Linear Algebra and its Applications 332334 (2001) 257263

www.elsevier.com/locate/laa
Stability of asynchronous two-dimensional
FornasiniMarchesini dynamical systems

A. Bhaya
a,
, E. Kaszkurewicz
a
, Y. Su
b
a
Department of Electrical Engineering, COPPE, Federal University of Rio de Janeiro, P.O. Box 68504,
Rio de Janeiro 21945-970, Brazil
b
Department of Mathematics, Fudan University, Shanghai, Peoples Republic of China
Received 1 November 1999; accepted 12 September 2000
Submitted by R.A. Horn
Abstract
Conditions for stability of two-dimensional (2D) FornasiniMarchesini (FM) models with
variable shifts are derived using nonnegative matrix theory. The conditions are shift inde-
pendent. The stability is proved under a general boundary condition (BC). 2001 Elsevier
Science Inc. All rights reserved.
Keywords: Stability; 2DFornasiniMarchesini model; Asynchronous 2Dsystems; Time shifts; Boundary
conditions; Saturation
1. Introduction
Fornasini and Marchesini [10] introduced a class of linear dynamical systems in
two discrete-time variables. This class nowbears their name (abbreviated to FM) and
the zero input two-dimensional (2D) linear FM model is given by the equation
x(i +1, j +1) = Ax(i +1, j) +Bx(i, j +1) +Cx(i, j). (1)
This system, as well as an equivalent version called the Roesser model [17], is used
in image processing (representations of lters), in the representation of discretized

This research was partially supported by the Brazilian Government agencies CAPES and CNPq as
well as MCT under the PRONEX program.

Corresponding author.
E-mail addresses: amit@coep.ufrj.br (A. Bhaya), eugenius@coep. ufrj.br (E. Kaszkurewicz),
yfsu@fudan.edu.cn (Y. Su).
0024-3795/01/$ - see front matter 2001 Elsevier Science Inc. All rights reserved.
PII: S 0 0 2 4 - 3 7 9 5 ( 0 0 ) 0 0 3 1 7 - 7
258 A. Bhaya et al. / Linear Algebra and its Applications 332334 (2001) 257263
partial differential equations [16], and in the models of different physical phenomena
such as compartmental systems, single-carriageway trafc ow, and river pollution
[8].
There is a well-developed stability theory for FM dynamical systems, using dif-
ferent approaches (see [4,912,15,18] and the references therein).
A sufcient condition for stability of the FM model, proved in [18], is
(|A| +|B| +|C|) < 1, (2)
where (A) denotes the spectral radius of the matrix A, and, given a matrix M =
(m
ij
), |M| denotes the nonnegative matrix (|m
ij
|).
In this paper, it is proved that this condition actually guarantees the asymptot-
ic stability of a more general 2D FM model with variable shifts and quantization
type nonlinearities. For simplicity, it will be supposed that C = 0 so that the model
considered is the following:
x(i +1, j +1) = Q(Ax(i +1, j d
2
(i, j)) +Bx(i d
1
(i, j), j +1)), (3)
where for all i, j, d
1
(i, j) and d
2
(i, j) are positive integers, referred to as shifts,
Q(x) = (q
1
(x
1
), . . . , q
n
(x
n
))
T
is any function satisfying the condition:
|q
m
()| || for all R and 1 m n. (4)
Many nonlinearities occurring in digital ltering are special cases of (4) (see [14])
and this nonlinearity is often referred to as a sector nonlinearity. The 2D system is
assumed to have boundary conditions (BC) satisfying
lim
i
x(i, 0) = lim
j
x(0, j) = 0. (5)
Stability of the zero solution of the dynamical system (3) is dened in the sense of
Liapunov [14]:
Denition 1.1. Equilibrium x
e
= 0 is stable if
> 0, () > 0 s.t. x(i, j) < i, j 0,
whenever x(i, 0) < , x(0, j) < for all i I, j J.
The motivation for introducing shifts into 2D models is as follows. First, FM
models with shifts may be used to model asynchronous iterative methods for the
solution of partial differential equations [1]. Second, the models and the correspond-
ing stability results are natural generalizations of the well-known ChazanMiranker
result [6] for a 1D system. In [6], the presence of varying shifts was indicated by the
use of the term chaotic iteration, but given the other connotations of the adjective
chaotic, it has become more standard in subsequent work to refer to iterations with
time varying delays [13] or asynchronous iterations [3]. It should also be pointed out
that, in the 1D case, a considerable amount of work has been done on this type of
model and its generalizations and the reader is referred to the papers [5,7,13,19,20]
to get some idea of the models and corresponding results that are available.
A. Bhaya et al. / Linear Algebra and its Applications 332334 (2001) 257263 259
2. Stability of variable shift FM models
In this section, it is shown that the asymptotic stability condition (2) actually
ensures asymptotic stability of the zero solution of the FM model with variable shifts
and sector nonlinearities (3).
Theorem 2.1. Consider the 2D system with variable shifts (3) above, subject to the
BCs satisfying (5). Assume that the variable shifts are bounded, i.e., there exists an
positive integer D such that
0 d
1
(i, j) D 1, 0 d
2
(i, j) D 1, (6)
and assume that
(|A| +|B|) < 1. (7)
Then the zero solution of (3) is asymptotically stable.
Some additional notation and a lemma is needed to proceed with the proof of
Theorem 2.1. For a pair of matrices or vectors A and B of the same size, the partial
order relation between them is denoted and dened as follows: A B means that
every element of A is less than or equal to the corresponding element of B; A < B
means that A B and A = B; A B means that every element of A is strictly less
than the corresponding element of B. Double indices (i, j) are interpreted as 1 2
matrices and the same partial order relation applied to them.
Lemma 2.2 [2]. For any matrices A, B R
nn
, (|A| +|B|) < 1 if and only if
there exists a vector v 0 and a scalar < 1 such that
(|A| +|B|)v v.
For a positive vector v 0, a generalized innity norm is dened as
x
v
= max
i
|x
i
|
v
i
,
where x
i
is the ith component of the vector x. It is clear that
x
v
< |x| < v.
Proof of Theorem 2.1. Since (|A| +|B|) < 1, from Lemma 2.2, there exists a
positive vector v 0 and a scalar : 0 < 1 such that
(|A| +|B|)v v.
For the positive integer D, a sequence S(k) of sets of indices is dened as follows:
S(k)=

(i, j) | 1 i kD, (k 1)D +1 j kD or

1 j kD, (k 1)D +1 i kD

, k = 1, 2, . . .
260 A. Bhaya et al. / Linear Algebra and its Applications 332334 (2001) 257263
Fig. 1. The sets S(k) and (k).
and a sequence of sets of boundary indices as follows (see Fig. 1):
(k) =

(i, 0), (0, j) | (k 1)D +1 i, j kD

, k = 1, 2, . . .
For k = 1, 2, . . . , dene
(k) = max
(i,j)S(k)
x(i, j)
v
,
(k) = max
(i,j)(k)
x(i, j)
v
and
c(k) = max

(k), (k), (k +1)

.
Let (i
0
+1, j
0
+1) be an index in S(k +1). From (6), it follows that (i
0
+1, j
0

d
2
(i
0
, j
0
)) and that (i
0
d
1
(i
0
, j
0
), j
0
+1) are indices in S(k +1) S(k) (k +
1) (k). Suppose that for all (i, j): (i, j) S(k +1) and (i, j) < (i
0
+1, j
0
+1),
inequality x(i, j)
v
c(k) holds. Then
|x(i
0
+1, j
0
+1)|
=

Ax(i
0
+1, j
0
d
2
(i
0
, j
0
)) +Bx(i
0
d
1
(i
0
, j
0
), j
0
+1)

Ax(i
0
+1, j
0
d
2
(i
0
, j
0
)) +Bx(i
0
d
1
(i
0
, j
0
), j
0
+1)

|A|

x(i
0
+1, j
0
d
2
(i
0
, j
0
))

+|B|

x(i
0
d
1
(i
0
, j
0
), j
0
+1)

c(k)|A|v +c(k)|B|v
c(k)v,
which is equivalent to
x(i
0
+1, j
0
+1)
v
c(k).
So, by induction on S(k +1) in both indices i and j, it follows that for all (i, j)
S(k +1), x(i, j)
v
c(k), i.e.,
(k +1) c(k). (8)
A. Bhaya et al. / Linear Algebra and its Applications 332334 (2001) 257263 261
Now, without loss of generality, suppose that (k +1) (k). Otherwise, one can
use
(k) = max{ (k), (k +1), . . .}
to replace (k). Thus, (8) can be written as
(k +1) max{(k), (k)}. (9)
Expanding (9) for k = 1, 2, . . . , gives
(1) = (1);
(2) max{(1), (1)};
(3) max

2
(1), (1), (2)

;
.
.
.
(k +1) max

k
(1),
k1
(1), . . . , (k 1), (k)

;
.
.
.
For any arbitrarily xed > 0, there exists some K (which depends on ) such that
(k) < ,
k
< for all k K,
and therefore, for all k 2K 1,
(k +1) max

k
(1),
k1
(1), . . . ,
K
(k K),

K1
(k K +1), . . . , (k)

max{(1), (1), }
= max{(1), (1), 1}.
So
lim
k
(k) = 0.
Some remarks on Theorem 2.1 are as follows.
Remark 2.3. Condition (7) is a generalization of the ChazanMiranker condition
for 1D systems with variable shifts (also called partial asynchronism) [3,6] to the 2D
case. The condition is also necessary in the following sense: if B = 0 and (|A|)
1, then there exists a 2D system (with a specially chosen sequence of shifts) which is
not asymptotically stable. Note that, since the stability condition derived in this paper
(condition (2) and/or (7)) actually coincides with the stability condition derived in
[18] for the system without shifts, it can be concluded that it is a very conservative
stability condition for the latter.
Remark 2.4. Another interpretation of Theorem 2.1 is that the iterative process
described by Eq. (1) with C = 0, under condition (7) will converge even in the cases
262 A. Bhaya et al. / Linear Algebra and its Applications 332334 (2001) 257263
in which the evaluation of a next state may depend not only on the closest states
in the (i, j)-plane, but also on states within some bounded neighborhood of the cur-
rent state. Alternatively, it may be afrmed that, under condition (7), the stability of
the zero solution is robust, in the sense that errors in the exact previous states are
tolerable.
Remark 2.5. The following BCs are found in the literature:
The nite memory BC:
x(i, 0) = x(0, j) = 0 for i, j large enough, (10)
which is used in most papers related to 2D models.
The nite sum BC:

i
x(i, 0) < + and

j
x(0, j) < +, (11)
which has been used to prove asymptotic stability using a Lyapunov function along
the diagonal lines of the (i, j) lattice. Obviously, the nite memory BC is the strong-
est one, since it implies the nite sum BC and implies, in turn, the asymptotic stabil-
ity BC assumed in this paper.
Remark 2.6. Theorem 2.1 is also valid for the stability of the more general shifted
2D system:
x(i +1, j +1)=Q

Ax(i d
1
(i, j), j d
2
(i, j))
+Bx(i d
3
(i, j), j d
4
(i, j))
+Cx(i d
5
(i, j), j d
6
(i, j))

,
under condition (2) and also requiring all shifts to be bounded.
3. Concluding remarks
This paper introduced a variant of the FM 2D dynamical system with variable
shifts. Stability was proved with a general BC. The result obtained here also gener-
alizes the well-known ChazanMiranker result for 1D systems (in the case of partial
asynchronism) to the case of 2D systems.
Acknowledgements
This research was partially supported by grants from CNPq, the Brazilian Na-
tional Council for Scientific and Technological Development, and PRONEX/MCT.
This research was carried out when Yangfeng Su was visiting the Department of
A. Bhaya et al. / Linear Algebra and its Applications 332334 (2001) 257263 263
Electrical Engineering at COPPE, Federal University of Rio de Janeiro as a CNPq
research fellow.
References
[1] D. Amitai, A. Averbuch, M. Israeli, S. Itzikowitz, E. Turkel, A survey of asynchronous nite-differ-
ence methods for parabolic PDEs on multiprocessors, Appl. Numer. Math. 12 (1993) 2745.
[2] A. Berman, B. Plemmons, Nonnegative Matrices in the Mathematical Sciences, Classics in Applied
in Applied Mathematical Sciences, SIAM, Philadelphia, PA, 1994.
[3] D.P. Bertsekas, J.N. Tsitsiklis, Parallel and Distributed Computation: Numerical Methods, Pren-
tice-Hall, Englewood Cliffs, NJ, 1989.
[4] T. Bose, D.A. Trautman, Twos complement quantization in two-dimensional state-space digital l-
ters, IEEE Tran. Signal Process. 40 (10) (1992) 25892592.
[5] R. Bru, L. Elsner, M. Neumann, Models of parallel chaotic iteration methods, Linear Algebra Appl.
103 (1988) 175192.
[6] D. Chazan, W.L. Miranker, Chaotic relaxation, Linear Algebra Appl. 2 (1969) 190222.
[7] L. Elsner, I. Koltracht, M. Neumann, Convergence of sequential and asynchronous nonlinear para-
contractions, Numer. Math. 62 (1992) 305319.
[8] E. Fornasini, A2-Dsystems approach to river pollution modelling, Multidimensional Systems Signal
Process. 2 (1991) 233265.
[9] E. Fornasini, G. Marchesini, State-space realization theory of two-dimensional lters, IEEE Trans.
Automat. Control AC-21 (4) (1976) 484491.
[10] E. Fornasini, G. Marchesini, Doubly-indexed dynamical systems: state-space models and structural
models, Math. Systems Theory 12 (1978) 5972.
[11] E. Fornasini, G. Marchesini, Stability analysis of 2-D systems, IEEE Trans. Circuits and Systems 27
(12) (1980) 12101217.
[12] T. Hinamoto, Stability of 2-D discrete systems described by the FornasiniMarchesini second model,
IEEE Trans. Circuits and Systems I: Fundamental Theory and Applications 44 (3) (1997) 254257.
[13] E. Kaszkurewicz, A. Bhaya, D.D. iljak, On the convergence of parallel asynchronous block-itera-
tive computations, Linear Algebra Appl. 131 (1990) 139160.
[14] D. Liu, A.N. Michel, Dynamical Systems with Saturation Nonlinearities, Springer, London, 1994.
[15] W.-S. Lu, On a Lyapunov approach to stability analysis of 2-D digital lters, IEEE Trans. Circuits
and Systems I: Fundamental Theory and Applications 41 (10) (1994) 665669.
[16] W. Marszalek, G.T. Kekkeris, Heat exchangers and linear image processing theory, Int. J. Heat Mass
Transfer 32 (12) (1989) 23632374.
[17] R.P. Roesser, A discrete state-space model for linear image processing, IEEE Trans. Automat. Con-
trol AC-20 (1975) 110.
[18] Y. Su, A. Bhaya, On the BoseTrautman condition for the stability of 2-D linear models, IEEE Trans.
Signal Process. 46 (7) (1998) 20692070.
[19] Y. Su, A. Bhaya, E. Kaszkurewicz, V.S. Kozyakin, Further results on the stability of asynchronous
discrete-time linear systems, Linear Algebra Appl. 281 (13) (1998) 1124.
[20] A. resin, M. Dubois, Asynchronous iterative algorithms: models and convergence, in: L. Kronsj,
D. Shumsheruddin (Eds.), Advances in Parallel Algorithms, Oxford, London, 1992, pp. 302342
(Chapter 10).