You are on page 1of 6

965

2D SYSTEMS THEORY - A COhIMON APPROACH TO PROBLEMS IN CIRCUITS,


CONTROL AND SIGNAL PROCESSING

E. Rogers*, K. Galkowski**, D.H. Owens***and N.Amann***


* University of Southa.mpton, U.K.
** Technical University of Wroclaw, Poland
*** University of Exeter, U.K.

Ab st ract 2 Basic nD Systemts Theory

The main aim of this paper is to give an overview of Despite the diversity of applications areas, a very large
2D systems theory, with particular emphasis on its role volume of literature exists on basic nD systems research
in control, circuits and signal processing applications. and is continually being added to. This often requires
mathematical tools outside those required in standard
1 Introduction (1D) linear systems theory. Space limitations obviously
prevent a comprehensive treatment of these techniques.
The past two to three decades, in particular, have Instead we give a summary of the main ones placing
seen a continually growing interest in so-called two- emphasis on where essential differences with the 1D
dimensional (2D) or , more generally, multi-dimensional case arise.
(nD) systems. This is clearly related to the wide vari- At an “abstract” level nD systems theory setss out
ety of applications of both practical and/or theoretical to examine the same basic questions as 1D theory, eg
interest. The key unique feature of an nD system is controllability, observability, causality, construction of
that the plant or process dynamics (input, output and state space models (realisation theory), stability and
state variables for example) depend on more than one stabilisation, feedback control, filtering. The basic
indeterminate and hence information is propagated in reason why generalising 1D results (generally) fails is
many independent directions. due to strong mathematical difficulties with the analysis
Many physical processes have a clear nD structure. tools employed. Also, as already noted, there are many
Also the nD approach is frequently used as an analysis key issues associated with nD systems which have no
tool to assist, or in some cases enable, the solution of 1D counterparts.
a wide variety of problems. A key point is that the Consider now the 2D linear systems case where trans-
applications areas for nD systems theory/ engineering fer function representations are to be used as the ana-
can be found within the general areas of circuits, control lysis base. Then difficulties immediately arise here due
and signal processing (and many ot,hers). to the complexity of t,he underlying ring structure, ie
An obvious approach to, for example, the control re- functions in two indeterminates where the underlying
lated analysis of nD systems, particularly at the con- ring does not have a division algorithm. The existence
ceptual level where some degree of similarity is often of a division algorithm for Euclidean rings forms the
apparent, is to simply extend standard, ie ID, tech- basis for the algorithmic derivation of’ many canonical
niques. This approach is, in general, incorrect since forms and solution techniques at the heart of 1D sys-
many common 1D techniques do not generalise. Also tems theory, eg the Smith form and the solution of 1D
there are many nD systems phenomena which have no polynomial equations.
1D systems counterparts. The main aims of this paper In the 1D case, coprimeness of polynomial matrices
are to : is a key analysis tool. There are three forms of coprime-
ness for an nD polynomialmatrix, termed factor, minor
(i) the differences between standard and zero respectively. These are all equivalent in the
(1D) and nD systems with particular emphasis on
1D case, but for the 2D case only minor and factor
the mathematical tools required;
coprimeness are equivalent. For the nD case ( n 2 3)
(ii) survey progress in a number of key areas which none of these concepts are equivalent!
stretch across the general areas of circuits, control As a simple example in the 2D case, consider the
and signal processing; and polynomials

(iii) to identify some key currently open research prob- z2) z1 - ’, b ( z 1 ,z 2 ) = 22 - 1


lems. These two polynomiails are factor coprime but have

UKACC International Conference on CONTROL ‘96,2-5


September 1996, Confereince Publication No. 427 0 IEE 1996
966

common zeros. Also the investigation of the stability ability of the state space model. This is another key
of 2D systems using the 2D transfer function descrip- area which does not generalise from 1D to nD where
tion is greatly influenced by this situation as we discuss only partial results are yet known.
next. Key results in feedback control theory for linear 1D
In simple terms, the difficulty is that the numerator systems are characterised by, for example, open loop
polynomial open loop can directly influence stability! controllability, eg the pole assignment problem. The
This key result was first reported by Goodman [I]after situation is again much more complex in the nD sys-
a considerable volume of literature had appeared on sta- tems case, where no such “crisp” theory exists! Gen-
bility tests based on 2D transfer function descriptions. erally speaking, the development of a rigorous control
To illustrate this point further,consider the following theory, and associated controller design algorithms, for
transfer functions in the discrete case 2D/nD systems is much less well developed than the
theory/ design of 2D/nD filters. Some aspects, such as
optimal control, have (not unexpectedly perhaps!) re-
ceived more attention than others. Substantial progress
has been made, however, for some applications driven
sub-classes where “the physics of the process involved”,
and hence simplifications in modelling, have been fully
where, in keeping with 2D systems convention, 21 and
exploited. See also the next section for further discus-
z2 are backward shift operators. These two transfer
sion of this obviously very important general a,rea.
functions have the same denominator but GI is stable
In the 1D case, similarity transforms applied to the
and G2 is not. The key point here is that these two
state vector have a key theoretical role. This is again
transfer functions have a non-essential singularity of
not true for nD systems where, for example, the prob-
the second kind at 21 = 22 = 1. There is no 1D ana-
lem of obtaining all possible state space realisations of
logue of these singularities!
a given state space model cannot be solved in a sim-
The a priori information available and the modelling
ilar manner. Also the synthesis problem which is at
objectives permit the choice of different model struc-
the heart of circuit theory and applications has not, yet
tures to describe 2D or nD systems. As a basic start-
been completely solved in the nD case.
ing point, these representations can be classified ac-
cording to whether or not (i) an input/output struc- Causality is a key feature of classical dynamical sys-
ture is included, and (ii) latent (or auxiliary) variables tems (it must be present for physical realisability). In
are included. This general area is discussed further in the 1D case, the general concept of “time” imposes a
Rocha[2] and the relevant references in this thesis. natural ordering into “past”, “present” and “future”.
As in the 1D case, state space models are a very im- Again the situation is different in the nD case due, in
portant class of internal representations. In this con- effect, to the fact that some of the indeterminates have
text,the concept of the state of a system can be defined a spatial rather than a temporal characteristic. Hence
(obviously) as the memory of the system, ie the past causality is not required since it is only necessary to be
and future evolutions are independent given the current able to recursively perform the required computations.
state. Hence the concept of a state depends on which or- One of the major consequences of this is that singu-
dering is considered on the underlying grid. For the dis- larity is of increased importance in the nD case. Also
crete case this is normally the product order defined as some of the spatial variables in an nD system can be
“naturally bounded” , ie they have bounded regions of
support. This, in turn, can lead to difficulties in formu-
lating a rigorous but practically implementable stabil-
ity theory. Often, however, it is possible to introduce
simplified stability definitions which (if they are applic-
Commonly used models for linear systems recursive able to a given problem) are easier to check (in relative
in the positive quadrant are the Roesser model [3] and terms) [5].
the Fornasini-Marchesini models [4]. The structure of
these models (and others) have been extensively invest- 3 Progress and Open Research Problems
igated. One of the most striking differences with the
1D case is the need to consider both a global, X , and The first significant work on nD systems appeared in
a local, 2 , state. Basically, the global state in a diag- the early 1960’s in the general area of circuit, analysis.
onal line L k , denoted xk , is defined as the collection of For example, Ansell [6] showed how positive real func-
all local states along L k . Concepts such as reachability tions in two or more variables could be used to study
and observability are then defined at both global and networks of transmission lines (distributed parameter
local levels and they can be quite distinct properties! elements and lumped reactances). The use of this the-
Minimality in the 1D linear case is completely charac- ory in studying transmission lines (and other applica-
terised by reachability (or controllability) and observ- tions) suggested that it may also be applicable to sys-
967

tems with delays. Significa.nt early developments on the approaches is the book by Dudgeon and Mersereau [12].
application of nD systems theory to time delay systems Design tools for multi-dimensional filters is an area
(and vice versa) are reviewed in the survey paper [7]. in which much work has been done and the filters im-
Following this early work, great interest was generated plemented! Also the literature continues to grow. The
in t.he synthesis of 2D and nD positive real functions or general area of control theory/ controller design for 2D/
mattrices under the usual circuits assumptions of passiv- nD systems is discussed at the end of this section.
ity, losslessness etc, and this general area is still open The general area of multi-dimensional filtering the-
in many respecm. ory remains a very active research topic. Prominent in
In parallel with these early circuit theory develop- this is work by Fettweis, see, for example, 1131 for sonir
ments, work began to appear on key systems theor- basic results, on the extension of classical analogue cir-
etic based properties of nD systems with the focus very cuit theory techniques (eg Kirchoff’s laws, passivity)
much on stability theory. For example, the classical together with wave filters to digital applications. This
Hurwitz theory was extended to the nD case. Follow- work holds much promise!
ing this a large volume of work appeared on stability In classical systems theory, one of the most iniport-
theory and tests,, much of it for 2D (and nD) linear sys- ant and powerful representations is, of course, the state
tems described in convolution form. A comprehensive space model (in its various forms). This general ap-
review of this work can be found in the survey paper proach was first employed in the general control area
by Jury[8]. and then moved to circuits. Hence it was to be expec-
A large numt’er of these tests in the discrete case are ted that a state space approach to nD systems (circuits,
based on examining the denominator polynomial of the control, signal processing) would emerge.
2D transfer function for the location of its “roots” rel- Focusing now on discrete 2D linear systems, two “in-
ative to the unii; bidisc, ie the natural generalisation of dependent” state space models duly emerged. The first,
the linear 1D case! The paper by Goodman [l] is very of these is credited to Roesser [3] and clearly has a first,
significant because it showed that in certain cases such order structure. Among the key features of t,his model
tests could produce an inconclusive result - return to is that the state vector is partitioned into sub vectors
the examples given in the previous section where it can - one for each of the two directions of information

be shown that the denominator polynomial (common to propagation (usually termed horizontal and vert,ical re-
both transfer functions) satisfies the stabilit,y test relat- spectively).
ive to the unit bidisc! This situation has no 1D counter- A key measure of the diversity of 2D/ nD systems
part and the sole reason for it is that GI and Ga have a over the 1D case, is the variety of state space descrip-
non-essential singularity of the second kind on the unit tions which can be used. One main alternative to
bidisc. To avoid this difficulty a large volume of 2D the Roesser model (certainly in terms of published re-
systems literature since 1977 assumes that such singu- search) are the Fornasini-Marchesini models and their
larities simply do not exist! Also there are continual variations/ generalisations [4]. Note, however, that
attempts to “properly characterise” these singularities, these model classes are not necessarily fully independ-
eg [91. ent and it is possible in some cases to transform from
Clearly the basic stability question for 2D systems is one to the other - see, for example, [14].
not as simple as the 1D case. There are, however, some It is straightforward to extend these state space mod-
classes of 2D systems where stability can be examined els from the 2D to nD case. Also they can be trans-
using standard LD tests! One such case is linear repet- formed to transfer function (matrix) representations us-
itive processes where such a theory has been developed ing the double z transform - a rigorous theory for
by Rogers and Owens [lo]. Recently this theory has this exists. Similarly, continuous versions of these
found use in the analysis and design of iterative learn- models can be defined where the corresponding trans-
ing control schemes [ll]and other applications (see also fer functions can be obtained by applying the multi-
below). The kej- point to note is that this success is to dimensional Laplace transform. Again in contrast to
a large extent due to inherent structural properties of the 1D case, it is necessary to consider models with a
this sub-class of 2D linear systems! so-called mixed structure. For example, Kaczorek [15]
The advances: in the general area of signal/ image has studied systems described by state space models
processing has (created many applications for, in par- which are discrete in one direction and continuous in
ticular, 2D and 3D filters. These include seismology, the other. Such models obviously have close links with
tomography and visual data communications applica- linear repetitive processes which have numerous prac-
tions. Early ana.lysis/ design tools where largely based tical applications.
on input/ output descriptions in the form of multi- As noted already here, a rigorous stability theory
dimensional difference equations with immediate links for these processes based on an abstract model in a
via the multi-dimensional z and Fourier transforms to Banach space setting has been developed [lo]. Also a
multi-dimensional transfer functions (or transfer func- cross section of the considerable volume of work already
tion matrices). .4 good starting point for studying such reported on the development of a, rigorous systems the-
968

ory (some of which exploits links with the Roesser and on which to base analysis is obviously a key question in
Fornasini-Marchesini state space models) can be found nD systems theory and one of many basic tasks is to at-
in the appropriate references in [lo] and, for example, tempt to characterise minimality. As noted previously,
[16, 171 - see also below. Recent work by Roberts, see there are several distinct definitions of controllability
for example [18, 191, reports an interesting new applic- and observability, say, for the Roesser model and hence
ation for repetitive processes. a characterisation based on a straightforward extension
In the 2D systems case, the state space represent- of the 1D case is unlikely (to say the least). This is in-
ations have prompted a huge volume of research on deed the case and substantial progress has only really
a wide range of systems theoretic questions. These been possible in certain special cases and even these
include controllability, observability, optimal control, do not yield simple tests [24]. Another approach to the
state feedback control, output feedback control. . . Quite realisation problem for nD systems is to study it, in a
simply, the literature on these aspects is far too large multi-dimensional polynomial framework [25]
to attempt any form of survey here. The key point to The construction of a range of state space realisations
stress again is that apart from very special cases (some for a 2D transfer function (filter) is very important in
of which have no real practical motivation) a straight- terms of applications since it is then possible (at least
forward generalisation of the 1D case is impossible. in principle) to “optimise” the structure used. For ex-
Another field for nD systems theory is processes ample, this can yield structures characterised by the
modelled by partial differential and/or difference equa- minimal number of specified circuit elements.
tions. Interesting applications reported to date include In the 1D case, it is only necessary to invoke the
river pollution modelling [20]. Such applications also concept of similarity and hence attention can be restric-
highlight another area of systems research which has ted to a subset of possible transformations in order to
attracted an increasing volume of attention in circuits, preserve specified key systems properties. For example,
control and signal processing which is briefly discussed use of orthogonal transformations preserves lossless-
next. ness, passivity and symmetry in the circuits case.
Often in modelling physical systems it is impossible Such problems are also of interest in nD systems
to construct standard equations. Instead only the so- where it is again natural to consider using similarity
called singular (or descriptor or implicit) form of the transformations. Early work by Sun-Yuan Kung et a1
state equations can be obtained, ie the derivative of the [26] treated this problem for the Roesser model and this
state vector is pre-multiplied by a square singular (or work was later extended to effect by Galkowski [14]
even rectangular) matrix. One implication of this is Recently Johnson and co-workers [23, 271 have pro-
that non-monic multi-dimensional polynomials arise in duced some new and extremely interesting results on
transfer function descriptions. This is well known in nD systems equivalence in a Rosenbrock system mat-
classic (1D) circuit theory but its importance increases rix setting. Galkowski, noting the difficulties with 2D
markedly in the nD systems case. Work on singular nD systems similarity, has developed some key results on
systems in a control systems framework can be found the use of 2D elementary operations to obtain various
in, for example, [21]. An important point to note is state space realisations. The so-called Elementary Op-
that here singularity may not be an intrinsic feature erations Algorithm has been developed [as]and the final
of the system but can, for example, be introduced by realisation matrix is obtained by applying these oper-
‘inadequate” identification of a model or by the dis- ations to the nD polynomial matrix (matrices) in turn
cretisation methods used. In some cases, it is possible together with augmentation of its size.
to avoid singularity by applying well defined algebraic Use of this algorithm can yield standard and singular
operations [22]. realisations in both the single input/ output and mul-
In 1D systems, minimality (and other key proper- tivariable cases. Also it gives some insights into the
ties such as the existence of a stabilising state feedback basic structural relationships between the Roesser and
control law) is completely characterised by the prop- Fornasini-Marchesini state space models [as].
erties of controllability and observability of the state Finally, consider the key area of a rigorous gener-
space model. Also the Rosenbrock system matrix plays ally applicable control theory, and attendant control-
a central role in understanding and characterising key ler design algorithms, for 2D/nD systems. Then gen-
systems theoretic properties. Such analysis makes ex- erally speaking this area is much less well developed
tensive use of elementary matrices, a sub-class of un- when compared to the theory/ design of 2D/nD fil-
imodular matrices over a ring, where the key point is ters. Many key questions are still essentially open and
that all unimodular matrices can be constructed as the in some areas this can (arguably) be attributed to not
product of elementary matrices. The situation in the being able to “appropriately formulate” the problem,
nD systems case is much more complex due to the eg the existence problem for a stabilising state feed-
complexity of the underlying ring structure ([23] for back control law. In the 1D linear systems case, poles
example). have a well defined (mathematical and physical) mean-
Constructing an appropriate state space description ing and hence the basic “pole assignment” problem is
969

“well defined” in these terms! What, if anything, is will have met the basic objectives laid down at the start!
a “pole” for 2D/nD linear systems and what does it It has, however, obviously been impossible to cover all
mean? Once itgain we see that even at the most basic aspects or contributions. For example, the recently ad-
level 1D resuhs and concepts do not (except in special vocated behavioural approach to 2D systems analysis
cases) generalise to 2D/nD systems! [2] has not been fully covered but (arguably) it is too
A considerable volume (continually growing) of liter- early to even attempt to assess its advantages and dis-
ature exists o n various approaches to control theory/ advantages.
controller design for 2D/nD systems. Optimal control
based strategies continue to receive much attention by,
for example, 13isiacco and co-workers [30]. Also there References
has been some very promising work begun to appear
[l] D.S. Goodman. Some stability properties of
on robustness issues (mainly stability) - see, for ex-
two-dimensional linear shift invariant digital fil-
ample, the special issue in Multidimensional Systems
ters. IEEE Trans. Circuits and Systems, CAS-
and Signal Processing, guest edited by Kogan [31].
24(4):201-208, 1977.
Some work has also begun to appear [32] on the ex-
tension of the well known and extensively used mat- [a] P. Rocha. Structure and Representation of 2-0
rix factorisatim based approach for 1D systems ana- Linear Systems. PhD thesis, University of Gronin-
lysis and design to the 2D/nD case. Note that this gen, The Netherlands, 1990.
work is based on achieving the weaker condition of so-
called practical stability [5] closed-loop. It is expected, [3] R.P. Roesser. A discrete state-space model for lin-
however, that this “restriction” can be removed. ear image processing. IEEE Trans. Auto. Contr.,
It is simply not possible to even attempt a more gen- 20:l-10, 1975.
eral survey of work on 2D/nD systems control theory/
[4] E. Fornasini and G. Marchesini. Double indexed
controller design in the space available here. The clear
dynamical systems : State models and structural
message is t h d much remains to be done, significant
properties. Mathematical Systems Theory, 12:59-
parts of it at i t very basic level relative to the 1D case.
7 2 , 1978.
Contrast this with the general 2D/nD filter case where
design techniciues are well established and some of the [5] P. Agthoklis and L. Bruton. Practical-BIB0 sta-
filters have been implemented! This is still a very active bility of n-dimensional discrete systems. Proceed-
research area --for example, the August 1995 issue [33] ings of the IEE, G:236-242, 1983.
of The IEEE Trans on Circuits and Systems 11: Ana-
logue and Digital Signal Processing contains 4 papers [6] H.G. Ansell. On certain two-variable generaliz-
in this general area. An obvious question to ask is: are ations of circuit theory, with applications to net-
any of these iesults/ techniques of use in the control works of transmission lines and lumped react-
case? ances. IEEE Trans. on Circuit Theory, CT-
Very substantial progress has been made on the de- 11(2):214-223, 1964.
velopment of a control theory/ controller design al- [7] E.D. Sontag. Linear systems over commutative
gorithms for special cases of 2D linear systems. For rings: A survey. Ricerche di Automatica, 7:1-34,
example, a feedback control theory for linear repetitive 1976.
processes based, in effect, on extending/ augmenting
well known lll schemes is emerging [34, 351. Also Well- [8] E.I. Jury. Stability of multidimensional scalar
stead and co-workers - see, for example, [36, 371 have and matrix polynomials. Proc. of the IEEE,
produced some very interesting results on self-tuning 66(9):1018-1047, 1978.
prediction ancl control for a class of 2D linear systems
with applications to a paper making plant. [9] S.A. Yost and P.H. Bauer. Shift invariant m-D
systems and singularities on T m :Implications for
robust stability. IEEE Trans. Circuits and Sys-
4 Conclusions tems, CAS-42(8):477-479, 1995.
Clearly substantial progress has been made on answer- [lo] E. Rogers and D.H. Owens. Stability Analysis for
ing key systeins theoretic questions for nD systems, Linear Repetitive Processes, volume 175 of Lec-
many of which do not arise in the 1D case! There are, ture Notes in Control and Information Sciences.
however, many, as yet, open or only partially answered Springer-Verlag , 1992.
questions - some of them at a very basic level (in
comparison to the 1D case). The increasing number of [ll]N.Amann, D.H. Owens, and E. Rogers. Iterative
applications areas serves to support increased research learning control using optimal feedback and feed-
in t,his general area. forward actions. International Journal of Control,
Hopefully this paper, plus the supporting references, 1996. Accepted for publication.
970

[12] D.E. Dudgeon and R.M. Mersereau. Multidimen- [25] K. Galkowski. The state-space realization of an
sional Digital Signal Processing. Prentice-Hall, n-dimensional transfer function. Int. Journal of
Englewood Cliffs, 1984. Carcuat Theory and Applns, 9:189-197, 1981.
[13] A. Fettweis. Simulation of hydromechanical par- [26] S . - Y . Kung, B.C. Levy, M. Morf, and T. Kailath.
tial differential equations by discrete passive dy- New results in 2D systems theory, pt 11: 2D state-
namical systems. In Kimura, H. et all editor, space models - realizations and the notions of
Recent Advances in Mathematical Theory of Sys- controllability, observability and minimality. Pro-
tems, Control, Networks and Signal Processing II, ceedings of the IEEE, 65(6):945-961, 1977.
pages 489-494. Mita Press, 1992.
[27] D.S. Johnson, E. Rogers, A.C. Pugh, G.E.
[I41 K. Galkowski. The Fornasini-Marchesini and the Hayton, and D.H. Owens. A polynomial matrix
Roesser model - recasting by similarity. IEEE theory for a certain class of 2-D linear systems.
Trans. Auto. Contr., AC-41:107-112, 1996. To appear in Linear Algebra and Its Applications,
[15] T. Kaczorek. 2-D continuous-discrete linear sys- 1996.
tems. In Proc. Tenth Int. Conf. on System Engin- [28] K. Galkowski. Elementary operations and equi-
eering ICSE'94, Coventry, 6-8 Sept., volume I, valence of 2-D systems. Int. Journal of Control,
pages 550-557, 1994. 6311129-1148, 1996.
[16] E. Rogers and D.H. Owens. Stability of linear re- [29] K . Galkowski. Elementary operation approach to
petitive processes - a delay-differential systems state-space realization of 2-D systems. Submitted
approach. IMA Journal of Mathematical Control to IEEE Trans. on Circuik and Systems, 1996.
and Information, 12(1):69-90, 1995.
[30] M. Bisiacco. New results in 2D optimal con-
[17] D.H. Owens and E. Rogers. Frequency domain trol theory. .Multidimensional Systems and Signal
Lyapunov equations and performance bounds for Processing, 6(3):189-222, 1995.
differential linear repetitive processes. Systems
and Control Letters, 26:65-68, 1995. [31] J. Kogan (guest editor). Robustness of multidi-
~~

mensional systems and signal processing. Multi-


[18] P.D. Roberts. Unit memory linear repetitive pro-
dimensional Systems and Signal Processing, 5(4),
cesses and iterative opt,imal control algorithms. In
1994.
Proc. Int. Conf. on Control 94, . pages
-
454-459,
Steveage, UK, 1994. IEE. [32] L. Xu, 0 . Saito, and K. Abe. The design of prac-
[19] P.D. Roberts. Unit memory repetitive process tically stable nD feedback systems. Automatica,
aspects of iterative optimal control algorithms. 30(9):1389-1397, 1994.
In PrOC 33rd ~ElLY Conference on Decision and [33] IEEE Trans. Circuits and Systems 11: Analog und
Control, pages 1394-1399, Piscataway, NJ, 1994. Digital Signal Processing, 42, August 1995.
IEEE.
[34] E. Rogers and D.H. Owens. Error actuated out-
[20] E. Fornasini. A 2-D systems approach to river put feedback control theory for differelltial lin-
pollution modelling. Multi-dimensional Systems ear repetitive processes. Int. Journal of Control,
and Signal Processing, 2:233-265, 1991. 65(5):981-998, 1995.
r211 F.L. Lewis. A review Of 2-D
[35] D.H. Owens and E. Rogers, I D controllers for a
Automatica, 28(2):345-254, 1992.
class of 2D linear systems with guaranteed per-
[22] K. Galkowski. Ti-ansformation of the transfer formance. Submitted to Automatica, 1996.
function variables of the singular n-dimensional
[36] W.P. Heath and P.E. Wellstead. Self-tuning pre-
Roesser model. Int. Journal of Circuit Theory and
diction and control for two-dimensional processes
Applns, 20:63-74, 1992.
part 1: Fixed parameter algorithms. Int. Journal
[23] D.S. Johnson. Coprameness in Multidimensional of Control, 62(1):65-108, 1995.
Systems and Symbolic Computation. PhD thesis,
Loughborough University of Technology, March [37] W.P. Heath and P.E. Wellstead. Self-tuning pre-
1993. diction and control for two-dimensional processes
part 2: Parameter estimation, set-point track-
[24] S.H. Zak, E.B. Lee, and W.S. Lu. Realizations ing and offset handling. Int. Journal o,f Control,
of 2-D filters and time delay systems. IEEE 62(2):239-270, 1995.
Trans. Circuits and Systems, CAS-33(12):1242-
1243, 1986.

You might also like