to

be congruent modulo $= ¢(s) (= a polynomial) [written as P= Q (mod 9)],
provided that the corresponding elements of P and Q have the same remain-
ders after division of ¢. Thus P = Q (mod ¢) only if there exists a poly-
nomial matrix R = R(s) such that P=Qvt+ @R.

Definition 4.3: Equivalencet Two rational matrices, G, = G,(s) and

G, = G,(s), of appropriate dimensions with least common denominators for
each of their entries as ¢, = $,(s) and $, = ¢2(s), respectively, are said
to be equivalent if ¢, = ¢, and there exist two unimodular matrices Uj(s)
and U,(s) such that 6G, = U,(@G,)U, (mod ¢), where ¢ = $,= ¢2-

Since two linear time-invariant systems are said to be equivalent

(Sec. 3.4) if their transfer functions are equivalent, Definition 4.3 forms
another definition of equivalence. This alternative definition, together with
the derivations in Secs. 4.3.1 and 4.3.2, forms the basis of the realization
procedure below.
Since the system S(A, B,C) is assumed to have nonrepeated eigen-
values, its controllable and observable form §(A, B,C) may be assumed to
consist of a set of controllable and observation subsystems §;(A, Bj, Cj)
with the corresponding Aj being expressed in the companion ona

*G. Birkhoff, Survey of Modern Algebra, Macmillan, New York, 1953.

T Called weakly equivalent in Ref. 18.
tControllable canonical form; subscript i is used here to distinguish this
representation from the canonical form in the previous section.
270 Transfer-Function Realizations

0 il 0
: 0 iors yO
aan hae ae (4.77)

-a ~a a
ni-1

associated with the characteristic (minimal) polynomial

ni i ni-1 i ni-2 i
@(8)
"= 8) Oe s ae i ge eae
i ni-1 ni-2

It can then be easily shown that for any ni < n (= the order of A) [18]

il

s me = Ss eee
adj [sI-A,] f [Y9(s) 7,(S) Vane
° See

ni-1 (aie)
Ss

0 0 0 oO

al 0 Q ©
- $; (8) 2 1 pen (ni - 1) (4.78)

gni-2 gni-3 dee

(ni - 1)

where Y,,(S) is a polynomial given by

ni-1-k ni-2-k
7,8) = s + Ce eee f as + we : eSE Wy dbp eee, ,ni-1
DI-

Using the definition of modulo, the equation above may be expressed as

adj[sl-A} =|: — |[y9(s) 7,(s) °** Y,4_1(S)] (mod ¢,)

ni-1

a
The transfer function G;(s) of the ith subsystem 8(A;, B C.) may therefore
be expressed as
Irreducible Realizations PU(AL

(SG(s)= C). Hy (s) 7,08) +** ¥,4_, (0B, (mod 4) (4.79)

The definition of modulo is now used to establish a relationship between

(4.79) and its Smith-McMillan form. Since G(s) is proper rational, it can
be expressed as

G(s) = 1G,(s)} = L(s)M(s)R(s) = {L,c (s)} | |diag

: v,(8)
}mai iR;is(s)}

or

v,(s)
G(s) = L.(8) R,(s)5 iel,...,q [qas in (4.61)] (4.80)
x)
where Iss is a column matrix and Rj (s) is a row matrix. By the definition
of modulo, then from (4.80),

be
o,G, = L. 1 (=) Re (mod $,) (4.81)
1

Thus combining (4.79) and (4.81) gives

L,Cc(s) (¥,(8)) R,ig(s) (mod ¢,)= C.}-

malls
| [vp(8) 7,8) *** ¥,45_1S)1B;“ (mod ¢,)
ae (4.82)
which may be split into two separate equations:

ip

A g c
C.\>, = L, (s)¥,(s) (mod ¢,) (4.83 a)

ni-1
Ss

[yy(s) 7,(8) *** Yy5_1 (IB, = R, (s) (mod ¢,) (4.83b)

= ais
272 Transfer-Function Realizations

Therefore, combining the derivations above and the definition of equivalence,

we have

= (¢ ae
{¢G.+ = 1L,(s)}Ho,M, HR, (s)} (mod ¢.)
or

oG = L{pM]R (mod ¢)
Il L[{M']R (mod ¢) Reels
which is consistent, since M is the Smith-Mc Millan form of G if ¢ is
assumed to be the least common multiplier of all the denominators of M
[i-e., if d(s) is the pole polynomial of G(s) or the minimal polynomial of A].
The significance of (4.83) is that while G is equivalent to its Smith-
McMillan form M, $G is equivalent to 6M, where $M is obtained by replac-
ing the numerator of each element of M by its remainder after division by
the corresponding denominator. This follows from the definition of modulo,
a further interpretation of which is as follows: If a/f is any quotient of
polynomials with 6 dividing $, then the remainder after dividing $(a/£)
by ¢ is just ¢(a@'/f), where a' is the remainder after dividing a by B.
A comparison of the results above and the derivation of the repeated-
eigenvalue case would suggest that the use of (4.83) stems from the need of
establishing a conceptually simple mechanism of obtaining the inverse of
(sI - A), where A = diag {A;}, A; being in companion form. As a consequence,
the resulting realization represents ¢(s)M(s), which by Definition 4.3 is
equivalent to M(s) and hence is equivalent to a realization of G(s). This
method of realization is illustrated below [18]. The order of minimal reali-
zation in this case is

q q
= ); 6{¢,(s)} = 2, ni
eal =a

Example 4.12: Irreducible realization of G(s)

it 2s-1 s-l
s(st 2) S(st 2) s(s+ 2)
G(s) = 1 Pa 1 (4. 84)

“s+2 ~(stij(s+2) (st1)(s+2)

Step 1: For this system, ¢ = ¢(s) = s(s+1)(s+2) =s?+ 3s? + 2s, and
the Smith-McMillan form is given by { t
a> Oy
Irreducible Realizations PATS)

1 0 eaten. 0 OMe wes

— aso

G(s) = 0 2 1 (4.85)
ae 0 Ae ear hh 0
—— | 94 Sst 2
a
eS ee
L(s) M(s) R(s)

From (4.83),

(s)G(s) = L(s)[¢(s)M(s)]R(s) (mod ¢) (4.86)

where

il
s*+2s :
[¢(s)M(s)] = (4.87)
0 -3s -2
s*+3s+2

(It can be seen that (4.87) represents the proper part of M(s)—which provides
another way of deriving [¢(s)M(s)])-

Step 1: Once (4.86) is established, realization of the subsystems cor-

responding to the q subsystems [G;(s)] follows the same procedure as for
repeated eigenvalue case.

Block 1: Second-order system, since $,(s) =s* + 2s = s*+als+ aj;

therefore,

A, = — nl=2 (4. 88a)

and as the ) System in (4.84) has three inputs and two outputs, the corre-
sponding B! and C! may be assumed to be

" by, by, bj, Ch Ch

1 = anda Ca (4. 88b)
ba ba, by Ch C22

Using (4.83), (4.85), and (4.88), we obtain

re eal 3 1
= L,%,(s) (mod ¢,) = [1]
1 1 :
ale Mardi 5
274 Transfer-Function Realizations

and equating coefficients of equal powers of s, this gives

Also,

bin bi, bis r

[Yo (S) vi(s)] = R, (s) (mod ¢,) = (1 2a=-1° ¢=1)
b}21 b}22 bi23

where

nl-1-k | on ght-2-k | Le 54ST

7,(s) = 8 nl-1-k k+1

nl=23; yo(s) =stal=s+2; y¥,(s)=1

Therefore, from above, substituting the corresponding values gives us

bh biz bis]
[s3n2) ly =o 2s si1}
ba bz. bag
which gives

B, =

1 -5 -3

Block 2: $2(s) =s? + 3s + 2= s* + a%s + a2; therefore, n2 = 2 and

0 1
oy =
-2 -3

and let

x bj, bj, bj ~ ch Ch,

B, = and C, =
2 * a
ba bé, bé Ch C3,

Following the same procedure as above:

Irreducible Realizations 275

Gh Pes id. 0
c
(1) = L2%2(s) (mod $2) = [-3s ) =2]
Con Cec hs sf

which gives

OO
C, =]
-2 -3

(2) For this block,

=
n2-1-k +
2) n2-2-k + e058 4
2
Yas Snteick: etd
Therefore, for n2=2, andk=0, ... n2-1=0, 1

vo (8) =stat=s+3; and y,(s) =1

Thus by substituting the above in (4.83), we have

[yo (s) 1 (s)]B, = R;(s) (mod ¢,)

or

bf, bi, by
[s+3 1] = | 2. aly
ba bg be
giving

® @
B, =
Omer Ze el!

This completes the computation since the third column of [p(s)M(s)] is null.
The complete irreducible realization of G(s) in (4.84) is therefore S(A, B,C),
where

I
1

;
A= | eee 4-------
|
!

(4. 89a)
:
|
!
!
276 Transfer-Function Realizations

Od 2a
B, 1) =a <3
B an (Onn +0 fee)
tr = = e 89b

B,
ie ea

6 =[C, 6] = (4. 89c)

Or eiaagiees
[If G(s) were proper, the transmission matrix D would be calculated from
D= lim... Gs)+]

4.4 REALIZATION FROM MATRIX FRACTIONS [8, 6, 10, 11, 19-25]

The realization procedures described in preceding sections are based on a

proper rational matrix G(s). These methods, which are applicable to a
majority of engineering systems where transfer functions (matrices) are
obtainable from input/output data, are not directly applicable to systems
that are described mathematically through a set of differential equations
(assuming zero initial conditions)

D(s)y(s) = N(s)u(s) (4.90a)

where u(s) is the m-input control, y(s) is the £-output variable, and N(s)
and D(s) are polynomial matrices of appropriate dimensions, respectively.
Furthermore, a realization procedure for (4.90a) provides an alternative
method when the available system description is in one of the two forms*

N(s)[D_ (s)]7u(s)
y(s) = (4.90b)
[D {s)]'N (s)als)
where N(s) and D(s) are relatively prime and 6{N(s)} 4 6{D(s)} in each
case. The realization methods from this matrix-fraction description of
multivariable systems are based on some basic relationships that may be
derived from the results of rational matrices outlined in Sec. 1.5; these
are considered in the following section. The subsequent sections use these
results to develop methods of obtaining canonical-form realizations from a
given description of the form in (4.90).

*The dimensions are N,(s): 2X m; D,(s): m X m and Dg(s): £X 23

No(s): 2Xm.
Realization from Matrix Fractions Pare

4.4.1 Matrix-Fraction Description

The derivations here are based on the concept of row/column reduction of

polynomial matrices (Sec. 1.5) and on the assumption that for any linear
system, the polynomial matrix associated with the output variable has full
rank [i-e., p[D(s)] = min (£,m)] and that it can be reduced to a row-proper
form, as in (1.107).

Theorem 4.4 Any m-square polynomial matrix D(s) is equivalent to

the following polynomial matrices:

dri-1
(1) Ty diag ALOE ASO IESmee i L..

1
dei s

(2) I. diag {s as L, diag

gaci-1

where the constant square matrices I and L have the following dimensions:
Ty, Fe: (m X m) and nonsingular; Ly: (2 dri x = dri); and Le(z dej x 2 dej)—
dri and dej being the highest degrees in any ith row (i € 1,.-.,m) and jth
column (j € 1,...,m) of D(s). The structure of the various diagonal matrices
is

dri
s

d dri
diag{s"" } = ae

drm
Ss

tne se

dri lp ame a eee gc

diri!
diag(t.e,7-4.8 11,8, ++:,8 4 END)
278 Transfer-Function Realizations

del
s

. e d .

digeteuunh = gd

dem
s

1 I
|
!
Ss !
1
1
1
1
I
!

del :
ehbes

— 1
ea
1 1
I1
over
diag/s \ = ai = ¥,(8)
ee asl
:
er
1 dej |
Seo (eae |
sie as
11
1
1
|s
es
1

1
i}
1
|
l
| dem
is

Proof: Since any given N(s) may be made to be row or column proper
through elementary row or column operation:

N(s) = N_(s)U_(s) row operation (4.91)

Il U(s)N_(s) column operation (4.92)

Realization from Matrix Fractions 279

where U;(s) and U,(s) are unimodular matrices; proof of the theorem then
follows directly if the matrices L,.(s) in (1.107) and L,(s) in (1.105) are
expressed as

dri-1 =
(1)iDY 1GL(s) ==diag
di {1,s,...,
ene }L, ==$,(s)L,, (4.93)
C

i
where L and L are
" s x 18 €
(2) L,() = L, diag : = L¥,8) constant matrices

ct (4.94)
Since L,(s) and L,(s) consists of lower-order terms (i-e., terms of power
less than dri and dcj, respectively, for any i,j), it is only necessary to
show that they can be broken into forms as in (4.93) and (4.94). Since the
rows or columns of N(s) can be rearranged such that dci> +++ > dcj > +:
> dem, or dri>-°:: > dri> drm, this follows directly from (1.104) and the
structure of ¢,(s) and y;(s)-.
An example of this representation is given below.

Example 4.13: Representation of D(s) as in Theorem 4.4

(s+2)*(st1) s*1
D(s) = (4.95)
0 (s+ 1)?

Ae ad,
8
tS LO 18° 2 OUD Ve ase Bele 340 :
<> (peasnea Bae A Se Uo ee pa a eae 5 00 (4.96)
Os babe Os tn. oat 0000. OMT se puree
—— NY oi Le an
SS ee
a 0 2
iP diag {s} p(s) bia
ig

ea)
1

s_i0
s° 1 0 re) Ams! Mapalele ht a2 ie
Si esac teae --+-- + (4.97)
0 «| gt{lo t1 0M 0 1 2ifo ja
—~ —~ <<
a a s

igo ngCc L Cc :
p (s)
280 Transfer-Function Realizations

Having established a method of "expanding" D(s), the general theorem on

matrix-fraction description may now be stated.

Theorem 4.5 A strictly proper transfer-function matrix Ggp(s) of

order 2X m may be expressed in either of the two forms:

(1) ce =" (i (s)icn Y.(8)P]"'Td,,(8)Q] (4-98)

(2) Gols) = [Qp (SI(s) + Py(s)1™ (4.99)

where

I(s) = diag fh for (1): ~ dimension £2 Xx 2

= diag ee. for (2): — dimension m X m

: g
(8) = diag {1,8,e.+-. ge i — dimension £ X ni
i=1

s m
p(s) = diag ( . — dimension » nj Xm
: a
nj-l

P and Q are constant (independent of s) matrices of dimensions:

£ £ : m m
12R yy) nix 23 Q: >) nix ms iP ims nj; Q: ox) nj
i=1 i=1 j=1 j=l

and ni (nj) is the degree of the denominator polynomials of the elements in

the ith row (column) of Ggp(s)-

Proof: Part (1). Let g;(s) be the monic least common multiple of the
denominators of all the elements in the ith row of Ggp(s)- Then
Realization from Matrix Fractions 281

814!) (8) Si m(5)

8, (s) 8, (s) 8, (s)

G =
g._il (Ss) ;
g..(s)
ij
[Link] (s) ;
sp) ~| 8.6) Bg 88) oe

8 91'S) 8 ;(S) 8 om '5)

& ,(s) ay g ,(s) a. g ,(s)

Since ni is the degree of g;(s), then 6 {g;(s)} = ni > 6{8,;(s)} for any
ici, ..., £2. Let A(s) be the diagonal matrix defined as

A(s) = diag {g,(s)}, Let pie 0 (4.101)

where g;(s) has the form

mi. i ni-1 i
g.(s) =
=s +
+*p,s8 + eee +
Pai (4.102)
°

Then Ggp(s) may be expressed as

G(s) == [A(s)]7!G(s)
-14 (4.103)

where G(s) = {8;;(s)}- Bovianyie 1 se ee lee jl sla, cys LOL 8i(S) have
the form

a ij_ni-1 ij_ni-2 : ij
8,,(8) = a, a qd, 8 ae zip qi (4.104 )

From (4.102) and (4.104), the denominator and the numerators of the ele-
ments of Gep(S) may be expressed as
282 Transfer-Function Realizations

Phi
i
gis) Sse
ni se PlLasiere Ss
niet, ] |Bit (nix1) =s
ni ay
+ Y (S)P
ei (4.105)
— eT

(1 X ni) i
Py
eeA ao ogy, UB aIKe
aly coGcy ial

ij
thi
ij
»
gj) =[Ls8,--.8
ni-n| ]] it ee
(niX 1) = v(s)a” (4.106)
— .

(1 X ni) ij
1
where ve is a row vector and p and q’ are column vectors of dimensions
as indicated. Combining (4.102) and (4.103), we have

A(s) = diag {6} + vy(s)P (4.107)

where
2 ni columns

v(8) = , 2 rows
Realization from Matrix Fractions 283

2 columns

1
Pai

1
0
Entci

il
Pi

i
Phi
i
a Pp. ;
Ns ni-1 > ni rows

i
Py

2g
Pre

g
0
Pro-1

g
Py
igs Rat i
a diag (Pap Paiv tt Pyh Le iE Cis aT g

Similarly, combining (4.104) and (4.106) gives

G = d(s)@ (4.108)

where Y(8) is as given above and Q is given by

284 Transfer-Function Realizations

id Bare
1) oo5 GG
1m
nl 1

11 1j 1m
ay 1 rey

il ij im
Sai os Gi
ij ; ; :
Q={q'}=| : : > ni rows
il ij im
qy creas aie kal

or Qj gm
n Te 7°" Ine

21 Qj fm
G49 eee qa, 1

m columns

Combination of (4.107) and (4.108) gives

G. (8) = LA@IG(s) = F(S) + 9(5)P]1

(8)@] (4.109a)
Part (2) of the theorem may be proved in a similar way by assuming
g.(s) as the monic least common multiple of denominators of all the elements
in the jth column of Ggp(s), jel, ..., m. Then, as in (4.108),

G.(6) = Gis){Ats)17
where

A(s) = diag {g,(s)}

J eriyie
2,Cope Sern (4.109b)
G(s) = {8,,(s)}
Realization from Matrix Fractions 285

Detailed derivation of (4.109b) is left to the reader.

In some cases it may be necessary to perform some permutation

(interchange of rows/columns) on the given transfer function (yielding an
equivalent system) to obtain the exact forms of (4.109a) and (4.109b). This
is illustrated below.

Example 4.14: Matrix-fraction description of G(s)

i eS perlnees
(s+ 1)?(s+ 2)? (s+ 2)?
G(s)= (4.110)
1 1
~ (s* 2)? ~ (8*2)?
(1) Left matrix-fraction description (4.98):

(s+1)2(s+ 2)? re Ale Gal (s+ 1)?

G(s) = [D(s)]“'N
£ 0 (s) = _
0 (s+2)? | |-1 -1

s*+6s3+ 138s?+12s+4 0
D ,(s) =
0 s*+4s+4

s# 0 6s? +13s2+12s+4 0
= +
0 s2 0 4s+4

Therefore,

3? 0
P(e) <=
0 s?

and, consequently,

ml 6 BF ee og ©
yw (s)= (4.111a)
le ie 0-00-10
t -8

giving
286 Transfer-Function Realizations

Ae x)

12 0

[Link] 18526126 bd 0 Dvpnsi>t SS 7aOn Ory |Loree

= BONG (4.111b)
0 4g +4 i Wy OM WW) ah
0 4

0 4
SS’
P

and

i a

i gear Rsiaril 1 s s4 s? 0 0 Wine,

N(s) = = Oat (4.111c)
-1 -1 0 0 OR ROee Ls 0 0

-1 -1

0 @
ed

Q
(2) Right matrix-fraction description (4.99):

ee 0 0 (s+1)*(s+2)
G(s) = N (s)[D_(s)I"? =
-1 s|\(s+2)? st2

0 s? +452+5s+2
DAs) =
s*+4s+4 Sar

s? 0 0 1 0 4s2+5s+2
= +
0 s4ii 1 0 4s+4 s+2
—_—_—

Te
where

I'(s) = diag {s3, s*}T (C., being a permutation matrix)

and
Realization from Matrix Fractions 287

TOO
s 0

wy (s) =| s? 0 (4.112a)
Ont
0 s

giving

ae 0)

0 4s? +5s+2 ots eve egy op poe

4s+4 s+2 i 2 ele 0 A 4 ta fea
ee (DEAR
p 0 s

and

10 70

1” 30 LONG .0) wl oe =
N (8) = et = pe a BC s* 0 (4.112c)
il
SE Setel ee
Q 0 s

Combining (4.112a), (4.112b), and (4.112c), we have

G(s) = [Qy (SITS\T, + Py(srJ7

= [Qy, (s)]K[I(s) + Py, (3)ia

which represent the same system with different numbering of the input/
output terminals. The derivations above and the results in Secs. 4.2.2 and
4.2.3 are used in the following sections to develop observable and control-
lable canonical-form realization methods.

4.4.2 Observable Canonical Form [20-24]

By extending the results in Sec. 3.2, it can be shown that if the output matrix
of a purely dynamic m-input/£-output nth-order system S(A, B, a has full
rank, it can be transformed into an observable canonical form §0 (AG, 1g Co)
288 Transfer-Function Realizations

through a ren transformation of the state vector, where the matrices

A, B, C and As Rie have appropriate dimensions and 8 has the follow-
ing structure:

Bait a oie ek Ani

Aoad Ajo0 ae

A, = || o n rows (4.118a)

ofl A599 : ol
a
te

n columns

with Aoi in an (ni X ni) obServable companion matrix [as in (4.9)]:

ni
i
ee ae estes
a 0 1 Susie 0 i
A...
Oll
= 5
Pri-g ni rows (4.113b)

0 0 eee 1 “Pi

See age Oe

ni columns

and Aoi i# j, is an (ni X nj) matrix given by

0 i) 2.070" 119) -p.

nj
i
0 Ore TO -p.
Tes nj-1 :
5 EE OW (4.1138c)
oij

Se eee
nj columns

The observable companion matrix Aoi has 1's occupying positions below
the diagonal elements and the last (ni)th column equal to some nonzero
column vector. Elements of this column vector constitute the characteristic
Realization from Matrix Fractions 289

polynomial* associated with Aoi: For i# j, Aoi; is an (ni X nj) matrix

which is identically equal to zero excepth for the final (nj)th column, as
shown above, ni, ic 1, ..., £, being the observability indices, obtained
during the transformation procedure. Let ai's be defined as

k
a0 =03; . ak = >; Td ee KenPl 2 move ore
i=1

that is,

AO =103 Foals Fo2)= nen?) os Ok = nL ee ct eke eoe-Pan ase rae’

Q@L=nlt+n2+-++- +nl=n

The output matrix Ge has the following form:

OMe. ONES 1S a.-0 O oO - © 5)

Ome ® @ © OF 1s 0 50
C = 10 “oo 80) £0 Oa DN 10 tee 0 2 (4.113d)
Om: v @ © Q @ Occ ©
O = Y @ © Oo OM @ - 1
SS SS ee See

TMI ak V4 ap OOO as fae a ial

The input matrix By is of order (n +n2+-+-++nf=n) mand assumes

no special form. For notational convenience the matrices above are often
expressed as

|
Oven 0 x 0 0 x
it 0 0 x 0 0 x

hie oiieal Gis Asoij aL ;

ni (4.114)
0 0 0 x 0°: 0 x

0 8 1 XxX 0 0 x
eS — ee

ni nj
290 Transfer-Function Realizations

O20 0 0 PW 0
0 0 -:- 0 0 @) goo i, ooo |)
Cris e Q (4.115)
(0) °

0 0 oe 0 0 0 eee 0 0 eee 1

SS eee Se

where X denotes a nonzero element.

To derive an observable realization procedure, let Ay be the constant
n X £ matrix consisting of the £ ordered ak columns of Ao [marked | in
(4.114)], Cy (= Ig) be the constant £ x £ matrix consisting of the 2 ordered
ak columns of Ce The following theorem (called the structure theorem
[20-24]) may now be stated.

Theorem 4.6 The strictly proper transfer-function matrix of an

m-input/ £-output nth-order (n> 2) completely observable system S(A, B, C)
may be expressed as

C(sI - AJB = G_ sp (s) = ©|[T(s)a - pie(s)A 2 J r y_(s)Bfo)

where I(s) and 7,(s) are as defined in (4.98).

Proof: Since S(A, B,C) is completely observable, there exists a

nonsingular matrix T such that

(yi Cis A) Be= {Olish Palin tap

Hl C [sl = Al Be (4.116)

The theorem may thus be proved by showing that

C [st - Ay? = © IF) - ¥(s)A

IY (8) (4.117)
which is equivalent to (as © g= to)

(ris) AIC, = $ (s)[sI - All (4.118)

By direct multiplication and using Theorem 4.5, the right-hand side of the
equation above becomes
Realization from Matrix Fractions 291

oll 012 olm

»_(s)Ist-
A] = diag {1,s,.. onan
~An ot “Aa 00 Be ra

8, (s) 0 0

le
=) % g,(s) 0 = A(s) (4.119)

0 0 8 ,(s)

The left-hand side of (4.118) may now be expanded as

x Sarees x

: ; x Pal el x
died cee enGlag 1-3) sneee ieblh. }C, (4.120)
Sk apo
Sg

2 nonzero columns of Ay

ae ini-l i ni-2 i
= diag {s™cles diag {-p,s - p,s Sete (4.121)
= diag {8,(s) } ree eee

=9/s(8)

This completes the proof. Since the structure of ae is known a priori,

Theorems 4.5 and 4.6 may be combined to obtain the following observable
realization theorem.
Theorem 4.7 The observable-form realization of the m-input/ £-output
strictly proper system described by the polynomial matrix representation

[diag {s™} - diag {1,s,. gti-1

IPly(s)=(Y(s)QIU(s) 4-122)
axe ech mae ee Be Se es

D(s) N(s)

is 89(Ao, Bos Ce); where the columns of the = ni x £ matrix P form the £
nonzero columns of the > ni-square matrix Agi Bo = Q and On consists of
292 Transfer-Function Realizations

k :
£ rows with 1's in any (ak)th element (ak = Zay Timp, Iu, Ba ooo, L2) hae!

all other elements in Cs are 0. The order of realization is 2 ni, ic l,..., &,
where ni is the degree of the least common multiplier of the elements of the
ith row of the polynomial matrix D(s)

The realization procedure* based on the preceding results is outlined

through the following illustrative example [23].
Example 4.15: Observable-form realization of

s*+3s%1 2s+3 yi (s) 1 0 u, (s)

~ (4.123)
s3+3s*+s 357+3s+6 | ly,(s) st1 s+3 | |u,(s)

Step 1—Transformation of D(s) into row-proper form: Since the

analytical results in Theorems 4.5 and 4.6 are based on the row-proper
form of D(s), this step is concerned with deriving an equivalent form of
(4.123), where the polynomial on the left is row proper. This is achieved
by the postmultiplication of a unimodular matrix U¢(s) such that D,(s) is
row proper:

U ,(s)D(s)y(s) = U,(s)N(s)u(s) —
D,(s) N,(s)

For the system in (4.124),

-s 1
Ue) =
1 0

giving

0 s*+6 i gs
D,(s) = and N,(s) = (4.125)
s*+3st1 2s+3 il 0
Step 2—Transformation of (4.125) into a form such that the elements
with highest degrees are in the principal diagonal positions: This is

*A system with strictly proper G(s) [6{D(s)} > 6{N(s)}] is considered here
to demonstrate the applicability of (4.117). Although Theorem 4.6 may be
extended to proper G(s) [24], such an extension is not essential, since in
any numerical algorithm the strictly proper part of G(s) may be easily iso-
lated from any given form. A method of checking the strict properness of
[D(s)]~*N(s) is given in step 9.
Realization from Matrix Fractions 293

achieved by appropriate column interchange in D,(s), that is, through a

matrix multiplication of the form

KD,(s)y(s) = KN,(s)u(s) (4.126)

—_—’ —_——’

D,(s) No(s)

The constant matrix K for (4.126) is

0 1
kKe=

es 0

giving

s2+6 0 C3} il
DS 7) |Ngis)= (4.127)
28s+3 s?+3st+1 0 1

Step 3—Identification of the elements in D)(s) in (4.126) with highest-

degree polynomials and their degrees: For the given system,

81 (S) I] 37 +6, nl=2

g.(s) =s?+3s+1, n2=2

Step 4—Decomposition of the matrix D,(s):

8,(S) d,, (s)

Des) =
da(S) 82(S)

-6 0

s2 0 1 s 0 0 0 0

0 s? Gene ieee Ea
—\— -2 -3

I(s) (8) 5 p a

Step 5—Decomposition of the matrix Np (s) as

Piha
ice Cao. its sco
Ne ele ver
0 0
1,6) ee
294 Transfer-Function Realizations

Step 6—Computation of ak k = 1, 2 and identification of c ie

Gui == yonth = 74

a2 =nltn2 =4

first second
column element
a1=2

|
" 0 il 0 0 f 1 0
C. = = C, =
0 0 0 1 0 i
t
=o.
a2=4
second fourth
column element

Step 7—Formulation of Bi

1
: 3 0
Poe a
0 0
Step 8—Formulation of A

known from the

structure of A,
’ a ‘ x 0 -6 0 0
il x 10 x al 0 0 0
Behe tg ae oe ee eae
coe a eae Onuic2ineil Pee
= second column

first column oes

Step 9—Computation of the transmission matrix D: Since gj;(s) are

monic, D can be directly computed as
Realization from Matrix Fractions 295

D = lim [diag {s™}] “*N,(s)

soo

1
=s o|fs+3s 1 0 0
oe = 4.129
shew | ok 0 0
S—co 1 ( . )

Step 10—Rearrangement of the columns of OF and Ba! Since the col-

umns of D)(s) and Nj) (s) are interchanged to bring the highest-degree poly-
nomial terms in the principal diagonal positions, rearrangement of the
columns of ce and BS is necessary to retain the original input/output num-
bering, giving

Cie KC | = (4.1380)

(4.131)

The observable canonical-form realization of (4.123) is therefore given by

S(Ap, Bo, Cg, D), where Ay = Ag and Bg, Co, and D as given in (4.128) to
(4.181).

4.4.3 Controllable Canonical Form [20-24]

If the transfer-function matrix representation of an m-input/2-output system

is given through a right inverse:

y(s) = N(s)[D(s)]~'U(s) (4.132)

[as opposed to the left inverse in (4.90b)], a controllable canonical-form

realization can be obtained by using Theorem 4.5. As before, the derivation
here is based on the results in Sec. 4.4.1. r
The controllable canonical form Sa(Ac; ee C,) of a completely con-
trollable purely dynamic system has the following form:
296 Transfer-Function Realizations

Aci Acie Oia

ir Acie Aces Acom

A, =|. n rows (4.133)

A
Set Poond cmm
ee
n columns

tt 9 0 0 O > 0

OF FO: Cady eee 0 O-+- 0

IN = 3 A = x ie ete
cii mJ cij . dee pia
0 O 1 O20 0

— |X x x ° Xx — |X x oS

nj ni

where X denotes a nonzero element. That is, Aci is a (nj X nj) companion
matrix with 1's occupying positions above the diagonal elements, and the
last [i-e., (nj)th] row equal to some nonzero row vector. For i # j, Koi is
an (nj X ni) matrix whose last row is nonzero and all of whose other rows
are null. As earlier, let

k
Be) bj, Ke 12, een (4.135)
jet
The input matrix here has the form
i) coo
O «++ Q

O err 0

a 0 4
oe 2 nj =n (4.136)

1 - 0

Om 0

WU). Dw sal
eta
Realization from Matrix Fractions 297

CG, has no special form in this representation.

Let Ava be the (n X m) matrix consisting of the m ordered 8k columns
of A (marked — above), and Br (= Im) be the (m x m) matrix consisting of
the m ordered Bk rows of Bo: The following theorem may now be stated.

Theorem 4.8 The strictly proper transfer function matrix of an

m-input/£-output nth-order (n > m) completely controllable system S(A, B, C)
may be expressed as

C(sI - A)-!B = G.
(8) = 6 vy(SIF) isA? (SB, (4.137)

where 7%, and I(s) are as defined in Theorem 4.5.

Theorem 4.9 The controllable canonical-form realization of the

m-input/-output strictly proper system is described by the matrix-fraction
representation

y(s) = [Qy(s)I[diag {s")} - Pdiag]- |“u(s) (4.138)

_ + a é

—_— °

N(s)
a

[D(s)]™
in 8BAe Bo cy where the columns of the m Xx - nj matrix P form the m
nonzero Eeturats of the 2 nj-square matrix As; 78), B. consists of the

m rows with 1's in any Bkth element (Bk = fsiMjy Is @ ily boop aM), eitoel gull

other elements of Bo are 0. The order of realization is 2 nj, j ce 1, 2,

-, m, where nj is the degree of the lcm of the elements of the jth column
of the polynomial matrix D(s).

Detailed proofs of the foregoing theorems are not considered here,

but the resulting realization procedure is illustrated below.

Example 4.16: Controllable-form realization of

a0 0 -(¢+48%+5s+2) {7}
N(s)[D(s)]~? = (4.139)
-s s7||s*+4s+4 st+2

Step 1—Transformation of D(s) into column-proper form: This can be

easily achieved by postmultiplication of a unimodular matrix U;,(s) such that
D,(s) is column proper, where
298 Transfer-Function Realizations

N(s)U_,(8) [D(s)U,,()]“* = Ny(s)[D3(s)]*

SN ee =

N,(s) D,(s)

D(s) in (4.139) is column proper, so this step is omitted [D(s) = D,(s)]-

Step 2—Rearrangement of the rows of D,(s) to bring the highest-degree

polynomials into their principal diagonal positions: This is achieved by a
postmultiplying D,(s) by a matrix L given by

Oe
L= [D,(s) = D(s) in this case]
-1 O

which transforms (4.139) into

Su
0 s||s?+4s2+5s+2 0
N,(s)L[D,(s)L]7* = (4.140)
ee soo Ss -(S+2) s*+45+4
N,(s D,(s — NS ee
o(S) o(S) N,(s) D,(s)

Step 3—Identification of the elements in D)(s) with highest-degree

polynomials in its columns:

gis) =.6° + 4s°s 58

1 2, “nt —3
go(s) = s*+48 +4, n2= 2

Step 4—Decomposition of Do(s):

i 2.0

33 0 EO 25al 242 Worego io .

D = =

ae 0 3? 2A 80 eed ; .
—\ —_~

I(s) P 0 s
S_Sm’
Realization from Matrix Fractions 299

Step 5—Decomposition of Nj(s):

1 0
OL kM Ol cde all on ae
N,(s) = Sse 0
Onno te Oe Cee

Q 0 s
SS’

v(Ss)
Step 6—Computation of 8k, k € 1, 2, and formulation of B

BY =e= 6
pg2 =nl+n2
=5
0 0

& ait if
Bors Md vO C103 Sine Bie
Cc m 0

ORO first third

62 east 1 row row

second fifth
column row

Step 7—Formulation of Cos

1 0 0 | — **
!
i]

Obs
On gO Omer OUhistae
A Sala EE I x x | ——first row ofP
A = if
i] (4.141a)
CGC 3 | ---------- SaaS
C= 00 708 Lola
i

a Ce | x x |+——<second row of P

*krows known from the structure of A,

300 Transfer-Function Realizations

BE Se

Oo Oo Ff OO fC
=|-2 -5 -4 0 0
0 OO @ —© al

2 1 0-4 +

Step 9—Computation of the transmission matrix D:

D = lim N,(s)[diag {s"J}]*

sS—oo

0 s ||s3 0 0
eee ~ (4.141b)
s—oo 32 -Ss 1 0 0
UE ar}
s

Step 10—Rearrangement of the rows of Ga and Bo? Since the columns

of D(s) and N(s) were interchanged through L, the original input/output
numbering is manifested by reverse transformation; that is,

0
0
B =BL=|0 -1 (4.141¢)
0
1

Clery = |. (4.141d)

The controllable canonical-form realization of (4.123) is therefore given by

Se(Ac,
Be, Ce,D) when Ac = Ag and Be, Ce, and D as given above.
Once a realization S(A, B, C) is obtained by using any of the preceding
methods, its minimality may be checked by using either the degree or the
Markov parameters of G(s). In the former method, as indicated earlier,
the order of minimal realization is

VON
min
5{G(s)}
2 614,(s)}
Realization from Matrix Fractions 301

p(s) being the invariant polynomials in the Smith-McMillan form of G(s).

This relationship is especially suitable for the methods in Sec. 4.3, where
the ~;(S) are available.
In the Markov parameter method, the order of minimal realization is

De = p(M)
p(M

where M is the infinite (block) Hankel matrix [a special form denoted by H

in (3.140)].

CB CAB CA2B
|CAB CA2B CA3B
SS 1 ayCave ewevista Maral Kas moe

The criterion above follows from controllability and observability proper-

ties, since minimality of S(A, B,C) implies complete state controllability
and complete observability. This method is particularly suitable if numerical
algorithms are used in the realization procedure.
The problems of translating one form of system representation into
another were first identified through the fundamental results that the input/
output description reveals only the controllable and observable modes and
that this part of the system defines the smallest state-space representation
among other representations with the same input/output relationships. These
results and associated realization methods (and algorithms) initiated the
creation of extensive literature on the subject. Work on this fundamental
problem of system theory highlighted the need for using the "right" relation-
ship between the input/output and state-space models for a particular prob-
lem. As a result, it is now well established that a transfer function matrix
can be used to study stability properties of a time-invariant multivariable
system only if the representation of the system being used in the analysis
is minimal. The effects of the "hidden" poles and zeros on system stability
are analyzed in the following chapter through the properties of decoupling
zeros introduced in earlier chapters. Another example of the importance of
realization theory in the context of feedback is the linear quadratic control
problem. Stability of the closed-loop optimized system can be assured only
if optimization is carried out by using a completely controllable and observ-
able model. These are considered in the following chapter.
302 Transfer-Function Realizations

REFERENCES

1. Mayne, D. Q., "Computational procedure for the minimal realization

of transfer-function matrices," Proc. IEE, 115: 1363-1368 (1968).
2. Kalman, R. E., 'Mathematical description of linear dynamical sys-
tems,'' SIAM J. Control, 1: 152-192 (1963).
3. Chen, C. T., Introduction to Linear System Theory, Holt, Rinehart
and Winston, New York, 1970.
4. Panda, S. P., and Chen, C. T., "Irreducible Jordan form realization
of a transfer-function matrix,'' Trans. IEEE, AC-14: 66-69 (1969).
5. Kuo, Y. L., "On the irreducible Jordan form realization and the
degree of a rational matrix,"’ Trans. IEEE, CT-17: 322-332 (1970).
6. Kailath, T., Linear Systems, Prentice-Hall, Englewood Cliffs, N.J.,
1980.
7. Fortmann, T. E., and Hitz, K. L., An Introduction to Linear Control
Systems, Marcel Dekker, New York, 1977.
8. Blackman, P. F., Introduction to State-Variables Analysis, Macmillan,
London, 1977.
9. Melsa, J. L., and Jones, S. K., Computer Programs for Computa-
tional Assistance in the Study of Linear Control Theory, McGraw-Hill,
New York, 1970.
10. Rosenbrock, H. H., State-Space and Multivariable Theory, Thomas
Nelson, London, 1970.
11. Rosenbrock, H. H., ''Computation of minimal representations of a
rational transfer-function matrix,"' Proc. IEE, 115: 325-327 (1968).
12. Munro, N., 'Minimal realization of transfer-function matrices using
the system matrix," Proc. IEE, 118: 1298-1301 (1971).
13. Gueguen, C. J., and Toumire, E., "Comments of 'Irreducible Jordan
form realization of a rational matrix','' Trans. IEEE, AC-14: 783-784
(1969).
14. Mayne, D. Q., 'An elementary derivation of Rosenbrock's minimal
realization algorithm,'' Trans. IEEE, AC-18: 306-307 (1973).
15. Kalman, R. E., 'Irreducible realization and the degree of a rational
matrix,'' SIAM J. Control, 18: 520-544 (1965).
16. Gilbert, E. G., "Controllability-observability in multivariable control
systems,'' SIAM J. Control, 2: 128-151 (1963).
17. Chen, C. T., "Representations of linear time-invariant composite
systems," Trans. IEEE, AC-13: 277-283 (1968).
18. Heymann, M., and Thorpe, J. A., ''Transfer equivalence of linear
dynamical systems,'' SIAM J. Control, 8: 19-40 (1970).
19. Rosenbrock, H. H., "On linear system theory,'' Proc. IEE, 114:
1353-1359 (1967).
20. Wolovich, W. A., and Falb, P. L., "On the structure of multivariable
systems," SIAM J. Control, 7: 437-451 (1969).
21. MacDuffee, C. C., Theory of Matrices, Chelsea, New York, 1959.
References 303

22. Wolovich, W. A., "On the numerators and zeros of rational transfer
matrices," Trans. IEEE, AC-18: 544-546 (1978).
23. Wolovich, W. A., "The determination of state-space representation of
linear multivariable systems,'' Automatica, 9: 97-106 (1973).
24. Wolovich, W. A., Linear Multivariable Systems, Springer-Verlag,
New York, 1974.
25. Silverman, L. M., "Realization of linear dynamical systems," Trans.
IEEE, AC-16: 554-567 (1971).
 oa
»<
<i ’
a aua .
ij Hh 7
~~

?
-_ 7
: — ye ‘4 7 “ =a “ oceania! 11 oreaeniay! nye |
7 — -
s ta ,
> © ' . iL’ i
- r at ay PEL ET? 1

= ide Fag rie tor

j i So a] i erm oe
= _ % e c.g ax rd

é
He ely
{ j me Te? ¢ Ret. S
® ’

v ‘ =|
- ry % b, 7 . = co]
itt '
j a) A Sim an |
ul ) be ° irs “mi! “A ’ of iy 6 @ip toi wal ~

A s iP Ls M
ne : ia? Wicreli exclauy ooruedliae
‘ > OTS (i nee
} 5 i ai f ares Liisa

hy Lid. Cape i626 liom

Lie amt vi ner’ "hyo ni
on ‘
i ‘ ' u (
: yt’: f :

, ‘a LP ig en

Pe, | —
ae

re