Automatic,

Vol. 12, pp. 497506.

Pergamon Press, 1976. Printed in Great Britain

Minimal Stable Partial Realization*?

G. LEDWICH$ and J. B. MOORE$

An asymptotically stable minimal order realization of a partial sequence of Markov parameters is achieved by reducing the problem to a standard but minimal one in decision algebra.

SummaryIn this paper two equivalent sets of necessary and sufficient conditions for the existence of an asymptotically stable partial realization are developed. Both sets are expressed as multivariable polynomial equations which may be tested for the existence of a solution in a finite number of rational steps via decision methods. Should a solution exist, it may be evaluated with the aid of polynomial factorization. The first set of conditions are based on results due to Ho and Kalman, and are useful for the case where the number of specified Markov parameters is greater than the order of the realization. For other cases, the second set of conditions which include results from a companion paper on minimal observers, require less computational effort to be tested. 1. INTRODUCTION

tice only the first 2n~.x parameters need be specified where n~.. is an upper bound on the order of such a realization. Let us consider, the linear time-invariant system with statespace equations
x(t) =Ax(t)+ Bu(t), y(t) =Cx(t)+Du(t)

(1.1) with u an r-vector input, y an m -vector output and x an n-vector state. Its transfer function is D + C(S1 A)-l?
=D+crBs-l +CABs-2+ CrA2Bs-3+ ...

SINCE THIS paper is concerned with asymptotically stable partial realizations of time invariant systems using decision methods, the relevant aspects of both these topics are now reviewed.
1.1 Asymptotically stable partial realizations

An important modelling or realization problem is to determine a state-space model of a given system from a particular input-output description of that system. Gilbert [l] first achieved minimal order realizations of a specified transfer function matrix, but his algorithm requires that the transfer function matrix elements have distinct poles. Kalman [2] developed an algorithm to reduce the state-space of any non-minimal realization to a minimal one. Ho and Kalman [3] use the Markov parameters Yj forj=l,2, . . . of the proper rational transfer function matrix G(s) to produce a minimal statespace realization. Other minimal realization algorithms exist which may be more computational efficient [4-7]. Note that in prac*Received 16 April 1975; revised 6 October 1975; revised 5 March 1976. The original version of this paper was presented Congress at the 6th IFAC which was held in Boston/Cambridge, MA. during August 1975. The published Proceedings of this IFAC Meeting may be ordered from: ISAInstrument Society of America, 400 Stanwix Street, Pittsburgh, PA 15222, U.S.A. This paper was recommended for publication in revised form by associate editor J. Ackermann. tWork supported by The Australian Research Grants Committee. $Electrical Engineering Department, University of Newcastle, New South Wales, 2308, Australia.

Recalling that the Markov transfer function G(s) are YO+Y,S-+Y2S-2+ . . . it system (1.1) is a realization
yO=D, Yj=CAi-B

parameters Yj of a defined from G(s) = is now clear that the of G(s) if and only if for j=l,2,. .. (1.2)

When only the first M Markov parameters are available and it is required to determine [C, A, B ] such that Yj = CAi-B for j = 1,2, . . . M, then we have what is known as a partial realization problem. An algorithm to determine a minimal partial realization, in general not unique, is given in Tether [8] or Kalman [9] (note also [10]). Unfortunately this algorithm will not guarantee asymptotic stability of the realization even when it is known that the system giving rise to the Markov parameters is asymptotically stable. Efforts to include a stability constrant on this form of canonical realization are described in [16]. The algorithm presented can be implemented for relatively simple systems but would require further research effort to give a precise procedure to implement the various tests for higher order multivariable systems. In this paper decidability theory and polynomial factorization are applied to achieve minimal asymptotically stable partial realizationsadmittedly at the expense of computational effort. 497

498

G. LEDWICHand J. B. MOORE stable partial realization are presented. In developing these conditions, the relevant results of realization theory and partial realization theory are reviewed. 2.1 Case Y,, Y,, . . . specified In the first instance we review realization results for the case where an infinite set of m x r Markov parameters Y,, YZ,. . . are specified see Ho and Kalman [3]. First Ho Kalman result. The transfer function G(s) (possibly non minimal) with specified Markov parameters Yj for j = 1,2, . . . has a denominator Sq+flq_lsq- +... + ~l~l~~~?and only ~;s) = Y#O+
Yj+l~, + . . . + Yj+q_,p .-l + Yj+q = o

1.2 Decision methods The decision methods of interest to us here determine whether or not a vector u exists such that ~(u) = Oand g(u)> O where ~() and g() are vectors with each element a polynomial in any given element of u. Such a decision can be made in a finite number of rational steps [11]. Unfortunately the number of steps increases exponentially with the number of unknowns, the elements of u, and the number of inequalities. Reference [12] provides a method for determining a solution v which satisfies the above equalities and inequalitiesgiven of course knowledge that such a u exists. This solution method involves polynomial factorization and thus in theory involves an infinite number of steps. Perhaps the key contribution of this paper consists of the efficient use of results due to Ho and Kalman, to formulate the stable partial realization problem in a form suitable for solution by decision methods [11, 12]. Of the two solutions presented, the approach using the Ho-Kalman results, is most efficient for the case when the number of specified Markov parameters A4 satisfies n ts M s 2n T where n T is the order of the minimal asymptotically stable partial realization. When M > 2n T the Ho-Kalman algorithm may be applied directly. But when M < n t some of the ideas of a companion paper [13] for designing special classes of minimal order observers are employed. (Actually it was the observation by T. E. Fortmann that certain observer problems could be expressed as stable partial realization problems that initiated our investigations of this topic.) There is intentionally some duplication of [13] in this paper for the case M < n t since for the reader whose chief interest is the stable partial realization problem, much of [13] would be obscure. Section 2 reviews some known realization results and extends these to derive a set of necessary and sufficient conditions for an asymptotically stable partial realization. A series of Hankel matrix properties are developed in Section 3 including a set of lower bounds for the order of a minimal realization. In Section 4, these properties are used to simplify the decision problem presented in Section 2 and a procedure is presented which may be used to evaluate the minimal order asymptotically stable partial realization. In Section 5 an alternative method to that described in Sections 2-4 is presented and is shown to be more efficient for the case M < n t.
2. REALIZATION THEORY

for

j=l,2,.

..

(2.1)

Before summarizing a second result from [3] it is useful to define the Hankel matrices Ha. and ~.. in terms of the specified Markov parameters Yjf0rj=l,2,. ..aS Y,
Y,

Y,
Y,

Yj Yj+,

Hij=

o .

. . Y,+l . 0
Yi+j-, 1

[~ Y,
Y,
Hij =

(2.2) Y,
Y.

. .

. .

Y,m Y,+,

1
Y,+,

Y,+* .

Y,+,

Second Ho -Kalman result. For a transfer function with Markov parameters Yj for j = 1,2,. . . and a denominator polynomial (3(s) of degree q, an irreducible realization is of order

n = rank H..

(2.3)

Recall that the rank condition rank H.q s p for some p < min {mq, rq } is equivalent to the determinant conditions
lH~+,l = O

for all t

(2.4)

where H~+, denotes the tth (P + 1)x (P + 1) submatrix of Hqa. With knowledge of the order n of the realization, a system realization may be determined using the HoKalman algorithm [3] or [4-7]. 2.2 Case Y,, Y,, . . . Y~ specijied When only the first M Markov

In this section necessary and sufficient conditions for the existence of an asymptotically

parameters

Minimal stable partial realization

YM are specified, we have what is Y,, Y2,... known as a partial realization problem. An algorithm described by Tether [8] and Kalman [9] achieves a realization for this case, but there is no guarantee that the realization is asymptotically stable. In this section we achieve asymptotically stable partial realizations using the HoKalman results as a starting point. It now proves convenient to introduce the following point of notation. Point of notation. An n th order realization having a transfer function with the first Al Markov parameters Y,, Y,, . . . Y~ and a q th order denominator polynomial ~(s) will be denoted as an [n, q, M] realization. An asymptotically stable [n, q, M] realization will be denoted an [n, q, M, As] realization. Now since an asymptotically stable partial realization is required, it is necessary that the characteristic polynomial /3(s) have left half plane zeros. Equivalently, there must exist parameters u,, w, and some positive constant q such that

499

result the transfer function denominator polynomial is ~ ( V~, s ) and since /3( Va, s ) is defined as in (2.5) the realization is asymptotically stable. For q = M equation (2.6) involves no restriction on ~(s). An arbitrary /3(s) maybe chosen to have strictly negative zero and 11~~ will be fully defined using (2.7), for any such choice. Using the second HoKalman result above it is clear that the order of an asymptotically stable partial realization must in this case satisfy n < min {Mm, Mr). The key lemma of this section, now stated, is an immediate consequence of the above claims and the second HoKalman result. Lemma 2.1. A [n, q, M, As] realization exists for some n s p if and only if the following condition (C 1) is satisfied. a Condition exists (C 1). There such that (2.6) is V., YM+I, YM+2, . . . Yd-, satisfied and, with Y4, Ya+,, . . . Y2, -, chosen satisfying (2.7), rank llq, s p (or equivalently for p < min {mq, rq} the (p + 1)x (p + 1) submatrices H~+, of Ha, satisfy (2.4)). (2.5) for
q

/wq>s)=YJi(vq)s
II [S2+(U? +6) S+(W?+ 6)] {Z*

q/2

even for
q

. I (s + W/+ E) qfl[S2+(U?+E)S is, +(w?+ E)] odd

where V, (of dimension q ) is used to denote the vector consisting of all vi and wi. Now we are in a position to make the following claim. A necessary and sufficient condition for a [., q, &f, As ] realization is that there exists a parameter vector V* such that with /3,( V, ) defined in (2.5) ~dvq)~ @l(vq)yl+l+. for . .+~.-l(vq)yj+q-l
+

Yq+j=o (2.6)

j=l,2,.

..(q)q)

Necessity follows by application of the first HoKalman result in the previous subsection. Sufficiency is seen by construction of such a . . reahzatlon. The parameters YM+,, YM+Z, . . Ya_, . (where Q = max {q + 1, M + 1}) are chosen arbitrarily while Ya, Yq+l, . . . maybe calculated from Yi=[po(vq +
Bq-1(vq)yi-.l

Remarks. (1) The condition (C 1) is in the form of multivariable polynomial equalities and thus the decision methods of [11, 12] may be applied. For this decision problem there are q + nn x (max{O,(q M)}) unknowns. In the next sections we show how both the number and complexity of the equalities (2.4) can be reduced significantly so that the computational effort in testing condition (C 1) can be reduced to manageable proportions. (2) For the case of scalar input realizations or scalar output realizations, Yi are either row or column vectors and thus m = 1 or r = 1, inevitably n = rank Hq. s q. This means that in the statement of Lemma 2.1, p = q and the inequality (rank H,. < n ) in (C 1) is automatically satisfied.
3. PROPERTIES OF H..

)Yj.q+/31( v.) Yi-q+l +... (2.7)

fori=~, ~+ l.... Having so defined an infinite series of Markov parameters, a realization may be constructed as in [3]. By the first HoKalman

In this section we explore some properties of Hankel matrices defined in (2.2) where the Markov parameters Y,, Y2, . . . satisfy (2. 1) for some q. In particular, we explore the properties of Ha, since these will be useful for simplifying the condition (C 1) o; the previous section. Property (P 1) [5]. (i) Dependence of the i th

500

G. LEDWICHand J. B. MOORE

row of Ha. on rows above the ith row of H.q implies dependence of the (i + km )th row on rows above the (i + km )th row in H~a for k=o, l,... (ii) Dependence on the ith column of H.. on columns to the left of the ith column of Hqa implies dependence of the (i + Irr)th column on columns to the left of the (i + kr)th column in H.. fork =0, l,.... The next property of Hankel matrices is given in terms of a special class of submatrices of Hqa denoted H(w, A) obtained by an ordered deletion of rows and columns of H.. defined from the p= integer indices A = [A,A*. . .A, ], ] [/.klP2. . . I.Lm as follows. Definition: H(p, A ) is obtained from H,. by deleting the rows numbered [i + (~i + k)m ] and the columns numbered [j + (Aj + k)r] for i = l,2,. ..m, j=l,2,. ..randk =0,1,. . .. Property (P2). Given that there exists an n such that rank H.a z n then there exist vectors A, p satisfying (3.1)

where Oif the jth row of H,,,~+,-,, is dependent on rows above it. { 1 otherwise Oif the jth columns of H~+l-.,,., is dependent on columns to the left of it. { 1 otherwise (3.4) The above choice of rows (or columns) is the same as the choice of independent rows (or columns) in the algorithm of [8]. The indices are the observability (controllability) p *(A*) Kronecker indices [14] of the minimal partial realization, and as a consequence we have the following property adapted from [8]. Property (P3). With Hqq defined in (2.2) specified and where Y,, Y,, . . . Y~ is Y M+l, Yhf+2, . . . YZM unspecified, and p * and A* defined above, a lower bound on the rank Hqq is given from rank
Hqq > n *

R, =

(3.3)

C, =

(3.5)

such that rank H(M, A) = n. Equivalently, given that rank H(p, A) # n for some n and all p, A satisfying (3. 1), then rank Hqq < n. Proof. The definition of rank tells us that the inequality rapk Hqq a n implies that there exists a full rank n x n matrix obtained from Hqq by deleting rows and columns according to the following rule. The ith row of Ha. is deleted if it is contained in the range space of the rows above it and the ith column of H.. is deleted if it is contained in the range space of the column to the left of it. Application of property (P 1) now ensures us that our full rank n x n matrix is in fact H(LL, A ) for some p and A satisfying (3.4) as desired. We now move on to a consideration of an important property of the Hankel matrix H~~ defined from (2.2) with Y,, Yz, . . . Y~ specified and Y~+,, Y~+2, . . . Y,~ unspecified. To determine a lower bound for the rank of such matrices Y&f+,, selection of the irrespective of Yhf+2,. . . Y,~, attention is focused on certain fully specified submatrices of H~~ namely the Hankel matrices H,,,~+l.,, and H~+I-~,,., where q is the smallest integer a j /m cl is the smallest integer z j/r Integer indices p.* = [p ?I.L ..1A ~] and A* = %. [A~A~... A?1 are defined as follows
ql p?=k~oRi+knt,A!=~ k -O al Ci+kr

where n*=~A?=~p?=
i-l i-l

rank

H(IA*, A*)

(3.6)

Some usefuf bounds. (1) The order of the minimal partial realization n * is clearly a lower bound for the order of the minimal stable partial realization n ~. (2) Let the Hankel matrix Hab be the smallest submatrix of Jhi which contains all rows for which Rj = 1 (i.e. independent rows) and all columns for which Cj = 1 (i.e. independent columns), also let q*= max (a, b ). With the first Ho-Kalman result in mind, upon reflection we see that q * is a lower bound for q in any [, q, M] realization and is also a lower bound for the q t of the minimal stable partial realization denoted [n t, q t, A4,As ].
4. MINIMAL ORDER ASYMPTOTICALLY PARTIAL REALIZATION STABLE

To determine a minimal order asymptotically denoted an partial realization stable [n t,q t,M, As ] realization, the value of n t, the minimal order itself, is first calculated being the following efficient procedure. 4.1 Steps to calculate nt 1. Test for the existence of a Step [, q, M, As] realization, for q = q*, q*+ 1,... by applying decision methods to determine whether or not for q = q*, q* + 1, . . . a parameter vector V. exists such that with pi ( V. ) defined in (2.5) polynomial equaIity condition (2.6) is

(3.2)

Minimal stable partial realization satisfied. The lowest integer q for which such a realization exists is denoted q~n. Note that q * is readily calculated as described in the previous sections. Step 2. Apply decision methods to the condition (C 1) for /3( V., s ) defined as in (2.5) to test for the existence of an [n, n, A4,As ] realization for n=ii, ii+ l,... where ii = max (qti., n *), until the conditions are satisfied. The lowest integer n is denoted n t. Note that n * is readily calculated as described in the previous section. There is a considerable saving in the above procedure for calculating n t in testing for an for n = ii, fi + 1,... [n, n, A4,As ] realization rather than in testing for an [n, q, M, As] q= realization for n=ii, fi+ l,... and The justification stems from the qmin,qmi.+1,. . .. following lemma. Lemma 4.1. The non-existence of an [n, n, M, As ] realization for some n implies the non-existence of an [n, q, M, As] realization for all q. The above lemma is established by application of the following two properties of realization: (a) The existence of an [n, q, M, As ] realization of G(s) implies the existence of an [n, q +<, M, As] realization of G(s)L (s)/L (s) = G(s) for any polynomial L(s) of degree /, where /=1 ,2,.... (b) The existence of an [n, q, M, As ] realization of G(s) for q > n implies the existence of an [n, q, M, As ] realization for q = n, since there must be at least (q n ) polezero cancellations possible in G(s). Remarks. (1) The algorithm must terminate since n Ts M x min (r, m). (2) For the case of scalar input or scalar output realizations, n t = qfin. To see this note that with step 1 implemented, lemma 2.1 tells us that there exists an [n, q~., M, As] realization for some n s qti.. But the property (b) above is violated unless n = q~in. Thus n t = qtin and step 2 above is not required. (3) For the multiple-input, multiple-output realization cases (m >1, r > 1), by setting qti = q * in step 2, step 1 can be eliminated at a possibly small increase in computational effort. It is shown in [17] that there exist cases where qti > n* so either term may be a more useful lower bound. (4) Having determined n, n t and q t may be evaluated by testing the conditions (Cl) for the existence of an [n T, q, M, As] . . reahzatlon for q = qfi., q tim+l,. . ., q t which is the lowest values of q for which a realization exists. Alternatively one may test for q = nt, nt 1,... until a realization fails to exist, depending on whether q is expected closer to n T than qfi. (5) To test the condition (C 1) there are of the order of (n !)2 determinant equalities.

501

Fortunately (C 1) may be considerably simplified, both in the number and complexity of the determinant tests as we now go to show. The simplifications we describe are crucial to achieve a practical algorithm but unfortunately the manipulations are intricate. 4.2 Simplification of condition (C 1) (This subsection may be omitted as a first reading of the paper). The properties (P 1)-(P3) of Hankel matrices are exploited to achieve a simplification of the condition (C 1) for the cases m >1, r >1. In particular the conditions (2.4) are simplified. For convenience we adapt the following convention. Convention. The columns and rows of a submatrix of H,~, possibly with its rows and columns interchanged, are assigned the numbers of the corresponding rows and columns of E&. Four steps are now described to achieve a simplification of (2.4) given that & is defined from (2.2) where the Markov parameters Y,, Y,, . . . satisfy (2. 1) for some q, and Y,, Y*, . . . Y~ are specified. Using the Convention above, the simplified equalities are readily stated although one has to be careful with bookkeeping details in practice. The four steps are now described along with a brief justification of each step. Step 1, description. Examine in turn the first m (M q + 1) rows of lZQQ linear dependence for on the rows above the row in question. If the i th row is dependent on the rows 1,2, . . . i 1, then delete this row. Likewise if the ith column is dependent on the columns 1,2,. . . i 1, then delete this column. Note that property (P 1) is used to accomplish this step. Our depleted matrix or submatrix is denoted R. Assertions (i) Rank A = rank 24... (ii) Property (P 1) holds with H.. replaced by ~ and the rows [columns] of E denoted according to the convention above. Justification. The above assertions are justified since only dependent rows [columns] of H.. are deleted, and so whether or not a row [column] of A is independent or not can be determined without knowledge of these dependent rows [columns]. Step 2, description. Calculate p * and A* via (3.5)-3.7) and thereby obtain H(p *, A *) a submatrix of H~~. (H(p *, A*) is also submatrix of H.. and ~) . Rearrange rows and columns of ~ to form

where both E4 and H(IA *, A *) are submatrices of H.

502
Remark.

G. LEDWICHand J. B. MOORE step 2 a full rank submatrix of H(P, *, A*) containing only specified elements. The simplification is achieved at the expense of not producing a least order ~, in step 3. Step 4, description. Form the submatrices ~:= where E,(/Z., X. ) of A. = [(l.L,-PY)(PzP!).. .(;: -p%)] and [(A, Af)(A, A%). . . (A, A?)] with (3.1) hold ing. (Here n = ~ u, = X A,.)
i-l i-l

The first two assertions above are immediate. The third assertion follows from property (P 1) holding for R and the properties of H(I.L*, A*). Thus in constructing [H,, ~,] from H, if the ith row is deleted, then the (i km )th row is also deleted for k = 1,2, . . .. As a consequence, property (P 1) part (i) holds with H., replaced by [~,, ~,] as desired. In a similar manner the remainder of assertion (iii) can be established. Form the matrix Step 3: Description.
Justification.

H, lth epaced by [1 H, ~4 [1
174

This operation merely groups the rows and columns of I-&. Assertions. (i) Rank H(I.L*, A*) = n*. (ii) Rank ~ = rank A = rank H... (iii) Property (P 1) part (i) holds with H.q replaced by [~,, ~,] and the rows of [~q, ~.J denoted according to the convention above. Also, property (P 1) (ii) holds independent and the columns of

denoted according to the convention above.

Assertions. (i) Given that there exists an n such that rank HI = n n*, then there exist vectors IL

and ~. as defined above with (3.1) holding such that rank H,(P., ~.) = (n n*). (ii) The conditions (2.4) are equivalent to the conditions lE@.+,, 1.+,)1= Ofor all possible F.+,, x.+, (4.1)

H,= Hg

H3H-(/L*,

A*)~2

Assertions. (i) Rank H.. = n*+ rank ~,. (ii) Property (P 1) holds with H~a replaced by ~, and the rows and columns of ~, designated by the numbers given to the corresponding rows of Ed when these are assigned according to the convention above. Justification. We have the following H(/L*, A*) [
H,

Justification. The arguments used above to demonstrate that property (P2) follows from (P 1) for Ha. also demonstrate that the above assertion (i) follows from assertions (i) and (ii) of step 3. Assertion (ii) follows from assertion (i). As a consequence of assertion (ii) above, we can restate condition (C 1) as Condition (C2). There exists a V,, Y~+,, YM+2, . . Yq_, such that (2.6) is satisfied and with . Yq, Ya+,, . . . Y,.., chosen to satisfy (2.7), the equations (4. 1) are satisfied. Remarks (1). Observe that the submatrices H~+, of (2.4) are of dimension (n + 1)x (n + 1)

27,
H,

I J
[o

column

H(p*, A*) g3

O H,

1 1
-1

row

operations

I
column

row operatiom

H(p*,

A*)

H, H, 1

operations

z=
[

o
H,

~o

Assertion (i) above is now immediate. From the above diagram note that [~J R,] is obtained from column operations on [~j ~~]. As noted earlier, property (P 1) part (i) holds for [~, Ed] and since the property is not affected by column operations, including deletions, we have that it also holds for [~, ~,] and ~,. Likewise it may be H, shown that (P 1)part (ii) holds for - and thus for H, [1 R,. We have thus shown that (P 1) parts (i) and (ii) hold for ~,. Remark. H(I.L *, A*) may contain elements which are polynomial functions of the unknown parameters making the expression of the determinant and adjoint of H(I.L *, A*) more difficult. The difficulty may be avoided by separating in

whereas the submatrices ~,(1.I .+,, ~.+,) of (4.1) are of dimension (n + 1 n*) x (n + 1 n*). This that a determinant means equality @,(& .+,, ~n+,)l = O is considerably simpler than the equality lH~+,l = O for even moderate n*. Moreover, the number of equalities in (4.1) is considerably less. The number of equalities in (2.4) is

24=(~Tl)(.~1)
where
A=A! B!(A ()B B)!

Minimal stable partial realization The number of equalities in (4.1) is no greater than
N.., = Example. M = 2, r = m = 2

503

nn*+rn-l ~n*

nn*+?l )( ~n*

where m is the number of w ~ z A4 q + 1 and f is the number of A~ = A4 q + 1. For the case n=6, n*=4, rn=?=3, q= 3, Nz,d=7392 and Nd., = 36-a spectacular reduction. (2) If If(p *, A*) contains some elements not fully specified, then, the determinant test (4. 1) will not equalities in the unknown be polynomial parameters. This problem is avoided by looking instead at determinants of submatrices of det [H(p *, A*)]~,. (3) For any A. and jL. for which a realization exists, this realization may be constructed having An(I-L. as the controllability ) (observability) indices of that realization. Of course at least one such set of indices for which a realization exists are Kronecker indices so these can be identified from the set of An or P. for which a realization exists and a canonical realization as in [15] constructed using these.

Assuming for the moment that c in (2.5) is zero, Y~ is defined in terms of Y, and YZ for q=2as Y,= giving 11 H., =
[ w*v* o Vz
W2V2

o 1
11

o
W2 V2

1 1 V2

01
1

W2V2 1

HQ. L*, A*)= o

[1

_~ ,n*=2

Now step 1 of the simplification of (C 1) is not relevant, step 2 is trivial as ~ = I&, and step 3 gives

I
W2V2

H, =
1

_ [

o o
stable

WW-[: -!l-[:1 :
w

W2+V2+1

V2 W2+U21

4.3 A minimal tial realization

order asymptotically

par-

With knowledge that an [n f, q, M, As ] realization exists for a known n T and some q T < q < n T obtained using the techniques of the previous subsection, such realization may be calculated as follows. Step 1. The polynomial equalities of (C2) are solved by polynomial factorization for the elements of Vq, YM+I, YM+2, . . . Y.. Step 2. The parameters Y,+,, Y4+z,. . . Y,. are calculated via (2.7). Thus Y,, Y2, . . . Y2~ and thereby 1-1.. and & are determined. Step 3. The HoKalman algorithm [3] or one of the simpler algorithms [47] is applied to H,. and & to yield a realization of order n T with the desired properties that it is minimal and asymptotically stable. Of course, in the event that scalar input or scalar output realizations are involved the full power of the multivariable algorithms are not required. For this case realizations in the observable or controllable canonical form are readily achieved once the /3, are calculated from V.. The steps of the algorithm and to simplification are now demonstrated on a simple example.

In this example it is clear that there are no v and w such that rank HI = Oso clearly n 7> n * = 2. Checking for rank ~, = 1 we examine 1~,1 which is zero for V2= 1 + W2, since (2.6) is trivially satisfied we may choose the solution W2= 1, V2= 2. This gives an extension

and

Y,=ol,

[1

53

thus 1 011 0101 HM~ = * 132 o 101

Using the Ackerman-Bucy procedure [4] a realization of third order may be calculated which turns out to be asymptotically stable. In particular,

504

G. LEDWICHand J. B. MOORE

For this example the minimal realization is of order 2 while the minimal stable realization is of order 3.
5. ALTERNATIVE APPROACH

Y, Y,. . . Y~ since CB = CDB, CAB = CD(AD)B , . . . CA~-B = CD(AD)~-B. Now the pair [AD, D C] is not completely observable since its observability matrix can be expressed as the product of two matrices one of which is the singular matrix D. Denoting the the system component of observable [AD, B, D C] as [&& ~~], it is clear that parameters has Markov [AD, BD, CD] Y,, Y,, . . . Y~ and is lower order than the realization [A, B, C]. It remains to be shown that [&, ED, ~~], for some choice of D, is an asymptotically stable realization, to complete the proof. In fact we show that {A,[iiD 1}for some D is a subset of {Ai[A ]}. Without loss of generality let A be upper triangular and let the n ith row of R be dependent on the rows below it. For this case

In this section we formulate our realization problem as a decision problem which may be solved by decision methods using a specialization of the results in [13]. The alternative approach includes fewer unknowns, in some cases when n t > M, than the methods of earlier sections. For the case Y,, YZ,. . . Y~ specified and Y M+,, Y M+, . . . unspecified, equation (1.2) re, duces to [Y, Y*... or v, 7 [C
Y~]=CR, R=[BAB. .. AB]B]

D=[;-i-

;]

=[;

i] (5.2)

For a fully specified sequence of Markov parameters, Kalman showed that any state realization is irreducible if and only if the system is completely controllable and observable. We now state a corresponding result for asymptotical y stable partial realizations. Lemma 5.1. A necessary condition for the system (1. 1) with system matrix [A, B, C], an asymptotically stable partial realization of the Markov parameters Y, Y,. . . Y~, to be of minimum order is that R and Q defined above be full rank matrices. Proof. Necessary conditions for minamality are that [A, B ] is completely controllable and [A, C] completely observable. For the case M < n the rank conditions on R and Q are equivalent to the controllability and observability conditions and the lemma is immediate. For the case M > n, we proceed by showing that if the lemma is not true we are led to an contradiction. In particular we show that if [A, 1?, C] is an asymptotically stable partial realization with R not of full rank then a lower order asymptotically stable partial realization can be constructed. For R not of full rank, we have for some singular D, that R = DR. Now this relationship [A, B, C] and system that the implies [AD, B, DC] have identical Markov parameters

where A, and AZ are upper triangular matrix of and x order (n i 1) and i, respectively denotes a submatrix of appropriate dimension. Manipulations yield the following forms for (AD) and & since it is the n - ith state of the system which is clearly unobservable.

since A 1 and Az are upper diagonal and ( submatrices of A this shows that {Al iiD)} is a subset of {A,(A )}. The above rank condition allows us to reduce the number of unknown parameters in (5.1) by considering the equivalent conditions. C = [Y, Y, . . . Y~]R(RR)
[Y, Y,... (5.3) = O (5.4) (5.5)

Y~][det (RR)Z R(RR)AR] det (RR)>

O

or B= [Y{... YLIQIQQl- (5.6)

[Y; . . . YL][det(QQ)l

Q(QQ)AQI =0 (5.7)

Minimal stable partial realization det(QQ)>O (5.8)

505

where XA denotes the adjoins of the matrix X. To determine the existence of an n th order asymptotically stable partial realization decision methods can be applied to the polynomial expressions (5.4) and (5,5) or (5.7) and (5.8) with the stability constraint on A. This test only involves the unknowns of A and B (or A and C) rather than all parameters of A, B and C as in (5.1). This decision problem is a specialisation of one studied in [13]. In [13] the above decision problem is greatly simplified by use of [A, C] in a special form with a set of structural indices m
r

order of the realization. For the special case m = 1 or r = 1, single output or input systems, it is only necessary to test one simple set of equalities involving n unknowns. If r >1 and m >1 the decision problem is inevitably more tedious to implement.

6. CONCLUSIONS

where Z n, = n. In [15] several structural forms in, are available with a reduced parameter set but the lower block triangular form for A described in [4] appears most desirable here. For this form the diagonal blocks of A are in the single output observable canonical form, thus the stability, constraint on A is satisfied by choosing the parameters describing the characteristic polynomial in (2.5). The number of unknowns in the decision problem is now ~ (r+ 1 i)n,. An i=l upper limit on the number of such unknowns is m. Similarly if [A, B ] is chosen in the special form the number of parameters would be less than mr. Thus an upper bound for the number of unknowns which are involved using these structural forms is min (m, mn ). Since the complexity of the decision problem depends largely on the numbers of unknowns, and the number of unknowns for the methods of the previous section totals (n M)rm for n > A4, the methods of the section should be used for the case (n M)rm 2 min (m, mn ). Since for multivariate systems it is possible for an n th order realization to exist for one set of structural indices ni and not for another set, it should be noted that the decision problem must be repeated for all possible choices of ni. This implication of the property for realizations is treated in [18]. We will (n+:-l) [A, C] form and have

In the paper we have pointed out that the decision methods of [12] can be applied to yield a solution to the problem of minimal order asymptotically stable partial realizations. In exploring the application of decision methods to this problem, the objective is to minimize by analytic means, the number of unknowns in the underlying polynomial equalities and so minimize the complexity of the decision problem and thereby achieve more efficient solutions than could otherwise be obtained as in [16]. Proceeding towards the objective has been fruitful, both in defining a relevant simplified decision problem and in giving insight into the nature of the minimal order asymptotically stable partial realization problem itself. Unfortunately the complexity of the completions for both procedures given in the paper for the design of such realization increases exponentially with the order of the realization. A curse of dimensionalthe work as in Dynamics ity overshadows Programming for control, Viterbi Algorithms for communications and shortest route methods for operations research. For the special case r = 1, m = 1, single input or output systems, there is no difficulty as it is only necessary to test one simple set of equation involving n unknowns. For r >1 and m >1 the decision problem is inevitably more tedious to implement.
REFERENCES [1] E. G.

GILBERT: Controllabilityand observability in multivariablecontrol systems. SIAM J. Control 2(1)

128-151 (1%3). [2] R. E. KALMAN: Mathematical description of linear dynamic systems. SIAMJ. Control 1(2) 152-192 (1%3). [3] B. L. Ho and R. E. KALMAN:Effective construction of linear state variable models from input output functions.
Proc. 3rd Ann. Allerton (1%5). [4] Conf. Circuit and System Theory

decision problems for the n+iW-1 for the [A, B] form: where n )

r)
Y=

x!
y!(xy)!

J. E. ACKERMANN R. S. BUCY: Canonical minimal and realization of a matrix of impulse response sequences. Irrf. Control 19, 224-231 (1971). [5] C. BONIVE~O, R. GUIOORZI nd G. MARRO: Irreducible a canonical realization from external data sequences. ht.
J. Control 17, 553-563 (1973). [6] J. RISSANEN:Recursive identification of linear system. SIAM J. Control 9, 420430 (1971).
[7]

Unfortunately the complexity of the computations for both procedures given in the paper for the design of minimal order stable partial realizations increases exponentially with the

[8]

J. RISSANEN: asis of invariants and canonical forms for B 10, 175-182 (1974). lineardynamicsystems.Automatic A. J. TSTHER: Construction of minimal linear state variable models from finite input-output data. IEEE AC
AC-15, (4)427436 (1970).

[9] R. E. KALMANand N. DE CLARIS: Aspects

oj Network

AUTO. ol.12,No.S-H V

506
and System Theory,

G. LEDWICH and J. B. MOORE

p. 385. Reinhartand Winston,New

realization

algorithm

for
AC-19,

mutual

sequences.

IEEE

York (1970). [10] S. S. GODBOLE: omments on construction of minimal C linear state variable models from finite input-output data, IEEE Trans. Aut. Control AC-17, 173-175 (1972). [11] A. SEIDENBERG: new decision method for elementary A algegra. Ann. Math. 60, 365-374 (1954). N [12] B. D. ANDERSON, . K. BOSE and E. S. JURY: Output feedback stabilization and related problemssolution via decision methods. IEEE Trans Aut. Control. AC-20,
53-66 (1975). M [13] J. B. MOORSand G. LEDWICH, inimal order observers for estimating linear functions of a state vector. IEEE Trans. Aut. Control AC-20, 623-632 (1975). M. A [14] B. W. DICKENSON, MORPand T. KAILATH: minimal

Trans. Aut. Control

31-38 (1974).

[15] D. G. LUENBERGER: anonical forms for multivariable C systems. IEEE Trans. Aut. Control AC-12, 290-293
(1%7). [16] J. R. ROMAN and T. E. BULLOCK: Minimal partial realizations in a canonical form. IEEE Trans. Aut. Control AC-20, 529-533 (1975). N [17] J. R. ROMAN: ew problems in partial realization theory. Thirteenth Allerton Conf. on Circuits and System Theory Proceedings. Monticello (1974).

[18] G. LEDWICH and T. E. FORTMANN: Comments on partial realizations, IEEE Trans. Aut. Control AC-19, 625-626
( 1974).