Linear Algebra and its Applications 413 (2006) 458473

www.elsevier.com/locate/laa
A polynomial matrix approach to the structural
properties of 2D positive systems
Ettore Fornasini
a,
, Maria Elena Valcher
a
, Raquel Pinto
b
a
Dipartimento di Ingegneria dellInformazione, Universit di Padova, via Gradenigo 6B, 35131 Padova, Italy
b
Department of Mathematics, University of Aveiro, 3810-193 Aveiro, Portugal
Received 21 December 2004; accepted 15 November 2005
Available online 10 January 2006
Submitted by G. de Oliveira
Abstract
Reachability and observability of two-dimensional (2D) discrete state-space models are introduced in
two different forms: a local form, which refers to single local states, and a global form, which pertains to the
innite set of local states lying on a separation set [M. Bisiacco, State and output feedback stabilizability
of 2D systems, IEEE Trans. Circ. Syst., CAS-32 (1985) 12461249; E. Fornasini, G. Marchesini, Global
properties and duality in 2-D systems, Syst. Control Lett. 2 (1) (1982) 3038]. While local reachability
and observability can be naturally characterized by resorting to classical state space techniques, their global
counterparts are better addressed by means of polynomial techniques. In this paper, reachability and observ-
ability are introduced in the context of 2D positive systems and their global versions investigated via a
polynomial approach. Necessary and sufcient conditions for the existence of these properties are provided
and, in particular, polynomial canonical forms for globally reachable/observable positive systems with scalar
inputs/scalar outputs are provided.
Published by Elsevier Inc.
Keywords: Positive 2D systems; Local reachability; Global reachability; Local observability; Global observability;
Polynomial methods

Corresponding author.
E-mail addresses: [email protected] (E. Fornasini), [email protected] (M.E. Valcher), [email protected] (R.
Pinto).
0024-3795/$ - see front matter Published by Elsevier Inc.
doi:10.1016/j.laa.2005.11.006
E. Fornasini et al. / Linear Algebra and its Applications 413 (2006) 458473 459
1. Introduction
2D positive systems, i.e. two-dimensional state-space models whose input, state and output
variables take nonnegative values, have been introduced in the nineties. The aim was that of
providing a unifying theoretical framework to a family of interesting problems. Specifically,
research interests in this eld have been stimulated by a series of contributions dealing with river
pollution modeling [4], modeling of a single-carriageway trafc ow[9], gas absorption and water
stream heating [14], diffusion of a tracer into a blood vessel [16], etc. These contributions share
two common features: on the one hand all internal variables are intrinsically nonnegative, as they
represent concentrations, pressures, numbers of vehicles, etc., on the other hand, the dynamics is
well described by a (quarter plane causal) 2D state-space model, as the system variables depend
on a time and a space coordinate and obey a quarter plane causality law.
Research efforts were rst oriented to extend positive matrix theory to pairs of matrices (see
e.g. [7,8]), thus leading to the analysis of the free state evolution and the asymptotic stability of
2D positive systems. More recently, research efforts in 2D positive systems have concentrated on
the analysis of their structural properties, and some results about reachability and controllability
have been presented in [1012], by assuming a traditional state-space, and hence geometric,
approach.
In the 2Dsetting, the concepts of reachability and observability are naturally introduced in two
different forms: a weak (local) form, which refers to single local states, and a strong (global)
form, which pertains to the innite set of local states lying on a separation set [1,5]. In this
paper, the aforementioned concepts are introduced and investigated in the context of 2D positive
systems, driven by nonnegative inputs and described by the following state-updating equation
[5]:
x(h +1, k +1) = A
1
x(h, k +1) +A
2
x(h +1, k)
+B
1
u(h, k +1) +B
2
u(h +1, k), (1.1)
y(h, k) = Cx(h, k), h, k Z, h +k 0, (1.2)
where the n-dimensional local states x(, ), the m-dimensional inputs u(, ) and the p-dimensional
outputs y(, ) take nonnegative values, A
1
and A
2
are nonnegative n n matrices, B
1
and B
2
are
nonnegative n m matrices, while C is a nonnegative p n matrix.
The initial conditions are assigned by specifying the (nonnegative) values of the state vectors
x(h, k) on the initial separation set S
0
:= {(h, k) Z Z : h +k = 0}, namely by assigning
all (the innitely many) local states constituting the so-called initial global state
X
0
:= {x(h, k) : (h, k) S
0
}.
While local properties easily lead to a point by point analysis in the discrete grid, and hence
to characterizations [10] based on the (innite family of) real matrices representing the way in
which any input sample at (h, k) contributes to the state evolution at each point of the causality
cone {(i, j) Z Z : i h, j k}, their global counterparts naturally suggest a separation
set viewpoint.
Moreover, the global state X
k
:= {x(h, k h) : h Z} can be represented either by means of
a 2D power series
X
k
(z
1
, z
2
) =

hZ
x(h, k h)z
h
1
z
kh
2
460 E. Fornasini et al. / Linear Algebra and its Applications 413 (2006) 458473
or (once the information about the separation set S
k
we are considering is known) by means of
a 1D power series
X
k
() =

hZ
x(h, k h)
h
,
and the state updating along the separation sets can be described, in turn, in 1D or 2D polynomial
terms. So, global reachability and observability properties may be fruitfully addressed starting
from a polynomial description.
With respect to standard 2D systems, the nonnegativity assumption rules out the possibility of
resorting to those linear algebra tools commonly used when dealing with vector spaces. Indeed,
all state, input and output variables lie in cones and this entails far deep consequences. On the one
hand, the results available are typically weaker, as cones are not necessarily nitely generated and
this occasionally prevents the existence of nite tests for structural properties. On the other hand,
nonnegativity allows to resort to a combinatorial approach, as reachability and observability tests
for 2D positive systems can be frequently translated into problems of path searching in suitable
directed graphs.
The paper is organized as follows: Section 2 introduces some notations and provides both local
and global reachability definitions. Local and global reachability are addressed in Sections 3 and 4,
respectively. Finally, local and global observability are introduced and characterized in Section 5.
2. Preliminary concepts
Before proceeding, it is convenient to include some basic definitions and preliminary concepts
that will be used in the paper. The Hurwitz products of two n n matrices A
1
and A
2
are
inductively dened [5] as
A
1
i

j
A
2
= 0, if either i or j is negative,
A
1
i

0
A
2
= A
i
1
, if i 0,
A
1
0

j
A
2
= A
j
2
, if j 0,
A
1
i

j
A
2
= A
1
(A
1
i1

j
A
2
) +A
2
(A
1
i

j1
A
2
), if i, j > 0.
Given a nonnegative vector v R
n
+
, we dene its nonzero pattern as the set p(v) := {i
{1, 2, . . . , n} : v
i
/ = 0}. Notice that the nonzero pattern of the sum of two nonnegative vectors
in R
n
+
is just the set union of their nonzero patterns, i.e., p(v
1
+v
2
) = p(v
1
) p(v
2
). Similar
definitions and remarks extend to nonnegative matrices. So, when the interest is just in the nonzero
patterns of the nonnegative vectors/matrices, and not in the specic values of their nonzero entries,
one can represent any vector v R
n
+
(any matrix M R
pn
+
) by means of the boolean vector
v
B
{0, 1}
n
(the boolean matrix M
B
{0, 1}
pn
) having the same nonzero pattern as v (as M).
We denote by 1
n
the n-dimensional real vector with all entries equal to 1.
Apolynomial vector v(z
1
, z
2
) R
+
[z
1
, z
2
] is said to be an ith p-monomial vector if v(1, 1) has
the same nonzero pattern as the ith canonical vector e
i
, i.e., p(v(1, 1)) = {i}, and its nonzero entry
is a monomial cz
h
1
z
k
2
in R
+
[z
1
, z
2
]. In the following, we will use the term ith p-monomial vector
also for row vectors v
T
(z
1
, z
2
) such that v(z
1
, z
2
) is an ith p-monomial vector. A p-monomial
matrix is a nonsingular (square) matrix whose columns are p-monomial vectors. P-monomial
vectors and p-monomial matrices in R
+
[] are dened in an analogous way. Standard monomial
vectors and monomial matrices in R
+
can be seen as special cases of their general polynomial
versions.
Local and global reachability properties are introduced in Definition 2.1, below.
E. Fornasini et al. / Linear Algebra and its Applications 413 (2006) 458473 461
Denition 2.1. A 2D state-space model (1.1)(1.2) is
(positively) locally reachable [5] if, upon assuming X
0
= 0, for every x

R
n
+
there exist
(h, k) Z Z, with h +k > 0, and a nonnegative input sequence u(, ) s.t. x(h, k) = x

.
When so, we will say that x

is (locally) reachable in h +k steps;

(positively) globally reachable [5] if, upon assuming X
0
= 0, for every global state X

with
entries in R
n
+
, there exist k Z
+
and a nonnegative input sequence u(, ) s.t. the global state
X
k
= {x(h, k h) : h Z} coincides with X

. When so, we will say that X

is (globally)
reachable in k steps.
Notice that all nonnegative input sequences involved have supports included in the half-plane
{(h, k) Z Z : h +k 0}. In the following, the specication positively will be omitted
when no ambiguities arise. Clearly, global reachability ensures local reachability.
3. Local reachability
When dealing with standard (i.e., not necessarily positive) 2D systems, local reachability is
easily tested by evaluating the column span of the reachability matrix in k steps [5], i.e.
R
k
= [B
1
B
2
A
1
B
1
A
1
B
2
+A
2
B
1
A
2
B
2
A
2
1
B
1
(A
1
1

1
A
2
)B
1
+A
2
1
B
2
A
k1
2
B
2
]
=
_
(A
1
i1

j
A
2
)B
1
+(A
1
i

j1
A
2
)B
2
_
i,j0, 0<i+jk
as k varies over the set of positive integers. Indeed, reachable states in k steps, i.e. local states
that can be reached in any assigned position of the separation set S
k
= {(h, k h) : h Z},
starting from X
0
= 0, constitute a linear subspace

X
k
R
n
, spanned by the columns of R
k
.
Clearly, the ascending chain of reachability subspaces

X
1


X
2


X
3
R
n
eventually
reaches stationarity and this necessarily happens, by the 2D CayleyHamilton theorem (see e.g.
[6]), in no more than n steps. As a consequence, if the 2D system is locally reachable, the point
(h, k) where x(h, k) attains the desired value x

(see Definition 2.1) can always be chosen on the

separation set S
n
.
Once we constrain the system and the input sequence to be nonnegative, the reachability
subspaces

X
k
, k N, are replaced by the reachability cones

X
+
k
, k N. In fact, the set

X
+
k
of
all local states that can be reached in any assigned position of the separation set S
k
, by means
of nonnegative inputs and starting from initial zero conditions (X
0
= 0), obviously coincides
with the set of all nonnegative combinations of the columns of R
k
, namely

X
+
k
= Cone(R
k
).
Consequently, a system is locally reachable if and only if there exists N N s.t. Cone(R
N
) =
R
n
+
.
It is worth remarking that for a not locally reachable 2Dpositive systemthe chain of reachability
cones does not necessarily reach stationarity and, indeed, certain positive states can be reached
only asymptotically. However, as the nonzero patterns of the vectors in R
n
+
constitute a nite set,
there exists a positive integer k such that the nonzero patterns of all reachable local states are
the set theoretic unions of the nonzero patterns of suitable columns of R
k
. Consequently, every
nonzero pattern can be reached either in no more than k steps or never.
Positive local reachability is trivially equivalent to the possibility of reaching (starting from
zero initial conditions) every vector of the canonical basis in R
n
by means of nonnegative inputs,
462 E. Fornasini et al. / Linear Algebra and its Applications 413 (2006) 458473
which in turn amounts to saying that there exists some N N s.t. the reachability matrix in N
steps, R
N
, includes an n n monomial submatrix.
Proposition 3.1 [10]. Given a 2D system (1.1)(1.2) the following facts are equivalent:
(i) the system is locally reachable;
(ii) there exist n pairs (h
i
, k
i
) Z
+
Z
+
, i = 1, 2, . . . , n, and n indices j = j (i)
{1, 2, . . . , m} s.t.
p
_
(A
1
h
i
1

k
i
A
2
)B
1
e
j
+(A
1
h
i

k
i
1
A
2
)B
2
e
j
_
= {i}; (3.1)
(iii) there exists N Nsuch that the reachability matrix in N steps R
N
has an n n monomial
submatrix.
Remarks. (i) For 2D systems with scalar inputs, the aforementioned condition (3.1) simply
becomes: there exist n pairs (h
i
, k
i
) Z
+
Z
+
, i = 1, 2, . . . , n, s.t.
p
_
(A
1
h
i
1

k
i
A
2
)B
1
+(A
1
h
i

k
i
1
A
2
)B
2
_
= {i}.
Notice, nally, that all pairs (h
i
, k
i
) are necessarily distinct, but the case may occur that h
i
+k
i
=
h
j
+k
j
for i / = j.
(ii) Differently from the 1D case, we cannot ensure that the local reachability index N is
bounded from above by the system dimension n [10].
4. Global reachability
When dealing with a polynomial description of the forced state evolution, the global state on
the kth separation set
X
k
(z
1
, z
2
) =

hZ
x(h, k h)z
h
1
z
kh
2
can be expressed in terms of the input sequences on the separation sets S
t
, 0 t k 1, as
follows:
X
k
(z
1
, z
2
) = R
k
(z
1
, z
2
)
_
_
_
_
_
U
k1
(z
1
, z
2
)
U
k2
(z
1
, z
2
)
.
.
.
U
0
(z
1
, z
2
)
_

_
, (4.1)
where
R
k
(z
1
, z
2
) =
_
(B
1
z
1
+B
2
z
2
) (A
1
z
1
+A
2
z
2
)(B
1
z
1
+B
2
z
2
)
(A
1
z
1
+A
2
z
2
)
k1
(B
1
z
1
+B
2
z
2
)
_
(4.2)
and
U
t
(z
1
, z
2
) =

hZ
u(h, t h)z
h
1
z
t h
2
, t = 0, 1, . . . , k 1.
Starting from this 2D polynomial description, we obtain the following characterization of global
reachability.
E. Fornasini et al. / Linear Algebra and its Applications 413 (2006) 458473 463
Proposition 4.1. The 2D system (1.1)(1.2) is globally reachable if and only if there exists some
nonnegative index N such that the 2D polynomial matrix R
N
(z
1
, z
2
) given in (4.2) includes an
n n p-monomial submatrix, i.e.
M diag{z

1
1
z

1
2
, z

2
1
z

2
2
, . . . , z

n
1
z

n
2
}, (4.3)
for some monomial matrix M and some
i
,
i
0, i = 1, 2, . . . , n.
Proof. It sufces to focus on the reachability of the elementary global states consisting of
all zero (local) states except for one of them, which coincides with the monomial vector e
i
,
i {1, 2, . . . , n}. Indeed, if the system is globally reachable then, in particular, all elementary
global states must be reachable. On the other hand, if all elementary global states are reachable,
each of themcan be reached by means of a suitable nite support nonnegative input sequence. So,
by superposing nonnegative combinations of such nite support input sequences, one can reach
every nonnegative global state. Consequently, the 2D system (1.1)(1.2) is globally reachable if
and only if, for every i {1, 2, . . . , n}, there exists k
i
N and h
i
Z such that
X
k
i
(z
1
, z
2
) = e
i
z
h
i
1
z
k
i
h
i
2
is a global state reachable in k
i
steps. This amounts to saying that
e
i
z
h
i
1
z
k
i
h
i
2
= R
k
i
(z
1
, z
2
)
_
_
_
_
_
_
_
U
(i)
k
i
1
(z
1
, z
2
)
U
(i)
k
i
2
(z
1
, z
2
)
.
.
.
U
(i)
0
(z
1
, z
2
)
_

_
,
for some U
(i)
t
(z
1
, z
2
), t = 0, 1, . . . , k
i
1. It entails no loss of generality assuming that each
U
(i)
t
(z
1
, z
2
) has nite support, namely it is a Laurent polynomial. The nonnegativity of the coef-
cients of all polynomial matrices and vectors involved ensures that the above condition holds
true if and only if there exists at least one column of R
k
i
(z
1
, z
2
) taking the following structure
e
i
c
i
z

i
1
z

i
2
, c
i
R
+
. So, the proposition statement holds for N = max{k
1
, k
2
, . . . , k
n
}.
Remark. The characterization given in Proposition 4.1 above, may be restated in terms of poly-
nomial reachability matrices in the single variable . Indeed, the 2Dsystem(1.1)(1.2) is globally
reachable if and only if there exists N N such that
R
N
() = [ (B
1
+B
2
) (A
1
+A
2
)(B
1
+B
2
) . . . (A
1
+A
2
)
N1
(B
1
+B
2
) ]
includes an n n p-monomial submatrix, i.e.:
M diag{

1
,

2
, . . . ,

n
},
for some monomial matrix M and some
i
0, i = 1, 2, . . . , n.
As an immediate corollary of the previous result, we get
Corollary 4.2. If the 2D system (1.1)(1.2) is globally reachable then
[ A
1
+A
2
B
1
+B
2
]
includes an n n p-monomial matrix.
Proof. The nonnegativity of the coefcients, together with the fact that the product of two poly-
nomials is a monomial if and only if they are both monomials, allows to saying that
464 E. Fornasini et al. / Linear Algebra and its Applications 413 (2006) 458473
(A
1
+A
2
)
k
i
1
(B
1
+B
2
)e
j
= e
i
(c
i

i
) (4.4)
c
i
R
+
,
i
Z
+
, implies either that the jth column of B
1
+B
2
is an ith p-monomial vector
(if k
i
= 1) or that some column of A
1
+A
2
is an ith p-monomial vector (if k
i
> 1). Since (4.4)
must be veried for every i {1, 2, . . . , n}, the result immediately follows.
The following lemma leads the way to further characterizations of global reachability.
Lemma 4.3. If the 2D system (1.1)(1.2) is globally reachable then the 1D positive system
described by the pair (A
1
+A
2
, B
1
+B
2
) is (positively) reachable.
Proof. By the previous Proposition 4.1, the 2D system (1.1)(1.2) is globally reachable if and
only if there exists some positive index N such that the 2D polynomial matrix R
N
() includes an
n n p-monomial submatrix, i.e.:
M diag{

1
,

2
, . . . ,

n
},
for some monomial matrix M and some
i
0, i = 1, 2, . . . , n. Since this condition holds true
for an arbitrary , then it must hold true for = 1. This means that the reachability matrix in N
steps of the pair (A
1
+A
2
, B
1
+B
2
) includes the monomial submatrix M and hence [3] the pair
(A
1
+A
2
, B
1
+B
2
) is positively reachable.
Lemma 4.4. If there exist an integer Z
+
and a nonzero polynomial p() R
+
[] such that
(A
1
+A
2
)

(B
1
+B
2
)e
j
= p()e
i
for some indices i {1, 2, . . . , n} and j {1, 2, . . . , m}, then there exists

Z
+
, 0

n
1, and a nonzero p() R
+
[] such that
(A
1
+A
2
)

(B
1
+B
2
)e
j
= p()e
i
.
Proof. We associate with the polynomial matrix pair (A
1
+A
2
, B
1
+B
2
) a directed graph
with n vertices and m sources: there is an arc connecting vertex j to vertex i if and only if the
(i, j)th entry of A
1
+A
2
is nonzero and, similarly, there is an arc connecting source j to vertex
i if and only if the (i, j)th entry of B
1
+B
2
is nonzero. Each arc is thus weighted by some
nonzero polynomial c +d, c, d R
+
.
In graph theoretic terms, condition
(A
1
+A
2
)

(B
1
+B
2
)e
j
= p()e
i
holds for some nonzero polynomial p() R
+
[] if and only if there exists a deterministic
path [10,15] from the jth source to the ith vertex in the directed graph associated with the
pair (A
1
+A
2
, B
1
+B
2
), namely a path starting from the jth source and reaching, after +1
steps, the vertex i and no other vertex. Such a condition holds, in turn, if and only if there exists
a deterministic path from the jth source to the ith vertex in the directed graph associated with
(A
1
+A
2
, B
1
+B
2
). But then, we may resort to the result obtained by Coxson and Larson in [2]
and say that if such a path exists, then there exists a path from source j to vertex i of length not
larger than n. This corresponds to saying that
(A
1
+A
2
)

(B
1
+B
2
)e
j
= c e
i
E. Fornasini et al. / Linear Algebra and its Applications 413 (2006) 458473 465
holds true for some

< n and some c > 0 and hence that
(A
1
+A
2
)

(B
1
+B
2
)e
j
= p()e
i
holds true for some

< n and some nonzero polynomial p() R
+
[].
Proposition 4.5. The 2D system (1.1)(1.2) of size n is globally reachable if and only if the 2D
polynomial matrix R
n
() includes an n n p-monomial submatrix.
Proof. By Proposition 4.1, the systemis globally reachable if and only if there exists some N N
such that the 2D polynomial matrix R
N
() includes an n n p-monomial submatrix. However,
by the previous Lemma 4.4, if
(A
1
+A
2
)

(B
1
+B
2
)e
j
= c

i
e
i
for some nonzero monomial c

i
R
+
[], then there exists

Z
+
, 0

n 1 and a
nonzero p() R
+
[] such that
(A
1
+A
2
)

(B
1
+B
2
)e
j
= p()e
i
.
This implies, in particular, that
(A
1
+A
2
)

[(A
1
+A
2
)

(B
1
+B
2
)e
j
] = (A
1
+A
2
)

[ p()e
i
] = c

i
e
i
.
As a consequence, the ith column of (A
1
+A
2
)

must be an ith p-monomial vector, p() is

necessarily a monomial, and there exists an ith monomial column p()e
i
in R
n
().
Remarks. (i) For 2D systems with scalar inputs, the proof of Proposition 4.5 is much easier, as
indeed Lemma 4.4 is unnecessary. Global reachability of a 2D system with scalar inputs ensures,
by Lemma 4.3, that the reachability matrix in n steps associated with the pair (A
1
+A
2
, B
1
+B
2
)
is monomial and hence
R
n
() = M diag{p
1
(), p
2
(), . . . , p
n
()}
for some monomial matrix M and some polynomials p
i
() R
+
[], i = 1, 2, . . . , n. But then,
one can follow the same reasoning adopted within the previous proof and say that
(A
1
+A
2
)

(B
1
+B
2
) = c

i
e
i
for some n implies that the column of R
n
() having a nonzero entry only in the ith position
is already an ith p-monomial vector.
(ii) The nonzero pattern of any global state X
k
on the separation set S
k
is dened as
P(X
k
) := {(i, j) {1, 2, . . . , n} Z : x
i
(j, k j) / = 0}.
Since all global states of a globally reachable 2D system can be obtained on the separation set
S
n
, every subset of {1, 2, . . . , n} Z represents the nonzero pattern of a global state that can be
reached on S
n
. On the other hand, if a 2D positive system is not globally reachable, the families
of the nonzero patterns of the reachable global states on S
k
may constitute a chain that strictly
increases when k goes to +, as shown by the following example.
Example. Consider the 2D positive system described by the following matrices:
A
1
=
_
_
_
_
0 0 0 0
1 1 0 0
0 0 0 0
0 0 1 1
_

_
, A
2
=
_
_
_
_
0 0 0 0
0 0 0 0
1 0 1 0
0 1 0 1
_

_
, B
1
=
_
_
_
_
1
0
0
0
_

_
, B
2
=
_
_
_
_
0
0
0
0
_

_
.
466 E. Fornasini et al. / Linear Algebra and its Applications 413 (2006) 458473
Since nonzero patterns of the global states are independent of the specic nonzero values taken
by the matrix entries and by the inputs sequences, for the sake of simplicity we will represent
nonzero entries as units or, equivalently, we will assume that all entries belong to the boolean
algebra {0, 1}. Consequently, the reachability matrix in k +2 steps is given by
R
k+2
() =
_
_
_
_
1 0 0 0 0
0 1 1 1 1
0
2

3

k+1
0 0 +
2
+
2
+ +
k
_

_
.
It is clear that for every k N the global state
e
2
+e
4
( +
2
+ +
k
) +e
3

k+1
can be reached only after k +2 steps, and therefore, as k increases, the nonzero patterns of the
reachable global states constitute a strictly ascending chain of subsets in {1, 2, 3, 4} Z.
A nice polynomial canonical form can be obtained for globally reachable systems with scalar
inputs.
Proposition 4.6. For a 2D system (1.1)(1.2) of size n with scalar inputs the following facts are
equivalent:
(i) the system is globally reachable;
(ii) there exists a permutation matrix P such that
P
T
(A
1
+A
2
)P =
_
_
_
_
_
_
_
_
a
12
0
0 a
23
0

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
a
n1,n
0 . . . 0 0
_

_
, P
T
(B
1
+B
2
) =
_
_
_
_
_
_
_
0
0
0
.
.
.
b
n
_

_
,
(4.5)
where a
i,i+1
, b
n
R
+
[] are nonzero monomials and denotes a polynomial in R
+
[] of
degree at most 1;
(iii) the 2D polynomial matrix R
n
() is a p-monomial matrix;
(iv) R
n
() R
+
[]
nn
is nonsingular and R
n
()
1
belongs to R
+
[
1
]
nn
.
Proof. (ii) (iii) (i) are obvious.
(i) (ii) By Corollary 4.2, global reachability ensures that the n (n +1) polynomial matrix
[ A
1
+A
2
B
1
+B
2
]
includes an n n p-monomial matrix. Suppose that B
1
+B
2
is not a monomial vector. Since
it cannot be zero, then it must either have at least two nonzero entries (case 1) or be a vector of
the following type p()e
i
, for some polynomial p() of lag 1 (case 2). On the other hand, since
the block matrix must include an n n p-monomial matrix, such a matrix must be A
1
+A
2
.
E. Fornasini et al. / Linear Algebra and its Applications 413 (2006) 458473 467
It is easily seen that, under these hypotheses, both in case 1 and in case 2, none of the vectors
(A
1
+A
2
)
k
(B
1
+B
2
) can be p-monomial. So, B
1
+B
2
is necessarily a p-monomial vector.
It entails no loss of generality assuming B
1
+B
2
= e
n
. In fact, we can always reduce ourselves
to this case by permuting either the vector components or the matrices B
1
and B
2
, possibly
both. Clearly, at most one column of A
1
+A
2
is not p-monomial and the set of the remaining
n 1 columns of A
1
+A
2
includes an ith p-monomial vector for i = 1, 2, . . . , n 1. Suppose
that the last column of A
1
+A
2
is not p-monomial. This implies that, on the one hand, all
remaining columns of A
1
+A
2
are p-monomial, on the other hand (A
1
+A
2
)(B
1
+B
2
) is
not p-monomial. For these reasons, both in case (A
1
+A
2
)(B
1
+B
2
) has at least two nonzero
entries (case 1) and in case it is a vector like p()e
i
, for some polynomial p() of lag 1 (case
2), also the following powers (A
1
+A
2
)
i
(B
1
+B
2
), i > 1, are not p-monomial. Suppose now
that (A
1
+A
2
)(B
1
+B
2
) is a p-monomial vector. Clearly the nonzero entry cannot be in the
last (namely nth) position, otherwise all powers (A
1
+A
2
)
i
(B
1
+B
2
) would have the same
structure, and it entails no loss of generality assuming that the only nonzero entry lies in the
n 1th row. We can now repeat the same reasoning we just applied to the last row and claim
that if the n 1th column would not be p-monomial then all the other columns in A
1
+A
2

would not be, and hence all remaining powers (A

1
+A
2
)
i
(B
1
+B
2
), i 2, would not be
p-monomial. In this way we have proven that (upon a suitable permutation) we can assume
that all columns of A
1
+A
2
, except possibly for the rst one, have to be p-monomial vectors
and
A
1
+A
2
=
_
_
_
_
_
_
_
_
a
12
0
0 a
23
0

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
a
n1,n
0 . . . 0 0
_

_
, B
1
+B
2
=
_
_
_
_
_
_
_
0
0
0
.
.
.
b
n
_

_
, (4.6)
where a
i,i+1
, b
n
R
+
[] are nonzero monomials and the entries denoted by are polynomials
in R
+
[] (of degree at most 1).
(iii) (iv) Suppose that R
n
() belongs to R
+
[]
nn
and its inverse R
n
()
1
to R
+
[
1
]
nn
.
So, at each point

R
+
R
n
(

) and R
n
(

)
1
are nonnegative matrices satisfying I
n
= R
n
(

)
R
n
(

)
1
. Since the only nonnegative square matrices endowed with nonnegative inverses are
monomial, this implies that R
n
(

) is monomial for every

R
+
. This is possible (if and) only
if R
n
() R
+
[]
nn
is p-monomial and hence satises (iii). The converse is obvious.
5. Local and global observability
Global and local reachability definitions have been introduced in Section 2 by referring to arbi-
trary global or local nonnegative states to be reached (starting from zero global initial conditions)
by means of nonnegative inputs. It can be shown that referring to the nonzero patterns of the
(global/local) states and of the input sequences (instead of considering their specic nonnegative
values) leads to reachability definitions that are exactly equivalent to those given in Section 2
and, consequently, a nonzero pattern approach to positive reachability is just an alternative way
for introducing the same concepts.
468 E. Fornasini et al. / Linear Algebra and its Applications 413 (2006) 458473
Denition 5.1. A 2D state-space model (1.1)(1.2) is
locally reachable if, upon assuming X
0
= 0, for every boolean vector x

B
{0, 1}
n
there exist
(h, k) Z Z, with h +k > 0, and a boolean input sequence u(, ) s.t. x(h, k) has the same
nonzero pattern as x

B
;
globally reachable if, upon assuming X
0
= 0, for every sequence of n-dimensional boolean
vectors {x
B
(h)}
hZ
, there exist k Z
+
and a boolean input sequence u(, ) determining a
global state X
k
= {x(h, k h), h Z} on S
k
with
P(X
k
) = P({x
B
(h)}
hZ
),
namely the nonzero patterns of the two sequences ordinately coincide.
Onthe other hand, if we aimat introducingobservabilitydefinitions startingfromthe free output
evolutions of 2D positive systems, and pretend that they provide reasonable dual properties w.r.t
local and global reachability, a nonzero pattern approach is somehow unavoidable.
Denition 5.2. A 2D state-space model (1.1)(1.2) is
locally observable if, upon assuming that the initial global state X
0
consists of a single nonzero
local state x(0, 0), the knowledge of the nonzero pattern of the free output evolution y

(h, k)
in every point (h, k) Z
+
Z
+
allows to uniquely determine the nonzero pattern of x(0, 0);
globally observable if the knowledge of the nonzero pattern of the free output evolution y

(h, k)
in every point (h, k) Z Z, h +k 0, allows to uniquely determine the nonzero pattern
of the initial global state X
0
.
It is easy to see that global observability trivially implies local observability, as, indeed, among
all possible initial global states one may consider those consisting of all zero local states except
at (0, 0), and, corresponding to that type of global states, the support of the free output evoluation
is included in the rst orthant.
In order to explore local observability, we introduce the observability matrix in k steps, i.e.
O
k
=
_
_
_
_
_
_
_
_
_
_
_
_
_
C
CA
1
CA
2
CA
2
1
C(A
1
1

1
A
2
)
CA
2
2
.
.
.
CA
k1
2
_

_
=
_
C(A
1
i

j
A
2
)
_
i,j0, 0i+j<k
,
where k is a positive integer. As a rst step, we provide a characterization of local observability.
Proposition 5.3. Given a 2D system (1.1)(1.2) the following facts are equivalent:
(i) the system is locally observable;
(ii) there exist n pairs (h
i
, k
i
) Z
+
Z
+
, i =1, 2, . . . , n, and n indices j =j (i) {1, 2, . . . ,
p} s.t.
E. Fornasini et al. / Linear Algebra and its Applications 413 (2006) 458473 469
p(e
T
j
C(A
1
h
i

k
i
A
2
)) = {i}; (5.1)
(iii) there exists N N s.t. the observability matrix in N steps O
N
has an n n monomial
submatrix.
Proof. (i) (ii) Suppose, by contradiction, that the system is locally observable but ii) does
not hold. This means that there exists {1, 2, . . . , n} s.t. none of the rows of the observability
matrix in k steps, for any k N, is an th monomial vector. It is easy to verify that the initial
states x(0, 0) = 1
n
and x(0, 0) = 1
n
e

have different nonzero patterns but produce free output

evolutions endowed with the same nonzero patterns. Thus the systemcannot be locally observable.
(ii) (i) If (ii) holds true, the ith entry of the local state x(0, 0) is nonzero if and only if
e
T
j
y

(h
i
, k
i
) / = 0, i = 1, 2, . . . , n. So, the system is locally observable.
The equivalence of (ii) and (iii) is obvious.
When dealing with 2Dsystems with scalar outputs, condition (ii) above simply becomes: there
exist n pairs (h
i
, k
i
) Z
+
Z
+
, i = 1, 2, . . . , n, s.t. C(A
1
h
i

k
i
A
2
) is an ith (row) monomial
vector. Notice that, also in this case, all pairs (h
i
, k
i
) are necessarily distinct, but the case may
occur that h
i
+k
i
= h
j
+k
j
for i / = j.
In order to address global observability by means of polynomial techniques, we express the
free output evolution on each separation set S
t
by means of a power series:
Y
t
(z
1
, z
2
) =

hZ
y(h, t h)z
h
1
z
t h
2
and relate it to the global initial conditions on the separation set S
0
as follows:
Y
t
(z
1
, z
2
) =

hZ
Cx(h, t h)z
h
1
z
t h
2
=

hZ
C
t

=0
(A
1

t
A
2
)x(h , h)z
h
1
z
t h
2
= C
t

=0
(A
1

t
A
2
)
_

hZ
x(h , h)z
h
1
z
h
2
_
z

1
z
t
2
= C
t

=0
(A
1

t
A
2
)z

1
z
t
2
X
0
(z
1
, z
2
)
= C(A
1
z
1
+A
2
z
2
)
t
X
0
(z
1
, z
2
).
Consequently
_
_
_
Y
0
(z
1
, z
2
)
.
.
.
Y
k1
(z
1
, z
2
)
_

_
= O
k
(z
1
, z
2
)X
0
(z
1
, z
2
), (5.2)
where
O
k
(z
1
, z
2
) :=
_
_
_
_
_
C
C(A
1
z
1
+A
2
z
2
)
.
.
.
C(A
1
z
1
+A
2
z
2
)
k1
_

_
. (5.3)
470 E. Fornasini et al. / Linear Algebra and its Applications 413 (2006) 458473
Starting from this 2D polynomial description, we can obtain a characterization of global observ-
ability.
Proposition 5.4. The 2D system(1.1)(1.2) is globally observable if and only if there exists some
nonnegative index N such that the 2D polynomial matrix O
N
(z
1
, z
2
) given in (5.3) includes an
n n p-monomial submatrix.
Proof. Of course, if there exists some index N Nsuch that the observability matrix O
N
(z
1
, z
2
)
includes an n n p-monomial submatrix, then there exists a suitable selection of separation sets
S
k
1
, S
k
2
, . . . , S
k
n
, k
i
Nanda correspondingsuitable choice of output components j
1
, j
2
, . . . ,
j
n
{1, 2, . . . , p} such that
_
_
_
_
_
_
e
T
j
1
Y
k
1
(z
1
, z
2
)
e
T
j
2
Y
k
2
(z
1
, z
2
)
.
.
.
e
T
j
n
Y
k
n
(z
1
, z
2
)
_

_
= M diag{z

1
1
z

1
2
, z

2
1
z

2
2
, . . . , z

n
1
z

n
2
}X
0
(z
1
, z
2
).
Since we have already seen that a p-monomial matrix in R
+
[]
nn
exhibits an inverse (in
R
+
[
1
]
nn
) having the same structure, it follows that
diag{z

1
1
z

1
2
, z

2
1
z

2
2
, . . . , z

n
1
z

n
2
}M
1

_
_
_
_
_
_
e
T
j
1
Y
k
1
(z
1
, z
2
)
e
T
j
2
Y
k
2
(z
1
, z
2
)
.
.
.
e
T
j
n
Y
k
n
(z
1
, z
2
)
_

_
= X
0
(z
1
, z
2
).
This allows an entry by entry identication of all local components of the initial global state, and
hence the identication of the nonzero pattern of X
0
.
Conversely, suppose by contradiction that the system is globally observable but there exists
some index {1, 2, . . . , n} such that none of the rows of the observability matrix in k steps
O
k
(z
1
, z
2
), for any k N, is an th p-monomial vector. Two cases may occur: either every
row having a nonzero th entry has also other nonzero entries, or all rows whose only nonzero
entry is the th one, exhibit a polynomial of strictly positive lag in the th position. If so, we
denote by L > 0 the smallest such lag. In the rst case, it is easy to see that the initial global
states
X
0
(z
1
, z
2
) =

hZ
1
n
z
h
1
z
h
2
,

X
0
(z
1
, z
2
) =

hZ
(1
n
e

) z
h
1
z
h
2
have different nonzero patterns but produce free output evolutions endowed with the same nonzero
patterns, thus contradicting global observability. Similarly, in the second case, the initial global
states
X
0
(z
1
, z
2
) =

hZ
1
n
z
h
1
z
h
2
,
E. Fornasini et al. / Linear Algebra and its Applications 413 (2006) 458473 471

X
0
(z
1
, z
2
) =

hZ
(1
n
e

)z
h
1
z
h
2
+

hZ
e

z
h(L+1)
1
z
h(L+1)
2
have different nonzero patterns but produce free output evolutions endowed with the same nonzero
patterns, contradicting again global observability.
Remark. As for global reachability, the equivalent characterization given in Proposition 5.4
above, may be restated in terms of polynomial reachability matrices in the single variable .
Indeed, the 2Dsystem(1.1)(1.2) is globally observable if and only if there exists N Nsuch that
O
N
() =
_
_
_
_
_
C
C(A
1
+A
2
)
.
.
.
C(A
1
+A
2
)
N1
_

_
includes an n n p-monomial submatrix.
Also, as an immediate corollary of the previous result, we get.
Corollary 5.5. If the 2D system (1.1)(1.2) is globally observable then
_
A
1
+A
2

C
_
includes an n n p-monomial matrix.
Starting fromProposition 5.4, it is straightforward to apply the same type of reasonings adopted
in Section 4 for global reachability, thus obtaining
Proposition 5.6. The 2D system (1.1)(1.2) of size n is globally observable if and only if the 2D
polynomial matrix O
n
() includes an n n p-monomial submatrix.
A nice polynomial canonical form can also be obtained for globally observable systems with
scalar outputs, by resorting to the results derived in this section and to the reasonings adopted
within the proof of Proposition 4.6.
Proposition 5.7. For a 2D system (1.1)(1.2) of size n with scalar outputs the following facts
are equivalent:
(i) the system is globally observable;
(ii) there exists a permutation matrix P such that
P
T
(A
1
+A
2
)P =
_
_
_
_
_
_
_
_
. . .
a
21
0 0 0
0 a
32
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0
0 0 . . . a
n,n1
0
_

_
,
CP = [ 0 0 0 . . . c
n
] ,
472 E. Fornasini et al. / Linear Algebra and its Applications 413 (2006) 458473
where a
i,i1
R
+
[] are nonzero monomials, c
n
R
+
, c
n
> 0, and denotes a polyno-
mial in R
+
[] of degree at most 1;
(iii) the 2D polynomial matrix O
n
() is a p-monomial submatrix;
(iv) O
n
() R
+
[]
nn
is nonsingular and its inverse O
n
()
1
belongs to R
+
[
1
]
nn
.
It is worthwhile, at this point, to briey comment on the duality relation existing between
(global/local) reachability and (global/local) observability. Clearly 2D systems lack the com-
pletely symmetric structure 1Dsystems are endowed with, as they have two input-to-state matrices
and a unique state-to-output matrix. However, if we assume (as in certain examples of [10])
B
1
= B and B
2
= 0, then all previous relations are clearly dual. In the general case, duality holds
in a weaker sense, by this meaning that there is an obvious correspondence between the various
characterizations, which can be obtained one from the other provided that we suitably replace the
pair (B
1
, B
2
) with the matrix C and suitably adjust the indices of the Hurwitz products.
6. Conclusions
In the paper, reachability and observability for 2D positive systems are investigated in their
local and global versions. Significantly enough, their characterizations exhibit strict relationships
both with the reachability and observability characterizations for standard (i.e. not necessarily
positive) 2D systems and with those available for 1D positive systems. Indeed, as in the standard
2D case, local reachability and observability naturally involve real matrices in their description,
while the corresponding global properties naturally rely on suitable polynomial matrices [5].
Moreover, as in the 1D case, all properties turn out to be related to the existence of a p-monomial
submatrix within the reachability and observability matrices [3,13].
The necessary and sufcient conditions obtained for the 4 properties take just the same form, but
pertain different (reachability/observability and real/polynomial) matrices, thus providing a nice
general framework where further investigations can be developed. The only significant difference
to remark is that, while global properties can be tested basing on certain block matrices whose
block number does not exceed the 2D system dimension, local properties may require to evaluate
quite large block matrices, and no upper bound on the block number is available up to now.
References
[1] M. Bisiacco, State and output feedback stabilizability of 2D systems, IEEE Trans. Circ. Syst. CAS-32 (1985)
12461249.
[2] P.G. Coxson, L.C. Larson, Monomial patterns in the sequence A
k
b, Linear Algebra Appl. 94 (1987) 89101.
[3] L. Farina, S. Rinaldi, Positive Linear Systems: Theory and Applications, Series on Pure and Applied Mathematics,
Wiley-Interscience, New York, 2000.
[4] E. Fornasini, A 2D systems approach to river pollution modelling, Multidim. Syst. Sign. Proc. 2 (1991) 233265.
[5] E. Fornasini, G. Marchesini, Global properties and duality in 2-D systems, Syst. Control Lett. 2 (1) (1982) 3038.
[6] E. Fornasini, G. Marchesini, Properties of pairs of matrices and state-models for 2D systems, in: C.R. Rao (Ed.),
Multivariate Analysis: Future Directions, North Holland Series in Probability and Statistics, vol. 5, 1993, pp. 131
180.
[7] E. Fornasini, M.E. Valcher, On the spectral and combinatorial structure of 2D positive systems, Linear Algebra
Appl. 245 (1996) 223258.
[8] E. Fornasini, M.E. Valcher, Directed graphs, 2D state models and characteristic polynomial of irreducible matrix
pairs, Linear Algebra Appl. 263 (1997) 275310.
[9] E. Fornasini, M.E. Valcher, Recent developments in 2D positive system theory, J. Appl. Math. Comput. Sci. 7 (4)
(1997) 713735.
E. Fornasini et al. / Linear Algebra and its Applications 413 (2006) 458473 473
[10] E. Fornasini, M.E. Valcher, Controllability and reachability of 2D positive systems: a graph theoretic approach,
IEEE Trans. Circ. Syst., Part I: Regular Papers CAS-52 (2005) 576585.
[11] T. Kaczorek, Reachability and controllability of 2D positive linear systems with state feedback, Control Cybernet.
29 (1) (2000) 141151.
[12] T. Kaczorek, Positive 1D and 2D Systems, Springer-Verlag, London (UK), 2002.
[13] H. Maeda, S. Kodama, Positive realization of difference equations, IEEE Trans. Circ. Syst. CAS-28 (1981) 3947.
[14] W. Marszalek, Two-dimensional state space discrete models for hyperbolic partial differential equations, Appl. Math.
Modell. 8 (1984) 1114.
[15] M.E. Valcher, Controllability and reachability criteria for discrete time positive systems, Int. J. Control 65 (1996)
511536.
[16] S. Vomiero, Unapplicazione dei sistemi 2D alla modellistica dello scambio sangue-tessuto, Ph.D. thesis, Univ. di
Padova, Italy, 1992 (in Italian).