SIAM J.

CONTROL
Vol. 12, No. 3, August 1974
Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

REALIZATION AND STRUCTURE THEORY OF BILINEAR

DYNAMICAL SYSTEMS*
PAOLO D’ALESSANDRO, ALBERTO ISIDORIf AND ANTONIO RUBERTI
Abstract. Starting from the description of the system provided by the series expansion of the zero-
state response, this paper develops the realization theory for bilinear systems. It is shown that the
condition of realizability corresponds to that of the factorizability of the kernels of this expansion.
The paper then analyzes the properties of the minimal and nonminimal factorizations and provides
a solution of the problem of characterizing the minimal realizations. Subsequently, developing the
structure analysis of the state space, it is shown that there exists a canonical form of the equations and
the results of the realization theory are interpreted on this basis. Lastly, the bases are laid for developing
realization procedures, and it is shown that the sequence of kernels of a bilinear system is completely
identified by a finite number of these kernels.

1. Introduction. In this paper we present a complete realization theory of

bilinear systems, a class of nonlinear systems which has raised considerable
interest in the last few years (see, for example, [8]).
The system model considered in this theory is the sequence of kernels
characterizing the Volterra series expansion of the zero-state response. It is
shown first that a necessary and sufficient condition for the existence of bilinear
realizations can be expressed as a sort of factorizability property ( 3). Then we
present quite a complete analysis of the sequences that satisfy this condition
(4), in order to derive the tools for studying minimal bilinear realizations. A
complete characterization of these realizations reveals the interest in considering,
besides the dimension of the state space, also the least number of multipliers needed
to implement the nonlinear map. The analysis shows that this number assumes
its minimum just over the minimal realizations. These, moreover, are a single
equivalence class ( 5).
As for linear systems, the minimal realization theory can be connected with
the structure analysis of the state space. This latter is developed here on the basis
of reachability (from the origin) and unobservability. As a result, we obtain a
canonical form of the equations and we show that the kernels of the Volterra
series expansion of the zero-state response depend only on one of the subsystems
identified in the decomposition. This subsystem provides, modulo an equivalence,
a minimal bilinear realization ( 6). Finally, we present an outline of the pro-
cedures for constructing minimal bilinear realizations ( 7).
For the sake of brevity and notational simplicity, the analysis is presented
here for single-input, single-output systems; the extension to the multidimensional
case can be found in [4]. Some results of the present paper (statement (c) of
Theorem 4 and Theorem 8) have been proved, simultaneously and independently,
by R. W. Brockett [1].

* Received by the editors June 8, 1972, and in revised form May 21, 1973. This work was supported
in part by the National Council of Research of Italy.

" Istituto di Automatica, University of Rome, 00184 Rome, Italy.

:I: Istituto di Automatica, University of Rome, 00184 Rome, Italy, and Fondazione U. Bordoni,
Rome, Italy.
517
 518 PAOLO D’ALESSANDRO, ALBERTO ISIDORI AND ANTONIO RUBERTI

2. Bilinear equations and their solutions. The input-state-output equations of

Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

a single-input single-output bilinear system are usually written in the form

(t)-- Ax(t) + Nx(t)u(t) + Bu(t),
(1)
y(t) Cx(
where u(t) R is the input, y(t) R is the output, and x(t) R" is the state at time t.
The matrices A, N, B, C are constant matrices of suitable dimensions.
If the system described by equations (1) is in the initial state Xo at time 0,
the response y(t) can be decomposed into the sum
(2) y(t) y,(t) + y,(t) + y,,(t),
in which the first two terms respectively represent the zero-input response and the
zero-state response.
The expression for yx(t) can be obtained immediately by putting u(t) 0 into
equations (1), i.e.,
(3) Yx(t) C ea’xo
If the input belongs to the class of functions bounded on finite intervals [2], the
response y,(t) can be expanded in a Volterra series with symmetrical kernels
according to the expression

(4) y,(t)
i= . wi(tl,

The symmetrical kernels of series (4) can be calculated from

ti)
k=,
u(t k) dtl dti.

(5) Wi(tl’ ti)-- 2 Vi(tl’

per
t,),
where the summation is carried out over all the i! permutations of the variables
l, ...,
ti. The functions vi(t, ...,
ti) are given by
(6a) Vl(tl) C eAt’B,
(6b) Vi(tl, ti) C eAtiN e A(ti- l-t,) N e A(t’-t2)
2
B IiI
k= 0
6_ l(tk + tk + 2 )] > 1,

where 6_ l(t) denotes the unit step function.

On the basis of the same assumptions, the response y,(t) can likewise be

(7) Yx,(t)
i=
.
expanded in a Volterra series according to the expression

zi(tl, N)ea{t-t*)xo

Here the expressions for the symmetrical kernels z(t, -.., t0, 1, .-.,
can be determined from the expressions (5) and (6) of the kernels w(t,...,
k=l
u(t tk) dt, dt i.

by replacing the matrix B by an n n identity matrix.

 BILINEAR DYNAMICAL SYSTEMS 519

3. Realizability conditions. Starting from expression (4) of the zero-state

Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

response of system (1), it is natural to assume the following definition of realiza-

bility.
DEFINITION 1. A sequence {wi(tl, ...,
ti)) of symmetrical kernels of a
Volterra series expansion is realizable by means of a constant bilinear dynamical
system with finite-dimensional state space (briefly: is bilinearly realizable) if there
exist four constant matrices A, N, B, C, respectively n n, n n, n 1, n,
such that the following relations are satisfied:

(8a) C ea’lB Wl(tl) for all tl,

Z C eAt’N e
per
A(t’-’-t’)
N eA(t’-t2)B H (- i(tk+
0
tlc+
(Sb)
wi(tl,... ti) for all tk, k 1,..., i, for all > 1.
The first result concerns a condition of realizability, and is expressed by the
following theorem.
THEOREM 1. A necessary and sufficient condition for a sequence {wi(t 1, "’", ti)}
of symmetrical kernels of a Volterra series expansion to be bilinearly realizable is:
(a) that wl(tl) have a proper rational Laplace transform;
(b) that there exist three matrices F(t), G(t), H(t), respectively m m, m 1,
m, of functions with proper ratibnal Laplace transforms, such that the following
relations are satisfied:
(9a) Wi(tl, "’", ti)-- H(ti)F(ti_l ti)"" F(t2 t3)G(t /-2)
on

(9b) Si {(t l, ti):tl > t2 > > ti} for all > 1.
Remark 1. The above condition also holds if the variables 1,..., ti in (9)
are permutated in any way.
Proof. Necessity. If the sequence {wi(tl,..., ti)} is realizable, then, by
hypothesis, there exist four matrices A, N, B, C such that equations (8) are satisfied.
The first of these implies the condition (a). Equation (8b), considered over the
sets Si defined above, reduces to
(10) C eAtiN e A(ti- -t,) N ea(t’-t2)B wi(tl, ti) for all > 1.
If the matrix N is now factorized in the form
(ll) U U’U"
where N’ is n m and N" is m n, it is immediately seen that (10) can be trans-
formed into the form (9), with the substitutions
(12) C eAS H(t), S" eAS
F(t), S" eAB G(t).
Since all these functions have proper rational Laplace transforms, it follows that
condition (b) is also satisfied.
 520 PAOLO D’ALESSANDRO, ALBERTO ISIDORI AND ANTONIO RUBERTI

Sufficiency. Suppose that (a) and (b) are both. satisfied, and consider the
Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

matrix

l/Wl(t) H(t)’
(13) L(t)
G(t) F(t)
Since all the elements of L(t) have proper rational Laplace transforms, this may
be interpreted as the weighting pattern of a constant linear system of finite order
with m + outputs and m + inputs. Consequently, there must exist three
matrices A, R, S, respectively n x n, n x (m + 1), (m + 1) x n, such that
(14) S eatR L(t).
By partitioning S and R in the form

(15) S R (R1 R2),

$2
where $1 is a x n matrix and R1 an n x matrix, one obtains
Wl(t) S1 eAtR1, H(t) $1 eAtR2,
(16)
G(t) S 2 eAtR1, F(t) S 2 eatR2
Now substitute (16) in (9) and let
(17) B R1, C S1, N R2S 2.
It follows that the bilinear system characterized by the matrix A and the matrices
N, B, C defined in (17) verifies the equations in (8) over the sets Si.
On the other hand, since wi(tl, ..., t) is symmetrical by definition, equations
(8) are satisfied for all values of the variables l, ..., ti. This completes the proof.
4. Factorizable sequences of kernels. The result established in Theorem
naturally leads to the consideration of those subsequences {wi(tl,..., t)} which
can be expressed in the form (9). In the sequel such sequences will be called
factorizable and the triplet {F(t), G(t), H(t)} a factorization.
In order to analyze the properties of these sequences, it is convenient to
introduce the following matrices (note that F(t) is m x m, G(t) is m x and
H(t) is x m):
(18) P[.F, G](tl,..., t)= [G(tl) F(t2)G(tl)... F(t)F(t_ 1)"" G(tl)],

H(tl)

(19) Qk[F,H](tl,
H(t
):F(t2)
H(t 1)"" F(tk- x)F(tk)/
).
We shall be concerned with the problem of testing whether the m rows of
Pk[F, G] (the m columns of Qk[F, HI) are linearly independent functions or not;
since F(t), G(t) and H(t), by hypothesis, are functions with proper rational Laplace
transform (see Theorem 1), it will be sufficient to test the linear independence
 BILINEAR DYNAMICAL SYSTEMS 521

over some subinterval of R k, e.g., over Ak {(t l, "’, tk)’O < ti <= 1, for all i}.
Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

This leads to introducing the Gramian matrices

(20) @[F, G] ; PP dt.., dt,

(21) k[F, HI | QQk dtl dtk.
The first basic results concerning matrices (20) and (18) are expressed by the
following lemmas.
LEMMA 1. There exists an integer k’ such that

(22) [] N[+ i] for all k < k’,

(23) N[] N[ + i] for all k >= k’,
and, furthermore,
(24) k’ __< m.

(25)
and that
(26) N[’k-,]
[ik]
-
Proof Relations (22) and (23) can be proved by first showing that
[ik + 1] for all k

[Nk] for some k N[’k] NiCk+ 1]

-
whose validity is a direct consequence of definitions (18) and (20). Lastly, (24)
follows from (22), (23) and
(27) [k] R" for all k.
LEMMA 2. /fonly rfi < m rows of Pm[F, G] (t 1,’", t,,) are linearly independent
over A,,, then there exists a constant nonsingular m x m matrix T such that

(28) TF(t)T_I= (F1;(t) F12(t)

F22(t)]
TG(t)=
Gl(t)
0
where the matrices F1 l(t) and G l(t) are respectively rfi rfi and rfi and
(t l,’", t) has its rfi rows linearly independent over A,.
Proof If only rfi rows of P, IF, G] are linearly independent over Am, there
exists a constant nonsingular m m matrix T such that
](tl, tm)’
(29) TPm[F, G] (t l, tin)
0
with the rfi rows of/(tl, .-., tin) linearly independent over
Bearing in mind definition (18), this proves the second equation of (28) and,
after the partition

(30) rF(t)r-l= (F11(t) F12(t)/

F2 l(t) V22(t)]’
 522 PAOLO D’ALESSANDRO, ALBERTO ISIDORI AND ANTONIO RUBERTI

also gives
Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

(31) P(tl, tm) Pm[Fll, G1](tl, ’’’, tm),

(32) F21(t)Pm_l[Fl G1](tl, tin_l) 0.
It follows from (31) that m[Fll, G1] has rankrfi; therefore, on the basis of
Lemma 1, ,[FI, G x] also has rank rh, or, what is the same, the rfi rows of
P,[F11, G1] (t 1,..., t,) are linearly independent. From this and from (32) it
follows that Fz1(t) 0 and this completes the proof.
Similar results can be proved for matrices (21) and (19).
LEMMA 3. There exists an integer k" such that
(33) [k] [k+ 1] for all k < k",
(34) ["k] ["k+ II for all k >= k",
and, furthermore,
(35) k" =< m.
LEMMA 4. If only fi < m columns of Qm[F, HI (t l, tin) are linearly inde-
pendent over Am, then there exists a constant nonsingular m x m matrix T such that

Flu(t) 0
(36) TF(t)T -1 H(t)T-’ (Hi(t) 0),
F21(t F22(t)
where the matrices F11(t) and Hl(t are respectively rfi x rfi and rfi, and
Qa,[F11, H1](t, ..., t,) has its rfi columns linearly independent over As,.
On the basis of these preliminary results, it is possible to examine the proper-
ties of the factorizable sequences of kernels and, particularly, to describe the set
of all factorizations. In this analysis the integer m will be called the dimension of
a factorization and, thus, a factorization will be minimal when its dimension
assumes the smallest value over the set of all factorizations of a given sequence.
The properties of the minimal factorizations are expressed by the following
theorem.
THEOREM 2. An m-dimensional factorization {F(t), G(t), H(t)} of a factorizable
sequence of kernels is minimal if and only if the rows of P[F, G] and the columns of
Qm[F, HI are linearly independent over A. The minimal factorizations are a single
equivalence class modulo the relation
Fl(t) TF2(t)T -1
(37) {Fl(t), Gl(t), H(t)} {F2(t), G2(t), H2(t)} Gl(t) TG2(t)
U,(t) H2(t)T- ’,
where T is a constant nonsingular m x m matrix.
Proof The proof will be divided into two parts, the first concerning the
minimality and the second the equivalence.
Part 1. Necessity. Suppose, for example, that the matrix Pm[F, G], correspond-
ing to a minimal factorization, has only rfi < m linearly independent rows. As a
consequence of Lemma 2, there will then exist a constant nonsingular m m
 BILINEAR DYNAMICAL SYSTEMS 523

matrix T such that

Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

(38) TF(t)T_ (Fx(t)

Considering also the partition
(39) H(t)T-’ (H,(t) H2(t)),
it is immediately seen that the triplet {F (t), G(t), H(t)}, which has a dimension
rfi < m, is also a factorization of the sequence {wi(tx,..., ti)} and this contra-
dicts the hypothesis of minimality. A similar proof, based on Lemma 4, can be
given for the matrix Q,,[F, HI.
Part 1. Sufficiency. Suppose that the factorization {F(t), G(t), H(t)}, for which

.
the matrices Pm[F, G] and Qm[F,H] have, respectively, their rows and their
columns linearly independent, is not a minimal one and that, consequently, there
exists a factorization {/(t), (t), (t)} of dimension rfi < m for the same sequence
{w(t, ti)} Considering now the matrices Pm[, ] and Qm[, ], it is readily
seen that (see (9))

Q[F, HI(t,,..., t)Pm[F, G](tm+ ,,..., t2m

(40) Qm[,/-](t,, tm)Pm[ ](tm+,,’’" t2m).

Premultiplying both sides of (40) by Q*[F,H](tx,..., tin), postmultiplying

by P*m[F, G](t,..., tin) and then integrating, one obtains

’m m I Q*m[F, H]Qm[ff ffI] dt, dtm

(41)
[. Pm[, 8]P*m[F, G] dtm+, dt2m.
The rank of the matrix on the L.H.S. is equal to m by hypothesis, while the
rank of the matrix on the R.H.S. does not exceed ; it follows that
(42) m __< rfi.
This contradicts the hypothesis that n5 < m and completes the proof of the first
part.
Part 2. If the triplet {F(t), G(t), H(t)} is a minimal factorization of the given
sequence of kernels, it can be verified immediately that the triplet { TF(t)T-1,
TG(t),H(t)T-} is also a minimal factorization (see (9)). Vice versa, let {F(t),
Gx(t),H(t)} and {Fz(t), Gz(t),Hz(t)} be any two minimal factorizations of the
given sequence of kernels. Considering the matrices Qm[F,H], Pm[F, G x],
Q,,[F2, H2] and P,,[F2, G2], one finds that
Qm[F,H,](tl, tm)" Pm[F1, G,](tm+l, t2m)
(43)
Qm[F2, Hz](t, tin)" Pro[F2, G2](tm+ ,..., t2m).
By virtue of the results proved in the first part, the m columns of Qm[F, Hx] and
Qra[F2, H2] and the m rows of Pm[F1, G] and Pm[F2, G2] are linearly independent
 524 PAOLO D’ALESSANDRO, ALBERTO ISIDORI AND ANTONIO RUBERTI

over A Consequently, generalizing the methods and results established by

Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

D. C. Youla in [9] with regard to functions defined in R 1, one finds that

Qm[FI, H,] QmEF2, H]T- ’,
(44)
Pm[FI, G,] TPm[F, G2],
where T is a constant nonsingular matrix. Bearing in mind the expressions (18)
and (19), it follows that
(45) Hi(t) H(t)T- ’,
(46) Gl(t) TG2(t).
As concerns the relation between Fl(t) and Fz(t), one can first of all write
Qm[F1, H,](t, tin). F,(tm+ ,). Pm[F1, Gl](tm+ 2,’", t2m+ 1)
(47)
Q,,[F2, H2](t,, , tin)" F2(tm+ 1)" P,,[F2, G2](t,,+ 2, "’", t2m+ 1).
Substituting the expressions (44) in the L.H.S. of (47), and bearing in mind the
linear independence of the columns of Qm[F2, H2] and of the rows of Pm[F2, G2],
one finally obtains
(48) Fl(t) TF2(t)T -1,
thus concluding the proof of the second part.
In addition to the result expressed by the preceding theorem, it is also conve-
nient to find a relation between any factorization and a minimal one. This is
expressed by the following theorem.
THEOREM 3. Any factorization of a factorizable sequence of kernels can be
written in the form
0 F13(t 0

(49)
F(t)
TIz!(t)
Go(t)
F22(t
0
0
F23(t
F33(t F(t
F43(t) F44(t)/

G(t) T H(t)=(Ho(t 0 H3(t 0)T-’,

where T is a constant nonsingular matrix, and the triplet {Fo(t), Go(t), Ho(t)} is a
minimal factorization of the same sequence.
Proof It is a direct consequence of Lemmas 2 and 4 and of Theorem 2.
We observe that this result also provides a rule for computing too.
COROLLARY 1. The dimension mo of a minimal factorization of a factorizable
sequence of kernels is given by
(50) mo {,,[F,H]m[F, G]},
rank
where {F(t), G(t), H(t)} is any given factorization of the same sequence.
 BILINEAR DYNAMICAL SYSTEMS 525

Before concluding this section it is important to stress that all the results
Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

proved here are valid even if the hypothesis that F(t), G(t) and H(t) are functions
with proper rational Laplace transform is relaxed. In fact, the only requirement
was the possibility of testing the linear independence of the rows of Pk[F, G!
(columns of Qk IF, HI) only on the interval A; for this, for example, it is sufficient
to require analyticity on F(t), G(t) and H(t).
5. Minimal realizations. We are now in a position to prove a basic result
concerning minimal bilinear realizations. Consistently with the case of linear
systems, we shall assume the following.
DEFINITIOrq 2. A bilinear realization {A, N, B, C} is minimal if the dimension
of the state space assumes the minimum value over the set of all realizations of a
given sequence of kernels.
Denoting by 6{L(t)} the order of the weighting pattern L(t), we then have the
following theorem.
THEOREM 4. Let {wi(tl,’’’ ti)} be a sequence ofbilinearly realizable kernels,
and let {Fo(t),Go(t),Ho(t)} be a minimal factorization of the subsequence
{wi(tl, ti)}. The minimal bilinear realizations are such that
(a) the state space dimension is given by

(51) no 6([(\ wl(t)

60(0 Ho(t)’}.
(b) the matrix N has minimum rank, given by
(52) ro dim {Fo(t), Go(t), Ho(t)} mo;
(c) they are a single equivalence class modulo the relation

Ax TA2 T-
N TNz T-1
(53) {AI,NI,B, C} {Az,N2,B2, C2} "
B TB 2
C1 C2 T- I,
where T is a constant nonsingular matrix.
Proof Let {F(t), G(t), H(t)} denote any factorization of {w,(t,..., t,)} and
consider the matrix
w,(t)
(54) L(t)
G(t) F(t)
From the proof of sufficiency in Theorem 1, it follows that every linear realiza-
tion of the matrix L(t) identifies a bilinear realization of {wi(t,..., ti)); con-
versely, from any bilinear realization it is possible to construct a factorization
and, consequently, a matrix such as (54). One can therefore conclude that all
bilinear realizations can be obtained by taking all linear realizations of all matrices
L(t) associated with the given sequence, and then applying (14), (15) and (17). It
follows that the dimension of the minimal bilinear realizations is equal to
(55) no min 6{L(t)}.
L(t)
 526 PAOLO D’ALESSANDRO, ALBERTO ISIDORI AND ANTONIO RUBERTI

If now the forms (49) are substituted in (54) and the matrix [10 0T] is factored on
Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

the left, and its inverse on the right, it is seen that

(56) {L(t)} > {Lo(t)} for all L(t),

where Lo(t) denotes the matrix

(57) Lo(t) [Wl(t) Ho(t)]

Go(t) Fo(t) ]"
This, since Lo(t) is an element of the set of matrices L(t), proves statement (a).
In order to prove statement (b), let {A, N, B, C} denote a minimal bilinear
realization and let
(58) N’N"= N
be a factorization of N such that N’ and N" have full rank r rank (N). Clearly,
the triplet
(59) F(t) N"eatN ’, G(t)= N"eAtB, H(t)= CeAtN
>=
is an r-dimensional factorization. From this it follows that r mo because, by
assumption, the dimension of.any factorization cannot be lower than too. We shall
also prove that the inequality r > mo yields a contradiction. In this case, in fact,
the triplet (59) would be a nonminimal factorization and therefore, according to
the results achieved in the preceding section, reducible. Assume, for instance, that
(59) is not minimal because only r < r rows of P[N"eatN ’, N"e’tB] are linearly
independent; then it is possible to find a nonsingular r x r matrix T that reduces
(59) to the r-dimensional triplet (see (38) and (39) in the proof of Theorem 2)
(60) Fll(t eA’i, G,(t)= eAtB, H(t)= CeA’i,
where and ] (respectively r x n and n r x) are defined by

(61) TN"= ,
Furthermore, this matrix T is such that
N’T-’=(Nq ).

(62) ge’B O, ge’gq O.

The triplet (60) is still a factorization and, therefore, the quadruplet

’’
{A, ’, B, C} is still a bilinear realization, minimal because its dimension is
unchanged. From this it follows that the triplet A,(B Nq), is a minimal

linear realization of

(63) el(t)
1(0 1
(since, otherwise, the quadruplet {A,q,B, C} would not be a minimal bi-
linear realization). As a consequence, the n rows of the matrix ea(B q) are
 BILINEAR DYNAMICAL SYSTEMS 527

linearly independent and this, applied to (62), implies N~"2 0. But this is in contra-
Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

diction with the assumption that N" has full rank r > mo and thus we conclude
that r mo.
As concerns statement (c), it is immediately verified that two quadruplets of
matrices equivalent according to (53) yield the same sequence {wi(tl,..’, ti)}.
Consider now any two minimal bilinear realizations of the same sequence
{AI,N1,Bx, C} and {Az,Nz,B2, C2}. It has been proved (statement (b))that
(64) rank {N } rank {N2} rno
and it is therefore always possible to construct from these realizations, according
to (12), two too-dimensional (i.e., minimal)factorizations {Fl(t), Gl(t),H(t)} and
{Fz(t), Gz(t), H2(t)}. These factorizations are equivalent by Theorem 2; therefore,
constructing for both factorizations the matrices (54), we have

(65) (l(t) H(t)]=(; 0M)Wl(t) Hz(t)(;

(t) F(t)
where M is a nonsingular mo mo matrix.
Gz(t) Fz(t) M
0

-
On the other hand, bearing in mind (59), the above expression (65) becomes
C1 eAlt(B,
(66) N’)= eA’(Bz N’2)
N’ N
In (66) there appear two minimal linear realizations of the same matrix.
These must therefore be related by

(67) (B1 Ni)

’ M

T(B2 N’2)
0
0 M -1
A TA2 T-l,
where T is a nonsingular n n matrix. It follows from this that the two arbitrarily
chosen minimal realizations are equivalent modulo the relation (53) and this
completes the proof of the theorem.
The relevance of the result proved in part (b) of this theorem will become clear
from the inspection of the simulation diagram of Fig. 1. Here the rank of N
assumes the meaning of the least number of multipliers needed for the simulation
by means of this diagram (note also that there does not exist any other diagram
providing a simulation with a lower number of multipliers). This consideration
suggests the interest in characterizing the simulation of a bilinear realization
through the number of integrators and of two input multipliers; from this point
of view it is nice to verify that, rather unexpectedly, the minimization of the first
one implies that of the second.

6. Structure analysis. The preceding sections have been concerned with the
problem of realizing the input-output description of bilinear systems. In analogy
 52,8 PAOLO D’ALESSANDRO, ALBERTO ISIDORI AND ANTONIO RUBERTI
Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

FIG.

with the theory of linear systems, and for the same reasons, it is now interesting
to develop a structure analysis from the state space description. As a result, we
shall show that it is possible to effect a decomposition of the state space and to
prove the existence of a canonical form of the equations, thereby laying the founda-
tions for a complete analysis of the correlation existing between the minimal
realizations and one of the subsystems resulting from the decomposition of the
state space.
This analysis will be based on suitable properties of input-state and, respec-
tively, state-output interaction. As concerns the former it is convenient to assume
the following.
DEFINITION 3. A state x of system (1) is reachable from the origin if there exists
an admissible input function that maps the origin of the state space into the state
x in a finite interval of time.
Due to the nonlinearity of system (1), the subset of all states reachable from
the origin is not a linear subspace. However, a state-space decomposition is still
possible if this subset is embedded into a suitable (i.e., the least one) linear sub-
space. For the purpose of evaluating this subspace, we introduce the sequence
of matrices
(68a) P1 B,
(68b) Pi (APi_I NP_ 1), 2, 3,...,
and we then have the following theorem.
THEOREM 5. The subset of all states of system (1) reachable from the origin
spans a subspace fv of R" which"
(a) is the least subspace invariant under A and N and containing [B];
(b) can be expressed as
(69)
Proof The proof of statement (a) is an extension of that used in the case of
linear systems. It is readily observed that, for any given subspace of R",
(70) x(t) for all [0, T] t(t) for all [0, T].
 BILINEAR DYNAMICAL SYSTEMS 529

By hypothesis, at least a basis {x l, ..., x,.} of 59p is reachable from the origin.
Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

Therefore, for (70),

(71) (A+Nuxj+Bu)659p for allu6R and for allj.
But, again for (70),
(72) Bu 59p for all u R
and, hence, from this and (71),
(73) (A + Nu)59, 59p for all u e R.

(74)
_
From (72) and (73) it follows that 59p contains [B] and is invariant under A and
N. Moreover, observing that any trajectory starting from the origin belongs to a
subspace invariant under A and N and containing [B], it is easy to prove, by

-
contradiction, that 59p is the least subspace with these properties.
As concerns statement (b), we easily verify that

and that
[PI Pi- ,] IEP1. P,] R"

[PI"’" Pk-1] [PI"’" Pk] for some k

(75)
= [PI"’" PAl [Pl"’" PA+,],

(76b)
and, then,
(76c)
59
__ -_
from which it follows that [P1

A59
P,] is invariant under A and N and contains
[B]. To prove that this is the least subspace with these properties, observe that
any subspace 59 of this type must satisfy
(76a)
A[P1], 59

59 [P1 P2.].
N59 NtEP1]

By iterating this construction it follows that any subspace invariant under A and
N and containing [B] contains, in turn, [P1" P,] which, consequently, is the
least of them. This completes the proof.
The state-output interaction property considered in the present analysis is
expressed by the following definition.
DEFINITION 4. A state x of systems (1) is unobservable if the component of the
response depending on the initial state is identically zero for every admissible
input function.
From (2), (3) and (7) it immediately follows that the set of all unobservable
states is a subspace. Introducing the matrix sequences
(77a) Q1 C,

(77b) Q’ Oi-lAI
Qi 1N
i- 2 3

we have the following theorem.

 530 PAOLO D’ALESSANDRO, ALBERTO ISIDORI AND ANTONIO RUBERTI

THEOREM 6. The subset of all the unobservable states of system (1) is a subspace
Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

of R" which
(a) is the largest subspace invariant under A and N and contained in .U[G];
(b) can be expressed as

(78)

Proof The proof can be developed along the same lines as that of Theorem 5.
Referring to the properties of teachability from the origin and unobserva-
bility, or, more exactly, to the subspaces ’p and q, it is possible to effect the

subspaces z’,
systems, i.e.,
,
decomposition of the state space Y" of the system (1) into the direct sum of four
cg and @, following the procedure adopted in the case of linear

(79)

On the basis of a decomposition of this type, which will be called a canonical de-
composition, it is possible to prove the following theorem.

subspaces
form
,,
THEOREM 7. Assuming as basis in the state space the union of bases of the four
c and of a canonical decomposition, equations (1) assume the

a(t)
+/-b(t)
/Aaa Aab Aac Aad
0 Abb 0
I [/Xa(t)
Xb(t)
Abd
&d(t) 0 0 0 Add/ \Xd(t)
Nab
N 0 Nd x(t)
]
+ u(t),
(80)
0
0
Ncc Ncd
0 Ndd
Xc(t)]U(t)+
Xd(t)/
o
o
x.(t)\
y(t)=(O Cb 0 Ca) x(t)
Xc(t)]
x(t)/
 BILINEAR DYNAMICAL SYSTEMS 531

where (xT,0 0 O) T, (0 x
0 O) T, (0 0 xr
O) w and (0 0 0 x) w are
Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

coordinates of vectors belonging respectively to the subspaces

(canonical form).
Moreover, within the equivalence relation induced by a constant coordinate
transformation in the state space, the quadruplets {A, N, B, C} which assume the
.form associated with (80) are an equivalence class modulo the transformation defined
by a matrix of the type

0
(81) T
0 0
0 0 0
where the partitions are consistent with those oj" the state vector.
Proo The proof of (80) is a standard consequence of statements (a) in
Theorems 5 and 6. The proof of the equivalence between canonical forms can be
deduced directly from that given by the authors in [3] for a similar result in the case
of linear systems.
A further direct analogy with well-known results in the case of linear systems
can be obtained by calculating the kernels w (t,..., t) of the Volterra series
expansion for the zero-state response. In fact, one finds that these kernels depend
only on the matrices Abb, Nbb, Bb, Cb associated with the "part b" resulting from
the canonical decomposition.
Furthermore, always on the basis of the analogy with the case of linear
systems, it becomes natural to establish a connection between the results of the
structure analysis and the realization theory, characterizing the minimality by
means of the properties that identify the "part b".
To this end it is convenient to show that the subspaces fp and fq, considered
in the structure analysis, can also be characterized in terms of the matrices (20)
and (21) on which the realization theory was based. In fact, we have
v
(s2)
(83) {[Ne Ce’]} ,
{[e’N, e’B]},

that are alternative to the ones provided by (69) and (78). To prove this result,
observe first that, from definition (20), we have

(84) k[etN, e’B] [etN, etB] + etN [eN, etB]N*eA*t dr.

k_

Again, by definition,
(85) N{[e’N, e’B]} N[B AB
from which (84) supplies

{2[eA’N, eA’B]} [B AB A 1B],

(86)
+ [N(B AB A B) A 1N(B AB .A"- 1B)].
 532 PAOLO D’ALESSANDRO, ALBERTO ISIDORI ,AND ANTONIO RUBERTI

By iterating this construction, and taking into account (68) and (69), it is possible
Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

to arrive at
(87) {Ee’N, e’B]} [P P... P] C.
Equation (83) is proved in a similar way.
We are now in a position to prove the following theorem.
THEOREM 8. A realization {A, N, B, C} of a bilinearly realizable sequence of
kernels is minimal if and only if its state space is both observable and spanned by the
states reachable from the origin.
Proof. Clearly, minimality implies the structure properties of the statement;
otherwise, it would be possible to find a realization with a lower dimension (i.e.,
the one defined by the quadruplet {Abb, Nb, B, C} of the canonical form (80)).
The converse may be proved in the following way. Let {A, N, B, C} and {A, N, B, C}
be any two realizations with their own state space reachable from the origin and
,
observable; let n, denote their dimensions. On the basis of definitions (18) and
(19) and from factorization (9), denoting max (n, ) with it can be seen that
Qn[ NeAt, CeAt](tl, t)Pn[eArN, eAtB](t+ 1, "’", ten)
(88)
Q[Ne at, e2t](tl, t)P[etN, egt](t+ 1, t2)"
The hypothesis that the state space of the quadruplet {A, N, B, C} is spanned
by the states reachable from the origin and observable implies, thanks to (82)
and (83), that the n columns of Q and the n rows of P in the L.H.S. of (88) are
linearly independent. The same can be said for the columns of Q and the fi rows
of P in the R.H.S. of (88). From this we conclude that n h. This, together with
the result proved in the first part of the proof, implies that any realization with
the aforesaid structure properties is minimal.
Remark 2. Since, by construction, the matrix quadruplet {Ab, N,B, C}
identified by the canonical decomposition satisfies the conditions of Theorem 8,
it can be considered as a minimal realization of the sequence of kernels associated
with the Volterra series expansion of the zero-state response of system (1).
We have also the following corollary.
COROLLARY 2. The dimension no of a minimal realization of a realizable
sequence of kernels is given by
(89) no rank l.,[Ne At, CeAt]:,[eAtN, eAtS]},
where {A, N, B, C} is any given realization of the same sequence.
7. Outline of realization methods. In the proof of the condition of bilinear
realizability for a sequence of kernels (Theorem 1), in the analysis of the properties
of the factorizations of such sequences (Theorem 2), and also in the study of the
connection between the structure analysis and the realization theory (Theorem 8),
there have emerged various procedures that can be considered as steps on the way
to the construction of minimal realizations, when a triplet {F(t), G(t),H(t)} is
available. In order to get an overall view of the various possibilities, it might
perhaps be useful to list the partial computational procedures which, as mentioned,
emerge from the previous analysis"
 BILINEAR DYNAMICAL SYSTEMS 533

(a) Reduction of a factorization, i.e., construction of a minimal factorization

Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

{Fo(t), Go(t),Ho(t)} from any given factorization {F(t), G(t), H(t)} (see proof of
Theorem 3).
(b) Construction of a linear realization from the matrix

L(O [Wl()
G(t) F(t)/
H(O

using one of the many existing procedures (see proof of Theorem 1).
(c) Reduction of a bilinear realization, i.e., construction of the quadruplet
{Abb, Nbb, Bb, Cb) from any realization {A, N, B, C} (see Theorem 7 and Remark 2).
Combining these procedures, various methods for the construction of
minimal bilinear realizations can be developed. For example, one could apply
procedure (b), without seeking a minimal realization, and subsequently procedure
(c). Alternatively, one could successively apply procedures (a) and (b), being careful
to ensure that the latter leads to a minimal realization.
It seems worthwhile to stress the considerable analogies that exist between
these procedures and the ones that are used in the realization theory of time-
invariant linear systems.
The outlined methods presuppose the knowledge of a triplet {F(t), G(t), H(t))
satisfying the factorizability condition (9) for all integers i. An interesting problem,
at this point, would be to determine whether the triplet may be computed from a
finite sequence of kernels, or not. This problem is not considered in full in the
present paper; we only observe that its actual significance relies upon the result
expressed by the following theorem.
,
THEOREM 9. A factorizable sequence of kernels {wi(tl,... ti} is uniquely
specified by the sequence {w(t,..., t)}22m+ where mo is the dimension of its
minimal factorizations.
Proof The proof will be constructive; in other words, it will consist m+ of a
procedure for determining uniquely the kernel WZmo+ 2 from {wi(tl, ti)}22 1.
Consider a minimal factorization {Fo(t), Go(t),Ho(t)} of the given sequence
and put
(90) ,,(t,, ..., t+) Q[Fo, H0](tl, "", t)P[Fo, Go](tk+ 1,’’’, tk+h),
where (see (9)) the element in position (i, j) is equal to the kernel of order +j
calculated for suitable arguments.
Proceed now by constructing:
(a) A factorization
(91) mo,,no(tl, t2mo)= R(t,,..., tno)S(t,no+ ,..., t2m0)
"",

such that the columns of R(t, ..., tno) and the rows of S(t, ..., tmo) are linearly
independent. By comparing this with (90), written for k h too, and recalling
Theorem 2, it follows (see [9]) that
R Qmo[Fo, HolM,
(92)
S M-Ipmo[Fo, Go],
where M is a constant nonsingular matrix.
 534 PAOLO D’ALESSANDRO, ALBERTO ISIDORI AND ANTONIO RUBERTI

(b) The nonsingular matrices

Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

(93) U
fA mO
R*R dtl dt, o,

(94) V
fA mO
SS* dtl dt,,o.
(c) The matrices

(95)
/(tl,"’, tmo+l) f/x mo+l,mo(tl, "’’, t2mo+l)
S*(tmo+ 2, t2mo+ 1) dtmo+ 2 dt2mo+ V- 1,

(96)
(tmo+ 2,’’’, t2mo+ 2) C-1 ;A mO
R*(t2,.", tmo+ 1)

mo,mo+ 1(t2, t2mo+ 2) dr2 dr,,o+ 1"

By direct substitution it is possible to verify that
(97) R(tl, tmo+ )( tmo+2, t2mo+2) 5gmo+ 1,mo+ l(t, t2mo+2).
The procedure outlined above made use of the kernels w2, "’’, W2mo+
which form part of the matrices 5eo,,, o, 5eo,,, + 1,5emo + 1,too" AS a result one obtains
the matrix 5emo + 1,mo+1 and this, as shown by (90), written for k h =mo + 1,
contains the additional element w2,,o + 2; this additional element is the very kernel
it was desired to construct. Since this procedure can be iterated, the proof is
complete.
The validity of this theorem does not rely (see the remarks at tile end of 3)
upon the assumption that F(t), G(t) and H(t) are functions with proper rational
Laplace transforms. If this is the case, it is possible to reduce further the amount
of information needed to specify the sequence {wi(tl,..., ti)} (see, for example,
[4]).
8. Conclusions. We end this paper with some concluding remarks. First of
all we would emphasize the strict analogy between the theory developed here
and that of the linear systems, whose well-known results can easily be obtained
merely by putting N 0. Another advantage of this theory is the systematic use
of linear algebra tools.
Results similar to the ones presented here in 6 are given in [5] for discrete-
time bilinear systems. A complete analysis including both discrete-time and
continuous-time systems is also presented in [4].
The present theory can be extended to cover the cases in which the initial state
is an arbitrary equilibrium state [7]. This makes it possible to handle bilinear
systems homogeneous in the state (i.e., B 0 in equation (1)).
Finally, for the sake of completeness, we also note that another approach is
possible to the problem of constructing minimal bilinear realizations from a given
nonlinear input/output map. This approach reduces the realization problem to
 BILINEAR DYNAMICAL SYSTEMS 535

that of matching an infinite sequence of input/output parameters: on this basis

Downloaded 01/03/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

it is possible to develop, for bilinear systems, a realization theory !-6] analogous to

the one originated after a well-known paper by B. L. Ho and R. E. Kalman.

REFERENCES
1] R.W. BROCKETT, On the algebraic structure ofbil&ear systems, Theory and Applications of Variable
Structure Systems, R. R. Mohler and A. Ruberti, eds., Academic Press, New York, 1972,
pp. 153-168.
[2] C. BRUNI, G. Dr PILLO AND G. KOCH, On the mathematical models of bilinear systems, Ricerche di
Automatica, 2 (1971), pp. 11-26.
[3] P. D’ALESSANDRO, A. ISIDORI AND A. RUaERTI, A new approach to the theory of canonical decomposi-
tion of linear dynamical systems, this Journal, (1973), pp. 148-158.
[4] Lectures on Bilinear System Theory, Notes for a Course held at C.I.S.M., Udine, Italy,
Springer-Verlag, Wien, 1972.
[5] P. D’ALESSANDRO, Structure properties, invariance and insensitivity of bilinear discrete-time systems,
Ricerche di Automatica, 3 (1972), pp. 158-169.
[6] A. ISlDORI, Direct construction of minimal bilinear realizations from nonlinear input/output maps,
IEEE Trans. Automatic Control. AC-18 (1973), to appear.
[7] A. ISlDORI AND A. RUBERTI, Realization theory ofbilinear systems, Geometric Methods in System
Theory, D. Q. Mayne and R. W. Brockett, eds., D. Reidel, Dordrecht, 1973.
[8] R. R. MOHLER, Natural bilinear control processes, IEEE Trans. SSC, SSC-6 (1970), pp. 192-197.
[9] D. C. YOULA, The synthesis of linear dynamical systems from prescribed weighting patterns, SIAM J.
Appl. Math., 14 (1966), pp. 527-549.