Are you sure?
This action might not be possible to undo. Are you sure you want to continue?
Original Russian Text c E.N. Gryazina, B.T. Polyak, A.A. Tremba, 2008, published in Avtomatika i Telemekhanika, 2008, No. 12, pp. 3–40.
REVIEWS
Ddecomposition Technique Stateoftheart
1
E. N. Gryazina, B. T. Polyak, and A. A. Tremba
Trapeznikov Institute of Control Sciences, Russian Academy of Sciences, Moscow, Russia
Received December 24, 2007
Abstract—It is a survey of recent extensions and new applications for the classical Ddecompo
sition technique. We investigate the structure of the parameter space decomposition into root
invariant regions for singleinput singleoutput systems linear depending on the parameters. The
Ddecomposition for uncertain polynomials is considered as well as the problem of describing
all stabilizing controllers of the certain structure (for instance, PIDcontrollers) that satisfy
given H
∞
criterion. It is shown that the Ddecomposition technique can be naturally linked
with M−Δ framework (a general scheme for analysis of uncertain systems) and it is applicable
for describing feasible sets for linear matrix inequalities. The problem of robust synthesis for
linear systems can be also treated via Ddecomposition technique.
PACS number: 02.30.Yy
DOI: 10.1134/S0005117908120011
1. INTRODUCTION
Consider a linear system depending on a vector parameter k with a characteristic polyno
mial a(s, k). The boundary of a stability domain (in the space k) is given by the equation
a(jω, k) = 0, ω ∈ (−∞, +∞), (1)
that is the imaginary axis (the boundary of instability in the root plane) is mapped into the
parameter space. If k ∈ R
2
(or k ∈ C) then we have two equations (real and imaginary part of (1))
in two variables and (in general) can deﬁne the parametric curve k(ω), −∞ < ω < ∞ deﬁning a
boundary of the stability domain. Moreover, the curve k(ω) divides the plane into root invariant
regions (i.e., regions with a ﬁxed number of stable and unstable roots of a(s, k)). This is the basic
idea of Ddecomposition approach. The idea can be traced to Vishnegradsky [85] who reduced a
cubic polynomial to the form a(s, k) = s
3
+ k
1
s
2
+ k
2
s + 1 and treated the coeﬃcients k
1
, k
2
as
parameters. Then Eq. (1) yields k
1
ω
2
= 1, ω(k
2
−ω
2
) = 0. Eliminating ω we get that Ddecompo
sition is given by the hyperbola k
1
k
2
= 1. The stability domain is the set k
1
k
2
> 1.
For the general case, similar ideas were exploited in [1, 27, 51] (the latter two papers deal with
timedelay systems). Moreover, Nyquist plot can be considered as the realization of the same
idea. But it was Yu. Neimark [13, 14] who developed the rigorous algorithm (and coined the name
“Ddecomposition”). In the Western literature the technique is described ﬁrst in [68]; the mapping
of contours other that imaginary axis was also proposed. This line of research was signiﬁcantly
developed by Siljak [80–82]. He extended the approach for nonlinear systems and for the case
of nonlinear parameter dependence. In his works Ddecomposition (which he calls the parameter
plane method) was broadened to become a useful tool for design purposes. Neimark’s method is also
1
This work was supported by the Russian Foundation for Basic Research, project nos. 080800371, 050801177
and Presidium RAS Programm no. 22. The preliminary version was presented at IX Chetaev Conference in the
12–16 of July 2007.
1991
1992 GRYAZINA et al.
investigated in [63], where some results on multiparameter cases can be found. The Ddecompo
sition technique is described in Soviet textbooks on automatic control, in the western literature
it can be found in [32, 38]. The papers [33, 39, 52, 61] show its high eﬃciency for the problems
of loworder controllers with H
∞
speciﬁcations. The analysis of the geometrical properties of the
Ddecomposition is given in [17–19].
In the 80s within the appearance of the problems with uncertainty (robust control) the
Ddecomposition seems to be eﬃcient for analysis of the families of linear systems [15]. Within
the stability investigation there is a question of interest: if the system remains stable since its
parameters are unknown constants belonging to a particular set. These parameters are called “un
certainties.” The most famous results in this ﬁeld are the Kharitonov theorem and its analogs
concerning the robust stability of interval polynomials as well as Tsypkin–Polyak locus for interval
and ellipsoidal restrictions on the polynomial coeﬃcients; see [21, 24, 30].
The Ddecomposition technique can be applied to the systems in state space form. More specif
ically, given a class / of r m matrices K, ﬁnd all the matrices K ∈ / such that A + BKC is
stable:
D = ¦K ∈ / : A + BKC is stable¦. (2)
Here A, B, C are given real matrices of dimensions n n, n r, m n respectively; stability is
understood either in continuoustime sense (all eigenvalues are in the open LHP) or discretetime
sense (all eigenvalues are in the open unit disc). Class / may be diﬀerent; below we analyze in
detail the simplest cases:
K = k ∈ R
n
or K = k
T
, k ∈ R
n
( m = 1 or r = 1), (3)
K = kI, k ∈ R or k ∈ C, m = r, (4)
K ∈ R
2×2
, (5)
where all calculations can be performed explicitly in the graphical form. Case (3) is equivalent to
the polynomial framework, two others are essentially matrix ones. Nevertheless we present general
description of Ddecomposition. It is closely related to the standard M −Δ setting.
Problem (2) arises in the design or robustness studies. For instance, to ﬁnd all stabilizing static
output controllers for the system
˙ x = Ax + Bu, y = Cx (6)
one can construct the set D (2) with / = R
r×m
; here K plays the role of the feedback gain. On
the other hand, if A is a nominal stable matrix and it is perturbed as A + BKC, where K is a
constant r m matrix; then (2) provides all admissible perturbations which preserve stability. It
is obvious that if we know the boundary of the stability domain ∂D, then we can ﬁnd the distance
to it:
ρ = min
K∈∂D
K. (7)
The quantity ρ
−1
is closely related to μ (structured singular value) [86]. If / is the set of all C
r×m
(R
r×m
) matrices, then ρ is complex (real) stability radius [55, 75]. Of course, the knowledge of
the entire set D provides much more information than the value of ρ. For instance, for design
purposes a designer can solve performance or speciﬁcation problems on the set of the stabilizing
controllers D.
Instead of direct optimization on stability domain D we can get more information on the roots of
the characteristic polynomial inside D by similar graphical tools. For instance, if one is interested
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
DDECOMPOSITION TECHNIQUE STATEOFTHEART 1993
in the stability degree σ > 0, then the regions of interest can be obtained replacing jω by −σ +jω
in the Ddecomposition equation. This approach has been proposed by Neimark [14] and it was
signiﬁcantly developed by Mitrovic [68] and Siljak [80–82]. Of course, we can deal with contours
other than shifted imaginary axis, for instance, with a sector or the other root placement. This
information on the root placement inside the stability domain is very helpful for the design purposes.
However we will not develop this line of research and deal with the simplest situations of the root
placement—the left halfplane (for continuoustime systems) and the unit disc (for discretetime
systems).
Additional detailed decomposition can be a subject of future research. Several examples below
illustrate how the technique can be extended to become a practical tool for a designer. On the
other hand, we restrict ourselves with a presentation of basic results of the proposed approach and
do not consider practical applications.
The natural question arises: what is the use of ﬁnding all root invariant domains, while in
practice we are interested in the stability domain only? The ﬁrst answer to the question is very
simple—we are unable to construct the stability domain separately, the only way suggested by
the proposed technique is to provide the complete decomposition of the parameter space and then
to pick out the stability regions (if they exist). On the other hand there are some situations
when eigenvalue invariant domains are also of interest. For instance, dealing with Nyquist diagram
under uncertainty one should guarantee that all transfer functions under consideration have the
same number of stable poles. In some cases, we can propose the method for distinguishing stability
regions among all the root invariant regions; see Section 2.2 for details.
In 1991 the problem of robust Ddecomposition was formulated and partially solved in [20].
In general, the problem is as follows: ﬁrst we select two real (or one complex) parameters, the
rest of the parameters are assumed to be uncertain but bounded by norm. The problem is to
describe the region in the plane of the selected parameters such that the polynomial is stable
for all feasible parameter values. This division for two parameter classes is very typical for the
problem of robust ﬁxedstructure controllers. The paper [28] contains analytical expressions of the
robust Ddecomposition for two real parameters for an aﬃne polynomial family with l
p
bounded
parametric uncertainties.
Another new application of the Ddecomposition technique is the theory of linear matrix in
equalities [2, 41]. For this problem statements, the technique allows deﬁning all the regions in the
parameter space such that the aﬃne family of symmetric matrices has a ﬁxed number of likesign
eigenvalues inside a region [25].
Finally, the randomization technique inside the stability domain seems promising for optimal
controller synthesis [74].
The rest of the paper is organized as follows. In Section 2 we consider the rankone case (i.e.,
singleinput or singleoutput systems), where the stability is determined by the location of the roots
of the characteristic polynomial.
Section 2.1 is addressed for the classical Ddecomposition for polynomials with one or two
real parameters. The main contribution here is the study of the Ddecomposition geometry. In
particular, we estimate the number of all root invariant regions as well as the number of simply
connected stability regions. Section 2.2 is refereed to the problem of aperiodic stability.
Section 3 is devoted to characteristic polynomials with aﬃne uncertain of a certain structure
and the arbitrary number of parameters. The stability domain for this polynomials is a union of
polytopes in the parameter space. We suggest the method to describe the stability components
and to calculate the stability radius for various norm of uncertainty for continuous and discrete
systems.
Section 4 contains the results on the robust Ddecomposition.
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
1994 GRYAZINA et al.
Section 5 describes the problem of the ﬁxedorder controller synthesis with H
∞
speciﬁcations.
We consider linear timeinvariant system with a plant given by scalar transfer function G(s) and
feedback controller C(s, k). The controller parameters k ∈ R
m
, m 3 deﬁne the feasible set of
loworder controllers. The structure of the controller is ﬁxed. The problem is to describe the set
of the parameters corresponding to all the controllers such that H(s, k)
∞
< γ, where H(s, k) is
a transfer function for the closedloop system.
In Section 6 we extend the Ddecomposition idea to the matrix case. Namely, we consider the
systems with matrix transfer functions. The ﬁrst subsection relates to the case, where a graphical
representation is possible. The number of parameters is restricted by two.
Section 7 is devoted to the application of the Ddecomposition in the theory of linear matrix
inequities. For a small number of parameters (1 or 2) on the line or on the plane we plot the regions
such that the aﬃne family of symmetric matrices has a ﬁxed number of likesign eigenvalues inside
a region.
In Section 8 it is shown that the Ddecomposition can be applied in the randomized algorithms
for hard problems of synthesis. Combined with HitandRun algorithm it allows us to generate
points asymptotically uniformly distributed on a multidimensional domain (generally speaking
nonconvex and not simply connected). In particular, it may be a stability domain.
Section 9 contains conclusive remarks.
2. DDECOMPOSITION FOR POLYNOMIALS
Consider a linear system with the characteristic polynomial of the form:
a(s, k) = a
0
(s) +
m
i=1
k
i
a
i
(s), (8)
where a
i
(s), i = 0, . . . , m are polynomials of degree no more than n and the parameters k ∈ R
m
.
Deﬁnition 1. The decomposition of the parameter space k into the regions D
l
= ¦k : a(s, k) has
l stable roots¦, l = 0, . . . , n, is called the Ddecomposition. The equation describing the boundary
of regions D
l
is called Ddecomposition equations. Thus D
n
is the stability domain D.
We parametrize the boundary of the stability domain on the root plane: for continuoustime
systems it is the imaginary axis jω, ω ∈ (−∞, +∞), for discretetime systems it is the unit circum
ference e
jω
, ω ∈ [0, 2π). Then the number of stable roots may change in one of the following cases:
the degree of polynomials changes, one real or two complex roots cross the stability boundary on
the root plane. For the discretetime systems the diﬀerence is that the decreasing of the degree
does not lead to the change of the stable root number.
Thus the Ddecomposition is given by the equations:
a(jω, k) = 0, ω ∈ (−∞, +∞) or a
(n)
(k) = 0; (9)
a(e
jω
, k) = 0, ω ∈ [0, 2π), (10)
where a
(n)
is the leading coeﬃcient of the polynomial a(s, k). Further, we deal with polynomials
with real coeﬃcients. Since for continuoustime systems a(jω, k) = 0 leads to a(−jω, k) = 0 (sim
ilarly for discretetime systems a(e
jω
, k) = 0 leads to a(e
−jω
, k) = 0), the interval of ω may be
reduced to [0, ∞) (correspondingly [0, π]).
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
DDECOMPOSITION TECHNIQUE STATEOFTHEART 1995
2.1. Development of Graphical Algorithms
In the section we consider cases when the number of parameters does not exceed two. Parameter
space is the plane and the root invariant regions can be graphically represented.
Linearity of Ddecomposition equation allows one to ﬁnd easily the stability radius (7) for
various l
p
norms of the vector k. This link between robust stability and Ddecomposition has been
emphasized in [15]. On the other hand, the connection between μanalysis and the known results
on parametric robustness [21, 32, 35, 38, 62] has been pointed out by Chen [43]. Scalar constant gain
problem has been solved semianalytically in [71], the result was extended to compute stabilizing
controllers of a given order [76]. We will not highlight these links but will focus on the simplest cases
when Eq. (8) provides graphical tools to describe Ddecomposition in the space of parameters k.
The case of the singleinput singleoutput systems with small number of parameters is the
classical framework of the Ddecomposition.
In many cases we can state that the stability domain is simply connected. But there exist
examples [14, 18], where it is not true. Remind that the set D is simply connected if for any
K
0
∈ D, K
1
∈ D there exist the continuous parametrization K(t) such that K(t) ∈ D, 0 t 1,
K(0) = K
0
, K(1) = K
1
. It is not clear what is the maximal number of connected stability regions.
Moreover, the total number of the root invariant regions is also unknown.
For singleinput singleoutput systems with a scalar transfer function H(s) =
b(s)
a(s)
, where a(s),
b(s) are polynomials of degree n, with the feedback gain k and characteristic polynomial p(s) =
a(s) + kb(s) the Ddecomposition is given by
a(jω) + kb(jω) = 0, ω ∈ [0, ∞). (11)
We avoid the situations when a(s), b(s) have a common imaginary (or zero) root. Thus (39) is
equivalent to −1/k = H(jω) or to the standard Nyquist diagram: the critical values of the gain k
(which correspond to a change of the stable roots number for the characteristic polynomial) are
deﬁned by the intersections of the Nyquist plot H(jω) and the real axis. Note that the alternative
way to analyze the situation is the root locus technique [3, 16, 49]. However, Ddecomposition
approach allows one to ﬁnd critical points analytically and to estimate their number.
Theorem 1 [52]. For polynomial family (11) the real axis may be divided by points −∞ < k
1
<
k
2
< . . . < k
m
< k
m−1
< +∞ into no more than m n + 2 intervals (−∞, k
1
), (k
1
, k
2
), . . . ,
(k
m−1
, +∞) such that for every interval k
i
< k < k
i+1
polynomial a(s) + kb(s) has a constant
number of stable roots ν
i
. Moreover the number of stability intervals (i.e., intervals (k
i
, k
i+1
) with
ν
i
= n) does not exceed
_
n
2
_
+ 2 (¸α is the maximal integer less than or equal to α).
Note that the maximal number of stability intervals may be attained for the case of odd n when
both (−∞, k
1
) and (k
m−1
, +∞) intervals are stability intervals.
The algorithm for ﬁnding the critical values k
i
and the numbers ν
i
is based on this theorem. We
formulate it here under assumption that polynomials a(s) and b(s) do not have imaginary roots
and all critical values ω are simple roots of the equation ImH(jω) = 0. The case of multiple ω
requires additional investigations and we omit it here.
Algorithm.
Step 1. Solve the polynomial equation ImH(jω) =0 in 0 ω <∞(due to symmetry ImH(jω) =
− ImH(−jω)). Denote the obtained solutions 0 = ω
0
< ω
1
< ω
2
< . . . < ω
m−2
and calculate
u
i
= Re H(jω
i
), i = 0, . . . , m−2, u
m−1
= b
(n)
/a
(n)
.
Step 2. Order quantities −
1
u
i
, i = 0, . . . , m − 1 and denote them k
1
< . . . < k
m−1
such that
k
s
= −
1
u
is
.
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
1996 GRYAZINA et al.
Step 3. Find the interval containing zero: 0 ∈ (k
, k
+1
) and equate ν
with the number of stable
roots of the polynomial a(s).
Step 4. The numbers ν
s
(s = 1, . . . , m) are calculated successfully:
ν
s±1
=
⎧
⎪
⎨
⎪
⎩
ν
s
∓2sgn ϕ(ω
is
), ω
is
,= 0, ϕ(ω
is
) ,= 0
ν
s
∓sgn ϕ(0), ω
is
= 0, ϕ(ω
is
) ,= 0
ν
s
∓sgn ϕ(Ω), k
s
= −a
(n)
/b
(n)
, Ω being large enough,
where ϕ(ω
is
) =
d
dω
Im H(jω
is
).
For the case of two real parameters, the Ddecomposition for the polynomial family
a(s, k) = a
0
(s) + k
1
a
1
(s) + k
2
a
2
(s), k
1
, k
2
∈ R, (12)
is given by a(jω, k
1
, k
2
) = 0.
Resolving this equation with the respect to the parameters we obtain the curve
k
1
= −
Δ
1
Δ
, k
2
= −
Δ
2
Δ
, (13)
where
Δ =
¸
¸
¸
¸
¸
Re a
1
(jω) Re a
2
(jω)
Ima
1
(jω) Ima
2
(jω)
¸
¸
¸
¸
¸
, Δ
1
=
¸
¸
¸
¸
¸
Re a
0
(jω) Re a
2
(jω)
Ima
0
(jω) Ima
2
(jω)
¸
¸
¸
¸
¸
, Δ
2
=
¸
¸
¸
¸
¸
Re a
1
(jω) Re a
0
(jω)
Ima
1
(jω) Ima
0
(jω)
¸
¸
¸
¸
¸
.
For Δ ,= 0 the formulae (13) deﬁne the parametric curve k
1
(ω), k
2
(ω), and two roots cross the
stability boundary with crossing this curve. It is symmetric in ω: k
i
(−ω) = k
i
(ω), i = 1, 2, thus it
suﬃces to take for continuoustime systems ω ∈ (0, ∞). Moreover, there are two straight lines
ω = 0 : a
0
(0) + k
1
a
1
(0) + k
2
a
2
(0) = 0,
ω = ∞ : a
(n)
0
+ k
1
a
(n)
1
+ k
2
a
(n)
2
= 0.
Curve (13) starts at a point on the ﬁrst line (for ω = 0) and terminates at a point on the other
line. The intersection of these lines speciﬁes one root (0 or ∞) to cross the stability boundary. We
call these lines boundary lines.
If Δ = 0 but Δ
1
,= 0 or Δ
2
,= 0, then curve (13) goes to inﬁnity; it may have discontinuity at
these points. The value of ω such that Δ = Δ
1
= Δ
2
= 0 is called singular frequency. In this case,
the Ddecomposition reduces into a straight line. We call these lines singular lines.
Note that the singular ω such that Δ = Δ
1
= Δ
2
= 0 may ﬁll a particular interval. For
instance, for the polynomial a(s, k) = k
1
s
2
+ k
2
+ 1 (proposed by E.M. Solnechnyi) the system of
the Ddecomposition equations reduces to one equation in two variables because Ima(jω, k) ≡ 0.
Thus the line −k
1
ω
2
+k
2
+1 = 0 is singular for all ω, and for any point from the region k
1
(k
2
+1) 0
the polynomial has a pair of imaginary roots.
Theorem 2 [6, 53]. For family (11) with one complex parameter, the number of Ddecomposition
regions does not exceed N (n − 1)
2
+ 2, and for family (12) with two real parameter, N
2n(n −1) + 3.
The above mentioned consideration is valid for the continuoustime systems. But using fractional
transformation
s =
z + 1
z −1
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
DDECOMPOSITION TECHNIQUE STATEOFTHEART 1997
(a)
2.0
1.5
1.0
0.5
0
–0.5
–1.0
–1.5
–2.0
2 1 0 –1 –2
0
2
1
3
4
5
Re
k
I
m
k
(b)
0.8
0
–0.2
2 1 0 –1 –2
k
1
k
2
0.6
0.4
0.2
–0.4
–0.6
–0.8
–1.0
6
6
2 2
4
3
5
3
5
6 6
Fig. 1. Ddecomposition in (a) Example 1 and (b) Example 2.
allows us to proceed to the discretetime systems and vice versa. Thus due to conformity the
results remain valid for the discretetime systems also. Most of the examples are demonstrative
in the discretetime version but the demonstrated features are preserved also for the continuous
time systems. We prefer discretetime systems because it is easier to make the Ddecomposition
analysis for the ﬁnite frequency interval ω ∈ [0, 2π) rather than inﬁnite interval for the continuous
time systems. Here and elsewhere we denote by a(s, k) the continuoustime polynomial, i.e., the
polynomials is stable when all its roots are in the left half plane, and a(z, k) being the discretetime
polynomial and it is stable when all its roots are inside the unit circumference.
Example 1 [6]. The Ddecomposition for the polynomial a(z, k) = z
n
+kz
n−1
+α, where k ∈ C,
has (n −1)
2
+1 root invariant regions for α > 1, and two root invariant regions for α < 1/(n −1).
The Ddecomposition is given by the parametric curve k(ω) = −e
jω
− αe
−jω(n−1)
, 0 ≤ ω < 2π,
depicted in Fig. 1a. In this ﬁgure (as well as in the further ﬁgures) the digit inside the region
marks the number of stable roots. This curve is generated by a moving point on the complex
plane. This motion is a superposition of two rotations. First rotation has radius 1 and frequency
2π while the second one has radius α and frequency (n −1)2π. For α > 1, this curve consists of n
arcs, each arc intersects any other twice. Thus the calculation of the number of intersections gives
N = n
2
−2n + 2; this is just one region less the upper bound given by Theorem 2.
Example 2 [18]. For the polynomial a(z, k) = z
n
+k
1
z
n−1
+αz
n−2
+ k
2
, 1 < α <
n
n−2
there are
n − 1 simply connected stability regions in the ¦k
1
, k
2
¦ plane. The structure of the regions for
n = 6 can be seen in Fig. 1b; three regions (marked by the digit 6) are in the loops of the boundary
curve and the other two regions are formed by the intersection of the curve and boundary lines
given by the equations a(e
j0
, k) = 0, a(e
jπ
, k) = 0 and plotted by solid lines.
2.2. Ddecomposition for Aperiodicity
Remind that the polynomial a(s) is called aperiodic if all its roots are simple real and negative [9].
For the discretetime system with the polynomial a(z) the deﬁnition is similar but “negative” should
be replaced by “absolute value less than one.” The aperiodic property of the linear system (i.e.,
the system with aperiodic characteristic polynomial) guarantees that the transient process is not
oscillating. This property is desirable in many cases so it is natural to formulate the problem of
describing aperiodic domain in the parameter space.
We consider the same aﬃne polynomial family (8). Deﬁnition 1 is modiﬁed as follows:
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
1998 GRYAZINA et al.
5
1
0
k
1
k
2
4
3
2
5 1 0 4 3 2
t
= –1
D
3
A
3
t
→
–
∞
t
→
–0
Fig. 2. Stability and aperiodicity regions for Visgnegradsky exmaple.
Deﬁnition 2. The decomposition of the parameter space k into the regions A
l
= ¦k : a(s, k) has
l real simple negative roots¦, l = 0, 1, . . . , n, is called Adecomposition.
Obviously, the transition from one A
l
to the other occurs only when a multiple real or zero root
appears or the degree of polynomial changes. Thus the boundary of the Adecomposition is deﬁned
by the system of equations
a(t, k) = 0,
a
(t, k) = 0,
(14)
where t ∈ (−∞, 0) is the value of the real multiple root,
denotes diﬀerentiation in t, or the equation
a(0, k) = 0,
or the condition
a
(n)
(k) = 0.
(The notations are the same as in the beginning of the section.)
Speciﬁcally, for the cubic polynomial in Vishnegradsky form a(s, k) = s
3
+ k
1
s
2
+ k
2
s + 1 the
boundary is deﬁned just by Eq. (14): t
3
+ k
1
t
2
+ k
2
t + 1 = 0, 3t
2
+ 2k
1
t + k
2
= 0. Expressing k
1
,
k
2
in terms of t obtain:
k
1
(t) =
1
t
2
−2t, k
2
(t) = t
2
−
2
t
.
Remind that t < 0, thus the whole curve lies in the positive ortant and it is the boundary of the
aperiodic stability region (Fig. 2).
This result was obtained by Vishnegradsky [85] and it is published in old textbooks on automatic
control.
Now consider the case of one real parameter, i.e., the polynomial of the form
p(s, k) = a(s) + kb(s), k ∈ R.
Equations (14) for this case can be rewritten as:
a(t) + kb(t) = 0,
a
(t) + kb
(t) = 0.
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
DDECOMPOSITION TECHNIQUE STATEOFTHEART 1999
Suppose that a(s) and b(s) do not have a common negative real root. Then we can eliminate k,
and obtain
q(t) = a
(t)b(t) −a(t)b
(t) = 0.
Let a(s), b(s) be the polynomials of degree n. Then q(t) is a polynomial of degree no more
than 2n −1 and it has no more than 2n−1 real negative roots t
1
< t
2
< < t
m
< 0, m 2n−1.
Thus the ray (−∞, 0) can be divided into no more than 2n intervals. This fact allows us to obtain
the analog of Theorem 1 and an algorithm for construction of the Adecomposition of the parameter
space; we don’t deep into details here.
The problem of the robust aperiodicity can be solved using the technique described in [21, 23].
3. CONSTRUCTION OF THE MULTIDIMENSIONAL STABILITY DOMAIN
In this section we proceed to the consideration of the singleinput or output systems with the
characteristic polynomial linearly depending on the parameters. In some special cases, the stability
domain can be described constructively in the multidimensional (not only in two dimensional)
parameter space. For the continuoustime systems, we consider the families of the form:
a(s, k) = a
0
(s) + f(s)
v
i=1
k
i
a
i
(s
2
), k = ¦k
1
, . . . , k
v
¦ ∈ R
v
, (15)
where a
i
, i = 0, 1, . . . , v, f are the polynomials with real coeﬃcients, deg(a
0
) = n
0
, deg(f) = n
1
,
max
i
¦deg(f) + deg(a
i
), n
0
¦ = n. Note that all the polynomials a
i
, i = 1, . . . , v contain only even
degrees of s.
The results of this section are valid also for the discretetime systems of the form
a(z, k) = a
0
(z) + f(z)
v
i=1
k
i
a
i
(z), k = ¦k
1
, . . . , k
v
¦ ∈ R
v
, (16)
where deg(a
0
) = n
0
, deg(f) = n
1
, and the polynomials a
i
(z), i = 1, 2, . . . , v are simultaneously
symmetric or antisymmetric as the polynomials of degree n n
0
−n
1
. We recall that the polyno
mial a
i
(z) is called the symmetric polynomial of degree n if a
i
(z) ≡ z
n
a
i
(z
−1
), and antisymmetric
polynomial if a
i
(z) ≡ −z
n
a
i
(z
−1
) (for instance, z
3
+ z is a symmetric polynomial of the fourth
degree). Here we consider only continuoustime systems, for more details see [8].
It seems that the works [64, 47] were the ﬁrst studies in this direction which considered the
continuoustime case with f(s) ≡ 1 and the stability domain described by a system of linear
inequalities. A similar result for the discretetime systems [48] is mentioned in the survey [10]
among other nonKharitonov approaches to robustness of the discretetime systems. It is also
worth mentioning the works of J. Ackermann [33], N. Munro [84], S.P. Bhattacharyya [38, 57],
where the stability domain was constructed for the PIDcontroller. Further, we show that an
algebraic method for detecting the stability domain discussed in the previous sections works for
the multidimensional parameter space.
We note that in the pure form the polynomials of this kind are encountered rarely, but the
problem often may be reduced to the considered family, for example, by ﬁxing the parameters at
the odd degrees. This approach recently proved well in the combined probabilistic/deterministic
methods [52, 45].
We apply the Ddecomposition idea for continuoustime system (15). The boundaries of the
regions with the constant number of stable roots according to (9) are described by the system of
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
2000 GRYAZINA et al.
equations:
Re a
0
(jω) + Re f(jω)
v
i=1
k
i
a
i
(−ω
2
) = 0, (17)
Ima
0
(jω) + Imf(jω)
v
i=1
k
i
a
i
(−ω
2
) = 0, ω ∈ [0, ∞]. (18)
For ω = ∞, the equations contain coeﬃcients with the degree n only.
Expressing
v
i=1
k
i
a
i
(−ω
2
) from one equation and substituting in into another, we obtain an
equation without parameters:
Re a
0
(jω)Imf(jω) −Re f(jω)Ima
0
(jω) = 0. (19)
The simple real positive roots of this equation 0 = ω
0
< ω
1
< . . . < ω
m−1
such that f(±jω) ,= 0
are called critical frequencies. Moreover, if the equality deg(a
0
) = max
i
¦deg(f) + deg(a
i
)¦ holds,
we denote ω
m
= ∞, this value corresponds to a change on the degree. For each critical frequency
Eqs. (17) and (18) are linearly dependent and describe the hyperplane in the parameter space which
is the boundary of the Ddecomposition:
p(ω
i
) +[f(jω
i
)[
2
v
=1
k
a
(−ω
2
i
) = 0, i = 0, 1, . . . , m, (20)
where p(ω) = Re a
0
(jω)Re f(jω) + Ima
0
(jω)Imf(jω).
Equation (20) is the real part of the polynomial a
∗
(jω, k) = a(jω, k)f(−jω). At the points
ω
i
, it is equivalent to the system of Eqs. (17) and (18). Note that Eq. (19) without parameters
corresponds to the imaginary part of the polynomial a
∗
(jω, k).
These hyperplanes divide the parameter space R
v
into the regions with constant number of
stable roots, any such region being a convex polytope. In turn, every region is a solution of the
system of linear inequalities with the respect to the parameters:
ξ
i
_
p(ω
i
) +[f(jω
i
)[
2
v
=1
k
a
(−ω
2
i
)
_
> 0, ξ
i
∈ ¦−1; 1¦, i = 0, 1, . . . , m. (21)
Assertion 1 [8]. Any root invariant region D
l
of polynomial (15) is a union of nonintersecting
polytopes.
Since the space of parameters is divided into domains with constant number of stable roots,
then there are several methods of determining the number of stable roots in each domain. The
simplest method is to take a point in every region and to calculate the roots of the corresponding
polynomials. The classical Ddecomposition technique prescribes to use shading that shows the
extent of changes in the number of stable roots at the crossing of the curves of Ddecomposition.
Eﬃcient use of shading in general form is possible only for a low dimensionality of the parameter
space, already for R
3
this technique becomes hindered. For the oneparameter family, an algebraic
method of determining the number of stable roots in various domains was proposed in [6, 53]. The
generalization of the Hermite–Biehler theorem [58] enables one to exploit the idea of shading for
the multidimensional case for the polynomials of a certain form (15).
Let some region in the parameter space be described by the system of linear inequalities (21)
for a ﬁxed set ¦ξ
i
¦, i = 0, 1, . . . , m, where ω
i
are the zeros of the imaginary part of the original
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
DDECOMPOSITION TECHNIQUE STATEOFTHEART 2001
0
100
–100
–100
0
100
–100
0
100
k
1
k
0
k
2
Fig. 3. Stability domain in the parameter space in Example 3.
polynomial which contains no parameters. Since by construction the region boundaries satisfy the
equations of Ddecomposition, within this region the polynomial a(s, k) has the certain number of
l stable and r unstable roots. The diﬀerence l−r (that is called the signature of the polynomial σ(a))
in this region [58] is deﬁned by the set of signs of the real part at the points ω
i
, which corresponds
to ξ
i
= sgnRe a(jω
i
), i = 0, 1, . . . , m, that is,
σ(a) =
_
(ξ
0
−2ξ
1
+ . . . + (−1)
m−1
2ξ
m−1
+ (−1)
m
ξ
m
)(−1)
m−1
sgn(Ima(j∞)), if n is even
(ξ
0
−2ξ
1
+ . . . + (−1)
m−1
2ξ
m−1
)(−1)
m−1
sgn(Ima(j∞)), if n is odd.
Let f(s) has l
1
roots to the left of the imaginary axis and r
1
roots to its right. Then, the
polynomial a(s, k) has l stable and n −l unstable roots if a
∗
(s, k) = a(s, k)f(−s) has l + r
1
stable
roots; at that, σ(a
∗
) = l + r
1
−(n −l) −l
1
= 2l −n + r
1
−l
1
. The generalization of the Hermite–
Biehler theorem allows one to establish what relation sgn Re a
∗
(jω
i
), i = 0, . . . , m provides given
signature, and these signs specify the choice of ξ
i
for the system of inequalities (21).
The advantages of the considered method manifest themselves with increase in the number of
parameters because solution of systems of linear inequalities still remains the main operations.
Example 3. Consider the polynomial a(s, k) = 0.1+f(s)(k
0
+k
1
s
2
+k
2
s
4
), where f(s) is a stable
polynomial; at that, f(jω) = T
m
(1−ω
2
) +jωT
m−1
(1−ω
2
), where T
i
(t) are the Chebyshev polyno
mials. These polynomials are deﬁned for all s ∈ C, but their root are located in the segment [−1; 1].
Thus without loss of generality we use the representation for the Chebyshev polynomials of the
form T
i
(t) = cos(i arccos t), 1 −ω
2
= t ∈ [−1; 1].
The imaginary part f(jω) has m diﬀerent simple positive roots, deg(f) = 2m, deg(a) = 2m + 4,
and due to combinatorial estimations there exist C
(m/2)+1
m
sets ¦ξ
i
¦ each deﬁning a system of
inequalities describing a component of the stability domain.
For m = 4, the number of permissible sets ¦ξ
i
¦ is four, and as shown in Fig. 3, solution of
inequalities for all such ¦ξ
i
¦ turns out to be nonempty.
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
2002 GRYAZINA et al.
If the polynomial a
0
(s) is stable the theorem holds:
Theorem 3 [8]. The component of the stability domain of family (15) containing the nominal
stable polynomial a
0
(s) obeys the system of inequalities:
sgn (p(ω
i
))
_
p(ω
i
) +[f(jω
i
)[
2
v
=1
k
a
(−ω
2
i
)
_
> 0, i = 0, 1, . . . , m,
where p(ω) = Re a
0
(jω)Re f(jω) + Ima
0
(jω)Imf(jω).
3.1. Stability Radius
Let in the original aﬃne family (15) which is certainly stable for k = 0, the parameters arbitrarily
remaining bounded in some weighted pnorm:
a(s, k) = a
0
(s) + f(s)
v
i=1
k
i
a
i
(s
2
), k
α
p
γ, (22)
where α = ¦α
1
, . . . , α
v
¦, α
i
> 0 are some weights k
α
p
=
_
v
i=1
[α
i
k
i
[
p
_1
p
.
With this formulation, for p = ∞the parameters can vary independently, each within its interval
α
i
[k
i
[ γ; for p = 1, the parameters can assume values from the simplex
v
i=1
α
i
[k
i
[ γ, that is, be
bounded in the socalled octahedral norm, and for p = 2, the elliptic constrains like
v
i=1
α
2
i
k
2
i
γ
2
are imposed on the parameters. the proposed description of the stability domain allows to handle
also an arbitrary pnorm.
We set down the inequalities specifying the stability region as:
ϕ
i
0
+
v
=1
k
ϕ
i
> 0, i = 0, . . . , m, (23)
where ϕ
i
are the values of the corresponding polynomials on the critical frequencies ω
i
, namely:
ϕ
i
0
= sgn (p(ω
i
))p(ω
i
) = [p(ω
i
)[, ϕ
i
= sgn (p(ω
i
))[f(jω
i
)[
2
a
(−ω
2
i
).
Now we are interested in the stability radius, that is, need to determine the maximum value γ
∗
such that family (22) is stable for all γ < γ
∗
. Stated diﬀerently, there exists a polyhedron deﬁned
by the system of linear inequalities (23), and we have to inscribe in it a sphere in the corresponding
norm of the maximal radius.
The dualnorm lemma enables us to answer this question for each hyperplane:
Lemma 1. Let in R
v
the halfspace / = ¦k : 1 + (k, ϕ) = 0¦ be given. Then
γ
∗
= inf¦k
α
p
: k ∈ /¦ =
1
ϕ
1/α
q
=
1
_
v
i=1
¸
¸
¸
ϕ
i
α
i
¸
¸
¸
q
_
1/q
,
where
1
p
+
1
q
= 1.
For the family under consideration, the boundary of the stability domain is deﬁned by several
hyperplanes. Therefore it suﬃces to take the minimum for all γ
∗
i
corresponding to each hyperplane
in order to answer the question of the stability radius. The aforementioned in summed up by the
following theorem.
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
DDECOMPOSITION TECHNIQUE STATEOFTHEART 2003
Theorem 4 [8]. The stability radius for the aﬃne families (22) obeys γ
∗
= min
i
ϕ
i
0
ϕ
i
1/α
q
, where
ϕ
i
0
+ (k, ϕ
i
) > 0, i = 0, . . . , m is the system of inequalities describing the stability domain.
This result is exceptionally simple in terms of computations. At ﬁrst, by solving Eq. (19)
we have to determine the critical frequencies and then set down the system of inequalities (23)
deﬁning the stability domain. Now, for each hyperplane we have to calculate the corresponding
norm of the normal vector ϕ
i
and the corresponding γ
∗
i
. The ﬁnal answer is obtained by taking
the minimum from the ﬁnite set of the elements γ
∗
i
whose number corresponds to the number of
critical frequencies.
Note that there are other classes of multiparametric polynomials that allows one to calculate
stability radius easily. For instance, for interval polynomials, the answer is obtained using Tsypkin–
Polyak locus [21, 22].
4. ROBUST DDECOMPOSITION FOR POLYNOMIALS
Consider systems with uncertainty. Suppose that a characteristic polynomial depends on two
types of parameters k and x. The number of parameters of the ﬁrst type does not exceed two while
for the others x ∈ R
the domain of variation A is given. This problem statement is natural for
the problem of controller synthesis under uncertainty, k being controller parameters.
A system is robustly stable (in x), if for a given k it is stable for all x ∈ A [24]. In this section
we discuss the methods for describing the stability domain in the parameter space k.
Further we consider continuoustime systems and characteristic polynomials with real coeﬃ
cients, for the discretetime systems (see remark).
The Ddecomposition idea is also eﬃcient in this case [20, 15]:
Deﬁnition 3. For the polynomial a(s, k, x) of the degree n with parameters k, robust Ddecompo
sition is the decomposition of the parameter space k into regions D
l
(l = 0, . . . , n) in each of which
the number of stable roots is the same for all admissible uncertainties:
D
l
= ¦k : a(s, k, x) has l stable roots, ∀x ∈ A¦ .
Thus D
n
is the robust stability domain. First we deﬁne the set Υ that separates the regions D
l
.
In contrast to the classical Ddecomposition it consists not of the curve and several lines but it is
a set given by two conditions from zero exclusion principe (compare to (9)):
Υ
s
=
_
k : a
(n)
(k, x) = 0, ∃x ∈ A
_
; (24)
Υ
Ω
= ¦a(jω, k, x) = 0, ∃x ∈ A, ∃ω ∈ [0, ∞)¦ =
_
ω∈[0,∞)
Υ(ω), (25)
where a
(n)
is the leading coeﬃcient, Υ(ω) = ¦k : a (j ω, k, x) = 0, ∃ x ∈ A¦. The set Υ
Ω
is a sort
of smear of the classical Ddecomposition curve into the strip, while Υ
s
is smear of the singular
line deﬁning by the order reduction condition. Within the stripes polynomial a(s, k, x) may have
various number of stable roots depending in the value of x ∈ A.
We consider robust Ddecomposition for two real or one complex parameters, i.e., k ∈ R
2
or
k ∈ C. In this case Υ(ω) is a parametric family of the ﬁgures in the plane, Υ
Ω
is a swept area, and
a boundary is the envelope of the family Υ(ω).
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
2004 GRYAZINA et al.
4.1. Robust Ddecomposition for Two Real Parameters
Consider the linear system with the characteristic polynomial linearly dependent on all the
parameters with uncertainty x bounded in l
p
norm (or weighted l
p
norm):
a(s, k, x) = a
0
(s) + k
1
a
1
(s) + k
2
a
2
(s) +
i=1
x
i
b
i
(s), x
p
=
_
i
[x
i
[
p
_
1/p
1. (26)
The maximal degree of polynomials a
i
, b
i
is n. In this case the set Υ
s
is bounded by two lines:
a
(n)
0
+ k
1
a
(n)
1
+ k
2
a
(n)
2
±b
(n)

q
= 0, (27)
where index
(n)
denotes the s
n
held coeﬃcients of the corresponding polynomials, b
(n)
∈ R
, q =
p/(p −1) is the dual norm index.
Set Υ(ω) for polynomial (26) in the nonsingular case is
Υ(ω) = −T
−1
(ω)B
X
(ω), (28)
where
T(ω)
.
=
_
Re a
1
(jω) Re a
2
(jω)
Ima
1
(jω) Ima
2
(jω)
_
,
B
X
(ω)
.
=
⎧
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎩
⎡
⎢
⎢
⎢
⎢
⎣
Re
_
a
0
(jω) +
i=1
x
i
b
i
(jω)
_
Im
_
a
0
(jω) +
i=1
x
i
b
i
(jω)
_
⎤
⎥
⎥
⎥
⎥
⎦
, x
p
1
⎫
⎪
⎪
⎪
⎪
⎬
⎪
⎪
⎪
⎪
⎭
.
(29)
Set B
X
depending on the uncertainty may be [28, 20]:
•
a 2polygon if x ∈ R
, p = ¦1, ∞¦;
•
an ellipse if x ∈ R
, p = 2;
•
a circle if x ∈ C
, p ∈ [1, ∞).
In the nonsingular case Υ(ω) is obtained by linear transformation B
X
(ω) (28), thus it is either
a polygon or an ellipse. For the polygon the problem of description the envelope is hard. For the
ellipse Υ(ω) is deﬁned by
Υ(ω) =
_
k : (T(ω)k −b
0
(ω))
T
M(ω)
−1
(T(ω)k −b
0
(ω)) ρ(ω)
2
_
,
where k = (k
1
, k
2
)
T
, b
0
= −(Re a
0
, Ima
0
)
T
and the particular from of M and ρ depend on the
uncertainty.
If one of the matrices T or M is singular then Υ(ω) may be an empty set, a line, s strip or the
whole plane. If both matrices are nonsingular the envelope of the elliptic family Υ(ω) is obtained
analytically as the solution of the system of two secondorder equations [12, 28].
Remark. In the case of discretetime systems, the substitution s = e
jω
, ω ∈ [0, 2π) is used and
variation of the degree of the polynomial does not lead to the change of the number of stable roots.
4.2. Robust Ddecomposition for One Complex Parameter
For the robust Ddecomposition for one complex parameter k of the aﬃne polynomial
a(s, k, x) = a
0
(s) + ka
1
(s) +
i=1
x
i
b
i
(s), x ∈ C
, x
p
1, (30)
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
DDECOMPOSITION TECHNIQUE STATEOFTHEART 2005
boundaries ∂Υ
s
and ∂Υ
Ω
of sets Υ
s
and Υ
Ω
can be obtained explicitly:
∂Υ
s
=
_
k :
¸
¸
¸a
(n)
0
+ ka
(n)
1
¸
¸
¸ =
_
_
_b
(n)
_
_
_
q
_
,
∂Υ
Ω
⊂ Υ
reg
Ω
∪ Υ
spec
Ω
,
where
Υ
reg
Ω
=
⎧
⎨
⎩
k
0
(ω) −ρ(ω)
k
0
(ω)
[k
0
(ω)
[
⎛
⎝
ρ(ω)
[k
0
(ω)
[
±j
¸
1 −
_
ρ(ω)
[k
0
(ω)
[
_
2
⎞
⎠
:
[k
0
(ω)
[ [ρ(ω)
[ > 0,
a
1
(jω) ,= 0, ω ∈ R
⎫
⎬
⎭
,
Υ
spec
Ω
=
_
k: [a
0
(jω) + ka
1
(jω)[ = b(jω)
q
,
if τ
0
(ω)
= [ρ(ω)
[ = 0, ω ∈ R
if these derivatives are not deﬁned
_
,
k
0
(ω)
.
= −a
0
(jω)/a
1
(jω), ρ(ω)
.
= b(jω)
q
/[a
1
(jω)[, q = p/(p −1).
Example 4. Let us consider the robust Ddecomposition for one complex parameter for the
discretetime system with polynomial (robust version of Example 1):
z
6
(1 + 0.05x
1
) + kz
5
+ 1.5 + 0.05(x
1
+ 2x
2
) = 0; z = e
jω
,
[x
i
[ 1, x
i
∈ C, i = 1, 2, k ∈ C.
(31)
The ﬁgure of Ddecomposition is shown in Fig. 4; regions D
l
are denoted by digits l. Ddecompo
sition without uncertainty is plotted by dotted curve; see also Fig. 1a.
5. SYNTHESIS OF LOWORDER CONTROLLERS WITH H
∞
SPECIFICATIONS
The problem of designing the loworder controllers is to determine a stabilizing controller such
that the closedloop system is stable and some addition quality criterion holds. Here we consider H
∞
criteria. We recall that the H
∞
norm is ﬁnite only for rational functions with stable denominator
(H(s) ∈ RH
∞
) and equals to H(s)
∞
= sup
ω
[H(jω)[.
The analytical theory of the design of the H
∞
optimal controllers has been developed in detail
in [5, 50, 86] but the resulting controllers may be of a very high order, sometimes even exceeding
that of the original system [11]. Additionally, stability of the closedloop system is very sensitive
to the controller parameters, their small variation often results in instability [59]. Since the H
∞
theory does not allow one to constrain the order of the designed controller, its direct application
to the design of controllers of a given structure faces substantial diﬃculties. The other approaches
to the ﬁxed order controller synthesis with H
∞
speciﬁcations are discussed in [52, 56, 60, 77, 78].
We present an alternative approach to the design of a given structure controllers, i.e., the order
of the controller is ﬁxed and it is the only choice of its parameters remains free. The results are
published in [7, 83]. Further we assume that there are only two or three adjustable controller
parameters that allows us to use widely the graphic methods. Thus the problem is to ﬁnd the
regions in the parameter space such that the corresponding controllers (i) stabilize the system and
(ii) satisfy the H
∞
criterion. It is important that our approach describes the whole set of controllers
with given speciﬁcations, if any. For some particular controllers the description of the whole domain
of admissible parameters in the parameter space is proposed in [40, 54, 70].
Let us consider a linear stationary onedimensional system, where the plane is given by the
scalar transfer function G(s). Let the system be closed by the controller C(s, k); the block diagram
of the system is shown in Fig. 5. Parameters k ∈ R
m
, m 3 deﬁne the set of admissible loworder
controllers. For instance, it may be the PIcontroller k
p
+
k
i
s
, the PIDcontroller k
p
+
k
i
s
+ k
d
s, and
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
2006 GRYAZINA et al.
1 0 2 –1 –2
Re
k
–2
–1
0
D
3
D
4
D
2
D
6
D
1
1
2
I
m
k
Fig. 4. Robust Ddecomposition in Example 4.
C
(
s
,
k
)
C
(
s
)
–
Fig. 5. Block diagram of the control system.
the ﬁrstorder controller
k
1
s+k
2
s+k
3
. The problem is to describe the set of the parameters common to
all controllers satisfying the performance criterion
H(s, k)
∞
< γ. (32)
We call these controllers the H
∞
controllers. The particular form of the transfer function H(s, k)
may be diﬀerent. H
∞
criterion arises in several formulations of the design problems of which we
present the most important examples as well as the corresponding functions H(s, k).
•
Determination of the controller C(s) guaranteeing the H
∞
performance of the closedloop
system,
W
1
(s)T(s)
∞
γ,
where W
1
(s) is a weight function (W
−1
1
∈ RH
∞
), T(s) =
C(s)G(s)
1+C(s)G(s)
is the transfer function of the
closedloop system (additional sensitivity), and γ is the given level of performance.
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
DDECOMPOSITION TECHNIQUE STATEOFTHEART 2007
•
Determination of a controller robustly stabilizing the plant family with additive uncertainty
G(s) = G
0
(s) +ΔG(s), where G
0
(s) is the transfer function of the nominal plant and the frequency
(nonparametric [38]) uncertainty ΔG(s) is bounded in the weighted norm W
2
(s)ΔG(s)
∞
1.
The problem comes to satisfying the criterion W
−1
2
(s)U(s)
∞
1, where U(s) =
C(s)
1+C(s)G(s)
,
W
2
∈ RH
∞
. The criterion W
−1
2
(s)T(s)
∞
1 corresponds to the multiplicative uncertainty
G(s) = G
0
(s)(1 + ΔG(s)); see [50, 86].
•
In the quantitative feedback theory (QFT) [5], the closedloop system must satisfy the following
constraints:
m
1
(ω) < [W(jω)T(jω)[ < m
2
(ω), ω ∈ [0, ω
1
],
[S(jω)[ < l
1
(ω), ω ∈ [0, ω
1
],
[T(jω)[ < l
2
(ω), ω ∈ [ω
2
, ∞),
where S(s) =
1
1+C(s)G(s)
is sensitivity and m
1
, m
2
, l
1
and l
2
are the limitative functions.
Assume that the controller stabilized the system, i.e., the denominator H(s) is stable. Then
H(s, k)
∞
= sup
ω∈[0,∞)
[H(jω, k)[, where H(s, k) =
Hn(s,k)
H
d
(s,k)
is the common rational transfer function.
Further we assume that the numerator H
n
and denominator H
d
depend linearly on the parame
ters k, which is the case for all considered types of the controllers (PI, PID, ﬁrstorder controller).
For instance, for the PIDcontroller and the sensitivity function H
n
(s, k) = sD(s), H
d
(s, k) =
sD(s) + (k
i
+ k
p
s + k
d
s
2
)N(s), where N(s) is the numerator of the transfer function of the plant
G(s) and D(s) is its denominator.
5.1. Ddecomposition for H
∞
Criterion
One of the approaches to deﬁne the admissible set of H
∞
controllers is based on the description
of the boundary of the set / = ¦k : [H
n
(jω, k)[ < γ[H
d
(jω, k)[, ∀ω ∈ [0, ∞)¦ in the explicit form.
The set of parameters k corresponding to H
∞
controllers is an intersection of / and the stability
domain for the characteristic polynomial. Similarly to the classical Ddecomposition approach, we
consider the crossing the boundary of /. There may be three essentially diﬀerent cases: either
the function [H(jω, k)[ reaches the value γ, its numerator and denominator vanish simultaneously,
or the degree of a denominator will decrease. The points where only the denominator is zero do
not belong to the boundary because their small neighborhood has not a single point where the
inequality [H
n
(jω, k)[ < γ[H
d
(jω, k)[ holds.
Whenever H(jω, k) is deﬁned we use suﬃcient extremum condition with the respect to ω to
obtain an easytouse theorem.
Theorem 5 [7]. The boundary of the set / is contained in the solutions of the systems
_
H
n
(jω, k) = 0
H
d
(jω, k) = 0,
ω ∈ [0, ∞), (33)
⎧
⎨
⎩
[H(jω, k)[
2
= γ
2
∂[H(jω, k)[
2
∂ω
= 0,
ω ∈ [0, ∞), (34)
and equations
[H(j∞, k)[ = γ, [H(0, k)[ = γ. (35)
h
d
(k) = 0, (36)
where h
d
is coeﬃcient at the higher order of a polynom H
d
.
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
2008 GRYAZINA et al.
1.0
0.5
0
–0.5
–1.0
k
d
–0.25 –0.20 –0.15 –0.10 –0.05 0 0.05
k
i
Fig. 6. H∞ region for PIDcontrollers.
For the ﬁxed k function [H(jω)[ may be at its maximum either at the solution of system (33)
or system (34) or at the solution of Eqs. (35). Equation (36) describes an opportunity of loss of
stability of the closedloop system.
For the two variables k = (k
1
, k
2
), it is the aggregate of points that are the solution of system (34).
The ﬁrst equation deﬁnes for each ω either a onedimensional curve or an empty set. The second
equation of the system enables one to specify such k on the curve the function [H[
2
reaches its
extremum for the given ω. Since the condition for equality to zero of the derivative keeps one from
discrimination whether it is the maximum or minimum that is reached at this point, points not
belonging to the desired boundary may occur among the solution.
Example 5. Consider the application of the proposed method for designing a ﬁxedorder PID
controller [46]. The plant is deﬁned by the transfer function G(s) =
s−1
s
2
+0.8s−0.2
, the controller, by
C(s, k
i
, k
d
) =
k
i
+kps+k
d
s
2
s
, k
p
= −0.35. It is required to determine all the controllers satisfying the
H
∞
criterion of performance W(s)T(s)
∞
1 with highfrequency ﬁlter W(s) =
s+0.1
s+1
.
The boundary of the domain / in the parameter space is speciﬁed by Theorem 5. Equation (33)
gives the vertical line k
i
= 0 (a part of this line is also the stability domain boundary), Eq. (35)
gives the horizontal line k
d
= −0.5, and system (34) gives the curves depicted in Fig. 6. Checking
each of the regions we ﬁnd out the desirable one marked with “.” Indeed, the whole this region
is inside the stability domain that is plotted with chain line.
6. DDECOMPOSITION FOR MATRICES
Consider real matrices A ∈ R
n×n
, B ∈ R
n×r
, C ∈ R
m×n
, and let / be a connected set of
real or complex matrices r m. Assume that the matrix A does not have zero or imaginary
eigenvalues for the continuoustime case, and does not have eigenvalues on the unit circumference
in the discretetime systems.
Deﬁne the matrix transfer function:
M(s) = C(A−sI)
−1
B (37)
for the continuoustime systems and
M(z) = C(A−zI)
−1
B (38)
for the discretetime systems, where the variables s and z are using to distinguish between these
two cases.
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
DDECOMPOSITION TECHNIQUE STATEOFTHEART 2009
Deﬁnition 4. The decomposition of the parameter space into the regions D
l
= ¦K ∈ / : A +
BKC has l stable eigenvalues¦, l = 0, . . . , n, is called Ddecomposition. The equation describing
the boundary of the regions D
l
is called the Ddecomposition equation. Thus D
n
is the set of
stabilizing matrices K.
The Ddecomposition technique is based on the following theorem.
Theorem 6 [53]. The equation
det(I + M(jω)K) = 0, ω ∈ (−∞, +∞) (39)
or
det(I + M(e
jω
)K) = 0, ω ∈ [0, 2π) (40)
deﬁnes the Ddecomposition for the class /, i.e., if Q ⊂ / is a connected set and det(I +
M(jω)K) ,= 0, ω ∈ (−∞, +∞), ∀K ∈ Q (continuoustime case) or det(I + M(e
jω
)K) ,= 0, ω ∈
[0, 2π), ∀K ∈ Q (discretetime case), then A+BKC has the same number of stable eigenvalues for
all matrices K from Q.
Equations (39) and (40) deﬁne the Ddecomposition in the implicit from. Here we consider some
particular cases when we can write the equations of the boundary of the Ddecomposition regions
explicitly. Note that this is in contrast with μanalysis, where the problem
min
K∈K, det(I+M(jω)K)=0
K
is under consideration (i.e., one is seeking for the largest ball contained in D). Better approximation
of D was proposed in [34]; it was the ellipsoid of the largest volume, inscribed in D.
Note that the system of Eqs. (39) and (40) deﬁnes the Ddecomposition in general form and
contains the polynomial case that is discussed in details in Section 2. Indeed, consider single
output system; i.e., r = 1. Then K is a row vector: K = [k
1
, . . . , k
m
] while M is a column vector
M = [M
1
, . . . , M
m
]
T
∈ C
m
. According to the formula det(I + ab
T
) = 1 +
m
i=1
a
i
b
i
with a, b ∈ C
m
the general Ddecomposition Eq. (39) is reduced to:
1 +
m
i=1
k
i
M
i
(jω) = 0 or a
0
(jω) +
m
i=1
k
i
a
i
(jω) = 0,
where a
0
(s) +
k
i
a
i
(s) is the characteristic polynomial of the degree n, M
i
(s) =
a
i
(s)
a
0
(s)
. Thus
the Eq. (39) is linear in K. Similarly, for the singleinput systems m = 1, K is a column vector
K = [k
1
, . . . , k
r
]
T
, and with have the same Ddecomposition Eq. (8).
However in general case Eqs. (39) and (40) are not linear in K, and the analysis requires the
extension of the Ddecomposition technique.
6.1. Graphical Algorithms
Here we consider classes of / such that the transfer function is not the ration of the polynomials
and classical Ddecomposition is not applicable. Though the modiﬁed Ddecomposition technique
allows us to describe the regions with constant number of stable eigenvalues.
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
2010 GRYAZINA et al.
We start with the class (4). In the terms of μanalysis this is class / with one scalar block. The
matrix K has the form K = kI, where k ∈ C or k ∈ R, I is a unit matrix mm. Then the matrix
A + BKC is equal to A + kBC and the problem is reduced to the simplest one: given n n real
matrices A and F, ﬁnd D(l) = ¦k ∈ C (or k ∈ R) : A + kF has l stable eigenvalues¦. The general
Ddecomposition Eq. (39) now reads
det(I + kM(jω)) = 0, ω ∈ (−∞, +∞). (41)
Denote the eigenvalues of M(jω) as λ
i
(ω), i = 1, . . . , n, Eqs. (41) split into 1 + kλ
i
(ω) = 0,
i = 1, . . . , n, and Ddecomposition boundary consists of n branches k
(i)
(ω):
k
(i)
(ω) = −
1
λ
i
(ω)
, i = 1, . . . , n. (42)
This equations can be obtained in a diﬀerent form with no use of the transfer function (37). If
A+kF (F = BC) has an imaginary eigenvalue then the matrix A+kF −jωI is singular for some
ω ∈ R, that is (A + kF − jωI)x = 0 for x ∈ C
n
or (A − jωI)x = −kFx. Thus we conclude that
parametric Ddecomposition boundary k(ω) is a generalized eigenvalue for the matrix pair A−jωI
and −F:
k(ω) = eig (A−jωI, −F). (43)
For the case of real k we should take into consideration only the intersection of the obtained
Ddecomposition picture and the real axis.
Theorem 7 [53]. For the real k the number of intervals preserving the same number of stable
eigenvalues of A + kBC does not exceed n(n + 1) + 1.
Now let us consider systems with two inputs and two outputs corresponding to class (5). The
transfer function now is the 2 2 matrix M =
_
m
1
m
2
m
3
m
4
_
, the matrix K contains in general
4 parameters K =
_
k
1
k
3
k
2
k
4
_
, and the general Ddecomposition Eq. (39) reduces to:
0 = det(I + MK) = 1 +
4
i=1
k
i
m
i
+ det M det K. (44)
To take the opportunity of the graphical representation we restrict ourselves by situations with
K depending on two parameters only. Consider the cases when it is possible.
(1) K =
_
k
1
0
0 k
2
_
.
In terms of μanalysis this structure of K corresponds to two scalar blocks. Equation (44) has
the form 0 = 1+k
1
m
1
+k
2
m
4
+k
1
k
2
(m
1
m
4
−m
2
m
3
). Denote m
i
= u
i
+jv
i
, i = 1, . . . , 4 and write
separately real and imaginary parts:
_
1 + k
1
u
1
+ k
2
u
4
+ αk
1
k
2
= 0
k
1
v
1
+ k
2
v
4
+ βk
1
k
2
= 0,
(45)
where α = u
1
u
4
−v
1
v
4
− u
2
u
3
+ v
2
v
3
, β = u
1
v
4
+ v
1
u
4
−u
2
v
3
− v
2
u
3
. Thus we get two quadratic
equation in two variables, the variables u
i
, v
i
, α, β depend on ω. For ω = 0 matrix M(jω) =
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
DDECOMPOSITION TECHNIQUE STATEOFTHEART 2011
C(A − jωI)
−1
B is real and v
i
(0) = 0, i = 1, . . . , 4 as well as β(0) = 0. Hence the second equation
vanishes and the ﬁrst equation: 1 + k
1
u
1
(0) + k
2
u
4
(0) + α(0)k
1
k
2
= 0. It is a hyperbola.
For every ω ,= 0, solution of system (45) speciﬁes the point of the Ddecomposition boundary. If
for some ω
∗
the solution is complex we ignore it because it does not belong to the Ddecomposition.
It means that there is no parameter value K
∗
such that A + BKC has eigenvalues ±jω
∗
.
(2) K =
_
k
1
k
2
−k
2
k
1
_
.
In terms of μanalysis this is a real 2 2 analog of a complex scalar (note that the eigenvalues
of such K are k
1
± jk
2
). For such K Eq. (44) reads 0 = 1 + k
1
(m
1
+ m
4
) − k
2
(m
2
− m
3
) + (k
2
1
+
k
2
2
) (m
1
m
4
−m
2
m
3
) and (45) is replaced with
_
1 + k
1
(u
1
+ u
4
) −k
2
(u
2
−u
3
) + α(k
2
1
+ k
2
2
) = 0
k
1
(v
1
+ v
4
) −k
2
(v
2
−v
3
) + β(k
2
1
+ k
2
2
) = 0,
(46)
where u
i
(ω), v
i
(ω), α(ω), β(ω) are the same as above. For ω = 0, we get v
i
(0) = 0, β(0) = 0, and
the second equation vanishes while the ﬁrst equation deﬁned the second order curve 1+k
1
(u
1
(0) +
u
4
(0)) −k
2
(u
2
(0) −u
3
(0)) +α(0)(k
2
1
+k
2
2
) = 0. For this case it is a circumference. Solving Eq. (46)
for ω ,= 0, we obtain k
1
(ω), k
2
(ω) corresponding to the boundary of the Ddecomposition. Thus
the Ddecomposition consists of the components of this curve and singular circumference.
(3) K =
_
−k
1
k
2
k
2
k
1
_
.
Similarly to the previous case, Eq. (44) is reduces to:
_
1 −k
1
(u
1
−u
4
) + k
2
(u
2
+ u
3
) −α(k
2
1
+ k
2
2
) = 0
−k
1
(v
1
−v
4
) + k
2
(v
2
+ v
3
) −β(k
2
1
+ k
2
2
) = 0.
(47)
The Ddecomposition consists of the parametric curve k
1
(ω), k
2
(ω), ω ,= 0 and the second order
curve corresponding to ω = 0. Again it the circumference 1+k
1
(u
1
(0)−u
4
(0))+k
2
(u
2
(0)+u
3
(0))−
α(0)(k
2
1
+ k
2
2
) = 0.
6.2. Examples
Example 6. This example demonstrates the attainability of the estimation of Theorem 7 for
n = 3. Let
A =
⎡
⎢
⎣
095 1 0
0 0 0.6
0 0 −0.95
⎤
⎥
⎦, B =
⎡
⎢
⎣
0 0 −0.22
0 −03 0
04 0 0
⎤
⎥
⎦, C = I.
In Fig. 7a the Ddecomposition is depicted. It consists of three branches (parametric curves cor
responding to the diﬀerent generalized eigenvalues). In the intersection with real axis k there are
13 intervals with constant number of stable eigenvalues, among them there 5 intervals corresponding
to stable matrices A + kB.
Example 7. Consider the discretetime system A + B
_
k
1
0
0 k
2
_
C, where
A =
_
−0.8848 0.4457
−0.8733 −0.9326
_
, B =
_
0.3914 0.2508
−0.5576 0.0266
_
, C =
_
0.1514 0.7854
−0.4255 −0.8148
_
,
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
2012 GRYAZINA et al.
2
3
2
1
0
–1
–2
–3
–6 –4 –2 0 2 4 6
2
3 3 3
2 2
2 2
1
0
1
3 3
1
Stability
Re
k
I
m
k
1
10
0
–10
–40
–50
–10 –5 0 5 10
2
2
1
0
1
k
1
k
2
1
0
1 0
–20
–30
Fig. 7. Ddecomposition in Examples 6 (left) and 7 (right).
5
0
–10
–20 –15 0 5
2
1
k
1
k
2
–5
–10 –5
3
1
0
–2
–4 –3 0 4
2
1
k
1
k
2
–1
–1
3
–3
–4
–5
–2 3 2 1
2
4
2
0
Fig. 8. Ddecomposition in Examples 8 (left) and 9 (right).
has a typical Ddecomposition structure. In Fig. 7b one can see two singular hyperbolas (subtle
lines) and two branches of the nonsingular curve (solid lines).
Example 8. This discretetime example is borrowed from [75, c. 889, Example 3]; n = 3,
m = r = 2. The optimal solution for (7) is supplied with K
∗
=
_
0.8483 0.5971
−0.5971 0.8483
_
; it has
the form
_
k
1
k
2
−k
2
k
1
_
. Thus we restrict ourselves with matrices K of this form and construct
Ddecomposition for such matrices (see Fig. 8a). It is generated by two singular circumferences
(subtle lines) and one parametric curve (solid line). Note that the nominal system (the origin
marked with “”) is close enough to the boundary of the stability domain and the distance to it, in
accordance with [75], equals 1.0374. However, other directions allow larger values of perturbations
preserving stability. For instance, if K = λ
_
−0.9680 −0.2508
0.2508 −0.9680
_
, then A + BKC remains stable
for 0 ≤ λ ≤ 19.6932.
Example 9. This continuoustime example is again originated in [75, c. 889, Example 2], and
the same problem of the stability radius is considered, n = 4, m = r = 2. The optimal solution
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
DDECOMPOSITION TECHNIQUE STATEOFTHEART 2013
found in [75] is the matrix of the form K =
_
−k
1
k
2
k
2
k
1
_
. The Ddecomposition of (k
1
, k
2
) plane
for the given class of matrices is shown in Fig. 8b. There are two disconnected components of the
parametric curve (solid lines) and one singular curve—the circumference (subtle line). Similar to
the previous example, for many directions preserve stability for perturbations larger than 0.5141
(stability radius).
7. DDECOMPOSITION FOR LINEAR MATRIX INEQUALITIES
This section is devoted to the following problem: In the space of parameters x ∈ R
, ﬁnd the
regions such that
A(x) = A
0
+
i=1
x
i
A
i
, A
i
∈ S
n×n
, (48)
has a ﬁxed number of negative eigenvalues. Here S
n×n
denotes the space of real, symmetric n n
matrices. In this section we follow standard for linear matrix inequalities notations for parameters x,
rather than k as it is in previous sections. The diﬀerence from the problem of the previous section
is the requirement for all matrices in (48) to be symmetric. It makes the problem signiﬁcantly
easier.
A special case of this problem is to ﬁnd a point x such that linear combination of symmetric
matrices is negative deﬁnite. This is the wellknown feasibility problem for linear matrix inequalities
(the common abbreviation LMI [2], [41]) A(x) < 0. By mednineties, the theory and methods for
linear matrix inequalities had shaped into a selfcontained discipline. In particular, optimization
problems under constrains in LMI form were called semideﬁnite programming. However the LMI
theory is not aimed at describing the whole feasible set, not to mention other signature invariant
regions; various problems involving model uncertainty often defy eﬃcient solution.
The primary goal here is not only to describe the whole feasible set ¦x ∈ R
: A(x) < 0¦ but also
to characterize all the regions in the parameter space such that the signature of the matrix A(x)
remains unaltered. It is accomplished using the Ddecomposition technique modiﬁed for the case
of symmetric matrices. But as it tool place in previous sections, for a small number of parameters,
= 2, the regions are easy to depict graphically on the plane. A complete description of the feasible
set gives a clue to new statements and solutions for the semideﬁnite programming problems, and
the determination of the other eigenvalue invariant regions is highly important for checking the
applicability conditions of the generalized Sprocedure; see, [36]. Moreover, this potentially leads
to relatively simple solution methods for systems of matrix inequalities with uncertainty in matrix
coeﬃcients.
7.1. One Parameter Problems
We start with the case of one scalar parameter:
A(x) = A + xB, A, B ∈ S
n×n
, x ∈ R.
The goal is to deﬁne the intervals on the parameter axis x such that the signature of the matrix
A(x) is constant:
D
l
=
_
x ∈ R: A(x) has l negative eigenvalues and n −l positive eigenvalues
_
.
All basic ideas of the proposed approach become apparent in this simplest scalar case.
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
2014 GRYAZINA et al.
Assume that for a certain real x
∗
, symmetric matrix A(x
∗
) = A + x
∗
B is nonsingular and
has l negative and n − l positive eigenvalues. As x varies, the number of likesign eigenvalues
can only alter if one of them (or several at once) crosses the origin, i.e., det(A + xB) = 0 for
some x ∈ R. This means that there exists a nonzero vector e such that (A + xB)e = 0, i.e.,
Ae = −xBe. In other words, x is a generalized eigenvalue of the pair of matrices A and −B, and
e is the associated generalized eigenvector. Therefore, the determination of the boundaries (points)
of the Ddecomposition regions in the scalar case reduces to ﬁnding all real generalized eigenvalues:
eig (A, −B) ∈ R.
It is seen that the maximal possible number of regions (intervals on the xaxis) is equal to n+1,
since the equation det(A+xB) = 0 has at most n real solutions. On the other hand, there might be
no real solutions at all, in this case the Ddecomposition is represented by a unique domain—the
number of negative eigenvalues of the matrix A + xB remains unaltered for all values of x.
Of a particular interest is the stability interval D
n
. The following theorem holds.
Theorem 8 [25]. Let A, B ∈ S
n×n
and B be nonsingular. Let λ
i
, e
i
, i = 1, . . . , n be the general
ized eigenvalues and the associated generalized eigenvectors of the pair (A, −B). Then
(1) If there exist complexvalued eigenvalues λ
i
then D
n
is empty;
(2) If all λ
i
are real, then denote
x =
_
max
i∈I
−
λ
i
, I
−
.
= ¦i : (Be
i
, e
i
) < 0¦ ,= ∅
−∞, I
−
= ∅;
x =
_
min
i∈I
+
λ
i
, I
+
.
= ¦i : (Be
i
, e
i
) > 0¦ ,= ∅
+∞, I
+
= ∅.
Then
D
n
=
_
(x, x), if x < x
∅, if x x.
If it is known that A < 0, the formulas above take the form:
x =
_
max
λ
i
<0
λ
i
−∞, provided all λ
i
> 0
and x =
_
min
λ
i
>0
λ
i
+∞, provided all λ
i
< 0.
(49)
7.2. Problems with Two and More Parameters
We now turn to the twoparameter case: A(x) = A
0
+ x
1
A
1
+ x
2
A
2
, A
i
∈ S
n×n
, i = 0, 1, 2. Let
one of the matrices A
1
, A
2
, say, A
2
be non singular. We ﬁx x
1
and denote A = A(x
1
)
.
= A
0
+x
1
A
1
and B
.
= A
2
. Then we are in the conditions of the previous section, and the critical values of
the parameter x
2
for the given x
1
are determined as the real generalized eigenvalues x
2
(x
1
) =
eig
_
A(x
1
), −B
_
. By varying x
1
we obtain the boundaries of the Ddecomposition regions. For
every x
1
equation det
_
A(x
1
) + x
2
B
_
= 0 has no more than n real roots, hence, the boundary of
Ddecomposition consists of no more than n branches.
In the general situation, the stability region
D
n
= ¦x ∈ R
2
: A(x) < 0¦
can be described separately (it is obviously convex). Namely, for every x
1
, determine the interval
(x
2
(x
1
), x
2
(x
1
)) according to Theorem 8 (it may happen to be empty for some or all values of x
1
);
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
DDECOMPOSITION TECHNIQUE STATEOFTHEART 2015
4
4
5
2
1
2
4
2
1.0
–1.0 –0.5 1.0
4
x
1
x
2
0 0.5
3
0.8
0.6
0.4
0.2
0
–0.2
–0.4
–0.6
–0.8
–1.0
5
6
4
5
4
(a)
6
–4 3
x
1
x
2
0 1
3
–2
–4
–6
(b)
4
2
0
–3 –2 –1 2 4
1
0
6
Fig. 9. Ddecomposition in Example 10.
as x
1
varies, the endpoints of the interval sweep the boundary of the stability region. Note that
D
n
is knowingly unbounded if one of the matrices A
i
, i = 1, . . . , , in LMI (48) is signdeﬁnite.
The marking can be performed in a standard way, i.e., upon constructing D
l
regions on the
plane, pick a point inside each of them and compute the eigenvalues of the corresponding matrix.
Example 10. In some special cases the description of the regions can be obtained explicitly. As
an example of this sort, consider the following n n, n = 2m matrices:
A
0
= A = −I; A
1
= B = diag(b
1
, . . . , b
m
, −b
m
, . . . , −b
1
);
A
2
= C = diag(c
1
, . . . , c
m
, c
m
, . . . , c
1
),
(50)
where diag denotes an antidiagonal matrix. The determinant is straightforward to compute:
det(A + x
1
B + x
2
C) = (−1)
m
(x
2
1
b
2
1
+ x
2
2
c
2
1
−1) . . . (x
2
1
b
2
m
+ x
2
2
c
2
m
−1),
so the boundaries of the Ddecomposition are represented by m ellipses x
2
1
b
2
i
+x
2
c
2
i
= 1 (they divide
the plane into n(n −2)/2 + 2 regions). The example of the Ddecomposition for 6 6 matrices of
this sort with parameters b
1
= 1, b
2
= 1.25, b
3
= 2; c
1
= 2, c
2
= 1.25, c
3
= 1 is depicted in Fig. 9a.
If we replace A and C in (50)
A = diag(c
1
, . . . , c
m
, c
m
, . . . , c
1
);
B = diag(b
1
, . . . , b
m
, −b
m
, . . . , −b
1
); C = −I,
(51)
then the determinant is
det(A + x
1
B + x
2
C) = (−1)
m
(x
2
1
b
2
1
−x
2
2
+ c
2
1
) . . . (x
2
1
b
2
m
−x
2
2
+ c
2
m
).
The boundary of the Ddecomposition is given by the family of m pairs of hyperbolas x
2
1
b
2
i
−x
2
2
+
c
2
i
= 0, however the total number of the regions does not exceed n
2
/2 + 1. Taking for b
i
, c
i
the
values form the previous example we depict the regions in Fig. 9b.
7.3. Linear Matrix Inequalities with Uncertainty
One of the valuable extensions of the proposed in the previous section approach is its modiﬁcation
to the presence of uncertainty. There are various formulation of LMI problems and assumptions on
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
2016 GRYAZINA et al.
uncertainties available in the literature; e.g., [41, 37]. Here we concentrate on the situation where
uncertainty is bounded in the spectral norm.
Let
A
i
(Δ
i
) = A
i
+ Δ
i
, A
i
, Δ
i
∈ S
n×n
, Δ
i
 ε
i
, i = 0, . . . , , (52)
where   is the spectral norm and ε
i
0 are given numbers. Note that in order to retain the
LMI structure of the problem, only symmetric perturbation Δ
i
are considered. We thus arrive at
the following uncertain linear function:
A(x, Δ) = A
0
(Δ
0
) +
i=1
x
i
A
i
(Δ
i
), Δ ∈ T
.
=
_
(Δ
0
, . . . , Δ
) : Δ
i
∈ S
n×n
, Δ
i
 ε
i
_
. (53)
The regions D
l
of robust Ddecomposition are now deﬁned as follows:
D
rob
l
=
_
x ∈ R
: A(x, Δ) has exactly l negative eigenvalues ∀ Δ ∈ T
_
; (54)
in particular, the robustly feasible region is deﬁned by:
D
rob
n
=
_
x ∈ R
: A(x, Δ) < 0 ∀ Δ ∈ T
_
.
The problem is to describe the boundaries of the regions of robust Ddecomposition; these bound
aries are now deﬁned as the values of the parameters x such that the matrix A(x, Δ) becomes
singular for some Δ ∈ T.
Similarly to the problem of robust Ddecomposition for polynomials, the boundaries of robust
Ddecomposition smear into “strips” inside which the matrix A(x, Δ) may have diﬀerent values of
signature depending on one or another admissible value of Δ.
Similarly to the analysis of the problem without uncertainty, we ﬁrst consider oneparameter
families in the simplest case where only the A matrix is subjected to additive uncertainty:
A(x, Δ) = (A + Δ) + xB; Δ ∈ T
.
= ¦Δ ∈ S
n×n
: Δ ε¦. (55)
The boundaries of the regions (intervals) of the robust Ddecomposition are deﬁned from the
condition that the matrix A+xB+Δ be singular for some Δ ∈ T, i.e., the problem is to determine
the radius of nonsingularity of the matrix A + xB. The theorem below is the main tool:
Theorem 9 [25]. For a nonsingular matrix M ∈ S
n×n
, its symmetric radius of nonsingularity
ρ(M)
.
= inf¦P: P ∈ S
n×n
, M + P is singular¦,
is equal to
ρ(M) = 1/M
−1
 = min
i
[λ
i
(M)[.
The critical value of P is given by P = −λee
T
, where λ is the minimal (in absolute value) eigenvalue
of M, and e is the associated eigenvector.
Theorem 9 is a counterpart of Theorem 3 in [73] for the symmetric case; for nonsymmetric
matrices, the radius of nonsingularity is given by the general formula in Theorem 3 [73].
By the theorem, the matrix (A + xB) + Δ with perturbations Δ ε remains robustly non
singular for the values of x satisfying
(A + xB)
−1
 <
1
ε
, (56)
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
DDECOMPOSITION TECHNIQUE STATEOFTHEART 2017
4
3
2
1
0
–1
–2
–3
–4 4 3 2 1 0 –1 –2 –3
x
1
x
2
D
3
D
2
D
2
rab
D
3
rab
D
n
rab
Fig. 10. Robust Ddecomposition for two parameters.
hence, introducing the function ϕ(x)
.
= (A + xB)
−1
, the intervals of robust nonsingularity are
found numerically as ¦x: ϕ(x) < 1/ε¦.
A more general problem where the uncertainty enters both matrices,
A(x, Δ) = (A + Δ
A
) + x(B + Δ
B
); Δ
A
 ε
A
; Δ
B
 ε
B
, (57)
is analyzed similarly. Represent A(x, Δ) = (A+xB)+(Δ
A
+xΔ
B
). For the perturbation Δ
A
+xΔ
B
of the matrix A + xB we have the estimate Δ
A
+ xΔ
B
 ε
A
+ [x[ε
B
, which is sharp (the
equality is attainable), since Δ
A
and Δ
B
are chosen independently. Introduce the function ϕ(x) =
(A + xB)
−1
; then by Theorem 9, the intervals of robust nonsingularity of equivalently, the
regions D
rob
l
are deﬁned by the condition
ϕ(x) <
1
ε
A
+[x[ε
B
.
Slightly changing the notation, we now address the case of twoparameter families:
A(x, Δ) = (A
0
+ Δ
0
) + x
1
(A
1
+ Δ
1
) + x
2
(A
2
+ Δ
2
); Δ
i
 ε
i
, i = 0, 1, 2.
We ﬁx x
1
and denote A = A
0
+x
1
A
1
and B = A
2
and also ε
A
.
= ε
0
+[x
1
[ε
1
and ε
B
= ε
2
. This leads
us to the setup of problem (57). Indeed, to determine the boundaries, we compose the functions
ϕ(x
2
)
.
= (A+x
2
B)
−1
 and ε(x
2
)
.
= 1/(ε
A
+[x
2
[ε
B
) and check the condition ϕ(x
2
) < ε(x
2
), which
deﬁned the segment of robustness in x
2
for a given value of x
1
. As x
1
varies, these intervals ﬁll
up twodimensional regions of robust Ddecomposition. On the other hand, the values of x
2
that
violate the condition ϕ(x
2
) < ε(x
2
) correspond to the absence of robustness; as x
1
varies, these
intervals of violation sweep certain twodimensional regions which are considered the boundaries of
robust Ddecomposition.
We also note that similarly to the uncertaintyfree case, the region D
rob
n
can be described sepa
rately from the rest D
rob
l
by combining the results of Theorems 8 and 9.
Example 11. For the twoparameter family in S
4×4
with matrices A
0
= −I, A
1
=
diag(−1, 1, 1, −1) and A
2
= diag(1, −1, 1, −1), the boundaries of Ddecomposition in the absence
of uncertainty are given by the two pairs of parallel lines in Fig 10.
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
2018 GRYAZINA et al.
Introducing uncertainty smears the boundaries into stripes and the D
l
regions shrink. In the
example considered, for equal relative levels of uncertainty Δ
i
 ε
i
= 0.2A
i
, i = 0, 1, 2, the
boundaries of robust Ddecomposition are given by the shaded region and the uncolored regions
represent the regions of robustness.
Within this framework we consider more general classes of uncertainties. The analysis is based
on the following result of Peterson.
Theorem 10 [72]. Let G = G
T
, M ,= 0, N ,= 0 be matrices of the size n n, n r, m n
correspondingly. Then inequality
G + MΔN + N
T
Δ
T
M
T
< 0
hold for all matrices Δ of size r m, Δ < 1 iﬀ there exist ε > 0 such that
G + εMM
T
+
1
ε
N
T
N < 0.
Various generalizations and applications of this result are addressed in [31]. For our case, if we
change the family (55) by
A(x, Δ) = A + Δ
A
+ xB, Δ
A
= MΔN + N
T
Δ
T
M
T
, Δ 1,
then A(x, Δ) < 0 for all Δ 1 if A + xB < −εMM
T
−
1
ε
N
T
N for some ε > 0. Then we can
ﬁnd the intervals in x that preserve robust signdeﬁniteness of A(x, Δ).
8. RANDOMIZED ALGORITHMS FOR SYNTHESIS USING DDECOMPOSITION
As it was shown in the previous sections the Ddecomposition idea in its classical form is applied
only in one or twodimensional space. There are only a few cases when the stability domain ban
be explicitly described for higher dimensional parameter space. One of these cases (a spacial
polynomial family (15)) is described in Section 3. However the general Ddecomposition Eq. (9)
or (10) for polynomials and (39) or (40) for matrices is valid also for the multidimensional case.
The problem is to propose the detailed enough description of the multidimensional stability domain
(or its boundary) in general situation.
One of the opportunities to achieve this goal is to generate point randomly (uniform generating is
preferable) inside the stability domain or on its boundary. It can be done using modern schemes of
Monte Carlo method. One of the methods (known as HitandRun) [29,79] generates approximately
uniformly distributed points inside the given domain, generally speaking nonconvex and not simply
connected. The second method (socalled ShakeandBake) [4, 44] allows one to generate point on
the boundary of the domain that is again non convex but connected. First we brieﬂy describe these
method in general, then demonstrate how they work for the description of the stability domain
(both for polynomials and matrices) and a domain speciﬁed by linear matrix inequalities.
For all the methods of this section the main tool is boundary oracle procedure. Similar to mem
bership oracle and separation oracle that are exploited in modern convex optimization theory [36],
for any x, y ∈ R
n
boundary oracle either provides intersection points of a ray x+λy and the bound
ary of the set / deﬁned implicitly or reports that the intersection is empty. Let / ⊂ R
n
be an
open bounded, generally speaking nonconvex and not simply connected (but consisting of a ﬁnite
number of connected components) set such that any intersection of the set / and a line consists of
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
DDECOMPOSITION TECHNIQUE STATEOFTHEART 2019
a ﬁnite number of intervals. For a certain point k ∈ / and a direction d ∈ R
n
, the boundary oracle
provides the intersection of the line k + λd, −∞< λ < +∞ and the set /, i.e., the set
S = ¦λ ∈ R : k + λd ∈ /¦.
This set, by assumption on /, consists of a ﬁnite number of intervals; suppose that the boundary
oracle is available (i.e., for any k, d the set S can be computed). Then HitandRun method works
as follows.
Step 1. Find a starting point k
0
∈ /.
Step 2. At the point k
i
∈ / generate a random direction d
i
∈ R
n
uniformly distributed on the
unit sphere (i.e., d
i
= ξ/ξ, ξ = randn(n, 1) the ndimensional vector with normally distributed
components).
Step 3. Apply boundary oracle procedure, i.e., deﬁne the set
S
i
= ¦λ ∈ R : k
i
+ λd
i
∈ /¦.
Step 4. Generate a point λ
i
uniformly distributed in S
i
(we recall that S
i
is a ﬁnite set of
intervals), and compute a new point
k
i+1
= k
i
+ λ
i
d
i
.
Step 5. Go to step 2 and increase i.
The general result on HitandRun method is due to Smith [79].
Theorem 11. Under the taken assumptions the distribution of the point k
i
tends to the uniform
one in / as i → ∞.
There is also the convergence estimate in the original paper [79, Theorem 3]; more precise estimate
for convex sets can be found in [65, 66].
We address the method for the polynomial stabilization problem ﬁrst. We consider the aﬃne
family of polynomials (8), i.e., the polynomials of the form
a(s, k) = a
0
(s) +
m
i=1
k
i
a
i
(s).
Our goal is to describe somehow the stability domain in the parameter space k, i.e.,
/ = D
n
= ¦k ∈ R
m
: a(s, k) is stable¦.
This set satisfy the above assumption, namely, / is open and consists of a ﬁnite number of com
ponents. Moreover, the boundary oracle is given by Algorithm from the Section 2.1. Suppose that
k
0
∈ / is known (for instance, if a
0
(s) is stable then h
0
can be chosen as k
0
= 0), then we may
apply HitandRun to generate points k
i
approximately uniformly ﬁlling the stability domain. Thus
it is the way to solve various synthesis problems. For instance, we have a performance index J(k)
(gain or phase margin, overshoot or other timeresponse characteristics, robustness margin, H
2
or
H
∞
norms). Calculating J(k
i
) for admissible point we ﬁnd the optimal one. In the neighborhood
of k
∗
we can repeat the process, taking the intersection of / with some trust region (say a ball
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
2020 GRYAZINA et al.
centered at k
∗
of radius ε) as new admissible set and k
∗
as the initial point. A number of examples
for various problems of synthesis are addressed in [74].
HitandRun is applicable not only for stabilization problem of polynomials but also for LMI’s
A(x) < 0 (48). The diﬃculty suﬃciently reduces because the boundary oracle of / provides an
interval (λ, λ) that can be computed as follows. We consider a point x ∈ R
and a direction y ∈ R
.
For λ ∈ R, we have
A(x + λy) = A
0
+
i=1
x
i
A
i
+ λ
i=1
y
i
A
i
,
and denoting
A
.
= A
0
+
i=1
x
i
A
i
, B
.
=
i=1
y
i
A
i
,
we are in the oneparameter case A(λ) = A + λB. Boundary points for a chosen direction y are
deﬁned as real generalized eigenvalues λ
i
for the matrix pair (A, −B). Generating directions uni
formly on the unit dimensional sphere in the form y = η/η, where η has standard dimensional
gaussian distribution, we obtain Ddecomposition boundary points x + λ
i
y, i.e., the boundary or
acle.
The proposed boundary oracle is especially eﬃcient for characterization of the stability do
main D
n
which is always convex for LMI. Let for some x ∈ R
matrix A(x) (48) be negative
deﬁnite, x ∈ D
n
, and y is a certain direction. Then critical values λ, λ such that signdeﬁniteness
of the matrix A(x + λy) is preserved are given by formulae (49). This approach is easytouse for
generating uniform distribution in D
n
(compare to [42]) as well as proposing new programming
methods.
Numerous application examples of this approach for optimization of linear functions subject to
LMI (for semideﬁnite programming, SDP [41, 69]) are given in [26]. These methods appear to be
quite competitive compare to the standard tools such as Yalmip [67] and LMI toolbox for Matlab.
Similarly HitandRun can be applied for robust LMI’s. Consider the construction of the bound
ary oracle for robust stability domain D
rob
n
(54) for the family (52) and (53) for > 2. We seek
for the intersection points of a ray and the boundary of robust Ddecomposition. Let x ∈ D
rob
n
be
robustly feasible point and y ∈ R
be a certain direction. Consider the straight line x + λy and
compute λ
rob
, λ
rob
, the minimal and maximal values of λ, which guarantee the negative deﬁniteness
of the matrix A(x + λy, Δ) for all Δ ∈ T. We have
A(x + λy, Δ) =
´
A(λ) + Δ(λ),
where it is denoted
´
A(λ)
.
= A
0
+
i=1
(x
i
+ λy
i
)A
i
, Δ(λ)
.
= Δ
0
+
i=1
(x
i
+ λy
i
)Δ
i
,
and by Theorem 9, the matrix
´
A(λ) + Δ(λ) remains nonsingular (hence, negative deﬁnite) for
all Δ ∈ T satisfying
_
_
_
_
_
´
A(λ)
_
−1
_
_
_
_
< 1/Δ(λ). Since the perturbations Δ
i
are independent, the
estimate Δ(λ) Δ
0
 +
i=1
[x
i
+λy
i
[ Δ
i
 = ε
0
+
i=1
[x
i
+λy
i
[ε
i
is sharp. hence, by considering
the two scalar functions
ϕ(λ)
.
=
_
_
_
_
_
_
_
A
0
+
i=1
(x
i
+ λy
i
)A
i
_
−1
_
_
_
_
_
_
, ε(λ)
.
=
1
ε
0
+
i=1
[x
i
+ λy
i
[ ε
i
, (58)
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
DDECOMPOSITION TECHNIQUE STATEOFTHEART 2021
the interval (λ
rob
, λ
rob
) of robust deﬁniteness of the family A(x + λy, Δ) can be found numerically
as ¦λ: ϕ(λ) ε(λ)¦.
Theorem 12 [25] . let A(x, 0) < 0. For any y ∈ R
, the maximal and minimum values of λ
retaining the negative deﬁniteness of the matrix A(x + λy, Δ) for all admissible perturbations Δ
are given by the two solutions of the equation ϕ(λ) = ε(λ) (58) over the segment [λ, λ], where
[λ, λ] are the bounds of negative deﬁniteness of the matrix A(x + λy, 0) (49).
Similarly, to solve a simpler problem of checking if x ∈ D
rob
n
for some x ∈ R
, it suﬃces to
consider the unperturbed matrix at the point x: A(x, 0) = A
0
+
i=1
x
i
A
i
and check if the inequality
ε
0
+
i=1
[x
i
[ε
i
< 1/(A(x, 0))
−1
 holds.
Another randomized method mentioned above is ShakeandBake that allows generating points
not inside but in the boundary of a set /. We assume the boundary oracle that provided some extra
information compare to one used for HitandRun. Namely, for the boundary point k ∈ ∂/ the
boundary oracle provides the internal normal N = N(k) for ∂/ at this point and for the internal
direction d ∈ R
n
, (d, N) > 0 computes λ = λ(k, d) = sup¦λ > 0 : k + td ∈ /, ∀t ∈ (0, λ)¦. In the
other words, λ is a point of the ﬁrst intersection of the ray k + λd, λ > 0 and the boundary /.
ShakeandBake algorithm is the following:
Step 1. Given an initial point k
0
∈ ∂/.
Step 2. At the point k
i
∈ ∂/ compute the unit internal normal N
i
for ∂/ and generate random
direction d
i
as follows: d
i
= α
i
N
i
+ c
i
, where c
i
is uniformly distributed in c
i
 = 1, (c
i
, N
i
) = 0;
α
i
> 0 is a scalar random variable, α =
_
1 −ξ
2
n−1
_
1/2
, where ξ = rand is uniformly random
in [0, 1].
Step 3. Using the boundary oracle ﬁnd λ
i
= λ(k
i
, d
i
) and generate a new point
k
i+1
= k
i
+ λ
i
d
i
.
Step 4. Go to Step 2 and increase i.
The behavior of this method in the general situation is given by the following
Theorem 13 [44]. Let / be a connected open bounded set with a normal mostly at every point.
Then the distribution of k
i
tends to the uniform distribution at ∂/ as i → ∞.
We note that the set / is not assumed to be convex. Figure 11 depicts the result of the
generating points at the boundary of the nonconvex set. Indeed, the points look approximately
uniformly distributed.
For the problems with stability domain /, ShakeandBake technique works, for instance, for
the matrix case considered in Section 6. Let
¯
/ = ¦K ∈ R
r×m
: A + BKC be stable¦,
/ be a simply connected component of
¯
/ that contains the certain K
0
(for instance, K
0
= 0 for
stable matrix A). The boundary oracle for ShakeandBake algorithm can be produced numerically
using Eq. (39) of the Ddecomposition for matrices. Finally, we obtain matrices K approximately
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
2022 GRYAZINA et al.
0.5
0.4
0.3
0.2
0.1
0
–0.1
–0,2
–0.3
–0.4
–0.5
k
2
–0.6 –0.4 –0.2 0 0.2 0.4 0.6
k
1
Fig. 11. Points generated in the boundary of nonconvex domain.
uniformly distributed on the boundary of /. ShakeandBake method also works for LMI; the
required boundary oracle is deﬁned explicitly.
Finally, proposed randomized algorithms allows one to ﬁll in the stability domain (or its bound
ary) with randomly generated points. This procedure may be regarded as an multidimensional
alternative for graphical algorithms.
9. CONCLUSION
We provided the simple and eﬀective techniques to describe the stability domain in the parameter
space. This is an extension of the Ddecomposition method for polynomials. Simultaneously with
the stability domain we construct all the root (eigenvalue) invariant regions, i.e., simply connected
regions of the domains with the invariant number of likesign roots (eigenvalues) of the system
matrix. This technique can be helpful for the loworder controllers design and for a detailed
robustness analysis.
After more than 50 years the Ddecomposition idea remains actual and there appear new appli
cations for various control problems. The conducted survey shows that the domain of applicability
for Ddecomposition is much wider than the original problem statement for polynomials.
Besides new results on the structure of Ddecomposition for polynomials, we discuss the method
for description of the root invariant regions for uncertain polynomials. The modiﬁcations of the
technique in order to describe the regions in the parameter space corresponding to the ﬁxedorder
controllers that satisfy some H
∞
speciﬁcations are considered. The main limitation is the small
number of parameters but in some spacial cases (see Section 3) is can be overcome. In these cases,
Ddecomposition allows characterizing the multidimensional stability domain as a system of linear
inequalities and it gives a straightforward solution of the stability radius problem.
The extension of the Ddecomposition for the systems with matrix transfer functions is of
particular importance. The condition of alteration of the number of stable eigenvalues of the
parameter depending matrices is formulated in terms of singularity of the new matrix. The result
has much in common with the basic Ddecomposition theorem that describes when the stable root
number changes but it is applicable for more general system structure. We pick out several classes
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
DDECOMPOSITION TECHNIQUE STATEOFTHEART 2023
of parameter depending matrices and get the analytical solution of Ddecomposition equations for
them. When the number of parameters do not exceed two the graphical representation of the result
is possible.
The Ddecomposition technique is applied in the theory of linear matrix inequalities. It is a
tool to construct the regions such that the aﬃne family of symmetric matrices has a ﬁxed number
of likesign eigenvalues inside a region. Finally, we propose randomized algorithms exploiting the
Ddecomposition idea in the multidimensional case.
ACKNOWLEDGMENTS
Authors are grateful to P.S. Shcherbakov for fruitful discussions and problems solution for Sec
tion 7, and also acknowledge E.M. Solnechnyi for useful remarks.
REFERENCES
1. Andronov, A.A. and Mayer, A.G., Simple Linear Systems with Delay, Avtom. Telemekh., 1946, no. 7,
pp. 95–106.
2. Balandin, D.V. and Kogan, M.M., Sintez zakonov upravleniya na osnove lineinykh matrichnykh ner
avenstv (Synthesis of Control Laws Based on Linear Matrix Inequalities), Moscow: Fizmatlit, 2007.
3. Bendrikov, G.A. and Teodorchik, K.F., Traektorii kornei lineinykh avtomaticheskikh sistem (Root Tra
jectories of Linear Automation Systems), Moscow: Nauka, 1964.
4. Borovkov, K.A., On a New Variant of the Monte Carlo Method, Teor. Veroyatn. Primen., 1991, vol. 36,
no 2, pp. 342–346.
5. Horowitz, I., Sintez system s obratnoi svyaz’yu (Synthesis of Feedback Systems), Moscow: Sovetskoe
Radio, 1970.
6. Gryazina, E.N., On the Ddecomposition Theory, Avtom. Telemekh., 2004, no. 12, pp. 15–28.
7. Gryazina, E.N., Polyak, B.T., and Tremba, A.A., Design of the LowOrder Controllers by the H
∞
Criterion: A Parametric Approach, Avtom. Telemekh., 2007, no. 3, pp. 93–105.
8. Gryazina, E.N. and Polyak, B.T., Multidimensional Stability Domain of Special Polynomial Families,
Avtom. Telemekh., 2007, no. 12, pp. 38–52.
9. Jury, E.I., Innory i ustoichivost’ (Innors and Dynamical System Stability), Moscow: Nauka, 1979.
10. Jury, E.I., Robustness of DiscreteTime Systems, Avtom. Telemekh., 1990, no. 5, pp. 3–28.
11. Kiselev, O.N. and Polyak, B.T., Design of the LowOrder Controllers by the H
∞
Criterion, Avtom.
Telemekh., 1999, no. 3, pp. 119–130.
12. Luzin, N.N., Diﬀerentsial’noe ischislenie (Diﬀerential Calculus), Moscow: Vysshaya Shkola, 1960.
13. Neimark, Yu.I., Search for the Parameter Values That Make Automatic Control System Stable, Avtom.
Telemekh., 1948, vol. 9, no. 3, pp. 190–203.
14. Neimark, Yu.I., Ustoichivost’ linearizovannykh system (Stability of the Linearized Systems), Leningrad:
LKVVIA, 1949.
15. Neimark, Yu.I., Robastnaya ustoichivost’ i Drazbienie (Robust Stability and Dpartition), Avtom. Tele
mekh., 1992, no. 7, pp. 10–18.
16. Nesenchuk, A.A., Analiz i sintez robastnykh dinamicheskikh sistem na osnove kornevogo godografa
(Analysis and Synthesis of the Robust Dynamical Systems Based on Root Approach), Minsk: OIPI
NAN Belarusi, 2005.
17. Nikolayev, Y.P., The Multidimensional Asymptotic Stability Domain of Linear Discrete Systems: Its
Symmetry and Other Properties, Avtom. Telemekh., 2001, no. 11, pp. 109–120.
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
2024 GRYAZINA et al.
18. Nikolayev, Y.P., The Set of Stable Polynomials of Linear Discrete Systems: Its Geometry, Avtom.
Telemekh., 2002, no. 7, pp. 44–54.
19. Nikolayev, Y.P., The Geometry of DDecomposition of a TwoDimensional Plane of Arbitrary Coeﬃ
cients of the Characteristic Polynomial of a Discrete System, Avtom. Telemekh., 2004, no. 12, pp. 49–61.
20. Petrov, N.P. and Polyak, B.T., Robust Ddecomposition, Avtom. Telemekh., 1991, no. 11, pp. 41–53.
21. Polyak, B.T. and Tsypkin, Ya.Z., Frequency Domain Criteria for Robust Stability and Aperiodicity of
Continuous Linear Systems, Avtom. Telemekh., 1990, no. 9, pp. 45–54.
22. Polyak, B.T. and Tsypkin, Ya.Z., Robust Stability of Linear DiscreteTime Systems, Dokl. Akad. Nauk
SSSR, 1991, vol. 316, no. 4, pp. 842–846.
23. Polyak, B.T. and Tsypkin, Ya.Z., Robust Aperiodicity, Dokl. Akad. Nauk SSSR, 1994, vol. 335, no. 3,
pp. 304–307.
24. Polyak, B.T. and Shcherbakov, P.S., Robastnaya ustoichivost’ i upravlenie (Robust Stability and Con
trol), Moscow: Nauka, 2002.
25. Polyak, B.T. and Shcherbakov, P.S., The Ddecomposition Technique for Linear Matrix Inequalities,
Avtom. Telemekh., 2006, no. 11, pp. 159–174.
26. Polyak, B.T. and Shcherbakov, P.S., Randomized Method for Semideﬁnite Programming, in Stochastic
Optimization in Informatics, Granichin, O.N., Ed., St. Petersburg: S.Peterburg. Gos. Univ., 2006,
pp. 38–70.
27. Sokolov, A.A., A Criterion of Stability for Linear Regulated Systems with Distributed Parameters, in
Inzhenernyi sbornik (Engineering Collection), Talitskikh, N.A., Ed., Leningrad: Akad. Nauk SSSR, 1946,
vol. II, no. 2, pp. 3–26.
28. Tremba, A.A., Robust Ddecomposition under l
p
bounded Parametric Uncertainties, Avtom. Telemekh.,
2006, no. 12, pp. 21–36.
29. Turchin, V.F., On the Computation of Multidimensional Integrals by the Monte Carlo Method, Teor.
Veroyatn. Primen., 1971, vol. XVI, no. 4, pp. 738–743.
30. Kharitonov, V.L., Asymptotic Stability of a Family of Systems of Linear Diﬀerential Equations, Diﬀ.
Uravn., 1978, vol. 1, no. 11, pp. 2086–2088.
31. Shcherbakov, P.S. and Khlebnikov, M.V., Petersen’s Lemma on Matrix Uncertainty and Its Generaliza
tions, Avtom. Telemekh., 2008, no. 11, pp. 125–139.
32. Ackermann, J., Robust Control: The Parameter Space Approach, London: Springer, 2002.
33. Ackermann, J. and Kaesbauer, D., Stable Polyhedra in Parameter Space, Automatica, 2003, vol. 39,
pp. 937–943.
34. Barmish, B. and Polyak, B.T., The Volumetric Singular Value and Robustness of Feedback Control
Systems, Proc. 32nd CDC, San Antonio, 1993, TX, pp. 521–523.
35. Barmish, B.R., New Tools for Robustness of Linear Systems, New York: MacMillan, 1995.
36. BenTal, A. and Nemirovski, A., Lectures on Modern Convex Optimization: Analysis, Algorithms and
Engineering Applications, Philadelphia: SIAM, 2001.
37. BenTal, A. and Nemirovski, A., On Tractable Approximations of Uncertain Linear Matrix Inequalities
Aﬀected by Interval Uncertainty, SIAM J. Optim., 2002, vol. 12, no. 3, pp. 811–833.
38. Bhattacharyya, S., Chapellat, H., and Keel L., Robust Control: The Parametric Approach, Upper Saddle
River: Prentice Hall, 1995.
39. Bhattacharyya, S.P., Tantaris, R.N., and Keel, L.H., Stabilization of DiscreteTime Systems by First
Order Controllers, IEEE Trans. Automat. Control, 2003, vol. 48, no. 5, pp. 858–861.
40. Blanchini, F., Lepschy, A., Miani, S., and Viaro, U., Characterization of PID and Lead/Lag Com
pensators Satisfying Given H
∞
Speciﬁcations, IEEE Trans. Automat. Control , 2004, vol. 49, no. 5,
pp. 736–740.
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
DDECOMPOSITION TECHNIQUE STATEOFTHEART 2025
41. Boyd, S.P., El Ghaoui, L., Feron, E., and Balakrishnan, V., Linear Matrix Inequalities in System and
Control Theory, Philadelphia: SIAM, 1994.
42. Calaﬁore, G., Random Walks for Probabilistic Robustness, Proc. 43rd Conf. Decision Control , Bahamas,
2004, pp. 5316–5321.
43. Chen, J., Fan, M.K.H., and Nett, C.N., Structural Singular Value and Stability of Uncertain Polynomials.
II, A Missing Link, Syst. Control Lett., 1994, vol. 23, no. 2, pp. 97–109.
44. Comets, F., Popov, S., Sch¨ utz, G.M., and Vachkovskaia, M., Billiards in a General Domain with Random
Reﬂections, URL: arXiv:math/0612799v3 [math.PR].
45. Dabbene, F., Polyak, B., and Tempo, R., On the Complete Instability of Interval Polynomials, Syst.
Control Lett., 2007, vol. 56, pp. 431–438.
46. Datta, A., Ho, M.T., and Bhattacharyya, S.P., Structure and Synthesis of PID Controllers, New York:
SpringerVerlag, 2002.
47. Delansky, J.F. and Bose, N.K., Real and Complex Polynomial Stability and Stability Domain Construc
tion via Network Realizability Theory, Int. J. Control , 1988, vol. 48, no. 3, pp. 1343–1349.
48. Delansky, J.F. and Bose, N.K., Schur Stability and Stability Domain Construction, Int. J. Control ,
1989, vol. 49, no. 4, pp. 1175–1183.
49. Evans, W.R., Control System Dynamics, New York: McGrawHill, 1954.
50. Francis, B., A Course in H
∞
Control Theory, Berlin: Springer, 1987.
51. Frazer, R.A. and Duncan, W.J., On the Criteria for Stability for Small Motions, Proc. Royal Soc., Ser. A,
1929, vol. 124, pp. 642–654.
52. Fujisaki, Y., Oishi, Y., and Tempo, R., A Mixed Probabilistic/Deterministic Approach to Fixed Order
H
∞
Controller Design, Proc. 45th CDC, 2006, pp. 3554–3559.
53. Gryazina, E.N. and Polyak, B.T., Stability Regions in the Parameter Space: Ddecomposition Revisited,
Automatica, 2006, vol. 42, no. 1, pp. 13–26.
54. G¨ uvenc, L. and Ackermann, J., Links between the Parameter Space and Frequency Domain Methods of
Robust Control, Int. J. Robust Nonlinear Control , 2001, vol. 11, pp. 1435–1453.
55. Hinrichsen, D. and Pritchard, A., Stability Radius for Structured Perturbations and the Algebraic Riccati
Equation, Syst. Control Lett., 1986, vol. 8, pp. 105–113.
56. Ho, M.T., Synthesis of H
∞
PID Controllers: A Parametric Approach, Automatica, 2003, vol. 39,
pp. 1069–1075.
57. Ho, M.T., Datta, A., and Bhattacharyya, S.P., A New Approach to Feedback Stabilization, Proc. 35th
CDC, 1996, pp. 4643–4648.
58. Ho, M.T., Datta, A., and Bhattacharyya, S.P., An Elementary Derivation of the Routh Hurwitz Crite
rion, IEEE Trans. Automat. Control , 1998, vol. 43, no. 3, pp. 405–409.
59. Keel, L. and Bhattacharyya, S.P., Robust, Fragile or Optimal?, IEEE Trans. Automat. Control , 1997,
vol. 42, no. 6, pp. 1098–1105.
60. Keel, L.H., Rego, J.I., and Bhattacharyya, S.P., A New Approach to Digital PID Design, IEEE Trans.
Automat. Control , 2003, vol. 48, no. 4, pp. 687–692.
61. Keel, L.H. and Bhattacharyya, S.P., PID Controller Synthesis Free of Analytical Models, Proc. 16 IFAC
World Congr., Prague, 2005.
62. Kogan, J., Robust Stability and Convexity, London: Springer, 1995.
63. Lehnigk, S.H., Stability Theorems for Linear Motions with an Introduction to Liapunov’s Direct Method,
Englewood Cliﬀs: Prentice Hall, 1966.
64. Lin, H. and Hollot, C.V., Results on Positive Pairs of Polynomials and Their Application to the Con
struction of the Stability Domains, Int. J. Control , 1987, vol. 45, no. 3, p. 153.
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
2026 GRYAZINA et al.
65. Lov´asz, L., HitandRun Mixes Fast, Mathem. Program., 1999, vol. 86, pp. 443–461.
66. Lov´asz, L. and Vempala, S., HitandRun from the Corner, SIAM J. Comput., 2006, vol. 35, no. 4,
pp. 985–1005.
67. Lofberg, J., Yalmip: A Toolbox for Modeling and Optimization in MATLAB, Proc. CACSD Conf.
Taipei, Taiwan, 2004, URL http://control.ee.ethz.ch/∼joloef/yalmip.php.
68. Mitrovic, D., Graphical Analysis and Synthesis of Feedback Control Systems. I. Theory and Analysis;
II. Synthesis; III. Sampleddata Feedback Control Systems, AIEE Trans. (Appl. and Industry), 1959,
vol. 77, pp. 476–496.
69. Nesterov, Yu. and Nemirovskii, A., InteriorPoint Polynomial Algorithms in Convex Programming,
Philadelphia: SIAM, 1994.
70. Odenthal, D. and Blue, P., Mapping of H
∞
Design Speciﬁcations into Parameter Space, Proc. 11th IFAC
Sympos. Robust Control Design, Prague, 2000.
71. Ozguler, A.B. and Kocan, A.A., An Analytic Determination of Stabilizing Feedback Gains, Report
Institut fur Dynamische Systeme, Universitat Bremen, 1994.
72. Petersen, I., A Stabilization Algorithm for a Class of Uncertain Systems, Syst. Control Lett., 1987, vol. 8,
pp. 351–357.
73. Polyak B.T. Robust Linear Algebra and Robust Aperiodicity, in Direct. in Mathemat. Syst. Theory and
Optim., Rantzer, A. and Byrnes, C.I., Eds., Berlin: Springer, 2003, pp. 249–260.
74. Polyak, B. and Gryazina, E., HitandRun: New Design Technique for Stabilization, Robustness and
Optimization of Linear Systems, Proc. 17th IFAC World Congress, Seoul, Korea, 2008.
75. Qiu, L., Bernhardsson, B., Rantzer, A., Davison, E.J., Young, P.M., and Doyle, J.C., A Formula for
Computation of the Real Stability Radius, Automatica, 1995, vol. 31, no. 6, pp. 879–890.
76. Saadaoui, K. and Ozguler, A.B., A New Method for the Computation of All Stabilizing Controllers of a
Given Order, Int. J. Control , 2005, vol. 78, no. 1, pp. 14–28.
77. Saeki, M. and Aimoto, K., PID Controller Optimization for Hinﬁnity Control by Linear Programming,
Int. J. Robust Nonlinear Control, 2000, vol. 10, pp. 83–99.
78. Saeki, M., Fixed Structure PID Controller Design for Standard H
∞
Control Problem, Proc. 16 IFAC
World Congr., Prague, 2005.
79. Smith, R.L., Eﬃcient Monte Carlo Procedures for Generating Points Uniformly Distributed over
Bounded Regions, Oper. Res., 1984, vol. 32, no. 6, pp. 1296–1308.
80. Siljak, D., Analysis and Synthesis of Feedback Control Systems in the Parameter Plane. I. Linear Contin
uous Systems, II. Sampleddata Systems, AIEE Trans. (Appl. and Industry), 1964, vol. 83, pp. 449–466.
81. Siljak, D., Generalization of the Parameter Plane Method, IEEE Trans. Automat. Control , 1966, AC11,
no. 1, pp. 63–70.
82. Siljak, D., Nonlinear Systems: The Parameter Analysis and Design, New York: Wiley, 1969.
83. Spelta, C. and Gryazina, E., Describing the H
∞
set in the Controller Parameters Space, Proc. Eur.
Control Conf., Kos, Greece, 2007, pp. 816–823.
84. Soylemez, M.T., Munro, N., and Baki, H., Fast Calculation of Stabilizing PID Controllers, Automatica,
2003, vol. 39, pp. 121–126.
85. Vishnegradsky, I., Sur la theorie generale des regulateurs, Comput. Rend. Acad. Sci., 1876, vol. 83,
pp. 318–321.
86. Zhou, K., Doyle, J.C., and Glover, K., Robust and Optimal Control , Upper Saddle River: Prentice Hall,
1996.
This paper was recommended for publication by A.P. Kurdyukov, a member of the Editorial
Board
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
1992
GRYAZINA et al.
investigated in [63], where some results on multiparameter cases can be found. The Ddecomposition technique is described in Soviet textbooks on automatic control, in the western literature it can be found in [32, 38]. The papers [33, 39, 52, 61] show its high eﬃciency for the problems of loworder controllers with H∞ speciﬁcations. The analysis of the geometrical properties of the Ddecomposition is given in [17–19]. In the 80s within the appearance of the problems with uncertainty (robust control) the Ddecomposition seems to be eﬃcient for analysis of the families of linear systems [15]. Within the stability investigation there is a question of interest: if the system remains stable since its parameters are unknown constants belonging to a particular set. These parameters are called “uncertainties.” The most famous results in this ﬁeld are the Kharitonov theorem and its analogs concerning the robust stability of interval polynomials as well as Tsypkin–Polyak locus for interval and ellipsoidal restrictions on the polynomial coeﬃcients; see [21, 24, 30]. The Ddecomposition technique can be applied to the systems in state space form. More specifically, given a class K of r × m matrices K, ﬁnd all the matrices K ∈ K such that A + BKC is stable: D = {K ∈ K : A + BKC is stable}. (2)
Here A, B, C are given real matrices of dimensions n × n, n × r, m × n respectively; stability is understood either in continuoustime sense (all eigenvalues are in the open LHP) or discretetime sense (all eigenvalues are in the open unit disc). Class K may be diﬀerent; below we analyze in detail the simplest cases: K = k ∈ Rn or K = kT , K = kI, k∈R k ∈ Rn
2×2
( m = 1 or r = 1), m = r,
(3) (4) (5)
or k ∈ C, ,
K ∈R
where all calculations can be performed explicitly in the graphical form. Case (3) is equivalent to the polynomial framework, two others are essentially matrix ones. Nevertheless we present general description of Ddecomposition. It is closely related to the standard M − Δ setting. Problem (2) arises in the design or robustness studies. For instance, to ﬁnd all stabilizing static output controllers for the system x = Ax + Bu, y = Cx ˙ (6)
one can construct the set D (2) with K = Rr×m; here K plays the role of the feedback gain. On the other hand, if A is a nominal stable matrix and it is perturbed as A + BKC, where K is a constant r × m matrix; then (2) provides all admissible perturbations which preserve stability. It is obvious that if we know the boundary of the stability domain ∂D, then we can ﬁnd the distance to it: ρ = min K .
K∈∂D
(7)
The quantity ρ−1 is closely related to μ (structured singular value) [86]. If K is the set of all Cr×m (Rr×m ) matrices, then ρ is complex (real) stability radius [55, 75]. Of course, the knowledge of the entire set D provides much more information than the value of ρ. For instance, for design purposes a designer can solve performance or speciﬁcation problems on the set of the stabilizing controllers D. Instead of direct optimization on stability domain D we can get more information on the roots of the characteristic polynomial inside D by similar graphical tools. For instance, if one is interested
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
DDECOMPOSITION TECHNIQUE STATEOFTHEART
1993
in the stability degree σ > 0, then the regions of interest can be obtained replacing jω by −σ + jω in the Ddecomposition equation. This approach has been proposed by Neimark [14] and it was signiﬁcantly developed by Mitrovic [68] and Siljak [80–82]. Of course, we can deal with contours other than shifted imaginary axis, for instance, with a sector or the other root placement. This information on the root placement inside the stability domain is very helpful for the design purposes. However we will not develop this line of research and deal with the simplest situations of the root placement—the left halfplane (for continuoustime systems) and the unit disc (for discretetime systems). Additional detailed decomposition can be a subject of future research. Several examples below illustrate how the technique can be extended to become a practical tool for a designer. On the other hand, we restrict ourselves with a presentation of basic results of the proposed approach and do not consider practical applications. The natural question arises: what is the use of ﬁnding all root invariant domains, while in practice we are interested in the stability domain only? The ﬁrst answer to the question is very simple—we are unable to construct the stability domain separately, the only way suggested by the proposed technique is to provide the complete decomposition of the parameter space and then to pick out the stability regions (if they exist). On the other hand there are some situations when eigenvalue invariant domains are also of interest. For instance, dealing with Nyquist diagram under uncertainty one should guarantee that all transfer functions under consideration have the same number of stable poles. In some cases, we can propose the method for distinguishing stability regions among all the root invariant regions; see Section 2.2 for details. In 1991 the problem of robust Ddecomposition was formulated and partially solved in [20]. In general, the problem is as follows: ﬁrst we select two real (or one complex) parameters, the rest of the parameters are assumed to be uncertain but bounded by norm. The problem is to describe the region in the plane of the selected parameters such that the polynomial is stable for all feasible parameter values. This division for two parameter classes is very typical for the problem of robust ﬁxedstructure controllers. The paper [28] contains analytical expressions of the robust Ddecomposition for two real parameters for an aﬃne polynomial family with lp bounded parametric uncertainties. Another new application of the Ddecomposition technique is the theory of linear matrix inequalities [2, 41]. For this problem statements, the technique allows deﬁning all the regions in the parameter space such that the aﬃne family of symmetric matrices has a ﬁxed number of likesign eigenvalues inside a region [25]. Finally, the randomization technique inside the stability domain seems promising for optimal controller synthesis [74]. The rest of the paper is organized as follows. In Section 2 we consider the rankone case (i.e., singleinput or singleoutput systems), where the stability is determined by the location of the roots of the characteristic polynomial. Section 2.1 is addressed for the classical Ddecomposition for polynomials with one or two real parameters. The main contribution here is the study of the Ddecomposition geometry. In particular, we estimate the number of all root invariant regions as well as the number of simply connected stability regions. Section 2.2 is refereed to the problem of aperiodic stability. Section 3 is devoted to characteristic polynomials with aﬃne uncertain of a certain structure and the arbitrary number of parameters. The stability domain for this polynomials is a union of polytopes in the parameter space. We suggest the method to describe the stability components and to calculate the stability radius for various norm of uncertainty for continuous and discrete systems. Section 4 contains the results on the robust Ddecomposition.
AUTOMATION AND REMOTE CONTROL Vol. 69 No. 12 2008
k) ∞ < γ. k) has l stable roots}. the interval of ω may be reduced to [0. (8) where ai (s). k) = 0). . The equation describing the boundary of regions Dl is called Ddecomposition equations. m loworder controllers. Section 9 contains conclusive remarks. we consider the systems with matrix transfer functions. Since for continuoustime systems a(jω. ω ∈ [0. k) = a0 (s) + i=1 ki ai (s). 69 No. where H(s. ω ∈ (−∞. We parametrize the boundary of the stability domain on the root plane: for continuoustime systems it is the imaginary axis jω. where a graphical representation is possible. π]). For the discretetime systems the diﬀerence is that the decreasing of the degree does not lead to the change of the stable root number. DDECOMPOSITION FOR POLYNOMIALS Consider a linear system with the characteristic polynomial of the form: m a(s. . +∞) or a(n) (k) = 0. k) is a transfer function for the closedloop system. The number of parameters is restricted by two. is called the Ddecomposition. +∞). Thus the Ddecomposition is given by the equations: a(jω. . k) = 0. k) = 0 leads to a(−jω. We consider linear timeinvariant system with a plant given by scalar transfer function G(s) and 3 deﬁne the feasible set of feedback controller C(s. k) = 0 (similarly for discretetime systems a(ejω . one real or two complex roots cross the stability boundary on the root plane. . Section 5 describes the problem of the ﬁxedorder controller synthesis with H∞ speciﬁcations. 12 2008 . where a(n) is the leading coeﬃcient of the polynomial a(s. AUTOMATION AND REMOTE CONTROL Vol. i = 0. 2. (9) (10) a(ejω . 2π). .1994 GRYAZINA et al. . it may be a stability domain. 2π). Section 7 is devoted to the application of the Ddecomposition in the theory of linear matrix inequities. . for discretetime systems it is the unit circumference ejω . k). k) = 0 leads to a(e−jω . For a small number of parameters (1 or 2) on the line or on the plane we plot the regions such that the aﬃne family of symmetric matrices has a ﬁxed number of likesign eigenvalues inside a region. In Section 8 it is shown that the Ddecomposition can be applied in the randomized algorithms for hard problems of synthesis. . k). The controller parameters k ∈ Rm . The ﬁrst subsection relates to the case. ω ∈ [0. Thus Dn is the stability domain D. ∞) (correspondingly [0. The decomposition of the parameter space k into the regions Dl = {k : a(s. Then the number of stable roots may change in one of the following cases: the degree of polynomials changes. The structure of the controller is ﬁxed. The problem is to describe the set of the parameters corresponding to all the controllers such that H(s. Namely. ω ∈ (−∞. m are polynomials of degree no more than n and the parameters k ∈ Rm . Combined with HitandRun algorithm it allows us to generate points asymptotically uniformly distributed on a multidimensional domain (generally speaking nonconvex and not simply connected). In Section 6 we extend the Ddecomposition idea to the matrix case. l = 0. n. In particular. Deﬁnition 1. k) = 0. we deal with polynomials with real coeﬃcients. Further.
1. Development of Graphical Algorithms In the section we consider cases when the number of parameters does not exceed two. . Note that the alternative way to analyze the situation is the root locus technique [3. Theorem 1 [52]. the result was extended to compute stabilizing controllers of a given order [76]. 18]. ω ∈ [0. .35. with the feedback gain k and characteristic polynomial p(s) = a(s) + kb(s) the Ddecomposition is given by a(jω) + kb(jω) = 0. (11) We avoid the situations when a(s). where a(s). Algorithm. um−1 = b(n) /a(n) . k1 ) and (km−1 . k2 < . . We will not highlight these links but will focus on the simplest cases when Eq. K(0) = K0 . K(1) = K1 .e. +∞) such that for every interval ki < k < ki+1 polynomial a(s) + kb(s) has a constant number of stable roots νi . < km < km−1 < +∞ into no more than m (km−1 . . . < km−1 such that ks = − u1 . Solve the polynomial equation Im H(jω) = 0 in 0 ω < ∞ (due to symmetry Im H(jω) = − Im H(−jω)). In many cases we can state that the stability domain is simply connected. m − 2. For polynomial family (11) the real axis may be divided by points −∞ < k1 < n + 2 intervals (−∞. But there exist examples [14. i = 0. The algorithm for ﬁnding the critical values ki and the numbers νi is based on this theorem. Thus (39) is equivalent to −1/k = H(jω) or to the standard Nyquist diagram: the critical values of the gain k (which correspond to a change of the stable roots number for the characteristic polynomial) are deﬁned by the intersections of the Nyquist plot H(jω) and the real axis. ki+1 ) with νi = n) does not exceed n + 2 ( α is the maximal integer less than or equal to α). i = 0. 16. We formulate it here under assumption that polynomials a(s) and b(s) do not have imaginary roots and all critical values ω are simple roots of the equation Im H(jω) = 0. . . Scalar constant gain problem has been solved semianalytically in [71]. b(s) are polynomials of degree n. Order quantities − ui . 12 2008 . However. .32. 2 Note that the maximal number of stability intervals may be attained for the case of odd n when both (−∞. . 69 No. This link between robust stability and Ddecomposition has been emphasized in [15]. k2 ). k1 ). i s AUTOMATION AND REMOTE CONTROL Vol. Step 1. K1 ∈ D there exist the continuous parametrization K(t) such that K(t) ∈ D. .38. Ddecomposition approach allows one to ﬁnd critical points analytically and to estimate their number. +∞) intervals are stability intervals. the total number of the root invariant regions is also unknown. ∞). (k1 . Remind that the set D is simply connected if for any K0 ∈ D.. Moreover the number of stability intervals (i. . b(s) have a common imaginary (or zero) root. m − 1 and denote them k1 < .62] has been pointed out by Chen [43]. . where it is not true. < ωm−2 and calculate ui = Re H(jωi ). b(s) For singleinput singleoutput systems with a scalar transfer function H(s) = a(s) . . It is not clear what is the maximal number of connected stability regions. Denote the obtained solutions 0 = ω0 < ω1 < ω2 < . intervals (ki . (8) provides graphical tools to describe Ddecomposition in the space of parameters k. The case of multiple ω requires additional investigations and we omit it here. Moreover. . the connection between μanalysis and the known results on parametric robustness [21.DDECOMPOSITION TECHNIQUE STATEOFTHEART 1995 2. 1 Step 2. 0 t 1. . Linearity of Ddecomposition equation allows one to ﬁnd easily the stability radius (7) for various lp norms of the vector k. Parameter space is the plane and the root invariant regions can be graphically represented. On the other hand. . The case of the singleinput singleoutput systems with small number of parameters is the classical framework of the Ddecomposition. . 49]. .
Δ1 = . 12 2008 . k2 (ω). 2. 53]. 69 No. and for any point from the region k1 (k2 +1) 0 the polynomial has a pair of imaginary roots. We call these lines singular lines. The numbers νs (s = 1. it may have discontinuity at these points. . ⎨ and equate ν with the number of stable Step 4. a0 + k1 a1 + k2 a2 (n) (n) (n) = 0. For the case of two real parameters. Solnechnyi) the system of the Ddecomposition equations reduces to one equation in two variables because Im a(jω. the Ddecomposition for the polynomial family a(s. . Theorem 2 [6. . Δ k2 = − Δ2 . Δ (13) For Δ = 0 the formulae (13) deﬁne the parametric curve k1 (ω). Curve (13) starts at a point on the ﬁrst line (for ω = 0) and terminates at a point on the other line. We call these lines boundary lines. k2 ) = 0. The intersection of these lines speciﬁes one root (0 or ∞) to cross the stability boundary. the Ddecomposition reduces into a straight line. s H(jωis ). Im a1 (jω) Im a2 (jω) Im a0 (jω) Im a2 (jω) Im a1 (jω) Im a0 (jω) Δ1 . It is symmetric in ω: ki (−ω) = ki (ω). Note that the singular ω such that Δ = Δ1 = Δ2 = 0 may ﬁll a particular interval. ϕ(ωis ) = 0 ωis = 0. (12) is given by a(jω. then curve (13) goes to inﬁnity. The value of ω such that Δ = Δ1 = Δ2 = 0 is called singular frequency. for the polynomial a(s. k) ≡ 0.1996 GRYAZINA et al. Find the interval containing zero: 0 ∈ (k . Ω being large enough.M. ⎪ s ⎩ ν ∓ sgn ϕ(Ω). i = 1. and two roots cross the stability boundary with crossing this curve. ⎧ ⎪ νs ∓ 2sgn ϕ(ωis ). In this case. Resolving this equation with the respect to the parameters we obtain the curve k1 = − where Δ= Re a1 (jω) Re a2 (jω) Re a0 (jω) Re a2 (jω) Re a1 (jω) Re a0 (jω) . k) = k1 s2 + k2 + 1 (proposed by E. The above mentioned consideration is valid for the continuoustime systems. Δ2 = . If Δ = 0 but Δ1 = 0 or Δ2 = 0. there are two straight lines ω=0: ω=∞: a0 (0) + k1 a1 (0) + k2 a2 (0) = 0. and for family (12) with two real parameter. But using fractional transformation z+1 s= z−1 AUTOMATION AND REMOTE CONTROL Vol. ωis = 0. thus it suﬃces to take for continuoustime systems ω ∈ (0. k1 . For instance. m) are calculated successfully: νs±1 = where ϕ(ωis ) = d dω Im ν ∓ sgn ϕ(0). For family (11) with one complex parameter. . ϕ(ωis ) = 0 ks = −a(n) /b(n) . k) = a0 (s) + k1 a1 (s) + k2 a2 (s). k2 ∈ R. Moreover. k1 . the number of Ddecomposition regions does not exceed N (n − 1)2 + 2. Thus the line −k1 ω 2 +k2 +1 = 0 is singular for all ω. k roots of the polynomial a(s). N 2n(n − 1) + 3. ∞). +1 ) Step 3.
This motion is a superposition of two rotations. n Example 2 [18]. the polynomials is stable when all its roots are in the left half plane. For the polynomial a(z. k) = 0 and plotted by solid lines..6 0.5 –2.” The aperiodic property of the linear system (i. This curve is generated by a moving point on the complex plane.5 Im k 0 –0. and a(z. In this ﬁgure (as well as in the further ﬁgures) the digit inside the region marks the number of stable roots. The structure of the regions for n = 6 can be seen in Fig. The Ddecomposition for the polynomial a(z.0 2 1997 (b) 2 6 6 5 3 4 6 6 5 3 k2 –2 –1 0 k1 1 2 Fig. i. 69 No. k) = z n + kz n−1 + α. this curve consists of n arcs. depicted in Fig.4 –0.6 –0. 0 ≤ ω < 2π.4 0. The Ddecomposition is given by the parametric curve k(ω) = −ejω − αe−jω(n−1) . First rotation has radius 1 and frequency 2π while the second one has radius α and frequency (n − 1)2π. k) being the discretetime polynomial and it is stable when all its roots are inside the unit circumference. has (n − 1)2 + 1 root invariant regions for α > 1. 1a. where k ∈ C. k) the continuoustime polynomial. Thus the calculation of the number of intersections gives N = n2 − 2n + 2. 2.0 –1. k) = 0. this is just one region less the upper bound given by Theorem 2. Ddecomposition for Aperiodicity Remind that the polynomial a(s) is called aperiodic if all its roots are simple real and negative [9]. Deﬁnition 1 is modiﬁed as follows: AUTOMATION AND REMOTE CONTROL Vol. Example 1 [6].2 0 –0. Most of the examples are demonstrative in the discretetime version but the demonstrated features are preserved also for the continuoustime systems. the system with aperiodic characteristic polynomial) guarantees that the transient process is not oscillating.DDECOMPOSITION TECHNIQUE STATEOFTHEART (a) 2.0 –2 –1 0 Re k 1 2 3 4 1 2 5 0 0.2 –0. We consider the same aﬃne polynomial family (8). 1b. 1 < α < n−2 there are n − 1 simply connected stability regions in the {k1 . three regions (marked by the digit 6) are in the loops of the boundary curve and the other two regions are formed by the intersection of the curve and boundary lines given by the equations a(ej0 . k) = z n + k1 z n−1 + αz n−2 + k2 . For the discretetime system with the polynomial a(z) the deﬁnition is similar but “negative” should be replaced by “absolute value less than one. k2 } plane. Thus due to conformity the results remain valid for the discretetime systems also. Here and elsewhere we denote by a(s. For α > 1..0 0. 2π) rather than inﬁnite interval for the continuoustime systems. a(ejπ .e. 1.8 –1. each arc intersects any other twice.5 1. 12 2008 .e.5 –1. Ddecomposition in (a) Example 1 and (b) Example 2. This property is desirable in many cases so it is natural to formulate the problem of describing aperiodic domain in the parameter space.2.0 1. allows us to proceed to the discretetime systems and vice versa. and two root invariant regions for α < 1/(n − 1).8 0. We prefer discretetime systems because it is easier to make the Ddecomposition analysis for the ﬁnite frequency interval ω ∈ [0.
t → –∞ 5 A3 4 3 k2 t = –1 2 D3 1 0 0 1 2 k1 Fig. or the condition a(n) (k) = 0. the transition from one Al to the other occurs only when a multiple real or zero root appears or the degree of polynomial changes. thus the whole curve lies in the positive ortant and it is the boundary of the aperiodic stability region (Fig. 3t2 + 2k1 t + k2 = 0. AUTOMATION AND REMOTE CONTROL Vol. Stability and aperiodicity regions for Visgnegradsky exmaple. (The notations are the same as in the beginning of the section. t → –0 3 4 5 Deﬁnition 2. . (14): t3 + k1 t2 + k2 t + 1 = 0. Expressing k1 . 12 2008 k ∈ R. . the polynomial of the form p(s. l = 0. . . k) = s3 + k1 s2 + k2 s + 1 the boundary is deﬁned just by Eq.) Speciﬁcally. k) = 0. is called Adecomposition. 2).e.1998 GRYAZINA et al. The decomposition of the parameter space k into the regions Al = {k : a(s. denotes diﬀerentiation in t. k) has l real simple negative roots}. Equations (14) for this case can be rewritten as: a(t) + kb(t) = 0. k) = 0. . 2. for the cubic polynomial in Vishnegradsky form a(s. n. or the equation a(0. 1. k) = 0. a (t. k2 in terms of t obtain: 1 2 k1 (t) = 2 − 2t. k2 (t) = t2 − . 0) is the value of the real multiple root. k) = a(s) + kb(s).. i. t t Remind that t < 0. a (t) + kb (t) = 0. Thus the boundary of the Adecomposition is deﬁned by the system of equations a(t. (14) where t ∈ (−∞. Obviously. Now consider the case of one real parameter. 69 No. This result was obtained by Vishnegradsky [85] and it is published in old textbooks on automatic control.
A similar result for the discretetime systems [48] is mentioned in the survey [10] among other nonKharitonov approaches to robustness of the discretetime systems. max{deg(f ) + deg(ai ). and antisymmetric polynomial if ai (z) ≡ −z n ai (z −1 ) (for instance. deg(f ) = n1 . f are the polynomials with real coeﬃcients. i = 0. . This approach recently proved well in the combined probabilistic/deterministic methods [52. and the polynomials ai (z). i = 1. v contain only even i degrees of s. We note that in the pure form the polynomials of this kind are encountered rarely. Thus the ray (−∞. . . This fact allows us to obtain the analog of Theorem 1 and an algorithm for construction of the Adecomposition of the parameter space. z 3 + z is a symmetric polynomial of the fourth degree). for example. Bhattacharyya [38. N. where the stability domain was constructed for the PIDcontroller. The boundaries of the regions with the constant number of stable roots according to (9) are described by the system of AUTOMATION AND REMOTE CONTROL Vol. (16) where deg(a0 ) = n0 . 0) can be divided into no more than 2n intervals. It is also worth mentioning the works of J. 3. Ackermann [33]. 69 No. by ﬁxing the parameters at the odd degrees. . 47] were the ﬁrst studies in this direction which considered the continuoustime case with f (s) ≡ 1 and the stability domain described by a system of linear inequalities. kv } ∈ Rv . S. deg(a0 ) = n0 . . Then q(t) is a polynomial of degree no more than 2n − 1 and it has no more than 2n − 1 real negative roots t1 < t2 < · · · < tm < 0. n0 } = n. 45]. . Let a(s). we show that an algebraic method for detecting the stability domain discussed in the previous sections works for the multidimensional parameter space. The results of this section are valid also for the discretetime systems of the form v a(z. 2. but the problem often may be reduced to the considered family. and obtain q(t) = a (t)b(t) − a(t)b (t) = 0. k = {k1 . k) = a0 (z) + f (z) i=1 ki ai (z). Here we consider only continuoustime systems. we consider the families of the form: v a(s. . We recall that the polynomial ai (z) is called the symmetric polynomial of degree n if ai (z) ≡ z n ai (z −1 ). m 2n − 1. deg(f ) = n1 . Further. b(s) be the polynomials of degree n. v. the stability domain can be described constructively in the multidimensional (not only in two dimensional) parameter space.DDECOMPOSITION TECHNIQUE STATEOFTHEART 1999 Suppose that a(s) and b(s) do not have a common negative real root. . . For the continuoustime systems. . . 23]. .P. 57]. . . . Then we can eliminate k. We apply the Ddecomposition idea for continuoustime system (15). In some special cases. It seems that the works [64. kv } ∈ Rv . v are simultaneously symmetric or antisymmetric as the polynomials of degree n n0 − n1 . . . . we don’t deep into details here. CONSTRUCTION OF THE MULTIDIMENSIONAL STABILITY DOMAIN In this section we proceed to the consideration of the singleinput or output systems with the characteristic polynomial linearly depending on the parameters. 12 2008 . . Munro [84]. k = {k1 . 1. k) = a0 (s) + f (s) i=1 ki ai (s2 ). i = 1. for more details see [8]. Note that all the polynomials ai . (15) where ai . The problem of the robust aperiodicity can be solved using the technique described in [21. .
1. 1}. 1. 69 No. (17) and (18) are linearly dependent and describe the hyperplane in the parameter space which is the boundary of the Ddecomposition: v i p(ωi ) + f (jωi )2 =1 2 k a (−ωi ) = 0. . i = 0. In turn. . For ω = ∞.2000 GRYAZINA et al. k) = a(jω. (17) and (18). < ωm−1 such that f (±jω) = 0 are called critical frequencies. m. (19) without parameters corresponds to the imaginary part of the polynomial a∗ (jω. (20) where p(ω) = Re a0 (jω)Re f (jω) + Im a0 (jω)Im f (jω). . The generalization of the Hermite–Biehler theorem [58] enables one to exploit the idea of shading for the multidimensional case for the polynomials of a certain form (15). Any root invariant region Dl of polynomial (15) is a union of nonintersecting polytopes. For each critical frequency Eqs. equations: v Re a0 (jω) + Re f (jω) i=1 v ki ai (−ω 2 ) = 0. we denote ωm = ∞. any such region being a convex polytope. . . ω ∈ [0. Equation (20) is the real part of the polynomial a∗ (jω. this value corresponds to a change on the degree. (17) (18) Im a0 (jω) + Im f (jω) i=1 ki ai (−ω 2 ) = 0. Let some region in the parameter space be described by the system of linear inequalities (21) for a ﬁxed set {ξi }. 53]. an algebraic method of determining the number of stable roots in various domains was proposed in [6. For the oneparameter family. Moreover. These hyperplanes divide the parameter space Rv into the regions with constant number of stable roots. . . ∞]. k)f (−jω). k). . The simplest method is to take a point in every region and to calculate the roots of the corresponding polynomials. (21) Assertion 1 [8]. . m. Since the space of parameters is divided into domains with constant number of stable roots. The classical Ddecomposition technique prescribes to use shading that shows the extent of changes in the number of stable roots at the crossing of the curves of Ddecomposition. i = 0. . m. every region is a solution of the system of linear inequalities with the respect to the parameters: v ξi p(ωi ) + f (jωi ) 2 =1 2 k a (−ωi ) > 0. if the equality deg(a0 ) = max{deg(f ) + deg(ai )} holds. where ωi are the zeros of the imaginary part of the original AUTOMATION AND REMOTE CONTROL Vol. it is equivalent to the system of Eqs. already for R3 this technique becomes hindered. the equations contain coeﬃcients with the degree n only. we obtain an equation without parameters: Re a0 (jω)Im f (jω) − Re f (jω)Im a0 (jω) = 0. 12 2008 . Note that Eq. . ξi ∈ {−1. Eﬃcient use of shading in general form is possible only for a low dimensionality of the parameter space. then there are several methods of determining the number of stable roots in each domain. i = 0. (19) The simple real positive roots of this equation 0 = ω0 < ω1 < . v Expressing i=1 ki ai (−ω 2 ) from one equation and substituting in into another. . . At the points ωi . . 1.
k)f (−s) has l + r1 stable roots. Example 3. The generalization of the Hermite– Biehler theorem allows one to establish what relation sgn Re a∗ (jωi ). . + (−1)m−1 2ξm−1 + (−1)m ξm )(−1)m−1 sgn(Im a(j∞)). . The advantages of the considered method manifest themselves with increase in the number of parameters because solution of systems of linear inequalities still remains the main operations. k) has the certain number of l stable and r unstable roots. 1 − ω 2 = t ∈ [−1. within this region the polynomial a(s. that is. + (−1)m−1 2ξm−1 )(−1)m−1 sgn(Im a(j∞)). f (jω) = Tm (1 − ω 2 ) + jωTm−1 (1 − ω 2 ). the polynomial a(s. Since by construction the region boundaries satisfy the equations of Ddecomposition. For m = 4. 69 No. Consider the polynomial a(s. . . m provides given signature. k) = a(s. Stability domain in the parameter space in Example 3. . Then.DDECOMPOSITION TECHNIQUE STATEOFTHEART 2001 100 k2 0 100 –100 –100 0 k0 100 –100 0 k1 Fig. The diﬀerence l−r (that is called the signature of the polynomial σ(a)) in this region [58] is deﬁned by the set of signs of the real part at the points ωi . at that. These polynomials are deﬁned for all s ∈ C. the number of permissible sets {ξi } is four. 1].1+f (s)(k0 +k1 s2 +k2 s4 ). Thus without loss of generality we use the representation for the Chebyshev polynomials of the form Ti (t) = cos(i arccos t). i = 0. (m/2)+1 sets {ξi } each deﬁning a system of and due to combinatorial estimations there exist Cm inequalities describing a component of the stability domain. at that. and as shown in Fig. k) = 0. 3. . which corresponds to ξi = sgnRe a(jωi ). 1]. Let f (s) has l1 roots to the left of the imaginary axis and r1 roots to its right. The imaginary part f (jω) has m diﬀerent simple positive roots. solution of inequalities for all such {ξi } turns out to be nonempty. if n is even (ξ0 − 2ξ1 + . if n is odd. AUTOMATION AND REMOTE CONTROL Vol. but their root are located in the segment [−1. k) has l stable and n − l unstable roots if a∗ (s. deg(a) = 2m + 4. 1. polynomial which contains no parameters. σ(a) = (ξ0 − 2ξ1 + . and these signs specify the choice of ξi for the system of inequalities (21). where f (s) is a stable polynomial. where Ti (t) are the Chebyshev polynomials. . . i = 0. . 12 2008 . . σ(a∗ ) = l + r1 − (n − l) − l1 = 2l − n + r1 − l1 . deg(f ) = 2m. . 3. m. .
3. need to determine the maximum value γ ∗ such that family (22) is stable for all γ < γ ∗ . For the family under consideration. Stated diﬀerently. be v 2 α2 ki i i=1 bounded in the socalled octahedral norm. 1 p (22) where α = {α1 . Stability Radius Let in the original aﬃne family (15) which is certainly stable for k = 0. the parameters can assume values from the simplex i=1 αi ki  γ. i = 0. (23) where ϕi are the values of the corresponding polynomials on the critical frequencies ωi . m. Then γ ∗ = inf{ k α p : k ∈ K} = 1 ϕ 1/α q = 1 v i=1 ϕi q αi 1/q . and for p = 2. . v With this formulation. . ϕ) = 0} be given. there exists a polyhedron deﬁned by the system of linear inequalities (23). for p = ∞ the parameters can vary independently. . We set down the inequalities specifying the stability region as: v ϕi + 0 =1 k ϕi > 0. v k α p γ. 0 Now we are interested in the stability radius. the boundary of the stability domain is deﬁned by several ∗ hyperplanes. 12 2008 . . The aforementioned in summed up by the following theorem. . the elliptic constrains like γ2 are imposed on the parameters. 69 No. i = 0. the proposed description of the stability domain allows to handle also an arbitrary pnorm. ϕi = sgn (p(ωi ))f (jωi )2 a (−ωi ). .2002 GRYAZINA et al. αv }. . k) = a0 (s) + f (s) i=1 ki ai (s2 ). AUTOMATION AND REMOTE CONTROL Vol. If the polynomial a0 (s) is stable the theorem holds: Theorem 3 [8]. each within its interval γ. Therefore it suﬃces to take the minimum for all γi corresponding to each hyperplane in order to answer the question of the stability radius. that is. Let in Rv the halfspace K = {k : 1 + (k. the parameters arbitrarily remaining bounded in some weighted pnorm: v a(s. . The dualnorm lemma enables us to answer this question for each hyperplane: Lemma 1. that is. 1. where p(ω) = Re a0 (jω)Re f (jω) + Im a0 (jω)Im f (jω). . . . and we have to inscribe in it a sphere in the corresponding norm of the maximal radius. for p = 1. . m.1. namely: 2 ϕi = sgn (p(ωi ))p(ωi ) = p(ωi ). where 1 p + 1 q = 1. The component of the stability domain of family (15) containing the nominal stable polynomial a0 (s) obeys the system of inequalities: v sgn (p(ωi )) p(ωi ) + f (jωi )2 =1 2 k a (−ωi ) > 0. αi > 0 are some weights k αi ki  α p = i=1 αi ki  p .
Further we consider continuoustime systems and characteristic polynomials with real coeﬃcients. For the polynomial a(s. . k being controller parameters. m is the system of inequalities describing the stability domain. Suppose that a characteristic polynomial depends on two types of parameters k and x. n) in each of which the number of stable roots is the same for all admissible uncertainties: Dl = {k : a(s. Υ (ω) = {k : a (j ω. The ﬁnal answer is obtained by taking ∗ whose number corresponds to the number of the minimum from the ﬁnite set of the elements γi critical frequencies. x) may have various number of stable roots depending in the value of x ∈ X . for interval polynomials. 15]: Deﬁnition 3. ∃x ∈ X . Note that there are other classes of multiparametric polynomials that allows one to calculate stability radius easily. (19) we have to determine the critical frequencies and then set down the system of inequalities (23) deﬁning the stability domain. . . . x) = 0. ΥΩ is a swept area. The number of parameters of the ﬁrst type does not exceed two while for the others x ∈ R the domain of variation X is given. AUTOMATION AND REMOTE CONTROL Vol. The set ΥΩ is a sort of smear of the classical Ddecomposition curve into the strip. k. In contrast to the classical Ddecomposition it consists not of the curve and several lines but it is a set given by two conditions from zero exclusion principe (compare to (9)): Υs = k : a(n) (k. 69 No. ∞)} = ω∈[0. while Υs is smear of the singular line deﬁning by the order reduction condition. Thus Dn is the robust stability domain.. and a boundary is the envelope of the family Υ(ω). x) = 0.∞) (24) Υ(ω). k. ∃ω ∈ [0. In this section we discuss the methods for describing the stability domain in the parameter space k. x) has l stable roots. where ϕi + (k. ROBUST DDECOMPOSITION FOR POLYNOMIALS Consider systems with uncertainty. k ∈ R2 or k ∈ C. . 22].DDECOMPOSITION TECHNIQUE STATEOFTHEART 2003 ϕi 0 ϕi i 1/α q Theorem 4 [8]. ∃x ∈ X . robust Ddecomposition is the decomposition of the parameter space k into regions Dl (l = 0. for each hyperplane we have to calculate the corresponding ∗ norm of the normal vector ϕi and the corresponding γi . 0 This result is exceptionally simple in terms of computations. The Ddecomposition idea is also eﬃcient in this case [20. A system is robustly stable (in x). For instance. k. 12 2008 . At ﬁrst. i. . x) of the degree n with parameters k. . ∃ x ∈ X }. if for a given k it is stable for all x ∈ X [24]. k.e. the answer is obtained using Tsypkin– Polyak locus [21. for the discretetime systems (see remark). Now. ∀x ∈ X } . (25) where a(n) is the leading coeﬃcient. In this case Υ(ω) is a parametric family of the ﬁgures in the plane. We consider robust Ddecomposition for two real or one complex parameters. x) = 0. 4. i = 0. First we deﬁne the set Υ that separates the regions Dl . k. This problem statement is natural for the problem of controller synthesis under uncertainty. ΥΩ = {a(jω. The stability radius for the aﬃne families (22) obeys γ ∗ = min . Within the stripes polynomial a(s. . by solving Eq. ϕi ) > 0.
x p 1. bi is n. 4. ⎫ ⎪ ⎪ ⎪ ⎪ ⎬ ⎤ ⎥ ⎥ ⎥. 69 No. ⎪ ⎪ ⎪ ⎪ ⎭ xi bi (jω) i=1 Set BX depending on the uncertainty may be [28. p = {1. x∈C . x) = a0 (s) + ka1 (s) + i=1 xi bi (s). For the ellipse Υ(ω) is deﬁned by Υ(ω) = k : (T (ω)k − b0 (ω))T M (ω)−1 (T (ω)k − b0 (ω)) ρ(ω)2 . a line.1. x) = a0 (s) + k1 a1 (s) + k2 a2 (s) + i=1 xi bi (s). In this case the set Υs is bounded by two lines: a0 + k1 a1 + k2 a2 ± b(n) (n) (n) (n) q = 0. where . b(n) ∈ R . k. ∞}. Robust Ddecomposition for One Complex Parameter For the robust Ddecomposition for one complex parameter k of the aﬃne polynomial a(s. ∞). • a circle if x ∈ C . 28]. Im a0 )T and the particular from of M and ρ depend on the uncertainty. T (ω) = ⎧⎡ ⎪ ⎪ Re ⎪ ⎪⎢ ⎨ . ⎢ BX (ω) = ⎢ ⎪⎢ ⎪⎣ ⎪ Im ⎪ ⎩ (28) Re a1 (jω) Re a2 (jω) Im a1 (jω) Im a2 (jω) a0 (jω) + a0 (jω) + xi bi (jω) i=1 . In the nonsingular case Υ(ω) is obtained by linear transformation BX (ω) (28). If one of the matrices T or M is singular then Υ(ω) may be an empty set.2004 GRYAZINA et al. (26) The maximal degree of polynomials ai . s strip or the whole plane. If both matrices are nonsingular the envelope of the elliptic family Υ(ω) is obtained analytically as the solution of the system of two secondorder equations [12. x p ⎥ ⎦ (29) 1 . Robust Ddecomposition for Two Real Parameters Consider the linear system with the characteristic polynomial linearly dependent on all the parameters with uncertainty x bounded in lp norm (or weighted lp norm): 1/p a(s. ω ∈ [0. In the case of discretetime systems. (27) where index (n) denotes the sn held coeﬃcients of the corresponding polynomials. p ∈ [1. (30) AUTOMATION AND REMOTE CONTROL Vol. where k = (k1 . Remark. x p = i xi  p 1. 4. k2 )T . q = p/(p − 1) is the dual norm index. For the polygon the problem of description the envelope is hard. b0 = −(Re a0 . 2π) is used and variation of the degree of the polynomial does not lead to the change of the number of stable roots.2. k. 20]: • a 2 polygon if x ∈ R . • an ellipse if x ∈ R . Set Υ(ω) for polynomial (26) in the nonsingular case is Υ(ω) = −T −1 (ω)BX (ω). 12 2008 . the substitution s = ejω . p = 2. thus it is either a polygon or an ellipse.
if τ0 (ω) = ρ(ω)  = 0. the block diagram of the system is shown in Fig. 83]. ⎭ a1 (jω) = 0. The results are published in [7. xi  1. i = 1. Let us consider the robust Ddecomposition for one complex parameter for the discretetime system with polynomial (robust version of Example 1): z 6 (1 + 0. 60. Let us consider a linear stationary onedimensional system. Further we assume that there are only two or three adjustable controller parameters that allows us to use widely the graphic methods. Ddecomposition without uncertainty is plotted by dotted curve. 1a. 5.. Additionally. 69 No. if any. . Ω Ω where Υreg Ω k0 (ω) ⎝ ρ(ω) ±j 1− = k0 (ω) − ρ(ω) ⎩ k0 (ω)  k0 (ω)  k : a0 (jω) + ka1 (jω) = b(jω) q ⎧ ⎨ ⎛ ρ(ω) k0 (ω)  2 ⎫ ⎬ ⎠ : k0 (ω)  ρ(ω)  > 0. 5. ∂ΥΩ ⊂ Υreg ∪ Υspec. Let the system be closed by the controller C(s. 78]. i.DDECOMPOSITION TECHNIQUE STATEOFTHEART 2005 boundaries ∂Υs and ∂ΥΩ of sets Υs and ΥΩ can be obtained explicitly: ∂Υs = k : a0 + ka1 (n) (n) = b(n) q .5 + 0. 70]. ω ∈ R if these derivatives are not deﬁned q = p/(p − 1). . 50. and s s AUTOMATION AND REMOTE CONTROL Vol. their small variation often results in instability [59].e.05x1 ) + kz 5 + 1. (31) The ﬁgure of Ddecomposition is shown in Fig. k0 (ω) = −a0 (jω)/a1 (jω). We recall that the H∞ norm is ﬁnite only for rational functions with stable denominator (H(s) ∈ RH∞ ) and equals to H(s) ∞ = sup H(jω). ω ∈ R ⎞ Υspec = Ω . It is important that our approach describes the whole set of controllers with given speciﬁcations. Thus the problem is to ﬁnd the regions in the parameter space such that the corresponding controllers (i) stabilize the system and (ii) satisfy the H∞ criterion. Parameters k ∈ Rm . see also Fig. xi ∈ C. Here we consider H∞ criteria. SYNTHESIS OF LOWORDER CONTROLLERS WITH H∞ SPECIFICATIONS The problem of designing the loworder controllers is to determine a stabilizing controller such that the closedloop system is stable and some addition quality criterion holds. 86] but the resulting controllers may be of a very high order. We present an alternative approach to the design of a given structure controllers. Since the H∞ theory does not allow one to constrain the order of the designed controller. stability of the closedloop system is very sensitive to the controller parameters. z = ejω . Example 4. 54. regions Dl are denoted by digits l. the order of the controller is ﬁxed and it is the only choice of its parameters remains free. For some particular controllers the description of the whole domain of admissible parameters in the parameter space is proposed in [40. sometimes even exceeding that of the original system [11]. For instance. 56. . the PIDcontroller kp + ki + kd s. it may be the PIcontroller kp + ki . ω The analytical theory of the design of the H∞ optimal controllers has been developed in detail in [5. 77. . 4. k).05(x1 + 2x2 ) = 0. 12 2008 . ρ(ω) = b(jω) q /a1 (jω). 2. The other approaches to the ﬁxed order controller synthesis with H∞ speciﬁcations are discussed in [52. k ∈ C. where the plane is given by the scalar transfer function G(s). its direct application to the design of controllers of a given structure faces substantial diﬃculties. m 3 deﬁne the set of admissible loworder controllers.
2 D1 1 Im k 0 D2 D6 D3 –1 D4 –2 –2 –1 0 Re k 1 2 Fig. 5. The particular form of the transfer function H(s. The problem is to describe the set of the parameters common to all controllers satisfying the performance criterion H(s. k). k) may be diﬀerent. k) ∞ < γ. H∞ criterion arises in several formulations of the design problems of which we present the most important examples as well as the corresponding functions H(s. T (s) = 1+C(s)G(s) is the transfer function of the closedloop system (additional sensitivity). k) C(s) Fig. Block diagram of the control system. 4. AUTOMATION AND REMOTE CONTROL Vol. Robust Ddecomposition in Example 4. C(s)G(s) −1 where W1 (s) is a weight function (W1 ∈ RH∞ ). and γ is the given level of performance. 69 No.2006 GRYAZINA et al. • Determination of the controller C(s) guaranteeing the H∞ performance of the closedloop system. – C(s. 1 s+k the ﬁrstorder controller ks+k3 2 . 12 2008 . W1 (s)T (s) ∞ γ. (32) We call these controllers the H∞ controllers.
. ∞). which is the case for all considered types of the controllers (PI.∞) Further we assume that the numerator Hn and denominator Hd depend linearly on the parameters k. the closedloop system must satisfy the following constraints: Determination of a controller robustly stabilizing the plant family with additive uncertainty G(s) = G0 (s) + ΔG(s). 12 2008 . The problem comes to satisfying the criterion W2 (s)U (s) ∞ m1 (ω) < W (jω)T (jω) < m2 (ω). k) = γ. k). (34) and equations H(j∞. k) = 0 Hd (jω. where N (s) is the numerator of the transfer function of the plant G(s) and D(s) is its denominator. Hd ω∈[0. is sensitivity and m1 . There may be three essentially diﬀerent cases: either the function H(jω. for the PIDcontroller and the sensitivity function Hn (s. where G0 (s) is the transfer function of the nominal plant and the frequency 1. l1 and l2 are the limitative functions. Theorem 5 [7]. The set of parameters k corresponding to H∞ controllers is an intersection of K and the stability domain for the characteristic polynomial. where U (s) = 1+C(s)G(s) .k) is the common rational transfer function. The points where only the denominator is zero do not belong to the boundary because their small neighborhood has not a single point where the inequality Hn (jω. The criterion W2 (s)T (s) ∞ 1 corresponds to the multiplicative uncertainty G(s) = G0 (s)(1 + ΔG(s)). k) = Hn(s. ω1 ]. For instance. k) ∞ = sup H(jω. its numerator and denominator vanish simultaneously. i. (35) (36) hd (k) = 0. ω ∈ [ω2 . 5. k) = sD(s). k) = 0. where hd is coeﬃcient at the higher order of a polynom Hd . k)2 = 0. k) < γHd (jω. ∀ω ∈ [0. ∞). where S(s) = 1 1+C(s)G(s) ω ∈ [0. ∂H(jω. Similarly to the classical Ddecomposition approach. or the degree of a denominator will decrease. Assume that the controller stabilized the system. we consider the crossing the boundary of K. m2 . 69 No. ∞). k) is deﬁned we use suﬃcient extremum condition with the respect to ω to obtain an easytouse theorem. k) < γHd (jω. ∂ω H(0. k).e. where H(s. The boundary of the set K is contained in the solutions of the systems Hn (jω. AUTOMATION AND REMOTE CONTROL Vol. Hd (s. Then (s. (nonparametric [38]) uncertainty ΔG(s) is bounded in the weighted norm W2 (s)ΔG(s) ∞ C(s) −1 1. k) = sD(s) + (ki + kp s + kd s2 )N (s). PID. (33) ⎧ ⎨ H(jω. see [50.DDECOMPOSITION TECHNIQUE STATEOFTHEART • 2007 −1 W2 ∈ RH∞ . ∞)} in the explicit form. ω ∈ [0. T (jω) < l2 (ω). Ddecomposition for H∞ Criterion One of the approaches to deﬁne the admissible set of H∞ controllers is based on the description of the boundary of the set K = {k : Hn (jω. ω1 ]. the denominator H(s) is stable. • In the quantitative feedback theory (QFT) [5]. Whenever H(jω. k)2 = γ 2 ⎩ ω ∈ [0. ﬁrstorder controller). S(jω) < l1 (ω).1. 86]. ω ∈ [0.k) H(s. k) reaches the value γ. k) holds. k) = γ.
It is required to determine all the controllers satisfying the H∞ criterion of performance W (s)T (s) ∞ 1 with highfrequency ﬁlter W (s) = s+0.5.2008 GRYAZINA et al. Checking each of the regions we ﬁnd out the desirable one marked with “×. ki . 69 No. (35). points not belonging to the desired boundary may occur among the solution. kd ) = i p s d . the controller. by C(s. Eq.20 –0.25 –0.0 0.05 Fig.05 ki 0 0. s+1 The boundary of the domain K in the parameter space is speciﬁed by Theorem 5.0 –0.35. Deﬁne the matrix transfer function: M (s) = C(A − sI)−1 B for the continuoustime systems and M (z) = C(A − zI)−1 B (38) (37) k +k s+k s2 for the discretetime systems.5 0 kd –0. Equation (36) describes an opportunity of loss of stability of the closedloop system. DDECOMPOSITION FOR MATRICES Consider real matrices A ∈ Rn×n .15 –0. B ∈ Rn×r . AUTOMATION AND REMOTE CONTROL Vol.2 .8s−0. 6. it is the aggregate of points that are the solution of system (34). 6. and does not have eigenvalues on the unit circumference in the discretetime systems. Equation (33) gives the vertical line ki = 0 (a part of this line is also the stability domain boundary). the whole this region is inside the stability domain that is plotted with chain line.1 . Example 5. The plant is deﬁned by the transfer function G(s) = s2 +0. For the ﬁxed k function H(jω) may be at its maximum either at the solution of system (33) or system (34) or at the solution of Eqs. and system (34) gives the curves depicted in Fig.10 –0. The second equation of the system enables one to specify such k on the curve the function H2 reaches its extremum for the given ω. C ∈ Rm×n .” Indeed. Assume that the matrix A does not have zero or imaginary eigenvalues for the continuoustime case. Since the condition for equality to zero of the derivative keeps one from discrimination whether it is the maximum or minimum that is reached at this point. where the variables s and z are using to distinguish between these two cases. and let K be a connected set of real or complex matrices r × m. 6. 1. 12 2008 . Consider the application of the proposed method for designing a ﬁxedorder PIDs−1 controller [46]. For the two variables k = (k1 . kp = −0. The ﬁrst equation deﬁnes for each ω either a onedimensional curve or an empty set. H∞ region for PIDcontrollers.5 –1. k2 ). (35) gives the horizontal line kd = −0.
if Q ⊂ K is a connected set and det(I + M (jω)K) = 0. 69 No. . where the problem K∈K. Here we consider some particular cases when we can write the equations of the boundary of the Ddecomposition regions explicitly. Similarly. it was the ellipsoid of the largest volume. AUTOMATION AND REMOTE CONTROL Vol. ω ∈ (−∞. then A + BKC has the same number of stable eigenvalues for all matrices K from Q. . However in general case Eqs. (39) is linear in K. . Better approximation of D was proposed in [34]. . According to the formula det(I + abT ) = 1 + the general Ddecomposition Eq. Indeed. km ] while M is a column vector M = [M1 . Though the modiﬁed Ddecomposition technique allows us to describe the regions with constant number of stable eigenvalues. The equation describing the boundary of the regions Dl is called the Ddecomposition equation. The decomposition of the parameter space into the regions Dl = {K ∈ K : A + BKC has l stable eigenvalues}. K is a column vector K = [k1 . . +∞).. (39) is reduced to: m m m i=1 ai bi with a. Note that the system of Eqs.1. l = 0. where a0 (s) + ki ai (s) is the characteristic polynomial of the degree n. (39) and (40) deﬁnes the Ddecomposition in general form and contains the polynomial case that is discussed in details in Section 2. i. det(I+M (jω)K)=0 min K is under consideration (i. i. .DDECOMPOSITION TECHNIQUE STATEOFTHEART 2009 Deﬁnition 4. . . (8). The Ddecomposition technique is based on the following theorem. Mi (s) = 6. and with have the same Ddecomposition Eq.e. 12 2008 . ω ∈ [0. consider singleoutput system. The equation det(I + M (jω)K) = 0. +∞) (39) deﬁnes the Ddecomposition for the class K. ∀K ∈ Q (continuoustime case) or det(I + M (ejω )K) = 0. is called Ddecomposition.e. and the analysis requires the extension of the Ddecomposition technique. Then K is a row vector: K = [k1 . n. . Note that this is in contrast with μanalysis. kr ]T . 2π). . Equations (39) and (40) deﬁne the Ddecomposition in the implicit from. one is seeking for the largest ball contained in D). Mm ]T ∈ Cm . inscribed in D. . ∀K ∈ Q (discretetime case). . . ω ∈ [0. . Thus Dn is the set of stabilizing matrices K. ai (s) .e. Thus a0 (s) the Eq.. . Theorem 6 [53]. 2π) (40) ω ∈ (−∞. for the singleinput systems m = 1. Graphical Algorithms Here we consider classes of K such that the transfer function is not the ration of the polynomials and classical Ddecomposition is not applicable. b ∈ Cm 1+ i=1 ki Mi (jω) = 0 or a0 (jω) + i=1 ki ai (jω) = 0. . (39) and (40) are not linear in K. or det(I + M (ejω )K) = 0.. r = 1.
For the real k the number of intervals preserving the same number of stable eigenvalues of A + kBC does not exceed n(n + 1) + 1. . . vi . The general Ddecomposition Eq. 69 No. Theorem 7 [53]. i = 1. . I is a unit matrix m × m. . For ω = 0 matrix M (jω) = AUTOMATION AND REMOTE CONTROL Vol. The matrix K has the form K = kI. (39) now reads det(I + kM (jω)) = 0. . i = 1.2010 GRYAZINA et al. The m1 m2 . Consider the cases when it is possible. (45) where α = u1 u4 − v1 v4 − u2 u3 + v2 v3 . 12 2008 . n. We start with the class (4). that is (A + kF − jωI)x = 0 for x ∈ Cn or (A − jωI)x = −kF x. −F ). . (41) Denote the eigenvalues of M (jω) as λi (ω). (44) To take the opportunity of the graphical representation we restrict ourselves by situations with K depending on two parameters only. the matrix K contains in general transfer function now is the 2 × 2 matrix M = m3 m4 4 parameters K = k1 k3 . . . (42) This equations can be obtained in a diﬀerent form with no use of the transfer function (37). . . . (43) For the case of real k we should take into consideration only the intersection of the obtained Ddecomposition picture and the real axis. +∞). . k1 0 . n. β depend on ω. β = u1 v4 + v1 u4 − u2 v3 − v2 u3 . λi (ω) i = 1. In the terms of μanalysis this is class K with one scalar block. ω ∈ (−∞. n. i = 1. Then the matrix A + BKC is equal to A + kBC and the problem is reduced to the simplest one: given n × n real matrices A and F . Eqs. (41) split into 1 + kλi (ω) = 0. Thus we get two quadratic equation in two variables. . . (39) reduces to: k2 k4 4 0 = det(I + M K) = 1 + i=1 ki mi + det M det K. the variables ui . α. and Ddecomposition boundary consists of n branches k(i) (ω): k(i) (ω) = − 1 . where k ∈ C or k ∈ R. 4 and write separately real and imaginary parts: 1 + k1 u1 + k2 u4 + αk1 k2 = 0 k1 v1 + k2 v4 + βk1 k2 = 0. Equation (44) has the form 0 = 1 + k1 m1 + k2 m4 + k1 k2 (m1 m4 − m2 m3 ). Thus we conclude that parametric Ddecomposition boundary k(ω) is a generalized eigenvalue for the matrix pair A − jωI and −F : k(ω) = eig (A − jωI. Now let us consider systems with two inputs and two outputs corresponding to class (5). ﬁnd D(l) = {k ∈ C (or k ∈ R) : A + kF has l stable eigenvalues}. Denote mi = ui + jvi . . (1) K = 0 k2 In terms of μanalysis this structure of K corresponds to two scalar blocks. and the general Ddecomposition Eq. . If A + kF (F = BC) has an imaginary eigenvalue then the matrix A + kF − jωI is singular for some ω ∈ R.
. −0. Thus the Ddecomposition consists of the components of this curve and singular circumference. In the intersection with real axis k there are 13 intervals with constant number of stable eigenvalues. vi (ω). Let 095 1 0 ⎢ ⎥ 0. among them there 5 intervals corresponding to stable matrices A + kB. Eq. 04 0 0 ⎡ ⎤ C = I.8733 −0.4457 . i = 1.95 ⎡ ⎤ 0 0 −0. It means that there is no parameter value K ∗ such that A + BKC has eigenvalues ±jω ∗ . β(0) = 0.DDECOMPOSITION TECHNIQUE STATEOFTHEART 2011 C(A − jωI)−1 B is real and vi (0) = 0.6 ⎦ . This example demonstrates the attainability of the estimation of Theorem 7 for n = 3. we obtain k1 (ω). −k1 k2 . k2 (ω). k1 0 0 k2 Example 7. Examples Example 6. 12 2008 C= AUTOMATION AND REMOTE CONTROL . For every ω = 0. It consists of three branches (parametric curves corresponding to the diﬀerent generalized eigenvalues). . (44) reads 0 = 1 + k1 (m1 + m4 ) − k2 (m2 − m3 ) + (k1 + 2 k2 ) × (m1 m4 − m2 m3 ) and (45) is replaced with 2 2 1 + k1 (u1 + u4 ) − k2 (u2 − u3 ) + α(k1 + k2 ) = 0 2 2 k1 (v1 + v4 ) − k2 (v2 − v3 ) + β(k1 + k2 ) = 0. For ω = 0.4255 −0. where 0. . solution of system (45) speciﬁes the point of the Ddecomposition boundary. −0. k2 (ω) corresponding to the boundary of the Ddecomposition.2. and the second equation vanishes while the ﬁrst equation deﬁned the second order curve 1 + k1 (u1 (0) + 2 2 u4 (0)) − k2 (u2 (0) − u3 (0)) + α(0)(k1 + k2 ) = 0.0266 Vol.9326 C. (46) for ω = 0. In Fig. It is a hyperbola. we get vi (0) = 0. Consider the discretetime system A + B −0. Solving Eq.8148 A= B= 0.8848 0. (3) K = k2 k1 Similarly to the previous case. 4 as well as β(0) = 0. (44) is reduces to: 2 2 1 − k1 (u1 − u4 ) + k2 (u2 + u3 ) − α(k1 + k2 ) = 0 2 2 −k1 (v1 − v4 ) + k2 (v2 + v3 ) − β(k1 + k2 ) = 0. β(ω) are the same as above. For this case it is a circumference. Again it the circumference 1+ k1 (u1 (0)− u4 (0))+ k2 (u2 (0)+ u3 (0))− 2 2 α(0)(k1 + k2 ) = 0. −0. 69 No.7854 . α(ω).3914 0. (2) K = −k2 k1 In terms of μanalysis this is a real 2 × 2 analog of a complex scalar (note that the eigenvalues 2 of such K are k1 ± jk2 ).1514 0. ω = 0 and the second order curve corresponding to ω = 0. k1 k2 . 6. . If for some ω ∗ the solution is complex we ignore it because it does not belong to the Ddecomposition. 7a the Ddecomposition is depicted. For such K Eq. (47) The Ddecomposition consists of the parametric curve k1 (ω).22 ⎢ ⎥ B = ⎣ 0 −03 0 ⎦. A=⎣ 0 0 0 0 −0.2508 . (46) where ui (ω). Hence the second equation vanishes and the ﬁrst equation: 1 + k1 u1 (0) + k2 u4 (0) + α(0)k1 k2 = 0.5576 0.
c.0374.5971 0. In Fig. if K = λ 0.2012 3 2 1 Im k 0 3 1 2 3 2 1 2 3 0 1 2 3 2 2 GRYAZINA et al. then A + BKC remains stable preserving stability.2508 −0. It is generated by two singular circumferences (subtle lines) and one parametric curve (solid line). 8. Example 3]. Ddecomposition in Examples 8 (left) and 9 (right).9680 for 0 ≤ λ ≤ 19. Thus we restrict ourselves with matrices K of this form and construct −k2 k1 Ddecomposition for such matrices (see Fig. However. it has m = r = 2. The optimal solution for (7) is supplied with K ∗ = −0. Example 9. Note that the nominal system (the origin marked with “×”) is close enough to the boundary of the stability domain and the distance to it. 7b one can see two singular hyperbolas (subtle lines) and two branches of the nonsingular curve (solid lines). 10 0 1 0 2 1 1 –10 k2 3 –20 –30 0 2 1 0 1 –1 –2 Stability –40 –50 –3 –6 –4 –2 0 Re k 2 4 6 –10 –5 0 k1 5 10 Fig.2508 . 1 5 0 0 –1 k2 –2 –3 –4 1 2 1 3 4 2 2 3 2 0 k2 –5 –10 –5 –15 –10 k1 –5 0 5 –4 –3 –2 –1 –20 0 k1 1 2 3 4 Fig. 69 No. 0. m = r = 2. 7.8483 0. 889. For instance. The optimal solution AUTOMATION AND REMOTE CONTROL Vol. has a typical Ddecomposition structure. other directions allow larger values of perturbations −0. This continuoustime example is again originated in [75. equals 1. n = 3. This discretetime example is borrowed from [75.9680 −0.5971 . in accordance with [75]. n = 4.8483 the form k1 k2 . 8a). 889. Ddecomposition in Examples 6 (left) and 7 (right). c. and the same problem of the stability radius is considered. 12 2008 . Example 2]. Example 8.6932.
for a small number of parameters. for many directions preserve stability for perturbations larger than 0. and the determination of the other eigenvalue invariant regions is highly important for checking the applicability conditions of the generalized Sprocedure.5141 (stability radius). Ai ∈ Sn×n . found in [75] is the matrix of the form K = 7. The goal is to deﬁne the intervals on the parameter axis x such that the signature of the matrix A(x) is constant: Dl = x ∈ R : A(x) has l negative eigenvalues and n − l positive eigenvalues All basic ideas of the proposed approach become apparent in this simplest scalar case. .DDECOMPOSITION TECHNIQUE STATEOFTHEART 2013 −k1 k2 . the regions are easy to depict graphically on the plane. k2 ) plane k2 k1 for the given class of matrices is shown in Fig. DDECOMPOSITION FOR LINEAR MATRIX INEQUALITIES This section is devoted to the following problem: In the space of parameters x ∈ R . The diﬀerence from the problem of the previous section is the requirement for all matrices in (48) to be symmetric. the theory and methods for linear matrix inequalities had shaped into a selfcontained discipline. There are two disconnected components of the parametric curve (solid lines) and one singular curve—the circumference (subtle line). not to mention other signature invariant regions. Similar to the previous example. ﬁnd the regions such that A(x) = A0 + i=1 xi Ai . The primary goal here is not only to describe the whole feasible set {x ∈ R : A(x) < 0} but also to characterize all the regions in the parameter space such that the signature of the matrix A(x) remains unaltered. symmetric n × n matrices. But as it tool place in previous sections. A complete description of the feasible set gives a clue to new statements and solutions for the semideﬁnite programming problems. [36]. this potentially leads to relatively simple solution methods for systems of matrix inequalities with uncertainty in matrix coeﬃcients. 12 2008 . The Ddecomposition of (k1 . various problems involving model uncertainty often defy eﬃcient solution. By mednineties. This is the wellknown feasibility problem for linear matrix inequalities (the common abbreviation LMI [2]. In this section we follow standard for linear matrix inequalities notations for parameters x. (48) has a ﬁxed number of negative eigenvalues. B ∈ Sn×n . A special case of this problem is to ﬁnd a point x such that linear combination of symmetric matrices is negative deﬁnite. 8b. One Parameter Problems We start with the case of one scalar parameter: A(x) = A + xB. However the LMI theory is not aimed at describing the whole feasible set. rather than k as it is in previous sections. It makes the problem signiﬁcantly easier. optimization problems under constrains in LMI form were called semideﬁnite programming. It is accomplished using the Ddecomposition technique modiﬁed for the case of symmetric matrices.1. 69 No. [41]) A(x) < 0. Moreover. AUTOMATION AND REMOTE CONTROL Vol. see. x ∈ R. = 2. In particular. Here Sn×n denotes the space of real. A. 7.
(49) 7. the number of likesign eigenvalues can only alter if one of them (or several at once) crosses the origin. i∈I+ i∈I− I+ = ∅. i = 1. hence.. . Assume that for a certain real x∗ . symmetric matrix A(x∗ ) = A + x∗ B is nonsingular and has l negative and n − l positive eigenvalues. We ﬁx x1 and denote A = A(x1 ) = A0 + x1 A1 . the determination of the boundaries (points) of the Ddecomposition regions in the scalar case reduces to ﬁnding all real generalized eigenvalues: eig (A. I+ = {i : (Bei . and the critical values of the parameter x2 for the given x1 are determined as the real generalized eigenvalues x2 (x1 ) = eig A(x1 ). 1. determine the interval (x2 (x1 ). min λi . The following theorem holds. Theorem 8 [25]. In other words. −B). For every x1 equation det A(x1 ) + x2 B = 0 has no more than n real roots. −B . This means that there exists a nonzero vector e such that (A + xB)e = 0. Ai ∈ Sn×n . Ae = −xBe. and B = A2 . (2) If all λi are real. x is a generalized eigenvalue of the pair of matrices A and −B. In the general situation. x2 (x1 )) according to Theorem 8 (it may happen to be empty for some or all values of x1 ). ei ) > 0} = ∅ −∞. B ∈ Sn×n and B be nonsingular.e. Problems with Two and More Parameters We now turn to the twoparameter case: A(x) = A0 + x1 A1 + x2 A2 . there might be no real solutions at all.2014 GRYAZINA et al. −B) ∈ R. . if x x. Of a particular interest is the stability interval Dn .. i. A2 . det(A + xB) = 0 for some x ∈ R. the formulas above take the form: x= max λi λi <0 −∞. since the equation det(A+ xB) = 0 has at most n real solutions. max λi . and e is the associated generalized eigenvector. n be the generalized eigenvalues and the associated generalized eigenvectors of the pair (A. ei . i = 0.e. provided all λi > 0 and x= λi >0 min λi +∞. If it is known that A < 0. 12 2008 . . x). Then we are in the conditions of the previous section. Let λi . A2 be non singular. Let A. On the other hand. . then denote x= x= Then Dn = (x. AUTOMATION AND REMOTE CONTROL Vol. Therefore. the boundary of Ddecomposition consists of no more than n branches. say. for every x1 . Namely. It is seen that the maximal possible number of regions (intervals on the xaxis) is equal to n + 1. . ei ) < 0} = ∅ I− = ∅. By varying x1 we obtain the boundaries of the Ddecomposition regions. I− = {i : (Bei . Let . 2. one of the matrices A1 . . in this case the Ddecomposition is represented by a unique domain—the number of negative eigenvalues of the matrix A + xB remains unaltered for all values of x. As x varies. provided all λi < 0. Then (1) If there exist complexvalued eigenvalues λi then Dn is empty. if x < x ∅. +∞. the stability region Dn = {x ∈ R2 : A(x) < 0} can be described separately (it is obviously convex). 69 No.2. i.
.0 –0. . .8 0. . 1 1 2 1 1 m 2 m so the boundaries of the Ddecomposition are represented by m ellipses x2 b2 +x2 c2 = 1 (they divide 1 i i the plane into n(n − 2)/2 + 2 regions). cm . . Example 10.2 –0. . Linear Matrix Inequalities with Uncertainty One of the valuable extensions of the proposed in the previous section approach is its modiﬁcation to the presence of uncertainty.4 0. c1 = 2. . . . −bm . . .25..2 0 –0. . then the determinant is det(A + x1 B + x2 C) = (−1)m (x2 b2 − x2 + c2 ) . There are various formulation of LMI problems and assumptions on AUTOMATION AND REMOTE CONTROL Vol. (50) A2 = C = diag(c1 . (x2 b2 + x2 c2 − 1). . The example of the Ddecomposition for 6 × 6 matrices of this sort with parameters b1 = 1. . . . The marking can be performed in a standard way. b3 = 2.e. Ddecomposition in Example 10. c2 = 1. 7. cm . Taking for bi . however the total number of the regions does not exceed n2 /2 + 1. 69 No. 9b. The determinant is straightforward to compute: det(A + x1 B + x2 C) = (−1)m (x2 b2 + x2 c2 − 1) . (x2 b2 − x2 + c2 ). . the endpoints of the interval sweep the boundary of the stability region. Note that Dn is knowingly unbounded if one of the matrices Ai .3.DDECOMPOSITION TECHNIQUE STATEOFTHEART (a) 1. If we replace A and C in (50) A = diag(c1 . ci the i values form the previous example we depict the regions in Fig. upon constructing Dl regions on the plane. in LMI (48) is signdeﬁnite. b2 = 1. . . i = 1.8 –1. . −bm . i. . where diag denotes an antidiagonal matrix.25. consider the following n × n. c3 = 1 is depicted in Fig. .0 –6 –4 –3 –2 –1 0 x1 1 2 3 4 Fig. .0 0. B = diag(b1 . . cm . bm . . . . c1 ). 12 2008 (51) . . 1 1 2 1 1 m 2 m The boundary of the Ddecomposition is given by the family of m pairs of hyperbolas x2 b2 − x2 + 1 i 2 c2 = 0. bm . −b1 ).5 0 x1 0. . . −b1 ). as x1 varies.6 0. C = −I. . As an example of this sort.4 –0. . 9. . .5 1. cm . In some special cases the description of the regions can be obtained explicitly.6 –0. . A1 = B = diag(b1 .0 4 5 6 3 4 2015 6 4 2 x2 5 4 (b) 4 6 4 2 5 4 3 1 0 2 1 2 x2 0 –2 –4 –1. n = 2m matrices: A0 = A = −I. . pick a point inside each of them and compute the eigenvalues of the corresponding matrix. . c1 ). . . . 9a. .
ρ(M ) = inf{ P : P ∈ Sn×n . Δ) has exactly l negative eigenvalues ∀ Δ ∈ D in particular. the robustly feasible region is deﬁned by: rob Dn = x ∈ R : A(x. Δ ∈ D = {Δ ∈ Sn×n : Δ The boundaries of the regions (intervals) of the robust Ddecomposition are deﬁned from the condition that the matrix A + xB + Δ be singular for some Δ ∈ D. . .. . 37]. Δ) = (A + Δ) + xB. Note that in order to retain the LMI structure of the problem. Δi εi . Δi εi . . By the theorem. Δ) = A0 (Δ0 ) + i=1 . Δ) becomes singular for some Δ ∈ D. i The critical value of P is given by P = −λeeT . we ﬁrst consider oneparameter families in the simplest case where only the A matrix is subjected to additive uncertainty: . . the matrix (A + xB) + Δ with perturbations Δ ε remains robustly nonsingular for the values of x satisfying 1 (A + xB)−1 < . Δ) < 0 ∀ Δ ∈ D . [41. ε}. . for nonsymmetric matrices. the radius of nonsingularity is given by the general formula in Theorem 3 [73]. Δ ∈ D = (Δ0 . only symmetric perturbation Δi are considered. M + P is singular}. Similarly to the analysis of the problem without uncertainty. Δ) may have diﬀerent values of signature depending on one or another admissible value of Δ. these boundaries are now deﬁned as the values of the parameters x such that the matrix A(x.g. Here we concentrate on the situation where uncertainty is bounded in the spectral norm. its symmetric radius of nonsingularity . where λ is the minimal (in absolute value) eigenvalue of M . Ai . Theorem 9 is a counterpart of Theorem 3 in [73] for the symmetric case. Δi ∈ Sn×n . . 69 No. (52) where · is the spectral norm and εi 0 are given numbers. .e. i = 0. i. Let Ai (Δi ) = Ai + Δi . ε AUTOMATION AND REMOTE CONTROL Vol. Similarly to the problem of robust Ddecomposition for polynomials. is equal to ρ(M ) = 1/ M −1 = min λi (M ). xi Ai (Δi ). . uncertainties available in the literature. For a nonsingular matrix M ∈ Sn×n . The theorem below is the main tool: Theorem 9 [25]. the boundaries of robust Ddecomposition smear into “strips” inside which the matrix A(x. (55) A(x.. . (54) The problem is to describe the boundaries of the regions of robust Ddecomposition. (53) The regions Dl of robust Ddecomposition are now deﬁned as follows: Dlrob = x ∈ R : A(x. 12 (56) 2008 . We thus arrive at the following uncertain linear function: A(x. Δ ) : Δi ∈ Sn×n . and e is the associated eigenvector.2016 GRYAZINA et al. e. the problem is to determine the radius of nonsingularity of the matrix A + xB.
For the twoparameter family in S4×4 with matrices A0 = −I. We ﬁx x1 and denote A = A0 + x1 A1 and B = A2 and also εA = ε0 + x1 ε1 and εB = ε2 . i = 0. A more general problem where the uncertainty enters both matrices. we now address the case of twoparameter families: εi . Represent A(x. Indeed. A(x. 1. ΔA εA . 1. 1. the intervals of robust nonsingularity are found numerically as {x : ϕ(x) < 1/ε}. AUTOMATION AND REMOTE CONTROL Vol. Δ) = (A0 + Δ0 ) + x1 (A1 + Δ1 ) + x2 (A2 + Δ2 ). 2. the region Dn can be described separob by combining the results of Theorems 8 and 9. A1 = diag(−1. . 1. we compose the functions . rately from the rest Dl Example 11. hence. introducing the function ϕ(x) = (A + xB)−1 . −1) and A2 = diag(1. then by Theorem 9. ϕ(x2 ) = (A + x2 B)−1 and ε(x2 ) = 1/(εA + x2 εB ) and check the condition ϕ(x2 ) < ε(x2 ).DDECOMPOSITION TECHNIQUE STATEOFTHEART 4 3 2 1 0 –1 –2 –3 –4 –3 –2 –1 0 x1 1 2 3 4 rab Dn rab D3 2017 D3 D2 rab D2 Fig. these intervals of violation sweep certain twodimensional regions which are considered the boundaries of robust Ddecomposition. since ΔA and ΔB are chosen independently. the values of x2 that violate the condition ϕ(x2 ) < ε(x2 ) correspond to the absence of robustness. the intervals of robust nonsingularity of equivalently. which deﬁned the segment of robustness in x2 for a given value of x1 . 10. which is sharp (the equality is attainable). 69 No. −1). these intervals ﬁll up twodimensional regions of robust Ddecomposition. Δi . . −1. as x1 varies. the boundaries of Ddecomposition in the absence of uncertainty are given by the two pairs of parallel lines in Fig 10. Δ) = (A+xB)+(ΔA +xΔB ). to determine the boundaries. As x1 varies. the regions Dlrob are deﬁned by the condition ϕ(x) < 1 . This leads us to the setup of problem (57). Δ) = (A + ΔA ) + x(B + ΔB ). ΔB εB . 12 2008 x2 . (57) is analyzed similarly. εA + xεB Slightly changing the notation. For the perturbation ΔA +xΔB of the matrix A + xB we have the estimate ΔA + xΔB εA + xεB . rob We also note that similarly to the uncertaintyfree case. Introduce the function ϕ(x) = (A + xB)−1 . A(x. On the other hand. Robust Ddecomposition for two parameters.
First we brieﬂy describe these method in general. m × n correspondingly.2 Ai . There are only a few cases when the stability domain ban be explicitly described for higher dimensional parameter space. Then inequality G + M ΔN + N T ΔT M T < 0 hold for all matrices Δ of size r × m. generally speaking nonconvex and not simply connected (but consisting of a ﬁnite number of connected components) set such that any intersection of the set K and a line consists of AUTOMATION AND REMOTE CONTROL Vol. Then we can ε ﬁnd the intervals in x that preserve robust signdeﬁniteness of A(x. Δ < 1 iﬀ there exist ε > 0 such that 1 G + εM M T + N T N < 0. Δ). RANDOMIZED ALGORITHMS FOR SYNTHESIS USING DDECOMPOSITION As it was shown in the previous sections the Ddecomposition idea in its classical form is applied only in one or twodimensional space. i = 0.2018 GRYAZINA et al. The second method (socalled ShakeandBake) [4. Let G = GT . the boundaries of robust Ddecomposition are given by the shaded region and the uncolored regions represent the regions of robustness. for any x. if we change the family (55) by A(x. However the general Ddecomposition Eq. 1 1 if A + xB < −εM M T − N T N for some ε > 0. then demonstrate how they work for the description of the stability domain (both for polynomials and matrices) and a domain speciﬁed by linear matrix inequalities. In the example considered. for equal relative levels of uncertainty Δi εi = 0. Δ 1. 44] allows one to generate point on the boundary of the domain that is again non convex but connected. Theorem 10 [72]. Within this framework we consider more general classes of uncertainties. For all the methods of this section the main tool is boundary oracle procedure.79] generates approximately uniformly distributed points inside the given domain. Δ) = A + ΔA + xB. 12 2008 . N = 0 be matrices of the size n × n. It can be done using modern schemes of Monte Carlo method. One of the methods (known as HitandRun) [29. 2. 1. 69 No. One of these cases (a spacial polynomial family (15)) is described in Section 3. (9) or (10) for polynomials and (39) or (40) for matrices is valid also for the multidimensional case. Let K ⊂ Rn be an open bounded. generally speaking nonconvex and not simply connected. The problem is to propose the detailed enough description of the multidimensional stability domain (or its boundary) in general situation. Introducing uncertainty smears the boundaries into stripes and the Dl regions shrink. Similar to membership oracle and separation oracle that are exploited in modern convex optimization theory [36]. n × r. M = 0. One of the opportunities to achieve this goal is to generate point randomly (uniform generating is preferable) inside the stability domain or on its boundary. The analysis is based on the following result of Peterson. For our case. ε Various generalizations and applications of this result are addressed in [31]. 8. then A(x. y ∈ Rn boundary oracle either provides intersection points of a ray x+ λy and the boundary of the set K deﬁned implicitly or reports that the intersection is empty. Δ) < 0 for all Δ ΔA = M ΔN + N T ΔT M T .
Step 4. For instance.1.e. deﬁne the set Si = {λ ∈ R : ki + λdi ∈ K}. k) = a0 (s) + i=1 ki ai (s). We address the method for the polynomial stabilization problem ﬁrst. 1) the ndimensional vector with normally distributed components). ξ = randn (n. This set satisfy the above assumption. Generate a point λi uniformly distributed in Si (we recall that Si is a ﬁnite set of intervals). Calculating J(ki ) for admissible point we ﬁnd the optimal one. Find a starting point k0 ∈ K. Theorem 11.e. 66]. by assumption on K. i. Step 3. i.e.. For a certain point k ∈ K and a direction d ∈ Rn . consists of a ﬁnite number of intervals. the boundary oracle provides the intersection of the line k + λd. At the point ki ∈ K generate a random direction di ∈ Rn uniformly distributed on the unit sphere (i. d the set S can be computed). Apply boundary oracle procedure. Our goal is to describe somehow the stability domain in the parameter space k. i.. we have a performance index J(k) (gain or phase margin. the boundary oracle is given by Algorithm from the Section 2. The general result on HitandRun method is due to Smith [79]. H2 or H∞ norms). if a0 (s) is stable then h0 can be chosen as k0 = 0). suppose that the boundary oracle is available (i. for any k. i. Under the taken assumptions the distribution of the point ki tends to the uniform one in K as i → ∞.DDECOMPOSITION TECHNIQUE STATEOFTHEART 2019 a ﬁnite number of intervals. Step 2. Moreover. K is open and consists of a ﬁnite number of components. robustness margin.. and compute a new point ki+1 = ki + λi di . Suppose that k0 ∈ K is known (for instance. k) is stable}.e. 12 2008 . more precise estimate for convex sets can be found in [65. Theorem 3].e. overshoot or other timeresponse characteristics.. K = Dn = {k ∈ Rm : a(s. the polynomials of the form m a(s. 69 No. −∞ < λ < +∞ and the set K. We consider the aﬃne family of polynomials (8). then we may apply HitandRun to generate points ki approximately uniformly ﬁlling the stability domain. There is also the convergence estimate in the original paper [79.. di = ξ/ ξ . This set. namely. Step 5. Thus it is the way to solve various synthesis problems. Then HitandRun method works as follows.. taking the intersection of K with some trust region (say a ball AUTOMATION AND REMOTE CONTROL Vol. Go to step 2 and increase i. In the neighborhood of k∗ we can repeat the process. Step 1. the set S = {λ ∈ R : k + λd ∈ K}.e.
(58) AUTOMATION AND REMOTE CONTROL No. Boundary points for a chosen direction y are deﬁned as real generalized eigenvalues λi for the matrix pair (A. negative deﬁnite) for all Δ ∈ D satisfying estimate Δ(λ) A(λ) < 1/ Δ(λ) . the i=1 Δ0 + i=1 xi + λyi  Δi = ε0 + −1 xi + λyi εi is sharp. We seek rob for the intersection points of a ray and the boundary of robust Ddecomposition. The proposed boundary oracle is especially eﬃcient for characterization of the stability domain Dn which is always convex for LMI. Let x ∈ Dn be robustly feasible point and y ∈ R be a certain direction. Δ) = A(λ) + Δ(λ). The diﬃculty suﬃciently reduces because the boundary oracle of K provides an interval (λ. Similarly HitandRun can be applied for robust LMI’s. A number of examples for various problems of synthesis are addressed in [74]. centered at k∗ of radius ε) as new admissible set and k∗ as the initial point. We consider a point x ∈ R and a direction y ∈ R . . Let for some x ∈ R matrix A(x) (48) be negative deﬁnite. the matrix A(λ) + Δ(λ) remains nonsingular (hence. 69 . hence. we have A(x + λy) = A0 + i=1 xi Ai + λ i=1 yi Ai . λ) that can be computed as follows. We have A(x + λy. we obtain Ddecomposition boundary points x + λi y. For λ ∈ R. B= i=1 yi Ai . A(λ) = A0 + i=1 −1 (xi + λyi )Ai . 69]) are given in [26]. Numerous application examples of this approach for optimization of linear functions subject to LMI (for semideﬁnite programming. HitandRun is applicable not only for stabilization problem of polynomials but also for LMI’s A(x) < 0 (48). ϕ(λ) = A0 + i=1 (xi + λyi )Ai . by considering the two scalar functions . and by Theorem 9. −B). where it is denoted . we are in the oneparameter case A(λ) = A + λB.e..2020 GRYAZINA et al. λ such that signdeﬁniteness of the matrix A(x + λy) is preserved are given by formulae (49). These methods appear to be quite competitive compare to the standard tools such as Yalmip [67] and LMI toolbox for Matlab. and y is a certain direction. This approach is easytouse for generating uniform distribution in Dn (compare to [42]) as well as proposing new programming methods. the boundary oracle. which guarantee the negative deﬁniteness of the matrix A(x + λy. x ∈ Dn . the minimal and maximal values of λ. Δ) for all Δ ∈ D. ε(λ) = ε0 + i=1 1 xi + λyi  εi Vol. Consider the straight line x + λy and compute λ rob . Since the perturbations Δi are independent. λ rob . Consider the construction of the boundrob ary oracle for robust stability domain Dn (54) for the family (52) and (53) for > 2. 12 2008 . . Then critical values λ. SDP [41. Δ(λ) = Δ0 + i=1 (xi + λyi )Δi . where η has standard dimensional gaussian distribution. i. A = A0 + i=1 xi Ai . . Generating directions uniformly on the unit dimensional sphere in the form y = η/ η . and denoting .
N i ) = 0. 1 − ξ n−1 2 1/2 . The behavior of this method in the general situation is given by the following Theorem 13 [44]. it suﬃces to consider the unperturbed matrix at the point x: A(x. 69 No. Namely. K 0 = 0 for stable matrix A). (ci . K be a simply connected component of K that contains the certain K 0 (for instance. the points look approximately uniformly distributed. for instance. Δ) for all admissible perturbations Δ are given by the two solutions of the equation ϕ(λ) = ε(λ) (58) over the segment [λ. Go to Step 2 and increase i. Step 2.DDECOMPOSITION TECHNIQUE STATEOFTHEART 2021 the interval (λrob . Another randomized method mentioned above is ShakeandBake that allows generating points not inside but in the boundary of a set K. d) = sup{λ > 0 : k + td ∈ K. 0) (49). Given an initial point k0 ∈ ∂K. In the other words. λrob ) of robust deﬁniteness of the family A(x + λy. The boundary oracle for ShakeandBake algorithm can be produced numerically using Eq. λ is a point of the ﬁrst intersection of the ray k + λd. For any y ∈ R . λ > 0 and the boundary K. For the problems with stability domain K. ShakeandBake algorithm is the following: Step 1. 0))−1 holds. (39) of the Ddecomposition for matrices. where [λ. 12 2008 . (d. N ) > 0 computes λ = λ(k. 0) = A0 + ε0 + i=1 xi Ai and check if the inequality i=1 xi εi < 1/ (A(x. ShakeandBake technique works. Let K be a connected open bounded set with a normal mostly at every point. λ]. Let K = {K ∈ Rr×m : A + BKC be stable}. we obtain matrices K approximately AUTOMATION AND REMOTE CONTROL Vol. λ)}. for the boundary point k ∈ ∂K the boundary oracle provides the internal normal N = N (k) for ∂K at this point and for the internal direction d ∈ Rn . At the point ki ∈ ∂K compute the unit internal normal N i for ∂K and generate random direction di as follows: di = αi N i + ci . di ) and generate a new point ki+1 = ki + λi di . λ] are the bounds of negative deﬁniteness of the matrix A(x + λy. where ξ = rand is uniformly random Step 3. 0) < 0. for the matrix case considered in Section 6. Δ) can be found numerically as {λ : ϕ(λ) ε(λ)}. rob Similarly. αi > 0 is a scalar random variable. We assume the boundary oracle that provided some extra information compare to one used for HitandRun. Theorem 12 [25] . 1]. We note that the set K is not assumed to be convex. Using the boundary oracle ﬁnd λi = λ(ki . Figure 11 depicts the result of the generating points at the boundary of the nonconvex set. to solve a simpler problem of checking if x ∈ Dn for some x ∈ R . the maximal and minimum values of λ retaining the negative deﬁniteness of the matrix A(x + λy. where ci is uniformly distributed in ci = 1. Then the distribution of ki tends to the uniform distribution at ∂K as i → ∞. ∀t ∈ (0. Indeed. α = in [0. Step 4. Finally. let A(x.
The condition of alteration of the number of stable eigenvalues of the parameter depending matrices is formulated in terms of singularity of the new matrix. CONCLUSION We provided the simple and eﬀective techniques to describe the stability domain in the parameter space.2 –0.2 0 k1 0.2 0.1 0 –0. After more than 50 years the Ddecomposition idea remains actual and there appear new applications for various control problems.4 0. The result has much in common with the basic Ddecomposition theorem that describes when the stable root number changes but it is applicable for more general system structure. i.4 –0. 12 2008 .5 –0. Simultaneously with the stability domain we construct all the root (eigenvalue) invariant regions.3 0. The extension of the Ddecomposition for the systems with matrix transfer functions is of particular importance.6 k2 GRYAZINA et al.4 –0. This is an extension of the Ddecomposition method for polynomials.1 –0. This technique can be helpful for the loworder controllers design and for a detailed robustness analysis.6 Fig. simply connected regions of the domains with the invariant number of likesign roots (eigenvalues) of the system matrix. In these cases. Besides new results on the structure of Ddecomposition for polynomials. The main limitation is the small number of parameters but in some spacial cases (see Section 3) is can be overcome. We pick out several classes AUTOMATION AND REMOTE CONTROL Vol. 9. proposed randomized algorithms allows one to ﬁll in the stability domain (or its boundary) with randomly generated points. Ddecomposition allows characterizing the multidimensional stability domain as a system of linear inequalities and it gives a straightforward solution of the stability radius problem. The conducted survey shows that the domain of applicability for Ddecomposition is much wider than the original problem statement for polynomials. 11.4 0. 69 No. The modiﬁcations of the technique in order to describe the regions in the parameter space corresponding to the ﬁxedorder controllers that satisfy some H∞ speciﬁcations are considered. Finally. ShakeandBake method also works for LMI. we discuss the method for description of the root invariant regions for uncertain polynomials. the required boundary oracle is deﬁned explicitly.e. This procedure may be regarded as an multidimensional alternative for graphical algorithms.2022 0. Points generated in the boundary of nonconvex domain.3 –0.. uniformly distributed on the boundary of K. –0.2 0.5 0.
1964. 2005. Sintez zakonov upravleniya na osnove lineinykh matrichnykh neravenstv (Synthesis of Control Laws Based on Linear Matrix Inequalities). Telemekh. pp. Telemekh. K. Telemekh. pp. E. Yu. 2007.N.. Moscow: Nauka. Avtom. 38–52. Moscow: Fizmatlit. 1992. no 2. Bendrikov. Leningrad: LKVVIA. 12. On the Ddecomposition Theory.N. Gryazina. no. D. K.F. Avtom. 3... 2004. 95–106. 4. E. and also acknowledge E. 11.. 5. Balandin. vol.. pp. 1946. 190–203. 9. we propose randomized algorithms exploiting the Ddecomposition idea in the multidimensional case. 1979. 1990. Telemekh. no. 12. 1999. ACKNOWLEDGMENTS Authors are grateful to P. Avtom. Gryazina.. Borovkov. O.A.. Telemekh. I..V.N. 16..I. 7. Moscow: Sovetskoe Radio. Minsk: OIPI NAN Belarusi. REFERENCES 1. 69 No. 11. pp. Nesenchuk. N. 109–120. and Kogan. 2001. Ustoichivost’ linearizovannykh system (Stability of the Linearized Systems). 7. 3. 1949. Nikolayev. 1960. 1991. E. Traektorii kornei lineinykh avtomaticheskikh sistem (Root Trajectories of Linear Automation Systems)... 1948. A. 6. 2007. Finally. A.I. and Polyak.P. Sintez system s obratnoi svyaz’yu (Synthesis of Feedback Systems). Jury. On a New Variant of the Monte Carlo Method. Telemekh. Avtom. pp. When the number of parameters do not exceed two the graphical representation of the result is possible. 14.A. 12. E.. 93–105. and Mayer. 9. and Polyak. no. Robustness of DiscreteTime Systems.T. 10. 10–18. Yu. pp. 3.A.M. Avtom. 15–28. 12 2008 . 342–346. 2. Analiz i sintez robastnykh dinamicheskikh sistem na osnove kornevogo godografa (Analysis and Synthesis of the Robust Dynamical Systems Based on Root Approach).. Innory i ustoichivost’ (Innors and Dynamical System Stability). vol.. Shcherbakov for fruitful discussions and problems solution for Section 7. no. 3–28. and Tremba. no. Neimark. Telemekh.I.. Telemekh. 2007. 3. A.G. A. Avtom. pp. Neimark. pp. no. no.. Neimark. Design of the LowOrder Controllers by the H∞ Criterion. E. 119–130. Luzin. It is a tool to construct the regions such that the aﬃne family of symmetric matrices has a ﬁxed number of likesign eigenvalues inside a region. Primen. B. 17. no. Avtom.DDECOMPOSITION TECHNIQUE STATEOFTHEART 2023 of parameter depending matrices and get the analytical solution of Ddecomposition equations for them. Polyak. Search for the Parameter Values That Make Automatic Control System Stable. Diﬀerentsial’noe ischislenie (Diﬀerential Calculus). 5. 13... Solnechnyi for useful remarks.I. Y. pp.N. B. Andronov. Design of the LowOrder Controllers by the H∞ Criterion: A Parametric Approach... M. Simple Linear Systems with Delay.. 1970.. Jury...S. Kiselev.. Yu. 7.M.I. G. Horowitz.A. Teor. 15. no.. Multidimensional Stability Domain of Special Polynomial Families.N. Avtom.. Moscow: Vysshaya Shkola. Gryazina. Robastnaya ustoichivost’ i Drazbienie (Robust Stability and Dpartition). Telemekh. 8. Avtom.T. Veroyatn.. Moscow: Nauka. The Multidimensional Asymptotic Stability Domain of Linear Discrete Systems: Its Symmetry and Other Properties..T. pp.A. AUTOMATION AND REMOTE CONTROL Vol. and Teodorchik. The Ddecomposition technique is applied in the theory of linear matrix inequalities. B. 36.
Nikolayev. 12. Petrov. pp. 1993. 3–26. vol. H.. 2002.. 738–743. Robust Stability of Linear DiscreteTime Systems. 30. no.. Tantaris.Z. New Tools for Robustness of Linear Systems. TX. Telemekh. 12 2008 . pp. Avtom. Avtom. 35. The Volumetric Singular Value and Robustness of Feedback Control Systems. 39. A. Dokl. Petersburg: S. Y. A. Moscow: Nauka. Avtom. P. 3.. 32nd CDC. Primen. in Inzhenernyi sbornik (Engineering Collection). Ya. no. F. pp. pp. vol. 2. 5.. B. Frequency Domain Criteria for Robust Stability and Aperiodicity of Continuous Linear Systems.. XVI. pp. pp.L. Randomized Method for Semideﬁnite Programming. O. 9. 1994. 32.. B. Polyak.. Polyak. Bhattacharyya. Chapellat.. and Kaesbauer.A.H. Y.N. no. SIAM J. Lepschy. vol. Blanchini. Ya. 304–307.. and Nemirovski. A. 521–523. and Shcherbakov. pp. 12. L. and Khlebnikov. 2003. 26. 1995. B. 2008. Kharitonov. 2002.P. 36. Shcherbakov. D. R. Polyak. and Keel.S. Bhattacharyya. and Tsypkin. 11. Avtom. pp. 2006. 2006. Univ. New York: MacMillan. 11.. 28. AUTOMATION AND REMOTE CONTROL Vol. 2086–2088. Uravn. Telemekh. Ed. 842–846. Automatica. S.. 49–61. no. IEEE Trans. Optim. Telemekh.. vol. and Viaro. Algorithms and Engineering Applications. 22. The Ddecomposition Technique for Linear Matrix Inequalities. Teor. J. Gos.. San Antonio. London: Springer. Petersen’s Lemma on Matrix Uncertainty and Its Generalizations. Robust Ddecomposition. Robust Aperiodicity. pp. Sokolov.R. Automat. P. 29.. A. 41–53. no. no. vol. vol.. Telemekh. Tremba. no. 2002.A.F.P. Philadelphia: SIAM. Automat. 736–740. 21.. no.Peterburg. Proc. Robastnaya ustoichivost’ i upravlenie (Robust Stability and Control).P. Granichin. B. Robust Ddecomposition under lp bounded Parametric Uncertainties. 316. 159–174. 21–36.. and Shcherbakov. 27..T. pp. 44–54. 38–70..A. 69 No... 1.Z. 3. Polyak. 125–139. 2003. 937–943. M. Ya.T. 1991. 2004. 1978. pp. 811–833. IEEE Trans. and Tsypkin. On the Computation of Multidimensional Integrals by the Monte Carlo Method. N. BenTal. Dokl. Robust Control: The Parametric Approach. 2004. 2006. S. Telemekh. B. V... pp. no. St.. 48. Upper Saddle River: Prentice Hall. vol.T. B. Ed.T. A. Polyak. Avtom.. Control. 31. 11. 1995. A Criterion of Stability for Linear Regulated Systems with Distributed Parameters. BenTal. Ackermann. 2002. pp.S.S. 5. Characterization of PID and Lead/Lag Compensators Satisfying Given H∞ Speciﬁcations. II.V.. 24.. vol. Robust Control: The Parameter Space Approach. Turchin. 34. Lectures on Modern Convex Optimization: Analysis. 4.T. 335. pp. B.. pp. no. 33. vol. U.. B. 20. S. Stable Polyhedra in Parameter Space. 858–861. pp.T. and Polyak. Nauk SSSR. Nikolayev. 1991. and Keel L. 1990. 37. in Stochastic Optimization in Informatics.T. Control .T. Miani.Z. Polyak. P. 19.2024 GRYAZINA et al.. 1971. 39.. The Set of Stable Polynomials of Linear Discrete Systems: Its Geometry. Nauk SSSR. J. B. The Geometry of DDecomposition of a TwoDimensional Plane of Arbitrary Coeﬃcients of the Characteristic Polynomial of a Discrete System. and Tsypkin. Leningrad: Akad.. Stabilization of DiscreteTime Systems by FirstOrder Controllers. and Shcherbakov. 23. Telemekh. pp. P.. Telemekh. no. 25. V. 40.. Barmish. A.S... Diﬀ. and Polyak. Akad. pp. 45–54. Barmish.P... 12. no. N. Akad. 49. Nauk SSSR. Talitskikh. 4.. 2001. no. On Tractable Approximations of Uncertain Linear Matrix Inequalities Aﬀected by Interval Uncertainty. Veroyatn.. and Nemirovski. 7.N. B. no.. Avtom. 1946. no. 38. Asymptotic Stability of a Family of Systems of Linear Diﬀerential Equations.. Ackermann... 11. 18. Avtom. A.
K. J. E. 48. 4643–4648. J. 61. 6. 4. Stability Regions in the Parameter Space: Ddecomposition Revisited. On the Complete Instability of Interval Polynomials.. Int. Robust.A. 105–113.. 2002. Calaﬁore. vol. and Bose. 69 No.. 42. Control . J. 153. Frazer. F. Automatica. Control . Results on Positive Pairs of Polynomials and Their Application to the Construction of the Stability Domains. 1435–1453.. 1098–1105.H. Kogan. vol. M. Robust Stability and Convexity.R. El Ghaoui. PID Controller Synthesis Free of Analytical Models... no. 1994. 2006. L. 2. C. Boyd.K. II. pp.N. Y.. Datta. M.. Schur Stability and Stability Domain Construction. 1988. Syst. vol. 1966. no. 1996. vol. A. M. Control . J. 1998. IEEE Trans. no.. 2006. 39. Syst. no. Ho. B. vol. 1. Hinrichsen. and Bhattacharyya. An Elementary Derivation of the Routh Hurwitz Criterion. 35th CDC . Datta. Proc.. 1989. 45th CDC . R. 3554–3559. 2003. and Hollot. Linear Matrix Inequalities in System and Control Theory. Delansky. Francis. pp. Structure and Synthesis of PID Controllers. 2007. 2004.H. Proc. URL: arXiv:math/0612799v3 [math. L. 1175–1183.P. B. S. Control Lett.. and Balakrishnan.T. Control . and Pritchard. Control . vol. Englewood Cliﬀs: Prentice Hall. W. 42. no. S.. IEEE Trans. vol.. S. R. AUTOMATION AND REMOTE CONTROL Vol. Oishi. 43. 2001. Real and Complex Polynomial Stability and Stability Domain Construction via Network Realizability Theory. Fujisaki. Evans. Keel. E. J. Int. vol. L... 53. 1995. G. L.P. J. Sch¨tz. 1987. Proc. Popov. Links between the Parameter Space and Frequency Domain Methods of u Robust Control. 1929. no. A New Approach to Feedback Stabilization. 43rd Conf. Control Lett. Int. 51. Decision Control . Polyak. Chen. 52.. 8. Control System Dynamics.. 55. A. 63..P. 59. Comets. N.M. 3.P. S. pp. 4. Random Walks for Probabilistic Robustness. M. S. 1986. and Ackermann. Dabbene. 42. and Bhattacharyya. S. N. 49. Ho. Robust Nonlinear Control . Fan.H. R.PR]. 45. 57. Rego.. 23. 1954. Feron. Automat.. pp. 16 IFAC World Congr.. pp.. and Tempo..J. and Bhattacharyya. pp.. Billiards in a General Domain with Random u Reﬂections. IEEE Trans. pp. 5316–5321. and Vachkovskaia. M. 124..DDECOMPOSITION TECHNIQUE STATEOFTHEART 2025 41.. 97–109. F. J. and Bose.H. 46. S. Automat. 48. B.P.P. 56. M. 54. Control . 1987. and Bhattacharyya. and Duncan. 405–409. S. 62.N. S. Proc. 1343–1349. A Course in H ∞ Control Theory.. 47. W. vol. On the Criteria for Stability for Small Motions. Stability Radius for Structured Perturbations and the Algebraic Riccati Equation. V.T.I. A. Datta. Delansky.F. New York: SpringerVerlag.. 58. J. 49. pp. G¨ venc. Fragile or Optimal?.. Control Lett. 45. Proc. G.. Synthesis of H∞ PID Controllers: A Parametric Approach. vol. Berlin: Springer. pp. Int. 11. Keel. 2005. 13–26. pp..T...T. Lehnigk.T. Stability Theorems for Linear Motions with an Introduction to Liapunov’s Direct Method . Gryazina.. no. Structural Singular Value and Stability of Uncertain Polynomials. A New Approach to Digital PID Design.F. Philadelphia: SIAM.. H. pp. New York: McGrawHill.. 64. 1994. A. 12 2008 . Ho. and Polyak. Royal Soc.. Automat. p. London: Springer.. and Nett. and Bhattacharyya. and Bhattacharyya. Automatica. 56. 687–692.. Syst. D. Prague. A.. 43. pp.. 44. vol. 3.. A Mixed Probabilistic/Deterministic Approach to Fixed Order H∞ Controller Design. pp. Bahamas. 60. 1069–1075.P.. Y. Ho. 1997. vol. 3. Ser. and Tempo. C..V... A Missing Link. J. Keel. pp.. 642–654..K. L. 431–438. 50. vol. 48.. no. 2003. J. pp. Lin.
L. 318–321. AC11. vol. vol. 78. K. CACSD Conf. AIEE Trans. vol.ch/∼joloef/yalmip.. Fast Calculation of Stabilizing PID Controllers. E. 2000.B. 14–28. 11th IFAC Sympos. 4. 86. 85. 10. M. Philadelphia: SIAM. 71.P. 84.. J. 83. A Stabilization Algorithm for a Class of Uncertain Systems. Seoul. a member of the Editorial Board AUTOMATION AND REMOTE CONTROL Vol. 70. Sampleddata Systems.. vol.. InteriorPoint Polynomial Algorithms in Convex Programming. SIAM J. HitandRun from the Corner. III.2026 GRYAZINA et al. 78. Sur la theorie generale des regulateurs. Berlin: Springer. and Ozguler. 1969. Control Conf. Yalmip: A Toolbox for Modeling and Optimization in MATLAB. 77. in Mathemat. D. pp.. Eur.J. 121–126. 6. A. Saeki. Saadaoui. pp. J. 2003. 75. Siljak. 1876. and Industry). Mitrovic. Automat. Upper Saddle River: Prentice Hall. Mapping of H∞ Design Speciﬁcations into Parameter Space...T.. 39. Siljak. Mathem. 77. 1959...T. Munro. and Vempala. Automatica. and Industry). no. Yu.. Lov´sz. Analysis and Synthesis of Feedback Control Systems in the Parameter Plane. M. no. vol. 1994. Rantzer. Lofberg. K. Prague. 449–466. 443–461. Control .. IEEE Trans. 72... Syst. Describing the H∞ set in the Controller Parameters Space. 249–260. 2004 . and Blue.. B.. P. vol. and Gryazina.. Program. no. Theory and Analysis.ethz. and Nemirovskii. and Doyle. pp. Korea. A. Robust and Optimal Control . and Kocan.. 1995. Graphical Analysis and Synthesis of Feedback Control Systems. Oper. Proc. Ozguler. 476–496. vol. a 66. and Byrnes. Fixed Structure PID Controller Design for Standard H∞ Control Problem. 2005. Kos. 74. Lov´sz.ee. J. and Aimoto. Qiu. 68. 86. vol.. pp.. (Appl. 1296–1308. Taipei..C. (Appl. 1984. Polyak. Petersen.php. Proc. Robust Nonlinear Control. I. a pp. E. 8. Proc. Control Lett.. L. Proc. 17th IFAC World Congress.M. Sampleddata Feedback Control Systems. Comput. pp. 1964. J. 2007. 2000. D. 67. Davison. N. An Analytic Determination of Stabilizing Feedback Gains. A Formula for Computation of the Real Stability Radius. Nonlinear Systems: The Parameter Analysis and Design. PID Controller Optimization for Hinﬁnity Control by Linear Programming. vol. C. pp. Nesterov. Prague. A. Kurdyukov. 1. 65. Automatica. Polyak B. 69 No. 82. A. pp... Robust Linear Algebra and Robust Aperiodicity. Vishnegradsky. This paper was recommended for publication by A. Syst. D. Theory and Optim. 79. Rantzer. II. pp. no. in Direct. Int. URL http://control. Int.. Report Institut fur Dynamische Systeme. Synthesis.. Saeki. HitandRun: New Design Technique for Stabilization.. Zhou. 1999. 1966.. 1987.. I. D. 816–823. and Glover. A New Method for the Computation of All Stabilizing Controllers of a Given Order. Res. II.. 2008. 31..C. M. 32. HitandRun Mixes Fast. New York: Wiley. 83. Sci. AIEE Trans.. 1.. Smith. and Baki. A. Eﬃcient Monte Carlo Procedures for Generating Points Uniformly Distributed over Bounded Regions. K. Acad. Universitat Bremen.. Spelta. Generalization of the Parameter Plane Method. pp. D... I. 2006. H. 1994. 35. 81. 985–1005. Taiwan. 879–890. 63–70. P. Doyle. pp. Young. 83–99. B. vol. E. 1996.. and Gryazina. 83. 12 2008 . vol.I.L. J. A. Linear Continuous Systems. C. pp. Robustness and Optimization of Linear Systems. 73. Comput. 351–357. Siljak. Eds.A. no.. pp.. Proc. Rend.. K. 2003. R. Robust Control Design. 2005. Greece. I.B. Soylemez. pp. S. Control . 6. 69. L. 76. Odenthal. Bernhardsson. 80. 16 IFAC World Congr.
This action might not be possible to undo. Are you sure you want to continue?
Use one of your book credits to continue reading from where you left off, or restart the preview.